content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
\label{sec:intro}
One of the central tasks in statistics is to model and sample from a multi-dimensional probability distribution. Classic statistics approaches this problem by fitting a model to the target distribution and then sampling from a fitted model via Markov Chain Monte Carlo (MCMC) techniques. Although such model-based methods are widely used, MCMC sampling often entails several technicalities. Beyond diagnosing whether the chain mixes, obtaining i.i.d.\ samples from MCMC methods is complex as one has to control correlations between successive samples or run parallel chains.
An alternative approach available in statistics, reserved for the one-dimensional case, is usually referred to as the (inverse) \textit{transform sampling}. Such an approach circumvents the calling for a parametric or nonparametric density and directly designs a sampler by transforming a simple uniform distribution. The idea is simple: one can transform a uniform measure $\mu = {\rm Unif}([0,1])$ to any one-dimensional target probability measure $\nu$ leveraging the following monotonic transformation $T \colon [0, 1] \rightarrow \R$ called the inverse Cumulative Distribution Function (CDF),
\begin{align}
\label{eqn:inv-cdf}
T(x)= \inf \{ y \in \R ~:~ \nu((-\infty, y]) \ge x \} \;.
\end{align}
Define the pushforward measure $T_\#\mu$ by $T_\#\mu(S) = \mu(\{ x : T(x) \in S \})$ for any Borel set $S\subseteq \R$, then one can easily check that $T_{\#} \mu = \nu$; namely, with a draw from the one-dimensional uniform distribution $x\sim \mu$, the transformed sample $T(x)$ has the target probability distribution $\nu$.
Recently, the \textit{transform sampling} idea has been extended to the multi-dimensional setting, as seen in both machine learning (specifically, generative modeling) and computational optimal transport. Again, given a target probability measure $\nu$ supported on $\cY$ and a user-specified probability measure $\mu$---that is easy to sample from such as a multivariate Gaussian---defined on $\cX$, we aim to find a measurable map $T \colon \cX \to \cY$ such that $T_{\#} \mu = \nu$, where $T_{\#} \mu$, the pushforward measure, is defined analogously to the one-dimensional case above. Such a map $T$, which is called a transport map from $\mu$ to $\nu$, transforms i.i.d.\ samples from $\mu$ into i.i.d.\ samples from $\nu$. Therefore, with a good estimate of the transformation $T$, the transform sampler operates and scales more efficiently than classic MCMC approaches.
Such transform sampling ideas have been leveraged in generative modeling by designing different criteria to learn a qualified transformation $T$; furthermore, remarkable empirical benchmarks have been documented. The essence of these methods can be summarized as follows. A transport map is obtained by minimizing $T \mapsto \cL(T_{\#} \widehat{\mu}, \widehat{\nu})$ over $\cF$, where $\cF$ is a map class that is rich enough to contain a transport map, $\widehat{\mu}$ and $\widehat{\nu}$ are empirical measures based on samples from $\mu$ and $\nu$, respectively, and $\cL$ measures certain discrepancy of two distributions. In summary, by properly designing a class of maps $\cF$ and collecting sufficiently many samples, we expect a minimizer $T$ that will satisfy $T_{\#} \mu \approx \nu$. In Generative Adversarial Networks (GAN) \citep{goodfellow_2014}, $\cF$ consists of neural networks, and $\cL$ is Jensen-Shannon divergence. Moreover, different choices of $\cL$ have led to several variants: $f$-divergences for $f$-GAN \citep{nowozin_2016}, Wasserstein distances for Wasserstein-GAN \citep{arjovsky_2017}, and Maximum Mean Discrepancies (MMD) for MMD-GAN \citep{dziugaite_roy_ghahramani_2015, li_swersky_zemel_2015}.
The Optimal Transport (OT) theory aims to identify the optimal transformation $T$, quantified by the transportation cost of moving mass from $\mu$ to $\nu$. When $\mu$ and $\nu$ both lie in the same space, say $\R^d$, Brenier \citep{brenier1991} proved, under mild regularity conditions, the following remarkable result that backs up the \textit{transform sampling} in the multi-dimensional setting. Consider the Wasserstein-$p$ distance $W_p(\mu, \nu)$ defined as
\begin{align*}
W_p(\mu, \nu) \coloneqq \inf_{\gamma \in \Pi(\mu, \nu)} \left( \int_{\R^d \times \R^d} \| x - y \|_2^p \dd{\gamma}(x, y) \right)^{1/p} \;,
\end{align*}
where $\Pi(\mu, \nu)$ denotes the set of all couplings between $\mu$ and $\nu$. Brenier established that for $p = 2$, there exists a unique optimal coupling $\gamma^{\star}$ and a unique optimal transport map $T^{\star}$ such that $\gamma^{\star} = (\mathrm{Id}, T^{\star})_{\#} \mu$; more importantly, $T^{\star}$ is the gradient of some convex function. As a result,
\begin{align*}
W_2^2(\mu, \nu) = \int_{\R^d \times \R^d} \|x - y\|_2^2 \dd{\gamma^{\star}}(x, y) = \int_{\R^d} \|x - T^\star(x)\|_2^2 \dd{\mu}(x) \;.
\end{align*}
Now let's contrast this result with the one-dimensional (inverse) transform sampling: when $\mu = {\rm Unif}([0,1])$, it turns out the inverse CDF map $T \colon [0, 1] \rightarrow \R$ in \eqref{eqn:inv-cdf} minimizes the transportation cost $W_p(\mu, \nu), p\geq 1$. Brenier's result significantly enriches the one-dimensional insight to the multi-dimensional case: now the multi-dimensional map $T \colon \R^d \rightarrow \R^d$ is the gradient of a convex function, as opposed to a monotonic map from $\R$ to $\R$.
Such an OT-based approach, however, can be inefficient in practice if the target $\nu$ is a high-dimensional embedding of some low-dimensional distribution. For instance, let $\nu$ be the distribution of handwritten digit images from the MNIST data set defined on $\R^{28 \times 28} \equiv \R^{784}$.\footnote{Each image is properly normalized and fit into a $28 \times 28$ pixel bounding box \citep{lecun1998gradient}.} To use the above OT-based approach, one must choose $\mu$ on $\R^{784}$ and find the map $T \colon \R^{784} \to \R^{784}$ discussed above. However, the support of $\nu$ is intrinsically low-dimensional (roughly $\R^{15}$ as in \citep{facco2017estimating}), hence other transform samplers with $\cX = \R^{15}$ yielding $T \colon \R^{15} \to \R^{784}$ are more efficient than the OT-based method in terms of estimating $T$ and computing $T(X)$ for $X \sim \mu$.
Motivated by this limitation, we propose a transform sampler having the best of both worlds: it is underpinned by OT theory, at the same time, operates when $\cX$ and $\cY$ are heterogeneous spaces. The key to our approach is to utilize the Gromov-Wasserstein (GW) distance between $\mu$ and $\nu$, well-defined for heterogeneous spaces $\cX$ and $\cY$. Given two continuous functions $c_{\cX} \colon \cX \times \cX \rightarrow \R$ and $c_{\cY} \colon \cY \times \cY \rightarrow \R$, the GW distance \citep{memoli_2011,chowdhury_memoli_2019} is
\begin{align}
\mathrm{GW}(\mu, \nu) \coloneqq \inf_{\gamma \in \Pi(\mu, \nu)} \left( \int_{\cX \times \cY} \int_{\cX \times \cY} \big( c_{\cX}(x, x') - c_{\cY}(y, y') \big)^2 \dd{\gamma} (x, y) \dd{\gamma}(x', y') \right)^{1/2} \;.
\end{align}
A few remarks regarding the comparison between Wasserstein and Gromov-Wasserstein are as follows. First, unlike Wasserstein which solves an infinite-dimensional linear program in the coupling $\gamma$, GW formulates a Quadratic Program (QP) in $\gamma$, which is known to be computationally hard \citep{cela1998QuadraticAssignment,LOIOLA2007657}. Second, GW aims to match the cost functions defined on two heterogeneous spaces, intending to identify isomorphism between spaces. Despite being an elegant notion of distance between metric measure spaces \citep[Definition~5.1]{memoli_2011}, GW is hard to compute in practice due to its QP nature; it is also unclear how to estimate ${\rm GW}(\mu, \nu)$ based on finite i.i.d.\ samples from $\mu$ and $\nu$, and how accurate such estimates are.
\paragraph{Main contributions}
This paper considers computational and statistical questions regarding Gromov-Wasserstein outlined above, and aims to design a new transform sampler as an approach to model and sample from multi-dimensional probability distributions given access to i.i.d.\ samples, circumventing the usual ways of modeling the density function or MCMC. Our transform sampler can also estimate good alignments between two heterogeneous metric measure spaces $(\cX, \mu, c_{\cX})$ and $(\cY, \nu, c_{\cY})$ from empirical data sets, with estimated maps that approximately pushforward one measure $\mu$ to the other $\nu$, and vice versa. Towards reaching these goals, we made the following specific contributions.
\begin{itemize}
\item We introduce a new notion, Reversible Gromov-Monge (RGM) distance, on metric measure spaces that majorizes the usual Gromov-Wasserstein distance. Furthermore, we show several analytic properties possessed by GW naturally carry over to RGM.
\item Our RGM formulation induces a transform sampler, as a relaxation of the usual GW formulation. Rather than solving a QP which is quadratic in the coupling $\gamma \in \Pi(\mu, \nu)$, we decouple the pair as $(\mathrm{Id}, F)_{\#} \mu$ and $(B, \mathrm{Id})_{\#} \nu$ with $F \colon \cX \rightarrow \cY$ and $B \colon \cY \rightarrow \cX$, respectively, and then bind them later via the constraint $(\mathrm{Id}, F)_{\#} \mu \approx (B, \mathrm{Id})_{\#} \nu$. Such a decoupling and binding idea will prove suitable for the statistical estimation problem based on finite i.i.d.\ samples. We will also show, from an operator viewpoint, such a decoupling and binding idea ensures that our RGM is an infinite-dimensional convex program in $F, B$ that admits a simple representation theorem, as opposed to the otherwise intractable infinite-dimensional QP in GW.
\item We derive non-asymptotic rates of convergence for the proposed RGM sampler using tools from empirical processes, for generic classes modeling the measurable maps $F$ and $B$. Based on our non-asymptotic results, concrete upper bounds can be easily spelled out in the cases where $F$ and $B$ are parametrized by deep neural networks. As mentioned earlier, the RGM sampler also promises to identify good alignments between metric measure spaces, and to learn approximate isomorphism when possible. We demonstrate such a point using numerical experiments on MNIST.
\end{itemize}
\paragraph{Organization}
The rest of the paper is organized as follows. First, we briefly review other related studies omitted in the discussion above. Then, in Section~\ref{sec:background}, preliminary background on optimal transport and Gromov-Wasserstein distance is outlined. Next, Section~\ref{sec:summary-of-results} summarizes the primary methodology and theory regarding our proposed Reversible Gromov-Monge sampler. Synthetic and real-world examples showcasing the effectiveness of the RGM sampler are demonstrated in Section~\ref{sec:numerical} as a proof of concept. The supplementary material collects details of the results in Sections \ref{sec:summary-of-results} and \ref{sec:numerical} along with extensive discussions.
\subsection{Related Literature}
Inferring the underlying probability distributions from data has been a central problem in statistics and the unsupervised machine learning since the invention of histograms by Pearson a century ago. Classic mathematical statistics explicitly models the density function in a parametric or a nonparametric way \citep{silverman_1986, wasserman_2006}, and studies the minimax optimality of directly estimating such density functions \citep{stone_1982}. It is also unclear how to proceed to sample from a possibly improper\footnote{Here we mean that the estimated density is not always non-negative and integrates to one.} density estimator, even with an optimal estimator at hand. One may employ Markov Chain Monte Carlo (MCMC) techniques for sampling from specific models. However, on the computational front, it is highly non-trivial how to ensure the mixing properties of MCMC for a designed sampler \citep[Chapter~7]{robert_casella_2004}.
Recent work in unsupervised machine learning proposes to learn complex, high-dimensional distributions via (deep) generative models, either explicitly by parametrizing the sufficient statistics of the exponential families \citep{doersch2016TutorialVariationalAutoencoders,kingma2013auto}, or implicitly by parametrizing the pushforward map transporting distributions \citep{dziugaite_roy_ghahramani_2015,goodfellow_2014}, with a focus on tractability in computation. Surprisingly, though lacking theoretical underpinning and optimality, the generative models' approach performs well empirically in large-scale applications where classical statistical procedures are destined to fail. There has been a growing literature on understanding distribution estimation with the implicit framework, with more general metrics and target distribution classes, to name a few, \cite{muandet_fukumizu_sriperumbudur_scholkopf_2017, li_swersky_zemel_2015, dziugaite_roy_ghahramani_2015} on MMDs, \cite{sriperumbudur2012empirical, liang2019EstimatingCertain} on integral probability metrics, and \cite{mroueh2017sobolev, arora2017generalization,liang2018HowWell,singh2018minimax,bai2018approximability, weed2019estimation, lei2019sgd, chen2020statistical} on generative adversarial networks. Last but not least, we emphasize that an alternative implicit distribution estimation approach using the simulated method of moments has been formulated in the econometrics literature since \cite{mcfadden1989MethodSimulated,pakes1989SimulationAsymptotics} and \cite{gourieroux1997SimulationbasedEconometric}.
Originally introduced as a tool for comparing objects in computer graphics, analytic properties of the Gromov-Wasserstein distance have been studied extensively \citep{memoli_2011,sturm_2012}; the most important one is that it defines a distance between metric measure spaces, namely, metric spaces endowed with probability measures. Since many real-world data sets can be modeled as metric measure spaces, the GW distance has been utilized in various problems such as shape correspondence \citep{solomon_etal_2016}, graph matching \citep{xu2019scalable}, and protein comparison \citep{gellert2019substrate}. Certain statistical aspects of comparing metric measure spaces have been studied in \cite{brecheteau2019statistical,weitkamp2020gromov}.
Computation of the GW distance amounts to a relaxation of the quadratic assignment problem \citep{koopmans1957AssignmentProblems}; both are known to be NP-hard \citep{cela1998QuadraticAssignment,LOIOLA2007657} in the worst case. Several approaches have been proposed for the approximate computation of the GW distance. \cite{memoli_2011} studies lower bounds on the GW distance that are easier to compute. \cite{peyre_etal_2016} adds an entropic regularization term to the GW distance, which leads to a fast iterative algorithm; \cite{scetbon2021linear} further modifies this by imposing a low-rank constraint on couplings. \cite{titouan_etal_2019} proposes the Sliced Gromov-Wasserstein distance defined by integrating GW distances over one-dimensional projections. Last but not least, recent papers \citep{xu2019scalable,blumberg2020mrec,chowdhury2021quantized} study scalable partitioning schemes to approximately compute GW distances.
\section{Background}
\label{sec:background}
In this section, we provide background on the Optimal Transport (OT) theory and the Gromov-Wasserstein distance. First, we start with some notations. Let $\|A\|$ denote the Frobenius norm of a matrix $A$ and $\|x\|$ denote the Euclidean norm of a vector $x$. Given a set $\cX$ and a function $f \colon \cX \to \R$, let $\|f\|_\infty = \sup_{x \in \cX} |f(x)|$ denote the sup norm. For an integer $n \in \N$, we define $[n] = \{1, \ldots, n\}$. For a metric space $\cX$, we denote its metric as $d_{\cX}$ and write $\cP(\cX)$ to denote the collection of all Borel probability measures on $\cX$; we call $\cX$ a Polish space if it is complete and separable. We call a pair $(\cX, \mu)$ a Polish probability space if $\cX$ is a Polish space and $\mu \in \cP(\cX)$. Given two Polish probability spaces $(\cX, \mu)$ and $(\cY, \nu)$, the collection of all transport maps from $\mu$ to $\nu$ is denoted as $\cT(\mu, \nu) \coloneqq \{ T \colon \cX \rightarrow \cY ~|~ T_\# \mu = \nu \}$; we call $\gamma \in \cP(\cX \times \cY)$ a coupling between $\mu$ and $\nu$ if $\gamma(A \times \cY) = \mu(A)$ and $\gamma(\cX \times B) = \nu(B)$ for all Borel subsets $A \subset \cX$ and $B \subset \cY$, and we denote the collection of all such couplings as $\Pi(\mu, \nu)$. For a sequence of numbers $a(n), b(n) \in \R$, we use $a(n)\precsim b(n)$ to denote the relationship that $a(n)/b(n)\leq C, \forall n$ with some universal constant $C>0$.
\subsection{A Brief Overview of Optimal Transport Theory}
A major goal of OT is minimizing the cost associated with the transport map between two Polish probability spaces, say $(\cX, \mu)$ and $(\cY, \nu)$. Consider a measurable function $c \colon \cX \times \cY \to \R_{+}$; we view $c(x, y)$ as the cost associated with $x \in \cX$ and $y \in \cY$. For each transport map $T \in \cT(\mu, \nu)$, we interpret $c(x, T(x))$ as a unit cost incurred by mapping each $x \in \cX$ to $T(x) \in \cY$. We define the average cost incurred by the transport map $T$ as the integration of all the unit costs with respect to $\mu$:
\begin{equation*}
\int_{\cX} c(x, T(x)) \dd{\mu}(x) \;.
\end{equation*}
Minimizing the cost over $\cT(\mu, \nu)$ is referred to as the Monge problem named after Gaspard Monge. If there exists a minimizer $T^\star$ to this problem, that is,
\begin{equation*}
T^\star \in \argmin_{T \in \cT(\mu, \nu)} \int_{\cX} c(x, T(x)) \dd{\mu}(x)\;,
\end{equation*}
we call $T^\star$ an optimal transport map.
Another important OT problem is minimizing the cost given by couplings. We define the average cost incurred by a coupling $\gamma \in \Pi(\mu, \nu)$ as the integration of the cost $c(x, y)$ with respect to $\gamma$:
\begin{equation*}
\int_{\cX \times \cY} c(x, y) \dd{\gamma}(x, y) \;.
\end{equation*}
Minimizing this cost over $\Pi(\mu, \nu)$ is called the Kantorovich problem credited to Leonid Kantorovich. If
\begin{equation*}
\gamma^\star \in \argmin_{\gamma \in \Pi(\mu, \nu)} \int_{\cX \times \cY} c(x, y) \dd{\gamma}(x, y) \;,
\end{equation*}
we call $\gamma^\star$ an optimal coupling.
The two OT problems are closely related: the Kantorovich problem is a relaxation of the Monge problem. To see this, for each $T \in \cT(\mu, \nu)$, define a map $(\mathrm{Id}, T) \colon \cX \to \cX \times \cY$ by $(\mathrm{Id}, T)(x) = (x, T(x))$. One can verify $(\mathrm{Id}, T)_{\#} \mu \in \Pi(\mu, \nu)$. Therefore, if we define $\Pi_{\cT} \coloneqq \{(\mathrm{Id}, T)_{\#} \mu : T \in \cT(\mu, \nu)\}$, then $\Pi_{\cT} \subset \Pi(\mu, \nu)$ and thus
\begin{equation*}
\inf_{T \in \cT(\mu, \nu)} \int_{\cX} c(x, T(x)) \dd{\mu}(x)
= \inf_{\gamma \in \Pi_{\cT}} \int_{\cX \times \cY} c(x, y) \dd{\gamma}(x, y)
\ge \inf_{\gamma \in \Pi(\mu, \nu)} \int_{\cX \times \cY} c(x, y) \dd{\gamma}(x, y)\;,
\end{equation*}
where the first equality follows from change-of-variables. In other words, two OT problems share the same objective function as a function of couplings; however, the Kantorovich problem has a larger constraint set.
Unlike the Monge problem, the Kantorovich problem has favorable properties. First, the objective function is linear in $\gamma$. Moreover, $\Pi(\mu, \nu)$ is compact in the weak topology of Borel probability measures defined on $\cX \times \cY$. This suggests that we can view the Kantorovich problem as an infinite-dimensional linear program.
Besides seeking optimal transport maps or couplings, another interesting aspect of OT problems is that the least possible cost can endow a metric structure among Polish probability spaces. If $\cX = \cY$ and $c = d_{\cX}^2$, the square root of the solution of the Kantorovich problem defines a distance between $\mu$ and $\nu$, known as the Wasserstein distance.
\begin{definition}
Given a Polish space $\cX$, the Wasserstein-2 distance between $\mu, \nu \in \cP(\cX)$ is defined as
\begin{equation*}
W_2(\mu, \nu) = \inf_{\gamma \in \Pi(\mu, \nu)} \left(\int_{\cX \times \cX} d_{\cX}^2(x, y) \dd{\gamma}(x, y)\right)^{1/2}\;.
\end{equation*}
\end{definition}
\begin{remark}
One can define the Wasserstein-$p$ distance by replacing the exponent $2$ above with $p \in [1, \infty]$. The Wasserstein-$p$ distance is known to satisfy the usual metric axioms.
\end{remark}
\subsection{Gromov-Wasserstein and Gromov-Monge Distances}
Although OT problems can be defined between arbitrary Polish probability spaces, in practice, it is unclear how to design a function $c \colon \cX \times \cY \to \R_+$ to represent meaningful cost associated with $x \in \cX$ and $y \in \cY$ in two heterogeneous spaces.
For instance, if $\cX = \R^p$ and $\cY = \R^q$ with $p \neq q$, there is no simple choice for a cost function $c$ over $\R^p \times \R^q$. As a result, classic OT theory (including Brenier's result) cannot be directly used for comparing heterogeneous Polish probability spaces.
M{\'e}moli's pioneering work \cite{memoli_2011} resolved this issue by considering a quadratic objective function of $\gamma$:
\begin{equation*}
\int_{\cX \times \cY} c(x, y) \dd{\gamma}(x, y) \Rightarrow \int_{\cX \times \cY}\int_{\cX \times \cY} (c_{\cX}(x, x') - c_{\cY}(y, y'))^2 \dd{\gamma}(x, y) \dd{\gamma}(x', y') \;,
\end{equation*}
where $c_{\cX}$ and $c_{\cY}$ are defined over $\cX \times \cX$ and $\cY \times \cY$, respectively. For instance, one can specify $c_{\cX} = d_{\cX}$ and $c_{\cY} = d_{\cY}$. Rather than considering a unit cost corresponding to each pair $(x, y) \in \cX \times \cY$, two pairs $(x, y)$ and $(x', y')$ in $\cX \times \cY$ are associated with the discrepancy of intra-space quantities $c_{\cX}(x, x')$ and $c_{\cY}(y, y')$. In summary, by switching from the integration $\dd{\gamma}$ to the double integration $\dd{\gamma} \dd{\gamma}$, we no longer need an otherwise inter-space quantity $c \colon \cX \times \cY \to \R_+$. Therefore, we can always define this objective function whenever we have proper $c_{\cX}$ and $c_{\cY}$ in each individual space, which leads to the following definition.
\begin{definition}
\label{def:mms}
A triple $(\cX, \mu, c_{\cX})$ is called a network space if $(\cX, \mu)$ is a Polish probability space such that $\mathrm{supp}(\mu) = \cX$ and $c_{\cX} \colon \cX \times \cX \to \R$ is continuous. The Gromov-Wasserstein distance between network spaces $(\cX, \mu, c_{\cX})$ and $(\cY, \nu, c_{\cY})$ is defined as
\begin{equation*}
\mathrm{GW}(\mu, \nu) = \inf_{\gamma \in \Pi(\mu, \nu)} \left( \int_{\cX \times \cY} \int_{\cX \times \cY} (c_{\cX}(x, x') - c_{\cY}(y, y'))^2 \dd{\gamma}(x, y) \dd{\gamma}(x', y') \right)^{1/2}\;.
\end{equation*}
\end{definition}
\begin{remark} \rm
On top of the network space definition introduced in \cite{chowdhury_memoli_2019}, we impose continuity of $c_{\cX}$ for a cleaner analysis. A network space $(\cX, \mu, c_{\cX})$ is called a metric measure space if $c_{\cX} = d_{\cX}$ as introduced in \cite{memoli_2011} and \cite{sturm_2012}. In short, a network space is a generalization of a metric measure space.
\end{remark}
Like the Wasserstein distance, the GW distance has metric properties; it satisfies symmetry and the triangle inequality, and $\mathrm{GW}(\mu, \nu) = 0$ if $(\cX, \mu, c_{\cX}) = (\cY, \nu, c_{\cY})$. However, the converse of this last statement does not hold in general: for its validity, a suitable equivalence relation needs to be defined on the collection of network spaces.
\begin{definition}
\label{def:iso}
Network spaces $(\cX, \mu, c_{\cX})$ and $(\cY, \nu, c_{\cY})$ are strongly isomorphic if there exists $T \in \cT(\mu, \nu)$ such that $T \colon \cX \to \cY$ is bijective and $c_{\cX}(x, x') = c_{\cY}(T(x), T(x'))$ for all $x, x' \in \cX$. In this case, we write $(\cX, \mu, c_{\cX}) \cong (\cY, \nu, c_{\cY})$ and such a transport map $T$ is called a strong isomorphism.
\end{definition}
One can easily check that $\cong$ is indeed an equivalence relation on the collection of network spaces. The following theorem states that the GW distance satisfies all metric axioms on the quotient space---under the equivalence relation $\cong$---of metric measure spaces.
\begin{theorem}[Lemma 1.10 of \cite{sturm_2012}]
\label{thm:1}
Let $\cM$ be the collection of all network spaces $(\cX, \mu, c_{\cX})$ such that $c_{\cX} = d_{\cX}$. Also, let $\cM/_{\cong}$ be the collection of all equivalence classes of $\cM$ induced by $\cong$. Then, GW satisfies the three metric axioms on $\cM/_{\cong}$.
\end{theorem}
Recall that the Monge problem is a restricted version of the Kantorovich problem with an additional constraint that couplings are given by a transport map; replacing $\Pi(\mu, \nu)$ in the Kantorovich problem with $\Pi_{\cT}$ yields the Monge problem. Imposing the same constraint on the definition of GW leads to the Gromov-Monge distance.
\begin{definition}
The Gromov-Monge distance between network spaces $(\cX, \mu, c_{\cX})$ and $(\cY, \nu, c_{\cY})$ is defined as
\begin{equation*}
\mathrm{GM}(\mu, \nu) = \inf_{T \in \cT(\mu, \nu)} \left( \int_{\cX} \int_{\cX} (c_{\cX}(x, x') - c_{\cY}(T(x), T(x')))^2 \dd{\mu}(x) \dd{\mu}(x') \right)^{1/2} \;.
\end{equation*}
\end{definition}
Loosely speaking, computing GM amounts to finding a transport map $T$ such that $c_{\cX}(x, x')$ best matches $c_{\cY}(T(x), T(x'))$ on average; we can view such a map $T$ as a surrogate for an isomorphism. See Section 3 of \cite{memoli2018distance} for more details of GM.
\section{Summary of Results}
\label{sec:summary-of-results}
Inspired by the Gromov-Wasserstein and Gromov-Monge distances, we propose a new metric---the reversible Gromov-Monge distance---between network spaces in this paper. Our formulation seeks a pair of transport maps $F \in \cT(\mu, \nu)$ and $B \in \cT(\nu, \mu)$ best approximating isomorphic relations between network spaces. We propose a novel transform sampling method that uses $F$ as a push-forward map to obtain i.i.d.\ samples from a target distribution $\nu$. We present two optimization formulations solving for such a pair $(F, B)$ in order: a potentially non-convex formulation that employs the standard gradient descent method to optimize, and an infinite-dimensional convex formulation where global optima can be found efficiently. For the former, we analyze the statistical rate of convergence for generic classes $\cF \times \cB$ parametrizing $(F, B)$. For the latter, we derive a new representer theorem on a suitable reproducing kernel Hilbert space (RKHS).
\subsection{Metric Properties of Reversible Gromov-Monge}
Our formulation is based on the following observation: for a coupling $\gamma$ such that $\gamma = (\mathrm{Id}, F)_{\#}\mu = (B, \mathrm{Id})_{\#} \nu$, which presents a binding constraint, we can simplify the objective function of GW as
\begin{equation*}
\int_{\cX \times \cY} (c_{\cX}(x, B(y)) - c_{\cY}(F(x), y))^2 \dd{\mu \otimes \nu} \;,
\end{equation*}
where $\dd{\mu \otimes \nu} \coloneqq \dd{\mu}(x) \dd{\nu}(y)$ denotes the product measure of $\mu$ and $\nu$.
Imposing the binding constraint on the definition of GW leads to the following definition.
\begin{definition}
For network spaces $(\cX, \mu, c_{\cX})$ and $(\cY, \nu, c_{\cY})$, we write $(F, B) \in \cI(\mu, \nu)$ if measurable maps $F \colon \cX \to \cY$ and $B \colon \cY \to \cX$ satisfy the binding constraint $(\mathrm{Id}, F)_{\#}\mu = (B, \mathrm{Id})_{\#} \nu$. We define the reversible Gromov-Monge (RGM) distance between $(\cX, \mu, c_{\cX})$ and $(\cY, \nu, c_{\cY})$ as
\begin{equation}
\label{eq:RGM}
\mathrm{RGM}(\mu, \nu) \coloneqq \inf_{(F, B) \in \cI(\mu, \nu)} \left(\int_{\cX \times \cY} (c_{\cX}(x, B(y)) - c_{\cY}(F(x), y))^2 \dd{\mu \otimes \nu}\right)^{1/2}\;.
\end{equation}
\end{definition}
\begin{remark}
\rm
A few remarks are in place for the binding constraint.
If $(\mathrm{Id}, F)_{\#}\mu = (B, \mathrm{Id})_{\#} \nu$, then $F_{\#} \mu = \nu$ and $B_{\#} \nu = \mu$ follow due to marginal conditions. However, the converse is not true in general. To see this, let $\mu = \nu = \mathrm{Unif}([0, 1])$, then $F_{\#} \mu = \nu$ and $B_{\#} \nu = \mu$ hold for $F(x) = B(x) = |2 x - 1|$. However, $(\mathrm{Id}, F)_{\#}\mu \neq (B, \mathrm{Id})_{\#} \nu$ because $(\mathrm{Id}, F)_{\#}\mu$ is a uniform measure on $\{(x, |2x-1|): x \in [0, 1]\}$, whereas $(B, \mathrm{Id})_{\#} \nu$ is a uniform measure on $\{(|2y-1|, y): y \in [0, 1]\}$.
\end{remark}
Roughly speaking, computing RGM consists in finding a pair $(F, B) \in \cI(\mu, \nu)$ such that $c_{\cX}(x, B(y))$ best matches $c_{\cY}(F(x), y)$ on average. Like a strong isomorphism, we can view such a pair as jointly capturing an isomorphic relation of $(\cX, \mu, c_{\cX})$ and $(\cY, \nu, c_{\cY})$. We will use this observation later to build a transform sampling method.
We will prove that RGM possesses metric properties similar to the Gromov-Wasserstein. Motivated by Theorem \ref{thm:1}, we derive the following result.
\begin{theorem}
\label{thm:metric}
Let $h \colon \R_+ \to \R$ be a continuous and strictly monotone function and $\cN^{h}$ be a collection of all network spaces $(\cX, \mu, c_{\cX})$ such that $c_{\cX} = h(d_{\cX})$. Then RGM satisfies the three metric axioms on $\cN^h/_{\cong}$ which is the collection of all equivalence classes of $\cN^h$ induced by $\cong$.
\end{theorem}
\begin{remark}
\rm
Suppose $\cX$ is a Euclidean space and $d_{\cX}$ is the standard Euclidean distance. If $h(x) = \exp(- \alpha x^2)$ with $\alpha > 0$, then $h(d_{\cX})$ is the radial basis function (RBF) kernel on $\cX$; we will use this in numerical experiments.
\end{remark}
We refer the proof of Theorem \ref{thm:metric} and details of the properties of RGM to Section \ref{sec:analytic-properties}.
\subsection{Transform Sampling via RGM}
\label{sec:RGM-sampler}
With the proposed notion of RGM, we design a transform sampling method in this section. The transform sampler is based on finding a minimizing pair $(F, B)$ of RGM, which can capture isomorphic relations between network spaces. To implement this method, we need to estimate $(F, B)$ using only i.i.d.\ samples from $\mu$ and $\nu$. Leveraging the Lagrangian form, we derive a minimization problem that can be implemented based on finite samples.
First, we rewrite the population minimization problem with the binding constraint as follows,
\begin{align*}
\min_{\substack{F \colon \cX \to \cY \\ B \colon \cY \to \cX}} \quad & \int_{\cX \times \cY} (c_{\cX}(x, B(y)) - c_{\cY}(F(x), y))^2 \dd{\mu \otimes \nu}\nonumber \\
\mathrm{s.t.} \quad & \cL_{\cX \times \cY}((\mathrm{Id}, F)_{\#} \mu, (B, \mathrm{Id})_{\#} \nu) = 0\;.
\end{align*}
Here, $\cL_{\cX \times \cY}$ is a suitable discrepancy measure on $\cP(\cX \times \cY)$ so that $\cL_{\cX \times \cY}((\mathrm{Id}, F)_{\#} \mu, (B, \mathrm{Id})_{\#} \nu) = 0$ is a surrogate for the original constraint $(\mathrm{Id}, F)_{\#}\mu = (B, \mathrm{Id})_{\#} \nu$. In practice, we do not require that $\cL_{\cX \times \cY} = 0$ implies $(\mathrm{Id}, F)_{\#}\mu = (B, \mathrm{Id})_{\#} \nu$; in fact, the former constraint can be a relaxation of the latter. The choice of $\cL_{\cX \times \cY}$ will be specified later. To solve this minimization problem, we propose utilizing the Lagrangian:
\begin{equation*}
\min_{\substack{F \colon \cX \to \cY \\ B \colon \cY \to \cX}} \int_{\cX \times \cY} (c_{\cX}(x, B(y)) - c_{\cY}(F(x), y))^2 \dd{\mu \otimes \nu} + \lambda \cdot \cL_{\cX \times \cY}((\mathrm{Id}, F)_{\#} \mu, (B, \mathrm{Id})_{\#} \nu)\;.
\end{equation*}
Given i.i.d.\ samples $\{x_i\}_{i=1}^{m}$ and $\{y_j\}_{j=1}^{n}$ from $\mu$ and $\nu$, respectively, we replace the population objective with its empirical estimates:
\begin{equation*}
\min_{\substack{F \colon \cX \to \cY \\ B \colon \cY \to \cX}} \frac{1}{m n} \sum_{i = 1}^{m} \sum_{j = 1}^{n} (c_{\cX}(x_i, B(y_j)) - c_{\cY}(F(x_i), y_j))^2 + \lambda \cdot \cL_{\cX \times \cY}((\mathrm{Id}, F)_{\#} \widehat{\mu}_m, (B, \mathrm{Id})_{\#} \widehat{\nu}_n)\;,
\end{equation*}
where $\widehat{\mu}_m$ and $\widehat{\nu}_n$ are the empirical measures based on $\{x_i\}_{i=1}^{m}$ and $\{y_j\}_{j=1}^{n}$, respectively. Empirically, we find that adding the following extra terms often enhance empirical results:
\begin{equation*}
\begin{split}
\min_{\substack{F \colon \cX \to \cY \\ B \colon \cY \to \cX}} \quad & \frac{1}{m n} \sum_{i = 1}^{m} \sum_{j = 1}^{n} (c_{\cX}(x_i, B(y_j)) - c_{\cY}(F(x_i), y_j))^2 + \lambda_1 \cdot \cL_{\cX \times \cY}((\mathrm{Id}, F)_{\#} \widehat{\mu}_m, (B, \mathrm{Id})_{\#} \widehat{\nu}_n) \\
& + \lambda_2 \cdot \cL_{\cX}(\widehat{\mu}_m, B_{\#} \widehat{\nu}_n) + \lambda_3 \cdot \cL_{\cY}(F_{\#} \widehat{\mu}_m, \widehat{\nu}_n)\;.
\end{split}
\end{equation*}
Like $\cL_{\cX \times \cY}$, we utilize suitable discrepancy measures $\cL_{\cX}$ and $\cL_{\cY}$ so that these additional terms help matching the marginals of $(\mathrm{Id}, F)_{\#} \widehat{\mu}_m$ and $(B, \mathrm{Id})_{\#} \widehat{\nu}_n$.
Lastly, we discuss the choice of $\cL_{\cX}, \cL_{\cY}$, and $\cL_{\cX \times \cY}$. We use the square of Maximum Mean Discrepancy (MMD) as the leading example.\footnote{This is merely a proof of concept. One may use other quantities in practice, described in Section \ref{sec:numerical}. MMD between two measures is a distance between their embeddings in some reproducing kernel Hilbert space (RKHS), which is indeed a metric under mild conditions \citep{muandet_fukumizu_sriperumbudur_scholkopf_2017}. Also, MMD is representable via the reproducing kernel of the RKHS, hence one may simply choose a kernel function to define it. Concretely, for any kernel $K_{\cX}$ on $\cX$, the square of MMD between $\widehat{\mu}_m$ and $B_{\#} \widehat{\nu}_n$ is
\begin{equation*}
\frac{1}{m^2} \sum_{i, i'} K_{\cX}(x_i, x_{i'}) + \frac{1}{n^2} \sum_{j, j'} K_{\cX}(B(y_j), B(y_{j'})) - \frac{2}{mn} \sum_{i, j} K_{\cX}(x_i, B(y_j))\;.
\end{equation*}
To utilize such a convenient closed form, we specify $\cL_{\cX}, \cL_{\cY}, \cL_{\cX \times \cY}$ as the square of corresponding MMDs by choosing kernels $K_{\cX}, K_{\cY}, K_{\cX \times \cY}$ on $\cX, \cY, \cX \times \cY$.} For the kernel $K_{\cX \times \cY}$ on the product space, we use the tensor product kernel $K_{\cX} \otimes K_{\cY}$ given as
\begin{equation*}
K_{\cX} \otimes K_{\cY}((x, y), (x', y')) = K_{\cX}(x, x') K_{\cY}(y, y')\;.
\end{equation*}
The tensor product notation is employed since the kernel on the product space inherits the feature map as the tensor product of two individual feature maps w.r.t. $K_{\cX}$ and $K_{\cY}$.
Denoting the MMD associated with a kernel $K$ as $\mathrm{MMD}_{K}$, we obtain the following minimization problem:
\begin{equation}
\label{eqn:1}
\begin{split}
\min_{\substack{F \colon \cX \to \cY \\ B \colon \cY \to \cX}} \quad & \frac{1}{m n} \sum_{i = 1}^{m} \sum_{j = 1}^{n} (c_{\cX}(x_i, B(y_j)) - c_{\cY}(F(x_i), y_j))^2 \\
& + \lambda_1 \cdot \mathrm{MMD}_{K_{\cX} \otimes K_{\cY}}^2((\mathrm{Id}, F)_{\#} \widehat{\mu}_m, (B, \mathrm{Id})_{\#} \widehat{\nu}_n)\\
& + \lambda_2 \cdot \mathrm{MMD}_{K_{\cX}}^2(\widehat{\mu}_m, B_{\#} \widehat{\nu}_n) + \lambda_3 \cdot \mathrm{MMD}_{K_{\cY}}^2(F_{\#} \widehat{\mu}_m, \widehat{\nu}_n) \;.
\end{split}
\end{equation}
Once we solve the problem above, the solution $\widehat{F} \colon \cX\to\cY$ will serve as an approximate isomorphism and facilitate transform sampling of the target $\nu$ from a known distribution $\mu$. The map $\widehat{B}$ possesses similar properties as $\widehat{F}$, whereas the map $\widehat{F}$ is of our primary interest for sampling purposes. The reverse map $\widehat{B} \colon \cY \rightarrow \cX$ also embeds point clouds in $\cY$ into $\cX$, with approximate isomorphism properties in the sense of Gromov-Monge.
\subsection{Statistical Rate of Convergence}
\label{subsec:stat_rate}
Like other transform sampling approaches for generative models, we consider \eqref{eqn:1} using vector-valued function classes $\cF$ and $\cB$ parametrized by neural networks, and then optimize using a gradient descent algorithm. We emphasize this minimization problem is much simpler than adversarial formulations as in GANs: variational problems of GANs consist of minimization over a class of generators and maximization over a class of discriminators, which requires complex saddle-point dynamics \citep{daskalakis2017training, liang2018interaction}. In contrast, our RGM only solves a single minimization problem in network parameters. Although generally non-convex in nature, the parameter minimization problem in neural networks can often be efficiently optimized by stochastic gradient descent, and can even provably achieve the global optima if the loss satisfies certain Polyak-\L ojasiewicz conditions \citep{bassily_belkin_ma_2018}.
We investigate the statistical rate of convergence for this minimization problem, assuming the empirical problem \eqref{eqn:1} can be solved accurately. First, define
\begin{equation}
\label{eqn:cost-def}
\begin{split}
C(\mu, \nu, F, B) \coloneqq & \int (c_{\cX}(x, B(y)) - c_{\cY}(F(x), y))^2 \dd{\mu \otimes \nu} \\
& + \lambda_1 \cdot \mathrm{MMD}_{K_{\cX} \otimes K_{\cY}}^2((\mathrm{Id}, F)_{\#} \mu, (B, \mathrm{Id})_{\#} \nu)\\
& + \lambda_2 \cdot \mathrm{MMD}_{K_{\cX}}^2(\mu, B_{\#} \nu) + \lambda_3 \cdot \mathrm{MMD}_{K_{\cY}}^2(F_{\#} \mu, \nu) \;.
\end{split}
\end{equation}
Then, the objective function of \eqref{eqn:1} is a plug-in estimator $C(\widehat{\mu}_m, \widehat{\nu}_n, F, B)$. We consider solving \eqref{eqn:1} over the transformation class $\cF\times\cB$ given as follows, for which we will state our non-asymptotic results in full generality. From now on, let $\cX$ and $\cY$ be subsets of Euclidean spaces of dimensions $\mathrm{dim}(\cX)$ and $\mathrm{dim}(\cY)$, respectively. $\cF$ (resp.\ $\cB$) is a collection of vector-valued measurable functions from $\cX$ to $\cY$ (resp.\ from $\cY$ to $\cX$). For each $F \in \cF$ and $k \in [\mathrm{dim}(\cY)]$, we write $F_k(x)$ to denote the $k$-th coordinate of $F(x)$. Accordingly, we define $\cF_k =\{F_k: \cX \rightarrow \R ~|~ F \in \cF\}$, namely, a collection of real-valued measurable functions defined on $\cX$ that are given as the $k$-th coordinate of $F \in \cF$. For $\ell \in [\mathrm{dim}(\cX)]$, we define $B_\ell$ and $\cB_{\ell} = \{B_\ell: \cY \rightarrow \R ~|~ B \in \cB\}$ analogously.
Then, solving \eqref{eqn:1} over $\cF \times \cB$ is written as
\begin{equation*}
\min_{(F, B) \in \cF \times \cB} C(\widehat{\mu}_m, \widehat{\nu}_n, F, B)\;.
\end{equation*}
We prove that the empirical solution leads to an approximate infimum of $(F, B) \mapsto C(\mu, \nu, F, B)$ evaluated with the population measures $\mu, \nu$, with sufficiently large sample sizes $m$ and $n$.
\paragraph{Overview of assumptions}
Before stating the next theorem, we present an overview of the assumptions. The complete statement of the assumptions and key definitions are designated to Sections~\ref{sec:statistical-theory}-\ref{sec:representation} in the supplementary material due to space constraints. Assumptions~\ref{a:bounded1} and \ref{a:lip_of_C} require the boundedness and Lipschitzness of the cost functions $c_{\cX}$ and $c_{\cY}$. Similarly, boundedness and Lipschitzness of the kernel functions $K_{\cX}, K_{\cY}$ corresponding to the MMD term are stated in Assumptions~\ref{a:bounded_kernels} and \ref{a:lip_kernels}, respectively. The last two assumptions are imposed on the set of transformations $F \colon \cX\rightarrow \cY$ and $B \colon \cY \rightarrow \cX$: Assumption~\ref{a:uniform_boundedness} requires the transformation class is bounded, and Assumption~\ref{a:separation} states that the classes should contain non-trivial maps. We shall employ a notion of combinatorial dimension to measure the complexity of real-valued function classes---the pseudo-dimension---formally stated in Definition~\ref{def:pdim}.
\begin{theorem}
\label{thm:stat}
Let $(\widehat{F}, \widehat{B})$ be a solution to the empirical RGM problem
\begin{align*}
& (\widehat{F}, \widehat{B}) \in \argmin_{(F, B) \in \cF \times \cB} C(\widehat{\mu}_m, \widehat{\nu}_n, F, B)\;,
\end{align*}
with $C \colon \cP(\cX) \times \cP(\cY) \times \cF \times \cB \rightarrow \R$ defined in \eqref{eqn:cost-def}. Under Assumptions \ref{a:bounded1}-\ref{a:separation}, the following inequality holds with probability $1- \delta$ on $\{x_i\}_{i = 1}^m$ and $\{y_j\}_{j = 1}^n$
\begin{align}
C(\mu, \nu, \widehat{F}, \widehat{B}) - \inf_{(F, B) \in \cF \times \cB} C(\mu, \nu, F, B) \precsim \cM(\cF, \cB, m, n, \delta)\;.
\end{align}
Here, $\cM(\cF, \cB, m, n, \delta)$ denotes a complexity measure of $(\cF, \cB)$ given in terms of pseudo-dimensions (Pdim) of $\cF_k$ and $\cB_\ell$ defined in Definition~\ref{def:pdim}:
\begin{equation*}
\cM(\cF, \cB, m, n, \delta)
\coloneqq
\sqrt{\frac{\log(\tfrac{m \vee n}{\delta})}{m\wedge n}} + \sqrt{\frac{\log(m \vee n)}{m\wedge n} \left( \sum_{k =1}^{{\rm dim}(\cY)} {\rm Pdim}(\cF_{k}) + \sum_{\ell = 1}^{{\rm dim}(\cX)} {\rm Pdim}(\cB_{\ell}) \right)}\;.
\end{equation*}
\end{theorem}
We provide required assumptions and the full proof of Theorem \ref{thm:stat} in Section \ref{sec:statistical-theory} along with the definition of the pseudo-dimension (Definition~\ref{def:pdim}). When $\cF$ and $\cB$ are parametrized by neural network classes (the ones we will use for numerical demonstrations in Section~\ref{sec:numerical}), tight pseudo-dimension bounds established in \cite{anthony_bartlett_1999, harvey2017NearlytightVCdimension} can be plugged in Theorem \ref{thm:stat} for concrete non-asymptotic rates.
\subsection{Convex Formulation and Representer Theorem}
\label{subsec:cvx-representer}
As the last bit of our contributions, we study a convex formulation of solving \eqref{eqn:1} by relaxing and lifting it to an infinite-dimensional space. There are two reasons behind our convex formulation: first, as a computational alternative to the possibly non-convex optimization; second, to point out a connection with the Nadaraya-Watson estimator in classic nonparametric statistics. The crux lies in relaxing optimizing over the map $F \colon \cX \rightarrow \cY$ to optimizing over its induced (dual) linear operator $\bF \colon L^2_\cY \rightarrow L^2_\cX$ that maps functions on $\cY$ to functions on $\cX$, where $L^2_{\cX}$ is the collection of real-valued measurable functions $f$ defined on $\cX$ such that $\int_{\cX} f^2 \dd{\pi_{\cX}} < \infty$ given a Borel measure $\pi_\cX$ on $\cX$; similarly, define $L^2_{\cY}$ given a Borel measure $\pi_{\cY}$ on $\cY$. Then, for a measurable map $F \colon \cX \to \cY$, we can define $\bF \colon L^2_{\cY} \to L^2_{\cX}$ by letting $\bF(g) = g \circ F$ for all $g \in L^2_{\cY}$. Similarly, we define $\bB \colon L^2_{\cX} \to L^2_{\cY}$ for each measurable map $B \colon \cY \to \cX$. We will see $\bF$ and $\bB$ are well-defined bounded linear operators in Section \ref{sec:representation} under a mild assumption.
To state the representer theorem, consider \eqref{eqn:1} with $c_{\cX} = K_{\cX}$ and $c_{\cY} = K_{\cY}$, same as kernel functions specified in MMD terms. We show that this problem can be reduced to a finite-dimensional convex optimization by proving a representer theorem. Since finite-dimensional convex optimization can be optimized globally with provable guarantees, such a formulation can be solved numerically in an efficient way.
Let us lay out more details to state the result. Due to Mercer's theorem, let $\{ \phi_k \in L^2_{\cX} \}_{k \in \N} $ and $\{ \psi_\ell \in L^2_{\cY} \}_{\ell \in \N}$ be countable orthonormal bases of $L^2_{\cX}$ and $L^2_{\cY}$ where the kernels admit the following spectral decompositions:
\begin{align}
\label{eqn:spectral}
K_{\cX}(x, x') = \sum_k \lambda_k \phi_k(x) \phi_k(x') \;, \quad K_{\cY}(y, y') = \sum_{\ell} \gamma_\ell \psi_\ell(y) \psi_\ell(y') \;,
\end{align}
with positive eigenvalues $\lambda_k, \gamma_\ell >0$.
Since $\bF \colon L^2_{\cY} \rightarrow L^2_{\cX}$ defines a bounded linear operator, one can represent $\bF$ (correspondingly $\bB$) under the orthonormal bases
\begin{align}
\bF[\psi_{\ell}] = \sum_{k = 1}^{\infty} \bF_{k \ell} \phi_k\;, \quad \bB[\phi_{k}] = \sum_{\ell = 1}^{\infty} \bB_{\ell k} \psi_\ell\;.
\end{align}
Here, $[\bF_{k \ell}]$ is a semi-infinite matrix with each column describing the $L^2_{\cX}$ representation of $\bF[\psi_{\ell}]$ under the basis $\{ \phi_k \in L^2_{\cX} \}_{k \in \N}$. With a slight abuse of notation, we will write $\bF$ and $\bB$ to denote these matrices $[\bF_{k \ell}]$ and $[\bB_{\ell k}]$. Then, we will prove in Section \ref{sec:representation} that the objective function in \eqref{eqn:1} with $c_{\cX} = K_{\cX}$ and $c_{\cY} = K_{\cY}$ is
\begin{equation*}
\begin{split}
\Omega(\bF, \bB)
&\coloneqq
\frac{1}{mn} \sum_{i, j} (\Psi_{y_j}^\top \bB \Lambda \Phi_{x_i} - \Phi_{x_i}^\top \bF \Gamma \Psi_{y_j})^2 \\
&+ \lambda_1 \cdot \Bigg(\frac{1}{m^2} \sum_{i, i'} \Phi_{x_i}^\top \Lambda \Phi_{x_i'} \Phi_{x_i}^\top \bF \Gamma \bF^\top \Phi_{x_{i'}} + \frac{1}{n^2} \sum_{j, j'} \Psi_{y_j}^\top \Gamma \Psi_{y_{j'}} \Psi_{y_j}^\top \bB \Lambda \bB^\top \Psi_{y_{j'}} \\
&\hspace{200pt}- \frac{2}{m n} \sum_{i, j} \Psi_{y_j}^\top \bB \Lambda \Phi_{x_i} \Phi_{x_i}^\top \bF \Gamma \Psi_{y_j} \Bigg) \\
&+ \lambda_2 \cdot \left(\frac{1}{m^2} \sum_{i, i'} \Phi_{x_i}^\top \Lambda \Phi_{x_i'} + \frac{1}{n^2} \sum_{j, j'} \Psi_{y_j}^\top \bB \Lambda \bB^\top \Psi_{y_{j'}} - \frac{2}{m n} \sum_{i, j} \Psi_{y_j}^\top \bB \Lambda \Phi_{x_i}\right) \\
&+ \lambda_3 \cdot \left(\frac{1}{m^2} \sum_{i, i'} \Phi_{x_i}^\top \bF \Gamma \bF^\top \Phi_{x_{i'}} + \frac{1}{n^2} \sum_{j, j'} \Psi_{y_j}^\top \Gamma \Psi_{y_{j'}} - \frac{2}{m n} \sum_{i, j} \Phi_{x_i}^\top \bF \Gamma \Psi_{y_j}\right) \;.
\end{split}
\end{equation*}
Here, $\bF$ and $\bB$ are the matrices denoting the operators induced by $F$ and $B$, respectively, $\Phi_x = [\cdots, \phi_k(x), \cdots ]^\top \in \R^\infty$ and $\Psi_y = [\cdots, \psi_\ell(y), \cdots]^\top \in \R^\infty$ for any $x \in \cX$ and $y \in \cY$, and $\Lambda = \mathrm{diag}(\lambda_1, \lambda_2, \dots)$ and $\Gamma = \mathrm{diag}(\gamma_1, \gamma_2, \dots)$ are diagonal matrices. Hence, \eqref{eqn:1} can be lifted to an infinite-dimensional optimization problem
\begin{equation}
\label{eqn:op}
\min_{(\bF, \bB) \in \cC} ~ \Omega(\bF, \bB)\;,
\end{equation}
where $\cC$ denotes the constraint set implying that $\bF$ and $\bB$ are matrices corresponding to bounded linear operators induced by some maps $F \colon \cX \to \cY$ and $B \colon \cY \to \cX$.
We will relax this problem by removing the constraint set $\cC$, namely, by considering all matrices in $\R^{\infty \times \infty}$ as the decision variables,
\begin{equation}
\label{eqn:relaxed}
\min_{\bF, \bB \in \R^{\infty \times \infty}} ~ \Omega(\bF, \bB)\;.
\end{equation}
In other words, this relaxed problem minimizes $\Omega$ over any pair of infinite-dimensional matrices. The next result, which we refer to as the representer theorem, shows that \eqref{eqn:relaxed} boils down to a finite-dimensional convex program.
\begin{theorem}
\label{thm:representer}
Consider the optimization \eqref{eqn:op} under the assumptions in Proposition \ref{prop:representer}. Then, for any minimizer $(\bF^\star, \bB^\star)$ to the relaxed problem \eqref{eqn:relaxed}, we can find finite-dimensional matrices $\mathsf{F}_{m,n}^{\star} \in \R^{m \times n}$ and $\mathsf{B}_{n, m}^{\star} \in \R^{n \times m}$ such that
\begin{align*}
\bF^{\star} = \Lambda \Phi_m \mathsf{F}_{m,n}^{\star} \Psi_n^\top \;, \quad \bB^{\star} = \Gamma \Psi_n \mathsf{B}_{n,m}^{\star} \Phi_m^\top \;,
\end{align*}
where $\Lambda = \mathrm{diag}(\lambda_1, \lambda_2, \dots)$, $\Gamma = \mathrm{diag}(\gamma_1, \gamma_2, \dots)$, and $\Phi_{m} \in \R^{\infty \times m}$ and $\Psi_{n} \in \R^{\infty \times n}$ are matrices whose elements are $\phi_k(x_i)$ and $\psi_{\ell}(y_j)$, as defined in \eqref{eqn:spectral}. In this case, $\Omega(\bF^\star, \bB^\star)$ can be rewritten as $\omega(\mathsf{F}_{m, n}^{\star}, \mathsf{B}_{n, m}^{\star})$ for some convex function $\omega$ defined over $\R^{m \times n} \times \R^{n \times m}$. Hence, by minimizing $\omega$ over $\R^{m \times n} \times \R^{n \times m}$, we obtain a relaxation of \eqref{eqn:relaxed}, that is,
\begin{equation*}
\min_{\bF, \bB \in \R^{\infty \times \infty}} \Omega(\bF, \bB)
\ge
\min_{\substack{\mathsf{F}_{m, n} \in \R^{m \times n} \\ \mathsf{B}_{n, m} \in \R^{n \times m}}} \omega(\mathsf{F}_{m, n}, \mathsf{B}_{n, m}) \;.
\end{equation*}
In particular, the RHS is a finite-dimensional convex optimization. Lastly, this relaxation is tight, that is,
\begin{equation*}
\min_{\bF, \bB \in \R^{\infty \times \infty}} \Omega(\bF, \bB)
=
\min_{\substack{\mathsf{F}_{m, n} \in \R^{m \times n} \\ \mathsf{B}_{n, m} \in \R^{n \times m}}} \omega(\mathsf{F}_{m, n}, \mathsf{B}_{n, m}) \;,
\end{equation*}
if kernel matrices $\bK_{\cX}$ and $\bK_{\cY}$ whose elements are $K_{\cX}(x_i, x_{i'})$ and $K_{\cY}(y_j, y_{j'})$, are positive definite.
\end{theorem}
\begin{remark}
\rm
Looking inside the proof of Theorem \ref{thm:representer}, we know the solution to the infinite-dimensional optimization is an operator taking form of $\bF^\star = \Lambda \Phi_m \mathsf{F}_{m,n}^\star \Psi_n^\top$, with a finite-dimensional matrix $\mathsf{F}_{m,n}^\star\in \R^{m \times n}$. Therefore, for any $g \in L^2_\cY$, we can deduce
\begin{align}
\label{eqn:nw-connection}
\bF^\star[g](x) = \underbrace{K_\cX(x, X_m)}_{1\times m} \underbrace{\mathsf{F}_{m,n}^\star}_{m\times n} \underbrace{g(Y_n)}_{n\times 1} \;,
\end{align}
where $K_{\cX}(x, X_m)$ maps each $x \in \cX$ to a row vector whose $i$-th element is $K_{\cX}(x, x_i)$ and $g(Y_n)$ denotes a column vector whose $j$-th element is $g(y_j)$.
Now let's draw a connection between the classic Nadaraya-Watson estimator and \eqref{eqn:nw-connection}. For now consider a special case: $(x_i, y_i)$'s are paired with $m = n$. In such a case, Nadaraya-Watson estimator takes the form
\begin{align}
\sum_{i,j} K_{\cX}(x, x_i) \cdot \tfrac{1}{m} \delta_{i=j} \cdot g(y_j) \; ;
\end{align}
Namely, for a new point $x$, the corresponding function value $g(y)$ evaluated on its coupled $y = F(x)$ is a weighted average of $g(y_j)$'s according to the affinity $K_{\cX}(x, x_i)$. Our solution \eqref{eqn:nw-connection} extends the above nonparametric smoothing idea to the decoupled data case, where the coupling weights $\mathsf{F}_{m,n}^\star$ is based on a solution to a convex program, with
\begin{align}
\eqref{eqn:nw-connection} = \sum_{i,j} K_{\cX}(x, x_i) \cdot \mathsf{F}_{m,n}^\star[i,j] \cdot g(y_j) \;.
\end{align}
Lastly, we draw another connection to the Monte-Carlo integration. One downstream task after learning the distribution $\nu$ is to perform numerical integration of $g \in L^2_\cY$ under the measure $\nu \in \cP(\cY)$. In our transform sampling framework, this amounts to evaluate $\E_{y \sim F^\star_\# \mu}[ g(y)] = \E_{x \sim \mu}[ g\circ F^\star (x)]$. The integration, casted in the induced operator form, has the expression
\begin{align}
\label{eqn:weights-MC}
\E_{x \sim \mu} \big[ \bF^\star[g](x) \big] = \E_{x \sim \mu} \big[ \underbrace{K_\cX(x, X_m) \mathsf{F}_{m,n}^\star}_{=: W(x) \in \R^n } g(Y_n) \big] = \E_{x \sim \mu} \big[ \sum_{j=1}^n W_j(x) g(y_j) \big]
\end{align}
where $W(x)$ can be interpreted as the importance weights in the Monte-Carlo integration. We conclude with one more remark: if plug in instead $x \sim \widehat{\mu}_m$ in \eqref{eqn:weights-MC}, one can verify that under mild conditions,
\begin{align}
\E_{x \sim \widehat{\mu}_m} \big[ \bF^\star[g](x) \big] = \frac{1}{n} \sum_{j=1}^n g(y_j) \;.
\end{align}
In other words, with the empirical measure as input, \eqref{eqn:weights-MC} outputs the simple sample average.
\end{remark}
\section{Experiments}
\label{sec:numerical}
This section examines the empirical performance of the reversible Gromov-Monge sampler. Following Section \ref{subsec:stat_rate}, we find a minimum $(\widehat{F}, \widehat{B})$ of \eqref{eqn:1} over a suitable class $\cF \times \cB$ via gradient descent; we inspect the quality of transform sampling $(\widehat{F}_{\#} \mu \approx \nu$) and space isomorphism. Complete technical details of the experiments are deferred to Section~\ref{sec:appendix2}.
\paragraph{Gaussian distributions}
Consider two strongly isomorphic Gaussian distributions on $\cX = \cY = \R^2$: the base measure $\mu = N(0, I_2)$ and the target distribution $\nu = N(0, \Sigma)$, where $I_2$ is the identity matrix and the entries of $\Sigma$ are $\Sigma_{1 1} = \Sigma_{2 2} = 1$ and $\Sigma_{1 2} = \Sigma_{2 1} = 0.7$. We let $c_{\cX}(x, x') = x^\top x'$ and $c_\cY(y, y') = y^\top \Sigma^{-1} y'$, then two network spaces are strongly isomorphic by design; indeed, any pair $(F, B)$ given by $F(x) = \Sigma^{1/2} Q x$ and $B(y) = Q^{\top} \Sigma^{-1/2} y$ for $Q \in O(2)$, where $O(2)$ is the orthogonal group, yields $c_{\cX}(x, B(y)) = c_{\cY}(F(x), y)$ for all $x, y \in \R^2$, hence $F$ and $B$ are strong isomorphisms. We aim at obtaining such a pair of (linear) isomorphisms by letting $\cF = \cB = \{x \mapsto W x : W \in \R^{2 \times 2}\}$, that is, the collection of all linear maps from $\R^2$ to $\R^2$. We set $K_{\cX} = K_{\cY}$ as a degree-2 polynomial kernel that maps $(x, y)$ to $(x^\top y + 1)^2$; the resulting MMD compares distributions by matching the first two moments, which is sufficient to distinguish Gaussian distributions. The resulting linear maps are given by $\widehat{F}(x) = \mathbf{F} x$ and $\widehat{B}(y) = \mathbf{B} y$ for some $\mathbf{F}, \mathbf{B} \in \R^{2 \times 2}$ satisfying
\begin{equation*}
\mathbf{F} \mathbf{F}^\top
=
\begin{pmatrix}
1.035 & 0.751 \\
0.751 & 1.094
\end{pmatrix}\;,
\quad \mathbf{B} \Sigma \mathbf{B}^\top
=
\begin{pmatrix}
0.940 & 0.001 \\
0.001 & 0.944
\end{pmatrix} \; ,
\quad \mathbf{F} \mathbf{B}
=
\begin{pmatrix}
0.966 & 0.029 \\
-0.020 & 1.029
\end{pmatrix} \; .
\end{equation*}
Since $\mathbf{F} \mathbf{F}^\top \approx \Sigma$, $\mathbf{B} \Sigma \mathbf{B}^\top \approx I_2$, and $\mathbf{F} \mathbf{B} \approx I_2$, the pair $(\widehat{F}, \widehat{B})$ can be seen as an instance of the pair of strong isomorphisms described above. Figure \ref{fig:Gaussian} illustrates that $\widehat{F}$ is a strong isomorphism (Definition \ref{def:iso}): (a) shows that $\widehat{F}_{\#} \mu \approx \nu$, that is, $\widehat{F}$ is roughly a transport map, and (b) implies that $c_{\cX}(x, x') \approx c_{\cY}(\widehat{F}(x), \widehat{F}(x'))$ holds.
\begin{figure}[ht]
\centering
\subfloat[\centering]{{\includegraphics[width=8cm]{img/Gaussian-pushforward.pdf}}}
\subfloat[\centering]{{\includegraphics[width=8cm]{img/Gaussian-cost-vs-cost.pdf}}}
\caption{Gaussian experiment: $m = n = 1000$ and $\lambda_1 = \lambda_2 = \lambda_3 = 1$. (a) shows $\{\tilde{y}_j\}_{j = 1}^{400}$ versus $\{\widehat{F}(\tilde{x}_i)\}_{i = 1}^{400}$, where $\{\tilde{y}_j\}_{j = 1}^{400}$ and $\{\tilde{x}_i\}_{i = 1}^{400}$ are i.i.d.\ from $\nu = N(0, \Sigma)$ and $\mu = N(0, I_2)$, respectively; they are new samples independent from $\{y_j\}_{j = 1}^{1000}$ and $\{x_i\}_{i = 1}^{1000}$ used in \eqref{eqn:1}. (b) shows the points $\{(c_{\cX}(\tilde{x}_i, \tilde{x}_{i'}), c_{\cY}(\widehat{F}(\tilde{x}_i), \widehat{F}(\tilde{x}_{i'})))\}_{i, i' = 1}^{40}$ and a straight line $y = x$.}
\label{fig:Gaussian}
\end{figure}
\paragraph{MNIST}
Next, let $\nu$ be a distribution of images corresponding to four digits (2, 4, 6, 7) from the MNIST data set, which is supported on $\R^{784}$. Recall from Section \ref{sec:intro} that the support $\cY$ of $\nu$ is low-dimensional \citep{facco2017estimating}, hence choosing $\cX = \R^{d}$ with $d \ll 784$ is reasonable. Here, we try an extreme embedding task with $d = 2$ and $\mu = N(0, I_2)$, that is, generate MNIST images by transforming two-dimensional Gaussian samples.
Unlike the Gaussian example where we design the cost functions in advance to make the two spaces strongly isomorphic, specifying them can be more complicated in general cases, which might affect the quality of the RGM sampler. Here, we briefly discuss some of the most commonly used cost functions: given a fixed exponent $p \in \N$ or constant $\alpha > 0$,
\begin{equation*}
(x, y) \quad \mapsto \quad
\underbrace{\|x - y\|^p}_{\text{distance-based}}
\quad \text{or} \quad \underbrace{\exp(-\alpha \|x - y\|^2)}_{\text{RBF kernel}} \;.
\end{equation*}
Clearly, $\|x - y\|^p$ is the most straightforward choice in Euclidean cases; $p = 1$ and $p = 2$ are indeed widely used in the literature \citep{peyre_etal_2016}. The RBF kernel, also referred to as the heat kernel, is a common choice in the object matching literature \citep{solomon_etal_2016}. In this MNIST example, we have found that these cost functions provide reasonable performance once they are scaled properly. Here, we will present the results based on the RBF kernel. Concretely, first define the RBF kernel $K_d(x, y) = \exp(-\|x - y\|^2 / d)$ for $d \in \N$ and $x, y \in \R^d$; here, the constant $(1 /d)$ serves as a scaling factor. Then, we define the cost functions as $c_\cX = (K_{2} - m_\cX) / \mathrm{sd}_\cX$ and $c_\cY = (K_{784} - m_\cY) / \mathrm{sd}_\cY$, where $m_\cX$ and $\mathrm{sd}_\cX$ are the median and the standard error of $\{K_\cX(x_i, x_{i'})\}_{i, i' = 1}^m$, respectively; $m_\cY$ and $\mathrm{sd}_\cY$ are defined analogously. This additional standardization process helps aligning the cost functions.
In the same vein, $K_{\cX}$ and $K_{\cY}$ must be properly specified; comparing the first two moments using the degree-2 polynomial kernel is no longer sufficient as the target distribution is non-Gaussian. We suggest using RBF kernels for the MMD terms as well; let $K_\cX = K_2$ and $K_\cY = K_{784}$. The MMD induced by the RBF kernel indeed defines a metric between distributions under mild assumptions \citep{muandet_fukumizu_sriperumbudur_scholkopf_2017}, which allows the resulting MMD terms to represent the original constraint of the RGM distance as mentioned in Section \ref{sec:RGM-sampler}.
For the function classes $\cF$ and $\cB$, we need richer classes instead of the linear maps used in the Gaussian case. To this end, we will use fully connected neural networks with three hidden layers, each of which consists of 50 neurons. Lastly, we let $m = n = 20000$ and $\lambda_1 = \lambda_2 = \lambda_3 = 100$. Figure \ref{fig:MNIST-images}(a) shows the images generated by applying the resulting map $\widehat{F}$ to new i.i.d.\ samples from $\mu = N(0, I_2)$. Though not perfect, we see that recognizable images can be generated by transforming two-dimensional Gaussian samples, efficient in computation.\footnote{Computational cost for obtaining $\widehat{F} \colon \R^2 \to \R^{784}$ and computing $\widehat{F}(X)$ from $X \sim \mu$ is far less than that of the OT-based sampler as explained in Section \ref{sec:intro}.}
\begin{figure}[ht]
\centering
\subfloat[MMD ($\R^2$)]{{\includegraphics[width=5cm]{img/mnist-MMD.pdf}}}%
\subfloat[Sinkhorn ($\R^4$)]{{\includegraphics[width=5cm]{img/mnist-Sinkhorn.pdf}}}%
\subfloat[Original]{{\includegraphics[width=5cm]{img/mnist-real.pdf}}}%
\caption{(a) and (b) are generated by transforming new i.i.d.\ samples from $\mu$ using $\widehat{F}$: (a) from $\mu = N(0, I_2)$ with MMDs and (b) from $\mu = N(0, I_4)$ with Sinkhorn divergences. (c) shows real MNIST images.}
\label{fig:MNIST-images}
\end{figure}
Meanwhile, the map $\widehat{B}$ shows how the MNIST images can be embedded in $\R^2$. Figure \ref{fig:MNIST-backward}(a) shows $\{\widehat{B}(\tilde{y}_j)\}_{j = 1}^{500}$, where $\{\tilde{y}_j\}_{j = 1}^{500}$ are i.i.d.\ from $\nu$ (125 $\times$ 4 digits), independent from $\{y_j\}_{j = 1}^{20000}$ used in \eqref{eqn:1}. We see that each digit forms a local cluster in $\R^2$, each of which is roughly representable according to the range of the angular coordinate. Lastly, though not perfect as in Figure \ref{fig:Gaussian}(b) (strongly isomorphic case), Figure \ref{fig:MNIST-backward}(b) shows that $\widehat{B}$ leads to a reasonable alignment of $c_\cX(\widehat{B}(y), \widehat{B}(y'))$ versus $c_\cY(y, y')$.
\begin{figure}[ht]
\centering
\subfloat[\centering]{{\includegraphics[height=5cm]{img/mnist-pullbackward.pdf}}}
\subfloat[\centering]{{\includegraphics[height=5cm]{img/mnist-cost-vs-cost.pdf}}}
\caption{(a) is generated by applying $\widehat{B}$ to $500$ out-of-sample MNIST images, i.i.d.\ $\{\tilde{y}_j\}_{j = 1}^{500}$ from $\nu$. (b) shows the points $\{(c_\cX(\widehat{B}(\tilde{y}_{j}), \widehat{B}(\tilde{y}_{j'})), c_\cY(\tilde{y}_j, \tilde{y}_{j'}))\}_{j, j' = 1}^{50}$ and a straight line $y = x$.}
\label{fig:MNIST-backward}
\end{figure}
We clarify that the current experiment with $\mu = N(0, I_2)$ is a proof of concept. Suppose one aims to obtain images comparable to those from dedicated MNIST generators. In that case, exhaustive tests should be done for tuning each component of the RGM sampler, which is beyond the scope of this paper. Instead, we highlight that the RGM sampler with a simple modification can indeed generate significantly improved images seen in Figure \ref{fig:MNIST-images}(b). These images are generated from the following settings that are fully introduced in Section~\ref{sec:appendix2}: $(\cX, \mu, c_{\cX}) = (\R^4, N(0, I_4), K_4)$ and MMD terms in \eqref{eqn:1} are replaced by Sinkhorn divergences \citep{genevay2018learning}. As such, the RGM sampler is amendable to other more general choices of $\cL_{\cX \times \cY}$ in its practical implementation.
\section{Discussions}
In this work, we proposed a novel distance between network spaces, called the Reversible Gromov-Monge distance, inspired by the Gromov-Wasserstein distance between metric measure spaces. Based on this, we designed a transform sampler using empirical data that can operate between distributions defined on heterogeneous spaces. In addition, we introduced two concrete optimization methods for computing RGM given finite samples and proved their properties. The resulting RGM transform sampler is more efficient than the classic optimal transport transform sampler in terms of computation: its optimization procedure is less complicated than that of GAN-type samplers.
Lastly, we briefly mention computational aspects of the RGM distance that are not fully investigated. We have mainly relied on a practical computation of the RGM distance using the Lagrangian form instead of the constrained form \eqref{eq:RGM}, which was sufficient for the good empirical performance of the RGM sampler as witnessed in Section \ref{sec:numerical}. However, one might be interested in the exact compuation instead and ask: when does minimizing the Lagrangian form lead to a close approximation to the RGM distance? Another important aspect is the comparison with the GW distance. Proposition \ref{prop:1} shows $\mathrm{GW}(\mu, \nu) \le \mathrm{RGM}(\mu, \nu)$ in theory; by approximating both quantities numerically, we can see how large the gap between the two distances is. We provide our findings and insights regarding these issues in Section \ref{sec:comparison-with-GW} for the interested readers; though not central to this paper, these can be of interest to the GW literature and we leave them as future research.
\section*{Acknowledgments}
Liang acknowledges the generous support from the NSF Career award (DMS-2042473), and the William S. Fishman Faculty Research Fund at the University of Chicago Booth School of Business. Liang wishes to thank Maxim Raginsky and Chris Hansen for discussions on simulation-based inference.
|
\section{Motivations and overview}
One of the most vexed problems that causes considerable confusion about gravitational waves is what their physical effects on matter really are. The confusion arose partly because one cannot define a localized energy-momentum for the gravitational field, and consequently the notion of stress-energy tensor of gravitational waves makes sense only as an average over several wavelengths in a ``coarse-grain'' sense (see Sections 20.4, 35.7, and 35.15 in \cite{Misner:1974qy}), rendering it somewhat unclear how the energy of a gravitational wave is transferred into matter.
Opinions on this issue had widely diverged until the \emph{sticky bead argument} was proposed (anonymously) by Feynman \cite{Feynman:a,Feynman:b} and by Bondi \cite{Bondi:1957dt}. The sticky bead argument suggests that, as a gravitational wave passes over two beads sliding with friction on a rigid rod, the beads will rub against the rod, thus absorbing some of the energy carried by the wave and dissipating it into heat.
A similar argument also applies to different classical mechanical systems, the response of which can be used to detect gravitational waves (see Chapter 37 of \cite{Misner:1974qy}). A notable example is the \emph{resonant mass detector}, which has been operated in various experiments (see \cite{Aguiar:2010kn} for a review) as alternatives to interferometric gravitational wave detectors (see \cite{Saulson:2017gqp,Reitze:2019nwo} for reviews).
To derive the response of a classical mechanical system to a gravitational wave, the equation of motion for mass elements of the system is dealt with in a standard Newtonian manner, except that, as a non-Newtonian effect, the tidal force produced by the gravitational wave provides the driving force against the Newtonian interacting force (e.g., friction in the sticky bead system, elastic and damping forces in the resonant mass detector, etc.) between mass elements (see Chapter 37, especially Section 37.2, of \cite{Misner:1974qy} for a detailed account).
This analysis is straightforward and easy to understand, but it might not manifest some subtle effects of gravitational waves not directly resulting from the tidal force.
Therefore, instead of phenomenologically considering the response of a classical mechanical system, it will yield valuable new insight into the gravitational wave effects on matter, if the response of a quantum system to gravitational waves can be studied from a more fundamental setting.
In this paper, we consider probably the simplest kind of such a theoretical quantum system --- the Unruh-DeWitt detector coupled to a massless real scalar field, and manage to investigate its response to a gravitational wave background for the two cases of a free-falling trajectory and a constant-accelerating trajectory. Our investigation shows that, in both cases, the transition rate of the Unruh-DeWitt detector is different from the result with no gravitational wave, and the leading-order correction due to the gravitational wave survives the long-wavelength limit --- an intriguing effect that cannot be explained out in terms of the gravitational wave tidal force.\footnote{The underlying mechanism of state transition of the Unruh-DeWitt detector is that the detector is coupled to quantum fluctuations of a quantum field of interest in the vacuum background, akin to spontaneous emission of an atom as a consequence of being coupled to quantum fluctuations of the electromagnetic field. Accordingly, simply by moving the Unruh-DeWitt detector in the vacuum background, it will yield a certain transition rate, which is measurable at least in principle. Experimentally speaking, it seems more realistic to model the detector as coupled to the electromagnetic field, instead of a massless scalar field, as Nature evidently has quantum fluctuations of the former in vacuum, but may not of the latter. Nevertheless, we adopt the model of a real scalar field, because not only it is theoretically the simplest but also it may still give the same measurement result of the detector that is coupled to the electromagnetic field but insensitive to its polarization. \secref{sec:Unruh-DeWitt detector} will further elaborate on the related issues concerning measurement.}
In the literature of the Unruh-DeWitt detector, it has been shown that the response of the Unruh-DeWitt detector is modified in the presence of boundaries \cite{Chiou:2016exd,Davies:1989me}, essentially because the boundary condition alters the mode expansion of the quantum field. In the limit that the length scale delimited by the boundaries goes to infinity, the modification becomes negligible.
In the case of a gravitational wave background, the detector's response is anticipated to change as well, since the gravitational wave also alters the mode expansion. However, in the formal limit that the wavelength of the gravitational wave goes to infinity, this change does not diminish as long as the amplitude of the gravitational wave remains finite. The gravitational wave effect on the Unruh-DeWitt detector is more involved than merely imposing a large length scale of the wavelength.
This paper is organized as follows.\footnote{Throughout this paper, we adopt the convention $(-,+,+,+)$ for the metric signature and use the natural units with both the Plank constant $\hbar$ and the speed of light $c$ set to unity.}
In \secref{sec:Unruh-DeWitt detector}, we give a brief review on the Unruh-DeWitt detector.\footnote{\secref{sec:Unruh-DeWitt detector} is based on Sec.~II and Appendix A of \cite{Chiou:2016exd}.}
In \secref{sec:scalar field}, we solve the equation of motion of a scalar field, i.e., the Klein-Gordon
equation, in a gravitational wave background. We then quantize the scalar field in the gravitational wave background using the light-front quantization \cite{Burkardt:1995ct} in \secref{sec:LF quantization}, and compute the Wightman function in \secref{sec:Wightman function}. With the Wightman function at hand, we compute the response of the Unruh-DeWitt detector along a free-falling trajectory and a constant-accelerating trajectory in \secref{sec:free-falling} and \secref{sec:constant-accelerating}, respectively. Finally, in \secref{sec:summary}, the results and their implications are summarized and remarked.
Additionally, we also explicitly solve the geodesic equation in a gravitation wave background in \appref{app:geodesic eq}.
\section{The Unruh-DeWitt detector}\label{sec:Unruh-DeWitt detector}
Whereas the notion of ``particles'' of a quantum field is clear to recognize and understand in flat spacetime, it is rather ambiguous in the context of quantum field theory in curved spacetime, as the particle content, quite surprisingly, turns out to be observer-dependent \cite{Fulling:1972md}. To have an unequivocal notion of particles, it thus requires an operational definition in terms of the response of a well-defined ``particle detector''. (The idea of a particle detector has already been considered for a different motivation in quantum optics by Glauber in 1963 \cite{Glauber:1963fi}.)
In 1976, Unruh proposed a theoretical model of such a particle detector and used it to address the problem of the particle content in relation to the observer's trajectory \cite{Unruh:1976db}. Unruh's detector is modeled as a point object in a small box coupled to the quantum field of interest, by which a particle of the quantum field is said to be detected if the object in the box is excited from its initial ground state to some excited state. (A similar model was also developed by S\'{a}nchez in 1981 \cite{Sanchez:1981xx}.)
In 1979, DeWitt \cite{DeWitt:1979} further improved Unruh's idea by simplifying the model as a two-level point monopole detector, which is now generally referred to as the \emph{Unruh-DeWitt detector} and widely used as a theoretical tool to probe quantum field effects in various settings of spacetime.
In this section, we briefly review the model of the Unruh-DeWitt detector, following closely the line of Sec.\ 3.3 in \cite{Birrell:1982ix}. Unlike \cite{Birrell:1982ix}, we consider the transition rates of both excitation ($\Delta E>0$) and de-excitation ($\Delta E<0$), and also take into account the switching function $\chi(\tau)$ as introduced in \cite{Louko:2006zv,Satz:2006kb,Louko:2007mu} in order to address the issue of regularization. We also briefly recap some passages in Appendix A of \cite{Chiou:2016exd} to address the measurement of the transition rate and the concept of detailed balance. For more about the Unruh-DeWitt detector and also the Unruh effect, see \cite{Birrell:1982ix,Wald:book,Padmanabhan:2003gd} and especially the comprehensive review article \cite{Crispino:2007eb}.
The Unruh-DeWitt detector is an idealized model with two energy levels, $\ket{E_0}$ and $\ket{E}$, coupled to a scalar field $\phi$ via a monopole interaction. If the detector moves along a world line $x^\mu(\tau)$, where $\tau$ is the detector's proper time, the Lagrangian for the monopole interaction is given by
\begin{equation}\label{monopole interaction}
\kappa\,\chi(\tau)\mu(\tau)\phi(x^\mu(\tau)),
\end{equation}
where $\kappa$ is a small coupling constant, $\mu(\tau)$ is the operator of the detector's monopole moment, and $\chi(\tau)$ is the switching function, which accounts for the switch-on and switch-off of the detector. As the switching function $\chi(\tau)$ can be modeled as a smooth enough function with a compact support as depicted in \figref{fig:switching function}, its inclusion introduces a finite timescale $\Delta$ for the switch-on period.
\begin{figure}
\centering
\includegraphics[]{switching_function.pdf}
\caption{A typical switching function $\chi(\tau)$ with a finite timescale $\Delta$ for the switch-on period.}
\label{fig:switching function}
\end{figure}
Moving along a given trajectory $x^\mu(\tau)$, the detector in general does not remain in its initial state $\ket{E_0}$ but can be excited (if $\Delta E := E-E_0 >0$) or de-excited (if $\Delta E<0$) to the other state $\ket{E}$, while at the same time the field $\phi$ makes a transition from the vacuum state $\ket{0}$ to an excited state $\ket{\Psi}$. By the first-order perturbation theory, the amplitude for the transition
\begin{equation}\label{transition}
\ket{0,E_0}\rightarrow\ket{\Psi,E}
\end{equation}
is given by
\begin{equation}\label{transition amplitude 0}
i\kappa\,\bra{\Psi,E}\int_{-\infty}^\infty \chi(\tau) \mu(\tau)\, \phi\left(x^\mu(\tau)\right) d\tau \ket{0,E_0},
\end{equation}
which leads to the factorized form
\begin{equation}\label{transition amplitude}
i\kappa \bra{E}\mu(0)\ket{E_0} \int_{-\infty}^\infty e^{i(E-E_0)\tau} \chi(\tau) \bra{\Psi}\phi\left(x^\mu(\tau)\right)\ket{0}\,d\tau
\end{equation}
by the equation of evolution for $\mu(\tau)$,
\begin{equation}
\mu(\tau)=e^{iH_0\tau}\mu(0)e^{-iH_0\tau},
\end{equation}
where $H_0$ is the Hamiltonian of the detector.
Summing the squared norm of the amplitude given in \eqref{transition amplitude} over all possible $\ket{\Psi}$,\footnote{Here, we use the completeness relation $\sum_{\ket{\Psi}}\ket{\Psi}\bra{\Psi}=\mathbbm{1}$, but note that, at the level of the first-order perturbation, only the one-particle states of $\ket{\Psi}$ contribute.} we obtain the transition probability of $\ket{E_0}\rightarrow\ket{E}$ as
\begin{equation}\label{transition probability}
\kappa^2\abs{\bra{E}\mu(0)\ket{E_0}}^2\ F(E-E_0),
\end{equation}
where
\begin{equation}\label{response function}
F(\Delta E) = \int_{-\infty}^\infty d\tau \int_{-\infty}^\infty d\tau'
e^{-i\Delta E(\tau-\tau')} \chi(\tau)\,\chi(\tau')\, D^+(x(\tau),x(\tau'))
\end{equation}
is the \emph{response function}, which depends on the trajectory but not the internal properties of the detector. The remaining factor, $c^2\abs{\bra{E}\mu(0)\ket{E_0}}^2$, represents the \emph{selectivity}, which depends only on the detector's internal properties.\footnote{In following sections, we will focus on the response function $F(\Delta E)$ and ignore the factor of selectivity.} The Wightman functions $D^{\pm}$ are defined as
\begin{subequations}
\begin{eqnarray}
D^+(x,x')&:=&\bra{0}\phi(x)\phi(x')\ket{0},\\
D^-(x,x')&:=&\bra{0}\phi(x')\phi(x)\ket{0}.
\end{eqnarray}
\end{subequations}
It should be noted that the first-order perturbation method is viable only for a short range of evolution time, since the squared norm of transition amplitude has to remain much smaller than unity. Therefore, \eqref{transition amplitude 0} with the unbounded integral $\int_{\infty}^\infty d\tau$ is problematic, unless a smooth enough switching function $\chi(\tau)$ with a finite switch-on duration is imposed. The imposition of $\chi(\tau)$ can be viewed as a prescription of regularization to make sense of the perturbation method.
The detector moving along a given trajectory $x(\tau)$ is said to be in equilibrium with the field $\phi$, if
\begin{equation}\label{equilibrium}
D^+(\tau,\tau')\equiv D^+(x(\tau),x(\tau')) = D^+(\Delta\tau), \quad \Delta\tau:=\tau-\tau',
\end{equation}
which depends only on $\Delta\tau$.
In this case, even though \eqref{response function} becomes infinite without the inclusion of $\chi(\tau)$, simply by setting $\chi(\tau)=1$, the (infinite) total transition probability divided by the (infinite) total proper time still sensibly yields a finite \emph{equilibrium transition rate} (i.e., probability per unit proper time) given as
\begin{equation}\label{transition rate}
R=
\kappa^2\abs{\bra{E}m(0)\ket{E_0}}^2
\dot{F}(\Delta E),
\end{equation}
where
\begin{equation}\label{F dot}
\dot{F}(\Delta E) := \int_{-\infty}^\infty d(\Delta\tau)
e^{-i\Delta E\Delta\tau} D^+(\Delta\tau).
\end{equation}
However, the burden of regularization is now carried over to the Wightman function $D^+(x,x')$, which will require some proper regularization procedure, such as the standard $i\epsilon$-regularization used in the case of a detector moving in Minkowski spacetime.\footnote{The standard $i\epsilon$-regularization can be replaced with different regularization procedures, e.g., by imposing a switching function $\chi(\tau)$ or introducing a spatial profile of the Unruh-DeWitt detector. See \cite{Louko:2006zv,Satz:2006kb,Louko:2007mu} for more discussions on the issue of regularization.}
We will make more comments on this point when we encounter the issue of regularization for the Wightman function in \secref{sec:Wightman function}.
On the other hand, if the detector is not in equilibrium with $\phi$ (i.e., $D^+(\tau,\tau')$ depends on both $\tau$ and $\tau'$ for the given trajectory), we can no longer make sense of the notion of equilibrium transition rate but can only refer to the \emph{total} transition probability, which now depends on the exact form of $\chi(\tau)$. Provided that $\chi(\tau)$ is smooth enough and its switch-on duration $\Delta$ is short enough (so that the first-order perturbation is viable), \eqref{transition probability} with \eqref{response function} is well defined and yields a finite total transition probability. By taking the time derivative of the total transition probability, we can still define the \emph{instantaneous transition rate} observed at a particular instant. We refer readers to \cite{Chiou:2016exd,Louko:2006zv,Satz:2006kb,Louko:2007mu} for more discussions on the non-equilibrium case, as we will only focus on the equilibrium case in this paper.
Both the equilibrium transition rate and the instantaneous transition rate in principle can be experimentally measured by deploying a large ensemble of identical Unruh-DeWitt detectors (with the same coupling constant and the same switching function, moving in the same trajectory). Measuring the ratio of the population of the detectors in the ensemble staying in the initial state $\ket{E_0}$ to that of the detectors excited or de-excited to the other state $\ket{E}$, one can deduce the transition rate. (See Appendix A of \cite{Chiou:2016exd} for more details.)
In some situations, the transition process \eqref{transition} and its reverse process $\ket{\Psi,E}\rightarrow\ket{0,E_0}$ can reach \emph{detailed balance}. If the detailed balance is established, the principle of detailed balance dictates that the transition rate $\dot{P}$ of \eqref{transition} and the transition rate $\dot{P}_r$ of its reverse process are both independent of $\tau$ and satisfy
\begin{equation}\label{detailed balance}
\frac{\dot{P}}{\dot{P}_r}=e^{-\beta\Delta E},
\end{equation}
where $1/\beta\equiv k_\mathrm{B}T$ is to be interpreted as the corresponding temperature. The ratio $\dot{P}/\dot{P}_r$ and therefore the temperature of detailed balance in principle can be measured again in terms of a large ensemble of identical detectors (but with a different measuring operation performed upon the ensemble).
Because the amplitudes of the transition process \eqref{transition} and its reverse process are complex conjugate to each other as a consequence of unitarity, the temperature of detailed balance is independent of the explicit design of the detector, as we can see in the examples below.
It should also be remarked that, whereas the condition that the trajectory is in equilibrium with the background field is necessary for detailed balance, it is unclear whether the condition is also sufficient. (See Appendix A of \cite{Chiou:2016exd} and Sec.\ III.A.4 of \cite{Crispino:2007eb}for more discussions about detailed balance.)
In the celebrated example of the Unruh-DeWitt detector moving with a constant acceleration in the Minkowski spacetime, the detailed balance relation is satisfied and the corresponding temperature is given by
\begin{equation}\label{Unruh temperature}
T=\frac{\abs{\text{acceleration}}}{2\pi k_\mathrm{B}},
\end{equation}
which is called the \emph{Unruh temperature}.
This is a consequence of the fact that the Minkowski vacuum is a thermal state of the right (left) Rindler modes (which are the modes of particles seen by the constant-accelerating observer) if the left (right) Rindler modes (which are the modes beyond the apparent event horizon of the constant-accelerating observer) are ignored. More precisely, tracing out the left (right) Rindler modes upon the Minkowski vacuum state gives rise to a density matrix for the many-particle system of the right (left) Rindler modes at the temperature \eqref{Unruh temperature}. (See Secs. III.A.2 and III.A.4 of \cite{Crispino:2007eb} for more details.)
From the perspective of the right (left) Rindler observer, we have
\begin{equation}\label{detailed balance 2}
\frac{\dot{P}}{\dot{P}_r} =\frac{\abs{\mathcal{A}}^2 n(\Delta E)}{\abs{\mathcal{A}_r}^2\left(1+n(\Delta E)\right)}
=\frac{n(\Delta E)}{1+n(\Delta E)}
=e^{-\beta\Delta E},
\end{equation}
where $\mathcal{A}$ and $\mathcal{A}_r$ are the amplitudes measured by the right (left) Rindler observer for the process \eqref{transition} and its reverse process, respectively, which are complex conjugate to each other, and
\begin{equation}
n(\omega) = \frac{1}{e^{\beta\omega}-1}
\end{equation}
is the Rindler particle number density for the thermal state at the Unruh temperature \eqref{Unruh temperature}. The factor $n(\Delta E)$ in the numerator in \eqref{detailed balance 2} is associated with the induced absorption of a Rindler particle from the thermal bath, and the factor $1+n(\Delta E)$ in the denominator is associated with the spontaneous and induced emissions of a Rindler particle to the thermal bath.\footnote{The amplitude of the transition of \eqref{transition} as $\ket{E_0}\rightarrow\ket{E}$ accompanied by the emission of a Minkowski-mode particle into the Minkowski vacuum can be reproduced from the Rindler observer's perspective as accompanied by the absorption of a Rindler-mode particle from the thermal bath (see Sec.~III.A.2 of \cite{Crispino:2007eb} for more details).}
For the case of an Unruh-DeWitt detector moving with a constant velocity in the Minkowski spacetime, from the perspective of a non-moving observer, one can easily compute
\begin{equation}\label{detailed balance 3}
\frac{\dot{P}}{\dot{P}_r} =\frac{\abs{\mathcal{A}}^2 \left(1+n(\Delta E)\right)}{\abs{\mathcal{A}_r}^2n(\Delta E)},
\end{equation}
where $\mathcal{A}$ and $\mathcal{A}_r$ are the amplitudes measured by the non-moving observer, which are complex conjugate to each other, and where
\begin{equation}
n(\omega_{\vec{k}}) := \bra{0} a^\dagger_{\vec{k}} a_{\vec{k}} \ket{0} = 0
\end{equation}
is the particle number density for the Minkowski vacuum state $\ket{0}$. The factor $1+n(\Delta E)$ is associated with the spontaneous and induced emissions of a particle to $\ket{0}$, and the factor $n(\Delta E)$ is associated with the induced absorption of a particle from $\ket{0}$. For $\Delta E<0$, it turns out that $\abs{\mathcal{A}}^2=\abs{\mathcal{A}_r}^2\neq0$, and consequently \eqref{detailed balance 3} yields $\dot{P}/\dot{P}_r=\infty$. Therefore, the detailed balance is satisfied in the trivial way corresponding to the zero temperature $T=0$ (i.e., $\beta=\infty)$. For $\Delta E>0$, it turns out that $\abs{\mathcal{A}}^2=\abs{\mathcal{A}_r}^2=0$, and the temperature is ill-defined.
In this paper, we study the Unruh-DeWitt detector moving along a free-falling trajectory or along a constant-accelerating trajectory in a monochromatic gravitational wave background, instead of the Minkowski background. As we will see, in various settings in the long-wavelength or short-wavelength limit of the gravitational wavelength, a free-falling or constant-accelerating detector is in equilibrium with $\phi$. The equilibrium transition rate depends on the amplitude of the gravitational wave, which can be viewed as a measurable effect of gravitational waves acting on a quantum system.
However, we do not consider detailed balance and the corresponding temperature. It is unclear whether detailed balance can be established in the presence of a gravitational wave. Even if detailed balance is established in certain settings, it is difficult to obtain the particle number density $n(\omega)$ in the gravitational wave background that can be used to compute the ratio $\dot{P}/\dot{P}_r$. We leave the issue of detailed balance for future research.
Finally, we remark that the underlying mechanism of the transition \eqref{transition} is that the Unruh-DeWitt detector is coupled to quantum fluctuations of $\phi$ in the vacuum $\ket{0}$, akin to spontaneous emission of an atom or a molecule as a consequence of being coupled to quantum fluctuations of the electromagnetic field. From the experimental perspective, it seems more realistic to model the detector as coupled to the electromagnetic field, rather than a scalar field $\phi$, as Nature evidently has quantum fluctuations in the vacuum of the electromagnetic field, but may not of a scalar field.
Indeed, generalized Unruh-DeWitt models coupled with different kinds of quantum fields have been formulated, including the electromagnetic field \cite{Boyer:1980wu,Boyer:1984yqq} and the Dirac field \cite{Iyer:1980yc}. However, in this paper, we adhere to the original Unruh-DeWitt model with a scalar field, as it is theoretically the simplest and its simplicity enables us to obtain the transition rate in a closed form that is easier to analyze.
The transition rate of the Unruh-DeWitt detector coupled to a non-scalar field in general depends not only on the detector's trajectory but also its orientation, as it can be sensitive to polarization of the filed. On the other hand, when a large ensemble of detectors is used to measure the transition rate, the ensemble as a whole may become insensitive to the field's degrees of polarization, if each detector in the ensemble is randomly oriented. The transition rate measured by the randomly-oriented ensemble as a whole can be represented by the simple model coupled to a scalar field (up to some detailed dependence on the explicit form of coupling). It is in this sense that the simple model with a scalar field is still relevant to realistic concern.
\section{Scalar field in a gravitational wave background}\label{sec:scalar field}
In this section, we solve the equation of motion of a real scalar field, i.e.\ the Klein-Gordon equation, in a gravitational wave background, which is otherwise a flat spacetime in the absence of gravitational waves. The gravitational waves are assumed to be weak enough so that the linearized theory, which neglects nonlinear gravitational wave effects, is adequate. In the linearized theory, any arbitrary gravitational wave can be decomposed into a linear superposition of plane waves. For simplicity, we only consider a monochromatic plane wave. To make the calculation simpler, we work in the transverse-traceless (TT) gauge.\footnote{For the issue that the TT gauge is always possible for any arbitrary gravitational wave, see Section 35.4 of \cite{Misner:1974qy} for more details.}
The action of a scalar field $\phi(x)$ in a curved spacetime is given by
\begin{equation}
S = \int \mathcal{L}[\phi(x)]\, d^4x
\end{equation}
with the Lagrangian density
\begin{equation}\label{L in curved spacetime}
\mathcal{L} =
\frac{1}{2}\sqrt{-g}
\left(- g^{\mu\nu} \nabla_\mu\phi \nabla_\nu\phi
- m^2\phi^2
- \xi R \phi^2
\right),
\end{equation}
where $m$ is the mass of the scalar particle, $R$ is the Ricci scalar, and $\xi$ is the coupling constant for the interaction between $\phi$ and $R$ (see e.g.\ \cite{Birrell:1982ix} for more details).
In order to obtain the Wightman function in a closed form, we consider the simplest case that $\phi$ is massless and does not couple with the curvature, i.e. $m=0$ and $\xi=0$. Variation with respect to $\phi$, i.e.\ $\delta S/\delta\phi=0$, then leads to the massless Klein-Gordon equation in curved spacetime,
\begin{equation}
\square \phi=0,
\end{equation}
where
\begin{eqnarray}\label{KG}
\square \phi
&:=& g^{\mu\nu}\nabla_\mu\nabla_\nu\phi
\equiv \frac{1}{\sqrt{-g}}\, \partial_\mu \left(\sqrt{-g}\, g^{\mu\nu} \partial_\nu \phi \right)\nonumber \\
&=&
g^{\mu\nu}\partial_\mu\partial_\nu\phi
+\partial_\mu g^{\mu\nu} \partial_\nu \phi
+\frac{1}{2}g^{\mu\nu} g^{\alpha\beta}\partial_\mu g_{\alpha\beta}\partial_\nu \phi.
\end{eqnarray}
The metric of the spacetime with a gravitational plane wave is given by
\begin{subequations}
\begin{eqnarray}
g_{\alpha\beta}(x)&=&\eta_{\alpha\beta}+h_{\alpha\beta}(x),\\
h_{\alpha\beta}(x)&=&A_{\alpha\beta}\,e^{-ik_\mu x^\mu}.
\end{eqnarray}
\end{subequations}
It follows that
\begin{subequations}
\begin{eqnarray}
g^{\alpha\beta} \partial_\mu g_{\alpha\beta}
&=& g^{\alpha\beta} \partial_\mu h_{\alpha\beta}= k_\mu g^{\alpha\beta}h_{\alpha\beta}= k_\mu h_{\alpha\beta}h^{\alpha\beta}\approx O(h^2),\\
\partial_\mu g^{\mu\nu}&=&\partial_\mu h^{\mu\nu}= k_\mu h^{\mu\nu} =0,
\end{eqnarray}
\end{subequations}
where we have applied the TT gauge to have $k_\mu h^{\mu\nu}=0$.
Consequently, up to the first order of $h$, the Klein-Gordon equation reads as
\begin{equation}
g^{\mu\nu}\partial_\mu\partial_\nu\phi=0.
\end{equation}
Given a gravitational plane wave prorogating along the $z$ direction, in the TT gauge, $h_{\mu\nu}$ takes the form
\begin{subequations}\label{h}
\begin{eqnarray}
h_{\mu\nu}(x)&=&\left(
\begin{array}{cccc}
0 & 0 & 0 & 0 \\
0 & h_{+} & h_{\times} & 0 \\
0 & h_{\times} & -h_{+} & 0 \\
0 & 0 & 0 & 0
\end{array}
\right),\\
h_{+/\times}(x)&=&A_{+/\times} \cos( kz- \omega t +\theta_{+/\times}),
\end{eqnarray}
\end{subequations}
where $\omega = \abs{k}$, and $h_+$ and $h_\times$ are the two independent modes of polarization with the amplitudes $A_+$ and $A_\times$ and phase shifts $\theta_+$ and $\theta_\times$, respectively.
Note that the metric given in \eqref{h} admits the Killing vector fields: $X=\partial_x$, $Y=\partial_y$, and $V=\partial_v=(\partial_t+\partial_z)/\sqrt{2}$. Accordingly, it is privileged to introduce the \emph{light-front} variables:
\begin{equation}\label{lc}
u=(t-z)/\sqrt{2},\quad v=(t+z)/\sqrt{2}.
\end{equation}
In terms of the coordinates $(u,v,x,y)$, the Klein-Gordon equation in the gravitational wave background reads as
\begin{eqnarray}\label{KG in LC}
&&\left[
-2 \frac{\partial^2}{\partial u\partial v}
+ \frac{\partial^2}{\partial x^2}
+ \frac{\partial^2}{\partial y^2}
- A_+\cos(w u+\theta_+)\left(\frac{\partial^2}{\partial x^2}-\frac{\partial^2}{\partial y^2}\right)
\right.\nonumber\\
&&
\left.
\qquad\qquad\qquad\qquad\qquad
\mbox{} - 2A_\times\cos(w u+\theta_\times) \frac{\partial^2}{\partial x\partial y}
\right] \phi
=0,
\end{eqnarray}
where we define the shorthand notation $w$ as
\begin{equation}\label{lc}
w:=\sqrt{2}\,\omega.
\end{equation}
As \eqref{KG in LC} is invariant under the translations along the $x$, $y$, and $v$ coordinates in accordance with the Killing vectors, we make the ansatz that the solution of $\phi$ takes the form
\begin{equation}\label{ansatz}
\phi(u,v,x,y) = e^{i(k_xx+k_yy -\omega v)} \chi(u),
\end{equation}
where $\chi(u)$ is a function of $u$ to be determined.
Substituting \eqref{ansatz} into \eqref{KG in LC}, we have
\begin{equation}
2i\,\omega \chi'(u)= \left[k_x^2 +k_y^2-A_+\cos(w u+\theta_+)(k_x^2- k_y^2) -2 A_\times\cos(w u+\theta_\times) k_x k_y\right] \chi(u).
\end{equation}
Integrating this equation then yields
\begin{equation}\label{chi}
\chi(u)
\propto
e^{-i k_u u } \,
e^{i k_u g_c(k_x,k_y)\frac{\sin{w u}}{w}}
e^{i k_u g_s(k_x,k_y)\frac{\cos{w u}}{w}},
\end{equation}
where $k_u:=(k_x^2+k_y^2)/{2\omega}$, or, equivalently, the mode frequency as a function of $k_u$, $k_x$, and $k_y$ is given by
\begin{equation}\label{omega k}
\omega\equiv
\omega_{k_u,k_x,k_y}
:=\frac{k_x^2+k_y^2}{2k_u},
\end{equation}
and
\begin{subequations}
\begin{eqnarray}
g_c(k_x,k_y)
&:=& \frac{1}{(k_x^2+k_y^2)}
\left[ A_+ (k_x^2-k_y^2)\cos\theta_+
+ 2A_\times k_xk_y \cos\theta_\times\right] , \\
g_s(k_x,k_y)
&:=& \frac{1}{(k_x^2+k_y^2)}
\left[ A_+ (k_x^2-k_y^2)\sin\theta_+
+ 2A_\times k_xk_y \sin\theta_\times\right].
\end{eqnarray}
\end{subequations}
For a given gravitational plane wave parametrized by $A_{+,\times}$, $\theta_{+,\times}$, and $w\equiv\sqrt{2}\,\omega$, we have obtained the eigenmode solutions of $\phi(u,v,x,y)$ parametrized by $k_x$, $k_y$, and $k_u$. Based on these eigenmodes, we can perform the field quantization in the next section.
\section{Light-front quantization}\label{sec:LF quantization}
The ordinary equal-time quantization scheme in curved spacetime requires a timelike Killing vector field to make sense of the notion of time \cite{Birrell:1982ix}. As the metrics given by \eqref{h} exhibits two spacelike and one lightlike Killing vector fields, but no timelike one, the ordinary scheme cannot apply. Instead, we adopt the \emph{light-front quantization} formalism, which is primarily used in the study of deep inelastic scattering in quantum chromodynamics (QCD) (see \cite{Burkardt:1995ct} for a review).
Under the light-front quantization scheme, the light-front direction in accordance with the lightlike Killing vector is treated as the direction of time. More precisely, we treat the light-front coordinate $v$ as the evolution parameter, and correspondingly define the frequency modes as eigenmodes of the Lie derivative via
\begin{equation}
\mathcal{L}_{V} \phi = -i \omega_k\phi, \quad \text{with}\, V=\partial_v,
\end{equation}
where $\omega_k$ is the frequency.
The Lagrangian density \eqref{L in curved spacetime} in the gravitational wave background given by \eqref{h} takes the form $\mathcal{L}=\sqrt{-g} \, (\partial_u \phi \partial_v \phi +\dots)$ in the coordinates $(u,v,x,y)$, where the part of ``$\dots$'' does not involve $\partial_v\phi$.
Consequently, the canonical momentum conjugate to $\phi$ is given by
\begin{equation}\label{pi}
\pi
:=\frac{\partial\mathcal{L}}{\partial(\partial_v\phi)}= \sqrt{-g}\,\partial_u\phi.
\end{equation}
The light-front quantization then demands the commutation relations given at equal light-front time $v$ as
\begin{equation}\label{CR 1}
[\phi(\bm{x},u,v),\pi(\bm{x}',u',v)]
=
\frac{i}{2} \sqrt{-g}\,\delta^2(\bm{x}-\bm{x}')\,\delta(u-u'),
\end{equation}
and
\begin{equation}\label{CR 2}
[\pi(\bm{x},u,v),\pi(\bm{x}',u',v)]=0.
\end{equation}
Note that \eqref{CR 1} implies the nonlocal commutation relation,
\begin{equation}
[\phi(\bm{x},u,v),\phi(\bm{x}',u',v)]
=
-\frac{i}{4} \sqrt{-g}\,\delta^2(\bm{x}-\bm{x}')\,\mathrm{sgn}(u-u'),
\end{equation}
which is a new feature that does not appear in the ordinary equal-time quantization scheme.\footnote{\label{foot:Dirac algorithm}Note that \eqref{pi} does not contain any time derivative $\partial_v$, and thus should be considered as a constraint equation. In other words, the phase space variables $\phi(x)$ and $\pi(x)$ at a given time $v$ are not completely independent of each other. In the presence of constraints, one has to apply the Dirac-Bergmann algorithm to arrive at a consistent Hamiltonian formalism, which then provides a proper starting point for the quantization procedure. Following the Dirac-Bergmann procedure, it is the Dirac bracket, instead of the Poisson bracket, that is to be promoted to the commutator $[\,\cdot,\cdot\,]$ for quantization. The difference between the Dirac bracket and the Poisson bracket gives rise to the extra factor $1/2$ in \eqref{CR 1}. See Appendix of \cite{Burkardt:1995ct} for more details.
In the light-front formalism, one has to address the additional issue arising from \emph{zero modes}, which correspond to the states that are independent of $u$ and have to be treated separately with special care. The resulting modified Dirac–Bergmann procedure could be very complicated, as the main difficulty lies in the fact that the constraint equation for zero modes is generally nonlinear. Fortunately, in our case as well as in many cases of free theories, the zero mode constraint does not get involved with the Hilbert space orthogonal to the zero modes, and thus can be simply projected out before the standard Dirac-Bergmann procedure is applied. For more about the zero-mode problem, see Appendix of \cite{Burkardt:1995ct} and the references therein.}
Here and hereafter, we use boldfaced letters to denote ``transverse'' vectors in shorthand: e.g., $\bm{x}:=(x,y)$, $\bm{k}:=(k_x,k_y)$, and $\bm{k}\cdot\bm{x}:= k_x x + k_yy$.
By virtue of \eqref{ansatz} and \eqref{chi}, the fields $\phi(x)$ can be cast in terms of the mode expansions as
\begin{equation}\label{phi of a and a dag}
\phi (u,v,\bm{x})
=\int \frac{d^3k}{(2\pi)^{3/2}} \,N_{k_u,\bm{k}}\,
\left( a_{k_u,\bm{k}} \,f_{k_u,\bm{k}}(\bm{x},u,v)
+ a^\dag_{k_u,\bm{k}} \,f^*_{k_u,\bm{k}}(\bm{x},u,v)\right),
\end{equation}
where
\begin{equation}
f_{k_u,\bm{k}} (u,v,\bm{x}):=e^{ i(\bm{k}\cdot \bm{x}-k_uu -\omega_{k_u,\bm{k}}v)} \chi(u)
\end{equation}
with $\chi(u)$ given by \eqref{chi}, and where $N_{k_u,\bm{k}}$ are normalization factors to be determined later.
The conjugate momentum field $\pi$ given by \eqref{pi} then reads as
\begin{eqnarray}
\pi (u,v,\bm{x})
&=&\sqrt{-g} \int \frac{d^3k}{(2\pi)^{3/2}}\, N_{k_u,\bm{k}}\,(-ik_u)\,g_{\bm k}(u)\\ \nonumber
&&\qquad\qquad\qquad \times\left( a_{k_u,\bm{k}} \,f_{k_u,\bm{k}}(\bm{x},u,v)
- a^\dag_{k_u,\bm{k}} \,f^*_{k_u,\bm{k}}(\bm{x},u,v)\right),
\end{eqnarray}
where
\begin{equation}
g_{\bm k}(u):=1- g_c(k_x,k_y)\cos(w u)+g_s(k_x,k_y)\sin(w u).
\end{equation}
Note that $k_u\geq0$ according to \eqref{omega k},\footnote{We have adopted the convention that $\omega_{k_u,\bm{k}}\geq0$. That is, positive-frequency modes (i.e., $\propto e^{i\omega_{k_u,\bm{k}}v}$) are associated with creation operators, while negative-frequency (i.e., $\propto e^{-i\omega_{k_u,\bm{k}}v}$) with annihilation operators in \eqref{phi of a and a dag}.} and the notation $\int d^3k$ is a shorthand for $\int_{-\infty}^{\infty} d^2k \int_{0}^{\infty} dk_u \equiv \int_{-\infty}^{\infty} dk_x \int_{-\infty}^{\infty} dk_y \int_{0}^{\infty} dk_u$.
By prescribing the commutation relations for $a_{k_u,\bm{k}}$ and $a^\dag_{k_u,\bm{k}}$ as
\begin{subequations}\label{a and a dag}
\begin{eqnarray}
[a_{k_u,\bm{k}},a^\dag_{k'_u,\bm{k}'}] &=& \delta^2(\bm{k}-\bm{k}')\delta(k_u-k'_u),\\
{[}a_{k_u,\bm{k}},a_{k'_u,\bm{k}'}{]} &=& [a^\dag_{k_u,\bm{k}},a^\dag_{k'_u,\bm{k}'}] = 0,
\end{eqnarray}
\end{subequations}
and the normalization factor as
\begin{equation}
N_{k_u,\bm{k}}=\frac{1}{\sqrt{2k_u}},
\end{equation}
the commutation relations \eqref{CR 1} and \eqref{CR 2} can be realized.
To show this, we first calculate
\begin{eqnarray}
&&[\phi(\bm{x},u,v),\pi(\bm{x}',u',v)]\nonumber\\
&=&\sqrt{-g}
\int \frac{d^3k}{(2\pi)^{3/2}} \int \frac{d^3k'}{(2\pi)^{3/2}} \,\frac{i g_{\bm k}(u')k_u'}{\sqrt{4k_uk'_u}} \,
\bigg\{ [a_{k_u,\bm{k}}, a^\dag_{k'_u,\bm{k}'} ]\,f_{k_u,\bm{k}}(\bm{x},u,v) f^*_{k'_u,\bm{k}'}(\bm{x}',u',v) \nonumber\\
&&\mbox{ } +
[a_{\bm{k}',k_u},a^\dag_{\bm{k},k_u } ]\,f_{k'_u,\bm{k}'}(\bm{x'},u',v) f^*_{k_u,\bm{k}}(\bm{x},u,v) \bigg\}\nonumber\\
&=&
\frac{i}{4}\sqrt{-g}\int_{-\infty}^{\infty} \frac{d^2k}{(2\pi)^2} \,g_{\bm k}(u')
\left( e^{ i\bm{k}\cdot (\bm{x}-\bm{x}')}+ e^{-i\bm{k}\cdot (\bm{x}-\bm{x}')} \right)
\int_{-\infty}^{\infty} \frac{dk_u}{2\pi} e^{-ik_u\lambda(u)}
\nonumber\\
&=&
\frac{i}{4}\sqrt{-g}\, \int_{-\infty}^{\infty} \frac{d^2k}{(2\pi)^2} \,\delta(\lambda(u))\, g_{\bm k}(u')
\left( e^{ i\bm{k}\cdot (\bm{x}-\bm{x}')}+ e^{-i\bm{k}\cdot (\bm{x}-\bm{x}')} \right),
\end{eqnarray}
where
\begin{equation}
\lambda(u) := u-u'+ w^{-1}\left[ g_c(k_x,k_y)(\sin{w u}-\sin{w u'})+ g_s(k_x,k_y)(\cos{w u}-\cos{w u'})\right].
\end{equation}
Since $\lambda(u)$ has a single root at $u=u'$, the identity
\begin{equation}
\delta(\lambda(u)) = \frac{\delta(u-u')}{\abs{\lambda'(u')}} =\frac{\delta(u-u')}{g_{\bm k}(u')}
\end{equation}
can be used to obtain
\begin{eqnarray}
[\phi(\bm{x},u,v),\pi(\bm{x}',u',v)]
&=& \frac{i}{4} \sqrt{-g}\,\delta(u-u') \int_{-\infty}^{\infty} \frac{d^2k}{(2\pi)^2} \,
\left( e^{ i\bm{k}\cdot (\bm{x}-\bm{x}')}+ e^{-i\bm{k}\cdot (\bm{x}-\bm{x}')} \right),\nonumber\\
&=& \frac{i}{2} \sqrt{-g}\,\delta^2(\bm{x}-\bm{x}')\,\delta(u-u'),
\end{eqnarray}
in agreement with \eqref{CR 1}.
Meanwhile, it is can be readily shown that \eqref{a and a dag} leads to \eqref{CR 2}.
In summary, the mode expansion of $\phi(x)\equiv\phi(u,v,\bm{x})$ in terms of creation and annihilation operators is given by
\begin{equation}\label{phi}
\phi(x)
=\int \frac{dk^3}{(2\pi)^{3/2}} \,\frac{1}{\sqrt{2k_u}}\,
\left( a_{k_u,\bm{k}} \,f_{k_u,\bm{k}}(u,v,\bm{x})
+ a^\dag_{k_u,\bm{k}} \,f^*_{k_u,\bm{k}}(u,v,\bm{x})\right).
\end{equation}
The Hilbert space is spanned by the Fock states in the form
\begin{eqnarray}
&&\ket{{}^1n_{k_{u1},\bm{k}_1},{}^2n_{k_{u2},\bm{k}_2},\dots,{}^jn_{k_{uj},\bm{k}_j}} \nonumber\\
&:=&({}^1n!\,{}^2n!\dots{}^jn!)^{-1/2}
(a^\dag_{k_{u1},\bm{k}_1})^{{}^1n}(a^\dag_{k_{u2},\bm{k}_2})^{{}^2n}\dots(a^\dag_{k_{uj},\bm{k}_j})^{{}^jn}
\ket{0},
\end{eqnarray}
which is a many-particle state with ${}^1n$ particles in the mode $(k_{u1},\bm{k}_1)$, ${}^2n$ particles in the mode $(k_{u2},\bm{k}_2)$, and so on. The no-particle state $\ket{0}$ is the vacuum, which is annihilated by all annihilation operators, i.e.,
\begin{equation}
a_{k_u,\bm{k}}\ket{0}=0,\quad \text{for all}\ (k_u,\bm{k}).
\end{equation}
\section{The Wightman function}\label{sec:Wightman function}
As we have successfully quantized the scalar field $\phi(x)$ in a monochromatic gravitational wave background, we are now ready to calculate the corresponding Wightman function.
According to \eqref{phi}, the Wightman function $D^+(x,x')$ takes the form
\begin{eqnarray}\label{Wightman}
D^+(x,x')
&:=& \bra{0}\phi(x)\phi(x')\ket{0} \nonumber\\
&=&
\int \frac{dk^3}{(2\pi)^{3/2}}
\int \frac{dk'^3}{(2\pi)^{3/2}} \,
\frac{1}{\sqrt{4k_u k'_u}} \,
\bra{0} a_{k_u,\bm{k}} a^\dag_{k'_u,\bm{k}'} \ket{0}\,
f_{k_u,\bm{k}}f^*_{k_u,\bm{k}}
\nonumber\\
&=&
\int \frac{dk^3}{(2\pi)^3}\frac{1}{2k_u}
e^{ i\bm{k}\cdot (\bm{x}-\bm{x}')} e^{-ik_u(u-u')} e^{-i\omega_{k_u,\bm{k}}(v-v')}\nonumber\\
&&\mbox{}\times
e^{ik_u g_c(\bm{k})(\sin{w u}-\sin{w u'})/w}\,
e^{ik_u g_s(\bm{k})(\cos{w u}-\cos{w u'})/w}.
\end{eqnarray}
This expression is complicated and does not have a closed form. Fortunately, it can be greatly simplified if we consider the two limiting situations: the \emph{long-wavelength limit} and the \emph{short-wavelength limit}.
The Unruh-DeWitt detector naturally provides two characteristic timescales. The first is ${\sim}1/\Delta E$, which characterizes the detector's response time for the two-level transition. The second is ${\sim}\Delta$, which characterizes the detector's switch-on period as shown in \figref{fig:switching function}.
As will be seen shortly, we will apply a particular form of $i\epsilon$-regularization to compute the Wightman function instead of specifying the switching function $\chi(\tau)$ with a finite switch-on period $\Delta$. In reality, however, a detector is always switched on only for a finite period $\Delta$, which lays down a timescale to be compared with the period $1/\omega$ of the gravitational wave. Therefore, we should keep in mind that the resulting transition rate of the Unruh-DeWitt detector computed from the $i\epsilon$-regularized Wightman function is legitimate only if the measurement performed at time $\tau$ is well within the switch-on period.\footnote{Accordingly, the condition $1/\Delta E\ll\Delta$ must be satisfied in order to yield a sensible result in agreement with the transition rate computed from the regularized Wightman function.}
The background gravitational wave is said to be in the long-wavelength limit, if the wavelength of the gravitational wave is so large that, within the whole switch-on period, the detector does not see any gravitational wave modulation, but effectively only sees a persisting gravitational wave amplitude. See \figref{fig:long-wavelength} for illustration.
More precisely, during the switch-on period, if the detector moves from $x^\mu(\tau)=(t,\vec{x})\equiv(t,x,y,z)$ to $x^\mu(\tau+\Delta)=(t',\vec{x}')\equiv(t',x',y',z')$, we have
\begin{equation}
t' \approx t + \frac{1}{1-v^2}\Delta,
\qquad
z' \approx z + \frac{v_z}{1-v^2}\Delta,
\end{equation}
where $\vec{v}=(v_x,v_y,v_z)$ is the detector's moving velocity (averaged over the switch-on period). The phase difference of the gravitational wave experienced by the detector during this period is given by
\begin{equation}
\delta\phi = (kz'-\omega t') - (kz -\omega t)
\approx \frac{\omega\Delta}{\sqrt{1-v^2}}(v_z-1),
\end{equation}
for a gravitational plane wave propagating in the $z$ direction ($k=\omega>0$). The precise condition for the long-wavelength limit is $\abs{\delta\phi}\ll1$, or equivalently,
\begin{equation}\label{long-wavelength condition}
\frac{1}{\Delta E} \ll \Delta \ll \frac{\sqrt{1-v^2}}{\abs{v_z-1}}\,\frac{1}{\omega}.
\end{equation}
It should be noted that, in the case of a detector moving in the $z$ direction at an extremely fast speed close to the speed of light, i.e., $v_z\approx1$, the condition \eqref{long-wavelength condition} is always satisfied even if the gravitational wavelength $1/\omega$ is small. The asymptotic behavior of the constant-accelerating trajectory given by \eqref{x a} is a typical example (see \figref{fig:long-wavelength}). At the opposite extreme, if the detector moves in the negative $z$ direction at a speed close to the speed of light, i.e., $v_z\approx-1$, the condition \eqref{long-wavelength condition} cannot be satisfied, no matter how long $1/\omega$ is.
If we neglect any corrections equal to or higher than the order of $O(\omega\Delta)$, we can simply take the formal limit $\omega\rightarrow0$ for the result of the long-wavelength limit.
In this formal limit, wherever the portion of the trajectory within the switch-on period is located in the spacetime, the phase of the gravitational wave upon this portion is to be treated as the same as that upon the hypersurface $u=0$. (See the right panel of \figref{fig:long-wavelength}, imaging that the wavelength becomes infinity.)
Therefore, the persisting amplitude the detector experiences during the switch-on period is given by $\left.h_{+/\times}(x)\right|_{u=0}\equiv\mathcal{A}_{+/\times}$ for $+$ and $\times$ modes, respectively, which is defined as
\begin{equation}\label{cal A}
\mathcal{A}_+:= A_+\cos \theta_+,\quad \mathcal{A}_\times:= A_\times\cos\theta_\times.
\end{equation}
In a real \emph{physical} setting (contrary to the \emph{formal} limit $\omega\rightarrow0$), if the long-wavelength condition \eqref{long-wavelength condition} satisfied, $\mathcal{A}_+$ and $\mathcal{A}_\times$ used in the formal limit are to be understood as representing the persisting amplitudes experienced by the detector during the switch-on period.
\begin{figure}
\centering
\begin{minipage}[b]{0.4\textwidth}
\includegraphics[width=\textwidth]{GW_background_a.pdf}
\end{minipage}
\hspace{0.7cm}
\begin{minipage}[b]{0.4\textwidth}
\includegraphics[width=\textwidth]{GW_background_b.pdf}
\end{minipage}
\caption{[\textit{Left}] Modulation of the phase of a gravitational plane wave propagating in the $z$ direction. The axes of $(t,z)$ and $(u,v)$ are both shown for reference. [\textit{Right}] A case of a very long wavelength is drawn to illustrate the long-wavelength condition \eqref{long-wavelength condition}. A world line given by \eqref{x a} (with $t_0=0$) is depicted as an example of the detector's trajectory, and the segment within the switch-on period, $\tau_0\lesssim\tau\lesssim\tau_0+\Delta$, is highlighted in a deeper shade (with $\tau_0$ chosen to be a certain positive value). The wavelength is so large that the phase of the gravitational wave is almost the same over the whole switch-on segment. Furthermore, if $\tau_0$ is chosen to be positive and very large, the switch-on segment will asymptote to the line of $u=0$, and the corresponding velocity will asymptote to $v_z\approx1$. In this asymptotic situation, the long-wavelength condition is satisfied even if $1/\omega$ is small.}
\label{fig:long-wavelength}
\end{figure}
On the other hand, the background gravitational wave is said to be in the short-wavelength limit, if the condition
\begin{equation}\label{short-wavelength condition}
\frac{\sqrt{1-v^2}}{\abs{v_z-1}}\,\frac{1}{\omega} \ll \frac{1}{\Delta E} \ll \Delta
\end{equation}
is satisfied. That is, the frequency of gravitational-wave modulation experienced by the moving detector is much higher than the frequency of the two-level energy difference. If we neglect any corrections equal to or higher than the order of $O(\Delta E/\omega)$, we can simply take the formal limit $\omega\rightarrow\infty$ for the result of the short-wavelength limit.
Although we cannot obtain a closed-form expression for the response of the Unruh-DeWitt detector in the case of arbitrary wavelengths, we can still learn a great deal from the two opposite limits.
\subsection{Long-wavelength limit}
We first consider the long-wavelength limit conditioned by \eqref{long-wavelength condition}. Neglecting any corrections in or higher than the order of $O(\omega\Delta)$, we take the formal limit $w\equiv\sqrt{2}\,\omega\rightarrow 0$ upon \eqref{Wightman}:
\begin{eqnarray}\label{L limit}
D^+_{\mathrm{lw}}(x-x')
&\equiv& \lim_{w\rightarrow0}D^+(x,x') \nonumber\\
&=&
\int \frac{d^3k}{(2\pi)^3}\frac{1}{2k_u}
e^{ i\bm{k}\cdot (\bm{x}-\bm{x}')}e^{-ik_u(1-g_c(k_x,k_y))(u-u')}e^{ -i\omega_{k_u,\bm{k}} (v-v')}.
\end{eqnarray}
By performing the change of variables,
\begin{subequations}
\begin{eqnarray}
k_u' &\equiv& k_u'(k_u,k_x,k_y)
=k_u (1-g_c(k_x,k_y)),\\
\omega_{k_u',\bm{k}}
&\equiv& \omega(k_u',k_x,k_y)
= \omega_{k_u,\bm{k}}
= \frac{k_x^2+k_y^2}{2k'_u}\left[1-g_c(k_x,k_y)\right]
\end{eqnarray}
\end{subequations}
\eqref{L limit} is simplified into the form
\begin{equation}\label{Wightman lw}
D^+_{\mathrm{lw}}(x-x')
=\int \frac{d^2k\,dk_u'}{(2\pi)^3}\frac{1}{2k_u'}
e^{ i\bm{k}\cdot (\bm{x}-\bm{x}')}e^{-ik_u'(u-u')}e^{ -i\omega_{k_u',\bm{k}} (v-v')},
\end{equation}
where we have used $dk_u=(1-g_c(k_x,k_y))^{-1}dk_u'$.
The Wightman function is in fact not a genuine function but a distribution. When substituted into \eqref{response function}, it yields an unambiguous result for the response function $F(\Delta E)$ as long as the switching function $\chi(\tau)$ is smooth enough and of compact support. Without specifying $\chi(\tau)$, however, the expression \eqref{Wightman lw} by itself is ambiguous and requires a proper regularization procedure to yield a sensible result in agreement with the condition of causality.\footnote{Recall the comments after \eqref{F dot}.} Here, we prescribe the particular form of $i\epsilon$-regularization (with an infinitesimal parameter $\epsilon>0$) as follows:
\begin{equation}\label{Wightman lw regularized}
D^+_{\mathrm{lw}}(x-x')
=\int \frac{d^2k\,dk_u'}{(2\pi)^3}\frac{1}{2k_u'}
e^{ i\bm{k}\cdot (\bm{x}-\bm{x}')}e^{-ik_u'(u-u'-i\epsilon)}e^{ -i\omega_{k_u',\bm{k}} (v-v'-i\epsilon)}.
\end{equation}
This $i\epsilon$-regularization conforms with the condition of causality, as we will see shortly that it reduces to the standard $i\epsilon$-regularization when the gravitational wave amplitude is turned off. This particular regularization can also be understood as providing a large-value cutoff for both $\omega_{k_\mu,\bm{k}}$ and $k_u$.
Applying the Gaussian integral formula
\begin{equation}
\int_{-\infty}^{\infty} dx e^{-ax^2+bx+c}=\sqrt{\frac{\pi}{a}}\, e^{\frac{b^2}{4a}+c}
\end{equation}
to the integration over $k_x$ and $k_y$, we have
\begin{eqnarray}\label{Iku}
\mathcal{I}(k'_u)
&:=&\int \frac{dk_x}{2\pi} \int\frac{dk_y}{2\pi}\,
e^{ i k_x \Delta x}\, e^{ i k_y \Delta y}\, e^{ -i\omega_{k'} (v-v'-i\epsilon)} \nonumber\\
&=&
\frac{k'_u}{2\pi i}
\frac{1}{\sqrt{1-\mathcal{A}^2}}
\frac{1}{(v-v'-i\epsilon)}\,
e^{i k'_u\, R(\Delta x,\Delta y)\,/2(v-v'-i\epsilon)},
\end{eqnarray}
where $\Delta x:=x-x'$, $\Delta y:=y-y'$, and $R(\Delta x,\Delta y)$ is defined as
\begin{equation}
R(\Delta x,\Delta y)
:=
\frac{(1+\mathcal{A}_+) \Delta x^2 + 2 \mathcal{A}_\times\Delta x\Delta y+ (1-\mathcal{A}_+)\Delta y^2}{1-\mathcal{A}^2},
\end{equation}
with $\mathcal{A}_+$ and $\mathcal{A}_\times$ defined in \eqref{cal A} and $\mathcal{A}^2$ defined as
\begin{equation}
\mathcal{A}^2:= \mathcal{A}_+^2+\mathcal{A}_\times^2.
\end{equation}
It then follows from \eqref{Wightman lw regularized} and \eqref{Iku} that
\begin{eqnarray}
D^+_{\mathrm{lw}}(x-x')
&=&
-i\int_{0}^{\infty}\frac{dk'_u}{2\pi}\, \frac{1}{2k'_u} \mathcal{I}(k'_u)\, e^{-i k'_u (u-u'-i \epsilon)}\nonumber\\
&=&
-\frac{1}{8\pi^2}\frac{1}{\sqrt{1-\mathcal{A}^2}}
\frac{1}{(u-u'-i \epsilon)(v-v'-i \epsilon)- R(\Delta x,\Delta y)/2}.
\end{eqnarray}
Finally, the result in the coordinates $(t,x,y,z)$ takes the form
\begin{equation}\label{D lw}
D^+_{\mathrm{lw}}(x-x')
= -\frac{1}{4\pi^2}
\frac{1}{\sqrt{1-\mathcal{A}^2}}\,
\frac{1}{(t-t'-i\epsilon)^2 -(z-z')^2-R(\Delta x,\Delta y) }.
\end{equation}
When $\mathcal{A}_+=\mathcal{A}_\times=0$, \eqref{D lw} reduces to the ordinary Wightman function in the Minkowski spacetime with the standard $i\epsilon$-regularization:
\begin{equation}\label{D standard}
D_\mathrm{M}^+(x-x')
= -\frac{1}{4\pi^2}\,
\frac{1}{(t-t'-i\epsilon)^2 -\abs{\vec{x}-\vec{x}'}^2},
\end{equation}
where $\vec{x}\equiv(x,y,z)$.
\subsection{Short-wavelength limit}
Next, we consider the short-wavelength limit conditioned by \eqref{short-wavelength condition}. Neglecting any corrections in or higher than the order of $O(\Delta E/\omega)$, we take the formal limit $w\equiv\sqrt{2}\,\omega\rightarrow\infty$ upon \eqref{Wightman}. The result simply reduces to
\begin{equation}
D^+_{\mathrm{sw}}(x-x')\equiv \lim_{w\rightarrow \infty} D^+(x,x')
= \int \frac{d^3k}{(2\pi)^3}\frac{1}{2k_u}
e^{ i\bm{k}\cdot (\bm{x}-\bm{x}')}e^{-ik_u(u-u')-i\epsilon}e^{ -i\omega_{k_u,\bm{k}} (v-v-i\epsilon)},
\end{equation}
where the $i\epsilon$-regularization is again prescribed.
The dependence of the gravitational wave amplitude disappears, and $D_{\mathrm{sw}}(x-x')$ in the coordinates $(t,x,y,z)$ reads as
\begin{equation}\label{D sw}
D^+_{\mathrm{sw}}(x,x')= -\frac{1}{4\pi^2}\,
\frac{1}{ (t-t'-i\epsilon)^2
-\abs{\vec{x}-\vec{x}'}^2 },
\end{equation}
which formally is identical to the ordinary Wightman function in the Minkowski spacetime as given in \eqref{D standard}.
Although \eqref{D sw} apparently looks the same as that in the Minkowski spacetime, the physics it implies can be quite different from the latter. For one thing, the geodesic equation in the short-wavelength limit is different from that in the Minkowski spacetime (see \appref{app:geodesic eq}).
Once the detector's trajectory is given and the Wightman function is known, we are ready to compute the transition rate by \eqref{F dot}. First, we consider the case that the detector follows a free-falling (i.e.\ geodesic) trajectory, and then the case that the detector moves with a constant acceleration in the $z$ direction.
\section{Free-falling trajectory}\label{sec:free-falling}
In this section, we study the response of an Unruh-DeWitt detector falling freely in a gravitational wave background.
The free-falling trajectory is given by the geodesic equation, which is solved explicitly in \appref{app:geodesic eq}.
\subsection{Long-wavelength limit}
In the long-wavelength limit $\omega\rightarrow0$, according to \eqref{const velocity sol}, a free-falling trajectory takes the form
\begin{equation}\label{x freefall}
x^\mu(\tau) = U^\mu\tau + x^\mu_0,
\end{equation}
where the 4-velocity
\begin{equation}
U^\mu=(u^t,u^x,u^y,u^z)
\end{equation}
is given by four constants, $u^t$, $u^x$, $u^y$, and $u^z$, subject to the constraint \eqref{velocity normalization 2}, and $x^\mu_0$ are displacement parameters.
Substituting \eqref{x freefall} into \eqref{D lw} yields
\begin{eqnarray}\label{D in freefall}
D^+_{\mathrm{lw}}(\tau,\tau')
&=& -\frac{1}{4\pi^2}\frac{1}{\sqrt{1-\mathcal{A}^2}}
\frac{1}{(u_t \Delta \tau -i\epsilon)^2-(u_z \Delta\tau)^2+\frac{1-u_t^2 +u_z^2}{(1-\mathcal{A}^2)}\Delta\tau^2 - R(\Delta x,\Delta y)} , \nonumber\\
&=& -\frac{1}{4\pi^2}\frac{1}{\sqrt{1-\mathcal{A}^2}}
\frac{1}{\Delta\tau^2((u^t)^2-(u^z)^2-i\epsilon')-\frac{\Delta\tau^2}{1-\mathcal{A}^2}((u^t)^2-(u^z)^2-1)}\nonumber\\
&=& -\frac{1}{4\pi^2}\frac{1-\mathcal{A}^2}{\sqrt{1-\mathcal{A}^2}}
\frac{1}{\Delta\tau^2[(1-\mathcal{A}^2)((u^t)^2-(u^z)^2-i\epsilon')-((u^t)^2-(u^z)^2)+1]}\nonumber\\
&=& -\frac{1}{4\pi^2}\frac{1-\mathcal{A}^2}{\sqrt{1-\mathcal{A}^2}}
\frac{1}{\Delta\tau^2[1-\mathcal{A}^2((u^t)^2-(u^z)^2)-i\epsilon'']}\nonumber\\
&=& -\frac{1}{4\pi^2}\frac{\sqrt{1-\mathcal{A}^2}}{1-\mathcal{A}^2((u^t)^2-(u^z)^2)}
\frac{1}{(\Delta\tau^2-i\epsilon''')^2},
\end{eqnarray}
where $\epsilon$, $\epsilon'$, $\epsilon''$, and $\epsilon'''$ are infinitesimal positive numbers (rescaled differently).
The fact that $D^+_{\mathrm{lw}}(\tau,\tau')$ depends only on $\Delta\tau\equiv\tau-\tau'$ suggests that a free-falling Unruh-DeWitt detector is in equilibrium with $\phi$ in the long-wavelength limit.
Substituting \eqref{D in freefall} into \eqref{F dot} then yields the equilibrium transition rate
\begin{equation}
\dot{F}(\Delta E)
= - \frac{1}{4\pi^2} \frac{\sqrt{1-\mathcal{A}^2}}{1-\mathcal{A}^2((u^t)^2-(u^z)^2)}
\int_{-\infty}^{\infty} d\Delta \tau\, \frac{e^{-i\Delta E \Delta\tau}}{(\Delta\tau-i\epsilon)^2}.
\end{equation}
If $\Delta \tau$ is considered to be a complex number, the transition rate can be calculated by a contour integral. The integrand has a pole of order 2 at $\Delta=i\epsilon$.
For $\Delta E>0$, the transition rate can be calculated by a contour integral along an infinite semicircle contour on the lower half of the $\Delta\tau$ plane. As the contour does not enclose the pole, the contour integral turns out to be zero.
For $\Delta E<0$, the integration can be calculated by a contour integral along an infinite semicircle contour on the upper half of the $\Delta\tau$ plane. The residue theorem applied to the pole at $\Delta\tau=i\epsilon$ gives
\begin{subequations}
\begin{eqnarray}
\dot{F}(\Delta E)
&=&- \frac{\Delta E}{2\pi}\frac{\sqrt{1-\mathcal{A}^2}}{1-\mathcal{A}^2((u^t)^2-(u^z)^2)}\\
&\equiv&- \frac{\Delta E}{2\pi}\frac{\sqrt{1-\mathcal{A}^2}}{1-\mathcal{A}^2(1+(u^x)^2+(u^y)^2 +\mathcal{A}_+\left((u^x)^2-(u^y)^2\right) +2\mathcal{A}_\times u^xu^y)},
\end{eqnarray}
\end{subequations}
where we have used \eqref{velocity normalization 2} to recast $(u^t)^2-(u^z)^2$ in terms of $u^x$ and $u^y$.\footnote{Under a rotation by $\theta$ around the $z$ axis, $u^{x}$, $u^y$, and $\mathcal{A}_{+/\times}$ transform as
\begin{eqnarray*}
u^{x\prime} &=& \cos\theta\, u^x + \sin\theta\, u^y, \qquad
u^{y\prime} = -\sin\theta\, u^x + \cos\theta\, u^y,\\
\mathcal{A}'_+ &=& \cos2\theta\mathcal{A}_+ + \sin2\theta\mathcal{A}_\times, \qquad
\mathcal{A}'_\times = -\sin2\theta\mathcal{A}_+ + \cos2\theta\mathcal{A}_\times.
\end{eqnarray*}
Note that $\mathcal{A}^2$, $(u^x)^2+(u^y)^2$, and $\mathcal{A}_+\left((u^x)^2-(u^y)^2\right) +2\mathcal{A}_\times u^xu^y$ are all invariant under this transformation.}
In summary, we have
\begin{subequations}\label{F v}
\begin{eqnarray}
\dot{F}(\Delta E)
&=& 0, \quad \text{for}\ \Delta E > 0, \\
&=& - \frac{\Delta E}{2\pi} \frac{\sqrt{1-\mathcal{A}^2}}{1-\mathcal{A}^2((u^t)^2-(u^z)^2)}, \quad \text{for}\ \Delta E < 0.
\end{eqnarray}
\end{subequations}
As expected, in the limit that the gravitational wave amplitude goes to zero, i.e., $\mathcal{A}^2\rightarrow0$, \eqref{F v} reduces to the ordinary result in the Minkowski spacetime:
\begin{equation}\label{F v Minkowski}
\dot{F}(\Delta E) \mathop{\longrightarrow}\limits_{\mathcal{A}^2\rightarrow0}
-\frac{\Delta E}{2\pi}\Theta(-\Delta E).
\end{equation}
Compared to the Minkowskian result, the transition rate \eqref{F v} is modified by an overall proportional factor that depends on the amplitude of the gravitational wave and the detector's velocity.
When $(u^t)^2 - (u^z)^2=1$ or, equivalently, $u^x=u^y=0$, \eqref{F v} in the case of $\Delta E < 0$ yields the maximum value:
\begin{equation}\label{F dot max}
\max_{U^\mu}\dot{F}(\Delta E)=- \frac{\Delta E}{2\pi} \frac{1}{\sqrt{1-\mathcal{A}^2}}.
\end{equation}
On the other hand, when $(u^t)^2 - (u^z)^2=0$ or, equivalently, $u^x$ and $u^y$ satisfy
\begin{equation}
1+(u^x)^2+(u^y)^2 +\mathcal{A}_+\left((u^x)^2-(u^y)^2\right) +2\mathcal{A}_\times u^xu^y=0,
\end{equation}
\eqref{F v} in the case of $\Delta E < 0$ yields the minimum value:
\begin{equation}
\min_{U^\mu}\dot{F}(\Delta E)=- \frac{\Delta E}{2\pi} \sqrt{1-\mathcal{A}^2}.
\end{equation}
It is instructive to compare \eqref{F v} with the case in flat spacetime with a compact
dimension. In a flat spacetime where the $z$ direction is compactified with a finite length $L$, the transition rate of the Unruh-DeWitt detector moving with the 4-velocity $U^\mu=(u^t,u^x,u^y,u^z)$ is given by (see \cite{Chiou:2016exd})
\begin{subequations}\label{rate for constant velocity}
\begin{eqnarray}
\label{rate for constant velocity a}
\dot{F}_L(\Delta E) &=& 0, \quad \text{for}\ \Delta E >0, \\
\label{rate for constant velocity b}
&=&
-\frac{\Delta E}{2\pi}
-\frac{i}{4\pi L u^t}
\ln
\left(
\frac{1-e^{i\frac{\Delta E L}{u^t+u^z}}}
{1-e^{-i\frac{\Delta E L}{u^t+u^z}}}
\,
\frac{1-e^{i\frac{\Delta E L}{u^t-u^z}}}
{1-e^{-i\frac{\Delta E L}{u^t-u^z}}}
\right),
\quad \text{for}\ \Delta E <0.
\end{eqnarray}
\end{subequations}
The correction due to the compact length $L$ in \eqref{rate for constant velocity} is scaled as $O(L^{-1})$, which vanishes in the formal limit $L\rightarrow\infty$. By contrast, the leading-order correction due to the gravitational wave is scaled as $O((1/\omega)^0)$ as shown in \eqref{F v}, which survives the formal limit $\omega\rightarrow0$, i.e., as the gravitational wavelength goes to infinity.
In both cases, the transition rates are different from the ordinary result in the Minkowski spacetime, essentially because the mode expansion of the quantum field $\phi$ is altered in the presence of the compact dimension or the gravitational wave. However, the gravitational wave effect is more involved and cannot be explained out simply by saying that the gravitational wave imposes a large length scale of the wavelength $1/\omega$ as well as the compact dimension does of the finite length $L$.
It is rather surprising that a detector that apparently has no spatial extent can sense the presence of a gravitational wave, as sensing tidal force requires a certain spatial extent. One might try to argue that the Unruh-DeWitt detector has an intrinsic energy scale $\Delta E$ and therefore, according to the energy-time uncertainty principle, exhibits a temporal scale $\sim1/\Delta E$, which in turn gives rise to a spatial scale $\sim v/\Delta E$, where $v<1$ is the velocity of the detector. However, even if the detector has a finite extent $\delta\ell$ (and even if $\delta\ell$ is much larger than ${\sim}1/\Delta E$ for whatever reason),\footnote{In fact, particle detectors with finite spatial extent have been discussed in the literature \cite{Grove:1983rp}.} the tidal force produced by a gravitational wave over a spatial separation of $\delta\ell$ is proportional to $(\delta\ell)\ddot{h}_{+/\times}\sim \mathcal{A}_{+/\times}\omega^2\delta\ell$, which vanishes in the limit $\omega\rightarrow0$. Since the correction in \eqref{F v} survives the formal limit $\omega\rightarrow0$, this gravitational wave effect on a quantum system is qualitatively different from that on a classical mechanical system, and cannot be understood in terms of gravitational wave tidal force. This is a genuine quantum effect that has no classical analogue.
\subsection{Short-wavelength limit}
The solution to the geodesic equation in a gravitational wave background is given by \eqref{geodesic eq sol}, which in general is very complicated.
The solution takes a simple form in the long-wavelength limit as given by \eqref{const velocity sol}, but it remains complicated under the short-wavelength condition \eqref{short-wavelength condition}.
Therefore, even though the Wightman function in the limit $\omega\rightarrow\infty$ as given in \eqref{D sw} apparently is identical to the ordinary Wightman function in the Minkowski spacetime, a free-falling Unruh-DeWitt detector in a gravitational wave background in general is \emph{not} in equilibrium with $\phi$, contrary to that in the Minkowski spacetime.
However, we do have a special geodesic solution given by \eqref{trivial sol}, which is simple and corresponds to a free-falling trajectory moving in the propagation direction of the gravitational wave, i.e.,
\begin{equation}\label{trivial sol x}
x^\mu(\tau) = (u^t\tau,0,0,u^z\tau) + x^\mu_0.
\end{equation}
Substituting this trajectory into $D^+_{\mathrm{sw}}(x,x')$ in \eqref{D sw}, we see that $D^+_{\mathrm{sw}}(\tau,\tau')$ depends only on $\Delta\tau\equiv\tau-\tau'$. Therefore, the Unruh-DeWitt detector that freely falls along the trajectory \eqref{trivial sol x} is in equilibrium with $\phi$ in the short-wavelength limit $\omega\rightarrow\infty$.
Since $D^+_{\mathrm{sw}}(x,x')$ is formally the same as the ordinary Wightman function in the Minkowski spacetime, the transition rate $\dot{F}(\Delta)$ along \eqref{trivial sol x} is the same as the ordinary result in the Minkowski spacetime, signaling no presence of the gravitational wave at all. This can be understood intuitively: since the oscillation of the gravitational wave is much faster than the detector's response time ${\sim}1/\Delta E$ for the two-level transition, the detector has no time to respond to the driving oscillation (provided that the gravitational wave is weak enough so that the linearized theory is legitimate).
\section{Constant-accelerating trajectory}\label{sec:constant-accelerating}
In this section, we study the response of an Unruh-DeWitt detector that moves with a constant acceleration $1/\alpha$ in the $z$ direction. The trajectory is given by
\begin{equation}\label{x a}
t= \alpha \sinh{\frac{\tau}{\alpha}}+t_0,\quad x= y=\mathrm{const},\quad z= \alpha\cosh{\frac{\tau}{\alpha}}+z_0,
\end{equation}
where $t_0$ and $z_0$ are displacement parameters.
\subsection{Long-wavelength limit}
In the long-wavelength limit, substituting \eqref{x a} into $D^+_\mathrm{lw}(x,x')$ in \eqref{D lw}, we have\footnote{The derivation involves some details, which can be found in Appendix C of \cite{Chiou:2016exd}.}
\begin{eqnarray}\label{D in acceleration}
D^+_\mathrm{lw}(\Delta \tau)
&=& -\frac{\alpha^2}{16\pi^2} \frac{1}{\sqrt{1-\mathcal{A}^2}}\frac{1}{\sinh^2{(\frac{\Delta\tau}{2\alpha}-\frac{i\epsilon}{2\alpha})}} \nonumber\\
&=& -\frac{1}{4\pi^2} \frac{1}{\sqrt{1-\mathcal{A}^2}}
\sum_{k=-\infty}^{\infty} \frac{1}{(\Delta \tau-i\epsilon+2\pi ik\alpha)^2},
\end{eqnarray}
where we have applied the identity
\begin{equation}
\csc^2{\pi x}=\frac{1}{\pi^2}\sum_{k=-\infty}^{\infty}\frac{1}{(x-k)^2}.
\end{equation}
As $D^+_{\mathrm{lw}}(\tau,\tau')$ depends only on $\Delta\tau\equiv\tau-\tau'$, the Unruh-DeWitt detector is in equilibrium with $\phi$.
Substituting \eqref{D in acceleration} into \eqref{F dot} and performing the contour integral, we obtain the transition rate
\begin{equation}\label{F a}
\dot{F}(\Delta E)= \frac{\Delta E}{2\pi}\,\frac{1}{\sqrt{1-\mathcal{A}^2}}\, \frac{1}{e^{2\pi\Delta E\alpha}-1},
\end{equation}
for both $\Delta E>0$ and $\Delta E<0$.
Except for the overall proportional factor $(1-\mathcal{A}^2)^{-1/2}$, this result is exactly the same as the ordinary result of a constant-accelerating Unruh-DeWitt detector moving in the Minkowski spacetime given by\footnote{It is often said that the transition rate \eqref{F a Minkowski} for a constant-accelerating detector moving in the Minkowski spacetime corresponds to the transition rate for a detector lying at rest in a thermal bath of particles of $\phi$ at the Unruh temperature $T=(2\pi k_\mathrm{B}\alpha)^{-1}$. This is common confusion, and it is only a coincidence solely for the case of $m=0$ and $\Delta E>0$ that the transition rate of the former happens to be identical to that of the latter. (See Sec.\ III.A.4 of \cite{Crispino:2007eb} for a detailed clarification for this confusion.) In fact, as discussed in \secref{sec:Unruh-DeWitt detector} (and more in Appendix~A of \cite{Chiou:2016exd}), the temperature of a detector, including the Unruh temperature \eqref{Unruh temperature}, is a notion of detailed balance between a transition process and its reverse process via \eqref{detailed balance}, whereas the transition rate \textit{per se} does not makes any sense of temperature.}
\begin{equation}\label{F a Minkowski}
\dot{F}(\Delta E)= \frac{\Delta E}{2\pi}\, \frac{1}{e^{2\pi\Delta E\alpha}-1}.
\end{equation}
The overall proportional factor is in perfect agreement with the free-falling case with $u^x=u^y=0$ as shown in \eqref{F dot max}.
It is again instructive to compare \eqref{F a} with the case in flat spacetime with a compact
dimension. In a flat spacetime where one spatial dimension perpendicular to the $z$ direction is compactified with a finite length $L$, the transition rate of the Unruh-DeWitt detector moving along \eqref{x a} is given by (see \cite{Chiou:2016exd})
\begin{equation}\label{rate for a in z}
\dot{F}_L(\Delta E)
= \frac{\Delta E}{2\pi} \frac{1}{e^{2\pi\Delta E\alpha}-1}
- \Theta(-\Delta E) \sum_{n=1}^\infty
\frac{\sin\left(2\alpha\Delta E\sinh^{-1}\frac{nL}{2\alpha}\right)}
{n\pi L\sqrt{1+\left(\frac{nL}{2\alpha}\right)^2}}.
\end{equation}
The correction due to the compact length $L$ in \eqref{rate for a in z} is scaled as $O(L^{-1})$, which vanishes in the formal limit $L\rightarrow\infty$. By contrast, the leading-order correction due to the gravitational wave is scaled as $O((1/\omega)^0)$ as shown in \eqref{F a}, which survives the formal limit $\omega\rightarrow0$.
Just as we have commented for the case of a free-falling detector, the gravitational wave effect on a constant-accelerating Unruh-DeWitt detector is a genuine quantum effect that cannot be understood in terms of gravitational wave tidal force.
\subsection{Short-wavelength limit}
In the short-wavelength limit, substituting the constant-accelerating trajectory \eqref{x a} into $D^+_\mathrm{sw}(x,x')$ in \eqref{D sw} yields the same result in the Minkowski spacetime.
Consequently, in the short-wavelength limit $\omega\rightarrow\infty$, the detector moving along \eqref{x a} is in equilibrium with $\phi$, and the transition rate is
the same as the ordinary result in the Minkowski spacetime.
Again, this can be understood intuitively: the gravitational oscillates so fast that the detector has no time to respond to it.
\section{Summary and remarks}\label{sec:summary}
Applying the techniques of light-front quantization used in the literature of QCD, we have successfully quantized the real scalar field $\phi(x)$ in a monochromatic gravitational wave background, obtaining the formulae \eqref{a and a dag} and \eqref{phi}. This enables us to compute the corresponding Wightman function $D^+(x,x')$ as given by \eqref{Wightman}, which is greatly simplified in the long-wavelength limit $\omega\rightarrow0$ and the short-wavelength limit $\omega\rightarrow\infty$, as given by \eqref{D lw} and \eqref{D sw}, respectively.
With the Wightman function at hand, we then investigate the response of the Unruh-DeWitt detector in a gravitational wave background for the two cases of a free-falling trajectory and a constant-accelerating trajectory moving in the propagation direction of the gravitational wave.
In the long-wavelength limit $\omega\rightarrow0$, the equilibrium transition rate of the detector moving along a free-falling trajectory \eqref{x freefall} is given by \eqref{F v}, and that along a constant-accelerating trajectory \eqref{x a} is given by \eqref{F a}.
These results are different from their corresponding counterparts \eqref{F v Minkowski} and \eqref{F a Minkowski}, respectively, in flat spacetime (without any gravitational wave) by an overall proportional factor, which depends on the amplitude of the gravitational wave but not the gravitational wave wavelength $1/\omega$. That is, in both cases, the leading-order correction due to the gravitational wave is of $O((\omega\Delta)^0)$, which survives the formal limit $\omega\rightarrow0$ as long as the gravitational wave amplitude remains finite. This suggests that the gravitational wave effect on the Unruh-DeWitt detector is more involved than merely imposing a large length scale of the wavelength as well as the presence of spatial boundaries does of the length scale delimited by the boundaries (see \cite{Chiou:2016exd,Davies:1989me}).
Furthermore, even if we suppose that the Unruh-DeWitt detector has a finite spatial extent, this effect is qualitatively different from that on a classical mechanical system and cannot be explained out in terms of gravitational wave tidal force. This is a genuine quantum effect that has no classical analogue.
On the other hand, in the short-wavelength limit $\omega\rightarrow\infty$, the Unruh-DeWitt detector following a free-falling trajectory in a gravitational wave background is \emph{not} in equilibrium with the field $\phi$ in general, except for the free-falling trajectory moving in the propagation direction of the gravitational wave as given by \eqref{trivial sol x}. The equilibrium transition rate along \eqref{trivial sol x} is the same as the ordinary result in flat spacetime, showing no response to the gravitational wave.
Furthermore, along a constant-accelerating trajectory given by \eqref{x a}, the equilibrium transition rate is again the same as the ordinary result in flat spacetime.
The fact that the Unruh-DeWitt detector in equilibrium with $\phi$ does not respond to the gravitational wave background in the limit $\omega\rightarrow\infty$ can be understood intuitively: the gravitational wave oscillates so fast that the detector has no time to respond to the driving oscillation within the timescale ${\sim}1/\Delta E$ for the two-level transition.
The results of our study also raise some open questions. We have demonstrated that the transition rate of an Unruh-DeWitt detector can be affected by the presence of a gravitational wave. It is unclear whether the gravitational wave is involved with energy transfer for the transition process \eqref{transition} as it is for the response of a classical mechanical system, or perhaps it merely acts as a ``catalyst'', which increases the transition efficiency but does not deposit or withdraw any net energy.
Neither does our study investigate the aspect of detailed balance. It is uncertain whether detailed balance can be established in some particular settings in a gravitational wave background. If detailed balance can be established after all, it is important to know whether the temperature of detailed balance is shifted by the gravitational wave background and whether the temperature shift can be understood in terms of energy transfer from the gravitational wave.
Furthermore, it was recently shown that the concurrence of transition probability of a pair of free-falling Unruh-DeWitt detectors, which serves as a probe of vacuum entanglement, responds to the presence of a gravitational wave and exhibits certain resonance effects \cite{Xu:2020pbj}. As our study considers arbitrary gravitational wave polarization and more general trajectories, including both free-falling and constant-accelerating ones, our results may help to investigate the interplay between vacuum entanglement and gravitational waves in broader settings.
The analysis in this paper is performed entirely in the framework of quantum field theory in curved spacetime. The scalar field $\phi$ is quantized in the gravitational wave background, and the Unruh-DeWitt detector is modeled as a quantum system coupled to $\phi$. The gravitational field, on the other hand, is treated completely as a classical background and not quantized at all. Although the gravitational wave effect on the Unruh-DeWitt detector is a quantum effect with no classical analogue, it is not a consequence of quantum gravity. Nevertheless, it is an intriguing open question whether the effect we found here can be understood in terms of a quantum detector coupled to both particles of $\phi$ and gravitons (quantized particles of the degrees of freedom of gravitational waves). This question might also be related to the aforementioned issue of energy transfer from the gravitational wave.
Finally, while it is conceptually important to understand the effects of a gravitational wave on a quantum system, it should be remarked that our investigation on the Unruh-DeWitt in response to a gravitational wave is mainly for theoretical concerns. Experimentally, measuring the response of the Unruh-DeWitt detector is extremely challenging, if not completely out of reach of current technology. In the case of a constant-accelerating trajectory, an experimentally reachable value for the acceleration $1/\alpha$ is extremely small compared to $\abs{\Delta E}$ (i.e., $\abs{\Delta E}\alpha\gg1$) for a typical two-level quantum system of which one can reliably measure the transition rate. This renders the transition rate given by \eqref{F a} experimentally indistinguishable from the result of a free-falling trajectory as given by \eqref{F v}.
With regard to the transition rate \eqref{F v} of a free-falling trajectory, its response to the gravitational wave seems to be measurable for $\Delta E<0$, provided that the amplitude of the gravitational wave is strong enough. However, as the transition rate, in principle, has to be measured by a large ensemble of identical Unruh-DeWitt detectors (see \secref{sec:Unruh-DeWitt detector} and Appendix A of \cite{Chiou:2016exd} for more discussions), the interactions between detectors of the ensemble and between the system and its surroundings will inevitably introduce noises that will spoil the signal in response to a gravitational wave that is extremely weak when arriving on earth.
Nevertheless, as the techniques of quantum measurement advance drastically in recent years, it might be possible to overcome the noise problem and eventually use an Unruh-DeWitt-type quantum system as a gravitational wave detector in the near future. Compared to resonance mass detectors and interferometric gravitational wave detectors, a quantum detector is expected to be sensitive to gravitational waves of much higher frequencies, since the characteristic timescale of a quantum system is typically much shorter than that of a resonance mass detector or an interferometric gravitational detector.
Furthermore, an Unruh-DeWitt-type detector may also have a quite wide bandwidth of sensitivity, because, as indicated by \eqref{F v}, the leading-order correction is insensitive to $\omega$ in the long-wavelength limit.
In order for the Unruh-DeWitt-type detector to be used as a gravitational wave detector, for the theoretical aspect, one will have to perform a detailed numerical analysis to know how exactly it responds to any arbitrary wavelength $1/\omega$, not only the results obtained in this paper for the long-wavelength and short-wavelength extremes.
\begin{acknowledgments}
The authors would like to thank an anonymous reviewer of the previous manuscript for raising some important issues, which have helped to improve this paper significantly. This work was supported in part by the Ministry of Science and Technology, Taiwan under the Grants No.\ 110-2112-M-002-016-MY3 and No.\ 110-2112-M-110-015.
\end{acknowledgments}
|
\section{Introduction}
The last few years have seen sweeping popularity of applying neural networks to a wide range of robotics problems \cite{sunderhauf2018limits}, such as perception \cite{kehl2017ssd, park2019deepsdf, florence2018dense}, reasoning \cite{fazeli2019see} and planning \cite{ichter2018learning}. In particular, researchers have had great success training control policies with neural networks on different robot platforms \cite{lee2020learning, tan2018sim, hwangbo2017control, kalashnikov2018qt}. Typically these control policies are obtained through reinforcement learning (RL) algorithms \cite{sutton2018reinforcement, schulman2015trust, haarnoja2018soft}. Although immensely successful, these neural-network controllers still generally lack theoretical guarantees on their performance, which could hinder their adoption in many safety-critical applications.
A crucial guarantee currently missing for neural-network controllers is the stability of the closed-loop system, especially Lyapunov stability. A system is regionally stable in the sense of Lyapunov if starting from any states within a region, the system eventually converges to an equilibrium. This region is called the region of attraction (ROA) \cite{slotine1991applied}. Lyapunov stability provides a strong guarantee on the asymptotic behavior of the system for any state within the region of attraction. It is well known that a system is Lyapunov stable if and only if there exists a \textit{Lyapunov function} \cite{slotine1991applied} that is strictly positive definite and strictly decreasing everywhere except at the goal equilibrium state. Therefore, our goal is to synthesize a pair: a neural-network controller to stabilize the system, and a Lyapunov function to certify its stability.
\begin{figure}
\centering
\begin{subfigure}{0.26\textwidth}
\includegraphics[width=0.98\textwidth]{figures/quadrotor3d_snapshots10.png}
\end{subfigure}
\begin{subfigure}{0.22\textwidth}
\includegraphics[width=0.98\textwidth]{figures/quadrotor3d_V_traj_snapshots.pdf}
\end{subfigure}
\caption{(left) Snapshots of stabilizing a 3D quadrotor with our neural-network controller to the hovering position at the origin (red snapshot) from different initial states. The green curves are the paths of the quadrotor center. (right) value of the neural-network Lyapunov functions along the simulated trajectories. The Lyapunov function has positive values, and decreases along the trajectories.}
\label{fig:quadrotor3d_snapshot6}
\end{figure}
In the absence of neural networks in the loop, a significant body of work from the control community provides tools to synthesize Lyapunov-stable controllers \cite{slotine1991applied, boyd1994linear}. For example, for a linear dynamical system, one can synthesize a linear LQR controller to achieve Lyapunov stability (with the quadratic Lyapunov function solved through the Riccati equation). For a control-affine system with polynomial dynamics, Javis-Wloszek et al. \cite{jarvis2003some} and Majumdar et al. \cite{majumdar2013control} have demonstrated that a Lyapunov-stable controller together with a Lyapunov function, both polynomial functions of the state, can be obtained by solving a sum-of-squares (SOS) program. Recently, for more complicated systems, researchers have started to represent Lyapunov functions (but not their associated controllers) using neural networks. For example, Chang et al. synthesized linear controllers and neural network Lyapunov functions for simple nonlinear systems \cite{chang2019neural}. In a similar spirit, there is growing interest to approximate the system dynamics with neural networks, such as for racing cars \cite{williams2017information}, actuators with friction/stiction \cite{hwangbo2019learning}, perceptual measurement like keypoints \cite{manuelli2020keypoints}, system with contacts \cite{ pfrommer2020contactnets}, and soft robots \cite{gillespie2018learning}, where an accurate Lagrangian dynamics model is hard to obtain, while the neural-network dynamics model can be extracted from rich measurement data. Hence we are interested in systems whose dynamics are given as a neural network.
Unlike previous work which is restricted to linear \cite{boyd1994linear, chang2019neural} or polynomial controllers \cite{jarvis2003some, majumdar2013control}, our paper provides a novel approach to synthesize a stable neural-network controller, together with a neural-network Lyapunov function, for a given dynamical system whose forward dynamics is approximated by another neural network. The overall picture together with a Lyapunov function is visualized in Fig. \ref{fig:feedback_diagram}.
\begin{figure}
\begin{subfigure}{0.3\textwidth}
\includegraphics[width=0.98\textwidth]{figures/networks_diagram.pdf}
\subcaption{}
\end{subfigure}
\begin{subfigure}{0.17\textwidth}
\includegraphics[width=0.98\textwidth]{figures/pendulum_V_3.pdf}
\subcaption{}
\end{subfigure}
\caption{(Left) The forward system\protect\footnotemark (which contains a neural network) is given, and we aim at finding the controller and a Lyapunov function to prove Lyapunov stability of the closed-loop system. Both the controller and Lyapunov function contain neural networks. (right) Visualization of a Lyapunov function for a 2-dimensional system. The Lyapunov function is usually a bowl-shaped function that is strictly positive except at the goal state.}
\label{fig:feedback_diagram}
\end{figure}
\footnotetext{The quadrotor picture is taken from \cite{bonna2015trajectory}.}
In order to synthesize neural-network controllers and Lyapunov functions, one has to first be able to verify that the neural-network functions satisfy the Lyapunov condition for \textit{all} states within a region. There are several techniques to verify certain properties of neural network outputs for all inputs within a range. These techniques can be categorized by whether the verification is exact, e.g., using Satisfiability Modulo Theories (SMT) solvers \cite{katz2017reluplex, chang2019neural, abate2020formal} or mixed-integer programs (MIP) solvers \cite{bunel2018unified, tjeng2018evaluating, dai2020counter, chen2020learning}, versus inexact verification by solving a relaxed convex problem \cite{bastani2016measuring, fazlyab2020safety, wong2018provable, Shen2020}. Another important distinction among these techniques is the activation functions used in the neural networks. For example, Abate et al. \cite{abate2020formal} and Chang et al. \cite{chang2019neural} learn neural-network Lyapunov functions with quadratic and \textit{tanh} activation functions respectively. On the other hand, the piecewise linear nature of (leaky) ReLU activation implies that the input and output of a (leaky) ReLU network satisfy mixed-integer linear constraints, and hence network properties can be exactly verified by MIP solvers \cite{tjeng2018evaluating, dai2020counter}. In this work, due to its widespread use, we choose the (leaky) ReLU unit for all neural networks. This enables us to perform exact verification of the Lyapunov condition without relaxation for safety-critical robot missions.
The verifiers (both SMT and MIP solvers) can either definitively certify that a given candidate function satisfies the Lyapunov condition everywhere in the region, or generate counter examples violating the Lyapunov condition. In this work, we solve MIPs to find the most adversarial counter examples, namely the states with the maximal violation of the Lyapunov condition. Then, in order to improve the satisfaction of the Lyapunov conditions, we propose two approaches to jointly train the controller and the Lyapunov function. The first approach is a standard procedure in counter-example guided training, where we add the counter examples to the training set and minimize a surrogate loss function of the Lyapunov condition violation on this training set \cite{abate2020formal, chang2019neural, ravanbakhsh2019learning}. The second approach is inspired by the bi-level optimization community \cite{bard2013practical, landry2019differentiable, dai2020counter}, where we directly minimize the maximal violation as a min-max problem through gradient descent.
Our contributions include: 1) we synthesize a Lyapunov-stable neural-network controller together with a neural-network Lyapunov function. To the best of our knowledge, this is the first work capable of doing this. 2) We compute an inner approximation of the region of attraction for the closed-loop system. 3) We present two approaches to improve the networks based on the counter examples found by the MIP verifier. 4) We demonstrate that our approach can successfully synthesize Lyapunov-stable neural-network controllers for systems including inverted pendulums, 2D and 3D quadrotors, and that they outperform a baseline LQR controller.
\section{Problem statement}
\label{sec:problem_definition}
We consider a discrete-time system whose forward dynamics is
\begin{subequations}
\begin{align}
x_{t+1} = f(x_t, u_t) = \phi_{\text{dyn}}(x_t, u_t) - \phi_{\text{dyn}}(x^*, u^*) + x^*\label{eq:forward_dynamics}\\
u_{\text{min}}\le u_t \le u_{\text{max}}\label{eq:input_limits}
\end{align}
\end{subequations}
where $x_t\in\mathbb{R}^{n_x}, u_t\in\mathbb{R}^{n_u}$, $u_{\text{min}}$ and $u_{\text{max}}$ are the lower/upper input limits. $\phi_{\text{dyn}}$ is a feedforward fully connected neural network with leaky ReLU activation functions \footnote{Since ReLU can be regarded as a special case of leaky ReLU, we present our work with leaky ReLU for generality.} \footnote{Our approach can also handle other architectures such as convolution. For simplicity of presentation we don't discuss them in this paper.}. $x^*/u^*$ are the state/control at the goal equilibrium. By definition the dynamics equation \eqref{eq:forward_dynamics} guarantees that at the equilibrium state/control $x_t = x^*, u_t = u^*$, the next state $x_{t+1}$ remains the equilibrium state $x_{t+1}=x^*$. Due to the universal approximation theorem \cite{leshno1993multilayer}, we can approximate an arbitrary smooth dynamical system written as \eqref{eq:forward_dynamics} with a neural network. Our goal is to find a control policy $u_t = \pi(x_t)$ and a Lyapunov function $V(x_t):\mathbb{R}^{n_x}\rightarrow \mathbb{R}$, such that the following Lyapunov conditions are satisfied:
\begin{subequations}
\begin{align}
V(x_t) > 0 \;\forall x_t\in\mathcal{S}, x_t\ne x^*\label{eq:lyapunov_positivity}\\
V(x_{t+1}) - V(x_t) \le -\epsilon_2 V(x_t)\;\forall x_t\in\mathcal{S}, x_t\ne x^*\label{eq:lyapunov_derivative}\\
V(x^*) = 0\label{eq:lyapunov_equlibrium}
\end{align}
\label{eq:lyapunov_condition}%
\end{subequations}
where $\mathcal{S}$ is a compact sub-level set $\mathcal{S} = \{x_t | V(x_t)\le \rho\}$, and $\epsilon_2 > 0$ is a given positive scalar. The Lyapunov conditions in \eqref{eq:lyapunov_condition} guarantee that starting from any state inside $\mathcal{S}$, the state converges exponentially to the equilibrium state $x^*$, and $\mathcal{S}$ is a region of attraction of the closed-loop system. In addition to the control policy and the Lyapunov function, we will find an inner approximation of the region of attraction. Note that condition \eqref{eq:lyapunov_derivative} is a constraint on the Lyapunov function $V(\cdot)$ as well as the control policy $\pi(\cdot)$, since $V(x_{t+1}) = V(f(x_t, \pi(x_t)))$ depends on both the control policy to compute $x_{t+1}$ together with the Lyapunov function $V(\cdot)$.
\section{Background on ReLU and MIP}
\label{sec:background}
In this section we give a brief overview of the mixed-integer linear formulation which encodes the input/output relationship of a neural network with leaky ReLU activation. This MIP formulation arises from the network output being a piecewise-affine function of the input, hence intuitively one can use linear constraints for each affine piece, and binary variables for the activated piece. Previously researchers have solved mixed-integer programs (MIP) to verify certain properties of the feedforward neural network in machine learning applications such as verifying image classifiers \cite{bunel2018unified, tjeng2018evaluating}.
\begin{figure}
\centering
\includegraphics[width=0.15\textwidth]{figures/leaky_relu.pdf}
\caption{A leaky ReLU activation function.}
\label{fig:leaky_relu}
\end{figure}
For a general fully-connected neural network, the input/output relationship in each layer is
\begin{subequations}
\begin{align}
z_i =& \sigma(W_{i}z_{i-1}+b_{i}), i=1,\hdots,n-1\\
z_n =& W_{n}z_{n-1}+b_{n},
z_0 = x,
\end{align}
\label{eq:fully_connected_network}%
\end{subequations}
where $W_{i}, b_{i}$ are the weights/biases of the i'th layer. The activation function $\sigma(\cdot)$ is the leaky ReLU function shown in Fig.\ref{fig:leaky_relu}, as a piecewise linear function $\sigma(y) = \max(y, cy)$ where $0 \le c<1$ is a given scalar. If we suppose that for one leaky ReLU neuron, the input $y\in\mathbb{R}$ is bounded in the range $y_{\text{lo}} \le y \le y_{\text{up}}$ (where $y_{\text{lo}}< 0$ and $y_{\text{up}}>0$), then we can use the big-M technique to write out the input/output relationship of a leaky ReLU unit $w=\sigma(y)$ as the following mixed-integer linear constraints
\begin{subequations}
\begin{align}
w\ge y,\quad
w \ge cy\\
w \le cy - (c-1)y_{\text{up}}\beta,\quad
w \le y - (c-1)y_{\text{lo}}(\beta-1)\\
\beta\in \{0, 1\},
\end{align}
\label{eq:leaky_relu_mix_integer}%
\end{subequations}
where the binary variable $\beta$ is active when $y\ge 0$. Since the only nonlinearity in the neural network \eqref{eq:fully_connected_network} is the leaky ReLU unit $\sigma(\cdot)$, by replacing it with constraints \eqref{eq:leaky_relu_mix_integer}, the relationship between the network output $z_n$ and input $x$ is fully captured by mixed-integer linear constraints.
We expect \textit{bounded} input to the neural networks since we care about states within a neighbourhood of the equilibrium so as to prove regional Lyapunov stability, and the system input $u_t$ is restricted within the input limits (Eq. \eqref{eq:input_limits}). With a bounded neural network input, the bound of each ReLU neuron input can be computed by either \textit{Interval Arithmetic} \cite{wong2018provable}, by solving a linear programming (LP) problem \cite{tjeng2018evaluating}, or by solving a mixed-integer linear programming (MILP) problem \cite{Cheng2017maximum, Fischetti2018}.
After formulating neural network verification as a mixed-integer program (MIP), we can efficiently solve MIPs to global optimality with off-the-shelf solvers, such as Gurobi \cite{gurobi} and CBC \cite{forrest2005cbc} via branch-and-cut method.
\section{Approach}
\label{sec:approach}
In this section we present our approach to finding a pair of neural networks as controller and the Lyapunov function. We will first use the technique described in the previous section \ref{sec:background}, and demonstrate that one can verify the Lyapunov condition \eqref{eq:lyapunov_condition} through solving MIPs. Then we will present two approaches to reduce the Lyapunov condition violation using the MIP results. Finally we explain how to compute an inner-approximation of the region of attraction.
\subsection{Verify Lyapunov condition via solving MIPs}
\label{subsec:lyapunov_mip}
We represent the Lyapunov function with a neural network $\phi_V:\mathbb{R}^{n_x}\rightarrow\mathbb{R}$ as
\begin{align}
V(x_t) = \phi_V(x_t) - \phi_V(x^*) + |R(x_t - x^*)|_1, \label{eq:nn_lyapunov}
\end{align}
where $R$ is a matrix with full column rank. $|R(x_t - x^*)|_1$ is the 1-norm of the vector $R(x_t-x^*)$. Eq. \eqref{eq:nn_lyapunov} guarantees $V(x^*)=0$, hence condition \eqref{eq:lyapunov_equlibrium} is trivially satisfied. Notice that even without the term $|R(x_t - x^*)|_1$ in \eqref{eq:nn_lyapunov}, the Lyapunov function would still satisfy $V(x^*) = 0$, but adding this 1-norm term assists $V(\cdot)$ in satisfying the Lyapunov condition $V(x_t)>0$. As visualized in Fig \ref{fig:lyapunov_add_l1_3}, $\phi_V(x_t) - \phi_V(x^*)$ is a piecewise-affine function of $x_t$ passing through the point $(x^*, 0)$. Most likely $(x^*, 0)$ is in the interior of one of the linear pieces, instead of on the boundary of a piece; hence locally around $x^*$, the term $\phi_V(x_t) - \phi_V(x^*)$ is a linear function of $x_t$, which will become negative away from $x^*$, violating the positivity condition $V(x_t)>0$ (\eqref{eq:lyapunov_positivity}). To remedy this, we add the term $|R(x_t - x^*)|_1$ to the Lyapunov function. Due to $R$ being full-rank, this 1-norm is strictly positive everywhere except at $x^*$. With sufficiently large $R$, we guarantee that at least locally around $x^*$ the Lyapunov function is positive. Notice that $V(x_t)$ is a piecewise-affine function of $x_t$.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{figures/lyapunov_add_l1_3.pdf}
\caption{The term $\phi_V(x_t) - \phi_V(x^*)$ is a piecewise-affine function that passes through $(x_t, V(x_t)) = (x^*, 0)$. Most likely $x^*$ is within the interior of a linear piece, but not at the boundary between pieces. This linear piece will go negative in the neighbourhood of $x^*$. By adding the 1-norm $|R(x_t-x^*)|_1$ (red lines), the Lyapunov function (blue lines) is at least locally positive around $x^*$.}
\label{fig:lyapunov_add_l1_3}
\end{figure}
Our approach will entail searching for both the neural network $\phi_V$ and the full column-rank matrix $R$ in \eqref{eq:nn_lyapunov}. To guarantee $R$ being full column-rank, we parameterize it as
\begin{align}
R = U \left(\Sigma + \text{diag}(r_1^2, \hdots, r_{n_x}^2)\right)V^T, \label{eq:R_svd}
\end{align} where $U, V$ are given orthonormal matrices, $\Sigma$ is a given diagonal matrix with strictly positive diagonal entries, and scalars $r_1, \hdots, r_{n_x}$ are free variables. The parameterization \eqref{eq:R_svd} guarantees $R$ being full column-rank since the minimal singular value of $R$ is strictly positive.
We represent the control policy using a neural network $\phi_\pi:\mathbb{R}^{n_x}\rightarrow\mathbb{R}^{n_u}$ as
\begin{align}
u_t = \pi(x_t) = \text{clamp}\left(\phi_\pi(x_t) - \phi_\pi(x^*) + u^*, u_{\text{min}}, u_{\text{max}}\right), \label{eq:control_network}
\end{align}
where $\text{clamp}(\cdot)$ clamps the value $\phi_\pi(x_t) - \phi_\pi(x^*) + u^*$ elementwisely within the input limits $[u_{\text{min}}, u_{\text{max}}]$, namely
\begin{align}
\text{clamp}(\alpha, \text{lo}, \text{up}) = \begin{cases} \text{up} \;\text{if } \alpha > \text{up}\\ \alpha \;\text{if } \text{lo}\le \alpha \le \text{up}\\ \text{lo}\; \text{if }\alpha < \text{lo}\end{cases}. \label{eq:clamp_u}
\end{align}
The control policy \eqref{eq:control_network} is a piecewise-affine function of the state $x_t$. Notice that \eqref{eq:control_network} guarantees that at the equilibrium state $x_t = x^*$, the control action is $u_t = u^*$.
It is worth noting that our approach is only applicable to systems that can be stabilized by \textit{regular} (e.g., locally Lipschitz bounded) controllers. Some dynamical systems, for example a unicycle, require non-regular controllers for stabilization, where our approach would fail. The readers can refer to \cite{sontag1999stability} for more background on regular controllers.
The Lyapunov condition \eqref{eq:lyapunov_condition}, in particular, \eqref{eq:lyapunov_positivity}, is a strict inequality. To verify this through MIP which only handles non-strict inequalities constraints $\ge$ and $\le$, we change condition \eqref{eq:lyapunov_positivity} to the following condition with $\ge$
\begin{align}
V(x_t) \ge \epsilon_1 |R(x_t-x^*)|_1\;\forall x\in\mathcal{S} \label{eq:lyapunov_positivity_l1},
\end{align}
where $0<\epsilon_1<1$ is a given positive scalar. Since $R$ is full column-rank, the right-hand side is 0 only when $x_t=x^*$. Hence the non-strict inequality constraint \eqref{eq:lyapunov_positivity_l1} is a sufficient condition for the strict inequality constraint \eqref{eq:lyapunov_positivity}. In Appendix \ref{subsec:l1_necessary} we prove that it is also a necessary condition.
In order to verify the Lyapunov condition \eqref{eq:lyapunov_condition} for a candidate Lyapunov function and a controller, we consider verifying the condition \eqref{eq:lyapunov_positivity_l1} and \eqref{eq:lyapunov_derivative} for a given bounded polytope $\mathcal{B}$ around the equilibrium state. The verifier solves the following optimization problems
\begin{subequations}
\begin{align}
\max_{x_t\in\mathcal{B}} \epsilon_1|R(x_t-x^*)|_1 - V(x_t) \label{eq:lyapunov_positivity_mip}\\
\max_{x_t\in\mathcal{B}} V(x_{t+1}) - V(x_t) + \epsilon_2 V(x_t),\label{eq:lyapunov_derivative_mip}
\end{align}
\label{eq:lyapunov_mip}%
\end{subequations}
where the objectives are the violation of condition \eqref{eq:lyapunov_positivity_l1} and \eqref{eq:lyapunov_derivative} respectively. If the optimal values of both problems are 0 (attained at $x_t=x^*$), then we certify the Lyapunov condition \eqref{eq:lyapunov_condition}. The objective in \eqref{eq:lyapunov_positivity_mip} is a piecewise-affine function of the variable $x_t$ since both $V(x_t)$ and $|R(x_t-x^*)|_1$ are piecewise-affine. Likewise in optimization problem \eqref{eq:lyapunov_derivative_mip}, since the control policy \eqref{eq:control_network} is a piecewise-affine functions of $x_t$, and the forward dynamics \eqref{eq:forward_dynamics} is a piecewise-affine function of $x_t$ and $u_t$, the next state $x_{t+1}=f(x_t, \pi(x_t))$, its Lyapunov value $V(x_{t+1})$ and eventually the objective function in \eqref{eq:lyapunov_derivative_mip} are all piecewise-affine functions of $x_t$. It is well known in the optimization community that one can maximize a piecewise-affine function within a bounded domain ($\mathcal{B}$ in this case) through solving an MIP \cite{vielma2010mixed}. In section \ref{sec:background} we have shown the MIP formulation on neural networks with leaky ReLU units; in Appendix \ref{subsec:l1_clamp_MIP} we present the MIP formulation for the 1-norm in $|R(x_t-x^*)|_1$ and the clamp function in the control policy.
By solving the mixed-integer programs in \eqref{eq:lyapunov_mip}, we either verify that the candidate controller is Lyapunov-stable with the candidate Lyapunov function $V(x_t)$ as a stability certificate; or we generate counter examples of $x_t$, where the objective values are positive, hence falsify the candidates. By maximizing the Lyapunov condition violation in the MIP \eqref{eq:lyapunov_mip}, we find not only \textit{a} counter example if one exists, but the \textit{worst} counter example with the largest violation. Moreover, since the MIP solver traverses a binary tree during branch-and-cut, where each node of the tree might find a counter example, the solver finds a list of counter examples during the solving process. In the next subsection, we use both the worst counter example and the list of all counter examples to reduce the Lyapunov violation.
\subsection{Trainer}
\label{subsec:trainer}
After the MIP verifier generates counter examples violating Lyapunov conditions, to reduce the violation, we use these counter examples to improve the candidate control policy and the candidate Lyapunov function. We present two iterative approaches. The first one minimizes a surrogate function on a training set, and the counter examples are appended to the training set in each iteration. This technique is widely used in the counter-example guided training \cite{chang2019neural, abate2020formal, chen2020learning, kapinski2014simulation}. The second approach minimizes the maximal Lyapunov condition violation directly by solving a min-max problem through gradient descent. In both approaches, we denote the parameters we search for as $\theta$, including
\begin{itemize}
\item The weights/biases in the controller network $\phi_\pi$;
\item The weights/biases in the Lyapunov network $\phi_V$;
\item $r_1, \hdots, r_{n_x}$ in the full column-rank matrix $R$ (Eq. \eqref{eq:R_svd}).
\end{itemize}
namely we optimize both the control policy and the Lyapunov function simultaneously, so as to satisfy the Lyapunov condition on the closed-loop system.
\subsubsection{Approach 1, growing training set with counter examples}
\label{subsubsec:train_counter_examples}
A necessary condition for satisfying the Lyapunov condition for \text{any} state in $\mathcal{B}$, is that the Lyapunov condition holds for many sampled states within $\mathcal{B}$. Hence we could reduce a surrogate loss function on a training set $\mathcal{X}$ containing sampled states. The training set $\mathcal{X}$ grows after each MIP solve by appending the counter examples generated from the MIP solve. Since the MIP \eqref{eq:lyapunov_positivity_mip} and the MIP \eqref{eq:lyapunov_derivative_mip} generate different counter examples, we keep two separate training sets $\mathcal{X}_1$ and $\mathcal{X}_2$ for MIP \eqref{eq:lyapunov_positivity_mip} and MIP \eqref{eq:lyapunov_derivative_mip} respectively.
We design a surrogate loss function for $\mathcal{X}_1, \mathcal{X}_2$ to measure the violation of the Lyapunov condition on the training set. We denote the violation of condition \eqref{eq:lyapunov_positivity_l1} on a sample state $x^{i}_1\in\mathcal{X}_1$ as $\eta_1(x^{i}_1)$, and the violation of condition \eqref{eq:lyapunov_derivative} on a sample state $x^{i}_2 \in\mathcal{X}_2$ as $\eta_2(x^{i}_2)$, defined as
\begin{subequations}
\begin{align}
\eta_1(x^{i}_1) =& \max(\epsilon_1|R(x^{i}_1-x^*)|_1 - V(x^{i}_1), 0)\label{eq:violation_1}\\
\eta_2(x^{i}_2) =& \max(V(f(x^{i}_2, \pi(x^{i}_2))) - V(x^{i}_2) + \epsilon_2V(x^{i}_2), 0),
\end{align}
\end{subequations}
We denote $\eta_1(\mathcal{X}_1)$ and $\eta_2(\mathcal{X}_2)$ as the vectors whose i'th entry is the violation on the i'th sample $\eta_1(x^{i}_1)$ and $\eta_2(x^i_2)$ respectively, then our surrogate function is defined as
\begin{align}
\text{loss}_\theta(\mathcal{X}_1, \mathcal{X}_2) = |\eta_1(\mathcal{X}_1)|_p + |\eta_2(\mathcal{X}_2)|_p, \label{eq:training_set_loss}
\end{align}
where $|\cdot|_p$ denotes the p-norm of a vector, such as 1-norm (mean of the violation), $\infty-$norm (maximal of the violation) and $4-$norm (a smooth approximation of the $\infty-$ norm). The subscript $\theta$ in the loss function \eqref{eq:training_set_loss} emphasizes its dependency on $\theta$, the parameters in both the controller and the Lyapunov function. We then minimize the surrogate loss function on the training set via standard batched gradient descent on $\theta$. The flow chart of this approach is depicted in Fig. \ref{fig:training_flow_chart}. Algorithm \ref{algorithm:train_on_sample} presents the pseudo-code.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{figures/training_flow_chart.pdf}
\caption{Flow chart of Algorithm \ref{algorithm:train_on_sample}.}
\label{fig:training_flow_chart}
\end{figure}
\begin{algorithm}
\caption{Train controller/Lyapunov function on training sets constructed from verifier}
\label{algorithm:train_on_sample}
\begin{algorithmic}[1]
\STATE Start with a candidate neural-network controller $\pi$, a candidate Lyapunov function $V$, and training sets $\mathcal{X}_1, \mathcal{X}_2$.
\WHILE{not converged}
\STATE Solve MIPs \eqref{eq:lyapunov_positivity_mip} and \eqref{eq:lyapunov_derivative_mip}.
\IF{MIP \eqref{eq:lyapunov_positivity_mip} or MIP \eqref{eq:lyapunov_derivative_mip} has maximal objective $ > 0$}
\IF{MIP \eqref{eq:lyapunov_positivity_mip} maximal objective $> 0$}
\STATE Add the counter examples from MIP \eqref{eq:lyapunov_positivity_mip} to $\mathcal{X}_1$.
\ENDIF
\IF{MIP \eqref{eq:lyapunov_derivative_mip} maximal objective $>0$}
\STATE Add the counter examples from MIP \eqref{eq:lyapunov_derivative_mip} to $\mathcal{X}_2$.
\ENDIF
\STATE Perform batched gradient descent on the parameters $\theta$ to reduce the loss function \eqref{eq:training_set_loss} on the training set $\mathcal{X}_1, \mathcal{X}_2$. Stop until either $\text{loss}_{\theta}(\mathcal{X}_1, \mathcal{X}_2)=0$, or reaches a maximal epochs.
\ELSE
\STATE converged = true.
\ENDIF
\ENDWHILE
\end{algorithmic}
\end{algorithm}
Since the surrogate loss function is the Lyapunov condition violation on just the sampled states, the batched gradient descent will overfit to the training set, and potentially cause large violation away from the sampled states. To avoid this overfitting problem, we consider an alternative approach without constructing the training sets.
\subsubsection{Approach 2, minimize the violation via min-max program}
\label{subsubsec:bilevel_approach}
Instead of minimizing a surrogate loss function on a training set, we can minimize the Lyapunov condition violation directly through the following min-max problem
\begin{equation}
\begin{split}
\min_{\theta}\left(\underbrace{\max_{x_t\in\mathcal{B}} \epsilon_1|R(x_t-x^*)|_1-V(x_t)}_{\text{MIP } \eqref{eq:lyapunov_positivity_mip}} \right.\qquad\qquad\\
\left.+ \underbrace{\max_{x_t\in\mathcal{B}}V(x_{t+1}) - V(x_t) + \epsilon_2 V(x_t)}_{\text{MIP }\eqref{eq:lyapunov_derivative_mip}}\right),\label{eq:bilevel_optimization}
\end{split}
\end{equation}
where $\theta$ are the parameters in the controller and the Lyapunov function, introduced at the beginning of this subsection \ref{subsec:trainer}. Unlike the traditional optimization problem, where the objective function is a closed-form expression of the decision variable $\theta$, in our problem \eqref{eq:bilevel_optimization} the objective function is the result of other maximization problems, whose coefficients and bounds of the constraint/cost matrices depend on $\theta$. In order to solve this min-max problem, we adopt an iterative procedure. In each iteration we first solve the inner maximization problem using MIP solvers, and then compute the gradient of the MIP optimal objective w.r.t the variables $\theta$, finally we apply gradient descent along this gradient direction, so as to reduce the objective in the outer minimization problem.
To compute the gradient of the maximization problem objective w.r.t $\theta$, after solving the inner MIP to optimality, we fix all the binary variables to their optimal solutions, and keep only the active linear constraints. The inner maximization problem can then be simplified to
\begin{subequations}
\begin{align}
\gamma(\theta)=\max_{s} c_\theta^Ts + d_\theta\\
\text{s.t } A_\theta s = b_\theta,
\end{align}
\label{eq:inner_maximization}%
\end{subequations}
where the problem coefficients/bounds $c_\theta, d_\theta, A_\theta, b_\theta$ are all explicit functions of $\theta$. $s$ contains all the continuous variables in the MIP, including $x_t$ and other slack variables. The optimal cost of \eqref{eq:inner_maximization} can be written in the closed form as $\gamma(\theta) = c_\theta^TA_\theta^{-1}b_\theta + d_\theta$, and then we can compute the gradient $\partial\gamma(\theta)/\partial\theta$ by back-propagating this closed-form expression. Note that this gradient is well defined if a tiny perturbation on $\theta$ changes only the optimal value of the continuous variables $s$, but not the set of active constraints or the optimal binary variable values (changing them would make the gradient ill-defined). This technique to differentiate the optimization objective w.r.t neural network parameters is becoming increasingly popular in the deep learning community. The interested readers can refer to \cite{amos2017optnet, agrawal2019differentiable} for a more complete treatment on differentiating an optimization layer.
Algorithm \ref{algorithm:bilevel} shows pseudo-code for this min-max optimization approach.
\begin{algorithm}
\caption{Train controller/Lyapunov function through min-max optimization}
\label{algorithm:bilevel}
\begin{algorithmic}[1]
\STATE Given a candidate control policy $\pi$ and a candidate Lyapunov function $V$.
\WHILE{not converged}
\STATE Solve MIP \eqref{eq:lyapunov_positivity_mip} and \eqref{eq:lyapunov_derivative_mip}.
\IF{Either of MIP \eqref{eq:lyapunov_positivity_mip} of \eqref{eq:lyapunov_derivative_mip} has maximal objective $>0$}
\STATE Compute the gradient of the MIP objectives w.r.t $\theta$, denote this gradient as $\partial \gamma/\partial\theta$.
\STATE $\theta = \theta - \text{StepSize}*\partial\gamma/\partial\theta$.
\ENDIF
\ENDWHILE
\end{algorithmic}
\end{algorithm}
\subsection{Computing region of attraction}
\label{subsec:roa}
After the training process in section \ref{subsec:trainer} converges to satisfy the Lyapunov condition for every state inside the bounded polytope $\mathcal{B}$, we compute an inner approximation of the region of attraction for the closed-loop system. (Notice that the verified region $\mathcal{B}$ is \textit{not} a region of attraction, since it's not an invariant set, while the sub-level sets of $V$ are guaranteed to be invariant). One valid inner approximation is the largest sub-level set $\mathcal{S} = \{x_t | V(x_t)\le \rho\}$ contained inside the verified region $\mathcal{B}$, as illustrated in Fig. \ref{fig:roa}. Since we already obtained the Lyapunov function $V(x_t)$ in the previous section, we only need to find the largest value of $\rho$ such that $\mathcal{S}\subset \mathcal{B}$.
Equivalently we can find $\rho$ through the following optimization problem
\begin{align}
\rho = \min_{x_t\in\partial \mathcal{B}} V(x_t), \label{eq:roa_mip}
\end{align}
where the compact set $\partial \mathcal{B}$ is the boundary of the polytopic region $\mathcal{B}$, and the constraint $x_t\in\partial\mathcal{B}$ can be formulated as mixed-integer linear constraints (with one binary variable for a face of the polytope $\mathcal{B}$). As explained previously, the Lyapunov function $V(x_t)$ is a piecewise-affine function of $x_t$, hence the optimization problem \eqref{eq:roa_mip} is again an MIP, and can be solved efficiently by MIP solvers.
It is worth noting that the size of this inner approximation of the region of attraction can be small, as we fix the Lyapunov function and only search for its sub-level set. To verify a larger inner approximation, one possible future research direction is to search for the Lyapunov function and the sub-level set simultaneously, as in \cite{richards2018lyapunov}.
\section{Results}
We synthesize stable controllers and Lyapunov functions on pendulum, 2D and 3D quadrotors. We use Gurobi as the MIP solver. All code runs on an Intel Xeon CPU. The sizes of the neural networks are shown in Table \ref{table:network_size} in Appendix \ref{subsec:network_structure}.
\subsection{Inverted pendulum}
\label{subsec:inverted_pendulum}
We first test our approach on an inverted pendulum. We approximate the pendulum Lagrangian dynamics using a neural network, by first simulating the system with many state/action pairs, and then approximating the simulation data through regression. To stabilize the pendulum at the top equilibrium $\theta=\pi, \dot{\theta}=0$, we synthesize a neural-network controller and a Lyapunov function using both Algorithm \ref{algorithm:train_on_sample} and \ref{algorithm:bilevel}. We verify the Lyapunov condition in the box region $0 \le \theta \le 2\pi, -5\le\dot{\theta}\le 5$. The Lyapunov function $V$ is shown in Fig. \ref{fig:pendulum_V}.
\begin{figure}
\begin{minipage}{0.21\textwidth}
\centering
\includegraphics[width=0.98\textwidth]{figures/roa.pdf}
\caption{An inner approximation of the region of attraction $\mathcal{S}$ is the largest sub-level set $V(x_t)\le \rho$ contained inside the verified region $\mathcal{B}$, where the Lyapunov function is positive definite and strictly decreasing.}
\label{fig:roa}
\end{minipage}
\hspace{2pt}
\begin{minipage}{0.26\textwidth}
\centering
\includegraphics[width=0.98\textwidth]{figures/pendulum_V_roa.pdf}
\caption{Heatmap of the Lyapunov function for the inverted pendulum. The red contour is the boundary of the verified inner approximation of the region of attraction, as the largest sub-level set contained in the verified box region $0\le \theta \le 2\pi, -5\le \dot{\theta}\le 5$.}
\label{fig:pendulum_V}
\end{minipage}
\end{figure}
We simulate the synthesized controller with the original pendulum Lagrangian dynamics model (not the neural network dynamics $\phi_{\text{dyn}}$). The result is shown in Fig. \ref{fig:pendulum_nn_simulate}. Although the neural network dynamics $\phi_{\text{dyn}}$ has approximation error, the simulation results show that the neural-network controller swings up and stabilizes the pendulum for not only the approximated neural network dynamics, but also for the original Lagrangian dynamics. Moreover, starting from many states outside the verified region of attraction, and even outside our verified box region, the trajectories still converge to the equilibrium. This suggests that the controller generalizes well. The small verified region of attraction suggests that in the future we can improve its size by searching over the Lyapunov function and the sub-level set simultaneously.
We start with a small box region $0.8 \pi \le \theta \le 1.2\pi, -1 \le \dot{\theta}\le 1$, and then gradually increase the verified region. We initialize the controller/Lyapunov network as the solution in the previous iteration on a smaller box region (at the first iteration, all parameters are initialized arbitrarily). For the smaller box $0.8\pi\le\theta\le1.2\pi, -1\le\dot{\theta}\le 1$, both algorihm \ref{algorithm:train_on_sample} and \ref{algorithm:bilevel} converge within a few minutes. For the larger box $0 \le\theta\le 2\pi, -5\le\dot{\theta}\le 5$, both algorithms converge within 3 hours.
\begin{figure}
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=0.98\textwidth]{figures/pendulum_phase_plot_roa.pdf}
\end{subfigure}
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=0.98\textwidth]{figures/pendulum_V_traj.pdf}
\end{subfigure}
\caption{(left) phase plot of simulating the pendulum Lagrangian dynamics with the neural-network controller. The red contour is the boundary of the verified region of attraction, as the largest level set within the verified box region $\mathcal{B}$ (black dashed box). All the simulated trajectories (even starting outside of the dashed box) converge to the goal state. (right) Lyapunov function value along the simulated trajectories. The Lyapunov function decreases monotonically along the trajectories.}
\label{fig:pendulum_nn_simulate}
\vspace{-12pt}
\end{figure}
\subsection{2D quadrotor}
\label{subsec:quadrotor_2d}
We synthesize a stabilizing controller and a Lyapunov function for the 2D quadrotor model used in \cite{tedrake2009underactuated}. Again we first train a neural network $\phi_{\text{dyn}}$ to approximate the Lagrangian dynamics. Our goal is to steer the quadrotor to hover at the origin. In Fig.\ref{fig:quadrotor2d_snapshot} we visualize the snapshots of the quadrotor stabilized by our neural-network controller. We verified the Lyapunov conditions in the region $[-0.75, -0.75, -0.5\pi, -4, -4, -2.75] \le [x, z, \theta, \dot{x}, \dot{z}, \dot{\theta}] \le [0.75, 0.75, 0.5\pi, 4, 4, 2.75]$.
\begin{figure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=0.97\textwidth]{figures/quadrotor2d_snapshot1.pdf}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=0.97\textwidth]{figures/quadrotor2d_snapshot2.pdf}
\end{subfigure}
\caption{Snapshots of the 2D quadrotor simulation (with the original Lagrangian dynamics) using our neural-network controller from different initial states. The red lines are the trajectories of the quadrotor body frame origin.}
\label{fig:quadrotor2d_snapshot}
\end{figure}
We sample 10000 initial states uniformly in the box $[-0.9, -0.9, -0.5\pi, -4.5, -4.5, -3] \le [x, z, \theta, \dot{x}, \dot{z},\dot{\theta}]\le[0.9, 0.9, 0.5\pi, 4.5, 4.5, 3]$. For each initial state we simulate the Lagrangian dynamics with the neural network and an LQR controller. We summarize the result in table \ref{table:quadrotor2d_sim} on whether the simulation converges to the goal state or not. More states can be stabilized by the neural-network controller than the LQR controller. Moreover, the off-diagonal entries in Table \ref{table:quadrotor2d_sim} demonstrates that the set of sampled states that are stabilized by the neural-network controller is a strict super-set of the set of states stabilized by the LQR controller. We believe there are two factors contributing to the advantage of our neural-network controller against an LQR: 1) the neural-network controller is piecewise linear while the LQR controller is linear; the latter can be a special case of the former. 2) the neural-network controller is aware of the input limits while the LQR controller is not.
\begin{table}
\centering
\begin{tabular}{|c | c | c|}
\hline
& NN succeeds & NN fails\\
\hline
LQR succeeds & 8078& 0\\
\hline
LQR fails& 1918& 4\\
\hline
\end{tabular}
\caption{Number of success/failure for $10,000$ simulations of 2D quadrotor with the neural network (NN) controller and an LQR controller. The simulation uses the Lagrangian dynamics.}
\label{table:quadrotor2d_sim}
\end{table}
We then focus on certain two dimensional slices of the state space, and sample many initial states on these slices. For each sampled initial state we simulate the Lagrangian dynamics using both the neural-network and the LQR controller. We visualize the simulation results in Fig. \ref{fig:quadrotor2d_lqr_vs_nn}. Each dot represents a sampled initial state, and we color each initial state based on whether the neural-network (NN)/LQR controllers succeed in stabilizing that initial state to the goal
\begin{itemize}
\item Purple: NN succeeds but LQR fails.
\item Green: both NN and LQR succeed
\item Red: both NN and LQR fail.
\end{itemize}
Evidently the large purple region suggests that the region of attraction with the neural-network controller is a strict super-set of that with the LQR controller. We observe that for the initial states where LQR fails, the LQR controller requires thrusts beyond the input limits. If we increase the input limits then LQR can stabilize many of these states. Hence by taking input limits into consideration, the neural-network controller achieves better performance than the LQR.
\begin{figure}
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=0.98\textwidth]{figures/quadrotor2d_lqr_vs_nn_result31.pdf}
\end{subfigure}
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=0.98\textwidth]{figures/quadrotor2d_lqr_vs_nn_result30.pdf}
\end{subfigure}
\caption{We sample 2500 initial states within the box region $[-8, -8]\le [\dot{x}, \dot{z}] \le [8, 8]$, with $[x, z, \theta, \dot{\theta}]$ fixed to $[-0.75, 0.3, 0.3\pi, 2]$ (left), and $[0.75, 0.5, -0.4\pi, 2]$ (right). We simulate from each initial state with the neural-network controller (NN) and the LQR controller, and color each initial state based on whether the simulation converges to the goal. All red dots (where the NN controller fails) are outside of the black box region within which we verified the Lyapunov conditions.}
\label{fig:quadrotor2d_lqr_vs_nn}
\end{figure}
Both algorithm \ref{algorithm:train_on_sample} and \ref{algorithm:bilevel} find the stabilizing controller. For a small box region $[-0.1, -0.1, -0.1\pi, -0.5, -0.5, -0.3] \le [x, z, \theta, \dot{x}, \dot{z}, \dot{\theta}]\le [0.1, 0.1, 0.1\pi, 0.5, 0.5, 0.3]$, both algorithms converge in 20 minutes. For the larger box used in Table \ref{table:quadrotor2d_sim}, the algorithms converge in 1 day.
\subsection{3D quadrotor}
We apply our approach to a 3D quadrotor model with 12 states \cite{mellinger2011minimum}. Again, our goal is to steer the quadrotor to hover at the origin. As visualized in Fig.\ref{fig:quadrotor3d_snapshot}, our neural-network controller can stabilize the system. Training this controller took 3 days.
\begin{figure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=0.98\textwidth]{figures/quadrotor3d_snapshot11.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=0.98\textwidth]{figures/quadrotor3d_snapshot12.png}
\end{subfigure}
\caption{Snapshots of simulating the quadrotor using our neural-network controller with the Lagrangian dynamics. The quadrotor converge to the hovering state at the origin (red).}
\label{fig:quadrotor3d_snapshot}
\end{figure}
The quadrotor Lagrangian dynamics is approximated by a neural network $\phi_{\text{dyn}}$ with comparatively large mean-squared error (MSE) around $10^{-4}$ (reducing MSE would require a neural network too large for our MIP solver), while other examples in this paper have MSE in the order of $10^{-6}$. Hence there are noticeable discrepancies between the simulation with Lagrangian dynamics and with the neural network dynamics $\phi_{\text{dyn}}$. In Fig \ref{fig:quadrotor3d_V_traj} we select some results to highlight the discrepancy, that the Lyapunov function always decreases along the trajectories simulated with neural-network dynamics, while it could increase with Lagrangian dynamics. Nevertheless, the quadrotor eventually always converges to the goal state. We note that the same phenomenon would also happen if we took a linear approximation of the quadrotor dynamics and stabilized the quadrotor with an LQR controller. If we were to plot the quadratic Lyapunov function (which is valid for the LQR controller and the linearized dynamics), that Lyapunov function could also increase along the trajectories simulated with the nonlinear Lagrangian dynamics (see Fig \ref{fig:quadrotor3d_lqr_V_traj} in the Appendix). Analogous to approximating the nonlinear dynamics with a linear one and stabilizing it with a linear LQR controller, our approach can be regarded as approximating the nonlinear dynamics with a neural network and stabilizing it with another neural-network controller.
\begin{figure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=0.98\textwidth]{figures/quadrotor3d_V_traj_newton_dynamics3.pdf}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=0.98\textwidth]{figures/quadrotor3d_V_traj_nn_dynamics3.pdf}
\end{subfigure}
\caption{3D quadrotor Lyapunov function value along the simulated trajectories with our neural-network controller. The quadrotor is simulated with Lagrangian dynamics (left) vs the dynamics approximated by a neural network $\phi_{\text{dyn}}$ (right). In both left and right sub-plots, the initial states are the same.}
\vspace{-10pt}
\label{fig:quadrotor3d_V_traj}
\end{figure}
Finally we compare the performance of Algorithm \ref{algorithm:train_on_sample} which appends counter examples to training sets, against Algorithm \ref{algorithm:bilevel} with min-max optimization. We take 10 runs of each algorithm, and report the average computation time for each algorithm in Table \ref{table:algorithm_time} \footnote{There are 2 failed runs with Algorithm \ref{algorithm:train_on_sample} on the 2D quadrotor example, that they time out after 6 hours, and are not included in Table \ref{table:algorithm_time}. For the 3D quadrotor we only include 3 runs as they are too time-consuming}. For the small-sized task (pendulum with 2 states), Algorithm \ref{algorithm:train_on_sample} is orders of magnitude faster than Algorithm \ref{algorithm:bilevel}, and they take about the same time on the medium-sized task (2D quadrotor with 6 states). On the large-sized task (3D quadrotor with 12 states), Algorithm \ref{algorithm:train_on_sample} doesn't converge while Algorithm \ref{algorithm:bilevel} can find the solution. We speculate this is because Algorithm \ref{algorithm:train_on_sample} overfits to the training set. For a small-sized task the overfitting is not a severe problem as a small number of sampled states are sufficient to represent the state space; while for a large-sized task it would require a huge number of samples to cover the state space. With the limited number of counter examples Algorithm \ref{algorithm:train_on_sample} overfits to these samples while causing large Lyapunov condition violation elsewhere. This is evident from the loss curve plot in Fig.\ref{fig:loss_for_algorithms} for a 2D quadrotor task. Although both algorithms converge, the loss curve decreases steadily with Algorithm \ref{algorithm:bilevel}, while it fluctuates wildly with Algorithm \ref{algorithm:train_on_sample}. We believe that the fluctuation is caused by overfitting to the training set in the previous iteration. Nevertheless, this comparison is not yet conclusive, and we are working to improve the performance of Algorithm \ref{algorithm:train_on_sample} on the large-size task.
\begin{table}
\centering
\begin{tabular}{|c|c|c|}
\hline
& Algorithm \ref{algorithm:train_on_sample} &Algorithm \ref{algorithm:bilevel}\\
\hline
Pendulum&8.4s& 224s\\
2D quadrotor &948.3s & 1004.7s\\
3D quadrotor &Time out after 5 days&65.7hrs\\
\hline
\end{tabular}
\caption{Average computation time of 10 runs for both algorithms. To speed up the computation, the verified region $\mathcal{B}$ is relatively small.}
\label{table:algorithm_time}
\end{table}
\begin{figure}
\centering
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=0.98\textwidth]{figures/counter_example_positivity_loss2.pdf}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=0.98\textwidth]{figures/counter_example_derivative_loss2.pdf}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=0.98\textwidth]{figures/bilevel_positivity_loss2.pdf}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=0.98\textwidth]{figures/bilevel_derivative_loss2.pdf}
\end{subfigure}
\caption{Loss curves on the 2D quadrotor task for Algorithm \ref{algorithm:train_on_sample} and \ref{algorithm:bilevel}}
\label{fig:loss_for_algorithms}
\vspace{-10pt}
\end{figure}
\section{Conclusion and Future work}
In this paper, we demonstrate a method to synthesize a neural-network controller to stabilize a robot system, as well as a neural-network Lyapunov function to prove the resulting stability. We propose an MIP verifier to either certify Lyapunov stability, or generate counter examples that can be used to improve the candidate controller and the Lyapunov function. We present another MIP to compute an inner approximation of the region of attraction. We demonstrate our approach on an inverted pendulum, 2D and 3D quadrotors, and showcase that it can outperform a baseline LQR controller.
Currently, the biggest challenge of our approach is scalability. The speed bottleneck lies in solving MIPs, where the number of binary variables scales linearly with the total number of neurons in the networks. In the worst case, the complexity of solving an MIP scales exponentially with the number of binary variables, when the solver has to check every node of a binary tree. However, in practice, the branch-and-cut process significantly reduces the number of nodes to explore. Recently, with the growing interest from the machine learning community, many approaches were proposed to speed up verifying neural networks through MIP by tightening the formulation \cite{anderson2020strong, tsay2021partition}. We plan to explore these approaches in the future.
Our proposed MIP formulation works for discrete-time dynamical systems. For continuous-time dynamical systems, neural networks have been previously used either to approximate the system dynamics \cite{lutter2018deep}, or to synthesize optimal controllers \cite{farshidian2019deep}. We plan to extend our approach to continuous-time dynamical systems. Moreover, we can readily apply our approach to systems whose dynamics are approximated by piecewise-affine dynamical systems, such as soft robots \cite{tonkens2020soft} and hybrid systems with contact \cite{marcucci2019mixed}, since piecewise-affine dynamic constraints can easily be encoded into MIP.
Many safety-critical missions also require the robot to avoid unsafe regions. We can readily extend our framework to synthesize barrier functions \cite{ames2016control} so that the robot certifiably stays within the safe region.
\section*{Acknowledgments}
Benoit Landry is sponsored by the NASA University Leadership initiative (grant $\#$80NSSC20M0163), and Lujie Yang is sponsored by Amazon Research Award $\#$6943503. This article solely reflects the opinions and conclusions of its authors and not any funding agencies. We would like to thank Vincent Tjeng and Shen Shen for the valuable discussion.
\bibliographystyle{plainnat}
|
\section{Introduction}
Interpolation of potential energy surfaces (PESs) is one of the most common applications of supervised machine learning (ML) methods.
Commonly, this is done with deep neural networks (DNNs) \cite{GP_vs_NN,NN_Tucker,NN_small_molecules,NN_highdim,many-body_NN,NN_environment-dependent_atom,NN_ANI-1,DTNN,SchNet,Yafu_diab_NN,Cormorant,PhysNet,PCA_NN_PES}, parametric models, or kernel models, such as Gaussian Processes (GPs) or kernel ridge regression (KRR) \cite{GP_PRL,GP_vs_NN,RAVH_NJP,GP_NaKNak_JCP,Jun_JCTC,Hiroki_JCP,Qingyong_GP,Multi-fidelity_GP,Gauss_PES,GP_Cui_JPhysB,GP_reactive_PES,GP_reactive_PES_2,GP_PES-Learn}.
While both methodologies have proven to be flexible enough, GPs require less "tunning" compared to NNs, where the search for an optimal architecture could be computationally demanding \cite{NN_NAS_2019,NN_NAS_2020}.
One of the advantages of GPs is their ability to quantify the uncertainty in their prediction, which is commonly used in applications where noisy data or the optimization of $f$ black-box functions. The latter application has proven to be a successful tool in physical-chemistry \cite{RAVH_NJP,RAVH_BODFT,Deng_BO_JCP,Phoenics_BO,BO_DFT_water,BO_Materials,MOBOpt_vargas}.\\
GPs and KRR mainly depend on two factors: training data, and the kernel function. The latter quantifies the similarity between a pair of points.
If the kernel function can generalize a similarity metric, GPs are efficient regression algorithms that require less training data than NNs \cite{GP_vs_NN}, and are also capable of extrapolating functions beyond the training data regime \cite{RAVH_PRL,RAVH_extrapolation,Jun_JCTC,duvenaud_bic,Krems_PRE}. To achieve more robust kernel functions, once can simply combine different simple kernel functions \cite{duvenaud_kerneladdition,gpbook}. However, to automate the search for an optimal kernel combination, one can use Bayesian model selection to search over possible priors, where a prior is determined by the kernel function form. Given the combinatorial nature of the space of possible kernel functions, sampling this distribution is hard and one must rely on heuristic search approaches. One such approach is the so called Automatic Bayesian Covariance Discovery (ABCD) \cite{duvenaud_bic,duvenaud_kerneladdition}, described in greater detail below. Recently, people have proposed the use of Bayesian optimization for this as well \cite{malkomes2016bayesian}. \\
Another technique that has proven to construct accurate PES is based on KRR using the atomic gradient information, also known as gradient-domain machine learning (GDML) \cite{GDML_1,GDML_2}.
Here, the kernel complexity is usually not modified. However, a combination of kernels could be easily implemented given the analytic derivatives of them \cite{GDML_Krems}; however, the training procedure is based on all possible combinations of kernels and the selection of them is through a cross-validation scheme.
The computational complexity of GDML also scales cubically with the number of training data.
In some cases, for large covariance matrices the Nystr\"om approximation is applied \cite{GP_BigData,GP_BigData_Rev}.\\
Although kernel structure search has been shown to accurately interpolate of PESs, they are not often applied to study high-dimensional physical systems due to a lack of accuracy and significant computational cost of ${\cal O}(N^3)$, where $N$ is the number of training points. \cite{GP_NaKNak_JCP,Jun_JCTC,Hiroki_JCP,Qingyong_GP}. Typically, approximations must be used to achieve the required scalability, which can degrade the predictive performance of GP models even further. For systems that require a higher number of training data points, multiple GPs must be trained in order to find the optimal kernel combination.
In this work, we present a different approach to construct kernel functions based on Bochners’ theorem \cite{gpbook} that are both more robust and computationally efficient.\\ In the following sections, we briefly introduce GPs and review how complex kernels can be constructed by using Bochners’ theorem, a strategy known as \emph{spectral kernel learning}.
We interpolate the PES of four different chemical systems, e.g., protonated Imidazole dimer \cite{Hiroki_JCP}, Benzene , Malonaldehyde, and Ethanol \cite{GDML_1,GDML_2}.
We also illustrate that by using GPytorch \cite{gpytorch} and KEOPS \cite{keops}, modern deep-learning libraries, combined with GPUs we can efficiently train full GPs with 20 000 training points without relying on the Nystr\"om approximation.
\section{Gaussian Processes}
Gaussian Processes are one of the most used non-parametric probabilistic ML models in physical sciences.
A GP is specified by its mean ($\mu$) and covariance ($\sigma$) functions, $f(\mathbf{x}) \sim {\cal GP}(\mu,\Sigma)$.
One of the key components in a GP model is the kernel function, $k(\mathbf{x}_i,\mathbf{x}_j)$, used to parametrized the covariance matrix.
For prediction, GPs have an exact posterior distribution whose mean and standard deviation are,
\begin{eqnarray}
\mu(\mathbf{x^*}) &=& \mathbf{k}(\mathbf{x^*},\mathbf{X})^\top \left [ K(\mathbf{X},\mathbf{X}) + \sigma_n 1 \right]^{-1} \mathbf{y} \label{eqn:gp_mu}\\
\sigma(\mathbf{x^*}) &=& k(\mathbf{x^*},\mathbf{x^*}) \left [ K(\mathbf{X},\mathbf{X}) + \sigma_n 1 \right]^{-1} \mathbf{k}(\mathbf{x^*},\mathbf{X}),\label{eqn:gp_cov}
\end{eqnarray}
where $\mathbf{X}$, for PESs, contains the geometry configurations and $\mathbf{y}$ is the energy for each configuration. $K(\cdot,\cdot)$ is known as the design or covariance matrix, whose elements are given by the kernel function, $K_{i,j} = k(\mathbf{x}_i,\mathbf{x}_j)$. $\mathbf{x}^*$ is the point where prediction is aimed, and $\mathbf{k}(\mathbf{x^*},\mathbf{X})$ is the similarity vector between $\mathbf{x^*}$ and all the training data.\\
Given the nature of GPs, it is possible to find the marginal distribution in closed form, $p(\mathbf{y}|\mathbf{X})$. It is common to use the logarithm of the marginal likelihood distribution, LMLL, to optimize the free-parameters of the kernel function ($\mathbf{\theta}$),
\begin{eqnarray}
\log p(\mathbf{y}|\mathbf{X},\mathbf{\theta}) &=& -\frac{1}{2}\mathbf{y}^\top\left [ K(\mathbf{X},\mathbf{X}) + \sigma_n \mathbb{1} \right]^{-1}\mathbf{y} \nonumber \\ &&- \frac{1}{2}\log \Big | K(\mathbf{X},\mathbf{X}) + \sigma_n \mathbb{1} \Big| -\frac{N}{2}\log 2\pi.
\label{eqn:lmll}
\end{eqnarray}
Some of the most common kernel functions are,
\begin{eqnarray}
k_{RBF}(\mathbf{x}_i,\mathbf{x}_j) &=& \exp \Big [ -\frac{1}{2}r^{2}(\mathbf{x}_i,\mathbf{x}_j) \Big] \label{eqn:k_rbf}\\
k_{RQ}(\mathbf{x}_i,\mathbf{x}_j) &=& \left ( 1 + \frac{|\mathbf{x}_i - \mathbf{x}_j|^2}{2\alpha\ell^2}\right )^{-\alpha} \label{eqn:k_rq}\\
k_{^{MAT}_{3/2}}(\mathbf{x}_i,\mathbf{x}_j) &=& \left (1 + \sqrt{3}r(\mathbf{x}_i,\mathbf{x}_j) \right ) \exp \Big [ -\sqrt{3}r(\mathbf{x}_i,\mathbf{x}_j) \Big] \nonumber \\ \label{eqn:k_mat_32}\\
k_{^{MAT}_{5/2}}(\mathbf{x}_i,\mathbf{x}_j) &=& \left (1 + \sqrt{5}r(\mathbf{x}_i,\mathbf{x}_j) + \frac{5}{3} r^{2}(\mathbf{x}_i,\mathbf{x}_j) \right ) \nonumber \\ &&\times \exp \Big [ -\sqrt{5}r(\mathbf{x}_i,\mathbf{x}_j) \Big] \label{eqn:k_mat}
\end{eqnarray}
where $\mathbf{r}^2(\mathbf{x}_i,\mathbf{x}_j) = (\mathbf{x}_i - \mathbf{x}_j)^\top M (\mathbf{x}_i - \mathbf{x}_j)$, where $M$ is a diagonal matrix parametrized with a $\ell_d$ length scale for each dimension of $\mathbf{x}$.
For more details about GPs we refer the reader to Ref. \citenum{gpbook}.
In previous works, it has been shown the accuracy of GPs increase by simply combining kernels to account for more complicated functions. One of the most common approaches was proposed in Refs. \citenum{duvenaud_kerneladdition,duvenaud_bic,RAVH_PRL}, where the combination of kernels was guided by selecting the kernel that has the maximum LMLL. To avoid selecting a kernel (${\cal M}_i$) with a large number of parameters, a term that penalizes the number of free parameters ($|{\cal M}_i|$) in the kernel was included,
\begin{eqnarray}
\text{BIC}({\cal M}_i) = \log p(\mathbf{y}|\mathbf{X},\mathbf{\theta}, {\cal M}_i) + \frac{1}{2} |{\cal M}_i|\log N.
\label{eqn:BIC}
\end{eqnarray}
This methodology is commonly named as the \emph{Bayesian information criterion} (BIC).
These kernel structure discovery methods have been demonstrated to extrapolate physical observables accurately enough to detect phase transitions \cite{RAVH_PRL,RAVH_extrapolation}.
Additionally, GPs with complex kernels have shown the possibility to predict accurate energies for PES trained only with low energy points \cite{Jun_JCTC,Hiroki_JCP}.
The computational complexity of Gaussian processes compounded with the fact that kernel search requires training many such models has resulted in an inability to use full available datasets \cite{Jun_JCTC,Hiroki_JCP,Qingyong_GP}.
Also, by considering highly complex kernel combinations, the optimization of the kernel parameters becomes harder, and given the greedy search strategy in practice one could end up with non-optimal kernels.
Finally, raising the complexity of the kernel combination in practice hits a plateau in the learning capacity meaning, adding more kernels do not increase the accuracy of the model \cite{Jun_JCTC}.\\
Another possibility to construct kernel functions automatically is through Bochners' theorem \cite{gpbook},
\begin{eqnarray}
k(\tau) = \int_{\mathbb{R}^{D}} e^{2\pi i s^\top\tau} S(\mathbf{s}) \;\mathrm{d} \mathbf{s},
\label{eqn:Bochners_theorem}
\end{eqnarray}
where $\mathbf{\tau} = \mathbf{x}_i - \mathbf{x}_j$ and $S(s)$ is the spectral density of the $k(\cdot,\cdot)$. $S(\cdot)$ and $k(\cdot)$ are Fourier duals.
In order for $k(\cdot)$ to be a valid kernel, $S(\cdot)$ must be integrable. For example, the spectral density of the SE kernel is also a Gaussian function.
This theorem ensures that the kernel $k(\tau)$ parametrizes a positive-definite covariance matrix.
The Matern family of covariance functions can also be derived using the Bochners' theorem \cite{gpbook}.
In Ref. \citenum{GP_SD}, L\'{a}zaro-Gredilla \emph{et al.} proposed a novel way to construct $S(\mathbf{s})$ by assuming it is proportional to a probability measure, $S(\mathbf{s})\propto p_{S}(\mathbf{S})$.
By doing so, the integral over the frequency domain can be computed by Monte Carlo,
\begin{eqnarray}
k(\mathbf{x}_i - \mathbf{x}_j) &=& \int_{\mathbb{R}^{D}} \; \mathrm{d} \mathbf{s} \;\; e^{2\pi i \mathbf{s}^\top\left( \mathbf{x}_i - \mathbf{x}_j\right )} S(\mathbf{s}) \nonumber \\ &=& \sigma_0^2 \int_{\mathbb{R}^{D}} \; \mathrm{d} \mathbf{s} \;\; e^{2\pi i \mathbf{s}^\top \mathbf{x}_i} e^{-2\pi i s^\top \mathbf{x}_j} p_{S}(\mathbf{s}) \nonumber \\
&=& \sigma_0^2 \mathbb{E}_{p_{S}}\left [ e^{2\pi i \mathbf{s}^\top \mathbf{x}_i} e^{-2\pi i \mathbf{s}^\top \mathbf{x}_j} \right],
\label{eqn:Bochners_theorem_2}
\end{eqnarray}
where $\sigma_0^2$ is a normalization constant, and $\mathbb{E}_{p_{S}}$ is the expectation value with respect to $p_{S}$. The samples from $p_{S}$ used to compute $k(\cdot,\cdot)$ are known as \emph{spectral points}.
It is possible to cancel the imaginary part of $k(\cdot,\cdot)$ in Eq. \ref{eqn:Bochners_theorem_2} by sampling a pair of $\{\mathbf{s}_r,-\mathbf{s}_r\}$. This Monte Carlo procedure is valid given that $S(\mathbf{s})$ is symmetric around zero.
By taking this into account, $k(\cdot,\cdot)$ has the following closed form,
\begin{eqnarray}
k(\mathbf{x}_i - \mathbf{x}_j) &\simeq& \frac{\sigma_0^2}{2|\delta|}\sum_{r=1}^{|\delta|} \left [e^{2\pi i \mathbf{s}_r^\top \mathbf{x}_i} e^{-2\pi i s_r^\top \mathbf{x}_j} + e^{-2\pi i \mathbf{s}_r^\top \mathbf{x}_i} e^{2\pi i s_r^\top \mathbf{x}_j} \right] \nonumber \\
&=& \frac{\sigma_0^2}{|\delta|}\sum_{r=1}^{|\delta|} \cos(2\pi\mathbf{s}_r^\top(\mathbf{x}_i-\mathbf{x}_j)),
\label{eqn:k_SD}
\end{eqnarray}
where $|\delta|$ is the total number of spectral points or samples used to approximate Eq. \ref{eqn:Bochners_theorem_2}. $\mathbf{s}_r$ are the frequencies that will be learned by maximizing Eq. \ref{eqn:lmll}. Because this kernel function corresponds to an explicit finite basis expansion,
inference can be done in ${\cal O}(N)$ time and space \cite{gpytorch}.
As it is stated in Ref. \citenum{GP_SD}, this approximation is similar to a set of Dirac deltas with amplitude $\sigma_0^2$ which are distributed accordingly to $p(\mathbf{s})$. For this work, we denote this kernel as the \emph{spectral delta} kernel, $k_{SD}(\cdot,\cdot)$. Figure \ref{fig:gp_sd} depicts a simple example on how to approximate $k_{RBF}$ with 100 spectral points.
It should be noted that $k_{SD}(\cdot,\cdot)$ is not the only possible kernel that can be derived from the Bochners' theorem.
In Ref. \citenum{GP_SMK}, it was shown that by assuming $S(\mathbf{s})$ as a linear combination of two Gaussians centered at $\mathbf{s}$ and $-\mathbf{s}$, the integral in the Fourier space has a close form where $k(\mathbf{\tau}) = \sum_{q=1}^{Q}\omega_q e^{-2\pi\mathbf{\tau}^\top M_q \tau} \cos(2\pi\mathbf{\tau}^\top\mu_q)$. $\omega_q$, $M_q$ and $\mu_q$ are the free parameters of this kernel.
\begin{figure}[h!]
\centering
\includegraphics[width=0.45\textwidth]{./GP_SD_RBF.png}
\includegraphics[width=0.45\textwidth]{./K_SD_RBF.png}
\caption{(left panel) We sample a GP with the $k_{RBF}$, black symbols, and train a second GP where the covariance function is $k_{SD}$ with $|\delta| = 100$. The mean of this GP, Eq. \ref{eqn:gp_mu}, is the blue solid curve. (right panel) Dashed blue curve is the target covariance function, $k_{RBF}$, where data was sampled, and the solid curve is the approximated one with the $k_{SD}$ kernel. We optimize the parameters of $k_{SD}$ by maximizing the LMLL, Eq. \ref{eqn:lmll}.
}
\label{fig:gp_sd}
\end{figure}
\section{Results and discussion}
Here, we first compare the accuracy of the PES for the protonated imidazole dimer \cite{Hiroki_JCP} interpolated with GPs with various kernel functions, i.g., Matern (Eq. \ref{eqn:k_mat}), Spectral Delta (Eq. \ref{eqn:k_SD}), and kernel combination through the BIC method.
For the latter, we used the kernel that was optimized through the BIC method, with 5 000 points, in Ref. \citenum{Hiroki_JCP},
\begin{eqnarray}
k_{H}(\mathbf{x}_i,\mathbf{x}_j) = a_0 k_{^{MAT}_{5/2}} + a_1k_{^{MAT}_{3/2}} + a_2k_{RBF} + a_3k_{RBF}, \nonumber \\
\label{eqn:k_Hiroki_simple}
\end{eqnarray}
where all $a_i$s and the internal parameters of all the kernels were optimized by maximizing Eq. \ref{eqn:lmll}.
The PES for the protonated imidazole dimer is a 51D surface with a range of energy points spanning from [0, 35 000] cm$^{-1}$.
For the $k_{SD}$, we considered a various number of spectral points, $|\delta|$, to study the impact on the accuracy of the model's prediction.
We quantified the accuracy of each model by computing the root-mean-square error (RMSE),
\begin{eqnarray}
\text{RMSE} = \sqrt{\frac{1}{n} \sum_i^{n} \left ( y_i - \hat{y}_i \right )^2}, \nonumber \\
\label{eqn:RMSE}
\end{eqnarray}
where $y_i$s are the predicted values with each GP, and $\hat{y}_i$ are the exact values computed at the MP2/6-31++G($d,p$) level of theory\cite{Hiroki_JCP}.
The test data set consists of 10 000 points spread throughout the same energy-range of the training points.
All kernel parameters for each different GP were optimized by maximizing LMLL in the GPytorch\cite{gpytorch} suite using the Adam optimizer\cite{Adam}.
To speed up the computation of the GPs with the $k_{^{MAT}_{5/2}}$ and $k_{H}$ kernels we used the KEOPS library\cite{keops}.
All calculations were carried in a single GPU, V100 SXM2 32GB.
For a GP with $k_{H}$, 18 000 training points was the maximum number of points that we could used before running out of memory.
The RMSEs for all different models, as a function of training points, are depicted in Fig. \ref{fig:Imidazole_simple}.
We found that in the low limit of training data, a GP with a simple kernel can predict a more accurate PES than a GP with the $k_{SD}$ kernel regardless of the number of spectral points.
However, as the number of training points increases so does the accuracy of a GP with $k_{SD}$.
This correlates with the idea that more points contain more information therefore more robust kernels can be designed.
There is a significant difference between a $k_{SD}$ with 500 and 1 000 spectral points, where the RMSE is almost half of the most simple $k_{SD}$.
The most accurate PES was achieved with a GP trained with 20 000 points using the $k_{SD}$ kernel with 5 000 spectral points, RMSE = 0.176 kcal/mol; twice more accurate than a GP with the $k_{H}$ kernel and 18 000 points (RMSE = 0.547 kcal/mol).
\begin{figure}[h!]
\centering
\includegraphics[width=0.5\textwidth]{./fig_rmse_vs_N_GP_Imidazole_simple}
\caption{RMSE of the protonated Imidazole dimer interpolated with a GP with different kernels. We considered the $k_{^{MAT}_{5/2}}$ (Eq. \ref{eqn:k_mat}), $k_{H}$ (Eq. \ref{eqn:k_Hiroki_simple}), and $k_{SD}$ (Eq. \ref{eqn:k_SD}) with a different number of $|\delta|$.
The RMSE for each model was computed with 10 000 energy points that were not including in the training set.
}
\label{fig:Imidazole_simple}
\end{figure}
Following the same procedure as in Ref. \citenum{Hiroki_JCP}, we split the total PES for Imidazole dimmer into three different PESs, one for each fragment of the system.
By fragmenting the entire PES, a more accurate model is achieved.
We denote each individual fragment as $\boldsymbol{\xi}_{i} \sim {\cal GP}(\mu,k_{ij})$.
In Table \ref{tbl:Imidazole}, we report the error for each fragment and the global PES. We compare the results produce with $k_{SD}$ with a GP with the $k_{^{MAT}_{5/2}}$ kernel and the ones optimized with the BIC method \cite{Hiroki_JCP}.
For fragments 1 and 2, individual molecule of Imidazol, the optimal kernel found with the BIC method for 5 000 energy points was,
\begin{eqnarray}
k_{1}(\mathbf{x}_i,\mathbf{x}_j) = k_{2}(\mathbf{x}_i,\mathbf{x}_j) =a_0 k_{^{MAT}_{5/2}} + a_1k_{^{MAT}_{3/2}} + a_2k_{RBF} .\nonumber \\
\label{eqn:k_Hiroki_1}
\end{eqnarray}
The third fragment describes the bridge between fragments 1 and 2. The optimal kernel optimized by the BIC method is,
\begin{eqnarray}
k_{12}(\mathbf{x}_i,\mathbf{x}_j) = \left (a_0 k_{^{MAT}_{5/2}} \times k_{RQ} + a_1k_{RBF} \right )\times k_{^{MAT}_{3/2}}. \nonumber \\
\label{eqn:k_Hiroki_12}
\end{eqnarray}
We found that, for $|\delta| = $ 15 000 and $N = $ 15 000 we achieved a total RMSE of 0.056 kcal/mol, where for fragments 1 and 2 the average error is 0.032 kcal/mol and an error of 0.086 kcal/mol for the fragment that describe the interaction between both Imidazole molecules.
For 5 000 training points, the $k_{SD}$ kernel can interpolate a more accurate PES than with a GP with the Matern kernel, even with 3 times more training points; Table \ref{tbl:Imidazole}.
\begin{table}
\caption{The RMSE of each fragment and total PES with 10,000 test energy points spanned in the same energy range from the training data, as a function of the spectral, $|\delta|$, and training points $N$. }
\label{tbl:Imidazole}
\begin{tabular}{ c | c | c c c | c}
& & \multicolumn{4}{c}{RMSE [kcal/mol]} \\
$|\delta|$ & $N$ & $\boldsymbol{\xi}_{1}$ & $\boldsymbol{\xi}_{2}$ & $\boldsymbol{\xi}_{12}$ & Total \\
\hline
\multirow{4}{3em}{10 000} & 5000 & 0.074 & 0.064 & 0.205 & 0.131 \\
&10000 & 0.048 & 0.041 & 0.117 & 0.077 \\
&15000 & 0.041 & 0.038 & 0.087 & 0.059 \\
& 20000 & 0.062 & 0.039 & 0.073 & 0.060 \\ \hline
\multirow{3}{3em}{15 000} & 5000 & 0.073 & 0.064 & 0.206 & 0.131 \\
& 10000 & 0.044 & 0.038 & 0.113 & 0.073 \\
& 15000 & 0.034 & 0.030 & 0.086 & \bf{0.056} \\ \hline \hline
$k_{^{MAT}_{5/2}}$ & 15000 & 0.227 & 0.162 & 0.102 & 0.171 \\
$k_{1}, k_{2}, k_{12}$\textsuperscript{\emph{a}} & 5000 & 0.124 & 0.099 & 0.124 & 0.116 \textsuperscript{\emph{b}}\\ \hline
\end{tabular}
\textsuperscript{\emph{a}} Eqs. (\ref{eqn:k_Hiroki_1} -- \ref{eqn:k_Hiroki_12}).\\
\textsuperscript{\emph{b}} From Ref. \cite{Hiroki_JCP}, the RMSE is 0.1815 kcal/mol.
\end{table}
We alaso consider the interpolation of the Benzene, Malonaldehyde, Ethanol, and Aspirin systems \cite{GDML_1,GDML_2}.
Our results illustrate that GPs, trained with the spectral density kernel, interpolate with high accuracy the PESs for high-dimensional molecular systems.
For all four systems, we computed the mean absolute error (MAE) (Eq. \ref{eqn:MAE}) for the entire data set of points, including the training set.
\begin{eqnarray}
\text{MAE} = \frac{1}{n} \sum^{n}_i | y_i - \hat{y_i}|
\label{eqn:MAE}
\end{eqnarray}
All calculations for GPs were carried in a single GPU, Tesla T4 16GB.
In Table \ref{tbl:gpsd_qml7}, we report the number of training and spectral points that lead to the most accurate GP. For all four systems, except for Malonaldehyde, we found that GPs with $k_{SD}$ interpolates more accurately that state-of-the-art deep-learning models; Table \ref{tbl:qml7}. We compared GPs with the $k_{SD}$ with deep-learning methods, e.g., deep-tensor NN (DTNN) \cite{DTNN}, PhysNet \cite{PhysNet}, and Cormorant \cite{Cormorant}, and KRR methods combined with gradients, e.g., GDML \cite{GDML_1,GDML_2}, and sGDML \cite{sGDML}.
Figure \ref{fig:qml7_cov} displays the value of the optimized $k_{SD}$ for each pair of geometries in the data sets for all four molecules.
Aspirin is the largest system considered here, a 57D PES described with 210 features.
We found that with 15 000 training points and 2 000 spectral points, the MAE of this GP is 0.127 kcal/mol. By increasing both, $N$ and $|\delta|$ we managed to reduce the error to 0.063 kcal/mol, Table \ref{tbl:gpsd_qml7}.
\begin{table}
\caption{The lowest error for four different molecular systems computed with a GP using the $k_{SD}$ kernel. The total error was computed for the entire data set\cite{GDML_1,GDML_2}, including the training points. }
\label{tbl:gpsd_qml7}
\begin{tabular}{ c | c c | c c | c c }
& \multicolumn{2}{c |}{$k_{SD}$} & \multicolumn{2}{c |}{MAE} & \multicolumn{2}{c}{RMSE} \\
& $N$ & $|\delta|$ & [meV] & [kcal/mol] & [meV] & [kcal/mol] \\ \hline
Benzene & 10 000 & 5 000 & 0.31 & 0.0071 & 0.41 & 0.0094 \\
Malonaldehyde & 20 000 & 5 000 & 2.91 & 0.067 & 4.44 & 0.102 \\
Ethanol & 20 000 & 5 000 & 1.87 & 0.043 & 4.16 & 0.096\\
Aspirin & 20 000 & 5 000 & 2.73 & 0.063 & 3.65 & 0.0841 \\\hline \hline
\end{tabular}
\end{table}
\begin{table}
\caption{MAE, reported in kcal/mol, for different ML models. The total error was computed for the entire data set\cite{GDML_1,GDML_2}, including the training points. }
\label{tbl:qml7}
\begin{tabular}{ c | c c | c }
& $N$ [$10^{3}$] & ML model & MAE [kcal/mol] \\\hline
\multirow{3}{3em}{Benzene} & 50 & SchNet\textsuperscript{\emph{a}} & 0.070 \\
& 50 & Cormorant\textsuperscript{\emph{b}} & 0.020 \\
& 10 & GP $k_{SD}$($|\delta| = $ 5K) & \bf{0.007} \\ \hline
\multirow{4}{7em}{Malonaldehyde} & 15 & GDML\textsuperscript{\emph{c}} & 0.08 \\
& 15 & sGDML\textsuperscript{\emph{d}} & 0.074 \\
& 15 & PhysNet\textsuperscript{\emph{e}} & \bf{0.072} \\
& 15 & GP $k_{SD}$($|\delta| = $ 6K) & 0.079 \\ \hline
\multirow{4}{7em}{Ethanol} & 15 & GDML\textsuperscript{\emph{c}} & 0.058 \\
& 15 & sGDML\textsuperscript{\emph{d}} & 0.051 \\
& 15 & PhysNet\textsuperscript{\emph{e}} & 0.050 \\
& 15 & GP $k_{SD}$($|\delta| = $ 6K) & \bf{0.049} \\ \hline
\multirow{4}{7em}{Aspirin} & 15 & \multirow{2}{4em}{GDML\textsuperscript{\emph{c}}} & 0.151 \\
& 50 & & 0.130 \\
& 15 & sGDML\textsuperscript{\emph{d}} & 0.131 \\
& 15 & \multirow{2}{4em}{PhysNet\textsuperscript{\emph{e}}} & 0.124 \\
& 50 & & 0.121 \\
& 50 & SchNet\textsuperscript{\emph{a}} & 0.12 \\
& 50 & Cormorant\textsuperscript{\emph{b}} & 0.098\\
& 20 & GP $k_{SD}$($|\delta| = $ 5K) & \bf{0.063} \\ \hline \hline
\end{tabular}
\textsuperscript{\emph{a}} Ref. \cite{SchNet},
\textsuperscript{\emph{b}} Ref.\cite{Cormorant},
\textsuperscript{\emph{c}} Refs.\cite{GDML_1,GDML_2},
\textsuperscript{\emph{d}} Ref.\cite{sGDML},
\textsuperscript{\emph{e}} Ref.\cite{PhysNet},
\end{table}
\begin{figure}[h!]
\centering
\includegraphics[width=0.35\textwidth]{./Cov_GP_SD_Benzene.png}
\includegraphics[width=0.35\textwidth]{./Cov_GP_SD_Malonaldehyde.png}\\
\includegraphics[width=0.35\textwidth]{./Cov_GP_SD_Ethanol.png}
\caption{For each molecule, we computed the value of spectral density covariance function, $k_{SD}$, for each pair of geometries ($\mathbf{x}_i,\mathbf{x}_j$) in the data set\cite{GDML_1,GDML_2}. $k_{SD}$ was optimized by maximizing the LMLL, Eq. \ref{eqn:lmll}, and the number of deltas for each system is reported in Table \ref{tbl:gpsd_qml7}. The displayed values were normalized for convenience.}
\label{fig:qml7_cov}
\end{figure}
For Benzene, we achieved a RMSE = $0.31$ meV with a GP trained with only 10 000 points and 5 000 spectral points.
However, a GP with only $N$ = 5 000 and $|\delta| = 200$ is capable of predicting a more accurate PES than GDML and DTNN.
All results are displayed in Fig. \ref{fig:Benzene}.
For only 800 training points, a GP's MAE is 8.13 meV, with 500 spectral points; while the MAE of GDML with 1 000 points is 3.0 meV.
We found that, in the limit of low number of training points, a GP's accuracy is not comparable with models like GDML or sGDML.
\begin{figure}[h!]
\centering
\includegraphics[width=0.5\textwidth]{./fig_mae_GP_SD_Benzene}
\caption{MAE of the Benzene surface computed with a GP with a $k_{SD}$ kernel. We considered different number of spectral points.
We also considered different number of training points, $N=[5000,10000]$, displayed with different colors.
The horizontal solid line indicates the MAE of GDML \cite{GDML_1,GDML_2}, MAE = 3.2 meV, and the dashed line is the MAE of DTNN, MAE = 1.7 meV.
}
\label{fig:Benzene}
\end{figure}
The results for Malonaldehyde are presented in Fig. \ref{fig:Malonaldehyde}.
For a GP trained with 15 000 points and 5 000 spectral points, the PES's accuracy is almost the same as the one predicted with GDML \cite{GDML_1,GDML_2}. However, by increasing the number of training points to 18 000, we achieved a more accurate prediction, MAE = 3.05, 0.35 meV more accurate than the one with GDML.
Any PES interpolated with a GP trained with $|\delta| \ge $ 500 and $N = $ 10 000 is more accurate than the one with DTNN.
\begin{figure}[h!]
\centering
\includegraphics[width=0.5\textwidth]{./fig_mae_GP_SD_Malonaldehyde}
\caption{MAE of the Malonaldehyde surface computed with a GP with a $k_{SD}$ kernel. We considered different number of spectral points.
We also considered different number of training points, displayed with different colors.
The dotted horizontal line is the MAE computed with a GP with $N = $ 18 000 and $|\delta| = $ 2 000, MAE = 3.05 meV.
The horizontal solid line indicates the MAE of GDML \cite{GDML_1,GDML_2}, MAE = 3.3 meV, and the dashed line is with DTNN, MAE = 8.2 meV.
}
\label{fig:Malonaldehyde}
\end{figure}
For Ethanol, the predicted surface of a GP with $N = $ 15 000 and 1 000 spectral points is comparable with the one computed with GDML, see Fig. \ref{fig:Ethanol}.
However, a GP with a larger number of $|\delta|$ is capable of interpolating this system more accurately; for a GP with $|\delta| = $ 6 000 and $N = $ 15 000 points, the MAE is still the same as the one with 5 000 spectral points; MAE = 2.16 meV.
\begin{figure}[h!]
\centering
\includegraphics[width=0.5\textwidth]{./fig_mae_GP_SD_Ethanol}
\caption{MAE of the Ethanol surface computed with a GP with the $k_{SD}$ kernel considering different number of spectral points.
We also considered different number of training points, displayed in different colors.
The horizontal solid line indicates the MAE of GDML \cite{GDML_1,GDML_2}, MAE = 2.4 meV.
}
\label{fig:Ethanol}
\end{figure}
\section{Outline}
We have presented an accurate GP model capable of interpolating high dimensional PESs, i.g., 51D for protonated imidazole dimer, 61D for Benzene, 9D for both Malonaldehyde and Ethanol (36 features for both), and 57D for Aspirin surfaces (210 features).
While combining simple kernels have proven to be a successful route, here we show an alternative path to enhance the accuracy of GPs based on the Bochners’ theorem.
This methodology can lead to highly accurate GPs capable of interpolating PESs for a variety of chemical systems.
The spectral delta kernel (Eq. \ref{eqn:k_SD}), derived from the Bochners’ theorem, can also be optimized by maximizing the log-marginal likelihood as is commonly done for vanilla GPs.
For the protonated Imidazole dimer, we managed to predict a PES with a RMSE of 0.22 kcal/mol using 15 000 training and 2 000 spectral points.
By splitting the total PES into three fragments, we achieved a PES with a total RMSE of 0.06 kcal/mol with $|\delta| = [5 000, 10 000]$.
We found that $k_{SD}$ is only more accurate than the BIC method or simple kernels for systems with more than 5 000 points.
Additionally, the optimal value of $|\delta|$ depends on the number of training points too; for example, for Benzene $|\delta| \ge $ 1 000 with N = 10 000 produce surfaces with MAEs lower than 1 meV ($\approx 0.023$ kcal/mol).
The largest dimensional system considered here was Aspirin. For this system, we managed to predict a global surface with a MAE lower than 0.07 kcal/mol with 20 000 points, which compared with deep-learning models is more accurate.
By eye-inspection, we found that the optimal $|\delta|$ is between $|\delta| = [N/3,N/2]$.
A possible future work is to study how well GPs with the $k_{SD}$ kernel can extrapolating quantum observables, e.g., the high energy points for PES.
In physical sciences, the interpolation of high-dimensional landscapes is not commonly done by GPs unless the full covariance matrix is approximated by a low-rank matrix, for example using the Nystr\"om approximation, or using deep-NN which architecture must be optimize for different systems.
Here, we illustrate that by combining modern deep-learning libraries such as GPytorch and KEOPS with GPUs, GPs are robust supervised ML algorithms capable of approximating high-dimensional complex functions without approximating the covariance matrix and still being training points efficient.
The work presented here makes GPs more suitable ML algorithms to study and simulate a wider variety of physical systems.
\begin{acknowledgements}
We thank R. Krems and H. Sugisawa for useful discussions.
This work was partially supported by the U.S. Air Force
Office of Scientific Research (AFOSR) in a grant,
FA9550-20-1-0354, to Professor P. Brumer, University
of Toronto.
\end{acknowledgements}
|
\section{Introduction}
\label{sec:intro}
The Hubbard model, which is a simple theoretical model to describe electron systems with repulsive Coulomb interactions, is expected to have rich phase structures so that it has been attracting the interest of not only condensed matter physicists but also particle physicists. It has been widely known that the Hubbard model has a similar path-integral form to the Nambu--Jona-Lasinio (NJL) model~\cite{Nambu:1961tp,Nambu:1961fr}, which is a low energy effective theory in Quantum Chromodynamics (QCD): Both consist of a hopping term and a four-fermi interaction term. Their similarity, unfortunately, leads to sharing the so-called sign problem, which is a notorious difficulty in the numerical analyses based on the Monte Carlo approach.
Recently the authors have successfully applied the tensor renormalization group (TRG) method\footnote{In this paper the TRG method or the TRG approach refers to not only the original numerical algorithm proposed by Levin and Nave \cite{Levin:2006jai} but also its extensions \cite{PhysRevB.86.045139,Shimizu:2014uva,Sakai:2017jwp,Adachi:2019paf,Kadoh:2019kqk,Akiyama:2020soe,adachi2020bondweighted,Kadoh:2021fri}.} to investigate the phase transition of the four-dimensional ($4d$) NJL model at high density and very low temperature~\cite{Akiyama:2020soe}. This work was followed by the application of the TRG method to analyze the metal-insulator transition of the $(1+1)d$ Hubbard model by calculating the electron density $\langle n\rangle$ as a function of the chemical potential $\mu$~\cite{Akiyama:2021xxr}. Our results for the critical chemical potential $\mu_{\rm c}$ and the critical exponent $\nu$ are consistent with an exact solution based on the Bethe ansatz~\cite{PhysRevLett.20.1445,LIEB20031}.
In this paper, we apply the TRG method to investigate the doping-driven metal-insulator transition in the $(2+1)d$ Hubbard model. \footnote{The model has also been investigated by the tensor network method based on the Hamiltonian formalism, like a fermionic PEPS, which is also free from the sign problem. For a recent study, see Ref.~\cite{Schneider:2021hqj}, for example.} The TRG method, which was originally proposed to study two-dimensional (2$d$) classical spin systems~\cite{Levin:2006jai}, has been developed to study wide varieties of fermionic models in particle physics~\cite{Shimizu:2014uva,Shimizu:2014fsa,Shimizu:2017onf,Takeda:2014vwa,Sakai:2017jwp,Yoshimura:2017jpk,Kadoh:2018hqq,Kadoh:2019ube,Kuramashi:2019cgs,Akiyama:2020ntf,Akiyama:2020soe,PhysRevD.101.094509}. It is also confirmed that the TRG method does not suffer from the sign problem by studying various quantum field theories~\cite{Shimizu:2014uva,Shimizu:2014fsa,Shimizu:2017onf,Takeda:2014vwa,Kawauchi:2017dnj,Kadoh:2018hqq,Kadoh:2019ube,Kuramashi:2019cgs,Akiyama:2020ntf,Akiyama:2020soe,Akiyama:2021xxr,Bloch:2021mjw,Nakayama:2021iyp}. We calculate the electron density $\langle n\rangle$ as a function of $\mu$ with three choices of $U=80$, $8$ and $2$.
The $\mu$ dependence of $\langle n\rangle$ allows us to determine the critical chemical potential $\mu_{\rm c}$ at the doping-driven metal-insulator transition from the half-filling plateau with $\langle n\rangle=1$ to the metallic state with $\langle n\rangle>1$. Our results at $U=80$, 8 and 2 show that $\vert \mu_{\rm c}- U/2\vert$ monotonically diminishes as $U$ decreases and seems to converge on $\vert \mu_{\rm c}- U/2\vert=0$ at $U=0$. This indicates the possibility that the model exhibits the metal-insulator transition over the wide region of the finite coupling.
This paper is organized as follows. In Sec.~\ref{sec:method} we define the Hubbard model in the path-integral formalism and give a brief description of the numerical algorithm. In Sec.~\ref{sec:results} we present the $\mu$ dependence of the electron density and determine the critical chemical potential $\mu_{\rm c}$ at the doping-driven metal-insulator transition. Section~\ref{sec:summary} is devoted to summary and outlook.
\section{Formulation and numerical algorithm}
\label{sec:method}
\subsection{(2+1)-dimensional Hubbard model in the path-integral formalism}
\label{subsec:action}
We consider the partition function of the Hubbard model in the path-integral formalism on an anisotropic lattice with the physical volume $V=L_x\times L_y\times \beta$, whose spatial extension is defined as $L_\sigma=aN_{\sigma}~(\sigma=x, y)$ with $a$ the spatial lattice spacing. $\beta$ denotes the inverse temperature, which is divided as $\beta=1/T=\epsilon N_\tau$.
Following Ref.~\cite{Akiyama:2021xxr}, the path-integral expression of the partition function is given by
\begin{align}
Z=\int\left(\prod_{n\in\Lambda_{2+1}}\prod_{s=\uparrow,\downarrow}{\rm d}\bar{\psi}_{s}(n){\rm d}\psi_{s}(n)\right){\rm e}^{-S},
\label{eq:Z}
\end{align}
where $n=(n_x,n_y,n_{\tau})\in\Lambda_{2+1}(\subset\mathbb{Z}^3)$ specifies a position on the lattice $|\Lambda_{2+1}|=N_x\times N_y\times N_\tau$. Since the Hubbard model describes the spin-1/2 fermions, they are labeled by $s=\uparrow,\downarrow$, corresponding to the spin-up and spin-down, respectively. Introducing the notation,
\begin{align}
\psi(n)=\left(
\begin{array}{c}
\psi_\uparrow(n)\\ \psi_\downarrow(n)
\end{array}
\right),
~\bar{\psi}(n)=\left(\bar{\psi}_\uparrow(n),\bar{\psi}_\downarrow(n)\right),
\end{align}
the action $S$ is defined as
\begin{align}
\label{eq:action}
S&=\sum_{n_\tau,n_x,n_y}\epsilon a^2\left\{\bar{\psi}(n)\left(\frac{\psi(n+{\hat \tau})-\psi(n)}{\epsilon}\right)\right.\nonumber\\
&\left.-t\sum_{\sigma=x,y}\left(\bar{\psi}(n+{\hat\sigma})\psi(n)+\bar{\psi}(n)\psi(n+{\hat\sigma})\right)+\frac{U}{2}\left(\bar{\psi}(n)\psi(n)\right)^2-\mu\bar{\psi}(n)\psi(n)\right\}.
\end{align}
The kinetic terms in the spatial directions contain the hopping parameter $t$. The four-fermi interaction term represents the Coulomb repulsion of electrons at the same lattice site. The chemical potential is denoted by the parameter $\mu$. Note that the half-filling is realized at $\mu=U/2$ in the current definition. We assume the periodic boundary condition in the spatial direction, $\psi(N_{x}+1,n_y,n_\tau)=\psi(1,n_y,n_\tau)$ and $\psi(n_x,N_{y}+1,n_\tau)=\psi(n_x,1,n_\tau)$, while the anti-periodic one in the temporal direction, $\psi(n_x,n_y,N_\tau+1)=-\psi(n_x,n_y,1)$. In the following discussion, we always set $a=1$.
\subsection{Numerical algorithm}
\label{subsec:algorithm}
Based on Ref.~\cite{Akiyama:2020sfo}, the tensor network representation of Eq.~\eqref{eq:Z} is immediately obtained. Set $d=2$ in the Appendix of Ref.~\cite{Akiyama:2021xxr} and one can find out the Grassmann tensor which generates the Grassmann tensor network of Eq.~\eqref{eq:Z}. The resulting Grassmann tensor $\mathcal{T}_{\Psi_{x}\Psi_{y}\Psi_{\tau}\bar{\Psi}_{\tau}\bar{\Psi}_{y}\bar{\Psi}_{x}}$ is of rank 6 and we evaluate $\mathrm{gTr}[\prod_{n}\mathcal{T}]$ with the anisotropic TRG (ATRG) algorithm \cite{Adachi:2019paf} whose extension to the Grassmann integrals, referred as the Grassmann ATRG (GATRG), is given in Ref.~\cite{Akiyama:2020soe}. We also follow the coarse-graining procedure employed in the previous study of $(1+1)d$ Hubbard model \cite{Akiyama:2021xxr}. Firstly, we carry out $m_{\tau}$ times of renormalization along with the temporal direction, which can be seen as the imaginary time evolution of the local Grassmann tensor. Secondly, $3d$ ATRG procedure is applied as the spacetime coarse-graining. As in the case of $(1+1)d$ Hubbard model, we have found that the optimal $m_{\tau}$ satisfies the condition $\epsilon 2^{m_{\tau}}\sim O(10^{-1})$ in the sense of preserved tensor norm.
\section{Numerical results}
\label{sec:results}
\subsection{Algorithmic-parameter dependence}
The partition function of Eq.~\eqref{eq:Z} is evaluated using the numerical algorithm explained above on lattices with the physical volume $V=L_x\times L_y\times \beta=N_x\times N_y\times \epsilon N_\tau$ with $(N_{x}, N_{y}, N_\tau)=(2^m,2^m,2^{m_{\tau}+m})~(m, m_{\tau}\in\mathbb{N})$.
We employ $U=80$, 8 and 2 for the four-fermi coupling with $t=1$ for the hopping parameter. In Fig.~\ref{fig:lnZ_delbeta} we plot the $\mu$ dependence of the thermodynamic potential $\ln Z/V$ at $U=8$ on $V=4096^2\times 1677.7216$ with the bond dimension $D=80$ in the GATRG algorithm choosing $\epsilon=2^{12}\times10^{-4},2^{8}\times10^{-4},2^{4}\times10^{-4},10^{-4}$. For each value of $\epsilon$, $m_{\tau}$ is chosen via the condition $\epsilon 2^{m_{\tau}}=2^{12}\times10^{-4}=O(10^{-1})$ following Ref.~\cite{Akiyama:2021xxr}. We find clear discretization effects for the coarsest case of $\epsilon=2^{12}\times10^{-4}$. On the other hand, the results with $\epsilon=2^{4}\times10^{-4}$ and $10^{-4}$ show good consistency. This means that the discretization effects with $\epsilon=10^{-4}$ are negligible. We employ $\epsilon=10^{-4}$ in the following calculations.
\begin{figure
\centering
\includegraphics[keepaspectratio,scale=0.5]{fig/lnz_eps.eps}
\caption{Thermodynamic potential at $U=8$ on $V=4096^2\times 1677.7216$ lattice. $\beta$ is divided with $\epsilon=2^{12}\times10^{-4}=0.4096$, $2^{8}\times10^{-4}=0.0256$, $2^{4}\times10^{-4}=0.0016$ and $10^{-4}=0.0001$. The bond dimension is set to be $D=80$.}
\label{fig:lnZ_delbeta}
\end{figure}
We investigate the convergence behavior of the thermodynamic potential by defining the quantity
\begin{align}
\delta=\left|\frac{\ln Z(D)-\ln Z(D=80)}{\ln Z(D=80)}\right|
\label{eq:delta}
\end{align}
on $V=4096^2\times 1677.7216$ lattice with $\epsilon=10^{-4}$. In Fig.~\ref{fig:lnZ_D}, we plot the $D$ dependence of $\delta$ at $(U,t)=(8,1)$ with the choices of $\mu=6.0, 7.5$ and 8.5. As we will see below, $\mu=6.0$ corresponds to $\langle n\rangle\approx1.0$ and $\mu=8.5$ does to $\langle n\rangle\approx1.5$. We observe that $\delta$'s at these values of $\mu$ decrease as a function of $D$, though some of them are fluctuating.
\begin{figure
\centering
\includegraphics[keepaspectratio,scale=0.5]{fig/delta.eps}
\caption{Convergence behavior of thermodynamic potential with $\delta$ of Eq.~\eqref{eq:delta} at $U=8$ with $\mu=6.0, 7.5$ and 8.5 as a function of $D$ on $V=4096^2\times 1677.7216$ lattice.}
\label{fig:lnZ_D}
\end{figure}
\subsection{Strong coupling limit}
We first consider the atomic limit at $(U,t)=(8,0)$. This case is analytically solvable. The electron density $\langle n\rangle$ is obtained by the numerical derivative of the thermodynamic potential in terms of $\mu$:
\begin{align}
\langle n\rangle=\frac{1}{V}\frac{\partial \ln Z(\mu)}{\partial \mu}\approx
\frac{1}{V}\frac{\ln Z(\mu+\Delta \mu)-\ln Z(\mu-\Delta \mu)}{2\Delta \mu}.
\end{align}
In Fig.~\ref{fig:edensity_t0} we compare the numerical and exact results for the $\mu$ dependence of $\langle n\rangle$ at $(U,t)=(8,0)$. Note that we set $m_{\tau}=24$ because this case is equivalent to the model defined on $V=1\times\beta$ lattice. Thanks to the vanishing hopping structure in the spatial direction, we can check the validity of the imaginary time evolution explained in Sec.~\ref{subsec:algorithm}.
The agreement of our numerical result with the exact solution shows that the imaginary time evolution carried out by the GATRG works precisely.
The deviation from the exact value defined by
\begin{align}
\delta_{\rm exact}(D)=\left|\frac{\ln Z(D)-\ln Z_{\rm exact}}{\ln Z_{\rm exact}}\right|
\label{eq:delta_exact}
\end{align}
is at most $O(10^{-4})$ in the range of $0\le \mu\le 16$. For $\mu<0$, the exact thermodynamic potential is equal to zero and the results obtained by the TRG are also equal to zero within a level of double precision.
Since the phase diagram of the metal-insulator transition in the (2+1)$d$ Hubbard model is not well known so far, we investigate the $\mu$ dependence of $\langle n\rangle$ choosing $U=80$ as a representative case in the strong coupling region. In Fig.~\ref{fig:edensity_U80} we plot the electron density $\langle n\rangle$ as a function of $\mu$ in the vicinity of $\mu\sim U$ with $D=80$. We have checked that the convergence behavior of $\delta$ at $U=80$ is better than that at $U=8$.
We observe that the electron density starts to increase from $\langle n\rangle=1$ at $\mu=77.0(2)$ or $\mu/U=0.9625(25)$ and reaches $\langle n\rangle=2$ with $\mu\gtrsim83.0$ or $\mu/U\gtrsim 1.04$. The $\mu$ dependence of $\langle n\rangle$ is smooth and continuous so that there is no signal of the first-order phase transition.
We expect that the critical chemical potential $\mu_{\rm c}$ at the doping-driven metal-insulator transition approaches to $\mu_{\rm c}/U=1$ toward the atomic limit and the transition from $\langle n\rangle=1$ to $\langle n\rangle=2$ becomes a step-function as a function of $\mu/U$.
\begin{figure
\centering
\includegraphics[keepaspectratio,scale=0.5]{fig/number_t0U8.eps}
\caption{Electron density $\langle n\rangle$ in the $(U,t)=(8,0)$ case at $\beta=1677.7216$ with $\epsilon=10^{-4}$ as a function of $\mu$. The solid line shows the exact solution and the blue circles are the results obtained by the TRG.}
\label{fig:edensity_t0}
\end{figure}
\begin{figure
\centering
\includegraphics[keepaspectratio,scale=0.5]{fig/number_t1U80.eps}
\caption{Electron density $\langle n\rangle$ as a function of $\mu$ at $(U,t)=(80,1)$ on $V=4096^2\times 1677.7216$ with $\epsilon=10^{-4}$. The bond dimension is set to be $D=80$.}
\label{fig:edensity_U80}
\end{figure}
\subsection{Critical chemical potential at $U=8$ and $2$}
\label{subsec:mu_c}
Now let us investigate the metal-insulator transition in the intermediate coupling region at $(U,t)=(8,1)$. There are a lot of previous work to investigate a possible superconducting phase expected in this coupling region.
Since we are interested in the thermodynamic and zero-temperature limit, we first check the volume dependence of the electron density $\langle n \rangle$.
In Fig.~\ref{fig:n_vol} we plot the $\mu$ dependence of $\langle n \rangle$ at $U=8$ changing the lattice sizes with $\epsilon=10^{-4}$, $m_{\tau}=12$ and $D=80$. We observe that the size of $(N_{x},N_{y},N_{\tau})=(2^{12},2^{12},2^{24})$, which corresponds to $V=4096^2\times 1677.7216$, is sufficiently large to be identified as the thermodynamic and zero-temperature limit. We observe the $\langle n \rangle=0$ plateau for $\mu\lesssim-4$ and the $\langle n\rangle=2$ one for $12\lesssim \mu$. The half-filling state is characterized by the plateau of $\langle n\rangle=1$ in the range of $2\lesssim \mu\lesssim 6$. These plateaus yield the vanishing compressibility $\kappa=\partial\langle n\rangle/\partial \mu$ indicating the insulating states.
\begin{figure
\centering
\includegraphics[keepaspectratio,scale=0.5]{fig/number_t1U8_vol.eps}
\caption{Electron density $\langle n\rangle$ at $U=8$ on two lattice sizes, $V=256^2\times 104.8576$ and $4096^2\times 1677.7216$, as a function of $\mu$. The bond dimension is set to be $D=80$.}
\label{fig:n_vol}
\end{figure}
\begin{figure
\centering
\includegraphics[keepaspectratio,scale=0.5]{fig/n_u8.eps}
\caption{Electron density $\langle n\rangle$ at $U=8$ on $V=4096^2\times 1677.7216$ with $\epsilon=10^{-4}$ as a function of $\mu$. The bond dimensions are $D=80$, $72$, $64$ and $56$. Fit results are drawn by the solid lines for each bond dimension.}
\label{fig:n_u8}
\end{figure}
In order to determine the critical chemical potential $\mu_{\rm c}$ in the limit of $D\rightarrow\infty$ at $U=8$ on $V=4096^2\times 1677.7216$ lattice we make a global fit of $\langle n\rangle$ with $D=80$, 72, 64 and 56 in the metallic phase near the transition point. In Fig.~\ref{fig:n_u8} we plot the results of $\langle n\rangle$ at $D=80$, $72$, $64$ and $56$ with a much finer resolution of $\Delta \mu$ than Fig.~\ref{fig:n_vol} focusing on the range of $6.0\le \mu\le8.2$, which covers the region of $1\le\langle n\rangle\le1.5$. This figure provides us a closer look at the $\mu$ dependence of $\langle n\rangle$ around $\mu_{\rm c}$. The results at $D=80$, $72$, $64$ and $56$ are almost degenerate indicating the small $D$ dependence. For the global fit we employ the following quadratic fitting function:
\begin{align}
\langle n\rangle=1+\alpha\left(\mu-\mu_{\rm c}(D)\right)+\beta\left(\mu-\mu_{\rm c}(D)\right)^2
\label{eq:n_fit}
\end{align}
with $\mu_{\rm c}(D)=\mu_{\rm c}(D=\infty)+\gamma/D$, where $\alpha$, $\beta$, $\gamma$ and $\mu_{\rm c}(D=\infty)$ are the fit parameters. The solid curves in Fig.~\ref{fig:n_u8} represent the fit results over the range of $6.3\le \mu\le8.0$. We obtain $\mu_{\rm c}(D=\infty)=6.43(4)$, which is presented in Table~\ref{tab:mu_c} together with other fitting results.
It may be instructive to compare Figs.~\ref{fig:n_vol}, \ref{fig:n_u8}, and the estimated location of $\mu_{\rm c}(D=\infty)$ with numerical data in Refs.~\cite{PhysRevB.67.085103,PhysRevB.87.035110}, though their calculations are carried out on very small lattice sizes and at low but finite temperatures.
\begin{table}[htb]
\caption{Critical chemical potential $\mu_{\rm c}(D=\infty)$ at $U=8$ and $2$ determined by the global fit of the data at $D=80$, $72$, $64$ and $56$.}
\label{tab:mu_c}
\begin{center}
\begin{tabular}{|c|cc|}\hline
$U$ & 8 & 2 \\
{\rm fit\; range} & [6.3, 8.0] & [1.2, 3.4] \\ \hline
$\mu_{\rm c}(D=\infty)$ & 6.43(4) & 1.30(6) \\ \hline
$\alpha$ & 0.372(9) & 0.39(1) \\ \hline
$\beta$ & $-$0.051(6) & $-$0.054(5) \\ \hline
$\gamma$ & $-$7(2) & $-$13(4) \\
\hline
\end{tabular}
\end{center}
\end{table}
We repeat the same analysis for the weak coupling case at $U=2$.
We apply the fit function of Eq.~\eqref{eq:n_fit} to four data sets with the bond dimensions of $D=80$, 72, 64 and 56.
Fit results are depicted in Fig.~\ref{fig:n_u2} and their numerical values are summarized in Table~\ref{tab:mu_c}.
Our results show that the deviation of $\vert\mu_{\rm c}(D)-U/2\vert$ diminishes as the Coulomb potential $U$ decreases. It is likely that $\vert\mu_{\rm c}(D)-U/2\vert$ vanishes only at $U=0$.
This means that the model exhibits the metal-insulator transition over the wide regime of the finite coupling, including the weak coupling region.
This conclusion may provide us a different scenario of the phase diagram from that predicted by the dynamical mean-field theory (DMFT)~\cite{RevModPhys.68.13}; there exists some $U_{\rm c}$ such that no metal-insulator transition occurs with $U<U_{\rm c}$.
\begin{figure
\centering
\includegraphics[keepaspectratio,scale=0.5]{fig/n_u2.eps}
\caption{Electron density $\langle n\rangle$ at $U=2$ on $V=4096^2\times 1677.7216$ with $\epsilon=10^{-4}$ as a function of $\mu$. The bond dimensions are $D=80$, $72$, $64$ and $56$. Fit results are drawn by the solid lines for each bond dimension.}
\label{fig:n_u2}
\end{figure}
\section{Summary and outlook}
\label{sec:summary}
We have investigated the doping-driven metal-insulator transition of the (2+1)$d$ Hubbard model in the path-integral formalism employing the TRG method. The electron density $\langle n\rangle$ is calculated in the wide range of $\mu$ corresponding to $0\le\langle n\rangle\le 2$.
We have also determined the critical chemical potential $\mu_{\rm c}$ at three values of $U$. Our results indicate that the deviation $\vert \mu_{\rm c}-U/2\vert$ vanishes only at $U=0$.
This means that the model exhibits the metal-insulator transition over the vast regime of the finite coupling $U$.
As a next step, it would be interesting to investigate the metal-insulator transition of the (3+1)$d$ Hubbard model.
\section*{Acknowledgment}
Numerical calculation for the present work was carried out with the supercomputer Fugaku provided by RIKEN (Project ID: hp200314) and also with the Oakforest-PACS (OFP) computer under the Interdisciplinary Computational Science Program of Center for Computational Sciences, University of Tsukuba. This work is supported in part by Grants-in-Aid for Scientific Research from the Ministry of Education, Culture, Sports, Science and Technology (MEXT) (No. 20H00148) and JSPS KAKENHI Grant Number JP21J11226 (S.A.).
\bibliographystyle{ptephy}
|
\section{Introduction}
This paper concerns convergence acceleration methods for fixed-point (FP) iterations of the type
\begin{equation}\label{eq:fixed-point}
x_{k+1}=q(x_{k}), \quad x_k\in\mathbb{R}^n, \quad k=0,1,2,\ldots, \tag{FP}
\end{equation}
that seek to approximate a fixed point $x^*=q(x^*)$.
Specifically, we consider the following nonlinear acceleration iteration with window size $m$:
\begin{equation}\label{eq:AA-iteration}
x_{k+1}= q(x_k) + \sum_{i=1}^{\min(k,m)}\beta_{i}^{(k)}(q(x_k)-q(x_{k-i})) \qquad k=0,1,2,\ldots,
\end{equation}
where the coefficients $\beta_{i}^{(k)}$ are determined by solving a small optimization
problem in every step $k$ that minimizes a linearized residual in the new iterate $x_{k+1}$. Method
\cref{eq:AA-iteration} is known as \emph{Anderson acceleration (AA)} \cite{anderson1965iterative}.
More precisely, defining the residuals $r(x)$ of the fixed-point iteration by
\begin{equation}\label{eq:resid}
r(x)=x-q(x),
\end{equation}
AA($m$), with window size $m$, solves in every iteration the optimization problem
\begin{equation}\label{eq:Andersonbetas}
\min_{\{\beta_i^{(k)} \}} \bigg\| r(x_k) + \sum_{i=1}^{\min(k,m)} \beta_i^{(k)} ( r(x_k) - r(x_{k-i})) \bigg\|,
\end{equation}
with up to $m$ variables, and optimization problem \cref{eq:Andersonbetas} is normally posed in the $2$-norm.
Our discussion will focus on the case where iteration \cref{eq:fixed-point} is, by itself, a convergent iteration, but this is, in fact, not necessary for Anderson iteration \cref{eq:AA-iteration} to converge or be effective.
Assume that $k>m$. Define $r_k =x_k-q(x_k)$ and
\begin{equation}\label{eq:beta-vector-form}
\boldsymbol{\beta}^{(k)} =\begin{bmatrix} \beta_1^{(k)} \\ \vdots \\ \beta_m^{(k)} \end{bmatrix},\quad
R_k = \begin{bmatrix} r_k-r_{k-1} & r_k-r_{k-2} & \cdots & r_k-r_{k-m} \end{bmatrix}.
\end{equation}
Then, using the 2-norm in \cref{eq:Andersonbetas}, the solution of the least-squares problem is given by
\begin{equation}\label{eq:AAm-beta-form}
\boldsymbol{\beta}^{(k)} = -(R_k^TR_k)^{-1} R_k^Tr_k,
\end{equation}
if $R_{k}^TR_k$ is invertible. Or more generally, we can write
\begin{equation}\label{eq:AAm-beta-form-pseudo}
\boldsymbol{\beta}^{(k)} = -R_k^{\dag}r_k,
\end{equation}
where $R_k^{\dag}$ is the pseudo-inverse of $R_k$, and we note that
\begin{equation}\label{eq:two-ways-pseudo-inverse}
R_k^{\dagger}= \big(R_k^TR_k\big)^{\dagger} R_k^T.
\end{equation}
This covers the case where $R_{k}^TR_k$ is not invertible, by taking $\boldsymbol{\beta}^{(k)}$ as
the minimum-norm solution of the least-squares problem in this case.
The specific case of AA($m$) with $m=1$ in \cref{eq:AA-iteration} reads
\begin{equation}
\label{eq:anderson-1-step}
x_{k+1} = (1+\beta_k) q(x_k) -\beta_k q(x_{k-1}),
\end{equation}
where we have defined
$$\beta_k=\beta_1^{(k)}.$$
When $m=1$ and $r_k \ne r_{k-1}$,
\begin{equation}\label{eq:AA-1-step-beta}
\beta_k = \displaystyle \frac{-r_k^T(r_k-r_{k-1})}{\|r_k-r_{k-1}\|^2}.
\end{equation}
When $r_k = r_{k-1}$, we can, according to \cref{eq:AAm-beta-form-pseudo}, take $\beta_k=0$.
Let $\boldsymbol{z}_k=\begin{bmatrix} x_k \\ x_{k-1}\end{bmatrix}$. Then we can in turn write AA(1) as
a fixed-point iteration,
\begin{equation}\label{eq:AA-fixed-point}
\boldsymbol{z}_{k+1} =\Psi(\boldsymbol{z}_k), \tag{AA}
\end{equation}
with
\begin{equation}\label{eq:AA(1)-system}
\Psi(\boldsymbol{z}_k)
= \begin{bmatrix}
q(x_k)+\beta(\boldsymbol{z}_k)\big( q(x_k)-q(x_{k-1})\big)\\
x_k
\end{bmatrix},
\end{equation}
where $\beta(\boldsymbol{z}_k) =\beta_k$. As we explain in some more detail below, AA($m$) for $m>1$ can also be written in the form of fixed-point iteration \cref{eq:AA-fixed-point} using a similar lifting approach, with $\boldsymbol{z}_k \in \mathbb{R}^{n(m+1)}$. In what follows, vectors such as $\boldsymbol{z}_k$ that live in the augmented space $\mathbb{R}^{n(m+1)}$ will be indicated by bold font.
In this paper, we are interested in how the asymptotic convergence speed of iteration \cref{eq:fixed-point} relates to the asymptotic convergence speed of the accelerated iteration \cref{eq:AA-fixed-point}. Specifically, we consider the case where $q(x)$ is differentiable at $x^*$, such that the asymptotic convergence of \cref{eq:fixed-point} is linear. We seek to investigate the improvement in asymptotic convergence speed resulting from the acceleration of \cref{eq:fixed-point} by \cref{eq:AA-fixed-point}.
For reasons that will become clear below, the relevant notion of convergence is root-linear (or \emph{r-linear}) convergence \cite{ortega2000iterative}:
\begin{definition}[r-linear convergence of a sequence]
Let $\{x_k\}$ be any sequence that converges to $x^*$. Define
\begin{equation*}
\rho_{\{x_k\}} = \limsup\limits_{k\rightarrow \infty}\|x^*-x_k\|^{\frac{1}{k}}.
\end{equation*}
We say $\{x_k\}$ converges r-linearly with r-linear convergence factor $\rho_{\{x_k\}}$ if $\rho_{\{x_k\}}\in(0,1)$, and $r$-superlinearly if $\rho_{\{x_k\}}=0$. The ``r-'' prefix stands for ``root''.
\end{definition}
Since the r-linear convergence factor of an iteration sequence resulting from $x_{k+1}=q(x_k)$, for a given iteration function $q(x)$, may depend on the initial guess $x_0$ and on the specific fixed point $x^*$ the iteration converges to in the case multiple fixed points exist, we need to consider the worst-case r-linear convergence factor for convergence of method $x_{k+1}=q(x_k)$ to a specific fixed point $x^*$:
\begin{definition}[r-linear convergence of a fixed-point iteration]\label{def:rho-method}
Consider fixed-point iteration $x_{k+1}=q(x_k)$. We define the set of iteration sequences that converge to a given fixed point $x^*$ as
\begin{equation*}
C(q, x^*)=\Big\{ \{x_k\}_{k=0}^{\infty} | \quad x_{k+1} = q(x_k) \textrm{ for } k=0,1,\ldots, \textrm{ and } \lim_{k \rightarrow \infty}x_k=x^*\Big\},
\end{equation*}
and the worst-case r-linear convergence factor over $C(q, x^*)$ is defined as
\begin{equation}\label{eq:r-factor-defi}
\rho_{q,x^*} = \sup \Big\{ \rho_{\{x_k\}} |\quad \{x_k\}\in C(q, x^*) \Big\}.
\end{equation}
We say that the FP method converges r-linearly to $x^*$ with r-linear convergence factor $\rho_{q,x^*}$ if $\rho_{q,x^*} \in(0,1)$.
\end{definition}
The following classical theorem (see, e.g., \cite{ortega2000iterative}) shows that, if the iteration function $q(x)$ in
\cref{eq:fixed-point} is differentiable at $x^*$, the worst-case r-linear convergence factor, $\rho_{q,x^*}$, is determined by the spectral radius of the Jacobian $q'(x)$ evaluated at $x^*$:
\begin{theorem}\label{thm:Ostrowski-Theorem}[Ostrowski Theorem]
Suppose that $q: D\subset \mathbb{R}^n \rightarrow \mathbb{R}^n$ has a
fixed point $x^*$ that is an interior point of $D$, and is differentiable at $x^*$. If the spectral radius
of $q'(x^*)$ satisfies $0<\rho(q'(x^*)) < 1$, then the FP method converges r-linearly with $\rho_{q,x^*} = \rho(q'(x^*))$.
\end{theorem}
It is useful to consider the special case where the iteration functions $q(x)$ in \cref{eq:fixed-point} is affine, i.e.,
$q(x)=M \, x+b$ and
\begin{equation}\label{eq:fixed-point-linear}
x_{k+1}=M \, x_{k} + b, \quad k=0,1,2,\ldots,
\end{equation}
where $x_k, b\in\mathbb{R}^n$ and $M\in\mathbb{R}^{n \times n}$.
Since the error propagation equation for iteration \cref{eq:fixed-point-linear} is the linear iteration
\begin{equation}\label{eq:fixed-point-linear-error}
e_{k+1}=M \, e_{k} \quad k=0,1,2,\ldots,
\end{equation}
where the error of iterate $x_k$ is defined by $e_k=x^*-x_k$,
we call iteration \cref{eq:fixed-point} linear when $q(x)$ is affine, and nonlinear otherwise.
It is well-known that, in the linear case, AA($m$) with infinite window size is essentially equivalent
to the GMRES iterative method applied to $(I-M) \, x=b$ \cite{walker2011anderson}.
Given an iteration \cref{eq:fixed-point} and a set of initial conditions $x_0$ that converge to a fixed point $x^*$, we can define the set of all sequences in $C(q, x^*)$ that converge with a smaller r-linear convergence factor than $\rho_{q,x^*}$ as
\begin{equation}\label{eq:S-definition}
S = \Big\{ \{x_k\}\in C(q, x^*) | \quad \rho_{\{x_k\}} <\rho_{q,x^*} \Big\}.
\end{equation}
In the linear case it is easy to see that, when $M$ is diagonalizable and has eigenvalues that are not all of equal magnitude, $\rho_{\{x_k\}}= \rho_{q,x^*}$, except for initial conditions $x_0$ that lie in a set of measure zero in $\mathbb{R}^n$. That is, in this case the set $S$ has measure zero ($|S|=0$) in $C(q, x^*)$.
At this point it is useful to consider a simple example to motivate the questions we address in this paper.
The simple linear example we consider is\\
\begin{problem}\label{prob:linear2x2}
\begin{equation}\label{eq:q-linear-2x2-simple}
x_{k+1} = M x_k, \qquad
M=\begin{bmatrix}
2/3 & 1/4\\
0 & 1/3
\end{bmatrix}.
\end{equation}
Clearly, the eigenvalues of $M$ are $\lambda_1=2/3$ and $\lambda_2=1/3$.
\end{problem}
\begin{figure}[h]
\centering
\includegraphics[width=.49\textwidth]{fig15_prob04_mc_rconv.pdf}
\includegraphics[width=.49\textwidth]{fig15_prob04_mc_betas_eratios.pdf}
\caption{\cref{prob:linear2x2} with initial guess $x_0=[0.2,0.1]^T$: (left panel) Root-averaged error $\sigma_k$ as a function of iteration number $k$ for FP iteration \cref{eq:fixed-point} and AA(1) iteration \cref{eq:anderson-1-step}. (right panel) AA(1) coefficient $\beta_k$ and error ratio $e_k/e_{k-1}$ as a function of iteration number $k$.}
\label{fig:simple-beta}
\end{figure}
\cref{fig:simple-beta} (left panel) shows convergence curves for the root-averaged error
\begin{equation}\label{eq:sigma}
\sigma_k = \|x^*-x_k\|^{ \frac{1}{k}},
\end{equation}
of both the FP iteration \cref{eq:q-linear-2x2-simple} and its AA(1) acceleration \cref{eq:anderson-1-step},
for initial condition $x_0=[0.2,0.1]^T$.
It is easy to see that, for all initial conditions, except when $x_0$ lies in the eigenvector direction of $\lambda_2=1/3$, $\sigma_k$ for FP iteration \cref{eq:q-linear-2x2-simple} must converge to $\rho(M)=\lambda_1=2/3$; the FP $\sigma_k$ convergence curve in \cref{fig:simple-beta} is consistent with this, and confirms that the sequence $\{x_k\}$ generated by FP iteration \cref{eq:q-linear-2x2-simple} converges r-linearly with convergence factor $\rho_{\{x_k\}}=\rho(M)=\lambda_1$. This is also consistent with \cref{thm:Ostrowski-Theorem}, with $q'(x^*)=M$.
The left panel of \cref{fig:simple-beta} also indicates that the AA(1) sequence $\{x_k\}$ converges r-linearly:
$\sigma_k$ for AA(1) appears to converge to a value $\in (0,1)$ that is smaller than $\rho(M)=2/3$, indicating asymptotic acceleration of
FP iteration \cref{eq:q-linear-2x2-simple} by AA(1).
The right panel of \cref{fig:simple-beta} shows, perhaps surprisingly, that the $\beta_k$ sequence of AA(1) does not converge as $k \rightarrow \infty$: it oscillates as $x_k$ converges to $x^*$. The figure indicates that this is related to oscillations in the error ratio $e_k/e_{k-1}$: the error ratio $e_k/e_{k-1}$ does not converge but oscillates as
$k \rightarrow \infty$ and $x_k \rightarrow x^*$, reflecting the well-known fact that q-linear convergence is often not obtained in cases where \cref{thm:Ostrowski-Theorem} guarantees r-linear convergence. Further numerical results in this paper will show that this convergence behavior is generic, both in the case of linear and nonlinear iterations \cref{eq:fixed-point}, and also for window size $m>1$: in most cases, $\{x_k\}$ for AA($m$) applied to \cref{eq:fixed-point} converges r-linearly, but $\boldsymbol{\beta}^{(k)}$ oscillates as $k \rightarrow \infty$. This paper will provide analysis of the AA($m$) fixed-point function $\Psi(\boldsymbol{z})$ in iteration \cref{eq:AA-fixed-point} that sheds light on the mechanism by which AA($m$) can converge r-linearly while $\boldsymbol{\beta}^{(k)}$ does not converge.
\begin{figure}[h]
\centering
\includegraphics[width=.49\textwidth]{fig1_prob04_mc_rconv.pdf}
\includegraphics[width=.49\textwidth]{fig1_prob04_mc_betas.pdf}
\includegraphics[width=.325\textwidth]{fig1_prob04_mc_inits.pdf}
\includegraphics[width=.325\textwidth]{fig1_prob04_mc_AAhist.pdf}
\includegraphics[width=.325\textwidth]{fig1_prob04_mc_qhist.pdf}
\caption{Convergence behavior for \cref{prob:linear2x2} with 1,000 random initial guesses. The green and red $\sigma_k$ curves in the top left panel correspond to the random initial conditions $x_0$ indicated in the top right panel. The middle panels show histograms for the final values of the red and green curves in the top-left panel at iteration $k=100$. The bottom panel shows $\beta_k$ sequences for the 1,000 initial conditions.} \label{fig:simple-MC}
\end{figure}
Next, \cref{fig:simple-MC} shows further numerical results for the simple \cref{prob:linear2x2} of \cref{eq:q-linear-2x2-simple}
that identify additional convergence properties of AA(1) which will be investigated in this paper. In the tests of \cref{fig:simple-MC}, we run the FP iteration \cref{eq:q-linear-2x2-simple} and its AA(1) acceleration \cref{eq:anderson-1-step} for 1,000 initial conditions $x_0$ that are chosen uniformly randomly within the square $[-0.25,0.25]^2$.
As expected, for the FP iteration \cref{eq:q-linear-2x2-simple} $\sigma_k$ converges to $\rho(M)=\lambda_1=2/3$ for all random initial guesses, corresponding to $S$ from \cref{eq:S-definition} satisfying $|S|=0$. For the AA(1) acceleration \cref{eq:anderson-1-step}, however, the numerical results indicate that the iteration sequences $\{x_k\}$ still each converge r-linearly, but the r-linear convergence factors $\rho_{\{x_k\}}$ now strongly depend on the initial condition, indicating that $|S|>0$ for AA(1) applied to the linear \cref{prob:linear2x2}.
Furthermore, the $\beta_k$ convergence curves for AA(1) depend on the initial condition, but they oscillate and do not converge.
And finally, the numerical results suggest that the accelerated iteration \cref{eq:AA-fixed-point} for \cref{prob:linear2x2}
may have an asymptotic r-linear convergence factor $\rho_{\Psi,x^*}$, see \cref{def:rho-method},
that is strictly smaller than $\rho_{q,x^*}=\rho(M)=2/3$ of iteration \cref{eq:fixed-point}, indicating an improved AA(1) asymptotic converge factor. As we explain in some more detail below, there are no known theoretical results that can quantify the asymptotic convergence improvement of AA($m$) compared to iteration \cref{eq:fixed-point}, not even for simple specific functions $q(x)$ such as the 2 $\times 2$ linear case of \cref{prob:linear2x2}. The numerical results in \cref{fig:simple-MC} are an indication that an upper bound for $\rho_{\Psi,x^*}$ that is $< \rho_{q,x^*}$ should exist for \cref{prob:linear2x2}, and further numerical results in this paper suggest the same for higher-dimensional and nonlinear iterations \cref{eq:fixed-point}. As explained below, this open question motivates the analysis of the AA($m$) fixed-point function $\Psi(\boldsymbol{z})$ in iteration \cref{eq:AA-fixed-point} that is the subject of this paper.
At this point it is useful to recall in some detail what is known about convergence of AA($m$).
While Anderson acceleration method \cref{eq:AA-iteration} dates back to 1965 \cite{anderson1965iterative} and has since been used to speed up convergence for fixed-point methods in many areas of scientific computing with often excellent results, very little was known about the convergence of AA($m$) until the 2015 paper \cite{toth2015convergence}. In this paper, Toth and Kelley proved two convergence results.
First, they showed that, for the linear case \cref{eq:fixed-point-linear} where $q(x)=Mx+b$, if $\|M\|=c<1$, $x_k$ converges at least r-linearly with r-linear convergence factor not worse than $c$, for any initial guess. This result is important in that it establishes convergence of AA($m$), but it does not provide information on AA($m$) actually accelerating the fixed-point convergence asymptotically. Also, it only covers the case where $\|M\|=c<1$, and thus excludes practically relevant cases such as when $\rho(M)<1$ with $\|M\|>1$.
Second, \cite{toth2015convergence} also considered the nonlinear case. Assuming that the AA($m$) coefficients are bounded, i.e., $\sum_{i=1}^{m}|\beta^{(k)}_i| < c_\beta \ \forall k$, that $q(x)$ is differentiable with $\|q'(x)\|\leq c<1$, and that $q'(x)$ is Lipschitz continuous, it was shown in \cite{toth2015convergence} that $x_k$ converges at least r-linearly with r-linear convergence factor not worse than $c$, for any initial guess sufficiently close to $x^*$. This result, however, also does not show an asymptotic improvement over $c$, and the boundedness of the $\beta^{(k)}_i$ remains as a strong assumption.
As far as we are aware, no further results have been obtained that can quantify the improvement in linear asymptotic convergence factor $\rho_{\Psi,x^*}$ that is often observed for AA($m$) with finite $m$, compared to the linear asymptotic convergence factor $\rho_{q,x^*}$ of iteration \cref{eq:fixed-point} when it converges r-linearly.
In more recent work, Evans et al.\ \cite{evans2020proof} were able to quantify the per-iteration convergence improvement of AA($m$).
They showed that, to first order, the convergence gain provided by AA in step $k$ is quantified by a factor $\theta_k \le 1$ that equals the ratio of the optimal value defined in \cref{eq:Andersonbetas} to $\|r(x_k)\|_2$.
However, $\theta_k \le 1$ may oscillate, and it is not clear how $\theta_k$ may be evaluated or bounded in practice or how it may translate to an improved linear asymptotic convergence factor $\rho_{\Psi,x^*}$ for AA($m$) compared to $\rho_{q,x^*}$ as observed, for example, in the numerical results of \cref{fig:simple-MC}.
The strong dependence in \cref{fig:simple-MC} of the AA(1) r-linear convergence factors $\rho_{\{x_k\}}$ on the initial guess may, at first, seem surprising. In light of \cref{thm:Ostrowski-Theorem}, one would expect, for the accelerated iteration \cref{eq:AA-fixed-point}, that the Jacobian of $\Psi(\boldsymbol{z})$ evaluated at the fixed point $\boldsymbol{z}^*$ would determine the linear convergence factor $\rho_{\{x_k\}}$ for most initial guesses, if $\Psi'(\boldsymbol{z})$ were differentiable at $\boldsymbol{z}^*$. The strong dependence of the r-linear convergence factors $\rho_{\{x_k\}}$ on the initial guess in \cref{fig:simple-MC} suggests otherwise. To shed light on numerical observations as in \cref{fig:simple-MC}, we analyze in this paper the differentiability properties of the AA($m$) fixed-point iteration function $\Psi(\boldsymbol{z})$ in iteration \cref{eq:AA-fixed-point}. We find, indeed, that, while being Lipschitz continuous, $\Psi(\boldsymbol{z})$ is not differentiable at $\boldsymbol{z}^*$. Further analysis reveals, however, that $\Psi(\boldsymbol{z})$ is directionally differentiable at $\boldsymbol{z}^*$ in all directions, and we obtain closed-form expressions for these directional derivatives. This allows us to compute the Lipschitz constant of $\Psi(\boldsymbol{z})$ at $\boldsymbol{z}^*$, and to investigate whether this Lipschitz constant may relate to numerically observed AA($m$) convergence factors as in \cref{fig:simple-MC}.
In addition to the fixed-point analysis presented in this paper, it has to be noted that further insight in the convergence behavior revealed by numerical tests as in \cref{fig:simple-MC} may be obtained from formulating AA($m$) with finite $m$ as a Krylov subspace method and deriving further theoretical properties of AA($m$) iterations from that formulation. Such a Krylov formulation is developed for AA($m$) in a companion paper \cite{AAKrylov} to the current paper. This leads to further explanations for numerical observations as in \cref{fig:simple-MC}, including, for example, the apparent gap in the AA(1) $\sigma_k$ spectrum that can be observed in the top-left panel of \cref{fig:simple-MC}.
The remainder of this paper is organized as follows.
\Cref{sec:AA1} provides a detailed analysis of the continuity and differentiability of the AA(1) fixed-point iteration function $\Psi(\boldsymbol{z})$ in \cref{eq:AA-fixed-point}, and \Cref{sec:AAm} extends this analysis to AA($m$). We investigate the continuity of the AA coefficients $\beta^{(k)}_i$ and of the fixed-point iteration function $\Psi(\boldsymbol{z})$ at the fixed point $\boldsymbol{z}^*$, and consider Lipschitz continuity, directional differentiability and differentiability of $\Psi(\boldsymbol{z})$. Some of the longer proofs are relegated to appendices.
\Cref{sec:numerics} contains numerical results that further illustrate how our theoretical findings relate to the asymptotic convergence of AA($m$), both for linear and nonlinear iterations \cref{eq:fixed-point} and in higher dimensions than in the simple $2 \times 2$ problem of \cref{fig:simple-MC}. To place our results in a broader context, we also compare AA($m$) convergence behavior in the linear case with GMRES and restarted GMRES($m$). We conclude in \Cref{sec:disc}.
\section{Analysis of $\Psi(\boldsymbol{z})$ for AA(1)}
\label{sec:AA1}
In this paper we analyze the continuity and differentiability properties of the AA($m$) fixed-point iteration function $\Psi(\boldsymbol{z})$ in iteration \cref{eq:AA-fixed-point}. To aid the analysis of $\Psi(\boldsymbol{z})$, it will also be useful to study the continuity of the AA($m$) coefficients $\boldsymbol{\beta}^{(k)}$ given in \cref{eq:AAm-beta-form-pseudo}.
We first consider, in this section, the case where $m=1$. We present results that cover general, nonlinear iteration functions $q(x)$. In \Cref{sec:AAm} we extend these results to $m>1$.
Throughout this paper, we assume that $q(x)$ in iteration \cref{eq:fixed-point} is continuously differentiable in a neighborhood of $x^*\in\mathbb{R}^{n}$, such that the Jacobian matrix $q'(x)\in\mathbb{R}^{n \times n}$ exists and is continous.
First, for AA(1), let us redefine $\Psi(\boldsymbol{z})$ in \cref{eq:AA(1)-system} for iteration \cref{eq:AA-fixed-point} as follows:
\begin{equation}\label{eq:Psi1}
\Psi(\boldsymbol{z}) =
\begin{bmatrix}
q(x) +\beta(\boldsymbol{z}) (q(x)-q(y))\\
x
\end{bmatrix},
\end{equation}
where
\begin{equation}\label{eq:z1}
\boldsymbol{z} =\begin{bmatrix} x\\ y\end{bmatrix} \quad \in\mathbb{R}^{2n},
\end{equation}
and $\beta_k$ in \cref{eq:AA-1-step-beta} is written as
\begin{equation}\label{eq:beta1}
\beta(\boldsymbol{z}) =
\begin{cases}
\displaystyle \displaystyle \frac{-r^T(x)(r(x)-r(y))}{\|r(x)-r(y)\|^2}, & \text{if}\,\, r(x)\neq r(y), \\
0, & \text{if}\,\, r(x)=r(y),
\end{cases}
\end{equation}
with $r(x) =x-q(x)$.
Note that when $\boldsymbol{z}=\boldsymbol{z}^*=\displaystyle \begin{bmatrix} x^*\\ x^*\end{bmatrix}$, $\Psi(\boldsymbol{z})=\boldsymbol{z}$.
For the linear case, we define $q(x)$ in \cref{eq:fixed-point} as
\begin{equation}\label{eq:linear-q-form}
q(x) = Mx + b, \quad M \in\mathbb{R}^{n \times n}, \quad x, \, b \in\mathbb{R}^{n}
\end{equation}
where one seeks to solve $A\,x=b$, with
\begin{equation}
A=I-M=I-q'(x).
\end{equation}
In the linear case, we will assume that matrix $A=I-q'(x)$ in $A\,x=b$ is nonsingular.
Similarly, we will usually assume in the nonlinear case that $r'(x)=I-q'(x)$ is nonsingular. We also exclude the trivial case where $A=I$ and $M=0$, or, more generally, $q'(x)=0$.
In the linear case \cref{eq:beta1} simplifies to
\begin{equation}
\beta(\boldsymbol{z}) =
\begin{cases}
\displaystyle \frac{-(Ax-b)^TA(x-y)}{(x-y)^TA^TA(x-y)}, & \text{if}\,\, x\neq y, \\
0, & \text{if}\,\, x=y,
\end{cases}
\end{equation}
and when $x\neq y$
\begin{equation}\label{eq:Psilin}
\Psi(\boldsymbol{z}) =
\begin{bmatrix}
(I-A)x+b- \displaystyle \frac{(Ax-b)^TA(x-y)}{(x-y)^TA^TA(x-y)}(I-A)(x-y)\\
x
\end{bmatrix}.
\end{equation}
When $x= y$,
\begin{equation}\label{eq:form-Psi-x=y-linear}
\Psi(\boldsymbol{z}) =
\begin{bmatrix}
q(x)\\
x
\end{bmatrix}=\begin{bmatrix}
Mx+b\\
x
\end{bmatrix}.
\end{equation}
Before providing our detailed analysis of the differentiability properties of $\Psi(\boldsymbol{z})$, we summarize our results in \cref{tab:continuity-differentiability-Lip}. While the proof for some of these results is elementary, the table provides a complete overview of the differentiability properties of $\Psi(\boldsymbol{z})$ which are useful to understand the convergence behavior of AA(1) viewed as the fixed-point method \cref{eq:AA-fixed-point}.
\newsavebox{\smlmat
\savebox{\smlmat}{$\small \boldsymbol{z} =\left[ \begin{array}{c} x \\ y \end{array}\right]$}
\begin{table}[h!]
\caption{Continuity and differentiability properties of $\beta(\boldsymbol{z})$ and $\Psi(\boldsymbol{z})$ at~\usebox{\smlmat}
for AA($m$) iteration \cref{eq:AA-fixed-point} with $m=1$, where $\beta(\boldsymbol{z})$
and $\Psi(\boldsymbol{z})$ are given by \cref{eq:beta1}
and \cref{eq:Psi1}.
}
\centering
\begin{tabular}{|l|c|c|c|}
\hline
&$r(x)\neq r(y)$ &$x=y, \ r(x) \neq 0$ & $x=y=x^*$ \\
\hline continuity of $\beta(\boldsymbol{z})$ &$\surd$ &$\times$ & $\times$ \\
\hline continuity of $\Psi(\boldsymbol{z})$ &$\surd$ &$\times$ & $\surd$ \\
\hline Lipschitz continuity of $\Psi(\boldsymbol{z})$ &$\surd$ &$\times$ & $\surd$ \\
\hline Gateaux-differentiability of $\Psi(\boldsymbol{z})$ &$\surd$ &$\times$ &$\surd$ \\
\hline differentiability of $\Psi(\boldsymbol{z})$ &$\surd$ &$\times$ & $\times$ \\
\hline
\end{tabular}\label{tab:continuity-differentiability-Lip}
\end{table}
\subsection{Continuity of $\beta(\boldsymbol{z})$ for AA(1)}
\begin{proposition}
$\beta(\boldsymbol{z})$ in \cref{eq:beta1} is continuous at $\boldsymbol{z} =\begin{bmatrix} x\\ y\end{bmatrix}$ when $r(x)\neq r(y)$.
\end{proposition}
\begin{proof}
Since $r(x)$ is a continuous function, $r(x)-r(y)$ and $\|(r(x)-r(y))\|^2$ are continuous functions. It follows that $\beta(\boldsymbol{z})$ in \cref{eq:beta1} is continuous when $r(x)\neq r(y)$.
\end{proof}
\begin{proposition}\label{prop:betaxx}
$\beta(\boldsymbol{z})$ in \cref{eq:beta1} is not continuous at $\boldsymbol{z} =\begin{bmatrix} x\\ y\end{bmatrix}$ when $x= y$ with $r(x) \neq 0$.
\end{proposition}
\begin{proof}
Consider $\boldsymbol{z}=\begin{bmatrix} x\\ y+d\end{bmatrix},$ where $x= y$ with $r(x) \neq 0$.
It is sufficient to find a path for $d$ along which $\beta(\boldsymbol{z})$ is not continuous as $d \rightarrow 0$.
Consider the case where $d=\epsilon\, e$, with $e$ a unit vector
in $\mathbb{R}^n$ and $r(x)-r(y+d(\epsilon))\neq 0$. Then
\begin{align*}
\beta(\boldsymbol{z}(\epsilon))& = - \displaystyle \frac{r(x)^T(r(x)-r(y+d(\epsilon)))}{\|r(x)-r(y+d(\epsilon))\|^2},\\
&= - \displaystyle \frac{r(x)^T(r(x)-r(y+\epsilon e))}{\|r(x)-r(y+\epsilon e)\|^2}.
\end{align*}
Since $r(y+\epsilon e)=r(y)+ r'(y)\epsilon e +Q(\epsilon e) \epsilon e$ with $\lim_{\epsilon \rightarrow 0} Q(\epsilon e)=0$,
we have
\begin{align*}
\lim_{\epsilon \rightarrow 0} \beta(\boldsymbol{z}(\epsilon))& =
\lim_{\epsilon \rightarrow 0} \displaystyle \frac{r(x)^T (r'(y) \epsilon e + Q(\epsilon e) \epsilon e) }{\|r'(y)\epsilon e +Q(\epsilon e) \epsilon e\|^2},
\end{align*}
where $r'(y)e +Q(\epsilon e) e\neq 0$ for sufficiently small $\epsilon$, since $r'(y)$ is nonsingular. Since $r'(y)$ is nonsingular, we have that $r(x)^T r'(y) e \neq 0$ for all unit vectors $e$, except for the unit vectors orthogonal to
$r'(y)^Tr(x)\neq 0$.
So for almost all unit vectors $e$ we have that
\begin{align*}
\beta(\boldsymbol{z}(\epsilon))& =
\displaystyle \frac{r(x)^T (r'(y) e + Q(\epsilon e) e) }{\epsilon \ \|r'(y) e +Q(\epsilon e) e\|^2} \rightarrow \pm \infty \textrm{ as } \epsilon \rightarrow 0.
\end{align*}
Thus, $\beta(\boldsymbol{z})$ is not continuous at $\boldsymbol{z} =\begin{bmatrix} x\\ y\end{bmatrix}$ when $x= y$ with $r(x) \neq 0$.
\end{proof}
\begin{remark}
\cref{prop:betaxx} also holds in the more general case when $r(x)= r(y) \neq r(x^*)$.
\end{remark}
\begin{proposition}\label{prop:betax*}
$\beta(\boldsymbol{z})$ in \cref{eq:beta1} is not continuous at $\boldsymbol{z} =\begin{bmatrix} x^*\\ x^*\end{bmatrix}$.
\end{proposition}
\begin{proof}
We investigate the limiting behavior of $\beta(\boldsymbol{z})$ along radial paths approaching $\boldsymbol{z}^* =\begin{bmatrix} x^*\\ x^*\end{bmatrix}$. We set $\boldsymbol{z}(\epsilon)=\begin{bmatrix} x^*+\epsilon d_1\\ x^*+\epsilon d_2\end{bmatrix}$.
Note that $\beta(\boldsymbol{z}(\epsilon))=0$ when $d_1 = d_2$. When $d_1 \neq d_2$ and $r(x^*+\epsilon d_1)-r(x^*+\epsilon d_2)\neq 0$, we have
\begin{align*}
\beta(\boldsymbol{z}(\epsilon))& = - \displaystyle \frac{r(x^*+\epsilon d_1)^T( r(x^*+\epsilon d_1)-r(x^*+\epsilon d_2))}{\|r(x^*+\epsilon d_1)-r(x^*+\epsilon d_2)\|^2},
\end{align*}
and, using $r(x^*+\epsilon d)=r(x^*)+ r'(x^*)\epsilon d +Q(\epsilon d) \epsilon d$ with $\lim_{\epsilon \rightarrow 0} Q(\epsilon d)=0$, we obtain
\begin{align*}
\beta(\boldsymbol{z}(\epsilon))& = - \displaystyle \frac{(r'(x^*)\epsilon d_1 +Q(\epsilon d_1) \epsilon d_1)^T( r'(x^*)\epsilon (d_1-d_2) +Q(\epsilon d_1) \epsilon d_1 - Q(\epsilon d_2) \epsilon d_2)}{\| r'(x^*)\epsilon (d_1-d_2) +Q(\epsilon d_1) \epsilon d_1 - Q(\epsilon d_2) \epsilon d_2 \|^2}\\
&=- \displaystyle \frac{(r'(x^*)d_1 +Q(\epsilon d_1) d_1)^T( r'(x^*) (d_1-d_2) +Q(\epsilon d_1) d_1 - Q(\epsilon d_2) d_2)}{\| r'(x^*) (d_1-d_2) +Q(\epsilon d_1) d_1 - Q(\epsilon d_2) d_2 \|^2},
\end{align*}
so
\begin{align*}
\lim_{\epsilon \rightarrow 0} \beta(\boldsymbol{z}(\epsilon))
&=- \displaystyle \frac{(r'(x^*)d_1)^T r'(x^*) (d_1-d_2)}{\| r'(x^*) (d_1-d_2)\|^2}.
\end{align*}
Note that, for example, $\lim_{\epsilon \rightarrow 0} \beta(\boldsymbol{z}(\epsilon))=0$ when $d_1=0$ and when $d_1=d_2$, and
$\lim_{\epsilon \rightarrow 0} \beta(\boldsymbol{z}(\epsilon))=-1$ when $d_2=0$.
Since the limit depends on the choice of $d_1$ and $d_2$, $\beta(\boldsymbol{z})$ is not continuous at $\boldsymbol{z} =\begin{bmatrix} x^*\\ x^*\end{bmatrix}$.
\end{proof}
\begin{remark}
It is interesting to note the difference in the limiting behavior of $\beta(\boldsymbol{z})$ along radial paths in the proofs of \cref{prop:betaxx} and \cref{prop:betax*}. The proof of \cref{prop:betax*} shows that, when approaching $\boldsymbol{z} =\begin{bmatrix} x^*\\ x^*\end{bmatrix}$ along radial paths, $\lim_{\epsilon \rightarrow 0} \beta(\boldsymbol{z}(\epsilon))$ is finite for any fixed $d_1$ and $d_2$. This is in contrast to the limiting behavior of $\beta(\boldsymbol{z})$ when approaching $\boldsymbol{z} =\begin{bmatrix} x\\ x\end{bmatrix}$ along radial paths with $x \neq x^*$ in \cref{prop:betaxx}, where $\beta(\boldsymbol{z}(\epsilon))$ grows without bound. The AA($m$) convergence proof in \cite{toth2015convergence} relies on the unproven assumption that $|\beta(\boldsymbol{z}_k)|$ is bounded above as $x_k \rightarrow x^*$.
\end{remark}
\subsection{Continuity and differentiability of $\Psi(\boldsymbol{z})$ for AA(1)}
In this section, we discuss the continuity and differentiability of $\Psi(\boldsymbol{z})$ at $\boldsymbol{z} =\begin{bmatrix} x\\ y\end{bmatrix}$. We consider the three cases of \cref{tab:continuity-differentiability-Lip}: $r(x)\neq r(y)$, $x = y$ with $r(x)\neq 0$, and $x=y=x^*$.
\begin{proposition}\label{prop:Psi-x-not-y}
$\Psi(\boldsymbol{z})$ in \cref{eq:Psi1} is continuous and differentiable at $\boldsymbol{z} =\begin{bmatrix} x\\ y\end{bmatrix}$ when $r(x)\neq r(y)$.
\end{proposition}
\begin{proof}
Recall that
\begin{equation*}
\Psi(\boldsymbol{z}) =
\begin{bmatrix}
q(x)+\beta(\boldsymbol{z})(q(x)-q(y))\\
x
\end{bmatrix},
\end{equation*}
where $\beta(\boldsymbol{z}) =\displaystyle \frac{-r^T(x)(r(x)-r(y))}{\|r(x)-r(y)\|^2}$ when $r(x)\neq r(y)$.
Since $q(x), q(y), r(x)$, and $r(y)$ are continuous and differentiable functions at $x$ and $y$,
$\Psi(\boldsymbol{z})$ is continuous and differentiable at $\boldsymbol{z} =\begin{bmatrix} x\\ y\end{bmatrix}$ when $r(x)\neq r(y)$.
\end{proof}
\begin{proposition}\label{prop:Psixx}
$\Psi(\boldsymbol{z})$ in \cref{eq:Psi1} is not continuous and not differentiable at $\boldsymbol{z} =\begin{bmatrix} x\\ y\end{bmatrix}$ when $x= y$ with $r(x) \neq 0$.
\end{proposition}
\begin{proof}
Let $\boldsymbol{z}(\epsilon)=\begin{bmatrix} x\\ y+\epsilon e\end{bmatrix},$ where $x= y$ with $r(x) \neq 0$, and $e$ is a unit vector in $\mathbb{R}^n$.
From the proof of \cref{prop:betaxx}, we have
\begin{align*}
\beta(\boldsymbol{z}(\epsilon))& =
\displaystyle \frac{r(x)^T (r'(y) e + Q(\epsilon e) e) }{\epsilon \ \|r'(y) e +Q(\epsilon e) e\|^2}.
\end{align*}
Plugging this into \cref{eq:Psi1} and using $q(y+\epsilon e)=q(y) + q'(y) \epsilon e + P(\epsilon e) \epsilon e$ with $\lim_{\epsilon \rightarrow 0} P(\epsilon e)=0$, we obtain
\begin{align*}
\Psi(\boldsymbol{z}(\epsilon)) &=
\begin{bmatrix}
q(x)+\displaystyle \frac{r(x)^T (r'(y) e + Q(\epsilon e) e) }{\epsilon \ \|r'(y) e +Q(\epsilon e) e\|^2}
(q(x)-q(y)- q'(y) \epsilon e - P(\epsilon e) \epsilon e)\\
x
\end{bmatrix},\\
&=
\begin{bmatrix}
q(x)+\displaystyle \frac{r(x)^T (r'(x) e + Q(\epsilon e) e) }{ \ \|r'(x) e +Q(\epsilon e) e\|^2}
(q'(x) e - P(\epsilon e) e)\\
x
\end{bmatrix},
\end{align*}
and
\begin{align*}
\lim_{\epsilon \rightarrow 0} \Psi(\boldsymbol{z}(\epsilon)) &=
\begin{bmatrix}
q(x)+\displaystyle \frac{r(x)^T r'(x) e}{ \ \|r'(x) e\|^2}
q'(x) e\\
x
\end{bmatrix}.
\end{align*}
While the limit is finite for any $e$, it depends on the choice of $e$, so $\Psi(\boldsymbol{z})$ is not continuous at $\boldsymbol{z}=\begin{bmatrix}x \\ x\end{bmatrix}$ where $r(x)\neq 0$. It follows that $\Psi(\boldsymbol{z})$ is not differentiable at $\boldsymbol{z}=\begin{bmatrix}x \\ x\end{bmatrix}$.
\end{proof}
The following proposition establishes the Lipschitz continuity of $\Psi(\boldsymbol{z})$ at $\boldsymbol{z}^* =\begin{bmatrix} x^*\\ x^*\end{bmatrix}$. The differentiability of $\Psi(\boldsymbol{z})$ at $\boldsymbol{z}^*$ is investigated in the next subsection.
\begin{proposition}\label{prop:Psix*}
$\Psi(\boldsymbol{z})$ in \cref{eq:Psi1} is Lipschitz continuous at $\boldsymbol{z} =\begin{bmatrix} x\\ y\end{bmatrix}$ when $x= y= x^*$ with global Lipschitz constant $L=( \|A^{-1}\| \|A\| + 1 ) \|I-A\|+1$ in the linear case, and with local Lipschitz constant $L=3+(4+4/c_r)\|r'(x^*)\|$ in the nonlinear case, where $c_r$ is a problem-dependent constant.
\end{proposition}
\begin{proof}
See \cref{app:proof-psi-lipsch}.
\end{proof}
\subsection{Differentiability of $\Psi(\boldsymbol{z})$ at $\boldsymbol{z}^*$ for AA(1)}
In this subsection we investigate the differentiability of $\Psi(\boldsymbol{z})$ at $\boldsymbol{z}^*$.
We first consider the directional derivatives of $\Psi(\boldsymbol{z})$ at $\boldsymbol{z}^*$.
\begin{definition}[Directional Derivative]
Let $F: U\subset \mathbb{R}^n \longrightarrow \mathbb{R}^m$ be a function on the open
set $U$. We call $\mathfrak{D}F(x, d)$ defined by
\begin{equation}\label{def-direction-diff}
\mathfrak{ D}F(x,d) = \lim_{h\downarrow 0}\displaystyle \frac{F(x+hd)-F(x)}{h}
\end{equation}
the directional derivative of $F$ at $x$ in direction $d$ if the limit exists. We say $F(x)$ is Gateaux differentiable in $x$ if the directional derivative of $F$ exists in $x$ for all directions.
\end{definition}
Note that, if $F$ is differentiable at $x$ with Jacobian $F'(x)$, then $\mathfrak{ D}F(x,d)=F'(x)d$.
\begin{theorem}\label{thm:x=y-directional-dif}
Consider $\Psi(\boldsymbol{z})$ at $\boldsymbol{z}^*=\begin{bmatrix} x^*\\ x^*\end{bmatrix}$ and direction $\boldsymbol{d}=\begin{bmatrix} d_1\\ d_2\end{bmatrix}$.
Let $M=q'(x^*)$ and $A= I-q'(x^*)$.
Then the directional derivative of $\Psi(\boldsymbol{z})$ at $\boldsymbol{z}^*$ in direction $\boldsymbol{d}$ is given by
\begin{equation}\label{AA(1)-direction-diff}
\mathfrak{ D} \Psi(\boldsymbol{z}^*,\boldsymbol{ d})
= \begin{bmatrix}
(1+\widehat{\beta}(\boldsymbol{d})) M & -\widehat{\beta}(\boldsymbol{d}) M\\
I & 0
\end{bmatrix}\boldsymbol{d},
\end{equation}
where
\begin{equation*}
\widehat{\beta}(\boldsymbol{d}) =
\begin{cases}
- \displaystyle \frac{d_1^TA^TA(d_1-d_2)}{(d_1-d_2)^TA^TA(d_1-d_2)}, & \text{if}\,\, d_1\neq d_2, \\
0, & \text{if}\,\, d_1=d_2.
\end{cases}
\end{equation*}
\end{theorem}
\begin{proof}
See \cref{app:proof-directional}.
\end{proof}
\begin{remark}\label{rmk-Psi-diff-n=1}
When $n=1$ and $d_1\neq d_2$, the result in \cref{eq:DPsi} simplifies considerably: since all quantities are scalar, $\widehat{\beta}(\boldsymbol{d})=\displaystyle -\frac{d_1}{d_1-d_2}$ and \cref{eq:DPsi} can be rewritten as:
\begin{align*}
\mathfrak{ D} \Psi(\boldsymbol{z}^*,\boldsymbol{ d})&=
\begin{bmatrix}
(1+\widehat{\beta}(\boldsymbol{d})) (1-a) & -\widehat{\beta}(\boldsymbol{d}) (1-a) \\
1 & 0
\end{bmatrix}\boldsymbol{d}\\
&=
\begin{bmatrix}
\frac{-d_2}{d_1-d_2} (1-a) & \frac{d_1}{d_1-d_2} (1-a) \\
1 & 0
\end{bmatrix}\boldsymbol{d},\\
& =\begin{bmatrix}
0\\
d_1
\end{bmatrix},\\
& =\begin{bmatrix}
0&0 \\
1 & 0
\end{bmatrix}\boldsymbol{d}.
\end{align*}
However, when $n=1$ and $d_1= d_2$, we get
\begin{align*}
\mathfrak{ D} \Psi(\boldsymbol{z}^*,\boldsymbol{ d})&=
\begin{bmatrix}
(1-a) &0 \\
1 & 0
\end{bmatrix}\boldsymbol{d}.
\end{align*}
This shows that, when $n=1$, $\Psi(\boldsymbol{z})$ is not differentiable at $\boldsymbol{z}^*$.
When $n>1$, $\Psi(\boldsymbol{z})$ is also not differentiable at $\boldsymbol{z}^*$, because
the matrix in \cref{eq:DPsi} depends on $\boldsymbol{d}$ and
$\mathfrak{D} \Psi(\boldsymbol{z}^*,\boldsymbol{d})$ cannot be written as $\Psi'(\boldsymbol{z}^*)\boldsymbol{d}$.
\end{remark}
Finally, it is interesting to consider the differentiability results of \cref{tab:continuity-differentiability-Lip} for AA(1) specifically for the scalar case, $n=1$. This is considered in \cref{app:scalar}.
\section{Analysis of $\Psi(\boldsymbol{z})$ for AA(m)}
\label{sec:AAm}
In this section, we extend the properties of AA(1) in \cref{tab:continuity-differentiability-Lip} to AA($m$).
First, let us extend $\Psi(\boldsymbol{z})$ in \cref{eq:AA(1)-system} for iteration \cref{eq:AA-fixed-point} to AA($m$) as follows:
\begin{equation}\label{eq:Psi-AAm}
\Psi(\boldsymbol{z}) =
\begin{bmatrix}
q(z_{m+1}) + \sum_{j=1}^{m}\beta_{j} (q(z_{m+1})-q(z_{m+1-j}))\\
z_{m+1}\\
z_m \\
\vdots \\
z_2
\end{bmatrix}=
\begin{bmatrix}
q(z_{m+1}) + Q(\boldsymbol{z})\boldsymbol{\beta}(\boldsymbol{z})\\
z_{m+1}\\
z_m\\
\vdots \\
z_2
\end{bmatrix},
\end{equation}
where
\begin{equation}\label{eq:def-Q-general}
Q(\boldsymbol{z}) = \begin{bmatrix} q(z_{m+1})-q(z_m) & q(z_{m+1})-q(z_{m-1}) & \ldots & q(z_{m+1})-q(z_{1}) \end{bmatrix}
\end{equation}
with
$
\boldsymbol{z} =\begin{bmatrix} z_{m+1}^T & z_m^T & \ldots & z_{1}^T\end{bmatrix}^T \in\mathbb{R}^{n(m+1)},
$
and $\boldsymbol{\beta}^{(k)}$ in \cref{eq:AAm-beta-form-pseudo} is written as
\begin{equation}\label{eq:beta-AAm}
\boldsymbol{\beta}(\boldsymbol{z}) =\begin{bmatrix} \beta_1 \\ \vdots \\ \beta_m \end{bmatrix}=
-R(\boldsymbol{z})^{\dag}r(z_{m+1}),
\end{equation}
where
\begin{equation}\label{eq:beta-continuous}
R(\boldsymbol{z}) = \begin{bmatrix} r(z_{m+1})-r(z_m) & r(z_{m+1})-r(z_{m-1}) & \ldots & r(z_{m+1})-r(z_{1}) \end{bmatrix},
\end{equation}
and $R(\boldsymbol{z})^{\dag}$ is the pseudo-inverse of $R(\boldsymbol{z})$.
Note that when $\boldsymbol{z}=\boldsymbol{z}^*=\displaystyle \begin{bmatrix} (x^*)^T & (x^*)^T & \ldots & (x^*)^T \end{bmatrix}^T$, $\Psi(\boldsymbol{z})=\boldsymbol{z}$. We can write AA($m$) as
the fixed-point iteration $\boldsymbol{z}_{k+1} =\Psi(\boldsymbol{z}_k)$ with $\Psi(\boldsymbol{z})$ as in \cref{eq:Psi-AAm}, where $\boldsymbol{z}_k =\begin{bmatrix} x_{k+m}^T & x_{k+m-1}^T & x_{k+m-2}^T & \ldots & x_{k}^T\end{bmatrix}^T$.
For convenience, we define the following operator:
\begin{equation}\label{eq:def-D-operator}
D(\boldsymbol{z})=\begin{bmatrix} z_{m+1}-z_m & z_{m+1}-z_{m-1} & \cdots & z_{m+1}-z_{1} \end{bmatrix}\quad \in\mathbb{R}^{n\times m}.
\end{equation}
For simplicity, we will sometimes denote $D(\boldsymbol{z}), R(\boldsymbol{z}), Q(\boldsymbol{z})$ and $\boldsymbol{\beta}(\boldsymbol{z})$ by $D,R,Q$ and $\boldsymbol{\beta}$.
Before providing our detailed analysis of the differentiability properties of $\Psi(\boldsymbol{z})$, we summarize our results in \cref{tab:continuity-differentiability-Lip-AAm}. While the proof for some of these results is elementary, the table provides a complete overview of the differentiability properties of $\Psi(\boldsymbol{z})$ which are useful to understand the convergence behavior of AA($m$) viewed as the fixed-point method \cref{eq:AA-fixed-point}.
\newsavebox{\smlmatt
\savebox{\smlmatt}{$\small \boldsymbol{z} =\begin{bmatrix} z_{m+1}^T & z_m^T & \ldots & z_{1}^T \end{bmatrix}^T$}
\begin{table}[H]
\caption{Continuity and differentiability properties of $\boldsymbol{\beta}(\boldsymbol{z})$ and $\Psi(\boldsymbol{z})$ at~\usebox{\smlmatt} for AA($m$) iteration
\cref{eq:AA-fixed-point} with $m\geq 1$, where $\boldsymbol{\beta}(\boldsymbol{z})$ and $\Psi(\boldsymbol{z})$ are given by \cref{eq:beta-AAm} and \cref{eq:Psi-AAm}. }
\centering
\begin{tabular}{|l|c|c|c|}
\hline
&$R(\boldsymbol{z})$ has full rank & $z_j=z \, \forall j, \ r(z) \neq 0$ & $\boldsymbol{z}=\boldsymbol{z}^*$ \\
\hline continuity of $\boldsymbol{\beta}(\boldsymbol{z})$ &$\surd$ & $\times$ & $\times$ \\
\hline continuity of $\Psi(\boldsymbol{z})$ &$\surd$ & $\times$ &$\surd$\tablefootnote{We only prove this in the linear case. \label{note1}} \\
\hline Lipschitz continuity of $\Psi(\boldsymbol{z})$ &$\surd$ & $\times$ & $\surd$\cref{note1} \\
\hline Gateaux-differentiability of $\Psi(\boldsymbol{z})$ &$\surd$ & $\times$ &$\surd$\tablefootnote{We prove this for almost all directions $\boldsymbol{d}$.} \\
\hline differentiability of $\Psi(\boldsymbol{z})$ &$\surd$ &$\times$ &$\times$ \\
\hline
\end{tabular}\label{tab:continuity-differentiability-Lip-AAm}
\end{table}
\subsection{Continuity of $\boldsymbol{\beta}(\boldsymbol{z})$ for AA($m$)}
\begin{proposition}\label{pro:beta-full-rank-continuous}
$\boldsymbol{\beta}(\boldsymbol{z})$ in \cref{eq:beta-AAm} is continuous at $\boldsymbol{z}$ when $R(\boldsymbol{z})$ has full rank.
\end{proposition}
\begin{proof}
Since $r(z)$ is a continuous function, $r(z_{m+1})-r(z_j)$ for $j=1,2,\ldots,m$ and $r(z_{m+1})$ are continuous functions, which means that $R(\boldsymbol{z})$ is continuous. If $R(\boldsymbol{z})$ has full rank, then $R^T(\boldsymbol{z})R(\boldsymbol{z})$ is invertible and the inverse is continuous and $\boldsymbol{\beta}(\boldsymbol{z})= -(R(\boldsymbol{z})^TR(\boldsymbol{z}))^{-1} R(\boldsymbol{z})^Tr(z_{m+1})$. It follows that $\boldsymbol{\beta}(\boldsymbol{z})$ in \cref{eq:beta-AAm} is continuous if $R(\boldsymbol{z})$ has full rank.
\end{proof}
\begin{proposition}\label{pro:beta-zj=z}
$\boldsymbol{\beta}(\boldsymbol{z})$ in \cref{eq:beta-AAm} is not continuous at $\boldsymbol{z}=\begin{bmatrix} z^T & z^T & \ldots & z^T\end{bmatrix}^T$ with $r(z)\neq 0$.
\end{proposition}
\begin{proof}
Let $\boldsymbol{z}_0=\begin{bmatrix} z^T & z^T & \ldots & z^T\end{bmatrix}^T$ with $r(z)\neq 0$ and $\boldsymbol{d} =\begin{bmatrix} d_{m+1}^T & d_m^T & \ldots & d_1^T\end{bmatrix}^T$, where all $d_j$ are zero except $d_m=d=\epsilon\, e$ with $e$ a unit vector
in $\mathbb{R}^n$ and $r(z)-r(z+d)\neq 0$. Then,
\begin{equation*}
R(\boldsymbol{z}_0+\boldsymbol{d}) = \begin{bmatrix} r(z)-r(z+d) & 0 & \ldots & 0\end{bmatrix}.
\end{equation*}
Using \cref{eq:beta-continuous} and \cref{eq:two-ways-pseudo-inverse}, we have
\begin{align*}
\boldsymbol{\beta}(\boldsymbol{z}_0+\boldsymbol{d}) &= -R(\boldsymbol{z}_0+\boldsymbol{d})^{\dag}r(z_{m+1}) ,\\
& = -\big(R(\boldsymbol{z}_0+\boldsymbol{d})^T R(\boldsymbol{z}_0+\boldsymbol{d})\big)^{\dag} R(\boldsymbol{z}_0+\boldsymbol{d})^T r(z),\\
& = \begin{bmatrix} \displaystyle \frac{-(r(z)-r(z+d))^T r(z)}{(r(z)-r(z+d))^T(r(z)-r(z+d))} & 0 & \ldots & 0 \end{bmatrix}^T.
\end{align*}
According to the proof of \cref{prop:betaxx}, $\displaystyle \frac{-(r(z)-r(z+d))^T r(z)}{(r(z)-r(z+d))^T(r(z)-r(z+d))}\rightarrow \pm \infty$ as $\epsilon \rightarrow 0$.
Thus, $\boldsymbol{\beta}(\boldsymbol{z})$ is not continuous at $\boldsymbol{z}_0$ with $r(z)\neq 0$.
\end{proof}
\begin{proposition}\label{pro:beta-z=z^*}
$\boldsymbol{\beta}(\boldsymbol{z})$ in \cref{eq:beta-AAm} is not continuous at $\boldsymbol{z}=\boldsymbol{z}^*$.
\end{proposition}
\begin{proof}
Consider $\boldsymbol{d} =\begin{bmatrix} (\epsilon d_1)^T & (\epsilon d_2)^T & (\epsilon d_1)^T &\ldots & ( \epsilon d_1)^T\end{bmatrix}^T$ with $\epsilon\neq 0$ and $ r(x^*+\epsilon d_1)-r(x^*+\epsilon d_2)\neq 0$. Then,
\begin{equation*}
R(\boldsymbol{z}^*+\boldsymbol{d}) = \begin{bmatrix} r(x^*+\epsilon d_1)-r(x^*+\epsilon d_2) & 0 & \ldots & 0\end{bmatrix}.
\end{equation*}
For simplicity, let $w=r(x^*+\epsilon d_1)-r(x^*+\epsilon d_2)$. Using \cref{eq:beta-continuous} and \cref{eq:two-ways-pseudo-inverse}, we have
\begin{align*}
\boldsymbol{\beta}(\boldsymbol{z}^*+\boldsymbol{d}) &= -R(\boldsymbol{z}^*+\boldsymbol{d})^{\dag}r(z_{m+1}) ,\\
& = -\big(R(\boldsymbol{z}^*+\boldsymbol{d})^T R(\boldsymbol{z}^*+\boldsymbol{d})\big)^{\dag} R(\boldsymbol{z}^*+\boldsymbol{d})^T r(x^*+\epsilon d_1),\\
& = \begin{bmatrix} \displaystyle \frac{-w^T r(x^*+\epsilon d_1)}{w^Tw} & 0 & \ldots & 0 \end{bmatrix}^T.
\end{align*}
Let $\displaystyle \Delta(\epsilon)=\frac{-w^T r(x^*+\epsilon d_1)}{w^Tw}$. According to the proof of \cref{prop:betax*}, $\lim_{\epsilon \rightarrow 0} \Delta(\epsilon)=0$ when $d_1=0$ and when $d_1=d_2$, and
$\lim_{\epsilon \rightarrow 0} \Delta(\epsilon)=-1$ when $d_2=0$.
Since the limit depends on the choice of $d_1$ and $d_2$, $\boldsymbol{\beta}(\boldsymbol{z})$ is not continuous at $\boldsymbol{z} =\boldsymbol{z}^*$.
\end{proof}
\subsection{Continuity and differentiability of $\Psi(\boldsymbol{z})$ for AA($m$)}
In this subsection, we discuss the continuity and differentiability of $\Psi(\boldsymbol{z})$ for AA($m$). We consider three cases at point $\boldsymbol{z}=\begin{bmatrix} z_{m+1}^T & z_m^T & \ldots & z_{1}^T\end{bmatrix}^T$:
\begin{itemize}
\item[(a)] $R(\boldsymbol{z})$ has full rank;
\item[(b)] $z_{j}=z, j=1,2,\ldots,m+1$, with $r(z) \neq 0$;
\item[(c)] $z_{m+1}=x^*$ and $R(\boldsymbol{z})$ is rank-deficient.
\end{itemize}
\begin{proposition}\label{prop:Psi-AAm-full-rank-diff}
$\Psi(\boldsymbol{z})$ is continuous at $\boldsymbol{z}$ when $R(\boldsymbol{z})$ has full rank. Furthermore, $\Psi(\boldsymbol{z})$ is differentiable at $\boldsymbol{z}$ when $R(\boldsymbol{z})$ has full rank.
\end{proposition}
\begin{proof}
Recall $\Psi(\boldsymbol{z})$ in \cref{eq:Psi-AAm}. When $R(\boldsymbol{z})$ in \cref{eq:beta-continuous} has full rank,
$\boldsymbol{\beta}(\boldsymbol{z})$ in \cref{eq:beta-AAm} is continuous by \cref{pro:beta-full-rank-continuous}, and $Q(\boldsymbol{z})$ in \cref{eq:def-Q-general} is continuous. It follows that $\Psi(\boldsymbol{z})$ is continuous. Furthermore, since $q(z_1), r(z_1), Q(\boldsymbol{z}), R(\boldsymbol{z})$, and $(R((\boldsymbol{z}))^TR(\boldsymbol{z}))^{-1}$ are differentiable, $\Psi(\boldsymbol{z})$ is differentiable.
\end{proof}
\begin{proposition}\label{AAm-prop-noncon-rank-case1}
$\Psi(\boldsymbol{z})$ is not continuous at $\boldsymbol{z}$ when $\boldsymbol{z}=\begin{bmatrix} z^T & z^T & \ldots & z^T \end{bmatrix}^T$ with $r(z)\neq 0$.
\end{proposition}
\begin{proof}
Let $\boldsymbol{z}_0 =\begin{bmatrix} z^T & z^T & \ldots & z^T \end{bmatrix}^T$ with $r(z) \neq 0$ and $\boldsymbol{d} =\begin{bmatrix} 0^T & d^T & 0& \ldots & 0^T \end{bmatrix}^T$, where $d=\epsilon\, e$ with $e$ a unit vector
in $\mathbb{R}^n$ and $r(z)-r(z+d)\neq 0$. Then,
\begin{equation*}
Q(\boldsymbol{z}_0+\boldsymbol{d}) = \begin{bmatrix} q(z)-q(z+d) & 0 & \ldots & 0\end{bmatrix}.
\end{equation*}
Let $w=r(z)-r(z+d)$. From the proof of \cref{pro:beta-zj=z}, we have
\begin{equation*}
Q(\boldsymbol{z}_0+\boldsymbol{d})\boldsymbol{\beta}(\boldsymbol{z}_0+\boldsymbol{d}) = \frac{-w^T r(z)}{w^Tw}\big(q(z)-q(z+d)\big).
\end{equation*}
From \cref{eq:Psi-AAm} we have
\begin{equation*}
\Psi(\boldsymbol{z}_0+\boldsymbol{d}) =
\begin{bmatrix}
q(z) -\displaystyle \frac{w^T r(z)}{w^Tw}\big(q(z)-q(z+d)\big)\\
z+d_{m+1}\\
\vdots \\
z+d_2
\end{bmatrix}.
\end{equation*}
From the proof of \cref{prop:Psixx}, we know that the limit of $q(z) -\displaystyle \frac{w^T r(z)}{w^Tw}\big(q(z)-q(z+d)\big)$ as $\epsilon\rightarrow 0$ depends on the choice of $e$. It follows that $\Psi(\boldsymbol{z}_0+\boldsymbol{d})$ is not continuous at $\boldsymbol{z}_0$.
\end{proof}
We analyze the continuity of case (c) for the linear case only, because it is not clear how to generalize \cref{prop:Psix*} for $m>1$ in the nonlinear case. Differentiability for case (c) is discussed in the next subsection.
\begin{proposition}\label{AAm-prop-noncon-rank-zm+1=x*}
In the linear case, $\Psi(\boldsymbol{z})$ is Lipschitz continuous at $\boldsymbol{z}$ when $R(\boldsymbol{z})$ is rank-deficient and $z_{m+1}=x^*$.
\end{proposition}
\begin{proof}
See \cref{app:proof-lipsch-m}.
\end{proof}
Note that \cref{AAm-prop-noncon-rank-zm+1=x*} contains the special case that $\boldsymbol{z}=\boldsymbol{z}^*$.
\subsection{Differentiability of $\Psi(\boldsymbol{z})$ at $\boldsymbol{z}^*$ for AA($m$)}
In this subsection we investigate the differentiability of $\Psi(\boldsymbol{z})$ at $\boldsymbol{z}^*$. We first consider the directional derivatives of $\Psi(\boldsymbol{z})$ at $\boldsymbol{z}^*$.
\begin{theorem}\label{thm:proof-directional-m}
Let $M=q'(x^*)$ and $A= r'(x^*)=I-q'(x^*)$. Then, the directional derivative of $\Psi(\boldsymbol{z})$ at $\boldsymbol{z}^*$ in any direction $\boldsymbol{d}$ such that $D(\boldsymbol{d})$ is full rank is given by
\begin{equation}\label{eq:-direct-nonlinear-AAm-matrix}
\mathfrak{ D} \Psi(\boldsymbol{z}^*,\boldsymbol{ d}) = \begin{bmatrix} (1+\sum_{j=1}^{m} \widehat{\beta}_j)M &- \widehat{\beta}_{1}M &\cdots & -\widehat{\beta}_{m-1}M &-\widehat{\beta}_{m}M\\
I& 0& & 0& 0\\
0& I& & 0& 0\\
\vdots & \vdots & &\vdots &\vdots\\
0& 0& \cdots& I& 0
\end{bmatrix}\boldsymbol{d},
\end{equation}
where $\widehat{\boldsymbol{\beta}}(\boldsymbol{d})=\begin{bmatrix} \widehat{\beta}_1 & \widehat{\beta}_2 & \cdots & \widehat{\beta}_m \end{bmatrix}^T= -( AD(\boldsymbol{d}))^{\dag}A d_{m+1}$
with $D(\boldsymbol{d})$ defined in \cref{eq:def-D-operator}.
\end{theorem}
\begin{proof}
See \cref{app:proof-directional-m}.
\end{proof}
\begin{remark}
The result in \cref{eq:-direct-nonlinear-AAm-matrix} holds for the linear case without the requirement that $D(\boldsymbol{d})$ is full rank because the term $P(hd_{m+1})$ in \cref{eq:expansion-r} vanishes.
For the nonlinear case, when $D(\boldsymbol{d})$ is rank-deficient, we do not know whether \cref{eq:limit-speudo-inverse} holds.
\end{remark}
\begin{remark}
For $m>1$, $\Psi(\boldsymbol{z})$ is not differentiable at $\boldsymbol{z}^*$, because
the matrix in \cref{eq:-direct-nonlinear-AAm-matrix} depends on $\boldsymbol{d}$ and
$\mathfrak{D} \Psi(\boldsymbol{z}^*,\boldsymbol{d})$ cannot be written as $\Psi'(\boldsymbol{z}^*)\boldsymbol{d}$.
\end{remark}
\section{Numerical results}
\label{sec:numerics}
In this section, we give further numerical results expanding on the AA(1) convergence patterns we identified in \cref{fig:simple-MC} for the simple $2 \times 2$ linear equation of \cref{prob:linear2x2}. We extend the numerical tests to larger linear problems and a nonlinear problem, for $m=1$ and $m>1$. We are also interested in comparing the convergence behavior of AA($m$) with GMRES for the linear problems: we compare the standard windowed version of AA($m$) with GMRES and a restarted version of AA($m$), which is essentially restarted GMRES($m$). We relate the numerical results to the theoretical findings of \Cref{sec:AA1,sec:AAm}.
\begin{figure}[h]
\centering
\includegraphics[width=.6\textwidth]{dir_deriv_Psi_norm.pdf}
\caption{
\cref{prob:linear2x2} (linear): Histogram of the norm of the directional derivative $\mathfrak{ D} \Psi(\boldsymbol{z},\boldsymbol{ d})$ of $\Psi(\boldsymbol{z})$ at $\boldsymbol{z}^*$ in unit vector direction $\boldsymbol{ d}$ (see Eq.\ \cref{AA(1)-direction-diff}), for $10^6$ unit vectors on a uniform polar grid in 4D space.} \label{prob2-dir_deriv}
\end{figure}
We first revisit the numerical results of \cref{fig:simple-MC} for the $2 \times 2$ linear equation of \cref{prob:linear2x2}.
As discussed before, \cref{fig:simple-MC} indicates that AA(1) sequences $\{x_k\}$ for \cref{prob:linear2x2}
converge r-linearly with a continuous spectrum of convergence factors $\rho_{\{x_k\}}$, and it appears that
a least upper bound $\rho_{\Psi,x^*}$ for $\rho_{\{x_k\}}$ exists for the AA(1) iteration \cref{eq:AA-fixed-point}
that is smaller than, say, 0.45, and substantially smaller than the r-linear convergence factor $\rho_{q,x^*}=2/3$ of
fixed-point iteration \cref{eq:fixed-point} by itself. Since there currently is no theory to establish the existence or
value of $\rho_{\Psi,x^*}$, it is interesting to investigate, in light of \cref{thm:x=y-directional-dif}, whether the existence of
all directional derivatives of $\Psi(\boldsymbol{z})$ at $\boldsymbol{z}^*$ may tell us something about the existence or value
of $\rho_{\Psi,x^*}$. As is well-known, in the case of an iteration function $\Psi(z)$ that is $L$-Lipschitz in a neighborhood of
$z^*$ with $L<1$, the FP iteration $z_{k+1}=\Psi(z_k)$ converges q-linearly with q-linear convergence factor not worse than $L$ \cite{kelley1995iterative}. In the case of AA(1), $\Psi(\boldsymbol{z})$ is not $L$-Lipschitz in a neighborhood of $\boldsymbol{z}^*$
(since, by \cref{prop:Psixx}, $\Psi(\boldsymbol{z})$ is not continuous at $\boldsymbol{z} =\begin{bmatrix} x\\ y\end{bmatrix}$ when
$x= y$ with $r(x) \neq 0$), but it is still interesting to investigate the size of the directional derivatives that we know by \cref{thm:x=y-directional-dif} exist in all directions at $\boldsymbol{z}^*$.
\cref{prob2-dir_deriv} shows a histogram for \cref{prob:linear2x2} of the norm of the directional derivative $\mathfrak{ D} \Psi(\boldsymbol{z},\boldsymbol{ d})$ of $\Psi(\boldsymbol{z})$ at $\boldsymbol{z}^*$ in unit vector direction $\boldsymbol{ d}$ (see Eq.\ \cref{AA(1)-direction-diff}), for $10^6$ unit vectors on a uniform polar grid in 4D space. The histogram indicates that the unit directional derivatives are bounded above by a value of about 1.6, but it is interesting that this value is greater than 1, which indicates that directional derivates or Lipschitz constants are not a useful avenue to prove the existence of a least upper bound $\rho_{\Psi,x^*}<\rho_{q,x^*}$ for AA(1).
\begin{figure}[h]
\centering
\includegraphics[width=.49\textwidth]{fig2_prob02_mc_rconv.pdf}
\includegraphics[width=.49\textwidth]{fig2_prob02_mc_betas.pdf}
\includegraphics[width=.325\textwidth]{fig2_prob02_mc_inits.pdf}
\includegraphics[width=.325\textwidth]{fig2_prob02_mc_AAhist.pdf}
\includegraphics[width=.325\textwidth]{fig2_prob02_mc_qhist.pdf}
\caption{FP and AA(1) results for nonlinear \cref{prob:nonlinear2x2} with 1,000 random initial guesses.} \label{prob3-mc}
\end{figure}
We next consider a nonlinear example:
\begin{problem}\label{prob:nonlinear2x2}
Consider the nonlinear system
\begin{align}
x_2= x_1^2 \label{eq:exm1-1}\\
x_1+(x_1-1)^2 +x_2^2 =1 \label{eq:exm1-2}
\end{align}
with solution $(x_1^*,x_2^*) = (0,0)$.
Let $x=[x_1 \ x_2]^T$ and define the FP iteration function
\begin{equation*}
q(x) = \begin{bmatrix} \displaystyle \frac{1}{2}(x_1+x_1^2+x_2^2) \\ \\ \displaystyle \frac{1}{2}(x_2+x_1^2) \end{bmatrix},
\end{equation*}
with Jacobian matrix
\begin{equation*}
q'(x)=
\begin{bmatrix}
x_1+\displaystyle \frac{1}{2}& x_2\\
x_1 & \displaystyle \frac{1}{2}
\end{bmatrix}.
\end{equation*}
We have
\begin{equation*}
q'(x^*) = \begin{bmatrix}
\displaystyle \frac{1}{2}& 0 \\
0 & \displaystyle \frac{1}{2}
\end{bmatrix}, \ \textrm{and} \quad
\rho(q'(x^*)) =\displaystyle \frac{1}{2}<1.
\end{equation*}
\end{problem}
\begin{figure}[h]
\centering
\includegraphics[width=.49\textwidth]{fig9_prob18__gmresInf_rconv.pdf}
\includegraphics[width=.49\textwidth]{fig9_prob18__gmresInf_inits.pdf}
\includegraphics[width=.325\textwidth]{fig9_prob18__gmresInf_qhist.pdf}
\includegraphics[width=.325\textwidth]{fig9_prob18__gmresInf_AAhist.pdf}
\includegraphics[width=.325\textwidth]{fig9_prob18__gmresInf_AArhist.pdf}
\caption{
\cref{prob:linear200}: linear problem with $M \in \mathbb{R}^{200 \times 200}$ and $\lambda_1=0.9$, $\lambda_2=-0.3$, $\lambda_3=0.3$, and $\lambda_4=-0.3$, for 200 random initial guesses. Comparison of AA(1) (red) with AA($\infty$) (blue), which is essentially equivalent to GMRES. It can be observed that the asymptotic linear convergence factors of the AA(1) sequences strongly depend on the initial guess, but the asymptotic linear convergence factors of the AA($\infty$) sequences and of the FP sequences do not depend on the initial guess.
} \label{prob18-gmresinf}
\end{figure}
\cref{prob3-mc} shows FP and AA(1) numerical results for the nonlinear \cref{prob:nonlinear2x2}.
The nonlinear results of \cref{prob3-mc} show convergence behavior that is qualitatively similar to the linear results of \cref{fig:simple-MC}:
the AA(1) sequences $\{x_k\}$ converge r-linearly, but the r-linear convergence factors $\rho_{\{x_k\}}$ depend on the initial guess on a set of nonzero measure. It appears that a least upper bound $\rho_{\Psi,x^*}$ for $\rho_{\{x_k\}}$ exists for the AA(1) iteration \cref{eq:AA-fixed-point}
that is smaller than the r-linear convergence factor $\rho_{q,x^*}=1/2$ of fixed-point iteration \cref{eq:fixed-point} by itself.
We also see that the $\beta_k$ sequences oscillate for this nonlinear problem as the AA(1) iteration approaches $x^*$, consistent with the
discontinuity of $\beta(\boldsymbol{z})$ at $\boldsymbol{z}^*$ shown in \cref{prop:betax*}.
We next consider a larger linear problem that we will use for comparison of AA($m$) to GMRES and a restarted version of AA($m$).
\begin{problem}\label{prob:linear200}
Consider the linear iteration
\begin{equation}\label{eq:q-linear200}
x_{k+1} = M x_k,
\end{equation}
with $M \in \mathbb{R}^{200 \times 200}$, where $M$ is diagonal except that $m_{1,2}=1$.
$M$ has 196 eigenvalues that are spaced uniformly between 0.29325 and 0.03, and 4 eigenvalues $\lambda_1$ to $\lambda_4$
that are specified such that $\lambda_1=0.9$ and $\lambda_1$ to $\lambda_3$ take on values that are specific to the problem instantiation (see results figures). In all cases, $\rho(M)=0.9$.
\end{problem}
\cref{prob18-gmresinf} shows results for \cref{prob:linear200} with eigenvalues $\lambda_1=0.9$, $\lambda_2=-0.3$, $\lambda_3=0.3$, and $\lambda_4=-0.3$. Comparing FP and AA(1) with AA($\infty$), which is essentially equivalent to GMRES. As in previous examples, the asymptotic linear convergence factors of the AA(1) sequences strongly depend on the initial guess, but it is interesting to observe that the asymptotic linear convergence factors of the AA($\infty$) sequences do not depend on the initial guess.
\begin{figure}[h]
\centering
\includegraphics[width=.49\textwidth]{fig10_prob18__gmres1_rconv.pdf}
\includegraphics[width=.49\textwidth]{fig10_prob18__gmres1_inits.pdf}
\includegraphics[width=.325\textwidth]{fig10_prob18__gmres1_qhist.pdf}
\includegraphics[width=.325\textwidth]{fig10_prob18__gmres1_AAhist.pdf}
\includegraphics[width=.325\textwidth]{fig10_prob18__gmres1_AArhist.pdf}
\caption{
\cref{prob:linear200}: linear problem with $M \in \mathbb{R}^{200 \times 200}$ and $\lambda_1=0.9$, $\lambda_2=0.3$, $\lambda_3=-0.3$, and $\lambda_4=-0.3$, for 200 random initial guesses. Comparison of (windowed) AA(1) (red) with restarted GMRES(1) (blue). It can be observed that the asymptotic linear convergence factors of the AA(1) sequences strongly depend on the initial guess, but the asymptotic linear convergence factors of the restarted GMRES(1) sequences do not depend on the initial guess.} \label{prob18-gmres1}
\end{figure}
We next compare AA($m$) with a restarted version of AA($m$). In the restarted version of AA($m$), we simply restart the entire AA($m$) iteration every $m$ steps. Since AA is essentially equivalent to GMRES, this restarted AA($m$) iterations is essentially equivalent to restarted GMRES($m$) with window size $m$.
\cref{prob18-gmres1} compares AA(1) for \cref{prob:linear200} with restarted AA(1), equivalent to GMRES(1).
It is interesting to see that the asymptotic convergence factors of restarted AA(1) do not appear to depend on the initial guess.
Also, most of the sequences $\{x_k\}$ for the standard windowed AA(1) (without restart) appear to have an r-linear convergence factor that is
smaller than the r-linear convergence factor of the restarted AA(1).
\begin{figure}[h]
\centering
\includegraphics[width=.49\textwidth]{fig11_prob19__gmres3_rconv.pdf}
\includegraphics[width=.49\textwidth]{fig11_prob19__gmres3_inits.pdf}
\includegraphics[width=.32\textwidth]{fig11_prob19__gmres3_qhist.pdf}
\includegraphics[width=.32\textwidth]{fig11_prob19__gmres3_AAhist.pdf}
\includegraphics[width=.32\textwidth]{fig11_prob19__gmres3_AArhist.pdf}
\caption{
\cref{prob:linear200}: linear problem with $M \in \mathbb{R}^{200 \times 200}$ and $\lambda_1=0.9$, $\lambda_2=-0.9$, $\lambda_3=0.3$, and $\lambda_4=-0.3$, for 200 random initial guesses.
Comparison of (windowed) AA(3) (red) with restarted AA(3) (blue), which is equivalent to GMRES(3). It can be observed that the asymptotic linear convergence factors of both the windowed AA(3) sequences and the restarted AA(3) sequences strongly depend on the initial guess. Windowed AA(3) converges faster than restarted AA(3).} \label{prob19-gmres3}
\end{figure}
Finally, \cref{prob19-gmres3} considers \cref{prob:linear200} with eigenvalues $\lambda_1=0.9$, $\lambda_2=-0.9$, $\lambda_3=0.7$, and $\lambda_4=-0.7$, comparing windowed AA(3) with restarted AA(3) (which is equivalent to GMRES(3)).
Interestingly, convergence factors for restarted AA(3) appear to depend strongly on the initial guess, similar to windowed AA(3), but unlike restarted AA(1) in \cref{prob18-gmres1}. We also see that windowed AA(3) generally converges faster than restarted AA(3), but this is not surprising because every iteration of windowed AA(3) uses information from three previous iterates (as soon as $k\ge3$), whereas iterations of restarted AA(3) use information from only two previous iterates on average.
The results of \cref{prob18-gmresinf,prob18-gmres1,prob19-gmres3} are interesting because they compare the asymptotic convergence speed of AA($m$) with GMRES and GMRES($m$), and the dependence of the r-linear convergence factor on the initial guess. Needless to say, these results raise many questions that require further investigation. For example, the dependence of GMRES($m$) convergence speed on the initial guess has been observed before and numerical results for small-size problems suggest dependence of GMRES(1) convergence with fractal patterns \cite{embree2003tortoise}, but as far as we know there are only limited theoretical results that explain, bound or quantify dependence of the asymptotic convergence factor on the initial guess for GMRES($m$).
\begin{figure}[h]
\centering
\includegraphics[width=.60\textwidth]{fig12_prob19__sweepm_supconv.pdf}
\caption{
\cref{prob:linear200}: linear problem with $M \in \mathbb{R}^{200 \times 200}$ and $\lambda_1=0.9$, $\lambda_2=-0.9$, $\lambda_3=0.3$, and $\lambda_4=-0.3$. Comparison of worst-case asymptotic convergence factors for (windowed) AA(m) (red) and restarted AA(m) (blue), for 200 random initial guesses in each test.} \label{prob19-sweepm}
\end{figure}
We conclude our discussion of \cref{prob:linear200} with \cref{prob19-sweepm}, which shows how the worst-case r-linear asymptotic convergence factors for windowed AA($m$) and restarted AA($m$) over 200 random initial guesses depend on the window size, $m$. The results of \cref{prob19-sweepm} are for \cref{prob:linear200} with eigenvalues $\lambda_1=0.9$, $\lambda_2=-0.9$, $\lambda_3=0.7$, and $\lambda_4=-0.7$. Since there are four eigenvalues that are much greater than the cluster of 196 eigenvalues between 0.3 and 0, both the windowed and restarted AA($m$) show large gains from increasing $m$ up to $m=4$, after which the improvement tapers off. When $m$ increases beyond 5 (for windowed AA($m$)), the improvements become smaller, because the 196 eigenvalues that are smallest in magnitude are clustered; nevertheless, increasing $m$ further continues to improve the estimated $\rho_{\Psi,x^*}$. This behavior is similar to what could be expected for the well-understood behavior of GMRES without restart, which approximately takes out isolated eigenvalues one-by-one. As before for $m=3$ in \cref{prob19-gmres3}, windowed AA($m$) converges faster than restarted AA($m$) with the same $m$, as expected. However, this does point to an advantage of windowed AA($m$) over restarted AA($m$), because the memory requirements for the two algorithms are the same, and the additional amount of work per step for windowed AA($m$) is usually small because the small least-squares problems solved in AA($m$) tend to be inexpensive relative to the evaluation of the $q(x)$ iteration function.
\section{Conclusion}
\label{sec:disc}
In this paper, we have investigated the continuity and differentiability properties of the iteration function $\Psi(\boldsymbol{z})$ and acceleration coefficient function $\boldsymbol{\beta}(\boldsymbol{z})$ for AA($m$), Andersen acceleration with window size $m$. We have established, for window size $m=1$ and $m>1$, the continuity and Gateaux-differentiability of $\Psi(\boldsymbol{z})$ at the fixed point $\boldsymbol{z}^*$, despite $\boldsymbol{\beta}(\boldsymbol{z})$ not being continuous at $\boldsymbol{z}^*$. These findings shed light on remarkable properties of the asymptotic converge of AA($m$) that we have revealed in numerical experiments, for linear and nonlinear problems. We find that AA($m$) sequences converge r-linearly but their r-linear converge factors depend on the initial guess on a set of nonzero measure, which is consistent with the non-differentiability of $\Psi(\boldsymbol{z})$ at $\boldsymbol{z}^*$. The discontinuity of $\boldsymbol{\beta}(\boldsymbol{z})$ at $\boldsymbol{z}^*$ is consistent with the observed oscillatory behaviour of $\boldsymbol{\beta}^{(k)}$ as $\{x_k\}$ converges to $x^*$.
In exact arithmetic, the rank-deficient case is handled properly by the pseudo-inverse formula of \cref{eq:AAm-beta-form-pseudo}
which computes the minimum-norm solution when the system is singular, and our analysis shows that, while
$\boldsymbol{\beta}(\boldsymbol{z})$ is not continuous at $\boldsymbol{z}^*$, $\Psi(\boldsymbol{z})$ is
continuous and Gateaux-differentiable at $\boldsymbol{z}^*$ so the discontinuity of $\boldsymbol{\beta}(\boldsymbol{z})$ does
not preclude convergence of $\{x_k\}$ to $x^*$.
It is interesting to also relate the findings of this paper to the results from \cite{desterck2020,wang2020} on asymptotic convergence for a stationary version of AA($m$). While, as we have seen in this paper, the asymptotic convergence factor $\rho_{\Psi,x^*}$ of AA($m$) in iteration \cref{eq:AA-fixed-point} cannot easily be computed since $\Psi(\boldsymbol{z})$ is not differentiable, \cite{desterck2020,wang2020} consider a stationary version of AA($m$) where the AA coefficients $\beta^{(k)}_i$ are fixed over all iterations $k$. With fixed coefficients $\beta_i$ in \cref{eq:AA-fixed-point}, $\Psi(\boldsymbol{z})$ in iteration \cref{eq:AA-fixed-point} is differentiable and the linear asymptotic convergence factor of the stationary AA iteration is computable as $\rho_{\Psi,x^*}=\rho(\Psi'(x^*))$. This enables choosing the stationary coefficients $\beta_i$ that minimize $\rho(\Psi'(x^*))$, if $x^*$ and $q'(x)$ are known. This approach is used in \cite{desterck2020,wang2020} to provide insight in the convergence improvement that results from the optimal stationary AA($m$) iteration, based on how AA($m$) improves the eigenvalue spectrum of $q'(x^*)$. Empirical results in \cite{desterck2020,wang2020} for AA($m$) acceleration of large canonical tensor decompositions by the alternating least-squares method, and of large machine learning optimization problems solved by the alternating direction method of multipliers, show that the convergence improvement obtained by the stationary AA($m$) iteration is similar to the convergence improvement provided by the non-stationary AA($m$). The work in \cite{desterck2020,wang2020}, however, as well as the fixed-point analysis of AA($m$) presented in this paper, leave open the question of determining $\rho_{\Psi,x^*}$ for the non-stationary AA($m$) that is widely used in science and engineering applications.
|
\section{introduction}
Engineering and controlling dipole-dipole interactions is at the heart of analog quantum simulation,
which can be realized through manipulating the field environment between emitters \cite{RevModPhys.86.153}.
A fascinating platform is the nanophotonic structure, wherein the dispersion relation can be designed on-demand,
and the atom-photon bound states can form within the band and band-gap when atoms are placed nearby
\cite{PhysRevLett.64.2418,Subwavelength2015AGT,Quantum2015Douglas,Hood10507,PhysRevA.97.043831,PhysRevLett.125.163602}.
The photonic component in the bound state is virtually excited, thereby can mediate long-range tunable dipole-dipole interactions without dissipations into the guided modes.
Usually, the photon bound state is exponentially localized and isotropic around the emitter, which limits the form of dipole-dipole interactions
\cite{Subwavelength2015AGT,Quantum2015Douglas}.
However, through properly modifying the photonic bath, photon bound states can be anisotropic
\cite{Belloeaaw0297,PhysRevX.11.011015,doi:10.1021/acsphotonics.8b01455,PhysRevLett.126.043602},
power-law scaling \cite{PhysRevA.97.043831,PhysRevLett.125.163602,PhysRevA.103.033511}, and even phase tunable
\cite{PhysRevResearch.2.023003,PhysRevLett.126.063601,PhysRevLett.126.103603,PhysRevLett.126.203601}.
These special bound states can be used to study more exotic many-body phases.
In addition, there are other new quantum phenomena when emitters interact with structured photonic baths,
which have potential applications in quantum information processing
\cite{PhysRevLett.115.063601,PhysRevLett.119.143602,PhysRevA.96.043811,Barik666,PhysRevLett.122.203603,PhysRevA.99.053852}.
A new paradigm to capture novel bound states is taking advantage of topological photonic lattices
\cite{PhysRevA.97.043831,Belloeaaw0297,PhysRevX.11.011015,PhysRevLett.125.163602,
PhysRevA.103.033511,PhysRevLett.126.063601,vega2021qubitphoton}.
In particular, in the dimer photonic lattice (a photonic analog of the Su-Schrieffer-Heeger (SSH) model),
there exist chiral photonic bound states with only one side envelope with respect to the emitter
\cite{Belloeaaw0297,PhysRevX.11.011015}.
These directional bound states can mediate directional dipole-dipole interactions.
As a natural extension, the trimer lattice allows for chiral edge states that appear at one side of the lattice \cite{PhysRevA.99.013833}.
These edge states have a direct connection with that of two-dimensional Aubry-Andr\'{e}-Harper (AAH) models
\cite{PhysRevA.96.032103},
thus is robust against disorders.
Using such chiral edge modes, the adiabatic topological pumping
has been predicted and experimentally observed in 1D phononic lattices
\cite{PhysRevLett.123.034301,PhysRevLett.126.095501}.
However, so far the interaction with emitters and emitter-emitter interactions in this type of photonic or phononic lattice have been unexplored.
On the other hand, phonon, the quanta of mechanical vibration, is regarded as an alternative quantum information carrier
\cite{Habraken_2012,Gustafsson207,PhysRevX.5.031031,PhysRevLett.117.015502,Chu199,PhysRevLett.120.213603,
Quantum2018Satzinger,Bienfait368,PhysRevLett.124.053601,PhysRevLett.125.153602,doi:10.1063/5.0024001,phononic2021Neuman}.
The spin-mechanical hybrid systems may overcome the shortcoming of vacuum radiations in nanophotonic structures,
because phonons do not decay into free-space.
One of the most promising platforms is diamond crystal
\cite{Burek:16,Sipahigil847,Evans662,CHIA2021219}, since in terms of nanofabrication,
it is capable of fabricating high-quality mechanical modes at the GHz frequency,
while integrating color centers as long-lived spin qubits
\cite{Balasubramanian2009Ultralong,Maurer1283,Bar2013Solid,PhysRevLett.112.036405,Tao2014Single,
Lee2017Topical,PhysRevLett.118.223603,PhysRevLett.119.223602,Quantum2019Bradac,Coupling2021Kazuhiro,Quantum2021Romain}.
So far, the strain coupling mechanism between diamond electric spins and mechanical modes has been widely explored,
providing new opportunities for realizing phonon networks in the strong-coupling regime
\cite{PhysRevLett.110.156402,PhysRevB.88.064105,PhysRevLett.111.227602,PhysRevLett.113.020503,
Dynamic2014Ovartchaiyapong,Strong2015Barfuss,PhysRevLett.116.143602,PhysRevX.6.041060,Controlling2018Sohn,
PhysRevB.97.205444,PhysRevX.8.041027,Coherent2020Maity,PhysRevA.101.042313,PhysRevA.103.013709}.
In particular, diamond phononic crystal waveguides can host phononic band gaps and the spin-phonon bound states can form
when the spin's frequency lies within this special frequency range
\cite{BandPeng,PhysRevResearch.2.013121,PhysRevResearch.3.013025}.
Such spin-phonon bound states can mediate long-range tunable spin-spin interactions for simulating spin models in hybrid quantum systems.
Moreover, the phononic topological states as well as their interactions with one or more emitters
at the quantum level are widely investigated in phononic crystals
\cite{PhysRevX.5.031011,Lemonde_2019,PhysRevB.101.085108,https://doi.org/10.1002/adfm.201904784}.
Combining topological phononic structures with solid-state spins may give rise to exotic spin-phonon bound states
and other interesting sound-matter interaction phenomena.
In this work, we consider sound-matter interactions in dimer and trimer phononic lattices,
where an array of solid-state defects are integrated into a diamond phononic crystal.
The dimer and trimer lattices used are an array of coupled phononic crystal cavities with spatially modulated hopping rates.
In contrast to previous works where the phononic edge modes are obtained from the breakdown of the time-reversal symmetry
with the assistance of optical modes \cite{PhysRevX.5.041002,Lemonde_2019,Two2020Ren},
here we obtain the edge states through the periodic modification of the hopping rates in the 1D phononic crystal.
In the dimer lattice, we show the formation of chiral spin-phonon bound states,
where the spin mimics the behaviour of a boundary to localize phonons on its one side.
Furthermore, when two spins are considered, the collective decay depends on the phononic waveguide structure between them,
and more interestingly, the phonon collective decay can be topology-dependent.
While in the trimer lattice,
the chiral bound states still exist with a number of six that corresponds to six chiral edge states,
occurring at certain energy and sublattices.
When the spin's frequency is resonant with the band, we show the spin relaxation becomes sublattice-dependent as a result of mirror symmetry breaking.
Furthermore, due to the topology or topological origin, the chiral spin-phonon bound states in both lattice structures
are robust to large amounts of disorder, which can mediate chiral and robust spin-spin interactions.
These interactions enable the exploration of exotic many-body phases such as double N\'{e}el ordered states
\cite{Belloeaaw0297,bello2021spin}.
The quantum control over phononic edge states is also considered in a finite system
in terms of population inversion between spins and edge states.
We also illustrate how to design the required lattice structure in a 1D diamond nanobeam
and show that our scheme is feasible under current experimental conditions.
This work provides a novel platform for manipulating phonons and is useful for phonon-based quantum applications.
\section{The setup}
\begin{figure}
\includegraphics[scale=0.14]{model.eps}
\caption{\label{fig1}(Color online) (a) Schematic of an array of phononic cavites in a diamond nanobeam
with integrated SiV centers. (b,c) Level structure of electronic ground state of single SiV centers
and strain coupling of SiV spins to mechanical modes via a Raman process (b) or directly (c).}
\end{figure}
We consider a full lattice of phononic cavities in diamond crystals,
where a single solid-state spin is integrated into each lattice site, as depicted in Fig.~1(a).
The acoustic resonators have identical resonance frequency $\omega_m$ and
each pair of adjacent ones are coupled via either direct near-field coupling or a phonon waveguide,
wherein the hopping rates are tunable via structure designs.
The Hamiltonian of this coupled phononic cavity waveguide is written as (setting $\hbar=1$)
\begin{eqnarray}\label{ME1}
\hat{H}_\text{ph}=\omega_m\sum_{n}\hat{d}_n^\dag\hat{d}_n-\sum_{n}(J_n\hat{d}_n^\dag\hat{d}_{n+1}+\text{H.c.}),
\end{eqnarray}
where $\hat{d}_n$ is the annihilation operator of the phononic mode,
and $J_n$ is the tunnelling strength between the $n$th cavity and the $(n+1)$th cavity.
The solid-state spins considered in this setup are SiV centers,
whose electric ground states have a splitting of $\sim46$ GHz
and possess a large strain susceptibility due to the spin-orbit coupling.
In the presence of a static magnetic field, the two orbit states are split into four energy levels, as shown in Fig.~1(b).
The sublevels $|g\rangle$ and $|e\rangle$ can be viewed as a long-lived spin,
which can be indirectly coupled to the acoustic modes with frequencies of $46$ GHz
via a Raman process (see Fig.~1(b)) or directly coupled to the acoustic modes with frequencies of several GHz (see Fig.~1(c))
\cite{PhysRevLett.120.213603,PhysRevB.97.205444}.
In both cases, the single phonon coupling strength can reach MHz such that the strong coupling condition can be satisfied.
Under the rotating wave approximation, the total Hamiltonian of the system is
\begin{eqnarray}\label{ME2}
\hat{H}_\text{tot}=\hat{H}_\text{ph}+\sum_{n}\omega_\sigma^{(n)}\hat{\sigma}_+^n\hat{\sigma}_-^n
+g\sum_{n}(\hat{\sigma}_{+}^n\hat{d}_n+\text{H.c.}),
\end{eqnarray}
with $\omega_\sigma^{(n)}$ the (effective) resonance frequency of the $n$th spin,
$\hat{\sigma}_+^n=|e\rangle^n\langle g|$,
and $g$ the (effective) spin-phonon coupling strength.
\section{Spin-phonon interactions in a dimer lattice}
\subsection{The Hamiltonian}
\begin{figure}
\includegraphics[scale=0.18]{dispersion.eps}
\caption{\label{fig1}(Color online) Schematic diagram of dimer (a) and trimer (b) lattice.
(c) Dispersion relations of the model in (a), with $J_1=J(1+\delta)$, $J_2=J(1-\delta)$ and $\delta=\pm0.3$.
(d) Dispersion relations of the model in (b), with $(J_a,J_b,J_c)=(1,4,3)J$.}
\end{figure}
When the hopping rates are spatially designed with a periodicity of $2$, the phononic crystal behaves as a dimer lattice,
which is the phononic analog of the Su-Schrieffer-Heeger (SSH) model
and allows for topological phonons at the quantum level.
The Hamiltonian in Eq.~(\ref{ME1}) can be rewritten as
\begin{eqnarray}\label{ME3}
\hat{H}_\text{ph}&=&\omega_m\sum_{n}(\hat{a}_n^\dag\hat{a}_n+\hat{b}_n^\dag\hat{b}_n)\notag\\
&-&\sum_{n}(J_1\hat{a}_n^\dag\hat{b}_n+J_2\hat{b}_n^\dag\hat{a}_{n+1}+\text{H.c.}),
\end{eqnarray}
where $\hat{a}_n$ and $\hat{b}_n$ are respectively the annihilation operator of phonon
at sublattice $A$ and sublattice $B$ of the $n$th unit cell.
The hopping rates can be replaced by $J_1=J(1+\delta)$, and $J_2=J(1-\delta)$,
with $\delta$ the dimerization strength.
To transform the Hamiltonian into momentum space, we
impose periodic boundary conditions and introduce the Fourier transformation
\begin{eqnarray}\label{ME4}
\hat{d}_k=\frac{1}{\sqrt{N}}\sum_{n=1}^{N}e^{-iknd_0}\hat{d}_n.
\end{eqnarray}
For simplicity, we take the lattice constant $d_0=1$ below.
Then the Hamiltonian of the phononic bath is transformed into
$\hat{H}_{\text{ph}}=\sum_k\hat{V}_k^\dag\hat{H}(k)\hat{V}_k$,
with $\hat{V}_k^\dag=(\hat{a}_k^\dag,\hat{b}_k^\dag)$ and the kernel
\begin{eqnarray}\label{ME5}
\hat{H}(k)=
{\left( \begin{array}{ccc}
\omega_m & -J_1-J_2e^{-ik}\\
-J_1-J_2e^{ik} & \omega_m
\end{array}\right )}.
\end{eqnarray}
Diagonalizing the Hamiltonian and taking $\omega_m$ as the energy reference,
we can get the eigenenergies, that is, the dispersion relation
\begin{eqnarray}\label{ME6}
\omega_{\alpha/\beta}(k)&=&\pm\sqrt{J_1^2+J_2^2+2J_1J_2\cos k}\notag\\
&=&\pm J\sqrt{2(1+\delta^2)+2(1-\delta^2)\cos k}.
\end{eqnarray}
The two dispersions are symmetric with respect to the zero energy due to chiral symmetry and we set $\omega_{\alpha}(k)=-\omega_{\beta}(k)\equiv\omega(k)$ below.
The band structure is shown in Fig.~2(c), where a middle bandgap with size $4J|\delta|$ is opened.
In particular, for $\delta<0$, it corresponds to a topological phase of sound
and there exists a pair of phononic edge states exponentially localized at the two ends when the system is finite
\cite{asboth2016short}.
We now consider the spins are coupled to the topological phonons in dimer lattices.
The interaction Hamiltonian of the spins and bath in $k$-space is given by
\begin{eqnarray}\label{ME7}
\hat{H}_{\text{int}}&=&\frac{g}{\sqrt{2N}}\sum_{k,m=A}e^{ikx_m}\hat{\sigma}_{+}^m(\hat{\alpha}_k-\hat{\beta}_k)\notag\\
&+&\frac{g}{\sqrt{2N}}\sum_{k,m=B}e^{i(kx_m-\phi(k))}\hat{\sigma}_{+}^m(\hat{\alpha}_k+\hat{\beta}_k)+\text{H.c.},
\end{eqnarray}
with $\hat{\alpha}_k$ and $\hat{\beta}_k$ the eigenoperators in the diagonalization basis, $x_m$ the unit cell position
and $\phi(k)\equiv\text{arg}(-J_1-J_2e^{-ik})$.
The first (second) term represents the interactions between the phonons and spins located at the $A$ ($B$) sublattice.
The single spins can acquire extra energy due to their interaction with the quantized phononic field,
which is the self-energy expressed as
\begin{eqnarray}\label{ME8}
\Sigma_e(z)=\sum_{k,i=j}\sum_{s=\alpha,\beta}\frac{\langle 0|\hat\sigma_-^j\hat{H}_\text{int}\hat{s}_k^\dag|0\rangle\langle 0|\hat{s}_k\hat{H}_\text{int}\hat\sigma_+^i|0\rangle}{z-\omega_s(k)},
\end{eqnarray}
with $|0\rangle=|g\rangle|\text{vac}\rangle$.
In particular, within the Markovian approximation,
$-2\text{Im}\Sigma_e(z)$ describes the decay rate and $\text{Re}\Sigma_e(z)$ represents the energy shift.
The interaction with the phononic bath can also lead to collective self-energy $\Sigma_{ij}(z)$ of the spins,
which can be obtained by simply replacing $i\neq j$ in the summation of $\Sigma_e$.
Since the phononic dimer lattice has sublattice (chiral) symmetry,
the pairwise collective self-energy can be classified into two categories:
one from the spins in the same sublattice and the other from the spins in the different sublattice
\begin{eqnarray}\label{ME9}
&\Sigma&_{ij}^{AA/BB}(z)=\pm\frac{-g^2zy_\text{min}^{|x_{ij}|}}{\sqrt{z^4-4J^2(1+\delta^2)z^2+16J^4\delta^2}}\\
&\Sigma&_{ij}^{AB}(z)=\pm\frac{g^2J[(1+\delta)y_\text{min}^{|x_{ij}|}+(1-\delta)y_\text{min}^{|x_{ij}+1|}]}
{\sqrt{z^4-4J^2(1+\delta^2)z^2+16J^4\delta^2}}\label{ME10},
\end{eqnarray}
where $x_{ij}=x_j-x_i$ is the cell distance between two spins,
$y_\text{min}=min(y_+,y_-)$ with respect to their absolute values and,
\begin{eqnarray}\label{ME11}
y_\pm=\frac{z^2-2J^2(1+\delta^2)\pm\sqrt{z^4-4J^2(1+\delta^2)z^2+16J^4\delta^2}}{2J^2(1-\delta^2)}.
\end{eqnarray}
The sign in these two self-energy expressions is positive when $y_\text{min}=y_+$ and negative when $y_\text{min}=y_-$.
The self-energy $\Sigma_e(z)=\Sigma_{i=j}^{AA/BB}(z)$ of single spins is also given here.
\subsection{Chiral spin-phonon bound states}
\begin{figure}
\includegraphics[scale=0.18]{dimerabba.eps}
\caption{\label{fig1}(Color online) Chiral spin-phonon bound states at zero energy.
(a,c,d) Single spins are coupled to the $A$ sublattice. (b) Single spins are coupled to the $B$ sublattice.
(c) With off-diagonal disorders. (d) In the presence of on-site disorders.
The disorder strength in both cases is $\xi_n\in[-J/4,J/4]$.
Note that the spin-phonon coupling strength is $g=0.3J$.}
\end{figure}
When the quantum emitters are coupled to this bath,
the related topological features will be imprinted into the emitter-bath interaction.
Specially, if the emitters' frequency is chosen to be strictly equal to the zero-energy modes,
the emitters can act as an effective edge of the lattice and localize the excitation on only one side
\cite{Belloeaaw0297,PhysRevX.11.011015}.
In this section, we show the formation of chiral spin-phonon bound states in this SiV-phononic crystal model.
We consider single spins coupled to the phononic dimer lattice, whose frequency lies within the bandgap.
The spin no longer decays into the propagating modes, but is dressed by the acoustic modes close to the bandedges.
The stationary state wavefunction in the single-excitation subspace can be expanded by
\begin{eqnarray}\label{ME12}
|\psi\rangle=(C_e\hat{\sigma}_{+}+\sum_{k}\sum_{s=a,b}C_{k,s}\hat{s}_k^\dag)|g\rangle|\text{vac}\rangle,
\end{eqnarray}
with $C_e$ the probability amplitude of the spin being in the upper state and
$C_{k,s}$ the probability amplitude for finding a phononic excitation in sublattice $A/B$ at the wavevector $k$.
The coefficients can be obtained by solving the secular equation
$\hat{H}_\text{tot}|\psi\rangle=E_\text{BS}|\psi\rangle$, yielding
\begin{eqnarray}\label{ME13}
C_{k,a}^{A(B)}&=&\frac{gC_eE_\text{BS}}{E_\text{BS}^2-\omega^2(k)}\quad \bigg(\frac{gC_e\omega(k)e^{i\phi(k)}}{E_\text{BS}^2-\omega^2(k)}\bigg)\\
C_{k,b}^{A(B)}&=&\frac{gC_e\omega(k)e^{-i\phi(k)}}{E_\text{BS}^2-\omega^2(k)}\quad
\bigg(\frac{gC_eE_\text{BS}}{E_\text{BS}^2-\omega^2(k)}\bigg)\label{ME14},
\end{eqnarray}
where the superscript in the wavefunction labels the sublattice (at the $j=0$ unit cell) to which the spin is coupled.
Doing the Fourier transformation, we can obtain the corresponding spatial distribution in sublattice $A/B$.
In this work, we mainly focus on the bound states at zero energy.
A direct observation of the expressions shows that, when setting $E_\text{BS}=0$,
$C_{j,a}^{A}=C_{j,b}^{B}=0$, with $j$ the unit cell index.
The other two spatial distributions can be given analytically, similar to
the calculation of the self-energy.
For the case of $\delta>0$ and the spin at the $j=0$ unit cell,
\begin{eqnarray}\label{ME15}
C_{j,b}^{A}=\left\{
\begin{aligned}
&\frac{gC_e(-1)^{j}}{J(1+\delta)}\bigg(\frac{1-\delta}{1+\delta}\bigg)^{j}, &j\geq0\\
&0, &j<0
\end{aligned}
\right.
\end{eqnarray}
and $C_{j,a}^{B}=C_{-j,b}^{A}$,
which indicates that the left (right) bound state vanishes when the spin is coupled to $A$\,($B$) sublattice.
The chirality of the bound states arises from the inversion symmetry breaking of the phononic bath
with respect to the coupling point, and reaches its maximum at $\omega_\sigma=\omega_m$.
These two perfect chiral bound states are shown in Fig.~3(a) and 3(b),
with numerical integrations of Eq.~(\ref{ME13}) and Eq.~(\ref{ME14}).
Note that the number of chiral bound states is equal to that of the edge states.
We next consider the robustness of this chiral spin-phonon bound states in the dimer lattice.
A finite chain with $40$ phononic cavities is used to simulate the SSH bath.
The chiral bound states in the presence of disorder in the phonon hopping strength are depicted in Fig.~3(c),
with adding the random terms $\hat{H}_d=\sum_{n}(\xi_{n}\hat{d}_n^\dag\hat{d}_{n+1}+\text{H.c.})$ to the total Hamiltonian.
The disorder strength $\xi_{n}$ is chosen within the range $[-W/2, W/2]$ for the $n$th lattice site.
We show that the chiral behavior of the bound states is protected, where the chiral symmetry persists.
As a contrast, the diagonal disorder is also taken into account, which is shown in Fig.~3(d).
We find that the middle bound state now has distributions in the A/B sublattices and both sides of the spin.
Besides, the energy of the bound state is no longer strictly fixed at zero.
\subsection{Topology-dependent collective radiation}
\begin{figure}
\includegraphics[scale=0.24]{abcoll.eps}
\caption{\label{fig1}(Color online)
The collective decay rate $\Gamma_{ij}$ of two spins, normalized to $\Gamma_e$, versus the spin's frequency in the upper band.
Here, the distance between the spins is fixed with $x_{ij}=2$,
$\delta=0.3$ for the trivial phase (purple line) and $\delta=-0.3$ for the topological phase (green line).}
\end{figure}
Though the main topological feature is most pronounced in the zero-energy modes,
the dynamics of spins in the phononic dimer lattice is different from that in the standard waveguide.
In the weak coupling limit, the spin dynamics is subject to Eq.~(\ref{ME9}) and Eq.~(\ref{ME10}),
therein the imaginary part represents the decay rate.
From Eq.~(\ref{ME9}) (in the case $i=j$), we show that the decay rate of the spins is sublattice-independent,
which can be roughly understood from the viewpoint that the spins couple to the two topology-different semi-infinite waveguides.
Without loss of generality, we consider two spins resonant with the phononic band.
There are three cases: i) the spins are coupled to $AA/BB$ sublattice;
ii) the spins are coupled to $AB$ sublattice (from left to right);
iii) the spins are coupled to $BA$ sublattice.
In these three situations, the waveguide structures between the spins
are standard-like, trivial and topological, respectively.
As a result, the phonon-induced collective decay rate is different for the three cases.
More interestingly, case II and case III are interchangeable
by only varying the sign of the parameter $\delta$.
Therefore, the collective radiation is topology-dependent.
The collective decay rate $-2\text{Im}\Sigma_{ij}$ can also be calculated by using
the identical relation $\lim_{y\rightarrow0^+}1/(x\pm iy)=P(1/x)\mp i\pi\delta(x)$,
which can provide a more intuitive insight to the features above.
In the standard waveguide, the collective decay rate oscillates versus the emitter's relative position, i.e.,
$\Gamma_{ij}=\Gamma_e\cos(kx_{ij})$.
In contrast, the collective decay rate in the SSH bath is given by
\begin{eqnarray}\label{ME16}
\Gamma_{ij}^\text{AA/BB}(\omega)&=&\text{sign}(\omega)\Gamma_e(\omega)\cos(k(\omega)x_{ij})\\
\Gamma_{ij}^\text{AB/BA}(\omega)&=&\text{sign}(\omega)\Gamma_e(\omega)\cos(k(\omega)x_{ij}\mp\phi(k))\label{ME17}.
\end{eqnarray}
Here, $x_{ij}$ is the cell-distance.
Obviously, the topology-dependent parameter (phase) $\phi(k)$ enters into Eq.~(\ref{ME17}).
In Fig.~4, we plot $\Gamma_{ij}^\text{AB}(\omega)$ as a function of spin frequency $\omega_\sigma$ inside the upper band,
with setting $x_{ij}=2$ and $\delta=\pm0.3$.
The case is opposite for $\omega_\sigma$ inside the lower band, i.e.,
$\Gamma_{ij}^\text{AB}(\omega)=-\Gamma_{ij}^\text{AB}(-\omega)$.
Perfect super or subradiance occurs when $\Gamma_{ij}^\text{AB}=\pm\Gamma_{e}$.
In particular, we find that the number of phononic super or subradiant points depends on the sign of $\delta$:
In the trivial phase ($\delta>0$), there are $x_{ij}-1$ perfect superradiant points inside the two passbands,
while in the topological phase ($\delta<0$), there are $x_{ij}$ frequency points.
\subsection{Chiral spin-spin interactions}
\begin{figure}
\includegraphics[scale=0.18]{abspinspin238.eps}
\caption{\label{fig1}(Color online)
Chiral and robust spin-spin interactions in the trivial phase ($\delta=0.3$) of the dimer lattice.
(a) Three spins are respectively coupled to the second, third and $8$th phononic cavity of a 12 site cavity chain.
Chiral spin-phonon bound states are plot.
(b,c) Population of three spins with time evolution, with the initial state $|\psi_0\rangle=|1\rangle_1|0\rangle_2|0\rangle_3$.
(b) With disorder on the hopping rates $J_1$ and $J_2$. (c) With diagonal disorder.
The disorder strength in both cases is $\xi_n\in[-J/2,J/2]$.}
\end{figure}
The chiral spin-phonon bound states discussed above can mediate spin-spin interactions
through the exchange of virtual phonons.
As a result, the spin-spin interactions will inherit the features appearing in the bound states:
i) decay exponentially with the spins' distance;
ii) the spin located in sublattice $A$\,($B$) only interacts with spin located in sublattice $B$\,($A$) on the one side;
iii) robust against the off-diagonal disorder.
For $\delta>0$, the interaction in the Markovion limit is given as
\begin{eqnarray}\label{ME18}
\hat{H}_{s-s}=\sum_{i<j}J_{ij}(\hat{\sigma}_+^i\hat{\sigma}_-^j+\hat{\sigma}_+^j\hat{\sigma}_-^i)
\end{eqnarray}
with the coupling strength $J_{ij}^{AA/BB}=0$ and
\begin{eqnarray}\label{ME19}
J_{ij}^{AB}=\left\{
\begin{aligned}
&\frac{g^2(-1)^{x_{ij}}}{J(1+\delta)}\bigg(\frac{1-\delta}{1+\delta}\bigg)^{x_{ij}}, &x_{ij}\geq0\\
&0, &x_{ij}<0
\end{aligned}
\right..
\end{eqnarray}
To numerically show these chiral spin-spin interactions, we use a finite lattice chain with $12$ phononic cavities
and consider three spins coupled to the second, third and $8$th cavity (see Fig.~5(a)).
The off-diagonal disorder is also taken into account.
The quantum dynamics of the three spins is plot in Fig.~5(b),
with the initial state $|\psi_0\rangle=|1\rangle_1|0\rangle_2|0\rangle_3$.
We find that the population of spin 3 is always zero and the population is oscillating between
spin 1 and spin 2 with an amplitude approaching one,
which reflects the directionality and robustness of the spin-spin interactions.
Similarly, the quantum dynamics of spins in the presence of diagonal disorder is plot in Fig.~5(c) for comparison.
Apart from extra population in spin 3, we find that the population of spin 1 can not reach zero, since
the energy value of the bound state is no longer pinned at zero in this case.
\subsection{Quantum control over phononic edge states}
\begin{figure}
\includegraphics[scale=0.23]{bbbccc.eps}
\caption{\label{fig1}(Color online)
Control of phononic edge states in dimer lattice
in a finite system composed of $12$ phononic cavities, with a single spin inside the $5$th cavity.
The time evolution of population in cavities array and the spin is plot. Here, we choose $\delta=-0.3$.}
\end{figure}
In the finite system, edge effects inevitably occur especially in the topological phase,
where a pair of edge states exponentially localize at the two ends of the lattice chain.
The phononic edge states in a phononic crystal can be controlled with the addition of solid-state spins.
In that case, the excitation can be transferred between spins and phononic edge states,
with eigenenergies $\pm\epsilon$ and eigenstates $|E_\pm\rangle$.
The dynamics is govern by the effective Hamiltonian
\begin{eqnarray}\label{ME20}
\hat{H}_\text{eff}&=&\sum_{\pm}\pm\epsilon|E_\pm\rangle\langle E_\pm|+(\omega_\sigma-\omega_m)|e\rangle\langle e|\notag\\
&+&\sum_{\pm}g_\pm(|e\rangle\langle E_\pm|+|E_\pm\rangle\langle e|),
\end{eqnarray}
with $g_\pm=g\langle E_\pm|\hat{d}_{x_0}^\dag|\text{vac}\rangle$
and $g_+=g_-$ ($g_+=-g_-$) when $\hat{d}_{x_0}^\dag=\hat{b}_{x_0}^\dag$ ($\hat{d}_{x_0}^\dag=\hat{a}_{x_0}^\dag$).
When setting $\omega_\sigma=\omega_m$ and applying the Schr\"{o}dinger equation,
we can obtain the coupled equations of the coefficients
\begin{eqnarray}\label{ME21}
i\dot{C}_e(t)&=&g_+C_+(t)+g_-C_-(t)\\
i\dot{C}_+(t)&=&\epsilon C_+(t)+g_+C_e(t)\\\label{ME22}
i\dot{C}_-(t)&=&-\epsilon C_-(t)+g_-C_e(t)\label{ME23},
\end{eqnarray}
with $C_e$ and $C_\pm$ the wavefuncions in the states $|e\rangle$ and $|E_\pm\rangle$, respectively.
Solving these equations, the spin's time evolution obeys
\begin{eqnarray}\label{ME24}
C_e(t)=\frac{\epsilon^2+2g_+^2\cos(\omega_0t)}{\epsilon^2+2g_+^2},
\end{eqnarray}
with $\omega_0=\sqrt{\epsilon^2+2g_+^2}$.
In the case of a large size cavity chain, $\epsilon\simeq0$ and
the spin exchanges energies between the left or right phononic edge state,
which are symmetry and antisymmetry superpositions of the states $|E_\pm\rangle$, respectively.
Concretely, we consider a single spin coupled to the $5$th phononic cavity of
a $12$ cavity chain designed with staggered hopping strengths and being in the topological phase.
The system's dynamics is shown in Fig.~6.
We show that the excitation is mainly transferred between the spin and the left phononic edge state,
in line with the prediction based on Eq.~(\ref{ME24}).
\section{Spin-phonon interactions in trimer lattices}
Previous work on trimer lattices have predicted the emergence of chiral edge states localized at one end of the system,
without a counterpart on the opposite edge at the same energy in the phase without inversion symmetry protection
\cite{PhysRevA.99.013833,PhysRevLett.123.034301,PhysRevLett.126.095501}.
Though there is no topology (no symmetry), these chiral edge states are still robust
against large amounts of disorder due to the topological origin.
Therefore, it's still possible to obtain chiral spin-phonon bound states where the spin acts as the effective edges, and
realize directional and robust quantum state transfer as well as quantum control over phononic edge states in a trimer lattice.
In addition, the emission dynamics of the spins in this lattice is worth investigating
since it is different from the one in a symmetry protected system.
\subsection{Hamiltonian}
Similar to the dimer lattice, a trimer lattice is modeled with alternating hopping rates and
having a periodicity of 3. The Hamiltonian of the phononic waveguide is rewritten as
\begin{eqnarray}\label{ME25}
\hat{H}_\text{ph}&=&\omega_m\sum_{n}(\hat{a}_n^\dag\hat{a}_n+\hat{b}_n^\dag\hat{b}_n+\hat{c}_n^\dag\hat{c}_n)\notag\\
&-&\sum_{n}(J_a\hat{a}_n^\dag\hat{b}_n+J_b\hat{b}_n^\dag\hat{c}_{n}+J_c\hat{c}_n^\dag\hat{a}_{n+1}+\text{H.c.}),
\end{eqnarray}
with $\hat{a}_n$, $\hat{b}_n$ and $\hat{c}_n$ the annihilation operators
for the $A$, $B$ and $C$ phononic modes at the $n$th unit cell.
In a common trimer lattice, the phonon tunneling strengths $J_a$, $J_b$ and $J_c$ are unequal and there is no symmetry.
The inversion symmetry only recovers in the special case of $J_a=J_b\neq J_c$.
Performing the same operation as in Sec.~III.A, the Hamiltonian can be transformed into momentum space,
with the kernel of Hamiltonian
\begin{eqnarray}\label{ME26}
\hat{H}(k)=
{\left( \begin{array}{ccc}
\omega_m & -J_a & -J_ce^{-ik}\\
-J_a & \omega_m & -J_b\\
-J_ce^{ik} & -J_b & \omega_m
\end{array}\right )}.
\end{eqnarray}
Unlike the dimer lattice where the analytical solution to the dispersion relation is given,
here we introduce a unitary matrix $M=(V_1,V_2,V_3)$ to diagonalize the kernel of Hamiltonian as
$\hat{D}(k)=M^\dag\hat{H}(k)M=\text{diag}(\omega_\alpha,\omega_\beta,\omega_\gamma)$,
with $V_i$ ($i\!=\!1,2,3$) the eigenvectors and $\omega_s$ ($s\!=\!\alpha,\beta,\gamma$) the eigenvalues.
The eigenoperators are given by
$(\hat\alpha_k\,\,\hat\beta_k\,\,\hat\gamma_k)^T=M^\dag(\hat{a}_k\,\,\hat{b}_k\,\,\hat{c}_k)^T$.
Then the Hamiltonian of the phononic trimer lattice in $k$-space reads
\begin{eqnarray}\label{ME27}
\hat{H}_\text{ph}=\sum_{k}\sum_{s=\alpha,\beta,\gamma}\omega_s(k)\hat{s}_k^\dag\hat{s}_k.
\end{eqnarray}
The band structure is numerically plotted in Fig.~2(d), with $J_a=J$, $J_b=4J$ and $J_c=3J$.
Working in $k$-space,
the interaction Hamiltonian for the spins and phonon modes in the trimer lattice becomes
\begin{eqnarray}\label{ME28}
\hat{H}_{\text{int}}=\frac{g}{\sqrt{N}}\sum_{k}\sum_{m,s}
e^{ikx_m}M(m,s)\hat{\sigma}_{+}^m\hat{s}_k+\text{H.c.},
\end{eqnarray}
where the interaction is projected to the new basis.
The self-energy of the spins is expressed as
\begin{eqnarray}\label{ME29}
\Sigma_e(z)=\frac{g^2}{2\pi}\int_{-\pi}^{\pi}dk\sum_{m=n,s}
\frac{\hat{M}(n,s)\hat{M}^{-1}(s,m)}{z-\omega_s(k)}.
\end{eqnarray}
The collective self-energy $\Sigma_{ij}$ of the spins can be obtained by
simply replacing $m\neq n$ in the summation of $\Sigma_e$
and adding a propagation factor $e^{ik(x_n-x_m)}$ to the integral.
\begin{figure}
\includegraphics[scale=0.13]{trimer.eps}
\caption{\label{fig1}(Color online)
Chiral spin-phonon bound states in a trimer lattice, with $(J_a,J_b,J_c)=(1,4,3)J$.
(a-f) Without disorder. (g) Disorder in the intracell hopping $J_a$ and $J_b$.
(h) Disorder in the intercell hopping $J_c$. (i) Disorder in the phononic frequency $\omega_m$.
The disorder strength in the three cases is $\xi_n\in[-J/2,J/2]$.}
\end{figure}
\subsection{Chiral spin-phonon bound states}
When the spin's frequency lies within the phononic bandgap, there exist
spin-phonon bound states in the single-excitation subspace.
The spatial distribution of the phononic part of the bound state can be calculated
by solving the stationary Schr\"{o}dinger equation, yielding
\begin{eqnarray}\label{ME30}
C_{j,n}^{m}=\frac{gC_e}{2\pi}\int_{-\pi}^{\pi}dke^{ikj}\sum_{s}\frac{\hat{M}(n,s)\hat{M}^{-1}(s,m)}{E_{BS}-\omega_s(k)}.
\end{eqnarray}
Here, $C_{j,n}^{m}$ is the wavefunction at $n=a/b/c$ sublattice of the $j$ unit cell
when the spin is coupled to the $m=A/B/C$ sublattice at the $j=0$ unit cell.
Inspired by the fact that chiral spin-phonon bound states appear at the zero-energy modes in a dimer lattice,
we start to find the chiral bound states in a trimer lattice
by setting the energy of the bound states equal to that of the chiral phononic edge states.
For a finite system with hopping rates $(J_a, J_b,J_c)=(1,4,3)J$,
there exist two edge states localized at the right end with energies $\pm4J$.
Thus, we plot the phononic part of the bound states in Fig.~7(a,d)
with setting $E_\text{BS}=\pm4J$ in Eq.~(\ref{ME30}).
We find the spin-phonon bound state is chiral only when the spin is coupled to the $A$ sublattice,
i.e., $C_{j,n}^A=0$ for $j\geq0$ and $C_{j,n}^A\neq0$ for $j<0$.
Also, these bound states have no components in the $A$ sublattice $C_{j,a}^A=0$,
while the amplitudes in the $B$ and $C$ sublattices are equal $|C_{j,b}^A|=|C_{j,c}^A|$.
Furthermore, the bound states are fractionally chiral when the spin is coupled to the $B$ and $C$ sublattices,
which is reflected in the components on the $A$ sublattice, $C_{j,a}^{B/C}=0$ for $j\leq0$ and $C_{j,a}^{B/C}\neq0$ for $j>0$.
However, the edge modes change when varying the boundary of the trimer lattice.
When the finite system is constructed by $(J_a, J_b,J_c)=(3,1,4)J$,
there are two pairs of edge states localized at the boundary, with energies $\pm3J$ and $\pm J$.
While for the finite system constructed by $(J_a, J_b,J_c)=(4,3,1)J$, there is no edge mode.
The bulk properties of the two cases are the same as that of $(J_a, J_b,J_c)=(1,4,3)J$.
Thus, we turn to study the spin-phonon bound states at these frequencies.
In Fig.~7(b,e) and 7(c,f), we show two pairs of chiral bound states located on the right/left side of the spin
with $E_\text{BS}=\pm3J/\pm J$, when the spin is coupled to the $B/C$ sublattice.
Note that the six chiral bound states are one-to-one corresponding to the six edge states.
We now consider the robustness of the chiral bound states in the trimer lattice.
As an example, we add disorder to the finite system with $20$ unit cells,
by setting $(J_a, J_b,J_c)=(1,4,3)J$, where a spin is coupled to the $B$ sublattice with energy $\omega_\sigma=\omega_m+3J$.
In Fig.~7(g), we allow disorder in the intracell hopping $J_a$ and $J_b$ independently.
While in Fig.~7(h), disorder in the intercell hopping $J_c$ is plotted.
On the other hand, bound states with on-site disorder acting in $\omega_m$ are shown in Fig.~7(i).
We observe several features from the numerical results.
First, the directionality is robust to disorder in the intracell hopping.
Second, the bound state has weight in each sublattices and both sides while the eigenenergies have a small deviation,
in the presence of disorder in the intercell hopping or on-site disorder.
Generally, the directional bound state is still robust when the disorder strength is on the order of the coupling/hopping strength.
All these results indicate that the directional spin-phonon bound states
can inherit the robustness of chiral edge states in the trimer lattice.
\begin{figure*}
\includegraphics[scale=0.22]{abcdec.eps}
\caption{\label{fig1}(Color online)
The decay rate $\Gamma_e$ of single spins coupled to $m=A/B/C$ sublattice in trimeried lattice.
(a) Inversion-symmetry broken phase $(J_a,J_b,J_c)=(1,4,3)J$.
(b) Inversion-symmetric phase $(J_a,J_b,J_c)=(1,1,3)J$. (c) $J_a=J_b=J_c=J$.}
\end{figure*}
\subsection{Sublattice-dependent spin relaxation}
From the discussion of chiral spin-phonon bound states in both lattices,
we can conclude that the spin indeed acts as
an effective boundary of the two semi-infinite waveguide structures.
In this case, the system can be divided into three parts:
the spin and two semi-infinite phononic waveguides.
Contrary to what happens in the dimer lattice
where the two semi-infinite structures are the same for spin at the $A/B$ sublattice,
there are six different semi-infinite structures in the trimer lattice
when the spin is respectively coupled to the $A/B/C$ sublattice.
Thus, we predict that $\Sigma_e^A\neq\Sigma_e^B\neq\Sigma_e^C$
in the inversion-symmetry breaking phase of the trimer lattice,
while $\Sigma_e^A=\Sigma_e^C\neq\Sigma_e^B$ in the inversion-symmetric phase of the trimer model.
We note that the real part of the self-energy is the energy shift, which is small compared to the resonance frequency.
Furthermore, the shift is zero when the spin's frequency lies within the band.
Therefore, we mainly focus on the spin relaxation when the frequency is resonant with the phononic waveguide.
Without loss of generality, we consider a single spin with frequency within the upper band,
which is coupled to the $m\!=\!A/B/C$ sublattice, respectively.
In the Markovian regime, the spin's decay rate $\Gamma_e^m(\omega_\sigma)=-2\text{Im}\Sigma_e^m$ can be written as
\begin{eqnarray}\label{ME31}
\Gamma_e^m(\omega_\sigma)=\frac{2g^2\hat{M}(m,0)\hat{M}^{-1}(0,m)}{v_g(\omega_\sigma)},
\end{eqnarray}
with $v_g(\omega_\sigma)=\partial\omega/\partial k|_{\omega=\omega_\sigma}$ the group velocity.
In Fig.~8, we numerically plot the decay rate of the spin coupled to different sublattices
when the spin's frequency lies within the upper band.
The hopping rates are $(J_a, J_b,J_c)=(1,4,3)J$, $(J_a, J_b,J_c)=(1,1,3)J$ and $J_a\!\!=\!\!J_b\!\!=\!\!J_c\!\!=\!\!J$
in Fig.~8(a), 8(b) and 8(c), respectively.
We find $\Gamma_e^A\!\neq\!\Gamma_e^B\!\neq\!\Gamma_e^C$ for $J_a\!\neq\!J_b\neq\!J_c$.
In the special case $J_a\!=\!J_b\!\neq\!J_c$ where inversion symmetry recovers, $\Gamma_e^A=\Gamma_e^C\neq\Gamma_e^B$.
Moreover, the standard 1D band-edge divergence of $\Gamma_e^B$ is canceled
in the inversion-symmetric phase of the trimer model,
when $\omega_\sigma$ approaches the lower band-edge of $\omega_\alpha(k)$.
For comparison, the situation that the spin interacts with a normal coupled cavity waveguide is shown in Fig.~8(c),
which predicts a decay rate $\Gamma_e=g^2/|\sin(k)|$.
In general, the spin relaxation in a trimer lattice is sublattice-dependent, which is consistent with the discussion above.
\subsection{Bound state-mediated spin-spin interactions}
\begin{figure}
\includegraphics[scale=0.18]{abcspinspin.eps}
\caption{\label{fig1}(Color online) Directional and relatively robust spin-spin interactions in trimer lattices.
(a,b) Three spins are coupled to a lattice chain with $12$ phononic cavities.
In (a), spins' position $(x_C,x_A,x_B)=(3,4,5)$. In (b), $(x_B,x_C,x_A)=(5,9,10)$.
(c,d) Time evolution of the population of three spins for the cases in (a) and (b), respectively.
The initial state in (c), $|\psi_0\rangle=|e\rangle_A|g\rangle_B|g\rangle_C$
and in (d), $|\psi_0\rangle=|g\rangle_A|e\rangle_B|g\rangle_C$.
The disorder in $J_a$ and $J_b$ is added independently with the same as that in Fig.~7(g).}
\end{figure}
We now discuss the spin-spin interactions mediated by
the chiral spin-phonon bound states appearing in the trimer lattice.
In the weak coupling limit, the interaction strength between two spins is
\begin{eqnarray}\label{ME32}
J_{ij}^{mn}=\frac{g^2}{2\pi}\int_{-\pi}^{\pi}dke^{ik(j-i)}\sum_{s}\frac{\hat{M}(n,s)\hat{M}^{-1}(s,m)}{E_{BS}-\omega_s(k)},
\end{eqnarray}
with $i,j$ the cell index, and $m,n$ the sublattice index.
As an example, we take advantage of the bound state shown in Fig.~7(b) involving three spins
with one at the $B$ sublattice and two neighbouring spins at the $A/C$ sublattices.
When the spins at the $A/C$ sublattices are placed on the left side of the spin $B$, as shown in Fig.~9(a),
the two neighbouring spins are decoupled to the spin $B$ due to the directionality of the bound state and
the population can be transferred between the two neighbouring spins.
For the case of the spins at the $A/C$ sublattices on the right side of the spin $B$, as shown in Fig.~9(b), the excitation is exchanged
between the spin at the $B$ sublattice and the antisymmetry superposition of the two neighbouring spins.
The dynamics of the three spins can be modeled with an effective Hamiltonian in the interaction picture as
\begin{eqnarray}\label{ME33}
\hat{H}_\text{eff}&=&J_{B}(\hat\sigma_+^B\hat\sigma_-^C-\hat\sigma_+^B\hat\sigma_-^A+\text{H.c.})
+J_{AC}(\hat\sigma_+^A\hat\sigma_-^C+\text{H.c.})\notag\\
&=&\sqrt{2}J_{B}(\hat\sigma_+^B\hat\sigma_-^\text{at}+\text{H.c.})
+J_{AC}(\hat\sigma_+^\text{st}\hat\sigma_-^\text{st}-\hat\sigma_+^\text{at}\hat\sigma_-^\text{at}),
\end{eqnarray}
with $\hat{\sigma}_-^\text{st/at}=(\hat\sigma_-^C\pm\hat\sigma_-^A)/\sqrt{2}$ and $J_B=J_{ij}^{BC}=J_{ij+1}^{BC}$.
In the first situation $J_B=0$ for $j\geq i$, while in the second situation $J_B\neq0$ for $j<i$.
This reflects the directionality of the spin-spin interaction when the spin is coupled to the $B$ sublattice.
When focusing on the case $J_B\neq0$,
we find that the transition is not resonant due to the interaction between the neighbouring spins.
Note that the detuning (the last term in the second line of Eq. (\ref{ME33})) can be removed by external driving fields;
thus the antisymmetric combinations of spin $A$ and spin $C$
acting as a single spin can be resonantly coupled to spin $B$.
Finally, we emphasize that the spin-spin interaction is directional only when one of the spins is at the $B$ sublattice.
We consider a resonator chain composed of $12$ lattice sites to numerically examine the two cases,
as shown in Fig.~9. The disorder in the intracell hopping rate $J_a$ and $J_b$ is also taken into account
and the spin's frequency is $\omega_\sigma\simeq3J$ in both cases.
In the first case, the three spins are placed at the third, $4$th and $5$th lattice sites
and the Rabi oscillation between spin $A$ and spin $C$ is observed in Fig.~9(c),
with the initial state $|\psi_0\rangle=|e\rangle_A|g\rangle_B|g\rangle_C$.
We show the population in spin $B$ keeps zero.
In the second case, the three spins are placed at the $5$th, $9$th and $10$th lattice sites
and the population is transferred between spin $B$ and spin $A/C$
with the initial state $|\psi_0\rangle=|g\rangle_A|e\rangle_B|g\rangle_C$, as shown in Fig.~9(d).
Though the energy of the bound state formed by spin $B$ is strictly pinned under disorders ($E_\text{BS}^B=3J$),
the energy of the bound states formed by spin $A$ and spin $C$ are not strictly pinned ($E_\text{BS}^{A/C}\approx3J$).
Therefore, the population of the spin $B$ can not reach zero with the time evolution.
\subsection{Quantum control over chiral phononic edge states}
\begin{figure}
\includegraphics[scale=0.23]{bbbccc2.eps}
\caption{\label{fig1}(Color online)
Control over the phononic edge states in the trimer lattice
composed of $12$ phononic cavities, with a single spin inside the $5$th cavity.
Time evolution of the population in the cavity array and the spin is plotted.
Here, we choose $(J_a,J_b,J_c)=(1,4,3)J$.}
\end{figure}
Because the edge modes in the trimer lattice remain in the bandgap,
the quantum control on these modes becomes possible with the addition of spins.
Furthermore, since the edge states in the inversion-symmetry breaking phase
of the trimer lattice have no degeneracy,
we can directly manipulate one of the phononic edge states
on the boundary of the phononic crystal with the addition of a single spin.
The resonance transfer can occur via tuning the frequency of the spin, with the interaction Hamiltonian
\begin{eqnarray}\label{ME34}
\hat{H}_{e-s}=g_{e-s}(|E\rangle\langle e|+|e\rangle\langle E|),
\end{eqnarray}
where $g_{e-s}\!=\!g\langle E|\hat{d}_{x_0}^\dag|\text{vac}\rangle$ is the coupling rate.
As an example, we consider the situation that a single spin is placed at the $5$th cavity of a $12$ cavity chain,
with alternating hopping strengths $(J_a,J_b,J_c)=(1,4,3)J$.
There exist two edge states localized on the right side of the system with energy $E_\text{edge}=\pm4J$.
In Fig.~10, we plot the time evolution of the system and show a transition
with the amplitude approaching one between the spin and one of the chiral phononic edge states.
\section{experimental feasibility}
In this section, we discuss the experimental feasibility of our system.
As illustrated in Fig.~1(a), an array of defect cavities can be fabricated in a diamond nanobeam.
As an example, we consider the phononic cavities with resonance frequency $\omega_m/2\pi\sim5$ GHz
and two adjacent phononic cavities are coupled via a phononic waveguide.
In the rotating frame, the cavity-waveguide-cavity Hamiltonian takes the form of
\begin{eqnarray}\label{ME35}
\hat{H}_\text{cwc}&=&\sum_{k}\delta_k\hat{f}_k^\dag\hat{f}_k\notag\\
&+&\sum_{k}\{(g_{k,1}\hat{d}_1+g_{k,2}(-1)^k\hat{d}_2)\hat{f}_k^\dag+\text{H.c.}\},
\end{eqnarray}
where $\delta_k$ is the detuning between the cavity and the $k$th waveguide mode,
$\hat{f}_k^\dag$ ($\hat{f}_k$) is the creation (annihilation) operator of the $k$th waveguide mode,
$g_{k,i}$ is the cavity-waveguide coupling and the term $(-1)^k$ accounts for the symmetry of the waveguide modes.
Consider that the waveguide is short and a single mode is nearly resonant with the cavity,
the Hamiltonian can be simplified to
\begin{eqnarray}\label{ME36}
\hat{H}_\text{cwc}&=&\delta\hat{f}^\dag\hat{f}
+g_1(\hat{d}_1\hat{f}^\dag+\hat{f}\hat{d}_1^\dag)\notag\\
&+&g_2(\hat{d}_2\hat{f}^\dag+\hat{f}\hat{d}_2^\dag).
\end{eqnarray}
When taking the large detuning limit $g_i\ll\delta$ and applying the perturbation theory,
we can arrive at the effective Hamiltonian describing hopping interactions between the two adjacent cavities
\begin{eqnarray}\label{ME37}
\hat{H}_\text{cc}=J_\text{hop}(\hat{d}_1^\dag\hat{d}_2+\hat{d}_2^\dag\hat{d}_1),
\end{eqnarray}
with $J_\text{hop}=g_1g_2/\delta$ the effective hopping rate between the phononic cavities.
The cavity-waveguide coupling is tunable via designing the mirror section of the nanocavity,
which connects the cavity and the waveguide.
Meanwhile, the detuning can be tuned via changing the structure of the connecting waveguide.
So far, the hopping strengths between adjacent mechanical resonators
have been designed and realized with a few megahertz
\cite{Generalized2017Fang,PhysRevX.8.041027}.
We now show how to design the lattice structure in this work.
As an example, we consider the case of the dimer lattice.
The phononic cavities are designed with different mirror sections on the left and right sides:
one with a large number of phononic crystal periods and the other with a smaller number of periods,
forming a strong mirror and a weak mirror respectively.
The strong/weak mirror can give rise to small/large cavity-waveguide coupling
\cite{PhysRevX.8.041027,Optical2016Fang,Generalized2017Fang,Mirhosseini2020Mirhosseini}.
Thus, if we always align the strong mirror section with the strong mirror section
and use identical waveguide to connect the nanocavities,
the hopping strengths are alternating with the strengths $J_1=g_w^2/\delta$ and $J_2=g_s^2/\delta$.
Note that the trimer lattice can be engineered with the same method.
For a typical nanobeam cavity shown in Refs.\!
\cite{Optical2016Fang,Generalized2017Fang,Mirhosseini2020Mirhosseini},
it has an intrinsic decay rate $\gamma_i$ on the order of MHz.
However, the phonons are virtually excited in the process of spin-spin interactions,
where the excited number of the phonons $\sim0.01$ is at a low level.
Furthermore, we can use phononic shielding to surround the nanobeam (the whole network),
which can provide a phononic bandgap to mitigate the leakage to the bulk
\cite{PhysRevX.8.041027}.
This protection of phonon modes can result in high quality factor $Q$ exceeding $10^7$,
and even ultra-high $Q\sim10^{10}$ of nanocavities in the 1D/2D optomechanical crystal
has been demonstrated in recent experiments
\cite{MacCabe840,Two2020Ren}.
Correspondingly, the intrinsic damping can be reduced to below kilohertz ($\gamma_i/2\pi<1$ kHz).
SiV centers are point defects in diamond with a silicon atom occupying two adjacent vacancies,
which can be embedded into diamond crystal via ion implantation technology.
As depicted in Fig.~1(c), the phononic modes can be directly coupled to the spin in a single SiV
through matching the spin's frequency to the phonon frequency,
where an external magnetic field tilted from the symmetry axis of the defect should be applied.
The spin-phonon coupling strength can be estimated by \cite{PhysRevLett.120.213603,PhysRevB.97.205444}
\begin{eqnarray}\label{ME38}
g=\frac{d_\text{spin}}{v}\sqrt{\frac{\hbar\omega_m}{2\rho V}}\xi(\vec{r}_\text{SiV}).
\end{eqnarray}
Here, $d_\text{spin}/2\pi\sim100$ THz/strain is the strain sensitivity,
$v\sim10^4$ m/s is the group velocity of acoustic waves in diamond,
$\omega_m/2\pi\sim5$ GHz, $\rho\sim3500$ kg/$\text{m}^3$,
$V$ is the volume of the optomechanical nanocavity and $\xi(\vec{r}_\text{SiV})\sim1$
is the strain distribution at the position of the SiV centers.
Thus, for a typical nanocavity, the single spin-phonon coupling strength can reach $g/2\pi\sim1$ MHz.
The main limit of the coherence of SiV centers is the dephasing time $T^*_2$,
which arises from the coupling to the thermal environment and the nuclear spins in diamond.
The thermal occupation can be frozen out in the low temperature
and the interactions with the nuclear spins can be suppressed using the dynamical decoupling technology
\cite{Hanson352,Preserving2009Du,PhysRevB.92.224419}.
For the dilution refrigeration temperature $T\sim10$ mk,
the spin dephasing $\gamma_s/2\pi\sim100$ Hz has been realized in the experiment
\cite{PhysRevLett.119.223602}.
In general, we choose $J/2\pi\sim3$ MHz and $g/2\pi\sim1$ MHz in the main text,
which leads to the spin-spin coupling strength $J_{ij}/2\pi\sim100$ kHz.
The dissipations of the spins and the phononic crystal can be neglected,
since $\{\gamma_i,\gamma_s\}\ll J_{ij}$.
\section{conclusion}
We have proposed a feasible scheme in the solid-state platform to
explore novel interactions between spins and phonons in dimer and trimer lattices,
with an array of SiV centers integrating into a 1D optomechanical crystal.
We have predicted a series of unconventional phenomena in this spin-phononic hybrid system.
In the dimer lattice, we show the formation of chiral
and robust spin-phonon bound states, which can mediate directional and robust spin-spin interactions.
We also show topology-dependent phononic collective radiation effects.
The coupling between a single spin and the phononic edge modes is discussed as well.
While in the trimer lattice, we show that the chiral and robust spin-phonon bound states still exist and the spin relaxation is sublattice-dependent.
The bound state-mediated spin-spin interactions exhibit chiral features and are relatively robust.
Quantum control over the chiral edge states, which are unique in the inversion-symmetry breaking phase of the trimer lattice,
is also demonstrated with the addition of spins.
Finally, we discuss the feasibility of this scheme under realistic parameters.
This work takes the advantage of phononic crystals and SiV centers,
thus providing a promising platform for manipulating phonons with spins,
and is useful for phonon-mediated quantum information applications.
\section*{Acknowledgments}
We gratefully acknowledge the use of the open source
Python numerical packages QuTiP
\cite{JOHANSSON20121760,JOHANSSON20131234}.
This work is
supported by the National Natural Science Foundation of
China under Grant No. 92065105 and
Natural Science Basic Research Program of Shaanxi
(Program No. 2020JC-02).
|
\section{Introduction}
In the last few decades, a significant amount of research has been invested in studying the curse of dimensionality of a distance measure. We have now a good understanding of the curse in terms of (i) concentration of measure \citep{Talagrand} and its consequence of being unable to find the nearest neighbor \citep{Beyer:1999}, and (ii) hubness \citep{HubsInSpace-JMLR-2010}.
Despite this advancement, breaking the curse is still elusive. The central question remains open: \textit{is it possible to find the single nearest neighbor of a query in high dimensions?}
The answer to this question is negative, based on the current analyses on distance \citep{Beyer:1999,Density-based-Indexing-KDD1999,Pestov}. Rather than finding the single nearest neighbor of a query, the issue has been diverted to finding all data points belonging to the same cluster as the query, in a distribution with well-separated clusters \citep{Beyer:1999,Density-based-Indexing-KDD1999}. These nearest neighbors are `meaningful' for the purpose of separating different clusters in a dataset. But the (single) nearest neighbor, in the true sense of the term, cannot be obtained using a distance measure, even in the special distribution of well-separated clusters described above \citep{Beyer:1999}.
Unlike many existing measures, a recently introduced Isolation Kernel \citep{ting2018IsolationKernel,IsolationKernel-AAAI2019} is derived directly from a finite dataset without learning and has no closed-form expression.
We show here that Isolation Kernel is not affected by the concentration effect, and it works well in high dimensions; and it does not need to use low intrinsic dimensions to explain why it works well in high dimensions.
Our theoretical analysis focuses on \textit{the ability of a feature map produced by kernel evaluation using a finite dataset to distinguish two distinct points in high dimensions} (\textbf{distinguishability} or \textbf{indistinguishability} hereafter). A kernel is said to have indistinguishability if its feature map is unable to distinguish two distinct points. This is manifested as being unable to find the single nearest neighbor of a query because there are many nearest neighbors which cannot be distinguished from each other, especially in high dimensions.
Our contributions are:
\vspace{-1mm}
\renewcommand{\labelenumi}{\Roman{enumi}}
\begin{enumerate}[itemsep=1ex, leftmargin=5mm]
\item Revealing that Isolation Kernel (IK) is the only measure that is provably free from the curse of dimensionality,
compared with existing metric-based Lipschitz continuous kernels.
\item Determining three key factors in IK that lead to breaking the curse of dimensionality.
(a) Space partitioning using an isolation mechanism is the key ingredient in measuring similarity between two points in IK.
(b) The probability of a point falling into a partition
is independent of data distribution, the distance measure used to create the partitions and the data dimensions.
(c) The unique feature map of IK has its dimensionality linked to a concatenation of these partitionings.
\item Proving that IK's unique feature map produced by a finite dataset has distinguishability,
independent of the number of data dimensions and data distribution.
Our theorem suggests that increasing the number of partitionings (i.e., the dimensionality of the feature map) leads to increased distinguishability.
\item Verifying that IK
breaks the curse of dimensionality in two sets of experiments. First, IK finds the single nearest neighbor in high dimensions; while many existing measures fail under the same condition. Second, IK enables consistently effective indexing, spectral \& density peaks clustering, SVM classification and t-SNE visualization in both low and high dimensions.
\item Showing that Euclidean distance, Gaussian and linear kernels have a mixed bag of poor and good results in high dimensions, reflecting the current understanding of existing measures.
\end{enumerate}
\section{Related work}
\subsection{The curse of dimensionality}
\label{sec_curse}
One of the early investigations of the curse of dimensionality is in terms of concentration of measure \cite{Mil_man_1972,Talagrand}. \citet{Talagrand} summarises the intuition behind the concentration of measure
as follows: ``a random variable that depends (in a `smooth' way) on the influence of many independent
random variables (but not too much on any of them) is essentially constant.''
Translating this statement in the context of a high dimensional space, a distance measure which depends on many independent dimensions (but not too much on any of them) is essentially constant.
\citet{Beyer:1999} analyze the conditions under which nearest neighbor is `meaningful' when a distance measure is employed
in high dimensions.
A brief summary of the conditions in high dimensions is given as follows:
\renewcommand{\labelenumi}{\alph{enumi})}
\begin{enumerate}
\item Finding the (single) nearest neighbor of a query is not meaningful because every point from the same cluster in the database has approximately the same distance to the query. In other words, the variance of distance distribution within a cluster approaches zero as the number of dimensions increases to infinity.
\item Despite this fact, finding the exact match is still meaningful.
\item The task of retrieving all points belonging to the same cluster as the query is meaningful because although variance of distances of all points within the cluster is zero, the distances of this cluster to the query are different from those of another cluster---making a distinction between clusters possible.
\item The task of retrieving the nearest neighbor of a query, which does not belong to any of the clusters in the database, is not meaningful for the same reason given in (a).
\end{enumerate}
\citet{Beyer:1999} reveal that finding (non-exact match) nearest neighbors is `meaningful' only if the dataset has clusters (or structure), where the nearest neighbors refer to any/all points in a cluster rather than a specific point which is the nearest among all points in the dataset. In short, a distance measure cannot produce the `nearest neighbor' in high dimensions, in the true sense of the term. Hereafter the term `nearest neighbor' is referred to the single nearest neighbor in a dataset, not many nearest neighbors in a cluster.
\citet{Beyer:1999} describe the notion of instability of a distance measure as follows: ``{\em A nearest neighbor query is unstable for a given $\epsilon$ if the distance from the query point to most data points is less than (1 + $\epsilon$) times the distance from the query point to its nearest neighbor.}''
Current analyses on the curse of dimensionality focus on distance measures \citep{Beyer:1999,NN-HighDim-VLDB-2000,Aggarwal-2001,HubsInSpace-JMLR-2010}. While random projection preserves structure in the data, it still suffers the concentration effect \citep{Ata-Kaban-2011}.
Current methods dealing with high dimensions are: dimension reduction \citep{PCA-Book2002,Hinton2008Visualizing}; suggestions to use non-metric distances, e.g., fractional distances \citep{Aggarwal-2001}; and angle-based measures, e.g., \citep{Angle-BasedAD-KDD2008}. None has been shown to break the curse.
Yet, the influence of the concentration effect in practice is less severe than suggested mathematically \citep{Concentration-Fractionaldistances}. For example, k-nearest neighbor (kNN) classifier still yields reasonable high accuracy in high dimensions \citep{Aryal2017}; while the concentration effect would have rendered kNN producing random prediction. One possible explanation of this phenomenon is that the high dimensional datasets employed have low intrinsic dimensions (e.g., \citep{Concentration-Fractionaldistances,LID-NIPS2011}.)
In a nutshell, existing works center around reducing the effect of concentration of measures,
knowing that the nearest neighbor could not be obtained in high dimensions.
A saving grace is that many high dimensional datasets used in practice have low intrinsic dimensions---leading to a significantly reduced concentration effect.
Our work shows that the concentration effect can be eliminated to get the nearest neighbor in high dimensions using a recently introduced measure called Isolation Kernel \citep{ting2018IsolationKernel,IsolationKernel-AAAI2019}, without the need to rely on low intrinsic dimensions. While three existing measures have a mixed bag of poor and good results in high dimensions, Isolation Kernel defined by a finite dataset yields consistently good results in both low and high dimensions.
\subsection{Isolation Kernel}
Unlike commonly used distance or kernels, Isolation Kernel \citep{ting2018IsolationKernel,IsolationKernel-AAAI2019} has no closed-form expression, and it is derived directly from a given dataset without learning; and the similarity between two points is computed based on the partitions created in data space.
The key requirement of Isolation Kernel (IK) is a space partitioning mechanism which isolates a point from the rest of the points in a sample set. This can be achieved in several ways, e.g., isolation forest \citep{liu2008isolation,ting2018IsolationKernel}, Voronoi Diagram \citep{IsolationKernel-AAAI2019} and isolating hyperspheres \citep{IDK-KDD2020}.
Isolation Kernel \citep{ting2018IsolationKernel,IsolationKernel-AAAI2019} is defined as follows.
Let $\mathbb{H}_\psi(D)$ denote the set of all admissible partitionings $H$ derived from a given finite dataset $D$, where each $H$ is derived from $\mathcal{D} \subset D$; and each point in $\mathcal{D}$ has the equal probability of being selected from $D$; and $|\mathcal{D}|=\psi$.
Each partition $\theta[{\bf z}] \in H$ isolates a point ${\bf z} \in \mathcal{D}$ from the rest of the points in $\mathcal{D}$. The union of all partitions of each partitioning $H$ covers the entire space.
\begin{definition}
Isolation Kernel of any two points ${\bf x},{\bf y} \in \mathbb{R}^d$ is defined to be
the expectation taken over the probability distribution on all partitionings $H \in \mathds{H}_\psi(D)$ that both ${\bf x}$ and ${\bf y}$ fall into the same isolating partition $\theta[{\bf z}] \in H$, where ${\bf z} \in \mathcal{D} \subset D$, $\psi=|\mathcal{D}|$:
\begin{eqnarray}
K_\psi({\bf x},{\bf y}\ |\ D) = {\mathbb E}_{\mathds{H}_\psi(D)} [\mathds{1}({\bf x},{\bf y} \in \theta[{\bf z}]\ | \ \theta[{\bf z}] \in H)],
\label{eqn_kernel}
\end{eqnarray}
\end{definition}
\noindent
where $\mathds{1}(\cdot)$ is an indicator function.
In practice, Isolation Kernel $K_\psi$ is constructed using a finite number of partitionings $H_i, i=1,\dots,t$, where each $H_i$ is created using randomly subsampled $\mathcal{D}_i \subset D$; and $\theta$ is a shorthand for $\theta[{\bf z}]$:
\begin{eqnarray}
K_\psi({\bf x},{\bf y}\ |\ D) & \simeq & \frac{1}{t} \sum_{i=1}^t \mathds{1}({\bf x},{\bf y} \in \theta\ | \ \theta \in H_i) \nonumber\\
& = & \frac{1}{t} \sum_{i=1}^t \sum_{\theta \in H_i} \mathds{1}({\bf x}\in \theta)\mathds{1}({\bf y}\in \theta).
\label{Eqn_IK}
\end{eqnarray}
This gives a good approximation of $K_\psi({\bf x},{\bf y}\ |\ D)$ when $|D|$ and $t$ are sufficiently large to ensure that the ensemble is obtained from a sufficient number of mutually independent $\mathcal{D}_i, i=1,\dots,t$.
Given $H_i$, let $\Phi_i({\bf x})$ be a $\psi$-dimensional binary column (one-hot) vector representing all $\theta_j \in H_i$, $j=1,\dots,\psi$; where ${\bf x}$ must fall into only one of the $\psi$ partitions.
The $j$-component of the vector is:
$\Phi_{ij}({\bf x})=\mathds{1}({\bf x}\in \theta_j\ |\ \theta_j\in H_i)$. Given $t$ partitionings, $\Phi({\bf x})$ is the concatenation of $\Phi_1({\bf x}),\dots,\Phi_t({\bf x})$.
\begin{prop}
\label{prop:featureMap}
\textbf{Feature map of Isolation Kernel.}
The feature mapping $\Phi: {\bf x}\rightarrow \mathbb \{0,1\}^{t\times \psi}$ of $K_\psi$, for point ${\bf x} \in \mathbb{R}^d$, is a vector that represents the partitions in all the partitionings $H_i\in \mathds{H}_\psi(D)$, $i=1,\dots,t$ that contain ${\bf x}$; where ${\bf x}$ falls into only one of the $\psi$ partitions in each partitioning $H_i$. Then, $\parallel {\Phi}({\bf x}) \parallel\ = \sqrt{t}$ holds, and Isolation Kernel always computes its similarity using its feature map as:
\[
K_\psi({\bf x},{\bf y}\ |\ D) \simeq \frac{1}{t} \left< {\Phi}({\bf x}|D), {\Phi}({\bf y}|D) \right>.
\]
\end{prop}
Let $\mathbbmtt{1}$ be a shorthand of $\Phi_i({\bf x})$ such that $\Phi_{ij}({\bf x})=1$ and $\Phi_{ik}({\bf x})=0, \forall k \neq j$ of any $j \in [1,\psi]$.
The feature map
${\Phi}({\bf x})$ is sparse because it has exactly $t$ out of $t \psi$ elements having 1 and the rest are zeros. This gives $\parallel {\Phi}({\bf x}) \parallel\ = \sqrt{t}$.
Note that while every point has $\parallel {\Phi}({\bf x}) \parallel\ = \sqrt{t}$ and $\Phi_i({\bf x}) = \mathbbmtt{1}$ for all $i \in [1,t]$, they are not all the same point.
Note that Isolation Kernel is a positive definite kernel with probability 1 in the limit of $t \rightarrow \infty$ because its Gram matrix is full rank as $\Phi(x)$ for all points $x \in D$ are mutually independent (see \cite{IDK-KDD2020} for details.) In other words, the feature map coincides with a reproducing kernel Hilbert space (RKHS) associated with the Isolation Kernel.
This finite-dimensional feature map has been shown to be the key in leading to fast runtime in kernel-based anomaly detectors \citep{IDK-KDD2020} and enabling the use of fast linear SVM \citep{IsolationSetKernel,IK-XFactor-2019}.
IK has been shown to be better than Gaussian and Laplacian kernels in SVM classification \citep{ting2018IsolationKernel}, better than Euclidean distance in density-based clustering \citep{IsolationKernel-AAAI2019}, and better than Gaussian kernel in kernel-based anomaly detection \citep{IDK-KDD2020}.
IK's superiority has been attributed to its data dependent similarity, given in the following lemma.
Let $D \subset \mathcal{X}_D \subseteq {\mathbb R}^d$ be a dataset sampled from an unknown distribution $\mathcal{P}_D$ where $\mathcal{X}_D$ is the support of $\mathcal{P}_D$
on ${\mathbb R}^d$; and
let $\rho_D({\bf x})$ denote the density of $\mathcal{P}_D$ at point ${\bf x} \in \mathcal{X}_D$, and $\ell_p$ be the $\ell_p$-norm.
The unique characteristic of Isolation Kernel \citep{ting2018IsolationKernel,IsolationKernel-AAAI2019} is:
\begin{lem}\label{lem_characteristic}
Isolation Kernel $K_\psi$ has the data dependent characteristic:
\vspace{-2mm}
\begin{equation}
K_\psi( {\bf x},{\bf y}\ |\ D) > K_\psi( {\bf x}',{\bf y}'\ |\ D)\nonumber
\end{equation}
for $\ell_p({\bf x}-{\bf y})\ =\-\ell_p({\bf x}'-{\bf y}')$,
$\forall {\bf x},{\bf y} \in \mathcal{X}_\mathsf{S}$ and $\forall {\bf x}',{\bf y}' \in \mathcal{X}_\mathsf{T}$
subject to $\forall {\bf z}\in \mathcal{X}_\mathsf{S}, {\bf z}'\in \mathcal{X}_\mathsf{T}, \ \rho_D({\bf z})<\rho_D({\bf z}')$.
\end{lem}
As a result, $K_\psi( {\bf x},{\bf y}\ |\ D)$ is not translation invariant, unlike the data independent Gaussian and Laplacian kernels.
We show in Section \ref{sec_distinguishability} that this feature map is instrumental in making IK the only measure that breaks the curse, compared with existing metric-based Lipschitz continuous kernels.
\section{Analyzing the curse of dimensionality in terms of (in)distinguishability}
\label{sec_break-curse}
In this section, we theoretically show that IK does not suffer from the concentration effect in any high dimensional data space;
while metric-based Lipschitz continuous kernels, including Gaussian and Laplacian kernels, do.
We introduce the notion of indistinguishability of two distinct points in a feature space $\mathcal{H}$ produced by kernel evaluation
using a finite dataset
to analyze the curse of dimensionality. Let $\Omega \subset \mathbb{R}^d$ be a data space, $D \subset \Omega$ be a finite dataset, and $\mathcal{P}_D$ be the probability distribution that generates $D$. The support of $\mathcal{P}_D$, denoted as $\mathcal{X}_D$, is an example of $\Omega$.
Given two
distinct points ${\bf x}_a,{\bf x}_b \in \Omega$ and their feature vectors
$\phi({\bf x}_a), \phi({\bf x}_b) \in \mathcal{H}_\kappa$ produced by a kernel $\kappa$ and a finite dataset $D = \{{\bf x}_i \in \mathbb{R}^d \; | \; i=1, \dots, n\}$ as $\phi({\bf x}_a) = [\kappa({\bf x}_a,{\bf x}_1),\dots,\kappa({\bf x}_a,{\bf x}_n)]^\top$ and
$\phi({\bf x}_b) = [\kappa({\bf x}_b,{\bf x}_1),\dots,\kappa({\bf x}_b,{\bf x}_n)]^\top$.
\begin{definition}
$\mathcal{H}_\kappa$ has \textbf{indistinguishability} if
there exists some $\delta_L \in (0,1)$ such that
the probability of $\phi({\bf x}_a) = \phi({\bf x}_b)$ is lower bounded by $\delta_L$ as follows:
\begin{equation}
P(\phi({\bf x}_a) = \phi({\bf x}_b)) \geq \delta_L. \label{indist}
\end{equation}
\end{definition}
On the other hand, given two
feature vectors
$\Phi({\bf x}_a), \Phi({\bf x}_b) \in \mathcal{H}_K$ produced by a kernel $K$ and a finite dataset $D$.
\begin{definition}
$\mathcal{H}_K$ has \textbf{distinguishability} if there exists some $\delta_U \in (0,1)$ such that $P(\Phi({\bf x}_a) = \Phi({\bf x}_b))$ is upper bounded by $\delta_U$ as follows:
\begin{equation}
P(\Phi({\bf x}_a) = \Phi({\bf x}_b)) \leq \delta_U. \label{eqn_distinguishability}
\end{equation}
\end{definition}
Note that these definitions of indistinguishability and distinguishability are rather weak. In practice, $\delta_L$ should be close to 1 for strong indistinguishability, and $\delta_U$ should be far less than 1 for strong distinguishability.
In the following, we first show in Section \ref{sec_indistinguishability} that a feature space $\mathcal{H}_\kappa$ produced by any metric-based Lipschitz
continuous kernel $\kappa$ and a finite dataset $D$ has indistinguishability when the number of dimensions $d$ of the data space is very large.
Then, in Section \ref{sec_distinguishability}, we show that a feature space $\mathcal{H}_K$ procudd by Isolation Kernel $K$ implemented using the Voronoi diagram \citep{IsolationKernel-AAAI2019}
has distinguishability for any $d$.
\subsection{$\mathcal{H}_\kappa$ of a metric-based Lipschitz
continuous kernel $\kappa$ has indistinguishability}
\label{sec_indistinguishability}
Previous studies have analyzed the concentration
effect in high dimensions~\citep{Pestov, Gromov, Ledoux}. Let a triplet
$(\Omega, m, F)$ be a probability metric space where $m: \Omega \times \Omega \mapsto \mathbb{R}$
is a metric, and
$F: 2^\Omega \mapsto \mathbb{R}$ is a probability measure. $\ell_p$-norm is an
example of $m$. $\Omega$ is the support of $F$, and thus $F$ is assumed to have non-negligible probabilities over the entire $\Omega$, i.e., points drawn from $F$ are widely distributed over the entire data space.
Further, let $f: \Omega \mapsto \mathbb{R}$ be a Lipschitz continuous function, i.e., $|f({\bf x}_a) - f({\bf x}_b)| \leq m({\bf x}_a,{\bf x}_b), \forall {\bf x}_a,{\bf x}_b \in \Omega$;
and $M$ be a median of $f$ defined as:
\begin{eqnarray}
F(\{{\bf x} \in \Omega \; | \; f({\bf x}) \leq M\}) = F(\{{\bf x} \in \Omega \; | \; f({\bf x}) \geq M\}).\nonumber
\end{eqnarray}
Then, the following proposition holds~\citep{Pestov, Gromov}.
\begin{prop}\label{prop1}
$F(\{{\bf x} \in \Omega | f({\bf x}) \in [M-\epsilon,M+\epsilon]\}) \geq 1-2\alpha(\epsilon)$ holds for any $\epsilon>0$, where
$\alpha(\epsilon) = C_1 e^{-C_2 \epsilon^2 d}$ with two constants $0 < C_1 \leq \frac{1}{2}$ and $C_2 > 0$.
\end{prop}
For our analysis on the indistinguishability of a feature space $\mathcal{H}_\kappa$ associated with any metric-based Lipschitz continuous kernel, we reformulate this proposition
by replacing $f({\bf x})$ and $M$ with $m({\bf x},{\bf y})$ and $M({\bf y})$, respectively, for
a given ${\bf y} \in \Omega$. That is:
\begin{eqnarray}
&&\hspace*{-5mm} |m({\bf x}_a,{\bf y}) - m({\bf x}_b,{\bf y})| \leq m({\bf x}_a,{\bf x}_b), \forall {\bf x}_a,{\bf x}_b \in \Omega, \mbox{ and}\nonumber\\
&&\hspace*{-5mm}F(\{{\bf x} \in \Omega \; | \; m({\bf x},{\bf y}) \leq M({\bf y})\}) = F(\{{\bf x} \in \Omega \; | \; m({\bf x},{\bf y}) \geq M({\bf y})\}).\nonumber
\end{eqnarray}
The inequality in the first formulae always holds because of the triangular inequality of $m$.
Then, we obtain the following corollary:
\begin{col}\label{col1}
Under the same condition in Proposition~\ref{prop1}, the inequality $F(A_\epsilon({\bf y})) \geq 1- 2\alpha(\epsilon)$ holds for any ${\bf y} \in \Omega$
and any $\epsilon > 0$, where
$A_\epsilon({\bf y}) = \{{\bf x} \in \Omega \; | \; m({\bf x},{\bf y}) \in [M({\bf y})-\epsilon, M({\bf y})+\epsilon]\}$.
\end{col}
This corollary represents the concentration effect where almost all
probabilities of the data are concentrated in the limited area $A_\epsilon({\bf y})$.
It provides an important fact on the indistinguishability of two distinct points using
a class of metric-based Lipschitz continuous kernels such as Gaussian kernel:
\begin{equation}
\kappa({\bf x},{\bf y}) = f_\kappa(m({\bf x},{\bf y})),\nonumber
\end{equation}
where $f_\kappa$ is Lipschitz continuous and monotonically decreasing for $m({\bf x},{\bf y})$.
\begin{lem}\label{lem1}
Given two distinct points ${\bf x}_a$, ${\bf x}_b$ i.i.d. drawn from $F$ on $\Omega$
and a data set $D = \{{\bf x}_i \in \mathbb{R}^d \; | \; i=1, \dots, n\} \sim G^n$
where every ${\bf x}_i$ is i.i.d drawn from any probability distribution $G$ on
$\Omega$, let the feature vectors of ${\bf x}_a$ and ${\bf x}_b$ be
$\phi({\bf x}_a) = [\kappa({\bf x}_a,{\bf x}_1),\dots,\kappa({\bf x}_a,{\bf x}_n)]^\top$ and
$\phi({\bf x}_b) = [\kappa({\bf x}_b,{\bf x}_1),\dots,\kappa({\bf x}_b,{\bf x}_n)]^\top$ in a feature space $\mathcal{H}_\kappa$. The following holds:
\begin{equation}
P\left(\ell_p(\phi({\bf x}_a)-\phi({\bf x}_b)) \leq 2L\epsilon n^{1/p} \right) \geq (1-2\alpha(\epsilon))^{2n}, \nonumber
\end{equation}
for any $\epsilon > 0$, where $\ell_p$ is an $\ell_p$-norm on $\mathcal{H}_\kappa$, and
$L$ is the Lipschitz constant of $f_\kappa$.
\end{lem}
\begin{thm}\label{thm1}
Under the same condition as Lemma~\ref{lem1}, the feature space $\mathcal{H}_\kappa$
has indistinguishability in the limit of $d \rightarrow \infty$.
\end{thm}
This result implies that the feature space $\mathcal{H}_\kappa$ produced by any finite dataset and any metric-based
Lipschitz continuous kernels such as Gaussian and Laplacian kernels have indistinguishability
when the data space has a high number of dimensions $d$. In other words, these kernels suffer
from the concentration effect.
\subsection{$\mathcal{H}_K$ of Isolation Kernel $K$ has distinguishability}
\label{sec_distinguishability}
We show that the feature map of Isolation Kernel implemented using Voronoi diagram \citep{IsolationKernel-AAAI2019}, which is not a metric-based
Lipschitz continuous kernel, has distinguishability.
We derive Lemma \ref{lem2} from the
fact given in \cite{Fukunaga}, before providing our theorem.
\begin{lem}\label{lem2}
Given ${\bf x} \in \Omega$ and a data set
$\mathcal D = \{ {\bf z}_1, \dots, {\bf z}_\psi \} \sim G^\psi$ where
every ${\bf z}_j$ is i.i.d. drawn from any probability distribution $G$ on
$\Omega$ and forms its Voronoi partition $\theta_j \subset \Omega$, the probability that ${\bf z}_j$ is the nearest neighbor
of ${\bf x}$ in $\mathcal D$ is given as: $P({\bf x} \in \theta_j) = 1/\psi$, for every $j=1,\dots,\psi$.
\end{lem}
This lemma points to a simple but nontrivial result that
$P({\bf x} \in \theta_j)$
is independent of $G$, $m({\bf x},{\bf y})$ and data dimension $d$.
\begin{thm}\label{thm2}
Given two distinct points ${\bf x}_a$, ${\bf x}_b$ i.i.d. drawn from $F$ on $\Omega$
and ${\mathcal D}_i = \{ {\bf z}_1, \dots, {\bf z}_\psi \} \sim G^\psi$ defined in
Lemma~\ref{lem2}, let the feature vectors of ${\bf x}_a$ and ${\bf x}_b$ be $\Phi({\bf x}_a)$ and $\Phi({\bf x}_b)$
in a feature space $\mathcal{H}_K$ associated with Isolation Kernel $K$, implemented using Voronoi diagram, as given in Proposition \ref{prop:featureMap}.
Then the following holds:
\[
P(\Phi({\bf x}_a)=\Phi({\bf x}_b)) \leq \cfrac{1}{\psi^t},
\]
and $\mathcal{H}_K$ always has strong distinguishability for large $\psi$ and $t$.
\end{thm}
Proofs of Lemmas \ref{lem1} \& \ref{lem2} and Theorems \ref{thm1} \& \ref{thm2} are given in Appendix \ref{sec_A_Proofs}.
In summary, \textbf{the IK implemented using Voronoi diagram is free from the curse of dimensionality for any probability distribution}.
\subsection{The roles of $\psi$ \& $t$ and the time complexity of Isolation Kernel}
The sample size (or the number of partitions in one partitioning) $\psi$ in IK has a function similar to the bandwidth parameter in Gaussian kernel, i.e., the higher $\psi$ is, the sharper the IK distribution \citep{ting2018IsolationKernel} (the small the bandwidth, the sharper the Gaussian kernel distribution.) In other words, $\psi$ needs to be tuned for a specific dataset to produce a good task-specific performance.
The number of partitionings $t$ has three roles. First, the higher $t$ is, the better Eq \ref{Eqn_IK} is in estimating the expectation described in Eq \ref{eqn_kernel}. High $t$ also leads to low variance of the estimation. Second, with its feature map, $t$ can be viewed as IK's user-definable effective number of dimensions in Hilbert space. Third, Theorem~\ref{thm2} reveals that increasing $t$ leads to increased distinguishability. The first two roles were uncovered in the previous studies \citep{ting2018IsolationKernel,IsolationKernel-AAAI2019,IDK-KDD2020}. We reveal the last role here in Theorem \ref{thm2}. These effects can be seen in the experiment reported in Appendix \ref{sec_partitiongs}.
In summary, \textbf{IK's feature map, which has its dimensionality linked to distinguishability, is unique among existing measures}. In contrast, Gaussian kernel (and many existing metric-based kernels) have a feature map with intractable dimensionality; yet, they have indistinguishability.
\textbf{Time complexity}:
IK, implemented using the Voronoi diagram, has time complexity $O(ndt\psi)$ to convert $n$ points in data space to $n$ vectors in its feature space. Note that each $\mathcal{D}$ implemented using one-nearest-neighbor produces a Voronoi diagram implicitly, no additional computation is required to build the Voronoi diagram explicitly (see \citet{IsolationKernel-AAAI2019} for details.) In addition, as this implementation is amenable to acceleration using parallel computing, it can be reduced by a factor of $w$ parallelizations to $O(\frac{n}{w}dt\psi)$.
The readily available finite-dimensional feature map of IK enables a fast linear SVM to be used. In contrast, Gaussian kernel (GK) must use a slow nonlinear SVM. Even adding the IK's feature mapping time, SVM with IK still runs up to two orders of magnitude faster than SVM with GK. See Section \ref{sec_traditional_tasks} for details.
\section{Empirical verification}
\label{sec_experiments}
We verify that IK breaks the curse of dimensionality by conducting two experiments.
In Section~\ref{sec_instability}, we verify the theoretical results in terms of instability of a measure, as used in a previous study of the concentration effect \citep{Beyer:1999}. This examines the (in)distinguishability of linear, Gaussian and Isolation kernels.
In the second experiment in Section \ref{sec_traditional_tasks}, we explore the impact of IK having distinguishability in three tasks: ball tree indexing \citep{omohundro1989five}, SVM classification \citep{CC01a} and t-SNE visualization \citep{Hinton2008Visualizing}.
\begin{figure*}[t]
\centering
\begin{subfigure}[b]{0.245\textwidth}
\centering
\includegraphics[width=1.5in]{cluster1.png}
\caption{Dataset used}
\end{subfigure}
\begin{subfigure}[b]{0.245\textwidth}
\centering
\includegraphics[width=1.6in]{GKIKvdmcen1.png}
\caption{Concentration effect}
\end{subfigure}
\begin{subfigure}[b]{0.245\textwidth}
\centering
\includegraphics[width=1.6in]{GKIKLKb.png}
\caption{${\bf q}_i$: between 2 clusters}
\end{subfigure}
\begin{subfigure}[b]{0.245\textwidth}
\centering
\includegraphics[width=1.6in]{GKIKLKcen.png}
\caption{${\bf q}_c$: sparse cluster center}
\end{subfigure}
\caption{The instability of a measure $\kappa$: Gaussian kernel (GK) and Linear kernel (LK) versus IK. The dataset used has 1,000 data points per cluster: the number of dimensions is increased from 10 to 10,000.
Query ${\bf q}_c$ is used in subfigure (b).
$\epsilon=0.005$ is used in subfigures (c) \& (d).
}
\label{fig_std-mean}
\end{figure*}
\begin{table*}[t]
\caption{The instability of a measure.
The dataset shown in Figure \ref{fig_std-mean}(a) is used here. $\epsilon=0.005$.}
\label{tbl_SNN_AG_IK}
\centering
\begin{tabular}{cc|cc}
\hline
\multicolumn{2}{c|}{SNN, AG and IK} & \multicolumn{2}{c}{$\ell_p$-norm for $p=0.1,0.5$ \& $2$ } \\ \hline
${\bf q}_i$: between 2 clusters & ${\bf q}_c$: sparse cluster center & ${\bf q}_i$: between 2 clusters & ${\bf q}_c$: sparse cluster center\\ \hline
\includegraphics[width=1.3in]{IKSNN2AGcen005.png} &
\includegraphics[width=1.3in]{IKSNN2AGspa005.png} &
\includegraphics[width=1.3in]{lpcen.png} & \includegraphics[width=1.3in]{lpspa.png}
\\\hline
\end{tabular}
\end{table*}
\subsection{Instability of a measure}
\label{sec_instability}
The concentration effect is assessed in terms of the variance of measure distribution $F_{m,d}$ wrt the mean of the measure, i.e., $var \left( \frac{F_{m,d}}{\mathbb{E}[F_{m,d}]} \right)$, as the number of dimensions $d$ increases,
based on a measure $m({\bf x},{\bf y}) = 1 - \kappa({\bf x},{\bf y})$.
A measure is said to have the concentration effect if the variance of the measure distribution in dataset $D$ approaches zero as $d \rightarrow \infty$.
Following \cite{Beyer:1999}, we use the same notion of instability of a measure, as a result of a nearest neighbor query, to show the concentration effect. Let $m({\bf q}|D) = \min_{{\bf y} \in D} 1- \kappa({\bf q},{\bf y})$ be the distance as measured by $\kappa$ from query ${\bf q}$ to its nearest neighbor in dataset $D$; and $N_\epsilon$ be the number of points in $D$ having distances wrt ${\bf q}$ less than $(1+\epsilon)m({\bf q}|D)$.
When there are many close nearest neighbors for any small $\epsilon>0$, i.e., $N_\epsilon$ is large, then the nearest neighbor query is said to be unstable. An unstable query is a direct consequence of the concentration effect.
Here we use $N_\epsilon$ as a proxy for indistinguishability: for small $\epsilon$, high $N_\epsilon$ signifies indistinguishability; and $N_\epsilon=1$ or close to 1 denotes distinguishability.
Figure \ref{fig_std-mean} shows the verification outcomes using the dataset shown in Figure \ref{fig_std-mean}(a).
Figure \ref{fig_std-mean}(b) shows the concentration effect, that is, $var \left( \frac{F_{m,d}}{\mathbb{E}[F_{m,d}]} \right) \rightarrow 0$ as $d \rightarrow \infty$ for Gaussian kernel (GK). Yet, using Isolation Kernel (IK), the variance approaches a non-zero constant under the same condition.
Figures \ref{fig_std-mean}(c) \& \ref{fig_std-mean}(d) show that $N_\epsilon$ increases as $d$ increases for GK for two different queries; and LK is unstable in one query. Yet, IK has almost always maintained a constant $N_\epsilon$ ($=1$) as $d$ increases.
\textbf{SNN, AG, Linear kernel and fractional distances}:
Here we examine the stability of Shared Nearest Neighbor (SNN) \citep{SNN,SNN-Defeat-Curse?}, Adaptive Gaussian kernel (AG) \citep{zelnik2005self} and fractional distances \cite{Aggarwal-2001}.
Our experiment using SNN has shown that it has query stability if the query is within a cluster; but it has the worst query instability when the query is outside any clusters when $k$ is set less than the cluster size (1000 in this case), even in low dimensions!
This result is shown in the first two subfigures in Table~\ref{tbl_SNN_AG_IK}. AG ($k=200$) has a similar behaviour as SNN that uses $k>1000$.
The last two subfigures in Table \ref{tbl_SNN_AG_IK} show a comparison between Euclidean distance and fractional distances ($\ell_p, p=0.1,0.5$). It is true that the fractional distances are better than Euclidean distance in delaying the concentration effect up to $d=1000$; but they could not resist the concentration effect for higher dimensions.
\textbf{Summary}. Many existing measures have indistinguishability in high dimensions that have prevented them from finding the nearest neighbor if the query is outside of any clusters.
\textbf{Only IK has distinguishability}. Note that IK uses distance to determine the (Voronoi) partition into which a point falls; but IK's similarity does not directly rely on distance. IK has a unique $t$-dimensional feature map that has increased distinguishability as $t$ increases, described in Section~\ref{sec_distinguishability}. No existing metric-based Lipschitz continuous kernels have a feature map with this property.
\subsection{The impact of (in)distinguishability on four traditional tasks}
\label{sec_traditional_tasks}
We use the same ten datasets in the following four tasks. All have been used in previous studies (see Appendix \ref{ap1}), except two synthetic datasets: Gaussians and $w$-Gaussians. The former has two well-separated Gaussians, both on the same 10000 dimensions; and the latter has two $w$-dimensional Gaussians which overlap at the origin only, as shown in Figure \ref{fig_SC}. Notice below that algorithms which employ distance or Gaussian kernel have problems with the $w$-Gaussians dataset in all four tasks; but they have no problems with the Gaussians dataset.
\begin{table*}[t]
\centering
\caption{Runtimes of exact 5-nearest-neighbors search. Brute force vs Ball tree index. Boldface indicates faster runtime between brute force and index; or better precision. The experimental details are in Appendix~\ref{ap1}. The Gaussians dataset is as used in Figure \ref{fig_std-mean}; $w$-Gaussians (where $w=5,000$) is as shown in Table \ref{fig:t-SNE}. The last two columns show the retrieval precision of 5 nearest neighbours. Every point in a dataset is used as a query; and the reported result is an average over all queries in the dataset.}
\begin{tabular}{|c|rrr|rr|rr|r|rr||rr|}
\hline
Dataset & $\#points$ & $\#dimen.$ & \#class & \multicolumn{2}{c|}{Distance} & \multicolumn{3}{c|}{IK} & \multicolumn{2}{c||}{LK} & \multicolumn{2}{c|}{Precision@5} \\\cline{5-13}
& & & & Brute & Balltr & Brute & Balltr & Map & Brute & Balltr & Distance & IK\\
\hline
Url & 3,000 & 3,231,961 & 2 & \textbf{38,702} & 41,508 & 112 & \textbf{109} & 9 & \textbf{39,220} & 43,620 & .88 & \textbf{.95}\\
News20 & 10,000 & 1,355,191 & 2
& \textbf{101,243} & 113,445 & 3,475 & \textbf{2,257} & 16 & \textbf{84,986} & 93,173 & .63 & \textbf{.88}\\
Rcv1 & 10,000 & 47,236 & 2 & \textbf{3,615} & 4,037 & 739 & \textbf{578} & 11 & \textbf{5,699} & 6,272 & .90 & \textbf{.94}\\
Real-sim & 10,000 & 20,958 & 2 & \textbf{2,541} & 2,863 & 4,824 & \textbf{4,558} & 21 & \textbf{2,533} & 2853 & .50 & \textbf{.63}\\
Gaussians & 2,000 & 10,000 & 2 & 46 & \textbf{41} & 229 & \textbf{213} & 46 & \textbf{45} & 54 & \textbf{1.00} & \textbf{1.00} \\
$w$-Gaussians & 2,000 & 10,000 & 2 & \textbf{53} & 77 & 210 & \textbf{205} & 45 & \textbf{56} & 73 & .90 & \textbf{1.00} \\
Cifar-10 & 10,000 & 3,072 & 10 & \textbf{340} & 398 & 7,538 & \textbf{7,169} & 67 & \textbf{367} & 439 & .25 & \textbf{.27}\\
Mnist & 10,000 & 780 & 10 & \textbf{58} & 72 & 1,742 & \textbf{1,731} & 2 & \textbf{87} & 106 & \textbf{.93} & \textbf{.93} \\
A9a & 10,000 & 122 & 2 & \textbf{10} & 13 & 5,707 & \textbf{5,549} & 1 & \textbf{12} & 16 & \textbf{.79} & \textbf{.79} \\
Ijcnn1 & 10,000 & 22 & 2 & 2.3 & \textbf{1.3} & 706 & \textbf{654} & 1 & 2.2 & \textbf{1.8} & \textbf{.97} & .96\\
\hline
\multicolumn{11}{|r}{\textit{Average}} & .77 & \textbf{.84} \\
\hline
\end{tabular}%
\label{tbl_indexing}%
\end{table*}%
\subsubsection{Exact nearest neighbor search using ball tree indexing}
\label{sec_ballTree}
Existing indexing techniques are sensitive to dimensionality, data size and structure of the dataset \citep{NN-indexibility-TODS2006,ExactNNSearch-PKDD2007}. They only work in low dimensional datasets of a moderate size where a dataset has clusters.
Table \ref{tbl_indexing} shows the comparison between the brute force search and an indexed search based on ball tree \citep{omohundro1989five}, in terms of runtime, using Euclidean distance, IK and linear kernel (LK).
Using LK, the ball tree index is always worse than the brute force search in all datasets, except the lowest dimensional dataset. Using Euclidean distance, we have a surprising result that the ball tree index ran faster than the brute force in one high dimensional dataset. However, this result cannot be `explained away' by using the argument that they may have low intrinsic dimensions (IDs) because the Gaussians dataset has 10,000 (theoretically true) IDs; while the indexed search ran slower in all other real-world high dimensional datasets where each has the median ID less than 25, as estimated using a recent local ID estimator \citep{LID-SIAM19} (see the full result in Appendix \ref{sec_ID}.)
Yet, with an appropriate value of $\psi$, IK yielded a faster search using the ball tree index than the brute force search in all datasets, without exception. Note that the comparison using IK is independent of the feature mapping time of IK because both the brute force and the ball tree indexing require the same feature mapping time (shown in the `Map' column in Table \ref{tbl_indexing}.)
It is interesting to note that the feature map of IK uses an effective number of dimensions of $t=200$ in all experiments. While one may tend to view this as a dimension reduction method, it is a side effect of the kernel, not by design. The key to IK's success in enabling efficient indexing is the kernel characteristic in high dimensions (in Hilbert space.) IK is not designed for dimension reduction.
Note that IK works not because it has used a low-dimensional feature map. Our result in Table \ref{tbl_indexing} is consistent with the previous evaluation (up to 80 dimensions only \citep{ExactNNSearch-PKDD2007}), i.e., indexing methods ran slower than brute-force when the number of dimensions is more than 16. If the IK's result is due to low dimensional feature map, it must be about 20 dimensions or less for IK to work. But all IK results in Table \ref{tbl_indexing} have used $t=200$.
Nevertheless, the use of $t=200$ effective dimensions has an impact on the absolute runtime. Notice that, when using the brute force method, the actual runtime of IK was one to two orders of magnitude shorter than that of distance in $d>40,000$; and the reverse is true in $d<40,000$.
We thus recommend using IK in high dimensional ($d>40,000$) datasets, not only because its indexing runs faster than the brute force, but IK runs significantly faster than distance too. In $d\le 40,000$ dimensional datasets, distance is preferred over IK because the former runs faster.
The last two columns in Table \ref{tbl_indexing} show the retrieval result in terms of precision of 5 nearest neighbors. It is interesting to note that IK has better retrieval outcomes than distance in all cases, where large differences can be found in News20, Realsim and $w$-Gaussians. The only exception is Ijcnn1 which has the lowest number of dimensions, and the difference in precision is small.
This shows that the concentration effect influences not only the indexing runtime but also the precision of retrieval outcomes.
\subsubsection{Clustering}
\label{sec_SC}
In this section, we examine the effect of IK versus GK/distance using a scalable spectral clustering algorithm (SC) \citep{chen2010parallel} and Density-Peak clustering (DP) \citep{DP2014clustering}.
We report the average clustering performance in terms of AMI (Adjusted Mutual Information) \citep{vinh2010information} of SC over 10 trials on each dataset because SC is a randomized algorithm; but only a single trial is required for DP, which is a deterministic algorithm.
\begin{table}[h]
\centering
\caption{Best AMI scores of Desity-Peak clustering (DP) and Spectral Clustering (SC) using Distance/Gaussian kernel (GK), Adaptive Gaussian kernel (AG) and IK. Note that every cluster in Cifar10 and Ijcnn1 has mixed classes. That is why all clustering algorithms have very low AMI.}
\begin{tabular}{|c|ccc|ccc|}
\hline
Dataset & \multicolumn{3}{c|}{DP} & \multicolumn{3}{c|}{SC} \\
\cline{2-7}
& Distance & AG & IK & GK & AG & IK \\
\hline
Url & .07 & \textbf{.19} & .16 & .04 & .04 & \textbf{.07}\\
News20 & .02 & .05 & \textbf{.27} & .16 & .15 & \textbf{.21} \\
Rcv1 & .18 & .19 & \textbf{.32} & \textbf{.19} & \textbf{.19} & \textbf{.19} \\
Real-sim & .02 & .03 & \textbf{.04} & .04 & .02 & \textbf{.05} \\
Gaussians & \textbf{1.00} & \textbf{1.00} & \textbf{1.00} & \textbf{1.00} & \textbf{1.00} & \textbf{1.00} \\
$w$-Gaussians & .01 & .01 & \textbf{.68} & .24 & .25 & \textbf{1.00} \\
Cifar-10 & .08 & .09 & \textbf{.10} & \textbf{.11} & \textbf{.11} & .10 \\
Mnist & .42 & \textbf{.73} & .69 & .49 & .50 & \textbf{.69} \\
A9a & .08 & .17 & \textbf{.18} & .15 & .14 & \textbf{.17} \\
Ijcnn1 & .00 & .00 & \textbf{.07} & \textbf{.00} & \textbf{.00} & \textbf{.00} \\
\hline
\textit{Average} & .19 & .25 & \textbf{.35} & .24 & .24 & \textbf{.35}\\
\hline
\end{tabular}%
\label{tbl_SC2}%
\end{table}%
\begin{table*}[t]
\centering
\caption{SVM classification accuracy \& runtime (CPU seconds). Gaussian kernel (GK), IK and Linear kernel (LK). $nnz\% = \#nonzero\_values / ((\#train+\#test) \times \#dimensions) \times 100$.
IK ran five trials to produce [mean]{\tiny $\pm$}[standard error]. The runtime of SVM with IK includes the IK mapping time. The separate runtimes of IK mapping and SVM are shown in Appendix \ref{sec_IK_mapping_time}.
\label{tbl:SVM-accuracy} }
\begin{tabular}{|c|rrrr|rrr||rrr|}\hline
Dataset & $\#train$ & $\#test$ & $\#dimen.$ & $nnz\%$ & \multicolumn{3}{c||}{Accuracy} & \multicolumn{3}{c|}{Runtime}\\
& & & & & LK & \multicolumn{1}{c}{IK} & GK & LK & IK & GK\\
\hline
Url & 2,000 & 1,000 & 3,231,961 & .0036 & \textbf{.98} & \textbf{.98}{\tiny $\pm$.001} & \textbf{.98} & 2 & 3 & 14 \\
News20 & 15,997 & 3,999 & 1,355,191 & .03
& .85 & \textbf{.92}{\tiny $\pm$.007} & .84 & 38 & 60 & 528 \\
Rcv1 & 20,242 & 677,399 & 47,236 & .16
& .96 & .96{\tiny $\pm$.013} & \textbf{.97} & 111 & 420 & 673\\
Real-sim & 57,848 & 14,461 & 20,958 & .24
& \textbf{.98} & \textbf{.98}{\tiny $\pm$.010} & \textbf{.98} & 49 & 57 & 2114\\
Gaussians & 1,000 & 1,000 & 10,000 & 100.0 & \textbf{1.00} & \textbf{1.00}{\tiny $\pm$.000} & \textbf{1.00} & 14 & 9 & 78\\
$w$-Gaussians & 1,000 & 1,000 & 10,000 & 100.0 & .49 & \textbf{1.00}{\tiny $\pm$.000} & .62 & 20 & 8 & 79 \\
Cifar-10 & 50,000 & 10,000 & 3,072 & 99.8
& .37 & \textbf{.56}{\tiny $\pm$.022} & .54 & 3,808 & 1087 & 29,322\\
Mnist & 60,000 & 10,000 & 780 & 19.3
& .92 & .96{\tiny $\pm$.006} & \textbf{.98} & 122 & 48 & 598\\
A9a & 32,561 & 16,281 & 123 & 11.3
& \textbf{.85} & \textbf{.85}{\tiny $\pm$.012} & \textbf{.85} & 1 & 31 & 100\\
Ijcnn1 & 49,990 & 91,701 & 22 & 59.1
& .92 & .96{\tiny $\pm$.006} & \textbf{.98} & 5 & 66 & 95\\
\hline
\end{tabular}
\end{table*}
Table \ref{tbl_SC2} shows the result.
With SC, IK improves over GK or AG \citep{zelnik2005self} in almost all datasets. Large improvements are found on $w$-Gaussians and Mnist. The only dataset in which distance and GK did well is the Gaussians dataset that has two well-separate clusters.
AG was previously proposed to `correct' the bias of GK \citep{zelnik2005self}. However, our result shows that AG is not better than GK in almost all datasets we have used. There are two reasons. First, we found that previous evaluations have searched for a small range of $\sigma$ of GK (e.g., \citep{chen2010parallel}.) Especially in high dimensional datasets, $\sigma$ shall be searched in a wide range (see Table \ref{tbl_SC_parameters} in Appendix \ref{ap1}.) Second, spectral clustering is weak in some data distribution even in low dimensional datasets. In this case, neither GK nor AG makes a difference. This is shown in the following example using $w$-Gaussians, where $w=1$, which contains two 1-dimensional Gaussian clusters which only overlap at the origin in the 2-dimensional space.
Figure \ref{fig_SC} shows the best spectral clustering results with IK and AG.
When the two Gaussians in the $w$-Gaussians dataset have the same variance, spectral clustering with AG (or GK) fails to separate the two clusters; but spectral clustering with IK succeeds.
\begin{figure}[h]
\centering
\begin{subfigure}[b]{0.22\textwidth}
\centering
\includegraphics[width=1.4in]{IK.png}
\caption{SC with IK (AMI=0.91)}
\end{subfigure}
\begin{subfigure}[b]{0.22\textwidth}
\centering
\includegraphics[width=1.4in]{AG2.png}
\caption{SC with AG (AMI=0.29)}
\end{subfigure}
\caption{The spectral clustering results with IK and AG on the $w$-Gaussians dataset (where $w=1$) that contains two 1-dimensional subspace clusters. Each cluster has 500 points, sampled from a $1$-dimensional Gaussian distribution $\mathcal{N}(0, 1)$. The density distribution of each cluster is shown in (a) wrt the corresponding 1-dimensional subspace. }
\label{fig_SC}
\end{figure}
When the variances of the two subspace clusters are substantially different, spectral clustering with AG yields a comparable good result to that with IK. Note that spectral clustering with GK cannot separate the two subspace clusters regardless. This is the specific condition, as used previously \citep{zelnik2005self} to show that AG is better than GK in spectral clustering.
In a nutshell, in general, IK performs better than GK and AG on spectral clustering in both low and high dimensional datasets.
This clustering result is consistent with the previous result \citep{IsolationKernel-AAAI2019} conducted using DBSCAN \citep{ester1996density} in both low and high dimensional datasets comparing between distance and IK.
Similar relative results with DP \cite{DP2014clustering} are shown in Table \ref{fig_SC}, where large differences in AMI in favor of IK over distance/AG can be found in News20 and $w$-Gaussians.
\begin{table*}[t]
\caption{t-SNE: Gaussian Kernel versus Isolation Kernel on the $w$-Gaussians datasets having two $w$-dimensional subspace clusters (shown in the first column.) Both clusters are generated using $\mathcal{N}(0, 1)$. ``$\times$'' is the origin---the only place in which the two clusters overlap in the data space. }
\label{fig:t-SNE}
\centering
\begin{tabular}{c|ccccc}
\hline
$w=1$ dataset & & t-SNE(GK) & t-SNE(IK) & FIt-SNE & q-SNE\\
\hline
\includegraphics[width=1.2in]{w1.png} & \begin{turn}{90} \quad $w=2$ \end{turn}&
\includegraphics[width=1.2in]{GK250.png} &
\includegraphics[width=1.2in]{2IK5.png} &
\includegraphics[width=1.2in]{FitSNE1.png} &
\includegraphics[width=1.2in]{qsne1.png}\\
& \begin{turn}{90} \quad $w=5000$ \end{turn}&
\includegraphics[width=1.2in]{GK5k.png} &
\includegraphics[width=1.2in]{IK5k.png} &
\includegraphics[width=1.2in]{f5k.png} &
\includegraphics[width=1.2in]{q5k.png} \\
\hline
\end{tabular}
\end{table*}
\begin{table*} [t]
\caption{t-SNE: GK versus IK on $w$-Gaussians datasets having two $w$-dimensional subspace clusters (different variances.) The red and blue clusters are generated using $\mathcal{N}(0, 1)$ and $\mathcal{N}(0, 32)$, respectively.
``$\times$'' is the origin---the only place in which the two clusters overlap in the data space. }
\label{fig:t-SNE-different-variance}
\centering
\begin{tabular}{c|ccccc}
\hline
$w=1$ dataset & & t-SNE(GK) & t-SNE(IK) & FIt-SNE & q-SNE\\
\hline
\includegraphics[width=1.2in]{w2.png} & \begin{turn}{90} \quad $w=2$ \end{turn}&
\includegraphics[width=1.2in]{GK_2_500_552.png} &
\includegraphics[width=1.2in]{IK_2_5_451.png} &
\includegraphics[width=1.2in]{f2.png} &
\includegraphics[width=1.2in]{q2.png} \\
& \begin{turn}{90} \quad $w=100$ \end{turn}&
\includegraphics[width=1.2in]{GK_100_500_004.png} &
\includegraphics[width=1.2in]{750.png} &
\includegraphics[width=1.2in]{f100.png} &
\includegraphics[width=1.2in]{q100.png} \\
\hline
\end{tabular}
\end{table*}
\subsubsection{SVM classification}
\label{sec_SVM-classification}
Table \ref{tbl:SVM-accuracy} shows the comparison between linear kernel, IK and Gaussian kernel using an SVM classifier in terms of accuracy and runtime. The experimental details are in Appendix~\ref{ap1}.
It is interesting to note that IK produced consistently high (or the highest) accuracy in all datasets. In contrast, both linear and Gaussian kernels have a mixed bag of low and high accuracies. For example, they both perform substantially poorer than IK on News20 and $w$-Gaussians (where $w=5,000$.) Each $w$-Gaussians dataset has two $w$-dimensional subspace clusters. The $w=1$ version of the dataset is shown in the first column in Table \ref{fig:t-SNE}.
The runtime result shows that, SVM using IK runs either comparably or faster than SVM using Linear Kernel, i.e., they are in the same order, especially in high dimensions. Because of employing LIBLINEAR \citep{fan2008liblinear}, both are up to two orders of magnitude faster in large datasets than SVM with GK which must employ the slower nonlinear LIBSVM \citep{CC01a}.
\subsubsection{Visualization using t-SNE}
\label{sec_visualization_experiment}
Studies in visualization \citep{MDS-wickelmaier2003introduction,Hinton2008Visualizing} often ignore the curse of dimensionality issue which raises doubt about the assumption made by visualization methods. For example, t-SNE \citep{Hinton2008Visualizing} employs Gaussian kernel as a means to measure similarity in the high dimensional data space. No studies have examined the effect of its use in the presence of the curse of dimensionality, as far as we know.
Here we show one example misrepresentation from t-SNE, due to the use of GK to measure similarity in the high dimensional data space on the $w$-Gaussians datasets.
The t-SNE visualization results comparing Gaussian kernel and Isolation Kernel are shown in the middle two columns in Table \ref{fig:t-SNE}. While t-SNE using GK has preserved the structure in low dimensional ($w=2$) data space, it fails completely to separate the two clusters in high dimensional ($w=5,000$) data space.
In contrast, t-SNE with IK correctly separates the two clusters in both low and high dimensional data spaces.
Recent improvements on t-SNE, e.g., FIt-SNE \citep{linderman2019fast} and q-SNE \citep{hakkinen2020qsne}, explore better local structure resolution or efficient approximation. As they employ Gaussian kernel, they also produce misrepresented structures, as shown in Table \ref{fig:t-SNE}.
Table \ref{fig:t-SNE-different-variance} shows the results on a variant $w$-Gaussians dataset where the two clusters have different variances. Here t-SNE using GK produces misrepresented structure in both low and high dimensions. It has misrepresented the origin in the data space as belonging to the red cluster only, i.e., no connection between the two clusters. Note that all points in the red cluster are concentrated at the origin in $w=100$ for GK. There is no such issue with IK.
In low dimensional data space, this misrepresentation is due solely to the use of a data independent kernel. To make GK adaptive to local density in t-SNE, the bandwidth is learned locally for each point. As a result, the only overlap between the two subspace clusters, i.e., the origin, is assigned to the dense cluster only.
The advantage of IK in low dimensional data space is due to the data dependency stated in Lemma \ref{lem_characteristic}. This advantage in various tasks has been studied previously \citep{ting2018IsolationKernel,IsolationKernel-AAAI2019,IsolationSetKernel,IDK-KDD2020}.
In high dimensional data space, the effect due to the curse of dimensionality obscures the similarity measurements made. Both this effect and the data independent effect collude in high dimensional data space when distance-based measures are used, leading to poor result in the transformed space.
\vspace{-1mm}
\subsubsection{Further investigation using the low dimensional $w$-Gaussians ($w=2$)}
It is interesting to note that (i) in SVM classification, both GK and IK could yield 100\% accuracy while LK still performed poorly at 51\% accuracy; and (ii) all three measures enable faster indexed search than brute force search and have equally high precision on this low dimensional dataset. But only IK can do well in the high dimensional $w$-Gaussians ($w=5,000$) in all four tasks, as shown in Tables \ref{tbl_indexing} to \ref{fig:t-SNE-different-variance}.
The $w$-Gaussians datasets provide an example condition in which GK/distance performed well in low dimensions but poorly in high dimensions in indexed search, SVM classification and t-SNE visualization.
This is a manifestation that the GK/distance used in these algorithms has indistinguishability in high dimensions.
\noindent
\textbf{Summary}:
\begin{itemize}[itemsep=0ex, leftmargin=5mm]
\item IK is the only measure that consistently (a) enabled faster indexed search than brute force search, (b) provided good clustering outcomes with SC and DP, (c) produced high accuracy in SVM classification, and (d) yielded matching structures in t-SNE visualization in both low and high dimensional datasets.
\item Euclidean distance or GK performed poorly in most high dimensional datasets in all tasks except SVM classification. We show one condition, using the artificial $w$-Gaussians datasets, in which they did well in low dimensions but performed poorly in high dimensions in all four tasks. The Gaussians dataset is one condition in which Euclidean distance or GK can do well in high dimensions.
\item Though low intrinsic dimensions may explain why GK/LK can do well in some high dimensional datasets in SVM classification, the explanation failed completely for all the slower indexed searches in high dimensions, and for the only faster indexed search on the Gaussians dataset as it has high intrinsic dimensions.
\end{itemize}
\section{Discussion}
\label{sec_discussion}
We have limited our discussion on the curse of dimensionality based on nearest neighbors, i.e., their (in)distinguishability. But the effect of the curse is broader than this. For example, it has effects on the consistency of estimators, an aspect not discussed in this article.
In addition to the concentration effect \citep{Talagrand,Beyer:1999}, the curse of dimensionality has also been studied under different phenomena, e.g., the correlation between attributes \citep{WhenNNisMeaningful-2009} and hubness \citep{HubsInSpace-JMLR-2010}.
It is interesting to analyze whether IK can deal with these issues, other than the concentration effect, effectively in high dimensions.
Our results in Tables \ref{tbl_indexing} to \ref{tbl:SVM-accuracy} show that some existing measures may perform well in high dimensions under special conditions. However, we are not aware of analyses that examine whether any existing measures can break the curse of dimensionality.
Previous work on intrinsic dimensionality (e.g., \cite{LID-NIPS2011}) has shown that nonparametric distance-based regressors could escape the curse of dimensionality if a high dimensional dataset has low intrinsic dimensions. This is consistent with the finding that the concentration effect depends more on the intrinsic dimensions $I_d$ than on the input dimensions $d > I_d$ \citep{Concentration-Fractionaldistances}. Our result in Table \ref{tbl_indexing} on the Gaussians dataset adds another element that needs an explanation for existing measures, i.e., why does distance perform well on datasets with high intrinsic dimensions?
Because our analysis on IK's distinguishability is independent of dimensionality and data distribution, IK does not rely on low intrinsic dimensions to perform well, as shown in the artificial datasets: both Gaussians and $w$-Gaussians have high ($\ge 5,000$ true) intrinsic dimensions in Tables \ref{tbl_indexing} to \ref{fig:t-SNE-different-variance}.
Our applications in ball tree indexing, two clustering algorithms, SVM classification and t-SNE here, and previously in DBSCAN clustering \citep{IsolationKernel-AAAI2019} and multi-instance learning \cite{IsolationSetKernel}, suggest that many existing algorithms can get performance improvement by simply replacing distance/kernel with IK. However, a caveat is in order: not all existing algorithms can achieve that. For example, OCSVM \citep{OCSVM2001} and OCSMM \citep{OCSMM2013} have been shown to work poorly with IK \citep{ting2020-IDK-GroupAnomalyDetection}. This is because the learning in these two algorithms is designed to have a certain geometry in Hilbert space in mind; and IK does not have such a geometry (see details in \cite{ting2020-IDK-GroupAnomalyDetection}.)
The isolating mechanism used has a direct influence on IK's distinguishability.
Note that the
proofs of Lemma \ref{lem2} and Theorem \ref{thm2} rely on the isolating partitions being the Voronoi diagram. It remains an open question whether IK with a different implementation could be proven to have distinguishability.
Kernel functional approximation is an influential approach to obtain an approximate finite-dimensional feature map from an infinite-dimensional feature map of a kernel such as Gaussian kernel. Its representative methods are Nystr\"{o}m \citep{Nystrom_NIPS2000} and random features \citep{RandomFeatures2007,OrthonogalRandomFeatures2016}. The former is often dubbed data dependent and the latter data independent. These terms refer to the use (or no use) of data samples to derive the approximate feature map. In either case, the kernel employed is data independent and translation invariant, unlike the data dependent IK in which the similarity depends local data distribution and it is not translation invariant.
VP-SVM \citep{VP-SVM-2016} employs Voronoi diagram to split the input space and then learns a local SVM with a local bandwidth of Gaussian kernel in each Voronoi cell. It aims to reduce the training time of SVM only, and do not address the curse of dimensionality at all.
A recent work \citep{IK-SNE-2021} has already shown that IK improves the effectiveness and efficiency of t-SNE, by simply replacing Gaussian kernel. However, it did not address the issue of curse of dimensionality, which is the focus of this paper.
The result of an additional investigation on the hubness effect \cite{HubsInSpace-JMLR-2010} can be found in Appendix \ref{App_hubness}.
\section{Conclusions}
We show for the first time that the curse of dimensionality can be broken using Isolation Kernel (IK). It is possible because (a) IK measures the similarity between two points based on the space partitionings produced from an isolation mechanism;
(b) the probability of a point falling into a partition in the isolation partitioning of Voronoi diagram is independent of data distribution, the distance measure used to create the Voronoi partitions and the data dimensions, implied in Lemma \ref{lem2}; and (c) IK's unique feature map has its dimensionality linked to a concatenation of these partitionings.
Theorem~\ref{thm2} suggests that increasing the number of partitionings $t$ (i.e., the dimensionality of the feature map) leads to increased distinguishability, independent of the number of dimensions and distribution in data space.
Isolation Kernel, with its feature map having the distinguishability stated in Theorem \ref{thm2}, is the key to consistently producing (i) faster indexed search than brute force search, and high retrieval precision, (ii) good clustering outcomes with SC and DP, (iii) high accuracy of SVM classification, and (iv) matching structures in t-SNE visualization in both low and high dimensional data spaces.
Euclidean distance, Gaussian and linear kernels have a mixed bag of poor and good results in high dimensions in our experiments. This is not a surprising outcome of the curse of dimensionality, echoing the current state of understanding of these measures, with or without the aid of intrinsic dimensions.
\newpage
|
\section{Introduction}
\lettrine{T}{he} increase of congestion and pollution levels in large metropolitan areas has recently incentivized the development of sustainable and green mobility solutions.
Urban Air Mobility (UAM) has the potential to address these problems by allowing air transport of goods and people \cite{easa}, thus reducing the pressure on urban traffic. \mcedit{It has been estimated that 160,000 air taxis could be in circulation worldwide by 2050, representing a USD 90 billion market \cite{roland-berger}.}
Tiltwing VTOL aircraft, with their capability to take-off and land in restricted spaces and their extended endurance, have recently received a lot of attention in a UAM context. Despite being \mcedit{investigated}
since the \mcedit{1950's} \cite{kuhn1959semiempirical, hargraves1961analytical, mccormick1967aerodynamics}, issues related to control and stability, as well as the mechanical complexity associated with the use of conventional internal combustion engines have \mcedit{prevented widespread adoption of this technology}. However, recent advances in battery and electric motor technology, combined with the development of modern control systems architectures, has spurred a new interest for these vehicles \cite{UMich19}. \mcedit{Some recent} prototypes based on tiltwing configurations \mcedit{are}
the Airbus A$^3$ Vahana, the NASA GL-10 \cite{rothhaar2014nasa}, or Rolls-Royce's eVTOL aircraft \cite{higgins2020aeroacoustic}.
The transition manoeuvre for a VTOL aircraft is the most critical phase of flight. Although heuristic strategies \mcedit{have been}
proposed based on smooth scheduling functions of the forward velocity or tilt angle \cite{me}, trajectory optimisation for the transition of VTOL aircraft is still an open problem as it involves determining the optimal combination of thrust and tiltwing angle to minimise an objective while meeting \mcedit{constraints on system states and control inputs},
which requires solving large-scale nonlinear optimisation problems.
The problem \mcedit{of determining minimum energy speed profiles}
for the forward transition manoeuvre of the Airbus A$^3$ Vahana \mcedit{was addressed in \cite{Iowa19}}, considering various phases of flight (cruise, transition, descent). A drawback of the approach is that the transition was assumed to occur instantly, which is \mcedit{unrealistic}
for such \mcedit{a} vehicle.
In \cite{UMich19}, the trajectory generation problem for take-off is formulated as a constrained optimisation problem \mcedit{and} solved using NASA's OpenMDAO framework and the SNOPT gradient-based optimiser. A case study based on the Airbus A$^3$ Vahana tiltwing aircraft is considered, including aerodynamic models of the wing beyond stall, and the effect of propeller-wing interaction.
Forward and backward optimal transition manoeuvres at constant altitude are computed in \cite{leo} for a tiltwing aircraft, considering leading-edge fluid injection active flow control and the use of a high-drag device.
All \mcedit{of these approaches consider optimising trajectories}
using general purpose NLP solvers without exploiting potentially
\mcedit{useful}
structures and simplifications.
\mcedit{This approach is}
computationally expensive \mcedit{and is not generally suitable for real-time implementation}.
We propose in this paper a \mcedit{convexification of the optimal trajectory generation}
problem based on a change of variables \mcedit{inspired by \cite{bobrow}}.
This translates the nonlinear equations of motion in \mcedit{the} time domain into \mcedit{a set of linear differential equations in the space}
domain along a prescribed path.
\mcedit{We thus derive}
an efficient method based on \mcedit{convex programming}. \mdsedit{Convex programming was successfully employed to solve optimisation problems related to energy management of hybrid-electric aircraft \cite{me2, me3} and spacecraft \cite{me4}.}
The paper is organized as follows. Section \ref{sec:modeling} introduces \mcedit{the} tiltwing VTOL aircraft dynamic and aerodynamic models. \mcedit{Section \ref{sub:convex_formulation} formulates the continuous time trajectory optimisation problem in terms of a pair of}
convex problems. \mcedit{These}
are discretised in Section \ref{sec:discrete} and Section \ref{sec:results} discusses simulation results obtained for a case study based on the Airbus A$^3$ Vahana.
\mcedit{Section~\ref{sec:conclusion} presents conclusions.}
\section{Modeling}
\label{sec:modeling}
A simplified longitudinal point-mass model of a tiltwing VTOL aircraft equipped with propellers is developed in this section. In order to account for the effect of the propeller wake on the wing, the flow velocity downstream is augmented by the induced velocity of the propeller.
The second-order dynamics of the tilting wing is also presented.
\mcedit{We}
consider a planar point-mass model of a VTOL aircraft as shown in Figure \ref{fig:diagram}, whose position with respect to an inertial frame $O_{XZ}$ is given by $(x,z)$ and velocity given by
\[
\dot{x} = V \cos{\gamma},
\qquad
\dot{z} = -V \sin{\gamma},
\]
where $V$ is the aircraft velocity magnitude and $\gamma$ the \mdsedit{flight path angle, defined as the angle of the velocity vector from horizontal}.
From \mcedit{Figure~\ref{fig:diagram}}
the point-mass equations of motion (EOM) in polar coordinates are
\begin{alignat}{2}
m \dot{V} &= T \cos{\alpha} - D -mg \sin{\gamma},
&\quad
V(t_0) &= V_{0},
\label{eq:eom1}
\\
m V \dot{\gamma} &= T \sin{\alpha} + L -mg \cos{\gamma},
&\quad
\gamma(t_0) &= \gamma_{0},
\label{eq:eom2}
\end{alignat}
where $m$ is the mass of the \mcedit{aircraft},
$g$ the acceleration due to gravity, $T$ the thrust magnitude, $\alpha$ the angle of attack.
The dynamics of the wing are given by
\begin{equation}
J_w \ddot{i}_w = M, \quad i_w(t_0) = i_{0}, \quad \dot{i}_w(t_0) = \Omega_{0},
\label{eq:tiltwing_dyna}
\end{equation}
where $J_w$ is the rotational inertia of the wing (\mcedit{about}
the $y$-axis), $M$ is the total torque delivered by the tilting actuators and $i_w$ is the tiltwing angle such that
\[
i_w + \theta = \alpha + \gamma .
\]
\mcedit{Here}
$\theta$ is the pitch angle, defined as the angle
\mcedit{of the fuselage axis from horizontal}.
\mcedit{For passenger comfort}, $\theta$ is regulated via the elevator to track a constant reference $\theta^* = 0$.
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth, trim={1.5cm 3.5cm 1.5cm 2cm}, clip]{img/diagram.pdf}
\caption{\mcedit{Force and velocity definitions for a VTOL aircraft}}
\label{fig:diagram}
\end{figure}
From momentum theory, the propeller generates an induced speed $v_i$ such that
\[
T = \rho A n (V \cos{\alpha} + v_i )( k_w v_i),
\]
\begin{comment}
\[
v_i = -\frac{V \cos{\alpha}}{2} + \sqrt{\frac{V^2 \cos^2{\alpha}}{4} + \frac{T}{2\rho A n}},
\]
\end{comment}
where $\rho$ is the air density, $A$ the rotor disk area, $n$ the number of \mcedit{propellers, and}
$k_w \approx 2$. The effective velocity $V_e$ and effective angle of attack $\alpha_e$ seen by the wing are given by
\begin{align*}
V_e \cos{\alpha_e} &= V \cos{\alpha} + k_w v_i ,
\\
V_e \sin{\alpha_e} &= V \sin{\alpha},
\end{align*}
\[
V_e^2 = V^2 + \frac{2T}{\rho A n}.
\]
The total lift and drag are modeled as the weighted sum of the blown and unblown counterparts
\begin{align*}
D &= (1-\mu) \tfrac{1}{2} \rho S(a_2 \alpha^2 + a_1 \alpha + a_0) V^2 \\
&\quad + \mu \tfrac{1}{2} \rho S(a_2 \alpha_e^2 +a_1 \alpha_e + a_0) V_e^2
\\
&\approx (1-\mu) \tfrac{1}{2} \rho S(a_1 \alpha + a_0) V^2 + \mu \tfrac{1}{2} \rho S(a_1 \alpha_e + a_0) V_e^2,
\\
L &= (1-\mu) \tfrac{1}{2} \rho S(b_1 \alpha + b_0) V^2 + \mu \tfrac{1}{2} \rho S(b_1 \alpha_e + b_0) V_e^2.
\end{align*}
\mcedit{Here}
$S$ is the wing area,
$\mu = 2 R n / l$ is the \mdsedit{blown-unblown ratio, $R$ is the radius of the rotor disk, $l$ is the wingspan},
\mcedit{$a_0, a_1, a_2$ and $b_0, b_1$ are constant parameters and
$\alpha, \alpha_e\ll 1$ is assumed so that terms involving $\alpha^2,\alpha_e^2$ are negligible}.
The following input and state constraints also apply
\begin{gather*}
0 \leq T \leq \overline{T}, \quad 0 \leq V \leq \overline{V},
\nonumber
\\
V(t_0) = V_0
\text{ and }
V(t_f) = V_f,
\nonumber
\\
\underline{a} \leq \dot{V} \leq \overline{a}, \quad \underline{M} \leq M \leq \overline{M}.
\nonumber
\end{gather*}
\mcedit{This paper considers how to optimise}
the transition between powered lift and cruise flight modes.
Minimising thrust is a natural choice as a proxy for minimising energy consumption,
\mcedit{suggesting the following objective function}
\begin{equation}
\label{eq:obj}
J = \int_{t_0}^{t_f} \left(T/\overline{T}\right)^2 \mathrm{d}t.
\end{equation}
The optimisation problem consists of minimising \eqref{eq:obj} while satisfying dynamical constraints, \mcedit{input and state constraints}.
\section{Convex formulation}
\label{sub:convex_formulation}
The \mcedit{aircraft dynamics}
derived in Section \ref{sec:modeling} contain
nonlinearities \mcedit{that make the}
trajectory optimisation problem \mcedit{nonconvex}.
However,
a change of \mcedit{variables (motivated by~\cite{bobrow})}
considerably \mcedit{simplifies} the structure of \mcedit{this problem}
and allows us to formulate the problem \mcedit{in terms of}
two convex programs. If we assume that a path $(x(s), z(s))$ parameterised by the curvilinear abscissa $s$ is already known \textit{a priori} (which \mcedit{is usually}
the case in a UAM context where flight corridors are prescribed) we can introduce \mcedit{via}
the chain rule the following change of differential operator,
\begin{equation}
\label{eq:CV}
\frac{\mathrm{d}}{\mathrm{d}t} = V \frac{\mathrm{d}}{\mathrm{d}s} .
\end{equation}
\mcedit{Denoting $\tfrac{\mathrm{d}\,\cdot}{\mathrm{d}s}=\cdot'$, the EOM in \eqref{eq:eom1} and \eqref{eq:eom2} are equivalent to}
\begin{align*}
m V V' &= T \cos{\alpha} - (1-\mu)\frac{1}{2}\rho S (a_1 \alpha + a_0 ) V^2
\\
&\quad -\mu\frac{1}{2}\rho S (a_1 \alpha_e + a_0) \Bigl(V^2 + \frac{2T}{\rho A n}\Bigr) -mg \sin{\gamma^*},
\\
m V^2 \gamma^{*\prime} &= T \sin{\alpha} + (1-\mu)\frac{1}{2}\rho S (b_1 \alpha + b_0) V^2
\\
&\quad +\mu\frac{1}{2}\rho S (b_1 \alpha_e + b_0 ) \Bigl(V^2 + \frac{2T}{\rho A n}\Bigr) -mg \cos{\gamma^*},
\end{align*}
where $\gamma^*$ is known \textit{a priori} from the path.
\mcedit{Specifically},
since $\mathrm{d}x= V \cos{\gamma} \, \mathrm{d}t$ \mcedit{and} $\mathrm{d}z=-V \sin{\gamma} \, \mathrm{d}t$, the flight path angle can be expressed in terms of the path variables \mcedit{as}
\[
\tan{\gamma^*} = -\frac{\mathrm{d}z}{\mathrm{d}x} .
\]
\mcedit{Defining new model states}
$E=V^2$ and $a=VV'$, \mcedit{so that}
\begin{equation}
\label{eq:diff}
E' = 2 a,
\end{equation}
the EOM reduce to
\begin{align}
\label{eq:eom7}
& m a = T \cos{\alpha} - (1-\mu)\frac{1}{2}\rho S\left(a_1 \alpha + a_0\right) E
\nonumber\\
&\quad- \mu\frac{1}{2}\rho S\left(a_1 \alpha_e + a_0\right)\left(E + \frac{2T}{\rho A n}\right) -mg \sin{\gamma^*},
\\
\label{eq:eom8}
& m E \gamma^{*\prime} = T \sin{\alpha} + (1-\mu)\frac{1}{2}\rho S\left(b_1 \alpha + b_0\right)E
\nonumber
\\
&\quad + \mu\frac{1}{2}\rho S\left(b_1 \alpha_e + b_0\right)\left(E + \frac{2T}{\rho A n}\right) -mg \cos{\gamma^*}.
\end{align}
\mcedit{Let} $\lambda = a_1/b_1$, \mcedit{then}
the linear combination
\eqref{eq:eom7} $+$ $\lambda$\eqref{eq:eom8}
yields
\begin{multline}
\!\!\!\!\! m a + \underbrace{(\lambda m \gamma^{*\prime} + \frac{1}{2}\rho S (a_0 -\lambda b_0))}_{c(\gamma^{*\prime})}E + \underbrace{mg( \sin{\gamma^*} + \lambda \cos{\gamma^*})}_{d(\gamma^*)}
\\
= \underbrace{T \cos{\alpha} + \lambda T \sin{\alpha} - \mu S^\star (a_0 - \lambda b_0 ) T}_{\tau} ,
\label{eq:nocvx}
\end{multline}
where
$S^\star=\frac{S}{AN}$
\mcedit{and $\tau$ is a virtual input defined by}
\begin{equation}
\tau = T \cos{\alpha} + \lambda T \sin{\alpha} - \mu S^\star (a_0 - \lambda b_0 ) T.
\label{eq:tau}
\end{equation}
Since $\gamma^*$ is prescribed \mcedit{by}
the path, \eqref{eq:nocvx} and \eqref{eq:diff} form a differential-algebraic system of linear equations and can thus be included as part of a convex program.
The constraints on the state \mcedit{variables} can be rewritten as
\begin{gather*}
0 \leq E \leq \overline{V}^2, \quad \underline{a} \leq a \leq \overline{a},
\\
E(s_0) = V_{0}^2
\text{ and }
E(s_f) = V_{f}^2.
\end{gather*}
The thrust constraint \mcedit{$ 0 \leq T \leq \overline{T}$}
cannot be expressed exactly as a function of $\tau$ only. However, assuming that $\alpha \ll 1$, and noting that $\lambda \ll 1$, and $S^\star (a_0 - \lambda b_0 ) \ll 1$ we have $\tau \approx T$ and
\mcedit{$ 0 \leq T \leq \overline{T}$ is therefore approximately equivalent to}
\[
0\leq \tau \leq \overline{T}.
\]
Finally the minimum thrust criterion in \eqref{eq:obj} can be \mcedit{approximated by}
a convex objective \mcedit{function under these conditions, since $\tau$ is then}
a good proxy for the thrust. By the change of differential operator in \eqref{eq:CV} we obtain
\[
J = \int_{s_0}^{s_f} \frac{(\tau/\overline{T})^2}{\sqrt{E}} \, \mathrm{d}s.
\]
We can thus state the following convex optimisation problem for a given path (i.e.~given $\gamma$ and $\gamma'$) as follows
\begin{equation}
\begin{aligned}
& \mathcal{P}_1 : \min_{\substack{\tau,\,E,\,a}} & & \int_{s_0}^{s_f} \frac{(\tau/\overline{T})^2}{\sqrt{E}} \, \mathrm{d}s,
\nonumber\\
& \text{ s.t.} & & m a + c(\gamma^{*\prime}) E + d(\gamma^*) = \tau ,
\nonumber\\
& & & E' = 2 a ,
\nonumber\\
& & & 0 \leq \tau \leq \overline{T},\ \underline{a} \leq a \leq \overline{a},
\nonumber\\
& & & 0 \leq E \leq \overline{V}^2,\ E(s_0) = V_{0}^2,\ E(s_f) = V_{f}^2.
\nonumber\\
\end{aligned}
\end{equation}
\mcedit{Solving $\mathcal{P}_1$ yields the optimal velocity}
along the path and a proxy for the \mcedit{optimal} thrust for sufficiently small values of the angle of attack ($\tau \approx T$).
\mcedit{However, a}
tiltwing angle \mcedit{profile that meets}
the dynamical constraints and follows the desired path with $\gamma \approx \gamma^*$ \mcedit{must also be computed}.
\mcedit{To achieve this}
we use the \mcedit{solution of $\mathcal{P}_1$ to define}
a new optimisation problem with variables $\gamma$, $\alpha$, $i_w$, and $M$ satisfying \mcedit{the constraints}
\begin{gather}
J_w (i_w' a + i_w'' E) = M,
\label{eq:d2i_w2}
\\
\underline{M} \leq M \leq \overline{M},
\nonumber
\\
i_w = \alpha + \gamma,
\nonumber
\end{gather}
\small
\begin{gather}
m E \gamma' = \tau \sin{\alpha} + (1-\mu)\frac{1}{2}\rho S (b_1 \alpha + b_0) E -mg \cos{\gamma}
\nonumber
\\
+\mu\frac{1}{2}\rho S \biggl[ b_1 \arcsin \biggl( \frac{\sqrt{E} \sin{\alpha}}{\sqrt{E + \smash{\frac{2\tau}{\rho A n}}\rule{0pt}{8.5pt}}}\biggr) + b_0\biggr] \Bigl(E + \frac{2\tau}{\rho A n}\Bigr),
\label{eq:eom_new}
\end{gather}
\normalsize
\mcedit{in which the objective is to minimise the cost function}
\begin{equation}
\label{eq:obj2}
J_{\gamma} = \int_{s_0}^{s_f} \frac{(\gamma - \gamma^*)^2}{\sqrt{E}} \mathrm{d}s.
\end{equation}
Note that to obtain \eqref{eq:d2i_w2} from \eqref{eq:tiltwing_dyna}, we \mcedit{have} used the result
$\frac{\mathrm{d^2}}{\mathrm{d}t^2} = V\frac{\mathrm{d}}{\mathrm{d}s} \left( V\frac{\mathrm{d}}{\mathrm{d}s} \right) =a \frac{\mathrm{d}}{\mathrm{d}s} + E \frac{\mathrm{d^2}}{\mathrm{d}s^2}$ inferred from the change of differential operator in \eqref{eq:CV}. To convexify the problem and \mcedit{to}
support the small angle approximation \mcedit{$\alpha,\alpha_e\ll 1$}
used \mcedit{in the derivation of the EOM},%
\footnote{We refer the reader to \cite{UMich19}, which shows that maintaining a small angle of attack during transition has negligible effect on performance, and this is preferable over operating the wing in dangerous near-stall regimes.}
we enforce bounds on $\alpha$ by imposing constraints $\underline{\alpha}~\leq~\alpha~\leq~\overline{\alpha}$,
\mcedit{for sufficiently small $\underline{\alpha}$, $\overline{\alpha}$,} \mbedit{chosen such that the wing is unstalled, and linear relationship with lift coefficient applies}.
This allows \mcedit{the dependence of \eqref{eq:eom_new} on $\alpha$} to be linearised
\small
\begin{multline}
m E \gamma' = \underbrace{\biggl(\tau + (1\!-\!\mu)\frac{1}{2}\rho S b_1 E +\mu\frac{1}{2}\rho S b_1 \sqrt{E^2 + \tfrac{2\tau E}{\rho A n}} \biggr)}_{p(E,\tau)}\alpha \\
+ \underbrace{(1 - \mu)\frac{1}{2}\rho S b_0 E +\mu\frac{1}{2}\rho S b_0 \Bigl(E + \frac{2\tau}{\rho A n}\Bigr)}_{q(E, \tau)} -mg \cos{\gamma}.
\label{eq:eom_small}
\end{multline}
\normalsize
\mcedit{Although $p(E,\tau)$ and $q(E,\tau)$ are determined by the solution of $\mathcal{P}_1$, \eqref{eq:eom_small} is nonconvex since} the last term is nonlinear in the \mcedit{decision} variable $\gamma$. However,
\mcedit{\eqref{eq:eom_small} can be omitted from the problem if the objective function \eqref{eq:obj2} is augmented}
\begin{equation}
\label{eq:obj3}
J^a_{\gamma} = \int_{s_0}^{s_f} \frac{(\gamma - \gamma^*)^2}{\sqrt{E}} +
\frac{(p \alpha + q - m E \gamma' -m g \cos{\gamma^*})^2}{(mg)^2\sqrt{E}} \mathrm{d}s .
\end{equation}
This \mcedit{objective attempts to minimise}
the error between the \mdsedit{guessed} \mcedit{and actual}
flight path angles \mcedit{while enforcing
\eqref{eq:eom_small} in the sense that, if a feasible solution exists satisfying both $\gamma = \gamma^\ast$ and \eqref{eq:eom_small}, then the optimal solution must satisfy \eqref{eq:eom_small}}.
We thus state the following optimisation problem, with $\tau$, $E$ and $a$ prescribed from problem $\mathcal{P}_1$
\begin{equation}
\begin{aligned}
& \mathcal{P}_2 : \min_{\substack{\alpha,\,\gamma,\,i_w,\, M}} & & \int_{s_0}^{s_f} \frac{(\gamma - \gamma^*)^2}{\sqrt{E}} \mathrm{d}s \\&&&+ \int_{s_0}^{s_f}\frac{(p \alpha + q - m E \gamma' -m g \cos{\gamma^*})^2}{(mg)^2\sqrt{E}} \mathrm{d}s
\nonumber\\
& \text{ s.t.} & & J_w (i_w' a + i_w'' E) = M,\ i_w(s_0) = i_{0},
\nonumber\\
& & & i_w'(s_0) \sqrt{E(s_0)} = \Omega_{0},
\nonumber\\
& & & i_w = \alpha + \gamma,
\nonumber\\
& & & \underline{M} \leq M \leq \overline{M},\ \underline{\alpha} \leq \alpha \leq \overline{\alpha}
\nonumber\\
& & & \underline{\gamma} \leq \gamma \leq \overline{\gamma},\ \underline{i_w} \leq i_w \leq \overline{i_w}.
\nonumber\\
\end{aligned}
\end{equation}
It should be emphasized that $\mathcal{P}_2$ may admit solutions that allow $\gamma$ to divert
from the \mdsedit{guessed} flight path angle $\gamma^*$. This occurs for example when the bounds on angle of attack do not allow
the desired path \mcedit{to be followed (see e.g.~Section \ref{sec:results}, scenario 2)}. In \mcedit{this case},
problems $\mathcal{P}_1$ and $\mathcal{P}_2$ can be
rerun sequentially after reinitializing with the newly obtained path ($\gamma^* \gets \gamma$). This process can be repeated iteratively until $\gamma$ is sufficiently close to $\gamma^*$.
Although we \mcedit{do not} provide theoretical guarantees of convergence \mcedit{for this} case, \mcedit{we note}
that the objective value \mcedit{has decreased by orders of magnitude}
after several iterations, as illustrated in Section \ref{sec:results}.
Only one EOM \mcedit{(\ref{eq:eom8})} was needed to constrain the variables of problem $\mathcal{P}_2$ since the EOM \mcedit{(\ref{eq:eom7})} can be enforced \textit{a posteriori} by choosing $T$ so that equation \eqref{eq:tau} is satisfied
\[
T = \frac{\tau}{ \cos{\alpha} + \lambda \sin{\alpha} - \mu S^\star (a_0 - \lambda b_0 )}.
\]
This guarantees that the solution satisfies both EOM within the required accuracy,
since problem $\mathcal{P}_1$ includes a linear combination of \mcedit{(\ref{eq:eom7}) and (\ref{eq:eom8}), and problem $\mathcal{P}_2$ enforces (\ref{eq:eom8})}.
\mcedit{Given an optimal set of} inputs and states as functions of the \mcedit{independent variable} $s$, the \mcedit{final} step is to convert back to time domain. This can be done by \mcedit{reversing} the change of differential operator in equation \eqref{eq:CV} and integratin
\[
t(\xi) = \int_{s_0}^\xi \frac{\mathrm{d}s }{V(s)}.
\]
\section{Discrete convex optimisation}
\label{sec:discrete}
\mcedit{The decision variables in} $\mathcal{P}_1$ and $\mathcal{P}_2$ \mcedit{are functions defined on the interval $[s_0,s_f]$}.
To obtain \mcedit{computationally tractable problems}, we
\mcedit{consider}
$N+1$ discretisation points $\{s_0, s_1, \ldots, s_{N}\}$ of the path, with spacing $\delta_k = s_{k+1} - s_{k}$, $k = 0, \ldots, N-1$ ($N$ steps).
The notation $\{u_0, \ldots, u_{N}\}$ is used for the sequence of the discrete values of a continuous variable $u$ evaluated at the discretisation points of the mesh, where $u_k = u(s_{k})$, $\forall k \in \{0, \ldots, N\}$.
Assuming a path $s_k \rightarrow (x_k, z_k)$, the prescribed flight path angle and rate are discretised as follows
\begin{align}
\gamma_k^* &= \arctan \Bigl(-\frac{z_{k+1}- z_{k}}{x_{k+1}- x_{k}}\Bigr), \quad k \in \{0, \ldots, N-1\},
\label{eq:gamma}
\\
\gamma_k^{*\prime} &= \begin{cases} (\gamma_{k+1}^*- \gamma_{k}^*)/{\delta_k}, & k \in \{0, \ldots, N-2\}, \\
\gamma_{N-2}^{*\prime}, & k = N-1.
\end{cases}
\label{eq:dgamma}
\end{align}
The resulting discrete time versions of $\mathcal{P}_1$ and $\mathcal{P}_2$ are
\begin{equation}
\begin{aligned}
& \mathcal{P}_1^\dagger : \min_{\substack{\tau,\,E,\,a}} & & \sum_{k=0}^{N-1} \frac{(\tau_k/\overline{T})^2}{\sqrt{E_k}} \, \delta_k,
\nonumber\\
& \text{ s.t.} & & m a_k + c(\gamma_k^{*\prime}) E_k + d(\gamma_k^*) = \tau_k ,
\nonumber\\
& & & E_{k+1} = E_{k} + 2 a_{k} \delta_{k},
\nonumber\\
& & & 0 \leq \tau_k \leq \overline{T}, \quad \underline{a} \leq a_k \leq \overline{a},
\nonumber\\
& & & 0 \leq E_k \leq \overline{V}^2, \quad E_0 = V_{0}^2, \quad E_{N} = V_{f}^2,
\nonumber\\
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
& \mathcal{P}_2^\dagger : \min_{\substack{\alpha,\,\gamma,\, \theta,\\i_w,\, \zeta,\, M}} & & \sum_{k=0}^{N-1} \frac{(\gamma_k - \gamma_k^*)^2}{\sqrt{E_k}}\delta_k \\ & & & + \sum_{k=0}^{N-1} \frac{(p \alpha_k + q - m E_k \Psi_k -m g \cos{\gamma_k^*})^2}{(mg)^2\sqrt{E_k}} \delta_k,
\nonumber\\
& \text{s.t.} & & \gamma_{k+1} = \gamma_{k} + \Psi_k \delta_k ,
\nonumber\\
& & & i_{w,k} = \alpha_k + \gamma_k,
\nonumber\\
& & & i_{w,k+1} = i_{w,k} + \zeta_k \delta_k, \quad i_{w,0} = i_{0},
\nonumber\\
& & & \zeta_{k+1} = \zeta_{k}\Bigl(1-\frac{a_k \delta_k}{E_k}\Bigr) + \frac{M_k \delta_k}{J_w {E_k}},
\nonumber \\
& & & \zeta_0 \sqrt{E}_0 = \Omega_{0},
\nonumber \\
& & & \underline{M} \leq M_k \leq \overline{M}, \quad \underline{\alpha} \leq \alpha_k \leq \overline{\alpha},
\nonumber\\
& & & \underline{\gamma} \leq \gamma_k \leq \overline{\gamma}, \quad \underline{i_w} \leq i_{w,k} \leq \overline{i_w}.
\nonumber\\
\end{aligned}
\end{equation}
Once
$\mathcal{P}_1^\dagger$ and $\mathcal{P}_2^\dagger$ have been solved, we check whether $|\gamma_k^*-\gamma_k| \leq \epsilon$ $\forall k \in \{0, \ldots, N\}$, where $\epsilon$ is a specified tolerance. If this condition is not met, the problem is reinitialized with the updated flight path angle and rate $\gamma^*_k \gets \gamma_k$, $\gamma^{*\prime}_k \gets \gamma_k'$.
and \mcedit{$\mathcal{P}_1^\dagger$ and $\mathcal{P}_2^\dagger$ are solved again}.
When \mcedit{the solution tolerance is met}
(or
\mcedit{the} maximum number of iterations \mcedit{is exceeded}) the problem is considered solved and the
input and state \mcedit{variables} are reconstructed
using
\begin{align*}
T_k &= \frac{\tau_k}{ \cos{\alpha_k} + \lambda \sin{\alpha_k} - \mu S^\star (a_0 - \lambda b_0 )}, \quad V_k &= \sqrt{E_k} ,
\end{align*}
\mcedit{and}
the time $t_k$ associated with each discretisation point
\mcedit{is computed, allowing solutions to be expressed}
as time series
\[
\label{eq:back2time}
t_k = \sum_{j=0}^{k-1} \frac{\delta_j}{V_j}.
\]
The procedure is summarised in Algorithm 1.
\begin{algorithm}
\caption{Convex trajectory optimisation}
\label{algo}
\mcedit{Compute $\gamma^*$, $\gamma^{*\prime}$ using \eqref{eq:gamma}, \eqref{eq:dgamma}} and \mcedit{initialise: $\gamma\gets\gamma^*$, $\gamma'\gets\gamma^{*\prime}$,} $\gamma^* \gets \infty$, $j \gets 0$\\
\While{$\displaystyle\max_{k \in \{0, \ldots, N\}}|\gamma_k^*-\gamma_k| > \epsilon$ $\&$ $j < $ MaxIters}{
$\gamma^{*} \gets \gamma$, $\gamma^{*\prime} \gets \gamma'$\\
Solve problem $\mathcal{P}_1^\dagger$ and problem $\mathcal{P}_2^\dagger$
\\
$j \gets j+1$
}
\For{$k\gets0$ \KwTo $N$}{
$T_k \gets \frac{\tau_k}{ \cos{\alpha_k} + \lambda \sin{\alpha_k} - \mu S^\star (a_0 - \lambda b_0 )}$ \\
$V_k \gets \sqrt{E_k}$\\
$t_k \gets \sum_{j=0}^{k-1} \frac{\delta_j}{V_j}$
}
\end{algorithm}
\section{Results}
\label{sec:results}
We consider a case study \mcedit{based on}
the Airbus A$^3$ Vahana. The \mcedit{aircraft}
parameters are reported in Table \ref{tab:param}.
We run Algorithm 1 \mcedit{using}
the convex programming software package CVX \cite{cvx} with \mcedit{the solver} Mosek \cite{mosek} to compute the optimal trajectory for 3 different transition manoeuvres, with boundary conditions given in Table \ref{tab:BC}.
The first \mcedit{scenario} is a forward smooth transition where the flight path angle varies slowly from $75^\circ$ to
\mcedit{a value close to zero},
as illustrated in Figure \ref{fig:traj1}. As the \mcedit{aircraft transitions}
from powered lift
to cruise, the velocity magnitude increases and the thrust decreases, illustrating the change in lift \mcedit{generation} from propellers to wing. The angle of attack remains small over the transition as it occurs smoothly.
The second \mcedit{scenario} is a constant altitude forward transition. This manoeuvre is more abrupt and would require a zero flight path angle throughout, thus exceeding the bounds on $\alpha$ at the beginning when the wing is tilted vertically. Solving problem $\mathcal{P}_2^\dagger$ with
\mcedit{constraints}
on $\alpha$
\mcedit{initially yields}
an optimal $\gamma$ profile
\mcedit{that differs}
from the \mcedit{anticipated}
zero flight path angle.
\mcedit{Reinitializing}
the problem with the \mcedit{computed}
flight path angle ($\gamma^* \gets \gamma$) and
\mcedit{iterating}
until a satisfactory match between $\gamma$ and $\gamma^*$
\mcedit{is}
achieved
\mcedit{results in convergence illustrated empirically in Figure \ref{fig:obj}}.
\mcedit{The solution thus obtained is shown}
in Figure \ref{fig:traj2}. Since the
flight path angle is not identically zero (as would be required for a constant altitude transition) the manoeuvre is characterised by an initial increase in altitude and overshoot of about 30\,m. This is
\mcedit{a consequence of}
limiting the angle of attack. More abrupt transitions with
\mcedit{smaller altitude variations} are possible, but this \mcedit{requires stalling}
the wing and \mcedit{operating} in flight envelopes \mcedit{for which our approximations are not valid}.
The third \mcedit{scenario} is a backward transition with an increase in altitude (Figure \ref{fig:traj3}).
\mcedit{This}
is characterised by \mcedit{an initial}
decrease in velocity magnitude and increase in thrust.
The strict bounds on angle of attack \mcedit{necessitate an increase in altitude of about 100\,m}.
A backward transition at constant altitude would require \mcedit{stalling}
the wing, which is prohibited in the present formulation, illustrating a limitation of our approach.
To achieve the backward transition,
a high-drag device or flaps \mcedit{are needed to provide braking forces}.
This can be modelled by adding a constant term $\delta$ to $c(\gamma^{*\prime})$ in problem $\mathcal{P}_1^\dagger$ when simulating the backward transition.
The time-complexity of the
\mcedit{proposed algorithm in scenario~1 is shown in Figure \ref{fig:time}}.
The time to completion \mcedit{as a function of problem size (number of discretisation points) is asymptotically quadratic, implying excellent scalability}. In particular, for $N<1500$, the time to completion is of the order of seconds. A real-time solution is thus possible
\mcedit{using modest computational resources}.
\mcedit{This is in stark contrast to the complexity of generic NLP approaches \cite{leo}.}
\mdsedit{Moreover, as shown in \cite{me3}, the time complexity for this type of problem can be reduced significantly with first-order solvers, paving the way for fast real-time implementations}.
\begin{comment}
\subsection{Forward smooth manoeuvre}
To be continued...
\subsection{Forward constant altitude manoeuvre}
To be continued...
\subsection{Backward manoeuvre}
To be continued...
\end{comment}
\begin{table}[ht]
\centering
\begin{tabular}{llll}
\hline
\textbf{Parameter} & \textbf{Symbol} & \textbf{Value} & \textbf{Units} \\ \hline
Mass & $m$ & $752.2$ & \si{kg} \\ \hline
Gravity acceleration & $g$ & $9.81$ & \si{m.s^{-2}} \\ \hline
Wing area & $S$ & $8.93$ & \si{m^2} \\ \hline
Disk area & $A$ & $2.83$ & \si{m^2} \\ \hline
Blown-unblown ratio & $\mu$ & $0.73$ & \si{-} \\ \hline
Wing inertia & $J_w$ & $1100$ & \si{kg.m^2} \\ \hline
Density of air & $\rho$ & $1.225$ & \si{kg.m^{-3}} \\ \hline
{Lift coefficients} & $b_0, b_1$ & $0.43, 0.11$ & \si{-}, \si{deg^{-1}} \\ \hline
{Drag coefficients} & $a_0, a_1$ & $0.029, 0.004$ & \si{-}, \si{deg^{-1}} \\ \hline
Maximum thrust & $\overline{T}$ & $8855$ & \si{N} \\ \hline
Angle of attack range & $\left[\underline{\alpha}, \overline{\alpha}\right]$ & $\left[-20, 20\right]$ & \si{deg} \\ \hline
Flight path angle range & $\left[\underline{\gamma}, \overline{\gamma}\right]$ & $\left[-90, 90\right]$ & \si{deg} \\ \hline
Tiltwing angle range & $\left[\underline{i}_w, \overline{i}_w\right]$ & $\left[0, 100\right]$ & \si{deg} \\ \hline
Acceleration range & $\left[\underline{a}, \overline{a}\right]$ & $\left[-0.3g, 0.3g\right]$ & \si{m . s^{-2}} \\ \hline
Velocity range & $\left[\underline{V}, \overline{V}\right]$ & $\left[0, 40\right]$ & \si{m/s} \\ \hline
Momentum range & $\left[\underline{M}, \overline{M}\right]$ & $\left[-50; 50\right]$ & \si{N.m} \\ \hline
Number of propellers & $n$ & $4$ & \si{-} \\ \hline
Discretisation points & $N$ & $1500$ & \si{-} \\ \hline
\end{tabular}
\vspace{1mm}\caption{Model parameters \mcedit{derived} from A$^3$ Vahana}
\label{tab:param}
\vspace{-3mm}
\end{table}
\begin{table}[ht]
\centering
\begin{tabular}{llll}
\hline
\textbf{Parameter} & \textbf{Symbol} & \textbf{Value} & \textbf{Units} \\ \hline
\multicolumn{4}{c}{{\cellcolor[rgb]{0.753,0.753,0.753}}\textbf{Forward transition}} \\
Velocity & $\left\{V_0; V_f\right\}$ & $\left\{0.5; 40\right\}$ & \si{m/s} \\ \hline
Tiltwing angle & $i_0$ & $75$ & \si{deg} \\ \hline
Tiltwing angle rate & $\Omega_0$ & $0$ & \si{deg/s} \\ \hline
Flight path angle & $\gamma_0$ & $75$ & \si{deg} \\ \hline
\multicolumn{4}{c}{{\cellcolor[rgb]{0.753,0.753,0.753}}\textbf{Backward transition}} \\
Velocity & $\left\{V_0; V_f\right\}$ & $\left\{40; 0.1\right\}$ & \si{m/s} \\ \hline
Tiltwing angle & $\left\{i_0; i_f\right\}$ & $\left\{0; 75\right\}$ & \si{deg} \\ \hline
Tiltwing angle rate & $\Omega_0$ & $0$ & \si{deg/s} \\ \hline
Flight path angle & $\gamma_0$ & $1.6$ & \si{deg} \\ \hline
\end{tabular}
\vspace{1mm}\caption{Boundary conditions for transitions}
\label{tab:BC}
\vspace{-3mm}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth, trim={0cm 0.3cm 0cm 0.5cm}, clip]{img/traj1.pdf}
\caption{Forward transition (scenario 1)}
\label{fig:traj1}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth, trim={0cm 0cm 0cm 0.9cm}, clip]{img/obj.pdf}
\caption{Convergence of objective value for $\mathcal{P}_2^\dagger$ (scenario 2)}
\label{fig:obj}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth, trim={0cm 0cm 0cm 0cm}, clip]{img/traj2.pdf}
\caption{Forward transition (scenario 2)}
\label{fig:traj2}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth, trim={0cm 0cm 0cm 0cm}, clip]{img/traj3.pdf}
\caption{Backward transition (scenario 3)}
\label{fig:traj3}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth, trim={0cm 0cm 0cm 0cm}, clip]{img/time_log.pdf}
\caption{Time to completion as a function of problem size}
\label{fig:time}
\end{figure}
\section{Conclusions}
\label{sec:conclusion}
A convex programming formulation of the minimum thrust transition problem for a \mcedit{tiltwing} VTOL aircraft with propeller-wing interaction was proposed and solved for various transition scenarios. The
approach can compute an optimal trajectory within seconds for large numbers of discretisation points and is particularly efficient at computing smooth transitions. A potential limitation of the approach is its reliance on small angles of attack, which may restrict the range of achievable manoeuvres. In particular, more abrupt manoeuvres with larger angles of attack are prohibited. \mcedit{Another drawback is that, for the case in which the initial guess of flight path angle is infeasible, the evidence for the convergence of the proposed iteration is empirical rather than theoretical}.
Future work will address these two issues using \mcedit{techniques}
from robust MPC theory.
A further extension to this work
\mcedit{is to develop a first-order solver for online trajectory generation}.
|
\section{Introduction}
This paper is concerned with the large time behavior of a solution
to the Cauchy problem for an inhomogeneous fractional diffusion equation
\begin{equation}
\label{eq:1.1}
\partial_t u+(-\Delta)^{\frac{\theta}{2}}u=f(x,t)\quad\mbox{in}\quad{\mathbb R}^N\times(0,\infty),
\quad
u(x,0)=\varphi(x)\quad\mbox{in}\quad{\mathbb R}^N,
\end{equation}
where $N\ge 1$, $\partial_t:=\partial/\partial t$, $0<\theta<2$ and
$$
\varphi\in L_K^1:=L^1({\mathbb R}^N,(1+|x|)^K\,dx)\quad\mbox{with}\quad K\ge 0.
$$
Here $(-\Delta)^{\theta/2}$ is the fractional power of the Laplace operator.
Inhomogeneous fractional diffusion equation~\eqref{eq:1.1}
appears in the study of various nonlinear problems with anomalous diffusion,
the Laplace equation with a dynamical boundary condition, and so on.
Under suitable integrability conditions on the inhomogeneous term $f$,
the solution~$u$ to problem~\eqref{eq:1.1} behaves like a suitable multiple of the fundamental solution~$G_\theta$
to the linear fractional diffusion equation
$$
\partial_t v+(-\Delta)^{\frac{\theta}{2}}v=0\quad\mbox{in}\quad{\mathbb R}^N\times(0,\infty)
$$
as $t\to\infty$.
In this paper we obtain the higher order asymptotic expansions (HOAE) of the large time behavior of
the solution~$u$. Furthermore, we study the precise description of the large time behavior of solutions to the Cauchy problem
for nonlinear fractional diffusion equations such as
\begin{equation}
\label{eq:1.2}
\partial_t u+(-\Delta)^{\frac{\theta}{2}}u=\lambda|u|^{p-1}u
\quad\mbox{in}\quad{\mathbb R}^N\times(0,\infty),
\quad
u(x,0)=\varphi(x)\quad\mbox{in}\quad{\mathbb R}^N,
\end{equation}
where $\lambda\in{\mathbb R}$, $p>1$ and $\varphi\in L_K^1$ with $K\ge 0$.
This paper is an improvement of \cite{IKM} and
it corresponds a fractional version of the papers~\cites{IIK, IK, IKK02}.
\vspace{5pt}
The large time behavior of solutions to nonlinear parabolic equations has been studied extensively
in many papers by various methods.
Here we just refer to the papers
\cites{AJY, BK, FK, FIK, FM, HKN, IIK, IK, IKK01, IKK02, IKM, Iwa, YM, YS01, YS02, YS03, YT1, YT2},
which are closely related to this paper.
Among others,
in \cites{IIK, IK, IKK02},
HOAE of solutions behaving like suitable multiples of the Gauss kernel
have already been well established. The property that
$$
\bigcup_{t>0}e^{t\Delta}L_K^1\subset L_K^1
\quad\mbox{for}\quad K\ge 0
$$
plays an important role in \cites{IIK, IK, IKK02} and it follows
from the exponential decay of the Gauss kernel at the space infinity.
For fractional diffusion equations,
if $0\le K<\theta$, then
\begin{equation}
\label{eq:1.3}
\bigcup_{t>0}e^{-t(-\Delta)^{\theta/2}}L_K^1\subset L_K^1
\end{equation}
holds and the arguments in \cites{IIK, IK, IKK02} are also applicable to fractional diffusion equations.
However, if $K\ge\theta$, then property~\eqref{eq:1.3} fails.
This fact prevents to establish analogous asymptotic expansions of solutions to the case of $\theta=2$.
In \cite{IKM} the authors of this paper and Michihisa studied a mechanism for property~\eqref{eq:1.3} to fail in the case of $K\ge\theta$,
and obtained HOAE of $e^{-(-\Delta)^{\theta/2}}\varphi$.
This argument is applicable to
the study of HOAE of solutions
to inhomogeneous fractional diffusion equations and nonlinear fractional diffusion equations,
however HOAE of \cite{IKM} to problem~\eqref{eq:1.1} do not have refined forms.
In this paper we improve and refine arguments in \cite{IKM} by
taking into an account of the Taylor expansion of the kernel $G_\theta$ with respect to
both of the space and the time variables, and obtain
HOAE of solutions to inhomogeneous fractional diffusion equations and nonlinear fractional parabolic equations.
Our arguments also reveal a mechanism
for the solution $u$ to problem~\eqref{eq:1.1} to break the property that
$u(t)\in L^1_K$ for $t>0$.
We introduce some notations.
Set $\mathbb N_0:={\mathbb N}\cup\{0\}$.
For any $k\ge 0$, let $[k]\in\mathbb N_0={\mathbb N}\cup\{0\}$
be such that $k-1<[k]\le k$.
Let $\nabla:=(\partial/\partial x_1,\dots,\partial/\partial x_N)$.
For any multi-index $\alpha\in{\mathbb M}:=\mathbb N_0^N$,
set
$$
|\alpha|:=\displaystyle{\sum_{i=1}^N}\alpha_i,\quad
\alpha!:=\prod_{i=1}^N\alpha_i!,\quad
x^\alpha:=\prod_{i=1}^N x_i^{\alpha_i},\quad
\partial_x^\alpha:=
\frac{\partial^{|\alpha|}}{\partial x_1^{\alpha_1}\cdots\partial x_N^{\alpha_N}}.
$$
For any $\alpha=(\alpha_1,\dots,\alpha_N)$, $\beta=(\beta_1,\dots,\beta_N)\in{\mathbb M}$,
we say
$\alpha\le\beta$
if $\alpha_i\le\beta_i$ for all $i\in\{1,\dots,N\}$.
Let $1\le q\le \infty$ and $K\ge 0$.
Let $\|\cdot\|_q$ be the usual norm of $L^q:=L^q({\mathbb R}^N)$.
Set
$$
|||f|||_{q,K}:=\|f_K\|_q\quad\mbox{with}\quad f_K(x):=|x|^K f(x).
$$
Let
$$
f\in L_K^q:=\left\{f\in L^q\,:\,\|f\|_{L^q_K}<\infty\right\},
\quad
\mbox{where}\quad
\|f\|_{L^q_K}:=\|f\|_q+|||f|||_{q,K}.
$$
For any $f\in L^1_K$ and $\alpha\in\mathbb M$ with $|\alpha|\le K$,
set
$$
M_\alpha(f):=\int_{\mathbb R^N}x^\alpha f(x)\,dx.
$$
We are ready to state our main results on the asymptotic expansions of solutions to inhomogeneous fractional diffusion equations.
In what follows, set $K_\theta:=[K/\theta]$. Furthermore, set
$$
g_{\alpha,m}(x,t):=\frac{(-1)^{|\alpha|+m}}{\alpha!m!}(\partial_t^m\partial_x^\alpha G_\theta)(x,t+1)
$$
for $(x,t)\in{\mathbb R}^N\times(0,\infty)$,
where $\alpha\in{\mathbb M}$ and $m\in{\mathbb N}_0$.
\begin{theorem}
\label{Theorem:1.1}
Let $N\ge1$, $0<\theta<2$, $0\le\ell\le K$, and $1\le q\le\infty$.
Let $\varphi\in L^1_K$ and $f$ be a measurable function in $\mathbb R^N\times(0,\infty)$ such that
\begin{equation}
\label{eq:1.4}
E_{K,q}[f]\in L^1_{{\rm loc}}(0,\infty),
\end{equation}
where
\begin{equation}
\label{eq:1.5}
\begin{aligned}
E_{K,q}[f](t)
& := (t+1)^{\frac{K}{\theta}}
\left[t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)}\|f(t)\|_q+\|f(t)\|_1\right]\\
& \qquad
+t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)}|||f(t)|||_{q,K}+|||f(t)|||_{1,K}
\quad\mbox{for}\quad t>0.
\end{aligned}
\end{equation}
Let $u\in C({\mathbb R}^N\times(0,\infty))$ be a solution to problem~\eqref{eq:1.1}, that is,
$u$ satisfies
$$
u(x,t)
=\int_{{\mathbb R}^N}G_\theta(x-y,t)\varphi(y)\,dy
+\int_0^t\int_{{\mathbb R}^N}G_\theta(x-y,t-s)f(y,s)\,dy\,ds
$$
for $(x,t)\in{\mathbb R}^N\times(0,\infty)$.
Then
\begin{equation}
\label{eq:1.6}
\sup_{0<t<\tau}\,t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)}|||u(t)-w(t)|||_{q,\ell}<\infty
\quad\mbox{for}\quad\tau>0,
\end{equation}
where
\begin{equation}
\label{eq:1.7}
w(x,t):=
\sum_{m=0}^{K_\theta}\sum_{|\alpha|\le K}
\bigg\{M_\alpha(\varphi)
+\int_0^t(s+1)^m M_\alpha(f(s))\,ds\bigg\}\,g_{\alpha,m}(x,t).
\end{equation}
Furthermore, there exists $C>0$ such that,
for any $\varepsilon>0$ and $T>0$,
\begin{equation}
\label{eq:1.8}
t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||u(t)-w(t)|||_{q,\ell}\\
\le\varepsilon t^{-\frac{K}{\theta}}
+Ct^{-\frac{K}{\theta}}\int_T^tE_{K,q}[f](s)\,ds
\end{equation}
holds for large enough $t>0$.
In particular, if
$$
\int_0^\infty E_{K,q}[f](s)\,ds<\infty,
$$
then
\begin{equation}
\label{eq:1.9}
\lim_{t\to\infty}t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)+\frac{K-\ell}{\theta}}|||u(t)-w(t)|||_{q,\ell}=0.
\end{equation}
\end{theorem}
Theorem~\ref{Theorem:1.1} corresponds to \cite{IKK02}*{Theorems~1.1, 1.2} for $\theta=2$
and it is an improvement of \cite{IKM}*{Theorem~3.1~(ii)}.
Our asymptotic profile $w$ has a pretty simpler form than that of \cite{IKM}.
(See Remarks~\ref{Remark:3.1} and \ref{Remark:5.2}.)
We also remark that, under condition~\eqref{eq:1.4},
both of $u(\cdot,t)$ and $w(\cdot,t)$ do not necessarily belong to $L^q_\ell$,
while $u(t)-w(t)\in L^q_\ell$.
In other words, the function~$w$ may break the property that $u(t)\in L^q_\ell$ for $t>0$.
Furthermore, we have:
\begin{corollary}
\label{Corollary:1.1}
Assume the same conditions as in Theorem~{\rm\ref{Theorem:1.1}}.
Let $u\in C({\mathbb R}^N\times(0,\infty))$ be a solution to problem~\eqref{eq:1.1}.
Then there exists $R>0$ such that
$$
u(t)\in \left\{h+\sum_{(\alpha,m)\in\Lambda_K^q}a_{\alpha,m}g_{\alpha,m}(x,t)\,:\,
\mbox{$h\in L^q_K$ with $\|h\|_{L^q_K}\le R$},\,\{a_{\alpha,m}\}\subset[-R,R]\right\}
$$
for $t>0$, where
$\Lambda_K^q:=\{(\alpha,m)\in{\mathbb M}\times{\mathbb N}_0\,:\,g_{\alpha,m}(\cdot,0)\not\in L^q_K\}$.
\end{corollary}
We explain the idea of the proof of Theorem~\ref{Theorem:1.1}.
We improve and refine arguments in the previous papers~\cites{IIK, IK, IKK02, IKM}
to obtain HOEA of the solution~$u$ to problem~\eqref{eq:1.1}, in particular, the integral term
$$
\int_0^t\int_{{\mathbb R}^N}G_\theta(x-y,t-s)f(y,s)\,dy\,ds.
$$
In \cite{IKM},
following the arguments in \cites{IIK, IK, IKK02},
the authors of this paper and Michihisa expanded
the integral kernel $G_\theta(x-y,t-s)$ by the Taylor expansions with respect to the space derivatives of $G_\theta(x,t-s)$.
Then the slow decay of $G_\theta(x,t)$
makes difficult to obtain refined HOAE of the solution~$u$ to problem~\eqref{eq:1.1}.
In this paper we expand the integral kernel $G_\theta(x-y,t-s)$
by the Taylor expansions with respect to both of space and time variable derivatives of $G_\theta(x,t)$.
(This is the same sprit as in \cite{FM}.)
Indeed, we introduce the following integral kernels by the use of the Taylor expansions of $G_\theta$:
\begin{equation}
\label{eq:1.10}
\begin{aligned}
{\mathcal S}^m_\ell(x,y,t) & :=(\partial_t^m G_\theta)(x-y,t)
-\sum_{|\alpha| \le \ell}\frac{(-1)^{|\alpha|}}{\alpha !}
(\partial_t^m\partial^\alpha_x G_\theta)(x,t)y^\alpha\\
& \,\, =\frac{1}{[\ell]!}\int_0^1(1-\tau)^{[\ell]}\frac{\partial^{[\ell]+1}}{\partial\tau^{[\ell]+1}}
(\partial_t^m G_\theta)(x-\tau y,t)\,d\tau,\\
{\mathcal T}(x,y,t,s) & :=G_\theta(x-y,t-s)-\sum_{m=0}^{K_\theta} \frac{(-1)^{m}}{m!}
(\partial_t^m G_\theta)(x-y,t)s^m\\
& \,\,=\frac{1}{K_\theta!}
\int_0^1(1-\tau)^{K_\theta}\frac{\partial^{K_\theta+1}}{\partial\tau^{K_\theta+1}}G_\theta(x-y,t-\tau s)\,d\tau,
\end{aligned}
\end{equation}
for $x$, $y\in{\mathbb R}^N$ and $0\le s<t$, where $0\le\ell\le K$ and $m\in{\mathbb N}_0$.
Then
$$
\begin{aligned}
& G_\theta(x-y,t-s)\\
& =\sum_{m=0}^{K_\theta} \frac{(-1)^{m}}{m!}(\partial_t^m G_\theta)(x-y,t)s^m+{\mathcal T}(x,y,t,s)\\
& =\sum_{m=0}^{K_\theta} \sum_{|\alpha| \le \ell}\frac{(-1)^{|\alpha|+m}}{\alpha ! m!}
(\partial_t^m\partial^\alpha_x G_\theta)(x,t)y^\alpha s^m
+\sum_{m=0}^{K_\theta} \frac{(-1)^m}{m!}{\mathcal S}_\ell^m(x,y,t) s^m+{\mathcal T}(x,y,t,s).
\end{aligned}
$$
Furthermore,
\begin{equation}
\label{eq:1.11}
\begin{split}
{\mathcal R}(x,y,t,s) & :={\mathcal T}(x,y,t,s)+\sum_{m=0}^{K_\theta} \frac{(-1)^m}{m!}{\mathcal S}_K^m(x,y,t) s^m\\
& \,=G_\theta(x-y,t-s)
-\sum_{m=0}^{K_\theta}\sum_{|\alpha| \le K}\frac{(-1)^{|\alpha|+m}}{\alpha ! m!}
(\partial_t^m\partial^\alpha_x G_\theta)(x,t)y^\alpha s^m.
\end{split}
\end{equation}
Then it follows from \eqref{eq:1.7} that
\begin{equation}
\label{eq:1.12}
\begin{aligned}
u(x,t)-w(x,t)
& =\int_{{\mathbb R}^N}{\mathcal R}(x,y,t+1,1)\varphi(y)\,dy\\
& +\int_0^t\int_{{\mathbb R}^N} {\mathcal R}(x,y,t+1,s+1)f(y,s)\,dy\,ds.
\end{aligned}
\end{equation}
Thanks to the decay of the derivatives of $G_\theta$ and \eqref{eq:1.10},
we see that
\begin{equation*}
\begin{aligned}
& {\mathcal T}(x,y,t,s)=O(|x-y|^{-N-K-\varepsilon})\quad\mbox{as}\quad |x-y|\to\infty,\\
& {\mathcal S}^m_K(x,y,t)=O(|x|^{-N-K-\varepsilon})\quad\mbox{as}\quad |x|\to\infty,
\end{aligned}
\end{equation*}
for some $\varepsilon>0$.
These decay of the integral kernels at the space infinity enables us
to establish HOAE of solutions to problem~\eqref{eq:1.1}
and to obtain Theorem~\ref{Theorem:1.1}.
These arguments require delicate integral estimates on the integral kernels
${\mathcal S}_\ell^m$ and ${\mathcal T}$.
\vspace{5pt}
Theorem~\ref{Theorem:1.1} is applicable to problem~\eqref{eq:1.2}
and it gives asymptotic profiles of solutions to problem~\eqref{eq:1.2}
as a linear combination of the derivatives of $G_\theta$ (see Theorem~\ref{Theorem:5.1}).
Furthermore,
taking a suitable approximation of the nonlinear term in problem~\eqref{eq:1.2},
we obtain refined asymptotic expansions of the solution to problem~\eqref{eq:1.1}
(see Theorem~\ref{Theorem:5.2}).
Here we state the following result, which is a variation of Theorem~\ref{Theorem:5.2}.
\begin{theorem}
\label{Theorem:1.2}
Let $N\ge 1$, $0<\theta<2$, $\lambda\in{\mathbb R}$, and $\varphi\in L^1_K\cap L^\infty$ with $K\ge0$.
Let $u\in C({\mathbb R}^N\times(0,\infty))$ be a solution to problem~\eqref{eq:1.2} with $p>1+\theta/N$
and satisfy
\begin{equation}
\label{eq:1.13}
\sup_{t>0}\,(t+1)^{\frac{N}{\theta}}\|u(t)\|_\infty<\infty.
\end{equation}
Then there exists $M_*\in{\mathbb R}$ such that
$$
M_*:=\lim_{t\to\infty}\int_{{\mathbb R}^N}u(x,t)\,dx
=\int_{{\mathbb R}^N}\varphi(x)\,dx+\int_0^\infty\int_{{\mathbb R}^N}F(u(x,t))\,dx\,dt,
$$
where $F(u(x,t)):=\lambda|u(x,t)|^{p-1}u(x,t)$.
Assume $N(p+\theta)>N+K$ and $\varphi\in L^\infty_k$ with $k=\min\{N+\theta,K\}$.
Let $1\le q\le\infty$.
Then
$$
\sup_{t>0}\,(t+1)^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||u(t)|||_{q,\ell}<\infty,
$$
where $0\le \ell \le K$ with $0\le \ell< \theta+N(1-1/q)$.
Furthermore, for any $\sigma>0$
$$
\begin{aligned}
&\sup_{t>0}\,t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}
|||u(t)-v(t)|||_{q,\ell}<\infty,
\\
&
t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||u(t)-v(t)|||_{q,\ell}
=o\left(t^{-\frac{K}{\theta}}\right)
+O\left(t^{-\frac{K}{\theta}}\int_1^t s^{\frac{K}{\theta}-A_p}h_\sigma(s)\,ds\right)
\quad\mbox{as}\quad t\to\infty,
\end{aligned}
$$
where $0\le\ell\le K$.
Here
$$
\begin{aligned}
& v(x,t):=
\sum_{m=0}^{K_\theta}\sum_{|\alpha|\le K}c_{\alpha,m}(t)g_{\alpha,m}(x,t)
+\int_0^t e^{-(t-s)(-\Delta)^{\frac{\theta}{2}}}F_\infty(s)\,ds,\\
& c_{\alpha,m}(t):=M_\alpha(\varphi)
+\int_0^t (s+1)^m M_\alpha(F(u(s))-F_\infty(s))\,ds,\\
& F_\infty(x,t):=F\left(M_*G_\theta(x,t+1)\right),\quad
h_\sigma(t):=t^{-(A_p-1)+\sigma}+t^{-1}+t^{-\frac{1}{\theta}}.
\end{aligned}
$$
\end{theorem}
Theorem~\ref{Theorem:1.2} corresponds to \cite{IK}*{Corollary~1.1} for $\theta=2$.
See Remark~\ref{Remark:5.1} for condition~\eqref{eq:1.13}.
\vspace{3pt}
The rest of this paper is organized as follows.
In Section~2 we collect some properties of the fundamental solution $G_\theta$.
In Section~3 we obtain some estimates on the integral kernels ${\mathcal S}_\ell^m(x,y,t)$
and ${\mathcal T}(x,y,t,s)$,
and prove Theorem~\ref{Theorem:1.1} and Corollary~\ref{Corollary:1.1}.
In Section~4 we apply Theorem~\ref{Theorem:1.1}
to obtain HOAE of solutions to
the Cauchy problem for a convection type inhomogeneous fractional diffusion equation.
In Section~5 we apply Theorem~\ref{Theorem:1.1} to
study HOAE of solutions to the Cauchy problem
for nonlinear fractional diffusion equations.
Furthermore, we prove Theorem~\ref{Theorem:1.2}.
\section{Preliminaries}
We recall some properties of the fundamental solution $G_\theta=G_\theta(x,t)$,
In what follows,
by the letter $C$
we denote generic positive constants (independent of $x$ and $t$)
and they may have different values also within the same line.
Let $0<\theta<2$.
The fundamental solution $G_\theta=G_\theta(x,t)$ is represented by
$$
G_\theta(x,t)=(2\pi)^{-\frac{N}{2}}\int_{{\mathbb R}^N}e^{ix\cdot \xi}e^{-t|\xi|^\theta}\,d\xi,
\quad
(x,t)\in{\mathbb R}^N\times(0,\infty).
$$
Then we have:
\begin{itemize}
\item[({\bf G})]
$G_\theta=G_\theta(x,t)$ is a positive smooth function in ${\mathbb R}^N\times(0,\infty)$ with the following properties:
\begin{itemize}
\item[(i)]
$\displaystyle{G_\theta(x,t)=t^{-\frac{N}{\theta}}G_\theta(t^{-\frac{1}{\theta}}x,1)}$
for $x\in{\mathbb R}^N$ and $t>0$;
\item[(ii)]
$\displaystyle{\sup_{x\in{\mathbb R}^N}\,(1+|x|)^{N+\theta+|\alpha|}
|(\partial_x^\alpha G_\theta)(x,1)|<\infty}$
for $\alpha\in{\mathbb M}$;
\item[(iii)]
$G_\theta(\cdot,1)$ is radially symmetric and decreasing with respect to $r:=|x|$. Furthermore,
$$
\liminf_{|x|\to+\infty}\,(1+|x|)^{N+\theta+j}(\partial_r^j G_\theta)(x,1)>0,\quad j\in\mathbb N_0;
$$
\item[(iv)]
$\displaystyle{G_\theta(x,t)=\int_{{\mathbb R}^N}G_\theta(x-y,t-s)\,G_\theta(y,s)\,dy}$
for $x\in\mathbb R^N$ and $t>s>0$;
\item[(v)]
$\displaystyle{\int_{\mathbb R^N}G_\theta(x,t)\,dx=1}$ for $t>0$.
\end{itemize}
\end{itemize}
See \cites{BT, BK}. (See also \cites{IKK01, IKK02, S}.)
Let $\alpha\in{\mathbb M}$ and $m\in{\mathbb N}_0$.
Let
$$
1\le q\le\infty,\quad
0\le \ell<\theta m'+|\alpha|+N\left(1-\frac{1}{q}\right)
\quad\mbox{with}\quad m':=\max\{m,1\}.
$$
It follows from ({\bf G})-(i), (ii) and \cite{IKM}*{Lemma~2.1} that
\begin{equation}
\label{eq:2.1}
|(\partial_t^m\partial_x^\alpha G_\theta)(x,t)|
\le Ct^{-\frac{N+|\alpha|}{\theta}-m}
\big(1+t^{-\frac{1}{\theta}}|x|\big)^{-(N+\theta m'+|\alpha|)}
\end{equation}
for $x\in{\mathbb R}^N$ and $t>0$.
This implies that
\begin{equation}
\label{eq:2.2}
\sup_{t>0}\,
t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)+\frac{|\alpha|-\ell}{\theta}+m}
|||(\partial_t^m\partial_x^\alpha G_\theta)(t)|||_{q,\ell}<\infty.
\end{equation}
\begin{lemma}
\label{Lemma:2.1}
Let $1\le q\le r\le\infty$, $\alpha\in{\mathbb M}$, and $m\in{\mathbb N}_0$.
Let
\begin{equation}
\label{eq:2.3}
0\le \ell<\theta m'+|\alpha|+N\left(\frac{1}{q}-\frac{1}{r}\right).
\end{equation}
Then there exists $C>0$ such that
$$
t^{\frac{N}{\theta}\left(\frac{1}{q}-\frac{1}{r}\right)+\frac{|\alpha|}{\theta}+m}
\biggr|\biggr|\biggr|\,\partial_t^m\partial_x^\alpha e^{-t(-\Delta)^{\theta/2}}\varphi\,
\biggr|\biggr|\biggr|_{r,\ell}
\le Ct^{\frac{\ell}{\theta}}\|\varphi\|_q+C|||\varphi|||_{q,\ell}
$$
for $\varphi\in L^q_\ell$ and $t>0$. Here
$$
\left[e^{-t(-\Delta)^{\theta/2}}\varphi\right](x):=\int_{{\mathbb R}^N}G_\theta(x-y,t)\varphi(y)\,dy,
\quad (x,t)\in{\mathbb R}^N\times(0,\infty).
$$
\end{lemma}
{\bf Proof.}
Assume \eqref{eq:2.3}. It follows that
$$
|x|^\ell\left[\partial_t^m\partial_x^\alpha e^{-t(-\Delta)^{\theta/2}}\varphi\right](x)
\le C\int_{{\mathbb R}^N}\left[|x-y|^\ell+|y|^\ell\right]
|(\partial_t^m\partial_x^\alpha G_\theta)(x-y,t)||\varphi(y)|\,dy
$$
for $(x,t)\in{\mathbb R}^N\times(0,\infty)$.
The Young inequality together with \eqref{eq:2.2} implies that
$$
\begin{aligned}
& \biggr|\biggr|\biggr|\,\partial_t^m\partial_x^\alpha e^{-t(-\Delta)^{\theta/2}}\varphi\,
\biggr|\biggr|\biggr|_{r,\ell}\\
& \le C|||\partial_t^m\partial_x^\alpha G_\theta(t)|||_{p,\ell}\|\varphi\|_q
+C\| \partial_t^m\partial_x^\alpha G_\theta(t)\|_p|||\varphi|||_{q,\ell}\\
& \le Ct^{-\frac{N}{\theta}\left(\frac{1}{q}-\frac{1}{r}\right)-\frac{|\alpha|}{\theta}-m+\frac{\ell}{\theta}}
\|\varphi\|_q
+Ct^{-\frac{N}{\theta}\left(\frac{1}{q}-\frac{1}{r}\right)-\frac{|\alpha|}{\theta}-m}|||\varphi|||_{q,\ell}
\end{aligned}
$$
for $t>0$, where $p\in[1,\infty]$ with $1/r=1/p+1/q-1$.
Then we obtain the desired inequality, and the proof is complete.
$\Box$
\section{Proof of Theorem~\ref{Theorem:1.1}}
In this section we prove Theorem~\ref{Theorem:1.1}.
We first prepare the following lemma.
\begin{lemma}
\label{Lemma:3.1}
Assume the same conditions as in Theorem~{\rm\ref{Theorem:1.1}}.
Then
$$
t^{\frac{N}{\theta}\left(1-\frac{1}{r}\right)}(t+1)^{\frac{K-\ell}{\theta}}|||f(t)|||_{r,\ell}\le E_{K,q}[f](t),
\quad t>0,
$$
where $0\le\ell\le K$ and $1\le r\le q$.
\end{lemma}
{\bf Proof.}
Let $0\le\ell\le K$ and $1\le q\le\infty$.
It follows that
$$
(t+1)^{-\frac{\ell}{\theta}}|x|^\ell\le C+C(t+1)^{-\frac{K}{\theta}}|x|^K,
\quad (x,t)\in{\mathbb R}^N\times(0,\infty).
$$
This together with \eqref{eq:1.5} implies that
$$
\begin{aligned}
(t+1)^{-\frac{\ell}{\theta}}|||f(t)|||_{r,\ell}
& \le C\|f(t)\|_r+C(t+1)^{-\frac{K}{\theta}}|||f(t)|||_{r,K}\\
& \le C\|f(t)\|_1^\lambda\|f(t)\|_q^{1-\lambda}
+C(t+1)^{-\frac{K}{\theta}}|||f(t)|||_{1,K}^\lambda|||f(t)|||_{q,K}^{1-\lambda}\\
& \le Ct^{-\frac{N}{\theta}\left(1-\frac{1}{r}\right)}(t+1)^{-\frac{K}{\theta}}E_{K,q}[f](t),
\quad t>0,
\end{aligned}
$$
where $1/r=\lambda+(1-\lambda)/q$.
Thus Lemma~\ref{Lemma:3.1} follows.
$\Box$
\vspace{5pt}
Next we prove a lemma on the integral kernel ${\mathcal S}^m_\ell(x,y,t)$.
\begin{lemma}
\label{Lemma:3.2}
Let $m\in{\mathbb N_0}$, $0\le\ell \le K$, $1\le q\le\infty$, and $j=0,1$.
\begin{itemize}
\item[{\rm (a)}]
There exists $C_1>0$ such that
$$
|||\nabla^j{\mathcal S}^m_\ell(\cdot,y,t)|||_{q,\ell}
\le C_1t^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)-m-\frac{j}{\theta}}|y|^\ell,
\quad (y,t)\in{\mathbb R}^N\times(0,\infty).
$$
\item[{\rm (b)}]
There exists $C_2>0$ such that
$$
\left|\left|\left|\,
\int_{\mathbb R^N}\nabla^j{\mathcal S}^m_K(\cdot,y,t)\varphi(y)\,dy\,\right|\right|\right|_{q,\ell}
\le C_2t^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)-m-\frac{K+j-\ell}{\theta}}
|||\varphi|||_{1,K},\quad t>0,
$$
for $\varphi\in L^1_K$.
\item[{\rm (c)}]
Let $\varphi\in L^1_K$. Then
$$
\lim_{t\to\infty}t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)+m+\frac{K+j-\ell}{\theta}}
\left|\left|\left|\,
\int_{\mathbb R^N}\nabla^j{\mathcal S}^m_K(\cdot,y,t)\varphi(y)\,dy\,
\right|\right|\right|_{q,\ell}=0.
$$
\end{itemize}
\end{lemma}
{\bf Proof.}
Let $0\le\ell\le K$, $1\le q\le\infty$, and $j=0,1$.
We prove assertion~(a).
Let $x$, $y\in{\mathbb R}^N$ and $t>0$.
It follows that
$$
|x-\tau y| \ge |x|-|y| \ge |x|/2
\quad\mbox{if}\quad
|x| \ge 2|y|\quad\mbox{and}\quad
0 \le \tau \le 1.
$$
Then, by \eqref{eq:1.10} and \eqref{eq:2.1} we have
\begin{equation*}
\begin{aligned}
&
|x|^\ell |\nabla^j{\mathcal S}^m_\ell(x,y,t)| \\
& \le C\int_0^1|x|^\ell
\left|(\partial_t^m\nabla^{[\ell]+j+1} G_\theta)(x-\tau y,t) \right| |y|^{[\ell]+1}\,d\tau \\
& \le C|y|^\ell \int_0^1|x|^{[\ell]+1} \ t^{-\frac{N}{\theta}-\frac{[\ell]+j+1}{\theta}-m}
\left(1+t^{-\frac{1}{\theta}}|x-\tau y| \right)^{-(N+\theta m'+[\ell]+j+1)}\,d\tau\\
& \le C|y|^\ell (t^{-\frac{1}{\theta}}|x|)^{[\ell]+1} \ t^{-\frac{N}{\theta}-m-\frac{j}{\theta}}
\left(1+t^{-\frac{1}{\theta}}\frac{|x|}{2} \right)^{-(N+\theta m'+[\ell]+j+1)}\\
& \le C|y|^\ell t^{-\frac{N}{\theta}-m-\frac{j}{\theta}}
\left(1+t^{-\frac{1}{\theta}}\frac{|x|}{2} \right)^{-(N+\theta m'+j)}
\end{aligned}
\end{equation*}
if $|x|\ge 2|y|$.
Similarly, by \eqref{eq:1.10} we have
\begin{equation*}
\begin{aligned}
|x|^\ell |\nabla^j{\mathcal S}^m_\ell(x,y,t)|
& \le |x|^\ell |(\partial_t^m \nabla^jG_\theta)(x-y,t)|
+C\sum_{|\alpha|\le\ell}|x|^\ell |(\partial_t^m\partial_x^\alpha \nabla^jG_\theta)(x,t)||y|^{|\alpha|}\\
& \le (2|y|)^\ell |(\partial_t^m \nabla^jG_\theta)(x-y,t)|
+C\sum_{|\alpha|\le\ell} |y|^\ell |x|^{|\alpha|}
|(\partial_t^m\partial_x^\alpha \nabla^jG_\theta)(x,t)|
\end{aligned}
\end{equation*}
if $|x|<2|y|$.
These together with \eqref{eq:2.2} imply that
$$
|||\nabla^j{\mathcal S}^m_\ell(\cdot,y,t)|||_{q,\ell}
\le Ct^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)-m-\frac{j}{\theta}}|y|^\ell.
$$
Thus assertion~(a) follows.
We prove assertions~(b) and (c).
Let $\varphi\in L^1_K$, $0\le\ell\le K$, and $R>0$.
It follows from \eqref{eq:1.10} that
\begin{equation*}
\begin{aligned}
& \biggr|\biggr|\biggr|\,\int_{\{|y|\ge R^{\frac{1}{\theta}}\}}
|\nabla^j{\mathcal S}^m_K(\cdot,y,t)||\varphi(y)|\,dy\,\biggr|\biggr|\biggr|_{q,\ell}\\
& \le\,\biggr|\biggr|\biggr|\,\int_{\{|y|\ge R^{\frac{1}{\theta}}\}}
|\nabla^j{\mathcal S}^m_\ell(\cdot,y,t)||\varphi(y)|\,dy\,\biggr|\biggr|\biggr|_{q,\ell}\\
& \qquad
+C\sum_{\ell<|\alpha|\le K}\biggr|\biggr|\biggr|\,\int_{\{|y|\ge R^{\frac{1}{\theta}}\}}
|(\partial_t^m\partial_x^\alpha \nabla^jG_\theta)(\cdot,t)||y|^{|\alpha|}|\varphi(y)|\,dy\,\biggr|\biggr|\biggr|_{q,\ell}\\
& \le \int_{\{|y|\ge R^{\frac{1}{\theta}}\}}
|||\nabla^j{\mathcal S}^m_\ell(\cdot,y,t)|||_{q,\ell}|\varphi(y)|\,dy\\
& \qquad
+C\sum_{\ell<|\alpha|\le K}\int_{\{|y|\ge R^{\frac{1}{\theta}}\}}
|||(\partial_t^m\partial_x^\alpha \nabla^jG_\theta)(\cdot,t)|||_{q,\ell}|y|^{|\alpha|}|\varphi(y)|\,dy.
\end{aligned}
\end{equation*}
This together with \eqref{eq:2.2} and assertion~(a) implies that
\begin{equation}
\label{eq:3.1}
\begin{aligned}
& t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)+m+\frac{j}{\theta}}\biggr|\biggr|\biggr|\,\int_{\{|y|\ge R^{\frac{1}{\theta}}\}}
|\nabla^j{\mathcal S}^m_K(\cdot,y,t)||\varphi(y)|\,dy\,\biggr|\biggr|\biggr|_{q,\ell}\\
& \le C\bigg(\int_{\{|y|\ge R^{\frac{1}{\theta}}\}}|y|^\ell|\varphi(y)|\,dy
+\sum_{\ell<|\alpha|\le K}t^{-\frac{|\alpha|-\ell}{\theta}}
\int_{\{|y|\ge R^{\frac{1}{\theta}}\}}|y|^{|\alpha|}|\varphi(y)|\,dy\bigg)\\
& \le C\bigg(\int_{\{|y|\ge R^{\frac{1}{\theta}}\}}|y|^\ell
\left(\frac{|y|}{R^{\frac{1}{\theta}}}\right)^{K-\ell}|\varphi(y)|\,dy\\
& \qquad\qquad\qquad\qquad\qquad
+\sum_{\ell<|\alpha|\le K}t^{-\frac{|\alpha|-\ell}{\theta}}
\int_{\{|y|\ge R^{\frac{1}{\theta}}\}}
\left(\frac{|y|}{R^{\frac{1}{\theta}}}\right)^{K-|\alpha|}|y|^{|\alpha|}|\varphi(y)|\,dy\bigg)\\
& =Ct^{-\frac{K-\ell}{\theta}}
\biggr((R^{-1}t)^{\frac{K-\ell}{\theta}}
+\sum_{\ell<|\alpha|\le K}(R^{-1}t)^{\frac{K-|\alpha|}{\theta}}\biggr)
\int_{\{|y|\ge R^{\frac{1}{\theta}}\}}|y|^K|\varphi(y)|\,dy.
\end{aligned}
\end{equation}
Similarly, by \eqref{eq:1.10} we have
\begin{equation}
\label{eq:3.2}
\begin{split}
& \biggr|\biggr|\biggr|\,\int_{\{|y|<R^{\frac{1}{\theta}}\}}
|\nabla^j{\mathcal S}^m_K(\cdot,y,t)||\varphi(y)|\,dy\,\biggr|\biggr|\biggr|_{q,\ell}\\
& \le C\biggr|\biggr|\biggr|\,\int_{\{|y|<R^{\frac{1}{\theta}}\}}
\int_0^1|(\partial_t^m\nabla^{[K]+j+1}G_\theta)(\cdot-\tau y,t)||y|^{[K]+1}|\varphi(y)|\,d\tau\,dy\,\biggr|\biggr|\biggr|_{q,\ell}\\
& \le C\int_{\{|y|<R^{\frac{1}{\theta}}\}}
\int_0^1 |||(\partial_t^m\nabla^{[K]+j+1}G_\theta)(\cdot-\tau y,t)|||_{q,\ell}|y|^{[K]+1}|\varphi(y)|\,d\tau\,dy.
\end{split}
\end{equation}
On the other hand, it follows that
\begin{equation*}
\begin{split}
|x|^\ell |(\partial_t^m\nabla^{[K]+j+1}G_\theta)(x-\tau y,t)|
& =|z+\tau y|^\ell |(\partial_t^m\nabla^{[K]+j+1}G_\theta)(z,t)|\\
& \le C(|z|^\ell+|y|^\ell)|(\partial_t^m\nabla^{[K]+j+1}G_\theta)(z,t)|
\end{split}
\end{equation*}
for $x$, $y\in{\mathbb R}^N$, $t>0$, and $\tau\in(0,1)$, where $z:=x-\tau y$.
This together with \eqref{eq:2.2} implies that
\begin{equation}
\label{eq:3.3}
\begin{split}
& |||(\partial_t^m\nabla^{[K]+j+1}G_\theta)(\cdot-\tau y,t)|||_{q,\ell}\\
& \le|||(\partial_t^m\nabla^{[K]+j+1}G_\theta)(t)|||_{q,\ell}
+|y|^\ell|||(\partial_t^m\nabla^{[K]+j+1}G_\theta)(t)|||_q\\
& \le Ct^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)-m-\frac{[K]+j+1-\ell}{\theta}}
+Ct^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)-m-\frac{[K]+j+1}{\theta}}|y|^\ell
\end{split}
\end{equation}
for $y\in{\mathbb R}^N$, $t>0$, and $\tau\in(0,1)$.
By \eqref{eq:3.2} and \eqref{eq:3.3} we obtain
\begin{equation}
\label{eq:3.4}
\begin{aligned}
& t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)+m+\frac{j}{\theta}}\biggr|\biggr|\biggr|\,\int_{\{|y|<R^{\frac{1}{\theta}}\}}
|\nabla^j{\mathcal S}^m_K(\cdot,y,t)||\varphi(y)|\,dy\,\biggr|\biggr|\biggr|_{q,\ell}\\
& \le C\int_{\{|y|<R^{\frac{1}{\theta}}\}}
(t^{-\frac{[K]+1-\ell}{\theta}}+t^{-\frac{[K]+1}{\theta}} |y|^{\ell})|y|^{[K]+1} |\varphi(y)|\,dy\\
& \le C\left(t^{-\frac{[K]+1-\ell}{\theta}}R^{\frac{[K]+1-K}{\theta}}
+t^{-\frac{[K]+1}{\theta}}R^{\frac{[K]+\ell+1-K}{\theta}}\right)
\int_{\{|y|<R^{\frac{1}{\theta}}\}}|y|^K |\varphi(y)|\,dy.
\end{aligned}
\end{equation}
Combining \eqref{eq:3.1} and \eqref{eq:3.4} and setting $R=t$,
we obtain
$$
t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)+m+\frac{j}{\theta}}\left|\left|\left|
\int_{{\mathbb R}^N}\nabla^j{\mathcal S}^m_K(\cdot,y,t)\varphi(y)\,dy\right|\right|\right|_{q,\ell}
\le Ct^{-\frac{K-\ell}{\theta}}|||\varphi|||_{1,K},\quad t>0,
$$
which implies assertion~(b).
Similarly, setting $R=\varepsilon t$ with $0<\varepsilon\le 1$,
we have
\begin{align*}
& t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)+m+\frac{j}{\theta}}\left|\left|\left|
\int_{{\mathbb R}^N}\nabla^j{\mathcal S}^m_K(\cdot,y,t)\varphi(y)\,dy\right|\right|\right|_{q,\ell}\\
& \le Ct^{-\frac{K-\ell}{\theta}}
\biggr((\varepsilon^{-1})^{\frac{K-\ell}{\theta}}
+\sum_{\ell<|\alpha|\le K}(\varepsilon^{-1})^{\frac{K-|\alpha|}{\theta}}\biggr)
\int_{\{|y| \ge(\varepsilon t)^{\frac{1}{\theta}}\}}|y|^K|\varphi(y)|\,dy\\
&
\qquad
+Ct^{-\frac{K-\ell}{\theta}}\left(\varepsilon^{\frac{[K]+1-K}{\theta}}
+\varepsilon^{\frac{[K]+\ell+1-K}{\theta}}\right)|||\varphi|||_{1,K}.
\end{align*}
This together with $\varphi\in L^1_K$ implies that
\begin{align*}
& \limsup_{t\to\infty}\,t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)+m+\frac{K+j-\ell}{\theta}}
\biggr|\biggr|\biggr|\,\int_{{\mathbb R}^N}\nabla^j{\mathcal S}^m_K(\cdot,y,t)\varphi(y)\,dy\,
\biggr|\biggr|\biggr|_{q,\ell}\\
& \le C\left(\varepsilon^{\frac{[K]+1-K}{\theta}}+\varepsilon^{\frac{[K]+\ell+1-K}{\theta}}\right)|||\varphi|||_{1,K}.
\end{align*}
Since $\varepsilon$ is arbitrary,
we obtain assertion~(c).
Thus Lemma~\ref{Lemma:3.2} follows.
$\Box$
\vspace{5pt}
By Lemmata~\ref{Lemma:3.1} and \ref{Lemma:3.2} we have:
\begin{lemma}
\label{Lemma:3.3}
Let $f$ be a measurable function in $\mathbb R^N\times(0,\infty)$.
Assume \eqref{eq:1.5} for some $K\ge 0$ and $1\le q\le\infty$.
Let $0\le\ell\le K$, $m\in{\mathbb N}_0$, and $j=0,1$.
Then there exists $C>0$ such that
\begin{equation}
\label{eq:3.5}
\begin{split}
& (t+1)^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)+\frac{K+j-\ell}{\theta}}
\biggr|\biggr|\biggr|\,\int_T^t\int_{{\mathbb R}^N}(s+1)^m\nabla^j{\mathcal S}^m_K(\cdot,y,t+1)f(y,s)\,dy\,ds\,
\biggr|\biggr|\biggr|_{q,\ell}\\
& \le C\int_T^t E_{K,q}[f](s)\,ds
\end{split}
\end{equation}
for $t>T\ge 0$. Furthermore,
$$
\lim_{t\to\infty}
t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)+\frac{K+j-\ell}{\theta}}
\biggr|\biggr|\biggr|\,\int_0^T\int_{{\mathbb R}^N}(s+1)^m\nabla^j{\mathcal S}^m_K(\cdot,y,t+1)f(y,s)\,dy\,ds\,
\biggr|\biggr|\biggr|_{q,\ell}=0
$$
for $T> 0$.
\end{lemma}
{\bf Proof.}
It follows from Lemma~\ref{Lemma:3.1} that
\begin{equation}
\label{eq:3.6}
|||f(s)|||_{1,K}\le CE_{K,q}[f](s),
\quad s>0.
\end{equation}
This together with Lemma~\ref{Lemma:3.2}~(c) implies that, for any $T>0$,
\begin{equation}
\label{eq:3.7}
\lim_{t\to\infty}
t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)+\frac{K+j-\ell}{\theta}}(s+1)^m
\left|\left|\left|\,
\int_{\mathbb R^N}\nabla^j{\mathcal S}^m_K(\cdot,y,t+1)f(y,s)\,dy\,\right|\right|\right|_{q,\ell}=0
\end{equation}
for $0<s<T$.
Furthermore, by \eqref{eq:3.6} with Lemma~\ref{Lemma:3.2}~(b) we see that
\begin{equation}
\label{eq:3.8}
\begin{split}
& (t+1)^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)+\frac{K+j-\ell}{\theta}}(s+1)^m\left|\left|\left|\,
\int_{\mathbb R^N}\nabla^j{\mathcal S}^m_K(\cdot,y,t+1)f(y,s)\,dy\,\right|\right|\right|_{q,\ell}\\
& \le C(t+1)^{-m}(s+1)^m\|f(s)\|_{1,K}
\le CE_{K,q}[f](s)
\end{split}
\end{equation}
for $0<s<t$.
Inequality~\eqref{eq:3.8} implies \eqref{eq:3.5}.
Furthermore, by \eqref{eq:3.7} and \eqref{eq:3.8} we apply the Lebesgue dominated convergence theorem
to obtain
$$
\lim_{t\to\infty}
t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)+\frac{K+j-\ell}{\theta}}
\biggr|\biggr|\biggr|\,
\int_0^T\int_{{\mathbb R}^N}
(s+1)^m \nabla^j{\mathcal S}_K^m(\cdot,y,t+1)f(y,s)\,dy\,ds
\,\biggr|\biggr|\biggr|_{q,\ell}=0.
$$
Thus Lemma~\ref{Lemma:3.3} follows.
$\Box$\vspace{5pt}
Next we prove the following lemma on the integral kernel ${\mathcal T}(x,y,t,s)$.
\begin{lemma}
\label{Lemma:3.4}
Let $1\le q\le\infty$, and $0\le \ell\le K$.
\begin{itemize}
\item[{\rm (a)}]
Let $\varphi\in L^1_K$ with $K\ge 0$ and $j=0,1$. Then there exists $C_1>0$ such that
\begin{equation}
\label{eq:3.9}
t^{\frac{N}{q}\left(1-\frac{1}{q}\right)+\frac{j}{\theta}}(t+1)^{\frac{K-\ell}{\theta}}
\biggr|\biggr|\biggr|\,\int_{{\mathbb R}^N}\nabla^j{\mathcal T}(\cdot,y,t+1,1)\varphi(y)\,dy\,
\biggr|\biggr|\biggr|_{q,\ell}
\le C_1\|\varphi\|_{L^1_K}
\end{equation}
for $t>0$.
Furthermore,
\begin{equation}
\label{eq:3.10}
\lim_{t\to\infty}t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)+\frac{K+j-\ell}{\theta}}
\biggr|\biggr|\biggr|\,\int_{{\mathbb R}^N}\nabla^j{\mathcal T}(\cdot,y,t+1,1)\varphi(y)\,dy\,
\biggr|\biggr|\biggr|_{q,\ell}=0.
\end{equation}
\item[{\rm (b)}]
Let $f$ be a measurable function in $\mathbb R^N\times(0,\infty)$
and satisfy \eqref{eq:1.5}.
Let $j=0$ if $0<\theta\le 1$ and $j\in\{0,1\}$ if $1\le\theta<2$.
Then there exists $C_2>0$ such that
\begin{equation}
\label{eq:3.11}
\begin{aligned}
&
t^{\frac{N}{q}\left(1-\frac{1}{q}\right)}(t+1)^{\frac{K-\ell}{\theta}}
\biggr|\biggr|\biggr|\,\int_T^t\int_{{\mathbb R}^N}\nabla^j {\mathcal T}(\cdot,y,t+1,s+1)f(y,s)\,dy\,ds\,
\biggr|\biggr|\biggr|_{q,\ell}\\
& \le C_2\int_T^t (t-s)^{-\frac{j}{\theta}}E_{K,q}[f](s)\,ds
\end{aligned}
\end{equation}
for $t> T\ge0$.
Furthermore,
\begin{equation}
\label{eq:3.12}
\lim_{t\to\infty}t^{\frac{N}{q}\left(1-\frac{1}{q}\right)+\frac{K-\ell}{\theta}}
\biggr|\biggr|\biggr|\,\int_0^T\int_{{\mathbb R}^N}{\mathcal T}(\cdot,y,t+1,s+1)f(y,s)\,dy\,ds\,
\biggr|\biggr|\biggr|_{q,\ell}=0
\end{equation}
for $T> 0$.
\end{itemize}
\end{lemma}
{\bf Proof.}
Let $0\le \ell\le K$ and $j=0,1$.
We find $\ell'>0$ such that
\begin{equation}
\label{eq:3.13}
\ell\le \ell',
\qquad
\theta K_\theta<\ell'<\theta(K_\theta+1).
\end{equation}
Let $x$, $y\in{\mathbb R}^N$ and $t>0$.
It follows that
\begin{equation}
\label{eq:3.14}
t^{-\frac{\ell}{\theta}}|x|^\ell
\le t^{-\frac{\ell}{\theta}}(|x-y|^{\ell}+|y|^\ell)
\le C\left(1+t^{-\frac{\ell'}{\theta}}|x-y|^{\ell'}+t^{-\frac{K}{\theta}}|y|^K\right).
\end{equation}
This together with \eqref{eq:1.10} implies that
\begin{equation}
\label{eq:3.15}
\begin{aligned}
&
t^{-\frac{\ell}{\theta}}|x|^\ell |\nabla^j{\mathcal T}(x,y,t,s)|\\
&
\le Ct^{-\frac{\ell'}{\theta}}s^{K_\theta+1}
\int_0^1|x-y|^{\ell'}|(\partial_t^{K_\theta+1}\nabla^jG_\theta)(x-y,t-\tau s)|\,d\tau
\\
&\quad
+C\bigg(1+t^{-\frac{K}{\theta}}|y|^K\bigg)
\left[|(\nabla^jG_\theta)(x-y,t-s)|+\sum_{m=0}^{K_\theta} s^m|(\partial_t^m\nabla^j G_\theta)(x-y,t)|\right]
\end{aligned}
\end{equation}
for $0<s<t$.
Let $\psi\in L^{r_1}_K\cap L^{r_2}_K$ with $1\le r_1, r_2\le q$. Let $1\le r_i'\le \infty$ $(i=1,2)$ be such that
$$
\frac{1}{q}=\frac{1}{r_i}+\frac{1}{r_i'}-1.
$$
Then we observe from the Young inequality, \eqref{eq:2.2} and \eqref{eq:3.15} that
\begin{equation}
\label{eq:3.16}
\begin{split}
&
t^{-\frac{\ell}{\theta}}
\biggr|\biggr|\biggr|\int_{{\mathbb R}^N}\nabla^j{\mathcal T}(\cdot,y,t,s)\psi(y)\,dy\,
\biggr|\biggr|\biggr|_{q,\ell}\\
&
\le Ct^{-\frac{\ell'}{\theta}}s^{K_\theta+1}
\int_0^1|||\partial_t^{K_\theta+1}\nabla^jG_\theta(t-\tau s)|||_{r'_1,\ell'}\|\psi\|_{r_1}\,d\tau
\\
&\qquad
+C\|\nabla^jG_\theta(t-s)\|_{r'_1}\|\varphi\|_{r_1}
+Ct^{-\frac{K}{\theta}}\|\nabla^jG_\theta(t-s)\|_{r'_1}|||\varphi|||_{r_1,K}\\
& \qquad\qquad
+C\sum_{m=0}^{K_\theta} s^m\bigg[\|\partial_t^m \nabla^jG_\theta(t)\|_{r'_2}\|\varphi\|_{r_2}
+t^{-\frac{K}{\theta}}\|\partial_t^m\nabla^j G_\theta(t)\|_{r'_2}|||\varphi|||_{r_2,K}\bigg]\\
&
\le Ct^{-\frac{\ell'}{\theta}}s^{K_\theta+1}\|\psi\|_{r_1}
\int_0^1(t-\tau s)^{-\frac{N}{\theta}\left(\frac{1}{r_1}-\frac{1}{q}\right)-(K_\theta+1)+\frac{\ell'-j}{\theta}}\,d\tau
\\
&\qquad
+C(t-s)^{-\frac{N}{\theta}\left(\frac{1}{r_1}-\frac{1}{q}\right)-\frac{j}{\theta}}
(\|\psi\|_{r_1}+t^{-\frac{K}{\theta}}|||\psi|||_{r_1,K})\\
& \qquad\qquad
+Ct^{-\frac{N}{\theta}\left(\frac{1}{r_2}-\frac{1}{q}\right)-\frac{j}{\theta}}
\bigg(1+\sum_{m=0}^{K_\theta} s^mt^{-m}\bigg)
(\|\psi\|_{r_2}+t^{-\frac{K}{\theta}}|||\psi|||_{r_2,K})
\end{split}
\end{equation}
for $t/2< s<t$.
On the other hand, it follows from \eqref{eq:3.13} that
\begin{equation*}
\begin{split}
\int_0^1(t-\tau s)^{-\frac{N}{\theta}\left(\frac{1}{r_1}-\frac{1}{q}\right)-(K_\theta+1)+\frac{\ell'-j}{\theta}}\,d\tau
& \le (t-s)^{-\frac{N}{\theta}\left(\frac{1}{r_1}-\frac{1}{q}\right)}
\int_0^1(t-\tau s)^{-(K_\theta+1)+\frac{\ell'-j}{\theta}}\,d\tau\\
& \le C(t-s)^{-\frac{N}{\theta}\left(\frac{1}{r_1}-\frac{1}{q}\right)-\frac{j}{\theta}}s^{-1}t^{-K_\theta+\frac{\ell'}{\theta}}
\end{split}
\end{equation*}
for $0<s<t$.
This together with \eqref{eq:3.16} implies that
\begin{equation}
\label{eq:3.17}
\begin{split}
& t^{-\frac{\ell}{\theta}}
\biggr|\biggr|\biggr|\int_{{\mathbb R}^N}\nabla^j{\mathcal T}(\cdot,y,t,s)\psi(y)\,dy\,
\biggr|\biggr|\biggr|_{q,\ell}\\
&
\le C(t-s)^{-\frac{N}{\theta}\left(\frac{1}{r_1}-\frac{1}{q}\right)-\frac{j}{\theta}}
(\|\psi\|_{r_1}+t^{-\frac{K}{\theta}}|||\psi|||_{r_1,K})\\
& \qquad\qquad
+Ct^{-\frac{N}{\theta}\left(\frac{1}{r_2}-\frac{1}{q}\right)}(t-s)^{-\frac{j}{\theta}}
(\|\psi\|_{r_2}+t^{-\frac{K}{\theta}}|||\psi|||_{r_2,K})
\end{split}
\end{equation}
for $t/2< s<t$.
Similarly,
by \eqref{eq:1.10} and \eqref{eq:3.14}
we have
\begin{equation}
\label{eq:3.18}
\begin{aligned}
t^{-\frac{\ell}{\theta}}|x|^\ell |\nabla^j{\mathcal T}(x,y,t,s)
& |\le Cs^{K_\theta+1}\left(1+t^{-\frac{\ell'}{\theta}}|x-y|^{\ell'}+t^{-\frac{K}{\theta}}|y|^K\right)\\
& \quad
\times\int_0^1|(\partial_t^{K_\theta+1}\nabla^jG_\theta)(x-y,t-\tau s)|\,d\tau
\end{aligned}
\end{equation}
for $0<s<t$.
It follows from the Young inequality, \eqref{eq:2.2}, and \eqref{eq:3.18} that
\begin{equation}
\label{eq:3.19}
\begin{split}
&
t^{-\frac{\ell}{\theta}}
\biggr|\biggr|\biggr|\int_{{\mathbb R}^N}\nabla^j{\mathcal T}(\cdot,y,t,s)\psi(y)\,dy\,
\biggr|\biggr|\biggr|_{q,\ell}\\
&
\le Cs^{K_\theta+1}
\int_0^1\|(\partial_t^{K_\theta+1}\nabla^j G_\theta)(t-\tau s)\|_{r'_1}\|\psi\|_{r_1}\,d\tau\\
& \qquad
+Cs^{K_\theta+1}t^{-\frac{\ell'}{\theta}}
\int_0^1|||(\partial_t^{K_\theta+1}\nabla^j G_\theta)(t-\tau s)|||_{r'_1,\ell'}\|\psi\|_{r_1}\,d\tau
\\
& \qquad\qquad
+Cs^{K_\theta+1}t^{-\frac{K}{\theta}}
\int_0^1\|(\partial_t^{K_\theta+1}\nabla^j G_\theta)(t-\tau s)\|_{r'_1}|||\psi|||_{r_1,K}\,d\tau\\
&
\le Cs^{K_\theta+1}\|\psi\|_{r_1}
\int_0^1(t-\tau s)^{-\frac{N}{\theta}\left(\frac{1}{r_1}-\frac{1}{q}\right)-(K_\theta+1)-\frac{j}{\theta}}\,d\tau\\
& \qquad
+Cs^{K_\theta+1}t^{-\frac{\ell'}{\theta}}\|\psi\|_{r_1}
\int_0^1(t-\tau s)^{-\frac{N}{\theta}\left(\frac{1}{r_1}-\frac{1}{q}\right)-(K_\theta+1)+\frac{\ell'-j}{\theta}}\,
d\tau\\
& \qquad\qquad
+Cs^{K_\theta+1}t^{-\frac{K}{\theta}}|||\psi|||_{r_1,K}
\int_0^1(t-\tau s)^{-\frac{N}{\theta}\left(\frac{1}{r_1}-\frac{1}{q}\right)-(K_\theta+1)-\frac{j}{\theta}}\,d\tau\\
& \le Cs^{K_\theta+1}t^{-\frac{N}{\theta}\left(\frac{1}{r_1}-\frac{1}{q}\right)-(K_\theta+1)}
(t-s)^{-\frac{j}{\theta}}
\bigg\{\|\psi\|_{r_1}+t^{-\frac{K}{\theta}}|||\psi|||_{r_1,K}\bigg\}\\
\end{split}
\end{equation}
for $0<s\le t/2$.
We prove assertion~(a).
Since $(t+1)/2\le 1$ for $0<t\le 1$,
by \eqref{eq:3.17} with $\psi=\varphi$ and $r_1=r_2=1$ we have
\begin{equation}
\label{eq:3.20}
\begin{split}
&
(t+1)^{-\frac{\ell}{\theta}}
\biggr|\biggr|\biggr|\int_{{\mathbb R}^N}\nabla^j{\mathcal T}(\cdot,y,t+1,1)\varphi(y)\,dy\,
\biggr|\biggr|\biggr|_{q,\ell}\\
&
\le Ct^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{j}{\theta}}
(\|\varphi\|_1+(t+1)^{-\frac{K}{\theta}}|||\varphi|||_{1,K})\\
&
\le Ct^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{j}{\theta}}\|\varphi\|_{L^1_K}
\end{split}
\end{equation}
for $0<t\le 1$.
On the other hand, since $(t+1)/2>1$ for $t>1$
by \eqref{eq:3.19} with $\psi=\varphi$ and $r_1=1$ we see that
\begin{equation*}
\begin{split}
& (t+1)^{-\frac{\ell}{\theta}}
\biggr|\biggr|\biggr|\int_{{\mathbb R}^N}\nabla^j{\mathcal T}(\cdot,y,t+1,1)\varphi(y)\,dy\,
\biggr|\biggr|\biggr|_{q,\ell}\\
& \le C(t+1)^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)-(K_\theta+1)}t^{-\frac{j}{\theta}}
\bigg\{\|\varphi\|_1+(t+1)^{-\frac{K}{\theta}}|||\varphi|||_{1,K}\bigg\}\\
& \le Ct^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{j}{\theta}}(t+1)^{-(K_\theta+1)}\|\varphi\|_{L^1_K}
\end{split}
\end{equation*}
for $t>1$.
This together with
$\theta(K_\theta+1)>K$ implies \eqref{eq:3.10} and
\begin{equation}
\label{eq:3.21}
t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)+\frac{j}{\theta}}(t+1)^{\frac{K-\ell}{\theta}}
\biggr|\biggr|\biggr|\int_{{\mathbb R}^N}{\mathcal T}(\cdot,y,t+1,1)\varphi(y)\,dy\,\biggr|\biggr|\biggr|_{q,\ell}
\le C\|\varphi\|_{L^1_K}
\end{equation}
for $t> 1$.
Combining \eqref{eq:3.20} and \eqref{eq:3.21}, we obtain \eqref{eq:3.9} and \eqref{eq:3.10}.
Thus assertion~(a) follows.
We prove assertion~(b).
Since $(t+1)/2<s+1$ for $t/2<s<t$,
by \eqref{eq:3.13} and \eqref{eq:3.17} with $\psi=f(s)$ and $(r_1,r_2)=(q,1)$ we have
$$
\begin{aligned}
&
(t+1)^{-\frac{\ell}{\theta}}
\biggr|\biggr|\biggr|\int_{{\mathbb R}^N}\nabla^j{\mathcal T}(\cdot,y,t+1,s+1)f(y,s)\,dy\,
\biggr|\biggr|\biggr|_{q,\ell}\\
&
\le C(t-s)^{-\frac{j}{\theta}}
(\|f(s)\|_q+(t+1)^{-\frac{K}{\theta}}|||f(s)|||_{q,K})\\
& \qquad
+Ct^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)}(t-s)^{-\frac{j}{\theta}}
(\|f(s)\|_1+(t+1)^{-\frac{K}{\theta}}|||f(s)|||_{1,K})
\end{aligned}
$$
for $t/2<s<t$.
This together with Lemma~\ref{Lemma:3.1} implies that
\begin{equation}
\label{eq:3.22}
\begin{split}
&
(t+1)^{-\frac{\ell}{\theta}}
\biggr|\biggr|\biggr|\int_{{\mathbb R}^N}\nabla^j{\mathcal T}(\cdot,y,t+1,s+1)f(y,s)\,dy\,\biggr|\biggr|\biggr|_{q,\ell}\\
&
\le C\left(s^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)}+t^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)}\right)
\left((s+1)^{-\frac{K}{\theta}}+(t+1)^{-\frac{K}{\theta}}\right)(t-s)^{-\frac{j}{\theta}}E_{K,q}[f](s)\\
&
\le Ct^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)}(t+1)^{-\frac{K}{\theta}}(t-s)^{-\frac{j}{\theta}}E_{K,q}[f](s)
\end{split}
\end{equation}
for $t/2<s<t$.
On the other hand,
by \eqref{eq:3.13} and \eqref{eq:3.17} with $\psi=f(s)$ and $r_1=r_2=1$ we have
$$
\begin{aligned}
&
(t+1)^{-\frac{\ell}{\theta}}
\biggr|\biggr|\biggr|\int_{{\mathbb R}^N}\nabla^j{\mathcal T}(\cdot,y,t+1,s+1)f(y,s)\,dy\,
\biggr|\biggr|\biggr|_{q,\ell}\\
&
\le Ct^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)}(t-s)^{-\frac{j}{\theta}}
(\|f(s)\|_1+(t+1)^{-\frac{K}{\theta}}|||f(s)|||_{1,K})
\end{aligned}
$$
for $0<s\le t/2$ with $(t+1)/2< s+1$.
This together with Lemma~\ref{Lemma:3.1} implies that
\begin{equation}
\label{eq:3.23}
\begin{aligned}
&
(t+1)^{-\frac{\ell}{\theta}}
\biggr|\biggr|\biggr|\int_{{\mathbb R}^N}\nabla^j{\mathcal T}(\cdot,y,t+1,s+1)f(y,s)\,dy\,
\biggr|\biggr|\biggr|_{q,\ell}\\
&
\le Ct^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)}
\left((s+1)^{-\frac{K}{\theta}}+(t+1)^{-\frac{K}{\theta}}\right)
(t-s)^{-\frac{j}{\theta}}E_{K,q}[f](s)
\\
&
\le Ct^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)}(t+1)^{-\frac{K}{\theta}}
(t-s)^{-\frac{j}{\theta}}E_{K,q}[f](s)
\end{aligned}
\end{equation}
for $0<s\le t/2$ with $(t+1)/2< s+1$.
Furthermore,
by \eqref{eq:3.19} with $\psi=f(s)$ and $r_1=1$
we have
$$
\begin{aligned}
&
(t+1)^{-\frac{\ell}{\theta}}
\biggr|\biggr|\biggr|\int_{{\mathbb R}^N}\nabla^j{\mathcal T}(\cdot,y,t+1,s+1)f(y,s)\,dy\,
\biggr|\biggr|\biggr|_{q,\ell}\\
&
\le C(s+1)^{K_\theta+1}
(t+1)^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)-(K_\theta+1)}(t-s)^{-\frac{j}{\theta}}
\times
\\
&\hspace{5cm}
\times
\bigg\{\|f(s)\|_1+(t+1)^{-\frac{K}{\theta}}|||f(s)|||_{1,K}\bigg\}
\end{aligned}
$$
for $0<s\le t/2$ with $(t+1)/2\ge s+1$.
This together with Lemma~\ref{Lemma:3.1} again implies that
\begin{equation}
\label{eq:3.24}
\begin{aligned}
&
(t+1)^{-\frac{\ell}{\theta}}
\biggr|\biggr|\biggr|\int_{{\mathbb R}^N}\nabla^j{\mathcal T}(\cdot,y,t+1,s+1)f(y,s)\,dy\,
\biggr|\biggr|\biggr|_{q,\ell}\\
&
\le Ct^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)}(s+1)^{K_\theta+1}
(t+1)^{-(K_\theta+1)}\left((s+1)^{-\frac{K}{\theta}}+(t+1)^{-\frac{K}{\theta}}\right)\times
\\
&\hspace{5cm}
\times
(t-s)^{-\frac{j}{\theta}}E_{K,q}[f](s)
\\
&
\le Ct^{-\frac{N}{\theta}\left(1-\frac{1}{q}\right)}(s+1)^{K_\theta+1-\frac{K}{\theta}}
(t+1)^{-(K_\theta+1)}(t-s)^{-\frac{j}{\theta}}E_{K,q}[f](s)
\end{aligned}
\end{equation}
for $0<s\le t/2$ with $(t+1)/2\ge s+1$.
Then, by \eqref{eq:3.24}, for any $T>0$, we observe from $\theta(K_\theta+1)>K$ that
\begin{equation}
\label{eq:3.25}
\lim_{t\to\infty}t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)+\frac{j}{\theta}}(t+1)^{\frac{K-\ell}{\theta}}
\biggr|\biggr|\biggr|\int_{{\mathbb R}^N}\nabla^j{\mathcal T}(\cdot,y,t+1,s+1)f(y,s)\,dy\,\biggr|\biggr|\biggr|_{q,\ell}=0
\end{equation}
for $0<s<T$.
Furthermore, by \eqref{eq:3.22}, \eqref{eq:3.23}, and \eqref{eq:3.24} we see that
\begin{equation}
\label{eq:3.26}
\begin{aligned}
&
t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)}(t+1)^{\frac{K-\ell}{\theta}}
\biggr|\biggr|\biggr|\int_{{\mathbb R}^N}\nabla^j{\mathcal T}(\cdot,y,t+1,s+1)f(y,s)\,dy\,\biggr|\biggr|\biggr|_{q,\ell}
\\
&
\le C(t-s)^{-\frac{j}{\theta}}E_{K,q}[f](s)
\end{aligned}
\end{equation}
for $0<s< t$.
This implies \eqref{eq:3.11}.
Furthermore, by \eqref{eq:3.25} and \eqref{eq:3.26}
we apply the Lebesgue dominated convergence theorem
to obtain \eqref{eq:3.12}.
Thus assertion~(b) follows.
The proof of Lemma~\ref{Lemma:3.4} is complete.
$\Box$\vspace{5pt}
Now we are ready to complete the proof of Theorem~\ref{Theorem:1.1}.
\vspace{5pt}
\newline
{\bf Proof of Theorem~\ref{Theorem:1.1}.}
Let $u$ and $w$ be as in Theorem~\ref{Theorem:1.1}.
Then, by \eqref{eq:1.11} and \eqref{eq:1.12} we have
$$
\begin{aligned}
& u(x,t)-w(x,t)\\
& =\int_{\mathbb R^N}{\mathcal T}(x,y,t+1,1)\varphi(y)\, dy
+\int_0^t\int_{{\mathbb R}^N} {\mathcal T}(x,y,t+1,s+1)f(y,s)\,dy\,ds\\
&
\qquad
+\sum_{m=0}^{K_\theta}\frac{(-1)^m}{m!}\int_{{\mathbb R}^N}{\mathcal S}^m_K(x,y,t+1)\varphi(y)\,dy\\
&
\qquad\qquad
+\sum_{m=0}^{K_\theta}\frac{(-1)^m}{m!}\int_0^t\int_{{\mathbb R}^N} (s+1)^m{\mathcal S}_K^m(x,y,t+1)f(y,s)\,dy\,ds.
\end{aligned}
$$
We apply Lemmata~\ref{Lemma:3.2}, \ref{Lemma:3.3}, and \ref{Lemma:3.4}
to obtain \eqref{eq:1.6} and \eqref{eq:1.8}.
Then we easily see that \eqref{eq:1.9} holds.
Thus Theorem~\ref{Theorem:1.1} follows.
$\Box$\vspace{3pt}
\newline
{\bf Proof of Corollary~\ref{Corollary:1.1}.}
By property (G)-(i) we see that
$g_{\alpha,m}(0)\in L^q_K$ is equivalent to $g_{\alpha,m}(t)\in L^q_K$ for $t\ge0$.
Then Corollary~\ref{Corollary:1.1} follows from Theorem~\ref{Theorem:1.1}.
$\Box$
\begin{remark}
\label{Remark:3.1}
{\rm(i)}
The arguments of \cites{IIK,IK,IKM} are in the frameworks of $L^q$ and $L^1_K$.
On the other hand, the arguments in the proof of Theorem~{\rm\ref{Theorem:1.1}}
are in the framework of $L^q_K$.
This improvement enables us to
obtain HOAE of solutions to the Cauchy problem
for nonlinear fractional diffusion equations such as \eqref{eq:1.2}.
See Section~{\rm 5}.
\vspace{3pt}
\newline
{\rm(ii)}
Let $0\le K<\theta$ and $\ell=0$.
By similar arguments to those in the proof of Theorem~$\ref{Theorem:1.1}$
we see that Theorem~{\rm\ref{Theorem:1.1}} holds with $E_{K,q}[f]$ replaced by
$$
E_{K,q}'[f](t)
:= (t+1)^{\frac{K}{\theta}}
\left[t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)}\|f(t)\|_q+\|f(t)\|_1\right]+|||f(t)|||_{1,K}.
$$
See also \cite{IKK02}*{Theorem~1.2}.
\end{remark}
\section{Fractional convection-diffusion equation}
In this section
we consider the Cauchy problem for a convection type inhomogeneous fractional diffusion equation
\begin{equation}
\label{eq:4.1}
\partial_t u+(-\Delta)^{\frac{\theta}{2}}u={\rm div} f(x,t)\quad\mbox{in}\quad{\mathbb R}^N\times(0,\infty),
\quad
u(x,0)=\varphi(x)\quad\mbox{in}\quad{\mathbb R}^N,
\end{equation}
where $1<\theta<2$, $\varphi\in L^1_K$ with $K\ge 0$, and $f=(f_1,\dots,f_N)$ is a vector-valued function
in ${\mathbb R}^N\times(0,\infty)$.
Similarly to Theorem~\ref{Theorem:1.1}, we have:
\begin{theorem}
\label{Theorem:4.1}
Let $N\ge1$, $1<\theta<2$, $K\ge 0$, and $1\le q\le\infty$.
Let $f=(f_1,\dots,f_N)$ be a vector-valued measurable function in $\mathbb R^N\times(0,\infty)$ satisfying \eqref{eq:1.5}.
Let $u\in C({\mathbb R}^N\times(0,\infty))$ be a solution to problem~\eqref{eq:4.1}, that is,
$u$ satisfies
$$
u(x,t)
=\int_{{\mathbb R}^N}G_\theta(x-y,t)\varphi(y)\,dy
+\int_0^t\int_{{\mathbb R}^N}\nabla G_\theta(x-y,t-s)\cdot f(y,s)\,dy\,ds
$$
for $(x,t)\in{\mathbb R}^N\times(0,\infty)$, where $\varphi\in L^1_K$.
Let $0\le \ell\le K$.
Then
\begin{equation}
\label{eq:4.2}
\sup_{0<t<\tau}\,t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)}|||u(t)-z(t)|||_{q,\ell}<\infty
\quad\mbox{for}\quad \tau>0,
\end{equation}
where
\begin{equation*}
\begin{split}
z(x,t) & :=
\sum_{m=0}^{K_\theta}\sum_{|\alpha|\le K}
M_\alpha(\varphi)\,g_{\alpha,m}(x,t)\\
& \qquad
+\sum_{m=0}^{K_\theta}\sum_{|\alpha|\le K}
\sum_{j=1}^N\left(\int_0^t(s+1)^m M_\alpha(f_j(s))\,ds\right)\partial_{x_j}g_{\alpha,m}(x,t).
\end{split}
\end{equation*}
Furthermore, there exists $C>0$ such that,
for any $\varepsilon>0$ and $T>0$,
\begin{equation}
\label{eq:4.3}
t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||u(t)-z(t)|||_{q,\ell}\\
\le\varepsilon t^{-\frac{K}{\theta}}
+Ct^{-\frac{K}{\theta}}\int_T^t(t-s)^{-\frac{1}{\theta}}E_{K,q}[f](s)\,ds
\end{equation}
holds for large enough $t>0$.
\end{theorem}
{\bf Proof of Theorem~\ref{Theorem:4.1}.}
Let $u$ and $z$ be as in Theorem~\ref{Theorem:4.1}.
Then, similarly to \eqref{eq:1.11} and \eqref{eq:1.12}, we have
$$
\begin{aligned}
&
u(x,t)-z(x,t)\\
&
=\int_{\mathbb R^N}{\mathcal T}(x,y,t+1,1)\varphi(y)\, dy
+\int_0^t\int_{{\mathbb R}^N} \nabla {\mathcal T}(x,y,t+1,s+1)\cdot f(y,s)\,dy\,ds\\
& \quad
+\sum_{m=0}^{K_\theta}\frac{(-1)^m}{m!}
\int_{{\mathbb R}^N}{\mathcal S}^m_K(x,y,t+1)\varphi(y)\,dy\\
& \quad
+\sum_{m=0}^{K_\theta}\frac{(-1)^m}{m!}\int_0^t\int_{{\mathbb R}^N}(s+1)^m\nabla {\mathcal S}_K^m(x,y,t+1)\cdot f(y,s)\,dy\,ds.
\end{aligned}
$$
Similarly to the proof of Theorem~\ref{Theorem:1.1},
we apply Lemmata~\ref{Lemma:3.2}, \ref{Lemma:3.3}, and \ref{Lemma:3.4}
to obtain \eqref{eq:4.2} and \eqref{eq:4.3}.
Thus Theorem~\ref{Theorem:4.1} follows.
$\Box$
\section{Nonlinear fractional diffusion equation}
Let $N\ge 1$, $0<\theta<2$, and $F\in C({\mathbb R}^N\times[0,\infty)\times{\mathbb R})$.
Consider the Cauchy problem
for a nonlinear fractional diffusion equation
\begin{equation}
\tag{P}
\partial_t u+(-\Delta)^{\frac{\theta}{2}}u=F(x,t,u)\quad\mbox{in}\quad{\mathbb R}^N\times(0,\infty),
\quad
u(x,0)=\varphi(x)\quad\mbox{in}\quad{\mathbb R}^N,
\end{equation}
where $\varphi\in L^1_K\cap L^\infty$ for some $K\ge 0$
under the following condition~(F):
\begin{itemize}
\item[(F)]
there exists $p>1+\theta/N$ such that
$$
|F(x,t,v)-F(x,t,w)|\le C(|v|+|w|)^{p-1}|v-w|
$$
for $(x,t,v,w)\in{\mathbb R}^N\times[0,\infty)\times{\mathbb R}^2$.
\end{itemize}
Let $u\in C({\mathbb R}^N\times(0,\infty))$ be a solution to problem~(P)
that is, $u$ satisfies
$$
u(x,t)=\left[e^{-t(-\Delta)^{\theta/2}}\varphi\right](x)
+\int_0^t \left[e^{-(t-s)(-\Delta)^{\theta/2}}F(\cdot,s,u(\cdot,s))\right](x)\,ds
$$
for $(x,t)\in{\mathbb R}^N\times(0,\infty)$.
In this section, under condition~\eqref{eq:1.13},
we obtain HOAE of the solution~$u$.
Theorem~\ref{Theorem:5.1} is an application of Theorem~\ref{Theorem:1.1}.
\begin{theorem}
\label{Theorem:5.1}
Let $N\ge 1$, $0<\theta<2$, and $\varphi\in L^1_K$ with $K\ge 0$.
Assume condition~{\rm (F)}.
Let $u\in C({\mathbb R}^N\times(0,\infty))$ be a solution to problem~{\rm (P)}.
Set
$$
F(x,t):=F(x,t,u(x,t)),\quad (x,t)\in{\mathbb R}^N\times(0,\infty).
$$
\begin{itemize}
\item[\rm (a)]
Assume that $\varphi\in L^\infty_k$ with $k=\min\{N+\theta,K\}$.
Let $u$ satisfy \eqref{eq:1.13}.
Then
\begin{equation}
\label{eq:5.1}
\sup_{t>0}\,(t+1)^{\frac{N}{q}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||u(t)|||_{q,\ell}<\infty
\end{equation}
for $1\le q\le\infty$ and $0\le\ell\le K$ with $\ell<\theta+N(1-1/q)$.
\item[\rm (b)]
Let $u$ satisfy \eqref{eq:5.1}.
If $p(N+\theta)>K+N$,
then
\begin{equation}
\label{eq:5.2}
E_{K,q}[F](t)\le C(t+1)^{\frac{K}{\theta}-A_p},\quad t>0,
\end{equation}
for $1\le q\le\infty$, where $A_p:=N(p-1)/\theta>1$.
\item[\rm(c)]
Assume that \eqref{eq:5.2} holds.
Set
\begin{equation}
\label{eq:5.3}
U_0(x,t):=\sum_{m=0}^{K_\theta}\sum_{|\alpha|\le K}
\left(M_\alpha(\varphi)+\int_0^t (s+1)^m M_\alpha(F(s))\,ds\right)\,g_{\alpha,m}(x,t).
\end{equation}
Then
\begin{equation}
\label{eq:5.4}
\sup_{t>0}\,t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)}(t+1)^{-\frac{\ell}{\theta}}
|||u(t)-U_0(t)|||_{q,\ell}<\infty
\end{equation}
and
\begin{equation}
\label{eq:5.5}
t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||u(t)-U_0(t)|||_{q,\ell}
=\left\{
\begin{array}{ll}
o(t^{-\frac{K}{\theta}})+O(t^{-A_p+1}) & \mbox{if}\quad A_p-1\not=K/\theta,\vspace{5pt}\\
O(t^{-\frac{K}{\theta}}\log t) & \mbox{if}\quad A_p-1=K/\theta,\vspace{3pt}
\end{array}
\right.
\end{equation}
as $t\to\infty$, for $1\le q\le\infty$ and $0\le \ell \le K$.
\end{itemize}
\end{theorem}
\begin{remark}
\label{Remark:5.1}
Let $N\ge 1$, $0<\theta<2$, and $\varphi\in L^\infty$.
Assume condition~{\rm (F)}.
%
\begin{itemize}
\item[\rm (i)]
There exists $\delta>0$ such that, if $\|\varphi\|_{L^{N(p-1)/\theta}}<\delta$,
then problem~{\rm (P)} possesses a solution~$u\in C({\mathbb R}^N\times(0,\infty))$
satisfying \eqref{eq:1.13}.
See \cites{IKK01, IKO, W}.
%
\item[\rm (ii)]
Let $F(x,t,u):=\lambda|u|^{p-1}u$ with $\lambda\le 0$.
Then the comparison principle implies that
$$
|u(x,t)|\le \left[e^{-t(-\Delta)^{\theta/2}}|\varphi|\right](x),\quad(x,t)\in{\mathbb R}^N\times(0,\infty).
$$
This together with Lemma~{\rm\ref{Lemma:2.1}} implies \eqref{eq:1.13}.
\end{itemize}
\end{remark}
We prepare the following lemma for the proof of Theorem~\ref{Theorem:5.1}.
\begin{lemma}
\label{Lemma:5.1}
Assume condition~{\rm (F)}. Let $K\ge 0$.
Let $v_1$ and $v_2$ be measurable functions in ${\mathbb R}^N\times(0,\infty)$
and $h$ in $(0,\infty)$ such that
\begin{equation}
\label{eq:5.6}
\begin{split}
& (t+1)^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||v_i(t)|||_{q,\ell}<\infty,\quad i=1,2,\\
& (t+1)^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||v_1(t)-v_2(t)|||_{q,\ell}\le h(t),
\end{split}
\end{equation}
for $t>0$, $1\le q\le\infty$, and $0\le\ell\le K$ with $\ell<\theta+N(1-1/q)$.
Assume that $p(N+\theta)>K+N$. Then
$$
E_{K,q}[F(v_1)-F(v_2)](t)\le C(t+1)^{-A_p+\frac{K}{\theta}}h(t),\quad t>0.
$$
\end{lemma}
{\bf Proof.}
Let $0\le\ell\le K$ and $1\le q\le\infty$.
Since $p(N+\theta)>K+N$, we find $\ell_1$, $\ell_2\ge 0$ such that
$$
0\le \ell_1<\theta+N,\quad 0\le \ell_2<\theta+N\left(1-\frac{1}{q}\right),
\quad \ell=(p-1)\ell_1+\ell_2.
$$
Then, by condition~(F) and \eqref{eq:5.6} we see that
$$
\begin{aligned}
& (t+1)^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||F(v_1(s))-F(v_2(s))|||_{q,\ell}\\
& \le C(t+1)^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}
(|||v_1(t)|||_{\infty,\ell_1}^{p-1}+|||v_2(t)|||_{\infty,\ell_1}^{p-1})|||v_1(t)-v_2(t)|||_{q,\ell_2}\\
& \le C(t+1)^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}-\frac{N(p-1)}{\theta}
+\frac{(p-1)\ell_1}{\theta}}|||v_1(t)-v_2(t)|||_{q,\ell_2}\\
& \le C(t+1)^{-\frac{\ell}{\theta}-\frac{N(p-1)}{\theta}+\frac{(p-1)\ell_1}{\theta}+\frac{\ell_2}{\theta}}h(t)
=C(t+1)^{-A_p}h(t),\quad t>0.
\end{aligned}
$$
Thus Lemma~\ref{Lemma:5.1} follows.
$\Box$\vspace{5pt}
\noindent
{\bf Proof of Theorem~\ref{Theorem:5.1}.}
We prove assertion~(a).
Since $A_p=N(p-1)/\theta>1$,
the comparison principle together with condition~(F) and \eqref{eq:1.13} implies that
$$
|u(x,t)|\le \exp\left(C\int_0^t(s+1)^{-A_p}\,ds\right)\left[e^{-t(-\Delta)^{\theta/2}}|\varphi|\right](x)
\le C\left[e^{-t(-\Delta)^{\theta/2}}|\varphi|\right](x)
$$
for $(x,t)\in{\mathbb R}^N\times(0,\infty)$.
This together with Lemma~\ref{Lemma:2.1} implies assertion~(a).
Furthermore, assertion~(b) follows from Lemma~\ref{Lemma:5.1} with $v_1=u$ and $v_2=0$.
On the other hand,
by Theorem~\ref{Theorem:1.1} with \eqref{eq:5.2} we obtain \eqref{eq:5.4}.
Furthermore,
for any $\varepsilon>0$ and $T>0$, we have
$$
t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||u(t)-U_0(t)|||_{q,\ell}
\le\varepsilon t^{-\frac{K}{\theta}}
+Ct^{-\frac{K}{\theta}}\int_T^t(s+1)^{\frac{K}{\theta}-A_p}\,ds
$$
for large enough $t>0$.
This implies \eqref{eq:5.5}. Thus assertion~(c) follows.
The proof of Theorem~\ref{Theorem:5.1} is complete.
$\Box$\vspace{5pt}
As a corollary of Theorem~\ref{Theorem:5.1}, we have:
\begin{Corollary}
\label{Corollary:5.1}
Let $N\ge 1$, $0<\theta<2$, and $\varphi\in L_K^1$ with $K\ge 0$.
Assume condition~{\rm (F)} and
\begin{equation}
\label{eq:5.7}
p>1+\frac{2K+\theta}{N}.
\end{equation}
Let $u\in C({\mathbb R}^N\times(0,\infty))$ be a solution to problem~{\rm (P)}
and satisfy \eqref{eq:5.1}.
Then there exists a set $\{M_{\alpha,m}\}\subset{\mathbb R}$,
where $m\in\{0,\dots,K_\theta\}$ and $\alpha\in{\mathbb M}$ with $|\alpha|\le K$, such that
\begin{equation}
\label{eq:5.8}
t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||u(t)-U_*(t)|||_{q,\ell}
=o\left(t^{-\frac{K}{\theta}}\right)
\quad\mbox{as}\quad t\to\infty
\end{equation}
for $1\le q\le\infty$ and $0\le \ell\le K$,
where
\begin{equation}
\label{eq:5.9}
U_*(x,t):=\sum_{m=0}^{K_\theta}\sum_{|\alpha|\le K}M_{\alpha,m}\,g_{\alpha,m}(x,t).
\end{equation}
\end{Corollary}
{\bf Proof.}
It follows from \eqref{eq:5.7} that
\begin{equation}
\label{eq:5.10}
p(N+\theta)>K+N,\qquad
A_p-1=\frac{N}{\theta}(p-1)-1>\frac{2K}{\theta}.
\end{equation}
By Theorem~\ref{Theorem:5.1} we have
\begin{equation}
\label{eq:5.11}
t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}
|||u(t)-U_0(t)|||_{q,\ell}=o\left(t^{-\frac{K}{\theta}}\right)
\quad\mbox{as}\quad t\to\infty
\end{equation}
for $1\le q\le\infty$ and $1\le \ell\le K$.
Here $U_0$ is as in Theorem~\ref{Theorem:5.1}.
Let $m\in\{0,\dots,K_\theta\}$ and $\alpha\in{\mathbb M}$ with $|\alpha|\le K$.
Assertion~(b) of Theorem~\ref{Theorem:5.1} implies that
\begin{equation}
\label{eq:5.12}
|M_\alpha(F(t))|\le C(t+1)^{-A_p+\frac{|\alpha|}{\theta}},\quad t>0.
\end{equation}
Then, by \eqref{eq:5.10} we find $M_{\alpha,m}\in {\mathbb R}$ such that
$$
M_{\alpha,0}=M_\alpha(\varphi)+\int_0^\infty M_{\alpha}(F(s))\,ds,
\quad
M_{\alpha,m}=\int_0^\infty (s+1)^mM_{\alpha}(F(s))\,ds\quad (m\ge1).
$$
Furthermore, by \eqref{eq:2.2}, \eqref{eq:5.3}, \eqref{eq:5.9}, \eqref{eq:5.10}, and \eqref{eq:5.12}
we have
$$
\begin{aligned}
& t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||U_0(t)-U_*(t)||_{q,\ell}\\
& \le t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}
\sum_{m=0}^{K_\theta}\sum_{|\alpha|\le K}
\left(\int_t^\infty (s+1)^m |M_\alpha(F(s))|\,ds\right)\,|||g_{\alpha,m}(t)|||_{q,\ell}\\
& \le C\sum_{m=0}^{K_\theta}\sum_{|\alpha|\le K}(t+1)^{-m-\frac{|\alpha|}{\theta}}
\int_t^\infty (s+1)^m(s+1)^{-A_p+\frac{|\alpha|}{\theta}}\,ds
\le Ct^{-A_p+1},\quad t\ge 1.
\end{aligned}
$$
This together with \eqref{eq:5.10} and \eqref{eq:5.11} implies \eqref{eq:5.8}.
Thus Corollary~\ref{Corollary:5.1} follows.
$\Box$\vspace{5pt}
Combining Theorems~\ref{Theorem:1.1} and \ref{Theorem:5.1},
we obtain a refined asymptotic expansion of the solution to problem~(P).
\begin{theorem}
\label{Theorem:5.2}
Assume the same conditions as in Theorem~$\ref{Theorem:5.1}$.
Let $u$ satisfy \eqref{eq:5.1}.
For $n=1,2,\dots$,
define a function $U_n=U_n(x,t)$ in ${\mathbb R}^N\times(0,\infty)$ inductively by
$$
\begin{aligned}
U_n(x,t) & :=U_0(x,t)+\int_0^t \left[e^{-(t-s)(-\Delta)^{\theta/2}}F_{n-1}(s)\right](x)\,ds\\
& \quad-\sum_{m=0}^{K_\theta}\sum_{|\alpha|\le K}
\left(\int_0^t (s+1)^m M_\alpha(F_{n-1}(s))\,ds\right)\,g_{\alpha,m}(x,t),
\end{aligned}
$$
where $U_0$ is as in Theorem~{\rm\ref{Theorem:5.1}} and $F_{n-1}(x,t):=F(x,t,U_{n-1}(x,t))$.
Then
\begin{equation}
\label{eq:5.13}
\sup_{t>0}\,t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}
|||u(t)-U_n(t)|||_{q,\ell}<\infty
\end{equation}
and
\begin{equation}
\label{eq:5.14}
\begin{aligned}
& t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||u(t)-U_n(t)|||_{q,\ell}\\
&\qquad
=\left\{
\begin{array}{ll}
o(t^{-\frac{K}{\theta}})+O(t^{-(n+1)(A_p-1)}) & \mbox{if}\quad (n+1)(A_p-1)\not=K/\theta,\vspace{5pt}\\
O(t^{-\frac{K}{\theta}}\log t) & \mbox{if}\quad (n+1)(A_p-1)=K/\theta,\\
\end{array}
\right.
\end{aligned}
\end{equation}
as $t\to\infty$, for $1\le q\le\infty$ and $0\le \ell\le K$.
\end{theorem}
{\bf Proof.}
Let $K\ge 0$.
By Theorem~\ref{Theorem:5.1} we have \eqref{eq:5.13} and \eqref{eq:5.14} with $n=0$.
Assume that \eqref{eq:5.13} and \eqref{eq:5.14} hold for some $n=k\in\{0,1,\dots\}$.
Then, by \eqref{eq:5.1} and \eqref{eq:5.13} with $n=k$ we have
\begin{equation}
\label{eq:5.15}
\begin{aligned}
& \sup_{t>0}\,t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||U_k(t)|||_{q,\ell}\\
& \le \sup_{t>0}\,t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||u(t)-U_k(t)|||_{q,\ell}
+\sup_{t>0}\,t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||u(t)|||_{q,\ell}<\infty
\end{aligned}
\end{equation}
for $1\le q\le\infty$ and $0\le \ell\le K$ with $\ell<\theta+N(1-1/q)$.
On the other hand, it follows that
\begin{equation}
\label{eq:5.16}
\begin{aligned}
& u(x,t)-\int_0^t \left[e^{-(t-s)(-\Delta)^{\theta/2}}F_k(s)\right](x)\,ds\\
& =\left[e^{-t(-\Delta)^{\theta/2}}\varphi\right](x)
+\int_0^te^{-(t-s)(-\Delta)^{\theta/2}}[F(s)-F_k(s)]\,ds
\end{aligned}
\end{equation}
for $(x,t)\in{\mathbb R}^N\times(0,\infty)$.
By \eqref{eq:5.2} and \eqref{eq:5.15}
we apply Lemma~\ref{Lemma:5.1} to obtain
$$
E_{K,q}[F-F_k]\in L^\infty(0,\tau)\quad\mbox{for}\quad\tau>0
$$
and
$$
\begin{aligned}
E_{K,q}[F-F_k](t)
& =\left\{
\begin{array}{ll}
o\left(t^{-A_p}\right) & \mbox{if}\quad (k+1)(A_p-1)<K/\theta,\vspace{5pt}\\
O\left(t^{-A_p}\log t\right) & \mbox{if}\quad (k+1)(A_p-1)=K/\theta,\vspace{5pt}\\
O\left(t^{-A_p+\frac{K}{\theta}-(k+1)(A_p-1)}\right) & \mbox{if}\quad (k+1)(A_p-1)>K/\theta,
\end{array}
\right.
\end{aligned}
$$
as $t\to\infty$.
Then we apply Theorem~\ref{Theorem:1.1} to \eqref{eq:5.16},
namely $f(x,t)=F(x,t)-F_k(x,t)$,
and obtain \eqref{eq:5.13} and \eqref{eq:5.14} with $n=k+1$.
Therefore, by induction we obtain \eqref{eq:5.13} and \eqref{eq:5.14} for $n=0,1,2,\dots$.
Thus Theorem~\ref{Theorem:5.2} follows.
$\Box$\vspace{5pt}
Similarly to the proof of Theorem~\ref{Theorem:5.2} for the case $n=1$,
we prove Theorem~\ref{Theorem:1.2}.
\vspace{5pt}
\newline
{\bf Proof of Theorem~\ref{Theorem:1.2}.}
Let $1\le q\le\infty$ and $0\le\ell\le K$ with $\ell<\theta+N(1-1/q)$.
Assume $p>1+\theta/N$ and
put $F(u(x,t)):=\lambda|u(x,t)|^{p-1}u(x,t)$.
Assertion~(a) follows from the similar argument to that of the proof of Corollary~\ref{Corollary:5.1}.
Furthermore, for any $\sigma>0$, by \eqref{eq:2.2} and \eqref{eq:5.12} we have
\begin{equation*}
\begin{aligned}
& t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||U_0(t)-M_*g(t)|||_{q,\ell}\\
& \le C\int_t^\infty |M_0(F(s))|\,ds+C\sum_{1\le |\alpha|\le K}t^{-\frac{|\alpha|}{\theta}}|M_\alpha(\varphi)|
+C\sum_{m=1}^{K_\theta}\sum_{|\alpha|\le K}t^{-m-\frac{|\alpha|}{\theta}}|M_\alpha(\varphi)|\\
&\qquad\qquad
+C\sum_{m=1}^{K_\theta}
\sum_{|\alpha|\le K}t^{-m-\frac{|\alpha|}{\theta}}\int_0^t (s+1)^m |M_\alpha(F(s))|\,ds\\
& \qquad\qquad\qquad\qquad
+C\sum_{1\le|\alpha|\le K}t^{-\frac{|\alpha|}{\theta}}\int_0^t |M_\alpha(F(s))|\,ds\\
& =O\left(t^{-(A_p-1)}\right)+O(t^{-\frac{1}{\theta}})+O(t^{-1})+O\left(t^{-1}\int_1^t s^{1-A_p}\,ds\right)
+O\left(t^{-\frac{1}{\theta}}\int_1^t s^{\frac{1}{\theta}-A_p}\,ds\right)\\
& =O\left(t^{-(A_p-1)+\sigma}\right)+O(t^{-1})+O(t^{-\frac{1}{\theta}})
=O(h_\sigma(t))
\end{aligned}
\end{equation*}
as $t\to\infty$.
This together with \eqref{eq:5.5} implies that
\begin{equation}
\label{eq:5.17}
t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||u(t)-M_*g(t)|||_{q,\ell}
=o\left(t^{-\frac{K}{\theta}}\right)+O(h_\sigma(t))
\quad\mbox{as}\quad t\to\infty.
\end{equation}
Furthermore,
combining Lemma~\ref{Lemma:5.1} and \eqref{eq:5.17}, we have
\begin{equation}
\label{eq:5.18}
E_{K,q}[F(u)-F_\infty](t)=o\left(t^{-A_p-\frac{K}{\theta}}\right)+O\left(t^{-A_p}h_\sigma(t)\right)
\end{equation}
as $t\to\infty$.
On the other hand, it follows that
$$
\begin{aligned}
w(x,t) & :=u(x,t)-\int_0^t e^{-(t-s)(-\Delta)^{\frac{\theta}{2}}}F_\infty(s)\,ds\\
& \,\,=\left[e^{-t(-\Delta)^{\frac{\theta}{2}}}\varphi\right](x)
+\int_0^t e^{-(t-s)(-\Delta)^{\frac{\theta}{2}}}[F(u(s))-F_\infty(s)]\,ds.
\end{aligned}
$$
Applying Theorem~\ref{Theorem:1.1},
for any $\varepsilon>0$ and $T>0$, we obtain
\begin{equation}
\label{eq:5.19}
\begin{aligned}
& t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||w(t)-w_*(t)|||_{q,\ell}\\
& \le\varepsilon t^{-\frac{K}{\theta}}
+C_*t^{-\frac{K}{\theta}}\int_T^t s^{\frac{K}{\theta}}E_{K,q}[F(u)-F_\infty](s)\,ds
\end{aligned}
\end{equation}
as $t\to\infty$, where $C_*$ is a positive constant independent of $\varepsilon$ and $T$ and
$$
w_*(x,t)=\sum_{m=0}^{K_\theta}\sum_{|\alpha|\le K}
\left(M_\alpha(\varphi)+\int_0^t (s+1)^mM_\alpha(F(u(s))-F_\infty(s))\,ds\right) g_{\alpha,m}(x,t).
$$
Since $\varepsilon$ is arbitrary,
by \eqref{eq:5.18} and \eqref{eq:5.19} we see that
$$
t^{\frac{N}{\theta}\left(1-\frac{1}{q}\right)-\frac{\ell}{\theta}}|||w(t)-w_*(t)|||_{q,\ell}
=o\left(t^{-\frac{K}{\theta}}\right)+O\left(t^{-\frac{K}{\theta}}
\int_T^t s^{\frac{K}{\theta}-A_p}h_\sigma(s)\,ds\right)
$$
as $t\to\infty$.
This implies assertion~(b).
Thus Theorem~\ref{Theorem:1.2} follows.
$\Box$
\begin{remark}
\label{Remark:5.2}
Let $u$ be a solution to the Cauchy problem for a nonlinear fractional diffusion equation
and possess the mass conservation law, that is,
$\int_{{\mathbb R}^N}u(x,t)\,dx$ is independent of $t$.
The mass conservation law has often played an important role
in the study of HOAE of solutions to various nonlinear problems,
see e.g. \cites{DC, EZ, FM, NY, YM, YS01, YS02, YS03, YT1, YT2}.
Then the arguments in the proof of Theorem~{\rm\ref{Theorem:4.1}} are valid
for the Cauchy problem.
\end{remark}
{\bf Acknowledgements.}
The authors of this paper were supported in part by JSPS KAKENHI Grant Number JP19H05599.
The second author was also supported in part by JSPS KAKENHI Grant Number JP20K03689.
\begin{bibdiv}
\begin{biblist}
\bib{AJY}{article}{
author={Achleitner, Franz},
author={J\"{u}ngel, Ansgar},
author={Yamamoto, Masakazu},
title={Large-time asymptotics of a fractional drift-diffusion-Poisson
system via the entropy method},
journal={Nonlinear Anal.},
volume={179},
date={2019},
pages={270--293},
}
\bib{AF}{article}{
author={Amann, H.},
author={Fila, M.},
title={A Fujita-type theorem for the Laplace equation with a dynamical
boundary condition},
journal={Acta Math. Univ. Comenian. (N.S.)},
volume={66},
date={1997},
pages={321--328},
}
\bib{BT}{article}{
author={Bogdan, Krzysztof},
author={Jakubowski, Tomasz},
title={Estimates of heat kernel of fractional Laplacian perturbed by
gradient operators},
journal={Comm. Math. Phys.},
volume={271},
date={2007},
pages={179--198},
}
\bib{BK}{article}{
author={Brandolese, Lorenzo},
author={Karch, Grzegorz},
title={Far field asymptotics of solutions to convection equation with
anomalous diffusion},
journal={J. Evol. Equ.},
volume={8},
date={2008},
pages={307--326},
}
\bib{DC}{article}{
author={Duro, Gema},
author={Carpio, Ana},
title={Asymptotic profiles for convection-diffusion equations with
variable diffusion},
journal={Nonlinear Anal.},
volume={45},
date={2001},
pages={407--433},
}
\bib{EZ}{article}{
author={Escobedo, Miguel},
author={Zuazua, Enrike},
title={Large time behavior for convection-diffusion equations in ${\mathbb R}^N$},
journal={J. Funct. Anal.},
volume={100},
date={1991},
pages={119--161},
}
\bib{FIK}{article}{
author={Fila, Marek},
author={Ishige, Kazuhiro},
author={Kawakami, Tatsuki},
title={Convergence to the Poisson kernel for the Laplace equation with a
nonlinear dynamical boundary condition},
journal={Commun. Pure Appl. Anal.},
volume={11},
date={2012},
pages={1285--1301},
}
\bib{FK}{article}{
author={Fino, Ahmad},
author={Karch, Grzegorz},
title={Decay of mass for nonlinear equation with fractional Laplacian},
journal={Monatsh. Math.},
volume={160},
date={2010},
pages={375--384},
}
\bib{FM}{article}{
author={Fujigaki, Yoshiko},
author={Miyakawa, Tetsuro},
title={Asymptotic profiles of nonstationary incompressible Navier-Stokes
flows in the whole space},
journal={SIAM J. Math. Anal.},
volume={33},
date={2001},
pages={523--544},
}
\bib{HKN}{article}{
author={Hayashi, Nakao},
author={Kaikina, Elena I.},
author={Naumkin, Pavel I.},
title={Asymptotics for fractional nonlinear heat equations},
journal={J. London Math. Soc. (2)},
volume={72},
date={2005},
pages={663--688},
}
\bib{IIK}{article}{
author={Ishige, Kazuhiro},
author={Ishiwata, Michinori},
author={Kawakami, Tatsuki},
title={The decay of the solutions for the heat equation with a potential},
journal={Indiana Univ. Math. J.},
volume={58},
date={2009},
pages={2673--2707},
}
\bib{IK}{article}{
author={Ishige, Kazuhiro},
author={Kawakami, Tatsuki},
title={Refined asymptotic profiles for a semilinear heat equation},
journal={Math. Ann.},
volume={353},
date={2012},
pages={161--192},
issn={0025-5831},
}
\bib{IKK01}{article}{
author={Ishige, Kazuhiro},
author={Kawakami, Tatsuki},
author={Kobayashi, Kanako},
title={Global solutions for a nonlinear integral equation with a
generalized heat kernel},
journal={Discrete Contin. Dyn. Syst. Ser. S},
volume={7},
date={2014},
pages={767--783},
}
\bib{IKK02}{article}{
author={Ishige, Kazuhiro},
author={Kawakami, Tatsuki},
author={Kobayashi, Kanako},
title={Asymptotics for a nonlinear integral equation with a generalized
heat kernel},
journal={J. Evol. Equ.},
volume={14},
date={2014},
pages={749--777},
}
\bib{IKM}{article}{
author={Ishige, Kazuhiro},
author={Kawakami, Tatsuki},
author={Michihisa, Hironori},
title={Asymptotic expansions of solutions of fractional diffusion
equations},
journal={SIAM J. Math. Anal.},
volume={49},
date={2017},
pages={2167--2190},
}
\bib{IKO}{article}{
author={Ishige, Kazuhiro},
author={Kawakami, Tatsuki},
author={Okabe, Shinya},
title={Existence of solutions to nonlinear parabolic equations via majorant integral kernel},
journal={preprint (arXiv:2101.06581)},
}
\bib{Iwa}{article}{
author={Iwabuchi, Tsukasa},
title={Global solutions for the critical Burgers equation in the Besov
spaces and the large time behavior},
journal={Ann. Inst. H. Poincar\'{e} Anal. Non Lin\'{e}aire},
volume={32},
date={2015},
pages={687--713},
}
\bib{NY}{article}{
author={Nagai, Toshitaka},
author={Yamada, Tetsuya},
title={Large time behavior of bounded solutions to a parabolic system of
chemotaxis in the whole space},
journal={J. Math. Anal. Appl.},
volume={336},
date={2007},
pages={704--726},
}
\bib{S}{article}{
author={Sugitani, Sadao},
title={On nonexistence of global solutions for some nonlinear integral
equations},
journal={Osaka Math. J.},
volume={12},
date={1975},
pages={45--51},
}
\bib{W}{article}{
author={Weissler, Fred B.},
title={Existence and nonexistence of global solutions for a semilinear
heat equation},
journal={Israel J. Math.},
volume={38},
date={1981},
pages={29--40},
}
\bib{YM}{article}{
author={Yamamoto, Masakazu},
title={Asymptotic expansion of solutions to the dissipative equation with
fractional Laplacian},
journal={SIAM J. Math. Anal.},
volume={44},
date={2012},
pages={3786--3805; {\it Erratum} in SIAM J. Math. Anal. {\bf 48} (2016), 3037--3038},
}
\bib{YS01}{article}{
author={Yamamoto, Masakazu},
author={Sugiyama, Yuusuke},
title={Asymptotic expansion of solutions to the drift-diffusion equation
with fractional dissipation},
journal={Nonlinear Anal.},
volume={141},
date={2016},
pages={57--87},
}
\bib{YS02}{article}{
author={Yamamoto, Masakazu},
author={Sugiyama, Yuusuke},
title={Spatial-decay of solutions to the quasi-geostrophic equation with
the critical and supercritical dissipation},
journal={Nonlinearity},
volume={32},
date={2019},
pages={2467--2480},
}
\bib{YS03}{article}{
author={Yamamoto, Masakazu},
author={Sugiyama, Yuusuke},
title={Optimal estimates for far field asymptotics of solutions to the
quasi-geostrophic equation},
journal={Proc. Amer. Math. Soc.},
volume={149},
date={2021},
pages={1099--1110},
}
\bib{YT1}{article}{
author={Yamada, Tetsuya},
title={Moment estimates and higher-order asymptotic expansions of
solutions to a parabolic system of chemotaxis in the whole space},
journal={Funkcial. Ekvac.},
volume={54},
date={2011},
pages={15--51},
}
\bib{YT2}{article}{
author={Yamada, T.},
title={Improvement of convergence rates for a parabolic system of
chemotaxis in the whole space},
journal={Math. Methods Appl. Sci.},
volume={34},
date={2011},
pages={2103--2124},
}
\end{biblist}
\end{bibdiv}
\end{document}
|
\section{Introduction} \label{sec:1}
\input{subtex/01intro}
\pagebreak
\section{Preliminaries} \label{sec:2}
\input{subtex/02prelim}
\section{Approach} \label{sec:3}
\input{subtex/03approach1}
\input{subtex/04approach2}
\vspace{-2pt}
\section{Experiments} \label{sec:4}
\input{subtex/05experiment}
\vspace{-2pt}
\section{Conclusion} \label{sec:6}
\vspace{-2pt}
We present an efficient higher-order optimization framework for training Neural ODEs.
Our method -- named \textbf{SNOpt} -- differs from existing second-order methods in various aspects.
While it leverages similar factorization inherited in Kronecker-based methods \citep{martens2015optimizing},
the two methodologies differ fundamentally in that
we construct analytic ODE expressions for higher-order derivatives (Theorem~\ref{prop:1})
and compute them through \texttt{\markblue{ODESolve}}.
This retains the favorable ${\cal O}(1)$ memory as opposed to their ${\cal O}({T})$.
It also enables a flexible rank-based factorization in Proposition~\ref{prop:2}.
Meanwhile,
our method extends the recent trend of OCP-inspired methods \citep{li2017maximum,liu2021dynamic}
to deep continuous-time models, yet using a rather straightforward framework
without imposing additional assumptions, such as Markovian or game transformation.
To summarize,
our work advances several methodologies to the emerging deep continuous-time models, achieving strong empirical results and
opening up new opportunities for analyzing models such as Neural SDEs/PDEs.
\newpage
\begin{ack}
The authors would like to thank Chia-Wen Kuo and Chen-Hsuan Lin for the meticulous proofreading,
and Keuntaek Lee for providing additional computational resources.
Guan-Horng Liu was supported by CPS NSF Award \#1932068, and Tianrong Chen was supported by ARO Award \#W911NF2010151.
\end{ack}
\bibliographystyle{icml2021}
\subsection{Results}
\begin{figure}[t]
\begin{minipage}{\textwidth}
\centering
\includegraphics[height=4.4cm]{fig/convergence2.pdf}
\vskip -0.05in
\caption{
Training performance in \emph{wall-clock} runtime,
averaged over 3 trials.
Our SNOpt achieves faster convergence against first-order baselines.
See Fig.~\ref{fig:convergence-full} in Appendix~\ref{app:5} for MNIST and Circle.
}
\label{fig:convergence}
\end{minipage}
\vskip 0.15in
\begin{minipage}{\textwidth}
\centering
\setlength\tabcolsep{4.5pt}
\captionsetup{type=table}
\caption{Test-time performance: accuracies for {\color{label1}image} and {\color{label3}time-series} datasets; NLL for {\color{label2}CNF} datasets}
\vskip -0.03in
\centering
\begin{tabular}{lccccccccc}
\toprule
& {MNIST} & {SVHN} & {CIFAR10}
& {SpoAD} & {ArtWR} & {CharT}
& Circle & {Gas} & {Miniboone} \\
\addlinespace[-0.25em]\arrayrulecolor{label1}
\cmidrule[1pt](lr){2-4}\corcmidrule\arrayrulecolor{label3}%
\cmidrule[1pt](lr){5-7}\corcmidrule\arrayrulecolor{label2}%
\cmidrule[1pt](lr){8-10}\corcmidrule
\addlinespace[0.4em]\arrayrulecolor{black}
\midrule
Adam & 98.83 & 91.92 & 77.41 & 94.64 & 84.14 & 93.29
& 0.90 & -6.42 & 13.10 \\[2pt]
SGD & 98.68 & 93.34 & 76.42 & \textbf{97.70} & 85.82 & 95.93
& 0.94 & -4.58 & 13.75 \\[1pt]
\midrule
\textbf{SNOpt}
& \textbf{98.99} & \textbf{95.77} & \textbf{79.11} & 97.41 & \textbf{90.23} & \textbf{96.63}
&\textbf{0.86} & \textbf{-7.55} & \textbf{12.50} \\
\bottomrule
\end{tabular} \label{table:accu}
\end{minipage}
\vskip 0.15in
\begin{minipage}{0.37\textwidth}
\vskip 0.05in
\centering
\includegraphics[height=3.5cm]{fig/complexity.pdf}
\caption{
Relative runtime and memory of our SNOpt compared to Adam (denoted by the dashed black lines) on all 9 datasets, where `Mn' is the shorthand for MNIST, and \textit{etc}.
}
\label{fig:complexity}
\end{minipage}
\hfill
\begin{minipage}{0.6\textwidth}
\centering
\includegraphics[height=3.8cm]{fig/point-cloud2.pdf}
\vskip -0.05in
\caption{
Sensitivity analysis where
each sample represents a training result using different optimizer and learning rate
(annotated by different symbol and color).
Our SNOpt achieves higher accuracies and is insensitive to hyper-parameter changes.
Note that x-axes are in $\log$ scale.
}
\label{fig:point-cloud}
\end{minipage}
\vskip -0.15in
\end{figure}
\textbf{Convergence and computation efficiency.}
Figures~\ref{fig:1} and \ref{fig:convergence}
report the training curves of each method measured by wall-clock time.
It is obvious that our SNOpt admits a superior
convergence rate compared to the first-order baselines,
and in many cases exceeds their performances by a large margin.
In Fig.~\ref{fig:complexity},
we report the computation efficiency of our SNOpt compared to Adam on each dataset,
and leave their numerical values in Appendix~\ref{app:4} (Table~\ref{table:runtime} and \ref{table:memory}).
For image and time-series datasets ({\ignorespaces\emph{i.e.}}{ } Mn{\texttildelow}CT),
our SNOpt
runs nearly as fast as first-order methods.
This is made possible
through a rigorous OCP analysis in Section~\ref{sec:3},
where we showed that second-order matrices %
can be constructed along with the \emph{same} backward integration when we compute the gradient.
Hence, only a minimal overhead is introduced.
As for CNF, which propagates the probability density additional to the vanilla state dynamics,
our SNOpt is roughly 1.5 to 2.5 times slower,
yet it still converges faster in the overall wall-clock time (see Fig.~\ref{fig:convergence}).
On the other hand,
the use of second-order matrices
increases the memory consumption of SNOpt by 10-40\%,
depending on the model and dataset.
However,
the actual increase in memory (less than 1GB for all datasets; see Table~\ref{table:memory})
remains affordable
on standard GPU machines.
More importantly, our SNOpt retains the ${\cal O}(1)$ memory throughout training.
\textbf{Test-time performance and hyper-parameter sensitivity.}
Table~\ref{table:accu} reports the test-time performance,
including the accuracies (\%) for image and time-series classification,
and the negative log-likelihood (NLL) for CNF.
On most datasets, our method achieves competitive results against standard baselines.
In practice, we also find that using the preconditioned updates
greatly reduce the sensitivity to hyper-parameters ({\ignorespaces\emph{e.g.}}{ } learning rate).
This is demonstrated in Fig.~\ref{fig:point-cloud},
where we sample distinct learning rates from a proper interval for each method (shown with different color bars)
and record their training results after convergence.
It is clear that
our method not only converges to higher accuracies with lower losses,
these values are also more concentrated on the plots.
In other words,
our method achieves better convergence in a more consistent manner across different hyper-parameters.
\begin{figure}[t]
\vskip -0.15in
\centering
\begin{minipage}{0.43\textwidth}
\setlength\tabcolsep{3pt}
\centering
\captionsetup{type=table}
\caption{
Performance of jointly optimizing the integration bound $t_1$ on CIFAR10
}
\vskip -0.07in
\centering
\begin{tabular}{lcc}
\toprule
Method & \specialcell{Train time (\%) \\ w.r.t. $t_1{=}1.0$ }
& \specialcell{Accuracy \\ (\%)}
\\
\midrule
ASM baseline
& 96 & 76.61 \\
\textbf{SNOpt (ours)} & \textbf{81} & \textbf{77.82} \\
\bottomrule
\end{tabular} \label{table:adap-t1}
\begin{center}
\captionsetup{type=figure}
\includegraphics[height=2cm]{fig/adapt-t1-result.pdf}
\caption{
Dynamics of $t_1$ over CIFAR10 training using different methods.
}
\label{fig:adap-t1b}
\end{center}
\end{minipage}
\hfill
\begin{minipage}{0.55\textwidth}
\setlength\tabcolsep{3pt}
\centering
\captionsetup{type=table}
\caption{
Measure of implicit regularization on SVHN
}
\vskip -0.07in
\centering
\begin{tabular}{lcc}
\toprule
& \# of function & Regularization \\
& evaluation (NFE) & ($ \int \norm{\nabla_{\mathbf{x}} F}^2 + \int \norm{F}^2$) \\
\midrule
Adam
& 42.1 & 323.88 \\
\textbf{SNOpt} & \textbf{32.6} & \textbf{199.1} \\
\bottomrule
\end{tabular} \label{table:reg}
\begin{center}
\captionsetup{type=figure}
\includegraphics[height=2cm]{fig/recursive-adjoint.pdf}
\caption{
Comparison between SNOpt and second-order recursive adjoint.
SNOpt is at least 2 times faster and improves the accuracies of baselines by 5-15\%.
}
\label{fig:recur-adj}
\end{center}
\end{minipage}
\vskip -0.1in
\end{figure}
\textbf{Joint optimization of the integration bound $t_1$.}
Table~\ref{table:adap-t1} and Fig.~\ref{fig:adap-t1b}
report the performance of optimizing $t_1$ along with its convergence dynamics.
Specifically, we compare our second-order feedback policy (\ref{eq:feedback}) derived in Section~\ref{sec:3.3}
to the first-order ASM baseline proposed in \citet{massaroli2020dissecting}.
It is clear that
our OCP-theoretic method
leads to substantially faster convergence,
and the optimized $t_1$ stably hovers around $0.5$ without deviation (as appeared for the baseline).
This drops the training time by nearly 20\% compared to the vanilla training, where we fix $t_1$ to $1.0$,
yet without sacrificing the test-time accuracy.
A similar experiment for MNIST (see Fig.~\ref{fig:app-adap-t1b} in Appendix~\ref{app:5})
shows a consistent result.
We highlight
these improvements as the benefit gained from introducing the well-established OCP principle
to these emerging deep continuous-time models.
\textbf{Comparison with recursive adjoint.}
Finally,
Fig.~\ref{fig:recur-adj} reports the comparison between our SNOpt and the recursive adjoint baseline (see Section~\ref{sec:2} and Table~\ref{table:adj2-error}).
It is clear that our method outperforms this second-order baseline by a large margin
in both runtime efficiency and test-time performance.
Note that we omit the comparison on CNF datasets since the recursive adjoint simply fails to converge.
\textbf{Remark} (Implicit regularization)\textbf{.}
In some cases ({\ignorespaces\emph{e.g.}}{ } SVHN in Fig.~\ref{fig:complexity}),
our method may run slightly faster
than first-order methods.
This is a distinct phenomenon arising exclusively from training these continuous-time models.
Since their forward and backward passes involve solving \emph{parameterized} ODEs (see Fig.~\ref{fig:2}),
the computation graphs are \emph{parameter-dependent}; hence adaptive throughout training.
In this vein, we conjecture that
the preconditioned updates in these cases may have guided the parameter to regions that
are numerically stabler (hence faster) for integration.\footnote{
In Appendix~\ref{app:4}, we provide some theoretical discussions (see Corollary~\ref{coro:4}) in this regard.
\label{footnote:4}
}
With this in mind, we report in Table~\ref{table:reg}
the value of Jacobian, $ \int \norm{\nabla_{\mathbf{x}} F}^2$, and Kinetic, $\int \norm{F}^2$, regularization \citep{finlay2020train} in SVHN training.
Interestingly, the parameter found by our SNOpt indeed has a substantially lower
value (hence stronger regularization and better-conditioned ODE dynamics) compared to the one found by Adam. This provides a plausible explanation of the reduction in the NFE when using our method, yet without hindering the test-time performance (see Table~\ref{table:accu}).
\iffalse
The numerical values are reported in Table~\ref{table:runtime}.
Notably,
\fi
\subsection{Efficient Second-order Preconditioned Update} \label{sec:3.2}
Theorem~\ref{prop:1} provides an attractive computational framework
that does not require recursive computation (as mentioned in Section~\ref{sec:2}) to obtain higher-order derivatives.
It suggests that
we can obtain first and second-order derivatives all at once with a single function call of \texttt{\markblue{ODESolve}}:
\begin{align}
\begin{split}
[{\bm{x}}_{t_0}, {Q_{\hvx}}(t_0), {Q_{\hvu}}(t_0), {Q_{\hvx \hvx}}(t_0), &\text{ }{Q_{\hvu \hvx}}(t_0), {Q_{\hvx \hvu}}(t_0), {Q_{\hvu \hvu}}(t_0)] \\ &=
\texttt{\markblue{ODESolve}(}[{\bm{x}}_{t_1}, \Phi_{ \bar{{\bm{x}}} }, \mathbf{0}, \Phi_{{ \bar{{\bm{x}}} }\hvx}, \mathbf{0}, \mathbf{0}, \mathbf{0}], t_1, t_0, \tilde{G} \texttt{)},
\label{eq:backward-ode2}
\end{split}
\end{align}
where $\tilde{G}$ augments the original dynamics $F$ in (\ref{eq:node}) with all 6 ODEs presented in (\ref{eq:cddp-q}).
Despite that this OCP-theoretic backward pass (\ref{eq:backward-ode2}) retains the same ${\cal O}(1)$ memory complexity as in (\ref{eq:backward-ode}),
the dimension of the new augmented state, which now carries second-order matrices,
can grow to an unfavorable size that dramatically slows down the numerical integration.
Hence, we must consider other representations of (\ref{eq:cddp-q}), if any, in order to proceed.
In the following proposition, we present one of which
that transforms (\ref{eq:cddp-q}) into a set of {vector} ODEs, so that we can compute them much efficiently.
\begin{proposition}[Low-rank representation of (\ref{eq:cddp-q})] \label{prop:2}
Suppose $\ell {:=} 0$ in (\ref{eq:ct-ocp})
and let ${Q_{\hvx \hvx}}(t_1){ =} \sum_{i=1}^R {\bm{y}}_i \otimes {\bm{y}}_i$ be a symmetric matrix of rank $R \le n$,
where ${\bm{y}}_i \in \mathbb{R}^m$ and $\otimes$ is the Kronecker product.
Then, for all $t \in [t_0, t_1]$, the second-order matrices appeared in (\ref{eq:cddp-qb}, \ref{eq:cddp-qc}) can be decomposed into
\begin{align*}
{Q_{\hvx \hvx}}(t) = \sum_{i=1}^R {\mathbf{q}}_i(t) \otimes {\mathbf{q}}_i(t), \quad
{Q_{\hvx \hvu}}(t) = \sum_{i=1}^R {\mathbf{q}}_i(t) \otimes {\mathbf{p}}_i(t), \quad
{Q_{\hvu \hvu}}(t) = \sum_{i=1}^R {\mathbf{p}}_i(t) \otimes {\mathbf{p}}_i(t),
\end{align*}
where the vectors ${\mathbf{q}}_i(t) \in \mathbb{R}^m $ and ${\mathbf{p}}_i(t) \in \mathbb{R}^n$ obey the following backward ODEs:
\begin{align}
- \fracdiff{{\mathbf{q}}_i(t)}{t} = {{F}_\hvx}(t)^\mathsf{T}{\mathbf{q}}_i(t), \quad
- \fracdiff{{\mathbf{p}}_i(t)}{t} = {{F}_\hvu}(t)^\mathsf{T}{\mathbf{q}}_i(t), \quad
\label{eq:vector-ode}
\end{align}
with the terminal condition given by $({\mathbf{q}}_i({t_1}),{\mathbf{p}}_i({t_1})) := ({\bm{y}}_i, \mathbf{0})$.
\end{proposition}
The proof is left in Appendix~\ref{app:2}.
Proposition~\ref{prop:2} gives a nontrivial conversion.
It indicates that the \emph{coupled matrix} ODEs presented in (\ref{eq:cddp-qb}, \ref{eq:cddp-qc}) can be disentangled into a set of \emph{independent vector} ODEs where each of them follows its
own dynamics (\ref{eq:vector-ode}).
As the rank $R$ determines the number of these vector ODEs,
this conversion will be particularly useful if the second-order matrices exhibit low-rank structures.
Fortunately, this is indeed the case for many Neural-ODE applications
which often propagate ${\mathbf{x}}_t$ in a latent space of higher dimension \citep{chen2018neural,grathwohl2018ffjord,kidger2020neural}.
Based on Proposition~\ref{prop:2}, the second-order precondition matrix ${\cal L}_{\theta\theta}$
is given by\footnote{
We drop the dependence on $t$ for brevity, yet all terms inside the integrations of (\ref{eq:Quu}, \ref{eq:q-kfac}) are time-varying.
}
\begin{align}
{\cal L}_{\theta\theta} \equiv {Q_{\hvu \hvu}}(t_0) =
\sum_{i=1}^R \left( { \int^{t_0}_{t_1} } {{F}_\hvu}^\mathsf{T}{\mathbf{q}}_i \text{ } { \mathrm{d} t } \right) \otimes \left( { \int^{t_0}_{t_1} } {{F}_\hvu}^\mathsf{T}{\mathbf{q}}_i \text{ } { \mathrm{d} t } \right),
\label{eq:Quu}
\end{align}
where ${\mathbf{q}}_i \equiv {\mathbf{q}}_i(t)$ follows (\ref{eq:vector-ode}).
Our final step is to facilitate efficient computation of (\ref{eq:Quu}) with Kronecker-based factorization,
which underlines many popular second-order methods for discrete DNNs \citep{grosse2016kronecker,martens2018kronecker}.
Recall that
the vector field $F$ is represented
\begin{wrapfigure}[7]{r}{0.4\textwidth}
\vspace{-7pt}
\setlength{\unitlength}{0.1\textwidth}
\begin{center}
\begin{picture}(4,1.5)
\put(0,0){\includegraphics[height=2.12cm]{fig/fig3-crop.pdf}}
\put(1.05, 1.27){{\color{darkgreen}\fontsize{8}{9}\selectfont$F({\cdot},{\cdot},{\theta}) \equiv F({\cdot},{\cdot},{{\mathbf{u}}_t}) $}}
\put(1.3, 0.93){{\fontsize{7}{9}\selectfont${\mathbf{z}}^n(t)$}}
\put(2.05, 0.93){{\fontsize{7}{9}\selectfont${\mathbf{z}}^{n{+}1}(t)$}}
\put(0.05, 1){{\fontsize{7}{9}\selectfont$(t{,}{\mathbf{x}}(t))$}}
\put(3.25, 1){{\fontsize{9.5}{10}\selectfont$\frac{{\textnormal{d}}{\mathbf{x}}(t)}{{ \mathrm{d} t }}$}}
\put(1.58, 0.27){{\fontsize{6}{7}\selectfont$\begin{cases} $\text{ }$ \\ $\text{ }$ \end{cases}$}}
\put(1.89, 0.37){{\fontsize{7.5}{9}\selectfont${\mathbf{h}}^n(t) = f{(}{\mathbf{z}}^n{(}t{)}{,}{\mathbf{u}}^n{(}t{)}{)}$}}
\put(1.7, 0.12){{\fontsize{7.5}{9}\selectfont${\mathbf{z}}^{n{+}1}(t) = \sigma({\mathbf{h}}^n(t))$}}
\end{picture}
\end{center}
\vskip -0.1in
\caption{The layer propagation inside the vector field $F$, where $f$ and $\sigma$ denote affine and nonlinear activation functions.}
\label{fig:3}
\end{wrapfigure}
by a DNN.
Let ${\mathbf{z}}^n(t)$, ${\mathbf{h}}^n(t)$, and ${\mathbf{u}}^n(t)$ denote
the activation vector, pre-activation vector, and the parameter of layer $n$
when evaluating $\fracdiff{{\mathbf{x}}}{t}$ at time $t$ (see Fig.~\ref{fig:3}),
then the integration in (\ref{eq:Quu}) can be broken down into each layer $n$,
\begin{align*}
{ \int^{t_0}_{t_1} } \left({{F}_\hvu}^\mathsf{T}{\mathbf{q}}_i\right) { \mathrm{d} t }
=& [\cdots, \textstyle { \int^{t_0}_{t_1} } \left(F_{{ \bar{{\bm{u}}} }^n}^\mathsf{T}{\mathbf{q}}_i\right) { \mathrm{d} t }, \cdots] \\
=& [\cdots, \textstyle { \int^{t_0}_{t_1} } \left( {\mathbf{z}}^n \otimes ({\fracpartial{F}{{\mathbf{h}}^n}}^\mathsf{T} {\mathbf{q}}_i) \right) { \mathrm{d} t }, \cdots],
\end{align*}
where the second equality holds by
$
F_{{ \bar{{\bm{u}}} }^n}^\mathsf{T} {\mathbf{q}}_i
= (\fracpartial{F}{{\mathbf{h}}^n} \fracpartial{{\mathbf{h}}^n}{{\mathbf{u}}^n}) ^\mathsf{T} {\mathbf{q}}_i
= {\mathbf{z}}^n \otimes ({\fracpartial{F}{{\mathbf{h}}^n}}^\mathsf{T} {\mathbf{q}}_i)$.
This is an essential step towards the Kronecker approximation of the layer-wise precondition matrix:
\begin{align*}
{\cal L}_{\theta^n\theta^n}
\equiv
Q_{{ \bar{{\bm{u}}} }^n { \bar{{\bm{u}}} }^n}(t_0)
&=
\sum_{i=1}^R
\left( { \int^{t_0}_{t_1} } \left( \textstyle {\mathbf{z}}^n \otimes ({\fracpartial{F}{{\mathbf{h}}^n}}^\mathsf{T} {\mathbf{q}}_i) \right) { \mathrm{d} t } \right)
\otimes
\left( { \int^{t_0}_{t_1} } \left( \textstyle
{\mathbf{z}}^n \otimes ({\fracpartial{F}{{\mathbf{h}}^n}}^\mathsf{T} {\mathbf{q}}_i) \right) { \mathrm{d} t } \right) \\
&\approx
{ \int^{t_0}_{t_1} } \underbrace{\mystrut{2.5ex} \left( \textstyle {\mathbf{z}}^n \otimes {\mathbf{z}}^n \right)}_{{\bm{A}}_n(t)} { \mathrm{d} t }
\otimes
{ \int^{t_0}_{t_1} } \underbrace{ \sum_{i=1}^R \left( \textstyle ({\fracpartial{F}{{\mathbf{h}}^n}}^\mathsf{T} {\mathbf{q}}_i) \otimes ({\fracpartial{F}{{\mathbf{h}}^n}}^\mathsf{T} {\mathbf{q}}_i) \right)}_{{\bm{B}}_n(t)} { \mathrm{d} t }.
\numberthis \label{eq:q-kfac}
\end{align*}
\begin{figure}[t]
\vskip -0.2in
\begin{minipage}{\textwidth}
\begin{algorithm}[H]
\vskip -0.05in
\caption{SNOpt: Second-order Neural ODE Optimizer} %
\label{alg:1}
\begin{algorithmic}[1]
\STATE {\bfseries Input:}
dataset ${\cal D}$,
parametrized vector field $F(\cdot,\cdot,\theta)$,
integration time $[t_0,t_1]$,
black-box ODE \\ $\quad$ solver \texttt{\markblue{ODESolve}},
learning rate $\eta$,
rank $R$, interval of the time grid $\Delta t$
\REPEAT
\STATE Solve ${\mathbf{x}}(t_1) =$ \texttt{\markblue{ODESolve}(}${\mathbf{x}}({t_0}), t_0, t_1, F$\texttt{)},
where ${\mathbf{x}}({t_0}) \sim {\cal D}$. \hfill \markgreen{$\rhd$ Forward pass}
\STATE Initialize $(\bar{{\bm{A}}}_n,\bar{{\bm{B}}}_n) := (\mathbf{0}, \mathbf{0})$ for each layer $n$
and set ${\mathbf{q}}_i(t_1) := {\bm{y}}_i$.
\FOR{$t^\prime$ {\bfseries in} $\{t_1, t_1-\Delta t, \cdots, t_0+\Delta t, t_0 \}$ }
\STATE
Set $t := t^\prime - \Delta t$ as the small integration step, then call \\[0.3ex]
$\quad [{\mathbf{x}}({{t}}), {Q_{\hvx}}({t}), {Q_{\hvu}}({t}), \{{\mathbf{q}}_i({t})\}_{i=1}^R]$ \\[0.3ex]
$\quad\quad = \texttt{\markblue{ODESolve}(}[{\mathbf{x}}({t^\prime}), {Q_{\hvx}}({t^\prime}), {Q_{\hvu}}({t^\prime}), \{{\mathbf{q}}_i({t^\prime})\}_{i=1}^R ], t^\prime, t, \widehat{G} \texttt{)}$,
\hfill \markgreen{$\rhd$ Backward pass\text{ }} \\
where $\widehat{G}$ augments the ODEs of state (\ref{eq:node}), first and second-order derivatives (\ref{eq:cddp-qa}, \ref{eq:vector-ode}).
\STATE Evaluate ${\mathbf{z}}^n(t)$, ${\mathbf{h}}^n(t)$, $F(t, {\mathbf{x}}_{{t}}, \theta)$, then
compute ${\bm{A}}_n(t), {\bm{B}}_n(t)$ in (\ref{eq:q-kfac}).
\STATE Update $\bar{{\bm{A}}}_n \leftarrow \bar{{\bm{A}}}_n + {\bm{A}}_n(t) \cdot \Delta t $
and $\bar{{\bm{B}}}_n \leftarrow \bar{{\bm{B}}}_n + {\bm{B}}_n(t) \cdot \Delta t $.
\ENDFOR
\STATE $\forall n$, apply
$ \theta^n \leftarrow \theta^n - \eta \cdot {\mathrm{vec}}(\bar{{\bm{B}}}_n^{-1}Q_{{ \bar{{\bm{u}}} }^n}(t_0)\bar{{\bm{A}}}_n^{-\mathsf{T}})$.
\hfill \markgreen{$\rhd$ Second-order parameter update\text{ }}
\UNTIL{ converges }
\end{algorithmic}
\end{algorithm}
\end{minipage}
\vskip 0.05in
\begin{minipage}{\textwidth}
\centering
\setlength{\unitlength}{\textwidth}
\begin{picture}(1,0.175)
\put(0.05,0){\includegraphics[height=2.5cm]{fig/fig4-crop.pdf}}
\put(0.053,0.105){{\fontsize{9}{10}\selectfont$\left[\begin{array}{c} $\quad\text{ }\text{ }$ \\ $\quad\text{ }\text{ }$ \\ $\quad\text{ }\text{ }$ \\ $\quad\text{ }\text{ }\text{ }$ \end{array}\right]$}}
\put(0.054,0.105){{\fontsize{8.5}{10}\selectfont$\begin{array}{c}~{\mathbf{x}}(t_0) \\ Q_{\bar{{\bm{x}}}}(t_0) \\ Q_{\bar{{\bm{u}}}}(t_0) \\ \cdot~\end{array}$}}
\put(0.598,0.105){{\fontsize{8.5}{10}\selectfont$\left[\begin{array}{c} $\qquad$ \\ $\qquad$ \\ $\qquad$ \\ $\qquad$ \end{array}\right]$}}
\put(0.598,0.105){{\fontsize{8.5}{10}\selectfont$\begin{array}{c}~{\mathbf{x}}(t_1) \\ Q_{\bar{{\bm{x}}}}(t_1) \\ Q_{\bar{{\bm{u}}}}(t_1) \\ \{{\bm{y}}_i\}_{i=1}^R ~\end{array}$}}
\put(0.31, 0.125){{\fontsize{8.5}{10}\selectfont$t_{j+1}$}}
\put(0.39, 0.11){{\fontsize{8.5}{10}\selectfont$t_{j}$}}
\put(0.901 , 0.078){{\fontsize{6}{10}\selectfont${\{}t_j{\}}$}}
\put(0.17, 0.044){{\fontsize{7.5}{10}\selectfont$ \begin{cases} \bar{{\bm{A}}}_n {=} {\sum_j} {\bm{A}}_n(t_j) {\cdot} \Delta t \\ \bar{{\bm{B}}}_n {=} {\sum_j} {\bm{B}}_n(t_j) {\cdot} \Delta t \end{cases}$}}
\put(0.387, 0.024){{\fontsize{8.5}{10}\selectfont$\widehat{G}$}}
\put(0.7425, 0.027){{\fontsize{5}{10}\selectfont$\widehat{G}$}}
\end{picture}
\vskip -0.1in
\caption{
Our second-order method, SNOpt, solves a new backward ODE,
{\ignorespaces\emph{i.e.}}{ } the $\widehat{G}$ appeared in line 6 of Alg.~\ref{alg:1},
which augments second-order derivatives,
while simultaneously collecting the matrices ${{\bm{A}}}_n(t_j)$ and ${{\bm{B}}}_n(t_j)$ on a sampled time grid $\{t_j\}$ for computing the
preconditioned update in (\ref{eq:layer-kfac}).
}
\label{fig:4}
\end{minipage}
\vskip -0.05in
\end{figure}
\vspace{-6pt}
We discuss the approximation behind (\ref{eq:q-kfac}), and also the one for (\ref{eq:layer-kfac}), in Appendix~\ref{app:2}.
Note that ${\bm{A}}_n(t)$ and ${\bm{B}}_n(t)$ are much smaller matrices in $\mathbb{R}^{m\times m}$ compared to the ones in (\ref{eq:cddp-q}), and
they can be efficiently computed with
automatic differentiation packages \citep{paszke2017automatic}.
Now, let
$\{ t_j \}$ be a time grid uniformly distributed over $[t_0,t_1]$
so that
$\bar{{\bm{A}}}_n {=} \textstyle \sum_j {\bm{A}}_n(t_j) \Delta t$ and
$\bar{{\bm{B}}}_n {=} \textstyle \sum_j {\bm{B}}_n(t_j) \Delta t$ approximate
the integrations in (\ref{eq:q-kfac}),
then our final preconditioned update law is given by
\begin{align}
\forall n, \quad
{\cal L}_{\theta^n\theta^n}^{-1} {\cal L}_{\theta^n}
\approx {\mathrm{vec}}\left( \bar{{\bm{B}}}_n^{-1} Q_{{ \bar{{\bm{u}}} }^n}(t_0) \bar{{\bm{A}}}_n^{-\mathsf{T}} \right),
\label{eq:layer-kfac}
\end{align}
where ${\mathrm{vec}}$ denotes vectorization.
Our second-order method -- named \textbf{SNOpt} -- is summarized in Alg.~\ref{alg:1},
with the backward computation ({\ignorespaces\emph{i.e.}}{ } line 4-9 in Alg.~\ref{alg:1}) illustrated in Fig.~\ref{fig:4}.
In practice, we also adopt
eigen-based amortization with Tikhonov regularization (\citet{george2018fast}; see Alg.~\ref{alg:2} in Appendix~\ref{app:4}),
which stabilizes the updates over stochastic training.
\textbf{Remark.}
The fact that Proposition~\ref{prop:2} holds only for degenerate $\ell$
can be easily circumvented in practice.
As $\ell$ typically represents weight decay, $\ell := \frac{1}{t_1-t_0} \norm{\theta}_2$,
which is time-independent,
it can be separated from the backward ODEs (\ref{eq:cddp-q}) and added after solving the backward integration, {\ignorespaces\emph{i.e.}}{ }
\begin{align*}
{Q_{\hvu}}(t_0) \leftarrow \gamma \theta + {Q_{\hvu}}(t_0), \quad
{Q_{\hvu \hvu}}(t_0)
\leftarrow \gamma {\bm{I}} + {Q_{\hvu \hvu}}(t_0),
\end{align*}
where $\gamma$ is the regularization factor.
Finally, we find that using the scaled \emph{Gaussian-Newton} matrix,
{\ignorespaces\emph{i.e.}}{ } ${Q_{\hvx \hvx}}(t_1) \approx \frac{1}{t_1 - t_0} \Phi_{ \bar{{\bm{x}}} } \otimes \Phi_{ \bar{{\bm{x}}} }$,
generally provides a good trade-off between the performance and runtime complexity.
As such, we adopt this approximation to Proposition~\ref{prop:2} for all experiments.
\subsection{Memory Complexity Analysis}
\begin{figure}[H]
\centering
\begin{minipage}{0.95\textwidth}
\vskip -0.05in
\centering
\captionsetup{type=table}
\caption{
Memory complexity at different stages of our derivation in terms of ${\mathbf{x}}_t \in \mathbb{R}^m$, $\theta \in \mathbb{R}^n$, and the rank $R$.
Note that \textit{all} methods have ${\cal O}(1)$ in terms of depth.
}
\vskip -0.05in
\begin{tabular}{rcccc}
\toprule
& Theorem~\ref{prop:1} & Proposition~\ref{prop:2} & \textbf{SNOpt} (Alg.~\ref{alg:1}) &
first-order adjoint
\\[1pt]
& Eqs.~(\ref{eq:cddp-q},\ref{eq:backward-ode2})
& Eqs.~(\ref{eq:vector-ode},\ref{eq:Quu})
& Eqs.~(\ref{eq:q-kfac},\ref{eq:layer-kfac})
& Eqs.~(\ref{eq:grad-node},\ref{eq:adjoint-node}) \\
\midrule
backward storage & ${\cal O}((m+n)^2)$ & ${\cal O}(Rm+Rn)$ & ${\cal O}(Rm+2n)$ & ${\cal O}(m+n)$ \\[3pt]
parameter update & ${\cal O}(n^2)$ & ${\cal O}(n^2)$ & ${\cal O}(2n)$ & ${\cal O}(n)$ \\
\bottomrule
\end{tabular} \label{table:memory-comp}
\end{minipage}
\vskip -0.05in
\end{figure}
Table~\ref{table:memory-comp} summarizes the memory complexity of different computational methods
that appeared along our derivation in Section~\ref{sec:3.1} and \ref{sec:3.2}.
Despite that \textit{all} methods retain ${\cal O}(1)$ memory as with the first-order adjoint method,
their complexity differs in terms of the state and parameter dimension.
Starting from our encouraging result in Theorem~\ref{prop:1},
which allows one to compute all derivatives with a single backward pass,
we first exploit their low-rank representation in Proposition~\ref{prop:2}.
This reduces the storage to ${\cal O}(Rm+Rn)$ and paves a way toward adopting Kronecker factorization,
which further facilitates efficient preconditioning.
With all these, our {SNOpt} is capable of performing efficient second-order updates while enjoying similar memory complexity (up to some constant) compared to first-order adjoint methods.
Lastly, for image applications where Neural ODEs often consist of convolution layers,
we adopt convolution-based Kronecker factorization \citep{grosse2016kronecker,gao2020trace},
which effectively makes the complexity to scale w.r.t. the number of feature maps ({\ignorespaces\emph{i.e.}}{ } number of channels) rather than the full size of feature maps.
\subsection{Extension to Architecture Optimization} \label{sec:3.3}
\begin{wrapfigure}[12]{r}{0.32\textwidth}
\vspace{-23pt}
\begin{center}
\includegraphics[height=2.35cm]{fig/adapt-t1-a.pdf}
\end{center}
\vskip -0.15in
\caption{
Training performance of CIFAR10 with Adam when using different $t_1$,
which motivates joint optimization of $t_1$.
Experiment setup is left in Appendix~\ref{app:4}.
}
\label{fig:adap-t1}
\end{wrapfigure}
Let us discuss an intriguing extension of our OCP framework to optimizing the architecture of Neural ODEs,
specifically the integration bound $t_1$.
In practice,
when problems
contain no prior information on the integration,
$[t_0, t_1]$ is typically set to
some trivial values (usually $[0,1]$) without further justification.
However,
these values can greatly affect both the performance and runtime.
Take CIFAR10 for instance (see Fig.~\ref{fig:adap-t1}),
the required training time decreases linearly as we drop $t_1$ from $1$, yet the accuracy retains mostly the same unless $t_1$ becomes too small.
Similar results also appear on MNIST (see Fig.~\ref{fig:app-adap-t1a} in Appendix~\ref{app:5}).
In other words, we may interpret the integration bound $t_1$
as an \emph{architectural parameter} that needs to be jointly optimized during training.
The aforementioned interpretation fits naturally into our OCP framework.
Specifically, we can consider the following extension of $Q$,
which introduces the terminal time $\mathrm{T}$ as a new variable:
\begin{align}
\widetilde{Q}(t,{\mathbf{x}}_t,{\mathbf{u}}_t, \mathrm{T}) := {\widetilde{\Phi}(\mathrm{T}, {\mathbf{x}}(\mathrm{T}))+ \int_t^{\mathrm{T}} \ell(\tau,{\mathbf{x}}_\tau,{\mathbf{u}}_\tau) \text{ } {\textnormal{d}} \tau },
\label{eq:Q-T}
\end{align}
where $\widetilde{\Phi}(\mathrm{T}, {\mathbf{x}}(\mathrm{T}))$ explicitly imposes the penalty for longer integration time,
{\ignorespaces\emph{e.g.}}{ } $\widetilde{\Phi} := {\Phi}({\mathbf{x}}(\mathrm{T})) + \frac{c}{2}\mathrm{T}^2$.
Following a similar procedure presented in Section~\ref{sec:3.1},
we can transform (\ref{eq:Q-T}) into its ODE form (as in (\ref{eq:Q-ode}))
then characterize its local behavior (as in (\ref{eq:cddp-q}))
along a solution path $({ \bar{{\bm{x}}} }_t,{ \bar{{\bm{u}}} }_t,\bar{T})$.
After some tedious derivations, which are left in Appendix~\ref{app:3},
we will arrive at the update rule below,
\begin{align}
\mathrm{T} \leftarrow \bar{T} - \eta \cdot \delta \mathrm{T}(\delta \theta), \quad \text{where}\quad
\delta \mathrm{T}(\delta \theta)
= [\widetilde{Q}_{\bar{T}\bar{T}}(t_0)]^{-1}\left( \widetilde{Q}_{\bar{T}}(t_0) + \widetilde{Q}_{\bar{T}{ \bar{{\bm{u}}} }}(t_0)\delta \theta \right).
\label{eq:feedback}
\end{align}
Similar to what we have discussed in Section~\ref{sec:3.1},
one shall view $\widetilde{Q}_{\bar{T}}(t_0) \equiv \fracpartial{{\cal L}}{\mathrm{T}}$ as the first-order derivative w.r.t. the terminal time $\mathrm{T}$. Likewise,
$\widetilde{Q}_{\bar{T}\bar{T}}(t_0) \equiv \fracpartial{^2{\cal L}}{\mathrm{T}\partial\mathrm{T}}$, and etc.
Equation (\ref{eq:feedback}) is a second-order \emph{feedback} policy
that adjusts its updates based on the change of the parameter $\theta$.
Intuitively, it moves in the descending direction of the preconditioned gradient
({\ignorespaces\emph{i.e.}}{ } $\widetilde{Q}_{\bar{T}\bar{T}}^{-1}\widetilde{Q}_{\bar{T}}$),
while accounting for the fact that \emph{$\theta$ is also progressing during training}
(via the feedback $\widetilde{Q}_{\bar{T}{ \bar{{\bm{u}}} }}\delta\theta$).
The latter is a distinct feature arising from the OCP principle.
As we will show later,
this update (\ref{eq:feedback}) leads to distinct behavior with superior convergence compared to first-order baselines \citep{massaroli2020dissecting}.
\subsection{Dynamics of Higher-order Derivatives using Continuous-time Optimal Control Theory} \label{sec:3.1}
\colorlet{ggreen}{green!60!black}
\colorlet{rred}{red!80!black}
\setul{0.2ex}{0.2ex}
OCP perspective is a recently emerging methodology for analyzing optimization of discrete DNNs.
Central to its interpretation is to treat the layer propagation of a DNN as discrete-time dynamics,
so that the training process, {\ignorespaces\emph{i.e.}}{ } finding an
\setulcolor{ggreen}
\ul{\textit{optimal parameter}} of
\setulcolor{rred}
\ul{\textit{a DNN}},
can be understood like an OCP, which searches for an
\setulcolor{ggreen}
\ul{\textit{optimal control}} subjected to
\setulcolor{rred}
\ul{\textit{a dynamical constraint}}.
This perspective
has provided useful insights on
characterizing the optimization process \citep{hu2019mean} and enhancing
principled algorithmic design \citep{liu2021differential}.
We leave a complete discussion in Appendix~{\ref{app:1}}.
Lifting this OCP perspective from discrete DNNs to Neural ODEs %
requires special treatments from continuous-time OCP theory \citep{todorov2006optimal}.
Nevertheless, we highlight that
training Neural ODEs and solving continuous-time OCP
are fundamentally intertwined since these models, by construction, represent continuous-time dynamical systems.
Indeed, the ASM used for deriving (\ref{eq:grad-node},~\ref{eq:adjoint-node}) originates from the celebrated Pontryagin's principle \citep{pontryagin1962mathematical},
which is an optimality condition to OCP.
Hence, OCP analysis is not only {motivated but principled} from an optimization standpoint.
We begin by first transforming (\ref{eq:node-prob}) to a form
that is easier to adopt the continuous-time OCP analysis.
\begin{equation}
\begin{split}
\min_\theta \br{\Phi({\mathbf{x}}_{t_1}) + { \int_{t_0}^{t_1} } \ell(t,{\mathbf{x}}_t, {\mathbf{u}}_t) { \mathrm{d} t } } \quad \text{subjected to }
\begin{cases}
\fracdiff{{\mathbf{x}}_t}{t} = F(t, {\mathbf{x}}_t, {\mathbf{u}}_t), \quad & {\mathbf{x}}_{t_0} = {\bm{x}}_{t_0} \\
\fracdiff{{\mathbf{u}}_t}{t} = \mathbf{0} , \quad & {\mathbf{u}}_{t_0} = \theta
\end{cases},
\label{eq:ct-ocp}
\end{split}
\end{equation}
where ${\mathbf{x}}(t) \equiv {\mathbf{x}}_t$, and etc.
It should be clear that (\ref{eq:ct-ocp}) describes (\ref{eq:node-prob}) without loss of generality
by having $(\Phi, \ell) := ({\cal L}, 0)$.
These functions are known as the terminal and intermediate costs in standard OCP.
In training Neural ODEs, $\ell$ can be used to describe either the weight decay, {\ignorespaces\emph{i.e.}}{ } $\ell \propto \norm{{\mathbf{u}}_t}$, or more complex regularization \citep{finlay2020train}.
The time-invariant ODE imposed for ${\mathbf{u}}_t$ makes the ODE of ${\mathbf{x}}_t$ equivalent to (\ref{eq:node}).
Problem (\ref{eq:ct-ocp}) shall be understood as
a particular type of OCP that searches for an optimal initial condition $\theta$ of a time-invariant control ${\mathbf{u}}_t$.
Despite seemly superfluous,
this is a necessary transformation that enables rigorous OCP analysis for the original training process (\ref{eq:node-prob}),
and it has also appeared in other control-related analyses \citep{zhong2020symplectic,chalvidal2021go}.
Next, define the accumulated loss from any time $t \in [t_0,t_1]$ to the integration end time $t_1$ as
\begin{align}
Q(t,{\mathbf{x}}_t,{\mathbf{u}}_t) := {\Phi({\mathbf{x}}_{t_1})+ \int_t^{t_1} \ell(\tau,{\mathbf{x}}_\tau,{\mathbf{u}}_\tau) \text{ } {\textnormal{d}} \tau },
\label{eq:Q}
\end{align}
which is also known in OCP as the \emph{cost-to-go} function.
Recall that our goal is to compute higher-order derivatives w.r.t. the parameter $\theta$ of Neural ODEs.
Under the new OCP representation (\ref{eq:ct-ocp}), the first-order derivative $\fracpartial{{\cal L}}{\theta}$ is identical to
$\fracpartial{Q(t_0, {\mathbf{x}}_{t_0}, {\mathbf{u}}_{t_0})}{{\mathbf{u}}_{t_0}}$.
This is because $Q(t_0, {\mathbf{x}}_{t_0}, {\mathbf{u}}_{t_0})$ accumulates all sources of losses between $[t_0,t_1]$ (hence it sufficiently describes ${\cal L}$)
and ${\mathbf{u}}_{t_0} = \theta$ by construction.
Likewise,
the second-order derivatives can be captured by the Hessian
$\fracpartial{^2 Q(t_0, {\mathbf{x}}_{t_0}, {\mathbf{u}}_{t_0})}{{\mathbf{u}}_{t_0}\partial{\mathbf{u}}_{t_0}} = \fracpartial{^2{\cal L}}{\theta\partial\theta} \equiv {\cal L}_{\theta\theta}$.
In other words,
we are only interested in obtaining the derivatives of $Q$ at the integration start time $t_0$.
To obtain these derivatives,
notice that we can rewrite (\ref{eq:Q}) as
\begin{align}
{0 = {\ell(t,{\mathbf{x}}_t,{\mathbf{u}}_t) + \fracdiff{Q(t,{\mathbf{x}}_t,{\mathbf{u}}_t)}{t}}}, \quad
Q({t_1},{\mathbf{x}}_{t_1}) = \Phi({\mathbf{x}}_{t_1}),
\label{eq:Q-ode}
\end{align}
since the definition of $Q$ implies that
$Q(t,{\mathbf{x}}_t,{\mathbf{u}}_t) ={\ell(t,{\mathbf{x}}_t,{\mathbf{u}}_t) { \mathrm{d} t } + Q(t+{ \mathrm{d} t },{\mathbf{x}}_{t+{ \mathrm{d} t }},{\mathbf{u}}_{t+{ \mathrm{d} t }})}$.
We now state our main result,
which provides a local characterization of (\ref{eq:Q-ode}) with a set of coupled ODEs expanded along a solution path.
These ODEs can be used to obtain {all second-order derivatives at $t_0$}.
\begin{theorem}[Second-order Differential Programming] \label{prop:1}
Consider a solution path $({ \bar{{\bm{x}}} }_t, { \bar{{\bm{u}}} }_t)$
that solves the ODEs in (\ref{eq:ct-ocp}).
Then the first and second-order derivatives of $Q(t,{\mathbf{x}}_t,{\mathbf{u}}_t)$,
expanded locally around this solution path,
obey the following backward ODEs:
\begin{subequations}
\begin{alignat}{2}
- \fracdiff{{Q_{\hvx}}}{t} &= {\ell_{\hvx} } + {{F}_\hvx^\T}{Q_{\hvx}}, \quad
&&\text{ }\text{ }- \fracdiff{{Q_{\hvu}}}{t} = {\ell_{\hvu} } + {{F}_\hvu^\T}{Q_{\hvx}}, \label{eq:cddp-qa} \\
- \fracdiff{{Q_{\hvx \hvx}}}{t} &= {{\ell}_{\hvx \hvx}} + {{F}_\hvx^\T}{Q_{\hvx \hvx}}+{Q_{\hvx \hvx}}{{F}_\hvx}, \quad
&&- \fracdiff{{Q_{\hvx \hvu}}}{t} = {{\ell}_{\hvx \hvu}} + {Q_{\hvx \hvx}}{{F}_\hvu} + {{F}_\hvx^\T}{Q_{\hvx \hvu}}, \label{eq:cddp-qb} \\
- \fracdiff{{Q_{\hvu \hvu}}}{t} &= {{\ell}_{\hvu \hvu}} + {{F}_\hvu^\T}{Q_{\hvx \hvu}} + {Q_{\hvu \hvx}}{{F}_\hvu}, \quad
&&- \fracdiff{{Q_{\hvu \hvx}}}{t} = {{\ell}_{\hvu \hvx}} + {{F}_\hvu^\T}{Q_{\hvx \hvx}} + {Q_{\hvu \hvx}}{{F}_\hvx}, \label{eq:cddp-qc}
\end{alignat} \label{eq:cddp-q}
\end{subequations}
where
${{F}_\hvx}(t) {\equiv} \fracpartial{F}{{\mathbf{x}}_t}|_{({ \bar{{\bm{x}}} }_t,{ \bar{{\bm{u}}} }_t)}$,
${Q_{\hvx \hvx}}(t) {\equiv} \fracpartial{^2Q}{{\mathbf{x}}_t\partial{\mathbf{x}}_t}|_{({ \bar{{\bm{x}}} }_t,{ \bar{{\bm{u}}} }_t)}$, and etc.
All terms in (\ref{eq:cddp-q})
are time-varying vector-valued or matrix-valued functions expanded at $({ \bar{{\bm{x}}} }_t, { \bar{{\bm{u}}} }_t)$.
The terminal condition is given by
\begin{align*}
{Q_{\hvx}}(t_1) = \Phi_{ \bar{{\bm{x}}} }, \quad
{Q_{\hvx \hvx}}(t_1) = \Phi_{{ \bar{{\bm{x}}} }\hvx}, \quad \text{ and } \quad
{Q_{\hvu}}(t_1) = {Q_{\hvu \hvu}}(t_1) = {Q_{\hvu \hvx}}(t_1) = {Q_{\hvx \hvu}}(t_1) = \mathbf{0}.
\end{align*}
\end{theorem}
The proof (see Appendix~\ref{app:2}) relies on rewriting (\ref{eq:Q-ode})
with \emph{differential states},
${\delta {\mathbf{x}}}_t := {\mathbf{x}}_t - { \bar{{\bm{x}}} }_t$,
which view the deviation from ${ \bar{{\bm{x}}} }_t$ as an \emph{optimizing variable}
(hence the name ``\emph{Differential Programming}'').
It can be shown that ${\delta {\mathbf{x}}}_t$ follows a linear ODE expanded along the solution path.
Theorem \ref{prop:1} has several important implications.
First, the ODEs in (\ref{eq:cddp-qa}) recover the original ASM computation (\ref{eq:grad-node},\ref{eq:adjoint-node}),
as one can readily verify that
${Q_{\hvx}}(t) \equiv {\mathbf{a}}(t)$ follows the same backward ODE in~(\ref{eq:adjoint-node})
and the solution of the second ODE in (\ref{eq:cddp-qa}), $Q_{{ \bar{{\bm{u}}} }}(t_0) = - { \int^{t_0}_{t_1} } {{F}_\hvu}^\mathsf{T}{Q_{\hvx}} { \mathrm{d} t } $,
gives the exact gradient in~(\ref{eq:grad-node}).
Meanwhile,
solving the coupled matrix ODEs presented in (\ref{eq:cddp-qb}, \ref{eq:cddp-qc})
yields the desired second-order matrix, ${Q_{\hvu \hvu}}(t_0) \equiv {\cal L}_{\theta\theta}$, for preconditioning the update.
Finally,
one can derive the dynamics of other higher-order tensors using the same Differential Programming methodology
by simply expanding (\ref{eq:Q-ode}) beyond the second order.
We leave some discussions in this regard in Appendix~\ref{app:2}.
|
\section{Introduction}
Museums, as one of the most important educational institutions outside schools, aim to exhibit and introduce historical and cultural artifacts from multi-sectors to the general public. However, in a time when most students are visual learners and are familiar with games, they are more attracted by and engaged in the content with active exploration and interaction \cite{Jukes01}. The young generation has gradually become impatient with traditional modes of instructions, such as using text or audio explanations, and disengaged with these conventional paradigms of exploration and knowledge acquisition \cite{eck06}. Today's museums, like other learning institutions, are trying to keep with technological and social changes and looking into ways to transform inactive exhibitions into interactive experiences to better meet their audiences' preferences, especially those of younger generations. Game-based learning is one of the interactive approaches that can satisfy their needs. A well-designed game-based learning context would balance gaming and learning and can benefit players twofold by (1) motivating them to learn the knowledge with a high level of engagement with the enjoyable game content \cite{Prensky01}, and (2) guiding them with a deeper understanding of complex settings in the games \cite{Gros01}. These advantages of game-based learning make games suitable to be embedded in the context of museums to make the exploration of artifacts active and engaging.
In the discovery of ancient artifacts, uncertainty is always a challenge for archaeologists. Excavation or unearthing of relics is one of the most important and exciting parts to eliminate uncertainty. Based on this feature, Henan Museum has launched "the archaeological blind box," a small brick with a random mini-artifact hidden inside it. It requires the visitors to use mini tools to sweep the dirt away to discover and collect the hidden artifact. By offering the young generation opportunities to stimulate them to explore the relics and implicitly encourage them to learn cultural knowledge, the blind boxes started a wave of amateur archaeology in China with 170,000 pieces sold from December 2020 to April 2021 \cite{labbrand01, Xinhua01}. On the other hand, uncertainty is also an essential element in video games. Well-designed uncertainty elements in a game could motivate players to start and continuously keep players engaged in the game \cite{Caillois01, Costikyan15, Johnson18, Power19}. The combination of archaeology and gaming provides a chance to blend knowledge and uncertainty together in a game to motivate players to acquire knowledge and enjoy the learning process.
Virtual reality (VR) games can be an ideal medium to achieve such a combination and enhance understanding of cultural knowledge through an immersive, active, and fun experience. VR expands the space and affordances of real-world museums and transforms them into immersive virtual environments that can incorporate creative and imaginative elements. 3D virtual gaming worlds can provide unlimited interactive possibilities and experiences at a low cost. People can access these experiences remotely, which is valuable when there are restrictions on time, cost, or distance, especially during the COVID-19 period. With its active experimentation and high immersion, VR games can enhance constructive, situated, and experimental learning \cite{Silva2017}.
We built a VR archaeology educational game, RelicVR, to enrich the museum visiting and learning experience. Our game provides a game-based learning environment, which emulates a simplified archaeological process of discovering ancient artifacts but with innovative gameplay elements. The archaeological process involves a set of actions such as chipping and digging away the earth that envelops a hidden artifact (see Figure \ref{fig:teaser}). The players need to wield tools in their hands and move around to chip away the clod of earth that wraps the relic. Since players initially have no knowledge of the artifact inside the clod, the uncertainty during gameplay is intended to heighten their senses and focus on the chipping process to prevent damaging the artifact. In short, RelicVR aims to make the exploration of ancient artifacts an active, fun, and engaging process. We conducted a short playtesting with eight players to evaluate the effects of gaming for learning about relics. Results showed that the game can provide a fun and engaging experience to players and is still be able to help them learn the features of the artifacts.
\section{Related Work}
\subsection{Games for Archaeology}
Developments in the field of archaeology have changed the way we present archaeological artifacts and discoveries today. The growing interest in digitalization and turning the process interactive has a considerable appeal to archaeologists wishing to present their research to the public. Early in 2002, Watrall \cite{Watrall02} stated that archaeologists must realize the potency of interactive media and embrace it. The educational potential of interactive games has been increasingly addressed over the past years, especially for pre-teen learners, as game-based learning can keep them more engaged and make their efficient learning \cite{eck06}. Some prior research has shown the overall success of applying educational games to enable students to learn about the archaeological investigation (e.g. \cite{Ardito12,Georgiadi}). At the University of Illinois, the VRchaelogy Lab \cite{ShacHuanCrai2018ra, ShacHuanCrai2019, Ramy16} had some surprising findings of how VR environments can enhance the learning of basic archaeology knowledge to college students and non-professionals.
\subsection{Games for Museums}
Efficient approaches that balance educational purposes and entertainment are desired in today's museums. The use of video games can increase general interests, especially from young generations, in exploring cultural artifacts and improve their experience \cite{Varinlioglu19}. Studies have indicated that a gamified application can increase participants' involvement, motivation, enjoyment, and overall exploratory experiences \cite{Brigham15, Seaborn15}.
Recent advances in VR have led to new possibilities in museum touring because of its capability to create immersive artificial worlds. There have been projects in which VR is used to recreate virtual living spaces in museums \cite{George03, Viking, BackToLife}. Jung et al. \cite{Jung01} found that entertainment experienced from VR can enhance the overall museum visits. However, most of the existing games for archaeology lack engaging interaction between players and the virtual environments. There are nearly no games that introduce artifacts from an archaeologist's discovery perspective. We argue that the scenery in VR alone is not enough to engage players and arouse their interests in the cultural heritage. In \textit{The Grasshopper: Games, Life, and Utopia Suits} \cite{cooper82}, Cooper defines the playing of a game as “the voluntary attempt to overcome unnecessary obstacles." To lead players to this self-motivated act, it is also necessary to include appropriate methods in the design of a successful game experience. Compared with existing games (e.g., \cite{ShacHuanCrai2018ra, ShacHuanCrai2019, Ramy16}), which are somewhat more oriented to teach more serious concepts and emphasize more on the learning side than on gameplay, RelicVR has more gaming features and is intended to be more interactive, encourage exploratory engagement and focus, and allow learning to be fun.
\subsection{Uncertainty}
Uncertain elements can make the gameplay experience engaging, enjoyable, and playful \cite{Caillois01, Costikyan15, Kumari2019}. According to Costikyan \cite{Costikyan15}, typical uncertain sources in a game include performative uncertainty, solver’s uncertainty, player unpredictability, and randomness. In terms of VR content, Xu et al. \cite{xu21} evaluated the effect of uncertainty in a VR exergame. They found that uncertainty could improve the exergame experience by increasing players’ energy exertion. We believe involving uncertainty elements in our game would give players a better gaming and learning experience. To achieve this effect, we obscure the object with a clod of earth that wraps the object entirely. Since the players do not know the structure and features of the hidden object, they need to carefully plan their actions and develop reasonable speculations based on the parts that have been revealed to avoid hitting the relic and damage it. In addition, if the object is hit, the game provides a dissonant sound, akin to negative feedback---something that players would want to avoid. According to Costikyan's classification, our approach could be classified as performative uncertainty because there is uncertainty in the outcome of the physical actions of the players and the gradual unearthing of the artifact. As the players reveal more hidden parts of the artifact, they are also going through a process of verifying their hypotheses or speculations in an iterative manner and continuously update their knowledge of the artifact.
\section{RelicVR}
We designed and developed RelicVR, a first-person perspective VR game that combines active exploration and knowledge acquisition of archaeological artifacts. The game intends to enrich the learning experience with such artifacts by embedding them into a game that simulates a simplified procedure of archaeological discovery. RelicVR is made to be run in the Oculus Rift S. It was developed using C\# on the Unity3D platform (version 2020.3.5f1) with the Oculus Integration SDK (28.0.0). In this section, we introduce its overall concepts, game world scenario, and interactions in the virtual environment (VE).
\subsection{Concept}
The principle behind RelicVR was inspired by archaeological activities in the real world. Instead of implementing the whole archaeological investigation approach, we mainly adapted the excavation phase in the game due to its interactive nature. In RelicVR, players start with little information of what is hidden in the clod of earth. When they use tools to remove the chunks of earth from the clod, they face a sense of uncertainty due to the unknown shape, structure, and features of the relic. As such, they need to be focused on the actions and observe the outcome of each physical activity. Such a process can be iteratively reinforced as more parts of the relic come to light.
We set up a few invisible trigger points in different places around the relic according to its features. When players hit a trigger point, a dialog box with text describing some features would appear and is accompanied with audio to capture players' attention. As more and more details of the artifact are exposed, players gradually increase their understanding by reading or listening to the prompted descriptions and careful observations of its visual details, now visible to the naked eye. In this first version of the game, we designed three dialog boxes for each relic to introduce it from different perspectives.
RelicVR aims to drive players to engage in active exploration and learning. It gives the players a similar role to an archaeologist to explore the relic actively. The players start out by observing a large clod of earth placed at the center of the VE, but can only have a limited understanding of the relic (see Figure~\ref{fig:relicvr}a). Driven by curiosity, the players are encouraged to discover what lies inside the clod. The game requires physical motions in the form of digging with selected tools and walking around to try to unearth the artifact within a limited time. This process requires continuous planning and problem-solving and frequent updates to the mental model of the object. During the exploration process, they also face similar challenges like archaeologists---they need to carefully plan their actions to avoid damaging the artifacts. After players have successfully excavated the artifact, they can get general information from an additional dialog box and put the relic into their collection. While not implemented in the current version, the collection can be shown to their friends, which can give them a sense of achievement.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{relicvrfigure2.PNG}
\caption{Game elements in RelicVR. (a) Players will see a clod at the beginning of the game. (b) Game environment is a mysterious temple to further immersive players in the activity. (c) Two available tools (a hammer and a shovel) to be used to excavate the artifact. (d) and (e) are two relics included in the first version of RelicVR. They are the Arhat (Luohan) sculpture and Sanxingdui bronze head with a gold mask, respectively.}
\Description{This figure provides some screenshots of the first version of RelicVR including the clod that hides the artifact, the game environment, digging tools, and two sample relics. }
\label{fig:relicvr}
\end{figure*}
\subsection{Game World}
In the first version of the game, we used a mysterious temple (Figure~\ref{fig:relicvr}b) as the gaming environment to increase the adaptability of artifacts from different eras and cultures. The temple does not refer to any cognizable mono-cultural motif. The daylight in the scene moves around the temple to give players a sense of time elapsing as the game progresses. Several spotlights for illuminating the artifact are hung high in the temple to focus players' attention. Fallen walls and rocks are randomly placed outside the temple to give a sense of ancientness and historical allure and mystery to the environment. As the players look around, they can see that the scene is a semi-enclosed space containing tools on a table and an unexplored artifact encased in a cubic clod of earth.
Two relics are included in the current version of RelicVR. The first one is the Arhat (Luohan) Sculpture from the Metropolitan Museum of Art in New York (Figure~\ref{fig:relicvr}d, referred to as Arhat hereafter) \cite{Luohan}. The second relic is the bronze head with a gold mask from Sanxingdui Museum in China (Figure~\ref{fig:relicvr}e, referred to as Gold Mask hereafter) \cite{GoldenMask}. We extracted the information from their official website and integrated it into the game. These two relics are scaled to approximately the same volume to minimize the differences in the gaming experience. Their sizes are about a human's height to increase the level of exploration. In addition, it provides players a chance of detailed observations of the relic which is less likely to be achieved in conventional museum visits.
The players can choose a hammer or a shovel on the workbench based on their preference, as shown in Figure~\ref{fig:relicvr}c. The hammer makes players feel more powerful and more flexible to handle, while the shovel is thinner but has a broad blade. The shovel can easily remove large areas of clod chunks but also increases the risk of hitting the artifacts. This is a trade-off between the two tools that the players need to balance carefully.
\subsection{Interactions}
The interactions in RelicVR mimic real-world actions to increase immersion and physical activity. Players need to grab one of the tools by pressing the grip button on the controller after pointing the virtual hand at the tool at a short distance or remotely at a longer distance. To simulate how digging is performed in real life, players are required to hold the grip button during the whole excavation activity. Players need to wave their arms to dig. The actions in the VE match the players' actions in the physical world. Since the artifacts and clod are designed to be of human size, players might need to raise their arms and squat to clean the dirt at the top and bottom regions.
In the virtual world, the players are encouraged to move physically around the clod enclosure to achieve coordinated movements in the virtual scene. The players can also use the joysticks on the controllers to move or turn around. However, such movements or turnings are only suggested when the player needs to make large positional adjustments in the VE.
Similar to real archaeological excavations, the players should control their actions carefully to avoid damaging the relic during the excavation process. Warnings are given using a (dissonant) sound that simulates striking the artifact. The choice of the negative sound feedback is to encourage players to avoid hearing it by not hitting the artifact.
The levels of difficulty in RelicVR can vary greatly according to the complexity of shapes and volume. For the current version, we selected models of similar complexity and calculated their approximate volume. We set a health bar and timer for the relic as elements of the game: the reduction of the health bar and timer has no effect on the artifacts and gameplay but would create tension and motivation for players to complete the task within the allotted time and minimize the reduction of health points. The health bar hovers above the artifact and always faces the player's view. A strike on the relic would reduce 1 point. Time begins to count down from when the game starts and is shown on the upper middle part of the screen. We found that for the current models, 7 minutes of playtime and 40 health are suitable for most players to keep a balance between having a degree of pressure and feeling motivated to complete the task.
\subsection{Innovation}
Recent advancements in graphical computation power and development of the marching cubes algorithm \cite{lorensen87} allow the deformation of objects that are non-trivial and are very detailed in real-time. To our knowledge, RelicVR is the first game that explores the use of dynamic voxel functions in VR to combine active gaming and exploratory learning together. Furthermore, instead of exploring a vast area of a VE such as \cite{DeepRock, Harbin}, RelicVR attempts to focus players in active explorations of a human-scale 3D area. The physical interaction afforded by VR devices enables physical activity by players during gameplay. Furthermore, the inclusion of uncertainty elements in the game can make players immersed and focused on playing the game.
\section{Playtesting}
\subsection{Experiment}
We have conducted a preliminary playtest with eight players (3 females and 5 males, aged between 19 to 25; \textit{M} = 21.5, \textit{SD} = 2.07). They were recruited from a local university. Half of them experienced the Arhat version, while the remaining experienced the Gold Mask version. The players were first introduced to the game and the use of the devices. They were asked to use tools to excavate the artifact enclosed within the clod. In addition, they were told to keep in mind the time they had and try to avoid damaging the artifact by not hitting it during the excavation process. At the same time, they needed to remember the details from the dialog boxes that would appear after discovering a key point, if possible. After experiencing the game, they were asked to complete a quiz with 6 fill-in-the-blank questions to assess how much they remembered after playing the game. Answers to these questions were derived from the content of the dialog boxes. In addition, we gave them two questionnaires, a core module of the Game Experience Questionnaire (GEQ) \cite{IJsselsteijn} and an adapted version of web-based learning tools evaluation scales (WBLT) \cite{Kay11} to elicit their perceived gaming and learning experience with RelicVR. At the end of the experiment, we conducted a semi-structured interview asking them about their feelings and thoughts about the game.
\subsection{Results and Discussion}
Overall, the eight players were satisfied with the experience in RelicVR. Figure~\ref{fig:questionnaire} summarizes the results of GEQ and WBLT. Based on the ratings, the players perceived a positive gaming experience in RelicVR in terms of competence, immersion, flow, and positive affect (the ratings in these categories were all above neutral), while they felt less tension/annoyance and negative affect (below neutral). On average, players gave lower-middle ratings in the challenge category, which shows a moderate level of difficulty in the game. Based on our results, RelicVR also provided a good learning experience. All eight players were able to answer most of the questions correctly (\textit{M} = 4.88, \textit{SD} = 0.83) after experiencing the game. As shown in Figure~\ref{fig:questionnaire}b, the players perceived RelicVR as a good learning tool in learning, design, and engagement (all above neutral).
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{Questionnaire.PNG}
\caption{Boxplots of the questionnaire results. (a) GEQ on a scale of 0 to 4. (b) WBLT on a scale of 1 to 5. For both questionnaires, the higher score represents a better experience in terms of the specific component. Scales representing neutral feeling are highlighted in both sub-figures. Component with asterisk symbol represents a reverse rating.}
\Description{Figure labelled 'a' shows the ratings of the 33-item core module of the GEQ. The ratings of the items were grouped into 7 categories: competence, immersion, flow, tension/annoyance, challenge, negative affect, and positive affect. Overall, RelicVR received positive feedback about game experience. Figure labelled 'b' shows the ratings of the 13-item WBLT evaluation scales. The ratings were grouped into 3 categories: learning, design and engagement. Most of the players gave high ratings to RelicVR.}
\label{fig:questionnaire}
\end{figure*}
During the interview, all eight players reported that RelicVR was more interesting than conventional museum visits, where artifacts could only be seen from a distance and are enclosed behind glass windows, a process that is typically non-interactive and impersonal. On the other hand, four players (P2, P4, P5, P6) mentioned that in RelicVR, the process of excavating by hand forced them to be close to the relic and be focused on it. Coupled with the assistance of text and audio introduction, they reported a deep impression of the relic and its information. P4 said, ``\textit{The dialog boxes at different key points were good to let me know about the artifact. When they came out, I could have some rest and time to watch and appreciate their physical appearance, of which I still have a clear memory after the game.}'' P3 and P7 said that they enjoyed the uncertainty in the exploration process. Based on our observations, they were always looking forward to seeing what the artifact would look like and trying carefully not to hit the object when clearing the earth.
P1 and P2 said that the game should make more explicit the connections between playing in the game and learning and have a better balance between them, though they thought RelicVR was a great learning tool. They mentioned that, by separating the description of the relics into multiple dialog boxes connected to the corresponding components of the relic, they got a better learning experience than reading a long description in one single place. However, they also argued that the learning experience should be further improved. Externally, because the timer was continuously counting down, they felt somewhat anxious to excavate the unfinished parts. Internally, they could not wait to explore more about the relic because the gaming elements were more appealing to them. To a certain extent, balancing the entertainment and education elements is often a challenge in games for learning. To provide a better learning experience, we plan to explore when the best time is to deliver knowledge related details in RelicVR; it should either be embedded in the game just like the current version, or be provided before or after the gaming part. On the other hand, it is also necessary to compare and derive the most suitable completion time which can give tension or pressure to players to complete the tasks but at the same not to let them feel overwhelmed or overly anxious. P4 expected more tools in the game, especially the tools that can be used to do elaborate and more precise excavations. In short, while improvements can be made to RelicVR, the players seem to have enjoyed the combination of active exploration and uncertainty of the discovery process that the game brings.
\section{Future Work}
For macro-level interaction, we plan to include the possibility of collecting the artifacts that players unearth as a first priority. As a shared value of professional archaeologists and Henan Museum's blind boxes show, the function will enhance the sense of achievement of players and could motivate them to play the game in the long-term, as the game could include a wide range of artifacts, some of which could be released for special occasions and in a limited way. Having more artifacts can also be used to provide levels in the game. These artifacts could be from different cultures of various sizes and shape complexities which would require players to perform more demanding activities and digging patterns to unearth them.
Moreover, a richer micro-level interaction is desirable. We plan to include more excavation tools and vary their effects on the clod to provide diverse experiences for players. In addition, to increase the level of difficulty and challenge, we can follow a similar approach to the different block types in Minecraft \cite{Minecraft}. We can apply various materials and mix them into the clod enclosure. Different mixes will lead players to use different strategies and elicit greater levels of exertion during gameplay since the required energy and feedback are different for each mix. On the other hand, a recent study has proposed an attachable device to the VR handheld controllers to provide weight and center of gravity simulation in VR \cite{monteiro2021icmi}. We think such device or the concept can be used in RelicVR to provide force feedback on users' hands or arms to simulate real excavation actions.
From a research perspective, RelicVR can be an ideal testbed for evaluating the effect of unknown and uncertain factors on gameplay and learning experiences in VR. In addition, it can be used to explore the effect of exertion levels \cite{xu21,xu2020chiplay}, body positions \cite{xu2020jmir}, and viewing perspectives \cite{xu2020g4h,monteiro2020ismar}. In addition, research has looked into the effects of display types for exergames \cite{xu21, xu2020g4h}. It is also interesting to evaluate and compare the experience and outcomes of game-based learning between VR and traditional displays (i.e., immersive vs. less-immersive experiences).
Finally, we also plan to refine other aspects of the game like optimizing further of completion condition of each artifact, letting players customize completion time, and adding a health bar system to balance game challenge and gameplay experience.
On a business level, RelicVR has great potential to be used in cooperation with museums to enhance the visitors' experience and promote popular science education. We have plans to contact local museums to explore ways that RelicVR can be used in any of their processes and enhance their visitors' experiences.
\section{Conclusion}
In this paper, we presented RelicVR, the first VR game that encourages players to discover historical artifacts through physical interaction and a game-based approach. By using dynamic voxel deformation in the game, it was able to add uncertain elements in the exploratory process---e.g., the unearthing of the artifact in a gradual but calibrated way to avoid hitting and damaging it. Our playtesters found that the uncertainty in the process helped enhance their experience. Also, the innovative interaction method has the potential to improve players' learning motivation and outcomes. Overall, RelicVR is an innovative game that can engage and immerse players in active exploration of archaeological artifacts.
\begin{acks}
The original inspiration of RelicVR came from a visit to Henan Museum invited by Mr. Haosen Zhao, a passionate young archaeologist. We want to express our thanks to him and wish him success in pursuing his dream to contribute to the field of archaeology. The authors also want to thank the participants who helped playtest our game and the reviewers for their insightful comments that have helped improved our paper. This work was supported in part by Xi'an Jiaotong-Liverpool University's Key Special Fund (\#KSF-A-03).
\end{acks}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
ACM's consolidated article template, introduced in 2017, provides a
consistent \LaTeX\ style for use across ACM publications, and
incorporates accessibility and metadata-extraction functionality
necessary for future Digital Library endeavors. Numerous ACM and
SIG-specific \LaTeX\ templates have been examined, and their unique
features incorporated into this single new template.
If you are new to publishing with ACM, this document is a valuable
guide to the process of preparing your work for publication. If you
have published with ACM before, this document provides insight and
instruction into more recent changes to the article template.
The ``\verb|acmart|'' document class can be used to prepare articles
for any ACM publication --- conference or journal, and for any stage
of publication, from review to final ``camera-ready'' copy, to the
author's own version, with {\itshape very} few changes to the source.
\section{Template Overview}
As noted in the introduction, the ``\verb|acmart|'' document class can
be used to prepare many different kinds of documentation --- a
double-blind initial submission of a full-length technical paper, a
two-page SIGGRAPH Emerging Technologies abstract, a ``camera-ready''
journal article, a SIGCHI Extended Abstract, and more --- all by
selecting the appropriate {\itshape template style} and {\itshape
template parameters}.
This document will explain the major features of the document
class. For further information, the {\itshape \LaTeX\ User's Guide} is
available from
\url{https://www.acm.org/publications/proceedings-template}.
\subsection{Template Styles}
The primary parameter given to the ``\verb|acmart|'' document class is
the {\itshape template style} which corresponds to the kind of publication
or SIG publishing the work. This parameter is enclosed in square
brackets and is a part of the {\verb|documentclass|} command:
\begin{verbatim}
\documentclass[STYLE]{acmart}
\end{verbatim}
Journals use one of three template styles. All but three ACM journals
use the {\verb|acmsmall|} template style:
\begin{itemize}
\item {\verb|acmsmall|}: The default journal template style.
\item {\verb|acmlarge|}: Used by JOCCH and TAP.
\item {\verb|acmtog|}: Used by TOG.
\end{itemize}
The majority of conference proceedings documentation will use the {\verb|acmconf|} template style.
\begin{itemize}
\item {\verb|acmconf|}: The default proceedings template style.
\item{\verb|sigchi|}: Used for SIGCHI conference articles.
\item{\verb|sigchi-a|}: Used for SIGCHI ``Extended Abstract'' articles.
\item{\verb|sigplan|}: Used for SIGPLAN conference articles.
\end{itemize}
\subsection{Template Parameters}
In addition to specifying the {\itshape template style} to be used in
formatting your work, there are a number of {\itshape template parameters}
which modify some part of the applied template style. A complete list
of these parameters can be found in the {\itshape \LaTeX\ User's Guide.}
Frequently-used parameters, or combinations of parameters, include:
\begin{itemize}
\item {\verb|anonymous,review|}: Suitable for a ``double-blind''
conference submission. Anonymizes the work and includes line
numbers. Use with the \verb|\acmSubmissionID| command to print the
submission's unique ID on each page of the work.
\item{\verb|authorversion|}: Produces a version of the work suitable
for posting by the author.
\item{\verb|screen|}: Produces colored hyperlinks.
\end{itemize}
This document uses the following string as the first command in the
source file:
\begin{verbatim}
\documentclass[sigconf]{acmart}
\end{verbatim}
\section{Modifications}
Modifying the template --- including but not limited to: adjusting
margins, typeface sizes, line spacing, paragraph and list definitions,
and the use of the \verb|\vspace| command to manually adjust the
vertical spacing between elements of your work --- is not allowed.
{\bfseries Your document will be returned to you for revision if
modifications are discovered.}
\section{Typefaces}
The ``\verb|acmart|'' document class requires the use of the
``Libertine'' typeface family. Your \TeX\ installation should include
this set of packages. Please do not substitute other typefaces. The
``\verb|lmodern|'' and ``\verb|ltimes|'' packages should not be used,
as they will override the built-in typeface families.
\section{Title Information}
The title of your work should use capital letters appropriately -
\url{https://capitalizemytitle.com/} has useful rules for
capitalization. Use the {\verb|title|} command to define the title of
your work. If your work has a subtitle, define it with the
{\verb|subtitle|} command. Do not insert line breaks in your title.
If your title is lengthy, you must define a short version to be used
in the page headers, to prevent overlapping text. The \verb|title|
command has a ``short title'' parameter:
\begin{verbatim}
\title[short title]{full title}
\end{verbatim}
\section{Authors and Affiliations}
Each author must be defined separately for accurate metadata
identification. Multiple authors may share one affiliation. Authors'
names should not be abbreviated; use full first names wherever
possible. Include authors' e-mail addresses whenever possible.
Grouping authors' names or e-mail addresses, or providing an ``e-mail
alias,'' as shown below, is not acceptable:
\begin{verbatim}
\author{Brooke Aster, David Mehldau}
\email{dave,judy,<EMAIL>}
\email{<EMAIL>}
\end{verbatim}
The \verb|authornote| and \verb|authornotemark| commands allow a note
to apply to multiple authors --- for example, if the first two authors
of an article contributed equally to the work.
If your author list is lengthy, you must define a shortened version of
the list of authors to be used in the page headers, to prevent
overlapping text. The following command should be placed just after
the last \verb|\author{}| definition:
\begin{verbatim}
\renewcommand{\shortauthors}{McCartney, et al.}
\end{verbatim}
Omitting this command will force the use of a concatenated list of all
of the authors' names, which may result in overlapping text in the
page headers.
The article template's documentation, available at
\url{https://www.acm.org/publications/proceedings-template}, has a
complete explanation of these commands and tips for their effective
use.
Note that authors' addresses are mandatory for journal articles.
\section{Rights Information}
Authors of any work published by ACM will need to complete a rights
form. Depending on the kind of work, and the rights management choice
made by the author, this may be copyright transfer, permission,
license, or an OA (open access) agreement.
Regardless of the rights management choice, the author will receive a
copy of the completed rights form once it has been submitted. This
form contains \LaTeX\ commands that must be copied into the source
document. When the document source is compiled, these commands and
their parameters add formatted text to several areas of the final
document:
\begin{itemize}
\item the ``ACM Reference Format'' text on the first page.
\item the ``rights management'' text on the first page.
\item the conference information in the page header(s).
\end{itemize}
Rights information is unique to the work; if you are preparing several
works for an event, make sure to use the correct set of commands with
each of the works.
The ACM Reference Format text is required for all articles over one
page in length, and is optional for one-page articles (abstracts).
\section{CCS Concepts and User-Defined Keywords}
Two elements of the ``acmart'' document class provide powerful
taxonomic tools for you to help readers find your work in an online
search.
The ACM Computing Classification System ---
\url{https://www.acm.org/publications/class-2012} --- is a set of
classifiers and concepts that describe the computing
discipline. Authors can select entries from this classification
system, via \url{https://dl.acm.org/ccs/ccs.cfm}, and generate the
commands to be included in the \LaTeX\ source.
User-defined keywords are a comma-separated list of words and phrases
of the authors' choosing, providing a more flexible way of describing
the research being presented.
CCS concepts and user-defined keywords are required for for all
articles over two pages in length, and are optional for one- and
two-page articles (or abstracts).
\section{Sectioning Commands}
Your work should use standard \LaTeX\ sectioning commands:
\verb|section|, \verb|subsection|, \verb|subsubsection|, and
\verb|paragraph|. They should be numbered; do not remove the numbering
from the commands.
Simulating a sectioning command by setting the first word or words of
a paragraph in boldface or italicized text is {\bfseries not allowed.}
\section{Tables}
The ``\verb|acmart|'' document class includes the ``\verb|booktabs|''
package --- \url{https://ctan.org/pkg/booktabs} --- for preparing
high-quality tables.
Table captions are placed {\itshape above} the table.
Because tables cannot be split across pages, the best placement for
them is typically the top of the page nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and the
table caption. The contents of the table itself must go in the
\textbf{tabular} environment, to be aligned properly in rows and
columns, with the desired horizontal and vertical rules. Again,
detailed instructions on \textbf{tabular} material are found in the
\textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed output of
this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more
desirable. Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
Always use midrule to separate table header rows from data rows, and
use it only for this purpose. This enables assistive technologies to
recognise table headers and support their users in navigating tables
more easily.
\section{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of the three are
discussed in the next sections.
\subsection{Inline (In-text) Equations}
A formula that appears in the running text is called an inline or
in-text formula. It is produced by the \textbf{math} environment,
which can be invoked with the usual
\texttt{{\char'134}begin\,\ldots{\char'134}end} construction or with
the short form \texttt{\$\,\ldots\$}. You can use any of the symbols
and structures, from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a few
examples of in-text equations in context. Notice how this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols and
structures available in \LaTeX\@; this section will just give a couple
of examples of display equations in context. First, consider the
equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\section{Figures}
The ``\verb|figure|'' environment should be used for figures. One or
more images can be placed within a figure. If your figure contains
third-party material, you must clearly identify it as such, as shown
in the example below.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample-franklin}
\caption{1907 Franklin Model D roadster. Photograph by Harris \&
Ewing, Inc. [Public domain], via Wikimedia
Commons. (\url{https://goo.gl/VLCRBB}).}
\Description{A woman and a girl in white dresses sit in an open car.}
\end{figure}
Your figures should contain a caption which describes the figure to
the reader.
Figure captions are placed {\itshape below} the figure.
Every figure should also have a figure description unless it is purely
decorative. These descriptions convey what’s in the image to someone
who cannot see it. They are also used by search engine crawlers for
indexing images, and when images cannot be loaded.
A figure description must be unformatted plain text less than 2000
characters long (including spaces). {\bfseries Figure descriptions
should not repeat the figure caption – their purpose is to capture
important information that is not already provided in the caption or
the main text of the paper.} For figures that convey important and
complex new information, a short text description may not be
adequate. More complex alternative descriptions can be placed in an
appendix and referenced in a short figure description. For example,
provide a data table capturing the information in a bar chart, or a
structured list representing a graph. For additional information
regarding how best to write figure descriptions and why doing this is
so important, please see
\url{https://www.acm.org/publications/taps/describing-figures/}.
\subsection{The ``Teaser Figure''}
A ``teaser figure'' is an image, or set of images in one figure, that
are placed after all author and affiliation information, and before
the body of the article, spanning the page. If you wish to have such a
figure in your article, place the command immediately before the
\verb|\maketitle| command:
\begin{verbatim}
\begin{teaserfigure}
\includegraphics[width=\textwidth]{sampleteaser}
\caption{figure caption}
\Description{figure description}
\end{teaserfigure}
\end{verbatim}
\section{Citations and Bibliographies}
The use of \BibTeX\ for the preparation and formatting of one's
references is strongly recommended. Authors' names should be complete
--- use full first names (``Donald E. Knuth'') not initials
(``D. E. Knuth'') --- and the salient identifying features of a
reference should be included: title, year, volume, number, pages,
article DOI, etc.
The bibliography is included in your source document with these two
commands, placed just before the \verb|\end{document}| command:
\begin{verbatim}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
ACM's consolidated article template, introduced in 2017, provides a
consistent \LaTeX\ style for use across ACM publications, and
incorporates accessibility and metadata-extraction functionality
necessary for future Digital Library endeavors. Numerous ACM and
SIG-specific \LaTeX\ templates have been examined, and their unique
features incorporated into this single new template.
If you are new to publishing with ACM, this document is a valuable
guide to the process of preparing your work for publication. If you
have published with ACM before, this document provides insight and
instruction into more recent changes to the article template.
The ``\verb|acmart|'' document class can be used to prepare articles
for any ACM publication --- conference or journal, and for any stage
of publication, from review to final ``camera-ready'' copy, to the
author's own version, with {\itshape very} few changes to the source.
\section{Template Overview}
As noted in the introduction, the ``\verb|acmart|'' document class can
be used to prepare many different kinds of documentation --- a
double-blind initial submission of a full-length technical paper, a
two-page SIGGRAPH Emerging Technologies abstract, a ``camera-ready''
journal article, a SIGCHI Extended Abstract, and more --- all by
selecting the appropriate {\itshape template style} and {\itshape
template parameters}.
This document will explain the major features of the document
class. For further information, the {\itshape \LaTeX\ User's Guide} is
available from
\url{https://www.acm.org/publications/proceedings-template}.
\subsection{Template Styles}
The primary parameter given to the ``\verb|acmart|'' document class is
the {\itshape template style} which corresponds to the kind of publication
or SIG publishing the work. This parameter is enclosed in square
brackets and is a part of the {\verb|documentclass|} command:
\begin{verbatim}
\documentclass[STYLE]{acmart}
\end{verbatim}
Journals use one of three template styles. All but three ACM journals
use the {\verb|acmsmall|} template style:
\begin{itemize}
\item {\verb|acmsmall|}: The default journal template style.
\item {\verb|acmlarge|}: Used by JOCCH and TAP.
\item {\verb|acmtog|}: Used by TOG.
\end{itemize}
The majority of conference proceedings documentation will use the {\verb|acmconf|} template style.
\begin{itemize}
\item {\verb|acmconf|}: The default proceedings template style.
\item{\verb|sigchi|}: Used for SIGCHI conference articles.
\item{\verb|sigchi-a|}: Used for SIGCHI ``Extended Abstract'' articles.
\item{\verb|sigplan|}: Used for SIGPLAN conference articles.
\end{itemize}
\subsection{Template Parameters}
In addition to specifying the {\itshape template style} to be used in
formatting your work, there are a number of {\itshape template parameters}
which modify some part of the applied template style. A complete list
of these parameters can be found in the {\itshape \LaTeX\ User's Guide.}
Frequently-used parameters, or combinations of parameters, include:
\begin{itemize}
\item {\verb|anonymous,review|}: Suitable for a ``double-blind''
conference submission. Anonymizes the work and includes line
numbers. Use with the \verb|\acmSubmissionID| command to print the
submission's unique ID on each page of the work.
\item{\verb|authorversion|}: Produces a version of the work suitable
for posting by the author.
\item{\verb|screen|}: Produces colored hyperlinks.
\end{itemize}
This document uses the following string as the first command in the
source file:
\begin{verbatim}
\documentclass[sigconf]{acmart}
\end{verbatim}
\section{Modifications}
Modifying the template --- including but not limited to: adjusting
margins, typeface sizes, line spacing, paragraph and list definitions,
and the use of the \verb|\vspace| command to manually adjust the
vertical spacing between elements of your work --- is not allowed.
{\bfseries Your document will be returned to you for revision if
modifications are discovered.}
\section{Typefaces}
The ``\verb|acmart|'' document class requires the use of the
``Libertine'' typeface family. Your \TeX\ installation should include
this set of packages. Please do not substitute other typefaces. The
``\verb|lmodern|'' and ``\verb|ltimes|'' packages should not be used,
as they will override the built-in typeface families.
\section{Title Information}
The title of your work should use capital letters appropriately -
\url{https://capitalizemytitle.com/} has useful rules for
capitalization. Use the {\verb|title|} command to define the title of
your work. If your work has a subtitle, define it with the
{\verb|subtitle|} command. Do not insert line breaks in your title.
If your title is lengthy, you must define a short version to be used
in the page headers, to prevent overlapping text. The \verb|title|
command has a ``short title'' parameter:
\begin{verbatim}
\title[short title]{full title}
\end{verbatim}
\section{Authors and Affiliations}
Each author must be defined separately for accurate metadata
identification. Multiple authors may share one affiliation. Authors'
names should not be abbreviated; use full first names wherever
possible. Include authors' e-mail addresses whenever possible.
Grouping authors' names or e-mail addresses, or providing an ``e-mail
alias,'' as shown below, is not acceptable:
\begin{verbatim}
\author{Brooke Aster, David Mehldau}
\email{dave,judy,<EMAIL>}
\email{<EMAIL>}
\end{verbatim}
The \verb|authornote| and \verb|authornotemark| commands allow a note
to apply to multiple authors --- for example, if the first two authors
of an article contributed equally to the work.
If your author list is lengthy, you must define a shortened version of
the list of authors to be used in the page headers, to prevent
overlapping text. The following command should be placed just after
the last \verb|\author{}| definition:
\begin{verbatim}
\renewcommand{\shortauthors}{McCartney, et al.}
\end{verbatim}
Omitting this command will force the use of a concatenated list of all
of the authors' names, which may result in overlapping text in the
page headers.
The article template's documentation, available at
\url{https://www.acm.org/publications/proceedings-template}, has a
complete explanation of these commands and tips for their effective
use.
Note that authors' addresses are mandatory for journal articles.
\section{Rights Information}
Authors of any work published by ACM will need to complete a rights
form. Depending on the kind of work, and the rights management choice
made by the author, this may be copyright transfer, permission,
license, or an OA (open access) agreement.
Regardless of the rights management choice, the author will receive a
copy of the completed rights form once it has been submitted. This
form contains \LaTeX\ commands that must be copied into the source
document. When the document source is compiled, these commands and
their parameters add formatted text to several areas of the final
document:
\begin{itemize}
\item the ``ACM Reference Format'' text on the first page.
\item the ``rights management'' text on the first page.
\item the conference information in the page header(s).
\end{itemize}
Rights information is unique to the work; if you are preparing several
works for an event, make sure to use the correct set of commands with
each of the works.
The ACM Reference Format text is required for all articles over one
page in length, and is optional for one-page articles (abstracts).
\section{CCS Concepts and User-Defined Keywords}
Two elements of the ``acmart'' document class provide powerful
taxonomic tools for you to help readers find your work in an online
search.
The ACM Computing Classification System ---
\url{https://www.acm.org/publications/class-2012} --- is a set of
classifiers and concepts that describe the computing
discipline. Authors can select entries from this classification
system, via \url{https://dl.acm.org/ccs/ccs.cfm}, and generate the
commands to be included in the \LaTeX\ source.
User-defined keywords are a comma-separated list of words and phrases
of the authors' choosing, providing a more flexible way of describing
the research being presented.
CCS concepts and user-defined keywords are required for for all
articles over two pages in length, and are optional for one- and
two-page articles (or abstracts).
\section{Sectioning Commands}
Your work should use standard \LaTeX\ sectioning commands:
\verb|section|, \verb|subsection|, \verb|subsubsection|, and
\verb|paragraph|. They should be numbered; do not remove the numbering
from the commands.
Simulating a sectioning command by setting the first word or words of
a paragraph in boldface or italicized text is {\bfseries not allowed.}
\section{Tables}
The ``\verb|acmart|'' document class includes the ``\verb|booktabs|''
package --- \url{https://ctan.org/pkg/booktabs} --- for preparing
high-quality tables.
Table captions are placed {\itshape above} the table.
Because tables cannot be split across pages, the best placement for
them is typically the top of the page nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and the
table caption. The contents of the table itself must go in the
\textbf{tabular} environment, to be aligned properly in rows and
columns, with the desired horizontal and vertical rules. Again,
detailed instructions on \textbf{tabular} material are found in the
\textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed output of
this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more
desirable. Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
Always use midrule to separate table header rows from data rows, and
use it only for this purpose. This enables assistive technologies to
recognise table headers and support their users in navigating tables
more easily.
\section{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of the three are
discussed in the next sections.
\subsection{Inline (In-text) Equations}
A formula that appears in the running text is called an inline or
in-text formula. It is produced by the \textbf{math} environment,
which can be invoked with the usual
\texttt{{\char'134}begin\,\ldots{\char'134}end} construction or with
the short form \texttt{\$\,\ldots\$}. You can use any of the symbols
and structures, from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a few
examples of in-text equations in context. Notice how this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols and
structures available in \LaTeX\@; this section will just give a couple
of examples of display equations in context. First, consider the
equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\section{Figures}
The ``\verb|figure|'' environment should be used for figures. One or
more images can be placed within a figure. If your figure contains
third-party material, you must clearly identify it as such, as shown
in the example below.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample-franklin}
\caption{1907 Franklin Model D roadster. Photograph by Harris \&
Ewing, Inc. [Public domain], via Wikimedia
Commons. (\url{https://goo.gl/VLCRBB}).}
\Description{A woman and a girl in white dresses sit in an open car.}
\end{figure}
Your figures should contain a caption which describes the figure to
the reader.
Figure captions are placed {\itshape below} the figure.
Every figure should also have a figure description unless it is purely
decorative. These descriptions convey what’s in the image to someone
who cannot see it. They are also used by search engine crawlers for
indexing images, and when images cannot be loaded.
A figure description must be unformatted plain text less than 2000
characters long (including spaces). {\bfseries Figure descriptions
should not repeat the figure caption – their purpose is to capture
important information that is not already provided in the caption or
the main text of the paper.} For figures that convey important and
complex new information, a short text description may not be
adequate. More complex alternative descriptions can be placed in an
appendix and referenced in a short figure description. For example,
provide a data table capturing the information in a bar chart, or a
structured list representing a graph. For additional information
regarding how best to write figure descriptions and why doing this is
so important, please see
\url{https://www.acm.org/publications/taps/describing-figures/}.
\subsection{The ``Teaser Figure''}
A ``teaser figure'' is an image, or set of images in one figure, that
are placed after all author and affiliation information, and before
the body of the article, spanning the page. If you wish to have such a
figure in your article, place the command immediately before the
\verb|\maketitle| command:
\begin{verbatim}
\begin{teaserfigure}
\includegraphics[width=\textwidth]{sampleteaser}
\caption{figure caption}
\Description{figure description}
\end{teaserfigure}
\end{verbatim}
\section{Citations and Bibliographies}
The use of \BibTeX\ for the preparation and formatting of one's
references is strongly recommended. Authors' names should be complete
--- use full first names (``Donald E. Knuth'') not initials
(``D. E. Knuth'') --- and the salient identifying features of a
reference should be included: title, year, volume, number, pages,
article DOI, etc.
The bibliography is included in your source document with these two
commands, placed just before the \verb|\end{document}| command:
\begin{verbatim}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
ACM's consolidated article template, introduced in 2017, provides a
consistent \LaTeX\ style for use across ACM publications, and
incorporates accessibility and metadata-extraction functionality
necessary for future Digital Library endeavors. Numerous ACM and
SIG-specific \LaTeX\ templates have been examined, and their unique
features incorporated into this single new template.
If you are new to publishing with ACM, this document is a valuable
guide to the process of preparing your work for publication. If you
have published with ACM before, this document provides insight and
instruction into more recent changes to the article template.
The ``\verb|acmart|'' document class can be used to prepare articles
for any ACM publication --- conference or journal, and for any stage
of publication, from review to final ``camera-ready'' copy, to the
author's own version, with {\itshape very} few changes to the source.
\section{Template Overview}
As noted in the introduction, the ``\verb|acmart|'' document class can
be used to prepare many different kinds of documentation --- a
double-blind initial submission of a full-length technical paper, a
two-page SIGGRAPH Emerging Technologies abstract, a ``camera-ready''
journal article, a SIGCHI Extended Abstract, and more --- all by
selecting the appropriate {\itshape template style} and {\itshape
template parameters}.
This document will explain the major features of the document
class. For further information, the {\itshape \LaTeX\ User's Guide} is
available from
\url{https://www.acm.org/publications/proceedings-template}.
\subsection{Template Styles}
The primary parameter given to the ``\verb|acmart|'' document class is
the {\itshape template style} which corresponds to the kind of publication
or SIG publishing the work. This parameter is enclosed in square
brackets and is a part of the {\verb|documentclass|} command:
\begin{verbatim}
\documentclass[STYLE]{acmart}
\end{verbatim}
Journals use one of three template styles. All but three ACM journals
use the {\verb|acmsmall|} template style:
\begin{itemize}
\item {\verb|acmsmall|}: The default journal template style.
\item {\verb|acmlarge|}: Used by JOCCH and TAP.
\item {\verb|acmtog|}: Used by TOG.
\end{itemize}
The majority of conference proceedings documentation will use the {\verb|acmconf|} template style.
\begin{itemize}
\item {\verb|acmconf|}: The default proceedings template style.
\item{\verb|sigchi|}: Used for SIGCHI conference articles.
\item{\verb|sigchi-a|}: Used for SIGCHI ``Extended Abstract'' articles.
\item{\verb|sigplan|}: Used for SIGPLAN conference articles.
\end{itemize}
\subsection{Template Parameters}
In addition to specifying the {\itshape template style} to be used in
formatting your work, there are a number of {\itshape template parameters}
which modify some part of the applied template style. A complete list
of these parameters can be found in the {\itshape \LaTeX\ User's Guide.}
Frequently-used parameters, or combinations of parameters, include:
\begin{itemize}
\item {\verb|anonymous,review|}: Suitable for a ``double-blind''
conference submission. Anonymizes the work and includes line
numbers. Use with the \verb|\acmSubmissionID| command to print the
submission's unique ID on each page of the work.
\item{\verb|authorversion|}: Produces a version of the work suitable
for posting by the author.
\item{\verb|screen|}: Produces colored hyperlinks.
\end{itemize}
This document uses the following string as the first command in the
source file:
\begin{verbatim}
\documentclass[sigconf]{acmart}
\end{verbatim}
\section{Modifications}
Modifying the template --- including but not limited to: adjusting
margins, typeface sizes, line spacing, paragraph and list definitions,
and the use of the \verb|\vspace| command to manually adjust the
vertical spacing between elements of your work --- is not allowed.
{\bfseries Your document will be returned to you for revision if
modifications are discovered.}
\section{Typefaces}
The ``\verb|acmart|'' document class requires the use of the
``Libertine'' typeface family. Your \TeX\ installation should include
this set of packages. Please do not substitute other typefaces. The
``\verb|lmodern|'' and ``\verb|ltimes|'' packages should not be used,
as they will override the built-in typeface families.
\section{Title Information}
The title of your work should use capital letters appropriately -
\url{https://capitalizemytitle.com/} has useful rules for
capitalization. Use the {\verb|title|} command to define the title of
your work. If your work has a subtitle, define it with the
{\verb|subtitle|} command. Do not insert line breaks in your title.
If your title is lengthy, you must define a short version to be used
in the page headers, to prevent overlapping text. The \verb|title|
command has a ``short title'' parameter:
\begin{verbatim}
\title[short title]{full title}
\end{verbatim}
\section{Authors and Affiliations}
Each author must be defined separately for accurate metadata
identification. Multiple authors may share one affiliation. Authors'
names should not be abbreviated; use full first names wherever
possible. Include authors' e-mail addresses whenever possible.
Grouping authors' names or e-mail addresses, or providing an ``e-mail
alias,'' as shown below, is not acceptable:
\begin{verbatim}
\author{Brooke Aster, David Mehldau}
\email{dave,judy,<EMAIL>}
\email{<EMAIL>}
\end{verbatim}
The \verb|authornote| and \verb|authornotemark| commands allow a note
to apply to multiple authors --- for example, if the first two authors
of an article contributed equally to the work.
If your author list is lengthy, you must define a shortened version of
the list of authors to be used in the page headers, to prevent
overlapping text. The following command should be placed just after
the last \verb|\author{}| definition:
\begin{verbatim}
\renewcommand{\shortauthors}{McCartney, et al.}
\end{verbatim}
Omitting this command will force the use of a concatenated list of all
of the authors' names, which may result in overlapping text in the
page headers.
The article template's documentation, available at
\url{https://www.acm.org/publications/proceedings-template}, has a
complete explanation of these commands and tips for their effective
use.
Note that authors' addresses are mandatory for journal articles.
\section{Rights Information}
Authors of any work published by ACM will need to complete a rights
form. Depending on the kind of work, and the rights management choice
made by the author, this may be copyright transfer, permission,
license, or an OA (open access) agreement.
Regardless of the rights management choice, the author will receive a
copy of the completed rights form once it has been submitted. This
form contains \LaTeX\ commands that must be copied into the source
document. When the document source is compiled, these commands and
their parameters add formatted text to several areas of the final
document:
\begin{itemize}
\item the ``ACM Reference Format'' text on the first page.
\item the ``rights management'' text on the first page.
\item the conference information in the page header(s).
\end{itemize}
Rights information is unique to the work; if you are preparing several
works for an event, make sure to use the correct set of commands with
each of the works.
The ACM Reference Format text is required for all articles over one
page in length, and is optional for one-page articles (abstracts).
\section{CCS Concepts and User-Defined Keywords}
Two elements of the ``acmart'' document class provide powerful
taxonomic tools for you to help readers find your work in an online
search.
The ACM Computing Classification System ---
\url{https://www.acm.org/publications/class-2012} --- is a set of
classifiers and concepts that describe the computing
discipline. Authors can select entries from this classification
system, via \url{https://dl.acm.org/ccs/ccs.cfm}, and generate the
commands to be included in the \LaTeX\ source.
User-defined keywords are a comma-separated list of words and phrases
of the authors' choosing, providing a more flexible way of describing
the research being presented.
CCS concepts and user-defined keywords are required for for all
articles over two pages in length, and are optional for one- and
two-page articles (or abstracts).
\section{Sectioning Commands}
Your work should use standard \LaTeX\ sectioning commands:
\verb|section|, \verb|subsection|, \verb|subsubsection|, and
\verb|paragraph|. They should be numbered; do not remove the numbering
from the commands.
Simulating a sectioning command by setting the first word or words of
a paragraph in boldface or italicized text is {\bfseries not allowed.}
\section{Tables}
The ``\verb|acmart|'' document class includes the ``\verb|booktabs|''
package --- \url{https://ctan.org/pkg/booktabs} --- for preparing
high-quality tables.
Table captions are placed {\itshape above} the table.
Because tables cannot be split across pages, the best placement for
them is typically the top of the page nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and the
table caption. The contents of the table itself must go in the
\textbf{tabular} environment, to be aligned properly in rows and
columns, with the desired horizontal and vertical rules. Again,
detailed instructions on \textbf{tabular} material are found in the
\textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed output of
this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more
desirable. Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
Always use midrule to separate table header rows from data rows, and
use it only for this purpose. This enables assistive technologies to
recognise table headers and support their users in navigating tables
more easily.
\section{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of the three are
discussed in the next sections.
\subsection{Inline (In-text) Equations}
A formula that appears in the running text is called an inline or
in-text formula. It is produced by the \textbf{math} environment,
which can be invoked with the usual
\texttt{{\char'134}begin\,\ldots{\char'134}end} construction or with
the short form \texttt{\$\,\ldots\$}. You can use any of the symbols
and structures, from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a few
examples of in-text equations in context. Notice how this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols and
structures available in \LaTeX\@; this section will just give a couple
of examples of display equations in context. First, consider the
equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\section{Figures}
The ``\verb|figure|'' environment should be used for figures. One or
more images can be placed within a figure. If your figure contains
third-party material, you must clearly identify it as such, as shown
in the example below.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample-franklin}
\caption{1907 Franklin Model D roadster. Photograph by Harris \&
Ewing, Inc. [Public domain], via Wikimedia
Commons. (\url{https://goo.gl/VLCRBB}).}
\Description{A woman and a girl in white dresses sit in an open car.}
\end{figure}
Your figures should contain a caption which describes the figure to
the reader.
Figure captions are placed {\itshape below} the figure.
Every figure should also have a figure description unless it is purely
decorative. These descriptions convey what’s in the image to someone
who cannot see it. They are also used by search engine crawlers for
indexing images, and when images cannot be loaded.
A figure description must be unformatted plain text less than 2000
characters long (including spaces). {\bfseries Figure descriptions
should not repeat the figure caption – their purpose is to capture
important information that is not already provided in the caption or
the main text of the paper.} For figures that convey important and
complex new information, a short text description may not be
adequate. More complex alternative descriptions can be placed in an
appendix and referenced in a short figure description. For example,
provide a data table capturing the information in a bar chart, or a
structured list representing a graph. For additional information
regarding how best to write figure descriptions and why doing this is
so important, please see
\url{https://www.acm.org/publications/taps/describing-figures/}.
\subsection{The ``Teaser Figure''}
A ``teaser figure'' is an image, or set of images in one figure, that
are placed after all author and affiliation information, and before
the body of the article, spanning the page. If you wish to have such a
figure in your article, place the command immediately before the
\verb|\maketitle| command:
\begin{verbatim}
\begin{teaserfigure}
\includegraphics[width=\textwidth]{sampleteaser}
\caption{figure caption}
\Description{figure description}
\end{teaserfigure}
\end{verbatim}
\section{Citations and Bibliographies}
The use of \BibTeX\ for the preparation and formatting of one's
references is strongly recommended. Authors' names should be complete
--- use full first names (``Donald E. Knuth'') not initials
(``D. E. Knuth'') --- and the salient identifying features of a
reference should be included: title, year, volume, number, pages,
article DOI, etc.
The bibliography is included in your source document with these two
commands, placed just before the \verb|\end{document}| command:
\begin{verbatim}
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction} \label{sec:intro}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Given an undirected simple graph,
a subset of the vertices of the graph is a {\em dominating set}
if every vertex not in the subset is adjacent to at least one vertex in the subset.
By imposing restrictions on dominating sets,
various variants are proposed.
One of the main variants is a {\em connected dominating set} ({\em CDS}).
A subset of the vertices of the graph is a {\em connected dominating set}
if the subset is a dominating set and the subgraph induced by the subset is connected.
Given an undirected simple graph,
the problem of computing a minimum CDS of the graph is called the {\em connected dominating set} (CDS) problem,
which was shown to be NP-hard~\cite{GJ1979}.
All the vertices composing a CDS are connected and the other vertices are adjacent to vertices in the CDS.
Computing a CDS is equivalent to computing some spanning tree
by regarding vertices in the CDS (not in the CDS, respectively) as non-leaves (leaves, respectively).
Namely,
the problem of computing a spanning tree with the maximum number of leaves is called the {\em maximum laves spanning tree} ({\em MLST}) problem,
which is equivalent to the CDS problem \cite{HL1984}.
The CDS and MLST problems have been extensively studied for various classes of graphs.
A few results on CDSs in grid graphs with restricted size have been shown in \cite{TF2003,LG2010,SN2021},
but a minimum CDS ({\em MCDS}) and the cardinality of an MCDS (called, the {\em connected domination number}) of any grid graph remain open and furthermore, no conjecture about the numbers of grid graphs has been posed.
\fi
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
\noindent
\textbf{Previous Results and Our Results.}~~~
Fujie~\cite{TF2003} proposed an integer linear programming approach to obtain a spanning tree with the maximum number of leaves for any graph.
For $m \times n$ grid graphs such that $m, n \leq 9$,
he showed computationally the maximum number of leaves (i.e., the connected domination number) using this approach.
Let $\gamma_{m,n}$ denote the connected domination number for an $m \times n$ grid graph.
Whether the following results are on the CDS problem or on the MLST problem,
we describe them using the connected domination number.
Fujie~\cite{TF2003} showed that for any $m,n$,
$\gamma_{m,n} \geq mn - \lfloor \frac{2mn}{3} \rfloor$.
Also,
for any $m,n \geq 4$,
he proved
$\gamma_{m,n} \leq \min \{ 2m + n-4 + \lfloor \frac{n-4}{3} \rfloor (m-2), 2n + m-4 + \lfloor \frac{m-4}{3} \rfloor (n-2) \}$.
Lie and Toulouse~\cite{LG2010} showed the following results:
\[
\gamma_{2, n}
=
\begin{cases}
2 & n = 2 \\
2 & n = 3 \\
n & n \geq 4
\end{cases}.
\]
For any $n \geq 3$,
$
\gamma_{3, n} = n
$.
For any $n \geq 4$,
\gamma_{4, n} = 2n - \lfloor \frac{n}{3} \rfloor
$.
For any $n \geq 5$,
\[
\gamma_{5, n} \leq
\begin{cases}
2n & (n \mod 3) = 0 \\
2n+1 & \mbox{otherwise}
\end{cases}.
\]
For any $n \geq 6$,
$
\gamma_{4, n} = 2n + 2
$.
For any $q \geq 2, n \geq 3q$,
$
\gamma_{3q, n} \leq qn + 2q - 2
$.
For any $q \geq 2, n \geq 3q+1$,
$
\gamma_{3q+1, n} \leq qn + n + 2q - 3 - \lfloor \frac{n}{3} \rfloor
$.
For any $q \geq 2, n \geq 3q+2$,
$
\gamma_{3q+2, n} \leq qn + n + 2q - 3 + i
$,
in which
\[
i =
\begin{cases}
0 & (i \mod 3) = 0 \\
1 & \mbox{otherwise}
\end{cases}.
\]
Srinivasan and Narayanaswamy~\cite{SN2021} showed for any $m, n$,
\[
\gamma_{m, n} \geq \lceil \frac{mn + \frac{ 2 \min \{ m, n \} }{3} }{3} \rceil.
\]
In this paper,
we determine the connected domination number of any grid graph.
Specifically, we show the following theorem:
\fi
\begin{THM} \label{THM:cdn}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
For any $m \geq 4, n \geq 4$,
\[
\gamma_{m, n} = \frac{mn - {a'}_{m,n} }{3} + \bar{r}'_{m,n} + {c'}_{m,n},
\]
in which
\[
{a'}_{m,n} = (m \mod 3) \cdot (n \mod 3),
\]
\[
\bar{r}'_{m,n} =
\begin{cases}
3 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 4 \\
2 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 2 \\
1 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 1 \\
0 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 0
\end{cases},
\]
\[
{c'}_{m,n} =
\begin{cases}
\min\{ \frac{m}{3}, \frac{n}{3} \} & (m \mod 3) = 0 \mbox{ and } (n \mod 3) = 0 \\
\frac{m}{3} & (m \mod 3) = 0 \mbox{ and } (n \mod 3) \ne 0 \\
\frac{n}{3} & (m \mod 3) \ne 0 \mbox{ and } (n \mod 3) = 0 \\
\lfloor \frac{m}{3} \rfloor + \lfloor \frac{n}{3} \rfloor - 1 & (m \mod 3) \ne 0 \mbox{ and } (n \mod 3) \ne 0 \\
\end{cases}.
\]
\fi
\end{THM}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
\noindent
\textbf{Our Ideas.}~~~
To determine the connected domination number,
we only have to show the matching upper and lower bounds on it.
We obtain upper bounds by providing instances of connected dominating sets.
The typical and advantageous approach to derive lower bounds for this kind of numbers
is to categorize vertices composing an MCDS into some patterns,
called blocks or tiles,
according to their characteristics
(e.g., the number of their adjacent vertices in the MCDS)
and show combinations of the patterns and lower bounds on the numbers of such patterns which are necessary to construct the MCDS.
Using the values of the lower bounds,
one can obtain lower bounds on the connected domination number.
Gon{\c{c}}alves et~al.~\cite{GPRT2011} used this kind of approach to their lower bounds
when they determined minimum dominating sets in grid graphs.
It is also used in the proofs of lower bounds for CDSs~\cite{SN2021}
and total dominating sets~\cite{SG2002,NS2010}.
However,
it is difficult for one to show that a CDS has the property of being connected
by the numbers of vertices in patterns and combinations of patterns.
Thus,
if one shows a lower bound using the above approach,
it sacrifices the tightness of the lower bound
and makes it difficult to match an upper bound.
Then,
not using such approach but using an algorithmic approach,
we show the matching upper and lower bounds on the connected domination number
in Secs.~\ref{sec:up} and \ref{sec:low}, respectively,
as the proof of Theorem~\ref{THM:cdn}.
For the lower bound proof,
we introduce a {\em regular} MCDS in Sec.~\ref{sec:regularity}.
Then,
a {\em completely} regular MCDS has very characteristic properties and thus,
we show that it is easy to evaluate a lower bound on the number of vertices composing the MCDS in Sec.~\ref{sec:regularityproperty}.
Moreover,
in Sec.~\ref{sec:routinefeasibility},
we show that
any MCDS can be converted algorithmically into a completely regular MCDS according to some routine,
which is defined in Sec.~\ref{sec:routine}.
\fi
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
\noindent
\textbf{Related Results.}~~~
The CDS problem is proved to be equivalent to the MLST problem~\cite{HL1984}.
Garey and Johnson~\cite{GJ1979} showed that the MLST problem on planar graphs with maximum degree four is NP-hard.
Research on the CDS problem (MLST problem) has been conducted for various classes of graphs in terms of complexity.
The CDS and MLST problems are shown to be NP-hard for several classes: bipartite graphs~\cite{PLH1983}, cubic graphs~\cite{PL1988} and unit disk graphs~\cite{CCJ1990}.
On the other hand,
polynomial-time algorithms were designed for other classes:
cographs~\cite{CP1984}, distance-hereditary graphs~\cite{DM1988} and cocomparability graphs~\cite{KS1993}
(see the detail in \cite{CF2020}).
For the problems which are NP-hard for some classes of graphs,
work on approximation algorithms for them has been done~\cite{GK1998,RDJWLK2004,DGPWWZ2008}.
For the CDS problem,
Du et~al.\ \cite{DGPWWZ2008} designed an approximation algorithm with a performance ratio $\epsilon (1 + \ln \Delta)$ for any $\epsilon > 1$,
in which
$\Delta$ is the maximum degree of a given graph.
For any $\epsilon' < 1$,
Guha and Khuller~\cite{GK1998} showed that there does not exist a polynomial-time approximation with a performance ratio $\epsilon' \ln \Delta$ unless $\mbox{NP} \subseteq \mbox{DTIME}(a^{O(\ln \ln a)})$,
in which $a$ is the number of vertices.
Note that approximation algorithms for the CDS problem are not used as those for the MLST problem in general.
For the MLST problem,
Lu and Ravi~\cite{LR1998} designed a 3-approximation algorithm and
Solis-Oba et~al.\ \cite{RS1998,SBL2017} improved it by showing a 2-approximation algorithm.
There are some approximation algorithms for other classes of graphs:
bipartite graphs~\cite{FA2007}, directed graphs~\cite{DT2009,DV2010,SSW2012} and cubic graphs~~\cite{LZ2002,CFMW2008,BZ2011}.
Refer to the comprehensive surveys about the CDS problem (e.g., \cite{DW2013,HHH2020}).
There is much work on important problems in a grid graph.
Consecutive studies~\cite{JK1983,TYC1992,CC1993,GPRT2011} had been conducted
on the dominating set problem in grid graphs for long
to determine the minimum size of a dominating set.
Also,
research on variants of the dominating set problem in a grid graph has been extensively done:
independent dominating sets~\cite{CO2015},
total dominating sets~\cite{SG2002,NS2010}
and
power dominating sets~\cite{DH2006}.
Other work on major problems on a grid graph except for variants of the dominating set problem has been conducted:
the independent set problem~\cite{CW1998},
the hamilton paths problem~\cite{IPS1982}
and
the shortest path problem~\cite{FH1977}.
\fi
\section{Notation and Definitions} \label{sec:notation}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
For any positive integers $m \geq 4$ and $n \geq 4$,
$G_{m,n} = (V, E)$ is an $m \times n$ grid graph,
in which
\[
V = \{ (x, y) \mid x \in [1, m], y \in [1, n] \}
\]
and
\[
E =
\{
\{ (x, y),(x, y+1) \} \mid x \in [1, m], y \in [1, n-1]
\}
\cup
\{
\{ (x, y),(x+1, y) \} \mid x \in [1, m-1], y \in [1, n]
\}.
\]
For any vertices $u, v \in V$ such that $\{ u, v \} \in E$,
we say that $u$ and $v$ are {\em adjacent}.
For any two vertices $v_{1}, v_{a} \in V$,
a {\em path} between $v_{1}$ and $v_{a}$ is a vertex sequence $v_{1} v_{2} \cdots v_{a}$
such that for any $i = 1, \ldots, a-1$, $v_{i}$ and $v_{i+1}$ are adjacent.
For any path $v_{1} v_{2} \cdots v_{a}$
such that for any $i, j \ne i$, $v_{i} \ne v_{j}$ holds,
we say that the path is {\em simple}.
For a vertex set $S \subseteq V$,
if for any vertex $v \in V$,
either $v \in S$ or $v$ is adjacent to a vertex in $S$,
then $S$ is a {\em dominating set} of $G_{m,n}$.
If a vertex set $S \subseteq V$ is a dominating set and
the subgraph induced by $S$ is connected,
namely, for any two vertices $u, v \in S$,
there exists a path between $u$ and $v$ of only vertices in $S$,
then
$S$ is a {\em connected dominating set} ({\em CDS}) of $G_{m,n}$.
{\em MCDS} denotes a CDS with minimum cardinality.
For any dominating set $D$ and any vertex $v \in D$,
we say that $v$ {\em dominates} $v$ itself and vertices adjacent to $v$.
For any set $S$,
$|S|$ denotes the number of elements in $S$.
\fi
\section{Upper Bounds} \label{sec:up}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
\fi
\begin{LMA}\label{LMA:UB}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
For any $m \geq 4, n \geq 4$,
\[
\gamma_{m, n} \leq \frac{mn - {a'}_{m,n} }{3} + \bar{r}'_{m,n}+ {c'}_{m,n},
\]
in which
\[
{a'}_{m,n} = (m \mod 3) \cdot (n \mod 3),
\]
\[
\bar{r}'_{m,n} =
\begin{cases}
3 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 4 \\
2 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 2 \\
1 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 1 \\
0 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 0
\end{cases}
\]
and
\[
{c'}_{m,n} =
\begin{cases}
\min\{ \frac{m}{3}, \frac{n}{3} \} & (m \mod 3) = 0 \mbox{ and } (n \mod 3) = 0 \\
\frac{m}{3} & (m \mod 3) = 0 \mbox{ and } (n \mod 3) \ne 0 \\
\frac{n}{3} & (m \mod 3) \ne 0 \mbox{ and } (n \mod 3) = 0 \\
\lfloor \frac{m}{3} \rfloor + \lfloor \frac{n}{3} \rfloor - 1 & (m \mod 3) \ne 0 \mbox{ and } (n \mod 3) \ne 0 \\
\end{cases}.
\]
\fi
\end{LMA}
\begin{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
We show an upper bound on the connected domination number by presenting a CDS instance for each $m \geq 4$ and $n \geq 4$.
First,
we consider the case of any $n$ and $(m \mod 3) = 0$.
We consider the vertex set
$S_{0,*} = A \cup B \cup C_{0,*}$,
in which
$A = \{ (1, 2), (2, 2), \cdots, (m, 2) \}$,
$B = \{ (2, 3), (2, 4), \cdots, (2, n) \}$, and
$C_{0,*} = \{ (x, 3), \cdots, (x, n) \mid x = 5, \ldots, m-1\}$
(see Fig.~\ref{fig:up}).
Vertices in $A$ ($B$ and $C_{0,*}$, respectively) dominate vertices in $\{ (1, 1), \cdots, (m, 1) \}$ ($\{ (1, 3), \cdots, (1, n) \}$ and $\{ (x-1, 3), \cdots, (x-1, n), (x+1, 3), \cdots, (x+1, n) \mid x = 5, \ldots, m-1 \}$, respectively).
Also,
since vertices in $S_{0,*}$ are connected as we can see from Fig.~\ref{fig:up},
$S_{0,*}$ is a CDS.
Since $|A| = m$,
$|B| = n-2$, and
$|C_{0,*}| = (n-2)(\frac{(m-1) - 5}{3} + 1) = \frac{(n-2)(m-3)}{3}$,
we have
\begin{equation} \label{LMA:UB:eq.1}
|S_{0,*}| = m + n-2 + \frac{(n-2)(m-3)}{3} = \frac{(n+1)m}{3}.
\end{equation}
Next in the case of any $m$ and $(n \mod 3) = 0$,
we consider the vertex set
$S_{*,0} = A \cup B \cup C_{*,0}$,
in which
$C_{*,0} = \{ (3, y), \cdots, (m, y) \mid y = 5, \ldots, n-1 \}$.
Vertices in $C_{*,0}$ dominate vertices in
$\{ (3, y-1), \cdots, (m, y-1), (3, y+1), \cdots, (m, y+1) \mid y = 5, \ldots, n-1 \}$.
Since vertices in $S_{*,0}$ are connected,
$S_{*,0}$ is a CDS.
$|A| = m$,
$|B| = n-2$ and
$|C_{*,0}| = (m-2)(\frac{(n-1) - 5}{3} + 1) = \frac{(m-2)(n-3)}{3}$,
and we have
\begin{equation} \label{LMA:UB:eq.2}
|S_{*,0}| = m + n-2 + \frac{(m-2)(n-3)}{3} = \frac{(m+1)n}{3}.
\end{equation}
By the above two instances,
we can have the following bounds.
\noindent
{\bf\boldmath $(m \mod 3) = 0$ and $(n \mod 3) = 0$:}
By Eqs.~(\ref{LMA:UB:eq.1}) and (\ref{LMA:UB:eq.2}),
\begin{equation} \label{LMA:UB:eq.3}
\gamma_{m,n} \leq \min \left \{ \frac{(n+1)m}{3}, \frac{(m+1)n}{3} \right \}
= \frac{mn}{3} + \min \left \{ \frac{m}{3}, \frac{n}{3} \right \}
= \frac{mn - {a'}_{m,n} }{3} + \bar{r}'_{m,n} + {c'}_{m,n}.
\end{equation}
Note that
\[
{a'}_{m,n} = (m \mod 3) \cdot (n \mod 3) = 0,
\]
\[
\bar{r}'_{m,n} = 0
\]
and
\[
{c'}_{m,n} = \min \left \{ \frac{m}{3}, \frac{n}{3} \right \}.
\]
\noindent
{\bf\boldmath $(m \mod 3) = 0$ and $(n \mod 3) \ne 0$:}
If $(m \mod 3) = 0$ and $(n \mod 3) \ne 0$,
it follows from (\ref{LMA:UB:eq.1}) that
\begin{equation} \label{LMA:UB:eq.4}
\gamma_{m,n} \leq \frac{(n+1)m}{3}
= \frac{mn - {a'}_{m,n} }{3} + \bar{r}'_{m,n} + {c'}_{m,n}.
\end{equation}
Note that
\[
{a'}_{m,n} = (m \mod 3) \cdot (n \mod 3) = 0,
\]
\[
\bar{r}'_{m,n} = 0
\]
and
\[
{c'}_{m,n} = \frac{m}{3}.
\]
\noindent
{\bf\boldmath $(m \mod 3) \ne 0$ and $(n \mod 3) = 0$:}
By Eq.(\ref{LMA:UB:eq.2}),
\begin{equation} \label{LMA:UB:eq.5}
\gamma_{m,n} \leq \frac{(m+1)n}{3}
= \frac{mn - {a'}_{m,n} }{3} + \bar{r}'_{m,n} + {c'}_{m,n}.
\end{equation}
Note that
\[
{a'}_{m,n} = (m \mod 3) \cdot (n \mod 3) = 0,
\]
\[
\bar{r}'_{m,n} = 0
\]
and
\[
{c'}_{m,n} = \frac{n}{3}.
\]
\noindent
{\bf\boldmath $(m \mod 3) = 1$ and $(n \mod 3) = 1$:}
We consider the vertex set
$S_{1,1} = A \cup B \cup C_{1,1} \cup D_{1,1} \cup E_{1,1}$,
in which
$C_{1,1} = \{ (3, y), \cdots, (m, y) \mid y = 5, \ldots, n-2 \}$,
$D_{1,1} = \{ (x, n-1), (x, n) \mid x = 5, \ldots, m-2 \}$
and
$E_{1,1} = \{ (m, n-1) \}$.
Vertices in $C_{1,1}$ ($D_{1,1}$, respectively)
dominate
vertices in $\{ (3, y-1), \cdots, (m, y-1), (3, y+1), \cdots, (m, y+1) \mid y = 5, \ldots, n-2 \}$
($\{ (x-1, n-1), (x-1, n), (x+1, n-1), (x+1, n) \mid x = 5, \ldots, m-2 \}$, respectively)
and
$E_{1,1}$ dominates $(m, n)$.
Since $S_{1,1}$ is connected,
$S_{1,1}$ is a CDS.
$|A| = m$,
$|B| = n-2$,
$|C_{1,1}| = (m-2)(\frac{(n-2) - 5}{3} + 1)$,
$|D_{1,1}| = 2 (\frac{(m-2) - 5}{3} + 1)$ and
$|E_{1,1}| = 1$.
Thus,
we have
\begin{eqnarray} \label{LMA:UB:eq.6}
|S_{1,1}| &=& m + n-2 + (m-2) \left (\frac{(n-2) - 5}{3} + 1 \right ) + 2 \left ( \frac{(m-2) - 5}{3} + 1 \right ) + 1 \nonumber \\
&=& m + n-1 + (m-2)\frac{n-4}{3} + 2 \frac{m-4}{3} \nonumber \\
&=& \frac{3m + 3n-3 + mn-4m-2n+8 +2m-8 }{3}
= \frac{mn + m + n - 3}{3} \nonumber \\
&=& \frac{mn-1}{3} + 1 + \frac{m-1}{3} + \frac{n-1}{3}
= \frac{mn-1}{3} + 1 + \lfloor \frac{m}{3} \rfloor + \lfloor \frac{n}{3} \rfloor - 1 \nonumber \\
&=& \frac{mn - {a'}_{m,n} }{3} + \bar{r}'_{m,n} + {c'}_{m,n}.
\end{eqnarray}
Note that
\[
{a'}_{m,n} = (m \mod 3) \cdot (n \mod 3) = 1,
\]
\[
\bar{r}'_{m,n} = 1,
\]
and
\[
{c'}_{m,n} = \lfloor \frac{m}{3} \rfloor + \lfloor \frac{n}{3} \rfloor - 1.
\]
\noindent
{\bf\boldmath $(m \mod 3) = 1$ and $(n \mod 3) = 2$:}
We consider the vertex set
$S_{1,2} = A \cup B \cup C_{1,2} \cup D_{1,2} \cup E_{1,2}$,
in which
$C_{1,2} = \{ (3, y), \cdots, (m, y) \mid y = 5, \ldots, n-3 \}$,
$D_{1,2} = \{ (x, n-2), (x, n-1), (x, n) \mid x = 5, \ldots, m-2 \}$
and
$E_{1,2} = \{ (m, n-2), (m, n-1) \}$.
Vertices in $C_{1,2}$ ($D_{1,2}$ and $E_{1,2}$, respectively)
dominate
vertices in $\{ (3, y-1), \cdots, (m, y-1), (3, y+1), \cdots, (m, y+1) \mid y = 5, \ldots, n-3 \}$
($\{ (x-1, n-2), (x-1, n-1), (x-1, n), (x+1, n-2), (x+1, n-1), (x+1, n) \mid x = 5, \ldots, m-2 \}$ and $\{ (m, n-1), (m, n) \}$, respectively).
Since $S_{1,2}$ is connected,
$S_{1,2}$ is a CDS.
$|C_{1,2}| = (m-2)(\frac{(n-3) - 5}{3} + 1)$,
$|D_{1,2}| = 3 (\frac{(m-2) - 5}{3} + 1)$ and
$|E_{1,2}| = 2$.
Hence,
we have
\begin{eqnarray} \label{LMA:UB:eq.7}
|S_{1,2}| &=& m + n-2 + (m-2) \left (\frac{(n-3) - 5}{3} + 1 \right ) + 3 \left (\frac{(m-2) - 5}{3} + 1 \right ) + 2 \nonumber\\
&=& m + n + (m-2) \frac{n-5}{3} + 3 \frac{m-4}{3} \nonumber \\
&=& \frac{3m + 3n + mn-5m-2n+10 + 3m-12}{3}
= \frac{mn + m + n - 2}{3} \nonumber\\
&=& \frac{mn-2}{3} + 2 + \frac{m-1}{3} + \frac{n-2}{3} - 1
= \frac{mn-2}{3} + 2 + \lfloor \frac{m}{3} \rfloor + \lfloor \frac{n}{3} \rfloor - 1 \nonumber \\
&=& \frac{mn - {a'}_{m,n} }{3} + \bar{r}'_{m,n} + {c'}_{m,n}.
\end{eqnarray}
Note that
\[
{a'}_{m,n} = (m \mod 3) \cdot (n \mod 3) = 2,
\]
\[
\bar{r}'_{m,n} = 2
\]
and
\[
{c'}_{m,n} = \lfloor \frac{m}{3} \rfloor + \lfloor \frac{n}{3} \rfloor - 1.
\]
\noindent
{\bf\boldmath $(m \mod 3) = 2$ and $(n \mod 3) = 1$:}
We consider the vertex set $S_{2,1} = A \cup B \cup C_{2,1} \cup D_{2,1} \cup E_{2,1}$,
in which
$C_{2,1} = \{ (3, y), \cdots, (m, y) \mid y = 5, \ldots, n-2 \}$,
$D_{2,1} = \{ (x, n-1), (x, n) \mid x = 5, \ldots, m-3 \}$
and
$E_{2,1} = \{ (m, n-1), (m, n) \}$.
Vertices in $C_{2,1}$ ($D_{2,1}$ and $E_{2,1}$, respectively)
dominate
$\{ (3, y-1), \cdots, (m, y-1), (3, y+1), \cdots, (m, y+1) \mid y = 5, \ldots, n-2 \}$
($\{ (x-1, n-1), (x-1, n), (x+1, n-1), (x+1, n) \mid x = 5, \ldots, m-3 \}$ and $\{ (m-1, n), (m, n) \}$, respectively).
Since $S_{2,1}$ is connected,
$S_{2,1}$ is a CDS.
$|A| = m$,
$|B| = n-2$,
$|C_{2,1}| = (m-2)(\frac{(n-2) - 5}{3} + 1)$,
$|D_{2,1}| = 2 (\frac{(m-3) - 5}{3} + 1)$ and
$|E_{2,1}| = 2$.
Thus,
we have
\begin{eqnarray} \label{LMA:UB:eq.8}
|S_{1,2}| &=& m + n-2 + (m-2) \left (\frac{(n-2) - 5}{3} + 1 \right ) + 2 \left (\frac{(m-3) - 5}{3} + 1 \right ) + 2 \nonumber\\
&=& m + n + (m-2) \frac{n-4}{3} + 2 \frac{m-5}{3} \nonumber\\
&=& \frac{3m + 3n + mn-4m-2n+8 + 2m-10}{3}
= \frac{mn + m + n - 2}{3} \nonumber\\
&=& \frac{mn-2}{3} + 2 + \frac{m-2}{3} + \frac{n-1}{3} - 1
= \frac{mn-2}{3} + 2 + \lfloor \frac{m}{3} \rfloor + \lfloor \frac{n}{3} \rfloor - 1 \nonumber\\
&=& \frac{mn - {a'}_{m,n} }{3} + \bar{r}'_{m,n} + {c'}_{m,n}.
\end{eqnarray}
Note that
\[
{a'}_{m,n} = (m \mod 3) \cdot (n \mod 3) = 2,
\]
\[
\bar{r}'_{m,n} = 2
\]
and
\[
{c'}_{m,n} = \lfloor \frac{m}{3} \rfloor + \lfloor \frac{n}{3} \rfloor - 1.
\]
\noindent
{\bf\boldmath $(m \mod 3) = 2$ and $(n \mod 3) = 2$:}
We consider the vertex set
$S_{2,2} = A \cup B \cup C_{2,2} \cup D_{2,2} \cup E_{2,2}$,
in which
$C_{2,2} = \{ (3, y), \cdots, (m, y) \mid y = 5, \ldots, n-3 \}$,
$D_{2,2} = \{ (x, n-2), (x, n-1), (x, n) \mid x = 5, \ldots, m-3 \}$
and
$E_{2,2} = \{ (m, n-2), (m, n-1), (m, n) \}$.
Vertices in $C_{2,2}$
($D_{2,2}$ and $E_{2,2}$, respectively)
dominate
$\{ (3, y-1), \cdots, (m, y-1), (3, y+1), \cdots, (m, y+1) \mid y = 5, \ldots, n-3 \}$
($\{ (x-1, n-2), (x-1, n-1), (x-1, n), (x+1, n-2), (x+1, n-1), (x+1, n) \mid x = 5, \ldots, m-2 \}$
and $\{ (m-1, n-1), (m-1, n), (m, n-1), (m, n) \}$, respectively).
Since $S_{2,2}$ is connected,
$S_{2,2}$ is a CDS.
$|A| = m$,
$|B| = n-2$,
$|C_{2,2}| = (m-2)(\frac{(n-3) - 5}{3} + 1)$,
$|D_{2,2}| = 3 (\frac{(m-3) - 5}{3} + 1)$ and
$|E_{2,2}| = 3$.
Hence,
we have
\begin{eqnarray} \label{LMA:UB:eq.9}
|S_{2,2}| &=& m + n-2 + (m-2) \left (\frac{(n-3) - 5}{3} + 1 \right ) + 3 \left (\frac{(m-3) - 5}{3} + 1 \right ) + 3 \nonumber\\
&=& m + n+1 + (m-2) \frac{n-5}{3} + 3 \frac{m-5}{3} \nonumber\\
&=& \frac{3m + 3n+3 + mn-5m-2n+10 +3m-15}{3}
= \frac{mn + m + n - 2}{3} \nonumber\\
&=& \frac{mn-4}{3} + 3 + \frac{m-2}{3} + \frac{n-2}{3} - 1
= \frac{mn-4}{3} + 3 + \lfloor \frac{m}{3} \rfloor + \lfloor \frac{n}{3} \rfloor - 1 \nonumber\\
&=& \frac{mn - {a'}_{m,n} }{3} + \bar{r}'_{m,n} + {c'}_{m,n}.
\end{eqnarray}
Note that
\[
{a'}_{m,n} = (m \mod 3) \cdot (n \mod 3) = 4,
\]
\[
\bar{r}'_{m,n} = 3
\]
and
\[
{c'}_{m,n} = \lfloor \frac{m}{3} \rfloor + \lfloor \frac{n}{3} \rfloor - 1.
\]
By the above equalities from (\ref{LMA:UB:eq.3}) to (\ref{LMA:UB:eq.9}),
for any $m \geq 4, n \geq 4$,
we have
\[
\gamma_{m,n} \leq |S_{i, j}|,
\]
in which
$(i, j) \in \{ (0, 0), (0, *), (*, 0), (1, 1), (1, 2), (2, 1), (2, 2) \}$.
Therefore,
we have shown the inequalities in the statement of this lemma.
\fi
\end{proof}
\ifnum \count12 > 0
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig_up.eps}
\end{center}
\caption{
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Upper bound instances of CDSs for any $m, n \in [9, 11]$.
Black circles denote vertices in the CDSs.
A vertex labeled with white letter A (B,C,D and E, respectively) belongs to the set $A$ ($B, C_{i,j}, D_{i,j}$ and $E_{i,j} \hspace{1mm} ((i,j) \in \{ (0, 0), (0, *), (*, 0), (1, 1), (1, 2), (2, 1), (2, 2) \}), respectively)$ for each $m,n$.
\fi
}
\label{fig:up}
\end{figure*}
\fi
\section{Lower Bounds} \label{sec:low}
\subsection{Regularity} \label{sec:regularity}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Let us define an MCDS satisfying key properties
to show lower bounds on the connected domination number.
For an MCDS $D$,
$p \in [1, m]$,
any $q \in [2, n]$,
$p' \in [p-1, m]$ and
$q' \in [q-1, n]$ such that
either $q' = q$ and $p' \ne p$ or
$p' = p$ and $q' \ne q$,
we say that $D$ is {\em \boldmath $(p', q')$-$(p, q)$-regular}
(Fig.~\ref{fig:regular})
if $D$ satisfies the following conditions:
When $q' = q$ and $p' \ne p$,
\begin{itemize}
\itemsep=-2.0pt
\setlength{\leftskip}{0pt}
\item[(Q1)]
$(p'+1, q), \cdots, (p, q) \in D$,
\item[(Q2)]
if $q \geq 2$ and $p'+1 \geq p-1$,
then $(p'+1, q-1), \cdots, (p-1, q-1) \notin D$,
\item[(Q3)]
if $q \geq 2$ and $p = m$,
then $(m, q-1) \notin D$,
\item[(Q4)]
if $q \geq 3$,
then $(p'+1, q-2), \cdots, (p, q-2) \notin D$,
\item[(Q5)]
if $q \geq 4$,
then $(p'+1, q-3), \cdots, (m, q-3) \in D$,
\item[(Q6)]
if $q \geq 4$ and $p' \geq 2$,
then $(p', q-3), \cdots, (p', n) \in D$,
\item[(Q7)]
if $q \geq 2$ and $p' \geq 2$,
then $(p'-1, q-1), \cdots, (p'-1, n) \notin D$
and
if $q \geq 3$ and $p' \geq 2$,
then $(p'-1, q-2), \cdots, (p'-1, n) \notin D$,
\item[(Q8)]
if $q \geq 5$,
then $(p'+1, q-4), \cdots, (m, q-4) \notin D$ and
\item[(Q9)]
if $q = n-1$,
then $(p'+1, n), \cdots, (p-1, n) \notin D$.
\end{itemize}
When $p' = p$ and $q' \ne q$,
\begin{itemize}
\itemsep=-2.0pt
\setlength{\leftskip}{0pt}
\item[(P1)]
$(p, q'+1), \cdots, (p, q) \in D$,
\item[(P2)]
if $p \geq 2$ and $q'+1 \geq q-1$,
then $(p-1, q'+1), \cdots, (p-1, q-1) \notin D$,
\item[(P3)]
if $p \geq 2$ and $q = n$,
then $(p-1, n) \notin D$
\item[(P4)]
if $p \geq 3$,
then $(p-2, q'+1), \cdots, (p-2, q) \notin D$
\item[(P5)]
if $p \geq 4$,
then $(p-3, q'+1), \cdots, (p-3, m) \in D$
\item[(P6)]
if $p \geq 4$,
then $(p-3, q'), \cdots, (m, q') \in D$, and
if $p = 2$,
then $(1, q'), \cdots, (m, q') \in D$,
\item[(P7)]
if $p \geq 2$ and $q' \geq 2$,
then $(p-1, q'-1), \cdots, (m, q'-1) \notin D$, and
if $p \geq 3$ and $q' \geq 2$,
then $(p-2, q'-1), \cdots, (m, q'-1) \notin D$,
\item[(P8)]
if $p \geq 5$,
then $(p-4, p'+1), \cdots, (p-4, n) \notin D$ and
\item[(P9)]
if $p = m-1$,
then $(m, q'+1), \cdots, (m, q-1) \notin D$.
\end{itemize}
Additionally,
if $D \subseteq V$ is $(p', q')$-$(p, q)$-regular,
then we say that vertices in $V$ satisfying the following conditions are {\em \boldmath $(p', q')$-$(p, q)$-regular}:
When $q' = q$ and $p' \ne p$,
\begin{itemize}
\itemsep=-2.0pt
\setlength{\leftskip}{10pt}
\item
$(p'+1, q), \cdots, (p, q)$,
\item
if $q \geq 2$,
then $(p'+1, q-1), \cdots, (p-1, q-1)$,
\item
if $q \geq 3$,
then $(p'+1, q-2), \cdots, (p, q-2)$,
\item
if $q \geq 4$,
then for each $y = 1, \ldots, q-3$, $(p'+1, y), \cdots, (m, y)$ and
\item
if $p' \geq 1$,
then for $y = 1, \ldots, n$, $(1, y), \cdots, (p', y)$.
\end{itemize}
When $p' = p$ and $q' \ne q$,
\begin{itemize}
\itemsep=-2.0pt
\setlength{\leftskip}{10pt}
\item
$(p, q'+1), \cdots, (p, q)$,
\item
if $p \geq 2$,
then $(p-1, q'+1), \cdots, (p, q-1)$
\item
if $p \geq 3$,
then $(p-2, q'+1), \cdots, (p, q)$
\item
if $p \geq 4$,
then for each $x = 1, \ldots, p-3$, $(x, q'+1), \cdots, (x, n)$ and
\item
if $q' \geq 1$,
then for each $x = 1, \ldots, m$, $(x, 1), \cdots, (x, q')$.
\end{itemize}
We will drop ``$(p', q')$-$(p, q)$-'' when it is clear from the context.
Also,
${R}(D)$ denotes the set of $(p', q')$-$(p, q)$-regular vertices in $D$.
We say that vertices in $V$ are not $(p', q')$-$(p, q)$-regular are {\em \boldmath $(p', q')$-$(p, q)$-irregular}.
We will drop ``$(p', q')$-$(p, q)$-'' when it is clear from the context as well.
${\overline{R}}(D)$ denotes the set of $(p', q')$-$(p, q)$-irregular vertices in $D$.
That is,
${\overline{R}}(D) = D \backslash {R}(D)$.
\fi
\ifnum \count12 > 0
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig_regular.eps}
\end{center}
\caption{
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Suppose that $D$ is a $(p', q')$-$(p, q)$-regular MCDS.
The left (right) figure denotes the case in which $q' = q$ and $p' \ne p$
($q' \ne q$ and $p' = p$).
Circles and crosses denote regular vertices in $D$ and not in $D$, respectively.
Vertices enclosed by dashed lines are irregular, and the others are regular.
Numbers on circles and crosses represent the numbers of the conditions for the regularity.
For example,
``6'' on the vertex $(q-3, p')$ in the left figure means that the vertex satisfies the condition (Q6).
\fi
}
\label{fig:regular}
\end{figure*}
\fi
\subsection{Overview of the Lower Bound Proof} \label{sec:overview}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
We evaluate the connected domination number for any $m(\geq 4), n(\geq 4)$ using the regularity.
We show the first lemma.
\fi
\begin{LMA}\label{LMA:L1}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
For any $m(\geq 4)$ and $n(\geq 4)$,
suppose that there exists an MCDS without the vertex $(1, 2)$
of the $m \times n$ grid graph.
Then,
there exists an MCDS containing $(1, 2)$ of the $n \times m$ grid graph.
\fi
\end{LMA}
\begin{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Any MCDS contains a vertex to dominate the vertex $(1, 1)$ and is connected.
Thus,
any MCDS contains either $(1, 2)$ or $(2, 1)$.
For some $m$ and $n$,
let $D$ be an MCDS without $(1, 2)$ of the $m \times n$ grid graph.
That is,
$D$ contains $(2, 1)$.
Now,
let $D'$ be the mirror image of $D$.
Specifically,
$D'$ contains a vertex $(y, x)$
if and only if $D$ contains the vertex $(x, y)$.
$D'$ is the MCDS of the $n \times m$ grid graph
and contains $(1, 2)$.
\fi
\end{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
By symmetry,
the connected domination number of an $m \times n$ grid graph is equal to that of an $n \times m$ grid graph, that is, $\gamma_{m,n} = \gamma_{n,m}$.
Hence,
for all MCDSs containing $(1, 2)$, deriving lower bounds on their elements, we can obtain the lower bounds for all MCDSs.
Then,
in what follows,
we consider only MCDSs containing $(1, 2)$.
Let us consider an MCDS containing the vertex $(1, 2)$.
The MCDS is $(0, 2)$-$(1, 2)$-regular by definition.
Starting from the MCDS,
we increase regular vertices in the MCDS sequentially according to some routine,
which is defined later.
Specifically,
the routine constructs a new vertex set $D'$ by removing some irregular vertices from an MCDS $D$ and adding vertices to $D$
such that the number of added vertices is equal to the removed vertices but the number of regular vertices in $D'$ is larger than that in $D$, and $D'$ is a CDS.
That is,
$D'$ is an MCDS with more regular vertices than $D$.
The routine repeats this construction according to some rules
until it cannot increase regular vertices,
and then it is finished.
Let us denote an MCDS at the time when the routine is finished as $D_{\tau}$,
which is referred to as the ``completely regular MCDS'' in Sec.~\ref{sec:intro}.
We give two definitions.
For a $(p', q')$-$(p, q)$-regular MCDS $D$, $x \in [2, m-1]$ and $y \in [2, n-1]$,
a {\em connector in $D$} is called a vertex $(x, y) \in {R}(D)$ satisfying one of the following two conditions:
\begin{itemize}
\itemsep=-2.0pt
\setlength{\leftskip}{0pt}
\item[(c-i)]
$(x+1, y), (x-1, y), (x-1, y-1), (x-1, y+1) \in {R}(D)$
and
$(x, y-1) \notin {R}(D)$ or
\item[(c-ii)]
$(x, y+1), (x, y-1), (x-1, y-1), (x+1, y-1) \in {R}(D)$
and
$(x-1, y) \notin {R}(D)$.
\end{itemize}
Also,
a {\em pre-connector in $D$} is called a vertex $(x, y) \in {R}(D)$ satisfying one of the following two conditions:
\begin{itemize}
\itemsep=-2.0pt
\setlength{\leftskip}{0pt}
\item[(p-i)]
$(x-1, y), (x-1, y-1), (x-1, y+1) \in {R}(D)$
and
$(x+1, y)$
or
\item[(p-ii)]
$(x, y+1), (x, y-1), (x-1, y-1), (x+1, y-1) \in {R}(D)$
and
$(x, y+1)$.
\end{itemize}
We call a vertex in ${R}(D)$ which is neither a connector nor a pre-connector a {\em dominator in $D$}.
\fi
\ifnum \count12 > 0
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig_cr.eps}
\end{center}
\caption{
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Black circles denote vertices in ${R}(D)$,
crosses denote vertices not in ${R}(D)$ and
gray triangles denote irregular vertices.
Vertices satisfying (c-i) or (c-ii) are connectors,
denoted by gray circles, and
vertices satisfying (p-i) or (p-ii) are pre-connectors,
denoted by double circles.
\fi
}
\label{fig:cr}
\end{figure*}
\fi
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
In Lemma~\ref{LMA:PCR},
we show that $D_{\tau}$ does not contain any pre-connector.
In Lemma~\ref{LMA:CR0},
we show that for any connector, vertices dominated by the connector are dominated by at least one dominator.
These two lemmas indicate that
the number of vertices by regular vertices in $D_{\tau}$ is equal to that of vertices dominated by dominators in $D_{\tau}$.
Furthermore,
in Lemma~\ref{LMA:DR},
we show that
dominators dominate $3 {d}_{m,n}$ vertices,
in which ${d}_{m,n}$ denotes the number of dominators in $D_{\tau}$.
Since the number of vertices in an $m \times n$ grid graph is $mn$,
we have
\begin{equation} \label{EQ:sec:overview.eq1}
mn = 3 {d}_{m,n} + {a}_{m,n},
\end{equation}
in which
${a}_{m,n}$ denotes the number of vertices which are not dominated by regular vertices in $D_{\tau}$
at the time when the routine is finished.
Let $\bar{r}_{m,n}$ denote the number of irregular vertices in $D_{\tau}$,
that is,
$\bar{r}_{m,n} = |{\overline{R}}(D_{\tau})|$.
Let ${c}_{m,n}$ denote the number of connectors in $D_{\tau}$.
Since regular vertices in $D_{\tau}$ are connectors or dominators by Lemma~\ref{LMA:PCR},
we have
\begin{equation} \label{EQ:sec:overview.eq2}
\gamma_{m,n} = {d}_{m,n} + {c}_{m,n} + \bar{r}_{m,n}.
\end{equation}
In Lemma~\ref{LMA:IR}, we show that
\begin{equation} \label{EQ:sec:overview.eq3}
a_{m,n} = (m \mod 3) \cdot (n \mod 3)
\end{equation}
and
\begin{equation} \label{EQ:sec:overview.eq4}
\bar{r}_{m,n} =
\begin{cases}
3 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 4 \\
2 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 2 \\
1 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 1 \\
0 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 0.
\end{cases}
\end{equation}
Also, in Lemma~\ref{LMA:CR1},
we show that
\begin{equation} \label{EQ:sec:overview.eq5}
{c}_{m,n} \geq
\begin{cases}
\min\{ \frac{m}{3}, \frac{n}{3} \} & (m \mod 3) = 0 \mbox{ and } (n \mod 3) = 0 \\
\frac{m}{3} & (m \mod 3) = 0 \mbox{ and } (n \mod 3) \ne 0 \\
\frac{n}{3} & (m \mod 3) \ne 0 \mbox{ and } (n \mod 3) = 0 \\
\lfloor \frac{m}{3} \rfloor + \lfloor \frac{n}{3} \rfloor - 1 & (m \mod 3) \ne 0 \mbox{ and } (n \mod 3) \ne 0. \\
\end{cases}
\end{equation}
By Eqs~(\ref{EQ:sec:overview.eq1}),(\ref{EQ:sec:overview.eq2}),(\ref{EQ:sec:overview.eq3}),(\ref{EQ:sec:overview.eq4}) and (\ref{EQ:sec:overview.eq5}),
we have the lower bound lemma:
\fi
\begin{LMA} \label{LMA:low}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
For any $m \geq 4$ and $n \geq 4$,
\[
\gamma_{m, n} \geq \frac{mn - {a'}_{m,n} }{3} + \bar{r}'_{m,n} + {c'}_{m,n},
\]
in which
\[
{a'}_{m,n} = (m \mod 3) \cdot (n \mod 3),
\]
\[
\bar{r}'_{m,n} =
\begin{cases}
3 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 4 \\
2 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 2 \\
1 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 1 \\
0 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 0
\end{cases}
\]
and
\[
{c'}_{m,n} =
\begin{cases}
\min\{ \frac{m}{3}, \frac{n}{3} \} & (m \mod 3) = 0 \mbox{ and } (n \mod 3) = 0 \\
\frac{m}{3} & (m \mod 3) = 0 \mbox{ and } (n \mod 3) \ne 0 \\
\frac{n}{3} & (m \mod 3) \ne 0 \mbox{ and } (n \mod 3) = 0 \\
\lfloor \frac{m}{3} \rfloor + \lfloor \frac{n}{3} \rfloor - 1 & (m \mod 3) \ne 0 \mbox{ and } (n \mod 3) \ne 0. \\
\end{cases}
\]
\fi
\end{LMA}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
\fi
\subsection{Regularization Routine} \label{sec:routine}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
We give some definitions to define the routine.
Suppose that $D$ is a $(p', q')$-$(p, q)$-regular CDS.
Also,
suppose that $D'$ is a CDS such that
$D' = (D \backslash U) \cup U' = D \backslash U \cup U'$,
in which
$U, U' \subseteq {\overline{R}}(D)$,
and
$|U| = |U'|$,
that is,
all the vertices in $U$ and $U'$ are irregular
and
the numbers of vertices in the two sets are equal.
Then,
we say that {\em $D$ is regularized into $D'$}.
Furthermore,
if $D'$ is a $(\hat{p}', \hat{q}')$-$(\hat{p}, \hat{q})$-regular CDS,
then we say that {\em $D$ is $(\hat{p}', \hat{q}')$-$(\hat{p}, \hat{q})$-regularized into $D'$}.
Note that if $D$ is an MCDS, then $D'$ is also an MCDS.
In the routine,
note that the conditions of Cases~3.2 and 3.3 are not exclusive.
Specifically,
if $p' \ne 0$, $q' \leq n-4$ and $p' \leq m-4$,
then the routine executes only one case of Cases~3.2 and 3.3.
In the beginning,
the routine initializes the variable {\tt D} as an MCDS containing $(1, 2)$.
The routine repeatedly executes Cases~1, 2, 3.1, 3.2 and 3.3 depending on the MCDS of {\tt D}
and regularizes it
until the routine executes Case~3.4 and is finished.
\fi
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
\noindent\vspace{-1mm}\rule{\textwidth}{0.5mm}
\vspace{-3mm}
{\sc Regularization Routine}\\
\rule{\textwidth}{0.1mm}
{\bf Initialization:}
{\tt D} $:=$ a $(0, 2)$-$(1, 2)$-regular MCDS (see Fig. ~\ref{fig:routine}). \\
Suppose that {\tt D} is $(p', q')$-$(p, q)$-regular.
Execute one of the following cases:
\\
{\bf\boldmath Case~1 ($q' = q$ and $p \leq m-1$):}\\
\hspace*{1mm}
{\tt D} $:=$ an MCDS into which {\tt D} is $(p', q')$-$(p+1, q)$-regularized
(guaranteed in Lemma~\ref{LMA:L4}).
\\
{\bf\boldmath Case~2 ($p' = p$ and $q \leq n-1$):}\\
\hspace*{1mm}
{\tt D} $:=$ an MCDS into which {\tt D} is $(p', q')$-$(p, q+1)$-regularized
(guaranteed in Lemma~\ref{LMA:L4}).
\\
{\bf\boldmath Case~3 (either $q'=q$ and $p=m$ or $p'=p$ and $q=n$):}
Execute one of the four cases: \\
\hspace*{2mm}
{\bf\boldmath Case~3.1 ($p' = 0$):}\\
\hspace*{3mm}
{\tt D} $:=$ an MCDS into which {\tt D} is $(2, 2)$-$(2, 3)$-regularized
(guaranteed in Lemma~\ref{LMA:L6}).
\\
\hspace*{2mm}
{\bf\boldmath Case~3.2 ($p' \ne 0$, $p' \leq m-2$ and $q' \leq n-4$):}\\
\hspace*{3mm}
{\tt D} $:=$ an MCDS into which {\tt D} is $(p', q'+3)$-$(p'+1, q'+3)$-regularized
(guaranteed in Lemma~\ref{LMA:C9}).
\\
\hspace*{2mm}
{\bf\boldmath Case~3.3 ($p' \ne 0$, $p' \leq m-4$ and $q' \leq n-2$):}\\
\hspace*{3mm}
{\tt D} $:=$ an MCDS into which {\tt D} is $(p'+3, q')$-$(p'+3, q'+1)$-regularized
(guaranteed in Lemma~\ref{LMA:C9}).
\\
\hspace*{2mm}
{\bf\boldmath Case~3.4
(either $p' = p = m-1$,
$q' = q = n-1$ or
both $m-3 \leq p' \leq m-2$ and $n-3 \leq q' \leq n-2$):}\\
\hspace*{3mm}
Finish.
\\
\noindent\vspace{-1mm}\rule{\textwidth}{0.5mm}
\fi
\ifnum \count12 > 0
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig_routine.eps}
\end{center}
\caption{
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Regularization routine.
Black circles and crosses denote regular vertices in {\tt D} and not in {\tt D}, respectively.
White circles and crosses denote regular vertices which are newly added into {\tt D} and are not added, respectively, when each case is executed.
Vertices enclosed by dashed lines are irregular and the other vertices are regular.
For each case,
the left figure on the left (right) of the arrow denotes the situation of {\tt D} immediately before (after) the routine execution.
\fi
}
\label{fig:routine}
\end{figure*}
\fi
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
If each case is feasible,
then the routine necessarily is finished
because the number of regular vertices is monotone increasing.
\fi
\subsection{Regularity Property} \label{sec:regularityproperty}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
In this section,
we show some properties of MCDSs processed by the routine
if the routine is feasible.
First,
we give some definitions.
Suppose that the routine is feasible.
Then,
let $\tau$ be the number of times that the routine is executed by the time that the routine is finished.
Let $D_{0}$ be an MCDS which is given to the routine at initialization.
$D_{i}$ denotes the value of the variable {\tt D} in the routine
immediately after the routine is executed for the $i (= 1, 2, \ldots, \tau)$-th time.
Let $p'_{i}, q'_{i}), p_{i}$ and $q_{i})$ be integers
such that $D_{i}$ is $(p'_{i}, q'_{i})$-$(p_{i}, q_{i})$-regular.
Also,
we define $p'_{0} = 0, p_{0} = 1$
and
$q'_{0} = q_{0} = 2$.
$D_{0}$ is $(0, 2)$-$(1, 2)$-regular by definition,
that is, $(p'_{0}, q'_{0})$-$(p_{0}, q_{0})$-regular.
Let $k$ be the number of times that the routine executes Case~3.1, 3.2, 3.3 or 3.4.
Let $t_{1}, \ldots, t_{k}$ be positive integers such that
the routine executes Case~3.1, 3.2, 3.3 or 3.4 for the $t_{1}, \ldots, t_{k}$-th time and
$t_{1} < t_{2} < \cdots < t_{k}$.
When the routine executes Case~3.4,
it is finished.
Thus,
the case which the routine executes for the $t_{k}$-th time is Case~3.4.
That is,
$t_{k} = \tau$.
\fi
\begin{LMA}\label{LMA:PCR}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Suppose that the routine is feasible.
Then,
$D_{\tau}$ does not contain any pre-connector.
\fi
\end{LMA}
\begin{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
By the supposition of this lemma,
the routine is feasible.
In the following,
we will show the following two properties by induction on the number $j \in [1, \tau]$ of times that the routine is executed:
\begin{itemize}
\itemsep=-2.0pt
\setlength{\leftskip}{0pt}
\item[(a)]
If the routine executes Case~3.1, 3.2 or 3.3 for the $j$-th execution,
then $D_{j}$ contains one pre-connector.
\item[(b)]
If the routine executes Case~1, 2 or 3.4 for the $j$-th execution,
then $D_{j}$ does not contain any pre-connector.
\end{itemize}
If we show that both (a) and (b) hold,
it implies that the statement of this lemma holds
because the routine is finished at the execution of Case~3.4
and (b) shows that the MCDS $D_{\tau}$ does not contain any pre-connector.
When $j = 0$,
that is,
at the initialization of the routine,
$D_{0} = \{ (1, 2) \}$ holds.
By the definition of a pre-connector,
$D_{0}$ does not contain a pre-connector
(see Fig.~\ref{fig:ldr}(1)).
We assume that (a) and (b) hold for the $j = t$-th execution of the routine
and prove that they also hold for the $t+1$-st execution.
We discuss each case depending on which Cases the routine executes for the $t$-th and $t+1$-st times.
Note that only irregular vertices are handled when $D_{t}$ is regularized into $D_{t+1}$,
and regular vertices are not removed from $D_{t}$ when the routine constructs $D_{t+1}$.
\noindent
{\bf\boldmath $t = 0$ or Case~1:}
If either $t = 0$ or
the routine executes Case~1 for the $t$-th execution,
then $D_{t}$ does not contain any pre-connector
immediately before the $t+1$-st execution by (b) in the induction hypothesis.
$q'_{t} = q_{t}$ and thus,
if $p_{t} \leq m-1$, then the routine executes Case~1 for the $t+1$-st execution and
if $p_{t} = m$, the routine executes Case~3.1, 3.2, 3.3 or 3.4.
That is,
the routine can execute Case~1, 3.1, 3.2, 3.3 or 3.4 for the $t+1$-st execution.
If the routine executes Case~1,
the regular vertices newly added into $D_{t+1}$ satisfy the definitions of neither connectors nor pre-connectors,
that is, the vertices are dominators
((2),(3),(10),(11) and (12) Fig.~\ref{fig:ldr}).
Hence,
in these cases, the number of pre-connectors does not change and
$D_{t+1}$ does not contain any pre-connector,
which means that (b) is true.
If the routine executes Case~3.1, 3.2 or 3.3,
the regular vertex newly added into $D_{t+1}$ is only one pre-connector ((4),(9) and (13) in Fig.~\ref{fig:ldr})
and (a) is true.
If the routine executes Case~3.4,
no vertex is added into $D_{t+1}$ and (b) is true.
\noindent
{\bf Case~2:}
We can show this case similarly to Case~1.
If the routine executes Case~2 for the $t$-th execution,
then $D_{t}$ does not contain any pre-connector immediately before the $t+1$-st execution by (b) in the induction hypothesis.
That is,
the routine can execute Case~2, 3.1, 3.2, 3.3 or 3.4 for the $t+1$-st execution.
If the routine executes Case~3.1, 3.2 or 3.3,
the regular vertex newly added into $D_{t+1}$ is only one pre-connector ((9),(13) in Fig.~\ref{fig:ldr}) and (a) is true.
If the routine executes Case~2,
the regular vertex newly added into $D_{t+1}$ is a dominator
((5),(6),(7),(14),(15),(16) in Fig.~\ref{fig:ldr}).
If the routine executes Case~3.4,
no vertex is added into $D_{t+1}$.
Hence,
the number of pre-connectors does not change in these cases and
$D_{t+1}$ does not contain any pre-connector.
Thus,
(b) is true.
\noindent
{\bf Case~3.1, 3.2, 3.3:}
If the routine executes Case~3.1, 3.2 or 3.3 for the $t$-th execution,
$D_{t}$ contains one pre-connector
immediately before the $t+1$-st execution by (a) in the induction hypothesis.
If the routine executes Case~1 or 2 for the $t+1$-st execution,
the pre-connector in $D_{t}$ becomes a connector in $D_{t+1}$
((5),(10),(14) in Fig.~\ref{fig:ldr}).
Thus,
$D_{t+1}$ contains any pre-connector.
In the following,
we show that the routine does not execute Case~3 (3.1, 3.2, 3.3) for the $t+1$-st execution.
If the routine executes Case~3.1 for the $t$-th execution,
then $(2, 3)$ in $D_{t}$ is a pre-connector.
Since we discuss the case of $n \geq 4$,
the routine executes Case~2 but not Case~3 for the $t+1$-st execution.
If the routine executes Case~3.2 for the $t$-th,
then $p'_{t} \leq m-2$ by the condition of Case~3.2.
Then,
the routine executes Case~1 but not Case~3 for the $t+1$-st
because $p_{t} = p'_{t+1} \leq m-1$.
Similarly,
if the routine executes Case~3.3 for the $t$-th,
then $q'_{t} \leq n-2$ by the condition of Case~3.3.
Then,
the routine executes Case~2 but not Case~3 for the $t+1$-st
because $q_{t} = q'_{t+1} \leq n-1$.
\noindent
{\bf Case~3.4:}
If the routine executes Case~3.4 for the $t$-th execution,
then (a) and (b) hold by the induction hypothesis.
We have shown the (a) and (b) are true.
\fi
\end{proof}
\ifnum \count12 > 0
\begin{landscape}
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig_lpq.eps}
\end{center}
\caption{
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Situations of MCDS $D_{t+1}$.
Circles and crosses denote regular vertices in $D_{t+1}$ and not in $D_{t+1}$, respectively.
Vertices not enclosed by dashed lines are regular.
White circles, double circles and white crosses become newly regular for the $t+1$-st execution.
Double circles and gray circles denote pre-connectors and connectors, respectively.
Each number above an arrow implies which Case the routine executes.
For example,
``3.1'' above the arrow between Figures~(3) and (4) means
that the routine executes Case~3.1 in the situation of (3) and
after that,
$D_{t+1}$ becomes the situation of (4).
The situations of $D_{t+1}$ when Case~3.4 is executed for the $t+1$-st execution are illustrated in Fig.~\ref{fig:lir}.
\fi
}
\label{fig:ldr}
\end{figure*}
\end{landscape}
\fi
\begin{LMA}\label{LMA:DR}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Suppose that the routine is feasible.
Then,
the number of vertices dominated by dominators in $D_{\tau}$ is $3d_{m,n}$.
\fi
\end{LMA}
\begin{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
For any $i = 0, \ldots, \tau$,
$d_{i}$ denotes the number of dominators in $D_{i}$.
By the supposition of this lemma,
the routine is feasible.
In what follows,
we will prove the following two properties by induction on the number $j \in [1, \tau]$ of the times that the routine is executed:
\begin{itemize}
\itemsep=-2.0pt
\setlength{\leftskip}{0pt}
\item[(a)]
If either $p'_{j} = p_{j}$ and $q_{j} \leq n-1$ or
$q'_{j} = q_{j}$ or $p_{j} \leq m-1$,
then the number of vertices dominated by either dominators or pre-connectors in $D_{j}$ is $3d_{j}+1$.
\item[(b)]
If either $p'_{j} = p_{j}$ and $q_{j} = n$ or
$q'_{j} = q_{j}$ and $p_{j} = m$,
then the number of vertices dominated by either dominators or pre-connectors in $D_{j}$ is $3d_{j}$.
\end{itemize}
When the routine executes Case~3.4 and is finished,
either $q'_{\tau} = q_{\tau}$ and $p_{\tau} = m$
or
$p'_{\tau} = p_{\tau}$ and $q_{\tau} = n$
by the condition of Case~3.
$D_{\tau}$ does not contain any pre-connector
by Lemma~\ref{LMA:PCR}.
Hence,
if these properties hold,
then (b) holds when the routine is finished, and
the number of vertices dominated by dominators in $D_{\tau}$ is $3d_{\tau}$.
By the definition of $d_{m,n}$,
the number of dominators at the time when the routine is finished is $d_{m,n}=d_{\tau}$
and we can show the statement of the lemma.
When $j = 0$, that is,
at the time of the initialization,
$D_{0}$ is $(0, 2)$-$(1, 2)$-regular
((1) in Fig.~\ref{fig:ldr}).
Since $D_{0} = \{ (1,2) \}$, $d_{0} = 1$.
The vertex $(1,2)$ dominates itself and three adjacent vertices $(1, 1), (1, 3)$ and $(2, 2)$, and the number of vertices dominated by a dominator in $D_{0}$ is four.
Hence,
when $j=0$, the property is true.
We assume that
(a) and (b) hold when the routine is executed for the $j=t$-th execution, and
we show that they also hold for the $t+1$-st execution.
Note that only irregular vertices are handled at the regularization,
and regular vertices are not removed from $D_{t}$ when the routine constructs $D_{t+1}$.
\noindent
{\bf Case~1:}
If the routine executes Case~1
for the $t+1$-st execution,
then $D_{t+1}$ is $(p'_{t}, q'_{t})$-$(p_{t}+1, q_{t})$-regular.
By definition,
the set of regular vertices in $D_{t+1}$ is the set of all the vertices in $D_{t}$ plus the vertex $(p_{t}+1, q_{t})$.
Since $(p_{t}, q_{t}-1) \notin R(D_{t+1})$ by the condition (Q2) of the regularity,
$(p_{t}+1, q_{t})$ satisfies the definition of neither connectors nor pre-connectors.
Thus,
$(p_{t}+1, q_{t})$ is a dominator and
we have
\begin{equation} \label{LMA:DR:eq.1}
d_{t+1} = d_{t} + 1.
\end{equation}
First,
we consider the case in which
$t = 0$ or
the routine executes Case~3.2 for the $t$-th execution,
that is,
$p_{t} = p'_{t}+1$
(i.e., $p_{t+1} = p'_{t}+2$)
and
$p_{t+1} = p_{t}+1 \leq m-1$
((10) in Fig.~\ref{fig:ldr}).
In this case,
it is sufficient that we prove (a) is true
because $p_{t}+1 \leq m-1$.
$(p_{t}+1, q_{t})$ dominates the five vertices $(p_{t}, q_{t}),(p_{t}+1, q_{t}),(p_{t}+1, q_{t}-1),(p_{t}+1, q_{t}+1)$ and $(p_{t}+2, q_{t})$.
By the definition of $(p'_{t}, q'_{t})$-$(p_{t}, q_{t})$-regularity,
vertices in $R(D_{t})$ do not dominate none of $(p_{t}+1, q_{t}-1),(p_{t}+1, q_{t}+1)$ and $(p_{t}+2, q_{t})$.
Thus,
the number of vertices which vertices in $R(D_{t+1})$ dominate increases by three vertices.
Since (a) is true for the $t$-th execution by the induction hypothesis,
dominators and a pre-connector in $D_{t}$ dominate $3 d_{t} +1$ vertices.
Thus,
we have using Eq.~(\ref{LMA:DR:eq.1}),
\[
3 d_{t} + 1 + 3 = 3(d_{t} + 1) + 1 = 3 d_{t+1} + 1,
\]
which implies that
(a) is also true for $j = t+1$
if $t = 0$ or
the routine executes Case~3.2 for the $t$-th execution.
Next,
we consider the case in which
the routine executes Case~1 for the $t$-th and $t+2$-nd executions,
that is,
$p_{t} > p'_{t}+1$
(i.e., $p_{t+1} > p'_{t}+2$)
and
$p_{t+1} = p_{t}+1 \leq m-1$
(Fig.~\ref{fig:ldr}(2),(3),(11)).
In this case,
it suffices to prove that (a) is true
because $p_{t}+1 \leq m-1$.
$(p_{t}+1, q_{t})$ dominates the five vertices $(p_{t}, q_{t}),(p_{t}+1, q_{t}),(p_{t}+1, q_{t}-1),(p_{t}+1, q_{t}+1)$ and $(p_{t}+2, q_{t})$, and
vertices in $R(D_{t})$ dominate none of $(p_{t}+1, q_{t}-1),(p_{t}+1, q_{t}+1)$ and $(p_{t}+2, q_{t})$ by the $(p'_{t}, q'_{t})$-$(p_{t}, q_{t})$-regularity.
Thus,
the number of vertices dominated by vertices in $R(D_{t+1})$ increase by three vertices
and by Eq.~(\ref{LMA:DR:eq.1}),
\[
3 d_{t} + 1 + 3 = 3(d_{t} + 1) + 1 = 3 d_{t+1} + 1.
\]
(a) is true for $j = t+1$
if the routine executes Case~1 for the $t$-th and $t+2$-nd executions.
Finally,
we consider the case in which
the routine executes Case~1 for the $t$-th execution and
does not execute Case~1 for the $t+2$-nd execution,
that is,
$p_{t+1} = p_{t}+1 = m$
(Fig.~\ref{fig:ldr}(3),(12)).
In this case,
it suffices to prove that (b) is true by the condition.
$(p_{t}+1, q_{t})$ dominates the four vertices $(p_{t}, q_{t}),(m, q_{t}),(m, q_{t}-1)$ and $(m, q_{t}+1)$ and
vertices in $R(D_{t})$ dominate neither $(m, q_{t}-1)$ nor $(m, q_{t}+1)$
by the $(p'_{t}, q'_{t})$-$(p_{t}, q_{t})$-regularity.
Thus,
the number of vertices dominated by vertices in $R(D_{t+1})$ increases by two vertices.
Using~Eq.~(\ref{LMA:DR:eq.1}),
we have
\[
3 d_{t} + 1 + 2 = 3(d_{t} + 1) = 3 d_{t+1},
\]
which implies
(b) is also true in this case when $j = t+1$.
\noindent
{\bf Case~2: }
We can prove this case similarly to the proof of Case~1.
If the routine executes Case~2
for the $t+1$-st execution,
then $D_{t+1}$ is $(p'_{t}, q'_{t})$-$(p_{t}, q_{t}+1)$-regular.
The set of regular vertices in $D_{t+1}$ is the set of all the vertices in $D_{t}$ plus the vertex $(p_{t}, q_{t}+1)$.
Since $(p_{t}-1, q_{t}) \notin R(D_{t+1})$ by the condition (P2) of the regularity,
$(p_{t}, q_{t}+1)$ satisfies the definition of neither connectors nor pre-connectors.
Thus,
$(p_{t}, q_{t}+1)$ is a dominator and
we have
\begin{equation} \label{LMA:DR:eq.2}
d_{t+1} = d_{t} + 1.
\end{equation}
First,
we consider the case in which
the routine executes Case~3.3 for the $t$-th execution,
that is,
$q_{t} = q'_{t}+1$
(i.e., $q_{t+1} = q'_{t}+2$)
and
$q_{t+1} = q_{t}+1 \leq n-1$
(Fig.~\ref{fig:ldr}(5),(14)).
In this case,
it is sufficient that we prove (a) is true
because $q_{t}+1 \leq n-1$.
$(p_{t}, q_{t}+1)$ dominates the five vertices $(p_{t}, q_{t}),(p_{t}, q_{t}+1),(p_{t}-1, q_{t}+1),(p_{t}+1, q_{t}+1)$ and $(p_{t}, q_{t}+2)$.
By the definition of $(p'_{t}, q'_{t})$-$(p_{t}, q_{t})$-regularity,
vertices in $R(D_{t})$ do not dominate none of $(p_{t}-1, q_{t}+1),(p_{t}+1, q_{t}+1)$ and $(p_{t}, q_{t}+2)$.
Thus,
the number of vertices which vertices in $R(D_{t+1})$ dominate increases by three vertices.
Since (a) is true for the $t$-th execution by the induction hypothesis,
dominators and a pre-connector in $D_{t}$ dominate $3 d_{t} +1$ vertices.
Thus,
we have using Eq.~(\ref{LMA:DR:eq.2}),
\[
3 d_{t} + 1 + 3 = 3(d_{t} + 1) + 1 = 3 d_{t+1} + 1,
\]
which implies that
(a) is also true for $j = t+1$
if the routine executes Case~3.3 for the $t$-th execution.
Next,
we consider the case in which
the routine executes Case~2 for the $t$-th and $t+2$-nd executions,
that is,
$q_{t} > q'_{t}+1$,
($q_{t+1} > q'_{t}+2$)
and
$q_{t+1} = q_{t}+1 \leq n-1$
(Fig.~\ref{fig:ldr}(6),(15)).
In this case,
it suffices to prove that (a) is true
because $q_{t}+1 \leq n-1$.
$(p_{t}, q_{t}+1)$ newly dominates the three vertices $(p_{t}-1, q_{t}+1),(p_{t}+1, q_{t}+1)$ and $(p_{t}, q_{t}+2)$, and
vertices in $R(D_{t})$ do not dominate them.
Thus,
the number of vertices dominated by vertices in $R(D_{t+1})$ increase by three vertices and by Eq.~(\ref{LMA:DR:eq.2}),
\[
3 d_{t} + 1 + 3 = 3(d_{t} + 1) + 1 = 3 d_{t+1} + 1.
\]
(a) is true for $j = t+1$
if the routine executes Case~2 for the $t$-th and $t+2$-nd executions.
Finally,
we consider the case in which
the routine executes Case~2 for the $t$-th execution and
does not execute Case~2 for the $t+2$-nd execution,
that is,
$q_{t+1} = q_{t}+1 = n$
(Fig.~\ref{fig:ldr}(7),(16)).
$(p_{t}, q_{t}+1)$ newly dominates the two vertices $(p_{t}-1, n)$ and $(p_{t}+1, n)$, and
vertices in $R(D_{t})$ do not dominate them
by the $(p'_{t}, q'_{t})$-$(p_{t}, q_{t})$-regularity.
Thus,
the number of vertices dominated by vertices in $R(D_{t+1})$ increases by two vertices.
Using~Eq.~(\ref{LMA:DR:eq.2}),
we have
\[
3 d_{t} + 1 + 2 = 3(d_{t} + 1) = 3 d_{t+1},
\]
which implies
(b) is also true in this case when $j = t+1$.
\noindent
{\bf Case~3.1:}
If the routine executes Case~3.1,
then a vertex which is not in $D_{t}$ but in $D_{t+1}$ is only $(2, 3)$,
which is a pre-connector, and dominators do not change
(Fig.~\ref{fig:ldr}(4)).
Thus,
\begin{equation} \label{LMA:DR:eq.3}
d_{t+1} = d_{t}.
\end{equation}
Also,
$(2, 3)$ dominates $(2, 4)$,
which is not dominated by vertices in $R(D_{t})$.
Since (b) is true for the $t$-th execution by the induction hypothesis,
dominators in $D_{t}$ dominate $3 d_{t}$ vertices.
Together with Eq.~(\ref{LMA:DR:eq.3}),
we have
\[
3 d_{t} + 1 = 3 d_{t+1} + 1,
\]
which implies that
(a) is true for $j = t+1$ in this case.
\noindent
{\bf Case~3.2:}
We can prove this case similarly to the previous case
(Fig.~\ref{fig:ldr}(9)).
If the routine executes Case~3.2,
then a vertex which is not in $D_{t}$ but in $D_{t+1}$ is only $(p_{t}, q_{t})$,
which is a pre-connector, and dominators do not change.
Thus,
\begin{equation} \label{LMA:DR:eq.4}
d_{t+1} = d_{t}.
\end{equation}
Also,
$(p_{t}, q_{t})$ dominates $(p_{t}+1, q_{t})$,
which is not dominated by vertices in $R(D_{t})$.
Since (b) is true for the $t$-th execution by the induction hypothesis,
dominators in $D_{t}$ dominate $3 d_{t}$ vertices.
Together with Eq.~(\ref{LMA:DR:eq.4}),
we have
\[
3 d_{t} + 1 = 3 d_{t+1} + 1,
\]
which implies that
(a) is true for $j = t+1$ in this case.
\noindent
{\bf Case~3.3:}
We omit the proof of this case
because we prove it similarly to that of Case~3.2
(Fig.~\ref{fig:ldr}(13)).
\noindent
{\bf Case~3.4:}
The routine does not do anything in Case~3.4, and $D_{t+1} = D_{t}$ hold.
Hence,
(a) and (b) hold by the induction hypothesis.
We have shown that the both properties hold for the $t+1$-st execution.
\fi
\end{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
\fi
\begin{LMA}\label{LMA:PQ}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Suppose that the routine is feasible.
Then,
the following properties hold:
\begin{itemize}
\itemsep=-2.0pt
\setlength{\leftskip}{0pt}
\item[(i)]
For any $i \in [1, t_{1}-1]$,
$p'_{i} = 0$
and
$q'_{i} = 2$.
\item[(ii)]
For any $i \geq t_{1}$,
$(p'_{i} \mod 3) = 2$
and
$(q'_{i} \mod 3) = 2$.
\end{itemize}
\fi
\end{LMA}
\begin{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
By definition,
$p'_{0} = 0, q'_{0} = 2, p_{0} = 1$ and $q_{0} = 2$
(Fig.~\ref{fig:ldr}(1)).
We can prove the following by induction on the number $i$ of times that the routine is executed:
For each $i = 1, \ldots, t_{1}-1$,
$q'_{i-1} = q_{i-1}$
and
$p_{i} \leq m-1$.
Also,
the routine executes Case~1 for the $i$-th execution,
$p'_{i} = p'_{0} = 0$,
$p_{i} = p_{i-1} + 1$ and
$q'_{i} = q_{i} = q'_{0} = 2$
(Fig.~\ref{fig:ldr}(2),(3)).
Thus,
(i) in the statement of this lemma holds.
Since $p'_{t_{1}-1} = p'_{0} = 0$,
$p_{t_{1}-1} = m$
and
$q'_{t_{1}-1} = q_{t_{1}-1} = q'_{0} = 2$,
the routine executes Case~3.1 for the $t_{1}$-the execution
(Fig.~\ref{fig:ldr}(4)).
Hence,
$p'_{t_{1}} = p_{t_{1}} = 2$,
$q'_{t_{1}} = 2$
and
$q_{t_{1}} = 3$.
We can prove the following by induction on the number $i$ of times that the routine is executed:
For each $i = t_{1} + 1, \ldots, t_{2}-1$,
the routine executes Case~2,
$p'_{i} = p_{i} = 2$,
$q'_{i} = 2$
and
$q_{i} = q_{i-1} + 1$
(Fig.~\ref{fig:ldr}(5),(6),(7)).
Since $p'_{t_{2}-1} = p_{t_{2}-1} = 2$,
$q'_{t_{2}-1} = 2$
and
$q_{t_{2}-1} = n$,
the routine executes either Case~3.2 or 3.3 for the $t_{2}$-th execution
(Fig.~\ref{fig:ldr}(8)).
First,
let us consider the case in which
the routine executes Case~3.2 for the $t_{j} \hspace{1mm} (j = 2, \ldots, k-1)$-th execution.
Then,
$p'_{t_{j}} = p'_{t_{j}-1}$,
$p_{t_{j}} = p'_{t_{j}-1} + 1$
and
$q'_{t_{j}} = q_{t_{j}} = q'_{t_{j}-1} + 3$
(Fig.~\ref{fig:ldr}(9)).
We can prove by induction on $i$ the following:
the routine executes Case~1 for the $i = t_{j} + 1, \ldots, t_{j+1}-2$-th execution,
$p'_{i} = p'_{t_{j}}$,
$p_{i} = p_{i-1} + 1$
and
$q'_{i} = q_{i} = q'_{t_{j}}$
(Fig.~\ref{fig:ldr}(10),(11)).
Moreover,
the routine executes Case~1 for the $t_{j+1}-1$-th execution,
$p'_{t_{j+1}-1} = p'_{t_{j}}$,
$p_{t_{j+1}-1} = m$
and
$q'_{t_{j+1}-1} = q_{t_{j+1}-1} = q'_{t_{j}}$
(Fig.~\ref{fig:ldr}(12)).
Next,
we consider the case in which
the routine executes Case~3.3 for the $t_{j} \hspace{1mm} (j = 2, \ldots, k-1)$-th execution.
Then,
$q'_{t_{j}} = q'_{t_{j}-1}$,
$q_{t_{j}} = q'_{t_{j}-1} + 1$
and
$p'_{t_{j}} = p_{t_{j}} = p'_{t_{j}-1} + 3$
(Fig.~\ref{fig:ldr}(13)).
We can prove by induction on $i$ the following:
the routine executes Case~2 for the $i = t_{j} + 1, \ldots, t_{j+1}-2$-th execution,
$q'_{i} = q'_{t_{j}}$,
$q_{i} = q_{i-1} + 1$
and
$p'_{i} = p_{i} = p'_{t_{j}}$
(Fig.~\ref{fig:ldr}(14),(15)).
Moreover,
the routine executes Case~2 for the $t_{j+1}-1$-th execution,
$q'_{t_{j+1}-1} = q'_{t_{j}}$,
$q_{t_{j+1}-1} = n$
and
$p'_{t_{j+1}-1} = p_{t_{j+1}-1} = p'_{t_{j}}$
(Fig.~\ref{fig:ldr}(16)).
By the above argument,
$p'_{t_{1}} = 2$
and
$q'_{t_{1}} = 2$.
Furthermore,
for any $i \geq t_{1} + 1$,
if the routine executes Case~3.2,
then the value of $q'_{i}$ increases by three.
Otherwise,
its value does not change.
Similarly,
if the routine executes Case~3.3,
then the value of $p'_{i}$ increases by three, and otherwise,
its value does not change.
Therefore,
(ii) in the statement holds,
which completes the proof.
\fi
\end{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
\fi
\begin{LMA}\label{LMA:CR0}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Suppose that the routine is feasible.
Then,
for any connector in $D_{\tau}$,
vertices dominated by the connector are dominated by at least one dominator in $D_{\tau}$.
\fi
\end{LMA}
\begin{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Suppose that the routine is feasible.
We prove the statement of this lemma by induction on the number $j$ of times that the routine is executed.
When $j = 0$,
that is,
at the initialization of the routine,
$D_{0}$ is $(0, 2)$-$(1, 2)$-regular.
At this point,
only $(2, 1)$ is regular and $D_{0}$ does not contain a connector.
Thus,
the statement is true
(Fig.~\ref{fig:ldr}(1)).
We assume that
the statement is true for the $j = t$-th execution of the routine and
show that it is also true for the $t+1$-st execution.
We will discuss each Case executed by the routine for the $t+1$-st execution.
Note that only irregular vertices are handled at the regularization,
and regular vertices are not removed from $D_{t}$ when the routine constructs $D_{t+1}$.
\noindent
{\bf Case~1:}
First,
we consider the case in which the routine executes Case~1 for the $t+1$-st execution.
In this case,
$D_{t+1}$ is $(p'_{t}, q'_{t})$-$(p_{t}+1, q_{t})$-regular.
If either $t = 0$ or
$p_{t} > p'_{t}+1$,
then connectors in $D_{t+1}$ are equal to those in $D_{t}$
and thus,
the statement is true by the induction hypothesis.
Second,
we consider the case in which
$t \ne 0$
and
$p_{t} = p'_{t}+1$,
that is,
$p_{t+1} = p'_{t}+2$
(Fig.~\ref{fig:ldr}(10)).
By definition,
$(p_{t}, q_{t})$ is a pre-connector in $D_{t}$ and is a connector in $D_{t+1}$.
$(p_{t}, q_{t})$ dominates the five vertices $(p_{t}, q_{t}), (p_{t}-1, q_{t}), (p_{t}, q_{t}-1), (p_{t}, q_{t}+1)$ and $(p_{t}+1, q_{t})$.
Also,
$(p_{t}, q_{t})$ is a connector and
the four vertices $(p_{t}+1, q_{t}), (p_{t}-1, q_{t}), (p_{t}-1, q_{t}-1)$ and $(p_{t}-1, q_{t}+1)$ belong to $D_{t+1}$.
The four vertices dominate all the vertices which $(p_{t}, q_{t})$ dominates.
Thus,
it suffices to show that
the four vertices are dominators.
Since $(p_{t}, q_{t}-1) \notin R(D_{t+1})$ by (Q2) in the regularity condition,
$(p_{t}+1, q_{t})$ satisfies the condition of neither a connector nor a pre-connector, and
is neither of them.
Similarly,
$(p_{t}-1, q_{t}+1)$ ($(p_{t}-1, q_{t}), (p_{t}-1, q_{t}-1)$) is neither a connector nor a pre-connector
because $(p_{t}-2, q_{t}), (p_{t}-2, q_{t}-1), (p_{t}-2, q_{t}-2) \notin R(D_{t+1})$
by the condition (Q7).
Thus,
vertices which $(p_{t}, q_{t})$ dominates are dominated by dominators.
The other connectors except for $(p_{t}, q_{t})$ do not change,
and thus, the statement is true by the induction hypothesis.
\noindent
{\bf Case~2:}
We prove this case similarly to the proof of Case~1.
When the routine executes Case~2 for the $t+1$-st execution,
$D_{t+1}$ is $(p'_{t}, q'_{t})$-$(p_{t}, q_{t}+1)$-regular.
If $q_{t} > q'_{t}+1$,
then connectors in $D_{t+1}$ are equal to those in $D_{t}$
and thus,
the statement is true by the induction hypothesis.
Second,
we consider the case in which
$q_{t} = q'_{t}+1$,
that is,
$q_{t+1} = q'_{t}+2$
(Fig.~\ref{fig:ldr}(14)).
By definition,
$(p_{t}, q_{t})$ is a pre-connector in $D_{t}$ and is a connector in $D_{t+1}$.
$(p_{t}, q_{t})$ dominates the five vertices $(p_{t}, q_{t}), (p_{t}-1, q_{t}), (p_{t}, q_{t}-1), (p_{t}, q_{t}+1)$ and $(p_{t}+1, q_{t})$.
Also,
$(p_{t}, q_{t})$ is a connector and
the four vertices $(p_{t}+1, q_{t}), (p_{t}-1, q_{t}), (p_{t}-1, q_{t}-1)$ and $(p_{t}-1, q_{t}+1)$ belong to $D_{t+1}$.
The four vertices dominate all the vertices which $(p_{t}, q_{t})$ dominates.
Thus,
it suffices to show that
the four vertices are dominators.
Since $(p_{t}-1, q_{t}) \notin R(D_{t+1})$ by (P2) in the regularity condition,
$(p_{t}, q_{t}+1)$ satisfies the condition of neither a connector nor a pre-connector, and
is neither of them.
Similarly,
$(p_{t}+1, q_{t}-1)$ ($(p_{t}, q_{t}-1), (p_{t}-1, q_{t}-1)$) is neither a connector nor a pre-connector
because $(p_{t}, q_{t}-2), (p_{t}-1, q_{t}-2), (p_{t}-2, q_{t}-2) \notin R(D_{t+1})$
by the condition (P7).
Thus,
vertices which $(p_{t}, q_{t})$ dominates are dominated by dominators.
The other connectors except for $(p_{t}, q_{t})$ do not change,
and thus, the statement is true by the induction hypothesis.
\noindent
{\bf Case~3.1, 3.2 or 3.3:}
If the routine executes Case~3.1, 3.2 or 3.3,
then a regular vertex newly add into $D_{t+1}$ is only a pre-connector and dominators do not change (Fig.~\ref{fig:ldr}(4),(9),(13)).
Thus,
the statement is true by the induction hypothesis.
\noindent
{\bf Case~3.4:}
If the routine executes Case~3.4,
then the routine does nothing and thus,
the statement is true by the induction hypothesis.
We have shown that the statement is also true for the $t+1$-st execution.
\fi
\end{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
\fi
\begin{LMA}\label{LMA:IR}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Suppose that the routine is feasible.
Then,
\[
{a}_{m,n} = (m \mod 3) \cdot (n \mod 3)
\]
and
\[
\bar{r}_{m,n} =
\begin{cases}
3 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 4 \\
2 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 2 \\
1 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 1 \\
0 & \hspace{10mm} (m \mod 3) \cdot (n \mod 3) = 0.
\end{cases}
\]
\fi
\end{LMA}
\begin{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Suppose that the routine is feasible.
If the routine executes Case~3.4 and is finished,
then one of the following conditions holds at that time:
(a)
$p'_{\tau} = p_{\tau} = m-1$,
(b)
$q'_{\tau} = q_{\tau} = n-1$ and
(c)
both $m-3 \leq p'_{\tau} \leq m-2$
and
$n-3 \leq q'_{\tau} \leq n-2$.
In what follows,
we evaluate the values of $a_{m,n}$ and $\bar{r}_{m,n}$
for each of these three cases.
\noindent
{\bf (a):}
First,
we consider the case of $p'_{\tau} = p_{\tau} = m-1$.
It follows from (ii) in Lemma~\ref{LMA:PQ}
that $(p'_{\tau} \mod 3) = 2$,
that is,
$(p'_{\tau} + 1 \mod 3) = 0$
(Fig.~\ref{fig:lir}).
Thus,
$(m \mod 3) = 0$ by the above equalities.
That is,
\[
(m \mod 3) \cdot (n \mod 3) = 0.
\]
On the other hand,
no irregular vertex exists when the routine is finished
because $D_{\tau}$ is $(m-1, q'_{\tau})$-$(m-1, n)$-regular.
Hence,
\[
\bar{r}_{m,n} = 0.
\]
Also,
since all the vertices are dominated by regular vertices in $D_{\tau}$,
we have
\[
a_{m,n} = 0,
\]
which implies that
\[
a_{m,n} = (m \mod 3) \cdot (n \mod 3).
\]
\noindent
{\bf (b):}
We can prove the case (b) similarly to the case (a).
It follows from (ii) in Lemma~\ref{LMA:PQ}
that $(q'_{\tau} \mod 3) = 2$,
that is,
$(q'_{\tau} + 1 \mod 3) = 0$.
Thus,
$(n \mod 3) = 0$ by the above equalities.
That is,
\[
(m \mod 3) \cdot (n \mod 3) = 0.
\]
On the other hand,
no irregular vertex exists when the routine is finished
because $D_{\tau}$ is $(p'_{\tau}, n-1)$-$(m, n-1)$-regular.
Hence,
\[
\bar{r}_{m,n} = 0.
\]
Also,
since all the vertices are dominated by regular vertices in $D_{\tau}$,
we have
\[
a_{m,n} = 0,
\]
which implies that
\[
a_{m,n} = (m \mod 3) \cdot (n \mod 3).
\]
\noindent
{\bf (c):}
Since $(p'_{\tau} \mod 3) = 2$ by (ii) in Lemma~\ref{LMA:PQ},
$(p'_{\tau} + 1 \mod 3) = 0$.
Hence,
if $p'_{\tau} = m-3$,
then
$(m-3 + 1 \mod 3) = (m-2 \mod 3) = 0$,
that is,
$(m \mod 3) = 2$.
If $p'_{\tau} = m-2$,
then
$(m \mod 3) = 1$.
Also,
$(q'_{\tau} + 1 \mod 3) = 0$ by (ii) in Lemma~\ref{LMA:PQ}.
Hence,
if $q'_{\tau} = n-3$,
then
$(n \mod 3) = 2$.
If $q'_{\tau} = n-2$,
then
$(n \mod 3) = 1$.
Let us call these facts Fact~(A).
In what follows,
we discuss each case for $m$ and $n$.
\noindent
{\bf\boldmath (c-1) $p'_{\tau} = m-3$ and $q'_{\tau} = n-3$:}
By Fact~(A),
\[
(m \mod 3) \cdot (n \mod 3) = 2 \cdot 2 = 4.
\]
Then,
$D_{\tau}$ is
$(m-3, n-3)$-$(m-3, n)$-regular or
$(m-3, n-3)$-$(m, n-3)$-regular.
Thus,
the irregular vertices in $D_{\tau}$ are
$
(m-2, n-2), (m-2, n-1), (m-2, n),
(m-1, n-2), (m-1, n-1), (m-1, n),
(m, n-2), (m, n-1)$ and $(m, n)
$
by definition.
The vertices which are not dominated by regular vertices in $D_{\tau}$ are
the four vertices $(m-1, n-1), (m-1, n), (m, n-1)$ and $(m, n)$,
which implies that
\[
a_{m,n} = 4.
\]
Hence,
\[
a_{m,n} = (m \mod 3) \cdot (n \mod 3).
\]
Additionally,
$(m-3,n-3),
(m-3,n-2), (m-3,n-1), (m-3,n),
(m-2,n-3), (m-1,n-3), (m,n-3)
\in D_{\tau}$
because $D_{\tau}$ is
$(m-3, n-3)$-$(m-3, n)$-regular or
$(m-3, n-3)$-$(m, n-3)$-regular.
We need three vertices to dominate
the vertices $(m-1, n-1), (m-1, n), (m, n-1)$ and $(m, n)$,
and hence
\[
\bar{r}_{m,n} = 3.
\]
\noindent
{\bf\boldmath (c-2) $p'_{\tau} = m-3$ and $q'_{\tau} = n-2$:}
By Fact~(A),
\[
(m \mod 3) \cdot (n \mod 3) = 2 \cdot 1 = 2.
\]
Then,
$D_{\tau}$ is
$(m-3, n-2)$-$(m-3, n)$-regular or
$(m-3, n-2)$-$(m, n-2)$-regular.
Thus,
the irregular vertices in $D_{\tau}$ are
$
(m-2, n-1), (m-2, n),
(m-1, n-1), (m-1, n),
(m, n-1)$ and $(m, n)
$
by definition.
The vertices which are not dominated by regular vertices in $D_{\tau}$ are
the two vertices $(m-1, n)$ and $(m, n)$,
which implies that
\[
a_{m,n} = 2.
\]
Hence,
\[
a_{m,n} = (m \mod 3) \cdot (n \mod 3).
\]
Additionally,
$(m-3,n-2),
(m-3,n-1), (m-3,n),
(m-2,n-2), (m-1,n-2), (m,n-2)
\in D_{\tau}$
because $D_{\tau}$ is
$D_{\tau}$ is
$(m-3, n-2)$-$(m-3, n)$-regular or
$(m-3, n-2)$-$(m, n-2)$-regular.
We need two vertices to dominate
the vertices $(m-1, n)$ and $(m, n)$,
and hence
\[
\bar{r}_{m,n} = 2.
\]
\noindent
{\bf\boldmath (c-3) $p'_{\tau} = m-2$ and $q'_{\tau} = n-3$:}
By Fact~(A),
\[
(m \mod 3) \cdot (n \mod 3) = 1 \cdot 2 = 2.
\]
Then,
$D_{\tau}$ is
$(m-2, n-3)$-$(m-2, n)$-regular or
$(m-2, n-3)$-$(m, n-3)$-regular.
Thus,
the irregular vertices in $D_{\tau}$ are
$
(m-1, n-2), (m-1, n-1), (m-1, n),
(m, n-2), (m, n-1)$ and $(m, n)
$
by definition.
The vertices which are not dominated by regular vertices in $D_{\tau}$ are
the two vertices $(m, n-1)$ and $(m, n)$,
which implies that
\[
a_{m,n} = 2.
\]
Hence,
\[
a_{m,n} = (m \mod 3) \cdot (n \mod 3).
\]
Additionally,
$(m-2,n-3),
(m-2,n-2), (m-2,n-1), (m-2,n),
(m-2,n-3), (m-1,n-3), (m,n-3)
\in D_{\tau}$
because $D_{\tau}$ is
$D_{\tau}$ is
$(m-2, n-3)$-$(m-2, n)$-regular or
$(m-2, n-3)$-$(m, n-3)$-regular.
We need two vertices to dominate
the vertices $(m, n-1)$ and $(m, n)$,
and hence
\[
\bar{r}_{m,n} = 2.
\]
\noindent
{\bf\boldmath (c-4) $p'_{\tau} = m-2$ and $q'_{\tau} = n-2$:}
By Fact~(A),
\[
(m \mod 3) \cdot (n \mod 3) = 1 \cdot 1 = 1.
\]
Then,
$D_{\tau}$ is
$(m-2, n-2)$-$(m-2, n)$-regular or
$(m-2, n-2)$-$(m, n-2)$-regular.
Thus,
the irregular vertices in $D_{\tau}$ are
$
(m-1, n-1), (m-1, n),
(m, n-1)$ and $(m, n)
$
by definition.
The vertex which are not dominated by regular vertices in $D_{\tau}$ is
the one vertex $(m, n)$,
which implies that
\[
a_{m,n} = 1.
\]
Hence,
\[
a_{m,n} = (m \mod 3) \cdot (n \mod 3).
\]
Additionally,
$(m-2,n-2),
(m-2,n-1), (m-2,n),
(m-1,n-2), (m,n-2)
\in D_{\tau}$
because $D_{\tau}$ is
$D_{\tau}$ is
$(m-2, n-2)$-$(m-2, n)$-regular or
$(m-2, n-2)$-$(m, n-2)$-regular.
We need one vertex to dominate
the vertex $(m, n)$,
and hence
\[
\bar{r}_{m,n} = 1.
\]
\fi
\end{proof}
\ifnum \count12 > 0
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig_lir.eps}
\end{center}
\caption{
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Vertices which are not dominated by regular vertices and irregular vertices when the routine is finished.
Black circles denote vertices in the MCDS $D_{\tau}$.
Vertices enclosed by dashed lines are irregular and the other vertices are regular.
In (a) and (b), there do not exist irregular vertices.
Gray triangles denote vertices which are not dominated by regular vertices in $D_{\tau}$.
\fi
}
\label{fig:lir}
\end{figure*}
\fi
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
\fi
\begin{LMA}\label{LMA:CR1}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
If the routine is feasible, then
\[
{c}_{m,n} \geq
\begin{cases}
\min\{ \frac{m}{3}, \frac{n}{3} \} & (m \mod 3) = 0 \mbox{ and } (n \mod 3) = 0 \\
\frac{m}{3} & (m \mod 3) = 0 \mbox{ and } (n \mod 3) \ne 0 \\
\frac{n}{3} & (m \mod 3) \ne 0 \mbox{ and } (n \mod 3) = 0 \\
\lfloor \frac{m}{3} \rfloor + \lfloor \frac{n}{3} \rfloor - 1 & (m \mod 3) \ne 0 \mbox{ and } (n \mod 3) \ne 0. \\
\end{cases}
\]
\fi
\end{LMA}
\begin{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Suppose that the routine is feasible.
The routine executes Case~3.1, 3.2 or 3.3 for the $i = t_{j} \hspace{1mm} (j = 1, \ldots, k-1)$-th execution by definition.
If the routine executes Case~3.1,
then the vertex $(p'_{t_{1}}, q'_{t_{1}}+1) (= (2, 3))$ becomes a pre-connector
(Fig.~\ref{fig:ldr}(4)).
If the routine executes Case~3.2,
then $(p'_{t_{j}}+1, q'_{t_{j}}+3)$ becomes a pre-connector
(Fig.~\ref{fig:ldr}(9)).
If the routine executes Case~3.3,
then $(p'_{t_{j}}+3, q'_{t_{j}}+1)$ becomes a pre-connector
(Fig.~\ref{fig:ldr}(13)).
Also,
the routine executes Case~1 or 2 for the $i = t_{j} + 1\hspace{1mm} (j = 1, \ldots, k-1)$-st execution.
If the routine executes Case~1 for $t_{j} + 1$-st,
then $(p_{t_{j}}, q_{t_{j}}) (= (p_{t_{j}+1}-1, q_{t_{j}+1}))$ which is a pre-connector becomes a connector.
If the routine executes Case~2,
$(p_{t_{j}}, q_{t_{j}}) (= (p_{t_{j}+1}, q_{t_{j}+1}-1))$ which is a pre-connector becomes a connector.
Hence,
the number of connectors at the time when the routine is finished is equal to the number of times that the routine executes Case~3.1, 3.2 or 3.3.
Let $c_{1}$ ($c_{2},c_{3}$, respectively) denote the number of times that the routine executes Case~3.1 (3.2, 3.3, respectively)
by the time when the routine is finished.
By definition,
\[
c_{m,n} = c_{1} + c_{2} + c_{3}.
\]
In what follows,
we show the inequalities in the statement by evaluating lower bounds on $c_{1},c_{2}$ and $c_{3}$.
The routine executes Case~1 for the $i \in [1, t_{1}-1]$-th execution.
Since we assume that $m \geq 4$ and $n \geq 4$ throughout this paper,
the routine executes Case~3.1 for the $t_{1}$-th execution.
Also,
the routine does not execute Case~3.1 after the $t_{1}$-th execution
because $(p'_{i} \mod 3) = 2$ for any $i > t_{1}$ by (ii) in Lemma~\ref{LMA:PQ}.
Thus,
$c_{1} = 1$ holds.
At the time when the routine executes Case~3.4,
one of the following three conditions holds:
(a)
$p'_{\tau} = p_{\tau} = m-1$,
(b)
$q'_{\tau} = q_{\tau} = n-1$ and
(c)
both $m-3 \leq p'_{\tau} \leq m-2$
and
$n-3 \leq q'_{\tau} \leq n-2$.
We evaluate lower bounds on $c_{2}$ and $c_{3}$ for each case of the three cases.
\noindent
{\bf (a):}
When the routine executes Case~3.1 for the $t_{1}$-th execution,
$p'_{t_{1}} = 2$ by the condition of the routine.
Using (ii) in Lemma~\ref{LMA:PQ},
we have for any $i > t_{1}$,
$(p'_{i} \mod 3) = 2$.
Only if the routine executes Case~3.3,
the value of $p'_{i}$ increases by three.
For the execution of the other cases,
it does not change.
Moreover,
$p'_{\tau} = m-1$
by the condition of (a).
Hence,
the number of times that the routine executes Case~3.3 is
\[
c_{3} = \frac{m-1 - 2}{3} = \frac{m}{3} - 1.
\]
Therefore,
\begin{equation} \label{LMA:CR1:eq.1}
c_{m,n} \geq c_{1} + c_{3} = 1 + \frac{m}{3} - 1 = \frac{m}{3}.
\end{equation}
\noindent
{\bf (b):}
We can prove this case similarly to that of (a).
When the routine executes Case~3.1 for the $t_{1}$-th execution,
$q'_{t_{1}} = 2$ by the condition of the routine.
Using (ii) in Lemma~\ref{LMA:PQ},
we have for any $i > t_{1}$,
Only if the routine executes Case~3.2,
the value of $q'_{i}$ increases by three.
For the execution of the other cases,
it does not change.
Moreover,
$q'_{\tau} = n-1$
by the condition of (b).
Hence,
the number of times that the routine executes Case~3.2 is
\[
c_{2} = \frac{n-1 - 2}{3} = \frac{n}{3} - 1.
\]
Therefore,
\begin{equation} \label{LMA:CR1:eq.2}
c_{m,n} \geq c_{1} + c_{2} = 1 + \frac{n}{3} - 1 = \frac{n}{3}.
\end{equation}
\noindent
{\bf (c):}
If $p'_{\tau} = m-3$ by the condition of (c),
we can evaluate the number of times that the routine executes Case~3.3 similarly to the case (a):
\[
c_{3} = \frac{m-3 - 2}{3} = \frac{m-2}{3} - 1.
\]
Since $(p'_{\tau}-2 \mod 3) = 0$ by (ii) in Lemma~\ref{LMA:PQ},
$\frac{m-2}{3} = \lfloor \frac{m}{3} \rfloor$,
which implies that
\begin{equation} \label{LMA:CR1:eq.3}
c_{3} = \frac{m-2}{3} - 1 = \lfloor \frac{m}{3} \rfloor - 1.
\end{equation}
Similarly,
if $p'_{\tau} = m-2$,
then
\begin{equation} \label{LMA:CR1:eq.4}
c_{3} = \frac{m-2 - 2}{3} - 1 = \lfloor \frac{m}{3} \rfloor - 1.
\end{equation}
If $q'_{\tau} = n-3$ by the condition of (c),
we can evaluate the number of times that the routine executes Case~3.2
similarly to the case (b) and
it is
\[
c_{2} = \frac{n-3 - 2}{3} = \frac{n-2}{3} - 1.
\]
Since $(q'_{\tau}-2 \mod 3) = 0$ by (ii) in Lemma~\ref{LMA:PQ},
\begin{equation} \label{LMA:CR1:eq.5}
c_{2} = \frac{n-2}{3} - 1 = \lfloor \frac{n}{3} \rfloor - 1.
\end{equation}
Similarly,
if $q'_{\tau} = n-2$,
then
\begin{equation} \label{LMA:CR1:eq.6}
c_{2} = \frac{n-2 - 2}{3} - 1 = \lfloor \frac{n}{3} \rfloor - 1.
\end{equation}
By Eqs.~(\ref{LMA:CR1:eq.3}), (\ref{LMA:CR1:eq.4}), (\ref{LMA:CR1:eq.5}) and (\ref{LMA:CR1:eq.6}),
\begin{equation} \label{LMA:CR1:eq.7}
c_{m,n} = c_{1} + c_{2} + c_{3}
= 1 + \lfloor \frac{m}{3} \rfloor - 1 + \lfloor \frac{n}{3} \rfloor - 1
= \lfloor \frac{m}{3} \rfloor + \lfloor \frac{n}{3} \rfloor - 1.
\end{equation}
By the above argument,
if $(m \mod 3) = 0$ and $(n \mod 3) = 0$,
then $(m-1 \mod 3) = 2$ and $(n-1 \mod 3) = 2$.
Also,
either $p'_{\tau} = m-1$ or
$q'_{\tau} = n-1$
because $(p'_{\tau} \mod 3) = 2$ and $(q'_{\tau} \mod 3) = 2$ by (ii) in Lemma~\ref{LMA:PQ}.
Thus,
either (a) or (b) is true
Then,
by Eqs.~(\ref{LMA:CR1:eq.1}) and (\ref{LMA:CR1:eq.2}),
\[
c_{m,n} \geq \min \left \{ \frac{m}{3}, \frac{n}{3} \right \}.
\]
In the case in which both $(m \mod 3) = 0$ and $(n \mod 3) \ne 0$,
$p'_{\tau} = m-1$ holds,
which satisfies the condition of (a).
Thus,
by Eq.~(\ref{LMA:CR1:eq.1}),
\[
c_{m,n} \geq \frac{m}{3}.
\]
In the case in which both $(m \mod 3) \ne 0$ and $(n \mod 3) = 0$,
$q'_{\tau} = n-1$ holds,
which satisfies the condition of (b).
Thus,
by Eq.~(\ref{LMA:CR1:eq.2}),
\[
c_{m,n} \geq \frac{n}{3}.
\]
The case in which $(m \mod 3) \ne 0$ and $(n \mod 3) \ne 0$
satisfies the condition of (c).
Thus,
by Eq.~(\ref{LMA:CR1:eq.7}),
\[
c_{m,n} = \lfloor \frac{m}{3} \rfloor + \lfloor \frac{n}{3} \rfloor - 1.
\]
\fi
\end{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
\fi
\subsection{Routine Feasibility} \label{sec:routinefeasibility}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
In this section,
we show that the routine is feasible and can obtain the MCDS $D_{\tau}$,
of which properties we have shown in the previous section.
Specifically,
we will complete the proof of the lower bound lemma by showing that
the routine can conduct regularization in each of Cases~1, 2, 3.1, 3.2 and 3.3.
We give a definition for the following lemmas.
For a CDS $D$,
a vertex $(x, y) \in {\overline{R}}(D)$ is called a {\em mobile in $D$}
if the vertex satisfies at least one of the following five conditions
(see Fig.~\ref{fig:mv}):
\begin{itemize}
\itemsep=-2.0pt
\setlength{\leftskip}{0pt}
\item[(i)]
$x \in [2, m-1]$,
$y \in [2, n-1]$
and
$(x+1, y), (x-1, y), (x-1, y-1), (x-1, y+1) \in D$,
\item[(ii)]
$x \in [2, m-1]$
and
$(x+1, n), (x-1, n), (x-1, n-1) \in D$,
\item[(iii)]
$x \in [2, m-1]$,
$y \in [2, n-1]$
and
$(x, y+1), (x, y-1), (x-1, y-1), (x+1, y-1) \in D$,
\item[(iv)]
$y \in [2, n-1]$
and
$(m, y+1), (m, y-1), (m-1, y-1) \in D$,
and
\item[(v)]
$(1, 4), (1, 2), (2, 2) \in D$.
\end{itemize}
For a CDS $D$,
when a vertex set $C$ is constructed by removing a mobile $v$ in $D$,
$C$ may not be connected, but still remains dominating.
Note that
if we obtain a new vertex set by adding another vertex into $C$ instead of $v$ to be connected,
then the new vertex set becomes a CDS different from $D$.
\fi
\ifnum \count12 > 0
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig_mv.eps}
\end{center}
\caption{
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Circles denote vertices in a CDS in the conditions of mobiles.
White circles denote mobile vertices.
\fi
}
\label{fig:mv}
\end{figure*}
\fi
\begin{LMA}\label{LMA:L2}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Suppose that $D$ is a $(p', q')$-$(p, q)$-regular CDS.
Also,
suppose that
for vertices $\overline{v} \in {\overline{R}}(D)$ and $u \in {R}(D)$,
there exists a simple path $P$ consisting of vertices in $D$ between $u$ and $\overline{v}$
such that
(i)
the number of vertices in ${\overline{R}}(D)$ in $P$ is minimum,
(ii)
$P$ does not contain $(p, q)$,
(iii)
if $q' = q$,
then $P$ does not contain $(p, q+1)$ and
(iv)
if $p' = p$,
then $P$ does not contain $(p+1, q)$.
Then,
$P$ contains a mobile in $D$.
\fi
\end{LMA}
\begin{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Let $D$ be a $(p', q')$-$(p, q)$-regular CDS.
Let $\overline{v} \in {\overline{R}}(D)$ and $u \in {R}(D)$ be vertices
such that
there exists a simple path $P$ consisting of vertices in $D$ between $u$ and $\overline{v}$ satisfying the four conditions in the statement of this lemma.
Since $\overline{v} \in {\overline{R}}(D)$ and $u \in {R}(D)$,
vertices
$\overline{v}' \in {\overline{R}}(D)$
and
$u' \in {R}(D)$
are contained in $P$
such that $\overline{v}'$ and $u'$ are adjacent.
Suppose that $u' = (x, y)$.
Let us consider the case in which $q = q'$.
Since $u'$ is adjacent to $\overline{v}'$, which is irregular,
there exist three cases with respect to the position of $u'$
(Fig.~\ref{fig:l2}):
(a)
both $y = q$ and $x \geq p'+1$,
(b)
both $y = q-3$ and $x \geq p+1$,
and
(c)
$y \geq q+1$, $x = p'$ and $p' \geq 2$.
We will discuss these three cases.
\noindent
{\bf\boldmath (a) $y = q$ and $x \geq p'+1$:}
Since the path $P$ does not contain $(p, q)$ by the definition of $P$,
$x \ne p$.
Thus,
$x \leq p-1$.
Let us discuss irregular vertices adjacent to $(x, q)$.
By the condition (Q2) of the regularity,
$(x, q-1) \notin R(D)$.
If $x \geq p'+2$,
then $(x-1, q) \in R(D)$
by the condition (Q1).
If $x = p'+1$
and
$q \geq 4$,
then $(x-1, q) \in R(D)$ by the condition (Q6).
If $x = p'+1$
and
$q \leq 3$,
then $q=2$ and $p'=0$
by (i) in Lemma~\ref{LMA:PQ},
which implies that
$(x, q) (= (1, 2))$ does not have an adjacent vertex on the left side.
Also,
if $x + 1 \leq m$,
$(x+1, q) \in R(D)$
by the condition (Q1).
Hence,
an irregular vertex which can be adjacent to $(x, q)$ is located at only $(x, q+1)$.
Then,
$(x, q+2) \in D$
holds.
The reason is as follows:
Assume that
$(x, q+2) \notin D$.
Then,
since $P$ does not contain $(p, q+1)$ by definition,
there exists $j \in [x+1, p-1]$ such that
$(j, q+2) \in D$, and
$(x+1, q+1), \ldots, (j, q+1) \in D$ or
there exists $j' \leq x-1$ such that
$(j', q+2) \in D$, and
$(j', q+1), \ldots, (x-1, q+1) \in D$.
However,
since $P$ is selected such that
the number of irregular vertices in $P$ is minimized,
either
$(j, q) = u'$
(i.e., $j = x$)
or
$(j', q) = u'$
(i.e., $j' = x$),
which contradicts the definitions of $j$ and $j'$.
Thus,
$(x, q+2) \in D$ and,
hence,
$(x, q+1) = \overline{v}'$
and
$\overline{v}' \in {\overline{R}}(D)$.
Then,
if $x = 1$,
then $\overline{v}'$ satisfies the definition (v) of a mobile,
and if $x = m$,
it satisfies the definition (iv).
Otherwise,
it satisfies the definition (iii).
\noindent
{\bf\boldmath (b) $y = q-3$ and $x \geq p+1$:}
Similarly to the case in which $y = q$,
we can show a mobile $\overline{v}' \in {\overline{R}}(D)$.
$(x, q-4) \notin R(D)$
by the condition (Q8).
Since $x \geq p+1$ by the condition of (b) and $p \geq p'+1$,
$x \geq p'+2$.
Hence,
$(x-1, q-3) \in R(D)$
by the condition (Q5).
Also,
if $x + 1 \leq m$,
$(x+1, q-3) \in R(D)$.
Hence,
an irregular vertex which can be adjacent to $(x, q-3)$ is located at only $(x, q-2)$.
Similarly to the proof of the case (a),
since $P$ is selected such that
the number of irregular vertices in $P$ is minimized,
$(x, q-1) \in D$.
Hence,
$(x, q-2) = \overline{v}'$
and
$\overline{v}' \in {\overline{R}}(D)$.
Then,
if $x = m$,
then $\overline{v}'$ satisfies the definition (iv) of a mobile.
Otherwise,
it satisfies the definition (iii).
\noindent
{\bf\boldmath (c) $x = p'$ and $y \geq q+1$:}
Similarly to the above cases,
we can show a mobile $\overline{v}' \in {\overline{R}}(D)$.
Specifically,
$(p'-1, y) \notin R(D)$ by the condition (Q7)
and
$(p', y+1), (p', y-1) \in R(D)$ by the condition (Q6).
Thus,
an irregular vertex which can be adjacent to $(p', y)$ is located at only $(p'+1, y)$.
Similarly to the proof of the case (a),
$(p'+2, y) \in D$.
Hence,
$(p'+1, y) = \overline{v}'$
and
$\overline{v}' \in {\overline{R}}(D)$.
Then,
if $y = m$,
$\overline{v}'$ satisfies the definition (ii) of a mobile
and otherwise,
it satisfies the definition (i).
Since we can prove the statement for the case in which $p = p'$ similarly to $q = q'$,
we omit the proof.
\fi
\end{proof}
\ifnum \count12 > 0
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig_l2.eps}
\end{center}
\caption{
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Figure for Lemma~\ref{LMA:L2}.
Suppose that $D$ is a $(p', q')$-$(p, q)$-regular MCDS.
The figures show that when $q'=q$,
white (gray, respectively) circles denote the positions of the vertex $u' = (x, y)$ ($\overline{v}'$, respectively).
Black circles denote regular vertices in $D$ and
vertices enclosed by dashed lines are irregular.
\fi
}
\label{fig:l2}
\end{figure*}
\fi
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
\fi
\begin{LMA}\label{LMA:MV}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Suppose that $D$ is a $(p', q')$-$(p, q)$-regular MCDS.
Then,
the following properties hold:
\begin{itemize}
\itemsep=-2.0pt
\setlength{\leftskip}{0pt}
\item[(i)]
Suppose that either $q'=q$ and $p=m$
or
$p'=p$ and $q=n$.
Also,
suppose that
$(x, y) \in R(D)$
and
$(x+2, y) \in {\overline{R}}(D)$.
Then,
an MCDS $D' = D \backslash \{ v \} \cup \{ (x+1, y) \}$ exists,
in which
$v$ is a mobile in $D$
(see Fig.~\ref{fig:lmv}).
\item[(ii)]
Suppose that either $q'=q$ and $p=m$
or
$p'=p$ and $q=n$.
Also,
suppose that
$(x, y) \in R(D)$
and
$(x, y+2) \in {\overline{R}}(D)$.
Then,
an MCDS $D' = D \backslash \{ v \} \cup \{ (x, y+1) \}$ exists,
in which
$v$ is a mobile in $D$.
\item[(iii)]
Suppose that $q'=q$ and $p \leq m-2$.
Also,
suppose that $P$ is a simple path consisting of $D$ between $(p-1, q)$ and $(p+2, q)$
such that
the number of vertices in ${\overline{R}}(D)$ in $P$ is minimum and
$P$ contains neither $(p, q)$ nor $(p, q+1)$.
Then,
an MCDS $D' = D \backslash \{ v \} \cup \{ (p+1, q) \}$ exists,
in which
$v$ is a mobile in $D$.
\item[(iv)]
Suppose that $p'=p$ and $q \leq n-2$.
Also,
suppose that $P$ is a simple path consisting of $D$ between $(p, q-1)$ and $(p, q+2)$
such that
the number of vertices in ${\overline{R}}(D)$ in $P$ is minimum and
$P$ contains neither $(p, q)$ nor $(p+1, q)$.
Then,
an MCDS $D' = D \backslash \{ v \} \cup \{ (p, q+1) \}$ exists,
in which
$v$ is a mobile in $D$.
\end{itemize}
\fi
\end{LMA}
\begin{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Suppose that $D$ is a $(p', q')$-$(p, q)$-regular MCDS.
\noindent
{\bf (i):}
First,
we consider the case of (i).
Suppose that
either
$q'=q$ and $p=m$ or
$p'=p$ and $q=n$.
Also,
suppose that $(x, y) \in R(D)$
and
$(x+2, y) \in {\overline{R}}(D)$.
By applying Lemma~\ref{LMA:L2} with $(x+2, y)$ and $(x, y)$ as $\overline{v}$ and $u$ in its statement, respectively,
there exists a mobile $\overline{v}' \in {\overline{R}}(D)$ in a path $P$,
in which
$P$ is a simple path of vertices in $D$ between $(x, y)$ and $(x+2, y)$
such that
the number of vertices in ${\overline{R}}(D)$ in $P$ is minimum.
Let us define
$D' = D \backslash \{ \overline{v}' \} \cup \{ (x+1, y) \}$.
Since $\overline{v}'$ is a mobile,
$D \backslash \{ \overline{v}' \}$ is dominating.
Then,
in what follows,
we will show that $D'$ is connected.
That is,
we will show that for any two vertices $a, b \in D'$,
there exists a path consisting of $D'$ between $a$ and $b$.
Let $D_{1}$ be a set of vertices $\tilde{v}$
such that there exists a path consisting of vertices in $D \backslash \{ \overline{v}' \}$ between $v$ and $(x, y)$.
Also,
let $D_{2}$ be the vertex set of vertices in $D$ except for those in $D_{1} \cup \{ \overline{v}' \}$.
That is,
$D_{2} = D \backslash (D_{1} \cup \{ \overline{v}' \})$.
There exists a path consisting of vertices in $D_{1}$ between $(x, y)$ and any vertex in $D_{1}$ by definition
(called Fact~(a)).
That is,
for any two vertices in $D_{1}$,
there exists a path of vertices in $D_{1}$ between them.
Also,
suppose that $\overline{v}' = (x', y')$ and
$\overline{v}'$ satisfies the condition (i) of a mobile,
that is,
$(x'+1, y'), (x'-1, y'), (x'-1, y'-1), (x'-1, y'+1) \in D$.
We will show that for any two vertices in $D_{2}$,
there exists a path of vertices in $D_{2}$ between them
by showing the number of subgraphs induced by $D \backslash \{ \overline{v}' \}$ is two.
If we can guarantee the existence of such a path,
then we have the following:
Since $D'$ contains $(x+1, y)$,
there exists a path of vertices in $D'$ between $(x, y)$ and $(x+2, y)$.
Hence,
for any $a \in D_{1}$ and any $b \in D_{2}$,
there exists a path of vertices in $D'$ between $a$ and $b$.
If both $(x', y'+1) \notin D$ and $(x', y'-1) \notin D$,
then the number of subgraphs induced by $D \backslash \{ \overline{v}' \}$ is two.
Then,
in the case in which $(x', y'+1) \in D$ ($(x', y'-1) \in D$, respectively),
we will show that $(x', y'+1)$ ($(x', y'-1)$, respectively) is contained in either $D_{1}$ or $D_{2}$
(called Property~(b)).
Let us consider the case of $(x', y'+1) \in D$.
$(x', y'+1)$ is adjacent to $(x'-1, y'+1)$,
that is,
the path $(x', y'+1)(x'-1, y'+1)(x'-1,y')$ consists of three vertices in $D \backslash \{ \overline{v}' \}$.
Thus,
if $(x'-1, y') \in D_{1}$, then $(x', y'+1) \in D_{1}$ and
if $(x'-1, y') \in D_{2}$, then $(x', y'+1) \in D_{2}$.
Similarly,
in the case of $(x', y'-1) \in D$,
if $(x'-1, y') \in D_{1}$, $(x', y'-1) \in D_{1}$,
and
if $(x'-1, y') \in D_{2}$, $(x', y'-1) \in D_{2}$.
Thus,
we have shown that Property~(b) is true,
which implies that
we have shown that $D'$ is an MCDS.
In the case in which $\overline{v}'$ satisfies the other conditions except for (i),
we can also prove that $D'$ is connected.
\noindent
{\bf (ii):}
We can prove (ii) similarly to the proof of (i).
Suppose that
either
$q'=q$ and $p=m$ or
$p'=p$ and $q=n$.
Also,
suppose that $(x, y) \in R(D)$
and
$(x, y+2) \in {\overline{R}}(D)$.
By applying Lemma~\ref{LMA:L2} with $(x, y+2)$ and $(x, y)$ as $\overline{v}$ and $u$, respectively,
there exists a mobile $\overline{v}' \in {\overline{R}}(D)$ in a path $P$,
in which
$P$ is a simple path of vertices in $D$ between $(x, y)$ and $(x, y+2)$
such that
the number of vertices in ${\overline{R}}(D)$ in $P$ is minimum.
Let us define
$D' = D \backslash \{ \overline{v}' \} \cup \{ (x, y+1) \}$.
Since $\overline{v}'$ is a mobile,
$D \backslash \{ \overline{v}' \}$ is dominating.
We omit the rest of the proof
because we can show that $D'$ is connected
similarly to the proof of (i).
\noindent
{\bf (iii):}
We can also prove (iii) similarly to the proof of (i).
Suppose that
$q'=q$ and $p \leq m-2$.
Also,
suppose that $P$ is a simple path between $(p-1, q)$ and $(p+2, q)$ which satisfies the conditions in statement (iii) of this lemma.
By the definition of the regularity,
$(p-1, q) \in R(D)$
and
$(p+2, q) \in {\overline{R}}(D)$.
By applying Lemma~\ref{LMA:L2} with $(p+2, q)$ and $(p-1, q)$ as $\overline{v}$ and $u$, respectively,
there exists a mobile $\overline{v}' \in {\overline{R}}(D)$ in the path $P$.
Let us define
$D' = D \backslash \{ \overline{v}' \} \cup \{ (p+1, q) \}$.
Since $\overline{v}'$ is a mobile,
$D \backslash \{ \overline{v}' \}$ is dominating.
We omit the rest of the proof
because we can show that $D'$ is connected
similarly to the proof of (i).
\noindent
{\bf (iv):}
We can also prove (iv) similarly to the proof of (i).
Suppose that
$p'=p$ and $q \leq n-2$.
Also,
suppose that $P$ is a simple path between $(p, q-1)$ and $(p, q+2)$ which satisfies the conditions in statement (iv).
By the definition of the regularity,
$(p, q-1) \in R(D)$
and
$(p, q+2) \in {\overline{R}}(D)$.
By applying Lemma~\ref{LMA:L2} with $(p, q+2)$ and $(p, q-1)$ as $\overline{v}$ and $u$, respectively,
there exists a mobile $\overline{v}' \in {\overline{R}}(D)$ in the path $P$.
Let us define
$D' = D \backslash \{ \overline{v}' \} \cup \{ (p, q+1) \}$.
Since $\overline{v}'$ is a mobile,
$D \backslash \{ \overline{v}' \}$ is dominating.
We omit the rest of the proof
because we can show that $D'$ is connected
similarly to the proof of (i).
\fi
\end{proof}
\ifnum \count12 > 0
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig_lmv.eps}
\end{center}
\caption{
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Suppose that $D$ is a $(p', q')$-$(p, q)$-regular MCDS.
Circles and crosses denote vertices in $D$ and not in $D$, respectively.
Lemma~\ref{LMA:L2} shows that for a path of vertices in $D$ between a white circle and a gray circle,
there exists a mobile in the path.
Note that by the regularity,
if $q' = q$, then $y = q'$,
and if $p' = p$, then $x = p'$.
\fi
}
\label{fig:lmv}
\end{figure*}
\fi
\begin{LMA}\label{LMA:LT}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Suppose that a vertex set $D$ is an MCDS which is either $(p, q)$-$(m, q)$-regular or $(p, q)$-$(p, n)$-regular.
Then,
the following properties are true:
\begin{itemize}
\itemsep=-2.0pt
\setlength{\leftskip}{0pt}
\item[(i)]
If for some $p \geq 2$ and some $q \geq 2$,
$(p+2, q+1) \in D$
and
$(p+3, q+1) \notin D$,
then
there exists an MCDS $D'$ into which $D$ is $(p+3, q)$-$(p+3, q+1)$-regularized
(see Fig.~\ref{fig:lt}).
\item[(ii)]
If for some $p \geq 2$ and some $q \geq 2$,
$(p+1, q+2) \in D$
and
$(p+1, q+3) \notin D$,
then
there exists an MCDS $D'$ into which $D$ is $(p, q+3)$-$(p+1, q+3)$-regularized.
\item[(iii)]
If $p = 0$,
$q = 2$,
$(1, 3) \in D$
and
$(2, 3) \notin D$,
then
there exists an MCDS $D'$ into which $D$ is $(2, 2)$-$(2, 3)$-regularized.
\end{itemize}
\fi
\end{LMA}
\begin{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
First,
we consider the cases (i) and (ii).
Suppose that a vertex set $D$ is an MCDS which is either $(p, q)$-$(m, q)$-regular or $(p, q)$-$(p, n)$-regular.
We will construct a CDS $\hat{D}$ into which $D$ is regularized.
Note that since $D$ is an MCDS,
$\hat{D}$ is also an MCDS.
If $D$ is $(p, q)$-$(m, q)$-regular,
then $(p, q), \ldots, (m, q) \in D$ by the conditions (Q1) and (Q6) of the regularity.
If $D$ is $(p, q)$-$(p, n)$-regular,
then $(p, q), \ldots, (m, q) \in D$ by the condition (P6).
These vertices are all regular and
are not removed from $D$ to construct $\hat{D}$.
Thus,
$\hat{D}$ satisfies the conditions (Q5) to be $(p, q+3)$-$(p+1, q+3)$-regular
and (P6) to be $(p+3, q)$-$(p+3, q+1)$-regular.
Similarly,
if $D$ is $(p, q)$-$(m, q)$-regular,
then $(p, q), \ldots, (p, n) \in D$ by the condition (Q6).
If $D$ is $(p, q)$-$(p, n)$-regular,
$(p, q), \ldots, (p, n) \in D$ by the conditions (P1) and (P6).
Thus,
$\hat{D}$ satisfies the conditions (Q6) to be $(p, q+3)$-$(p+1, q+3)$-regular
and (P5) to be $(p+3, q)$-$(p+3, q+1)$-regular.
Also,
since $(p+1, q), (p+2, q), (p,q+1), (p, q+2) \in D$,
these vertices dominate all the vertices which $(p+1, q+1)$ can dominate and
$(p+1, q+1) \notin D$.
Thus,
$\hat{D}$ satisfies the conditions (Q4) and (P4).
Since $(p+1, q-1), \ldots, (m, q-1) \notin D$ by the conditions (Q2) and (P7),
$\hat{D}$ satisfies the conditions (Q8) to be $(p, q+3)$-$(p+1, q+3)$-regular
and (P7) to be $(p+3, q)$-$(p+3, q+1)$-regular.
Since $(p-1, q+1), \ldots, (p-1, n) \notin D$ by the conditions (Q7) and (P2),
$\hat{D}$ satisfies the conditions (Q7) to be $(p, q+3)$-$(p+1, q+3)$-regular
and (P8) to be $(p+3, q)$-$(p+3, q+1)$-regular.
By the above argument,
$\hat{D}$ satisfies all the conditions except for (Q1) ((P8), respectively) to be $(p, q+3)$-$(p+1, q+3)$-regular ($(p+3, q)$-$(p+3, q+1)$-regular, respectively).
Then,
in order to satisfy these two conditions,
we will show that
if $(p+2, q+1) \in D$,
then we can obtain $\hat{D}$ such that $(p+3, q+1) \in \hat{D}$,
and
if $(p+1, q+2) \in D$,
we can obtain $\hat{D}$ such that $(p+1, q+3) \in \hat{D}$
by regularizing $D$.
Moreover,
we will show that
$\hat{D}$ is dominating and connected.
We can show that this lemma is true by showing them.
We first consider the case (i), that is,
the case in which
$(p+2, q+1) \in D$
and
$(p+3, q+1) \notin D$.
In this case,
one of the following two cases holds
(see Fig.~\ref{fig:lt}):
\begin{itemize}
\itemsep=-2.0pt
\setlength{\leftskip}{0pt}
\item[(a)]
For some $y \in [q+1, n]$,
$(p+2, q+1), \cdots, (p+2, y) \in D$
and
$(p+3, q+1), \cdots, (p+3, y) \notin D$.
Also,
If $y+1 \leq n$,
$(p+2, y+1), (p+3, y+1) \in D$.
\item[(b)]
For some $y' \in [q+1, n-1]$,
$(p+2, q+1), \cdots, (p+2, y') \in D$,
$(p+3, q+1), \cdots, (p+3, y') \notin D$
and
$(p+2, y'+1) \notin D$.
\end{itemize}
We will investigate each case.
\noindent
{\bf (a): }
For each $i = q+1, \ldots, y$,
we construct a vertex set $D'$ by removing $(p+2, i)$ from $D$ and adding $(p+3, i)$ to $D'$.
That is,
$D' = D \backslash
\{ (p+2, j) \mid j \in [q+1, y] \} \cup
\{ (p+3, j) \mid j \in [q+1, y] \}$.
We show that $D'$ is dominating.
The vertices dominated by $(p+2, i)$ are
$(p+1, i), (p+2, i), (p+3, i), (p+2, i-1)$ and $(p+2, i+1)$,
and
we confirm that these five vertices are also dominated by vertices in $D'$.
If $i \in [q+2, y-1]$,
then $(p+2, i), (p+3, i), (p+2, i-1)$ and $(p+2, i+1)$
are dominated by
the newly added vertices
$(p+3, i), (p+3, i), (p+3, i-1)$ and $(p+3, i+1)$, respectively.
If $i = q+1$,
then $(p+2, q+1), (p+3, q+1)$ and $(p+2, q+2)$
are dominated by
the new vertices $(p+3, q+1), (p+3, q+1)$ and $(p+3, q+2)$, respectively.
Since $D$ is
$(p, q)$-$(m, q)$-regular or
$(p, q)$-$(p, n)$-regular
by the supposition of this lemma,
$(p+2, q) \in D$
by the condition (Q1) to be $(p, q)$-$(m, q)$-regular and
the condition (P6) to be $(p, q)$-$(p, n)$-regular,
which results in $(p+2, q) \in D'$.
If $i = y$
and
$y+1 \leq n$,
then
$(p+2, y+1) \in D'$
because $(p+2, y+1) \in D$
by the condition of the case~(a).
If $i = y$ and
$y = n$,
$(p+2, n+1)$ does not exist.
Also,
since
$p \geq 2$ by the condition of (a)
and
$D$ is $(p, q)$-$(m, q)$-regular or
$(p, q)$-$(p, n)$-regular by the supposition of the lemma,
$(p, i) \in D$
by the conditions (Q6) to be $(p, q)$-$(m, q)$-regular and
(P1) to be $(p, q)$-$(p, n)$-regular.
Thus,
$(p, i) \in D'$
and
$(p, i)$ dominates $(p+1, i)$.
Next,
we show that $D'$ is connected.
$(p+2, q), (p+3, q), (p+2, q), (p+3, y+1) \in D'$
because $(p+2, q), (p+3, q), (p+2, q), (p+3, y+1) \in D$.
For any two vertices $a, b \in D$,
suppose that there exists a path $P$ between $a$ and $b$ consisting of vertices in $D$
such that $P$ contains some vertex $(p+2, i') \hspace{1mm} (i' \in [q+1, y])$.
For any $i'' \in [q+1, y+1]$,
$D$ does not contain $(p+1, i'')$
because of the minimality of $D$.
That is,
$P$ necessarily contains either $(p+2, y+1)$ or $(p+3, y+1)$.
Hence,
there exists a path of vertices in $D'$ between $a$ and $b$,
which implies that
$D'$ is connected.
Therefore,
$(p+3, q+1) \in D'$
and
$D'$ is a $(p+3, q)$-$(p+3, q+1)$-regular MCDS.
\noindent
{\bf (b): }
For each $i = q+1, \ldots, y'$,
we construct a vertex set $D'$ by removing $(p+2, i)$ from $D$ and adding $(p+3, i+1)$.
That is,
$D' = D \backslash
\{ (p+2, j) \mid j \in [q+1, y'] \} \cup
\{ (p+3, j+1) \mid j \in [q+1, y'] \}$.
In what follows,
we will show that $D'$ is an MCDS.
After that,
by applying (ii) in Lemma~\ref{LMA:MV} with
$(p+3, q) \in R(D')$ and $(p+3, q+2) \in {\overline{R}}(D')$
as $(x, y)$ and $(x, y+2)$, respectively,
there exists an MCDS $D''$ such that $D'' = D' \backslash \{ v' \} \cup \{ (p+3, q+1) \}$,
in which $v' \in {\overline{R}}(D')$ is a mobile.
Since $(p+3, q+1) \in D''$,
$D''$ is $(p+3, q)$-$(p+3, q+1)$-regular.
First,
we show that $D'$ is dominating.
For each $i = q+1, \ldots, y'$,
the vertices dominated by $(p+2, i)$ are
$(p+1, i), (p+2, i), (p+3, i), (p+2, i-1)$ and $(p+2, i+1)$.
If $i \ne q+1$,
then $(p+2, i), (p+3, i), (p+2, i-1)$ and $(p+2, i+1)$ are
dominated by
$(p+3, i), (p+3, i), (p+3, i-1)$ and $(p+3, i+1)$, respectively.
Also,
if $i = q+1$,
then
$(p+2, q+2)$ is dominated by
$(p+3, q+2)$.
Since $D$ is
$(p, q)$-$(m, q)$-regular or
$(p, q)$-$(p, n)$-regular
by the supposition of this lemma,
$(p+2, q), (p+3, q) \in D$ by the conditions (Q1) and (P6).
Hence,
$(p+2, q), (p+3, q) \in D'$,
which dominate $(p+2, q+1), (p+3, q+1)$ and $(p+2, q)$.
Next,
we show that $D'$ is connected.
If $y' \leq n-2$,
then either $(p+4, y'+1) \in D$
or
$(p+3, y'+2) \in D$
because there exists a vertex to dominate $(p+3, y'+1)$ and
$D$ is connected.
If $y' = n-1$,
then $(p+4, y'+1) \in D$ for the same reason.
Hence,
either $(p+4, y'+1) \in D'$
or $(p+3, y'+2) \in D'$.
We now define
$A = \{ (p+3, j+1) \mid j \in [q+1, y'] \}$
and
$B = D' \backslash A$.
That is,
$B = D \backslash \{ (p+2, j) \mid j \in [q+1, y'] \}$.
In what follows,
we show that for any two vertices $a, b \in D'$,
there exists a path consisting of vertices in $D'$ between $a$ and $b$.
If $a, b \in B$,
then such a path exists because $D$ is connected.
If $a, b \in A$,
such a path exists by the definition of $A$.
If $a \in A$ and $b \in B$,
then there exists a path of vertices in $D'$
between $a$ and either $(p+4, y'+1)$ or $(p+3, y'+2)$.
Since the vertices $(p+4, y'+1)$ and $(p+3, y'+2)$ belong to $B$,
there exists a path of vertices in $D'$ between $b$ and either $(p+4, y'+1)$ or $(p+3, y'+2)$.
Hence,
there also exists such a path between $a$ and $b$ in this case.
By the above argument,
there exists such a path of vertices in $D'$ between $a$ and $b$.
That is, $D'$ is connected.
Therefore,
$D'$ is an MCDS and
we have shown that $D''$ is a $(p+3, q)$-$(p+3, q+1)$-regular MCDS by the initial argument.
The case of (ii),
that is,
the case in which
$(p+1, q+2) \in D$
and
$(p+1, q+3) \notin D$ is symmetric to the case (i) and thus,
we omit the proof of this case.
Since the case~(iii) is a special case of the case~(i),
we can prove this case in the same way as that of (i) and omit the proof.
\fi
\end{proof}
\ifnum \count12 > 0
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig_lt.eps}
\end{center}
\caption{
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Figure for Lemma~\ref{LMA:LT}.
Suppose that $D$ is an MCDS which is either $(p, q)$-$(m, q)$-regular or $(p, q)$-$(p, n)$-regular.
Circles and crosses denote vertices in $D$ and not in $D$, respectively.
\fi
}
\label{fig:lt}
\end{figure*}
\fi
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
\fi
\begin{LMA}\label{LMA:L4}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Suppose that a vertex set $D$ is a $(p', q')$-$(p, q)$-regular MCDS.
Then,
the following properties hold:
\begin{itemize}
\itemsep=-2.0pt
\setlength{\leftskip}{0pt}
\item[(i)]
If $q' = q$
and
$p \leq m-1$,
there exists an MCDS $D'$ into which $D$ is $(p', q')$-$(p+1, q)$-regularized.
\item[(ii)]
If $p' = p$
and
$q \leq n-1$,
there exists an MCDS $D'$ into which $D$ is $(p', q')$-$(p, q+1)$-regularized.
\end{itemize}
\fi
\end{LMA}
\begin{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Suppose that a vertex set $D$ is a $(p', q')$-$(p, q)$-regular MCDS.
First, we consider the case (i) in which
$q' = q$
and
$q \leq n-2$.
We will construct a vertex set $\hat{D}$ by regularizing $D$.
Note that $\hat{D}$ is also an MCDS
because $D$ is an MCDS.
When $(p', q')$-$(p+1, q)$-regularizing $D$,
the conditions (Q5),(Q6),(Q7) and (Q8) of the regularity of $D$ are not changed.
Thus,
$\hat{D}$ also satisfies the conditions (Q5),(Q6),(Q7) and (Q8) to be $(p', q')$-$(p+1, q)$-regular.
Also,
$(p'+1, q), \ldots, (p, q) \in D$ ($(p'+1, q-1), \ldots, (p-1, q-1) \notin D$, $(p'+1, q-2), \ldots, (p, q-2) \notin D$, respectively)
by the condition (Q1) ((Q2), (Q4), respectively) of the regularity of $D$.
Thus,
$(p'+1, q), \ldots, (p, q) \in \hat{D}$,
$(p'+1, q-1), \ldots, (p-1, q-1) \notin \hat{D}$
and
$(p'+1, q-3), \ldots, (p, q-1) \notin \hat{D}$.
In what follows,
we will construct $\hat{D}$ by regularizing $D$
such that
$(p+1, q) \in \hat{D}$,
$(p, q-1) \notin \hat{D}$ and
$(p+1, q-2) \notin \hat{D}$
in order that
$\hat{D}$ satisfies the conditions (Q1),(Q2) and (Q4).
Also,
we will show that
if $p + 1 = m$,
then $(m, q-1) \notin \hat{D}$
to satisfy the condition (Q3)
and show that
if $q = n-1$,
then $(p, q) \notin \hat{D}$
to satisfy the condition (Q9).
Moreover,
we will show that
$\hat{D}$ is dominating and connected.
We can prove (i) in the statement of this lemma by showing them.
We first consider the case in which $(p+1, q) \in D$
(see Case~1 in Fig.~\ref{fig:routine}).
If $q' = q \leq 2$,
then $q' = q = 2$ by (i) in Lemma~\ref{LMA:PQ}.
Since the condition (Q1) is satisfied,
if $p \geq 2$,
then $(p-1, 2), (p, 2), (p+1, 2) \in D$,
and
if $p = 1$,
then $(1, 2), (2, 2) \in D$.
Thus,
all the vertices which $(p, 1)$ can dominate are dominated by other vertices in $D$.
$D$ does not contain $(p, q-1) (= (p, 1))$
because $D$ is minimum.
If $q' = q > 2$,
then $q' \geq 5$ by (ii) in Lemma~\ref{LMA:PQ}.
Then,
if $p' \geq 2$ and $p \geq p'+2$,
then $(p-1, q), (p, q), (p+1, q) \in D$
by the condition (Q1).
If $p' \geq 2$ and $p = p'+2$,
then $(p-1, q), (p, q), (p+1, q) \in D$ as well
by the conditions (Q1) and (Q6).
It follows from the condition (Q1) that
if $p' \leq 1$,
then $(p-1, q), (p, q), (p+1, q) \in D$
and
if $p = 1$,
then $(p, q)(=(1, q)), (p+1, q)(=(2, q)) \in D$.
Moreover,
if $q' \geq 5$,
$(p, q-3) \in D$ by the condition (Q5).
Thus,
all the vertices which $(p, q-1)$ can dominate are dominated by other vertices in $D$.
$(p, q-1) \notin D$
because $D$ is minimum.
By the above argument,
$(p, q-1) \notin \hat{D}$,
which implies that
$\hat{D}$ satisfies the condition (Q2).
Similarly,
$(p, q-3),(p+1, q-3),(p+2, q-3) \in D$ by the condition (Q5)
and
$(p+1, q) \in D$ by the above supposition.
Thus,
$(p, q-2) \notin D$
because $D$ is minimum.
$\hat{D}$ satisfies the condition (Q4).
Therefore,
if $(p+1, q) \in D$,
$D$ satisfies the conditions (Q1),(Q2) and (Q4) to be $(p', q')$-$(p+1, q)$-regular.
If $p = m-1$,
then $(m, q-1) \notin D$
because $D$ satisfies the condition (Q1) and is minimum.
That is,
$D$ satisfies the condition (Q3).
If $q = n-1$,
then $(p'+1, n), \cdots, (p-1, n) \notin D$
because $D$ satisfies the condition (Q1) and is minimum.
That is,
$D$ satisfies the condition (Q9).
Therefore,
regarding $D$ as $\hat{D}$ has completed the $(p', q')$-$(p+1, q)$-regularization.
Next, we consider the case in which
$(p+1, q) \notin D$.
If $p + 2 \leq m$ and $(p+2, q) \in D$,
let $P$ be a simple path consisting of $D$ between
$(p+2, q)$ and $(p-1, q)$
such that
the number of vertices in ${\overline{R}}(D)$ in $P$ is minimum.
We will discuss the following three cases:
\begin{itemize}
\itemsep=-2.0pt
\setlength{\leftskip}{0pt}
\item[(a)]
$p+2 \leq m$,
$(p+2, q) \in D$ and
$P$ contains neither $(p, q)$ nor $(p, q+1)$,
\item[(b)]
$p+2 \leq m$, $(p+2, q) \in D$
and
$P$ contains at least of $(p, q)$ and $(p, q+1)$, and
\item[(c)]
either $p+1 = m$ or
both $p + 2 \leq m$ and $(p+2, q) \notin D$.
\end{itemize}
We will show that for each of (a),(b) and (c),
$D$ can be regularized into some MCDS which contains $(p+1, q)$.
Then,
we can show that the MCDS is $(p', q')$-$(p+1, q)$-regular similarly to the proof of the above discussed case,
that is, the case of $(p+1, q) \in D$.
\noindent
{\bf (a):}
By applying Lemma~\ref{LMA:MV}~(iii)
with $(p+2, q) \in {\overline{R}}(D)$ and $(p-1, q) \in R(D)$
as $(p-1, q)$ and $(p+2, q)$, respectively,
MCDS $D' = D \backslash \{ v \} \cup \{ (p+1, q) \}$, in which
$v \in {\overline{R}}(D)$ is a mobile.
Thus,
$D'$ satisfies the statement of this lemma.
\noindent
{\bf (b):}
In this case,
a vertex which can dominate $(p+1, q-1)$ is as follows:
if $q \geq 3$,
then $(p+1, q-1)$, $(p+1, q-2)$, $(p, q-1)$ or $(p+2, q-1)$,
and
if $q = 2$,
then $(p+1, q-1)$, $(p, q-1)$ or $(p+2, q-1)$.
First,
we consider the case in which $(p+1, q-2) \in D$ or $(p, q-1) \in D$
(Fig.~\ref{fig:l4}(2-1),(2-2)).
By the conditions (Q1) and (Q6),
all the vertices except for $(p+1, q-1)$ which $(p+1, q-2)$ or $(p, q-1)$ can dominate are dominated by regular vertices in $D$.
Hence,
$D$ contains exactly one of $(p+1, q-2)$ and $(p, q-1)$
because $D$ is minimum.
Let $v$ be the vertex contained in $D$
and let us construct $D'$ by removing $v$ from $D$ and adding $(p+1, q)$ to $D$ instead.
Then,
$D'$ is still dominating.
By the conditions (Q1) and (Q6),
for any two vertices except for $v$ in $D$,
no simple path consisting of vertices in $D$ between the two vertices contains $v$,
which implies that
$D'$ is connected.
Thus,
$D'$ is an MCDS.
Second,
let us consider the case in which
$(p+1, q-2) \notin D$ and $(p, q-1) \notin D$
(Fig.~\ref{fig:l4}(4)).
That is,
$(p+1, q-1) \in D$ or $(p+2, q-1) \in D$.
However,
$(p+2, q-1) \in D$
because $D$ is minimum.
By the condition of (b),
$P$ contains at least one of $(p, q)$ and $(p, q+1)$.
Also,
since $(p, q-1) \notin D$,
$P$ contains $(p, q+1)$ if $P$ contains $(p, q)$.
It implies that $(p, q+1) \in D$.
Let us consider the case in which $P$ contains $(p, q+2)$.
That is,
$(p, q+2) \in D$
(Fig.~\ref{fig:l4}(6)).
Then, let us define
$D' = D \backslash \{ (p, q+1) \} \cup \{ (p+1, q) \}$
(Fig.~\ref{fig:l4}(7)).
By the condition (Q1) and the fact that $(p, q+2) \in D$,
a vertex in $D$ which only $(p, q+1)$ can dominate is $(p+1, q+1)$.
$(p+1, q+1)$ dominates the vertex in $D'$.
Thus,
$D'$ is dominating.
We next show that $D'$ is connected.
Let us define $C = D \backslash \{ (p, q+1) \}$,
and define the vertex set $C_{1}$ of vertices each of which
there exists a path consisting of vertices in $C$ from $(p+2, q)$ to.
Also,
we define $C_{2} = C \backslash C_{1}$.
We now show that $D'$ is connected by showing that
for any $v \in C_{1}$ and any $u \in C_{2}$,
there exists a path of vertices in $D'$ between $v$ and $u$.
If $(p, q) \in C_{1}$,
there exists a path of vertices in $C$ different from $P$ between $(p+2, q)$ and $(p, q)$ and all the vertices in $C$ belong to $C_{1}$.
Hence,
$D'$ is connected.
If $(p, q) \in C_{2}$,
there exists a path $P'$ of vertices in $C_{2}$ between $(p, q)$ and $u$.
Then,
let us consider the path comprising the following three paths:
a path of vertices in $C_{1}$ between $v$ and $(p+2, q)$,
the path $(p+2, q)(p+1, q)(p, q)$ and $P'$.
This path consists of vertices in $D'$,
which implies that
$D'$ is connected.
Next we consider the case in which
$q \leq n-2$ and $P$ does not contain $(p, q+2)$.
We assume that $P$ does not contain $(p+1, q+1)$.
Since $P$ contains $(p, q+1)$,
$P$ also contains $(p-1, q+1)$ and
there exists some $x \leq p-1$
such that $P$ contains $(x, q+1)$ and $(x, q+2)$,
which contradicts the definition that the number of vertices in ${\overline{R}}(D)$ in $P$ is minimum,
and thus $P$ contains $(p+1, q+1)$
(Fig.~\ref{fig:l4}(8)).
In the case of $q = n-1$,
we assume again that $P$ does not contain $(p+1, q+1) (=(p+1, n))$.
Since $(p'+1, n), \cdots, (p-1, n) \notin D$ by the condition (Q9),
$P$ does not contain $(p, q+1) (=(p, n))$,
which contradicts the fact discussed above.
Hence,
if $P$ contains $(p, q+1)$ but does not contain $(p, q+2)$,
$P$ contains $(p+1, q+1)$.
Then,
we consider the case in which
$(p+1, q-2) \notin D$,
$(p, q-1) \notin D$,
$P$ does not contain $(p, q+2)$
and
$(p+2, q+1) \in D$
(Fig.~\ref{fig:l4}(9)).
If $q + 2 \leq n$ and $p + 3 \leq m$,
then we define $D' = D \backslash \{ (p, q+1), (p+2, q+1), (p+2, q-1) \} \cup \{ (p+1, q), (p+1, q+2), (p+3, q) \}$
(Fig.~\ref{fig:l4}(10-1)).
If $q + 2 \leq n$ and $p + 3 > m$ (i.e., $p + 2 = m$),
then we define
$D' = D \backslash \{ (p, q+1), (p+2, q+1) \} \cup \{ (p+1, q), (p+1, q+2) \}$
(Fig.~\ref{fig:l4}(10-2)).
If $q + 2 > n$ (i.e., $q + 1 = n$),
then we define
$D' = D \backslash \{ (p, q+1) \} \cup \{ (p+1, q) \}$
(Fig.~\ref{fig:l4}(10-3)).
For each $D'$ in these cases,
we can show that $D'$ is dominating and connected
similarly to the proof of the case in which $P$ contains $(p, q+2)$.
Moreover,
if $q = n-1$,
then $(p-1, n) \notin D'$ because $D'$ is minimum,
which satisfies the condition (Q9).
Thus,
$D'$ is $(p', q')$-$(p+1, q)$-regular.
Finally,
we consider the case in which
$(p+1, q-2) \notin D$,
$(p, q-1) \notin D$,
$P$ does not contain $(p, q+2)$ and
$(p+2, q+1) \notin D$
(Fig.~\ref{fig:l4}(11)).
Then,
$P$ contains $(p+1, q+2)$,
that is,
$(p+1, q+2) \in D$.
Let us define $D' = D \backslash \{ (p, q+1) \} \cup \{ (p+1, q) \}$
and we can show that $D'$ is dominating and connected
similarly to the proof of the above case.
Thus,
$D'$ is $(p', q')$-$(p+1, q)$-regular.
\noindent
{\bf (c):}
If either $(p+1, q-2) \in D$ or $(p, q-1) \in D$,
we can prove the statement in the same way as the proof of the case (b)
(Fig.~\ref{fig:l4}(15)).
We consider the case in which $(p+1, q-2) \notin D$
and
$(p, q-1) \notin D$.
Then,
$(p+2, q-1) \in D$ in the same way as the case (b)
(Fig.~\ref{fig:l4}(16)).
We consider whether $(p+2, q-2) \in D$ or not
(Fig.~\ref{fig:l4}(17)).
If $p+2 = m$,
then $(p+3, q-1) \in D$ is impossible.
Thus,
note that $(p+2, q-2) \in D$
because
$(p+2, q-2) \in D$ mush hold
so that $(p+2, q-1)$ is connected to any other vertex in $D$.
We consider the case of $(p+2, q-2) \in D$
(Fig.~\ref{fig:l4}(18)).
If $p+3 \geq m$,
then we define $D' = D \backslash \{ (p+2, q-1), (p+2, q-2) \} \cup \{ (p+1, q), (p+3, q-2) \}$
(Fig.~\ref{fig:l4}(19-1)).
If $p+2 = m$,
then we define $D' = D \backslash \{ (p+2, q-1) \} \cup \{ (p+1, q) \}$
(Fig.~\ref{fig:l4}(19-2)).
For each $D'$ in the both cases,
we can show that $D'$ is dominating and connected
similarly to the proof of the above case.
Thus,
$D'$ is $(p', q')$-$(p+1, q)$-regular.
We consider the case of $(p+2, q-2) \notin D$
(Fig.~\ref{fig:l4}(20)).
$(p+3, q-1) \in D$ holds
so that $(p+2, q-1)$ with respect to $D$ is connected to any other vertex.
We define $D' = D \backslash \{ (p+2, q-1) \} \cup \{ (p+1, q) \}$
(Fig.~\ref{fig:l4}(21)).
We can show that $D'$ in each of these cases is dominating and connected
similarly to the proof of the above case.
Thus,
$D'$ is $(p', q')$-$(p+1, q)$-regular.
We omit the proof of (ii)
because it can be proved similarly to the proof of the case (i).
\fi
\end{proof}
\ifnum \count12 > 0
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig_l4.eps}
\end{center}
\caption{
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Figure for Lemma~\ref{LMA:L4}.
Suppose that $D$ is a $(p', q')$-$(p, q)$-regular MCDS.
Circles and crosses denote vertices in $D$ and not in $D$, respectively.
White circles denote vertices in $D$ which are handled at the regularization.
White crosses denote vertices not in $D$
or vertices removed from $D$ at the regularization.
\fi
}
\label{fig:l4}
\end{figure*}
\fi
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
\fi
\begin{LMA}\label{LMA:L6}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Suppose that a vertex set $D$ is a $(0, 2)$-$(m, 2)$-regular MCDS.
Then,
there exists an MCDS which $D$ is $(2, 2)$-$(2, 3)$-regularized into
(Fig.~\ref{fig:l6}(2)).
\fi
\end{LMA}
\begin{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Suppose that a vertex set $D$ is a $(0, 2)$-$(m, 2)$-regular MCDS.
We will construct a vertex set $\hat{D}$ by regularizing $D$.
Note that $\hat{D}$ is also an MCDS
because $D$ is an MCDS.
$(1, 2), \ldots, (m, 2) \in D$ by the condition (Q1) for $D$ to be $(0, 2)$-$(m, 2)$-regular,
which are regular vertices and are not removed from $D$ to construct $\hat{D}$.
Hence,
$(1, 2), \ldots, (m, 2) \in \hat{D}$
and
$\hat{D}$ satisfies the condition (P6) to be $(2, 2)$-$(2, 3)$-regular.
Similarly,
$(1, 1), \cdots, (m-1, 1) \notin D$ and $(m, 1) \notin D$
by the conditions (Q2) and (Q3) of $D$.
Hence,
$(1, 1), \ldots, (m, 1) \notin \hat{D}$
and
$\hat{D}$ satisfies the condition (P2).
That is,
$\hat{D}$ into which $D$ is regularized into satisfies all the conditions except for (P1) to be $(2, 2)$-$(2, 3)$-regular.
Now we will show that we can obtain $\hat{D}$ such that $(2, 3) \in \hat{D}$ to satisfy the condition (P1) by the regularization.
Also,
we will show that $\hat{D}$ is dominating and connected,
which leads to the statement is true.
If $(2, 3) \in D$ (Fig.~\ref{fig:l6}(2)),
then $D$ is $(2, 2)$-$(2, 3)$-regular and
$D$ can be regarded as the MCDS $\hat{D}$.
In what follows,
we consider the case in which $(2, 3) \notin D$
(Fig.~\ref{fig:l6}(3)).
If $(1, 3) \in D$ (Fig.~\ref{fig:l6}(4)),
then we can have a $(2, 2)$-$(2, 3)$-regular MCDS from $D$ using Lemma~\ref{LMA:LT}(iii).
We next consider the case in which
$(1, 3) \notin D$
and
$(2, 4) \in D$
(Fig.~\ref{fig:l6}(6)).
Since $(2, 2) \in R(D)$ and $(2, 4) \in {\overline{R}}(D)$,
there exists an MCDS $D' = D \backslash \{ v' \} \cup \{ (2, 3) \}$,
in which $v' \in {\overline{R}}(D)$ is a mobile,
by applying Lemma~\ref{LMA:MV}(ii)
with $(2, 2)$ and $(2, 4)$ as $(x, y)$ and $(x, y+2)$, respectively.
$(2, 3) \in D'$
and thus,
$D'$ is $(2, 2)$-$(2, 3)$-regular.
We next consider the case in which
$(1, 3) \notin D$
and
$(2, 4) \notin D$
(Fig.~\ref{fig:l6}(7)).
$D$ contains either $(1, 4)$ or $(1, 5)$ to dominate $(1, 4)$, and
$(1, 5) \in D$ holds because $D$ is connected.
Then,
since $D$ is $(0, 5)$-$(1, 5)$-regular,
$D'$ which is $(0, 5)$-$(m, 5)$-regular can be constructed from $D$
by applying Lemma~\ref{LMA:L4} $m-1$ times
(Fig.~\ref{fig:l6}(8)).
Then,
$(1, 4), \ldots, (m, 4) \notin D$ by the conditions (Q2) and (Q3), which contradicts that the constructed CDS is connected.
Therefore,
both
$(1, 3) \notin D$
and
$(2, 4) \notin D$
do not hold.
We have completed the proof.
\fi
\end{proof}
\ifnum \count12 > 0
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig_l6.eps}
\end{center}
\caption{
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Figure for Lemma~\ref{LMA:L6}.
Suppose that $D$ is a $(0, 2)$-$(m, 2)$-regular MCDS.
Circles and crosses denote vertices in $D$ and not in $D$, respectively.
\fi
}
\label{fig:l6}
\end{figure*}
\fi
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
\fi
\begin{LMA}\label{LMA:L9}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
For $p \in [2, m-3]$ and $q \in [2, n-3]$,
suppose that a vertex set $D$ is an MCDS which is $(p, q)$-$(m, q)$-regular or $(p, q)$-$(p, n)$-regular.
Then,
there exists an MCDS $D'$ into which $D$ is $(p, q)$-$(\hat{p}, \hat{q})$-regularized
such that
either $(\hat{p}, \hat{q}) = (p+3, q+1)$
or
$(\hat{p}, \hat{q}) = (p+1, q+3)$.
\fi
\end{LMA}
\begin{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
The proof of this lemma is similar to that of Lemma~\ref{LMA:L6}.
First, for $p \in [2, m-3]$ and $q \in [2, n-3]$,
suppose that a vertex set $D$ is an $(p, q)$-$(m, q)$-regular MCDS.
We will construct a vertex set $\hat{D}$ by regularizing $D$.
Note that $\hat{D}$ is also an MCDS
because $D$ is an MCDS.
By the conditions (Q1) and (Q6) for $D$ to be $(p, q)$-$(m, q)$-regular,
$(p+1, q), \ldots, (m, q) \in D$
and
$(p, q), \ldots, (p, n) \in D$, respectively.
Since these vertices are regular,
they are not removed from $D$ when $D$ is regularized into $\hat{D}$.
Thus,
$(p+1, q), \ldots, (m, q) \in \hat{D}$
and
$(p, q), \ldots, (p, n) \in \hat{D}$,
which implies that
$\hat{D}$ satisfies the conditions (Q5), (P5), (Q6) and (P6) to be $(p, q)$-$(\hat{p}, \hat{q})$-regular
such that
$(\hat{p}, \hat{q}) = (p+3, q+1)$
or
$(\hat{p}, \hat{q}) = (p+1, q+3)$.
Similarly,
by the conditions (Q7) and (Q2),
$(p-1, q+1), \ldots, (p-1, n) \notin \hat{D}$
and
$(p+1, q-1), \ldots, (m, q-1) \notin \hat{D}$, respectively.
Thus,
$\hat{D}$ satisfies the conditions (Q7), (P7), (Q8) and (P8) to be $(p, q)$-$(\hat{p}, \hat{q})$-regular.
Since $D$ is minimum,
$(p+1, q+1) \notin D$, that is, $(p+1, q+1) \notin \hat{D}$.
Hence,
$\hat{D}$ satisfies the conditions (Q4) and (P4) to be $(p, q)$-$(\hat{p}, \hat{q})$-regular.
By the above argument,
$\hat{D}$ into which $D$ is regularized into satisfies all the conditions except for (Q1) and (P1).
Now we will show that
we can obtain $\hat{D}$ such that $(p+3, q+1) \in \hat{D}$ ($(p+1, q+3) \in \hat{D}$, respectively) to satisfy the condition (Q1) ((P1), respectively) by regularizing $D$.
If $(p+3, q+1) \in D$ or $(p+1, q+3) \in D$,
then $D$ can be regarded as the MCDS $\hat{D}$
(see Fig.~\ref{fig:l9}(2-1),(2-2)).
In what follows,
we consider the case in which
$(p+3, q+1) \notin D$
and
$(p+1, q+3) \notin D$
(Fig.~\ref{fig:l9}(3-1),(3-2)).
If $(p+2, q+1) \in D$,
then $D$ can be $(p, q)$-$(p+3, q+1)$-regularized into an MCDS $D'$
by applying Lemma~\ref{LMA:LT}(i) with $p$ and $q$ as $p$ and $q$ in its statement, respectively
(Fig.~\ref{fig:l9}(4-1)).
Also,
if $(p+1, q+2) \in D$,
$D$ can be $(p, q)$-$(p+1, q+3)$-regularized into an MCDS $D'$ by applying Lemma~\ref{LMA:LT}(ii) with $p$ and $q$ as $p$ and $q$, respectively
(Fig.~\ref{fig:l9}(4-2)).
Hence,
$D'$ can be regarded as $\hat{D}$.
Second,
we discuss the case in which
$(p+2, q+1) \notin D$
and
$(p+1, q+2) \notin D$.
Since $D$ contains vertices to dominate $(p+2, q+2)$ and is connected,
either $(p+3, q+2) \in D$ or $(p+2, q+3) \in D$
(Fig.~\ref{fig:l9}(5-1),(5-2)).
If $(p+3, q+2) \in D$,
then
there exists an MCDS $D' = D' \backslash \{ v' \} \cup \{ (p+3, q+1) \}$,
in which $v' \in {\overline{R}}(D)$ is a mobile,
by applying Lemma~\ref{LMA:MV} (ii) with $(p+3, q) \in R(D)$ and $(p+3, q+2) \in {\overline{R}}(D)$ as $(x, y)$ and $(x, y+2)$, respectively.
Since $(p+3, q+1) \in D'$, $D'$ is $(p+3, q)$-$(p+3, q+1)$-regular,
which implies that the statement is true.
Similarly,
if $(p+2, q+3) \in D$,
there exists an MCDS $D' = D' \backslash \{ v' \} \cup \{ (p+1, q+3) \}$,
in which $v' \in {\overline{R}}(D)$ is a mobile,
by applying Lemma~\ref{LMA:MV} (i) with $(p, q+3) \in R(D)$ and $(p+2, q+3) \in {\overline{R}}(D)$ as $(x, y)$ and $(x+2, y)$, respectively.
Since $(p+1, q+3) \in D'$, $D'$ is $(p, q+3)$-$(p+1, q+3)$-regular,
which implies that the statement is true.
We omit the proof of the case in which $D$ is $(p, q)$-$(p, n)$-regular
because this case is symmetric to the case discussed above.
\fi
\end{proof}
\ifnum \count12 > 0
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig_l9.eps}
\end{center}
\caption{
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
The top figure illustrates the situation of the statement of Lemma~\ref{LMA:L9}.
Suppose that $D$ is an MCDS which is $(p, q)$-$(m, q)$-regular or $(p, q)$-$(p, n)$-regular.
Circles and crosses denote vertices in $D$ and not in $D$, respectively.
White circles denote vertices in $D$ which are handled at the regularization.
White crosses denote vertices not in $D$
or vertices removed from $D$ at the regularization.
\fi
}
\label{fig:l9}
\end{figure*}
\fi
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
\fi
\begin{LMA}\label{LMA:C9}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
For $p \geq 2$ and $q \geq 2$,
suppose that a vertex set $D$ is an MCDS which is $(p, q)$-$(m, q)$-regular or $(p, q)$-$(p, n)$-regular.
Then,
the following properties hold:
\begin{itemize}
\itemsep=-2.0pt
\setlength{\leftskip}{0pt}
\item[(i)]
if $p \leq m-4$
and
$q \leq m-4$,
then there exists an MCDS $D'$
into which $D$ is $(p+3, q)$-$(p+3, q+1)$-regularized or $(p, q+3)$-$(p+1, q+3)$-regularized,
\item[(ii)]
if $p \leq m-4$ and $q \in \{ n-3, n-2 \}$,
then there exists an MCDS $D'$ into which $D$ is $(p+3, n-2)$-$(p+3, n-1)$-regularized,
and
\item[(iii)]
if $q \leq n-4$
and
$p \in \{ m-3, m-2 \}$,
then there exists an MCDS $D'$ into which $D$ is $(m-2, q+3)$-$(m-1, q+3)$-regularized.
\end{itemize}
\fi
\end{LMA}
\begin{proof}
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
For $p \geq 2$ and $q \geq 2$,
suppose that a vertex set $D$ is an MCDS which is $(p, q)$-$(m, q)$-regular or $(p, q)$-$(p, n)$-regular.
\noindent
{\bf (i):}
If $p \leq m-4$
and
$q \leq n-4$,
then there exists an MCDS $D'$ into which
$D$ is $(p+3, q)$-$(p+3, q+1)$-regularized or $(p, q+3)$-$(p+1, q+3)$-regularized
by Lemma~\ref{LMA:L9},
which satisfies the statement of this lemma.
\noindent
{\bf (ii):}
If $q = n-3$,
then there exists an MCDS $D'$ into which
$D$ is either $(p+3, q)$-$(p+3, q+1)$-regularized,
that is,
$(p+3, n-3)$-$(p+3, n-2)$-regularized or
$(p, q+3)$-$(p+1, q+3)$-regularized,
that is,
$(p, n)$-$(p+1, n)$-regularized.
Hence,
if $D'$ is $(p+3, n-3)$-$(p+3, n-2)$-regular, then the statement is true
(Fig.~\ref{fig:c9}(2)).
Then,
we discuss the case in which $D'$ is $(p, n)$-$(p+1, n)$-regular
(Fig.~\ref{fig:c9}(3)),
that is,
the case in which $(p+1, n) \in D'$.
In this case,
$D'$ satisfies all the conditions except for (Q1) similarly to the proof of Lemma~\ref{LMA:L9}.
Thus,
we prove that there exists an MCDS into which $D'$ is regularized and satisfies (P1), that is,
which contains $(p+3, n-2)$.
We can obtain a $(p, n)$-$(p+3, n)$-regular MCDS $D''$ from ~\ref{LMA:L4} by applying Lemma~\ref{LMA:L4} to ~\ref{LMA:L4} twice
(Fig.~\ref{fig:c9}(4)).
Since $(p+1, n), (p+2, n), (p+3, n) \in D''$,
we define
$D''' = D'' \backslash \{ (p+1, n), (p+2, n) \} \cup \{ (p+3, n-2), (p+3, n-1) \}$.
Since $D'''$ is still dominating and connected by Fig.~\ref{fig:c9}(5),
$D'''$ is an MCDS.
Moreover,
since $(p+1, n), (p+2, n), (p+3, n-2)$ and $(p+3, n-1)$ are irregular vertices in $D'$ and $D$,
$D$ can be $(p+3, n-3)$-$(p+3, n-2)$-regularized into $D'''$.
We consider the case in which $q = n-2$.
If $(p+3, n-1) \in D$,
the statement is clearly true
(Fig.~\ref{fig:c9}(7)).
If $(p+3, n-1) \notin D$
and
$(p+2, n-1) \in D$
(Fig.~\ref{fig:c9}(9)),
then the statement is true
because $D$ can be $(p+3, n-2)$-$(p+3, n-1)$-regularized into $D'$ by Lemma~\ref{LMA:LT}(i).
If $(p+3, n-1) \notin D$,
$(p+2, n-1) \notin D$
and
$(p+1, n) \in D$
(Fig.~\ref{fig:c9}(11)),
we define $D' = D \backslash \{ (p+1, n) \} \cup \{ (p+2, n-1) \}$.
$D'$ is dominating and connected by the figure.
Thus,
$(p+3, n-1) \notin D'$
and
$(p+2, n-1) \in D'$,
which implies that the statement is true
because this situation is the same as the previously discussed case.
Finally,
we consider the case in which
$(p+3, n-1) \notin D$,
$(p+2, n-1) \notin D$
and
$(p+1, n) \notin D$.
Since $D$ contains vertices to dominate $(p+2, n)$ and is connected,
$(p+3, n) \in D$
(Fig.~\ref{fig:c9}(12)).
There exists an MCDS $D' = D \backslash \{ v' \} \cup \{ (p+3, n-1) \}$,
in which $v' \in {\overline{R}}(D)$ is a mobile in $D$,
by applying Lemma~\ref{LMA:MV}~(ii) with $(p+3, n-2) \in R(D)$ and $(p+3, n) \in {\overline{R}}(D)$ as $(x, y)$ and $(x+2, y)$, respectively.
$D'$ is $(p+3, n-2)$-$(p+3, n-1)$-regular
because $(p+3, n-1) \in D'$.
\noindent
{\bf (iii):}
We omit the proof of this case
because we can prove it similarly to that of the case (ii).
\fi
\end{proof}
\ifnum \count12 > 0
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig_c9.eps}
\end{center}
\caption{
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
Lemma~\ref{LMA:C9}.
Suppose that $D$ is an MCDS which is $(p, q)$-$(m, q)$-regular or $(p, q)$-$(p, n)$-regular.
Circles and crosses denote vertices in $D$ and not in $D$, respectively.
White circles denote vertices in $D$ which are handled at the regularization.
White crosses denote vertices not in $D$
or vertices removed from $D$ at the regularization.
\fi
}
\label{fig:c9}
\end{figure*}
\fi
\ifnum \count10 > 0
\fi
\ifnum \count11 > 0
\fi
|
\section{Introduction}
\label{INTRO}
Stochastic geometry was initially stimulated by applications to biology,
astronomy and material sciences, then introduced seriously in the late
nineties to the field of wireless communications \cite{Baccelli10}. It is
particularly suited for modelling large scale wireless communication
networks, where a network is treated as a realization (snapshot) of a
spatial point process in the entire Euclidean plane \cite{Haenggi12}. In
particular, it is a natural approach to describe node locations in randomly
formed networks, e.g. ad hoc networks, or irregular and regular networks,
e.g. cellular networks. It provides a natural way of computing macroscopic
properties, by averaging over the potential geometrical patterns for all
network nodes to obtain key performance characteristics, such as
connectivity, stability, and capacity, as functions of a relatively small
number of parameters, e.g. the intensity of the underlying point process and
the operational parameters. By a spatial average, it is meant an average
calculated over many points in the considered domain. These locations in our
context are the network elements at the time when the snapshot is taken.
Example spatial averages are the fraction of nodes which transmit, the
fraction of space which is covered or connected, the fraction of nodes that
transmit their data successfully, and the average geographic progress
achieved by a node forwarding a packet towards some destination \cite{ElSawy17}.
The most common function of stochastic geometry in wireless communications is to characterize
the signal to interference and noise ratio (SINR), which can be used to
calculate many cellular performance metrics, such as outage probability,
coverage probability, spatial opportunity, spatial throughput, network
throughput, medium access probability and spectral efficiency \cite%
{Okegbile21}, in both downlink (DL) and uplink (UL) directions. It should be
noted, however, that in light the huge of influx of wireless emissions in
recent years, the impact of noise now pales in comparison with interference
\cite{Chun20}. As such, there is a growing modelling trend (see, for example, \cite{Tang20, Liu20, Haroon20, Kouzayha21}) to replace SINR by SIR,
and we will follow this trend in the present article. In all cases, the most
relevant assumption made when stochastic geometry is used in modelling
cellular networks is that base stations (BS) are deployed in the Euclidean
space as a Poisson Point Process (PPP), although many variant processes,
e.g. Thomas cluster processes or Mat\'{e}rn cluster processes, have been
suggested as well \cite{Blaszczyszyn18}.
The stochastic geometric model of cellular networks, portrayed in Figure \ref%
{fig:Voronoi}, has established itself confidently in the past decade as a
replacement to the once popular hexagonal grid model, in which the base
stations were placed at the centers of the hexagonal lattices. It has been
shown \cite{Lee13} that the PPP approach provides much more accurate results
than the hexagonal grid model when both are used to model real world
cellular installation. It has also been shown \cite{Andrews10} that the PPP
model gives lower bounds, whereas the hexagonal model gives upper bounds, of
the coverage probability, which means that the former is safer to rely on.
\begin{figure}[H]
\centering
\begin{tikzpicture}[xscale=1.4,yscale=1.4]
\begin{axis}[
ticks=none,
scale only axis,
xmin=0,
xmax=8.8,
ymin=0,
ymax=7.5,
axis line style={line width=2pt},
legend pos=outer north east,
legend cell align=left,
legend style={draw=none,font=\small,row sep=0.2cm,/tikz/every odd column/.append style={column sep=0.1cm,text width=8em}},
]
\addplot[only marks, mark=BS,mark size=2.5pt, red] coordinates{(1.9, 0.68)};
\addplot[only marks,seagreen,mark size=1.5pt,very thick] coordinates {
(7.5,0.48)
};
\addplot[mark=BS,mark size=2.5pt, red] coordinates{(6.3, 1.15)};
\addplot[mark=BS,mark size=2.5pt, red] coordinates{(7.1, 0.21)};
\addplot[mark=BS,mark size=2.5pt, red] coordinates{(1, 3.9)};
\addplot[mark=BS,mark size=2.5pt, red] coordinates{(0.12, 4.66)};
\addplot[mark=BS,mark size=2.5pt, red] coordinates{(1.25, 6.77)};
\addplot[mark=BS,mark size=2.5pt, red] coordinates{(4.9, 2.7)};
\addplot[mark=BS,mark size=2.5pt, red] coordinates{(7.77, 7)};
\addplot[mark=BS,mark size=2.5pt, red] coordinates{(8.09, 2.55)};
\addplot[no markers, cyan!70!green] coordinates {
(0,2)
(2.6,2.5)
(4.1,0.94)
(4.25,0)
};
\addplot[no markers,cyan!70!green] coordinates {
(2.6,2.5)
(3.75,5.15)
(1.85,5.3)
(0,3.85)
};
\addplot[no markers,cyan!70!green] coordinates {
(1.85,5.3)
(0,5.98)
};
\addplot[no markers,cyan!70!green] coordinates {
(3.75,5.15)
(4.6,5.65)
(4.49,7.5)
};
\addplot[no markers,cyan!70!green] coordinates {
(4.1,0.94)
(6.47,2.47)
(6.62,4.73)
(4.6,5.65)
};
\addplot[no markers,cyan!70!green] coordinates {
(6.47,2.47)
(7.8,1.35)
(5.6,0)
};
\addplot[no markers,cyan!70!green] coordinates {
(7.8,1.35)
(9.5,0.8)
};
\addplot[no markers,cyan!70!green] coordinates {
(6.62,4.73)
(9.5,4.9)
};
\addplot[only marks,seagreen,mark size=1.5pt,very thick] coordinates {
(0.26 ,0.35)
(0.92 , 0.62)
(1.84 , 1.57)
(2.76 ,0.59)
};
\addplot[only marks,seagreen,mark size=1.5pt,very thick] coordinates {
(0.32, 2.7)
(0.94, 2.52)
(1.62, 2.75)
(1.3, 3.25)
(1.25, 4.51)
(1.93, 4.8)
(2.28, 4.25)
(2.35, 4.75)
};
\addplot[only marks,seagreen,mark size=1.5pt,very thick] coordinates {
(1, 5.1)
};
\addplot[only marks,seagreen,mark size=1.5pt,very thick] coordinates {
(1.92, 6.2)
(3.1, 6.95)
(4.3, 6.55)
};
\addplot[only marks,seagreen,mark size=1.5pt,very thick] coordinates {
(3.3, 1.98)
(4.1, 3.38)
(4.32, 5.15)
(5.25, 4.92)
(6.05, 4.45)
(5.65, 3.6)
};
\addplot[only marks,seagreen,mark size=1.5pt,very thick] coordinates {
(5.6 , 1.35)
(6.4 , 0.89)
(6.6 , 2.1)
};
\addplot[only marks,seagreen,mark size=1.5pt,very thick] coordinates {
(5.68, 7.1)
(6.3, 6.25)
(7.55, 5.72)
(7.75, 6.55)
(8.15, 6.9)
};
\addplot[only marks,seagreen,mark size=1.5pt,very thick] coordinates {
(6.95, 2.35)
(6.86, 3.55)
(6.7, 4.3)
(8.13, 4.07)
(8.53, 2.03)
};
\node [right] at (axis cs: 4.13, 3.4) {{\scriptsize{\color{black} UE}}};
\node [right] at (axis cs: 5, 2.7) {{\scriptsize{\color{black} BS}}};
\end{axis}
\end{tikzpicture}
\caption[Vor]{A Poisson-Voronoi tessellation model of a cellular network. The BSs, induced by a PPP, partition the plane into Voronoi cells, each having one BS and a number of UEs, with the property that a UE in some cell is closer to the BS of that cell than to the BS of any other cell. This model is so natural and realistic for cellular networks that it has superseded the once popular model of hexagonal cells, which was artificial and idealistic.}
\label{fig:Voronoi}
\end{figure}
This spatial consideration afforded by stochastic geometry is a key paradigm
shift in communications systems performance evaluation, where time averages
have traditionally reigned supreme. Before the advent of stochastic
geometry, performance analysis of wireless networks was performed using
either damaging mathematical simplifications or exhaustive simulations with
the aim to average out the many sources of randomness, such as BS and user
equipment (UE) locations and fading phenomena. The mathematical
simplifications gave rise to poor analytical results.
An important technique used in cellular networks, and considered in the
present article, is fractional power control (FPC), where each UE adjusts
its power level in the UL direction under control of its serving BS \cite%
{Haroon20}. More specifically, each UE controls its transmit power such that
the received signal power at its serving BS is equal to a predefined
threshold. This control benefits the UE by saving its energy and benefits
the BS by decreasing its SIR. The optimal levels of transmit power in a
network depend on path loss, shadowing, and multipath fading, as well as the
network configuration. It should be noted, however, that while FPC in
homogeneous cellular networks, it is less effective in heterogeneous
networks (HetNets), where there are more than one tier of BSs, because far
UEs, especially those at the edge of the cell, will use more transmit power,
causing more interference to neighbor cells.
We prove in this article two theorems that expose a prevalent flaw prevalent in the stochastic geometric characterization of coverage
probability, namely that the latter is dependent of BS density. The
Theorems, one for DL and one for UL, prove that the opposite is true; namely
that the coverage probability is independent of BS density. This fact
has only been alluded to in some articles previously \cite{Bai15,Herath18}. The
authors of \cite{Bai15} in the context of analyzing DL coverage probability
of millimeter-wave cellular networks noted that \textquotedblleft coverage
does not scale with BS density." Also, the authors of \cite{Herath18}, while
analyzing UL FPC, noted that coverage is \textquotedblleft \ invariant to the
density of deployment of BSs when the shadowing is mild and power control is
fractional." To the best of our knowledge, the present article is the first
to provide a rigorous mathematical proof that coverage probability is
independent of BS density.
The rest of the article is organized as follows. Section \ref{RW} reviews the
stochastic geometric models of cellular networks in the recent literature. Section \ref{MODEL} provides the
modelling process of a cellular system in both directions, downlink (DL) and
uplink (UL). The simulation algorithms developed to validate the mathematical models are presented in Section \ref{SIM}. Some numerical results for sample systems are presented in Section \ref{EW}. Section \ref{CONC} has the conclusions.
\section{Related work}
\label{RW}
Perhaps the precursor of using stochastic geometry as a modelling vehicle
for wireless systems was the work in \cite{Gilbert61}, approximately half a
century ago. However, the real momentum, especially for cellular networks,
has been gained over the past decade which witnessed an outbreak of hectic
research in the field. This research can be categorized in many ways, but
the direction of transmission, i.e. DL, UL or both, seems suitable enough.
Much work has focused on DL coverage probability. For example, in \cite{Li11}%
, the problem of DL data transmission scheduling in wireless networks is
studied, proposing two algorithms concerning power assignment. In \cite%
{Bai15}, the authors study DL coverage probability in the context of
millimeter-wave (mmWave) cellular networks and note that mmWave coverage
does not scale with BS density. In \cite{Chen19}, DL coverage probability is
studied with composite $\kappa-\mu$ shadowed and lognormal shadowed fading,
also known as doubly-selective
(DS) fading. In \cite{Qiong20}, the authors derive an expression for DL
coverage probability, taking into account a queue setup. The static
properties of the physical layer of the network are studied by stochastic
geometry and the dynamic properties of the queue setup are studied with a
discrete time Markov chain. Upper and lower bounds on the dynamic coverage
probability are derived. In \cite{Liu20}, the authors study DL in a setting
of an ultra-dense network, a promising technology for 5G cellular networks,
where lots of low power Small Base Stations (SBSs) overlapping with Macro
Base Stations (MBSs) are deployed. The authors derive the coverage
probability.
Many authors have analyzed the DL coverage probability in the context of
HetNets. For example, in \cite{Ouamri20}, the authors characterize the DL
coverage probability in the context of a 2-tier HetNet, under Line of Site
(LOS) and Non Line of Site (NLOS) Path loss models with different path loss
exponents, using stochastic geometry. Also working in HetNets are the
authors of \cite{Fadoul20b}, who consider interference to be a major
performance bottleneck. They consider $K$-tier transmission modeled by
factorial moment and stochastic geometry to obtain expressions for the
coverage probability. They also compare the model with a single-tier,
traditional hexagonal grid model, concluding the superiority of the former.
Also, in \cite{Lei18}, DL is studied using the stochastic geometry for a
two-tier ultradense HetNet with small-cell base stations (SBSs) and UEs
densely deployed in a traditional macrocell network. Performance is studied
in terms of the association probability, average link spectral efficiency
(SE), and average DL throughput. In \cite{Fadoul20}, the authors study DL
for HetNets, where in each typical macrocell, many
femtocells and few picocells are deployed. They reveal that the coverage probability of the DL direction
decreases when the FPC increases in the UL direction,
mentioning nothing of the deployment density. In \cite{Bouras20}, the authors
note that in current cellular networks, cell association is heavily based on
the DL signal power and that all devices are associated with the same BS in
both DL and UL, noting that this technique is inappropriate in HetNets as
transmission levels significantly vary from BS to BS. They suggest
decoupling DL and UL as a solution, where the UL cell association is not
necessarily based on the same criteria as the DL association. The authors of
\cite{Arif20} consider also the same decoupling solution to alleviate the
load imbalance problem in HetNets. The BSs in each tier are modeled by an
independent, homogeneous PPP, whereas the spatially clustered UEs are
modeled by a Mat\'{e}rn cluster process. Analytical expressions for the coverage
probability are derived, showing the merit of the solution.
Like DL, a huge body of research work has focused on UL coverage probability
in different scenarios. For example, in \cite{Kouzayha20}, the authors study
UL for a radio frequency (RF) wake-up solution for Internet of Things (IoT)
devices over cellular networks. When the IoT device has no data to transmit,
it turns its main circuitry OFF and switches to a sleep mode. The transition
back to the active mode is only achieved upon receiving enough power at the
device's front end. They derive an expression for the UL coverage
probability after successful wake up. In \cite{Chun20}, the authors study UL
in a densely deployed cellular network in an interference limited scenario,
meaning that the thermal noise power is very small if compared with the
interference. The authors characterize the SIR, noting it is much more
tractable than SINR in a Poisson cellular network. In \cite{Mariam21}, UL is
studied in the context of mmWave cellular networks. The coverage
probability is characterized under the assumption that the BSs
and UEs are modeled as two independent homogeneous PPPs. Interestingly, the
authors find no clear and consistent correlation between the coverage probability and BS
density, noting that the probability increases with increasing BS density at
first, then decreases as the network becomes denser and eventually collapse
as the density is further increased. In \cite{Kamiya20}, the authors employ
stochastic geometry to study the SINR for UL Poisson cellular network, where
truncated fractional transmit power control is performed. They obtains the
UL SINR distribution under channel-adaptive user scheduling, including cases
in which edge users are both allowed and not allowed to transmit at maximum
transmit power. In \cite{Herath18}, the authors analyze three UL transmit
power control schemes, assuming composite Rayleigh-lognormal fading. Using
stochastic geometry, they derive network coverage probability, noting that
it is highly dependent on the severity of shadowing, the power control
scheme, but \emph{invariant} of the density of deployment of base stations
when the shadowing is mild and power control is fractional. This note is
relevant to and consistent with the theorems proved in the present work.
Also, like DL, many authors have studied UL coverage probability for
HetNets. For example, in \cite{Haroon20}, the authors study SIR for HetNets
leveraging FPC. They propose nonuniform SBS deployment (NU-SBSD) to reduce
interference, where SBS deployment (SBSD) near macro base station (MBS) is
avoided, and MBS coverage edge area is enriched with ultra-dense SBSD. Also,
in \cite{Onireti20}, the authors characterize the UL outage probability,
which is the complement of coverage probability, in multitier millimeter
wave cellular networks, using stochastic geometry. In each tier, the BSs are
assumed to have their spatial density, antenna gain, receiver sensitivity,
blockage parameter, and pathloss exponents. The results show that imposing a
maximum power constraint on the user significantly affects the SINR and
outage probability. Moreover, in \cite{Jia19}, the authors indicate that in
the conventionally coupled UL and DL association (CUDA), the UL performance
is limited greatly by the DL parameters such as the density and power of
base stations (BSs). To overcome this issue, they focus on a three-tier
HetNet as well as global performance evaluation, where the cross-tier
dual-connectivity (DC) and decoupled UL and DL association (DUDA) are
integrated. In \cite{Ali19}, UL is studied for a two-tier dense HetNet,
where decoupled UL coverage probability is computed using multi-slope path
loss model. The authors find that the decoupled UL coverage probability is
higher when incorporating multi-slope path loss model as compared to
single-slope path loss model, while the decoupled UL spectral efficiency is
observed to be lower when incorporating dual slope path loss model.
Finally, some research work, such as ours, focuses on both DL and UL,
targeting comprehensive treatment. For example, in \cite{Andrews16}, an
authoritative tutorial on cellular network analysis using stochastic
geometry is provided, characterizing DL and UL coverage probabilities after
a lavish introduction to stochastic geometry. Also, in \cite{Gao19}, both DL
and UL are investigated in the context of mmWave communication, known for
being sensitive to blockage. The authors consider a Time-Division Duplex
(TDD) mode able to provide dynamic UL-DL configurations, and compute the
coverage probability based on stochastic geometry. In \cite{Kundu20}, both
DL and UL are studied in the context of full duplex (FD) for a UE capable of
transmitting and receiving data simultaneously in the same frequency
resource. The authors obtain the coverage probabilities, revealing that
while FD improves DL performance, it severely hurts UL performance.
Furthermore, in \cite{Sadeghabadi20}, the authors study both DL and UL for
massive multiple-input multiple-output (M-MIMO), promising for increasing
the spectral efficiency. They analyze an asynchronous DL M-MIMO system in
terms of the coverage probability by means of stochastic geometry.
Some authors investigate coverage probability for both DL and UL, but in the
context of HetNets. An interesting example is in \cite{Wang19} where the
authors consider DL simultaneous information transmission and power transfer
and UL information transmission for an unmanned aerial vehicle
(UAV)-assisted mmWave cellular network. Distinguishing features of mmWave
communications, such as different path loss models and directional
transmissions are taken into account. DL association probability and energy
coverage of the two tiers of UAVs and ground base stations (GBSs) are
investigated.
Still, some other authors study both DL and UL for the purpose of
decoupling, aiming to alleviate BS load imbalance and also ease
interference. For example, in \cite{Bouras20}, the authors note that in
current cellular networks, cell association is heavily based on the DL
signal power and that all devices are associated with the same BS in both DL
and UL. While this practice is adequate in homogeneous networks, where all
BSs have similar transmission levels, it can fail in dense HetNets as
transmission levels significantly vary from BS to BS. They propose
decoupling DL and UL as a solution, where the UL cell association is not
necessarily based on the same criteria as the DL association. Also, the
authors of \cite{Arif20} consider DL and UL decoupling is investigated to
alleviate the load imbalance problem in HetNets, aiming to increase the
coverage probabilities and data rates. The BSs in each tier are modeled by
an independent, homogeneous PPP, whereas the spatially clustered UEs are
modeled by a Mat\'{e}rn process. Analytical expressions for the coverage
probability, are derived.
\section{modelling}
\label{MODEL}
The subject of stochastic geometry and its use in modelling wireless systems
is vast \cite{Blaszczyszyn18}. The key aspect of the present study is that
all the BSs are located according to a PPP $\Phi $, which effectively means
they are randomly scattered in the Euclidean plane with independent
locations. At any given time, only one UE can be active communicating with
the BS on any time/frequency resource. Before delving even more into the
modelling process, some definitions used throughout the article are in order.
\begin{description}
\item \textbf{Definition 1 (BS-UE association)}: BS-UE association is the assignment of a UE to a BS, for both to
establish a communications session.
\item \textbf{Definition 2 (Serving BS)}: Once a UE is associated with a BS, the latter is said to be the serving BS of the UE.
\item \textbf{Definition 3 (Typical receiver)}: The typical receiver is the receiving device (UE or BS) where the SIR is to be assessed. It is always placed at the origin of the Euclidean plane in the
model, or the origin of the simulation window in the simulation.
\item \textbf{Definition 4 (Tagged transmitter)}: The tagged transmitter is transmitting device (UE or BS) associated with the typical receiver.
\item \textbf{Definition 5 (Typical circle)}: The typical circle is the circle centered at the typical
receiver and having the tagged transmitter on its circumference.
\item \textbf{Definition 6 (Signal)}: A signal is the transmission arriving at, and intended for,
the typical receiver.
\item \textbf{Definition 7 (Interference)}: An interference is the transmission arriving at, but not intended for, the typical
receiver.
\item \textbf{Definition 8 (Interferer)}: An interferer is a transmitter causing interference at the typical receiver. That is, it is any transmitter in the network other than
the tagged transmitter.
\item \textbf{Definition 9 (Signal to Interference Ratio (SIR))}: The quotient of
the signal at the typical receiver and the sum of all
interferences at the typical receiver.
\item \textbf{Association rule}: A UE will associate with the BS nearest it.
\end{description}
As per Definitions 3 and 4, in DL the typical receiver is a UE and the
tagged transmitter is a BS, whereas in UL the typical receiver is a BS and
the tagged transmitter is a UE.
As per the association rule, a UE will be nearer to its serving BS than to
any other BS in the cellular network. Strictly speaking, however, the UE
associates with the BS that provides the highest average SIR. Surprisingly,
depending on the fading and shadowing conditions, a more distant BS could
provide a higher instantaneous SIR, hence gets associated with the UE. That
is why, we rely on the average, which invariably translates to association
with the nearest BS. We will denote the distance between the associated
BS-UE pair throughout by $R$. If the BS density is $\lambda $, it can be
shown that $R$ is a random variable (RV) with the Rayleigh distribution\
\begin{equation}
f_{R}(r)=2\lambda \pi re^{-\lambda \pi r^{2}},\qquad r\geq 0
\label{Reighlay}
\end{equation}
Along the same line, the notation used throughout the article is provided in
Table \ref{table:Notation}. The notation of the PPP, pivotal in our work,
depends on its setting and the field it is being applied in and on the
interpretation of the process. For example, a simple PPP $\Phi $ may be
considered as a random set, which suggests the notation $x\in \Phi $,
implying that $x$ is a random point belonging to or being an element of the
Poisson random set $N$. Another, more general, interpretation is to consider
a Poisson or any other point process as a random counting measure, so one
can write the number of points of a PPP $\Phi $ being found or located in
some (Borel measurable) region $B$ as $\Phi (B)$, which is an RV (having a
Poisson distribution, in the case of a PPP) \cite{Blaszczyszyn18}.
\begin{table}[H]
\caption{Notation used in the model and simulation.}
\centering
\begin{tabular}{c l }
\hline\hline
Parameter & Description\\ [0.5ex]
\hline
BS & Base station\\
UE & User Equipment (can be a mobile phone, tablet, laptop, etc.)\\
$\Phi$ & Poisson point process (PPP) of BSs\\
$\Psi$ & Point process of UEs (not Poisson) \\
$\lambda$ & Density of BS (per m$^2$), i.e. intensity of PPP $\Phi$\\
$\alpha$ & Path-loss exponent (per m) \\
SIR & Signal to interference ratio (dB) \\
$\xi$ & SIR threshold (dB)\\
$G$ & Rayleigh channel gain of tagged transmitter ($G \sim Exp(1)$)\\
p & Transmit power (Watts)\\
$p_d$ & DL coverage probability\\
$p_u$ & UL coverage probability\\
\hline
$\mathcal{N}$ & Number of simulation runs (PPP realizations)\\
$S$ & Side of the simulation square window (m)\\
$N$ & Number of BSs in a simulation run, $N \sim Pois(\lambda S^2)$).\\
($x,y$) & Location of a BS in a simulation run\\
($u,v$) & Location of a UE in a simulation run\\
Covered & No. of times typical receiver is covered by tagged transmitter\\
\hline
\end{tabular}
\label{table:Notation}
\end{table}
Figure \ref{fig:DU} shows a stochastic geometric construction to
characterize the SIR at the typical receiver of a cellular network, which with the tagged transmitter,
defines the typical circle. In part \ref{fig:DL}, we can see
the DL model, where the typical receiver is a UE, and the tagged transmitter
is a BS at distance $R$. All the BSs outside the typical circle are
interferers to the typical UE. The typical circle defines an exclusion zone,
in the sense that there can be no BS inside of it (or else the typical UE
would associate with it.) In part \ref{fig:UL}, we can see the UL model, where the typical
receiver is a BS, and the tagged transmitter is a UE at distance $R$. All
the UEs except the tagged are interferers to the typical BS. The
typical circle does not define an exclusion zone, as there can be UEs inside
of it as we can see (the association problem of the DL model is not present here.)
\begin{figure}[H]
\centering
\begin{subfigure}{0.2\textwidth}
\caption{ \vspace{-1cm} \hspace{-10cm}\small (a)}
\begin{tikzpicture}[scale=1.5]
\hspace{-3cm}
\begin{axis}[,
height=8.5cm,
width=8cm,
hide axis,
legend style={at={(0.55,-.01)}, anchor=north,legend columns=-3,mark options={scale=0.75}},
]
\definecolor{sienna}{rgb}{0.53, 0.18, 0.09}
\addplot [only marks, mark size=2.5pt, thin, color=red, mark=BS,mark options={scale=1}]
coordinates {
(0,0)
(1.5,4.14)
(-1.7,3.59)
(-0.46,2.71)
(2.5,3.22)
(-2.4,1.28)
(-0.24,1.30)
(1.05,1.53)
(3.3,1.86)
(3,0.36)
(-2.6,-0.62)
(-0.7,-0.28)
(3,-1)
(-0.57,-1.5)
(1.7,-1.39)
};
\addplot [only marks, mark size=1.5,color=seagreen,very thick,mark options={scale=1}] coordinates {
(-1.05,4.54)
(2.3,4.54)
(0.08,3.69)
(2.15,3)
(-1.64,1.02)
(-0.31,0.65)
(1.30,0.80)
(3.07,1.28)
(1.22,0.10)
(2.18,0.07)
(-0.28,-0.56)
(3.87,-0.48)
(-3.44,-1.24)
(-0.06,-1.17)
(2.37,-1.61)
};
\addplot [only marks, mark size=5,color=black, mark=+] coordinates {
(1.22,0.10)
};
\draw[cyan, thick=1pt] (axis cs:1.22,0.10) circle (0.9cm);
\draw[blue,thick=1pt] (axis cs:0,0) -- node[centered, above=-1mm,black]{\scriptsize$R$} (axis cs:1.22,0.10);
\draw[dashed,thick=1pt,seagreen] (axis cs:1.22,0.10) -- node[centered, above right=-2.3mm,black]{\scriptsize $D_z$} (axis cs:-0.49,2.71);
\draw[red,thick=1pt] (axis cs:0.08,3.69) -- node[centered, above left= -2mm,black]{\scriptsize $R_z$} (axis cs:-0.46,2.71);
\put(105,202){\shortstack{$z$}}
\end{axis}
\end{tikzpicture}
\label{fig:DL}
\end{subfigure}
\rule{.7\textwidth}{0.01in}
\vspace{0.6cm}
\hspace{-0.35cm}
\begin{subfigure}{0.2\textwidth}
\vspace{-1.2cm}
\caption{ \vspace{-1cm} \hspace{-10cm}\small (b)}
\begin{tikzpicture}[scale=1.5]
\hspace{-3cm}
\pgfplotsset{
every axis legend/.append style={ at={(1.05,0.95)}, anchor=north west,legend columns = 1}}
\begin{axis}[,
height=8.5cm,
width=8cm,
hide axis,
]
\addplot [only marks, mark size=2.5pt, thin, color=red, mark=BS]
coordinates {
(0,0)
(1.5,4.14)
(-1.7,3.59)
(-0.46,2.71)
(2.5,3.22)
(-2.4,1.28)
(-0.24,1.30)
(1.05,1.53)
(3.3,1.86)
(3,0.36)
(-2.6,-0.62)
(-0.7,-0.28)
(3,-1)
(-0.57,-1.5)
(1.7,-1.39)
};
\addplot [only marks, mark size=1.5,color=seagreen, very thick] coordinates {
(-1.05,4.54)
(2.3,4.54)
(0.08,3.69)
(2.15,3)
(-1.64,1.02)
(-0.31,0.65)
(1.30,0.80)
(3.07,1.28)
(1.22,0.10)
(2.18,0.07)
(-0.28,-0.56)
(3.87,-0.48)
(-3.44,-1.24)
(-0.06,-1.17)
(2.37,-1.61)
};
\addplot [only marks, mark size=5,color=black, mark=+] coordinates {
(0,0)
};
\definecolor{sienna}{rgb}{0.53, 0.18, 0.09}
\draw[cyan, thick=1pt] (axis cs:0,0) circle (0.9cm);
\draw[ thick=1pt, blue] (axis cs:0,0) -- node[centered, above= -1mm,black]{\scriptsize$R$} (axis cs:1.22,0.10);
\draw[ thick=1pt, dashed, seagreen] (axis cs:0,0) -- node[centered, above right= -1mm,black]{\scriptsize $U_{\mathfrak{z}}$} (axis cs:0.08,3.69);
\draw[thick=1pt, red] (axis cs:0.08,3.69) -- node[centered, above left= -2mm,black]{\scriptsize $R_{\mathfrak{z}}$} (axis cs:-0.46,2.71
);
\put(138,240){\shortstack{$\mathfrak{z}$}}
\end{axis}
\end{tikzpicture}
\label{fig:UL}
\end{subfigure}
\caption[T]{Stochastic geometric models for assessing the SIR at a typical receiver at the origin of the cellular network. (a) The DL model, where the typical receiver is a UE. The typical circle defines an exclusion zone, as it cannot contain a BS inside. The BSs outside the typical circle, which cause interference at the typical UE, form a PPP $\Phi$. (b) The UL model, where the typical receiver is a BS. The typical circle does not define an exclusion zone, as it can contain UEs inside. All the UEs form a PP $\Psi$. Since the UEs are {\emph satellites} to their serving BSs, as per the association rule, $\Psi$ is not Poisson, a major challenge to the analysis. We mitigate this challenge by relocating each UE, except the tagged, to the position of its serving BS. That is, we relocate all the interferers.
}
\label{fig:DU}
\end{figure}
The RV $R_{z}$ are identically distributed but not independent in general.
The dependence is induced by the structure of Poisson-Voronoi tessellation
and the restriction that only one BS can lie in each cell. To visualize this
dependence, recall that the presence of a BS in a particular Voronoi cell
forbids the presence of any other BS in that cell. However, as discussed in
detail in the next section, this dependence is weak, motivating the
approximation we will make in the UL model below. On the other hand, the RV $%
R_{z}$ is upper bounded by the RV $D_{z}$, even for the pairs inside the
typical circle (or else the BS at $z$ would be closer to the typical UE,
violating the association rule.). The UEs are satellites to the points of
the PPP $\Phi $, based on the association rule, so they are not themselves a
PPP, making the UL model a challenge. This challenge is mitigated in the
present work through an elegant approximation.
We can relate Figure \ref{fig:Voronoi} and Figure \ref{fig:DU} by
recognizing that in the present article, we consider orthogonal
communications. This means that in each cell there can be only one active UE
on any time/frequency resource. Accordingly, Figure \ref{fig:Voronoi} is a
snapshot of the UEs that are active in all the cells at the same time. Thus,
we see a random number of UEs in each cell. Figure \ref{fig:DU}, on the
other hand, is a snapshot of the UEs that are active on the same frequency
in all the cells at the same time. Thus, we see exactly one UE per cell. We
focus in the model on one such resource. That is, every BS-UE associated
pair in Figure \ref{fig:DU} is operating on the same resource, hence is the
interference that we are going to characterize.
Given a homogeneous PPP with intensity $\lambda $, the number of process
points in the an area of size $A$, denoted by $\Phi (A)$, is Poisson
distributed%
\begin{equation*}
\mathbb{P}[\Phi (A)=k]=\frac{e^{-\lambda |A|}(\lambda |A|)^{k}}{k!}
\end{equation*}%
That is, a homogeneous PPP is completely characterized by a single number, $%
\lambda $. It is both stationary (translation invariant) and isotropic
(rotation invariant).
Similarly, if $F$ is a continuous RV, we can also write its Laplace transform as a
(continuous) expectation, namely
\begin{equation}
\mathcal{L}_{F}(s)=\mathbb{E}\left[ e^{-sF}\right] \label{LapAsExpect}
\end{equation}
The probability generating function of a non-negative integer-valued RV
leads to the probability generating functional (PGFL) with respect to any
non-negative bounded function $f$ for a point process defined on $\mathbb{R}%
^{d}$, with $0\leq f(x)\leq 1$. The need for the PGFL, which we will invoke
twice in this article, arises naturally when one wishes to find the Laplace
transform of a sum of functions. As the Laplace transform of a RV is already
an expectation of an exponential function, with the RV its exponent, then if
the RV is a sum of independent and identically distributed (iid) RVs, the
exponential becomes a product lending itself readily to the invocation of
PGFL. Due to this intimate relationship, some authors call the PGFL in this
context a Laplace functional.
We will also use in the present work Slivnyak's theorem \cite{Blaszczyszyn18}%
. It states that for a PPP, because of the independence between its points,
conditioning on a point at some location $x$ does not change the
distribution of the rest of the process. That is, removing an
infinitesimally small area from the underlying plane of the PPP, leaves
still a PPP, since the distribution of points in all nonoverlapping areas
are independent for the PPP. Consequently, any property seen from some
location $x$ remains the same whether or not there is a point at $x$. Though
simple, this theorem allows us to add and remove points to/from the PPP
liberally. In our work, for example, it allows us in UL to place a BS from
those imposed by the PPP $\Phi $\ and place it at the origin, treating the
remaining points, $\Phi \backslash x$, still as a PPP (for example, when
invoking a PGFL.)
Before starting the modelling process, a word on random channels is in
order. In our work, random channel effects are incorporated by
multiplicative RVs, namely $G$ for the signal and $G_{i}$ for each
interferer $i$. For simplicity we assume these all correspond to Rayleigh
fading with mean $1$, denoted $\exp (1)$, i.e. $\mu =1$ in $f_{G_{i}}(x)=\mu
e^{-\mu x}$. We consider small-scale Rayleigh fading, and assume $G$ and the
$G_{i}$ iid RVs following an exponential distribution with mean $1$.
\subsection{Downlink model}
The key assumptions of the DL system model are as follows:
\begin{itemize}
\item BSs are located according to a homogeneous PPP $\Phi $ of intensity $%
\lambda $ in the Euclidean plane.
\item A UE wishing to start a communications session associates with the BS
that is closer to it than any other BS in the cellular network (association
rule.)
\item The BS transmits at a fixed power $p$ to a designated UE on a
particular time-frequency resource, i.e. orthogonal multiple access
communications within the cell take place. The consequence of orthogonality
is that the UE sees interference from all other BSs in the plane using the
same resource.
\item Signals attenuate with distance according to the standard power-law
path loss propagation model, with path loss exponent $\alpha >2$. That is,
the average received power at distance $r$ from a transmitter of power $p$
is $pr^{-\alpha }$.
\item Random channel effects are incorporated by a multiplicative RVs $G$
for the signal and $G_{z}$ for every interferer at $z$. For simplicity we
assume these all to correspond to Rayleigh fading with mean $1$, so $G$ and
the $G_{z}$ are iid RVs, having exponential distribution with mean $1$.
\end{itemize}
Let $I_{d}$ denote the interference experienced at the typical UE. The interference is due to every BS $z$ in the plane, except the tagged BS, denoted by $\mathfrak{b}$, at
distance $D_{z}$ from the typical UE, as shown in Figure \ref{fig:DL},. That is, the interference $I_{d}$ is created
by a PPP with intensity $\lambda$ outside the typical circle, and is given by
\begin{equation}
I_{d}=\sum_{z\in \Phi \backslash \{ \mathfrak{b}\}}pG_{z}D_{z}^{-\alpha},
\label{id}
\end{equation}%
where $p$ is the power of the BS at point $z$, and $D_{z}$ is a RV representing the
distance from the BS at $z$a and the typical UE. The
SIR at the typical UE is then given by
\begin{equation}
\mathtt{SIR}_{\text{UE}}=\frac{pGR^{-\alpha }}{I_{d}} \label{SIRd}
\end{equation}%
Note that the distance $R_{z}$ is not used in anything in DL, but it will be used
in the UL model. Note also that the typical circle defines an exclusion zone. That is, there can never be a BS closer to the
typical UE than the tagged BS. For if there were such a
BS, it would be the tagged BS associated with the typical UE. Later in the UL
model, we will see that we cannot define an exclusion zone by any BS-UE
pair.
The goal now is to derive the DL coverage probability $p_{d}$ in the
cellular network, which is exactly the complementary cumulative distribution function (CCDF) of SIR over the entire
network (Recall that the CDF gives $\mathbb{P}[$SIR$\leq \xi ]$). The
coverage probability can also be visualized as the probability that a randomly
chosen user can achieve a target SIR $\xi $, the average fraction of users
who at any time achieve SIR $\xi $, or the average fraction of network area
that is in \textquotedblleft coverage\textquotedblright \ at any time.
We
have two routes now for obtaining $p_{d}$ as far as the manipulation of RVs is
concerned. Either use RVs throughout and keep resolving them via the $%
\mathbb{E}$ operator vehicle, or condition the SIR expression at the very
beginning, using particular values for those RVs, then keep on
deconditioning. We will take the former route. At the outset, We will invoke the concept of total probability, using (\ref{id}) and (\ref{SIRd}), to get
\begin{eqnarray}
p_{d|R} &=&\mathbb{P}[\mathtt{SIR}_{\text{UE}}>\xi ] \notag \\
&=&\mathbb{P}\left[ \frac{pGR^{-\alpha }}{I_{d}}>\xi \right] \notag \\
&=&\mathbb{P}\left[ G>\frac{\xi }{p}R^{\alpha }I_{d}\right] \notag \\
\overset{(\text{a})}{=} &&\mathbb{E}%
_{I_{d}}\left[ \mathbb{P}\left[ G>\frac{\xi }{p}R^{\alpha }I_{d}\right] %
\right] \notag \\
\overset{(\text{b})}{=} &&\mathbb{E}%
_{I_{d}}\left[ e^{-\frac{\xi }{p}R^{\alpha }I_{d}}\right] \notag \\
\overset{(\text{c})}{=} &&\mathcal{L%
}_{I_{d}}(\frac{\xi }{p}R^{\alpha }) \label{First}
\end{eqnarray}%
where%
\begin{equation}
\mathcal{L}_{A}(s)=\int_{0}^{\infty }e^{-st}f_{A}(t)dt=\mathbb{E}\left[
e^{-sA}\right] \label{LT}
\end{equation}%
is the Laplace transform of the RV $I_{d}$\ conditioned on the RV $R$\ between the typical UE and the tagged BS. In (a) we utilized the
fact that we can write a probability $\mathbb{P}\left[ A>B)\right] $ as $%
\mathbb{E}_{B}\left[ \mathbb{P}\left[ A>B\right] \right] $ (or $\mathbb{E}%
_{A}\left[ \mathbb{P}\left[ A>B\right] \right] $), in (b) we benefited from
the fact that $G\sim \exp (1)$, i.e. $f_{G}(r)=e^{-r}$, and in (c) we used
the Laplace transform definition (\ref{LT}).
Clearly, the DL coverage probability $p_{d}$ in (\ref{First}) is conditioned
on $R$, the distance between the typical UE and the tagged BS. We will now embark on deconditioning $p_{d}$ (basically by obtaining its
expectation with respect to $R$.) Since $R$ is the distance between the typical UE
and the closest BS (the tagged BS), it is Rayleigh distributed, i.e. and $%
f_{R}(r)=2\lambda \pi re^{-\lambda \pi r^{2}}$ from (\ref{Reighlay}).
Further, $R$ ranges from an arbitrarily small positive real number greater
than $0$ (to exclude the typical UE) to $\infty $. Thus, the
conditional coverage probability
\begin{eqnarray}
p_{d|R} &=&\mathbb{E}_{R}\left[ p_{d|R}\right] \notag \\
&=&\mathbb{E}_{R}\left[ \mathcal{L}_{I_{d}}(\frac{\xi }{p}R^{\alpha })\right]
\notag \\
&=&\int_{0}^{\infty }\mathcal{L}_{I_{d}}(\frac{\xi }{p}r^{\alpha })f_{R}(r)dr
\notag \\
&=&2\lambda \pi \int_{0}^{\infty }e^{-\lambda \pi r^{2}}\mathcal{L}_{I_{d}}(%
\frac{\xi }{p}r^{\alpha })rdr \label{pcDecond}
\end{eqnarray}
Next, we will embark on finding the Laplace transform $\mathcal{L}_{I_{d}}$
of the DL interference $I_{d}$. Using (\ref{id}) and (\ref{LT}), we get%
\begin{eqnarray}
\mathcal{L}_{I_{d}}(s) &=&\mathbb{E}\left[ e^{-sI_{d}}\right] \notag\\
&=&\mathbb{E}_{\Phi ,G_{z}}\left[ e^{-s\sum_{z\in \Phi \backslash \{%
\mathfrak{b}\}}pG_{z}D_{z}^{-\alpha }}\right] \notag\\
&=&\mathbb{E}_{\Phi ,G_{z}}\left[ \prod \limits_{z\in \Phi \backslash \{%
\mathfrak{b}\}}e^{-spG_{z}D_{z}^{-\alpha }}\right] \notag\\
\overset{(\text{a})}{=} &&\mathbb{E}_{\Phi }\left[ \prod \limits_{z\in \Phi
\backslash \{ \mathfrak{b}\}}\mathbb{E}_{G_{z}}\left[ e^{-spG_{z}D_{z}^{-%
\alpha }}\right] \right] \notag\\
\overset{(\text{b})}{=} &&\mathbb{E}_{\Phi }\left[ \prod \limits_{z\in \Phi
\backslash \{ \mathfrak{b}\}}\mathcal{L}_{G_{z}}(spD_{z}^{-\alpha })\right]
\notag\\
\overset{(\text{c})}{=} &&\exp \left( -\lambda \int_{\mathbb{R}%
^{2}\backslash D(o,r)}(1-\mathcal{L}_{G_{z}}(spD_{z}^{-\alpha }))\right)
\label{LIds}
\end{eqnarray}%
where $D(o,r)$ is a disc centered at the origin and has a radius $r$. In (a)
we benefited from the independence of the $G_{z}$, which are iid and in (b)
we used the definition (\ref{LT}) of the Laplace transform. In (c), to
decondition on $D_{z}$ which is distributed \emph{differently} for each
point $z$ of the PPP, we invoked the PGFL $\mathbb{E}_{\Phi }\left[
\prod \limits_{z\in \Phi }f(x)\right] $, with $f(x)=\mathcal{L}%
_{G_{z}}(spx^{-\alpha })$, of the PPP $\Phi $. Note that $\mathbb{E}_{G_{z}}$
will be taken at the level of a single point $z$, whereas $\mathbb{E}_{\Phi }
$ will be taken at the level of the PPP $\Phi $. However, for the \emph{%
expectation with respect to} $\Phi $, we will apply it to $D_{z}^{-\alpha }$%
, as the latter involves all the PPP points. Note finally that some authors
call the PGFL the Laplace functional, when the former is invoked to find a
Laplace transform as is the case here.
Note that $D_{z}$ ranges from $r^{+}$ (to exclude the tagged BS) to $\infty $. That is, the integral is carried out outside the
typical circle, namely from $r$ to $\infty $. Switching to
polar coordinates, with the interferer now at $(x,\theta )\in \mathbb{R}^{2}$, then using the fact that $G_{z}\sim \exp (1)$, i.e. $f_{G_{z}}(t)=e^{-t}$, then Using (\ref{LIds}) yields
\begin{eqnarray}
\mathcal{L}_{I_{d}}(s) &=&\exp \left( -\lambda \int_{0}^{2\pi
}\int_{r}^{\infty }(1-\mathcal{L}_{G_{z}}(spx^{-\alpha }))xdxd\theta \right)
\text{ } \notag \\
&=&\exp \left( -2\pi \lambda \int_{r}^{\infty }(\frac{spx^{-\alpha }}{%
1+spx^{-\alpha }})xdx\right) \label{Lap3}
\end{eqnarray}%
For use in (\ref{pcDecond}), we write this result as%
\begin{eqnarray*}
\mathcal{L}_{I_{d}}(\frac{\xi }{p}r^{\alpha }) &=&\exp \left( -2\pi \lambda
\int_{r}^{\infty }(\frac{(\frac{\xi }{p}r^{\alpha })px^{-\alpha }}{1+(\frac{%
\xi }{p}r^{\alpha })px^{-\alpha }})xdx\right) \\
&=&\exp \left( -\pi \lambda r^{2}\xi ^{\frac{2}{\alpha }}\int_{\xi
^{-2/\alpha }}^{\infty }\frac{1}{1+u^{\alpha /2}}du\right)
\end{eqnarray*}%
where $u=\left( x/r\right) ^{2}\xi ^{-\frac{2}{\alpha }}$. Substituting this
in (\ref{pcDecond}), we get
\begin{equation}
p_{d}=2\widetilde{\lambda }\int_{0}^{\infty }e^{-\widetilde{\lambda }%
r^{2}}e^{-\widetilde{\lambda }r^{2}\sqrt[\kappa ]{\xi }\int_{\frac{1}{\sqrt[%
\kappa ]{\xi }}}^{\infty }\frac{1}{1+u^{\kappa }}du}rdr \label{pdWithLampda}
\end{equation}%
where $\widetilde{\lambda }=\lambda \pi $ and $\kappa =\alpha /2$.
From (\ref{pdWithLampda}), it \emph{appears} that the DL coverage probability $p_{d}$ is dependent on the BS density parameter $\lambda $. Indeed, this parameter exists in a plethora of coverage probability (and derivatives) expressions all over the literature on stochastic geometric modelling of DL cellular systems (See, for example,
\cite{ElSawy17, Chun20, Liu20, Kouzayha21, Bai15, Chen19, Qiong20, Ouamri20, Lei18, Fadoul20, Gao19, Ali19, Andrews16, Kundu20, Sadeghabadi20, Wang19}). However, this dependence is false, as we will show in the next
Theorem.
\begin{description}
\item \textbf{Theorem 1}: Under the stochastic geometric model of the cellular DL
system, the DL coverage probability $p_{d}$ is independent of the BS density
$\lambda $.
\end{description}
\textbf{Proof}: The proof is attained through two changes of
variables. Starting with (\ref{pdWithLampda}), use the substitution $x=r^{2}$
to get
\begin{eqnarray*}
p_{d} &=&2\widetilde{\lambda }\int_{0}^{\infty }e^{-\widetilde{\lambda }%
r^{2}}e^{-\widetilde{\lambda }r^{2}\xi ^{\frac{2}{\alpha }}\int_{\xi ^{-%
\frac{2}{\alpha }}}^{\infty }\frac{1}{1+u^{\kappa }}du}rdr \\
&=&\widetilde{\lambda }\int_{0}^{\infty }e^{-\widetilde{\lambda }x\left( 1+%
\sqrt[\kappa ]{\xi }\int_{\frac{1}{\sqrt[\kappa ]{\xi }}}^{\infty }\frac{1}{%
1+u^{\kappa }}du\right) }dx
\end{eqnarray*}%
Now use the substitution $z=\widetilde{\lambda }x$ to get
\begin{eqnarray}
p_{d} &=&\int_{0}^{\infty }e^{-z\left( 1+\sqrt[\kappa ]{\xi }\int_{\frac{1}{%
\sqrt[\kappa ]{\xi }}}^{\infty }\frac{1}{1+u^{\kappa }}du\right) }dz
\notag \\
&=&\frac{1}{1+\sqrt[\kappa ]{\xi }\int_{\frac{1}{\sqrt[\kappa ]{\xi }}%
}^{\infty }\frac{1}{1+u^{\kappa }}du}
\label{pdWithoutLampda}
\end{eqnarray}%
where $\lambda $ has totally disappeared, proving the theorem. $\blacksquare$
Theorem 1 then calls for a revisit to the DL coverage probability expressions (and their derivatives)
that include $\lambda $. Those expressions exist in a huge number of articles, such as those cited just above the Theorem. It is worth examining them for possible elimination of a parameter that may exist superfluously.
\textbf{DL Special cases}:
First, we will consider the special case of $\alpha =4$, i.e. $\kappa =2$. From (\ref{pdWithoutLampda}), we can obtain (using a computer algebra
system, e.g. Maxima (R)), a simple closed form expression for the DL coverage
probability, namely%
\begin{eqnarray}
p_{d} &=&\frac{1}{1+\sqrt{\xi }\int_{\frac{1}{\sqrt{\xi }}}^{\infty }\frac{1%
}{1+u^{2}}du} \notag\\
p_{d} &=&\frac{1}{1+\sqrt{\xi }\left( \frac{\pi }{2}-\arctan (\frac{1}{\sqrt{%
\xi }})\right) }
\label{pd4}
\end{eqnarray}%
Second, we will consider the special case of $\alpha =6$, i.e. $\kappa =3$. From (\ref{pdWithoutLampda}), we can obtain (using a computer algebra
system, e.g. Maxima (R)), a not-so-simple closed form expression for the DL coverage
probability, namely%
\begin{eqnarray}
p_{d} &=&\frac{1}{1+\sqrt[3]{\xi }\int_{\frac{1}{\sqrt[3]{\xi }}}^{\infty }%
\frac{1}{1+u^{3}}du} \notag\\
p_{d} &=&\frac{6c}{6c+\ln \left( c^{2}-c+1\right) +\sqrt{3}\left( {\pi }%
-2\arctan \left( \frac{2c-1}{\sqrt{3}}\right) \right) -2\ln \left(
c+1\right) }
\label{pd6}
\end{eqnarray}%
where $c=1/\sqrt[3]{\xi }$. To further attest to the accuracy of the last
two results, they will be plotted later together with the corresponding
simulation counterparts.
\subsection{Uplink model}
The net interference at the typical BS is the sum of the received transmissions
from all the UEs except the tagged, including those inside the typical circle. Referring to Figure \ref{fig:UL}, for each UE $\mathfrak{z}\in \Psi $, we denote its distance to its serving BS by $%
R_{\mathfrak{z}}$. Although the RVs $R_{\mathfrak{z}}$ are identically
distributed, they are not independent. However, as shown later in this
section, this dependence is weak and we will henceforth assume the $R_{\mathfrak{z}}$ iid. Under this assumption, we will first derive the coverage probability for the general distribution of $R_{\mathfrak{z}}$.
The set of interferers are the points of $\Psi $, which is not a PPP. The reason is
that they are not scattered uniformly in $R^{2}$, but rather associated to a
PPP as satellites. This poses difficulty, as we will not have the luxury of
using such useful tools as the PGFL here. However, we can get around this
difficutly as follows. Note that each point of $\Psi $ is associated, by
being closest, to a point in the PPP $\Phi $ of BSs, which we used above in
the down link analysis. More importantly, as each point in $\Psi $ is
located somewhere around the corresponding point in $\Phi $, we can
approximate the "spatial" average of the former to be the latter.
Consequently, we can approximate the locations of the interfering UEs $\Psi $
by the points of $\Phi $. Accordingly, below we will carry out the sums and
products involved in the stochastic geometry analysis over $\Phi $, rather
than $\Psi $. Specifically, for the sake of calculating the interference, we
will consider that each interfering UE is placed exactly at
its serving BS's location. Consequently, referring to Figure \ref{fig:UL},
we will employ the distance $R_{\mathfrak{z}}$ between this UE and its
serving BS to calculate its emitted power. We will then consider this as
interference at the typical BS at distance $D_{z}$ away, not $U_{\mathfrak{z}}$.
In UL, FPC leads to amplifying the transmit power $p$ at the UE based on its
distance to the serving BS. If the distance is $R$ and the FPC factor is $%
\epsilon $, with values in $[0,1]$, then $p$ is amplified by $R^{\epsilon
\alpha }$ to offset the path loss, which is $R^{-\alpha }$, where $\alpha $
is the path loss exponent, with values greater than $2$. Reducing the power
of a UE as it gets closer to its serving BS is useful for two reasons. First
it saves the UE's battery. Second, it makes the UE less of an interferer to
UEs of neighbor cells. Combining the effects of FPC, power loss and fading,
the amount of power reaching the serving BS from the UE will be $%
pGR^{-\alpha (1-\epsilon )}$.
Referring to Figure \ref{fig:UL}, the RV $R_{\mathfrak{z}}$ is upper bounded by $U_{\mathfrak{z}}$, otherwise
the sample UE at $\mathfrak{z}$ would associate with the typical BS. Moreover, as can be seen, the
typical circle does not define an exclusion zone, as UEs can be located inside of it without breaking the associatiation rule. Let $I_{u}$ denote the interference caused by all the UEs, except the tagged, at the typical BS. Accordingly,
\begin{equation}
I_{u}=\sum_{\mathfrak{z}\in \Psi \backslash \{ \mathfrak{u}\}}pG_{\mathfrak{z}%
}R_{\mathfrak{z}}^{\alpha \epsilon }U_{\mathfrak{z}}^{-\alpha }
\label{Iu}
\end{equation}%
In the UL, the SIR of the typical BS, at distance $R$
from the tagged UE, is%
\begin{equation}
\mathtt{SIR}_{\text{BS}}=\frac{pGR^{-\alpha (1-\epsilon )}}{I_{u}}
\label{SIRu}
\end{equation}%
Consequently, the probability $p_{u}$ of UL coverage is
\begin{equation*}
p_{u}=\mathbb{P}[\mathtt{SIR}_{\text{BS}}>\xi ]\text{.}
\end{equation*}%
This probability can be visualized as being the average area or the average
fraction of users in coverage. As noted earlier, we perform analysis on a
randomly chosen BS assumed to be\ located at the origin associated with the
closest UE. Under assumption 1, the distribution of the distance $R$\ of the
closest mobile from the randomly chosen BS can be assumed Rayleigh given by (%
\ref{Reighlay}). This chosen BS is placed at the origin, and is henceforth
called the typical BS.
Referring to Figure \ref{fig:UL}, both $R$ and $R_{\mathfrak{z}}$ are Rayleigh
distributed, i.e. $f_{R}(r)=f_{R_{\mathfrak{z}}}(r)=2\lambda \pi
re^{-\lambda \pi r^{2}}$. Thus, $R_{\mathfrak{z}}$ ranges from $0$ to $R$,
with $R$ ranging from an arbitrarily small positive real number to $\infty $. Now, the conditional UL coverage probability is defined as%
\begin{eqnarray*}
p_{u|R} &=&\mathbb{P}[\mathtt{SIR}_{\text{BS}}>\xi ] \\
&=&\mathbb{E}\left[ \mathbb{P}\left[ \frac{pGR^{-\alpha (1-\epsilon )}}{I_{u}%
}>\xi \right] \right] \\
&=&\mathbb{E}\left[ \mathbb{P}\left[ G>\xi p^{-1}R^{\alpha (1-\epsilon
)}I_{u}\right] \right] \\
&\overset{(\text{a})}{=}&\mathbb{E}\left[ e^{-\xi p^{-1}I_{u}R^{\alpha
(1-\epsilon )}}\right] \\
&=&\mathcal{L}_{I_{u}}(\xi p^{-1}R^{\alpha (1-\epsilon )})
\end{eqnarray*}%
where $\mathcal{L}_{I_{u}}$ is the Laplace transform of the distribution of
the $I_{u}$ RV. In (a), we used the fact that $G\sim \exp (1)$, i.e. $%
f_{G}(x)=e^{-x}$, which implies that $\mathbb{P}[G>x]=e^{-x}$. Now, we
decondition on $R$, getting%
\begin{eqnarray}
p_{u} &=&\int_{0}^{\infty }\left. \mathcal{L}_{I_{u}}(\xi p^{-1}R^{\alpha
(1-\epsilon )})\right \vert _{R=r}f_{R}(r)dr \notag \\
&=&\int_{0}^{\infty }2\widetilde{\lambda }re^{-\widetilde{\lambda }r^{2}}%
\mathcal{L}_{I_{u}}(\xi p^{-1}r^{\alpha (1-\epsilon )})dr \label{pcMAIN}
\end{eqnarray}%
where $\widetilde{\lambda }=\pi \lambda $. We integrate from a point just
outside the origin, to skip the typical BS that resides there, to $\infty $
where the closest UE can possibly exist.
Next, we will embark on finding $\mathcal{L}_{I_{u}}$, the Laplace transform
of the distribution of the RV $I_{u}$.$\mathcal{\ }$Substituting for $I_{u}$
from (\ref{Iu}), gives
\begin{eqnarray}
\mathcal{L}_{I_{u}}(s) &=&\mathbb{E}[e^{-sI_{u}}] \notag \\
&=&\mathbb{E}\left[ \exp \left( s\sum_{\mathfrak{z}\in \Psi }-pG_{\mathfrak{z%
}}R_{\mathfrak{z}}^{\alpha \epsilon }U_{\mathfrak{z}}^{-\alpha }\right) %
\right] \notag \\
&=&\mathbb{E}\left[ \prod \limits_{\mathfrak{z}\in \Psi }\exp \left( -spG_{%
\mathfrak{z}}R_{\mathfrak{z}}^{\alpha \epsilon }U_{\mathfrak{z}}^{-\alpha
}\right) \right] \label{EIz}
\end{eqnarray}%
In (\ref{EIz}), for each point $\mathfrak{z}\in \Psi $ there are three RVs: $%
G_{\mathfrak{z}},R_{\mathfrak{z}},U_{\mathfrak{z}}$. The $G_{\mathfrak{z}}$
are independent of the $R_{\mathfrak{z}}$ and of the $U_{\mathfrak{z}}$.
However, $U_{\mathfrak{z}}$ and $R_{\mathfrak{z}}$ are dependant in that $R_{%
\mathfrak{z}}<U_{\mathfrak{z}}$ (Recall that $R_{z}$ is the distance between
an interfering UE $\mathfrak{z}$ and its typical BS, and $U_{\mathfrak{z}}$
is the distance between the same interfering UE $\mathfrak{z}$ and the
typical BS at the origin). That is $\mathbb{P}[R_{\mathfrak{z}}<x|U_{%
\mathfrak{z}}=x]=1$, since if $U_{\mathfrak{z}}<R_{\mathfrak{z}}$ the
interfering UE $\mathfrak{z}$ would associate with the typical BS at the
origin.
We will next apply the expectation operator $\mathbb{E}$ to the three RVs
in (\ref{EIz}), one at a time, starting with the one that is independent of the
other two, namely $G_{\mathfrak{z}}$. Only \emph{one} of these expectations,
the one with respect to the PPP $\Psi $, will be evaluated using a PGFL and
the others using the standard definition of an expectation. Note that $\mathbb{E}_{G_{\mathfrak{z}}}$, $%
G_{\mathfrak{z}}\sim {Exp}(1)$, i.e. $f_{G_{\mathfrak{z}}}(x)=e^{-x}$.
\begin{eqnarray}
\mathcal{L}_{I_{u}}(s) &=&\mathbb{E}_{\Psi ,R_{\mathfrak{z}},G_{\mathfrak{z}%
}}\left[ \prod \limits_{\mathfrak{z}\in \Psi }e^{-spG_{\mathfrak{z}}R_{%
\mathfrak{z}}^{\alpha \epsilon }U_{\mathfrak{z}}^{-\alpha }}\right] \notag
\\
\overset{(\text{a})}{=} &&\mathbb{E}_{\Psi ,R_{\mathfrak{z}}}\left[
\prod \limits_{\mathfrak{z}\in \Psi }\mathbb{E}_{G_{\mathfrak{z}}}\left[
e^{-spG_{\mathfrak{z}}R_{\mathfrak{z}}^{\alpha \epsilon }U_{\mathfrak{z}%
}^{-\alpha }}\right] \right] \notag \\
\overset{(\text{b})}{=} &&\mathbb{E}_{\Psi ,R_{\mathfrak{z}}}\left[
\prod \limits_{\mathfrak{z}\in \Psi }\int_{0}^{\infty }e^{-\left( 1+spR_{%
\mathfrak{z}}^{\alpha \epsilon }U_{\mathfrak{z}}^{-\alpha }\right) x}dx%
\right] \notag \\
&=&\mathbb{E}_{\Psi ,R_{\mathfrak{z}}}\left[ \prod \limits_{\mathfrak{z}\in
\Psi }\frac{1}{1+spR_{\mathfrak{z}}^{\alpha \epsilon }U_{\mathfrak{z}%
}^{-\alpha }}\right] \label{LIus2}
\end{eqnarray}%
In (a) we used the fact that the $G_{\mathfrak{z}}$ are iid and in (b) we
used the fact that $f_{G_{\mathfrak{z}}}(x)=e^{-x}$.
Now, we consider the expectation with respect to $\Psi $, to uncondition on $%
U_{\mathfrak{z}}$, the distance between every point $\mathfrak{z}\in \Psi $
and the origin. We will use for this expectation a PGFL, since $U_{\mathfrak{%
z}}$\ is distributed differently for each point $\mathfrak{z}\in \Psi $.
Based on the approximation we are proposing, the points of the
non-Poisson $\Psi $ PP are now relocated to the points of the (homogeneous)
Poisson $\Phi $ PP. In particular, each point $\mathfrak{z}\in \Psi $ will
be relocated to the position of the associated point $z\in \Phi $. That is,
we will consider each interfering UE at point $z\in \Phi $ emitting power $%
pR_{\mathfrak{z}}^{\alpha \epsilon }$, but causing interference with this same power at the typical BS, at a distance $D_{z}$ based on the UE relocation. This allows us to write%
\begin{equation*}
\mathbb{E}_{\Psi }\left[ \prod \limits_{\mathfrak{z}\in \Psi }f(\mathfrak{z})%
\right] \approx \mathbb{E}_{\Phi }\left[ \prod \limits_{z\in \Phi }f(z)\right]
=e^{-\lambda \int_{\mathbb{R}^{2}}(1-f(x))}\text{.}
\end{equation*}%
Substituting for $f(y)$ from (\ref{LIus2}), converting to polar coordinates,
and substituting for the angle integral by $2\pi $, then%
\begin{eqnarray}
\mathcal{L}_{I_{u}}(s) &=&\mathbb{E}_{R_{\mathfrak{z}}}\left[ \mathbb{E}%
_{\Phi }\left[ \prod \limits_{z\in \Phi }\frac{1}{1+spR_{\mathfrak{z}%
}^{\alpha \epsilon }D_{z}^{-\alpha }}\right] \right] \notag\\
&=&\mathbb{E}_{R_{\mathfrak{z}}}\left[ e^{-2\pi \lambda \int_{0}^{\infty }%
\frac{1}{1+\left( sp\right) ^{-1}R_{\mathfrak{z}}^{-\alpha \epsilon
}x^{\alpha }}xdx}\right]
\label{LIZ(s)}
\end{eqnarray}%
The integral includes the entire Euclidean space, so $r$ goes from $0$,
"very near" the origin to avoid having a BS at the origin where interference
is assessed, to infinity. However, we approximate this nearness by $0$, not
fearing a singularity in the denominator of the integrand since there is a $1$.
Note that $R_{\mathfrak{z}}$ is lower bounded by $U_{\mathfrak{z}}$, for if $R_{\mathfrak{z}}<U_{\mathfrak{z}}$, the UE at $\mathfrak{z}$ would associate with the typical BS. But note that $U_{\mathfrak{z}}$ has been
replaced now, through PP relocation, by $D_{\mathfrak{z}}$, which after
instantiation at the deconditioning step above becomes $x$. Therefore, when
we decondition on $R_{\mathfrak{z}}$ below, the integral will be from $0$ to $x$.
We will apply the last expectation, $\mathbb{E}_{R_{\mathfrak{z}}}$, using the
property that the expectation of an integral equals the integral
of the expectation of the integrand. The integral will extend from $R=r$ as
a minimum to $\infty $. Note that the distribution
of $R_{\mathfrak{z}}$ is Rayleigh, as it is an association distance (based on the closest distance).
In light of (\ref{LIZ(s)}), using the Rayleigh
distribution $f_{R_{\mathfrak{z}}}(y)=2\lambda \pi ye^{-\lambda \pi y^{2}}$,
we have
\begin{eqnarray}
\mathcal{L}_{I_{u}}(s) &=&e^{-2\pi \lambda \int_{0}^{\infty }\left( \mathbb{E%
}_{R_{\mathfrak{z}}}\left[ \frac{1}{1+\left( sp\right) ^{-1}R_{\mathfrak{z}%
}^{-\alpha \epsilon }x^{\alpha }}\right] \right) xdx} \notag \\
&=&e^{-2\pi ^{2}\lambda ^{2}\int_{0}^{\infty }x\int_{0}^{x^{2}}\frac{%
e^{-\lambda \pi u}}{1+\left( sp\right) ^{-1}u^{-\alpha \epsilon /2}x^{\alpha
}}dudx} \label{LIus}
\end{eqnarray}%
where $u=y^{2}$. Recall that $D_{z}^{-\alpha }$ and $R_{\mathfrak{z}}$ are
dependent in that if $D_{z}=x$ then $R_{\mathfrak{z}}<x$. That is, the
distance $R_{\mathfrak{z}}$ between an interferer and the typical BS is
upper bounded by the distance $D_{z}$ between the (relocated) interfering UE
and the typical BS. From (\ref{Iu}) and (\ref{pcMAIN}), it follows that
\begin{eqnarray}
\mathcal{L}_{I_{u}}(\xi p^{-1}r^{\alpha (1-\epsilon )}) &=&e^{-2\pi
^{2}\lambda ^{2}\int_{0}^{\infty }x\int_{0}^{x^{2}}\frac{e^{-\lambda \pi u}}{%
1+\left( sp\right) ^{-1}u^{-\alpha \epsilon /2}x^{\alpha }}dudx} \notag \\
&=&e^{-2\pi ^{2}\lambda ^{2}\int_{0}^{\infty }x\int_{0}^{x^{2}}\frac{\xi
r^{\alpha (1-\epsilon )}e^{-\lambda \pi u}}{\xi r^{\alpha (1-\epsilon
)}+u^{-\alpha \epsilon /2}x^{\alpha }}dudx} \label{Liu}
\end{eqnarray}%
From (\ref{pcMAIN}) and (\ref{Liu}), we get
\begin{equation}
p_{u}=2\widetilde{\lambda }\int_{0}^{\infty }re^{-\widetilde{\lambda }%
r^{2}}e^{-2\widetilde{\lambda }^{2}\xi r^{2\kappa (1-\epsilon
)}\int_{0}^{\infty }x\int_{0}^{x^{2}}\frac{e^{-\widetilde{\lambda }u}}{\xi
r^{2\kappa (1-\epsilon )}+u^{-\epsilon \kappa }x^{2\kappa }}dudx}dr
\label{puWithLampda}
\end{equation}%
where $\widetilde{\lambda }=\pi \lambda $ and $\kappa =\alpha /2$.
From (\ref{puWithLampda}), it \emph{appears} that the UL coverage probability $p_{u}
$ is dependent on the BS density $\lambda $. Indeed, this belief is popular
research work on stochastic geometric models of the cellular UL system (See,
for example, \cite{ElSawy17, Chun20, Haroon20, Herath18, Kouzayha20, Mariam21, Andrews16, Jia19, Gao19, Ali19, Kundu20, Sadeghabadi20, Wang19}). However, we
will show in Theorem 2 below that this belief is flawed. Specifically, we will show that the presence of $\lambda $ in (\ref{puWithLampda}), as in a large number of similar
expressions in the literature of stochastic geometric modelling of cellular
networks, is superfluous.
\begin{description}
\item \textbf{Theorem 2}: Under the stochastic geometric model of the cellular UL
system, the UL coverage probability $p_{u}$ is independent of the BS density
$\lambda $.
\end{description}
\textbf{Proof}: The proof is attained through a sequence of changes of
variables. Starting with (\ref{puWithLampda}), use the substitution $v=r^{2}$
to get%
\begin{equation*}
p_{u}=\widetilde{\lambda }\int_{0}^{\infty }e^{-\widetilde{\lambda }v}e^{-2%
\widetilde{\lambda }^{2}\xi v^{\kappa (1-\epsilon )}\int_{0}^{\infty
}x\int_{0}^{x^{2}}\frac{e^{-\widetilde{\lambda }u}}{\xi v^{\kappa
(1-\epsilon )}+u^{-\epsilon \kappa }x^{2\kappa }}dudx}dv
\end{equation*}%
Use $y=x^{2}$ to get%
\begin{equation*}
p_{u}=\widetilde{\lambda }\int_{0}^{\infty }e^{-\widetilde{\lambda }v}e^{-%
\widetilde{\lambda }^{2}\xi v^{\kappa (1-\epsilon )}\int_{0}^{\infty
}\int_{0}^{y}\frac{e^{-\widetilde{\lambda }u}}{\xi v^{\kappa (1-\epsilon
)}+u^{-\epsilon \kappa }y^{\kappa }}dudy}dv
\end{equation*}%
Use $x=\widetilde{\lambda }u$ to get%
\begin{equation*}
p_{u}=\widetilde{\lambda }\int_{0}^{\infty }e^{-\widetilde{\lambda }v}e^{-%
\widetilde{\lambda }\xi v^{\kappa (1-\epsilon )}\int_{0}^{\infty }\int_{0}^{%
\widetilde{\lambda }y}\frac{e^{-x}}{\xi v^{\kappa (1-\epsilon )}+\left(
\frac{x}{\widetilde{\lambda }}\right) ^{-\epsilon \kappa }y^{\kappa }}dxdy}dv
\end{equation*}%
Use $z=\widetilde{\lambda }v$ to get
\begin{equation*}
p_{u}=\int_{0}^{\infty }e^{-z}e^{-\widetilde{\lambda }\xi z^{\kappa
(1-\epsilon )}\int_{0}^{\infty }\int_{0}^{\widetilde{\lambda }y}\frac{e^{-x}%
}{\xi z^{\kappa (1-\epsilon )}+x^{-\epsilon \kappa }\left( \widetilde{%
\lambda }y\right) ^{\kappa }}dxdy}dz
\end{equation*}%
Finally, use $u=\widetilde{\lambda }y$ to get
\begin{equation}
p_{u}=\int_{0}^{\infty }e^{-z\left( 1+\xi z^{\kappa (1-\epsilon
)-1}\int_{0}^{\infty }\int_{0}^{u}\frac{e^{-x}}{\xi z^{\kappa (1-\epsilon
)}+x^{-\epsilon \kappa }u^{\kappa }}dxdu\right) }dz
\label{puWithoutLampda}
\end{equation}%
where $\lambda $ has totally disappeared, proving the theorem. $\blacksquare$
This is the final expression for the UL coverage probability. The striking
observation about the above $p_{u}$ is that for all $\alpha $, the UL
coverage probability is independent of the density $\lambda $ of the BSs. A possible interpretation is that as the BS density gets higher, the Voronoi cells get smaller, and the UEs get closer to their serving BSs, and vice versa. This means that the increase in interference due to the increase of BS density, will be offset by an increase in signal due to the nearness of the tagged UE to the typical BS, keeping SIR, hence the coverage probability, unchanged.
Just like Theorem 1, Theorem 2 calls for a revisit to the UL coverage probability expressions (and their derivatives)
that include $\lambda $. Those expressions exist in a huge number of articles, such as those cited just above Theorem 2. It is worth examining them for possible elimination of a parameter that may exist misleadingly.
\textbf{UL Special cases}:
First, we will consider the special case of $\epsilon =0$, i.e. there is no FPC by the UE. For this case, we get from (\ref{puWithoutLampda}) that
\begin{eqnarray}
p_{u} &=&\int_{0}^{\infty }e^{-z\left( 1+\xi z^{\kappa -1}\int_{0}^{\infty
}\int_{0}^{u}\frac{e^{-x}}{\xi z^{\kappa }+u^{\kappa }}dxdu\right) }dz
\notag \\
&=&\int_{0}^{\infty }e^{-z\left( 1+\xi z^{\kappa -1}\int_{0}^{\infty }\frac{%
1-e^{-u}}{\xi z^{\kappa }+u^{\kappa }}du\right) }dz \label{puk}
\end{eqnarray}
Using a computer algebra system (e.g., Maxima (R)) or a table of integrals
(e.g., \cite{Gradshteyn15}: p. 325, \#3.241], one can find%
\begin{equation*}
\int_{0}^{\infty }\frac{1}{\xi z^{\kappa }+u^{\kappa }}du=\frac{1}{\kappa
\xi ^{1-1/\kappa }z^{\kappa -1}}B\left( 1-\frac{1}{\kappa },\frac{1}{\kappa }%
\right) ,\qquad 0<\frac{1}{\kappa }<1
\end{equation*}%
where $B(x,y)$ is the Bessel function of $x$ and $y$. Substituting in (\ref%
{puk}), we get%
\begin{equation*}
p_{u}=\int_{0}^{\infty }e^{-z\left( 1+\left( \frac{\sqrt[\kappa ]{\xi }}{%
\kappa }B\left( 1-\frac{1}{\kappa },\frac{1}{\kappa }\right) -\xi z^{\kappa
-1}\int_{0}^{\infty }\frac{e^{-u}}{\xi z^{\kappa }+u^{\kappa }}du\right)
\right) }dz
\end{equation*}
For $\alpha =4$ ($\kappa =2$), one can find
\begin{equation*}
\int_{0}^{\infty }\frac{1}{\xi y^{2}+x^{2}}dx=\frac{1}{2y\sqrt{\xi }}B\left(
0.5,0.5\right) =\frac{\pi }{2y\sqrt{\xi }}
\end{equation*}%
Substituting in (\ref{puk}), we get%
\begin{equation}
p_{u}=\int_{0}^{\infty }e^{-z\left( 1+\left( \frac{\sqrt{\xi }}{2}B\left(
\frac{1}{2},\frac{1}{2}\right) -\xi z\int_{0}^{\infty }\frac{e^{-u}}{\xi
z^{2}+u^{2}}du\right) \right) }dz \label{puk=2}
\end{equation}%
Using a computer algebra system (e.g., Maxima (R)) or a table of integrals
(e.g., \cite{Gradshteyn15}: p. 343, \#3.354], one can find the integral at
the exponent%
\begin{equation}
\int_{0}^{\infty }\frac{e^{-u}}{\xi z^{2}+u^{2}}dx=\frac{1}{\sqrt{\xi }z}%
\left[ \mathop{\rm ci}(\sqrt{\xi }z)\sin (\sqrt{\xi }z)-\mathop{\rm si}(%
\sqrt{\xi }z)\cos (\sqrt{\xi }z)\right] \label{si}
\end{equation}%
where%
\begin{equation*}
\mathop{\rm si}(x)=\mathop{\rm Si}(x)-\frac{\pi }{2}=-\int_{x}^{\infty }%
\frac{\sin t}{t}dt
\end{equation*}%
and%
\begin{equation*}
\mathop{\rm ci}(x)=\mathop{\rm Ci}(x)=-\int_{x}^{\infty }\frac{\cos t}{t}dt
\end{equation*}%
are the sine and cosine integrals, respectively. Using (\ref{si}) and the
fact that $B\left( \frac{1}{2},\frac{1}{2}\right) =\pi $ in (\ref{puk=2}),
the coverage probability for $\alpha =4$ becomes%
\begin{equation}
p_{u}=\int_{0}^{\infty }e^{-z\left( 1+\sqrt{\xi }\left( \frac{\pi }{2}-%
\mathop{\rm ci}(\sqrt{\xi }z)\sin (\sqrt{\xi }z)+\mathop{\rm si}(\sqrt{\xi }%
z)\cos (\sqrt{\xi }z)\right) \right) }dz \label{puAlpha=4}
\end{equation}
Second, we will consider the special case of $\epsilon =1$, i.e. there is channel inversion by the UE. For this case, we get from (\ref{puWithoutLampda}) that
\begin{eqnarray*}
p_{u} &=&\int_{0}^{\infty }e^{-z\left( 1+\xi z^{-1}\int_{0}^{\infty
}\int_{0}^{u}\frac{x^{\kappa }e^{-x}}{\xi x^{\kappa }+u^{\kappa }}%
dxdu\right) }dz \\
&=&\int_{0}^{\infty }e^{-z}e^{-\xi \int_{0}^{\infty }\int_{0}^{u}\frac{%
x^{\kappa }e^{-x}}{\xi x^{\kappa }+u^{\kappa }}dxdu}dz \\
&=&\exp \left( -\xi \int_{0}^{\infty }\int_{0}^{u}\frac{x^{\kappa }e^{-x}}{%
\xi x^{\kappa }+u^{\kappa }}dxdu\right)
\end{eqnarray*}
At $\kappa =2$:
\begin{equation*}
p_{u}=\exp \left( -\xi \int_{0}^{\infty }\int_{0}^{u}\frac{x^{2}e^{-x}}{\xi
x^{2}+u^{2}}dxdu\right)
\end{equation*}%
which could not be simplified further because the inner integral could not
be evaluated.
\section{Simulation}
\label{SIM}
Despite diligent search in the literature, we have not spotted
any formal stochastic geometric simulation algorithms. What is
available is only code, and even that is rare, poorly documented, and sporadic.
In this section, we provide the simulation algorithms we have developed for the present
article. Although the algorithms were primarily intended to validate our
analytical results, we feel they can be of value in
similar stochastic geometric research. Since DL and UL are different, as we
have seen in the modelling, their simulation algorithms are different, and
therefore will be introduced separately. However, we will give below some
points common to both.
Although we have proved that the BS density $\lambda $ has no bearing on the
coverage probability in either DL or UL, we have included it in the two
simulation algorithms. The reason for that is to allow the interested reader
to verify practically what we have proven theoretically. Indeed, one can use
the simulation algorithms below to ascertain that for any value of $\lambda $%
, the coverage probability remains the same for the same path loss exponent $%
\alpha $ and SIR threshold $\xi $. Actually, it was experimentation of this
sort that led us to discover the irrelevance of $\lambda $ and motivated the
theoretical proof presented earlier.
\subsection{Downlink}
The modelling of the DL mode was undoubtedly simple. Its simulation is just
as simple. After scattering the BSs of a realization in the simulation
window, we select the closest BS to the origin and make it the tagged BS,
considering that the typical UE is at the origin already. That is, we do not
need to generate any UEs. The transmission received at the origin from the
tagged BS will be then considered signal and every other transmission
received there considered interference. It is a simple matter then to
calculate the SIR at the typical UE and use it to determine if the typical
UE is covered by the tagged BS for this realization, for the given $\xi $.
The coverage probability is evaluated after all $\mathcal{N}$ realizations
have been generated, by dividing the number of realizations where the
typical UE was covered by the total number $\mathcal{N}$ of realizations.
The algorithm then proceeds to calculate the SIR at the typical UE, using (%
\ref{SIRd}).
\subsection{Uplink}
UL simulation is more involved than DL simulation, as was the case with
their modelling. A key element for a successful UL simulation is the
association table, $A$, an example of which is shown in Table \ref{table:A}.
For a network of $N$ BSs, this $N\times 2$ table has one column for the BSs
and another for the UEs. In particular, the table stores the coordinates of
each BS and those of its associated UE (or conversely, each UE and its
serving BS) in one row. As such, the table not only tells which UE is
associated to which BS, but also allows calculating the distance between any
BS or UE and any other BS or UE. These distances are crucial to calculate
the interference at a given receiver. For example, referring to Table \ref%
{table:A}, the first BS is at the origin and its associated UE is $734.2$ m
west and $628.4$ m north, i.e. the BS and UE are $\sqrt{734.2^{2}+628.4^{2}}%
=966.4$ m apart. We note that the third BS has no associated UE yet.
\begin{table}[H]
\caption{Example BS-UE Association table $A$ for UL simulation.}
\centering
\begin{tabular}{c c }
\hline\hline
BS & UE\\
\hline
(0,0) & (-734.2,628.4)\\
(1243.2,-221.4) & (973.2,1628.4)\\
(-345.2,928.4) & NULL \\
$\ldots$ & $\ldots$\\
\hline
\end{tabular}
\label{table:A}
\end{table}
The association table $A$ is built by the simulation algorithm as follows.
First, the BS column is filled at once by generating $N$ pairs of random
numbers (drawn from a uniform distribution) in the interval $[-S,S]$, where $%
S$ is the simulation window (square) side. The first number in the pair is
taken as the $x$ coordinate, and the second as the $y$ coordinate. Then the
UEs are added incrementally, from top to bottom, one at a time based on the
BS-UE association rule, which says:\ a UE must be closer to its serving BS
than to any other BS in the cellular network. Adding the first UE to the
table is simplest. A pair of random numbers is first generated (drawn from a
uniform distribution) in the interval $[-S,S]$ to determine the location of
the UE in the simulation window. Then the distances between this UE and all
the BSs in the network, whose locations are in column $1$ of the table, are
calculated. The UE is assigned immediately to the closest BS.
\begin{algorithm
\SetKwInOut{Input}
\SetKwInOut{\textbf{Input}: $\mathcal{N}$, $\xi$, $p$, $\lambda$, $S$, $\alpha$}\\
\SetKwInOut{Output}
\SetKwInOut{\textbf{Output}: $p_d$}\\
Covered := 0\\
\For{$i:= 1$ \textnormal{to} $\mathcal{N}$}
{
\tcc{Find number $N$ of BSs to deploy in simulation window:}
Generate a Possion distributed random number $N$, $N \sim$ Pois($\lambda S^2$) \\
\tcc{Scatter $N$ BSs over an $S \times S$ square window, storing the location at a 1 column array $D$:}
\For{$i: = 1$ \textnormal{to} $N$} {
Generate two uniformly distributed random numbers $(x,y)$, where $x,y \sim$ U($-S,S$), for BS $i$ location.\\
$D(i)=(x,y)$
}
Identify BS $n$ nearest the origin and designate it tagged BS\\
\tcc{Find SIR at typical UE and check if the latter is covered by tagged BS for the given $\xi$:}
Generate a Poisson random number $G$, where $G \sim$ Pois(1) \\
Calculate distance $R$ between the tagged BS and origin\\
NUM $= p G R^{-\alpha}$ \\
DENUM $= 0$ \\
\For {$i:=1$ \textnormal{to} $N$}
{
\If{$i \ne n$}
{
Generate a Poisson random number $G_i$, where $G_i \sim$ Pois(1) \\
Use array $D$ to calculate the distance $D_i$ between BS $i$ and origin\\
DENUM $=$ DENUM $+ p G_i {D_i}^{-\alpha}$\\
}
}
SIR := NUM/DENUM \\
\If{\textnormal{SIR} $> \xi$}
{
Covered := Covered + 1
}
}
\tcc{Calculate DL coverage probability $p_d$ for this threshold $\xi$:}
$p_d$ := Covered $/ \mathcal{N}$
\caption{Simulation of cellular DL system}
\end{algorithm}
\label{alg:DL}
\hspace{2cm}
\begin{algorithm
\SetKwInOut{Input}
\SetKwInOut{\textbf{Input}: $\mathcal{N}$, $\xi$, $p$, $\lambda$, $S$, $\alpha$, $\epsilon$}\\
\SetKwInOut{Output}
\SetKwInOut{\textbf{Output}: $p_u$}\\
Covered := 0\\
\For{$i:= 1$ \textnormal{to} $\mathcal{N}$}
{
\tcc{Find number $N$ of BSs to deploy in simulation window:}
Generate a Possion distributed random number $N$, $N \sim$ Pois($\lambda S^2$) \\
\tcc{Build table $A$ for a network of $N$ BSs and $N$ UEs:}
\For{$i: = 1$ \textnormal{to} $N$} {
$A(i,1)=A(i,2)=$ NULL
\tcc{Empty the association table, $A$}
}
\tcc{1- Fill in column 1 of $A$ with BSs, making BS 1 typical.}
$A(1,1)=(0,0)$
\For{$i: = 2$ \textnormal{to} $N$} {
Generate random location $(x,y)$, where $x,y \sim$ U($-S,S$), for BS $i$.\\
$A(i,1)=(x,y)$
}
\tcc{2- Fill in column 2 of table $A$ with UEs.}
\While{\textnormal{($\exists$ \textnormal{NULL} $\in A$)}}
{
Generate random location $(u,v)$, where $u,v \sim$ U($-S,S$), for a UE.\\
Calculate the $N$ distances between the UE and every BS $i$\\
Identify BS $n$, the BS nearest the UE\\
\If{\textnormal{$A(n,2)=$\textnormal{NULL}}}
{
$A(n,2)=(u,v)$
\tcc{Assocate UE with BS $n$ if available.}
}
}
\tcc{Find SIR at typical BS, update `Covered' \& calculate $p_u$:}
Generate a Poisson random number $G$, where $G \sim$ Pois(1) \\
Calculate from table $A$ distance $R$ between typical BS and tagged UE\\
NUM $= p G R^{-\alpha(1-\epsilon)}$ \\
DENUM $= 0$ \\
\For {$i:=2$ \textnormal{to} $N$}{
Generate a Poisson random number $G_i$, where $G_i \sim$ Pois(1) \\
Calculate from table $A$ distance $R_i$ between UE $i$ and serving BS\\
Calc. from table $A$ distance $U_i$ between UE $i$ and typical BS\\
DENUM $=$ DENUM $+ p G_i {R_i}^{\alpha \epsilon} {U_i}^{-\alpha}$\\
}
SIR := NUM/DENUM \\
\If{\textnormal{SIR} $> \xi$}
{
Covered := Covered + 1
}
}
$p_u$ := Covered $/ \mathcal{N}$
\caption{Simulation of cellular UL system}
\label{alg:UL}
\end{algorithm}
The addition of the second UE gets a little harder. First, we generate the
location and calculate the distance as we did for the first UE. However, now
there is a possibility that the second UE is closest to the BS to which the
first UE was associated. Here the attempt fails. We keep making attempts
till we find a nearest BS that is available, to which we make the
association by recording the coordinates of the UE in the same row of that
nearest BS. Of course, adding UEs keeps getting harder and harder as the
number of associations increases since the chance of finding a BS that is
both closest to the UE and at the same time available becomes increasingly
small. For this probabilistic behavior, building the association table is
the most time consuming part of the algorithm. All the other parts are just
fast computations based on distances obtained from the numbers of this table.
The algorithm then proceeds to calculate the SIR at the typical BS, using (%
\ref{SIRu}). Note that in the simulation we can calculate exactly the two
distances: $R_{\mathfrak{z}}$, the distance between the interfering UE and
its serving BS, and $U_{\mathfrak{z}}$, the distance between the interfering
UE and the typical BS. That is, we can do in an exact way in the simulation,
what we did in an approximate way in the modelling. Interestingly, however,
the results of both the modelling and simulation conform impressively,
testifying to the validity of the approximation made in the modelling.
Assessing the time complexity of both algorithms is difficult since the
number $N$ of BSs, i.e. PPP points, generated in each realization is not
constant, but a Poisson distributed RV ranging from $0$ to $\infty $. This
difficulty increases even more in the UL case, where the way the association
table $A$ is built is also random, as explained earlier.
\section{Experimental Work}
\label{EW}
We implemented the simulation algorithms in the Matlab (R) language as two
separate programs, one for DL and the other for UL. We coded the algorithms
in Matlab (R). The simulation time took around 1 hr on a Laptop of $8$ GB
RAM and Core $5$ CPU at $2.4$ GHz, for $3000$ runs, on a simulation window
in the form of a square of side $2000$ m. Matlab (R) was also used to
perform the computations of the analytical results, and proved particularly
efficient with numerical integration of formidable functions.
An important implementation tip is in order. The pseudo code shown in the two
simulation algorithms calculates the coverage probability for only one value
of the threshold, $\xi $. To sketch a smooth curve over a reasonable range
of the threshold, such as that of the coverage probability figures below, where the range is from $-15$ dB to $15$ dB, we should
then run the algorithm $\mathcal{N}$ times, with $\mathcal{N}$ typically
exceeding $3000$ for good convergence, for each of the $31$ dB values. To
save simulation time, however, we did something interesting that yielded the
same results nonetheless. In each realization, we tested the resulting SIR
with all the values of $\xi $ in the range, starting from $-15$ dB and going
upwards, incrementing the variable `Covered' each time the former exceeds
the latter for this particular value of $\xi $, for which an array, rather
than a single variable, is created. Once the former stops exceeding the
latter, we quit the comparisons for this realization. This trick reduces the
simulation time significantly, yet produces the same results. The same
thing, by the way, applies in UL when using more than one value of the power
control factor, $\epsilon $. Instead of running $\mathcal{N}$ realizations
for each $\epsilon $, we use the same realization to get results for as many
$\epsilon $ values as desired. Of course, one would use arrays for these
multiple results, instead of the variables in the algorithms below which are
intended for only one $\xi $ value and one $\epsilon $ value (in case of
UL.) We always calculate the SIR at the typical receiver placed at the
origin (which is a UE in DL and a BS in UL.)
From our experience, there are two tricks that can make the running of the
algorithms extremely fast. First, we use the same realization to check for
all required thresholds. For example, in our experiments we use 31
thresholds, from $-15$ dB to $15$ dB, in increments of $1$ dB. Then, when we
obtain the SIR for a realization, we keep checking (using a loop) it against
the thresholds from the smaller going towards the bigger, within the same
realization. Once the SIR fails to exceed a threshold, we exit (or break)
the loop, since there is no point in checking the SIR against greater
thresholds. Moreover, in the UL system in particular, we use also the same
realization for all the values of the power control factor $\epsilon $ under
consideration. In some implementations we have seen, there is one
realization per threshold value per control factor value, totalling $93$
realizations if there are $31$ threshold values and $3$ power factor values.
These $93$ realizations are replaced by only $1$ realization, a huge savings
in terms of simulation time.
Furthermore, in our UL algorithm we used a trick in building the association
table that accelerated the simulation immensely. Unlike popular
implementations which insist on building the association table strictly from
top down, starting by associating the first BS, then the second BS, and so
on, we associate with the BS that turns out to be closest to the UE just
generated, provided the BS is available. We note in passing that we start in
UL by placing one BS at the origin, designating it to be the typical BS.
This BS can in fact be at any row of the association table $A$, so it might
as well be at the first row, for easy reference.
\begin{figure}[H]
\begin{center}
\begin{tikzpicture}[scale=1.1]
\begin{axis}[
xlabel={\scriptsize{SIR threshold, $\xi$ (dB)}} ,ylabel={\scriptsize{DL coverage probability, $p_d$ }},
height=11cm,
width=12cm,
grid=major,
axis lines=left,
ymax=1,
ymin=0,
y tick label style={
font=\scriptsize},
xmin=-15,
xmax=15,
x tick label style={
font=\scriptsize},
xscale=1,
legend pos=north east,
]
\addplot [no marks,mark= *, color=black] coordinates {
(-15,0.96965)
(-14,0.96219)
(-13,0.95301)
(-12,0.94178)
(-11,0.92814)
(-10,0.9117)
(-9,0.89208)
(-8,0.86894)
(-7,0.842)
(-6,0.81113)
(-5,0.77636)
(-4,0.73793)
(-3,0.69632)
(-2,0.65223)
(-1,0.6065 )
(0,0.5601)
(1,0.51396)
(2,0.46895)
(3,0.4257)
(4, 0.38499)
(5,0.34694)
(6,0.3118)
(7,0.27963)
(8,0.25038)
(9,0.22391)
(10,0.20005)
(11,0.17861)
(12,0.15939)
(13,0.14219)
(14,0.12681)
(15,0.11308)
};
\addlegendentry{\scriptsize{Analaytic}}
\addplot [only marks,mark size=2pt, color=red] coordinates {
(-15, 0.971)
(-10,0.9139)
(-5,0.7868)
(0,0.5664)
(5,0.3495)
(10,0.206)
(15,0.1133)
};
\addlegendentry{\scriptsize{Simulation}}
\addplot [no marks,mark= *, color=black] coordinates {
(-15,0.98462)
(-14,0.98078)
(-13,0.97602)
(-12,0.97014)
(-11,0.96291)
(-10, 0.95409)
(-9,0.94339)
(-8,0.93052)
(-7,0.91519)
(-6,0.89714 )
(-5,0.87617)
(-4,0.85215)
(-3,0.82511)
(-2,0.79519)
(-1,0.76268)
(0,0.72804)
(1,0.6918)
(2,0.65457)
(3,0.61697)
(4,0.57955)
(5, 0.54283)
(6, 0.50721)
(7,0.47299)
(8,0.44037)
(9,0.40949)
(10,0.3804)
(11,0.35311)
(12,0.32759)
(13,0.30378)
(14,0.28161)
(15,0.26099)
};
\addplot [only marks,mark size=2pt, color=red] coordinates {
(-15, 0.9849)
(-10,0.9579)
(-5,0.8798)
(0,0.7304)
(5,0.5488)
(10,0.3913)
(15,0.2662)
};
\node [right] at (axis cs: 4, 0.6) {{{\scriptsize{\color{black} $\alpha=6$ ($\kappa=3$)}}}};
\node [right] at (axis cs: -2.5, 0.4) {{\scriptsize{\color{black} $\alpha=4$ ($\kappa=2$)}}};
\end{axis}
\end{tikzpicture}
\caption[state,]{DL coverage probability, $p_d$, for two values of $\alpha$}
\label{fig:DLP}
\end{center}
\end{figure}
\begin{figure}[H]
\centering
\begin{center}
\begin{tikzpicture}[scale=1.1]
\begin{axis}[
xlabel={\scriptsize{SIR threshold, $\xi$ (dB)}} ,ylabel={\scriptsize{UL coverage probability, $p_u$ }},
height=11cm,
width=12cm,
grid=major,
axis lines=left,
ymax=1,
ymin=0,
y tick label style={
font=\scriptsize},
xmin=-15,
xmax=15,
x tick label style={
font=\scriptsize},
xscale=1,
legend pos=north east,
]
\addplot [no marks,mark= *, color=black] coordinates {
(-15,0.89563)
(-14,0.88447)
(-13,0.87243)
(-12,0.85951)
(-11,0.84569)
(-10,0.83097)
(-9,0.81535)
(-8,0.79886)
(-7,0.78152)
(-6,0.76337)
(-5,0.74444)
(-4,0.72481)
(-3,0.70451)
(-2,0.68363)
(-1,0.66225)
(0,0.64043)
(1,0.61828)
(2,0.59587)
(3,0.57329)
(4,0.55064)
(5,0.52801)
(6,0.50547)
(7,0.48312)
(8,0.46103)
(9,0.43928)
(10,0.41792)
(11,0.39703)
(12,0.37665)
(13,0.35683)
(14,0.33761)
(15,0.31903)
};
\addlegendentry{\scriptsize{Analaytic}}
\addplot [only marks,mark size=2pt, color=red] coordinates {
(-15, 0.893)
(-10,0.833)
(-5,0.748)
(0,0.6416)
(5,0.5226)
(10,0.4105)
(15,0.311)
};
\addlegendentry{\scriptsize{Simulation}}
\addplot [no marks,mark= *, color=black] coordinates {
(-15,0.91528)
(-14, 0.9017)
(-13,0.88642)
(-12,0.86937)
(-11,0.85047)
(-10, 0.8297)
(-9,0.80705)
(-8,0.78257)
(-7,0.75634)
(-6,0.72848)
(-5,0.69915)
(-4,0.66856)
(-3,0.63694)
(-2,0.60455)
(-1,0.57165)
(0,0.53853)
(1,0.50547)
(2,0.47273)
(3,0.44057)
(4,0.40922)
(5, 0.37886)
(6, 0.34968)
(7,0.32179)
(8,0.29531)
(9,0.27031)
(10,0.24682)
(11,0.22486)
(12,0.20442)
(13,0.18547)
(14,0.16797)
(15,0.15187)
};
\addplot [only marks,mark size=2pt, color=red] coordinates {
(-15, 0.9203)
(-10,0.8428)
(-5,0.7135)
(0,0.5493)
(5,0.381)
(10,0.2465)
(15,0.1508)
};
\node [right] at (axis cs: 3, 0.6) {{\scriptsize{\color{black} $\alpha=6$ ($\kappa=3$)}}};
\node [right] at (axis cs: -2, 0.4) {{\scriptsize{\color{black} $\alpha=4$ ($\kappa=2$)}}};
\end{axis}
\end{tikzpicture}
\caption[state]{UL coverage probability, $p_u$, for two values of $\alpha$, and with no FPC, i.e. $\epsilon=0$. }
\label{fig:ULP}
\end{center}
\end{figure}
We sketch in Figures \ref{fig:DLP} and \ref{fig:ULP} the analytical and
simulation values of the DL and UL coverage probabilities, respectively, for
a range of a threshold, $\xi $, extending from $-15$ dB to $15$ dB. For Figure \ref{fig:DLP}, we used (\ref{pd4}) and (\ref{pd6}) to sketch the analytical results for the cases of $\alpha=4$ and $\alpha=6$, respectively, and Aglorithm 1 to sketch the simulation results. On the other hand, for Figure \ref{fig:ULP}, we used (\ref{puWithoutLampda}) and (\ref{puAlpha=4}) to sketch the analytical results for the cases of $\alpha=6$ and $\alpha=4$, respectively, and Algorithm 2 to sketch the simulation results. The UL coverage probability is sketched for the case of no FPC, i.e. $\epsilon=0$.
As can
be seen the agreement between the two types of values is excellent. From the
actual values we used for graphig the curves, the difference between the
analytical and simulation results is within only $1\%$, out of as few as $%
3000$ simulation runs (realizations) on a $2000\times 2000$ m simulation
window.
\section{Conclusions}
\label{CONC}
In this article we have investigated thoroughly the modelling and simulation
of cellular DL and UL cellular channels using stochastic geometry. A number
of contributions have been made in the process. First, we introduced the DL
and UL models concisely together so as to expose their similarities and
differences vividly. In addition, we were able to derive closed form
expressions for the DL coverage probability in two special cases. Then, for
UL, known to be challenging, we introduced an approximation that
circumvented the challenge and yet yielded excellent results validated later
by simulation. For UL also, we obtained for the coverage probability an
expression, though not closed form, but simple enough to calculate easily.
For the simulation dimension of our study, we presented two effecient Monte
Carlo simulation algorithms designed to validate the models, but can be
useful to anyone using stochastic geometric in the communications field or
otherwise. Finally, we have proved two theorems that go against established
belief in the cellular modelling community. Namely, we have proved that under the
stochastic geometric model, the coverage probability in either DL and UL is
independent of the BS density. This finding calls for a revisit to a large
body of results published in the past decade with the BS density present superfluously in them.
|
\section{Introduction}
\label{sec:intro}
Although superficially similar, finding journeys in public transit networks is very different from finding routes in road networks.
This is evident from the difficulties in adapting algorithms for road networks to public transit networks~\cite{Berger2009}.
Two of the main reasons for this are time-dependency and multiple criteria.
Arguable the most common criteria when planning routes in both road and public transit networks is travel time.
For road networks, this is often sufficient, but not so for public transit networks.
While multiple criteria are not a requirement for public transit routing algorithms, there is one criteria besides travel time that many travellers care about:
The number of required transfers.
This is because changing between trains is inconvenient, especially when travelling with luggage.
Additionally, there's the risk of missing a connection when the earlier train is delayed.
Therefore, most algorithms for public transit routing are multi-criteria in at least travel time and the number of transfers, and return a set of Pareto-optimal results.
The user can then choose between faster or more convenient travel.
Although multi-criteria Pareto optimization is NP-hard in general, this particular combination of criteria is efficiently tractable in public transit networks~\cite{Muller-Hannemann2006}.
The more important aspect of public transit networks is the inherent time-dependency.
While road networks can be time-dependent when one considers traffic data or road closures, it is one of the core properties of public transit networks.
Travel is only possible at certain times; in more remote areas, travel may even be impossible during the night or on certain days.
This time-dependency has an important implication:
An algorithm can only answer queries that fall within the time period for which it has data.
Many state-of-the-art algorithms require preprocessing to enable fast queries.
If this preprocessing is too costly, it may not be feasible to process a long time period, which means users may not be able to plan journeys in advance.
\subsection{Related Work}
\label{sec:related}
While many public transit routing algorithms have been proposed, especially in recent years, to the best of our knowledge, none of them explicitly tackle the problem of long (in the time sense) timetables.
For an extensive survey of both classic and modern routing algorithms, please refer to Bast et~al.~\cite{Bast2014a}.
Pyrga et~al.~\cite{Pyrga2008} solve the problem of finding optimal journeys by finding shortest paths in graphs.
They introduce the time-extended and time-dependent model as well as speed-up techniques such as goal directed search.
In general, however, running variants of Dijkstra's algorithm on these graph structures is outperformed by current state-of-the-art techniques.
Berger et~al.~\cite{Berger2010} propose SUBITO and k-flags, two speed-up techniques that Pareto-optimize both travel time and number of transfers.
Their techniques allow multi-criteria queries with dynamic update, but have the restriction that they cannot add a new arc to the network.
As they note, this is not critical in train networks, but if one also considers buses (especially as rail replacements), new arcs are much more likely to appear in updates.
Their timetable seems to span a single day.
RAPTOR~\cite{Delling2012} does not explicitly model the data as a graph, instead operating directly on the timetable data.
Their experiments cover the usual single day, although their model suggests that it may be able to be extended to longer periods.
The Connection Scan Algorithm (CSA)~\cite{Dibbelt2018} also foregoes graphs and instead works on an ordered array of connections.
This simple model makes it easy to adjust to changes in the timetable, but it is not obvious if this still holds for large timetables.
In addition, CSA's performance relies on efficient use of CPU caches, and cache efficiency is expected to degrade with increasing amounts of data.
Public Transit Labelling (PTL)~\cite{Delling2015} uses a hub labelling approach.
An existing hub labelling algorithm is used to compute departure and arrival labels for each stop, which can, at query time, be used to reconstruct optimal journeys.
Those labels are then optimized by exploiting the properties of public transit networks and queries.
PTL achieves very low query times, even for multi-criteria queries, at the cost of extensive preprocessing and high memory footprint.
Transfer Patterns (TP)~\cite{Bast2010,Bast2016,Bast2014} is a speed-up technique that precomputes transfer patterns between all stops in the network.
Transfer patterns are defined as the sequence of stops where passengers transfer between vehicles.
At query time, these patterns are applied to efficiently find all Pareto-optimal journeys.
Similar to this work, TP uses bit sets to indicate traffic days for trips, although it is not clear from the original publication how many days are considered.
[Bast2014] specifies that only single day and seven day periods are evaluated.
\subsection{Our Contribution}
\label{sec:contribution}
In this work, we present how the trip-based public transit routing algorithm first introduced in~\cite{Witt2015} can be modified to be able to efficiently handle time periods of arbitrary length.
Section~\ref{sec:prelim} introduces necessary definitions, before Section~\ref{sec:basic} gives a high-level introduction to the basic trip-based algorithm.
Section~\ref{sec:extend} proposes changes to the algorithm for efficient handling of large time periods,
with Section~\ref{sec:updating} describing how the data can be updated without complete re-computation when changes to the timetable occur.
Finally, Section~\ref{sec:experiment} contains the experimental evaluation of the proposed changes.
\section{Preliminaries}
\label{sec:prelim}
This section introduces required definitions and notation used in the remainder of this work.
We define a public transit network via an aperiodic \textbf{timetable}, which consists of a set of stops, a set of footpaths, and a set of trips.
A \textbf{stop} is a location where passengers can enter or exit a vehicle, such as a bus or train.
A stop may represent anything between an entire train station and a single platform, or even part of a platform, depending on the granularity of the model.
Less fine-grained models often assign a \textbf{minimum change time} $\chTime{\cdot}$ to each stop.
This is the minimum time that must pass between the arrival of one vehicle and the departure of another for passengers to be able to transfer between them.
This tries to account for the time people need to go from their arrival platform to the departure platform.
It can also serve as buffer time to prevent minor delays from making a journey unfeasible.
A \textbf{footpath} connects two stops.
A footpath indicates that passengers are expected to travel between these two stops on foot.
This may be between two nearby stations, or -- on more fine-grained models -- between two platforms at the same station.
We use the most general model of directed, non-transitive footpaths, which means that walking is only considered between stops directly linked via a footpath.
The time required to walk from stop $s_1$ to stop $s_2$ is denoted by $\fpTime{s_1}{s_2}$.
If no footpath between $s_1$ and $s_2$ exists, we define $\fpTime{s_1}{s_2} := \infty$, for ease of definitions.
A \textbf{trip} represents a vehicle, usually a bus, train, or similar.
Each trip travels at a specific time along a sequence of stops, denoted by $\left<\nthStopOn{t}{1},\nthStopOn{t}{2},\ldots,\nthStopOn{t}{n}\right>$.
For each $\nthStopOn{t}{i}$, the timetable contains the arrival time $\arrTime{t}{i}$ and the departure time $\depTime{t}{i}$ of trip $t$ at that stop.
We define $\arrTime{t}{1} := 0$ and $\depTime{t}{n} := \infty$.
Note that a trip may visit the same stop more than once, such as on circular routes.
Trips are partitioned into \textbf{routes}, such that all trips belonging to the same route visit the same sequence of stops and none overtakes any other
-- no trip departs later at one stop and arrives earlier at a later stop than any other trip of the same route.
More specifically, we require that all trips of a route can be totally ordered with respect to
\begin{equation}
t_1 \preceq t_2 \iff \forall i \in \mathopen{}\mathclose\bgroup\left[1, n\aftergroup\egroup\right]:
\arrTime{t_1}{i} \leq \arrTime{t_2}{i} \land
\depTime{t_1}{i} \leq \depTime{t_2}{i} \text{.}
\end{equation}
Trips that violate this requirement are assigned to different routes, even though they share the same stop sequence.
This ``no overtaking'' rule ensures that it is never beneficial to wait for a later trip of the same route.
Note that these routes do not necessarily correspond to any e.g.\ subway lines as defined by the transport provider, they are simply a way to group trips in a way useful for route planning.
A \textbf{transfer} is a connection between two trips.
Whenever passengers can exit one trip and board another, subject to the constraints given by footpaths and minimum change times, we say that a transfer exists.
This means that for each transfer $\transfer{t_1}{e}{t_2}{b}$,
\begin{equation}
\arrTime{t_1}{e} + \chTime{\nthStopOn{t_1}{e}} \leq \depTime{t_2}{b}\label{eqn:transfer_at_stop}
\end{equation}
must hold; if the transfer involves a footpath, the requirement is instead
\begin{equation}
\arrTime{t_1}{e} + \fpTime{\nthStopOn{t_1}{e}}{\nthStopOn{t_2}{b}} \leq \depTime{t_2}{b} \text{.}\label{eqn:transfer_via_footpath}
\end{equation}
A \textbf{journey} is, essentially, a sequence of trips and interspersed footpaths.
The result of a query is one or more journeys telling the user how to get from their source stop to their destination.
\section{Basic Trip-Based Public Transit Routing}
\label{sec:basic}
This section describes the basic trip-based public transit routing algorithm first introduced in \cite{Witt2015}.
The core idea of the algorithm is to put the focus on the fundamental building blocks of public transit networks, the trips.
This is most apparent during the query phase:
Whereas other algorithms track progress by recording which stops can be reached, the trip-based algorithms records which \emph{trips} can be reached.
\subsection{Preprocessing}
Before queries can be performed, a short preprocessing phase is required.
The preprocessing can itself be divided into two steps: Transfer computation and transfer reduction.
During transfer computation, possible transfers between pairs of trips are identified.
The process is fairly straightforward: For each trip $t$, we iterate over all its stops.
At each stop $\nthStopOn{t}{i}$ we find, for each route $r$ visiting that stop, the first trip $u$ that can be reached from our current trip, i.e., the earliest trip such that equation \ref{eqn:transfer_at_stop} is satisfied.
If there is such a trip, we record the existence of a transfer $\transfer{t}{i}{u}{j}$, where $j$ is the index of that stop in the stop sequence of $u$.
We then repeat this for each stop reachable via footpath from $\nthStopOn{t}{i}$, except that equation \ref{eqn:transfer_via_footpath} must be satisfied.
Since each trip is processed independently, this is trivially parallelizable.
There are two main advantages to this reification of transfers.
Firstly, since all possible transfers are identified during preprocessing, the query algorithm does not need to evaluate minimum change times or examine footpaths.
Secondly, this allows arbitrarily complex transfer constraints.
For example, it may be known that two routes always stop at neighbouring platforms, so the minimum change time at this stop may be lower for transfers between trips of these specific routes.
Any constraints (or relaxations) that affect the existence of a transfer is dealt with during this preprocessing step, and has therefore no further performance cost at query time.
The second, optional preprocessing step is transfer reduction.
In this step, transfers which do not lead to non-dominated journeys are removed from the previously computed set.
This significantly reduces the number of transfers, which reduces memory usage and improves query performance.
As previously, we process each trip independently, allowing trivial parallelization.
For each trip, we keep track of reached stops together with arrival time, as well as reached trips together with the index in their stop sequence where we reach them.
Stops are processed in reverse order, starting with the last stop visited by the trip.
At each stop $\nthStopOn{t}{e}$, we evaluate each transfer $\transfer{t}{e}{\cdot}{\cdot}$.
Transfers which improve arrival times or allow us to reach an earlier trip, or an earlier stop on a trip, are kept.
All other transfers are removed.
\subsection{Query}
The query algorithm is reminiscent of a breadth-first search, with trips as graph nodes and transfers as edges between them.
First, we identify routes that visit the destination stop or stops connected to the destination stop via a footpath.
These are the destination routes and are recorded, together with the stop index where passengers should exit and the length of the footpath, if applicable.
For an earliest-arrival query, we then find, for each route at the source stop or a stop reachable via footpath from the source stop, the first reachable trip.
These trips, and the stop indices where they are reached, are put into a queue.
The entries $(t, i)$ of this queue are then processed as follows.
First, if we already found a journey, we might be able to prune this entry.
To do so, we compare the previously found arrival time at the destination stop with $\arrTime{t}{i+1}$.
If this entry's time is later, it clearly cannot lead to an improved journey, and can therefore be skipped.
Otherwise, we check if this entry's trip belongs to one of the destination routes and reaches the destination.
If so, we update the arrival time at the destination stop using the recorded data for this route.
Next, we examine all transfers $\transfer{t}{j}{\cdot}{\cdot}$, with $j > i$.
If a transfer leads to an earlier trip of a route or an earlier stop on a trip than previously reached, a new entry is added to the queue.
To determine this, we keep -- for each route -- track of the entries that were previously added to the queue.
Once the queue is empty, we report the non-dominated journeys found.
A profile or range query uses the procedure described above as a subroutine.
We start by finding all reachable trips in the given departure time interval and ordering them by departure time.
We then process them in reverse order, starting with the latest possible departure.
For each distinct departure time, we perform an earliest-arrival query as described above.
However, we do not reset the tracking data structure between each of these.
This means that each (earlier) departure will not re-explore the paths taken by later departures, which would only lead to dominated journeys.
For this to work, we require only small changes to the algorithm described above:
We now also track the number of transfers required to reach a trip, and also compare against that when determining whether a transfer leads to a new entry.
In addition, the pruning rule now also takes the number of transfers into account, because a journey with a later arrival time may still be optimal if it requires fewer transfers.
\section{Extending the Time Horizon}
\label{sec:extend}
Although in theory, the basic trip-based algorithm can be applied to timetables of arbitrary time intervals, it is not efficient to do so.
The reason for this is that while timetables are aperiodic, they usually do have some regularity:
Many trips operate at the same time on most days, with exceptions for weekends, holidays, etc.
Treating each instance of this trip as independent is correct, but not optimal.
This section introduces enhancements to the trip-based algorithm to allow it to efficiently handle large time intervals, such as timetables spanning an entire year.
\subsection{Changes to the Model}
\label{sec:extend_model}
The underlying model of the timetable remains largely unchanged, except for one crucial addition:
Each trip now also has an associated bit set that indicates on which days the trip is active,
i.e., a $1$ at the $i$-th position of the set means that the trip operates on the $i$-th day of whatever time interval the timetable covers.
\subsection{Preprocessing}
\label{sec:extend_preprocessing}
Preprocessing is extended to account for the new bit sets.
For transfer computation, this simply means that transfers are only allowed when both trips have at least one active day in common.
One consequence of this is that a trip may have multiple transfers to different trips of the same route, because the earliest reachable trip may not be active on all days.
For transfer reduction, the changes are slightly more complicated.
Here, we additionally keep track of the days on which each transfer is required, again by using a bit set.
During the query stage, we can then select only those transfers that are valid during the requested day(s).
This increases memory usage, but improves query performance.
\subsection{Query}
\label{sec:extend_query}
Although the query stage works the same conceptually, there are some differences in how the data is managed.
The query now must also specify a date on which the journey should take place.
The first step then is to extract the reduced transfers for this date and the following days, in case of long journeys.
This adds some runtime overhead, however, the extracted data can easily be cached, which avoids paying this cost repeatedly.
After this, we proceed to find reachable trips as before, paying attention to the days of operation of each trip.
The remainder of the query stage stays the same -- we don't need to pay any more attention to the bit sets, because we only look at valid transfers, which means that only active trips are reached.
The main challenges are on the engineering side, where data and code has to be structured differently to account for the altered model.
The most important change is in how trips are identified: Whereas before, trips were simply numbered from $0$ to the total number of trips, it's now more complicated, since a query might encounter more than one instance of a trip.
It's also unfeasible when updating the timetable data, which we will discuss in the next section (\ref{sec:updating}).
Instead of a simple integer, trips are now identified by a triple of route ID, trip ID, and a day offset.
The trip IDs are scoped to each route and ordered according to $\cdot\leq\cdot$ (i.e., time).
The day offset is relative to the query day: The day given by the query corresponds to $1$, the next day to $2$, etc.
An offset of $0$ is used for the day before the query, to allow the use of trips that start on the previous day, but cross into the current one.
The trip ID and the day offset is stored as a single integer, with the offset in the higher bits.
This enables efficient comparisons when testing which trip is earlier.
However, since the number of trips is no longer fixed, some of the optimizations used in~\cite{Witt2015} no longer work.
In particular, the original paper proposed ``unrolling'' the data structure used to track reached trips, which makes updates slightly slower (because all following trips also have to be updated),
but lookups much faster (because only a single, known memory location has to be read).
While we could apply the same trick in this work if we limited the query to a low enough number of days, we chose not to.
Instead, we use explicit (Pareto-)sets using tuples of trip ID, stop index, and number of transfers.
This allows us to be more flexible, both in the number of days a query may span, and with regard to possible extensions with more optimization criteria.
Overall, the more complex data layout comes with a small performance cost, compared to the more streamlined original version.
However, it is easy to extract arbitrary time periods into the flat data structures used by the original algorithm, trading extra memory for reduced running time.
If desired, one could adopt a hybrid model in practice, were frequently queried periods (such as the next few days) use the flat structures, while the rest is served by the algorithm proposed in this work.
\section{Updating Timetable Data}
\label{sec:updating}
With an increased time period covered by the timetable, it is likely that occasional changes are required.
Construction work or outages may close some stops and add new ones, require alternate routes for trip, or disable some routes completely.
On top of that, delays and outages may happen no matter how long the time period is.
For short timetables, a complete reprocessing of the data may be feasible, but for longer timetables, we can do better.
While there are many different types of alterations possible, the two fundamental ones are adding and removing trips.
For example, a delay can be incorporated by removing a trip and re-adding it with updated arrival and departure times.
Removing a trip requires updating trips that had a transfer to that trip.
Adding a trip requires updating trips that may have a transfer to the new trip, in addition to computing transfers from the new trip.
In our experiments, we simply re-compute and re-reduce all transfers from the updated trips.
More fine-grained updates may be possible, but would require more sophisticated algorithms to identify which transfers need to be added or removed.
\section{Experimental Evaluation}
\label{sec:experiment}
We ran experiments to measure the required time for preprocessing and queries, as well as for dynamic updates.
We performed our experiments using a 64-core AMD Epyc 7702P processor and 1024\,GB DDR4 RAM.
Preprocessing and each individual dynamic update is done in parallel, using up to 128 threads.
Queries are performed sequentially.
We evaluated five different real-world data sets, covering public transit networks of varying sizes:
Germany, provided to us by Deutsche Bahn, Switzerland, (acquired from \texttt{gtfs.cheops.ch}), the Netherlands (\texttt{gtfs.ovapi.nl}), Sweden (\texttt{trafiklab.se}), and Madrid (\texttt{emtmadrid.es}).
These data instances are summarised in Table~\ref{tab:data_sets}.
\begin{table}[tp]
\centering
\caption{Instances used for experiments.}\label{tab:data_sets}
\begin{tabular}{l r r r r r r}
\hline
Instance & Days & Stops & Trips & Routes & Footpaths \\\hline
Germany & $363$ & $249$\,k & $1\,703$\,k & $240$\,k & $400$\,k \\
Switzerland & $363$ & $30$\,k & $959$\,k & $52$\,k & $94$\,k \\
Netherlands & $145$ & $73$\,k & $636$\,k & $17$\,k & $329$\,k \\
Sweden & $97$ & $51$\,k & $257$\,k & $22$\,k & $61$\,k \\
Madrid & $372$ & $4.6$\,k & $193$\,k & $1.4$\,k & $22$\,k \\
\hline
\end{tabular}
\end{table}
Figures for preprocessing can be found in Table~\ref{tab:preprocessing}.
We can see that the time required for transfer computation is negligible.
Transfer reduction takes a noticeable, but still fairly low amount of time, and results in a substantially lower number of transfers,
which affects both memory usage and query performance positively.
\begin{table}[tp]
\centering
\caption{Preprocessing time and number of transfers.
Listed are the total number of transfers computed, the number of transfers remaining after reduction,
the ratio of the two, and the time required to compute and reduce the transfers (in parallel).
}\label{tab:preprocessing}
\begin{tabular}{l r r r r r r}
\hline
Instance & Total transfers & Reduced transfers & Ratio & Comp.\ time & Reduction time \\\hline
Germany & $2\,357$\,million & $335$\,million & $14.2\,\%$ & $15$\,s & $26.2$\,min \\
Switzerland & $903$\,million & $167$\,million & $17.4\,\%$ & $11$\,s & $14.1$\,min \\
Netherlands & $544$\,million & $78$\,million & $14.4\,\%$ & $5$\,s & $3.6$\,min \\
Sweden & $247$\,million & $51$\,million & $21.0\,\%$ & $2$\,s & $1.3$\,min \\
Madrid & $442$\,million & $37$\,million & $8.5\,\%$ & $8$\,s & $1.3$\,min \\
\hline
\end{tabular}
\end{table}
Before we can perform queries, we need to extract the transfers relevant to that query, as described in Section~\ref{sec:extend_query}.
The evaluations of this process can be found in Table~\ref{tab:preparation}.
When executed in parallel, the total time for extraction is fairly negligible -- at most up to $10$ seconds for Germany.
Ideally, this extraction is done once in advance and then cached, which is what we did for our query experiments.
Of course, this requires a certain amount of memory, but not unreasonably so.
If RAM is short, one could use, e.g., an LRU caching strategy.
The final column in Table~\ref{tab:preparation} lists the time required to transform the data so it can be consumed by the original trip-based algorithm proposed in~\cite{Witt2015}.
This ``flat'' format is simpler and allows a more efficient query, as we will see next.
\begin{table}[tp]
\centering
\caption{Query preparation.
Listed are the (sequential) times required to extract transfers for all days and for a single day,
followed by the total memory usage for all extracted transfers.
The final column lists the time required to ``flatten'' two consecutive days so the original trip-based algorithm~(\cite{Witt2015}) can be used.
}\label{tab:preparation}
\begin{tabular}{l r r r r r r}
\hline
Instance & Total extr.\ time & Extr.\ time per day & Memory usage & Flatten time \\\hline
Germany & $6.7$\,min & $1\,110$\,ms & $163$\,GB & $3\,096$\,ms \\
Switzerland & $2.3$\,min & $377$\,ms & $37$\,GB & $645$\,ms \\
Netherlands & $30$\,s & $207$\,ms & $9.6$\,GB & $400$\,ms \\
Sweden & $18$\,s & $186$\,ms & $7.4$\,GB & $505$\,ms \\
Madrid & $56$\,s & $151$\,ms & $52$\,GB & $599$\,ms \\
\hline
\end{tabular}
\end{table}
For query evaluation, we performed $10\,000$ full-day profile (24 hours from midnight to midnight) queries, with source stop, destination stop, and query day chosen uniformly at random.
However, not all of these queries can be answered with a valid journey, because at any given time, the network may not connected.
This is because some stops may be seasonable, e.g., only operated in summer or in winter, or they may be relocated stops that only exist during a time of construction at the actual stop location.
These kinds of queries can often be answered very quickly with a negative (empty) result.
In order to not skew the results too much, we disregard queries that do not return at least one valid journey in the following.
The results of the query evaluation can be found in Figure~\ref{fig:query_eval}.
Note that the y-axis scale is logarithmic.
The left box plot for each instance corresponds to the queries performed on the full timetable as described in this paper.
For comparison, the right box plot shows the query times using the original algorithm.
The factor between the two is around $4$, with Madrid a notable outlier at $8$.
The reason for the difference in execution time is that the original algorithm has much simpler (flatter) data structures, which allows more streamlining in the query execution.
The more sophisticated data structures used in this work are used to keep the data more compact, and to make updates to that data more efficient.
It's worth noting that queries on the full timetable finds some journeys that the flat version misses, because we limited each query to two consecutive days.
\begin{figure}[tbp]
\centering
\input{combined_query_times_successful.fig}
\caption{Query times, on logarithmic scale.
\texttt{Full} refers to queries on the full timetable as presented in this paper.
\texttt{Flat} refers to queries using the original trip-based algorithm on a relevant portion of the timetable (cf.~\ref{sec:extend_query}).
}\label{fig:query_eval}
\end{figure}
For dynamic updates, we delay trips, chosen uniformly at random, by removing them, adding a random delay, and adding them again.
While this is only one kind of updates amongst many possible ones, as noted in Section~\ref{sec:updating}, most changes to the timetable can be implemented by removing and adding trips.
Furthermore, the costly part of doing updates is not the actual change itself, but the re-computation and re-reduction of transfers of neighbouring trips, which may gain or lose transfers to new or removed trips.
Figure~\ref{fig:update_eval} shows the time required for such an update for each instance.
Madrid is the most expensive one, and the only one where a majority of updates took multiple seconds.
This is because unlike the other four instances, Madrid is a metropolitan network, and much denser than the others.
Therefore, each trips has many neighbours, so each update touches more trips.
This becomes clearer when we normalize the values by the number of trips for which we recompute the transfers~(Figure~\ref{fig:update_eval2}).
A consequence of this is that batched updates of a local part of the network are much more efficient than individual ones.
When making multiple changes to neighbouring trips, it is likely that the sets of trips requiring new transfers have large overlap.
By delaying the computation of these transfers until all pending changes have been applied, updates can become much more efficient.
\begin{figure}[tbp]
\centering
\begin{minipage}[t]{0.48\textwidth}%
\input{update_times.fig}%
\caption{Update times, on logarithmic scale.}\label{fig:update_eval}%
\end{minipage}\hfill%
\begin{minipage}[t]{0.48\textwidth}%
\input{update_time_per_trip.fig}%
\caption{Update times, normalized to show time per updated trip, on logarithmic scale.}\label{fig:update_eval2}%
\end{minipage}
\end{figure}
\section{Conclusion}
\label{sec:conclusion}
We introduced an extension of the trip-based public transit routing algorithm presented by~\cite{Witt2015}, to allow efficient covering of timetables of arbitrary length.
Our experiments showed that this increased flexibility comes at a runtime cost.
However, this gap can be closed by falling back on the original algorithm for hotspots where this is beneficial, by quickly transforming the necessary data.
An important aspect when covering larger periods of time is the ability to update the data without starting from scratch.
We showed that thanks to careful modelling of the data, a simple approach achieves update performance acceptable for real-time scenarios.
An interesting open problem is to combine this work with the speed-up technique proposed by~\cite{Witt2016}.
|
\section{Introduction}
\label{sec:Intro}
In their seminal 1969 paper \cite{DS1969}, Davenport and Schmidt introduce
a novel approach to study how well a given real number $\xi$
may be approximated by algebraic integers of degree at most $n+1$, for a
given positive integer $n$.
Using geometry of numbers, they show that if, for some $c>0$ and $\lambda>0$,
there are arbitrarily large values of $X$ for which the conditions
\begin{equation}
\label{intro:eq1}
|x_0|\le X \quad\mbox{and}\quad \max_{1\le k\le n}|x_0\xi^k-x_k|\le cX^{-\lambda}
\end{equation}
admit no non-zero integer solution $\mathbf{x}=(x_0,\dots,x_n)\in\mathbb{Z}^{n+1}$,
then for some $c'>0$ there are infinitely many algebraic integers $\alpha$
of degree at most $n+1$ such that
\[
|\xi-\alpha|\le c' H(\alpha)^{-1/\lambda-1}
\]
where $H(\alpha)$ stands for the height of $\alpha$, namely the largest
absolute value of the coefficients of its irreducible polynomial
over $\mathbb{Z}$. Assuming that $\xi$ is not itself an algebraic number of
degree at most $n$, they further show that admissible values for $\lambda$ are
$\lambda=1$ if $n=1$, $\lambda=(-1+\sqrt{5})/2$ if $n=2$, $\lambda=1/2$
if $n=3$ and $\lambda=1/\lfloor n/2\rfloor$ if $n\ge 4$. Since then,
their results have been extended to many other settings which include
approximation to $\xi$ by algebraic integers of a given degree
\cite{BT2000}, by algebraic units of a given degree \cite{Te2001} and
by conjugate algebraic integers \cite{RW2004}. A $p$-adic analog is given
in \cite{Te2002} and an extension to a variety of inhomogeneous problems
is proposed in \cite{BL2005a}. Refined values for
$\lambda$ have also been established by Laurent \cite{La2003}, by
Schleischitz \cite{Schl2020, Schl2021} and by Badziahin \cite{Ba2021}.
For each $\xi\in\mathbb{R}$ and each integer $n\ge 1$, let $\widehat{\lambda}_n(\xi)$
(resp.\ $\lambda_n(\xi)$) denote the supremum of all $\lambda \ge 0$
such that, for $c=1$, the conditions \eqref{intro:eq1} admit a non-zero
integer solution ${\bf x}=(x_0,\dots,x_n)\in\mathbb{Z}^{n+1}$ for each sufficiently
large $X$ (resp.\ for arbitrarily large values of $X$). Further, let
$\tau_{n+1}(\xi)$ denote the supremum of all $\tau\ge 0$ for which
there exist infinitely many algebraic integers $\alpha$ of degree at
most $n+1$ with $|\xi-\alpha| \le H(\alpha)^{-\tau}$. Then, in crude
form, the observation of Davenport and Schmidt is that
\begin{equation}
\label{intro:eq:dualite}
\tau_{n+1}(\xi) \ge \widehat{\lambda}_n(\xi)^{-1}+1.
\end{equation}
Thus any upper bound on $\widehat{\lambda}_n(\xi)$ yields a lower bound on $\tau_{n+1}(\xi)$.
Assume now that $\xi$ is not itself an algebraic number of degree at most $n$,
namely that $1,\xi,\dots,\xi^n$ are linearly independent over $\mathbb{Q}$, or equivalently
that $[\mathbb{Q}(\xi):\mathbb{Q}]>n$. Then a result of Dirichlet \cite[\S II.1, Theorem 1A]{Schm1980}
yields $1/n\le \widehat{\lambda}_n(\xi)$ and, expecting the equality, it is natural to conjecture that
$\tau_{n+1}(\xi) \ge n+1$ as in \cite[\S 5]{Schm1983}. For $n=1$, we have
$\widehat{\lambda}_1(\xi)=1$ and the conjectured lower bound $\tau_2(\xi)\ge 2$ follows.
However, for $n=2$, the upper bound $\widehat{\lambda}_2(\xi)\le (-1+\sqrt{5})/2\cong 0.618$
from \cite[Theorem 1a]{DS1969} is best possible by \cite[Theorem 1.1]{Ro2004}, and the corresponding lower bound $\tau_3(\xi)\ge (3+\sqrt{5})/2\cong 2.618$ is also best possible by
\cite[Theorem 1.1]{Ro2003} (see also \cite{RZ2011}). This disproves the natural
conjecture for $n=2$ and suggests that it might be false as well for each $n\ge 3$.
Any counterexample $\xi$ would have $[\mathbb{Q}(\xi):\mathbb{Q}]>n$ and $\widehat{\lambda}_n(\xi)>1/n$.
So, it would be transcendental over $\mathbb{Q}$ according to Schmidt's subspace theorem
\cite[\S VI.1, Corollary 1E]{Schm1980}. Although the existence of such number remains
an open problem for $n\ge 3$, we know by contrast large families of transcendental real
numbers $\xi$ with $\widehat{\lambda}_2(\xi)>1/2$. In chronological order, they are the extremal
numbers $\xi$ of \cite{Ro2004}, the Sturmian continued fractions of \cite{BL2005b},
Fischler's numbers from \cite{Fi2007}, the Fibonacci type numbers
of \cite{Ro2007} and the Sturmian type numbers of \cite{Po2020}, all contained in the very
general class of numbers studied in \cite{Po2021}. In particular, we know
by \cite[Corollary]{Ro2007} that the values $\widehat{\lambda}_2(\xi)$ with $\xi$ real and
transcendental form a dense subset of the interval $[1/2,(-1+\sqrt{5})/2]$.
For $n\ge 3$, recent progresses have been made on upper bounds for $\widehat{\lambda}_n(\xi)$.
The estimates $\widehat{\lambda}_3(\xi)\le 1/2$ and $\widehat{\lambda}_n(\xi)\le 1/\lfloor n/2\rfloor$ for
$n\ge 4$ from \cite[Theorems 2a and 4a]{DS1969} have been refined by Laurent \cite{La2003}
to $\widehat{\lambda}_n(\xi)\le 1/\lceil n/2\rceil$ for each $n\ge 3$, together with an important
simplification in the proof. When $n=3$, the best computed upper bound (yet not
optimal) remains that of \cite{Ro2008},
\begin{equation}
\label{intro:eq:lambda_3}
\widehat{\lambda}_3(\xi)\leq \alpha =0.4245\cdots,
\end{equation}
where $\alpha$ is the root of the polynomial $1-3x+4x^3-x^4$ in the interval $[1/3,1/2]$.
For even integers $n=2m\ge 4$, Schleischitz \cite{Schl2020, Schl2021}
refined the upper bound $\widehat{\lambda}_n(\xi)\le 1/m$ by reducing to the case
where $\lambda_n(\xi)\le 1/m$ and then by using a transference inequality
of Marnat and Moshchevitin \cite{MM2020} relating $\lambda_n(\xi)$ and
$\widehat{\lambda}_n(\xi)$.
Nevertheless, all those refinements, including those of the recent preprint
of Baziahin \cite{Ba2021}, are of the form $\widehat{\lambda}_n(\xi)\le 1/(n/2+c_n)$
with $0 < c_n < 1$. Our main result below improves significantly
on this when $n$ is large.
\begin{theorem}
\label{intro:thm:main}
For any integer $n\ge 2$ and any $\xi\in\mathbb{R}$ with $[\mathbb{Q}(\xi):\mathbb{Q}]>n$, we have
\begin{equation}
\label{intro:thm:main:eq}
\widehat{\lambda}_n(\xi) \leq \frac{1}{n/2+a\sqrt{n}+1/3}.
\end{equation}
where $a=(1-\log(2))/2\cong 0.1534$.
\end{theorem}
Note that the multiplicative constant $a$ in the denominator is not optimal and
could be improved with additional work. The same applies to the additive
constant $1/3$ given the actual choice of $a$. In view of \eqref{intro:eq:dualite},
this gives
\[
\tau_{n+1}(\xi)\ge n/2+a\sqrt{n}+4/3
\]
for the same $n$ and $\xi$.
As explained by Bugeaud in \cite[Prop.~3.3]{Bu2004}, the arguments of Davenport and
Schmidt leading to \eqref{intro:eq:dualite} can also be adapted to Wirsing's problem
of approximating real numbers $\xi$ by algebraic numbers, yielding
$\omega^*_n(\xi)\ge \widehat{\lambda}_n(\xi)^{-1}$ for any integer $n\ge 1$, where
$\omega^*_n(\xi)$ denotes the supremum of all $\omega>0$ for which there
exist infinitely many algebraic numbers $\alpha$ of degree at most $n$
with $|\xi-\alpha|\le H(\alpha)^{-\omega-1}$. Thus
when $[\mathbb{Q}(\xi):\mathbb{Q}]>n$, the inequality \eqref{intro:thm:main:eq} implies that
$\omega_n^*(\xi)\ge n/2+a\sqrt{n}+1/3$. However, this is superseded by the recent
breakthrough of Badziahin and Schleischitz who showed in \cite{BS2021} that
$\omega_n^*(\xi)>n/\sqrt{3}$ when $[\mathbb{Q}(\xi):\mathbb{Q}]>n\ge 4$. Previous to their work,
the best lower bounds for large values of $n$ were of the form $n/2+\GrO(1)$.
For small values of $n$, namely for $n$ odd with $5\le n\le 49$ and for $n$ even
with $4\le n\le 100$, we obtain the following estimates which improve on Theorem \ref{intro:thm:main}.
\begin{theorem}
\label{intro:thm:impair}
Suppose that $n=2m+1\ge 5$ is odd. Then for each $\xi\in\mathbb{R}$ with $[\mathbb{Q}(\xi):\mathbb{Q}]>n$, we have
\[
\widehat{\lambda}_{2m+1}(\xi) \leq \alpha_m,
\]
where $\alpha_m$ is the positive root of the polynomial $P_m(x)=1 -(m+1)x - mx^2$.
\end{theorem}
\begin{theorem}
\label{intro:thm:pair}
Suppose that $n =2m\ge 4$ is even. Then for each $\xi\in\mathbb{R}$ with $[\mathbb{Q}(\xi):\mathbb{Q}]>n$, we have
\begin{equation*}
\widehat{\lambda}_{2m}(\xi) \leq \beta_m
\end{equation*}
where $\beta_m$ is the positive root of the polynomial
\[
Q_m(x)
=\begin{cases}
1-mx-mx^2-m(m-1)x^3 &\text{if $m\ge3$,}\\
1-3x+x^2-2x^3-2x^4 &\text{if $m=2$.}
\end{cases}
\]
\end{theorem}
Again, these upper bounds could be slightly improved with more work, at least for each $n\geq 6$
(even or odd). Note that they are relatively close to $1/(m+2)$ as one finds
\[
\frac{1}{m+2}<\alpha_m<\frac{1}{m+2}+\frac{2}{(m+2)^3}
\quad\mbox{and}\quad
\frac{1}{m+2}<\beta_m<\frac{1}{m+2}+\frac{7}{(m+2)^3}
\]
for each $m\ge 2$. The table below shows how they compare to those of Laurent (L.) for odd $n\le 13$
and to those of Schleischitz (S.) and Badziahin (B.) for even $n\le 12$.
\begin{figure}[H]
\label{intro:table}
{\renewcommand{\arraystretch}{1.2}
\centering
$\begin{array}{|c|| c | c | c |c|}
\hline
\mathbf{n} & \mathbf{L.} & \mathbf{S.} & \mathbf{B.} & \mathbf{new} \\ \hline
4 & & 0.3706\cdots & 0.3660\cdots & 0.3370\cdots \\ \hline
5 & 0.3333\cdots & & & 0.2807\cdots \\ \hline
6 & & 0.2681\cdots & 0.2637\cdots & 0.2444\cdots \\ \hline
7 & 0.2500\cdots & & & 0.2152\cdots \\ \hline
8 & & 0.2107\cdots & 0.2071\cdots & 0.1919\cdots \\ \hline
9 & 0.2000\cdots & & & 0.1753\cdots \\ \hline
10 & & 0.1737\cdots & 0.1708\cdots & 0.1587\cdots \\ \hline
11 & 0.1666\cdots & & & 0.1483\cdots \\ \hline
12 & & 0.1478\cdots & 0.1454\cdots & 0.1357\cdots \\ \hline
13 & 0.1428\cdots & & & 0.1286\cdots \\ \hline
\end{array}$}
\caption{Upper bounds for $\widehat{\lambda}_n$}
\label{intro:table1}
\end{figure}
For the proof we develop new tools. The main one concerns the behavior of the
function $f(\ell)=\dim\mathcal{U}^\ell(A)$ for $\ell\in\{0,1,\dots,n+1\}$ where $A$ is any
subspace of $\mathbb{R}^{n+1}$ and $\mathcal{U}^\ell(A)$ stands for the subspace of
$\mathbb{R}^{n-\ell+1}$ spanned by the images of $A$ through the projections
$(x_0,\dots,x_n)\mapsto (x_k,\dots,x_{k+n-\ell})$ for $k=0,\dots,\ell$,
with the convention that $\mathcal{U}^{n+1}(A)=0$.
In Sections \ref{sec:three} and \ref{sec:proofs}, we show that such a function is
concave and monotone increasing as long as $\mathcal{U}^\ell(A)\neq\mathbb{R}^{n-\ell+1}$.
We also study the degenerate cases where $f(\ell)<\dim(A)+\ell \le n-\ell+1$.
In Section \ref{sec:minimal}, we form a sequence of minimal points
$(\mathbf{x}_i)_{i\ge 0}$ for $\xi$ in $\mathbb{Z}^{n+1}$, and recall how the exponents
$\widehat{\lambda}_n(\xi)$ and $\lambda_n(\xi)$ can be computed from this data.
Given integers $0\le j,\ell\le n$,
we say that Property $\mathcal{P}(j,\ell)$ holds if, for any subspace $A=\Span{\mathbf{x}_i,\dots,\mathbf{x}_q}$
of dimension at most $j+1$ spanned by consecutive minimal points with a large enough
initial index $i$, we have $\dim\mathcal{U}^\ell(A)\ge\dim(A)+\ell$. This is a crucial notion
with the remarkable feature that $\mathcal{P}(j,\ell)$ implies $\mathcal{P}(j+1,\ell-1)$ when $\ell\ge 1$.
In Sections \ref{sec:P0} and \ref{sec:P1}, we establish some consequences of Properties
$\mathcal{P}(0,\ell)$ and $\mathcal{P}(1,\ell)$ respectively and we provide lower bounds on
$\widehat{\lambda}_n(\xi)$ which ensure that these properties hold. In Section \ref{sec:P0}, we
also study the general situation where $\mathcal{P}(j,\ell-1)$ holds but not $\mathcal{P}(j,\ell)$ for
some integer $\ell\ge 1$. In Sections \ref{sec:first} and
\ref{sec:alt}, we provide two types of upper bounds for the height of $\mathcal{U}^\ell(A)$
when $\mathcal{P}(j,\ell)$ holds and $A=\Span{\mathbf{x}_i,\dots,\mathbf{x}_q}$ has dimension $j+1$.
In particular, the estimate of Section \ref{sec:alt} yields a strong constraint on the growth
of the norms $X_i=\norm{\mathbf{x}_i}$. These tools are combined in Section \ref{sec:main} to
prove Theorem \ref{intro:thm:main}.
Finally Theorems \ref{intro:thm:impair} and \ref{intro:thm:pair} are proved respectively
in Sections~\ref{sec:odd} and~\ref{sec:even}, with the help of a new construction
presented in Section~\ref{sec:new}.
We start in the next section by fixing some notation, including our notion of height
for the subspaces of $\mathbb{R}^m$ defined over $\mathbb{Q}$.
\section{Heights}
\label{sec:heights}
For each integer $m\ge 1$, we view $\mathbb{R}^m$ as an Euclidean space
for the usual scalar product of points $\mathbf{x},\,\mathbf{y}\in\mathbb{R}^m$ written
$\mathbf{x}\cdot\mathbf{y}$, and we denote by $\norm{\mathbf{x}}=\sqrt{\mathbf{x}\cdot\mathbf{x}}$
the Euclidean norm of a point $\mathbf{x}\in\mathbb{R}^m$. For
each integer $k=1,\dots,m$, we also identify $\bigwedge^k\mathbb{R}^m$
with $\mathbb{R}^{\binom{m}{k}}$ via a choice of ordering of the Pl\"ucker
coordinates and we denote by $\norm{\boldsymbol{\alpha}}$ the resulting
Euclidean norm of a point $\boldsymbol{\alpha}\in\bigwedge^k\mathbb{R}^m$.
For any subset $A$ of $\mathbb{R}^m$, we denote by $\langle A\rangle$
the linear subspace of $\mathbb{R}^m$ spanned by $A$ over $\mathbb{R}$. When
$A$ is a finite set $\{\mathbf{x}_1,\dots,\mathbf{x}_k\}$, we simply write
$\langle \mathbf{x}_1,\dots,\mathbf{x}_k \rangle$.
Let $V$ be an arbitrary vector subspace of $\mathbb{R}^m$ defined over $\mathbb{Q}$.
If $V\neq 0$, we define its height $H(V)$ as the covolume of the lattice
$V\cap\mathbb{Z}^m$ inside $V$. Explicitly, if $\dim(V)=k$ and if
$\{\mathbf{x}_1,\dots,\mathbf{x}_k\}$ is a basis of $V\cap\mathbb{Z}^m$ over $\mathbb{Z}$, then
\begin{equation}
\label{heights:eq:H(V)}
H(V)=\norm{\mathbf{x}_1\wedge\cdots\wedge\mathbf{x}_k}.
\end{equation}
For $V=0$, we set $H(0)=1$. Then, we have the duality relation
\begin{equation}
\label{heights:eq:duality}
H(V)=H(V^\perp)
\end{equation}
where $V^\perp$ denotes the orthogonal complement of $V$ in $\mathbb{R}^m$
\cite[Chapter I, \S8]{Schm1991}. In particular, this gives
$H(\mathbb{R}^m)=H(0^\perp)=1$. If $\mathbf{x}\in\mathbb{Z}^m$ is \emph{primitive}, namely
if the gcd of its coordinates is $1$, we have
$H(\langle\mathbf{x}\rangle)=H(\langle\mathbf{x}\rangle^\perp)=\norm{\mathbf{x}}$.
We will also need the following important inequality of Schmidt
\begin{equation}
\label{heights:eq:schmidt}
H(U\cap V)H(U+V)\le H(U)H(V),
\end{equation}
valid for any subspaces $U$ and $V$ of $\mathbb{R}^m$
defined over $\mathbb{Q}$ \cite[Chapter I, Lemma 8A]{Schm1991}.
Finally, given $\xi\in\mathbb{R}$ and $\mathbf{x}=(x_0,\dots,x_m)\in
\mathbb{Z}^{m+1}\setminus\{0\}$, we define
\[
L_\xi(\mathbf{x})=\max_{1\le j\le m}|x_0\xi^j-x_j|,
\]
and note that
\[
L_\xi(\mathbf{x}) \asymp \norm{\Xi_m\wedge\mathbf{x}}
\quad
\text{where}
\quad
\Xi_m=(1,\xi,\dots,\xi^{m})
\]
with implicit constants depending only on $\xi$ and $m$.
The latter relation is instructive since the product
$\norm{\Xi_m\wedge\mathbf{x}}\,\norm{\Xi_m}^{-1}\norm{\mathbf{x}}^{-1}$ represents
the sine of the angle between $\Xi_m$ and $\mathbf{x}$. We will
repeatedly use the following generalization of \cite[Lemma 9]{DS1969}.
\begin{lemma}
\label{heights:lemma}
Suppose that $\mathbf{x}_1,\dots,\mathbf{x}_k\in\mathbb{Z}^{m+1}$ are linearly
independent. Then
\[
H(\langle\mathbf{x}_1,\dots,\mathbf{x}_k\rangle)
\le \norm{\mathbf{x}_1\wedge\cdots\wedge\mathbf{x}_k}
\ll \sum_{i=1}^k\norm{\mathbf{x}_i}\prod_{j\neq i}L_\xi(\mathbf{x}_j)
\]
with an implied constant which depends only on $\xi$ and\/ $m$.
\end{lemma}
This follows from \eqref{heights:eq:H(V)} by writing $\mathbf{x}_j=x_{j,0}\Xi_m+\Delta_j$
for $j=1,\dots,k$, where $x_{j,0}$ stands for the first coordinate of
$\mathbf{x}_j$, and then by expanding the exterior product upon noting that
$\norm{\Delta_j}\asymp L_\xi(\mathbf{x}_j)$.
\section{Three crucial propositions}
\label{sec:three}
Let $\ell,n$ be integers with $0\le \ell\le n$. For each
$\mathbf{x}=(x_0,\dots,x_n)\in\mathbb{R}^{n+1}$, we denote by
\[
\mathbf{x}^{(k,\ell)} = (x_{k},\dots,x_{k+n-\ell})\in \mathbb{R}^{n+1-\ell}
\quad (0\le k\le \ell),
\]
the points consisting of $n+1-\ell$ consecutive coordinates of $\mathbf{x}$,
and we denote by
\[
\mathcal{U}^{\ell}(\mathbf{x}) = \Span{ \mathbf{x}^{(0,\ell)},\dots, \mathbf{x}^{(\ell,\ell)}}
\subseteq \mathbb{R}^{n+1-\ell},
\]
the vector subspace of $\mathbb{R}^{n+1-\ell}$ which they generate.
In general, for each non-empty subset $A$ of $\mathbb{R}^{n+1}$, we define
\[
\mathcal{U}^\ell(A)=\sum_{\mathbf{x}\in A}\mathcal{U}^\ell(\mathbf{x})
\quad\mbox{and}\quad
\mathcal{U}^{n+1}(A)=0.
\]
Then, we have $\mathcal{U}^k(\mathcal{U}^{\ell-k}(A)) = \mathcal{U}^\ell(A)
= \mathcal{U}^\ell(\langle A\rangle)$ for any integers
$0\le k\le \ell\le n+1$.
Our interest in the truncated points $\mathbf{x}^{(k,\ell)}$ comes from
the fact that, when $\mathbf{x}\in\mathbb{Z}^{n+1}$ and $\ell<n$, they belong to
$\mathbb{Z}^{n+1-\ell}$ and, for given $\xi\in \mathbb{R}$, they satisfy
\begin{equation}
\label{three:eq:xkl}
\norm{\mathbf{x}^{(k,\ell)}} \le \norm{\mathbf{x}}
\quad\mbox{and}\quad
L_\xi(\mathbf{x}^{(k,\ell)}) \ll L_\xi(\mathbf{x})
\end{equation}
with implied constants depending only on $\xi$ and $n$.
So Lemma \ref{heights:lemma} yields
\begin{equation}
\label{three:eq:HUx}
H(\mathcal{U}^\ell(\mathbf{x}))\ll \norm{\mathbf{x}}L_\xi(\mathbf{x})^{d-1}
\quad
\text{if}
\quad
d=\dim\mathcal{U}^\ell(\mathbf{x})>0.
\end{equation}
In general, when $A$ is a subspace of $\mathbb{R}^{n+1}$ defined
over $\mathbb{Q}$, the subspace $\mathcal{U}^\ell(A)$ of $\mathbb{R}^{n+1-\ell}$
is also defined over $\mathbb{Q}$ and, as the above example shows,
we need some information on its dimension in order to estimate
its height.
In this section, we state three propositions concerning
$\mathcal{U}^\ell(A)$ as a function of $\ell$, for
a fixed subspace $A$ of $\mathbb{R}^{n+1}$, but postpone their proofs
to the next section. In order to state the first one,
we recall that a function $f:\{0,\dots,n+1\}\rightarrow \mathbb{R}$ is
\emph{convex} if it satisfies the following equivalent conditions
\begin{itemize}
\item[(1)] $f(i)-f(i-1) \le f(i+1)-f(i)$ \ for $i=1,\dots,n$;
\medskip
\item[(2)] $\displaystyle \frac{f(j)-f(i)}{j-i} \le \frac{f(k)-f(j)}{k-j}$ \
whenever $0\leq i<j<k\le n+1$.
\end{itemize}
We say that $f$ is \emph{concave} if $-f$ is convex. We also fix a
positive integer $n$.
\begin{prop}
\label{three:prop:concave}
Let $A$ be a subspace of $\mathbb{R}^{n+1}$. Then $f(\ell)=\dim\mathcal{U}^\ell(A)$
is a concave function of $\ell\in\{0,\dots,n+1\}$. Moreover,
there is an integer $m\in\{0,\dots,n+1\}$ for which the function $f$ is
monotonically increasing on $\{0,\dots,m\}$, while strictly
decreasing with $f(\ell)=n-\ell+1$ for $\ell\in\{m,\dots,n+1\}$.
\end{prop}
Figure \ref{three:fig} illustrates this result. Taking it for granted,
we deduce a useful corollary.
\begin{figure}[H]
\begin{tikzpicture}[xscale=0.5,yscale=0.4]
\draw[-stealth, semithick] (-0.15,0)--(22,0) node[below]{$\ell$};
\draw[-stealth, semithick] (0,-0.15)--(0,12) node[right]{$f(\ell)=\dim\mathcal{U}^\ell(A)$};
\node[below] at (0,-0.15) {$0$};
\node[draw,circle,inner sep=1.25pt,fill] at (0,2) {};
\node[draw,circle,inner sep=1.25pt,fill] at (1,5) {};
\node[draw,circle,inner sep=1.25pt,fill] at (5,10) {};
\node[draw,circle,inner sep=1.25pt,fill] at (9,10) {};
\node[draw,circle,inner sep=1.25pt,fill] at (19,0) {};
\node[left] at (0,2) {$\dim(A)$};
\node[below] at (19,0) {$n+1$};
\draw[thick] (0,2)--(1,5)--(1.5,6);
\draw[thick,dotted] (1.5,6)--(2,7);
\draw[thick,dotted] (3,8)--(4,9);
\draw[thick] (4,9)--(5,10)--(9,10)--(19,0);
\draw[dashed] (9,10) -- (9,0);
\draw[thick] (9,0.15)--(9,-0.15) node[below]{$m$};
\draw[dashed] (5,10) -- (0,10);
\draw[thick] (0.15,10)--(-0.15,10) node[left]{$n-m+1$};
\draw[->] (16,6) to [out=270,in=0, looseness=1] (15,5);
\node[above] at (16,6) {$f(\ell)=n-\ell+1$};
\end{tikzpicture}
\caption{Graph of the piecewise linear function interpolating the
values $f(\ell)=\dim\mathcal{U}^\ell(A)$ at integers $\ell\in\{0,\dots,n+1\}$.}
\label{three:fig}
\end{figure}
\begin{corollary}
\label{three:cor}
Let $A$ be a subspace of $\mathbb{R}^{n+1}$, and let $\ell\in\{1,\dots,n\}$. Then
\begin{itemize}
\item[(i)]
$\min\{\dim\mathcal{U}^{\ell}(A),\,\dim(A)+\ell-1\} \le \dim\mathcal{U}^{\ell-1}(A)$;
\item[(ii)]
$\min\{\dim\mathcal{U}^{\ell-1}(A),\,n-\ell+1\} \le \dim\mathcal{U}^{\ell}(A)$.
\end{itemize}
\end{corollary}
\begin{proof}
Let $f$ and $m$ be as in Proposition \ref{three:prop:concave}.
If $f(\ell-1)=n-\ell+2$, then $\ell-1\ge m$, so $f(\ell)=n-\ell+1$
and we are done. Otherwise the function $f$ is concave and
monotonically increasing on $\{0,\dots,\ell\}$, hence $f(\ell-1)
\le f(\ell)$, so (ii) holds. If $f(\ell-1)=f(\ell)$ or if
$\ell=1$, then (i) also holds since $f(0)=\dim\mathcal{U}^0(A)=\dim(A)$.
So we may further assume that $f(\ell-1) < f(\ell)$ and that
$\ell > 1$. By concavity of $f$, we deduce that
\begin{align*}
\frac{f(\ell-1)-f(0)}{\ell-1} \geq f(\ell)-f(\ell-1) \geq 1,
\end{align*}
hence $f(\ell-1)\geq f(0)+\ell-1 = \dim(A)+\ell-1$,
and (i) holds again.
\end{proof}
The second proposition provides additional information in the
degenerate situation where $\dim\mathcal{U}^\ell(A)<\dim(A)+\ell$.
\begin{prop}
\label{three:prop:height}
Let $j,\ell\ge 0$ be integers with $j+2\ell\le n$, and let
$A$ be a subspace of $\mathbb{R}^{n+1}$ of dimension $j+1$ defined
over $\mathbb{Q}$. Suppose that
\[
d:=\dim\mathcal{U}^\ell(A) \le j+\ell,
\]
and set $V=\mathcal{U}^{n-d}(A)$. Then, we have $0\le d-j-1< \ell\le n-d$ and
\[
\dim\mathcal{U}^t(A)=d
\quad\mbox{and}\quad
H(\mathcal{U}^t(A))\asymp H(V)^{n-d-t+1}
\]
for each $t=d-j-1,\dots,n-d$, with implied constants depending
only on $n$. Moreover, for such
$t$ and for $\mathbf{x}\in\mathbb{R}^{n+1}$, the condition
$\mathcal{U}^t(\mathbf{x})\subseteq\mathcal{U}^t(A)$ is equivalent to
$\mathcal{U}^{n-d}(\mathbf{x})\subseteq V$, thus independent of $t$.
\end{prop}
The last result exhibits the generic behavior of a family of linear maps.
\begin{prop}
\label{three:prop:avoiding}
Let $\ell\in\{0,\dots,n\}$ and let $V$ be a subspace
of $\mathbb{R}^{n-\ell+1}$. Suppose that a vector subspace $A$ of
$\mathbb{R}^{n+1}$ satisfies $\dim(A)\le n-\ell+1$
and $\mathcal{U}^\ell(A)\not\subseteq V$. Then, there exists
a point $\mathbf{a}=(a_0,\dots,a_\ell)\in\mathbb{Z}^{\ell+1}$ with
$\sum_{k=0}^\ell |a_k|\le (n+1)^\ell$
such that the linear map
\begin{equation}
\label{three:eq:tau}
\begin{array}{rcl}
\tau_\mathbf{a}\colon \mathbb{R}^{n+1} &\longrightarrow &\mathbb{R}^{n-\ell+1}\\
\mathbf{x}&\longmapsto&\sum_{k=0}^\ell a_k\mathbf{x}^{(k,\ell)}
\end{array}
\end{equation}
is injective on $A$ with $\tau_\mathbf{a}(A)\not\subseteq V$.
\end{prop}
\section{Proofs of the three propositions}
\label{sec:proofs}
Our goal is to prove the statements of the preceding section by
re-interpreting them in a polynomial setting similar to
that of \cite{RW2004}. In particular, we will connect the
function $f$ of Proposition \ref{three:prop:concave} to
the Hilbert-Samuel function of a graded module over a
polynomial ring in two variables. To this end, we start
by fixing some notation.
Let $E=\mathbb{R}[T_0,T_1]$ denote the ring of polynomials in two variables
$T_0$ and $T_1$ over $\mathbb{R}$, and let $D=\mathbb{R}[\delta_0,\delta_1]$
denote the subring of $\mathrm{End}_\mathbb{R}(E)$ spanned by the partial derivatives
\[
\delta_0=\frac{\partial}{\partial T_0}
\quad\mbox{and}\quad
\delta_1=\frac{\partial}{\partial T_1}
\]
restricted to $E$. It is easily seen that these commuting linear
operators are algebraically independent over $\mathbb{R}$. Thus, $D$ is
a commutative ring isomorphic to $E$. In particular, both $D$
and $E$ are graded rings (by the degree) as well as unique
factorization domains. Moreover, $E$ is a $D$-module for the
natural action of the differential operators of $D$ on $E$.
For each integer $n\ge 0$, we denote by $E_n=\mathbb{R}[T_0,T_1]_n$ and
$D_n=\mathbb{R}[\delta_0,\delta_1]_n$ the homogeneous parts of $E$ and
$D$ of degree $n$. We also denote by $\psi_n\colon\mathbb{R}^{n+1}\to
E_n$ the linear isomorphism sending a point $\mathbf{x}=(x_0,\dots,x_n)
\in\mathbb{R}^{n+1}$ to the polynomial
\[
\psi_n(\mathbf{x}) = \sum_{i=0}^n \binom{n}{i} x_i T_0^{n-i}T_1^i.
\]
When $n\ge 1$, we find that
\begin{equation}
\label{proofs:eq0}
\delta_0\psi_n(\mathbf{x})=n\psi_{n-1}(\mathbf{x}^{(0,1)})
\quad\mbox{and}\quad
\delta_1\psi_n(\mathbf{x})=n\psi_{n-1}(\mathbf{x}^{(1,1)}),
\end{equation}
thus $D_1\psi_n(\mathbf{x})=\psi_{n-1}(\mathcal{U}^1(\mathbf{x}))$. We deduce that,
for any subspace $A$ of $\mathbb{R}^{n+1}$, we have $D_1\psi_n(A)
= \psi_{n-1}(\mathcal{U}^1(A))$ and so, by induction,
\begin{equation}
\label{proofs:eq1}
D_\ell\psi_n(A) = \psi_{n-\ell}(\mathcal{U}^\ell(A))
\quad \text{for each $\ell\in\{0,\dots,n\}$.}
\end{equation}
Thus, if we identify $\mathbb{R}^{n+1}$ with $E_n$ for each $n\ge 0$,
then $\mathcal{U}^\ell(A)$ becomes simply $D_\ell A$ for each subspace
$A$ of $E_n$ and each $\ell=0,\dots,n+1$, including $\ell=n+1$
because $D_{n+1}A=0$.
From now on, we fix a positive integer $n$, a subspace $A$
of $\mathbb{R}^{n+1}$, and a spanning set $\{\mathbf{x}_1,\dots,\mathbf{x}_s\}$
of $A$ as a vector space over $\mathbb{R}$. We set $P_i=\psi_n(\mathbf{x}_i)$
for each $i=1,\dots,s$, and form the $D$-module homomorphism
$\varphi\colon D^s \to E$ given by
\[
\varphi(\mathbf{d}) = d_1P_1+\cdots+d_sP_s
\]
for each $\mathbf{d}=(d_1,\dots,d_s)\in D^s$. Then $M:=\ker(\varphi)$
is a graded submodule of $D^s$. Define
\[
f(\ell)=\dim(\varphi(D_\ell^s))
\quad\mbox{and}\quad
g(\ell)=\dim(M_\ell)
\quad
\text{for each $\ell\in\{0,1,\dots,n+1\}$,}
\]
where $M_\ell=M\cap D_\ell^s$ stands for the homogeneous part of $M$
of degree $\ell$, and the dimensions are taken over $\mathbb{R}$. Then, we have
\begin{equation}
\label{proofs:eq2}
f(\ell)=\dim(D_\ell^s)-\dim(M_\ell)=(\ell+1)s-g(\ell)
\quad
(0\le \ell\le n+1).
\end{equation}
By \eqref{proofs:eq1}, we also have $\varphi(D_\ell^s)=D_\ell\psi_n(A)
=\psi_{n-\ell}(\mathcal{U}^\ell(A))$ for $\ell=0,\dots,n$. Comparing
dimensions, this gives
\begin{equation}
\label{proofs:eq3}
f(\ell)=\dim(\mathcal{U}^\ell(A)) \quad
(0\le \ell\le n+1)
\end{equation}
upon noting that for $\ell=n+1$ both sides vanish. Finally, we define
\begin{equation}
\label{proofs:eq4}
h(\ell)=g(\ell+1)-g(\ell)=\dim(M_{\ell+1}/\delta_1M_\ell)
\quad
\text{for each $\ell\in\{0,1,\dots,n\}$.}
\end{equation}
With this notation, our main observation is the following.
\begin{lemma}
\label{proofs:lemma:h}
For each $\ell\in\{1,\dots,n\}$, we have $h(\ell-1)\le h(\ell)$
with equality if and only if $M_{\ell+1}=D_1M_\ell$.
\end{lemma}
\begin{proof}
Fix an integer $\ell\in\{1,\dots,n\}$, and consider the linear
map $\nu\colon M_\ell \to M_{\ell+1}/\delta_1M_\ell$ given by
$\nu(\mathbf{d})=\delta_0\mathbf{d}+\delta_1M_\ell$ for each $\mathbf{d}\in M_\ell$.
If $\mathbf{d}\in\ker(\nu)$, then $\delta_0\mathbf{d}=\delta_1\mathbf{u}$ for some
$\mathbf{u}\in M_\ell\subseteq D_\ell^s$. Hence $\delta_0$ divides
$\delta_1\mathbf{u}$ in $D^s$, so $\mathbf{u}=\delta_0\mathbf{v}$ for some
$\mathbf{v}\in D_{\ell-1}^s$, and then $\mathbf{d}=\delta_1\mathbf{v}$. Since
$\mathbf{d},\mathbf{u}\in M_\ell$, we find $0=\varphi(\mathbf{u})=\delta_0\varphi(\mathbf{v})$
and $0=\varphi(\mathbf{d})=\delta_1\varphi(\mathbf{v})$, thus $\varphi(\mathbf{v})=0$.
This means that $\mathbf{v}\in M_{\ell-1}$, and so $\mathbf{d}=\delta_1\mathbf{v}
\in\delta_1M_{\ell-1}$. This shows
that $\ker(\nu)\subseteq\delta_1M_{\ell-1}$. As the reverse
inclusion is clear, we conclude that $\ker(\nu)=\delta_1M_{\ell-1}$,
and so $\nu$ induces an injective map from $M_\ell/\delta_1M_{\ell-1}$
to $M_{\ell+1}/\delta_1M_\ell$. Comparing dimensions, we deduce
that $h(\ell-1)\le h(\ell)$. Moreover, we have the equality
$h(\ell-1)=h(\ell)$ if and only if $\nu$ is surjective, a condition
which amounts to $M_{\ell+1}=\delta_0M_\ell+\delta_1M_\ell$ or
equivalently to $M_{\ell+1}=D_1M_\ell$.
\end{proof}
As a consequence, we deduce the first assertion of
Proposition \ref{three:prop:concave}.
\begin{corollary}
\label{proofs:cor:lemma:h}
The function $g(\ell)=\dim(M_\ell)$ is convex on
$\{0,1,\dots,n+1\}$, while the function
$f(\ell)=\dim(\mathcal{U}^\ell(A))$ is concave on the same set.
\end{corollary}
\begin{proof}
The assertion for $g$ follows directly from the lemma
and the definition of $h$ in \eqref{proofs:eq4}. Then
\eqref{proofs:eq2} gives $f(\ell)$ as the sum of two
concave functions of $\ell$ on $\{0,\dots,n+1\}$,
namely $(\ell+1)s$ and $-g(\ell)$, thus $f$ is concave.
\end{proof}
For each $m\ge 0$, we note that the action of $D$ on $E$
induces a non-degenerate bilinear form
\[
\begin{array}{rcl}
D_m\times E_m &\longrightarrow &\mathbb{R}\\
(\delta,P) &\longmapsto &\delta P
\end{array}
\]
which identifies $D_m$ with the dual of $E_m$, the
dual of the natural basis $(T_0^{m-i}T_1^i)_{0\le i\le m}$
of $E_m$ being
\[
\left(
\frac{\delta_0^{m-i}\delta_1^i}{(m-i)!i!}
\right)_{0\le i\le m}.
\]
So, for each subspace $W$ of $E_m$ (resp.\ $W$ of $D_m$), its orthogonal
space
\[
W^\perp=\{\delta\in D_m\,;\,\delta W=0\}\subseteq D_m
\quad
\big(\text{resp.\ } W^\perp=\{P\in E_m\,;\,WP=0\}\subseteq E_m\,\big)
\]
satisfies $\dim(W^\perp)=m+1-\dim(W)$ and $(W^\perp)^\perp=W$.
We can now prove the following.
\begin{lemma}
\label{proofs:lemma:injective}
Let $W$ be a proper subspace of $E_m$ for some $m\ge 0$. Then
there are at most $m$ real numbers $a$ for which the differential
operator $\delta=\delta_0+a\delta_1$ is not injective on $W$.
\end{lemma}
\begin{proof}
Without loss of generality, we may assume that $W$ has codimension
$1$ in $E_m$. Then $W^\perp=\langle \gamma\rangle$ for some non-zero
$\gamma\in D_m$. Suppose that $\delta=\delta_0+a\delta_1\in D_1$ is not
injective on $W$, for some $a\in\mathbb{R}$. Then $\delta W$ is a proper
subspace of $E_{m-1}$ and so $\beta\delta W=0$ for some non-zero
$\beta\in D_{m-1}$. Therefore $\beta\delta$ belongs to $W^\perp$, and
so is proportional to $\gamma$. This means that $\delta$ divides
$\gamma$ in $D$. As $\gamma$ has degree $m$, it admits at most
$m$ non-associate divisors of degree $1$. Thus $a$ belongs to
a set of at most $m$ numbers.
\end{proof}
The next result is the second part of Proposition \ref{three:prop:concave}.
\begin{corollary}
\label{proofs:cor:lemma:injective}
There is a smallest integer $m\in\{0,\dots,n+1\}$ for which
$f(m)=n-m+1$. For this choice of $m$, the function $f$ is
monotonically increasing on $\{0,\dots,m\}$, while strictly
decreasing with $f(\ell)=n-\ell+1$ for $\ell\in\{m,\dots,n+1\}$.
\end{corollary}
\begin{proof}
The existence of $m$ follows from the fact that $f(n+1)=0$.
For that $m$, we have $\varphi(D_m^s)=E_{n-m}$. Thus for each
$\ell\in\{m,\dots,n+1\}$, we find $\varphi(D_\ell^s) =
D_{\ell-m}E_{n-m} = E_{n-\ell}$ and so $f(\ell)=n-\ell+1$.
It remains to prove that $f$ is monotonically increasing on
$\{0,\dots,m\}$. This is automatic if $m=0$. Otherwise,
since $f$ is concave on $\{0,\dots,n+1\}$, this amounts to showing
that $f(m-1)\le f(m)$. By the choice of $m$, the vector space
$W=\varphi(D_{m-1}^s)$ is a proper subspace of $E_{n-m+1}$.
Then, by Lemma \ref{proofs:lemma:injective}, there is some
$\delta\in D_1$ which is injective on $W$, thus $f(m)=\dim(D_1W)\ge
\dim(\delta W)=\dim(W)=f(m-1)$.
\end{proof}
Similarly, we will derive Proposition \ref{three:prop:avoiding}
from the following result.
\begin{prop}
\label{proofs:prop:avoiding}
Let $\ell\in\{0,\dots,n\}$ and let $S$ be a subspace
of $E_{n-\ell}$. Suppose that a subspace $B$ of
$E_n$ satisfies $\dim(B)\le n-\ell+1$
and $D_\ell B\not\subseteq S$. Then, there exist
$a_0,\dots,a_\ell\in\mathbb{Z}$ with
$\sum_{k=0}^\ell |a_k| \le (n+1)^\ell$
such that the differential operator
$\delta=\sum_{k=0}^\ell a_k\delta_0^{\ell-k}\delta_1^k\in D_\ell$
is injective on $B$ with $\delta B\not\subseteq S$.
\end{prop}
\begin{proof}
We proceed by induction on $\ell$. For $\ell=0$, the result
is automatic, it suffices to take $a_0=1$. Suppose that
$\ell\ge 1$ and set $S^*=\{P\in E_{n-\ell+1}\,;\,D_1P\subseteq S\}$.
Then $S^*$ is a subspace of $E_{n-\ell+1}$ and $D_{\ell-1}B
\not\subseteq S^*$. So, by induction, we may assume the existence
of $a^*_0,\dots,a^*_{\ell-1}\in\mathbb{Z}$ with
$\sum_{k=0}^{\ell-1} |a^*_k| \le (n+1)^{\ell-1}$ for which $\delta^* =
\sum_{k=0}^{\ell-1} a^*_k\delta_0^{\ell-1-k}\delta_1^k\in D_{\ell-1}$
is injective on $B$ with $\delta^* B\not\subseteq S^*$. Then,
$W=\delta^*B$ is a proper subspace of $E_{n-\ell+1}$ because its
dimension is $\dim(B)\le n-\ell+1$. Since $W\not\subseteq S^*$, we
also have $D_1W\not\subseteq S$, and so there is at most one
$a\in\mathbb{R}$ for which $\gamma=\delta_0+a\delta_1$ satisfies
$\gamma W\subseteq S$. By Lemma \ref{proofs:lemma:injective}, we
can therefore choose $a\in\mathbb{Z}$ with $|a|\le n-\ell+1\le n$ such
that $\gamma$ is injective on $W$ with $\gamma W\not\subseteq S$.
Then $\delta=\gamma\delta^*$ has the required properties.
\end{proof}
To deduce Proposition \ref{three:prop:avoiding}, we simply apply
the above result with $S=\psi_{n-\ell}(V)$, $B=\psi_n(A)$ and
note that, by virtue of \eqref{proofs:eq0}, we have $\delta\circ\psi_n
= n(n-1)\cdots(n-\ell+1)\psi_{n-\ell}\circ\tau_\mathbf{a}$, with
$\mathbf{a}=(a_0,\dots,a_\ell)$.
Finally, for the proof of Proposition \ref{three:prop:height},
we need to extend the notion of height on the homogeneous
components of $D$ and $E$. For each $m\ge 0$, we denote by
$\psi_m^*\colon\mathbb{R}^{m+1}\to D_m$ the isomorphism which is
dual to $\psi_m\colon\mathbb{R}^{m+1}\to E_m$ in the sense that
$\psi_m^*(\mathbf{y})\psi_m(\mathbf{x})=\mathbf{y}\cdot\mathbf{x}$ for any $\mathbf{x},\mathbf{y}
\in\mathbb{R}^{m+1}$. We say that a subspace $S$ of $D_m$
(resp.\ $E_m$) is defined over $\mathbb{Q}$ if it is generated
by elements of $\mathbb{Q}[\delta_0,\delta_1]_m$ (resp.\
$\mathbb{Q}[T_0,T_1]_m$) or equivalently if $(\psi^*_m)^{-1}(S)$
(resp.\ $(\psi_m)^{-1}(S)$ is a subspace $V$ of $\mathbb{R}^{m+1}$
defined over $\mathbb{Q}$, and we define its height $H(S)$
to be $H(V)$. Then the formula \eqref{heights:eq:duality}
translates into $H(S)=H(S^\perp)$ for each subspace $S$
of $E_m$ defined over $\mathbb{Q}$, and its orthogonal space
$S^\perp$ in $D_m$. Moreover, for each non-zero
$\delta\in \mathbb{Q}[\delta_0,\delta_1]_m$ and each $k\ge 0$,
the subspace $D_k\delta$ of $D_{k+m}$ is defined over $\mathbb{Q}$
of dimension $k+1$ and, by \cite[Proposition 5.2]{RW2004},
its height satisfies
\begin{equation}
\label{proofs:eq5}
H(D_k\delta)\asymp H(\langle\delta\rangle)^{k+1}
\end{equation}
with implied constants that do not depend on $\delta$.
In this setting, Proposition \ref{three:prop:height} follows
immediately from the next result upon setting $B=\psi_n(A)$ and
noting that our choice of height yields $H(D_t B)=H(\mathcal{U}^t(A))$
for each $t=0,\dots,n$.
\begin{prop}
\label{proofs:prop:height}
Let $j,\ell\ge 0$ be integers with $j+2\ell\le n$, and let
$B$ be a subspace of $E_n$ of dimension $j+1$ defined over
$\mathbb{Q}$. Set $d=\dim(D_\ell B)$ and suppose that $d \le j+\ell$.
Then, we have $0\le d-j-1< \ell\le n-d$ and there exists
a non-zero operator $\delta\in \mathbb{Q}[\delta_0,\delta_1]_d$
such that
\[
D_tB=(D_{n-d-t}\delta)^\perp,
\quad
\dim(D_tB)=d
\quad\mbox{and}\quad
H(D_tB)\asymp H(\langle\delta\rangle)^{n-d-t+1}
\]
for $t=d-j-1,\dots,n-d$, with implied constants depending
only on $n$. Moreover, for such $t$ and for $P\in E_n$, the
condition $D_tP\subseteq D_tB$ is independent of\/ $t$ and
amounts to $\delta P=0$.
\end{prop}
\begin{proof}
We may assume that $B=\psi_n(A)$ (with $A$ defined over
$\mathbb{Q}$), and then $f(t)=\dim(D_tB)$ for $t=0,\dots,n+1$.
Let $m$ be as in Corollary
\ref{proofs:cor:lemma:injective}, so that $f$ is
monotonically increasing on $\{0,\dots,m\}$. Since
$f(\ell) = d \le j+\ell < n-\ell+1$, this corollary gives
$\ell<m$. We deduce that
\[
f(0)=j+1 \le f(\ell)=d \le f(m)=n-m+1,
\]
thus $1\le d-j$, while the hypotheses yield $d-j\le\ell\le n-d$.
In turn this gives
\[
f(d-j) \le f(\ell)=f(0)+(d-j)-1.
\]
So, $f$ is not strictly increasing on $\{0,\dots,d-j\}$. Being
concave, it is therefore constant on $\{d-j-1,\dots,m\}$, equal
to $f(\ell)=d$. In particular, we obtain $m=n+1-f(m)=n-d+1$,
and so $f(n-d)=d$. This means that $D_{n-d}B$ has codimension
$1$ in $E_d$. As it is defined over $\mathbb{Q}$, we deduce that
$(D_{n-d}B)^\perp=\langle\delta\rangle$
for some non-zero $\delta\in \mathbb{Q}[\delta_0,\delta_1]_d$.
For each $t=d-j-1,\dots,n-d$,
the subspace $D_tB$ of $E_{n-t}$ has dimension $d$, while
$D_{n-d-t}\delta$ has codimension $d$ in $D_{n-t}$. Since
their product is
\[
(D_{n-d-t}\delta)(D_tB)=\delta D_{n-d}B=0,
\]
we deduce that $D_tB=(D_{n-d-t}\delta)^\perp$. Thus, using
\eqref{proofs:eq5}, we obtain
\[
H(D_tB)=H(D_{n-d-t}\delta)
\asymp H(\langle\delta\rangle)^{n-d-t+1}.
\]
Finally, for $P\in E_n$,
the condition $D_tP\subseteq D_tB$ may be rewritten as
$(D_{n-d-t}\delta)(D_t P)=0$, so it is equivalent to
$\delta P \in D_{n-d}^\perp=0$ (inside $E_{n-d}$).
\end{proof}
\section{Minimal points and properties $\mathcal{P}(j,\ell)$}
\label{sec:minimal}
From now on, we fix a positive integer $n$ and a real number $\xi$ with
$[\mathbb{Q}(\xi):\mathbb{Q}] > n$. Our goal is to establish an upper bound for
$\widehat{\lambda}_n(\xi)$ which depends only on $n$.
Since $[\mathbb{Q}(\xi):\mathbb{Q}] > n$, non-zero points $\mathbf{x},\mathbf{y}\in\mathbb{Z}^{n+1}$
which satisfy $L_\xi(\mathbf{x})=L_\xi(\mathbf{y})<1$ have a non-zero first
coordinate and come by pairs $\mathbf{y}=\pm \mathbf{x}$. Thus, for each large enough
real number $X\ge 1$, there is a unique pair of non-zero points $\pm \mathbf{x}$
in $\mathbb{Z}^{n+1}$ with $\norm{\mathbf{x}}\le X$ for which $L_\xi(\mathbf{x})<1$ is minimal.
We choose the one whose first coordinate is positive and, like Davenport
and Schmidt in \cite{DS1969}, we call it the \emph{minimal point
corresponding to $X$}. This differs slightly from their own definition,
but it plays the same role.
We order these minimal points in a
sequence $(\mathbf{x}_i)_{i\ge 0}$ by increasing norm. Then,
\begin{itemize}
\item their norms $X_i=\norm{\mathbf{x}_i}$ are positive and strictly increasing,
\item the quantities $L_i=L_\xi(\mathbf{x}_i)$ are strictly decreasing,
\item if $L_\xi(\mathbf{x})<L_i$ for some $i\ge 0$ and some non-zero $\mathbf{x}\in\mathbb{Z}^{n+1}$,
then $\norm{\mathbf{x}}\ge X_{i+1}$.
\end{itemize}
In terms of the associated sequences $(X_i)_{i\ge 0}$ and $(L_i)_{i\ge 0}$,
we have the well-known formulas
\begin{equation}
\label{minimal:eq:lambda}
\lambda_n(\xi)
= \limsup_{i\to\infty} \frac{-\log(L_i)}{\log(X_i)}
\quad\mbox{and}\quad
\widehat{\lambda}_n(\xi)
= \liminf_{i\to\infty} \frac{-\log(L_i)}{\log(X_{i+1})}
\end{equation}
which follow from the definition of these exponents given in the
introduction. In particular, if $\widehat{\lambda}_n(\xi)>\lambda$ for some
$\lambda\in\mathbb{R}$, then $L_i=o(X_{i+1}^{-\lambda})$ and a fortiori
\begin{equation}
\label{minimal:eq:Li}
L_i\ll X_{i+1}^{-\lambda},
\end{equation}
where from now on all implicit multiplicative constants are independent of $i$.
By construction, each minimal point $\mathbf{x}_i$ is primitive and so we have
\[
H(\Span{\mathbf{x}_i})=X_i.
\]
For subspaces spanned by two consecutive minimal points, a simple adaptation
of the proofs of \cite[Lemma~2]{DS1967} and \cite[Lemma~4.1]{Ro2004}
yields the following estimate.
\begin{lemma}
\label{minimal:lemma:HA1}
For each $i\ge 0$, we have $H\big(\Span{\mathbf{x}_i,\mathbf{x}_{i+1}}\big) =
\norm{\mathbf{x}_i\wedge\mathbf{x}_{i+1}}\asymp X_{i+1}L_i$.
\end{lemma}
More generally, we are
interested in the subspaces $\langle\mathbf{x}_i,\dots,\mathbf{x}_q\rangle$
of $\mathbb{R}^{n+1}$ spanned by minimal points with consecutive indices,
as in \cite[\S 3]{NPR2020} (see also \cite{MM2020}). It is well-known that,
for each $i\ge 0$, we have
\begin{equation}
\label{minimal:sum}
\langle\mathbf{x}_i,\mathbf{x}_{i+1},\dots\rangle
= \sum_{k=i}^\infty \langle \mathbf{x}_k\rangle
= \mathbb{R}^{n+1}
\end{equation}
because these are subspaces of $\mathbb{R}^{n+1}$ defined over $\mathbb{Q}$ which
contain $\lim_{k\to\infty}\norm{\mathbf{x}_k}^{-1}\mathbf{x}_k=\norm{\Xi}^{-1}\Xi$
where $\Xi=(1,\xi,\dots,\xi^n)$ has $\mathbb{Q}$-linearly independent coordinates.
This justifies the following construction.
\begin{definition}
\label{minimal:def:AY}
For each $i\ge 0$ and each $j=0,\dots,n-1$, we set
\[
\sigma_j(i)=q,
\quad
A_j(i)=\langle\mathbf{x}_i,\dots,\mathbf{x}_q\rangle
\quad\mbox{and}\quad
Y_j(i)=X_{q+1}
\]
where $q\ge i$ is the largest index for which $\dim\langle\mathbf{x}_i,\dots,\mathbf{x}_q\rangle=j+1$.
We also set
\[
A_n(i)=\mathbb{R}^{n+1} \quad\mbox{and}\quad Y_{-1}(i)=X_i.
\]
\end{definition}
So, for $j=0,\dots,n-1$, we have
\[
\dim(A_j(i))=j+1
\quad\mbox{and}\quad
A_{j+1}(i)=\langle\mathbf{x}_i,\dots,\mathbf{x}_q,\mathbf{x}_{q+1}\rangle
\quad
\text{where $q=\sigma_j(i)$.}
\]
For $j=0$, we note that $\sigma_0(i)=i$, $A_0(i)=\langle\mathbf{x}_i\rangle$
and $Y_0(i)=X_{i+1}$. For $j\ge 1$, the following notation is useful.
\begin{definition}
\label{minimal:def:I}
We denote by $I$ the set of indices $i\geq 1$ such that
$\mathbf{x}_{i-1},\mathbf{x}_i,\mathbf{x}_{i+1}$ are linearly independent. We say
that $i<j$ are \emph{consecutive elements} of $I$ or that $j$
is the \emph{successor} of $i$ in $I$ if $j$ is the smallest
element of $I$ with $j>i$.
\end{definition}
When $n=1$, the set $I$ is empty. However, when $n>1$, we
may form $q=\sigma_j(i)$ for each $i\ge 1$ and each
$j=1,\dots,n-1$. Since $\Span{\mathbf{x}_{q-1},\mathbf{x}_q}\subseteq
A_j(i)$ and $\mathbf{x}_{q+1} \notin A_j(i)$, we deduce that
$q\in I$. Thus $I$ is infinite. Moreover, if $i<j$ are
consecutive elements of $I$, we have
\[
\Span{\mathbf{x}_i,\mathbf{x}_{i+1}}=\dots=\Span{\mathbf{x}_{j-1},\mathbf{x}_j}
\neq \Span{\mathbf{x}_j,\mathbf{x}_{j+1}},
\]
Applying Lemma \ref{minimal:lemma:HA1}, we obtain the following useful
estimate.
\begin{lemma}
\label{minimal:XLXL}
Suppose that $n\ge 2$. Then $I$ is an infinite set and for each pair $i<j$ of
consecutive elements of $I$, we have $X_jL_{j-1}\asymp X_{i+1}L_i$.
\end{lemma}
The above also shows that, for each integer $i\ge 0$, we have
$\sigma_1(i)=j$ where $j$ is the smallest element of $I$ with $j>i$, so
$A_1(i)=A_1(j-1)$ and $A_2(i)=A_2(j-1)=\Span{\mathbf{x}_{j-1},\mathbf{x}_j,\mathbf{x}_{j+1}}$.
In the next sections, we will provide upper bound estimates for the height of
the subspaces $\mathcal{U}^\ell(A_j(i))$ when the following condition is fulfilled.
\begin{definition}
\label{minimal:def:Pjl}
Let $j,\ell\in\{0,\dots,n\}$. We say that property $\mathcal{P}(j,\ell)$ holds
if, for each sufficiently large integer $i\ge 0$ and each $m=0,\dots,j$, we have
$\dim\mathcal{U}^\ell(A_m(i))\ge m+\ell+1$.
\end{definition}
Of course, this depends on our fixed choice of $\xi$ and $n$. Clearly
$\mathcal{P}(n,0)$ holds, because $\mathcal{U}^0(A_m(i))=A_m(i)$
has dimension $m+1$ for each $i\ge 0$ and each $m=0,\dots,n$.
Moreover $\mathcal{P}(j,\ell)$ implies $\mathcal{P}(j-1,\ell)$ if $j>0$. The next result
provides a further crucial implication.
\begin{prop}
\label{minimal:prop:Pjl}
Suppose that property $\mathcal{P}(j,\ell)$ holds for some $j,\ell\in\{0,\dots,n\}$.
Then we have $j+2\ell\le n$. If moreover $\ell>0$, then $\mathcal{P}(j+1,\ell-1)$
holds as well.
\end{prop}
\begin{proof}
By hypothesis, $\mathcal{U}^\ell(A_j(i))$ is a subspace of $\mathbb{R}^{n-\ell+1}$ of dimension
at least $j+\ell+1$ for each large enough $i$. Comparing dimensions yields
$j+2\ell\le n$. Now, suppose that $\ell>0$, and set $A=A_m(i)$ for a choice
of integers $m\in\{0,\dots,j+1\}$ and $i\ge 0$. By Corollary \ref{three:cor}(i),
we have
\[
\min\{\dim\mathcal{U}^\ell(B), m+\ell\} \le \dim\mathcal{U}^{\ell-1}(A)
\]
for any subspace $B$ of $A$. If $m\le j$, we choose $B=A_m(i)$. Otherwise,
we choose $B=A_{m-1}(i)$. Then, assuming $i$ large enough, we have
$\dim\mathcal{U}^\ell(B) \ge \dim(B)+\ell \ge m+\ell$ because of $\mathcal{P}(j,\ell)$,
and so $\dim\mathcal{U}^{\ell-1}(A) \ge m+\ell$. Thus $\mathcal{P}(j+1,\ell-1)$ holds.
\end{proof}
\section{Property $\mathcal{P}(0,\ell)$}
\label{sec:P0}
We keep the notation of the preceding section, and fix a real
number $\lambda$ with $0<\lambda<\widehat{\lambda}_n(\xi)$, thus
$\lambda<1$. In this section, we derive useful consequences
of the assumption that, for some
$j\ge 0$ and $\ell\ge 1$, we have $\mathcal{P}(j,\ell-1)$ but not
$\mathcal{P}(j,\ell)$. As an example of application, we recover an
important result of Badziahin and Schleischitz from \cite{BS2021}
which yields $\mathcal{P}(0,\ell)$ under a simple condition on $\ell$.
We recall our convention that all implicit
multiplicative constants are independent of $i$, and we start with
two general lemmas.
\begin{lemma}
\label{P0:lemma:A}
Let $0\le j,\ell\le n$ be integers. Suppose that
\[
d:=\liminf_{i\to\infty} \dim\mathcal{U}^\ell(A_j(i)) \le n-\ell.
\]
Then there are arbitrarily large integers $i\ge 1$ for which
\begin{equation}
\label{P0:lemma:A:eq}
\dim\mathcal{U}^\ell(A_j(i))=d \quad\mbox{and}\quad {\mathcal{U}}^\ell(\mathbf{x}_{i-1})\not\subseteq \mathcal{U}^\ell(A_j(i)).
\end{equation}
\end{lemma}
\begin{proof}
Set $V_i=\mathcal{U}^\ell(A_j(i))$ for each $i\ge 0$, and let $E$ denote the infinite
set of integers $i\ge 0$ for which $\dim(V_i)=d$. Choose $p\in E$ large enough so that
$\dim(V_i)\ge d$ for each $i\ge p$. Using \eqref{minimal:sum},
we find
\[
\sum_{i=p}^\infty V_i
= {\mathcal{U}}^{\ell}(\Span{\mathbf{x}_p,\mathbf{x}_{p+1},\dots})
= {\mathcal{U}}^{\ell}(\mathbb{R}^{n+1})
= \mathbb{R}^{n-\ell+1}.
\]
Thus, there exists a smallest $q\in E$ with $q\ge p$ such that
$\sum_{i=p}^q V_i = \mathbb{R}^{n-\ell+1}$. Since $V_p$ has dimension
$d\le n-\ell$, it is a proper subspace of $\mathbb{R}^{n-\ell+1}$. So, we
must have $q>p$, and thus $V_{q-1}\neq V_{q}$ by the choice of $q$.
As $\dim(V_{q-1})\ge d=\dim(V_{q})$, this means that
$V_{q-1}\not\subset V_{q}$ and therefore
\eqref{P0:lemma:A:eq} holds for $i=q>p$.
\end{proof}
For any integer $m\ge 1$, any subspace of $\mathbb{R}^m$ defined over
$\mathbb{Q}$ has height at least $1$. The next lemma provides an instance
where this lower bound can be improved.
\begin{lemma}
\label{P0:lemma:S}
Let $\ell\in\{0,\dots,n\}$ and let $V$ be a subspace of $\mathbb{R}^{n+1-\ell}$
defined over $\mathbb{Q}$. Suppose that ${\mathcal{U}}^\ell(\mathbf{x}_i)\subseteq V$ and
that ${\mathcal{U}}^\ell(\mathbf{x}_{i-1})\not\subseteq V$ for some $i\ge 1$. Then,
$1\ll H(V)L_{i-1}$.
\end{lemma}
\begin{proof}
Choose $m\in\{0,\dots,\ell\}$ such that $\mathbf{y}:=\mathbf{x}_{i-1}^{(m,\ell)}
\notin V$ and set $U=\langle \mathbf{y},\mathbf{z}\rangle$ where
$\mathbf{z}:=\mathbf{x}_{i}^{(m,\ell)}$. Then, we have $U\cap V=\langle \mathbf{z}\rangle$
and so $H(U\cap V)=g^{-1}\|\mathbf{z}\|\asymp g^{-1}X_i$ where $g$
denotes the gcd of the coordinates of $\mathbf{z}$. Since $\mathbf{y}$ and $g^{-1}\mathbf{z}$
are integer points of $U$, we also find
\[
H(U)
\le \|\mathbf{y}\wedge g^{-1}\mathbf{z}\|
=g^{-1}\|\mathbf{y}\wedge\mathbf{z}\|
\ll g^{-1}X_iL_{i-1},
\]
thus $1\le H(U+V) \ll H(U\cap V)^{-1}H(U)H(V)\ll H(V)L_{i-1}$.
\end{proof}
We now come to the main result of this section.
\begin{prop}
\label{P0:prop:I}
Let $j\ge 0$ and $\ell\ge 1$ be integers with $j+2\ell\le n$. Suppose that
$\mathcal{P}(j,\ell-1)$ holds but not $\mathcal{P}(j,\ell)$. Then, there are arbitrarily large
values of $i\ge 1$ for which
\begin{equation}
\label{P0:prop:I:eq1}
\dim\mathcal{U}^\ell(A_j(i)) = \ell+j \quad\mbox{and}\quad
{\mathcal{U}}^\ell(\mathbf{x}_{i-1}) \not\subseteq \mathcal{U}^\ell(A_j(i)) \varsubsetneq \mathbb{R}^{n-\ell+1}.
\end{equation}
For those $i$, we further have
\begin{equation}
\label{P0:prop:I:eq2}
1\ll H\big(\mathcal{U}^{\ell-1}(A_j(i))\big) L_{i-1}^{n-j-2\ell+2}.
\end{equation}
\end{prop}
\begin{proof}
Since $\mathcal{P}(j,\ell-1)$ holds, Corollary \ref{three:cor}(ii)
gives
\[
\dim\mathcal{U}^\ell(A_j(i)) \ge \min\{ \ell+j,\, n-\ell+1 \} = \ell+j
\]
for each sufficiently large $i$. Since $\mathcal{P}(j,\ell)$ does not hold,
we conclude that
\[
\liminf_{i\to\infty} \dim\mathcal{U}^\ell(A_j(i)) =\ell+j \le n-\ell.
\]
Then, Lemma \ref{P0:lemma:A} provides infinitely
many $i$ for which \eqref{P0:prop:I:eq1} holds.
For such $i$, Proposition \ref{three:prop:height} applies
with $A=A_j(i)$, $d=\ell+j$ and $t\in\{\ell-1,\ell\}$. Setting
$V={\mathcal{U}}^{n-\ell-j}(A)$, it gives
\[
H({\mathcal{U}}^{\ell-1}(A)) \asymp H(V)^{n-j-2\ell+2}
\quad\mbox{and}\quad
{\mathcal{U}}^{n-\ell-j}(\mathbf{x}_{i-1})\not\subseteq V.
\]
Since ${\mathcal{U}}^{n-\ell-j}(\mathbf{x}_i)\subseteq V$, Lemma \ref{P0:lemma:S}
yields $1\ll H(V)L_{i-1}$. Then \eqref{P0:prop:I:eq2} follows.
\end{proof}
The following restatement of \cite[Lemma 3.1]{BS2021} is
central to the present paper.
\begin{prop}[Badziahin-Schleischitz, 2021]
\label{P0:prop:BS}
Suppose that $\widehat{\lambda}_n(\xi)>1/(n-\ell+1)$ for some integer $\ell$ with
$0\le \ell\le n/2$. Then $\mathcal{P}(0,\ell)$ holds. More precisely, we have
$\dim{\mathcal{U}}^\ell(\mathbf{x}_i)=\ell+1$ for each sufficiently large $i$ .
\end{prop}
The proof given in \cite{BS2021} is based on a method of Laurent from \cite{La2003}.
To illustrate Proposition \ref{P0:prop:I}, we give the following alternative
argument.
\begin{proof}
Since $\mathcal{P}(0,0)$ holds, there is a largest integer $m\ge 0$ for which
$\mathcal{P}(0,m)$ holds. Suppose that $m<\ell$. Since $\mathcal{P}(0,m+1)$ does not hold
and $2(m+1)\le n$, Proposition \ref{P0:prop:I} shows the existence of
arbitrarily large values of $i$ for which
\[
1 \ll H({\mathcal{U}}^m(\mathbf{x}_i))L_{i-1}^{n-2m}.
\]
As \eqref{three:eq:HUx} gives $H({\mathcal{U}}^m(\mathbf{x}_i))\ll X_iL_i^m$,
this yields $1\ll X_iL_{i-1}^{n-m}$. Then, by the formulas
\eqref{minimal:eq:lambda}, we deduce that $\widehat{\lambda}_n(\xi)
\le 1/(n-m)\le 1/(n-\ell+1)$. This contradiction shows that
$\mathcal{P}(0,\ell)$ holds. The second assertion of the lemma follows since
$\dim{\mathcal{U}}^\ell(\mathbf{x}_i) \le \ell+1$ for each $i\ge 0$.
\end{proof}
If $\mathcal{P}(0,\ell)$ holds for some $\ell\ge 1$, then $\dim \mathcal{U}^\ell(\mathbf{x}_i)
=\ell+1$ for each large enough $i$ and, for these $i$, we obtain
$1\le H(\mathcal{U}^\ell(\mathbf{x}_i))\ll X_iL_i^\ell \ll X_iX_{i+1}^{-\ell\lambda}$
by \eqref{three:eq:HUx} and our choice of $\lambda$.
This implies that $\lambda\le 1/\ell$ and so
$\widehat{\lambda}_n(\xi)\le 1/\ell$. Combined with Proposition \ref{P0:prop:I},
this observation yields
\[
\widehat{\lambda}_n(\xi)\le \max\{1/(n-\ell+1),1/\ell\}
\]
for each integer $\ell$ with $1\le \ell\le n/2$. Thus, if $n\ge 2$, we
obtain $\widehat{\lambda}_n(\xi)\le 1/\lfloor n/2\rfloor$, which is the estimate
of Davenport and Schmidt from \cite{DS1969} mentioned in the introduction.
We conclude with another consequence of $\mathcal{P}(0,\ell)$.
\begin{lemma}
\label{P0:lemma:Fbis}
Suppose that $\mathcal{P}(0,\ell)$ holds for some integer $\ell\ge 1$ with $2\ell<n$. Then,
there are infinitely many $i\ge 1$ for which ${\mathcal{U}}^\ell(\mathbf{x}_{i-1})\not\subseteq
{\mathcal{U}}^\ell(\mathbf{x}_i)$ and, for those $i$, we have
\begin{equation*}
\label{P0:lemma:Fbis:eq}
X_{i+1}^\theta \ll X_i
\quad\text{where}\quad
\theta=\frac{\ell\lambda}{1-\lambda}
\end{equation*}
\end{lemma}
\begin{proof}
Since $\mathcal{P}(0,\ell)$ holds, we have $\dim{\mathcal{U}}^\ell(\mathbf{x}_i)=\ell+1\le n-\ell$
for each sufficiently large $i$. By Lemma \ref{P0:lemma:A}, we deduce that
both $\dim{\mathcal{U}}^\ell(\mathbf{x}_i)=\ell+1$ and ${\mathcal{U}}^\ell(\mathbf{x}_{i-1})\not\subseteq
{\mathcal{U}}^\ell(\mathbf{x}_i)$ for infinitely many $i\ge 1$. For those $i$, we have
$H({\mathcal{U}}^\ell(\mathbf{x}_i))\ll X_iL_i^\ell$ by \eqref{three:eq:HUx}, and Lemma
\ref{P0:lemma:S} gives
\[
1 \ll H({\mathcal{U}}^\ell(\mathbf{x}_i))L_{i-1} \ll X_iL_i^\ell L_{i-1}
\ll X_i^{1-\lambda}X_{i+1}^{-\ell\lambda},
\]
so $X_{i+1}^\theta \ll X_i$.
\end{proof}
\section{Property $\mathcal{P}(1,\ell)$}
\label{sec:P1}
With the notation of the preceding sections (including the choice of
$\lambda$), we provide below a sufficient condition on $\ell$ and $\lambda$
for property $\mathcal{P}(1,\ell)$ to hold. We also establish consequences of
$\mathcal{P}(1,\ell)$. The results of this section provide a first step towards the
proof of Theorems \ref{intro:thm:impair} and \ref{intro:thm:pair}. We start
with a crude height estimate.
\begin{lemma}
\label{P1:lemma1}
If $\mathcal{P}(1,\ell)$ holds for some $\ell\ge 0$, then
$H(\mathcal{U}^\ell(\mathbf{x}_i,\mathbf{x}_{i+1}))\ll X_{i+1}^{1-(\ell+1)\lambda}$.
\end{lemma}
\begin{proof}
For given $i\ge 0$ and $\ell\in\{0,\dots,n\}$, the subspace
$\mathcal{U}^\ell(\mathbf{x}_i,\mathbf{x}_{i+1})$ of $\mathbb{R}^{n-\ell+1}$ is generated by
the set $\{\mathbf{x}_i^{(0,\ell)},\dots,\mathbf{x}_i^{(\ell,\ell)},
\mathbf{x}_{i+1}^{(0,\ell)},\dots,\mathbf{x}_{i+1}^{(\ell,\ell)}\}$ which consists
of integer points $\mathbf{y}$ with $\norm{\mathbf{y}}\le X_{i+1}$ and
$L_\xi(\mathbf{y})\ll L_i$. If $\mathcal{P}(1,\ell)$ holds
and $i$ is large enough, this space has dimension $d\ge \ell+2$,
and so Lemma \ref{heights:lemma} gives $H(\mathcal{U}^\ell(\mathbf{x}_i,\mathbf{x}_{i+1}))
\ll X_{i+1}L_i^{d-1}\le X_{i+1}^{1-(\ell+1)\lambda}$.
\end{proof}
By the above, property $\mathcal{P}(1,\ell)$ implies that
$\lambda\le 1/(\ell+1)$ and so $\widehat{\lambda}_n(\xi)\le 1/(\ell+1)$. The
next lemma provides finer estimates.
\begin{lemma}
\label{P1:lemma2}
Suppose that $n\ge 2$ and that $\mathcal{P}(1,\ell)$ holds for some integer
$\ell\ge 0$. Then, for each pair of consecutive elements $i<j$ of $I$,
we have
\[
H(\mathcal{U}^\ell(\mathbf{x}_i,\mathbf{x}_{i+1}))
\ll X_{j+1}^{-\ell\lambda}X_{i+1}^{1-\lambda}
\quad\mbox{and}\quad
H(\mathcal{U}^\ell(\mathbf{x}_{i-1},\mathbf{x}_i,\mathbf{x}_{i+1}))
\ll X_{j+1}^{-\ell\lambda}X_{i+1}^{1-\lambda}X_i^{-e\lambda},
\]
where $e=\dim\mathcal{U}^\ell(\mathbf{x}_{i-1},\mathbf{x}_i,\mathbf{x}_{i+1})-\ell-2$.
\end{lemma}
\begin{proof}
For a pair of consecutive elements $i<j$ of $I$, we have $\Span{\mathbf{x}_i,\mathbf{x}_{i+1}}=
\Span{\mathbf{x}_{j-1},\mathbf{x}_j}$. Thus we have a chain of subspaces
\begin{align*}
U=\mathcal{U}^\ell(\mathbf{x}_j)
&\subseteq
V=\mathcal{U}^\ell(\mathbf{x}_i,\mathbf{x}_{i+1})=U+\mathcal{U}^\ell(\mathbf{x}_{j-1}) \\
&\subseteq
W=\mathcal{U}^\ell(\mathbf{x}_{i-1},\mathbf{x}_i,\mathbf{x}_{i+1})=V+\mathcal{U}^\ell(\mathbf{x}_{i-1}).
\end{align*}
If $i$ is large enough, then by property $\mathcal{P}(1,\ell)$ we have $\dim(U)=\ell+1$,
$\dim(V)=\dim(U)+a$ and $\dim(W)=\dim(V)+b$ for some integers $a\ge 1$
and $b\ge0$. As each subspace $\mathcal{U}^\ell(\mathbf{x}_h)$ is generated by integer
points $\mathbf{y}$ with $\norm{\mathbf{y}}\le X_h$ and $L_\xi(\mathbf{y})\ll L_h$, Lemma
\ref{heights:lemma} gives
\[
H(V) \ll X_j L_j^\ell L_{j-1}^a
\quad\mbox{and}\quad
H(W) \ll X_j L_j^\ell L_{j-1}^a L_{i-1}^b.
\]
Finally, by Lemma \ref{minimal:XLXL}, we have $X_jL_{j-1}\asymp X_{i+1}L_i$.
Since $a\ge 1$, we deduce that
\[
H(V) \ll L_j^\ell X_j L_{j-1} \asymp L_j^\ell X_{i+1}L_i
\quad\mbox{and}\quad
H(W) \ll L_j^\ell X_j L_{j-1} L_{i-1}^e \asymp L_j^\ell X_{i+1}L_i L_{i-1}^e,
\]
where $e=a+b-1=\dim(W)-\ell-2\ge 0$. The conclusion follows.
\end{proof}
Under the hypotheses of Lemma \ref{P1:lemma2}, we have
$1\le H(\mathcal{U}^\ell(\mathbf{x}_i,\mathbf{x}_{i+1})) \ll X_{j+1}^{-\ell\lambda}X_{i+1}^{1-\lambda}$
and so $X_{j+1}^\theta\ll X_{i+1}$ with $\theta=\ell\lambda/(1-\lambda)$,
for each pair of consecutive elements $i<j$ of $I$. A fortiori, this implies that
$X_{i+1}^\theta\ll X_{i}$ for each $i\ge 0$. The following result yields a weaker estimate but assumes $\mathcal{P}(1,\ell-1)$ instead of $\mathcal{P}(1,\ell)$.
\begin{lemma}
\label{P1:lemma3}
Suppose that $\mathcal{P}(1,\ell-1)$ holds for some integer $\ell$ with
$1\le \ell\le n/2$, and that $\widehat{\lambda}_n(\xi) > 1/(2\ell)$.
Then, for each $i\ge 0$, we have $X_iL_i^\ell\gg 1$ and so
$X_{i+1}^{\ell\lambda}\ll X_i$ .
\end{lemma}
\begin{proof}
By definition of property $\mathcal{P}(1,\ell-1)$ we have,
\[
\dim\mathcal{U}^{\ell-1}(\mathbf{x}_{i-1},\mathbf{x}_i)\ge \ell+1
\quad\mbox{and}\quad
\dim\mathcal{U}^{\ell-1}(\mathbf{x}_i)\ge \ell
\]
for each sufficiently large $i\ge 1$. Fix such an integer $i$. Then,
Corollary \ref{three:cor}(ii) gives
\[
\dim\mathcal{U}^{\ell}(\mathbf{x}_i)\ge \min\{\dim\mathcal{U}^{\ell-1}(\mathbf{x}_i),n-\ell+1\}\ge \ell.
\]
If $\dim\mathcal{U}^{\ell}(\mathbf{x}_i)\ge \ell+1$, then $1\le H(\mathcal{U}^{\ell}(\mathbf{x}_i)) \ll X_iL_i^{\ell}$
and we are done. Otherwise, $\mathcal{U}^\ell(\mathbf{x}_i)$ has dimension $\ell$. Then,
Proposition \ref{three:prop:height} applies with $A=\Span{\mathbf{x}_i}$,
$j=0$, $d=\ell$, $V=\mathcal{U}^{n-\ell}(\mathbf{x}_i)$ and $t=\ell-1$. It gives
\[
\dim\mathcal{U}^{\ell-1}(\mathbf{x}_i)=\ell
\quad\mbox{and}\quad
H(\mathcal{U}^{\ell-1}(\mathbf{x}_i)) \asymp H(V)^{n-2\ell+2} \ge H(V)^2 .
\]
Moreover, since $\dim\mathcal{U}^{\ell-1}(\mathbf{x}_{i-1},\mathbf{x}_i)>\ell$, we have
$\mathcal{U}^{\ell-1}(\mathbf{x}_{i-1}) \not\subseteq\mathcal{U}^{\ell-1}(\mathbf{x}_i)$ and so the
proposition also gives $\mathcal{U}^{n-\ell}(\mathbf{x}_{i-1}) \not\subseteq V$.
By Lemma \ref{P0:lemma:S}, this in turn yields $H(V)L_{i-1}\gg 1$, thus
we find
\[
1\ll H(\mathcal{U}^{\ell-1}(\mathbf{x}_i))L_{i-1}^2 \ll X_iL_i^{\ell-1}L_{i-1}^2 \ll X_i^{1-2\lambda}L_i^{\ell-1}.
\]
As $\widehat{\lambda}_n(\xi) > 1/(2\ell)$, we may assume that $\lambda\ge 1/(2\ell)$.
This gives $1\ll X_i^{1-1/\ell}L_i^{\ell-1}$ and so $1\ll X_iL_i^\ell$.
\end{proof}
We now come to the main result of this section.
\begin{prop}
\label{P1:prop}
Suppose that, for some integer $\ell\ge 1$ with $1+2\ell\le n$, property $\mathcal{P}(0,\ell)$
holds but not $\mathcal{P}(1,\ell)$. Then we have $0\le 1-(n-\ell)\lambda-\ell\lambda^2$.
\end{prop}
\begin{proof}
Since $\mathcal{P}(0,\ell)$ holds, we have $\ell+1=\dim\mathcal{U}^\ell(A_0(i))
\le \dim\mathcal{U}^\ell(A_1(i))$ for each sufficiently large $i\ge 0$. Since $\mathcal{P}(1,\ell)$
does not hold, we also have $\dim\mathcal{U}^\ell(A_1(i))\le \ell+1$ for arbitrarily large
values of $i$. Thus, Lemma \ref{P0:lemma:A} applies with $j=1$ and $d=\ell+1$,
and so there are arbitrarily large integers $i\ge 1$ with
\[
\dim\mathcal{U}^\ell(A_1(i))=\ell+1
\quad\mbox{and}\quad
\mathcal{U}^\ell(\mathbf{x}_{i-1})\not\subseteq \mathcal{U}^\ell(A_1(i))\subseteq\mathbb{R}^{n-\ell+1}.
\]
For these $i$, Proposition \ref{three:prop:height} applies with $A=A_1(i)$,
$j=1$, $d=\ell+1$ and any $t\in\{\ell-1,\ell\}$. Setting
$V=\mathcal{U}^{n-\ell-1}(A_1(i))$, it gives
\[
H\big(\mathcal{U}^{\ell-1}(A_1(i))\big)\asymp H(V)^{n-2\ell+1}
\quad\mbox{and}\quad
H\big(\mathcal{U}^{\ell}(A_1(i))\big)\asymp H(V)^{n-2\ell}.
\]
Since $\mathcal{U}^\ell(\mathbf{x}_{i-1})\not\subseteq \mathcal{U}^\ell(A_1(i))$, it also gives
$\mathcal{U}^{n-\ell-1}(\mathbf{x}_{i-1})\not\subseteq V$. By Lemma \ref{P0:lemma:S},
this implies that $1\ll H(V)L_{i-1}\ll H(V)X_i^{-\lambda}$ and so
\[
X_i^\lambda \ll H(V).
\]
Because of $\mathcal{P}(0,\ell)$, the subspace $\mathcal{U}^\ell(\mathbf{x}_i)$ of
$\mathcal{U}^\ell(A_1(i))$ has dimension $\ell+1$ when $i$ is large enough, and
then it coincides with $\mathcal{U}^\ell(A_1(i))$. Moreover, $\mathcal{P}(1,\ell-1)$ holds
by Proposition \ref{minimal:prop:Pjl}. Thus using \eqref{three:eq:HUx} and
Lemma \ref{P1:lemma1}, the above estimates imply
\begin{align*}
X_i^{(n-2\ell)\lambda}
&\ll H(\mathcal{U}^\ell(\mathbf{x}_i))\ll X_iL_i^\ell\ll X_iX_{i+1}^{-\ell\lambda},\\
X_i^{(n-2\ell+1)\lambda}
&\ll H(\mathcal{U}^{\ell-1}(\mathbf{x}_i,\mathbf{x}_{i+1}))\ll X_{i+1}^{1-\ell\lambda}.
\end{align*}
The second row of estimates implies $1-\ell\lambda>0$ and provides a
lower bound for $X_{i+1}$ in terms of $X_i$. Substituting it in the first row
and comparing powers of $X_i$, we deduce that
\[
(1-\ell\lambda)(n-2\ell)\lambda
\le (1-\ell\lambda)-(\ell\lambda)(n-2\ell+1)\lambda,
\]
which after simplications reduces to $0\le 1-(n-\ell)\lambda-\ell\lambda^2$.
\end{proof}
\begin{cor}
\label{P1:cor}
Suppose that $n\ge 3$. Let $\ell$ be an integer with $1\le \ell < n/2$
and let $\rho$ denote the unique positive root of the polynomial
$P(x)=1-(n-\ell)x-\ell x^2$. Then, we have
$\rho>1/(n-\ell+1)$. If $\widehat{\lambda}_n(\xi)>\rho$, then $\mathcal{P}(1,\ell)$ holds.
\end{cor}
\begin{proof}
Let $k=n-\ell+1$. Since $P(1/k)=(n-2\ell+1)/k^2>0$, we have $\rho>1/k$.
If $\widehat{\lambda}_n(\xi)>\rho$, we may assume that $\lambda>\rho$. Then
we have $\lambda>1/k=1/(n-\ell+1)$ and thus property $\mathcal{P}(0,\ell)$ holds
by Proposition \ref{P0:prop:BS}. Since $\lambda>\rho$, we also have
$P(\lambda)<0$, and so the preceding proposition implies that $\mathcal{P}(1,\ell)$
holds.
\end{proof}
We remarked after Lemma \ref{P1:lemma1} that property $\mathcal{P}(1,\ell)$ implies
$\widehat{\lambda}_n(\xi)\le1/(\ell+1)$. Thus, with the notation and hypotheses of the
above corollary, we obtain
\[
\widehat{\lambda}_n(\xi)\le \max\{1/(\ell+1),\rho\}.
\]
If $n=2m+1\ge 3$ is odd, we may choose $\ell=m$. Then $\rho$ is the positive
root of $P(x)=1-(m+1)x-mx^2$, denoted by $\alpha_m$ in the statement of
Theorem \ref{intro:thm:impair}. As $P(1/(m+1))<0$, this yields
$\widehat{\lambda}_n(\xi)\le 1/(m+1)=\lceil n/2\rceil^{-1}$ which is the main
result of Laurent in \cite{La2003}.
\section{First general height estimates}
\label{sec:first}
With the notation of the preceding section (including the choice of
$\lambda$), we first show that property $\mathcal{P}(j,\ell)$ yields
special bases for the subspaces $\mathcal{U}^\ell(A_j(i))$. Then, we
deduce an upper bound on the height of these subspaces in terms of
the quantities $Y_j(i)$ from Definition \ref{minimal:def:AY}.
\begin{lemma}
\label{first:lemma:basis}
Suppose that $\mathcal{P}(j,\ell)$ holds for some integers $j,\ell\in\{0,\dots,n\}$.
For each large enough integer $i\ge 0$ and each integer $q\ge i$ such that
$A_j(i)=\langle \mathbf{x}_i,\dots,\mathbf{x}_q\rangle$, there exist an integer $e\ge 0$
and a basis $\{\mathbf{y}_0,\mathbf{y}_1,\dots,\mathbf{y}_{\ell+j+e}\}$ of $\mathcal{U}^\ell(A_j(i))$
made of points of $\mathbb{Z}^{n+1-\ell}$ of norm $\le X_q$ with
\[
\begin{cases}
L_\xi(\mathbf{y}_m)\ll X_{q+1}^{-\lambda} &\text{for \ $0\le m\le \ell$,}\\
L_\xi(\mathbf{y}_{\ell+m})\ll Y_{j-m}(i)^{-\lambda} &\text{for \ $1\le m\le j$,}\\
L_\xi(\mathbf{y}_{\ell+j+m})\ll Y_{0}(i)^{-\lambda} &\text{for \ $1\le m\le e$.}
\end{cases}
\]
\end{lemma}
\begin{proof}
We proceed by induction on $j$. If $j=0$, we have $q=i$, and
$\mathcal{U}^{\ell}(A_0(i))={\mathcal{U}}^\ell(\mathbf{x}_i)$ has dimension $\ell+1$ for $i$
large enough. Then the points $\mathbf{y}_m=\mathbf{x}_i^{(m,\ell)}$ for $m=0,\dots,\ell$
have the required properties. Now suppose that $j\ge 1$. For $i$ large enough,
$\mathcal{U}^\ell(A_j(i))$ has dimension $\ge \ell+j+1$. Choose $q\ge i$ such that
$A_j(i)=\langle \mathbf{x}_i,\dots,\mathbf{x}_q\rangle$, and choose $p$ minimal with
$i<p\le q$ such that $\dim\langle \mathbf{x}_p,\dots,\mathbf{x}_q\rangle=j$ or equivalently
such that $A_{j-1}(p)=\langle \mathbf{x}_p,\dots,\mathbf{x}_q\rangle$. Since
$\mathcal{P}(j-1,\ell)$ holds, we may assume by induction that,
when $i$ is large enough, the vector space
$\mathcal{U}^\ell(A_{j-1}(p))$ contains linearly independent points
$\mathbf{y}_0,\mathbf{y}_1,\dots,\mathbf{y}_{\ell+j-1}$ of $\mathbb{Z}^{n+1-\ell}$ of norm $\le X_q$
with $L_\xi(\mathbf{y}_m)\ll X_{q+1}^{-\lambda}$ for $0\le m\le \ell$ and
$L_\xi(\mathbf{y}_{\ell+m})\ll Y_{j-1-m}(p)^{-\lambda}$ for $1\le m\le j-1$.
Since $\mathbf{x}_{p-1}\notin\langle\mathbf{x}_p,\dots,\mathbf{x}_q\rangle$, we have
\[
\dim\langle\mathbf{x}_i,\dots,\mathbf{x}_r\rangle
\ge 1+\dim\langle\mathbf{x}_p,\dots,\mathbf{x}_r\rangle
\]
for each $r=p,\dots,q$. Thus, for each integer $m$ with $1\le m\le j-1$,
we have $\sigma_m(i) \le \sigma_{m-1}(p) < q$, and so $Y_m(i)\le Y_{m-1}(p)$.
By the above, this means that $L_\xi(\mathbf{y}_{\ell+m})\ll Y_{j-m}(i)^{-\lambda}$
for $1\le m\le j-1$. Since $\{\mathbf{y}_0,\mathbf{y}_1,\dots,\mathbf{y}_{\ell+j-1}\}$
is a linearly independent subset of $\mathcal{U}^\ell(A_j(i))$, we may complete
it to a basis $\{\mathbf{y}_0,\mathbf{y}_1,\dots,\mathbf{y}_{\ell+j+e}\}$ for some $e\ge 0$
by adding $e+1$ points of the form $\mathbf{x}_h^{(s,\ell)}$ with $i\le h\le q$
and $0\le s\le \ell$. These new points belong to $\mathbb{Z}^{n+1-\ell}$,
have norm $\le X_q$, and satisfy $L_\xi(\mathbf{y}_{\ell+j+m})\ll L_i
\ll Y_{0}(i)^{-\lambda}$ for $0\le m\le e$.
\end{proof}
\begin{prop}
\label{first:prop}
Suppose that $\mathcal{P}(j,\ell)$ holds for some integers $j,\ell\in\{0,\dots,n\}$
with $j\ge 1$. For each large enough $i\ge 0$, we have
\[
H\big(\mathcal{U}^\ell(A_j(i))\big)
\ll Y_{j-1}(i)^{1-\ell\lambda}
\Big( \prod_{m=1}^{j} Y_{j-m}(i)^{-\lambda} \Big)
Y_0(i)^{-e\lambda}
\]
with $e=\dim\mathcal{U}^\ell(A_j(i))-\ell-j-1\ge 0$.
\end{prop}
\begin{proof}
For given $i\ge 0$, choose $q\ge i$ minimal such that
$A_j(i)=\langle\mathbf{x}_i,\dots,\mathbf{x}_q\rangle$. Then, assuming $i$ large
enough so that $e\ge 0$, consider the basis $\{\mathbf{y}_0,\mathbf{y}_1,\dots,\mathbf{y}_{\ell+j+e}\}$
of $\mathcal{U}^\ell(A_j(i))$ provided by Lemma \ref{first:lemma:basis}.
By the choice of $q$, we have $X_q=Y_{j-1}(i)\le X_{q+1}$ and so
$L_\xi(\mathbf{y}_m)\ll Y_{j-1}(i)^{-\lambda}$ for $0\le m\le \ell$.
Since this basis consists of integer points, we also have
\[
H\big(\mathcal{U}^\ell(A_j(i))\big)
\le \|\mathbf{y}_0\wedge\mathbf{y}_1\wedge\cdots\wedge\mathbf{y}_{\ell+j+e}\|.
\]
We conclude by applying Lemma \ref{heights:lemma} along with
the estimates of Lemma \ref{first:lemma:basis}.
\end{proof}
When $\ell=0$, property $\mathcal{P}(j,\ell)$ holds and we obtain the following
estimate.
\begin{corollary}
\label{first:cor:HA}
$\displaystyle H(A_j(i)) \ll Y_{j-1}(i) \prod_{m=0}^{j-1} Y_m(i)^{-\lambda}$
for all $j=0,\dots,n-1$ and all $i\ge 0$.
\end{corollary}
\begin{cor}
\label{first:cor:T}
Suppose that $\mathcal{P}(j,\ell)$ holds for some integers $j\ge 1$ and $\ell\ge 0$ with
$j+2\ell<n$. Then there are arbitrarily large integers $i\ge 1$ for which
$\mathcal{U}^\ell(\mathbf{x}_{i-1})\not\subseteq\mathcal{U}^\ell(\mathbf{x}_i)$ and
\begin{equation}
\label{first:cor:T:eq}
1\ll Y_{j-1}(i)^{1-\ell\lambda}
\Big(\prod_{m=1}^{j}Y_{j-m}(i)^{-\lambda}\Big)
X_i^{-\lambda}.
\end{equation}
\end{cor}
\begin{proof}
Let $d=\liminf_{i\to\infty}\dim\mathcal{U}^\ell(A_j(i))$. If $d=\ell+j+1$, then we
have $d\le n-\ell$ and Lemma \ref{P0:lemma:A} provides infinitely many
$i\ge 1$ for which ${\mathcal{U}}^\ell(\mathbf{x}_{i-1})\not\subseteq \mathcal{U}^\ell(A_j(i))$.
For those $i$, we have $\mathcal{U}^\ell(\mathbf{x}_{i-1})\not\subseteq\mathcal{U}^\ell(\mathbf{x}_i)$
and Lemma \ref{P0:lemma:S} gives $1\ll H(V)L_{i-1}$
with $V=\mathcal{U}^\ell(A_j(i))$. Then \eqref{first:cor:T:eq} follows using
$L_{i-1}\ll X_i^{-\lambda}$ and the upper bound for $H(V)$ provided by
Proposition \ref{first:prop}. Otherwise, for each sufficiently large $i$,
we have $\dim\mathcal{U}^\ell(A_j(i)) > \ell+j+1$ and \eqref{first:cor:T:eq}
follows directly from Proposition \ref{first:prop} using
$H\big(\mathcal{U}^\ell(A_j(i))\big)\ge 1$, $Y_0(i)=X_{i+1}\ge X_i$ and
$e\ge 1$. Moreover, as $2\ell<n$, Lemma \ref{P0:lemma:A} also gives
$\mathcal{U}^\ell(\mathbf{x}_{i-1})\not\subseteq\mathcal{U}^\ell(\mathbf{x}_i)$ for infinitely many $i\ge 1$.
\end{proof}
\section{An alternative height estimate}
\label{sec:alt}
Keeping the same notation, we derive a second height estimate for $\mathcal{U}^\ell(A_j(i))$ by
an indirect process, as in the proof of Lemma \ref{P0:lemma:S},
namely by writing this space as a sum of two subspaces with a well-chosen one
dimensional intersection, and then by applying Schmidt's height inequality
\eqref{heights:eq:schmidt}.
\begin{prop}
\label{alt:prop}
Suppose that $\mathcal{P}(j,\ell)$ holds for some integers $1\le j\le \ell<n$.
For each $i\ge 0$, we have
\[
H\big(\mathcal{U}^\ell(A_j(i))\big)
\ll H\big(A_j(i)\big) Y_{j}(i)^{-(\ell-j+1)\lambda}
\prod_{m=1}^{j-1} Y_m(i)^{-\lambda}.
\]
\end{prop}
The main feature of this estimate is that it involves a negative power of $Y_j(i)$
and so, as we will see, it yields an upper bound for $Y_j(i)$ in terms of
$Y_0(i),\dots,Y_{j-1}(i)$. The proof requires the following simple observation.
\begin{lemma}
\label{alt:lemma}
Let $\tau_\mathbf{a}\colon\mathbb{R}^{n+1}\to\mathbb{R}^{n-\ell+1}$ be the linear map
given by \eqref{three:eq:tau} for fixed $\ell\in\{0,\dots,n\}$
and $\mathbf{a}=(a_0,\dots,a_\ell)\in \mathbb{Z}^{\ell+1}\setminus\{0\}$.
Then we have $\norm{\tau_\mathbf{a}(\mathbf{x}_i)}\asymp X_i$ as $i\to\infty$.
\end{lemma}
\begin{proof}
Since $\lim_{i\to\infty} X_i^{-1}\mathbf{x}_i =\norm{\Xi}^{-1}\Xi$ where
$\Xi=(1,\xi,\dots,\xi^n)$, we find
\[
\lim_{i\to\infty} X_i^{-1}\tau_\mathbf{a}(\mathbf{x}_i)
=\norm{\Xi}^{-1}\tau_\mathbf{a}(\Xi)
=\norm{\Xi}^{-1}(a_0+a_1\xi+\cdots+a_\ell\xi^\ell)(1,\xi,\dots,\xi^{n-\ell}).
\]
As $[\mathbb{Q}(\xi):\mathbb{Q}]>n\ge \ell$, this limit is non-zero, and the conclusion follows.
\end{proof}
\begin{proof}[Proof of Proposition \ref{alt:prop}]
We may assume that $i$ is large enough so that Lemma \ref{first:lemma:basis}
applies. Choose $q\ge i$ maximal such that
$A_j(i)=\langle\mathbf{x}_i,\dots,\mathbf{x}_q\rangle$, and to simplify notation set
$A=A_j(i)$. Then, by definition, we have $X_{q+1}=Y_j(i)$. Moreover,
we have $j+2\ell\le n$ by $\mathcal{P}(j,\ell)$ (see Proposition \ref{minimal:prop:Pjl}),
thus $\dim(A) =j+1\le n$, and so $A$ is a proper subspace of $\mathbb{R}^{n+1}$.
Using the basis of ${\mathcal{U}}^\ell(A)$ provided by Lemma \ref{first:lemma:basis}
for the present choice of $q$, we set
\[
W=\langle \mathbf{y}_0,\mathbf{y}_1,\dots,\mathbf{y}_{\ell+j-1}\rangle
\subseteq \mathbb{R}^{n+1-\ell}.
\]
Since $\dim(W)=\ell+j<\dim{\mathcal{U}}^\ell(A)$, we have ${\mathcal{U}}^\ell(A)\not\subseteq W$.
As $A$ is a proper subspace of $\mathbb{R}^{n+1}$, Proposition \ref{three:prop:avoiding}
provides a non-zero
point $\mathbf{a}=(a_0,\dots,a_\ell)\in\mathbb{Z}^{\ell+1}$ with $\|\mathbf{a}\|\ll 1$ such that
the linear map $\tau_\mathbf{a}\colon\mathbb{R}^{n+1}\to\mathbb{R}^{n+1-\ell}$ is injective
on $A$ with $\tau_\mathbf{a}(A)\not\subseteq W$. Thus we have
$\dim(\tau_\mathbf{a}(A)\cap W) \le j$ and so there are $\ell$ points
$\mathbf{z}_0,\dots,\mathbf{z}_{\ell-1}$ among $\mathbf{y}_0,\mathbf{y}_1,\dots,\mathbf{y}_{\ell+j-1}$ such
that
\[
\tau_\mathbf{a}(A) \cap \langle \mathbf{z}_0,\dots,\mathbf{z}_{\ell-1} \rangle = 0.
\]
By construction, these are integer points of norm $\le X_q$ and, since $j\le \ell$,
we may order them so that
\begin{equation}
\label{lemma:D:eq1}
\begin{cases}
L_\xi(\mathbf{z}_m)\ll X_{q+1}^{-\lambda}=Y_{j}(i)^{-\lambda}
&\text{for $0\le m\le \ell-j$,}\\
L_\xi(\mathbf{z}_{\ell-j+m})\ll Y_{j-m}(i)^{-\lambda}
&\text{for $1\le m\le j-1$.}
\end{cases}
\end{equation}
Since $\dim {\mathcal{U}}^\ell(A) =\ell+j+1+e$ for some integer $e\ge 0$,
we may complete $\{\mathbf{z}_0,\dots,\mathbf{z}_{\ell-1}\}$ to a maximal linearly
independent subset $\{\mathbf{z}_0,\dots,\mathbf{z}_{\ell-1+e}\}$ of ${\mathcal{U}}^\ell(A)$
such that
\[
{\mathcal{U}}^\ell(A)
= \tau_\mathbf{a}(A) \oplus \langle \mathbf{z}_0,\dots,\mathbf{z}_{\ell-1+e} \rangle
\]
by adding integer points of the form $\mathbf{x}_h^{(s,\ell)}$ with $i\le h\le q$
and $0\le s\le \ell$. These new points have norm $\le X_q$ and satisfy
\begin{equation}
\label{lemma:D:eq2}
L_\xi(\mathbf{z}_{\ell-1+m}) \ll L_i \ll Y_0(i)^{-\lambda}
\quad \text{for $1\le m\le e$.}
\end{equation}
Define
\[
U=\tau_\mathbf{a}(A), \quad \mathbf{z}=\tau_\mathbf{a}(\mathbf{x}_q)
\quad\mbox{and}\quad V=\langle \mathbf{z},\mathbf{z}_0,\dots,\mathbf{z}_{\ell-1+e} \rangle,
\]
so that
\[
U+V={\mathcal{U}}^\ell(A)
\quad\mbox{and}\quad
U\cap V=\Span{\mathbf{z}}.
\]
Since $\norm{\mathbf{a}}\ll 1$, we find
\[
H(U)\ll H(A), \quad \norm{\mathbf{z}}\asymp X_q
\quad\mbox{and}\quad L_\xi(\mathbf{z}) \ll L_q\ll X_{q+1}^{-\lambda}=Y_j(i)^{-\lambda},
\]
where the middle estimate $\norm{\mathbf{z}}\asymp X_q$ comes from
Lemma \ref{alt:lemma}. We deduce that
\[
H(U\cap V)=g^{-1}\|\mathbf{z}\|\asymp g^{-1}X_q
\]
where $g$ denotes the content of $\mathbf{z}$. Since $g^{-1}\mathbf{z}$ is an integer point,
we further have
\[
H(V) \le \|g^{-1}\mathbf{z}\wedge\mathbf{z}_0\wedge\cdots\wedge\mathbf{z}_{\ell-1+e}\|
= g^{-1}\|\mathbf{z}\wedge\mathbf{z}_0\wedge\cdots\wedge\mathbf{z}_{\ell-1+e}\|.
\]
Applying Lemma \ref{heights:lemma} with the estimates \eqref{lemma:D:eq1},
\eqref{lemma:D:eq2}, this implies
\begin{align*}
H(V)
&\ll g^{-1} X_q Y_j(i)^{-(\ell-j+1)\lambda}
\Big(\prod_{m=1}^{j-1} Y_{j-m}(i)^{-\lambda}\Big) Y_0(i)^{-e\lambda}\\
&\ll H(U\cap V) Y_j(i)^{-(\ell-j+1)\lambda}
\prod_{m=1}^{j-1} Y_{j-m}(i)^{-\lambda}.
\end{align*}
The conclusion follows since $H({\mathcal{U}}^\ell(A))\ll H(U)H(V)/H(U\cap V)$
by \eqref{heights:eq:schmidt}.
\end{proof}
Combining Proposition \ref{alt:prop} with the crude estimate
$H\big(\mathcal{U}^\ell(A_j(i))\big)\ge 1$ and the upper bound for $H(A_j(i))$
given by Corollary \ref{first:cor:HA}, we obtain the following upper bound
for $Y_j(i)$.
\begin{corollary}
\label{alt:cor:E}
Suppose that $\mathcal{P}(j,\ell)$ holds for some integers $1\le j\le \ell<n$.
Then, for each $i\ge 0$, we have
\[
Y_{j}(i)^{(\ell-j+1)\lambda}
\ll Y_{j-1}(i) \Big(\prod_{m=1}^{j-1} Y_m(i)^{-2\lambda} \Big)
Y_0(i)^{-\lambda}.
\]
\end{corollary}
When $j=1$, this can be reformulated as follows.
\begin{corollary}
\label{alt:cor:F}
Suppose that $\mathcal{P}(1,\ell)$ holds for some integer $1\le \ell<n$.
Then the ratio
\begin{equation}
\label{eq:theta}
\theta=\frac{\ell\lambda}{1-\lambda}
\end{equation}
satisfies $0<\theta\le 1$ and we have $Y_1(i)^\theta\ll Y_0(i)$ and
$Y_0(i)^\theta\ll Y_{-1}(i)$ for each $i\ge 0$.
\end{corollary}
\begin{proof}
By Corollary \ref{alt:cor:E}, we have $Y_1(i)^{\ell\lambda}\ll
Y_0(i)^{1-\lambda}$ for each $i\ge 0$. Since $\lambda<1$,
this yields $Y_1(i)^\theta\ll Y_0(i)$ for all $i\ge 0$ and thus
$\theta\le 1$. For $i\ge 1$, this in turn gives
\[
Y_0(i)^\theta=X_{i+1}^\theta\le Y_1(i-1)^\theta
\ll Y_0(i-1)=X_i=Y_{-1}(i).
\qedhere
\]
\end{proof}
More generally, Corollary \ref{alt:cor:E} admits the following consequence.
\begin{cor}
\label{alt:cor:G}
Suppose that $\mathcal{P}(j,\ell)$ holds for some integers $1\le j\le \ell<n$,
and that $\theta^{j-1}+\theta^j\ge 1$ where $\theta$ is given by \eqref{eq:theta}.
Then we have $Y_m(i)^\theta\ll Y_{m-1}(i)$ for each each $i\ge 0$ and
each $m=0,1,\dots,j$.
\end{cor}
\begin{proof}
Since $\mathcal{P}(1,\ell)$ holds, Corollary \ref{alt:cor:F} gives $0<\theta\le 1$
and $Y_m(i)^\theta\ll Y_{m-1}(i)$ for $m=0,1$. So, we are done if $j=1$.
Suppose now that $j\ge 2$. Since $\theta\le 1$, the hypothesis
$\theta^{j-1}+\theta^j\ge 1$ implies that $\theta^{j-2}+\theta^{j-1}\ge 1$.
Since $\mathcal{P}(j-1,\ell)$ holds, we may
assume by induction that $Y_m(i)^\theta\ll Y_{m-1}(i)$ for each $i\ge 0$
and each $m=0,\dots,j-1$. Then, using Corollary \ref{alt:cor:E},
we obtain $Y_j(i)^{(\ell-j+1)\lambda} \ll Y_{j-1}(i)^\rho$ where
\begin{align*}
\rho
&= 1 - 2 \lambda\Big(\sum_{m=1}^{j-1} \theta^{j-1-m}\Big)
- \lambda\theta^{j-1}\\
&= 1 - \lambda - \lambda \sum_{m=1}^{j-1} (\theta^{m-1}+\theta^m)
\le \frac{\ell\lambda}{\theta} - \lambda \sum_{m=1}^{j-1} \frac{1}{\theta}
= \frac{(\ell-j+1)\lambda}{\theta},
\end{align*}
and so $Y_j(i)^\theta \ll Y_{j-1}(i)$.
\end{proof}
\section{Main proposition and proof of Theorem \ref{intro:thm:main}}
\label{sec:main}
We first prove the following general statement and then optimize the choice of parameters
to deduce Theorem \ref{intro:thm:main}. The notation, including the choice of $\lambda$
is as in the preceding sections.
\begin{prop}
\label{main:prop}
Let $1\le k\le \ell$ be integers with $k+2\ell=n$. Suppose that
\[
\frac{1}{2}\le \theta^k \le 1
\quad\text{where}\quad
\theta=\frac{\ell\lambda}{1-\lambda}.
\]
Then we have
$\lambda \le \eta^{-1}$ where $\eta=\ell+\theta+\theta^2+\cdots+\theta^{k+1}$.
\end{prop}
\begin{proof}
Assume on the contrary that $\lambda>\eta^{-1}$. Since $\theta\le 1$,
we have $\eta\le \ell+k+1=n-\ell+1$, thus $\lambda>(n-\ell+1)^{-1}$,
and so Proposition \ref{P0:prop:BS} shows that
$\mathcal{P}(0,\ell)$ holds. By Lemma \ref{P0:lemma:Fbis}, this implies that
${\mathcal{U}}^\ell(\mathbf{x}_{i-1})\not\subseteq {\mathcal{U}}^\ell(\mathbf{x}_i)$ for infinitely many
$i\ge 1$, and that
\begin{equation}
\label{main:prop:eq1}
Y_0(i)^\theta=X_{i+1}^\theta\ll Y_{-1}(i)=X_i
\quad
\text{whenever}
\quad
{\mathcal{U}}^\ell(\mathbf{x}_{i-1})\not\subseteq {\mathcal{U}}^\ell(\mathbf{x}_i).
\end{equation}
Let $j$ be the largest integer with $0\le j\le k$ for which $\mathcal{P}(j,\ell)$ holds.
Since $\theta\le 1$, we have $\theta^{j-1}+\theta^j
\ge \theta^{k-1}+\theta^k\ge 1$. So, if $j\ge 1$, Corollary \ref{alt:cor:G}
gives
\begin{equation}
\label{main:prop:eq2}
Y_m(i)^\theta\ll Y_{m-1}(i)
\quad\text{for each}\quad m=0,1,\dots,j\quad\mbox{and}\quad i\ge 0.
\end{equation}
Suppose that $j<k$. Then $\mathcal{P}(j+1,\ell)$ does not hold. However,
by Proposition \ref{minimal:prop:Pjl}, $\mathcal{P}(j+1,\ell-1)$ holds
because $\mathcal{P}(j,\ell)$ does. Thus, by Proposition \ref{P0:prop:I},
there are arbitrarily large $i\ge 1$ for which
\[
{\mathcal{U}}^\ell(\mathbf{x}_{i-1})\not\subseteq \mathcal{U}^\ell(A_{j+1}(i))
\quad\mbox{and}\quad
1 \ll H\big(\mathcal{U}^{\ell-1}(A_{j+1}(i))\big)L_{i-1}^{n-j-2\ell+1}.
\]
Using Proposition \ref{first:prop} to estimate from above the height of
$\mathcal{U}^{\ell-1}(A_{j+1}(i))$ and recalling that $n=k+2\ell$ and
$L_{i-1}\ll Y_{-1}(i)^{-\lambda}$, this gives
\[
1 \ll Y_{j}(i)^{1-(\ell-1)\lambda}
\Big( \prod_{m=0}^{j} Y_{j-m}(i)^{-\lambda} \Big)
Y_{-1}(i)^{-(k-j+1)\lambda}.
\]
For those $i$, we have ${\mathcal{U}}^\ell(\mathbf{x}_{i-1})\not\subseteq {\mathcal{U}}^\ell(\mathbf{x}_i)$.
Thus, using \eqref{main:prop:eq1} if $j=0$ and \eqref{main:prop:eq2} else, we find
\begin{align*}
0 &\le 1 - \lambda(\ell-1) - \lambda\Big(\sum_{m=0}^{j} \theta^{m}\Big)
- \lambda(k-j+1)\theta^{j+1} \\
&\le 1 - \lambda\Big(\ell-1+\sum_{m=0}^{k+1} \theta^{m}\Big)
= 1- \lambda\eta,
\end{align*}
against the hypothesis that $\lambda>\eta^{-1}$.
The above contradiction shows that $j=k$. Thus $\mathcal{P}(k,\ell)$ holds and so
$\mathcal{P}(k+1,\ell-1)$ holds as well (by Proposition \ref{minimal:prop:Pjl}).
As $(k+1)+2(\ell-1)=n-1<n$, Corollary \ref{first:cor:T} provides arbitrarily large
values of $i$ for which
\[
1 \ll Y_{k}(i)^{1-(\ell-1)\lambda}
\Big(\prod_{m=0}^{k} Y_{k-m}(i)^{-\lambda}\Big) Y_{-1}(i)^{-\lambda}.
\]
Since $j=k\ge 1$, \eqref{main:prop:eq2} applies. So, we conclude that
\[
0 \le 1 - \lambda\Big( \ell-1+\sum_{m=0}^{k+1} \theta^{m} \Big)
= 1 - \lambda\eta,
\]
which again contradicts the hypothesis that $\lambda>\eta^{-1}$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{intro:thm:main}]
For $2\le n\le 3$, the upper bound for $\widehat{\lambda}_n(\xi)$ provided by the
theorem is weaker than the prior ones mentioned in the introduction. For
$4\le n\le 11$, they are also weaker than those coming from Theorems
\ref{intro:thm:pair} and \ref{intro:thm:impair}, and listed on Table
\ref{intro:table1}. So, we may assume that $n\ge 12$.
Suppose, by contradiction, that the upper bound for $\widehat{\lambda}_n(\xi)$ given
by Theorem \ref{intro:thm:main} does not hold for such $n$. Then, we may
assume that
\[
\lambda=\Big(\frac{n}{2}+a\sqrt{n}+\frac{1}{3}\Big)^{-1}
\]
where $a=(1-\log(2))/2$. Define
\begin{align*}
&\ell =\left\lfloor \frac{n}{2}-\frac{\log(2)}{2}\sqrt{n} + 1\right\rfloor,
\quad
k= n -2\ell,
\quad
\theta= \frac{\ell\lambda}{1-\lambda}
\quad\mbox{and}\quad
\eta= \ell - 1 + \sum_{m=0}^{k+1}\theta^m.
\label{main:proof:eq:eta}
\end{align*}
Since $n\ge 9$, we find that $1\le \ell <n/2$ and so $k\ge 1$.
We will show further that
\begin{equation}
\label{main:proof:eq:conditions}
\frac{1}{2}\le \theta^k <1
\quad\mbox{and}\quad
\eta>\lambda^{-1}.
\end{equation}
So this choice of parameters fulfills all the hypotheses of
Proposition \ref{main:prop} but not its conclusion. This will
conclude the proof, by contradiction.
A quick computer computation shows that
\eqref{main:proof:eq:conditions} holds for $12\le n< 900$. So, we
may assume that $\sqrt{n}\ge 30$. This assumption will simplify the
estimates below.
By choice of $\ell$, there is some $t\in(0,2]$ such that
\[
\ell=(n-\log(2)\sqrt{n} + t)/2\,,
\quad\hbox{and then}\quad
k=\log(2) \sqrt{n} - t.
\]
Using the actual value of $\lambda$, this gives
\[
\theta=\frac{n-k}{2\big(\lambda^{-1}-1\big)}
=\frac{n-\log(2)\sqrt{n}+t}{n+2a\sqrt{n}-4/3}.
\]
Define
\begin{equation}
\label{main:proof:eq:epsilon}
\epsilon=1-\theta=\frac{\sqrt{n}-(4/3+t)}{n+2a\sqrt{n}-4/3}.
\end{equation}
As $n\ge 12$, we have $0<\epsilon<1$, thus $0<\theta<1$, and so $\theta^k<1$.
We set
\[
r=-\frac{\log(2)}{\log(1-\epsilon)}
\]
so that $\theta^r=(1-\epsilon)^r=1/2$. Since $0<\epsilon<1$, we
find
\[
\frac{\epsilon}{1-\epsilon/2}
=\sum_{i=1}^\infty\frac{\epsilon^i}{2^{i-1}}
\le \sum_{i=1}^\infty\frac{\epsilon^i}{i}
=-\log(1-\epsilon)
\le \sum_{i=1}^\infty \epsilon^i
=\frac{\epsilon}{1-\epsilon},
\]
thus, by definition of $r$,
\begin{equation}
\label{main:proof:estimates:r}
\log(2)\left(\frac{1}{\epsilon}-1\right)
\le r
\le \log(2)\left(\frac{1}{\epsilon}-\frac{1}{2}\right).
\end{equation}
Moreover, \eqref{main:proof:eq:epsilon} yields
\[
\frac{1}{\epsilon}
= \sqrt{n} + b + \frac{c}{\sqrt{n}-(4/3+t)}
\]
with $b=2a+4/3+t$ and $c=b(4/3+t)-4/3$. As $c>0$, this implies that
\begin{equation}
\label{main:proof:estimates:1/epsilon}
\frac{1}{\epsilon}
\ge \sqrt{n} + b + \frac{c}{\sqrt{n}} \ge \sqrt{n}+1.64,
\end{equation}
so $r>\log(2)\sqrt{n}\ge k$, and thus $\theta^k\ge \theta^r=1/2$.
This proves the first condition in \eqref{main:proof:eq:conditions}.
To verify the second condition $\eta>\lambda^{-1}$, we first note that
\begin{equation}
\label{main:proof:majoration:1/epsilon}
\frac{1}{\epsilon}
\le \sqrt{n} + b + \frac{c}{30-(4/3+2)} \le \sqrt{n}+4.05,
\end{equation}
using the hypothesis $\sqrt{n}\ge 30$ and $t\le 2$. We set further
\[
\delta=k+2-r \quad\mbox{and}\quad E=\frac{1-\theta^\delta}{1-\theta}-\delta.
\]
As $\theta^r=1/2$, we find
\begin{align*}
\eta
=\ell-1+\frac{1-\theta^{k+2}}{1-\theta}
&=\ell-1+\frac{1-\theta^{r}}{1-\theta}+\theta^r\frac{1-\theta^{\delta}}{1-\theta}\\
&=\frac{n-k}{2}-1+\frac{1}{2\epsilon}+\frac{1}{2}(E+\delta)
=\frac{n}{2}-\frac{r}{2}+\frac{1}{2\epsilon}+\frac{E}{2}.
\end{align*}
Using the upper bound for $r$ given by \eqref{main:proof:estimates:r}
and then the lower bound for $1/\epsilon$ given by
\eqref{main:proof:estimates:1/epsilon}, we deduce that
\begin{align*}
\eta-\lambda^{-1}
&\ge \frac{n}{2}+\frac{a}{\epsilon}+\frac{\log(2)}{4}+\frac{E}{2}
-\left(\frac{n}{2}+a\sqrt{n}+\frac{1}{3}\right)\\
&\ge 1.64\,a+\frac{\log(2)}{4}-\frac{1}{3}+\frac{E}{2}\\
&\ge 0.09+\frac{E}{2}.
\end{align*}
So it remains to show that $E>-0.18$.
Lagrange remainder theorem gives
\[
\theta^\delta
= (1-\epsilon)^\delta
= 1 - \delta\epsilon
+ \frac{\delta(\delta-1)}{2}(1-\epsilon')^{\delta-2}\epsilon^2
\]
for some real number $\epsilon'$ with $0<\epsilon'<\epsilon$, and thus
\[
E = \frac{1-\theta^\delta}{\epsilon} - \delta
= -\frac{\delta(\delta-1)}{2}(1-\epsilon')^{\delta-2}\epsilon.
\]
As $k<r$, we have $\delta<2$. Moreover, the upper bound for $r$ given
by \eqref{main:proof:estimates:r} together with that of $1/\epsilon$
given by \eqref{main:proof:majoration:1/epsilon} yields
\[
\delta \ge k+2-\log(2)\left(\frac{1}{\epsilon}-\frac{1}{2}\right)
\ge -2.47.
\]
Since by \eqref{main:proof:estimates:1/epsilon} we have
$\epsilon <1/\sqrt{n}\le 1/30$, we conclude that
\[
|E| \le \frac{2.47\times3.47}{2}\left(1-\frac{1}{30}\right)^{\displaystyle -4.47}\frac{1}{30}
\le 0.17.
\qedhere
\]
\end{proof}
\section{A new construction}
\label{sec:new}
In this section, we introduce a new construction which in some cases
yields $\lambda_k(\xi)>1$ for an integer $k\ge 1$. We will use it
in the proof of Theorems \ref{intro:thm:impair} and \ref{intro:thm:pair}
in combination with the following results of Schleischitz
\cite[Theorems 1.6 and 1.12]{Schl2016}.
\begin{theorem}[Schleischitz, 2016]
\label{new:thm:Schleischitz}
Let $\xi\in\mathbb{R}\setminus\mathbb{Q}$. For each integer $n\ge 1$, we have
\[
\widehat{\lambda}_n(\xi) \le \max\{1/n,1/\lambda_1(\xi)\}.
\]
Moreover, if $\lambda_k(\xi)>1$ for some integer $k\ge 1$, then
$\lambda_1(\xi)=k-1+k\lambda_k(\xi)$.
\end{theorem}
In fact, we will simply need the following weaker consequence.
\begin{cor}
\label{new:cor}
Let $\xi\in\mathbb{R}\setminus\mathbb{Q}$ and let $\lambda\in(1/n,1)$. Suppose that, for some
integer $k\ge 1$, there exist non-zero points $C\in\mathbb{Z}^{k+1}$ for which
$L_\xi(C)$ is arbitrarily small while the products $\norm{C}L_\xi(C)^\lambda$
remain bounded from above. Then, we have $\widehat{\lambda}_n(\xi)\le \lambda$.
\end{cor}
\begin{proof}
The hypothesis implies that $\lambda_k(\xi)\ge 1/\lambda>1$. By the above
theorem, we conclude that $\lambda_1(\xi)\ge 1/\lambda$ and so
$\widehat{\lambda}_n(\xi)\le\max\{1/n,\lambda\}=\lambda$.
\end{proof}
For each positive integer $k$ and each non-zero subspace $V$ of $\mathbb{R}^{k+1}$
defined over $\mathbb{Q}$, it is natural to define
\[
L_\xi(V)=\norm{\mathbf{z}_1\wedge\cdots\wedge\mathbf{z}_s\wedge\Xi_k}
\]
where $\{\mathbf{z}_1,\dots,\mathbf{z}_s\}$ is a basis of $V\cap\mathbb{Z}^{k+1}$ over
$\mathbb{Z}$, and where $\Xi_k=(1,\xi,\dots,\xi^k)$. This is independent of the choice
of the basis, like for the height $H(V)=\norm{\mathbf{z}_1\wedge\cdots\wedge\mathbf{z}_s}$
of $V$. In particular, if $\mathbf{z}$ is a primitive point of $\mathbb{Z}^{k+1}$, we find
\[
L_\xi(\Span{\mathbf{z}})=\norm{\mathbf{z}\wedge\Xi_k}\asymp L_\xi(\mathbf{z})
\]
with an implied constant that depends only on $k$ and $\xi$ (see \S\ref{sec:heights}).
In general, if $\{\mathbf{y}_1,\dots,\mathbf{y}_s\}$ is a maximal linearly independent
subset of $V\cap\mathbb{Z}^{k+1}$, then arguing as in the proof
of Lemma \ref{heights:lemma}, we find
\[
L_\xi(V)\le\norm{\mathbf{y}_1\wedge\cdots\wedge\mathbf{y}_s\wedge\Xi_k}
\ll L_\xi(\mathbf{y}_1)\cdots L_\xi(\mathbf{y}_s),
\]
with an implied constant of the same nature. We can now present our construction.
\begin{prop}
\label{new:prop}
Let $\xi\in\mathbb{R}\setminus\mathbb{Q}$, let $k,\ell\ge 1$ be integers, and let $n=k+\ell$.
Suppose that $V$ is a subspace of $\mathbb{R}^{k+1}$ of dimension $k$, and that
$\mathbf{x}\in\mathbb{Z}^{n+1}$ satisfies $\mathcal{U}^\ell(\mathbf{x})\not\subseteq V$. Finally, let
$\{\mathbf{z}_1,\dots,\mathbf{z}_k\}$ be a basis of $V\cap \mathbb{Z}^{k+1}$. Then the point
\[
C=\left(\det(\mathbf{z}_1,\dots,\mathbf{z}_k,\mathbf{x}^{(0,\ell)}),
\dots,
\det(\mathbf{z}_1,\dots,\mathbf{z}_k,\mathbf{x}^{(\ell,\ell)})\right)\in\mathbb{Z}^{\ell+1}
\]
is non-zero. It satisfies
\[
\norm{C} \ll \norm{\mathbf{x}}L_\xi(V)+H(V)L_\xi(\mathbf{x})
\quad\mbox{and}\quad
L_\xi(C) \ll H(V)L_\xi(\mathbf{x})
\]
with implied constants that depend only on $k$ and $\xi$.
\end{prop}
We will write $C(V,\mathbf{x})$ to denote this point $C$, although it is determined
by $V$ and $\mathbf{x}$ only up to multiplication by $\pm 1$. In practice, this is
no problem since this ambiguity does not affect the quantities $\norm{C}$
and $L_\xi(C)$.
\begin{proof} Write $\mathbf{x}=(x_0,\dots,x_n)$ and $C=(C_0,\dots,C_\ell)$.
Then we have $\mathbf{x}=x_0\Xi_n+\Delta$ with $\norm{\Delta}\asymp L_\xi(\mathbf{x})$
and for each $j=0,\dots,\ell$ we find
\[
C_j = x_0\xi^j\det(\mathbf{z}_1,\dots,\mathbf{z}_k,\Xi_k) + \det(\mathbf{z}_1,\dots,\mathbf{z}_k,\Delta^{(j,\ell)}).
\]
The estimates for $\norm{C}$ and $L_\xi(C)=\max_{1\le j\le k}|C_j-C_0\xi^j|$ follow.
\end{proof}
\section{Small odd degree}
\label{sec:odd}
This section is devoted to the proof of Theorem~\ref{intro:thm:impair}.
So, we suppose that
\[
n=2m+1 \geq 5
\]
is an odd integer. We argue by contradiction, assuming that $\widehat{\lambda}_n(\xi)
> \alpha$ where $\alpha=\alpha_m$ is the unique positive root of
\[
P_m(x)=1-(m+1)x-mx^2.
\]
Then $\mathcal{P}(1,m)$ holds by Corollary \ref{P1:cor}, and so $\mathcal{P}(2,m-1)$
holds as well by Proposition \ref{minimal:prop:Pjl}. Thus, for each large
enough $i\in I$, the vector space
\[
V_i=\mathcal{U}^{m-1}(\mathbf{x}_{i-1},\mathbf{x}_i,\mathbf{x}_{i+1}) \subseteq \mathbb{R}^{m+3}
\]
has dimension at least $m+2$ (see the definitions and remarks in Section \ref{sec:minimal}).
\begin{lemma}
\label{odd:lemma1}
With the above hypotheses, we have
\begin{equation}
\label{odd:lemma1:eq1}
X_{j+1}^\theta\ll X_{i+1}
\quad\mbox{and}\quad
X_{i+1}^\theta\ll X_i
\quad
\text{where}
\quad
\theta=\frac{m\alpha}{1-\alpha}=\frac{1}{1+\alpha},
\end{equation}
for each pair of consecutive elements $i<j$ of $I$. If $i$ is large enough,
then $\dim(V_i)=m+2$,
\begin{equation}
\label{odd:lemma1:eq2}
H(V_i)\ll X_{j+1}^{-(m-1)\alpha}X_{i+1}^{1-\alpha}X_i^{-\alpha}
\quad\mbox{and}\quad
L_\xi(V_i)\ll X_{j+1}^{-m\alpha}X_{i+1}^{-\alpha}X_i^{-\alpha}.
\end{equation}
\end{lemma}
\begin{proof}
Since $\widehat{\lambda}_n(\xi) > \alpha$, we can choose $\epsilon>0$ such that
$\alpha+\epsilon<\widehat{\lambda}_n(\xi)$. Then the results of the preceding sections
apply with $\lambda=\alpha+\epsilon$. In particular, for each
pair of consecutive elements $i<j$ of $I$, Lemma \ref{P1:lemma2} gives
\[
H(\mathcal{U}^m(\mathbf{x}_i,\mathbf{x}_{i+1}))
\ll X_{j+1}^{-m\alpha}X_{i+1}^{1-\alpha}
\quad\mbox{and}\quad
H(V_i) \ll X_{j+1}^{-(m-1)\alpha}X_{i+1}^{1-\alpha-\epsilon}X_i^{-e(i)\alpha}
\]
where $e(i)=\dim(V_i)-(m+1)$, because both $\mathcal{P}(1,m)$ and $\mathcal{P}(1,m-1)$ hold.
Since we have $H(\mathcal{U}^m(\mathbf{x}_i,\mathbf{x}_{i+1}))\ge 1$, the first estimate yields
$X_{j+1}^\theta\ll X_{i+1}$ with $\theta$ as in \eqref{odd:lemma1:eq1}
(this also follows from Corollary \ref{alt:cor:F}). So, if $i$ is large enough to
admit a predecessor $h<i$ in $I$, we also have $X_{i+1}^\theta\ll X_{h+1}\le X_i$,
thus $X_{i+1}\ll X_i^{1+\alpha}$. This proves \eqref{odd:lemma1:eq1}.
If furthermore $V_i=\mathbb{R}^{m+3}$, then $e(i)=2$ and using $X_{j+1}\ge X_{i+1}$,
we obtain
\[
1=H(V_i) \ll X_{i+1}^{1-m\alpha-\epsilon}X_i^{-2\alpha}
\ll X_i^{(1+\alpha)(1-m\alpha-\epsilon)-2\alpha}
=X_i^{-(1+\alpha)\epsilon},
\]
which forces $i$ to be bounded. So, if $i$ is large enough, we have $\dim(V_i)=m+2$,
thus $e(i)=1$ and the estimate for $H(V_i)$ in \eqref{odd:lemma1:eq2} follows.
Finally, the proof of Lemma \ref{P1:lemma2} shows that $V_i$ admits a basis
of the form
\[
\big\{\mathbf{x}_j^{(0,m-1)},\dots,\mathbf{x}_j^{(m-1,m-1)},\mathbf{x}_{j-1}^{(p,m-1)},
\mathbf{x}_h^{(q,m-1)}\big\}
\]
for some $p,q\in\{0,1,\dots,m-1\}$ and some $h\in\{i-1,j-1\}$. So
the considerations of the preceding section provide
$L_\xi(V_i) \ll L_j^mL_{j-1}L_{i-1}
\le X_{j+1}^{-m\alpha}X_{i+1}^{-\alpha}X_i^{-\alpha}$.
\end{proof}
It is now an easy matter to complete the proof of Theorem \ref{intro:thm:impair}.
By the preceding lemma, there are arbitrarily large pairs of successive elements
$i<j$ of $I$ for which $\dim(V_i)=m+2$ and $V_j\not\subseteq V_i$. The latter
condition means that $\mathcal{U}^{m-1}(\mathbf{x}_{j+1})\not\subseteq V_i$. So, for these
pairs, Proposition \ref{new:prop} shows that the point
\[
C_i=C(V_i,\mathbf{x}_{j+1})\in\mathbb{Z}^m
\]
is non-zero. Using the estimates of the preceding lemma, it also gives
\begin{align*}
L_\xi(C_i)
&\ll H(V_i)L_{j+1}
\ll X_{j+1}^{-m\alpha}X_{i+1}^{1-\alpha}X_i^{-\alpha},\\
\norm{C_i}
&\ll X_{j+1}L_\xi(V_i)+H(V_i)L_{j+1}
\ll X_{j+1}^{1-m\alpha}X_{i+1}^{-\alpha}X_i^{-\alpha}.
\end{align*}
Using \eqref{odd:lemma1:eq1}, we find that
\begin{align*}
\norm{C_i}L_\xi(C_i)^\alpha
\ll X_{j+1}^{1-m\alpha-m\alpha^2}X_{i+1}^{-\alpha^2}X_i^{-\alpha-\alpha^2}
&= X_{j+1}^{\alpha}X_{i+1}^{-\alpha^2}X_i^{-\alpha/\theta}\\
&\ll X_{j+1}^{\alpha}X_{i+1}^{-\alpha^2-\alpha}
= X_{j+1}^{\alpha}X_{i+1}^{-\alpha/\theta}
\ll 1
\end{align*}
is bounded from above, and that
\begin{align*}
L_\xi(C_i)
\ll X_{j+1}^{-m\alpha}X_{i+1}^{1-\alpha}X_i^{-\alpha}
\ll X_{j+1}^{-m\alpha}X_{i+1}^{1-\alpha-\alpha\theta}
\le X_{j+1}^{-m\alpha+1-\alpha-\alpha\theta}
= X_{j+1}^{-\alpha^2\theta}
\end{align*}
tends to $0$ as $i$ goes to infinity. By Corollary \ref{new:cor}, this implies that
$\widehat{\lambda}_n(\xi)\le \alpha$.
\section{Small even degree}
\label{sec:even}
We conclude, in this section, with the proof of Theorem~\ref{intro:thm:pair}.
So, we assume that
\[
n=2m \geq 4
\]
is an even integer. For the proof, define $\beta=\beta_m$ to be the single positive root of
\begin{align*}
Q_m(x)&=\begin{cases}
1-mx-mx^2-m(m-1)x^3 &\text{if $m\ge3$,}\\
1-3x+x^2-2x^3-2x^4 &\text{if $m=2$,}
\end{cases}\\
\intertext{as in the statement of the theorem, then write}
\gamma&=\begin{cases}
\displaystyle \frac{m+4}{m^2+6m+2} &\text{if $m\ge4$,}\\[8pt]
8/33 &\text{if $m=3$,}\\
1/3 &\text{if $m=2$,}
\end{cases}\\
\intertext{and define $\delta$ to be the single positive root of}
R_m(x)&=1-(m+1)x-(m-1)x^2.
\end{align*}
It is a simple matter to check that
\[
1/(m+2) < \delta < \gamma < \beta.
\]
We will prove that $\widehat{\lambda}_n(\xi)\le \beta$ through the following chains of implications
\[
\begin{array}{lclcl}
\widehat{\lambda}_n(\xi)>\delta
&\Longrightarrow & \text{$\mathcal{P}(1,m-1)$ holds}
&\Longrightarrow & \text{$\mathcal{P}(2,m-2)$ holds}\\[5pt]
\widehat{\lambda}_n(\xi)>\gamma
&\Longrightarrow & \text{$\mathcal{P}(2,m-1)$ does not hold}
&\Longrightarrow & \widehat{\lambda}_n(\xi)\le \beta.
\end{array}
\]
Recall that $\lambda$ represents a fixed real number with
$0<\lambda<\widehat{\lambda}_n(\xi)$.
\begin{lemma}
\label{even:lemma1}
Suppose that $\widehat{\lambda}_n(\xi) > \delta$. Then both $\mathcal{P}(1,m-1)$
and $\mathcal{P}(2,m-2)$ hold. Moreover, we have
$X_{i+1}^{m\lambda}\ll X_i$ for each $i\ge 0$.
\end{lemma}
\begin{proof}
For the choice of $\ell=m-1$, we have $1\le \ell<n/2$ and
$R_m(x)=1-(n-\ell)x-\ell x^2$. Thus, by Corollary \ref{P1:cor}, our
hypothesis implies $\mathcal{P}(1,m-1)$ which in turn implies $\mathcal{P}(2,m-2)$,
by the general Proposition \ref{minimal:prop:Pjl}. Since
$\delta\ge 1/(m+2)\ge 1/(2m)$, the growth estimate follows
from Lemma \ref{P1:lemma3}.
\end{proof}
\begin{lemma}
\label{even:lemma2}
Suppose that $\mathcal{P}(2,m-1)$ holds and choose $\theta\in(0,1]$ such that
\begin{equation}
\label{even:lemma2:eq}
\frac{m-1}{\theta}+\theta > \frac{1}{\lambda}-1.
\end{equation}
Then we have $X_{j+1}^\theta \le X_{i+1}$ for any large enough pair of consecutive
elements $i<j$ of $I$.
\end{lemma}
\begin{proof}
Property $\mathcal{P}(2,m-1)$ means that $\mathcal{U}^{m-1}(\mathbf{x}_{i-1},\mathbf{x}_i,\mathbf{x}_{i+1})=\mathbb{R}^{m+2}$
for each large enough $i\in I$. By definition, it also implies property $\mathcal{P}(1,m-1)$.
Thus, for all but finitely many triples $h<i<j$ of consecutive elements of $I$, we
obtain, by Lemma \ref{P1:lemma2},
\[
1=H(\mathcal{U}^{m-1}(\mathbf{x}_{i-1},\mathbf{x}_i,\mathbf{x}_{i+1}))
\ll X_{j+1}^{-(m-1)\lambda}X_{i+1}^{1-\lambda}X_{h+1}^{-\lambda}
\]
using the crude estimate $X_i\ge X_{h+1}$. Taking logarithms, this gives
\[
(m-1)\frac{\log(X_{j+1})}{\log(X_{i+1})}+\frac{\log(X_{h+1})}{\log(X_{i+1})}
\le \frac{1-\lambda}{\lambda} + {\mathcal{O}}\Big(\frac{1}{\log(X_{i+1})}\Big).
\]
Let $\mu\in[0,1]$ denote the inferior limit of the ratio $\log(X_{i+1})/\log(X_{j+1})$
as $(i,j)$ runs through the pairs of consecutive elements $i<j$ of $I$. Then, there
is a sequence of triples $(h,i,j)$, with $i$ going to infinity and $h<i<j$ consecutive
in $I$, for which the ratio $\log(X_{i+1})/\log(X_{j+1})$ converges to $\mu$.
Over that sequence, the inferior limit of $\log(X_{h+1})/\log(X_{i+1})$ is at
least $\mu$, and so the above inequality implies that $\mu>0$ and
\[
\frac{m-1}{\mu}+\mu \le \frac{1}{\lambda}-1.
\]
As $m\ge 2$, the expression $(m-1)/x+x$ is a strictly decreasing function
of $x$ on $(0,1]$. We conclude that our choice of $\theta$ satisfies
$0<\theta<\mu$, and so $\theta \le \log(X_{i+1})/\log(X_{j+1})$ for
any pair of consecutive elements $i<j$ of $I$ with $i$ large enough.
\end{proof}
\begin{lemma}
\label{even:lemma3}
Suppose that $\widehat{\lambda}_n(\xi)>\gamma$. Then $\mathcal{P}(2,m-1)$ does not hold.
\end{lemma}
\begin{proof}
Suppose on the contrary that $\mathcal{P}(2,m-1)$ holds. By the hypothesis,
we may assume that $\lambda>\gamma$. A short
computation shows that $(m-1)/\theta+\theta\ge 1/\gamma-1$ for
\[
\theta=\begin{cases}
1 &\text{if $m=2$,}\\
15/17 &\text{if $m=3$,}\\
(m+4)/(m+5) &\text{if $m\ge 4$.}
\end{cases}
\]
Thus this choice of $\theta$ fulfills the main condition \eqref{even:lemma2:eq}
of Lemma \ref{even:lemma2}, and so we have $X_{j+1}^\theta\le X_{i+1}$
for each large enough pair of consecutive elements $i<j$ of $I$. For $m=2$, this
becomes $X_{j+1}\le X_{i+1}$ which
is already a contradiction because the sequence $(X_i)_{i\ge 0}$ is strictly increasing.
Thus, we may assume that $m\ge 3$.
Since $\mathcal{P}(2,m-1)$ holds, Proposition \ref{minimal:prop:Pjl} implies
that $\mathcal{P}(3,m-2)$ holds as well. So, for each large enough $i$, the subspace
$V_i=\mathcal{U}^{m-2}(A_3(i))$ of $\mathbb{R}^{m+3}$ has dimension at least $m+2$
and Proposition \ref{first:prop} gives
\begin{equation}
\label{even:lemma3:eq1}
H(V_i)\ll Y_2(i)^{1-(m-1)\lambda}Y_1(i)^{-\lambda}Y_0(i)^{-(1+e(i))\lambda}
\end{equation}
where $e(i)=\dim(V_i)-m-2\in\{0,1\}$. If $V_i\neq \mathbb{R}^{m+3}$ for infinitely many
$i$, then Lemma \ref{P0:lemma:A} provides arbitrarily large $i\in I$ for which
$\mathcal{U}^{m-2}(\mathbf{x}_{i-1})\not\subseteq V_i$. For those $i$, we have $e(i)=0$ and
Lemma \ref{P0:lemma:S} gives $1\ll H(V_i)L_{i-1}$. Then, we obtain
\begin{equation}
\label{even:lemma3:eq2}
1 \ll Y_2(i)^{1-(m-1)\lambda}Y_1(i)^{-\lambda}Y_0(i)^{-\lambda}X_i^{-\lambda}.
\end{equation}
Otherwise, we have $e(i)=1$ for all sufficiently large $i\in I$ and the above estimate
follows directly from \eqref{even:lemma3:eq1} since $1\le H(V_i)$ and
$Y_0(i)=X_{i+1}>X_i$. Thus \eqref{even:lemma3:eq2} holds for infinitely many
$i\in I$. Viewing such $i$ as part of a triple of consecutive elements
$h<i<j$ of $I$, we have $Y_1(i)=X_{j+1}$, $Y_0(i)=X_{i+1}$ and
$X_i\ge X_{h+1}$, thus
\begin{equation}
\label{even:lemma3:eq3}
Y_1(i)^\theta\ll Y_0(i) \quad\mbox{and}\quad Y_0(i)^\theta\ll X_i
\end{equation}
by our initial observation at the beginning of the proof.
So, we deduce that
\begin{equation}
\label{even:lemma3:eq4}
Y_2(i)^{1-(m-1)\lambda}
\gg Y_1(i)^{\lambda(1+\theta+\theta^2)} \gg Y_1(i)^{3\lambda\theta}
\end{equation}
for arbitrarily large $i\in I$. Since $m\ge 3$, we may also apply Corollary
\ref{alt:cor:E} with $j=2$ and $\ell=m-1$. Using \eqref{even:lemma3:eq3},
this gives
\[
Y_2(i)^{(m-2)\lambda}
\ll Y_1(i)^{1-2\lambda}Y_0(i)^{-\lambda}
\ll Y_1(i)^{1-2\lambda-\lambda\theta}
\]
for each $i\in I$. Substituting this upper bound for $Y_2(i)$ into
\eqref{even:lemma3:eq4} and then comparing powers of $Y_1(i)$, we conclude
that
\[
3(m-2)\lambda^2\theta \le (1-(m-1)\lambda)(1-2\lambda-\lambda\theta),
\]
and thus $3(m-2)\theta \le (1/\gamma-(m-1))(1/\gamma-2-\theta)$,
as $\lambda$ can be taken arbitrarily close to $\gamma$. However, this inequality
is false for the actual values of $\gamma$ and $\theta$. This contradiction shows
that $\mathcal{P}(2,m-1)$ does not hold.
\end{proof}
\subsection*{Proof of Theorem \ref{intro:thm:pair}}
We may assume that $\gamma<\lambda<\widehat{\lambda}_n(\xi)$. Then $\mathcal{P}(1,m-1)$ and
$\mathcal{P}(2,m-2)$ hold while $\mathcal{P}(2,m-1)$ does not hold, by Lemmas \ref{even:lemma1}
and \ref{even:lemma3}. In particular, Proposition \ref{P0:prop:I} applies
with $j=2$ and $\ell=m-1$, and so there are infinitely many integers $i\ge 1$
for which
\begin{equation}
\label{even:proof:eq1}
\dim\mathcal{U}^{m-1}(A_2(i))=m+1
\quad\mbox{and}\quad
\mathcal{U}^{m-1}(\mathbf{x}_{i-1})
\not\subseteq \mathcal{U}^{m-1}(A_2(i))
\varsubsetneq \mathbb{R}^{m+2}.
\end{equation}
For those $i$, Lemma \ref{P0:lemma:S} and Proposition \ref{P0:prop:I}
further give
\[
1\ll H(\mathcal{U}^{m-1}(A_2(i)))L_{i-1}
\quad\mbox{and}\quad
1\ll H(\mathcal{U}^{m-2}(A_2(i)))L_{i-1}^2.
\]
Any such $i$ belongs to $I$ and, upon denoting by $j$ its successor in $I$,
Proposition \ref{first:prop} gives
\[
H(\mathcal{U}^{m-2}(A_2(i)))
\ll Y_1(i)^{1-(m-1)\lambda}Y_0(i)^{-\lambda}
= X_{j+1}^{1-(m-1)\lambda}X_{i+1}^{-\lambda}.
\]
By $\mathcal{P}(1,m-1)$, we also have $\dim\mathcal{U}^{m-1}(A_1(i))\ge m+1$
if $i$ is large enough. Comparing with \eqref{even:proof:eq1}, this
implies that $\mathcal{U}^{m-1}(A_2(i))=\mathcal{U}^{m-1}(A_1(i))$ and so Lemma
\ref{P1:lemma2} gives
\[
H(\mathcal{U}^{m-1}(A_2(i)))
=H(\mathcal{U}^{m-1}(A_1(i)))
\ll X_{j+1}^{-(m-1)\lambda}X_{i+1}^{1-\lambda}.
\]
By Lemma \ref{even:lemma1}, we also have $L_{i-1}\ll X_i^{-\lambda}
\ll X_{i+1}^{-m\lambda^2}$. Combining all the above inequalities, we obtain
\begin{equation}
\label{even:proof:eq2}
1\ll X_{j+1}^{-(m-1)\lambda}X_{i+1}^{1-\lambda-m\lambda^2}
\quad\mbox{and}\quad
1\ll X_{j+1}^{1-(m-1)\lambda}X_{i+1}^{-\lambda-2m\lambda^2}.
\end{equation}
As we can take $i$ arbitrarily large, this in turn implies that
\[
0\le (1-(m-1)\lambda)(1-\lambda-m\lambda^2)-(m-1)\lambda(\lambda+2m\lambda^2).
\]
If $m\ge 3$, the right hand side of this inequality simplifies to $Q_m(\lambda)$.
So, in that case, we obtain $\lambda\le \beta$, thus $\widehat{\lambda}_n(\xi)\le \beta=\beta_m$
as needed.
For the case $m=2$, we look more closely at the vector spaces
\[
V_i=\mathcal{U}^1(A_1(i-1))=\mathcal{U}^1(\mathbf{x}_{i-1},\mathbf{x}_{i})\subseteq\mathbb{R}^4
\]
for each integer $i\ge 1$. Since $\mathcal{P}(1,1)$ holds, we have $\dim\mathcal{U}^1(\mathbf{x}_i)=2$
and $\dim(V_i)\ge 3$ for each large enough $i$. When $V_i=\mathbb{R}^4$, we find
$1=H(V_i)\ll X_iL_{i-1}^3\ll X_i^{1-3\lambda}$ since $V_i$ is generated by
points $\mathbf{y}\in\mathbb{Z}^4$ with $\norm{\mathbf{y}}\le X_i$ and $L_\xi(\mathbf{y})\ll L_{i-1}$.
As $\lambda>\gamma=1/3$, we conclude that both $\dim\mathcal{U}^1(\mathbf{x}_i)=2$
and $\dim(V_i)=3$ for each large enough $i$, say for $i\ge i_0$. Then, $V_i$
admits a basis of the form $\{\mathbf{x}_{i-1}^{(p,1)},\mathbf{x}_i^{(0,1)},\mathbf{x}_i^{(1,1)}\}$
for some $p\in\{0,1\}$, thus
\begin{equation*}
\label{even:proof:eq3}
H(V_i)\ll X_iL_iL_{i-1}\ll X_{i+1}^{-\lambda}X_i^{1-\lambda}
\quad\mbox{and}\quad
L_\xi(V_i)\ll L_i^2L_{i-1}\ll X_{i+1}^{-2\lambda}X_i^{-\lambda}.
\end{equation*}
For each $i\ge i_0$ for which \eqref{even:proof:eq1} holds, we have
$\mathcal{U}^1(\mathbf{x}_{i-1})\not\subseteq V_{i+1}$, thus $V_i\cap V_{i+1}=\mathcal{U}^1(\mathbf{x}_i)$
and so $\mathcal{U}^1(\mathbf{x}_{i+1})\not\subseteq V_i$. Then, by Proposition \ref{new:prop},
the point $C_i=C(V_i,\mathbf{x}_{i+1})\in\mathbb{Z}^2$ is non-zero with
\[
\norm{C_i}
\ll H(V_i)L_{i+1}+X_{i+1}L_\xi(V_i)
\ll X_{i+1}^{1-2\lambda}X_i^{-\lambda}.
\]
Moreover, by Lemma \ref{minimal:lemma:HA1}, the pair $\{\mathbf{x}_i,\mathbf{x}_{i+1}\}$
is a basis of $A_1(i)\cap\mathbb{Z}^5$ over $\mathbb{Z}$. So, letting $j$ denote the successor
of $i$ in $I$, we may write $\mathbf{x}_j=a\mathbf{x}_i+b\mathbf{x}_{i+1}$ for some $a,b\in\mathbb{Z}$
with $b\neq 0$. Since $C(V_i,\mathbf{x})$ is linear in $\mathbf{x}$ (for a fixed basis of $V_i$)
and since $C(V_i,\mathbf{x}_i)=0$, we find that $C(V_i,\mathbf{x}_j)=b C_i$. Thus, by
Proposition \ref{new:prop}, we obtain
\[
L_\xi(C_i)
\le L_\xi(C(V_i,\mathbf{x}_j))
\ll H(V_i)L_j
\ll X_{j+1}^{-\lambda}X_{i+1}^{-\lambda}X_i^{1-\lambda}.
\]
In particular, $L_\xi(C_i)\ll X_i^{1-3\lambda}$ converges to $0$ as $i\to\infty$, since
$\lambda>1/3$. Thus, by Corollary \ref{new:cor}, the product $\norm{C_i}L_\xi(C_i)^{\lambda}$ tends to infinity with $i$. So we have
\[
1 \ll \norm{C_i}L_\xi(C_i)^{\lambda}
\ll X_{j+1}^{-\lambda^2}X_{i+1}^{1-2\lambda-\lambda^2}X_i^{-\lambda^2}.
\]
Using the estimate $X_{i+1}^{2\lambda}\ll X_i$ from Lemma \ref{even:lemma1},
we conclude that
\begin{equation}
\label{even:proof:eq4}
X_{j+1}^{\lambda^2} \ll X_{i+1}^{1-2\lambda-\lambda^2-2\lambda^3}
\end{equation}
for each pair of consecutive elements $i<j$ of $I$ with $i\ge i_0$, for which
\eqref{even:proof:eq1} holds. For these, the two estimates \eqref{even:proof:eq2}
also apply. In particular, the second one yields
\[
X_{i+1}^{\lambda+4\lambda^2} \ll X_{j+1}^{1-\lambda}.
\]
Combining this with \eqref{even:proof:eq4}, we conclude that
$\lambda^2(\lambda+4\lambda^2)\le (1-\lambda)(1-2\lambda-\lambda^2-2\lambda^3)$,
which simplifies to $Q_2(\lambda)\ge 0$. This gives $\lambda\le \beta$ and thus
$\widehat{\lambda}_4(\xi)\le \beta=\beta_2$.
|
\section{Introduction}
\label{sec:Intro}
Phase transitions describe a fundamental change in the behavior of a system as its parameters are changed. Understanding why these systems exhibit different observable features in different physical scenarios and how they transition between them is a fundamental problem in many branches of physics. In quantum systems, the interest is in phase transitions, which occur near zero temperature and are driven by quantum fluctuations~\cite{Greiner:2002wt,Heyl_2018}. We are concerned here with an open quantum system, where particles of an ensemble may be lost to the environment. Open quantum systems constitute a realistic scenario from an experimental perspective, where lost particles can be measured to monitor the system as it evolves. The underlying non-unitary evolution can produce dissipation and decoherence not exhibited by their closed counterparts.
For both open and closed quantum systems, quantum fluctuations can become negligible in the thermodynamic limit of large particle number (e.g. atoms and/or photons). Their time evolution can be described by a so-called \textit{semiclassical model} or mean-field approximation \cite{OzoriodeAlmeidaA.2011}, which takes the form of a closed set of ordinary differential equations (ODEs) for associated averaged quantities. Such an approach has long been common in quantum optics to study systems with large numbers of photons, such as lasers, and it is the starting point of our analysis here. The attractors of the semiclassical ODE, that is, its stable solutions, therefore, give insight into the observable behaviour of the underlying quantum system. In other words, phase transitions exhibited by the quantum system can be identified as changes in the stability of solutions of the limiting semiclassical model, which are examples of \textit{bifurcations}. Thus, bifurcation analysis of the semiclassical ODE --- with a combination of analytical and advanced numerical tools --- allows one to systematically map out the attractors and phase transitions in the thermodynamical limit; see for example, \cite{guckenheimer1983nonlinear,Kuz1,redbook} as entry points to bifurcation theory and associated numerical methods. It has long been common in quantum optics to study systems with large numbers of photons, such as lasers, by means of their semiclassical descriptions.
We are interested here in the relationship between dynamics and bifurcations of the semiclassical ODE model and the observable behaviour of the quantum system when the number of particles of the considered system is relatively small. In this case, the system is far from the thermodynamic limit, and quantum fluctuations may be significant and cannot simply be neglected. Quantum simulations of the underlying Hamiltonian, which are only feasible computationally for small numbers of particles, can be used to investigate the predictive power of semiclassical models for novel types of quantum systems that operate with very few atoms and/or photons. It has been shown by means of quantum simulations that fingerprints of semiclassical predictions can still be found in probabilistic features of quantum systems for surprisingly low numbers of particles~\cite{Carmichael15,PhysRevA.98.023804,Vukics19,PhysRevResearch.Kevin2021}. As a recent example, the semiclassical limit of the unbalanced Dicke model \cite{PhysRevA.Kevin2021} predicts parameter regimes of superradiant switching, quantum hysteresis, and oscillations which have been observed in its quantum optical description \cite{PhysRevResearch.Kevin2021}.
This paper probes the connections between the semiclassical and quantum regime for the specific example of the open two-site Bose-Hubbard dimer \cite{someBHrefs1,someBHrefs2,Kordas2015}. This well-known quantum mechanical model describes the dynamics of bosonic particles in a lattice, where the behaviour is determined by the interplay of the hopping rate of particles between lattice sites and their on-site interactions. Different experimental realisations of this system have been achieved in the form of semiconductor microcavities~\cite{lagoudakis_coherent_2010,abbarchi_macroscopic_2013,rodriguez_interaction-induced_2016} and superconducting circuits~\cite{raftery_observation_2014,eichler_quantum-limited_2014}. Of particular interest to us is an optical realisation in the form of two lossy coupled photonic crystal resonators~\cite{Hamel_2015,garbin2021spontaneous}. In this setup, good agreement between experimental measurements and the semiclassical description has been obtained in a regime up to moderate strength of the optical drive signal, in particular, concerning the observation of spontaneous symmetry breaking~\cite{GBKYA_2020}. Recent theoretical results over a much wider range of parameters have uncovered a rich variety of behaviour of the semiclassical open Bose-Hubbard dimer, including chaotic behaviour with non-switching, regular switching and chaotic switching between the two sites; see \cite{GBK_2021} for more details. We focus here specifically on the comparison between the quantum and semiclassical descriptions of the two-site Bose-Hubbard model. Such comparisons have been performed before under the excitation of the anti-bonding mode~\cite{Casteels2017}, and under excitation of the bonding mode for the case of positive intermode coupling~\cite{BinMahmud,Lledo2019}. Here we consider the situation where the bonding mode is only excited while the intermode coupling of the cavities is negative. In this way, we extend earlier comparisons to the specific parameter constellations that most resembles the experimental setup in \cite{garbin2021spontaneous} and, to our knowledge, have not been studied before.
Our approach is as follows: we choose parameter values from \cite{GBKYA_2020} to obtain a bifurcation diagram of the semiclassical ODE model with different qualitative behaviour, specifically, symmetry breaking, periodic motion and multi-stability of asymmetric states. We then obtain temporal traces of the open Bose--Hubbard dimer by means of quantum trajectory simulations \cite{CarmichaelHoward2007Smiq}. We then compare the two. By performing simulations of the quantum system for an increasing number of photons, we are able to observe the emergence of semiclassical attractors in a fully quantum realisation even with low photon number. Furthermore, we consider signatures for the existence of anti-bunching \cite{PhysRevLett.108.183601} and entanglement \cite{PhysRevLett.2004_Howard,PhysRevLett.88.197901} in the quantum system as different bifurcations of the semiclassical approximation are encountered. Overall, our results showcase the role a semiclassical bifurcation analysis of a quantum system may play in providing a roadmap of interesting fundamental behaviour of interest, both from a theoretical and an experimental perspective. Specifically for the open two-site Bose--Hubbard dimer, this demonstrates that it might be feasible to investigate the quantum footprint of even quite complex dynamical behaviour, such as different types of chaotic switching~\cite{GBK_2021}.
The computations of equilibria and local bifurcations presented here are implemented in and performed with the pseudo-arclength continuation package \textsc{Auto07p} \cite{Doe1,Doe2} and its extension \textsc{HomCont} \cite{san2}. The quantum trajectory simulations were carried out in the software package \textsc{QuTiP} \cite{JohanssonJ.Qutip2012,JohanssonJ.Qutip2013}. Visualisation and post-processing of the data are performed with \textsc{Matlab}\textsuperscript{\textregistered}.
\section{Quantum Hamiltonian and semiclassical ODE model}
\label{sec:BHmodel}
The Hamiltonian of the Bose--Hubbard model for two cavities takes the form \cite{BinMahmud}
\begin{equation} \label{eq:Hamiltonian}
\begin{aligned}
\hat{H} = -J &(\hat{a}_1^{\dagger}\hat{a}_2 +
\hat{a}_2^{\dagger}\hat{a}_1 ) + \sum_{j=1,2}\left( \omega_c
\hat{a}_j^{\dagger}\hat{a}_j + U
\hat{a}_j^{\dagger}\hat{a}_j^{\dagger}\hat{a}_j \hat{a}_j \right)
\\
&+ \sum_{j=1,2}\left( F e^{-i\omega_p t}\hat{a}_j + F^* e^{i\omega_p t}\hat{a}^{\dagger}_j \right),
\end{aligned}
\end{equation}
where $\hat{a}_j^{\dagger}$ is the creation operator at cavity $j$. The intermode coupling is represented by $J$, the frequency inside the cavities by $\omega_c$ and the on-site energy by $U$. The last term of \eqref{eq:Hamiltonian} represents a driving field with frequency $\omega_p$ and amplitude $F$, which we consider equal for both cavities. In what follows, we work in a co-moving frame with respect of $\omega_p$; in this way, the time dependency of the Hamiltonian~\eqref{eq:Hamiltonian} is dropped. Since we are interested in two lossy mutually coupled cavities, the Lindblad master equation, which couples a thermal bath with the dynamics induced by \eqref{eq:Hamiltonian}, takes the form
\begin{equation} \label{eq:LindForm}
i\frac{d\hat{\rho}}{dt} = [ \hat{H}, \hat{\rho}] + i
\frac{\gamma}{2}\sum_{j=1,2}[2 \hat{a}_j \hat{\rho}
\hat{a}_j^{\dagger}-\hat{a}_j^{\dagger}\hat{a}_j \hat{\rho}-
\hat{\rho} \hat{a}_j^{\dagger}\hat{a}_j],
\end{equation}
where $\hat{\rho}$ is the density matrix and $\gamma$ is the loss rate.
System~\eqref{eq:LindForm} can be evolved numerically by computing quantum trajectories~\cite{CarmichaelHoward2007Smiq}, where the wavefunction is evolved with a Monte Carlo algorithm. Such a computation is performed for a non-Hermitian effective Hamiltonian
\begin{equation} \label{eq:Hamiltonian2}
\begin{aligned}
\hat{H} = -J &(\hat{a}_1^{\dagger}\hat{a}_2 +
\hat{a}_2^{\dagger}\hat{a}_1 ) + \sum_{j=1,2}\left( - \Delta \hat{a}_j^{\dagger}\hat{a}_j + U \hat{a}_j^{\dagger}\hat{a}_j^{\dagger}\hat{a}_j \hat{a}_j \right)
\\
&+ \sum_{j=1,2}\left( F \hat{a}_j + F^* \hat{a}^{\dagger}_j \right) - \frac{i\gamma}{2} \left( \hat{a}_1^\dagger \hat{a}_1+ \hat{a}_2^\dagger \hat{a}_2 \right).
\end{aligned}
\end{equation}
where the non-Hermitian terms account for null-measurement back-action; moreover, Eq.~\eqref{eq:Hamiltonian2} is expressed in a co-moving frame with respect to the frequency $\omega_p$ of the driving field, and $\Delta=\omega_c-\omega_p$ is the detuning. The coherent evolution is interrupted by `jumps', which correspond to photon emission from the cavities implemented by the action of $\hat{a}_{1}$ or $\hat{a}_2$ on the wavefunction. The timing of these jumps is stochastic, and each trajectory corresponds to a particular \textit{unravelling} of the master equation given by system~\eqref{eq:LindForm}. This approach is quite powerful as one can obtain temporal information regarding observables, which mimics the situation of an observer taking measurements of the outputs of the cavities; see \cite{CarmichaelHoward2007Smiq} for more details. We consider here the observables $\langle a^\dagger_1 a_1 \rangle$ and $\langle a^\dagger_2 a_2 \rangle$, the photon number expectation values in each of the two cavities.
In the co-moving frame with respect the driving frequency $\omega_p$, the mean-field approximation
\begin{equation} \label{eq:Coupled}
\begin{aligned}
i\dfrac{d \alpha_1}{d \tau} & = \left( -\Delta - i \dfrac{\gamma}{2} +2U|\alpha_1|^2 \right) \alpha_1 -J \alpha_2 + F, \\
i\dfrac{d \alpha_2}{d \tau} & = \left( -\Delta - i \dfrac{\gamma}{2} +2U|\alpha_2|^2 \right) \alpha_2 - J \alpha_1 + F,
\end{aligned}
\end{equation}
is derived from \eqref{eq:LindForm} by considering the expectations (mean-values) $\alpha_1 = \langle \hat{a}_1 \rangle$ and $\alpha_2 = \langle \hat{a}_2 \rangle$ under the assumption that their product factorises, meaning that, for example, $\langle \hat{a}_1\hat{a}_2 \rangle =\langle \hat{a}_1\rangle \langle \hat{a}_2 \rangle$; see \cite{Kordas2015} for details.
It is mathematically convenient to rescale time and the variables of \eqref{eq:Coupled} by introducing
$$\tau = 2t/\gamma \quad \mbox{and} \quad
(A,B) =(-2i \overline{\alpha_1}\sqrt{|U|/\gamma},-2i \overline{\alpha_2}\sqrt{|U|/\gamma})$$
to obtain the system
\begin{equation}
\label{eq:CoupledAB}
\begin{aligned}
\dfrac{d A}{d t} & = -A + i\left(\delta + \xi(U) |A|^2 \right) A +i \kappa B + f, \\
\dfrac{d B}{d t} & =-B + i\left(\delta + \xi(U) |B|^2 \right) B+ i \kappa A +f.
\end{aligned}
\end{equation}
for the rescaled complex field electric amplitudes $A$ and $B$, where
\begin{equation}
\label{eq:parscalings}
\xi(U)=\text{sign}(U), \quad \delta = -\frac{2\Delta}{\gamma}, \kappa=-\frac{2J}{\gamma} \quad \mbox{and} \quad f=4F\frac{\sqrt{|U|}}{\gamma^{3/2}}.
\end{equation}
System~\eqref{eq:CoupledAB} is the semiclassical ODE model of the open Bode-Hubbard dimer, and we will investigate how its bifurcations manifest themselves in the evolution as described by the Lindblad master equation~\eqref{eq:LindForm} for small photon numbers in the two cavities. Note that we are considering here the symmetric case where both cavities are pumped with the same intensity $F$; hence, system~\eqref{eq:LindForm} is invariant under the exchange of the two cavities given by $(\hat{a}_1, \hat{a}_2) \rightarrow (\hat{a}_2, \hat{a}_1) $. This reflectional symmetry manifests itself as $\mathbb{Z}_2$-equivariance \cite{Golubitsky1985} of the semiclassical system~\eqref{eq:CoupledAB} with respect to the map $(A,B) \to (B,A)$.
Mathematically, system \eqref{eq:CoupledAB} is a system of four real-valued ODEs, which can be written out either in terms of real and imaginary parts or amplitudes and phases of both $A$ and $B$; hence, it defines a vector field with a four-dimensional phase space. The sign $\xi(U)=\text{sign}(U)$ appears in \eqref{eq:CoupledAB} to allow for convenient comparisons with different instances of the Bose-Hubbard dimer in the literature; however, the transformation
$$(A,B,U,\delta,\kappa) \mapsto (\overline{A},\overline{B},-U,-\delta,-\kappa)$$
implies that all results for $\xi(U)=1$ directly translate to those for $\xi(U)=-1$. From now on, we set $\xi(U)=\text{sign}(U)=1$ and consider positive $\kappa$ as the experimentally relevant case \cite{Hamel_2015,garbin2021spontaneous,Haddadi_2014} whose bifurcations have been studied in considerable detail in \cite{GBKYA_2020,GBK_2021}. More specifically, we also set $\delta=-4.5$ and consider the one-parameter bifurcation diagram in the pump strength $f$, for which one finds different types of phase transitions, including symmetry breaking, multistability and the onset of periodicity.
We compare the dynamics of the semiclassical ODE model~\eqref{eq:CoupledAB} with quantum trajectory computations, requiring that we set $\left( J,\Delta,U,\gamma\right) = \left( -3.5,4.5,0.5,2.0\right)$ in system~\eqref{eq:LindForm} so that the two sets of parameter values agree. Regarding pump strength $F$, we note that any scaling of the form
\begin{equation}
\label{eq:scaleF}
(U,F) \rightarrow \left(U_\mu,F_\mu\right) = \left(U / \mu,\sqrt{\mu} F\right)
\end{equation}
with $\mu>0$ does not change the value of $f$ in system~\eqref{eq:CoupledAB} as given by \eqref{eq:parscalings}; hence, the rescaling \eqref{eq:scaleF} leaves the observed semiclassical behaviour unchanged. Classically, this can be understood as simultaneously changing both the strength of the nonlinearity and the optical
intensity. On the other hand, in the quantum regime, the scaling factor $\mu$ plays a very important role: it encodes the number of photons in the cavities. As such, it enters as a scaling when one compares intensities; namely $\langle a^\dagger_1 a_1 \rangle=\mu |A|^2$ and $\langle a^\dagger_2 a_2 \rangle=\mu |B|^2$, which follows from the scaling in~\eqref{eq:parscalings} to obtain the pump strength $f$ of the semiclassical model~\eqref{eq:CoupledAB}. Increasing $\mu$ increases the number of photons and, hence, reduces the importance of cavity field fluctuations. In other words, the scaling factor $\mu$ allows one to simulate the quantum system for different numbers of photons trapped in the cavities, while the nature of the limiting semiclassical dynamics remains the same. In particular, by considering different values of $\mu$, one can investigate the importance of quantum fluctuations. We remark that this approach has been used in \cite{Lledo2019} to investigate system~\eqref{eq:Hamiltonian} for the case $J>0$.
\section{Bifurcation diagram and tracked quantum trajectory simulation}
\label{sec:Phase}
\begin{figure}[t!]
\centering
\includegraphics{./figures/GMBK_QuanTraj_fig01.eps}
\caption{Bifurcation diagram of system~\eqref{eq:CoupledAB} in the pump strength $f$ compared with quantum trajectories of system~\eqref{eq:LindForm}. Panel~(a) shows branches of equilibria and periodic solutions represented by the intensities $|A|^2$ and $|B|^2$, namely: stable equilibria (blue), saddle equilibria with one unstable eigenvalue (cyan) and with two unstable eigenvalues (orange), and stable periodic solutions (dark green); these solutions bifurcate at points of pitchfork bifurcation $\mathbf{P}$, of saddle-node bifurcation of asymmetric equilibria $\mathbf{S_1^*}$ and $\mathbf{S_2^*}$, of saddle-node bifurcation of asymmetric equilibria $\mathbf{S}^*$, and of Andronov--Hopf bifurcation $\mathbf{H}$. Panel~(b) shows quantum trajectory realisations of system~\eqref{eq:Hamiltonian2} as $F_\mu$ with $\mu=3.0$ is slowly ramped linearly as $F_{\mu}\approx 0.216\,t$; here the observables $\langle \hat{a}^{\dagger}_{1} \hat{a}_{1} \rangle$ (red curve) and $\langle \hat{a}^{\dagger}_{2} \hat{a}_{2} \rangle$ (blue curve) are superimposed on the branches of solutions from panel~(a) (dark and light grey curves). Throughout, $(\delta,\kappa)=(-4.5, 3.5)$ for system~\eqref{eq:CoupledAB} and $\left( J,\Delta,U,\gamma\right) = \left( -3.5,4.5,0.5,2.0\right)$ for system~\eqref{eq:Hamiltonian2}; the vertical grey-dashed lines indicate the $f$-values considered in \sref{sec:ComF}.}
\label{fig:BifDiag}
\end{figure}
We first compare in \fref{fig:BifDiag} the one-parameter bifurcation diagram in the pump current $f$ of system~\eqref{eq:CoupledAB} for $(\delta,\kappa)=(-4.5, 3.5)$ with quantum trajectories of system~\eqref{eq:LindForm} computed for slowly increasing pump current with corresponding parameter values. \Fref{fig:BifDiag}(a) shows the branches and bifurcations of the equilibrium and periodic solutions of system~\eqref{eq:CoupledAB} in terms of their intensities $|A|^2$ and $|B|^2$ over the relevant range of the pump strength $f$. Here, one finds stable equilibria along blue curves, while cyan and orange curves represent unstable equilibria with different numbers of positive eigenvalues. The green curves represent stable periodic solutions; specifically, these curves trace out the maxima (only) in $|A|^2$ and $|B|^2$ of these periodic solutions. Throughout, there exists a symmetric equilibrium with $|A|^2=|B|^2$ and $\arg{(A)}=\arg{(B)}$, and one finds a number of bifurcations that give rise to other branches of solutions. The $f$-range is divided in this way into intervals~\textbf{(i)}--\textbf{(vii)} with different qualitative behaviour as follows:
\begin{enumerate}
\item[\textbf{(i)}]
$f=0$ to the pitchfork bifurcation $\mathbf{P_1}$: there exists only the symmetric equilibrium and it is stable.
\item[\textbf{(ii)}]
$\mathbf{P_1}$ to the Andronov--Hopf bifurcation $\mathbf{H}$: the symmetric equilibrium is now unstable and a pair of stable asymmetric equilibria exist; these bifurcate at $\mathbf{P_1}$ with $|A|^2>|B|^2$ and $|A|^2<|B|^2$, respectively.
\item[\textbf{(iii)}]
between the Andronov--Hopf bifurcations $\mathbf{H_1}$ and $\mathbf{H_2}$:
the asymmetric equilibria are unstable, and there is now a pair of stable periodic orbits that emerge and disappear at $\mathbf{H_1}$ and $\mathbf{H_2}$ --- one near each equilibrium with $|A|^2>|B|^2$ and $|A|^2<|B|^2$, respectively---.
\item[\textbf{(iv)}]
$\mathbf{H_2}$ to the saddle-node bifurcation of asymmetric states $\mathbf{S^*_1}$: the asymmetric equilibria are again stable and the situation is as in interval~\textbf{(ii)}.
\item[\textbf{(v)}]
between the saddle-node bifurcations of asymmetric states $\mathbf{S^*_1}$ and $\mathbf{S^*_2}$: two additional pairs of asymmetric equilibria exist, one stable and one unstable, so that there are now two pairs of stable asymmetric equilibria. The pair of unstable equilibria emerges or disappears with the respective pair of stable periodic orbits at the points $\mathbf{S^*_1}$ and $\mathbf{S^*_2}$.
\item[\textbf{(vi)}]
$\mathbf{S^*_2}$ to the pitchfork bifurcation $\mathbf{P_2}$: there is again a single pair of stable asymmetric equilibria as well as the unstable symmetric equilibrium, as in intervals~\textbf{(ii)} and~\textbf{(iv)}.
\item[\textbf{(vii)}]
beyond $\mathbf{P_2}$: the symmetric equilibrium is the only solution and stable again, as in interval~\textbf{(i)}.
\end{enumerate}
\Fref{fig:BifDiag}(b) shows the realisation of one quantum trajectory of system~\eqref{eq:LindForm} as $F_{\mu}$ with $\mu=3.0$ is increased linearly at a slow rate of approximately $0.216$ per unit of time $t$, where $\left( J,\Delta,U,\gamma\right) = \left( -3.5,4.5,0.5,2.0\right)$. The quantum trajectory is shown in terms of the two observables $\langle a^\dagger_1 a_1 \rangle$ and $\langle a^\dagger_2 a_2 \rangle$, and superimposed is the bifurcation diagram from panel~(a) after the corresponding scaling by $\mu$ of the intensities $|A|^2$ and $|B|^2$; here stable branches are dark grey and unstable branches light grey. Notice in \fref{fig:BifDiag}(b) how both observables follow the branch of symmetric equilibria in interval~\textbf{(i)} and then split into an asymmetric situation when the semiclassical system exhibits the pitchfork bifurcation~$\mathbf{P_1}$. Near and beyond this transition, in intervals~\textbf{(i)} and~\textbf{(iii)}, there are increased fluctuations that include switching between which of $\langle a^\dagger_1 a_1 \rangle$ or $\langle a^\dagger_2 a_2 \rangle$ is the larger observable. Past the Andronov--Hopf bifurcation $\mathbf{H_2}$, in interval~\textbf{(iv)}, the quantum trajectory is clearly localised at one of the asymmetric stable equilibria; more precisely, the one with $\langle a^\dagger_2 a_2 \rangle>\langle a^\dagger_1 a_1 \rangle$. As $F_{\mu}$ increases, the quantum trajectory jumps to a different asymmetric equilibrium, namely in between the asymmetric saddle-node bifurcations $\mathbf{S^*_1}$ and $\mathbf{S^*_2}$ where two pairs of asymmetric equilibria exist, which is interval~\textbf{(v)}. As $F_{\mu}$ increases further, $\langle a^\dagger_1 a_1 \rangle$ and $\langle a^\dagger_2 a_2 \rangle$ remain localised near this new asymmetric equilibrium in interval~\textbf{(vi)}. The two observables then come together near the pitchfork bifurcation $\mathbf{P_2}$ and then remain localised at the symmetric equilibrium that is stable again in parameter interval~\textbf{(vii)}.
Overall, \fref{fig:BifDiag} illustrates that the slowly ramped quantum trajectory effectively follows the stable branches of the one-parameter bifurcation diagram, meaning that it remains localised near one of the stable equilibrium solutions, but with a considerable level of fluctuations. Increased sensitivity and jumps induced by these fluctuations are observed near bifurcation points, in the $F_{\mu}$-range between $\mathbf{P_1}$ and $\mathbf{H_2}$ as well as that bounded by $\mathbf{S^*_1}$ and $\mathbf{S^*_2}$ with bistability between different asymmetric equilibria.
\section{Quantum trajectories at specific values of $f$}
\label{sec:ComF}
We proceed by investigating how these observed properties of the quantum trajectory manifest themselves in the parameter intervals with different limiting semiclassical dynamics; specifically, at the values of $f = F_\mu/\sqrt{\mu}$ in intervals \textbf{(i)}--\textbf{(vi)}, as indicated by the dashed vertical lines in \fref{fig:BifDiag}(a). Here, we consider two values of the scaling factor $\mu$, that is, two different photon numbers, namely $\mu=1.0$ and $\mu=3.0$. More specifically, we present for each case:
\begin{itemize}
\item[(1)]
temporal traces for a single realisation of a quantum trajectory of system~\eqref{eq:LindForm} represented by the observables $\langle a^\dagger_1a_1 \rangle$ and $\langle a^\dagger_2 a_2 \rangle$, shown over the corresponding range (that depends on $\mu$) with the respective equilibria of system~\eqref{eq:CoupledAB}.
\item[(2)]
the associated temporal traces of the ratio $$O=\dfrac{\langle a^\dagger_1 a^\dagger_2 a_1 a_2 \rangle}{\langle a^\dagger_1 a_1 \rangle \langle a^\dagger_2 a_2 \rangle},$$
which is a measure of the validity of the factorisation property of the quantum system used to derive the semiclassical ODE; note that $O=1$ means that the factorisation is exact.
\item[(3)]
two-dimensional histogram in the $(\langle a^\dagger_1a_1 \rangle, \langle a^\dagger_2 a_2 \rangle)$-plane, where the bins are given by a $200\times200$ grid of the corresponding shown ranges and with the respective equilibria and periodic solutions of system~\eqref{eq:CoupledAB}. The histograms are constructed from three different realisations of quantum trajectories by sampling 50,000 equidistant points in the time interval $(0,10^4)$; moreover, the invariance of system~\eqref{eq:LindForm} under the permutation of the sites is used to double the number of points and to correct asymmetric bias introduced by the sampling.
\end{itemize}
\subsection{Comparison in intervals \textbf{(i)} to \textbf{(iii)}}
\label{sec:Int_i_to_iii}
\begin{figure}[t!]
\centering
\includegraphics{./figures/GMBK_QuanTraj_fig02.eps}
\caption{Quantum trajectories of system~\eqref{eq:LindForm} in interval~\textbf{(i)} for $F_\mu/\sqrt{\mu} = f = 2.0$ with $\mu=1.0$ (left column) and with $\mu=3.0$ (right column). Panels~(a1) and~(b1) show the temporal trace of the observables $\langle \hat{a}^{\dagger}_{1} \hat{a}_{1} \rangle$ (red curve) and $\langle \hat{a}^{\dagger}_{2} \hat{a}_{2} \rangle$ (blue curve), and panels~(a2) and~(b2) show the associated evolution of the ratio $O=\langle a^\dagger_1 a^\dagger_2 a_1 a_2 \rangle / \langle a^\dagger_1 a_1 \rangle \langle a^\dagger_2 a_2 \rangle$. Panels~(a3) and~(b3) show histograms for a $200\times200$ grid in the $(|A|^2, |B|^2)$-plane constructed from three different quantum trajectories; the data is symmetrised and the diagonal symmetry line is shown in grey. Also shown are the equilibria of system~\eqref{eq:CoupledAB}, as dashed lines in the temporeal traces and as dots in the histograms, where colour reflects their stability as in \fref{fig:BifDiag}.}
\label{fig:Quant_Int_i}
\end{figure}
In interval~\textbf{(i)} there exists only the single symmetric stable equilibrium of the semiclassical ODE~\eqref{eq:CoupledAB}, which is stable and attracts all initial conditions. As \fref{fig:Quant_Int_i} shows, the respective quantum trajectory stays close to this equilibrium but is subject to clear fluctuations. This localisation is illustrated in panels~(a1) and (b1) with the temporal traces of the two observables $\langle \hat{a}^{\dagger}_{1} \hat{a}_{1} \rangle$ and $\langle \hat{a}^{\dagger}_{2} \hat{a}_{2} \rangle$, which can be seen to fluctuate around the corresponding equilibrium intensity-value. Here the ranges have been chosen to agree with the scaling by $\mu$ so that a direct comparison of the observables can be made, including with their limiting semiclassical behaviour. Notice that the fluctuations are relatively larger for $\mu=1$ in panel~(a1) than those for $\mu=3$ in panel~(b1). Similarly, the observable $O$, whilst subject to fluctuations in both cases, is on average further from its limiting value of $1.0$ for $\mu=1$ in panel~(a2) compared to $\mu=3$ in panel~(b2). It is interesting to note that, despite the factorisation assumption being relatively inaccurate in this regime, the two-dimensional histograms are still well centred around the stable equilibrium. The two-dimensional histograms in the respective ranges of the $(\langle \hat{a}^{\dagger}_{1} \hat{a}_{1} \rangle,\langle \hat{a}^{\dagger}_{2} \hat{a}_{2} \rangle)$-plane in \fref{fig:Quant_Int_i}(a3) and (b3) show distributions that are well centered around the stable equilibrium on the symmetry axis. Note that the colour map is scaled to account for the dependence of the size of the bins on $\mu$; this also allow for a direct comparison of the (relative) heights of the histograms for different values of $\mu$. There is considerable spread due to the fluctuations in the quantum trajectories, which are indeed comparable with the temporal traces in panels~(a1) and (b1). The spread is smaller and more symmetrical around the equilibrium for $\mu=3$ in panel~(b3) compared to that for $\mu=1$ in panel~(a3). This illustrates that the distribution becomes more Gaussian with a smaller variance as $\mu$ is increased.
\begin{figure}[t!]
\centering
\includegraphics{./figures/GMBK_QuanTraj_fig03.eps}
\caption{Quantum trajectories of system~\eqref{eq:LindForm} in interval~\textbf{(ii)} for $F_\mu/\sqrt{\mu} = f = 4.0$ as represented in \fref{fig:Quant_Int_i}.}
\label{fig:Quant_Int_ii}
\end{figure}
In interval~\textbf{(ii)}, past the pitchfork bifurcation $\mathbf{P_1}$, the symmetric equilibrium of system~\eqref{eq:CoupledAB} is now unstable, and there is a pair of stable asymmetric equilibria, each with their own basin of attraction. \Fref{fig:Quant_Int_ii} shows that quantum trajectories of~\eqref{eq:LindForm} display switching between the two asymmetric states in between (quite short) epochs of localisation near one of them. The switching appears to be dominant in the temporal trace for $\mu=1$ in panel~(a1), and epochs of localisation (while still short) are visible more clearly for $\mu=3$ in panel~(b1). Overall, the role of fluctuations seems to be much more important here than in interval~\textbf{(i)}, as they drive switching between the two stable solutions. These observations are represented in the two-dimensional histograms in panels~(a3) and (b3) by the fact that the distributions are now bimodal, quite broad and not sharply focused around the two stable asymmetric equilibria. Notice also the existence of a `bridge' between the areas of localisation near the asymmetric equilibria, which reflects the likely route for the switching driven by the quantum fluctuations. Again, the histogram is less broad, and its features are crisper for $\mu=3$ in panel~(b3) compared to $\mu=1$ in panel~(a3).
\begin{figure}[t!]
\centering
\includegraphics{./figures/GMBK_QuanTraj_fig04.eps}
\caption{Quantum trajectories of system~\eqref{eq:LindForm} in interval~\textbf{(iii)} for $F_\mu/\sqrt{\mu} = f = 6.0$ as represented in \fref{fig:Quant_Int_i}; also shown in the histograms is the pair of stable asymmetric periodic solutions (green closed curves).}
\label{fig:Quant_Int_iii}
\end{figure}
\Fref{fig:Quant_Int_iii} illustrates the situation in interval~\textbf{(iii)}, in between the two Hopf bifurcation points $\mathbf{H_1}$ and $\mathbf{H_2}$, with the new feature of a pair of attracting periodic solutions of system~\eqref{eq:CoupledAB} near the now unstable asymmetric equilibria. The quantum trajectories of~\eqref{eq:LindForm} still display switching between these two asymmetric periodic states with epochs near one of them, with a considerable level of fluctuations. As before, the level of fluctuations is relatively higher for $\mu=1$ in panels~(a1) and~(a2) than for $\mu=3$ in panels~(b1) and~(b2). Comparison with \fref{fig:Quant_Int_ii} shows that the fluctuations are larger compared to the situation in interval~\textbf{(ii)}. In particular, the fluctuations during epochs of localisation are now larger since the quantum trajectories are no longer attracted to a steady state. The histograms in \fref{fig:Quant_Int_iii}(a3) and~(b3) show distributions that illustrate the properties of quantum trajectories differently. While the maxima are near the stable periodic solutions of the semiclassical ODE, there appears to be no definite fingerprint of periodicity of the quantum trajectories in the $(\langle \hat{a}^{\dagger}_{1} \hat{a}_{1} \rangle,\langle \hat{a}^{\dagger}_{2} \hat{a}_{2} \rangle)$-plane. Notice that there is again a clear `bridge' of preferred switching, which is considerably sharper for $\mu=1$ compared to $\mu=3$, but the histograms do not appear to identify the pair of periodic attractors.
\begin{figure}[t!]
\centering
\includegraphics{./figures/GMBK_QuanTraj_fig05.eps}
\caption{Comparison of oscillatory behaviour in interval~\textbf{(iii)} for $F_\mu/\sqrt{\mu} = f = 6.0$. Panels~(a1), (b1) and (c1) show temporal traces of the observable $S_o=\langle \hat{a}^{\dagger}_{1} \hat{a}_{1} \rangle + \langle \hat{a}^{\dagger}_{2} \hat{a}_{2} \rangle$ of quantum trajectories of system~\eqref{eq:LindForm} with $\mu=1$, $\mu=3$ and $\mu=20$, respectively. Panels~(a2), (b2) and (c2) show the corresponding power spectra $|\mathcal{F}(S_o)|$ (green data); here the vertical purple line indicates the frequency of the corresponding periodic solution of system~\eqref{eq:LindForm}, which is shown in panel~(d) as a temporal traces of $S = |A|^2 + |B|^2$.}
\label{fig:Fourier}
\end{figure}
The above discussion shows that the frequent switching between the two localised oscillations obscures the possible periodicity of the observables $\langle \hat{a}^{\dagger}_{1} \hat{a}_{1} \rangle$ and $\langle \hat{a}^{\dagger}_{2} \hat{a}_{2} \rangle$. To identify the fingerprint of these semiclassical periodic oscillations in the quantum realm, we therefore now show in \fref{fig:Fourier} temporal traces of the observable $S_o=\langle \hat{a}^{\dagger}_{1} \hat{a}_{1} \rangle + \langle \hat{a}^{\dagger}_{2} \hat{a}_{2} \rangle$ of the quantum trajectories as well as their spectra, for $\mu=1$, $\mu=3$ and also for $\mu=20$. The observable $S_o$ (subject to the same scaling by $\mu$) is the quantum analogue of the total intensity $S = |A|^2 + |B|^2$ of the semiclassical ODE. Due to the symmetry properties of the Bose-Hubbard dimer, switching between symmetric states manifests itself much less in $S_o$. More specifically, due to its invariance under the permutation of the two sites, $S_o$ minimises fluctuation transients driven by the quantum system. This is why one can observe signs of periodicity in the temporal traces in \fref{fig:Fourier}(a1) to~(c1). The temporal trace of the semiclassical periodic orbit is shown in \fref{fig:Fourier}(d) for comparison; note that it is close to being sinusoidal, which is due to the periodic orbit still being close to the Hopf bifurcation. As $\mu$ is increased, fluctuations are reduced, and the periodicity in the temporal trace of $S_o$ becomes crisper. Indeed, the temporal trace for $\mu=20$ in panel~(c1) is recognised as a `noisy version' of the periodic signal in panel~(d) and, hence, clearly contains fingerprints of the semiclassical periodic solution. This observation is quantified by power spectra $|\mathcal{F}(S_o)|$ of the respective temporal traces. Already for $\mu=1$ in panel~(a2) the spectrum shows a recognisable peak near the main frequency of the semiclassical periodic temporal trace. For $\mu=3$ in panel~(b2), the spectrum is sharper and its frequency closer to that of the semiclassical oscillation, and this is even more the case for $\mu=20$ in panel~(c2). We conclude that \fref{fig:Fourier} clearly shows the emergence of periodicity in the quantum realm provided $\mu$, that is, the photon number, is taken to be sufficiently large (but still moderate). We remark that this phenomenon has also been observed recently in quantum trajectory simulations of the unbalanced Dicke model \cite{PhysRevResearch.Kevin2021}.
\subsection{Comparison in intervals \textbf{(iv)} to \textbf{(vi)}}
\label{sec:Int_vi_to_vi}
\begin{figure}[t!]
\centering
\includegraphics{./figures/GMBK_QuanTraj_fig06.eps}
\caption{Quantum trajectories of system~\eqref{eq:LindForm} in interval~\textbf{(iv)} for $F_\mu/\sqrt{\mu} = f = 10.0$ as represented in \fref{fig:Quant_Int_i}.}
\label{fig:Quant_Int_iv}
\end{figure}
In interval~\textbf{(iv)}, past the second Hopf bifurcation $\mathbf{H_2}$, the two asymmetric equilibria are again stable and the only attractors of system~\eqref{eq:CoupledAB}. In other words, the situation is topologically the same as in interval~\textbf{(ii)}. However, as \fref{fig:Quant_Int_iv} shows, we find marked differences in the observed behaviour of quantum trajectories of~\eqref{eq:LindForm}. The temporal trace for $\mu=1$ in panel~(a1) show long epochs where the quantum trajectory is localised near one of the stable steady states, with much more occasional switchings between them. For $\mu=3$ in panel~(b1), the relative strength of quantum fluctuations is now so low that not a single switching occurs in the time window presented. Comparison with \fref{fig:Quant_Int_ii} shows that the overall level of fluctuation is comparable in intervals~\textbf{(ii)} and~\textbf{(iv)}. However, the observable $O$ is now seen to fluctuate around unity in panels~(a2) and~(b2), meaning that the factorization assumption is reasonable in interval~\textbf{(iv)}. Notice further that switching events manifest themselves in the observable $O$ in panel~(a2) as sudden larger spikes away from its average. The clear observation that quantum trajectories linger much longer near one of the two stable equilibria in interval~\textbf{(iv)} is explained by the fact that these equilibria are more attracting and also further apart from each other and from the unstable symmetric equilibrium from which they bifurcated. More specifically, the two stable equilibria represent a situation of extreme symmetry breaking where one site has practically all photons of the overall coupled system, while the other has near-zero photons --- this in spite of the fact that both sites are pumped symmetrically. This distance between the two attractors and the increased residence times of quantum trajectories are illustrated very clearly by the histograms in panels~(a3) and~(b3). Already for $\mu=1$ in panel~(a3), the distribution is very bimodal with pronounced peaks near the two attracting equilibria system~\eqref{eq:CoupledAB}, which are very close to the coordinate axes owing to the fact that one of the two intensities is practically zero; the `bridge' corresponding to preferred switching is now much weakened. For $\mu=3$ in panel~(b3), there is no longer a discernible bridge due to the occurrence of very few transitions between the two attractors. The histogram is now quite sharply focused on the pair of stable asymmetric equilibria, with very well defined peaks. That notwithstanding, there are still sufficiently many switchings due to quantum fluctuations to ensure that the computed histogram captures both attractors.
\begin{figure*}
\centering
\includegraphics{./figures/GMBK_QuanTraj_fig07.eps}
\caption{Quantum trajectories of system~\eqref{eq:LindForm} in interval~\textbf{(v)} for $F_\mu/\sqrt{\mu} = f = 13.0$ as represented in \fref{fig:Quant_Int_i}.}
\label{fig:Quant_Int_v}
\end{figure*}
The behaviour in interval~\textbf{(v)}, in between the two saddle-node bifurcations $\mathbf{S^*_1}$ and $\mathbf{S^*_2}$ is characterised by the existence of an additional pair of stable asymmetric equilibria of system~\eqref{eq:CoupledAB}, as well as a pair of unstable asymmetric equilibria. This means that there are now a total of four attractors with their respective basins of attraction. The additional stable equilibria are characterised by a relatively small imbalance between the sites compared to the other pair, which still represent the scenario where practically all photons are at one of the two sites. \Fref{fig:Quant_Int_v} shows that quantum trajectories of~\eqref{eq:LindForm} remain localised near one of the four stable equilibria for a certain amount of time and then switch to being localised near another stable equilibrium and so on. The temporal trace for $\mu=1$ in panel~(a1) shows quite a number of switchings. While for $\mu=3$ in panel~(b1), residence times are longer, and the number of switchings per time interval is decreased. Interestingly, the overall level of fluctuation is relatively quite low compared to earlier cases. Note also that, as seen in panels~(a2) and ~(b2), localisation near the new equilibria with smaller values of the intensity is associated with especially low fluctuations and around an average value of $O=1.0$, implying the factorization assumption is accurate in this regime. These attractors feature significant photon numbers in both cavities, such that the role of quantum fluctuations is less important. However, whilst the other attractors feature a large population in one cavity, the other, almost empty, cavity is strongly impacted by fluctuations. The four different attractors are also clear features of the histograms in panels~(a3) and~(b3). The distributions show clear peaks near each of the stable equilibria, which are considerable crisper for $\mu=3$ compared to $\mu=1$. Moreover, the histograms are considerably larger near the pair of equilibria where one of the intensities is practically zero. Note that switching events follow `weak bridges' between the attractors, meaning that switching between neighbouring attractors in the $(\langle \hat{a}^{\dagger}_{1} \hat{a}_{1} \rangle,\langle \hat{a}^{\dagger}_{2} \hat{a}_{2} \rangle)$-plane are by far the ones that are most likely to occur.
\begin{figure}[t!]
\centering
\includegraphics{./figures/GMBK_QuanTraj_fig08.eps}
\caption{Quantum trajectories of system~\eqref{eq:LindForm} in interval~\textbf{(vi)} for $F_\mu/\sqrt{\mu} = f = 17.0$ as represented in \fref{fig:Quant_Int_i}.}
\label{fig:Quant_Int_vi}
\end{figure}
\Fref{fig:Quant_Int_vi} shows the situation in interval~\textbf{(vi)}, where there are again only two stable asymmetric equilibria of~\eqref{eq:CoupledAB}, namely the ones near the symmetry line with a relatively small imbalance between the sites. As panels~(a1) and~(b1) show, the quantum trajectories of~\eqref{eq:LindForm} remain near both of these equilibria, with negligible residence times near each attractor and many switchings per time interval. Residence times are notably larger for $\mu=3$ that for $\mu=1$, but remain very small for either case compared to the (topologically equivalent) situation in interval~\textbf{(iv)} in \fref{fig:Quant_Int_iv}. The level of fluctuations appears to be relatively small, and the observable $O$ in \fref{fig:Quant_Int_vi}(a2) and~(b2) remains very close to $1.0$ throughout; again, any noticeable fluctuations of $O$ appear to be due to switchings between the two stable equilibria. However, the role of these fluctuations is important, as the frequent switching means that the associated histograms in panels~(a3) and~(b3) are characterised by a single peak with a maximum at a point in between the two stable asymmetric equilibria. In other words, the two nearby attractors are not resolved: while the histogram for $\mu=3$ in panel~(b3) appears to be more elongated around the two attracting equilibria, it still has only a single peak. This is a direct reflection of the very low residence times of localisation of as measured by the observables $\langle \hat{a}^{\dagger}_{1} \hat{a}_{1} \rangle$ and $\langle \hat{a}^{\dagger}_{2} \hat{a}_{2} \rangle$. As we will see below, it is possible to detect the (weak) localisation of the quantum trajectories already for $\mu=3$, namely by considering the difference between these two observables; see already \fref{fig:Violin}(b2).
\section{Evolution of histograms with $F_{\mu}$}
\label{sec:violin}
\begin{figure}[t!]
\centering
\includegraphics{./figures/GMBK_QuanTraj_fig09.eps}
\caption{Scaled histograms of system~\eqref{eq:LindForm} for $\mu=1$ (left column) and for $\mu=3$ (right column) of the observables $\langle \hat{a}^{\dagger}_{i} \hat{a}_{i} \rangle$ (top row) and of their difference $D_o=\langle \hat{a}^{\dagger}_{1} \hat{a}_{1} \rangle - \langle \hat{a}^{\dagger}_{2} \hat{a}_{2} \rangle$ (bottom row) at different values of $F_\mu$ superimposed on the bifurcation diagram of system~\eqref{eq:CoupledAB}. Here $F_{\mu}$ is increased from $2.0\mu$ to $15.75\mu$ in steps of $0.5\mu$, and the histograms are scaled to have a maximum of $0.5\mu$. The contrast in the background shading of the panels highlights intervals~\textbf{(i)} to~~\textbf{(vi)}.}
\label{fig:Violin}
\end{figure}
\Fref{fig:Violin} provides an overview of how the statistical properties of quantum trajectories of system~\eqref{eq:LindForm} develop with the pump strength $F_\mu$, and how this compares with the corresponding bifurcation diagram of the semiclassical ODE~\eqref{eq:CoupledAB}. Here, we again consider the two cases $\mu=1$ and $\mu=3$. In this representation, histograms are shown for $F_{\mu}$ starting from $2.0\mu$ in steps of $0.5\mu$ and up to $15.75\mu$, where the maximum of each histogram is scaled to the $F_{\mu}$-size of $0.5\mu$. As for the two-dimensional histograms shown in previous figures, these histograms are constructed from three different realisations of quantum trajectories by sampling 50,000 equidistant points in the time interval $(0,10^4)$, where we now use 80 uniform bins over the shown range of the respective observable. The invariance of system~\eqref{eq:LindForm} under the permutation of the sites is again used to double the number of points and symmetrise the data; the maxima are then scaled to $0.5\mu$.
\Fref{fig:Violin} shows histogram plots for two different observables. Panels~(a2) and~(b2) show histograms for $\langle \hat{a}^{\dagger}_{i} \hat{a}_{i} \rangle$, which corresponds to the projection of the respective two-dimensional histogram onto the $\langle \hat{a}^{\dagger}_{1} \hat{a}_{1} \rangle$-axis (or equivalently the $\langle \hat{a}^{\dagger}_{2} \hat{a}_{2} \rangle$-axis). To obtain additional information regarding semiclassical fingerprints in the distributions, \fref{fig:Violin}(a2) and~(b2) show histograms for the difference $D_o=\langle \hat{a}^{\dagger}_{1} \hat{a}_{1} \rangle - \langle \hat{a}^{\dagger}_{2} \hat{a}_{2} \rangle$; note that considering this observable corresponds to the projection of the respective two-dimensional histogram onto the antidiagonal and implies the symmetry of panels~(a2) and~(b2) with respect to the $F_{\mu}$-axis.
The evolution of the histograms for $\mu=1$ is shown in \fref{fig:Violin}(a1) and~(a2). There are noticeable changes in the statistical properties of the two observables associated with the transitions through the different bifurcations, while the distributions remain more or less the same in the intervals~\textbf{(i)} to~\textbf{(vi)} that are covered by the shown range of $F_{\mu}$; compare with \fref{fig:BifDiag}. Starting at low $F_{\mu}$ in interval~\textbf{(i)}, one first observes that the histogram widens as the first pitchfork bifurcation is approached and subsequently becomes bimodal in interval~\textbf{(ii)}, with a pair of peaks near the two stable asymmetric equilibria. Notice the large component of the distribution in between these two equilibria, which corresponds to frequent switchings between them. This `bridge' and, hence, the number of switchings in the time interval clearly become much less pronounced past interval~\textbf{(iii)}. The existence of stable periodic orbits, on the other hand, is not evident in the histograms. When the second Hopf bifurcation is reached, the distribution is strongly bimodal and remains clearly localised near the re-stabilised asymmetric equilibria throughout interval~\textbf{(iv)}. Notice that in \fref{fig:Violin}(a1) for the observable $\langle \hat{a}^{\dagger}_{i} \hat{a}_{i} \rangle$ the solution is less well resolved near the upper branch compared to the lower branch, due to the larger fluctuations for the site with more photons; this issue clearly does not arise for the symmetric observable $D_o$ in panel~(a2). The strong localisation extends well into interval~\textbf{(v)} with the additional pair of stable equilibria. There is a somewhat gradual change of the histogram to localisation around the other pair of stable equilibria in this interval. Notice that the distinction between these two equilibria is quite weak initially and quickly becomes nonexistent, with a distribution with a maximum in between the two attractors, even for the observable $D_o$.
The evolution of the histograms for $\mu=3$ in \fref{fig:Violin}(b1) and~(b2) is quite similar, but the distributions are more clearly resolved, that is, more concentrated at the respective attractor. However, there are noticeable differences in the upper range of $F_{\mu}$. In interval~\textbf{(v)}, the switching to a different pair of attractors now manifests itself as a quite sudden change of the histogram; this reflects the scarcity of switchings in a finite time series. Moreover, the two new asymmetric equilibria are now distinguished by the histogram, especially clearly for the observable $D_o$ in panel~(b2). Overall, we conclude that the histogram plots of \fref{fig:Violin}, especially those in panels~(a2) and~(b2) for the symmetric observable $D_o$, provide a good summary of how the statistical properties of quantum trajectories change both with the pump strength $F_\mu$ as well as with increasing numbers of photons as represented by the scaling parameter $\mu$. Indeed, this representation agrees with the results in \sref{sec:ComF} regarding the behaviour for representative values of $F_\mu$ in the intervals~\textbf{(i)} to~\textbf{(vi)} --- but it also provides insight into how the distributions of quantum trajectories change from interval to interval, as semiclassical bifurcations are encountered.
\section{Antibunching and entanglement}
\begin{figure}[t!]
\centering
\includegraphics{./figures/GMBK_QuanTraj_fig10.eps}
\caption{The minimum $\min\{g_{11}^{(2)}(0), g_{22}^{(2)}(0)\}$ of the averaged $g^{(2)}$-functions of the two sites~(a) and the averaged von Neumann entropy $E$~(b) as a function of pump strength $F_\mu$ as functions of the pump strength $F_\mu$. Each curve was computed from a single ramped quantum trajectory of system~\eqref{eq:LindForm}, where $\mu =0.5$ (cyan), $\mu =1.0$ (blue), $\mu =2.0$ (purple), and $\mu =3.0$ (lilac). Red-dashed vertical lines are at the bifurcations of~\eqref{eq:CoupledAB}, and the corresponding intervals~\textbf{(i)} to~\textbf{(vii)} are highlighted by contrasting background shading.}
\label{fig:AntiBunEnt}
\end{figure}
We now investigate whether there is evidence of quantum phenomena in system~\eqref{eq:LindForm} for low photon numbers as the pump strength $F_\mu$ is varied. To this end, \fref{fig:AntiBunEnt} shows the two indicator functions $\min\{g_{11}^{(2)}(0), g_{22}^{(2)}(0)\}$ and $E$ over the relevant range of $F_\mu$, where the semiclassical bifurcations are shown and the associated intervals~\textbf{(i)} to~\textbf{(vii)} of different behaviour are highlighted.
Antibunching of photons refers to the emission of a photon reducing the probability of a subsequent emission, and is a strictly quantum phenomena. In the present context, it can be detected by the condition that $\min\{g_{11}^{(2)}(0), g_{22}^{(2)}(0)\}<1$, where the function $g^{(2)}_{ii}(0)$ is the second-order correlation function for the light in cavity $i$
$$g^{(2)}_{ii}(0) = \frac{\langle \hat{a}^\dagger_i \hat{a}^\dagger_i \hat{a}_i \hat{a}_i \rangle}{\left(\langle \hat{a}^\dagger_i \hat{a} \rangle\right)^2}.$$
Plotting the smaller of the time averaged second-order correlation functions of the two sites, $g^{(2)}_{11}(0)$ and $g^{(2)}_{22}(0)$, for a single quantum trajectory allows us to identify antibunching \cite{PhysRevLett.108.183601}. \Fref{fig:AntiBunEnt}(a) shows time-averaged values of $\min\{g_{11}^{(2)}(0), g_{22}^{(2)}(0)\}$ for $\mu =0.5$, $\mu =1.0$, $\mu =2.0$, and $\mu =3.0$ at different values of $F_\mu$ for single trajectories of length $t = 500$. In interval~\textbf{(i)}, both sites display slight bunching, less so for larger $\mu$, implying a thermal photon number distribution. Past the pitchfork bifurcation~$\mathbf{P_1}$ and the appearance of asymmetric attracting states of system~\eqref{eq:CoupledAB}, we find $\min\{g_{11}^{(2)}(0), g_{22}^{(2)}(0)\}<1$: consistently in intervals~\textbf{(iii)} to~\textbf{(vii)} for all shown values of $\mu$, and including interval~\textbf{(ii)} for $\mu =0.5$ and $\mu =1.0$. Hence, the site with more photons is always anti-bunched from about $F_\mu = 5.0$ and higher. As one would expect, antibunching is reduced as $\mu$ increases and the closer the system is to its semiclassical limit, and $\min\{g_{11}^{(2)}(0), g_{22}^{(2)}(0)\}$ appears to approach the value $1.0$ as $\mu$ is increased.
To investigate the existence of entanglement between the sites, we consider the time-averaged von Neumann entropy of one of the cavities, given by~\cite{PhysRevLett.2004_Howard,PhysRevLett.88.197901}
$$E = - \overline{\mbox{trace}(\rho_1(t) \ln \rho_1(t))},$$
where $\rho_1(t)$ is the density matrix for site 1 at time $t$ (or equivalently site 2) and the overline represents time averaging. \Fref{fig:AntiBunEnt}~(b) shows $E$ as a function of $F_\mu$ for $\mu =0.5$, $\mu =1.0$, $\mu =2.0$, and $\mu =3.0$, as computed from the same quantum trajectories described above. Entanglement is identified by the condition that $E>0$, which means that the states are not `pure', and we conclude that there is entanglement throughout the entire range of $F_\mu$. Notice that the level of entanglement as measured by $E$ increases in interval~\textbf{(i)} and then, past the first pitchfork bifurcation~$\mathbf{P_1}$, reaches and stays on a plateau throughout interval~\textbf{(ii)} to~\textbf{(v)}. The indicator $E$ then increases quite steeply in interval~\textbf{(vi)} with maximal multistability between asymmetric attractors, reaches a maximum near the second pitchfork bifurcation~$\mathbf{P_2}$ and then decreases equally steeply in interval~\textbf{(vii)} where the symmetric equilibrium is again the only attractor of the limiting system~\eqref{eq:CoupledAB}. The von Neumann entropy in the plateau appears to show slowly reducing entanglement as $\mu$ increases, suggesting that entanglement goes away in the thermodynamic limit. However, the maximum of $E$ appears to become larger and more narrow with increasing $\mu$. It is a known phenomenon that, near certain types of phase transitions, there is a divergence in entanglement at the critical point in the thermodynamic limit~\cite{Osterloh:2002wc,PhysRevLett.92.073602,Reslen_2005,PhysRevA.71.064101}. Whether this is the explanation for the sharpening of the von Neumann energy near $\mathbf{P_2}$ remains an interesting question for future research.
\section{Conclusions and outlook}
\label{sec:concl}
The case study of the open two-site Bose-Hubbard dimer presented here shows that it is possible to identify recognisable fingerprints of phase transitions --- that is, bifurcations --- of the limiting semiclassical (mean-field) model in the quantum realm. More specifically, we considered the case of negative intermode coupling when the semiclassical model features a transition, as the pump strength is increased, from symmetric dynamics via symmetry breaking at a pitchfork bifurcation to oscillatory dynamics and then to multistability between different types of asymmetric states. These features were recognised reliably in the statistical properties of different observables of quantum trajectories, even quite far from the semiclassical limit, that is, for low numbers of photons at each site.
These theoretical results are fundamental in nature and suggest that finding quantum signatures of such rather complex nonlinear phenomena may be possible in an experimental context. For the system under study here, this has been achieved up to a moderate level of the pump strength, allowing for the experimental verification of spontaneous symmetry breaking in good agreement with a bifurcation study of the semiclassical model \cite{GBKYA_2020}. We believe that the experimental identification of more complicated dynamics with additional levels of multistability --- including that involving different types of localised and non-localised chaotic dynamics \cite{GBK_2021} --- should be possible. For the Bose-Hubbard dimer or other quantum systems such as the Dicke model, such experiments are challenging but would present an opportunity for studying how semiclassical chaos arises in quantum systems~\cite{PhysRevA.34.482,PhysRevLett.107.100401}.
\section*{Acknowledgement}
We thank Bruno Garbin, Ariel Levenson and Alejandro Yacomotti for sharing their insights on the open Bose--Hubbard dimer model, especially concerning their experiments. We also thank Ricardo Gutierrez-Jauregui for helpful comments.
\bibliographystyle{siam}
|
\section*{Supplementary Information}
We use $\vect{\mathbb{R}} $ to denote the set of real numbers and $\vect{\mathbb{N}} $ to denote the set of natural numbers. We use $\|.\|$ to denote the Euclidean norm for vectors and the Frobenius norm for matrices, while using bold symbols to illustrate matrices and vectors. We define an interval $[0, K ), K \in \vect{\mathbb{N}} $ and let $p (\cT )$ be the distribution over all the tasks observed in this interval. For any $k \in [0, K ) ,$ we define a parametric model $g (. )$ with ${\vect{y}}_{k} = g (\vect{x}_{k}; {\vect{\theta}}_{k} )$, where ${\vect{\theta}}_{k}$ is a vector comprising all parameters of the model with $\vect{x}_{k} \in \cX_{k}$. Let $n$ be the number of samples and $m$ be the number of dimensions. Suppose a task at $k | k \in [0, K )$ is observed and denoted as $\cT_{k} : \cT_{k} \sim p (\cT )$, where $\cT_{k} =\{\cX_{k}, \ell_{k}\}$ is a tuple with $\cX_{k} \in \vect{\mathbb{R}} ^{n m}$ being the input data and $\ell_{k}$ quantifies the loss incurred by $\cX_{k}$ using the model $g$ for the task at $k$. We denote a sequence of ${\vect{\theta}}_{k}$ as $\vect{u}_{k:K} = \{{\vect{\theta}}_{\tau} \in \Omega_{\theta}, k \leq \tau \leq K \},$ with $\Omega_{\theta}$ being the compact~ (feasible ) set for the parameters. We denote the optimal value with a superscript ${ (* )};$ for instance, we use ${\vect{\theta}}_k^{ (* )}$ to denote the optimal value of ${\vect{\theta}}_{k}$ at task $k.$ In this paper we use balance point, equilibrium point, and saddle point to refer to the point of balance between generalization and forgetting. We interchange between these terms whenever convenient for the discussion. We will use $\nabla_{ (j )} i$ to denote the gradient of $i$ with respect to $j$ and $\Delta i$ to denote the first difference in discrete time.
\section{Additional Results}
We define the cost~ (combination of catastrophic cost and generalization cost ) at any instant $k$ as $J_{k} ({\vect{\theta} }_{k} ) = \gamma_{k} \ell_{k} + \sum_{\tau = 0}^{k-1} \gamma_{\tau} \ell_{\tau},$ where $\ell_{\tau}$ is computed on task $\cT_{\tau}$ with $\gamma_{\tau}$ describing the contribution of $\cT_{\tau}$ to this sum. We will show that for any fixed $k$, the catastrophic forgetting cost~$J_k ({\vect{\theta} }_{k} )$ is divergent in the limit $k \rightarrow \infty$ if equal contribution from each task is expected.
\begin{lemma}
For any $k \in \vect{\mathbb{N}} ,$ define $J_k ({\vect{\theta} }_{k} ) = \sum_{\tau = 0}^{k} \gamma_{\tau} \ell_{\tau}.$ For all $\tau,$ assume $\ell_{\tau}$ to be continuous with $L \geq \ell_{\tau} \geq \epsilon, \forall \tau, \epsilon >0$ and let $\gamma_{\tau} =1$. Then $J_k ({\vect{\theta} }_{k} )$ is divergent as $k \rightarrow \infty.$
\end{lemma}
\begin{proof}[Proof of Lemma 1]
With $J_k ({\vect{\theta} }_{k} ) = \sum_{\tau = 0}^{k} \gamma_{\tau} \ell_{\tau}$ we write $\lim_{k \rightarrow \infty} \sum_{\tau = 0}^{k} \gamma_{\tau} \ell_{\tau} \geq \lim_{k \rightarrow \infty} \sum_{\tau = 0}^{k} \gamma_{\tau} \epsilon $, where $\gamma_{\tau}=1$ which provides $ \lim_{k \rightarrow \infty} \sum_{\tau = 0}^{k} \epsilon = \infty.$ Therefore, $J_k ({\vect{\theta} }_{k} )$ is divergent.
\end{proof}
When $\ell_{\tau} \geq \epsilon$ with $\epsilon >0$ implies that each task incurs a nonzero cost. Furthermore, $\gamma_{\tau} =1,$ it implies that each task provides equal contribution to the catastrophic forgetting cost and contributed nonzero value to $J_k ({\vect{\theta} }_{k} )$. The aforementioned lemma demonstrates that equivalent performance~ (no forgetting on all tasks ) cannot be guaranteed for an infinite number of tasks when each task provides a nonzero cost to the sum~ (you have to learn for all the tasks ). However, if the task contributions are prioritized based on knowledge about the task distribution, the sum can be ensured to be convergent as shown in the next corollary.
\begin{corr}
For any $k \in \vect{\mathbb{N}} ,$ define $J_k ({\vect{\theta} }_{k} ) = \sum_{\tau=0}^{k} \gamma_{\tau} \ell_{\tau}$ where $\ell_{\tau}$ is continuous and bounded such that $\epsilon \leq \ell_{\tau} \leq L, \forall \epsilon >0.$ Define $N= \frac{1}{k}$ and choose $\gamma_{N}$ such that $\gamma_{N} \rightarrow 0, N \rightarrow \infty$ and assume when there are infinite number of tasks, $\lim_{ N \rightarrow \infty} \sum_{N} \gamma_{N} \leq M.$ Under these assumptions, $J_k ({\vect{\theta} }_{k} )$ is convergent.
\end{corr}
\begin{proof}[Proof of Corollary 1]
Since $\ell_{\tau} \leq L$, $J_k ({\vect{\theta} }_{k} ) = lim_{k \rightarrow \infty} \sum_{\tau=0}^{k} \gamma_{\tau} \ell_{\tau} \leq lim_{k \rightarrow \infty} \sum_{\tau=0}^{k} \gamma_{\tau} L \leq L lim_{k \rightarrow \infty} \sum_{\tau = 0}^{k} \gamma_{\tau}.$
Since $\lim_{k \rightarrow \infty} \sum_{\tau = 0}^{k} \gamma_{\tau} = \lim_{ N \rightarrow \infty} \sum_{N} \gamma_{N}$ as $N =\frac{1}{k}$, therefore $\lim_{k \rightarrow \infty} \sum_{\tau = 0}^{k} \gamma_{\tau} \leq M$ and $J_k ({\vect{\theta} }_{k} )$ is upper bounded by $ L M$. As a result, $J_k ({\vect{\theta} }_{k} )$ is convergent since $J_k ({\vect{\theta} }_{k} )$ is a monotone.
\end{proof}
To solve the problem at $k$, we seek ${\vect{\theta} }_{k}$ to minimize $J_{k} ({\vect{\theta} }_{k} )$. Similarly, to solve the problem in the complete interval $[0, K )$, we seek a ${\vect{\theta} }_{k}$ to minimize $J_{k} ({\vect{\theta}}_{k} )$ for each $k \in [0,K ).$ In other words we seek to obtain ${\vect{\theta} }_{k}$ for each task such that the cost $J_{k} ({\vect{\theta} }_{k} )$ is minimized. The optimization problem for the overall CL problem~ (overarching goal of CL ) is then provided as the minimization of the cumulative cost $V_{k} (\vect{u}_{k:K} ) = \sum_{\tau=k}^{K} \beta_{\tau} J_{\tau} ( {\vect{\theta}}_{\tau} )$ such that $V_k^{ (* )},$ is given as \begin{equation} V_k^{ (* )} = min_{\vect{u}_{k:K} } V_{k} (\vect{u}_{k:K} ), \label{op} \end{equation} with $0 \leq \beta_{\tau} \leq 1$ being the contribution of $J_{\tau}$ and $\vect{u}_{k:K}$ being a weight sequence of length $K-k.$ We will now derive the difference equation for our cost formulation.
\begin{prop}
For any $k \in [0, K ),$ define $ V_k = \sum_{\tau=k}^{K} \beta_{\tau} J_{\tau} ( {\vect{\theta}}_{\tau} )$ with ${\vect{\theta}}_{\tau} \in \Omega.$ Define $\vect{u}_{k:K} = \{{\vect{\theta}}_{\tau} \in \Omega, k \leq \tau \leq K \},$, with $\Omega$ being the compact~ (feasible ) set as a sequence of parameters with length $K-k$ and $V_k^{ (* )} = min_{\vect{u}_{k:K} } \sum_{\tau=k}^{K} \beta_{\tau} J_{\tau} ( {\vect{\theta}}_{\tau} ).$ Then, the following is true
\begin{equation}
\begin{aligned}
\Delta V^{ (* )}_{k} = - min_{{\vect{\theta}}_{k} \in \Omega} \big[ \beta_k J_{k} ( {\vect{\theta}}_{k} ) + \big ( \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}_{k} \rangle +\langle \nabla_{\vect{x}_{k}} V_{k}^{ (* )}, \Delta \vect{x}_{k} \rangle \big )\big],&
\end{aligned}
\label{eq_opt}
\end{equation}
where $\Delta V^{ (* )}_{k}$ represents the first difference due to the introduction of a task, $\Delta {\vect{\theta}}_{k} $ due to parameters and $\nabla_{\vect{x}_{k}}$ due to the task data with $\beta_k \in \mathbb{R} \cup [0,1], \forall k$ and $J_{k} ({\vect{\theta} }_{k} ) = \gamma_{k} \ell_{k} + \sum_{\tau = 0}^{k-1} \gamma_{\tau} \ell_{\tau}$.
\label{Der_HJB}
\end{prop}
\begin{proof}
Given $V_k^{ (* )} = min_{\vect{u}_{k:K} } \sum_{\tau=k}^{K} \beta_{\tau} J_{\tau} ( {\vect{\theta}}_{\tau} ),$ we split the interval $[k, K )$ as $[k, k+1 )$ and $[k+1, K )$ to write $$V^{ (* )}_{k}= min_{{\vect{\theta}}_{\tau} \in \Omega} \big[ \beta_k J_{k} ( {\vect{\theta}}_{k} )\big] + min_{\vect{u}_{k+1:K} }\big[ \sum_{\tau=k+1}^{K} \beta_{\tau} J_{\tau} ( {\vect{\theta}}_{\tau} ) \big].$$
$V_{k}= \sum_{\tau=k}^{K} \beta_{\tau} J_{\tau} ({\vect{\theta}}_{\tau} )$ provides $\sum_{\tau=k+1}^{K} \beta_{\tau} J_{\tau} ( {\vect{\theta}}_{\tau} )$ is $V_{k+1}$ therefore $min_{\vect{u}_{k+1:K} }\big[ \sum_{\tau=k+1}^{K} \beta_{\tau} J_{\tau} ( {\vect{\theta}}_{\tau} ) \big]$ is $V^{ (* )}_{k+1}.$ We then achieve
$$V^{ (* )}_{k} = min_{{\vect{\theta}}_{k} \in \Omega} \big[ \beta_k J_{k} ( {\vect{\theta}}_{k} ) + V^{ (* )}_{k+1} \big].$$
Since the minimization is with respect to $k$ now, the terms in $k+1$ can be pulled into of the bracket without any change to the minimization problem. We then approximate $V^{ (* )}_{k+1}$ using the information provided at $k.$ Since $V^{ (* )}_{k+1}$ is a function of $\vect{y}_{k},$ which is then a function of $ (k, \vect{x}_{k}, {\vect{\theta}}_{k} ),$ and all changes in $\vect{y}_{k}$ can be summarized through $ (k, \vect{x}_{k}, {\vect{\theta}}_{k} ).$ Therefore, a Taylor series of $V^{ (* )}_{k+1}$ around $ (k, \vect{x}_{k}, {\vect{\theta}}_{k} )$ provides
\begin{equation}
\begin{aligned}
&V^{ (* )}_{k+1} = V^{ (* )}_{k} + \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}_{k} \rangle &\\ &+ \langle \nabla_{\vect{x}_{k}} V_{k}^{ (* )}, \Delta \vect{x}_{k} \rangle + \langle \nabla_{k} (V^{ (* )}_{k} ), \Delta k \rangle + \cdots, &
\end{aligned}
\label{eq_a_1_5}
\end{equation}
where $\cdots$ summarizes all the higher order terms. As $k \in \mathbb{N}$ and $\langle \nabla_{k} (V^{ (* )}_{k} ), \Delta k \rangle$ represents the first difference in $V^{ (* )}_{k}$ hitherto denoted by $\Delta V^{ (* )}_{k}.$ We therefore achieve
\begin{equation}
\begin{aligned}
V^{ (* )}_{k+1} &= V^{ (* )}_{k} + \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )}, \Delta~\vect{\theta}_{k} \rangle \\
&+ \langle \nabla_{\vect{x}_{k}} V_{k}^{ (* )}, \Delta \vect{x}_{k} \rangle + \Delta V^{ (* )}_{k} + \cdots,
\end{aligned}
\label{eq_a_1_5}
\end{equation}
Substitute into the original equation to get
\begin{equation}
\begin{aligned}
& V^{ (* )}_{k} = min_{{\vect{\theta}}_{k} \in \Omega} \big[ \beta_k J_{k} ( {\vect{\theta}}_{k} ) \big] + \big ( V^{ (* )}_{k} + \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}_{k} \rangle & \\ &+ \langle \nabla_{\vect{x}_{k}} V_{k}^{ (* )}, \Delta \vect{x}_{k} \rangle +\Delta V^{ (* )}_{k} \big ) + \cdots,&
\end{aligned}
\label{eq_a_1_5}
\end{equation}
Cancel common terms and assume that the higher order terms~ ($\cdots$ ) are negligible to obtain
\begin{equation}
\begin{aligned}
\Delta V^{ (* )}_{k} = - min_{{\vect{\theta}}_{k} \in \Omega} \big[ \beta_k J_{k} ( {\vect{\theta}}_{k} ) + \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}_{k} \rangle +\langle \nabla_{\vect{x}_{k}} V_{k}^{ (* )}, \Delta \vect{x}_{k} \rangle \big].&
\end{aligned}
\label{eq_opt}
\end{equation}
which is a difference equation in $V^{ (* )}_{k}.$
\end{proof}
Note that $V^{ (* )}_{k}$ is the minima for the overarching CL problem and $\Delta V^{ (* )}_{k}$ represents the change in $V^{ (* )}_{k}$ upon introduction of a task~ (we hitherto refer to this as perturbations ). Zero perturbations $ (\Delta V^{ (* )}_{k}=0 )$ implies that the introduction of a new task does not impact our current solution; that is, the optimal solution on all previous tasks is optimal on the new task as well.
The solution of the CL problem can directly be obtained by solving Eq. \eqref{eq_opt} using all the available data. Thus, $ min_{{\vect{\theta}}_{k} \in \Omega} \big[ H (\Delta \vect{x}_{k}, \vect{\theta}_{k} ) \big] \quad \text{yields } \Delta V^{ (* )}_{k} \approx 0$ for $\beta > 0,$ with $H (\Delta \vect{x}_{k}, \vect{\theta}_{k} ) = \beta_k J_{k} ( {\vect{\theta}}_{k} ) + \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}_{k} \rangle + \langle \nabla_{\vect{x}_{k}} V_{k}^{ (* )}, \Delta \vect{x}_{k} \rangle.$ Essentially, minimizing $H (\Delta \vect{x}_{k}, \vect{\theta}_{k} )$ would minimize the perturbations introduced by any new task.
We simulate worst-case discrepancy by iteratively updating $\Delta \vect{x}_{k}$ through gradient ascent, thus maximizing generalization. Next, we minimize forgetting under maximum generalization by iteratively updating $\vect{\theta}_{k}$ through gradient descent. To formalize our idea, let us indicate the iteration index at $k$ by $i$ and write $\Delta \vect{x}_{k}$ as $\Delta \vect{x}^{ (i )}_{k}$ and ${\vect{\theta}}_{k}$ as ${\vect{\theta}}^{ (i )}_{k}$ with $H (\Delta \vect{x}_{k}, \vect{\theta}_{k} )$ as $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} )$~ (for simplicity of notation, we will denote $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} )$ as $H$ whenever convenient ). Towards these updates, we will first get an upper bound on $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} )$ and solve the upper bounding problem.
\begin{prop}
Let $k \in [0, K )$ and define $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} ) = \beta_k J_{k} ( {\vect{\theta}}^{ (i )}_{k} ) + \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}^{ (i )}_{k} \rangle + \langle \nabla_{\vect{x}^{ (i )}_{k}} V_{k}^{ (* )}, \Delta \vect{x}^{ (i )}_{k} \rangle$ assume that $ \nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )} \leq \nabla_{{\vect{\theta}}_{k}} J_{k} ({\vect{\theta}}^{ (i )}_{k} ).$ Then the following approximation is true:
\begin{equation}
\begin{aligned}
& H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} ) \leq \beta_k J_{k} (\vect{\theta}^{ (i )}_{k} ) + (J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} ) - J_{k} (\vect{\theta}^{ (i )}_{k} ) ) + ( J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} ) - J_{k} (\vect{\theta}^{ (i )}_{k} ) ),&
\end{aligned}
\label{eq_a_1_6}
\end{equation}
where $\beta_k \in \mathbb{R} \cup [0,1], \forall k$ and $\zeta \in \mathbb{N}$ and $J_{k+\zeta}$ indicates $\zeta$ updates on $\Delta \vect{x}^{ (i )}_{k}$ and $\vect{\theta}^{ (i+\zeta )}_{k}$ indicates $\zeta$ updates on ${\vect{\theta}}^{ (i )}_{k}.$
\label{prop2}
\end{prop}
\begin{proof}
Consider $ H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} ) = \beta_k J_{k} ( {\vect{\theta}}^{ (i )}_{k} ) + \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}^{ (i )}_{k} \rangle + \langle \nabla_{\vect{x}^{ (i )}_{k}} V_{k}^{ (* )}, \Delta \vect{x}^{ (i )}_{k} \rangle.$ Assuming $\nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )} \leq \nabla_{{\vect{\theta}}_{k}} J_{k} ({\vect{\theta}}^{ (i )}_{k} )$ we may write through finite difference approximation as
\begin{equation}
\begin{aligned}
\langle \nabla_{{\vect{\theta}}^{ (i )}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}^{ (i )}_{k} \rangle &\leq& \langle \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} ({\vect{\theta}}^{ (i )}_{k} ) , \Delta {\vect{\theta}}^{ (i )}_{k} \rangle,\\
&\leq& (J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} ) - J_{k} (\vect{\theta}^{ (i )}_{k} ) )
\end{aligned}
\label{eq_a_1_6}
\end{equation}
Similarly, we may write
\begin{equation}
\begin{aligned}
\langle \nabla_{\vect{x}_{k}} V_{k}^{ (* )}, \Delta \vect{x}_{k} \rangle &\leq&
\langle \nabla_{\vect{x}_{k}} J_{k} ({\vect{\theta}}^{ (i )}_{k} ), \Delta \vect{x}_{k} \rangle,\\
&\leq& ( J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} ) - J_{k} (\vect{\theta}^{ (i )}_{k} ) ).
\end{aligned}
\label{eq_a_1_6}
\end{equation}
Upon substitution, we have our result:
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} ) \leq \beta_k J_{k} (\vect{\theta}^{ (i )}_{k} ) + (J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} ) - J_{k} (\vect{\theta}^{ (i )}_{k} ) ) + (J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} ) - J_{k} (\vect{\theta}^{ (i )}_{k} ) ).
\end{aligned}
\label{eq_a_1_6}
\end{equation}
\end{proof}
Our cost to be analyzed will be given as
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} ) = \beta_k J_{k} ( {\vect{\theta}}^{ (i )}_{k} ) + \langle \nabla_{{\vect{\theta}}^{ (i )}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}^{ (i )}_{k} \rangle + \langle \nabla_{\vect{x}^{ (i )}_{k}} V_{k}^{ (* )}, \Delta \vect{x}^{ (i )}_{k} \rangle.
\end{aligned}
\label{eq_approx}
\end{equation}
and use this definition of $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} )$ from here on.
\section{Main results}
\begin{figure}
\includegraphics[width = \columnwidth]{Figures/Proof_Il.png}
\caption{Illustration of proofs. $\Delta \vect{x}$~ (player~1 ) is the horizontal axis and the vertical axis indicates $\vect{\theta}$~ (player~2 ) where the curve indicates H. If we start from red circle for player~1~ (player~2 is fixed at the blue circle ) H is increasing~ (goes from a grey circle to a red asterisk ) with player 1 reaching the red star. Next, start from the blue circle~ ($\vect{\theta}$ is at the red star ), the cost decreases.}
\label{fig:proof}
\end{figure}
We will define two compact sets $\Omega_{\theta}, \Omega_{x}$ and seek to show existence and stability of a saddle point~$ (\Delta \vect{x}^{ (i )}_{k} , \vect{\theta}_k^{ (i )} )$ for a fixed $k.$ To illustrate the theory, we refer to Fig.~\ref{fig:proof}, for each $k,$ the initial values for the two players are characterized by the pair $\{\vect{\theta}^{ (i )}_{k} \text{ (blue circle )}, \Delta \vect{x}^{ (i )}_{k} \text{ (red circle )} \},$ and $H ( \Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} )$ is indicated by the grey circle on the cost curve~ (the dark blue curve ). Our proofing strategy is as follows.
First, we fix $\vect{\theta}_k^{ (. )} \in \Omega_{\theta}$ and construct $\mathcal{M}_k= \{\vect{\theta}^{ (. )}_{k}, \Omega_{x} \},$ to prove that H is maximizing with respect to $\Delta \vect{x}^{ (i )}_{k}.$
\begin{lemma}
For each $k \in [0, K ),$ fix $\vect{\theta}_k^{ (. )} \in \Omega_{\theta}$ and construct $\mathcal{M}_k = \{\Omega_{x}, \vect{\theta}^{ (. )}_{k}\}$ with $\Omega_{\theta}, \Omega_{x}$ being the sets of all feasible $\vect{\theta}^{ (. )}_{k}$ and $\vect{x}^{ (i )}_{k}$ respectively. Define $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )$ as in Eq. \eqref{eq_approx} for $ (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) \in \mathcal{M}_k $ and consider $$ \Delta \vect{x}^{ (i+1 )}_{k} - \Delta \vect{x}^{ (i )}_{k} = \alpha_{k}^{ (i )} \nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) ) / \| \nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) \|^2.$$ Consider the assumptions $\nabla_{ \vect{x}^{ (i )}_{k}} V_{k}^{ (* )} \leq \nabla_{ \vect{x}^{ (i )}_{k}} J_{k}$ and $\langle \nabla_{\vect{x}^{ (i )}_{k}} J_k, \nabla_{\vect{x}^{ (i )}_{k}} J_k \rangle>0$, and let $\alpha_{k}^{ (i )} \rightarrow 0, i \rightarrow \infty.$ It follows that $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )$ converges asymptotically to a maximizer.
\label{lem:lem_max}
\end{lemma}
\begin{proof}
Fix $\vect{\theta}^{ (. )}_{k} \in \Omega_{\theta}$ and construct $\mathcal{M}_k$ such that $\mathcal{M}_k = \{\vect{\theta}^{ (. )}_{k}, \Omega_{x} \}$ which we call a neighborhood. Therefore, for $ (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (. )}_{k} ), (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) \in \mathcal{M}_k$ we may write a first-order Taylor series expansion of $H (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (. )}_{k} )$ around $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )$ as
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (. )}_{k} ) = H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) + \langle \nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ), \Delta \vect{x}^{ (i+1 )}_{k} - \Delta \vect{x}^{ (i )}_{k} \rangle .
\end{aligned}
\end{equation}
We substitute the update as $\alpha_{k}^{ (i )} \frac{\nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )}{\| \nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) \|^2}$ to get
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (. )}_{k} ) - H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )= \langle \nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ), \alpha_{k}^{ (i )} \frac{\nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )}{\| \nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )\|^2} \rangle.
\end{aligned}
\end{equation}
The derivative $\nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )$ can be written as
\begin{equation}
\begin{aligned}
& \nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) ) \leq \nabla_{\Delta \vect{x}^{ (i )}_{k}} \Big[ \beta_k J_{k} ( {\vect{\theta}}^{ (. )}_{k} ) + \langle \nabla_{{\vect{\theta}}^{ (. )}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}^{ (. )}_{k} \rangle & \\ &+ \langle \nabla_{\vect{x}^{ (i )}_{k}} V_{k}^{ (* )}, \Delta \vect{x}^{ (i )}_{k} \rangle \Big] = \nabla_{ \vect{x}^{ (i )}_{k}} V_{k}^{ (* )} \leq \nabla_{ \vect{x}^{ (i )}_{k}} J_{k}.&
\end{aligned}
\end{equation}
Substitution reveals
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (. )}_{k} ) - H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) = \alpha_{k}^{ (i )} \frac{ \langle \nabla_{\vect{x}^{ (i )}_{k}} J_{k}, \nabla_{\vect{x}^{ (i )}_{k}} J_{k} \rangle
}{\|\nabla_{\vect{x}^{ (i )}_{k}} J_{k} \|^2}
\end{aligned}
\end{equation}
for $\alpha_{k}^{ (i )} >0$; and under the assumption that $ \langle \nabla_{\vect{x}^{ (i )}_{k}} J_{k}, \nabla_{ \vect{x}^{ (i )}_{k}} J_{k} \rangle >0$ we obtain
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (. )}_{k} ) - H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) = \alpha_{k}^{ (i )} \frac{ \langle \nabla_{\vect{x}^{ (i )}_{k}} J_{k}, \nabla_{\vect{x}^{ (i )}_{k}} J_{k}
\rangle }{\| \nabla_{\vect{x}^{ (i )}_{k}} J_{k} \|^2 } \geq 0.
\end{aligned}
\end{equation}
Let $ B_{x} = \alpha_{k}^{ (i )} \frac{ \langle \nabla_{\vect{x}_{k}} J_{k}, \nabla_{ \vect{x}_{k}} J_{k} \rangle }{\| \nabla_{ \vect{x}^{ (i )}_{k}} J_{k} \|^2 } \leq \alpha_{k}^{ (i )}$ and therefore $0 \leq H (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (. )}_{k} ) - H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )\leq \alpha^{ (i )}_{k}.$ We therefore have $H (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (. )}_{k} ) - H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) \geq 0$ and $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )$ is maximizing with respect to $\Delta \vect{x}^{ (i )}_{k}.$ Furthermore, under the assumption that $\alpha^{ (i )}_{k} \rightarrow 0, k \rightarrow \infty$, we have $H (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (. )}_{k} ) - H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) \rightarrow 0$ asymptotically and we have our result.
\end{proof}
Similarly, we fix $\Delta \vect{x}^{ (. )}_{k} \in \Omega_x$ and construct $\mathcal{N}_k= \{ \Omega_{\theta}, \Delta \vect{x}^{ (. )}_{k} \},$ to prove that $H$ is minimizing with respect to $\vect{\theta}^{ (i )}_{k}.$
\begin{lemma}
For each $k \in [0, K ),$ fix $\Delta \vect{x}^{ (. )}_{k} \in \Omega_{x}$ and construct $\mathcal{N}_k= \{\Delta \vect{x}^{ (. )}_{k}, \Omega_{\theta} \}.$ Then for any $ (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) \in \mathcal{N}_k$ define $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )$ as in Eq. \eqref{eq_approx} with Proposition.~\ref{prop2} being true and let $\vect{\theta}^{ (i+1 )}_{k} - \vect{\theta}^{ (i )}_{k} = -\alpha_{k}^{ (i )} \nabla_{{\vect{\theta}}^{ (i )}_{k}} H (\Delta \vect{x}^{ (. )}_{k}, \vect{\theta}^{ (i )}_{k} ) )$. Assume that $\| \nabla_{\vect{\theta}^{ (i )}_{k}} J_{k} ( {\vect{\theta}}^{ (i+\zeta )}_{k} )\| \leq L_1$ and $\| \nabla_{\vect{\theta}^{ (i )}_{k}} J_{k+\zeta} ( {\vect{\theta}}^{ (i )}_{k} )\| \leq L_2$ and let $\alpha_{k}^{ (i )} \rightarrow 0, i \rightarrow \infty.$ Then $\vect{\theta}^{ (i )}_{k}$ converges to a local minimizer.
\label{lem:lem_min}
\end{lemma}
\begin{proof}
First, we fix $\Delta \vect{x}^{ (. )}_k \in \Omega_{x}$ and construct $\mathcal{N}_k= \{ \Omega_{\theta} , \Delta \vect{x}^{ (. )}_{k} \}.$ For any $ (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ), (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) \in\mathcal{N}_k$ we write a first-order Taylor series expansion of $H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} )$ around $H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} )$ to write
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) = H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ) + \langle \nabla_{\vect{\theta}^{ (i )}_{k}} H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ), \vect{\theta}^{ (i+1 )}_{k} - \vect{\theta}^{ (i )}_{k} \rangle .
\end{aligned}
\end{equation}
We then substitute
$\vect{\theta}^{ (i+1 )}_{k} - \vect{\theta}^{ (i )}_{k} = -\alpha_{k}^{ (i )} \nabla_{{\vect{\theta}}^{ (i )}_{k}} H (\Delta \vect{x}^{ (. )}_{k}, \vect{\theta}^{ (i )}_{k} ) $ to get
\begin{equation}
\begin{aligned}
&H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ) = -\alpha_{k}^{ (i )}\langle \nabla_{\vect{\theta}^{ (i )}_{k}} H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ), \nabla_{{\vect{\theta}}^{ (i )}_{k}} H (\Delta \vect{x}^{ (. )}_{k}, \vect{\theta}^{ (i )}_{k} )\rangle.&
\end{aligned}
\end{equation}
Following Proposition 2, the derivative $\nabla_{{\vect{\theta}}^{ (i )}_{k}} H (\Delta \vect{x}^{ (. )}_{k}, \vect{\theta}^{ (i )}_{k} )$ can be written as
\begin{equation}
\begin{aligned}
\nabla_{{\vect{\theta}}^{ (i )}_{k}} H (\Delta \vect{x}^{ (. )}_{k}, \vect{\theta}^{ (i )}_{k} )\leq \nabla_{\vect{\theta}_{k}} [\beta_k J_{k} ( {\vect{\theta}}^{ (i )}_{k} ) + (J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} ) - J_{k} (\vect{\theta}^{ (i )}_{k} ) ) + (J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} ) - J_{k} (\vect{\theta}^{ (i )}_{k} ) )]
\end{aligned}
\end{equation}
Simplification reveals
\begin{equation}
\begin{aligned}
\nabla_{{\vect{\theta}}^{ (i )}_{k}} H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ) )\leq \nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 ) J_{k} ( {\vect{\theta}}^{ (i )}_{k} ) + \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} ) + \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} ).
\end{aligned}
\end{equation}
Substitution therefore provides
\begin{align}
&H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ) \nonumber\\ &\leq
- \alpha_{k}^{ (i )} \langle \nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 ) J_{k} ( {\vect{\theta}}^{ (i )}_{k} ) + \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} ) + \nabla_{\vect{\theta}^{ (i )}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} ), \nonumber
\\ & \nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 ) J_{k} ( {\vect{\theta}}^{ (i )}_{k} ) + \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} ) + \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} ) \rangle .
\end{align}
Opening the square with Cauchy's inequality provides
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} )
&\leq& - \alpha_{k}^{ (i )} \Big[ \|\nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 ) J_{k} ( {\vect{\theta}}_{k} ) \|^2 \\
&+& \| \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} )\|^2
+ \| \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} )\|^2
\\
&+& 2\|\nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 ) J_{k} ( {\vect{\theta}}_{k} ) \| \| \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} )\| \\
&+& 2\| \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} )\|\| \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} )\|
\\
&+& 2\|\nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 ) J_{k} ( {\vect{\theta}}_{k} ) \| \| \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} )\| \Big].
\end{aligned}
\end{equation}
We simplify with Young's inequality to achieve
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} )
&\leq& - \alpha_{k}^{ (i )} \Big[ \|\nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 )J_{k} ( {\vect{\theta}}_{k} ) \|^2 \\
&+& \| \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} )\|^2
\\ &+& \| \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} )\|^2
\\ &+& \|\nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 ) J_{k} ( {\vect{\theta}}_{k} ) \|^2 \\ &+& \| \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} )\|^2 \\
&+& \| \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} )\|^2
\\ &+& \| \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} )\|^2
\\ &+& \|\nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 ) J_{k} ( {\vect{\theta}}_{k} ) \|^2
\\ &+& \| \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} )\|^2 \Big].
\end{aligned}
\end{equation}
Further simplification results in
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} )
&\leq& - \alpha_{k}^{ (i )} \Big[ 3 \|\nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 ) J_{k} ( {\vect{\theta}}_{k} ) \|^2 \\
&+& 3 \| \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} )\|^2
+ 3 \| \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} )\|^2 \Big].
\end{aligned}
\end{equation}
With the assumption that $\| \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} )\| \leq L_1$ and $\| \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} )\|\leq L_2,$, we may write
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ) &\leq& -\alpha_{k}^{ (i )} B_{\theta},
\end{aligned}
\end{equation}
where $B_{\theta} = \Bigg[ ( (\sqrt{3}\beta_k-2\sqrt{3} )^2 + 3 )L_1^2 + 3L_2^2\Bigg].$ Assuming that $\alpha_{k}^{ (i )}$ is chosen such that $\alpha^{ (i )}_{k} \rightarrow 0$, we obtain $H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ) \rightarrow 0$ as $i \rightarrow \infty$ and $H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ) < 0.$ Therefore H converges to a local minimizer.
\end{proof}
From here on, we will define our cost function as H whereever convinient for simplicity of notations. Since, for any k, there exists a local maximizer $\Delta \vect{x}_k^{ (* )} \in \Omega_x,$ we may define $\mathcal{N}^{ (* )}_k = \{\Delta \vect{x}_k^{ (* )}, \Omega_\theta\}$ where the set $\Omega_{x}$ is comprised of a local maximizer $\Delta \vect{x}_k^{ (* )}$
\begin{lemma}
For any $k \in [0, K )$, let $\vect{\theta}^{ (* )}_{k} \in \Omega_\theta,$ be the minimizer of $H$ according to Lemma \ref{lem:lem_min} and define $\mathcal{M}^{ (* )}_k = \{\Omega_{x}, \vect{\theta}^{ (* )}_{k}\}.$ Then for $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ), (\Delta \vect{x}_k^{ (i )}, \vect{\theta}^{ (* )}_k ) \in \mathcal{M}^{ (* )}_k,$ $H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ) \geq H (\Delta \vect{x}_k^{ (i )}, \vect{\theta}^{ (* )}_k ),$ where $\Delta \vect{x}_k^{ (* )}$ is a maximizer for H according to Lemma. \ref{lem:lem_max}.
\label{lem:lem_max_opt}
\end{lemma}
\begin{proof}
By Lemma \ref{lem:lem_min}, for each $k \in [0, K ),$ there exists a minimizer $\vect{\theta}^{ (* )}_{k} \in \Omega_\theta$ such that $\mathcal{M}^{ (* )}_k = \{ \Omega_{x}, \vect{\theta}^{ (* )}_{k} \}.$ Therefore by Lemma \ref{lem:lem_max}, $H (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (* )}_{k} ) - H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (* )}_{k} ) \geq 0$ for $ (\Delta \vect{x}_k^{ (i+1 )}, \vect{\theta}^{ (* )}_k ), (\Delta \vect{x}_k^{ (i )}, \vect{\theta}^{ (* )}_k ) \in \mathcal{M}^{ (* )}_k.$ Let $\Delta \vect{x}_k^{ (* )} \in \Omega_x $ be the converging point according to Lemma \ref{lem:lem_max}. Then, for $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ), (\Delta \vect{x}_k^{ (i )}, \vect{\theta}^{ (* )}_k ) \in \mathcal{M}^{ (* )}_k$ a $H (\Delta \vect{x}^{ (* )}_{k}, \vect{\theta}^{ (* )}_{k} ) - H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (* )}_{k} ) \geq 0.$ by Lemma~\ref{lem:lem_max} which provides the result.
\end{proof}
\begin{lemma}
For any $k \in [0, K )$, let $\Delta \vect{x}_k^{ (* )} \in \Omega_x,$ be the maximizer of $H$ according to Lemma \ref{lem:lem_max} and define $\mathcal{N}^{ (* )}_k = \{\Delta \vect{x}^{ (* )}_{k}, \Omega_{\theta}\}.$ Then for $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ), (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ) \in \mathcal{N}^{ (* )}_k,$ $H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ) \leq H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ),$ where $\vect{\theta}^{ (* )}_{k}$ is a minimizer for H according to Lemma. \ref{lem:lem_max}.
\label{lem:lem_min_opt}
\end{lemma}
\begin{proof}
By Lemma \ref{lem:lem_max}, for each $k \in [0, K ),$ there exists a maximizer $\Delta \vect{x}_k^{ (* )} \in \Omega_x,$ such that $\mathcal{N}^{ (* )}_k = \{\Delta \vect{x}^{ (* )}_{k}, \Omega_{\theta}\}.$ Therefore by Lemma \ref{lem:lem_min}, $H (\Delta \vect{x}^{ (* )}_{k}, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (* )}_{k}, \vect{\theta}^{ (i )}_{k} ) \leq 0$ for $ (\Delta \vect{x}^{ (* )}_{k}, \vect{\theta}^{ (i+1 )}_{k} ), (\Delta \vect{x}^{ (* )}_{k}, \vect{\theta}^{ (i )}_{k} ) \in \mathcal{N}^{ (* )}_k.$ Let $\vect{\theta}_k^{ (* )} \in \Omega_{\theta} $ be the converging point according to Lemma \ref{lem:lem_min}. Then, for $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ), (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ) \in \mathcal{M}^{ (* )}_k,$ $H (\Delta \vect{x}^{ (* )}_{k}, \vect{\theta}^{ (* )}_{k} ) - H (\Delta \vect{x}^{ (* )}_{k}, \vect{\theta}^{ (i )}_{k} ) \leq 0$ by Lemma~\ref{lem:lem_min} which provides the result.
\end{proof}
Next, we prove that the union of the two neighborhoods for each k $\mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k,$ is non-empty.
\begin{lemma}
For any $k \in [0, K )$, let $\vect{\theta}^{ (* )}_{k} \in \Omega_\theta,$ be the minimizer of $H$ according to Lemma \ref{lem:lem_min} and define $\mathcal{M}^{ (* )}_k = \{\Omega_{x}, \vect{\theta}^{ (* )}_{k}\}.$ Similarly, let $\Delta \vect{x}_k^{ (* )} \in \Omega_x,$ be the maximizer of $H$ according to Lemma \ref{lem:lem_max} and define $\mathcal{N}^{ (* )}_k = \{\Delta \vect{x}^{ (* )}_{k}, \Omega_{\theta}\}.$ Then, $\mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$ is nonempty.
\label{lem:lem_nonE}
\end{lemma}
\begin{proof}
Let $\mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$ be empty. Then, for any $ (\Delta \vect{x}^{ (i+1 )}_k, \vect{\theta}^{ (. )}_{k} ), (\Delta \vect{x}^{ (i )}_k,\vect{\theta}^{ (. )}_{k} ) \in \mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$, $H (\Delta \vect{x}^{ (i+1 )}_k, \vect{\theta}^{ (. )}_{k} ) - H (\Delta \vect{x}^{ (i )}_k,\vect{\theta}^{ (. )}_{k} )$ is undefined because the union is empty. This contradicts Lemma \ref{lem:lem_min_opt}. Similarly, $H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (. )}_k,\vect{\theta}^{ (i )}_{k} )$ for $ (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ), (\Delta \vect{x}^{ (. )}_k,\vect{\theta}^{ (i )}_{k} ) \in \mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$ also contradicts Lemma \ref{lem:lem_max_opt}. Therefore, by contradiction, $\mathcal{M}_k \cup \mathcal{N}_k$ cannot be empty.
\end{proof}
\subsection{Final Results}
We are now ready to present the main results. We show that there exists an equilibrium point~ (Theorem~1 ) and that the equilibrium point is stable~ (Theorem~2 ).
\begin{thm}[Existence of an Equilibrium Point]
For any $k \in [0, K )$, let $\vect{\theta}^{ (* )}_{k} \in \Omega_\theta,$ be the minimizer of $H$ according to Lemma \ref{lem:lem_min_opt} and define $\mathcal{M}^{ (* )}_k = \{\Omega_{x}, \vect{\theta}^{ (* )}_{k}\}.$ Similarly, let $\Delta \vect{x}_k^{ (* )} \in \Omega_x,$ be the maximizer of $H$ according to Lemma \ref{lem:lem_max_opt} and define $\mathcal{N}^{ (* )}_k = \{\Delta \vect{x}^{ (* )}_{k}, \Omega_{\theta}\}.$ Further, let $\mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$ be nonempty according to Lemma.~\ref{lem:lem_nonE}, then $ (\Delta \vect{x}^{ (* )}_{k},\vect{\theta}^{ (* )}_{k} ) \in \mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$ is a local equilibrium point.
\end{thm}
\begin{proof}
By Lemma \ref{lem:lem_min_opt} we have at $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ), (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ) \in \mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$ that
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ) &\leq& H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ).
\end{aligned}
\end{equation}
Similarly, according to Lemma \ref{lem:lem_max_opt}, at $ (\Delta \vect{x}^{ (* )}_k, \vect{\theta}^{ (* )}_{k} ), (\Delta \vect{x}^{ (i )}_k, \vect{\theta}^{ (* )}_{k} ) \in \mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$ we have
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (* )}_k, \vect{\theta}^{ (* )}_{k} ) \geq H (\Delta \vect{x}^{ (i )}_k, \vect{\theta}^{ (* )}_{k} ).
\end{aligned}
\end{equation}
Putting these inequalities together, we get
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ) \geq H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ) \geq H (\Delta \vect{x}_k^{ (i )}, \vect{\theta}^{ (* )}_k ),
\end{aligned}
\end{equation}
which is the saddle point condition, and therefore $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{(*)}_k )$ is a local equilibrium point in $\mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k.$
\end{proof}
According to the preceeding theorem, there is at least one equillibrium point for the game summarized by $H$.
\begin{thm} [Stability of the Equilibrium Point]
For any $k \in [0,K)$, $\Delta \vect{x}^{ (i )}_{k} \in \Omega_x$ and $\vect{\theta}^{ (i )}_{k} \in \Omega_\theta$ be the initial values for $\Delta \vect{x}^{ (i )}_{k}$ and $\vect{\theta}^{ (i )}_{k}$ respectively. Define $\mathcal{M}_k = \{\Omega_{x}, \Omega_{\theta}\}$ with $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} )$ given by Proposition \ref{prop2}. Let $\Delta \vect{x}^{ (i+1 )}_{k} - \Delta \vect{x}^{ (i )}_{k} = \alpha_{k}^{ (i )}\times (\nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) )/\| \nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) \|^2 )$ and $\vect{\theta}^{ (i+1 )}_{k} - \vect{\theta}^{ (i )}_{k} = -\alpha_{k}^{ (i )}\times \nabla_{{\vect{\theta}}^{ (i )}_{k}} H (\Delta \vect{x}^{ (. )}_{k}, \vect{\theta}^{ (i )}_{k} ).$ Let the existence of an equilibrium point be given by Theorem~1, then as a consequence of Lemma \ref{lem:lem_max} and \ref{lem:lem_min} $ (\Delta \vect{x}^{ (* )}_{k},\vect{\theta}^{ (* )}_{k} ) \in \mathcal{M}_k$ is a stable equilibrium point for $H$.
\label{thm:thm_st}
\end{thm}
\begin{proof}
Consider now the order of plays by the two players. By Lemma~\ref{lem:lem_max}, a game starting at $ (\Delta \vect{x}_k^{ (i )}, \vect{\theta}^{ (i )}_k ) \in \mathcal{M}_k$ will reach $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k )$ which is a maximizer for $H.$ Now, define $\mathcal{N}_k = (\Delta \vect{x}_k^{ (* )}, \Omega_{\theta} ) \subset \mathcal{M}_k$ then a game starting at $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ) \in \mathcal{N}_k$ will converge to $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ) \in \mathcal{N}_k$ according to Lemma~\ref{lem:lem_min}. Since, $\mathcal{N}_k \subset \mathcal{M}_k,$ our result follows.
\end{proof}
\section{Experimental Details}
Much of this information is a repetition of details provided in \cite{Hsu18_EvalCL,vandeven2019generative}.
\begin{enumerate}
\item \textit{Incremental Domain Learning~ (IDL ):} Incremental domain refers to the scenario when each new task introduces changes in the marginal distribution of the inputs. This scenario has been extensively discussed in the domain adaptation literature, where this shift in domain is typically referred to as ``non-stationary data distribution" or domain shift. Overall, we aim to transfer knowledge from the old task to a new task where each task can be different in the sense of their marginal distribution.
\item \textit{Incremental Class Learning~ (ICL ):} In this scenario, each task contains an exclusive subset of classes. The number of output nodes in a model equals the number of total classes in the task sequence. For instance, tasks could be constructed by using exactly one class from the MNIST data set where we aim to transfer knowledge from one class to another.
\item \textit{Incremental Task Learning~ (ITL ):} In this setup, the output spaces are disjoint between tasks[ for example, the previous task can be a classification problem of five classes, while the new task can be a regression. This scenario is the most generic and allows for the tasks to be defined arbitrarily. For each tasks, a model requires task-specific identifier$t$.
\end{enumerate}
\textbf{Split-MNIST}
For split-MNIST, the original MNIST-data set is split into five partitions where each partition is a two-way classification. We pick 60000 images for training (6000 per digit ) and 10000 images for test, i.e. (1000 per digit ). For the incremental task learning in the split-MNIST experiment, the ten digits are split into five two-class classification tasks (the model has five output heads, one for each task ) and the task identity (1 to 5 ) is given for test. For the incremental class learning setup, we require the model to make a prediction over all classes (digits 0 to 9 ). For the incremental domain learning, the model always predicts over two classes.
\textbf{Permuted-MNIST}
For permuted-MNIST, we permute the pixels in the MNIST data to create tasks where each task is a ten-way classification. The three CL scenarios that are generated for the permuted-MNIST are similar to the Split-MNIST data set except for the idea that the different tasks are now generated by applying random pixel permutations to the images. For incremental task learning, we use a multi-output strategy, and each task is attached to a task identifier. For incremental domain and class, we use a single output strategy and each task as one where one of the 10 digits are predicted. In incremental class learning, for each new task 10 new classes are generated by permuting the MNIST data set. For incremental task and domain, we use a total of 10 tasks whereas for incremental classes, we generate a total of 100 tasks.
\textbf{Network Architecture}
We keep our architecture identical to what is provided in \cite{Hsu18_EvalCL,vandeven2019generative}. The loss function is categorical cross-entropy for classification. All models were trained for $2$ epochs per task with a minibatch size of $128$ using the Adam optimizer ($\beta_1 = 0.9$, $\beta_2 = 0.999$, learning rate$= 0.001$ ) as the default. For BCL, the size of the buffer (i.e., a new task array ) $\vect{\mathcal{D}} _N (k )$ and a task memory array (samples from all the previous tasks ) $\vect{\mathcal{D}} _{P} (k ) )$ is kept equivalent to naive rehearsal and other memory-driven approaches such as GEM and MER~ ($16,000$ samples ).
\textbf{Comparison Methods -- Baseline Strategies}
Additional details can be found from \cite{Hsu18_EvalCL,vandeven2019generative}
\begin{enumerate}
\item A sequentially-trained neural network with different optimizers such as SGD, Adam~\cite{kingma2014adam}, and Adagrad~\cite{duchi2011adaptive}.
\item A standard $L_2-$regularized neural network where each task is shown sequentially.
\item Naive rehearsal strategy (experience replay ) where a replay buffer is initialized and data corresponding to all the previous tasks are stored. The buffer size is chosen to match the space overhead of online EWC and SI.
\end{enumerate}
\textbf{Comparison Method-CL}
We compared the following CL methods:
\begin{enumerate}
\item \textbf{EWC~\cite{kirkpatrick2017overcoming} / Online EWC~\cite{schwarz2018progress} / SI~\cite{zenke2017continual}}: For these methods, a regularization term is added to the loss, with a hyperparameter used to control the regularization strength such that: $L (total ) = L (current ) + \lambda L (regularization ).$ $\lambda$ is set through a hyperparameter.
\item \textbf{LwF~\cite{LwF} / DGR~\cite{DGR}} Here, we set the loss to be $L (total ) = \alpha L (current ) + (1 − \alpha )L (replay )$ where hyperparameter $\alpha$ is chosen according to how many tasks have been seen by the model.
\item For \textbf{RtF}~\cite{vandeven2019generative}, MAS~\cite{aljundi2018memory}, GEM\cite{lopez2017gradient} and MER\cite{riemer2018learning}, we refer to the respective publication for details.
\end{enumerate}
Additional details about the experiments can be found in \cite{Hsu18_EvalCL} as our paper retains their hyper-parameters and the experimental settings.
\bibliographystyle{plain}
\section{Introduction}
In continual learning~(CL), we incrementally adapt a model to learn tasks~(defined according to the problem at hand) observed sequentially. CL has two main objectives: maintain long-term memory~(remember previous tasks) and navigate new experiences continually~(quickly adapt to new tasks). An important characterization of these objectives is provided by the stability-plasticity dilemma~\cite{carpenter1987massively}, where the primary challenge is to balance network stability~(preserve past knowledge; minimize catastrophic forgetting) and plasticity~(rapidly learn from new experiences; generalize quickly). This balance provides a natural objective for CL: \textit{balance forgetting and generalization.}
Traditional CL methods either minimize catastrophic forgetting or improve quick generalization but do not model both. For example, common solutions to the catastrophic forgetting issue include (1) representation-driven approaches~\cite{yoon2017lifelong, ke2020continual}, (2) regularization approaches~\cite{kirkpatrick2017overcoming, aljundi2018memory, mirzadeh2020understanding, farajtabar2019orthogonal, yin2020optimization, yin2020optimization, jung2020continual, pan2020continual, chaudhry2020continual, titsias2019functional}, and (3) memory/experience replay~\cite{lin1992self, lopez2017gradient, chaudhry2019continual, chaudhry2019tiny, fini2020online}. Solutions to the generalization problem include representation-learning approaches~(matching nets~\cite{vinyals2016matching}, prototypical networks~\cite{snell2017prototypical}, and metalearning approaches~\cite{finn2017model, finn2019online, Caccia2020OnlineFA, yao2020don}). More recently, several approaches \cite{nagabandi2019deep,farajtabar2019orthogonal,DBLP:journals/corr/abs-1806-06928, joseph2020metaconsolidation, yin2020optimization, ebrahimi2020adversarial} have been introduced that combine methods designed for quick generalization with frameworks designed to minimize forgetting.
The aforementioned CL approaches naively minimize a loss function~(combination of forgetting and generalization loss) but do not explicitly account for the trade-off in their optimization setup. The first work to formalize this trade-off was presented in meta-experience replay~(MER)~\cite{riemer2018learning}, where the forgetting-generalization trade-off was posed as a gradient alignment problem. Although MER provides a promising methodology for CL, the balance between forgetting and generalization is enforced with several hyperparameters.
Therefore, two key challenges arise: (1)~lack of theoretical tools that study the existence~(\textit{under what conditions does a balance point between generalization and forgetting exists?}) and stability~(\textit{can this balance be realistically achieved?}) of a balance point and (2)~lack of a systematic approach to achieve the balance point. We address these challenges in this paper.
We describe a framework where we first formulate CL as a sequential decision-making problem and seek to minimize a cost function summed over the complete lifetime of the model. At any time $k,$ given that the future tasks are not available, the calculation of the cost function becomes intractable. To circumvent this issue, we use Bellman's principle of optimality~\cite{bellman2015adaptive} and recast the CL problem to model the catastrophic forgetting cost on the previous tasks and generalization cost on the new task. We show that equivalent performance on an infinite number of tasks is not practical~(Lemma~1 and Corollary~1) and that tasks must be prioritized.
\begin{figure}[!h]
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width = \columnwidth]{Figures/MCL_Discrete.png}
\label{fig:ills}
\end{subfigure}\\
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width = \columnwidth ]{Figures/Method.png}
\label{fig:met}
\end{subfigure}
\caption{(left) Exemplary CL problem: the lifetime of the model can be split into three intervals. At $k =1$ we seek to recognize lions; at $k=2$ we seek to recognize both lions and cats; and at $k=3$ we seek to recognize cats, lions, and dogs. (right) Illustration of the proposed method: our methodology comprises an interplay between two players. The first player maximizes generalization by simulating maximum discrepancy between two tasks. The second player minimizes forgetting by adapting to maximum discrepancy.}
\label{fig:ills}
\end{figure}
To achieve a balance between forgetting and generalization, we pose the trade-off
as a saddle point problem where we designate one player for maximizing the generalization cost~(player 1) and another for minimizing the forgetting cost~(player 2). We prove mathematically that there exists at least one saddle point between generalization and forgetting for each new task~(Theorem~1). Furthermore, we show that this saddle point can be attained asymptotically~(Theorem~2) when player strategies are chosen as gradient ascent-descent. We then introduce balanced continual learning~(BCL), a new algorithm to achieve this saddle point. In our algorithm~(see Fig.~\ref{fig:ills} for a description of BCL), the generalization cost is computed by training and evaluating the model on given new task data. The catastrophic forgetting cost is computed by evaluating the model on the task memory (previous tasks). We first maximize the generalization cost and then minimize the catastrophic forgetting cost to achieve the balance. We compare our approach with other methods such as elastic weight consolidation~(EWC)~\cite{kirkpatrick2017overcoming}, online EWC~\cite{schwarz2018progress}, and MER~\cite{riemer2018learning} on continual learning benchmark data sets~\cite{Hsu18_EvalCL} to show that BCL is better than or comparable to the state-of-the-art methods. Moreover, we also show in simulation that our theoretical framework is appropriate for understanding the continual learning problem. The contributions of this paper are (1) a theoretical framework to study the CL problem, (2) BCL, a method to attain balance between forgetting and generalization, and (3) advancement of the state of the art in CL.
\section{Problem Formulation}
We use $\vect{\mathbb{R}} $ to denote the set of real numbers and $\vect{\mathbb{N}} $ to denote the set of natural numbers. We use $\|.\|$ to denote the Euclidean norm for vectors and the Frobenius norm for matrices, while using bold symbols to illustrate matrices and vectors. We define an interval $[0, K), K \in \vect{\mathbb{N}} $ and let $p(\cT)$ be the distribution over all the tasks observed in this interval. For any $k \in [0, K) ,$ we define a parametric model $g(.)$ with ${\vect{y}}_{k} = g(\vect{x}_{k}; {\vect{\theta}}_{k})$, where ${\vect{\theta}}_{k}$ is a vector comprising all parameters of the model with $\vect{x}_{k} \in \cX_{k}$. Let $n$ be the number of samples and $m$ be the number of dimensions. Suppose a task at $k | k \in [0, K)$ is observed and denoted as $\cT_{k} : \cT_{k} \sim p(\cT)$, where $\cT_{k} =\{\cX_{k}, \ell_{k}\}$ is a tuple with $\cX_{k} \in \vect{\mathbb{R}} ^{n \times m}$ being the input data and $\ell_{k}$ quantifies the loss incurred by $\cX_{k}$ using the model $g$ for the task at $k$. We denote a sequence of ${\vect{\theta}}_{k}$ as $\vect{u}_{k:K} = \{{\vect{\theta}}_{\tau} \in \Omega_{\theta}, k \leq \tau \leq K \},$ with $\Omega_{\theta}$ being the compact~(feasible) set for the parameters. We denote the optimal value with a superscript ${(*)};$ for instance, we use ${\vect{\theta}}_k^{(*)}$ to denote the optimal value of ${\vect{\theta}}_{k}$ at task $k.$ In this paper we use balance point, equilibrium point, and saddle point to refer to the point of balance between generalization and forgetting. We interchange between these terms whenever convenient for the discussion. We will use $\nabla_{(j)} i$ to denote the gradient of $i$ with respect to $j$ and $\Delta i$ to denote the first difference in discrete time.
An exemplary CL problem is described in Fig.~\ref{fig:ills} where we address a total of $K=3$ tasks. To particularize the idea in Fig. \ref{fig:ills}, we define the cost~(combination of catastrophic cost and generalization cost) at any instant $k$ as
\[J_{k}({\vect{\theta} }_{k}) = \gamma_{k} \ell_{k} + \sum_{\tau = 0}^{k-1} \gamma_{\tau} \ell_{\tau},\]
where $\ell_{\tau}$ is computed on task $\cT_{\tau}$ with $\gamma_{\tau}$ describing the contribution of $\cT_{\tau}$ to this sum.
To solve the problem at $k$, we seek ${\vect{\theta} }_{k}$ to minimize $J_{k}({\vect{\theta} }_{k})$. Similarly, to solve the problem in the complete interval $[0,K]$, we seek a ${\vect{\theta} }_{k}$ to minimize $J_{k}({\vect{\theta} }_{k})$ for each $k \in [0,K].$ In other words we seek to obtain ${\vect{\theta} }_{k}$ for each task such that the cost $J_{k}({\vect{\theta} }_{k})$ is minimized. Therefore, the optimization problem for the overall CL problem~(overarching goal of CL) is provided as the minimization of the cumulative cost \[ V_{k}(\vect{u}_{k:K}) = \sum_{\tau=k}^{K} \beta_{\tau} J_{\tau}( {\vect{\theta}}_{\tau}) \] such that $V_k^{(*)},$ is given as \begin{equation} V_k^{(*)} = min_{\vect{u}_{k:K} } V_{k}(\vect{u}_{k:K}), \label{op_1} \end{equation} with $0 \leq \beta_{\tau} \leq 1$ being the contribution of $J_{\tau}$ and $\vect{u}_{k:K}$ being a weight sequence of length $K-k.$
Within this formulation, two parameters determine the contributions of tasks: $\gamma_{\tau}$, the contribution of each task in the past, and $\beta_{\tau}$, the contribution of tasks in the future. To successfully solve the optimization problem, $V_{k}(\vect{u}_{k:K})$ must be bounded and differentiable, typically ensured by the choice of $\gamma_{\tau}, \beta_{\tau}.$ Lemma~1~(full statement and proof in Appendix A) states that \textit{equivalent performance cannot be guaranteed for an infinite number of tasks}. Furthermore, Corollary~1~(full statement and proof in Appendix A) demonstrates that \textit{if the task contributions are prioritized, the differentiability and boundedness of $J_\tau({\vect{\theta}}_{\tau})$ can be ensured}. A similar result was proved in \cite{knoblauch2020optimal}, where a CL problem with infinite memory was shown to be NP-hard from a set theoretic perspective. These results~(both ours and in \cite{knoblauch2020optimal}) demonstrate that a CL methodology cannot provide perfect performance on a large number of tasks and that tasks must be prioritized.
Despite these invaluable insights, the data corresponding to future tasks~(interval $[k,K]$) is not available, and therefore $V_{k}(\vect{u}_{k:K})$ cannot be evaluated. The optimization problem in Eq.~\eqref{op_1} naively minimizes the cost~(due to both previous tasks and new tasks) and does not provide any explicit modeling of the trade-off between forgetting and generalization. Furthermore, $\vect{u}_{k:K},$ the solution to Eq. \eqref{op_1} is a sequence of parameters, and it is not feasible to maintain $\vect{u}_{k:K}$ for a large number of tasks. Because of these three issues, the problem is theoretically intractable in its current form.
We will first recast the problem using tools from dynamic programming~\cite{lewis2012optimal}, specifically Bellman's principle of optimality, and derive a difference equation that summarizes the complete dynamics for the CL problem. Then, we will formulate a two-player differential game where we seek a saddle point solution to balance generalization and forgetting.
\section{Dynamics of Continual Learning}
Let \[ V_k^{(*)} = min_{\vect{u}_{k:K} } \sum_{\tau=k}^{K} \beta_{\tau} J_{\tau}( {\vect{\theta}}_{\tau});\] the dynamics of CL~(the behavior of optimal cost with respect to $k$) is provided as
\begin{equation}
\begin{aligned}
\Delta V^{(*)}_{k} = - min_{{\vect{\theta}}_{k} \in \Omega_{\theta}} \big[ \beta_k J_{k}( {\vect{\theta}}_{k}) + \big( \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{(*)} , \Delta {\vect{\theta}}_{k} \rangle +\langle \nabla_{\vect{x}_{k}} V_{k}^{(*)}, \Delta \vect{x}_{k} \rangle \big)\big].\\
\end{aligned}
\label{eq_M_DES}
\end{equation}
The derivation is presented in Appendix A~(refer to Proposition~1). Note that $V^{(*)}_{k}$ is the minima for the overarching CL problem in Eq. \eqref{eq_M_DES}and $\Delta V^{(*)}_{k}$ represents the change in $V^{(*)}_{k}$ upon introduction of a new task~(we hitherto refer to this as perturbations). Zero perturbations $(\Delta V^{(*)}_{k}=0)$ implies that the introduction of a new task does not impact our current solution; that is, the optimal solution on all previous tasks is optimal on the new task as well. Therefore, the smaller the perturbations, the better the performance of a model on all tasks, thus providing our main objective: minimize the perturbations~($\Delta V^{(*)}_{k} $). In Eq. \ref{eq_M_DES}, $\Delta V^{(*)}_{k} $ is quantified by three terms: the cost contribution from all the previous tasks and the new task~$J_{k}( {\vect{\theta}}_{k});$ the change in the optimal cost due to the change in the parameters $\langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{(*)}, \Delta {\vect{\theta}}_{k} \rangle$; and the change in the optimal cost due to the change in the input (introduction of new task)~$\langle \nabla_{\vect{x}_{k}} V_{k}^{(*)}, \Delta \vect{x}_{k}\rangle$.
The first issue with the cumulative CL problem~(Eq. \eqref{op_1}) can be attributed to the need for information from the future. In Eq. \eqref{eq_M_DES}, all information from the future is approximated by using the data from the new and the previous tasks. Therefore, the solution of the CL problem can directly be obtained by solving Eq. \eqref{eq_M_DES} using all the available data. Thus, \[ min_{{\vect{\theta}}_{k} \in \Omega} \big[ H(\Delta \vect{x}_{k}, \vect{\theta}_{k}) \big] \quad \text{yields } \Delta V^{(*)}_{k} \approx 0 \] for $\beta > 0,$ with \[ H(\Delta \vect{x}_{k}, \vect{\theta}_{k}) = \beta_k J_{k}( {\vect{\theta}}_{k}) + \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{(*)} , \Delta {\vect{\theta}}_{k} \rangle + \langle \nabla_{\vect{x}_{k}} V_{k}^{(*)}, \Delta \vect{x}_{k} \rangle.\] Essentially, minimizing $H(\Delta \vect{x}_{k}, \vect{\theta}_{k})$ would minimize the perturbations introduced by any new task $k$.
In Eq. \eqref{eq_M_DES}, the first and the third term quantify generalization and the second term quantifies forgetting. A model exhibits generalization when it successfully adapts to a new task~(minimizes the first and the third term in Eq. \eqref{eq_M_DES}). The degree of generalization depends on the discrepancy between the previous tasks and the new task~(numerical value of the third term in Eq. \eqref{eq_M_DES}) and the worst-case discrepancy prompts maximum generalization. Quantification of generalization is provided by $\Delta \vect{x}_{k}$ that summarizes the discrepancy between subsequent tasks. However, $\Delta \vect{x}_{k} = \vect{x}_{k+1} - \vect{x}_{k}$, and $\vect{x}_{k+1}$ is unknown at $k.$ Therefore, we simulate worst-case discrepancy by iteratively updating $\Delta \vect{x}_{k}$ through gradient ascent in order to maximize $H(\Delta \vect{x}_{k}, \vect{\theta}_{k})$; thus maximizing generalization. However, large discrepancy increases forgetting, and worst-case discrepancy yields maximum forgetting. Therefore, once maximum generalization is simulated, minimizing forgetting~(update $\vect{\theta}_{k}$ by gradient descent) under maximum generalization provides the balance.
To formalize our idea, let us indicate the iteration index at $k$ by $i$ and write $\Delta \vect{x}_{k}$ as $\Delta \vect{x}^{(i)}_{k}$ and ${\vect{\theta}}_{k}$ as ${\vect{\theta}}^{(i)}_{k}$ with $H(\Delta \vect{x}_{k}, \vect{\theta}_{k})$ as $H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k})$~(for simplicity of notation, we will denote $H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k})$ as $H$ whenever convenient). Next, we write
\begin{equation}
\begin{aligned}
\centering
&\underset{{\vect{\theta}}^{(i)}_{k} \in \Omega_{\theta} }{min} \bigg[ H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k}) \bigg] = \underset{{\vect{\theta}}^{(i)}_{k} \in \Omega_{\theta}}{min} \big[ \beta_k J_{k}( {\vect{\theta}}^{(i)}_{k}) + \langle \nabla_{{\vect{\theta}}^{(i)}_{k}} V_{k}^{(*)} , \Delta {\vect{\theta}}^{(i)}_{k}\rangle + \langle \nabla_{\vect{x}^{(i)}_{k}} V_{k}^{(*)}, \Delta \vect{x}^{(i)}_{k} \rangle \big]& \\ &\leq \underset{{\vect{\theta}}^{(i)}_{k} \in \Omega_{\theta}}{min} \big[\beta_k J_{k}( {\vect{\theta}}^{(i)}_{k}) + \langle \nabla_{{\vect{\theta}}^{(i)}_{k}} V_{k}^{(*)} , \Delta {\vect{\theta}}^{(i)}_{k}\rangle + \underset{\Delta \vect{x}^{(i)}_{k} \sim p(\cT)}{max} \langle \nabla_{\vect{x}^{(i)}_{k}} V_{k}^{(*)}, \Delta \vect{x}^{(i)}_{k} \rangle \big] &\\ & \leq \underset{{\vect{\theta}}^{(i)}_{k} \in \Omega_{\theta} }{min} \quad \underset{\Delta \vect{x}^{(i)}_{k} \sim p(\cT)}{max} \big[ H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k})\big]. & \label{eq_op_2}
\end{aligned}
\end{equation}
In Eq. \eqref{eq_op_2}, we seek the solution pair $(\Delta \vect{x}_k^{(*)}, \vect{\theta}_k^{(*)} ) \in (\Omega_{\theta}, \Omega_{\Delta \vect{x}_k^{(*)}}),$ where $\Delta \vect{x}_k^{(*)}$ maximizes $H$~(maximizing player, player 1) while $\vect{\theta}_k^{(*)}$ minimizes $H$~(minimizing player, player 2) where $(\Omega_{\theta}, \Omega_{\Delta \vect{x}_k^{(*)}})$ are the feasible sets for $\Delta \vect{x}_k^{(i)}$ and $\vect{\theta}_k^{(i)}$ respectively. The solution is attained, and $(\Delta \vect{x}_k^{(*)}, \vect{\theta}_k^{(*)})$ is said to be the equilibrium point when it satisfies the following condition:
\begin{equation}
\begin{aligned}
H(\Delta \vect{x}_k^{(*)}, \vect{\theta}^{(i)}_{k})
\geq H(\Delta \vect{x}_k^{(*)}, \vect{\theta}_k^{(*)} ) \geq H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}_k^{(*)} ).
\end{aligned}
\label{eq_condition}
\end{equation}
\subsection{Theoretical Analysis\label{theory}}
\begin{wrapfigure}[17]{l}{0.51\textwidth}
\includegraphics[width = \columnwidth]{Figures/Proof_Il.png}
\caption{Illustration of the proofs. $\Delta \vect{x}$~(player~1) is the horizontal axis, and the vertical axis indicates $\vect{\theta}$~(player~2) where the curve indicates H. If we start from the red circle for player~1~(player~2 is fixed at the blue circle), H is increasing~(goes from a grey circle to a red asterisk) with player 1 reaching the red asterisk. Next, start from the blue circle~($\vect{\theta}$ is at the red asterisk), the cost decreases. }
\label{fig:proof}
\end{wrapfigure} With our formulation, two key questions arise: Does our problem setup have an equilibrium point satisfying Eq. \eqref{eq_condition}? and how can one attain this equilibrium point? We answer these questions with Theorems~1 and 2, respectively. Full statements and proofs are provided in Appendix~A.
To illustrate the theory, we refer to Fig.~\ref{fig:proof}, where the initial values for the two players are characterized by the pair $\{\vect{\theta}^{(i)}_{k} \text{(blue circle)}, \Delta \vect{x}^{(i)}_{k} \text{(red circle)} \}$ and the cost value at $\{\vect{\theta}^{(i)}_{k}, \Delta \vect{x}^{(i)}_{k} \}$ is indicated by $H( \Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k})$~(the grey circle on the cost curve~(the dark blue curve)). Our proofing strategy is as follows. First, we fix $\vect{\theta}_k^{(.)} \in \Omega_{\theta}$ and construct a neighborhood $\mathcal{M}_k= \{\Omega_{x}, \vect{\theta}^{(.)}_{k}\}.$ Within this neighborhood we prove in Lemmas~2 and 4 that if we search for $\Delta \vect{x}^{(i)}_{k}$ through gradient ascent, we can converge to a local maximizer, and $H$ is maximizing with respect to $\Delta \vect{x}^{(i)}_{k}.$
Second, we let $\Delta \vect{x}^{(.)}_{k} \in \Omega_x$ be fixed, and we search for $ \vect{\theta}^{(i)}_{k}$ through gradient descent. Under this condition, we demonstrate two ideas in Lemmas 3 and 5: (1) we show that $H$ is minimizing in the neighborhood $\mathcal{N}_k: \mathcal{N}_k= \{\Omega_{\theta} , \Delta \vect{x}^{(.)}_{k} \}$; and (2) we converge to the local minimizer in the neighborhood $\mathcal{N}_k.$ Third, in the union of the two neighborhoods $\mathcal{M}_k \cup \mathcal{N}_k,$ (proven to be nonempty according to Lemma~6), we show that there exists at least one local equilibrium point~(Theorem~1); that is, there is at least one balance point.
\begin{thm}[Existence of an Equilibrium Point]
For any $k \in [0,K]$, let $\vect{\theta}^{(*)}_{k} \in \Omega_\theta,$ be the minimizer of $H$ according to Lemma 5 and define $\mathcal{M}^{(*)}_k = \{\Omega_{x}, \vect{\theta}^{(*)}_{k}\}.$ Similarly, let $\Delta \vect{x}_k^{(*)} \in \Omega_x,$ be the maximizer of $H$ according to Lemma 4 and define $\mathcal{N}^{(*)}_k = \{\Delta \vect{x}^{(*)}_{k}, \Omega_{\theta}\}.$ Further, let $\mathcal{M}^{(*)}_k \cup \mathcal{N}^{(*)}_k$ be nonempty according to Lemma.~6, then $(\Delta \vect{x}^{(*)}_{k},\vect{\theta}^{(*)}_{k}) \in \mathcal{M}^{(*)}_k \cup \mathcal{N}^{(*)}_k$ is a local equilibrium point.
\end{thm}
\begin{proof}
By Lemma \ref{lem:lem_min_opt} we have at $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ), (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ) \in \mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$ that
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ) &\leq& H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ).
\end{aligned}
\end{equation}
Similarly, according to Lemma \ref{lem:lem_max_opt}, at $ (\Delta \vect{x}^{ (* )}_k, \vect{\theta}^{ (* )}_{k} ), (\Delta \vect{x}^{ (i )}_k, \vect{\theta}^{ (* )}_{k} ) \in \mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$ we have
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (* )}_k, \vect{\theta}^{ (* )}_{k} ) \geq H (\Delta \vect{x}^{ (i )}_k, \vect{\theta}^{ (* )}_{k} ).
\end{aligned}
\end{equation}
Putting these inequalities together, we get
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ) \geq H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ) \geq H (\Delta \vect{x}_k^{ (i )}, \vect{\theta}^{ (* )}_k ),
\end{aligned}
\end{equation}
which is the saddle point condition, and therefore $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{(*)}_k )$ is a local equilibrium point in $\mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k.$ since $\mathcal{M}^{(*)}_k \cup \mathcal{N}^{(*)}_k$ be nonempty according to Lemma.~6.
\end{proof}
We next show that this equilibrium point is stable~(Theorem~2) under a sequential play. Specifically, we show that when player 1 plays first and player~2 plays second, we asymptotically reach a saddle point pair $(\Delta \vect{x}^{(*)}_{k},\vect{\theta}^{(*)}_{k})$ for $H.$ At this saddle point, both players have no incentive to move, and the game converges.
\begin{thm}[Stability of the Equilibrium Point]
For any $k \in [0,K]$, $\Delta \vect{x}^{(i)}_{k} \in \Omega_x$ and $\vect{\theta}^{(i)}_{k} \in \Omega_\theta$ be the initial values for $\Delta \vect{x}^{(i)}_{k}$ and $\vect{\theta}^{(i)}_{k}$ respectively. Define $\mathcal{M}_k = \{\Omega_{x}, \Omega_{\theta}\}$ with $H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k})$ given by Proposition 2. Let $\Delta \vect{x}^{(i+1)}_{k} - \Delta \vect{x}^{(i)}_{k} = \alpha_{k}^{(i)}\times (\nabla_{\Delta \vect{x}^{(i)}_{k}} H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(.)}_{k}) )/\| \nabla_{\Delta \vect{x}^{(i)}_{k}} H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(.)}_{k}) \|^2)$ and $\vect{\theta}^{(i+1)}_{k} - \vect{\theta}^{(i)}_{k} = -\alpha_{k}^{(i)}\times \nabla_{{\vect{\theta}}^{(i)}_{k}} H(\Delta \vect{x}^{(.)}_{k}, \vect{\theta}^{(i)}_{k}).$ Let the existence of an equilibrium point be given by Theorem~1, then, as a consequence of Lemmas 2 and 3, $(\Delta \vect{x}^{(*)}_{k},\vect{\theta}^{(*)}_{k}) \in \mathcal{M}_k$ is a stable equilibrium point for $H$ given
\label{thm:thm_st}.
\end{thm}
\begin{proof}
Consider now the order of plays by the two players. By Lemma~\ref{lem:lem_max}, a game starting at $ (\Delta \vect{x}_k^{ (i )}, \vect{\theta}^{ (i )}_k ) \in \mathcal{M}_k$ will reach $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k )$ which is a maximizer for $H.$ Now, define $\mathcal{N}_k = (\Delta \vect{x}_k^{ (* )}, \Omega_{\theta} ) \subset \mathcal{M}_k$ then a game starting at $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ) \in \mathcal{N}_k$ will converge to $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ) \in \mathcal{N}_k$ according to Lemma~\ref{lem:lem_min}. Since, $\mathcal{N}_k \subset \mathcal{M}_k,$ our result follows.
\end{proof}
In this game, the interplay between these two opposing players~(representative of generalization and forgetting, respectively) introduces the dynamics required to play the game. Furthermore, the results presented in this section are local to the task. In other words, we prove that we can achieve a balance between generalization and forgetting for each task $k$~(neighborhoods are task dependent, and we achieve a local solution given a task $k$).
Furthermore, our game is sequential; that is, there is a leader~(player 1) and a follower~(player 2). The leader~$(\Delta \vect{x}^{(i)}_{k})$ plays first, and the follower~$(\vect{\theta}^{(i)}_{k})$ plays second with complete knowledge of the leader's play. The game is directed by $\Delta \vect{x}^{(i)}_{k}$, and any changes in the task~(reflected in $\Delta \vect{x}^{(i)}_{k})$ will shift the input and thus the equilibrium point. Consequently, the equilibrium point varies with respect to a task, and one will need to attain a new equilibrium point for each shift in a task. Without complete knowledge of the tasks~(not available in a CL scenario), only a local result is possible. This highlights one of the key limitations of this work. Ideally, we would like a balance between forgetting and generalization that is independent of tasks. However, this would require learning a trajectory of the equilibrium point~(How does the equilibrium point change with the change in the tasks?) and is beyond the scope of this paper. One work that attempts to do this is~\cite{rusu2016progressive}, where the authors learn a parameter per task. For a large number of tasks, however, such an approach is computationally prohibitive.
These results are valid only under certain assumptions: (1) the Frobenius norm of the gradient is bounded, always positive; (2) the cost function is bounded and differentiable; and (3) the learning rate goes to zero as $i$ tends to infinity. The first assumption is reasonable in practice, and gradient clipping or perturbation strategies can be used to ensure it. The boundedness of the cost (second assumption) can be ensured by prioritizing the contributions of the task~(Lemma~1 and Corollary~1). The third assumption assumes a decaying learning rate. Learning rate decay is a common strategy and is employed widely. Therefore, all assumptions are practical and reasonable.
\begin{wrapfigure}[19]{r}{0.44\columnwidth}
\vspace{-11mm}
\begin{algorithm}[H]
\SetCustomAlgoRuledWidth{0.40\columnwidth}
Initialize ${\vect{\theta}}, D_{P}, D_{N}$ \\
\While{$k=1,2,3,... K$}{
j = 0\\
\While{$j < \rho$}{
Get $\vect{b}_{N} \in D^{N}_{k}$
Get $\vect{b}_{P} \in D^{P}_{k}$
Get $\vect{b}_{PN} = \vect{b}_{P} \cup \vect{b}_{N}$
Copy $\vect{b}_{PN}$ into $\vect{x}^{PN}_{k}$
i = 0
\While{$i+1 <= \zeta$}{
Update~$\vect{x}^{PN}_{k}$ with $J_{k}({\vect{\theta}}_k)$ using gradient ascent
i = i+1 }
Calculate $J_{k+\zeta}(\vect{\theta}^{(i)}_{k}) - J_{k}(\vect{\theta}^{(i)}_{k})$
Copy ${\vect{\theta}}^{(i)}_k$ into ${\vect{\theta}}^{B}_k$
i = 0
\While{$i+1 <= \zeta$}{
Update~${\vect{\theta}}^{B}_{k}$ with $J_{k}(\vect{\theta}^{B}_{k})$
i = i+1 }
Calculate $(J_{k}({\vect{\theta}}^{B}_{k}) - J_{k}(\vect{\theta}^{(i)}_{k}))$
Calculate $ H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k})$
Update ${\vect{\theta}}^{(i)}_k$ using gradient descent
}
j= j+1
}
Update ${D_{P}}$ with ${D_{N}}$
\caption{BCL \label{alg1a}}
\end{algorithm}
\end{wrapfigure}
\subsection{Balanced Continual Learning~\label{BCL}} Equipped with the theory, we develop a new CL method to achieve a balance between forgetting and generalization. By Proposition~2, the cost function can be upper bounded as $ H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k}) \leq \beta_k J_{k}(\vect{\theta}^{(i)}_{k}) + (J_{k}(\vect{\theta}^{(i+\zeta)}_{k}) - J_{k}(\vect{\theta}^{(i)}_{k})) + ( J_{k+\zeta}(\vect{\theta}^{(i)}_{k}) - J_{k}(\vect{\theta}^{(i)}_{k}) ),$ where $J_{k+\zeta}$ indicates $\zeta$ updates on player 1 and $\vect{\theta}^{(i+\zeta)}_{k}$ indicates $\zeta$ updates on player 2.
\begin{equation}
\begin{aligned}
\underbrace{\frac{\alpha_{k}^{(i)} \nabla_{\Delta \vect{x}_{k}} E[ H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k})] }{\| \nabla_{\Delta \vect{x}_{k}} H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k}) \|^2}}_{\text{Player~1}}, \\
\underbrace{-\alpha_{k}^{(i)}\times \nabla_{{\vect{\theta}}_{k}} E[H(\Delta \vect{x}^{(*)}_{k}, \vect{\theta}^{(i)}_{k}) )]}_{\text{Player~2}},
\end{aligned}
\label{eq:eq_Strat}
\end{equation}
The strategies for the two players $\Delta \vect{x}_{k}, {\vect{\theta}}_{k}$ are chosen in Eq. \eqref{eq:eq_Strat} with $E$ being the expected value operator. We can approximate the required terms in our update rule~(player strategies) using data samples (batches). Note that the approximation is performed largely through one-sided finite difference, which may introduce an error and is another potential drawback. The pseudo code of the BCL is shown in Algorithm~\ref{alg1a}. We define a new task array $\vect{\mathcal{D}} _N(k)$ and a task memory array~$\vect{\mathcal{D}} _{P}(k) \subset \cup_{\tau = 0}^{k-1}\cT_{\tau}$~(samples from all previous tasks). For each batch $b_{N} \in \vect{\mathcal{D}} _N(k)$, we sample $b_{P}$ from~$\vect{\mathcal{D}} _P(k),$ combine to create $b_{PN}(k)= b_{P}(k) \cup b_{N}(k)$, and perform a sequential play. Specifically, for each task the first player initializes $x_k^{PN} = b_{PN}(k)$ and performs $\zeta$ updates on $x_k^{PN}$ through gradient ascent. The second player, with complete knowledge of the first player's strategy, chooses the best play to reduce $H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k})$. To estimate player~2's play, we must estimate different terms in $H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k}).$ This procedure involves three steps. First, we use the first player's play and approximate $( J_{k+\zeta}(\vect{\theta}^{(i)}_{k}) - J_{k}(\vect{\theta}^{(i)}_{k}) ).$ Second, to approximate $(J_{k}(\vect{\theta}^{(i+\zeta)}_{k}) - J_{k}(\vect{\theta}^{(i)}_{k})):$ (a), we copy $\hat{\vect{\theta}}$ into $\hat{\vect{\theta}}_{B}$~(a temporary network) and perform $\zeta$ updates on $\hat{\vect{\theta}}_{B}$; and (b) we compute $J_{k}(\vect{\theta}^{(i+\zeta)}_{k})$ using $\hat{\vect{\theta}}_{B}(k+\zeta)$ and evaluate $(J_{k}(\vect{\theta}^{(i+\zeta)}_{k}) - J_{k}(\vect{\theta}^{(i)}_{k})).$ Third, equipped with these approximations, we compute $H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k})$ and obtain the play for the second player. Both these players perform the steps repetitively for each piece of information~(batch of data). Once all the data from the new task is exhausted, we move to the next task.
\subsection{Related Work}
Traditional solutions to the CL focus on either the forgetting issue~\cite{rusu2016progressive, yoon2017lifelong, yao2020don, caccia2021online, kirkpatrick2017overcoming, zenke2017continual, aljundi2018memory, lin1992self, lopez2017gradient, chaudhry2019continual} or the generalization issue~\cite{vinyals2016matching, snell2017prototypical, finn2017model, finn2019online, Caccia2020OnlineFA}. Common solutions to the forgetting problem involve dynamic architectures and flexible knowledge representation such as \cite{rusu2016progressive, yoon2017lifelong, yao2020don, caccia2021online}, regularization approaches including \cite{kirkpatrick2017overcoming, zenke2017continual, aljundi2018memory} and memory/experience replay \cite{lin1992self, lopez2017gradient, chaudhry2019continual}.
Similarly, quick generalization to a new task has been addressed through few-shot and one-shot learning approaches such as matching nets \cite{vinyals2016matching} and prototypical network \cite{snell2017prototypical}. More recently, the field of metalearning has approached the generalization problem by designing a metalearner that can perform quick generalization from very little data \cite{finn2017model, finn2019online, Caccia2020OnlineFA}.
In the past few years, metalearners for quick generalization have been combined with methodologies specifically designed for reduced forgetting \cite{javed2019meta, beaulieu2020learning}. For instance, the approaches in \cite{javed2019meta, beaulieu2020learning} adapt the model-agnostic metalearning~(MAML) framework in \cite{finn2017model} with robust representation to minimize forgetting and provide impressive results on CL. However, both these approaches require a pretraining phase for learning representation. Simultaneously, Gupta et al.~\cite{gupta2020lamaml} introduced LA-MAML---a metalearning approach where the impact of learning rates on the CL problem is reduced through the use of per-parameter learning rates. LA-MAML~\cite{gupta2020lamaml} also introduced episodic memory to address the forgetting issue. Other approaches also have attempted to model both generalization and forgetting. In \cite{farajtabar2019orthogonal}, the gradients from new tasks are projected onto a subspace that is orthogonal to the older tasks, and forgetting is minimized. Similarly, Joseph and Balasubramanian \cite{joseph2020metaconsolidation} utilized a Bayesian framework to consolidate learning across previous and current tasks, and Yin et al.~\cite{yin2020optimization} provided a framework for approximating loss function to summarize the forgetting in the CL setting. Furthermore, Abolfathi et al.~\cite{abolfathi2021coachnet} focused on sampling episodes in the reinforcement learning setting, and Elrahimi et al.~\cite{ebrahimi2020adversarial} introduced a generative adversarial network-type structure to progressively learn shared features assisting reduced forgetting and improved generalization. Despite significant progress, however, these methods~\cite{javed2019meta, beaulieu2020learning, gupta2020lamaml,farajtabar2019orthogonal, joseph2020metaconsolidation, yin2020optimization, abolfathi2021coachnet, ebrahimi2020adversarial} are still inherently tilted toward maximizing generalization or minimizing forgetting because they naively minimize the loss function. Therefore, the contribution of different terms in the loss function becomes important. For instance, if the generalization cost is given more weight, a method would generalize better. Similarly, if forgetting cost is given more weight, a method would forget less. Therefore, the resolution of the trade-off inherently depends on an hyperparameter.
The first work to formalize the trade-off in CL was MER, where the trade-off was formalized as a gradient alignment problem. Similar to MER, Doan et al.~(\cite{doan2021theoretical}) studied forgetting as an alignment problem. In MER, the angle between the gradients was approximated by using Reptile \cite{nichol2018firstorder}, which promotes gradient alignment by reducing weight changes. On the other hand, \cite{doan2021theoretical} formalized the alignment as an eigenvalue problem and introduced a PCA-driven method to ensure alignment. Our approach models this balance as a saddle point problem achieved through stochastic gradient such that the saddle point~(balance point or equilibrium point) resolves the trade-off.
Our approach is the first in the CL literature to prove the existence of the saddle point~(the balance point) between generalization and forgetting given a task in a CL problem. Furthermore, we are the first to theoretically demonstrate that the saddle point can be achieved reasonably under a gradient ascent-descent game. The work closest to ours is \cite{ebrahimi2020adversarial}, where an adversarial framework is described to minimize forgetting in CL by generating task-invariant representation. However, \cite{ebrahimi2020adversarial} is not model agnostic~(the architecture of the network is important) and requires a considerable amount of data at the start of the learning procedure. Because of these issues, \cite{ebrahimi2020adversarial} is not suitable for learning in the sequential scenario.
\section{Experiments}
We use the CL benchmark \cite{Hsu18_EvalCL} for our experiments and retain the experimental settings~(hyperparameters) from \cite{Hsu18_EvalCL, vandeven2019generative}. For comparison, we use the split-MNIST, permuted-MNIST, and split-CiFAR100 data sets while considering three scenarios: incremental domain learning~(IDL), incremental task learning~(ITL), and incremental class learning~(ICL). The splitting and permutation strategies when applied to the MNIST or CiFAR100 data set can generate task sequences for all three scenarios~(illustrated in Figure~1 and Appendix: Figure~2 of \cite{Hsu18_EvalCL}). For comparing our approach, we use three baseline strategies---standard neural network with Adam~\cite{kingma2014adam}, Adagrad~\cite{duchi2011adaptive}, and SG---and use $L_2$-regularization and naive rehearsal~(which is similar to experience replay). For CL approaches, we use EWC~\cite{kirkpatrick2017overcoming}, online EWC~\cite{schwarz2018progress}, SI~\cite{zenke2017continual}, LwF~\cite{LwF}, DGR~\cite{DGR}, RtF~\cite{vandeven2019generative}, MAS~\cite{aljundi2018memory}, MER~\cite{riemer2018learning}, and GEM~\cite{GEM}. We utilize data preprocessing as provided by \cite{Hsu18_EvalCL}. Additional details on experiments can be found in Appendix~B and \cite{Hsu18_EvalCL, vandeven2019generative}. All experiments are conducted in Python 3.4 using the pytorch $1.7.1$ library with the NVIDIA-A100 GPU for our simulations.
\begin{table}[bt]
\vspace{-5pt}
\tiny
\centering
\resizebox{\columnwidth}{!}{
\begin{tabular}{c|ccc|ccc}
\hline \hline
\multirow{3}{*}{\textbf{Method}}&\multicolumn{3}{c|}{\textbf{split-MNIST}} &
\multicolumn{3}{c}{\textbf{permuted-MNIST}} \\ \cline{2-7}
& Incremental & Incremental & Incremental
& Incremental & Incremental & Incremental\\
& task learning & domain learning & class learning
& task learning & domain learning & class learning\\
& [ITL] & [IDL] & [ICL]
& [ITL] & [IDL] & [ICL] \\ \hline
Adam& $95.52 \pm 2.14$ & $54.75 \pm2.06$ &$19.72\pm0.03$
& $93.42 \pm 0.56$ & $77.87 \pm1.27$ &$14.02\pm1.25$\\
SGD& $97.65 \pm0.28$ & $62.80 \pm0.34$ &$19.36\pm0.02$
& $90.95 \pm0.20$ & $78.17 \pm1.16$ &$12.82\pm0.95$\\
Adagrad& $98.37 \pm0.29$ & $57.59 \pm2.54$ &$19.59\pm0.17$
& $92.45 \pm0.16$ & $91.59 \pm0.46$ &$29.09\pm1.48$\\
$L_2$& $97.62 \pm0.69$ & $66.84 \pm3.91$. &$22.92\pm1.90$
& $94.87 \pm0.38$ & $92.81 \pm0.32$ &$13.92\pm1.79$\\
Naive rehearsal& $99.32 \pm0.10$ & $94.85 \pm0.80$ &$90.88\pm0.70$
& $96.23 \pm0.04$ & $95.84 \pm0.06$ &$96.25\pm0.10$\\
Naive rehearsal-C& $99.41 \pm0.04$ & $97.13 \pm0.37$ &$94.92\pm0.63$
& $97.13 \pm0.03$ & $96.75 \pm0.03$ &$97.24\pm0.05$\\
\hline \hline
EWC& $96.59 \pm0.99$ &$57.31\pm1.07$ &$19.70\pm0.14$
& $95.38 \pm0.33$ &$89.54\pm0.52$ &$26.32\pm4.32$\\
Online~EWC& $99.01 \pm0.12$ &$58.25\pm1.23$ &$19.68\pm0.05$
& $95.15 \pm0.49$ &$93.47\pm0.01$ &$42.58\pm6.50$\\
SI& $99.10 \pm0.16$ &$64.63\pm1.67$ &$19.67\pm0.25$
& $94.35 \pm0.51$ &$91.12\pm0.93$ &$58.52\pm4.20$\\
MAS& $98.88 \pm0.14$ &$61.98\pm7.17$ &$19.70\pm0.34$
& $94.74 \pm0.52$ &$93.22\pm0.80$ &$50.81\pm2.92$\\
GEM& $98.32 \pm0.08$ &$97.37\pm0.22$ &$93.04\pm0.05$
& $95.44 \pm0.96$ &$96.86\pm0.02$ &$96.72\pm0.03$\\
DGR& $99.47 \pm0.03$ &$95.74\pm0.23$ &$91.24\pm0.33$
& $92.52 \pm0.08$ &$95.09\pm0.04$ &$92.19\pm0.09$\\
RtF&$\textbf{99.66}\pm\textbf{0.03}$&$97.31\pm0.11$&$92.56\pm0.21$
&$97.31\pm0.01$ &$97.06\pm0.02$ &$96.23\pm0.04$\\
MER&$97.12\pm0.10$ &$92.16\pm0.35$ &$93.20\pm0.12$
&$97.15\pm0.08$ &$96.11\pm0.31$ &$91.71\pm0.03$\\ \cline{2-7}
BCL~(With Game) &$99.52\pm0.07$ & $\textbf{98.71 }\pm\textbf{0.06}$ &$\textbf{97.32}\pm\textbf{0.17}$
& $\textbf{97.41} \pm \textbf{0.01}$ & $\textbf{97.51} \pm \textbf{0.05}$ &$\textbf{97.61}\pm\textbf{0.01}$ \\
BCL~(Without Game) & $97.73 \pm 0.03$ & $96.43 \pm0.29$ &$91.88\pm0.55$
& $96.16 \pm 0.03$ & $96.08 \pm 0.06$ &$95.96\pm0.06$\\
\hline
\hline
\end{tabular}}
\caption{Performance of our approach compared with other methods in the literature. We record the mean and standard deviation of the retained accuracy for the different methods. The best scores are in bold.}
\label{tab:main_res}
\vspace{-8mm}
\end{table}
\textbf{Comparison with the state of the art:} The results for our method are summarized in Table~\ref{tab:main_res} and \ref{tab:cifar_100}. The efficiency for any method is calculated by observing the average accuracy~(retained accuracy (RA)~\cite{riemer2018learning}) at the end of each repetition and then evaluating the mean and standard deviation of RA across different repetitions. For each method, we report the mean and standard deviation of RA over five repetitions of each experiment. In each column, we indicate the best-performing method in bold. For the split-MNIST data set, we obtain $99.52\pm0.07$ for ITL, $98.71\pm0.06$ for IDL, and $97.32\pm 0.17$ for ICL. Similarly, with the permuted-MNIST data set, we obtain $97.41\pm0.01$ for ITL, $97.51\pm0.05$ for IDL, and $97.61\pm0.01$ for ICL. Furthermore, with the split-CiFAR100 data set, we obtain $81.82 \pm 0.17$ for ITL, $62.11\pm 0.00$ for IDL, and $69.27 \pm 0.03$ for ICL. BCL is the best-performing methodology for all cases~(across both data sets) except RtF for ITL~($0.14 \%$ drop) with the split-MNIST data set.
\begin{wraptable}[16]{r}{0.61\columnwidth}
\tiny
\centering
\begin{tabular}{c|ccc}
\hline \hline
\multirow{3}{*}{\textbf{Method}}&\multicolumn{3}{c}{\textbf{split-CiFAR100}}\\ \cline{2-4}
& Incremental & Incremental & Incremental \\
& task learning & domain learning & class learning \\ \hline
Adam& $30.53 \pm0.58$ & $19.65 \pm0.14$ &$17.20\pm0.06$\\
SGD& $43.77 \pm1.15$ & $19.17 \pm0.12$ &$17.18\pm0.12$\\
Adagrad& $36.27 \pm0.43$ & $19.06 \pm0.14$ &$15.83\pm0.20$\\
$L_2$& $51.73 \pm1.30$ & $19.96 \pm0.15$ &$17.12\pm0.04$\\
Naive rehearsal& $70.20 \pm0.17$ & $35.94 \pm0.39$ &$34.33\pm0.19$\\
Naive rehearsal-C& $78.41 \pm0.37$ & $51.81 \pm0.18$ &$51.28\pm0.17$\\
\hline \hline
EWC& $61.11 \pm1.43$ &$19.76\pm0.12$ &$19.70\pm0.14$\\
Online~EWC& $63.22 \pm0.97$ &$20.03\pm0.10$ &$17.16\pm0.09$\\
SI& $64.81 \pm1.00$ &$20.26\pm0.09$ &$17.26\pm0.11$\\
MAS& $64.77 \pm0.78$ &$19.99\pm0.16$ &$17.07\pm0.12$\\ \cline{2-4}
BCL(With Game) & $\textbf{81.82} \pm \textbf{0.17}$ &$\textbf{62.11 }\pm\textbf{0.00}$ & $\textbf{69.27} \pm \textbf{0.03}$ \\
BCL(Without Game) & $69.17 \pm0.12$ &$51.82\pm0.19$ & $52.82 \pm 0.01$ \\
\hline
\hline
\end{tabular}
\caption{Performance of BCL for the split-CiFAR100 data set. We record the retained accuracy for the different methods. We obtained RA scores for all methods except BCL from \cite{Hsu18_EvalCL}.}
\label{tab:cifar_100}
\end{wraptable} Generally, ITL is the easiest learning scenario~\cite{Hsu18_EvalCL}, and all methods therefore perform well on ITL~(the performance is close). For the ITL scenario with the split-MNIST data set, BCL is better than most methods; but several methods, such as naive rehearsal, naive rehearsal-C, RtF, and DGR, attain close RA values~(less than $1\%$ from BCL). Note that both DGR and RtF involve a generative model pretrained with data from all the tasks. In a sequential learning scenario, one cannot efficiently train generative models because data corresponding to all the tasks is not available beforehand. Although RtF provides improved performance for split-MNIST~(ITL), the improvement is less than $1\%$~(not significant). In fact, RtF performance is poorer for BCL in ICL by $4.76\%$~(a significant drop in performance) and in IDL by $1.4 \%.$
Two additional observations can be made about the split-MNIST data set. First, Adagrad, SGD, and L2 achieve better performance than does Adam. Therefore, in our analysis Adagrad appears more appropriate although Adam is popularly used for this task. Second, naive rehearsal~(both naive rehearsal and naive rehearsal-C approaches) achieves performance equivalent to the state-of-the-art methods with similar memory overhead. Furthermore, naive rehearsal performs much better than online EWC and SI, especially in the ICL scenario. These limitations indicate that regularization-driven approaches are not much better than baseline models and in fact perform poorer than methods involving memory~(naive rehearsals, MER, BCL, etc.). In~\cite{knoblauch2020optimal}, it was shown theoretically that memory-based approaches typically do better than regularization-driven approaches, as is empirically observed in this paper, too. Another interesting observation is that EWC and online EWC require significant hyperparameter tuning, which would be difficult to do in real-world scenarios. Other regularization-based methods, such as SI and MAS, also suffer from the same issue.
The observations from the split-MNIST carry forward to the permuted-MNIST data set. Moreover, RA values for the permutation MNIST data set are better for the split-MNIST data set across the board, indicating that the permutation MNIST data set presents an easier learning problem. Similar to the observations made with the split-MNIST data set, BCL is better than all methods for the permuted-MNIST dataset, with naive rehearsal and RtF providing RA values that are close~(less than $1 \%$). The only methodology in the literature that attempts to model the balance between forgetting and generalization is MER, an extension of GEM. From our results, we observe that BCL is better than MER in all cases~(Split-MNIST--$2.4 \%$ improvement for ITL, $6.55 \%$ improvement for IDL, $4.12 \%$ improvement for ICL and Permuted-MNIST--$0.26 \%$ improvement for ITL, $1.4 \%$ improvement for IDL and $5.9 \%$ improvement for ICL).
The substantial improvements obtained by BCL are more evident in Table \ref{tab:cifar_100} where the results on the split-CiFAR100 data set are summarized. BCL is clearly the best-performing method. The next best-performing method is naive rehearsal-C, where BCL improves performance by $3.41 \%$ for ITL, $10.3 \%$ for IDL, and $17.99 \%$ for ICL. Other observations about the regularization methods and the rest of the baseline methods carry forward from Table~\ref{tab:main_res}. However, one key difference is that while AdaGrad is observed to be better than Adam for the MNIST data set, Adam is comparable to SGD and Adagrad for split-CiFAR100. In summary, BCL is comparable to or better than the state of the art in the literature for both the MNIST and CiFAR100 data sets.
\begin{wrapfigure}[30]{l}{0.49\textwidth}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width = \columnwidth]{Figures/tradeoff_loss.pdf}
\caption{}
\label{fig:all_task}
\end{subfigure}\\
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width = \columnwidth]{Figures/tradeoff_loss_one_task.pdf}
\caption{}
\label{fig:one_task}
\end{subfigure}
\caption{(Top) Progression of different terms in Eq. \eqref{eq_M_DES} with respect to update iterations, where the index on the x axis is calculated as $k\times 300$. The tasks boundaries~(at what $i$ the tasks are introduced) are illustrated through shades of grey. (down) Illustration of the cost at $k =3$.}
\end{wrapfigure}
\textbf{Do we need a game to achieve this performance?} To provide additional insights into the benefits of the game, we compare RA values with and without the game. In our setup, $\Delta \vect{x}^{(i)}_{k}$ aims at increasing the cost, and $\vect{\theta}^{(i)}_{k}$ aims at reducing the cost. If we hold the play for $\Delta \vect{x}^{(i)}_{k}$, then the dynamics required to play the game do not exist, thus providing a method that can perform CL without the game. Therefore, we induce the absence of a game by fixing $\Delta \vect{x}^{(i)}_{k}$ and perform each of the nine experiments for five repetitions~(ICL, IDL, and ITL for split-MNIST, permuted-MNIST, and split-CiFAR100). We summarize these results in the last two rows of Tables \ref{tab:main_res} and \ref{tab:cifar_100}. Consequently, we make two observations. First: even without the game, we achieve RA values comparable to the state of the art. This is true for both the MNIST and CiFAR100 data sets. Second, with the introduction of the game, we observe improved RA values across the board~(at least by $1 \%$ with MNIST). The difference is clearer with the CiFAR100 data set where we observe a substantial improvement in RA values~(at least $10 \%$).
\textbf{Does the theory appropriately model the continual learning problem?} In Fig.~\ref{fig:all_task} we plot the progression of $J_{k}( {\vect{\theta}}_{k})+ \langle \nabla_{\vect{x}_{k}} V_{k}^{(*)}, \Delta \vect{x}_{k}\rangle$~(blue curve), $\langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{(*)}, \Delta {\vect{\theta}}_{k} \rangle$~(red curve), and the sum, namely, $\Delta V^{(*)}_{k} $~(green curve). A total of six tasks, sampled from the permutation MNIST data set within the ICL setting, are illustrated in Fig. \ref{fig:all_task}. These tasks are introduced every 300 update steps. From Fig. \ref{fig:all_task} we make two important observations. First, as soon as tasks 1, 3, 4, and 6 are introduced, all three curves~(red, blue, and green) indicate a bump. Second, when tasks 2 and 5 are introduced, there is no change. On a closer look at task 3 in Fig.~\ref{fig:one_task}, we observe that when the task 3 is introduced, the blue curve exhibits a large positive bump~(the introduction of the new tasks increases the first and the third terms in Eq. \eqref{eq_M_DES}). The increase implies that task 3 forced the model to generalize and increased forgetting on tasks 1 and 2~(observed by the increase in the green curve). To compensate for this increase, we require the model~$\vect{\theta}_k$ to behave adversarially and introduce a large enough negative value in $\langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{(*)}, \Delta {\vect{\theta}}_{k} \rangle$~(red curve) to cancel out the increase in the blue curve. In Fig.~\ref{fig:one_task} the red curve demonstrates a large negative value~(expected behavior) and eventually~(as $i$ increases) forces the blue curve~(by consequence, green: the sum of red and blue) to move toward zero~(the model compensates for the increase in forgetting). As observed, the blue and the red curves behave opposite to each other and introduce a push-pull behavior that stops only when the two cancel each other and the sum~(green) is zero. Once the sum has reached zero, there is no incentive for the red and green to be nonzero, and therefore they remain at zero; thus, all three curves (green, red, and blue) remain at zero once converged until a task 4 is introduced~(when there is another bump, as seen in Fig. \ref{fig:all_task}). However, this increase in the blue curve is not observed when tasks 2 and 5 are introduced. When new tasks are similar to the older tasks, it is expected that $\Delta V^{(*)}_{k} =0,$ as is observed in Fig. \ref{fig:all_task}.
All of these observations are fully explained by Eq. \ref{eq_M_DES}, which illustrates that the solution to the CL problem is obtained optimally only when $\Delta V^{(*)}_{k} =0$~(observed in Figs. \ref{fig:all_task} and \ref{fig:one_task}). The term $\Delta V^{(*)}_{k} $ is quantified by $J_{k}( {\vect{\theta}}_{k}), \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{(*)}, \Delta {\vect{\theta}}_{k} \rangle$ and $\langle \nabla_{\vect{x}_{k}} V_{k}^{(*)}, \Delta \vect{x}_{k}\rangle$. Our theory suggests that there exists an inherent trade-off between different terms in Eq. \ref{eq_M_DES}. Therefore every time a new task is observed, it is expected that~$J_{k}( {\vect{\theta}}_{k})+ \langle \nabla_{\vect{x}_{k}} V_{k}^{(*)}, \Delta \vect{x}_{k}\rangle$ increases~(increase in blue curve when tasks 1, 3, 4, and 6 are introduced) and $\langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{(*)}, \Delta {\vect{\theta}}_{k} \rangle$ compensates to cancel this increase~(red curve exhibits a negative jump). In Theorems~1 and 2 we demonstrate the existence of this balance point~(for each task, as $i$ increases, $\Delta V^{(*)}_{k} $ tends to zero, as observed in Fig. \ref{fig:one_task}) and $\Delta V^{(*)}_{k} $ remains zero~(the balance point is stable, proved in Theorem~2) until a new task increases forgetting. Furthermore, our theory claims the existence of a solution with respect to each task. This is also observed in Fig.~\ref{fig:all_task} as, for each task, there is an increase in cost, and BCL quickly facilitates convergence. \emph{These observations indicate that our dynamical system in Eq. \ref{eq_M_DES}
accurately describes the dynamics of the continual learning problem}. Furthermore, the assumptions under which the theory is developed are practical and are satisfied by performing continual learning on the permuted MNIST problem.
\section{Conclusion}
We developed a dynamic programming-based framework to enable the methodical study of key challenges in CL. We show that an inherent trade-off between generalization and forgetting exists and must be modeled for optimal performance. To this end, we introduce a two-player sequential game that models the trade-off. We show in theory and simulation that there is an equilibrium point that resolves this trade-off~(Theorem~1) and that this saddle point can be attained~(Theorem~2). However, we observe that any change in the task modifies the equilibrium point. Therefore, a global equilibrium point between generalization and forgetting is not possible, and our results are valid only in a neighborhood~(defined given a task). To attain this equilibrium point, we develop BCL and demonstrate state-of-the-art performance on a CL benchmark~\cite{Hsu18_EvalCL}. In the future, we will extend our framework for nonEuclidean tasks.
\section{Broader Impact}
\textit{Positive Impacts:} CL has a wide range of applicability. It helps avoids retraining, and it improves the learning efficiency of learning methods. Therefore, in science applications where the data is generated sequentially but the data distribution varies with time, our theoretically grounded method provides the potential for improved performance.
\textit{Negative Impacts:} Our theoretical framework does not have direct adverse impacts. However, the potential advantages of our approach can improve the efficiency of adverse ML systems such as fake news, surveillance, and cybersecurity attacks.
\begin{ack}
This work was supported by the U.S.\ Department of Energy, Office of Science, Advanced Scientific Computing Research, under Contract DE-AC02-06CH11357 and by a DOE Early Career Research Program award. We are grateful for the computing resources from the Joint Laboratory for System Evaluation and Leadership Computing Facility at Argonne. We also are grateful to Dr. Vignesh Narayanan, assistant professor, University of South Carolina, and Dr. Marieme Ngom, Dr. Sami Khairy --postdoctoral appointees, Argonne National Laboratory, for their insights.
\end{ack}
\bibliographystyle{plain}
\section*{Supplementary Information}
We use $\vect{\mathbb{R}} $ to denote the set of real numbers and $\vect{\mathbb{N}} $ to denote the set of natural numbers. We use $\|.\|$ to denote the Euclidean norm for vectors and the Frobenius norm for matrices, while using bold symbols to illustrate matrices and vectors. We define an interval $[0, K ), K \in \vect{\mathbb{N}} $ and let $p (\cT )$ be the distribution over all the tasks observed in this interval. For any $k \in [0, K ) ,$ we define a parametric model $g (. )$ with ${\vect{y}}_{k} = g (\vect{x}_{k}; {\vect{\theta}}_{k} )$, where ${\vect{\theta}}_{k}$ is a vector comprising all parameters of the model with $\vect{x}_{k} \in \cX_{k}$. Let $n$ be the number of samples and $m$ be the number of dimensions. Suppose a task at $k | k \in [0, K )$ is observed and denoted as $\cT_{k} : \cT_{k} \sim p (\cT )$, where $\cT_{k} =\{\cX_{k}, \ell_{k}\}$ is a tuple with $\cX_{k} \in \vect{\mathbb{R}} ^{n m}$ being the input data and $\ell_{k}$ quantifies the loss incurred by $\cX_{k}$ using the model $g$ for the task at $k$. We denote a sequence of ${\vect{\theta}}_{k}$ as $\vect{u}_{k:K} = \{{\vect{\theta}}_{\tau} \in \Omega_{\theta}, k \leq \tau \leq K \},$ with $\Omega_{\theta}$ being the compact~ (feasible ) set for the parameters. We denote the optimal value with a superscript ${ (* )};$ for instance, we use ${\vect{\theta}}_k^{ (* )}$ to denote the optimal value of ${\vect{\theta}}_{k}$ at task $k.$ In this paper we use balance point, equilibrium point, and saddle point to refer to the point of balance between generalization and forgetting. We interchange between these terms whenever convenient for the discussion. We will use $\nabla_{ (j )} i$ to denote the gradient of $i$ with respect to $j$ and $\Delta i$ to denote the first difference in discrete time.
\section{Additional Results}
We define the cost~ (combination of catastrophic cost and generalization cost ) at any instant $k$ as $J_{k} ({\vect{\theta} }_{k} ) = \gamma_{k} \ell_{k} + \sum_{\tau = 0}^{k-1} \gamma_{\tau} \ell_{\tau},$ where $\ell_{\tau}$ is computed on task $\cT_{\tau}$ with $\gamma_{\tau}$ describing the contribution of $\cT_{\tau}$ to this sum. We will show that for any fixed $k$, the catastrophic forgetting cost~$J_k ({\vect{\theta} }_{k} )$ is divergent in the limit $k \rightarrow \infty$ if equal contribution from each task is expected.
\begin{lemma}
For any $k \in \vect{\mathbb{N}} ,$ define $J_k ({\vect{\theta} }_{k} ) = \sum_{\tau = 0}^{k} \gamma_{\tau} \ell_{\tau}.$ For all $\tau,$ assume $\ell_{\tau}$ to be continuous with $L \geq \ell_{\tau} \geq \epsilon, \forall \tau, \epsilon >0$ and let $\gamma_{\tau} =1$. Then $J_k ({\vect{\theta} }_{k} )$ is divergent as $k \rightarrow \infty.$
\end{lemma}
\begin{proof}[Proof of Lemma 1]
With $J_k ({\vect{\theta} }_{k} ) = \sum_{\tau = 0}^{k} \gamma_{\tau} \ell_{\tau}$ we write $\lim_{k \rightarrow \infty} \sum_{\tau = 0}^{k} \gamma_{\tau} \ell_{\tau} \geq \lim_{k \rightarrow \infty} \sum_{\tau = 0}^{k} \gamma_{\tau} \epsilon $, where $\gamma_{\tau}=1$ which provides $ \lim_{k \rightarrow \infty} \sum_{\tau = 0}^{k} \epsilon = \infty.$ Therefore, $J_k ({\vect{\theta} }_{k} )$ is divergent.
\end{proof}
When $\ell_{\tau} \geq \epsilon$ with $\epsilon >0$ implies that each task incurs a nonzero cost. Furthermore, $\gamma_{\tau} =1,$ it implies that each task provides equal contribution to the catastrophic forgetting cost and contributed nonzero value to $J_k ({\vect{\theta} }_{k} )$. The aforementioned lemma demonstrates that equivalent performance~ (no forgetting on all tasks ) cannot be guaranteed for an infinite number of tasks when each task provides a nonzero cost to the sum~ (you have to learn for all the tasks ). However, if the task contributions are prioritized based on knowledge about the task distribution, the sum can be ensured to be convergent as shown in the next corollary.
\begin{corr}
For any $k \in \vect{\mathbb{N}} ,$ define $J_k ({\vect{\theta} }_{k} ) = \sum_{\tau=0}^{k} \gamma_{\tau} \ell_{\tau}$ where $\ell_{\tau}$ is continuous and bounded such that $\epsilon \leq \ell_{\tau} \leq L, \forall \epsilon >0.$ Define $N= \frac{1}{k}$ and choose $\gamma_{N}$ such that $\gamma_{N} \rightarrow 0, N \rightarrow \infty$ and assume when there are infinite number of tasks, $\lim_{ N \rightarrow \infty} \sum_{N} \gamma_{N} \leq M.$ Under these assumptions, $J_k ({\vect{\theta} }_{k} )$ is convergent.
\end{corr}
\begin{proof}[Proof of Corollary 1]
Since $\ell_{\tau} \leq L$, $J_k ({\vect{\theta} }_{k} ) = lim_{k \rightarrow \infty} \sum_{\tau=0}^{k} \gamma_{\tau} \ell_{\tau} \leq lim_{k \rightarrow \infty} \sum_{\tau=0}^{k} \gamma_{\tau} L \leq L lim_{k \rightarrow \infty} \sum_{\tau = 0}^{k} \gamma_{\tau}.$
Since $\lim_{k \rightarrow \infty} \sum_{\tau = 0}^{k} \gamma_{\tau} = \lim_{ N \rightarrow \infty} \sum_{N} \gamma_{N}$ as $N =\frac{1}{k}$, therefore $\lim_{k \rightarrow \infty} \sum_{\tau = 0}^{k} \gamma_{\tau} \leq M$ and $J_k ({\vect{\theta} }_{k} )$ is upper bounded by $ L M$. As a result, $J_k ({\vect{\theta} }_{k} )$ is convergent since $J_k ({\vect{\theta} }_{k} )$ is a monotone.
\end{proof}
To solve the problem at $k$, we seek ${\vect{\theta} }_{k}$ to minimize $J_{k} ({\vect{\theta} }_{k} )$. Similarly, to solve the problem in the complete interval $[0, K )$, we seek a ${\vect{\theta} }_{k}$ to minimize $J_{k} ({\vect{\theta}}_{k} )$ for each $k \in [0,K ).$ In other words we seek to obtain ${\vect{\theta} }_{k}$ for each task such that the cost $J_{k} ({\vect{\theta} }_{k} )$ is minimized. The optimization problem for the overall CL problem~ (overarching goal of CL ) is then provided as the minimization of the cumulative cost $V_{k} (\vect{u}_{k:K} ) = \sum_{\tau=k}^{K} \beta_{\tau} J_{\tau} ( {\vect{\theta}}_{\tau} )$ such that $V_k^{ (* )},$ is given as \begin{equation} V_k^{ (* )} = min_{\vect{u}_{k:K} } V_{k} (\vect{u}_{k:K} ), \label{op} \end{equation} with $0 \leq \beta_{\tau} \leq 1$ being the contribution of $J_{\tau}$ and $\vect{u}_{k:K}$ being a weight sequence of length $K-k.$ We will now derive the difference equation for our cost formulation.
\begin{prop}
For any $k \in [0, K ),$ define $ V_k = \sum_{\tau=k}^{K} \beta_{\tau} J_{\tau} ( {\vect{\theta}}_{\tau} )$ with ${\vect{\theta}}_{\tau} \in \Omega.$ Define $\vect{u}_{k:K} = \{{\vect{\theta}}_{\tau} \in \Omega, k \leq \tau \leq K \},$, with $\Omega$ being the compact~ (feasible ) set as a sequence of parameters with length $K-k$ and $V_k^{ (* )} = min_{\vect{u}_{k:K} } \sum_{\tau=k}^{K} \beta_{\tau} J_{\tau} ( {\vect{\theta}}_{\tau} ).$ Then, the following is true
\begin{equation}
\begin{aligned}
\Delta V^{ (* )}_{k} = - min_{{\vect{\theta}}_{k} \in \Omega} \big[ \beta_k J_{k} ( {\vect{\theta}}_{k} ) + \big ( \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}_{k} \rangle +\langle \nabla_{\vect{x}_{k}} V_{k}^{ (* )}, \Delta \vect{x}_{k} \rangle \big )\big],&
\end{aligned}
\label{eq_opt}
\end{equation}
where $\Delta V^{ (* )}_{k}$ represents the first difference due to the introduction of a task, $\Delta {\vect{\theta}}_{k} $ due to parameters and $\nabla_{\vect{x}_{k}}$ due to the task data with $\beta_k \in \mathbb{R} \cup [0,1], \forall k$ and $J_{k} ({\vect{\theta} }_{k} ) = \gamma_{k} \ell_{k} + \sum_{\tau = 0}^{k-1} \gamma_{\tau} \ell_{\tau}$.
\label{Der_HJB}
\end{prop}
\begin{proof}
Given $V_k^{ (* )} = min_{\vect{u}_{k:K} } \sum_{\tau=k}^{K} \beta_{\tau} J_{\tau} ( {\vect{\theta}}_{\tau} ),$ we split the interval $[k, K )$ as $[k, k+1 )$ and $[k+1, K )$ to write $$V^{ (* )}_{k}= min_{{\vect{\theta}}_{\tau} \in \Omega} \big[ \beta_k J_{k} ( {\vect{\theta}}_{k} )\big] + min_{\vect{u}_{k+1:K} }\big[ \sum_{\tau=k+1}^{K} \beta_{\tau} J_{\tau} ( {\vect{\theta}}_{\tau} ) \big].$$
$V_{k}= \sum_{\tau=k}^{K} \beta_{\tau} J_{\tau} ({\vect{\theta}}_{\tau} )$ provides $\sum_{\tau=k+1}^{K} \beta_{\tau} J_{\tau} ( {\vect{\theta}}_{\tau} )$ is $V_{k+1}$ therefore $min_{\vect{u}_{k+1:K} }\big[ \sum_{\tau=k+1}^{K} \beta_{\tau} J_{\tau} ( {\vect{\theta}}_{\tau} ) \big]$ is $V^{ (* )}_{k+1}.$ We then achieve
$$V^{ (* )}_{k} = min_{{\vect{\theta}}_{k} \in \Omega} \big[ \beta_k J_{k} ( {\vect{\theta}}_{k} ) + V^{ (* )}_{k+1} \big].$$
Since the minimization is with respect to $k$ now, the terms in $k+1$ can be pulled into of the bracket without any change to the minimization problem. We then approximate $V^{ (* )}_{k+1}$ using the information provided at $k.$ Since $V^{ (* )}_{k+1}$ is a function of $\vect{y}_{k},$ which is then a function of $ (k, \vect{x}_{k}, {\vect{\theta}}_{k} ),$ and all changes in $\vect{y}_{k}$ can be summarized through $ (k, \vect{x}_{k}, {\vect{\theta}}_{k} ).$ Therefore, a Taylor series of $V^{ (* )}_{k+1}$ around $ (k, \vect{x}_{k}, {\vect{\theta}}_{k} )$ provides
\begin{equation}
\begin{aligned}
&V^{ (* )}_{k+1} = V^{ (* )}_{k} + \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}_{k} \rangle &\\ &+ \langle \nabla_{\vect{x}_{k}} V_{k}^{ (* )}, \Delta \vect{x}_{k} \rangle + \langle \nabla_{k} (V^{ (* )}_{k} ), \Delta k \rangle + \cdots, &
\end{aligned}
\label{eq_a_1_5}
\end{equation}
where $\cdots$ summarizes all the higher order terms. As $k \in \mathbb{N}$ and $\langle \nabla_{k} (V^{ (* )}_{k} ), \Delta k \rangle$ represents the first difference in $V^{ (* )}_{k}$ hitherto denoted by $\Delta V^{ (* )}_{k}.$ We therefore achieve
\begin{equation}
\begin{aligned}
V^{ (* )}_{k+1} &= V^{ (* )}_{k} + \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )}, \Delta~\vect{\theta}_{k} \rangle \\
&+ \langle \nabla_{\vect{x}_{k}} V_{k}^{ (* )}, \Delta \vect{x}_{k} \rangle + \Delta V^{ (* )}_{k} + \cdots,
\end{aligned}
\label{eq_a_1_5}
\end{equation}
Substitute into the original equation to get
\begin{equation}
\begin{aligned}
& V^{ (* )}_{k} = min_{{\vect{\theta}}_{k} \in \Omega} \big[ \beta_k J_{k} ( {\vect{\theta}}_{k} ) \big] + \big ( V^{ (* )}_{k} + \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}_{k} \rangle & \\ &+ \langle \nabla_{\vect{x}_{k}} V_{k}^{ (* )}, \Delta \vect{x}_{k} \rangle +\Delta V^{ (* )}_{k} \big ) + \cdots,&
\end{aligned}
\label{eq_a_1_5}
\end{equation}
Cancel common terms and assume that the higher order terms~ ($\cdots$ ) are negligible to obtain
\begin{equation}
\begin{aligned}
\Delta V^{ (* )}_{k} = - min_{{\vect{\theta}}_{k} \in \Omega} \big[ \beta_k J_{k} ( {\vect{\theta}}_{k} ) + \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}_{k} \rangle +\langle \nabla_{\vect{x}_{k}} V_{k}^{ (* )}, \Delta \vect{x}_{k} \rangle \big].&
\end{aligned}
\label{eq_opt}
\end{equation}
which is a difference equation in $V^{ (* )}_{k}.$
\end{proof}
Note that $V^{ (* )}_{k}$ is the minima for the overarching CL problem and $\Delta V^{ (* )}_{k}$ represents the change in $V^{ (* )}_{k}$ upon introduction of a task~ (we hitherto refer to this as perturbations ). Zero perturbations $ (\Delta V^{ (* )}_{k}=0 )$ implies that the introduction of a new task does not impact our current solution; that is, the optimal solution on all previous tasks is optimal on the new task as well.
The solution of the CL problem can directly be obtained by solving Eq. \eqref{eq_opt} using all the available data. Thus, $ min_{{\vect{\theta}}_{k} \in \Omega} \big[ H (\Delta \vect{x}_{k}, \vect{\theta}_{k} ) \big] \quad \text{yields } \Delta V^{ (* )}_{k} \approx 0$ for $\beta > 0,$ with $H (\Delta \vect{x}_{k}, \vect{\theta}_{k} ) = \beta_k J_{k} ( {\vect{\theta}}_{k} ) + \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}_{k} \rangle + \langle \nabla_{\vect{x}_{k}} V_{k}^{ (* )}, \Delta \vect{x}_{k} \rangle.$ Essentially, minimizing $H (\Delta \vect{x}_{k}, \vect{\theta}_{k} )$ would minimize the perturbations introduced by any new task.
We simulate worst-case discrepancy by iteratively updating $\Delta \vect{x}_{k}$ through gradient ascent, thus maximizing generalization. Next, we minimize forgetting under maximum generalization by iteratively updating $\vect{\theta}_{k}$ through gradient descent. To formalize our idea, let us indicate the iteration index at $k$ by $i$ and write $\Delta \vect{x}_{k}$ as $\Delta \vect{x}^{ (i )}_{k}$ and ${\vect{\theta}}_{k}$ as ${\vect{\theta}}^{ (i )}_{k}$ with $H (\Delta \vect{x}_{k}, \vect{\theta}_{k} )$ as $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} )$~ (for simplicity of notation, we will denote $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} )$ as $H$ whenever convenient ). Towards these updates, we will first get an upper bound on $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} )$ and solve the upper bounding problem.
\begin{prop}
Let $k \in [0, K )$ and define $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} ) = \beta_k J_{k} ( {\vect{\theta}}^{ (i )}_{k} ) + \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}^{ (i )}_{k} \rangle + \langle \nabla_{\vect{x}^{ (i )}_{k}} V_{k}^{ (* )}, \Delta \vect{x}^{ (i )}_{k} \rangle$ assume that $ \nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )} \leq \nabla_{{\vect{\theta}}_{k}} J_{k} ({\vect{\theta}}^{ (i )}_{k} ).$ Then the following approximation is true:
\begin{equation}
\begin{aligned}
& H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} ) \leq \beta_k J_{k} (\vect{\theta}^{ (i )}_{k} ) + (J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} ) - J_{k} (\vect{\theta}^{ (i )}_{k} ) ) + ( J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} ) - J_{k} (\vect{\theta}^{ (i )}_{k} ) ),&
\end{aligned}
\label{eq_a_1_6}
\end{equation}
where $\beta_k \in \mathbb{R} \cup [0,1], \forall k$ and $\zeta \in \mathbb{N}$ and $J_{k+\zeta}$ indicates $\zeta$ updates on $\Delta \vect{x}^{ (i )}_{k}$ and $\vect{\theta}^{ (i+\zeta )}_{k}$ indicates $\zeta$ updates on ${\vect{\theta}}^{ (i )}_{k}.$
\label{prop2}
\end{prop}
\begin{proof}
Consider $ H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} ) = \beta_k J_{k} ( {\vect{\theta}}^{ (i )}_{k} ) + \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}^{ (i )}_{k} \rangle + \langle \nabla_{\vect{x}^{ (i )}_{k}} V_{k}^{ (* )}, \Delta \vect{x}^{ (i )}_{k} \rangle.$ Assuming $\nabla_{{\vect{\theta}}_{k}} V_{k}^{ (* )} \leq \nabla_{{\vect{\theta}}_{k}} J_{k} ({\vect{\theta}}^{ (i )}_{k} )$ we may write through finite difference approximation as
\begin{equation}
\begin{aligned}
\langle \nabla_{{\vect{\theta}}^{ (i )}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}^{ (i )}_{k} \rangle &\leq& \langle \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} ({\vect{\theta}}^{ (i )}_{k} ) , \Delta {\vect{\theta}}^{ (i )}_{k} \rangle,\\
&\leq& (J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} ) - J_{k} (\vect{\theta}^{ (i )}_{k} ) )
\end{aligned}
\label{eq_a_1_6}
\end{equation}
Similarly, we may write
\begin{equation}
\begin{aligned}
\langle \nabla_{\vect{x}_{k}} V_{k}^{ (* )}, \Delta \vect{x}_{k} \rangle &\leq&
\langle \nabla_{\vect{x}_{k}} J_{k} ({\vect{\theta}}^{ (i )}_{k} ), \Delta \vect{x}_{k} \rangle,\\
&\leq& ( J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} ) - J_{k} (\vect{\theta}^{ (i )}_{k} ) ).
\end{aligned}
\label{eq_a_1_6}
\end{equation}
Upon substitution, we have our result:
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} ) \leq \beta_k J_{k} (\vect{\theta}^{ (i )}_{k} ) + (J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} ) - J_{k} (\vect{\theta}^{ (i )}_{k} ) ) + (J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} ) - J_{k} (\vect{\theta}^{ (i )}_{k} ) ).
\end{aligned}
\label{eq_a_1_6}
\end{equation}
\end{proof}
Our cost to be analyzed will be given as
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} ) = \beta_k J_{k} ( {\vect{\theta}}^{ (i )}_{k} ) + \langle \nabla_{{\vect{\theta}}^{ (i )}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}^{ (i )}_{k} \rangle + \langle \nabla_{\vect{x}^{ (i )}_{k}} V_{k}^{ (* )}, \Delta \vect{x}^{ (i )}_{k} \rangle.
\end{aligned}
\label{eq_approx}
\end{equation}
and use this definition of $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} )$ from here on.
\section{Main results}
\begin{figure}
\includegraphics[width = \columnwidth]{Figures/Proof_Il.png}
\caption{Illustration of proofs. $\Delta \vect{x}$~ (player~1 ) is the horizontal axis and the vertical axis indicates $\vect{\theta}$~ (player~2 ) where the curve indicates H. If we start from red circle for player~1~ (player~2 is fixed at the blue circle ) H is increasing~ (goes from a grey circle to a red asterisk ) with player 1 reaching the red star. Next, start from the blue circle~ ($\vect{\theta}$ is at the red star ), the cost decreases.}
\label{fig:proof}
\end{figure}
We will define two compact sets $\Omega_{\theta}, \Omega_{x}$ and seek to show existence and stability of a saddle point~$ (\Delta \vect{x}^{ (i )}_{k} , \vect{\theta}_k^{ (i )} )$ for a fixed $k.$ To illustrate the theory, we refer to Fig.~\ref{fig:proof}, for each $k,$ the initial values for the two players are characterized by the pair $\{\vect{\theta}^{ (i )}_{k} \text{ (blue circle )}, \Delta \vect{x}^{ (i )}_{k} \text{ (red circle )} \},$ and $H ( \Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} )$ is indicated by the grey circle on the cost curve~ (the dark blue curve ). Our proofing strategy is as follows.
First, we fix $\vect{\theta}_k^{ (. )} \in \Omega_{\theta}$ and construct $\mathcal{M}_k= \{\vect{\theta}^{ (. )}_{k}, \Omega_{x} \},$ to prove that H is maximizing with respect to $\Delta \vect{x}^{ (i )}_{k}.$
\begin{lemma}
For each $k \in [0, K ),$ fix $\vect{\theta}_k^{ (. )} \in \Omega_{\theta}$ and construct $\mathcal{M}_k = \{\Omega_{x}, \vect{\theta}^{ (. )}_{k}\}$ with $\Omega_{\theta}, \Omega_{x}$ being the sets of all feasible $\vect{\theta}^{ (. )}_{k}$ and $\vect{x}^{ (i )}_{k}$ respectively. Define $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )$ as in Eq. \eqref{eq_approx} for $ (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) \in \mathcal{M}_k $ and consider $$ \Delta \vect{x}^{ (i+1 )}_{k} - \Delta \vect{x}^{ (i )}_{k} = \alpha_{k}^{ (i )} \nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) ) / \| \nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) \|^2.$$ Consider the assumptions $\nabla_{ \vect{x}^{ (i )}_{k}} V_{k}^{ (* )} \leq \nabla_{ \vect{x}^{ (i )}_{k}} J_{k}$ and $\langle \nabla_{\vect{x}^{ (i )}_{k}} J_k, \nabla_{\vect{x}^{ (i )}_{k}} J_k \rangle>0$, and let $\alpha_{k}^{ (i )} \rightarrow 0, i \rightarrow \infty.$ It follows that $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )$ converges asymptotically to a maximizer.
\label{lem:lem_max}
\end{lemma}
\begin{proof}
Fix $\vect{\theta}^{ (. )}_{k} \in \Omega_{\theta}$ and construct $\mathcal{M}_k$ such that $\mathcal{M}_k = \{\vect{\theta}^{ (. )}_{k}, \Omega_{x} \}$ which we call a neighborhood. Therefore, for $ (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (. )}_{k} ), (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) \in \mathcal{M}_k$ we may write a first-order Taylor series expansion of $H (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (. )}_{k} )$ around $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )$ as
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (. )}_{k} ) = H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) + \langle \nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ), \Delta \vect{x}^{ (i+1 )}_{k} - \Delta \vect{x}^{ (i )}_{k} \rangle .
\end{aligned}
\end{equation}
We substitute the update as $\alpha_{k}^{ (i )} \frac{\nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )}{\| \nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) \|^2}$ to get
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (. )}_{k} ) - H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )= \langle \nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ), \alpha_{k}^{ (i )} \frac{\nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )}{\| \nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )\|^2} \rangle.
\end{aligned}
\end{equation}
The derivative $\nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )$ can be written as
\begin{equation}
\begin{aligned}
& \nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) ) \leq \nabla_{\Delta \vect{x}^{ (i )}_{k}} \Big[ \beta_k J_{k} ( {\vect{\theta}}^{ (. )}_{k} ) + \langle \nabla_{{\vect{\theta}}^{ (. )}_{k}} V_{k}^{ (* )} , \Delta {\vect{\theta}}^{ (. )}_{k} \rangle & \\ &+ \langle \nabla_{\vect{x}^{ (i )}_{k}} V_{k}^{ (* )}, \Delta \vect{x}^{ (i )}_{k} \rangle \Big] = \nabla_{ \vect{x}^{ (i )}_{k}} V_{k}^{ (* )} \leq \nabla_{ \vect{x}^{ (i )}_{k}} J_{k}.&
\end{aligned}
\end{equation}
Substitution reveals
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (. )}_{k} ) - H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) = \alpha_{k}^{ (i )} \frac{ \langle \nabla_{\vect{x}^{ (i )}_{k}} J_{k}, \nabla_{\vect{x}^{ (i )}_{k}} J_{k} \rangle
}{\|\nabla_{\vect{x}^{ (i )}_{k}} J_{k} \|^2}
\end{aligned}
\end{equation}
for $\alpha_{k}^{ (i )} >0$; and under the assumption that $ \langle \nabla_{\vect{x}^{ (i )}_{k}} J_{k}, \nabla_{ \vect{x}^{ (i )}_{k}} J_{k} \rangle >0$ we obtain
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (. )}_{k} ) - H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) = \alpha_{k}^{ (i )} \frac{ \langle \nabla_{\vect{x}^{ (i )}_{k}} J_{k}, \nabla_{\vect{x}^{ (i )}_{k}} J_{k}
\rangle }{\| \nabla_{\vect{x}^{ (i )}_{k}} J_{k} \|^2 } \geq 0.
\end{aligned}
\end{equation}
Let $ B_{x} = \alpha_{k}^{ (i )} \frac{ \langle \nabla_{\vect{x}_{k}} J_{k}, \nabla_{ \vect{x}_{k}} J_{k} \rangle }{\| \nabla_{ \vect{x}^{ (i )}_{k}} J_{k} \|^2 } \leq \alpha_{k}^{ (i )}$ and therefore $0 \leq H (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (. )}_{k} ) - H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )\leq \alpha^{ (i )}_{k}.$ We therefore have $H (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (. )}_{k} ) - H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) \geq 0$ and $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )$ is maximizing with respect to $\Delta \vect{x}^{ (i )}_{k}.$ Furthermore, under the assumption that $\alpha^{ (i )}_{k} \rightarrow 0, k \rightarrow \infty$, we have $H (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (. )}_{k} ) - H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) \rightarrow 0$ asymptotically and we have our result.
\end{proof}
Similarly, we fix $\Delta \vect{x}^{ (. )}_{k} \in \Omega_x$ and construct $\mathcal{N}_k= \{ \Omega_{\theta}, \Delta \vect{x}^{ (. )}_{k} \},$ to prove that $H$ is minimizing with respect to $\vect{\theta}^{ (i )}_{k}.$
\begin{lemma}
For each $k \in [0, K ),$ fix $\Delta \vect{x}^{ (. )}_{k} \in \Omega_{x}$ and construct $\mathcal{N}_k= \{\Delta \vect{x}^{ (. )}_{k}, \Omega_{\theta} \}.$ Then for any $ (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) \in \mathcal{N}_k$ define $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} )$ as in Eq. \eqref{eq_approx} with Proposition.~\ref{prop2} being true and let $\vect{\theta}^{ (i+1 )}_{k} - \vect{\theta}^{ (i )}_{k} = -\alpha_{k}^{ (i )} \nabla_{{\vect{\theta}}^{ (i )}_{k}} H (\Delta \vect{x}^{ (. )}_{k}, \vect{\theta}^{ (i )}_{k} ) )$. Assume that $\| \nabla_{\vect{\theta}^{ (i )}_{k}} J_{k} ( {\vect{\theta}}^{ (i+\zeta )}_{k} )\| \leq L_1$ and $\| \nabla_{\vect{\theta}^{ (i )}_{k}} J_{k+\zeta} ( {\vect{\theta}}^{ (i )}_{k} )\| \leq L_2$ and let $\alpha_{k}^{ (i )} \rightarrow 0, i \rightarrow \infty.$ Then $\vect{\theta}^{ (i )}_{k}$ converges to a local minimizer.
\label{lem:lem_min}
\end{lemma}
\begin{proof}
First, we fix $\Delta \vect{x}^{ (. )}_k \in \Omega_{x}$ and construct $\mathcal{N}_k= \{ \Omega_{\theta} , \Delta \vect{x}^{ (. )}_{k} \}.$ For any $ (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ), (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) \in\mathcal{N}_k$ we write a first-order Taylor series expansion of $H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} )$ around $H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} )$ to write
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) = H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ) + \langle \nabla_{\vect{\theta}^{ (i )}_{k}} H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ), \vect{\theta}^{ (i+1 )}_{k} - \vect{\theta}^{ (i )}_{k} \rangle .
\end{aligned}
\end{equation}
We then substitute
$\vect{\theta}^{ (i+1 )}_{k} - \vect{\theta}^{ (i )}_{k} = -\alpha_{k}^{ (i )} \nabla_{{\vect{\theta}}^{ (i )}_{k}} H (\Delta \vect{x}^{ (. )}_{k}, \vect{\theta}^{ (i )}_{k} ) $ to get
\begin{equation}
\begin{aligned}
&H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ) = -\alpha_{k}^{ (i )}\langle \nabla_{\vect{\theta}^{ (i )}_{k}} H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ), \nabla_{{\vect{\theta}}^{ (i )}_{k}} H (\Delta \vect{x}^{ (. )}_{k}, \vect{\theta}^{ (i )}_{k} )\rangle.&
\end{aligned}
\end{equation}
Following Proposition 2, the derivative $\nabla_{{\vect{\theta}}^{ (i )}_{k}} H (\Delta \vect{x}^{ (. )}_{k}, \vect{\theta}^{ (i )}_{k} )$ can be written as
\begin{equation}
\begin{aligned}
\nabla_{{\vect{\theta}}^{ (i )}_{k}} H (\Delta \vect{x}^{ (. )}_{k}, \vect{\theta}^{ (i )}_{k} )\leq \nabla_{\vect{\theta}_{k}} [\beta_k J_{k} ( {\vect{\theta}}^{ (i )}_{k} ) + (J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} ) - J_{k} (\vect{\theta}^{ (i )}_{k} ) ) + (J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} ) - J_{k} (\vect{\theta}^{ (i )}_{k} ) )]
\end{aligned}
\end{equation}
Simplification reveals
\begin{equation}
\begin{aligned}
\nabla_{{\vect{\theta}}^{ (i )}_{k}} H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ) )\leq \nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 ) J_{k} ( {\vect{\theta}}^{ (i )}_{k} ) + \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} ) + \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} ).
\end{aligned}
\end{equation}
Substitution therefore provides
\begin{align}
&H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ) \nonumber\\ &\leq
- \alpha_{k}^{ (i )} \langle \nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 ) J_{k} ( {\vect{\theta}}^{ (i )}_{k} ) + \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} ) + \nabla_{\vect{\theta}^{ (i )}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} ), \nonumber
\\ & \nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 ) J_{k} ( {\vect{\theta}}^{ (i )}_{k} ) + \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} ) + \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} ) \rangle .
\end{align}
Opening the square with Cauchy's inequality provides
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} )
&\leq& - \alpha_{k}^{ (i )} \Big[ \|\nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 ) J_{k} ( {\vect{\theta}}_{k} ) \|^2 \\
&+& \| \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} )\|^2
+ \| \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} )\|^2
\\
&+& 2\|\nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 ) J_{k} ( {\vect{\theta}}_{k} ) \| \| \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} )\| \\
&+& 2\| \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} )\|\| \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} )\|
\\
&+& 2\|\nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 ) J_{k} ( {\vect{\theta}}_{k} ) \| \| \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} )\| \Big].
\end{aligned}
\end{equation}
We simplify with Young's inequality to achieve
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} )
&\leq& - \alpha_{k}^{ (i )} \Big[ \|\nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 )J_{k} ( {\vect{\theta}}_{k} ) \|^2 \\
&+& \| \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} )\|^2
\\ &+& \| \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} )\|^2
\\ &+& \|\nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 ) J_{k} ( {\vect{\theta}}_{k} ) \|^2 \\ &+& \| \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} )\|^2 \\
&+& \| \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} )\|^2
\\ &+& \| \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} )\|^2
\\ &+& \|\nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 ) J_{k} ( {\vect{\theta}}_{k} ) \|^2
\\ &+& \| \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} )\|^2 \Big].
\end{aligned}
\end{equation}
Further simplification results in
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} )
&\leq& - \alpha_{k}^{ (i )} \Big[ 3 \|\nabla_{{\vect{\theta}}^{ (i )}_{k}} (\beta_k-2 ) J_{k} ( {\vect{\theta}}_{k} ) \|^2 \\
&+& 3 \| \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} )\|^2
+ 3 \| \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} )\|^2 \Big].
\end{aligned}
\end{equation}
With the assumption that $\| \nabla_{{\vect{\theta}}^{ (i )}_{k}} J_{k} (\vect{\theta}^{ (i+\zeta )}_{k} )\| \leq L_1$ and $\| \nabla_{\vect{\theta}_{k}} J_{k+\zeta} (\vect{\theta}^{ (i )}_{k} )\|\leq L_2,$, we may write
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ) &\leq& -\alpha_{k}^{ (i )} B_{\theta},
\end{aligned}
\end{equation}
where $B_{\theta} = \Bigg[ ( (\sqrt{3}\beta_k-2\sqrt{3} )^2 + 3 )L_1^2 + 3L_2^2\Bigg].$ Assuming that $\alpha_{k}^{ (i )}$ is chosen such that $\alpha^{ (i )}_{k} \rightarrow 0$, we obtain $H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ) \rightarrow 0$ as $i \rightarrow \infty$ and $H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i )}_{k} ) < 0.$ Therefore H converges to a local minimizer.
\end{proof}
From here on, we will define our cost function as H whereever convinient for simplicity of notations. Since, for any k, there exists a local maximizer $\Delta \vect{x}_k^{ (* )} \in \Omega_x,$ we may define $\mathcal{N}^{ (* )}_k = \{\Delta \vect{x}_k^{ (* )}, \Omega_\theta\}$ where the set $\Omega_{x}$ is comprised of a local maximizer $\Delta \vect{x}_k^{ (* )}$
\begin{lemma}
For any $k \in [0, K )$, let $\vect{\theta}^{ (* )}_{k} \in \Omega_\theta,$ be the minimizer of $H$ according to Lemma \ref{lem:lem_min} and define $\mathcal{M}^{ (* )}_k = \{\Omega_{x}, \vect{\theta}^{ (* )}_{k}\}.$ Then for $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ), (\Delta \vect{x}_k^{ (i )}, \vect{\theta}^{ (* )}_k ) \in \mathcal{M}^{ (* )}_k,$ $H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ) \geq H (\Delta \vect{x}_k^{ (i )}, \vect{\theta}^{ (* )}_k ),$ where $\Delta \vect{x}_k^{ (* )}$ is a maximizer for H according to Lemma. \ref{lem:lem_max}.
\label{lem:lem_max_opt}
\end{lemma}
\begin{proof}
By Lemma \ref{lem:lem_min}, for each $k \in [0, K ),$ there exists a minimizer $\vect{\theta}^{ (* )}_{k} \in \Omega_\theta$ such that $\mathcal{M}^{ (* )}_k = \{ \Omega_{x}, \vect{\theta}^{ (* )}_{k} \}.$ Therefore by Lemma \ref{lem:lem_max}, $H (\Delta \vect{x}^{ (i+1 )}_{k}, \vect{\theta}^{ (* )}_{k} ) - H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (* )}_{k} ) \geq 0$ for $ (\Delta \vect{x}_k^{ (i+1 )}, \vect{\theta}^{ (* )}_k ), (\Delta \vect{x}_k^{ (i )}, \vect{\theta}^{ (* )}_k ) \in \mathcal{M}^{ (* )}_k.$ Let $\Delta \vect{x}_k^{ (* )} \in \Omega_x $ be the converging point according to Lemma \ref{lem:lem_max}. Then, for $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ), (\Delta \vect{x}_k^{ (i )}, \vect{\theta}^{ (* )}_k ) \in \mathcal{M}^{ (* )}_k$ a $H (\Delta \vect{x}^{ (* )}_{k}, \vect{\theta}^{ (* )}_{k} ) - H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (* )}_{k} ) \geq 0.$ by Lemma~\ref{lem:lem_max} which provides the result.
\end{proof}
\begin{lemma}
For any $k \in [0, K )$, let $\Delta \vect{x}_k^{ (* )} \in \Omega_x,$ be the maximizer of $H$ according to Lemma \ref{lem:lem_max} and define $\mathcal{N}^{ (* )}_k = \{\Delta \vect{x}^{ (* )}_{k}, \Omega_{\theta}\}.$ Then for $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ), (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ) \in \mathcal{N}^{ (* )}_k,$ $H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ) \leq H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ),$ where $\vect{\theta}^{ (* )}_{k}$ is a minimizer for H according to Lemma. \ref{lem:lem_max}.
\label{lem:lem_min_opt}
\end{lemma}
\begin{proof}
By Lemma \ref{lem:lem_max}, for each $k \in [0, K ),$ there exists a maximizer $\Delta \vect{x}_k^{ (* )} \in \Omega_x,$ such that $\mathcal{N}^{ (* )}_k = \{\Delta \vect{x}^{ (* )}_{k}, \Omega_{\theta}\}.$ Therefore by Lemma \ref{lem:lem_min}, $H (\Delta \vect{x}^{ (* )}_{k}, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (* )}_{k}, \vect{\theta}^{ (i )}_{k} ) \leq 0$ for $ (\Delta \vect{x}^{ (* )}_{k}, \vect{\theta}^{ (i+1 )}_{k} ), (\Delta \vect{x}^{ (* )}_{k}, \vect{\theta}^{ (i )}_{k} ) \in \mathcal{N}^{ (* )}_k.$ Let $\vect{\theta}_k^{ (* )} \in \Omega_{\theta} $ be the converging point according to Lemma \ref{lem:lem_min}. Then, for $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ), (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ) \in \mathcal{M}^{ (* )}_k,$ $H (\Delta \vect{x}^{ (* )}_{k}, \vect{\theta}^{ (* )}_{k} ) - H (\Delta \vect{x}^{ (* )}_{k}, \vect{\theta}^{ (i )}_{k} ) \leq 0$ by Lemma~\ref{lem:lem_min} which provides the result.
\end{proof}
Next, we prove that the union of the two neighborhoods for each k $\mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k,$ is non-empty.
\begin{lemma}
For any $k \in [0, K )$, let $\vect{\theta}^{ (* )}_{k} \in \Omega_\theta,$ be the minimizer of $H$ according to Lemma \ref{lem:lem_min} and define $\mathcal{M}^{ (* )}_k = \{\Omega_{x}, \vect{\theta}^{ (* )}_{k}\}.$ Similarly, let $\Delta \vect{x}_k^{ (* )} \in \Omega_x,$ be the maximizer of $H$ according to Lemma \ref{lem:lem_max} and define $\mathcal{N}^{ (* )}_k = \{\Delta \vect{x}^{ (* )}_{k}, \Omega_{\theta}\}.$ Then, $\mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$ is nonempty.
\label{lem:lem_nonE}
\end{lemma}
\begin{proof}
Let $\mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$ be empty. Then, for any $ (\Delta \vect{x}^{ (i+1 )}_k, \vect{\theta}^{ (. )}_{k} ), (\Delta \vect{x}^{ (i )}_k,\vect{\theta}^{ (. )}_{k} ) \in \mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$, $H (\Delta \vect{x}^{ (i+1 )}_k, \vect{\theta}^{ (. )}_{k} ) - H (\Delta \vect{x}^{ (i )}_k,\vect{\theta}^{ (. )}_{k} )$ is undefined because the union is empty. This contradicts Lemma \ref{lem:lem_min_opt}. Similarly, $H (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ) - H (\Delta \vect{x}^{ (. )}_k,\vect{\theta}^{ (i )}_{k} )$ for $ (\Delta \vect{x}^{ (. )}_k, \vect{\theta}^{ (i+1 )}_{k} ), (\Delta \vect{x}^{ (. )}_k,\vect{\theta}^{ (i )}_{k} ) \in \mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$ also contradicts Lemma \ref{lem:lem_max_opt}. Therefore, by contradiction, $\mathcal{M}_k \cup \mathcal{N}_k$ cannot be empty.
\end{proof}
\subsection{Final Results}
We are now ready to present the main results. We show that there exists an equilibrium point~ (Theorem~1 ) and that the equilibrium point is stable~ (Theorem~2 ).
\begin{thm}[Existence of an Equilibrium Point]
For any $k \in [0, K )$, let $\vect{\theta}^{ (* )}_{k} \in \Omega_\theta,$ be the minimizer of $H$ according to Lemma \ref{lem:lem_min_opt} and define $\mathcal{M}^{ (* )}_k = \{\Omega_{x}, \vect{\theta}^{ (* )}_{k}\}.$ Similarly, let $\Delta \vect{x}_k^{ (* )} \in \Omega_x,$ be the maximizer of $H$ according to Lemma \ref{lem:lem_max_opt} and define $\mathcal{N}^{ (* )}_k = \{\Delta \vect{x}^{ (* )}_{k}, \Omega_{\theta}\}.$ Further, let $\mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$ be nonempty according to Lemma.~\ref{lem:lem_nonE}, then $ (\Delta \vect{x}^{ (* )}_{k},\vect{\theta}^{ (* )}_{k} ) \in \mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$ is a local equilibrium point.
\end{thm}
\begin{proof}
By Lemma \ref{lem:lem_min_opt} we have at $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ), (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ) \in \mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$ that
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ) &\leq& H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ).
\end{aligned}
\end{equation}
Similarly, according to Lemma \ref{lem:lem_max_opt}, at $ (\Delta \vect{x}^{ (* )}_k, \vect{\theta}^{ (* )}_{k} ), (\Delta \vect{x}^{ (i )}_k, \vect{\theta}^{ (* )}_{k} ) \in \mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$ we have
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (* )}_k, \vect{\theta}^{ (* )}_{k} ) \geq H (\Delta \vect{x}^{ (i )}_k, \vect{\theta}^{ (* )}_{k} ).
\end{aligned}
\end{equation}
Putting these inequalities together, we get
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ) \geq H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ) \geq H (\Delta \vect{x}_k^{ (i )}, \vect{\theta}^{ (* )}_k ),
\end{aligned}
\end{equation}
which is the saddle point condition, and therefore $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{(*)}_k )$ is a local equilibrium point in $\mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k.$
\end{proof}
According to the preceeding theorem, there is at least one equillibrium point for the game summarized by $H$.
\begin{thm} [Stability of the Equilibrium Point]
For any $k \in [0,K)$, $\Delta \vect{x}^{ (i )}_{k} \in \Omega_x$ and $\vect{\theta}^{ (i )}_{k} \in \Omega_\theta$ be the initial values for $\Delta \vect{x}^{ (i )}_{k}$ and $\vect{\theta}^{ (i )}_{k}$ respectively. Define $\mathcal{M}_k = \{\Omega_{x}, \Omega_{\theta}\}$ with $H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (i )}_{k} )$ given by Proposition \ref{prop2}. Let $\Delta \vect{x}^{ (i+1 )}_{k} - \Delta \vect{x}^{ (i )}_{k} = \alpha_{k}^{ (i )}\times (\nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) )/\| \nabla_{\Delta \vect{x}^{ (i )}_{k}} H (\Delta \vect{x}^{ (i )}_{k}, \vect{\theta}^{ (. )}_{k} ) \|^2 )$ and $\vect{\theta}^{ (i+1 )}_{k} - \vect{\theta}^{ (i )}_{k} = -\alpha_{k}^{ (i )}\times \nabla_{{\vect{\theta}}^{ (i )}_{k}} H (\Delta \vect{x}^{ (. )}_{k}, \vect{\theta}^{ (i )}_{k} ).$ Let the existence of an equilibrium point be given by Theorem~1, then as a consequence of Lemma \ref{lem:lem_max} and \ref{lem:lem_min} $ (\Delta \vect{x}^{ (* )}_{k},\vect{\theta}^{ (* )}_{k} ) \in \mathcal{M}_k$ is a stable equilibrium point for $H$.
\label{thm:thm_st}
\end{thm}
\begin{proof}
Consider now the order of plays by the two players. By Lemma~\ref{lem:lem_max}, a game starting at $ (\Delta \vect{x}_k^{ (i )}, \vect{\theta}^{ (i )}_k ) \in \mathcal{M}_k$ will reach $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k )$ which is a maximizer for $H.$ Now, define $\mathcal{N}_k = (\Delta \vect{x}_k^{ (* )}, \Omega_{\theta} ) \subset \mathcal{M}_k$ then a game starting at $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ) \in \mathcal{N}_k$ will converge to $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ) \in \mathcal{N}_k$ according to Lemma~\ref{lem:lem_min}. Since, $\mathcal{N}_k \subset \mathcal{M}_k,$ our result follows.
\end{proof}
\section{Experimental Details}
Much of this information is a repetition of details provided in \cite{Hsu18_EvalCL,vandeven2019generative}.
\begin{enumerate}
\item \textit{Incremental Domain Learning~ (IDL ):} Incremental domain refers to the scenario when each new task introduces changes in the marginal distribution of the inputs. This scenario has been extensively discussed in the domain adaptation literature, where this shift in domain is typically referred to as ``non-stationary data distribution" or domain shift. Overall, we aim to transfer knowledge from the old task to a new task where each task can be different in the sense of their marginal distribution.
\item \textit{Incremental Class Learning~ (ICL ):} In this scenario, each task contains an exclusive subset of classes. The number of output nodes in a model equals the number of total classes in the task sequence. For instance, tasks could be constructed by using exactly one class from the MNIST data set where we aim to transfer knowledge from one class to another.
\item \textit{Incremental Task Learning~ (ITL ):} In this setup, the output spaces are disjoint between tasks[ for example, the previous task can be a classification problem of five classes, while the new task can be a regression. This scenario is the most generic and allows for the tasks to be defined arbitrarily. For each tasks, a model requires task-specific identifier$t$.
\end{enumerate}
\textbf{Split-MNIST}
For split-MNIST, the original MNIST-data set is split into five partitions where each partition is a two-way classification. We pick 60000 images for training (6000 per digit ) and 10000 images for test, i.e. (1000 per digit ). For the incremental task learning in the split-MNIST experiment, the ten digits are split into five two-class classification tasks (the model has five output heads, one for each task ) and the task identity (1 to 5 ) is given for test. For the incremental class learning setup, we require the model to make a prediction over all classes (digits 0 to 9 ). For the incremental domain learning, the model always predicts over two classes.
\textbf{Permuted-MNIST}
For permuted-MNIST, we permute the pixels in the MNIST data to create tasks where each task is a ten-way classification. The three CL scenarios that are generated for the permuted-MNIST are similar to the Split-MNIST data set except for the idea that the different tasks are now generated by applying random pixel permutations to the images. For incremental task learning, we use a multi-output strategy, and each task is attached to a task identifier. For incremental domain and class, we use a single output strategy and each task as one where one of the 10 digits are predicted. In incremental class learning, for each new task 10 new classes are generated by permuting the MNIST data set. For incremental task and domain, we use a total of 10 tasks whereas for incremental classes, we generate a total of 100 tasks.
\textbf{Network Architecture}
We keep our architecture identical to what is provided in \cite{Hsu18_EvalCL,vandeven2019generative}. The loss function is categorical cross-entropy for classification. All models were trained for $2$ epochs per task with a minibatch size of $128$ using the Adam optimizer ($\beta_1 = 0.9$, $\beta_2 = 0.999$, learning rate$= 0.001$ ) as the default. For BCL, the size of the buffer (i.e., a new task array ) $\vect{\mathcal{D}} _N (k )$ and a task memory array (samples from all the previous tasks ) $\vect{\mathcal{D}} _{P} (k ) )$ is kept equivalent to naive rehearsal and other memory-driven approaches such as GEM and MER~ ($16,000$ samples ).
\textbf{Comparison Methods -- Baseline Strategies}
Additional details can be found from \cite{Hsu18_EvalCL,vandeven2019generative}
\begin{enumerate}
\item A sequentially-trained neural network with different optimizers such as SGD, Adam~\cite{kingma2014adam}, and Adagrad~\cite{duchi2011adaptive}.
\item A standard $L_2-$regularized neural network where each task is shown sequentially.
\item Naive rehearsal strategy (experience replay ) where a replay buffer is initialized and data corresponding to all the previous tasks are stored. The buffer size is chosen to match the space overhead of online EWC and SI.
\end{enumerate}
\textbf{Comparison Method-CL}
We compared the following CL methods:
\begin{enumerate}
\item \textbf{EWC~\cite{kirkpatrick2017overcoming} / Online EWC~\cite{schwarz2018progress} / SI~\cite{zenke2017continual}}: For these methods, a regularization term is added to the loss, with a hyperparameter used to control the regularization strength such that: $L (total ) = L (current ) + \lambda L (regularization ).$ $\lambda$ is set through a hyperparameter.
\item \textbf{LwF~\cite{LwF} / DGR~\cite{DGR}} Here, we set the loss to be $L (total ) = \alpha L (current ) + (1 − \alpha )L (replay )$ where hyperparameter $\alpha$ is chosen according to how many tasks have been seen by the model.
\item For \textbf{RtF}~\cite{vandeven2019generative}, MAS~\cite{aljundi2018memory}, GEM\cite{lopez2017gradient} and MER\cite{riemer2018learning}, we refer to the respective publication for details.
\end{enumerate}
Additional details about the experiments can be found in \cite{Hsu18_EvalCL} as our paper retains their hyper-parameters and the experimental settings.
\bibliographystyle{plain}
\section{Introduction}
In continual learning~(CL), we incrementally adapt a model to learn tasks~(defined according to the problem at hand) observed sequentially. CL has two main objectives: maintain long-term memory~(remember previous tasks) and navigate new experiences continually~(quickly adapt to new tasks). An important characterization of these objectives is provided by the stability-plasticity dilemma~\cite{carpenter1987massively}, where the primary challenge is to balance network stability~(preserve past knowledge; minimize catastrophic forgetting) and plasticity~(rapidly learn from new experiences; generalize quickly). This balance provides a natural objective for CL: \textit{balance forgetting and generalization.}
Traditional CL methods either minimize catastrophic forgetting or improve quick generalization but do not model both. For example, common solutions to the catastrophic forgetting issue include (1) representation-driven approaches~\cite{yoon2017lifelong, ke2020continual}, (2) regularization approaches~\cite{kirkpatrick2017overcoming, aljundi2018memory, mirzadeh2020understanding, farajtabar2019orthogonal, yin2020optimization, yin2020optimization, jung2020continual, pan2020continual, chaudhry2020continual, titsias2019functional}, and (3) memory/experience replay~\cite{lin1992self, lopez2017gradient, chaudhry2019continual, chaudhry2019tiny, fini2020online}. Solutions to the generalization problem include representation-learning approaches~(matching nets~\cite{vinyals2016matching}, prototypical networks~\cite{snell2017prototypical}, and metalearning approaches~\cite{finn2017model, finn2019online, Caccia2020OnlineFA, yao2020don}). More recently, several approaches \cite{nagabandi2019deep,farajtabar2019orthogonal,DBLP:journals/corr/abs-1806-06928, joseph2020metaconsolidation, yin2020optimization, ebrahimi2020adversarial} have been introduced that combine methods designed for quick generalization with frameworks designed to minimize forgetting.
The aforementioned CL approaches naively minimize a loss function~(combination of forgetting and generalization loss) but do not explicitly account for the trade-off in their optimization setup. The first work to formalize this trade-off was presented in meta-experience replay~(MER)~\cite{riemer2018learning}, where the forgetting-generalization trade-off was posed as a gradient alignment problem. Although MER provides a promising methodology for CL, the balance between forgetting and generalization is enforced with several hyperparameters.
Therefore, two key challenges arise: (1)~lack of theoretical tools that study the existence~(\textit{under what conditions does a balance point between generalization and forgetting exists?}) and stability~(\textit{can this balance be realistically achieved?}) of a balance point and (2)~lack of a systematic approach to achieve the balance point. We address these challenges in this paper.
We describe a framework where we first formulate CL as a sequential decision-making problem and seek to minimize a cost function summed over the complete lifetime of the model. At any time $k,$ given that the future tasks are not available, the calculation of the cost function becomes intractable. To circumvent this issue, we use Bellman's principle of optimality~\cite{bellman2015adaptive} and recast the CL problem to model the catastrophic forgetting cost on the previous tasks and generalization cost on the new task. We show that equivalent performance on an infinite number of tasks is not practical~(Lemma~1 and Corollary~1) and that tasks must be prioritized.
\begin{figure}[!h]
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width = \columnwidth]{Figures/MCL_Discrete.png}
\label{fig:ills}
\end{subfigure}\\
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width = \columnwidth ]{Figures/Method.png}
\label{fig:met}
\end{subfigure}
\caption{(left) Exemplary CL problem: the lifetime of the model can be split into three intervals. At $k =1$ we seek to recognize lions; at $k=2$ we seek to recognize both lions and cats; and at $k=3$ we seek to recognize cats, lions, and dogs. (right) Illustration of the proposed method: our methodology comprises an interplay between two players. The first player maximizes generalization by simulating maximum discrepancy between two tasks. The second player minimizes forgetting by adapting to maximum discrepancy.}
\label{fig:ills}
\end{figure}
To achieve a balance between forgetting and generalization, we pose the trade-off
as a saddle point problem where we designate one player for maximizing the generalization cost~(player 1) and another for minimizing the forgetting cost~(player 2). We prove mathematically that there exists at least one saddle point between generalization and forgetting for each new task~(Theorem~1). Furthermore, we show that this saddle point can be attained asymptotically~(Theorem~2) when player strategies are chosen as gradient ascent-descent. We then introduce balanced continual learning~(BCL), a new algorithm to achieve this saddle point. In our algorithm~(see Fig.~\ref{fig:ills} for a description of BCL), the generalization cost is computed by training and evaluating the model on given new task data. The catastrophic forgetting cost is computed by evaluating the model on the task memory (previous tasks). We first maximize the generalization cost and then minimize the catastrophic forgetting cost to achieve the balance. We compare our approach with other methods such as elastic weight consolidation~(EWC)~\cite{kirkpatrick2017overcoming}, online EWC~\cite{schwarz2018progress}, and MER~\cite{riemer2018learning} on continual learning benchmark data sets~\cite{Hsu18_EvalCL} to show that BCL is better than or comparable to the state-of-the-art methods. Moreover, we also show in simulation that our theoretical framework is appropriate for understanding the continual learning problem. The contributions of this paper are (1) a theoretical framework to study the CL problem, (2) BCL, a method to attain balance between forgetting and generalization, and (3) advancement of the state of the art in CL.
\section{Problem Formulation}
We use $\vect{\mathbb{R}} $ to denote the set of real numbers and $\vect{\mathbb{N}} $ to denote the set of natural numbers. We use $\|.\|$ to denote the Euclidean norm for vectors and the Frobenius norm for matrices, while using bold symbols to illustrate matrices and vectors. We define an interval $[0, K), K \in \vect{\mathbb{N}} $ and let $p(\cT)$ be the distribution over all the tasks observed in this interval. For any $k \in [0, K) ,$ we define a parametric model $g(.)$ with ${\vect{y}}_{k} = g(\vect{x}_{k}; {\vect{\theta}}_{k})$, where ${\vect{\theta}}_{k}$ is a vector comprising all parameters of the model with $\vect{x}_{k} \in \cX_{k}$. Let $n$ be the number of samples and $m$ be the number of dimensions. Suppose a task at $k | k \in [0, K)$ is observed and denoted as $\cT_{k} : \cT_{k} \sim p(\cT)$, where $\cT_{k} =\{\cX_{k}, \ell_{k}\}$ is a tuple with $\cX_{k} \in \vect{\mathbb{R}} ^{n \times m}$ being the input data and $\ell_{k}$ quantifies the loss incurred by $\cX_{k}$ using the model $g$ for the task at $k$. We denote a sequence of ${\vect{\theta}}_{k}$ as $\vect{u}_{k:K} = \{{\vect{\theta}}_{\tau} \in \Omega_{\theta}, k \leq \tau \leq K \},$ with $\Omega_{\theta}$ being the compact~(feasible) set for the parameters. We denote the optimal value with a superscript ${(*)};$ for instance, we use ${\vect{\theta}}_k^{(*)}$ to denote the optimal value of ${\vect{\theta}}_{k}$ at task $k.$ In this paper we use balance point, equilibrium point, and saddle point to refer to the point of balance between generalization and forgetting. We interchange between these terms whenever convenient for the discussion. We will use $\nabla_{(j)} i$ to denote the gradient of $i$ with respect to $j$ and $\Delta i$ to denote the first difference in discrete time.
An exemplary CL problem is described in Fig.~\ref{fig:ills} where we address a total of $K=3$ tasks. To particularize the idea in Fig. \ref{fig:ills}, we define the cost~(combination of catastrophic cost and generalization cost) at any instant $k$ as
\[J_{k}({\vect{\theta} }_{k}) = \gamma_{k} \ell_{k} + \sum_{\tau = 0}^{k-1} \gamma_{\tau} \ell_{\tau},\]
where $\ell_{\tau}$ is computed on task $\cT_{\tau}$ with $\gamma_{\tau}$ describing the contribution of $\cT_{\tau}$ to this sum.
To solve the problem at $k$, we seek ${\vect{\theta} }_{k}$ to minimize $J_{k}({\vect{\theta} }_{k})$. Similarly, to solve the problem in the complete interval $[0,K]$, we seek a ${\vect{\theta} }_{k}$ to minimize $J_{k}({\vect{\theta} }_{k})$ for each $k \in [0,K].$ In other words we seek to obtain ${\vect{\theta} }_{k}$ for each task such that the cost $J_{k}({\vect{\theta} }_{k})$ is minimized. Therefore, the optimization problem for the overall CL problem~(overarching goal of CL) is provided as the minimization of the cumulative cost \[ V_{k}(\vect{u}_{k:K}) = \sum_{\tau=k}^{K} \beta_{\tau} J_{\tau}( {\vect{\theta}}_{\tau}) \] such that $V_k^{(*)},$ is given as \begin{equation} V_k^{(*)} = min_{\vect{u}_{k:K} } V_{k}(\vect{u}_{k:K}), \label{op_1} \end{equation} with $0 \leq \beta_{\tau} \leq 1$ being the contribution of $J_{\tau}$ and $\vect{u}_{k:K}$ being a weight sequence of length $K-k.$
Within this formulation, two parameters determine the contributions of tasks: $\gamma_{\tau}$, the contribution of each task in the past, and $\beta_{\tau}$, the contribution of tasks in the future. To successfully solve the optimization problem, $V_{k}(\vect{u}_{k:K})$ must be bounded and differentiable, typically ensured by the choice of $\gamma_{\tau}, \beta_{\tau}.$ Lemma~1~(full statement and proof in Appendix A) states that \textit{equivalent performance cannot be guaranteed for an infinite number of tasks}. Furthermore, Corollary~1~(full statement and proof in Appendix A) demonstrates that \textit{if the task contributions are prioritized, the differentiability and boundedness of $J_\tau({\vect{\theta}}_{\tau})$ can be ensured}. A similar result was proved in \cite{knoblauch2020optimal}, where a CL problem with infinite memory was shown to be NP-hard from a set theoretic perspective. These results~(both ours and in \cite{knoblauch2020optimal}) demonstrate that a CL methodology cannot provide perfect performance on a large number of tasks and that tasks must be prioritized.
Despite these invaluable insights, the data corresponding to future tasks~(interval $[k,K]$) is not available, and therefore $V_{k}(\vect{u}_{k:K})$ cannot be evaluated. The optimization problem in Eq.~\eqref{op_1} naively minimizes the cost~(due to both previous tasks and new tasks) and does not provide any explicit modeling of the trade-off between forgetting and generalization. Furthermore, $\vect{u}_{k:K},$ the solution to Eq. \eqref{op_1} is a sequence of parameters, and it is not feasible to maintain $\vect{u}_{k:K}$ for a large number of tasks. Because of these three issues, the problem is theoretically intractable in its current form.
We will first recast the problem using tools from dynamic programming~\cite{lewis2012optimal}, specifically Bellman's principle of optimality, and derive a difference equation that summarizes the complete dynamics for the CL problem. Then, we will formulate a two-player differential game where we seek a saddle point solution to balance generalization and forgetting.
\section{Dynamics of Continual Learning}
Let \[ V_k^{(*)} = min_{\vect{u}_{k:K} } \sum_{\tau=k}^{K} \beta_{\tau} J_{\tau}( {\vect{\theta}}_{\tau});\] the dynamics of CL~(the behavior of optimal cost with respect to $k$) is provided as
\begin{equation}
\begin{aligned}
\Delta V^{(*)}_{k} = - min_{{\vect{\theta}}_{k} \in \Omega_{\theta}} \big[ \beta_k J_{k}( {\vect{\theta}}_{k}) + \big( \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{(*)} , \Delta {\vect{\theta}}_{k} \rangle +\langle \nabla_{\vect{x}_{k}} V_{k}^{(*)}, \Delta \vect{x}_{k} \rangle \big)\big].\\
\end{aligned}
\label{eq_M_DES}
\end{equation}
The derivation is presented in Appendix A~(refer to Proposition~1). Note that $V^{(*)}_{k}$ is the minima for the overarching CL problem in Eq. \eqref{eq_M_DES}and $\Delta V^{(*)}_{k}$ represents the change in $V^{(*)}_{k}$ upon introduction of a new task~(we hitherto refer to this as perturbations). Zero perturbations $(\Delta V^{(*)}_{k}=0)$ implies that the introduction of a new task does not impact our current solution; that is, the optimal solution on all previous tasks is optimal on the new task as well. Therefore, the smaller the perturbations, the better the performance of a model on all tasks, thus providing our main objective: minimize the perturbations~($\Delta V^{(*)}_{k} $). In Eq. \ref{eq_M_DES}, $\Delta V^{(*)}_{k} $ is quantified by three terms: the cost contribution from all the previous tasks and the new task~$J_{k}( {\vect{\theta}}_{k});$ the change in the optimal cost due to the change in the parameters $\langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{(*)}, \Delta {\vect{\theta}}_{k} \rangle$; and the change in the optimal cost due to the change in the input (introduction of new task)~$\langle \nabla_{\vect{x}_{k}} V_{k}^{(*)}, \Delta \vect{x}_{k}\rangle$.
The first issue with the cumulative CL problem~(Eq. \eqref{op_1}) can be attributed to the need for information from the future. In Eq. \eqref{eq_M_DES}, all information from the future is approximated by using the data from the new and the previous tasks. Therefore, the solution of the CL problem can directly be obtained by solving Eq. \eqref{eq_M_DES} using all the available data. Thus, \[ min_{{\vect{\theta}}_{k} \in \Omega} \big[ H(\Delta \vect{x}_{k}, \vect{\theta}_{k}) \big] \quad \text{yields } \Delta V^{(*)}_{k} \approx 0 \] for $\beta > 0,$ with \[ H(\Delta \vect{x}_{k}, \vect{\theta}_{k}) = \beta_k J_{k}( {\vect{\theta}}_{k}) + \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{(*)} , \Delta {\vect{\theta}}_{k} \rangle + \langle \nabla_{\vect{x}_{k}} V_{k}^{(*)}, \Delta \vect{x}_{k} \rangle.\] Essentially, minimizing $H(\Delta \vect{x}_{k}, \vect{\theta}_{k})$ would minimize the perturbations introduced by any new task $k$.
In Eq. \eqref{eq_M_DES}, the first and the third term quantify generalization and the second term quantifies forgetting. A model exhibits generalization when it successfully adapts to a new task~(minimizes the first and the third term in Eq. \eqref{eq_M_DES}). The degree of generalization depends on the discrepancy between the previous tasks and the new task~(numerical value of the third term in Eq. \eqref{eq_M_DES}) and the worst-case discrepancy prompts maximum generalization. Quantification of generalization is provided by $\Delta \vect{x}_{k}$ that summarizes the discrepancy between subsequent tasks. However, $\Delta \vect{x}_{k} = \vect{x}_{k+1} - \vect{x}_{k}$, and $\vect{x}_{k+1}$ is unknown at $k.$ Therefore, we simulate worst-case discrepancy by iteratively updating $\Delta \vect{x}_{k}$ through gradient ascent in order to maximize $H(\Delta \vect{x}_{k}, \vect{\theta}_{k})$; thus maximizing generalization. However, large discrepancy increases forgetting, and worst-case discrepancy yields maximum forgetting. Therefore, once maximum generalization is simulated, minimizing forgetting~(update $\vect{\theta}_{k}$ by gradient descent) under maximum generalization provides the balance.
To formalize our idea, let us indicate the iteration index at $k$ by $i$ and write $\Delta \vect{x}_{k}$ as $\Delta \vect{x}^{(i)}_{k}$ and ${\vect{\theta}}_{k}$ as ${\vect{\theta}}^{(i)}_{k}$ with $H(\Delta \vect{x}_{k}, \vect{\theta}_{k})$ as $H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k})$~(for simplicity of notation, we will denote $H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k})$ as $H$ whenever convenient). Next, we write
\begin{equation}
\begin{aligned}
\centering
&\underset{{\vect{\theta}}^{(i)}_{k} \in \Omega_{\theta} }{min} \bigg[ H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k}) \bigg] = \underset{{\vect{\theta}}^{(i)}_{k} \in \Omega_{\theta}}{min} \big[ \beta_k J_{k}( {\vect{\theta}}^{(i)}_{k}) + \langle \nabla_{{\vect{\theta}}^{(i)}_{k}} V_{k}^{(*)} , \Delta {\vect{\theta}}^{(i)}_{k}\rangle + \langle \nabla_{\vect{x}^{(i)}_{k}} V_{k}^{(*)}, \Delta \vect{x}^{(i)}_{k} \rangle \big]& \\ &\leq \underset{{\vect{\theta}}^{(i)}_{k} \in \Omega_{\theta}}{min} \big[\beta_k J_{k}( {\vect{\theta}}^{(i)}_{k}) + \langle \nabla_{{\vect{\theta}}^{(i)}_{k}} V_{k}^{(*)} , \Delta {\vect{\theta}}^{(i)}_{k}\rangle + \underset{\Delta \vect{x}^{(i)}_{k} \sim p(\cT)}{max} \langle \nabla_{\vect{x}^{(i)}_{k}} V_{k}^{(*)}, \Delta \vect{x}^{(i)}_{k} \rangle \big] &\\ & \leq \underset{{\vect{\theta}}^{(i)}_{k} \in \Omega_{\theta} }{min} \quad \underset{\Delta \vect{x}^{(i)}_{k} \sim p(\cT)}{max} \big[ H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k})\big]. & \label{eq_op_2}
\end{aligned}
\end{equation}
In Eq. \eqref{eq_op_2}, we seek the solution pair $(\Delta \vect{x}_k^{(*)}, \vect{\theta}_k^{(*)} ) \in (\Omega_{\theta}, \Omega_{\Delta \vect{x}_k^{(*)}}),$ where $\Delta \vect{x}_k^{(*)}$ maximizes $H$~(maximizing player, player 1) while $\vect{\theta}_k^{(*)}$ minimizes $H$~(minimizing player, player 2) where $(\Omega_{\theta}, \Omega_{\Delta \vect{x}_k^{(*)}})$ are the feasible sets for $\Delta \vect{x}_k^{(i)}$ and $\vect{\theta}_k^{(i)}$ respectively. The solution is attained, and $(\Delta \vect{x}_k^{(*)}, \vect{\theta}_k^{(*)})$ is said to be the equilibrium point when it satisfies the following condition:
\begin{equation}
\begin{aligned}
H(\Delta \vect{x}_k^{(*)}, \vect{\theta}^{(i)}_{k})
\geq H(\Delta \vect{x}_k^{(*)}, \vect{\theta}_k^{(*)} ) \geq H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}_k^{(*)} ).
\end{aligned}
\label{eq_condition}
\end{equation}
\subsection{Theoretical Analysis\label{theory}}
\begin{wrapfigure}[17]{l}{0.51\textwidth}
\includegraphics[width = \columnwidth]{Figures/Proof_Il.png}
\caption{Illustration of the proofs. $\Delta \vect{x}$~(player~1) is the horizontal axis, and the vertical axis indicates $\vect{\theta}$~(player~2) where the curve indicates H. If we start from the red circle for player~1~(player~2 is fixed at the blue circle), H is increasing~(goes from a grey circle to a red asterisk) with player 1 reaching the red asterisk. Next, start from the blue circle~($\vect{\theta}$ is at the red asterisk), the cost decreases. }
\label{fig:proof}
\end{wrapfigure} With our formulation, two key questions arise: Does our problem setup have an equilibrium point satisfying Eq. \eqref{eq_condition}? and how can one attain this equilibrium point? We answer these questions with Theorems~1 and 2, respectively. Full statements and proofs are provided in Appendix~A.
To illustrate the theory, we refer to Fig.~\ref{fig:proof}, where the initial values for the two players are characterized by the pair $\{\vect{\theta}^{(i)}_{k} \text{(blue circle)}, \Delta \vect{x}^{(i)}_{k} \text{(red circle)} \}$ and the cost value at $\{\vect{\theta}^{(i)}_{k}, \Delta \vect{x}^{(i)}_{k} \}$ is indicated by $H( \Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k})$~(the grey circle on the cost curve~(the dark blue curve)). Our proofing strategy is as follows. First, we fix $\vect{\theta}_k^{(.)} \in \Omega_{\theta}$ and construct a neighborhood $\mathcal{M}_k= \{\Omega_{x}, \vect{\theta}^{(.)}_{k}\}.$ Within this neighborhood we prove in Lemmas~2 and 4 that if we search for $\Delta \vect{x}^{(i)}_{k}$ through gradient ascent, we can converge to a local maximizer, and $H$ is maximizing with respect to $\Delta \vect{x}^{(i)}_{k}.$
Second, we let $\Delta \vect{x}^{(.)}_{k} \in \Omega_x$ be fixed, and we search for $ \vect{\theta}^{(i)}_{k}$ through gradient descent. Under this condition, we demonstrate two ideas in Lemmas 3 and 5: (1) we show that $H$ is minimizing in the neighborhood $\mathcal{N}_k: \mathcal{N}_k= \{\Omega_{\theta} , \Delta \vect{x}^{(.)}_{k} \}$; and (2) we converge to the local minimizer in the neighborhood $\mathcal{N}_k.$ Third, in the union of the two neighborhoods $\mathcal{M}_k \cup \mathcal{N}_k,$ (proven to be nonempty according to Lemma~6), we show that there exists at least one local equilibrium point~(Theorem~1); that is, there is at least one balance point.
\begin{thm}[Existence of an Equilibrium Point]
For any $k \in [0,K]$, let $\vect{\theta}^{(*)}_{k} \in \Omega_\theta,$ be the minimizer of $H$ according to Lemma 5 and define $\mathcal{M}^{(*)}_k = \{\Omega_{x}, \vect{\theta}^{(*)}_{k}\}.$ Similarly, let $\Delta \vect{x}_k^{(*)} \in \Omega_x,$ be the maximizer of $H$ according to Lemma 4 and define $\mathcal{N}^{(*)}_k = \{\Delta \vect{x}^{(*)}_{k}, \Omega_{\theta}\}.$ Further, let $\mathcal{M}^{(*)}_k \cup \mathcal{N}^{(*)}_k$ be nonempty according to Lemma.~6, then $(\Delta \vect{x}^{(*)}_{k},\vect{\theta}^{(*)}_{k}) \in \mathcal{M}^{(*)}_k \cup \mathcal{N}^{(*)}_k$ is a local equilibrium point.
\end{thm}
\begin{proof}
By Lemma \ref{lem:lem_min_opt} we have at $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ), (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ) \in \mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$ that
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ) &\leq& H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ).
\end{aligned}
\end{equation}
Similarly, according to Lemma \ref{lem:lem_max_opt}, at $ (\Delta \vect{x}^{ (* )}_k, \vect{\theta}^{ (* )}_{k} ), (\Delta \vect{x}^{ (i )}_k, \vect{\theta}^{ (* )}_{k} ) \in \mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k$ we have
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}^{ (* )}_k, \vect{\theta}^{ (* )}_{k} ) \geq H (\Delta \vect{x}^{ (i )}_k, \vect{\theta}^{ (* )}_{k} ).
\end{aligned}
\end{equation}
Putting these inequalities together, we get
\begin{equation}
\begin{aligned}
H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ) \geq H (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ) \geq H (\Delta \vect{x}_k^{ (i )}, \vect{\theta}^{ (* )}_k ),
\end{aligned}
\end{equation}
which is the saddle point condition, and therefore $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{(*)}_k )$ is a local equilibrium point in $\mathcal{M}^{ (* )}_k \cup \mathcal{N}^{ (* )}_k.$ since $\mathcal{M}^{(*)}_k \cup \mathcal{N}^{(*)}_k$ be nonempty according to Lemma.~6.
\end{proof}
We next show that this equilibrium point is stable~(Theorem~2) under a sequential play. Specifically, we show that when player 1 plays first and player~2 plays second, we asymptotically reach a saddle point pair $(\Delta \vect{x}^{(*)}_{k},\vect{\theta}^{(*)}_{k})$ for $H.$ At this saddle point, both players have no incentive to move, and the game converges.
\begin{thm}[Stability of the Equilibrium Point]
For any $k \in [0,K]$, $\Delta \vect{x}^{(i)}_{k} \in \Omega_x$ and $\vect{\theta}^{(i)}_{k} \in \Omega_\theta$ be the initial values for $\Delta \vect{x}^{(i)}_{k}$ and $\vect{\theta}^{(i)}_{k}$ respectively. Define $\mathcal{M}_k = \{\Omega_{x}, \Omega_{\theta}\}$ with $H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k})$ given by Proposition 2. Let $\Delta \vect{x}^{(i+1)}_{k} - \Delta \vect{x}^{(i)}_{k} = \alpha_{k}^{(i)}\times (\nabla_{\Delta \vect{x}^{(i)}_{k}} H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(.)}_{k}) )/\| \nabla_{\Delta \vect{x}^{(i)}_{k}} H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(.)}_{k}) \|^2)$ and $\vect{\theta}^{(i+1)}_{k} - \vect{\theta}^{(i)}_{k} = -\alpha_{k}^{(i)}\times \nabla_{{\vect{\theta}}^{(i)}_{k}} H(\Delta \vect{x}^{(.)}_{k}, \vect{\theta}^{(i)}_{k}).$ Let the existence of an equilibrium point be given by Theorem~1, then, as a consequence of Lemmas 2 and 3, $(\Delta \vect{x}^{(*)}_{k},\vect{\theta}^{(*)}_{k}) \in \mathcal{M}_k$ is a stable equilibrium point for $H$ given
\label{thm:thm_st}.
\end{thm}
\begin{proof}
Consider now the order of plays by the two players. By Lemma~\ref{lem:lem_max}, a game starting at $ (\Delta \vect{x}_k^{ (i )}, \vect{\theta}^{ (i )}_k ) \in \mathcal{M}_k$ will reach $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k )$ which is a maximizer for $H.$ Now, define $\mathcal{N}_k = (\Delta \vect{x}_k^{ (* )}, \Omega_{\theta} ) \subset \mathcal{M}_k$ then a game starting at $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (i )}_k ) \in \mathcal{N}_k$ will converge to $ (\Delta \vect{x}_k^{ (* )}, \vect{\theta}^{ (* )}_k ) \in \mathcal{N}_k$ according to Lemma~\ref{lem:lem_min}. Since, $\mathcal{N}_k \subset \mathcal{M}_k,$ our result follows.
\end{proof}
In this game, the interplay between these two opposing players~(representative of generalization and forgetting, respectively) introduces the dynamics required to play the game. Furthermore, the results presented in this section are local to the task. In other words, we prove that we can achieve a balance between generalization and forgetting for each task $k$~(neighborhoods are task dependent, and we achieve a local solution given a task $k$).
Furthermore, our game is sequential; that is, there is a leader~(player 1) and a follower~(player 2). The leader~$(\Delta \vect{x}^{(i)}_{k})$ plays first, and the follower~$(\vect{\theta}^{(i)}_{k})$ plays second with complete knowledge of the leader's play. The game is directed by $\Delta \vect{x}^{(i)}_{k}$, and any changes in the task~(reflected in $\Delta \vect{x}^{(i)}_{k})$ will shift the input and thus the equilibrium point. Consequently, the equilibrium point varies with respect to a task, and one will need to attain a new equilibrium point for each shift in a task. Without complete knowledge of the tasks~(not available in a CL scenario), only a local result is possible. This highlights one of the key limitations of this work. Ideally, we would like a balance between forgetting and generalization that is independent of tasks. However, this would require learning a trajectory of the equilibrium point~(How does the equilibrium point change with the change in the tasks?) and is beyond the scope of this paper. One work that attempts to do this is~\cite{rusu2016progressive}, where the authors learn a parameter per task. For a large number of tasks, however, such an approach is computationally prohibitive.
These results are valid only under certain assumptions: (1) the Frobenius norm of the gradient is bounded, always positive; (2) the cost function is bounded and differentiable; and (3) the learning rate goes to zero as $i$ tends to infinity. The first assumption is reasonable in practice, and gradient clipping or perturbation strategies can be used to ensure it. The boundedness of the cost (second assumption) can be ensured by prioritizing the contributions of the task~(Lemma~1 and Corollary~1). The third assumption assumes a decaying learning rate. Learning rate decay is a common strategy and is employed widely. Therefore, all assumptions are practical and reasonable.
\begin{wrapfigure}[19]{r}{0.44\columnwidth}
\vspace{-11mm}
\begin{algorithm}[H]
\SetCustomAlgoRuledWidth{0.40\columnwidth}
Initialize ${\vect{\theta}}, D_{P}, D_{N}$ \\
\While{$k=1,2,3,... K$}{
j = 0\\
\While{$j < \rho$}{
Get $\vect{b}_{N} \in D^{N}_{k}$
Get $\vect{b}_{P} \in D^{P}_{k}$
Get $\vect{b}_{PN} = \vect{b}_{P} \cup \vect{b}_{N}$
Copy $\vect{b}_{PN}$ into $\vect{x}^{PN}_{k}$
i = 0
\While{$i+1 <= \zeta$}{
Update~$\vect{x}^{PN}_{k}$ with $J_{k}({\vect{\theta}}_k)$ using gradient ascent
i = i+1 }
Calculate $J_{k+\zeta}(\vect{\theta}^{(i)}_{k}) - J_{k}(\vect{\theta}^{(i)}_{k})$
Copy ${\vect{\theta}}^{(i)}_k$ into ${\vect{\theta}}^{B}_k$
i = 0
\While{$i+1 <= \zeta$}{
Update~${\vect{\theta}}^{B}_{k}$ with $J_{k}(\vect{\theta}^{B}_{k})$
i = i+1 }
Calculate $(J_{k}({\vect{\theta}}^{B}_{k}) - J_{k}(\vect{\theta}^{(i)}_{k}))$
Calculate $ H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k})$
Update ${\vect{\theta}}^{(i)}_k$ using gradient descent
}
j= j+1
}
Update ${D_{P}}$ with ${D_{N}}$
\caption{BCL \label{alg1a}}
\end{algorithm}
\end{wrapfigure}
\subsection{Balanced Continual Learning~\label{BCL}} Equipped with the theory, we develop a new CL method to achieve a balance between forgetting and generalization. By Proposition~2, the cost function can be upper bounded as $ H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k}) \leq \beta_k J_{k}(\vect{\theta}^{(i)}_{k}) + (J_{k}(\vect{\theta}^{(i+\zeta)}_{k}) - J_{k}(\vect{\theta}^{(i)}_{k})) + ( J_{k+\zeta}(\vect{\theta}^{(i)}_{k}) - J_{k}(\vect{\theta}^{(i)}_{k}) ),$ where $J_{k+\zeta}$ indicates $\zeta$ updates on player 1 and $\vect{\theta}^{(i+\zeta)}_{k}$ indicates $\zeta$ updates on player 2.
\begin{equation}
\begin{aligned}
\underbrace{\frac{\alpha_{k}^{(i)} \nabla_{\Delta \vect{x}_{k}} E[ H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k})] }{\| \nabla_{\Delta \vect{x}_{k}} H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k}) \|^2}}_{\text{Player~1}}, \\
\underbrace{-\alpha_{k}^{(i)}\times \nabla_{{\vect{\theta}}_{k}} E[H(\Delta \vect{x}^{(*)}_{k}, \vect{\theta}^{(i)}_{k}) )]}_{\text{Player~2}},
\end{aligned}
\label{eq:eq_Strat}
\end{equation}
The strategies for the two players $\Delta \vect{x}_{k}, {\vect{\theta}}_{k}$ are chosen in Eq. \eqref{eq:eq_Strat} with $E$ being the expected value operator. We can approximate the required terms in our update rule~(player strategies) using data samples (batches). Note that the approximation is performed largely through one-sided finite difference, which may introduce an error and is another potential drawback. The pseudo code of the BCL is shown in Algorithm~\ref{alg1a}. We define a new task array $\vect{\mathcal{D}} _N(k)$ and a task memory array~$\vect{\mathcal{D}} _{P}(k) \subset \cup_{\tau = 0}^{k-1}\cT_{\tau}$~(samples from all previous tasks). For each batch $b_{N} \in \vect{\mathcal{D}} _N(k)$, we sample $b_{P}$ from~$\vect{\mathcal{D}} _P(k),$ combine to create $b_{PN}(k)= b_{P}(k) \cup b_{N}(k)$, and perform a sequential play. Specifically, for each task the first player initializes $x_k^{PN} = b_{PN}(k)$ and performs $\zeta$ updates on $x_k^{PN}$ through gradient ascent. The second player, with complete knowledge of the first player's strategy, chooses the best play to reduce $H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k})$. To estimate player~2's play, we must estimate different terms in $H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k}).$ This procedure involves three steps. First, we use the first player's play and approximate $( J_{k+\zeta}(\vect{\theta}^{(i)}_{k}) - J_{k}(\vect{\theta}^{(i)}_{k}) ).$ Second, to approximate $(J_{k}(\vect{\theta}^{(i+\zeta)}_{k}) - J_{k}(\vect{\theta}^{(i)}_{k})):$ (a), we copy $\hat{\vect{\theta}}$ into $\hat{\vect{\theta}}_{B}$~(a temporary network) and perform $\zeta$ updates on $\hat{\vect{\theta}}_{B}$; and (b) we compute $J_{k}(\vect{\theta}^{(i+\zeta)}_{k})$ using $\hat{\vect{\theta}}_{B}(k+\zeta)$ and evaluate $(J_{k}(\vect{\theta}^{(i+\zeta)}_{k}) - J_{k}(\vect{\theta}^{(i)}_{k})).$ Third, equipped with these approximations, we compute $H(\Delta \vect{x}^{(i)}_{k}, \vect{\theta}^{(i)}_{k})$ and obtain the play for the second player. Both these players perform the steps repetitively for each piece of information~(batch of data). Once all the data from the new task is exhausted, we move to the next task.
\subsection{Related Work}
Traditional solutions to the CL focus on either the forgetting issue~\cite{rusu2016progressive, yoon2017lifelong, yao2020don, caccia2021online, kirkpatrick2017overcoming, zenke2017continual, aljundi2018memory, lin1992self, lopez2017gradient, chaudhry2019continual} or the generalization issue~\cite{vinyals2016matching, snell2017prototypical, finn2017model, finn2019online, Caccia2020OnlineFA}. Common solutions to the forgetting problem involve dynamic architectures and flexible knowledge representation such as \cite{rusu2016progressive, yoon2017lifelong, yao2020don, caccia2021online}, regularization approaches including \cite{kirkpatrick2017overcoming, zenke2017continual, aljundi2018memory} and memory/experience replay \cite{lin1992self, lopez2017gradient, chaudhry2019continual}.
Similarly, quick generalization to a new task has been addressed through few-shot and one-shot learning approaches such as matching nets \cite{vinyals2016matching} and prototypical network \cite{snell2017prototypical}. More recently, the field of metalearning has approached the generalization problem by designing a metalearner that can perform quick generalization from very little data \cite{finn2017model, finn2019online, Caccia2020OnlineFA}.
In the past few years, metalearners for quick generalization have been combined with methodologies specifically designed for reduced forgetting \cite{javed2019meta, beaulieu2020learning}. For instance, the approaches in \cite{javed2019meta, beaulieu2020learning} adapt the model-agnostic metalearning~(MAML) framework in \cite{finn2017model} with robust representation to minimize forgetting and provide impressive results on CL. However, both these approaches require a pretraining phase for learning representation. Simultaneously, Gupta et al.~\cite{gupta2020lamaml} introduced LA-MAML---a metalearning approach where the impact of learning rates on the CL problem is reduced through the use of per-parameter learning rates. LA-MAML~\cite{gupta2020lamaml} also introduced episodic memory to address the forgetting issue. Other approaches also have attempted to model both generalization and forgetting. In \cite{farajtabar2019orthogonal}, the gradients from new tasks are projected onto a subspace that is orthogonal to the older tasks, and forgetting is minimized. Similarly, Joseph and Balasubramanian \cite{joseph2020metaconsolidation} utilized a Bayesian framework to consolidate learning across previous and current tasks, and Yin et al.~\cite{yin2020optimization} provided a framework for approximating loss function to summarize the forgetting in the CL setting. Furthermore, Abolfathi et al.~\cite{abolfathi2021coachnet} focused on sampling episodes in the reinforcement learning setting, and Elrahimi et al.~\cite{ebrahimi2020adversarial} introduced a generative adversarial network-type structure to progressively learn shared features assisting reduced forgetting and improved generalization. Despite significant progress, however, these methods~\cite{javed2019meta, beaulieu2020learning, gupta2020lamaml,farajtabar2019orthogonal, joseph2020metaconsolidation, yin2020optimization, abolfathi2021coachnet, ebrahimi2020adversarial} are still inherently tilted toward maximizing generalization or minimizing forgetting because they naively minimize the loss function. Therefore, the contribution of different terms in the loss function becomes important. For instance, if the generalization cost is given more weight, a method would generalize better. Similarly, if forgetting cost is given more weight, a method would forget less. Therefore, the resolution of the trade-off inherently depends on an hyperparameter.
The first work to formalize the trade-off in CL was MER, where the trade-off was formalized as a gradient alignment problem. Similar to MER, Doan et al.~(\cite{doan2021theoretical}) studied forgetting as an alignment problem. In MER, the angle between the gradients was approximated by using Reptile \cite{nichol2018firstorder}, which promotes gradient alignment by reducing weight changes. On the other hand, \cite{doan2021theoretical} formalized the alignment as an eigenvalue problem and introduced a PCA-driven method to ensure alignment. Our approach models this balance as a saddle point problem achieved through stochastic gradient such that the saddle point~(balance point or equilibrium point) resolves the trade-off.
Our approach is the first in the CL literature to prove the existence of the saddle point~(the balance point) between generalization and forgetting given a task in a CL problem. Furthermore, we are the first to theoretically demonstrate that the saddle point can be achieved reasonably under a gradient ascent-descent game. The work closest to ours is \cite{ebrahimi2020adversarial}, where an adversarial framework is described to minimize forgetting in CL by generating task-invariant representation. However, \cite{ebrahimi2020adversarial} is not model agnostic~(the architecture of the network is important) and requires a considerable amount of data at the start of the learning procedure. Because of these issues, \cite{ebrahimi2020adversarial} is not suitable for learning in the sequential scenario.
\section{Experiments}
We use the CL benchmark \cite{Hsu18_EvalCL} for our experiments and retain the experimental settings~(hyperparameters) from \cite{Hsu18_EvalCL, vandeven2019generative}. For comparison, we use the split-MNIST, permuted-MNIST, and split-CiFAR100 data sets while considering three scenarios: incremental domain learning~(IDL), incremental task learning~(ITL), and incremental class learning~(ICL). The splitting and permutation strategies when applied to the MNIST or CiFAR100 data set can generate task sequences for all three scenarios~(illustrated in Figure~1 and Appendix: Figure~2 of \cite{Hsu18_EvalCL}). For comparing our approach, we use three baseline strategies---standard neural network with Adam~\cite{kingma2014adam}, Adagrad~\cite{duchi2011adaptive}, and SG---and use $L_2$-regularization and naive rehearsal~(which is similar to experience replay). For CL approaches, we use EWC~\cite{kirkpatrick2017overcoming}, online EWC~\cite{schwarz2018progress}, SI~\cite{zenke2017continual}, LwF~\cite{LwF}, DGR~\cite{DGR}, RtF~\cite{vandeven2019generative}, MAS~\cite{aljundi2018memory}, MER~\cite{riemer2018learning}, and GEM~\cite{GEM}. We utilize data preprocessing as provided by \cite{Hsu18_EvalCL}. Additional details on experiments can be found in Appendix~B and \cite{Hsu18_EvalCL, vandeven2019generative}. All experiments are conducted in Python 3.4 using the pytorch $1.7.1$ library with the NVIDIA-A100 GPU for our simulations.
\begin{table}[bt]
\vspace{-5pt}
\tiny
\centering
\resizebox{\columnwidth}{!}{
\begin{tabular}{c|ccc|ccc}
\hline \hline
\multirow{3}{*}{\textbf{Method}}&\multicolumn{3}{c|}{\textbf{split-MNIST}} &
\multicolumn{3}{c}{\textbf{permuted-MNIST}} \\ \cline{2-7}
& Incremental & Incremental & Incremental
& Incremental & Incremental & Incremental\\
& task learning & domain learning & class learning
& task learning & domain learning & class learning\\
& [ITL] & [IDL] & [ICL]
& [ITL] & [IDL] & [ICL] \\ \hline
Adam& $95.52 \pm 2.14$ & $54.75 \pm2.06$ &$19.72\pm0.03$
& $93.42 \pm 0.56$ & $77.87 \pm1.27$ &$14.02\pm1.25$\\
SGD& $97.65 \pm0.28$ & $62.80 \pm0.34$ &$19.36\pm0.02$
& $90.95 \pm0.20$ & $78.17 \pm1.16$ &$12.82\pm0.95$\\
Adagrad& $98.37 \pm0.29$ & $57.59 \pm2.54$ &$19.59\pm0.17$
& $92.45 \pm0.16$ & $91.59 \pm0.46$ &$29.09\pm1.48$\\
$L_2$& $97.62 \pm0.69$ & $66.84 \pm3.91$. &$22.92\pm1.90$
& $94.87 \pm0.38$ & $92.81 \pm0.32$ &$13.92\pm1.79$\\
Naive rehearsal& $99.32 \pm0.10$ & $94.85 \pm0.80$ &$90.88\pm0.70$
& $96.23 \pm0.04$ & $95.84 \pm0.06$ &$96.25\pm0.10$\\
Naive rehearsal-C& $99.41 \pm0.04$ & $97.13 \pm0.37$ &$94.92\pm0.63$
& $97.13 \pm0.03$ & $96.75 \pm0.03$ &$97.24\pm0.05$\\
\hline \hline
EWC& $96.59 \pm0.99$ &$57.31\pm1.07$ &$19.70\pm0.14$
& $95.38 \pm0.33$ &$89.54\pm0.52$ &$26.32\pm4.32$\\
Online~EWC& $99.01 \pm0.12$ &$58.25\pm1.23$ &$19.68\pm0.05$
& $95.15 \pm0.49$ &$93.47\pm0.01$ &$42.58\pm6.50$\\
SI& $99.10 \pm0.16$ &$64.63\pm1.67$ &$19.67\pm0.25$
& $94.35 \pm0.51$ &$91.12\pm0.93$ &$58.52\pm4.20$\\
MAS& $98.88 \pm0.14$ &$61.98\pm7.17$ &$19.70\pm0.34$
& $94.74 \pm0.52$ &$93.22\pm0.80$ &$50.81\pm2.92$\\
GEM& $98.32 \pm0.08$ &$97.37\pm0.22$ &$93.04\pm0.05$
& $95.44 \pm0.96$ &$96.86\pm0.02$ &$96.72\pm0.03$\\
DGR& $99.47 \pm0.03$ &$95.74\pm0.23$ &$91.24\pm0.33$
& $92.52 \pm0.08$ &$95.09\pm0.04$ &$92.19\pm0.09$\\
RtF&$\textbf{99.66}\pm\textbf{0.03}$&$97.31\pm0.11$&$92.56\pm0.21$
&$97.31\pm0.01$ &$97.06\pm0.02$ &$96.23\pm0.04$\\
MER&$97.12\pm0.10$ &$92.16\pm0.35$ &$93.20\pm0.12$
&$97.15\pm0.08$ &$96.11\pm0.31$ &$91.71\pm0.03$\\ \cline{2-7}
BCL~(With Game) &$99.52\pm0.07$ & $\textbf{98.71 }\pm\textbf{0.06}$ &$\textbf{97.32}\pm\textbf{0.17}$
& $\textbf{97.41} \pm \textbf{0.01}$ & $\textbf{97.51} \pm \textbf{0.05}$ &$\textbf{97.61}\pm\textbf{0.01}$ \\
BCL~(Without Game) & $97.73 \pm 0.03$ & $96.43 \pm0.29$ &$91.88\pm0.55$
& $96.16 \pm 0.03$ & $96.08 \pm 0.06$ &$95.96\pm0.06$\\
\hline
\hline
\end{tabular}}
\caption{Performance of our approach compared with other methods in the literature. We record the mean and standard deviation of the retained accuracy for the different methods. The best scores are in bold.}
\label{tab:main_res}
\vspace{-8mm}
\end{table}
\textbf{Comparison with the state of the art:} The results for our method are summarized in Table~\ref{tab:main_res} and \ref{tab:cifar_100}. The efficiency for any method is calculated by observing the average accuracy~(retained accuracy (RA)~\cite{riemer2018learning}) at the end of each repetition and then evaluating the mean and standard deviation of RA across different repetitions. For each method, we report the mean and standard deviation of RA over five repetitions of each experiment. In each column, we indicate the best-performing method in bold. For the split-MNIST data set, we obtain $99.52\pm0.07$ for ITL, $98.71\pm0.06$ for IDL, and $97.32\pm 0.17$ for ICL. Similarly, with the permuted-MNIST data set, we obtain $97.41\pm0.01$ for ITL, $97.51\pm0.05$ for IDL, and $97.61\pm0.01$ for ICL. Furthermore, with the split-CiFAR100 data set, we obtain $81.82 \pm 0.17$ for ITL, $62.11\pm 0.00$ for IDL, and $69.27 \pm 0.03$ for ICL. BCL is the best-performing methodology for all cases~(across both data sets) except RtF for ITL~($0.14 \%$ drop) with the split-MNIST data set.
\begin{wraptable}[16]{r}{0.61\columnwidth}
\tiny
\centering
\begin{tabular}{c|ccc}
\hline \hline
\multirow{3}{*}{\textbf{Method}}&\multicolumn{3}{c}{\textbf{split-CiFAR100}}\\ \cline{2-4}
& Incremental & Incremental & Incremental \\
& task learning & domain learning & class learning \\ \hline
Adam& $30.53 \pm0.58$ & $19.65 \pm0.14$ &$17.20\pm0.06$\\
SGD& $43.77 \pm1.15$ & $19.17 \pm0.12$ &$17.18\pm0.12$\\
Adagrad& $36.27 \pm0.43$ & $19.06 \pm0.14$ &$15.83\pm0.20$\\
$L_2$& $51.73 \pm1.30$ & $19.96 \pm0.15$ &$17.12\pm0.04$\\
Naive rehearsal& $70.20 \pm0.17$ & $35.94 \pm0.39$ &$34.33\pm0.19$\\
Naive rehearsal-C& $78.41 \pm0.37$ & $51.81 \pm0.18$ &$51.28\pm0.17$\\
\hline \hline
EWC& $61.11 \pm1.43$ &$19.76\pm0.12$ &$19.70\pm0.14$\\
Online~EWC& $63.22 \pm0.97$ &$20.03\pm0.10$ &$17.16\pm0.09$\\
SI& $64.81 \pm1.00$ &$20.26\pm0.09$ &$17.26\pm0.11$\\
MAS& $64.77 \pm0.78$ &$19.99\pm0.16$ &$17.07\pm0.12$\\ \cline{2-4}
BCL(With Game) & $\textbf{81.82} \pm \textbf{0.17}$ &$\textbf{62.11 }\pm\textbf{0.00}$ & $\textbf{69.27} \pm \textbf{0.03}$ \\
BCL(Without Game) & $69.17 \pm0.12$ &$51.82\pm0.19$ & $52.82 \pm 0.01$ \\
\hline
\hline
\end{tabular}
\caption{Performance of BCL for the split-CiFAR100 data set. We record the retained accuracy for the different methods. We obtained RA scores for all methods except BCL from \cite{Hsu18_EvalCL}.}
\label{tab:cifar_100}
\end{wraptable} Generally, ITL is the easiest learning scenario~\cite{Hsu18_EvalCL}, and all methods therefore perform well on ITL~(the performance is close). For the ITL scenario with the split-MNIST data set, BCL is better than most methods; but several methods, such as naive rehearsal, naive rehearsal-C, RtF, and DGR, attain close RA values~(less than $1\%$ from BCL). Note that both DGR and RtF involve a generative model pretrained with data from all the tasks. In a sequential learning scenario, one cannot efficiently train generative models because data corresponding to all the tasks is not available beforehand. Although RtF provides improved performance for split-MNIST~(ITL), the improvement is less than $1\%$~(not significant). In fact, RtF performance is poorer for BCL in ICL by $4.76\%$~(a significant drop in performance) and in IDL by $1.4 \%.$
Two additional observations can be made about the split-MNIST data set. First, Adagrad, SGD, and L2 achieve better performance than does Adam. Therefore, in our analysis Adagrad appears more appropriate although Adam is popularly used for this task. Second, naive rehearsal~(both naive rehearsal and naive rehearsal-C approaches) achieves performance equivalent to the state-of-the-art methods with similar memory overhead. Furthermore, naive rehearsal performs much better than online EWC and SI, especially in the ICL scenario. These limitations indicate that regularization-driven approaches are not much better than baseline models and in fact perform poorer than methods involving memory~(naive rehearsals, MER, BCL, etc.). In~\cite{knoblauch2020optimal}, it was shown theoretically that memory-based approaches typically do better than regularization-driven approaches, as is empirically observed in this paper, too. Another interesting observation is that EWC and online EWC require significant hyperparameter tuning, which would be difficult to do in real-world scenarios. Other regularization-based methods, such as SI and MAS, also suffer from the same issue.
The observations from the split-MNIST carry forward to the permuted-MNIST data set. Moreover, RA values for the permutation MNIST data set are better for the split-MNIST data set across the board, indicating that the permutation MNIST data set presents an easier learning problem. Similar to the observations made with the split-MNIST data set, BCL is better than all methods for the permuted-MNIST dataset, with naive rehearsal and RtF providing RA values that are close~(less than $1 \%$). The only methodology in the literature that attempts to model the balance between forgetting and generalization is MER, an extension of GEM. From our results, we observe that BCL is better than MER in all cases~(Split-MNIST--$2.4 \%$ improvement for ITL, $6.55 \%$ improvement for IDL, $4.12 \%$ improvement for ICL and Permuted-MNIST--$0.26 \%$ improvement for ITL, $1.4 \%$ improvement for IDL and $5.9 \%$ improvement for ICL).
The substantial improvements obtained by BCL are more evident in Table \ref{tab:cifar_100} where the results on the split-CiFAR100 data set are summarized. BCL is clearly the best-performing method. The next best-performing method is naive rehearsal-C, where BCL improves performance by $3.41 \%$ for ITL, $10.3 \%$ for IDL, and $17.99 \%$ for ICL. Other observations about the regularization methods and the rest of the baseline methods carry forward from Table~\ref{tab:main_res}. However, one key difference is that while AdaGrad is observed to be better than Adam for the MNIST data set, Adam is comparable to SGD and Adagrad for split-CiFAR100. In summary, BCL is comparable to or better than the state of the art in the literature for both the MNIST and CiFAR100 data sets.
\begin{wrapfigure}[30]{l}{0.49\textwidth}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width = \columnwidth]{Figures/tradeoff_loss.pdf}
\caption{}
\label{fig:all_task}
\end{subfigure}\\
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width = \columnwidth]{Figures/tradeoff_loss_one_task.pdf}
\caption{}
\label{fig:one_task}
\end{subfigure}
\caption{(Top) Progression of different terms in Eq. \eqref{eq_M_DES} with respect to update iterations, where the index on the x axis is calculated as $k\times 300$. The tasks boundaries~(at what $i$ the tasks are introduced) are illustrated through shades of grey. (down) Illustration of the cost at $k =3$.}
\end{wrapfigure}
\textbf{Do we need a game to achieve this performance?} To provide additional insights into the benefits of the game, we compare RA values with and without the game. In our setup, $\Delta \vect{x}^{(i)}_{k}$ aims at increasing the cost, and $\vect{\theta}^{(i)}_{k}$ aims at reducing the cost. If we hold the play for $\Delta \vect{x}^{(i)}_{k}$, then the dynamics required to play the game do not exist, thus providing a method that can perform CL without the game. Therefore, we induce the absence of a game by fixing $\Delta \vect{x}^{(i)}_{k}$ and perform each of the nine experiments for five repetitions~(ICL, IDL, and ITL for split-MNIST, permuted-MNIST, and split-CiFAR100). We summarize these results in the last two rows of Tables \ref{tab:main_res} and \ref{tab:cifar_100}. Consequently, we make two observations. First: even without the game, we achieve RA values comparable to the state of the art. This is true for both the MNIST and CiFAR100 data sets. Second, with the introduction of the game, we observe improved RA values across the board~(at least by $1 \%$ with MNIST). The difference is clearer with the CiFAR100 data set where we observe a substantial improvement in RA values~(at least $10 \%$).
\textbf{Does the theory appropriately model the continual learning problem?} In Fig.~\ref{fig:all_task} we plot the progression of $J_{k}( {\vect{\theta}}_{k})+ \langle \nabla_{\vect{x}_{k}} V_{k}^{(*)}, \Delta \vect{x}_{k}\rangle$~(blue curve), $\langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{(*)}, \Delta {\vect{\theta}}_{k} \rangle$~(red curve), and the sum, namely, $\Delta V^{(*)}_{k} $~(green curve). A total of six tasks, sampled from the permutation MNIST data set within the ICL setting, are illustrated in Fig. \ref{fig:all_task}. These tasks are introduced every 300 update steps. From Fig. \ref{fig:all_task} we make two important observations. First, as soon as tasks 1, 3, 4, and 6 are introduced, all three curves~(red, blue, and green) indicate a bump. Second, when tasks 2 and 5 are introduced, there is no change. On a closer look at task 3 in Fig.~\ref{fig:one_task}, we observe that when the task 3 is introduced, the blue curve exhibits a large positive bump~(the introduction of the new tasks increases the first and the third terms in Eq. \eqref{eq_M_DES}). The increase implies that task 3 forced the model to generalize and increased forgetting on tasks 1 and 2~(observed by the increase in the green curve). To compensate for this increase, we require the model~$\vect{\theta}_k$ to behave adversarially and introduce a large enough negative value in $\langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{(*)}, \Delta {\vect{\theta}}_{k} \rangle$~(red curve) to cancel out the increase in the blue curve. In Fig.~\ref{fig:one_task} the red curve demonstrates a large negative value~(expected behavior) and eventually~(as $i$ increases) forces the blue curve~(by consequence, green: the sum of red and blue) to move toward zero~(the model compensates for the increase in forgetting). As observed, the blue and the red curves behave opposite to each other and introduce a push-pull behavior that stops only when the two cancel each other and the sum~(green) is zero. Once the sum has reached zero, there is no incentive for the red and green to be nonzero, and therefore they remain at zero; thus, all three curves (green, red, and blue) remain at zero once converged until a task 4 is introduced~(when there is another bump, as seen in Fig. \ref{fig:all_task}). However, this increase in the blue curve is not observed when tasks 2 and 5 are introduced. When new tasks are similar to the older tasks, it is expected that $\Delta V^{(*)}_{k} =0,$ as is observed in Fig. \ref{fig:all_task}.
All of these observations are fully explained by Eq. \ref{eq_M_DES}, which illustrates that the solution to the CL problem is obtained optimally only when $\Delta V^{(*)}_{k} =0$~(observed in Figs. \ref{fig:all_task} and \ref{fig:one_task}). The term $\Delta V^{(*)}_{k} $ is quantified by $J_{k}( {\vect{\theta}}_{k}), \langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{(*)}, \Delta {\vect{\theta}}_{k} \rangle$ and $\langle \nabla_{\vect{x}_{k}} V_{k}^{(*)}, \Delta \vect{x}_{k}\rangle$. Our theory suggests that there exists an inherent trade-off between different terms in Eq. \ref{eq_M_DES}. Therefore every time a new task is observed, it is expected that~$J_{k}( {\vect{\theta}}_{k})+ \langle \nabla_{\vect{x}_{k}} V_{k}^{(*)}, \Delta \vect{x}_{k}\rangle$ increases~(increase in blue curve when tasks 1, 3, 4, and 6 are introduced) and $\langle \nabla_{{\vect{\theta}}_{k}} V_{k}^{(*)}, \Delta {\vect{\theta}}_{k} \rangle$ compensates to cancel this increase~(red curve exhibits a negative jump). In Theorems~1 and 2 we demonstrate the existence of this balance point~(for each task, as $i$ increases, $\Delta V^{(*)}_{k} $ tends to zero, as observed in Fig. \ref{fig:one_task}) and $\Delta V^{(*)}_{k} $ remains zero~(the balance point is stable, proved in Theorem~2) until a new task increases forgetting. Furthermore, our theory claims the existence of a solution with respect to each task. This is also observed in Fig.~\ref{fig:all_task} as, for each task, there is an increase in cost, and BCL quickly facilitates convergence. \emph{These observations indicate that our dynamical system in Eq. \ref{eq_M_DES}
accurately describes the dynamics of the continual learning problem}. Furthermore, the assumptions under which the theory is developed are practical and are satisfied by performing continual learning on the permuted MNIST problem.
\section{Conclusion}
We developed a dynamic programming-based framework to enable the methodical study of key challenges in CL. We show that an inherent trade-off between generalization and forgetting exists and must be modeled for optimal performance. To this end, we introduce a two-player sequential game that models the trade-off. We show in theory and simulation that there is an equilibrium point that resolves this trade-off~(Theorem~1) and that this saddle point can be attained~(Theorem~2). However, we observe that any change in the task modifies the equilibrium point. Therefore, a global equilibrium point between generalization and forgetting is not possible, and our results are valid only in a neighborhood~(defined given a task). To attain this equilibrium point, we develop BCL and demonstrate state-of-the-art performance on a CL benchmark~\cite{Hsu18_EvalCL}. In the future, we will extend our framework for nonEuclidean tasks.
\section{Broader Impact}
\textit{Positive Impacts:} CL has a wide range of applicability. It helps avoids retraining, and it improves the learning efficiency of learning methods. Therefore, in science applications where the data is generated sequentially but the data distribution varies with time, our theoretically grounded method provides the potential for improved performance.
\textit{Negative Impacts:} Our theoretical framework does not have direct adverse impacts. However, the potential advantages of our approach can improve the efficiency of adverse ML systems such as fake news, surveillance, and cybersecurity attacks.
\begin{ack}
This work was supported by the U.S.\ Department of Energy, Office of Science, Advanced Scientific Computing Research, under Contract DE-AC02-06CH11357 and by a DOE Early Career Research Program award. We are grateful for the computing resources from the Joint Laboratory for System Evaluation and Leadership Computing Facility at Argonne. We also are grateful to Dr. Vignesh Narayanan, assistant professor, University of South Carolina, and Dr. Marieme Ngom, Dr. Sami Khairy --postdoctoral appointees, Argonne National Laboratory, for their insights.
\end{ack}
\bibliographystyle{plain}
|
\section{Introduction}
\IEEEPARstart{L}{ung} cancer is the leading cause of cancer deaths around the world. In 2018, the global number of new cases and deaths were 2.09 million and 1.76 million, respectively~\cite{Bray2018}.
Radiation therapy is one of the most effective and widely used treatment for cancer patients~\cite{Baskar2012}. With the development of such treatment technologies, lung cancer death rate dropped 45\% from 1990 to 2015 among men, and 19\% from 2002 to 2015 among women~\cite{Siegel2018}. However, about half of all cancer patients who receive radiation therapy during their course of illness will suffer from Radiation-Induced Injuries (RII) to the hematopoietic tissue, skin, lung and gastrointestinal (GI) systems~\cite{Baskar2012}. To prevent, mitigate or treat the RII plays an important role in improving the quality of radiation therapies.
For lung cancer, the most common RII is the radiation-induced Pulmonary Fibrosis (PF)~\cite{Wilson2009}, i.e., inflammation and subsequent scarring of lung tissues caused by radiation, which could lead to breathing problems due to lung damage and even lung failure~\cite{Giuranno2019}. Observation and assessment of the PF lesions using Computed Tomography (CT) imaging is critical for diagnosis and treatment follow-up of this disease~\cite{Christe2019}. For an accurate and quantitative measurement of PF, it is desirable to segment the PF lesions from 3D CT scans.
The segmentation results can provide detailed spatial distribution and accurate volumetric measurement of the lesions, which is important for treatment decision making, PF progress modeling, treatment effect assessment and prognosis prediction.
\begin{figure*}[t]
\centering
\includegraphics[width=0.9\textwidth]{figures/pf_examples.pdf}
\caption{
Examples of radiation-induced Pulmonary Fibrosis (PF) of the Rhesus Macaque. The first row shows lung CT images, and the second row shows manual segmentation results of PF lesions. Note that the ambiguous boundary, irregular shape and various size and position make the segmentation task challenging. }
\label{fig:image_example}
\end{figure*}
As manual segmentation of lesions from 3D images is time-consuming, labor-intensive and faced with inter- and intra-observer varabilities, automatic PF lesion segmentation from CT images is highly desirable~\cite{VanRikxoort2013}.
However, this is challenging due to several reasons. Firstly, with different severity of the disease, PF lesions have a large variation of size and shape. A small lesion may only contain few pixels, while a large lesion can occupy a lung segment. The irregular shapes make it difficult to use a statistical model for the segmentation task~\cite{Wang2015c}. Secondly, the lesions have a complex spatial distribution, and can be scattered in different segments of the lung. Thirdly, PF lesions often adhere to lung structures including vessels, airways and the pleura, and other lesions like lung nodules and pneumonia lesions with similar appearance may also exist. These factors, alongside with the low contrast of soft tissues in CT images, make it hard to delineate the boundary of PF lesions. Fig.~\ref{fig:image_example} shows two examples of lung CT scans with PF, and it demonstrates the difficulties of accurate segmentation.
Recently, Convolutional Neural Networks (CNNs) have been increasingly used for automatic medical image segmentation~\cite{Shen2017}. By automatically learning features from a large set of annotated images, they have outperformed most traditional segmentation methods using hand-crafted features~\cite{VanRikxoort2013}, such as for recognition of lung nodules~\cite{Xie2019,Wang2017e},
lung lobes~\cite{Xie2020} and COVID-19 infection lesions~\cite{Fan2020,Wang2020c}.
However, to the best of our knowledge, CNNs for PF lesion segmentation have rarely been investigated so far.
Besides the above challenges, existing CNNs may obtain suboptimal performance for the PF lesion segmentation task due to the following reasons. First, lung CT images usually have anisotropic 3D resolutions with high inter-slice spacing and low intra-slice spacing. Existing CNNs using pure 2D convolutions or pure 3D convolutions have limited ability to learn effective 3D features from such images, as 2D CNNs~\cite{Ronneberger2015,Zhou2018,Roy2019,Wang2015c} can only learn intra-slice features, and most 3D networks~\cite{Abdulkadir2016,Milletari2016,Liwenqi2017,Huang2019} are designed with isotropic receptive field in terms of voxels. When dealing with 3D images with large inter-slice spacing, they have an imbalanced physical receptive field (in terms of mm) along each axis, i.e., the physical receptive field in the through-plane direction is much larger than that in in-plane directions, which may limit effective learning of 3D features.
Second, existing CNNs often use position invariant convolutions without spatial awareness, which makes it difficult to handle objects with various position and size. Attention mechanisms have recently been proposed to improve the spatial awareness~\cite{Oktay2018,Roy2019, Wang2017c}, but their obtained spatial attention does not match the target region well, and their performance on PF lesions has not been investigated.
What's more, current success of deep learning methods for segmentation relies highly on a large set of annotated images for training~\cite{Shen2017}. For 3D medical images, acquiring pixel-level annotations in segmentation tasks is extremely time-consuming and difficult, as accurate annotations could be only provided by experts with domain knowledge~\cite{Tajbakhsh2019}. For the PF lesion segmentation task, annotation of a CT volume could take several hours, and the complex shape and appearance of PF lesions further increase the efforts and time needed for annotation, which makes it difficult to annotate a large set of 3D pulmonary CT scans for training.
To deal with these problems, we propose a novel semi-supervised framework with a novel 2.5D CNN based on multi-scale attention for the segmentation of PF lesions from CT scans with large inter-slice spacing. The contribution is three-fold. First, we propose a novel network for PF lesion segmentation (i.e., PF-Net), which employs multi-scale guided dense attention to deal with lesions with various size and position, and combines 2D and 3D convolutions to achieve balanced physical receptive field along different axes to better learn 3D features from medical images with anisotropic resolution.
Second, a novel semi-supervised learning framework using Iterative Confidence-based Refinement And Weighting of pseudo Labels (I-CRAWL) is proposed, \textcolor{black}{where uncertainty estimation is employed to assess the quality of pseudo labels of unannotated images in both pixel level and image level. We propose Confidence-Aware Refinement (CAR) based on pixel-level uncertainty to refine pseudo labels, and introduce confidence-based image weighting according to image-level uncertainty to suppress low-quality pseudo labels.} Thirdly, we apply our proposed method to radiation-induced PF lesion segmentation from CT scans, and extensive experimental results show that our method outperformed state-of-the-art semi-supervised methods and existing 2D, 3D and 2.5D CNNs for segmentation. As far as we know, this is the first work on PF lesion segmentation based on deep learning, and our method has a potential to reduce the annotation burden for large-scale 3D image datasets in the development of automatic segmentation models with high performance.
\section{Related Works}
\subsection{CNNs for Medical Image Segmentation}
CNNs have achieved state-of-the-art performance for many medical image segmentation tasks~\cite{Shen2017}. Most widely used segmentation CNNs are inspired by U-Net~\cite{Ronneberger2015}, which is based on an encoder-decoder structure to learn features at multiple scales. UNet++~\cite{Zhou2018} extends U-Net with a series of nested, dense skip pathways for a higher performance. Attention U-Net~\cite{Oktay2018} introduced an attention gate using high-level features to calibrate low-level features. Spatial and channel ``Squeeze and Excitation" (scSE)~\cite{Roy2019} enables a 2D network to focus on the most relevant features for better performance.
Typical networks for segmentation of 3D volumes include 3D U-Net~\cite{Abdulkadir2016}, V-Net~\cite{Milletari2016} and HighRes3DNet~\cite{Liwenqi2017}.
They assume that the input volume has an isotropic 3D resolution to learn 3D features, and are not suitable for images with large inter-slice spacing. To better deal with such images, a 3D anisotropic hybrid network that uses a pre-trained 2D encoder and a decoder with anisotropic convolutions was proposed in~\cite{Liu2018c}. Jia et al.~\cite{Jia2020} designed a pyramid anisotropic CNN based on decomposition of 3D convolutions. The nnU-Net~\cite{Isensee2021} automatically configures network structures and training strategies, where different types of convolution kernels can be adaptively combined for a given dataset. In~\cite{Lei2019}, a lightweight CNN combining inter-slice and intra-slice convolutions was proposed to for segmentation of CT images. In~\cite{Wang2019b}, 2D and 3D convolutions were combined in a single network for Vestibular Schwannoma segmentation from images with anisotropic resolution. However, these networks have limited ability to segment lesions with various size and position.
\subsection{Segmentation of Lung CT Images}
CNNs have been widely used for segmentation of lung structures from CT images~\cite{VanRikxoort2013}.
In~\cite{Xie2020}, cascaded CNNs with non-local modules~\cite{Wang2017c} were proposed to leverage structured relationships for pulmonary lobe segmentation. In~\cite{Nadeem2020}, a CNN was combined with freeze-and-grow propagation for airway segmentation. For lung lesions, a central focused CNN~\cite{Wang2017e} was proposed to segment lung nodules from heterogeneous CT images, and Fan et al.~\cite{Fan2020} employed reverse attention and edge attention to segment COVID-19 lung infection. Wang et al.~\cite{Wang2020c} developed a noise-robust framework for automatic segmentation of COVID-19 pneumonia lesions by learning from non-expert annotations. Despite the large amount of works on lung structure and lesion segmentation so far, there is a lack of deep learning models for the challenging task of radiation-induced pulmonary fibrosis segmentation.
\subsection{Semi-Supervised Learning}
To reduce the burden for annotation, semi-supervised learning methods have been increasingly employed for medical image segmentation by using a limited number of annotated images and a large amount of unannotated images~\cite{Tajbakhsh2019}. Existing semi-supervised methods mainly have two categories. The first category is based on pseudo labels~\cite{Lee2013,Bai2017a,Fan2020}, where a model trained with annotated images obtains pseudo labels for unannotaed images that are then used to update the segmentation model. Lee~\cite{Lee2013} used such a strategy for classification problems, and Bai et al.~\cite{Bai2017a} updated the pseudo segmentation labels and network parameters alternatively and used Conditional Random Field (CRF) to refine the pseudo labels. Fan et al.~\cite{Fan2020} progressively enlarged the training set with unlabeled data and their pseudo labels for learning. However, this method ignores the quality of pseudo labels, which may limit the performance of the learned model.
The second category is to learn from annotated and unannotated images simultaneously, and they often consist of a supervised loss function for annotated images and an unsupervised regularization loss function for all the images. The regularization can be based on teacher-student consistency~\cite{Cui2019,Yu2019}, transformation consistency~\cite{Li2020}, multi-view consistency~\cite{Xia2020} and reconstruction-based auxiliary task~\cite{Chen2019}. Adversarial learning~\cite{Zhao2018}
also regularizes the segmentation model by minimizing the distribution difference between segmentation results of annotated images and those of unannotaed images. However, adversarial models are hard to train, and the complex size and shape of PF lesions make it difficult to capture the true distribution of lesion masks when only a small set of annotated images are available.
\section{Method}
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{figures/pf-net.pdf}
\caption{
Structure of our proposed PF-Net. To deal with 3D images with large inter-slice spacing, the first two scales use 2D convolutions while the other scales use 3D convolutions. $\tilde{P}_s$ is the predicted spatial attention at scale $s$, and it is sent to all lower scales with dense connection in the decoder, as shown by green lines. }
\label{fig:network}
\end{figure*}
In this section, we first introduce our proposed Pulmonary Fibrosis segmentation Network (PF-Net) with multi-scale guided dense attention, and then describe how it is used in our I-CRAWL framework for semi-supervised learning.\textbf{}
\subsection{PF-Net: A 2.5D Network with \textcolor{black}{Multi-Scale Guided Dense Attention}}
Our proposed PF-Net is illustrated in Fig.~\ref{fig:network}. It employs an encoder-decoder backbone structure that is commonly used and effective for medical image segmentation~\cite{Ronneberger2015,Milletari2016,Zhou2018}. The encoder contains five scales, and each is implemented by a convolutional block followed by a max-pooling layer for down-sampling. \textcolor{black}{In each block, we use two convolutional layers each followed by a Batch Normalization (BN)~\cite{Ioffe2015} layer} and a parametric Rectified Linear Unit (pReLU), and a dropout layer is inserted before the second convolutional layer. The decoder uses the same type of convolutional blocks as the encoder. We extend this backbone in the following aspects:
\subsubsection{2.5D Network Structure}
To deal with the anisotropic 3D resolution with high inter-slice spacing and low intra-slice spacing, we combine 2D (i.e., intra-slice) and 3D convolutions so that the network has an approximately balanced physical receptive field along each axis. Let $S$ represent the number of scales in the network ($S=5$ in this paper), we use 2D convolutions and 2D max-poolings for the first $M$ scales in the encoder, and employ 3D convolutions and 3D max-poolings for the last $S-M$ scales in the encoder. Each resolution level in the decoder contains the same type of 2D or 3D convolutional blocks as in the encoder. We use trilinear interpolation for upsampling in the decoder.
As our lung CT images have a resolution around 0.3$\times$0.3$\times$1.25 mm, i.e., the in-plane resolution is about four times of the through-plane resolution, we set $M=2$ so that the 2D max-pooling layers in the first two scales make the resulted feature maps have a near-isotropic 3D resolution, as shown in Fig.~\ref{fig:network}. Note that the first and $S$-th scales have the highest and lowest spatial resolution, respectively. We use $F^e_s$ and $F^d_s$ to denote the feature maps obtained by the last convolutional block at scale $s$ in the encoder and decoder, respectively.
Note that $F^e_S$ = $F^d_S$ in the bottleneck block.
\subsubsection{\textcolor{black}{Multi-Scale Guided Dense Attention}}\label{sec:method:mdac}
To better deal with PF lesions with various position and size, we use multi-scale attention to improve the network's spatial awareness, and propose dense attention to leverage multi-scale contextual information for the segmentation task. Specifically, at each scale $s$ of the decoder, we use a convolutional layer to get a spatial attention map $\tilde{P}_s$, and use $\tilde{P}_s$ as a high-level attention signal to guide the learning in all lower scales of the decoder. The input of the decoder at scale $s$ is a concatenation of three parts: $F^e_s$ from scale $s$ of the encoder, an upsampled version of the decoder feature map $F^d_{s+1}$ and $\tilde{P}^s_{s+1}\oplus \tilde{P}^s_{s+2} \oplus ... \oplus \tilde{P}^s_{S}$, where $\oplus$ is the concatenation operation and $\tilde{P}^s_{s+1}$ is the upsampled version of $\tilde{P}_{s+1}$ so that it has the same spatial resolution as $F^e_s$. Therefore, a lower scale accepts the attention maps from all the higher scales as input, which is referred to as \textcolor{black}{Multi-Scale Dense Attention (MSDA)}. The decoder thus takes advantage of multi-scale contextual information that enables the network to pay more attention to the target region.
To better learn the spatial attention, we propose \textcolor{black}{Multi-Scale Guided Attention (MSGA)} to explicitly supervise $\tilde{P}_1, \tilde{P}_2, ..., \tilde{P}_S$ at different scales. Let $P_s$ denote the softmaxed output of $\tilde{P}_s$ and $Y$ denote the ground truth (i.e., one-hot probability map) of a training sample $X$. Unlike a common deep supervison strategy that upsamples $\tilde{P_s}$ or $P_s$ at different scales to the same spatial resolution as $Y$~\cite{Dou2016miccai,Zhang2018f,Wang2019a}, we down-sample $Y$ to obtain multi-scale ground truth for spatial attention, which makes the loss calculation more efficient and can directly supervise the spatial attention maps at different scales. Let $Y_s$ denote the down-sampled ground truth at scale $s$. The multi-scale loss function for a single image is:
\begin{equation}\label{eq:loss1}
\ell(\mathcal{P}, Y) = \sum_{s = 1}^{S}\alpha_s L_s(P_s, Y_s)
\end{equation}
where $\mathcal{P}=\{P_1, P_2, ..., P_S\}$. $L_s()$ is a base loss function for image segmentation, such as the Dice loss~\cite{Milletari2016}. $\alpha_s$ is the weight of $L_s()$ at scale $s$.
\subsection{Semi-Supervised Learning using I-CRAWL}
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{figures/framework.pdf}
\caption{
Our proposed I-CRAWL (Iterative Confidence-based Refinement And Weighting of pseudo Labels) framework for semi-supervised segmentation. }
\label{fig:icrawl}
\end{figure*}
Generating pseudo labels for unannotated images has been shown effective for semi-supervised segmentation~\cite{Lee2013,Bai2017a,Fan2020}. However, the pseudo labels often contain some incorrect regions, and low-quality pseudo labels can largely limit the performance of the learned segmentation model. To prevent the learning process from being corrupted by inaccurate pseudo labels, we propose an Iterative Confidence-based Refinement And Weighting of pseudo Labels (I-CRAWL) framework for semi-supervised segmentation.
Assume that the entire training set consists of one subset $\mathcal{D}^L$ with $N^L$ labeled images and another subset $\mathcal{D}^U$ with $N^U$ unlabeled images. For an image $X^L_i \in \mathcal{D}^L$, its ground truth label $Y^L_i$ is known, while for an image $X^U_i \in \mathcal{D}^U$, its ground truth label is not provided, and we use $Y^U_i$ to denote its estimated pseudo label. As the pseudo labels may have a large range of quality, we introduce an image-level weight $w^U_i \in [0, 1]$ for each pair of $X^U_i,Y^U_i$ for learning, and define $w^U_i$ based on the confidence (or uncertainty, inversely) of $Y^U_i$. The weight for a labeled image $X^L_i$ from $\mathcal{D}^L$ can be similarly denoted as $w^L_i$, and we set $w^L_i = 1$ as the corresponding label $Y^L_i$ is clean and reliable. Therefore, the labeled subset can be denoted as $\mathcal{D}^L = \{(X^L_i, Y^L_i, w^L_i), ...\}$ ($i=1, 2, ..., N^L$), and the unlabeled subset with pseudo labels can be denoted as $\mathcal{D}^U = \{(X^U_i, Y^U_i, w^U_i), ...\}$ ($i=1, 2, ..., N^U$).
Our I-CRAWL is illustrated in Fig.~\ref{fig:icrawl}, and it is an iterative learning process with $K$ rounds. Each round has four steps: 1) inference for unannotated images with uncertainty estimation, 2) confidence-aware refinement of pseudo labels, 3) confidence-based image weighting, and 4) network update where the current pseudo labels and image-level weights are used to train the network. These steps are detailed as follows.
\subsubsection{Inference for Unannotated Images with Uncertainty Estimation}
With the network parameter $\theta_{k-1}$ obtained in the last round, in round $k$, we first use $\theta_{k-1}$ to predict provisional pseudo labels for unannotated images in $\mathcal{D}^U$ and the associated uncertainty estimation.
Note that in the first round (i.e., $k = 1$), the initial network parameter $\theta_0$ is obtained by pre-training with the annotated images in $\mathcal{D}^L$.
With $\theta_{k-1}$, we employ the Monte Carlo (MC) Dropout~\cite{Gal2016} that has shown to be an effective method for estimation of epistemic uncertainty caused by the lack of training data~\cite{Yu2019,Xia2020}. MC Dropout feeds an image $X^U_i$ into the network $R$ times with random dropout, which leads to $R$ predictions, i.e., $R$ foreground probability maps~\cite{Gal2016}. The average of these $R$ foreground probability maps is taken as the provisional probability map $P^U_i$, a binarization of which leads to a provisional pseudo segmentation label $Y^U_i$.
At the same time, the statistical variance of the $R$ foreground probability maps is taken as the uncertainty map $V^U_i$, which gives voxel-level uncertainty. As uncertainty information can indicate potentially wrong segmentation results~\cite{Yu2019,Wang2019i}, we treat the pseudo labels with high uncertainty (i.e., low confidence) values as unreliable labels for images in $\mathcal{D}^U$, and propose a Confidence-Aware Refinement (CAR) method to improve the pseudo labels' quality.
\subsubsection{Confidence-Aware Refinement of Pseudo Labels}
Given the provisional pseudo label $Y^U_i$ with the uncertainty map $V^U_i$ for an unannotated image $X^U_i\in\mathcal{D}^U$, and let $\mathbf{x}$ denote a voxel, we split the voxels in $Y^U_i$ into three sets according to the status: high-confidence foreground voxels $\mathbb{F} = \{\mathbf{x}| Y^U_{i\mathbf{x}} = 1, V^U_{i\mathbf{x}} \le \epsilon\}$, high-confidence background voxels $\mathbb{B} = \{\mathbf{x}| Y^U_{i\mathbf{x}} = 0, V^U_{i\mathbf{x}} \le \epsilon\}$ and undetermined voxels $\mathbb{U} = \{\mathbf{x}|V^U_{i\mathbf{x}} > \epsilon\}$, where $\epsilon=0.05$ is a small threshold value.
For undetermined voxels in $\mathbb{U}$, we refine their labels according to a contextual regularization considering inter-voxel connections and softened probabilities for these voxels.
Our CAR for pseudo label refinement has two steps: probability map softening and contextual regularization. First, we soften the foreground probabilities for uncertain voxels, which will degrade the influence of the network's prediction for these voxels in the following contextual regularization step. Let $p$ denote a foreground probability value and $u$ denote the corresponding uncertainty value, the softening function is:
\begin{equation}
g(p,u)=0.5 + (p - 0.5)\times(1-u)
\end{equation}
where the softened foreground probability get closer to 0.5 when $u$ is larger. Let $\dot{P}^U_{i\mathbf{x}}$ denote the softened foreground probability of voxel $\mathbf{x}$, and is obtained by:
\begin{align}
\dot{P}^U_{i\mathbf{x}}=\begin{cases}
P^U_{i\mathbf{x}} & \text{if } {\mathbf{x}\in \mathbb{F}\cup\mathbb{B}}
\\
g(P^U_{i\mathbf{x}},V^U_{i\mathbf{x}}) & \text{if } {\mathbf{x}\in \mathbb{U}}
\end{cases}
\end{align}
Then, we use contextual regularization taking $\dot{P}^U_i$ as input to refine the pseudo labels, which is implemented by a fully connected Conditional Random Field (CRF)~\cite{Krahenbuhl2011}. For simplicity, we denote $X^U_{i\mathbf{x}}$, $Y^U_{i\mathbf{x}}$ and $\dot{P}^U_{i\mathbf{x}}$ as $x_{\mathbf{x}}$, $y_{\mathbf{x}}$ and $\dot{p}_{\mathbf{x}}$, respectively. The energy function of CRF is:
\begin{align}\label{eq:crf}
E(Y^U_i)=\sum_{\mathbf{x}}\phi(y_{\mathbf{x}}) +
\sum_{\mathbf{x,y}}\psi(y_{\mathbf{x}},y_{\mathbf{y}})
\end{align}
where $\phi(y_{\mathbf{x}}) = -y_{\mathbf{x}}\text{log}(\dot{p}_{\mathbf{x}}) - (1-y_{\mathbf{x}})(1-\dot{p}_{\mathbf{x}})$ constrains the output to be consistent with the softened probability map, and this constraint for the low-confidence voxels in $\mathbb{U}$ is weak. The second term in Eq.~\eqref{eq:crf} is a pairwise potential that encourages the label's contextual consistency:
\begin{multline}\label{eq:crf_psi}
\psi(y_{\mathbf{x}},y_{\mathbf{y}}) =
\mu(y_{\mathbf{x}},y_{\mathbf{y}})
\Big[ w_1\text{exp}(-\frac{||\mathbf{x} - \mathbf{y}||^2}{2\sigma_\alpha^2}-\frac{||x_{\mathbf{x}} - x_{\mathbf{y}}||^2}{2\sigma_\beta^2}) \\
+w_2\text{exp}(-\frac{||\mathbf{x} - \mathbf{y}||^2}{2\sigma_\gamma^2}) \Big]
\end{multline}
where $\mu(y_{\mathbf{x}},y_{\mathbf{y}})=1$ if $y_{\mathbf{x}}\neq y_{\mathbf{y}}$ and 0 otherwise. Minimization of Eq.~\eqref{eq:crf} leads to a refined pseudo label $Y^U_i$ for $X^U_i$.
\subsubsection{Confidence-based \textcolor{black}{Image Weighting} of Pseudo Labels}\label{sec:method:weight}
With the new pseudo label $Y^U_i$, we further employ the confidence to update its \textcolor{black}{image-level} weight $w^U_i$ to suppress low-quality pseudo labels at the image level. We first define an \textcolor{black}{image-level uncertainty} $\mathbf{v}_i$ based on uncertainty map $V^U_i$:
\begin{align}
\mathbf{v}_i = \frac{\sum_{\mathbf{x}} V^U_{i\mathbf{x}} }{\sum_{\mathbf{x}} Y^U_{i\mathbf{x}} + \eta}
\end{align}
where $\mathbf{v}_i$ is the sum of voxel-level uncertainty normalized by the segmented lesion's volume in the image. $\eta = 10^{-5}$ is a small number for numerical stability. Let $v_{max}$ and $v_{min}$ denote the maximal and minimal values of $\mathbf{v}_i$ among all the unannotated images, we map $\mathbf{v}_i$ to the range of [0, 1]:
\begin{align}
\mathbf{v}'_i = \frac{\mathbf{v}_i - \mathbf{v}_{min} }{\mathbf{v}_{max} - \mathbf{v}_{min} }
\end{align}
Finally, the \textcolor{black}{image-level weight} $w^U_i$ is defined as:
\begin{align}\label{eq:weight}
w^U_i = (1.0 - \mathbf{v}'_i)^{1/\gamma}
\end{align}
where $\gamma \geq 1.0$ is a hyper parameter to control the non-linear mapping between the image-level uncertainty $\mathbf{v}'_i$ and the weight. We do not use $\gamma < 1.0$ as it will lead the weight of most samples to be very small (close to 0.0).
In contrast, $\gamma > 1.0$ leads the weight for most samples close to 1.0, and only samples with a high uncertainty will be strongly suppressed. In the experiment, we set $\gamma = 3.0$ according to the best performance on the validation set.
\subsubsection{Model Update with Batch Training}
With the refined pseudo labels and image-level weights obtained above, we train the network based on $\mathcal{D}^L$ and the current pseudo labels for images in $\mathcal{D}^U$, where each image is weighted by $w^{L}_i$ or $w^{U}_i$ in the segmentation loss function. The weighted loss for the entire training set is:
\begin{equation}\label{eq:loss2}
\mathcal{L}= \sum_{i = 1}^{N^L}w^L_i \ell(\mathcal{P}^L_i, Y^L_i) +
\sum_{i = 1}^{N^U}w^U_i \ell(\mathcal{P}^U_i, Y^U_i)
\end{equation}
where $\ell()$ is defined in Eq.~\eqref{eq:loss1}. $\mathcal{P}^L_i$ and $\mathcal{P}^U_i$ are the multi-scale predictions obtained by PF-Net for an annotated image and an unannotated image, respectively.
\section{Experiments and Results}
\subsection{Experimental Setting}
\subsubsection{Data}
Thoracic CT scans of 41 Male Rhesus Macaques with radiation-induced lung damage were collected with ethical committee approval. Once irradiated, each individual underwent serial CT scans for assessment of PF around every 30 days in 3 to 8 months. 133 CT scans with PF were used for experiments of the segmentation task. The CT scans have a slice thickness of 1.25~mm, with image size 512 $\times$ 512 and pixel size ranging from 0.20~mm $\times$ 0.20~mm to 0.38~mm $\times$ 0.38~mm. We randomly split the dataset at individual level into 86 scans from 27 individuals for training, 15 scans from 4 individuals for validation, and 32 scans from 10 individuals for testing. Manual annotations given by an experienced radiologist were used as the segmentation ground truth. In the training set, we used 18 scans with annotations as $\mathcal{D}^L$ and the other 68 scans as unannotated images $\mathcal{D}^U$ for semi-supervised learning, and also investigated the performance of our method with different ratios of annotated images. For preprocessing, we crop the lung region and normalize the intensity to [0, 1] using a window/level of 1500/-650.
\subsubsection{Implementation and Evaluation Metrics}
Our PF-Net\footnote{\url{https://github.com/HiLab-git/PF-Net}} and I-CRAWL framework were implemented in Pytorch with PyMIC\footnote{\url{https://github.com/HiLab-git/PyMIC}}~\cite{Wang2020c} library on a Ubuntu desktop with an NVIDIA GTX 1080 Ti GPU. The channel number parameter $N$ in PF-Net was set as 16. Dropout was only used in the encoder. The dropout rate at the first two scales of the encoder was 0 due to the small number of feature channels, and that for the last three scales was 0.3, 0.4 and 0.5, respectively.
We set the base loss function $L_s()$ as Dice loss~\cite{Milletari2016}. PF-Net was trained with Adam optimizer, weight decay of $10^{-5}$, patch size of $48 \times 192 \times 192$ and batch size of 2.
\textcolor{black}{The round number for our I-CRAWL was $K = 3$, and we use round 0 to refer to the pre-training with annotated images. In each of the following round, the pseudo labels of unannotated images acquired by CAR were kept fixed, and we used Adam optimizer to train the network for tens of thousands of iterations until the performance on validation set stopped to increase.} The learning rate was initialized as $10^{-3}$ and halved every 10k iterations. Uncertainty estimation was based on 10 forward passes of MC dropout. We used the SimpleCRF library\footnote{\url{https://github.com/HiLab-git/SimpleCRF}} to implement fully connected CRF~\cite{Krahenbuhl2011}. Following~\cite{Kamnitsas2017}, image intensity was rescaled from [0, 1] to [0, 255] before the image was sent into the CRF, and the CRF parameters were: $w_1=3$, $w_2=10$, $\sigma_\alpha=10$, $\sigma_\beta=20$ and $\sigma_\gamma=15$, which was tuned based on the validation set. $\gamma$ in Eq.~\eqref{eq:weight} was 3.0, and the performance based on different $\gamma$ values is shown in Table~\ref{tab:valid_gamma}.
For quantitative evaluation of the segmentation, we used Dice score, Relative Volume Error (RVE) and the 95-th percentile of Hausdorff Distance (HD$_{95}$) between the segmented PF lesions and the ground truth in 3D volumes. Paired t-test was used to see if two methods were significantly different.
\begin{comment}
Let $R_s$ represent the segmented region obtained by a CNN and $S_s$ denote the corresponding surface point set. The ground truth region and surface point set are denoted by $R_g$ and $S_g$ respectively.
\begin{align}
\text{Dice}(\mathcal{R}_s, \mathcal{R}_g)=\frac{2|\mathcal{R}_s\cap \mathcal{R}_g|}{|\mathcal{R}_s|+|\mathcal{R}_g|}
\end{align}
\begin{align}
\text{RVE}(\mathcal{R}_s, \mathcal{R}_g)=\frac{abs(|\mathcal{R}_s| - | \mathcal{R}_g|)}{|\mathcal{R}_g|}
\end{align}
\begin{align}
\hat{d}(\mathcal{S}_s,\mathcal{S}_g)=\max_{i\in \mathcal{S}_s}\min_{j\in\mathcal{S}_g} ||i - j ||_2
\label{eq:hd1}
\end{align}
\begin{align}
HD(\mathcal{S}_s,\mathcal{S}_g)=\max\Big(\hat{d}(\mathcal{S}_s,\mathcal{S}_g), \hat{d}(\mathcal{S}_g,\mathcal{S}_s)\Big)
\label{eq:hd2}
\end{align}
HD$_{95}$ is a variant of HD, and it uses the 95-th percentile instead of the maximal value in Eq.~\eqref{eq:hd1}.
\end{comment}
\subsection{Performance of PF-Net}
In this section, we investigate the performance of our PF-Net for pulmonary fibrosis segmentation only using the 18 annotated images for training. The results of semi-supervised learning will be demonstrated in Section~\ref{sec:result_semi}.
\subsubsection{Comparison with Existing Networks}
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{figures/net_compare.pdf}
\caption{
Qualitative comparison between different networks for PF lesion segmentation. (a) and (b) are from two different individuals. Blue arrows indicate some mis-segmented regions.}
\label{fig:net_compare}
\end{figure*}
\begin{table}
\centering
\caption{Quantitative comparison of different networks for PF lesion segmentation. \textcolor{black}{$^*$ denotes significant difference from the others ($p$-values $<$ 0.05).}
}
\label{tab:net_compare_others}
\begin{tabular
{l|c|c|c}
\hline
& Dice (\%) & RVE (\%) & HD$_{95}$ (mm) \\ \hline
2D U-Net~\cite{Ronneberger2015} & 59.31$\pm$15.66 & 38.05$\pm$19.17 & 15.73$\pm$15.57
\\
Attention U-Net~\cite{Oktay2018} & 59.22$\pm$14.55 &31.93$\pm$20.47 & 18.64$\pm$15.58
\\ \hline
3D U-Net~\cite{Abdulkadir2016} & 57.19$\pm$13.47 & 40.09$\pm$22.45 & 18.93$\pm$12.98
\\
3D scSE-Net~\cite{Roy2019} & 59.54$\pm$12.24 & 41.95$\pm$19.59 & 19.18$\pm$15.57
\\
3D V-Net~\cite{Milletari2016} & 60.37$\pm$13.33 & 35.03$\pm$27.95 & 19.22$\pm$14.65
\\ \hline
\textcolor{black}{nnU-Net~\cite{Isensee2021}} & 64.51$\pm$15.25 & 33.19$\pm$21.54 & 18.46$\pm$15.70 \\
AH-Net~\cite{Liu2018c} & 60.50$\pm$13.27 & 35.77$\pm$23.96 & 17.90$\pm$13.66 \\
VS-Net~\cite{Wang2019b} & 67.45$\pm$9.70 & 30.03$\pm$24.78 & 18.35$\pm$17.67
\\
PF-Net & \textbf{70.36$\pm$10.14$^*$} & \textbf{27.96$\pm$21.72} & \textbf{10.87$\pm$9.28$^*$} \\
\hline
\end{tabular}
\end{table}
Our PF-Net was compared with three different categories of network structures: 1) 2D CNNs including the typical 2D U-Net~\cite{Ronneberger2015} and more advanced attention U-Net~\cite{Oktay2018} that leverages spatial attention to focus more on the segmentation target; 2) 3D CNNs including 3D U-Net~\cite{Abdulkadir2016}, 3D V-Net~\cite{Milletari2016} and scSE-Net~\cite{Roy2019} that combines a 3D U-Net~\cite{Abdulkadir2016} backbone with spatial and channel ``squeeze and excitation" modules; and 3) existing networks designed for dealing with volumetric images with large inter-slice spacing (i.e., anisotropic resolution): \textcolor{black}{nnU-Net~\cite{Isensee2021} that automatically configures the network structure so that it is adapted to the given dataset,} AH-Net~\cite{Liu2018c} that transfers features learned from 2D images to 3D anisotropic volumes, and VS-Net~\cite{Wang2019b} that uses a mixture of 2D and 3D convolutions with spatial attention. Quantitative evaluation results of these methods are shown in Table~\ref{tab:net_compare_others}. We found that 3D U-Net~\cite{Abdulkadir2016} achieved a Dice score of 57.19\%, which was the lowest among the compared methods. 2D U-Net~\cite{Ronneberger2015}, Attention U-Net~\cite{Oktay2018}, 3D V-Net~\cite{Milletari2016} and 3D scSE-Net~\cite{Roy2019} achieved similar performance with Dice score around 60\%. The nnU-Net~\cite{Isensee2021}, AH-Net~\cite{Liu2018c} and VS-Net~\cite{Wang2019b} designed to deal with anisotropic resolution generally performed better than these 2D and 3D networks. Among these existing methods, VS-Net achieved the best Dice that was 67.45\%. Our PF-Net achieved Dice, RVE and HD$_{95}$ of 70.36\%, 27.96\% and 10.87 mm, respectively, where the Dice and HD$_{95}$ were significantly better than those of the other compared methods.
Fig.~\ref{fig:net_compare} shows a qualitative comparison between these networks, where (a) and (b) are from two individuals, and axial and coronal views are shown in each case. In Fig.~\ref{fig:net_compare}(a), it can be seen that 2D U-Net~\cite{Ronneberger2015}, 3D V-Net~\cite{Milletari2016} and AH-Net~\cite{Liu2018c} lead to obvious under-segmentation, as highlighted by blue arrows in the first row. VS-Net~\cite{Wang2019b} performs better than them, but the result of our PF-Net is closer to the ground truth than that of VS-Net. From the coronal view of Fig.~\ref{fig:net_compare}(b), we can observe that the result of 2D U-Net~\cite{Ronneberger2015} lacks inter-slice consistency. 3D V-Net~\cite{Milletari2016} achieves better inter-slice consistency, but it has a poor segmentation in the upper and lower regions of the lesion, as indicated by the blue arrow in the third column. AH-Net~\cite{Liu2018c}, VS-Net~\cite{Wang2019b} and our PF-Net that consider anisotropic resolution obtain better performance than the above networks purely using 2D or 3D convolutions. What's more, Fig.~\ref{fig:net_compare} shows that the lesions have complex and irregular sizes and shapes, and the proposed PF-Net is able to achieve more accurate segmentation in these cases than AH-Net~\cite{Liu2018c} and VS-Net~\cite{Wang2019b}.
\subsubsection{Ablation Study}
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{figures/crf_compare.pdf}
\caption{
Visual comparison between CRF and our proposed Confidence-Aware Refinement (CAR) for update of pseudo labels. In the uncertainty map, dark green and red colors represent low and high uncertainty values, respectively. Blue arrows highlight some local differences. }
\label{fig:crf_compare}
\end{figure*}
\begin{table}
\centering
\caption{Quantitative comparison of different variants of baseline network and our PF-Net. \textcolor{black}{Baseline ($M$) means 2D and 3D convolutions are used in the first $M$ and the other convolutional blocks, respectively. The second section is based on Baseline ($M$ = 2), and $^*$ denotes significant improvement from it ($p$-value $<$ 0.05).} Note that the last row is our PF-Net.
}
\label{tab:net_compare_ablation}
\begin{tabular
{l|c|c|c}
\hline
& Dice (\%) & RVE (\%) & HD$_{95}$ (mm) \\ \hline
\textcolor{black}{Baseline ($M$ = 0)} & 60.83$\pm$14.93 & 36.09$\pm$21.29 & 19.80$\pm$11.48
\\
\textcolor{black}{Baseline ($M$ = 1)} & 62.42$\pm$13.01 & 33.74$\pm$24.50 & 18.87$\pm$12.03
\\
Baseline ($M$ = 2) & \textbf{67.87$\pm$9.96} & \textbf{30.79$\pm$20.37} & \textbf{15.31$\pm$15.88}
\\
\textcolor{black}{Baseline ($M$ = 3)} & 65.98$\pm$12.27 & 31.45$\pm$21.84 & 16.92$\pm$12.66
\\
\textcolor{black}{Baseline ($M$ = 4)} & 65.93$\pm$14.01 & 32.37$\pm$23.17 & 17.35$\pm$12.67
\\
\textcolor{black}{Baseline ($M$ = 5)} & 63.11$\pm$10.90 & 32.71$\pm$22.35 & 17.60$\pm$12.28
\\ \hline
\textcolor{black}{+ non-local} & 68.63$\pm$15.49 & 29.82$\pm$22.27 & 14.80$\pm$13.99
\\
+ deep supervision & 68.57$\pm$10.30 & 30.73$\pm$21.24 & 17.56$\pm$19.22
\\
+ MSGA & 69.32$\pm$13.20 & \textbf{27.55$\pm$21.67} & 13.85$\pm$16.68
\\
+ MSGA + MSDA$^\circ$ & 70.25$\pm$10.09$^*$ & 28.09$\pm$19.67 & 12.28$\pm$12.12 \\
+ MSGA + MSDA & \textbf{70.36$\pm$10.14$^*$} & 27.96$\pm$21.72 & \textbf{10.87$\pm$9.28$^*$}\\
\hline
\end{tabular}
\end{table}
To investigate the effectiveness of each component of our PF-Net, we set the baseline as a naive 2.5D U-Net that extends 2D U-Net~\cite{Ronneberger2015} by using pReLU, replacing 2D convolutions with 3D convolutions at the three lowest resolution levels, and adding dropout layers to each convolutional block. \textcolor{black}{To justify the choice of using 2D convolutions at the first two resolution levels and 3D convolutions at the other resolution levels, we set $M$ to 0-5 respectively. Note that $M = 0$ and $M = 5$ correspond to pure 2D and pure 3D networks, respectively. Comparison between these variants are listed in the first section of Table~\ref{tab:net_compare_ablation}, which shows that the performance increases as $M$ changes from 0 to 2, and decreases when $M$ is 3 and larger. This is in line with our motivation to set $M$ to 2 due to the fact that the in-plane resolution is around four times of the through-plane resolution. Thus, we use baseline ($M$ = 2) in the following ablation study.}
The proposed PF-Net is referred to as baseline + \textcolor{black}{MSDA + MSGA, where MSDA is our proposed multi-scale dense attention and MSGA is the proposed multi-scale guided attention.} We compared the baseline and our PF-Net with: 1) \textcolor{black}{Baseline + non-local~\cite{Wang2017c}, where the non-local is a self-attention block inserted at the bottleneck (scale 5) of the baseline network, and it was not used at the lower scales with higher resolution due to memory constraint; } 2) Baseline + deep supervision, where the baseline network was only combined with a typical deep supervision strategy as implemented in~\cite{Dou2016miccai,Zhang2018f}; \textcolor{black}{3) Baseline + MSGA, without using MSDA; 4) Baseline + MSGA + MSDA$^\circ$, where MSDA$^\circ$ is a variant of MSDA} and it refers to $\tilde{P}_{s}$ is only sent to its next lower scale $s-1$, rather than all the lower scales in the decoder of PF-Net.
Quantitative evaluation results of these variants are listed in the second section of Table~\ref{tab:net_compare_ablation}. It shows that
compared with the baseline ($M = 2$), deep supervision improved the Dice score from 67.87\% to 68.57\%, but our MSGA was more effective with Dice of 69.32\%. Using MSDA$^\circ$ could further improve the performance, but it was less effective than our MSDA. Table~\ref{tab:net_compare_ablation} demonstrates that our PF-Net was better than the compared variants, and it significantly outperformed the baseline in terms of Dice and HD$_{95}$. \textcolor{black}{Fig.~\ref{fig:attention} presents a visualization of multi-scale attention maps of PF-Net. It shows that attention maps across scales are consistent with each other and they change from coarse to fine as the spatial resolution increases. }
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth]{figures/attention.pdf}
\caption{
\textcolor{black}{Visualization of multi-scale attention maps of PF-Net. (a)-(b) CT image and segmentation result. (c)-(f) Attention maps $~\tilde{P}_5$, $~\tilde{P}_4$, $~\tilde{P}_3$ and $~\tilde{P}_2$, respectively.}}
\label{fig:attention}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figures/uncertainty_compare.pdf}
\caption{
\textcolor{black}{Visual comparison of uncertainty obtained by different methods. (b)-(f) show two sub-regions from (a), as indicated by red and green rectangles, respectively.}}
\label{fig:uncertain_compare}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=0.8\textwidth]{figures/uncertainty_rounds.pdf}
\caption{
\textcolor{black}{Change of uncertainty maps at different rounds of I-CRAWL. (a) input image. (g) sub-region of the input. (b) ground truth (just for reference here, not used during training). (c)-(f) and (h)-(k) show the pseudo labels and the corresponding uncertainty maps of I-CRAWL as the round increases. }}
\label{fig:uncertain_rounds}
\end{figure*}
\subsection{Results of Semi-supervised Learning}\label{sec:result_semi}
\begin{table}
\centering
\caption{Performance on the validation set based on different $\gamma$ values in the confidence-based sampling weighting of pseudo labels in round 1 of I-CRAWL. ``Initial" refers to model pre-trained on annotated images (round 0). \textcolor{black}{$^*$ denotes significant improvement from it ($p$-value $<$ 0.05).}}
\label{tab:valid_gamma}
\begin{tabular
{l|c|c|c}
\hline
& Dice (\%) & RVE (\%) & HD$_{95}$ (mm) \\ \hline
Initial & 66.62$\pm$8.53 & 23.91$\pm$15.14 & 11.11$\pm$6.00
\\
No weighting & 67.83$\pm$8.25 & 22.74$\pm$17.82 & 9.12$\pm$4.48
\\
$\gamma = 1.0$ & 67.36$\pm$8.75 & 24.73$\pm$18.83 & 10.02$\pm$5.24
\\
$\gamma = 2.0$ & 68.00$\pm$8.23 & 23.01$\pm$16.90 & \textbf{8.69$\pm$4.60$^*$} \\
$\gamma = 3.0$ & \textbf{68.75$\pm$8.00$^*$} & \textbf{22.42$\pm$17.11} & 8.97$\pm$4.45$^*$\\
$\gamma = 4.0$ & 68.54$\pm$8.64$^*$ & 22.45$\pm$18.06 & 8.87$\pm$4.72$^*$\\
$\gamma = 5.0$ & 66.86$\pm$8.74 & 23.44$\pm$17.95 & 8.75$\pm$4.51$^*$\\
\hline
\end{tabular}
\end{table}
With PF-Net as the segmentation network structure, we further validate our proposed I-CRAWL for semi-supervised training. In Section~\ref{sec:icrawl_hyper-para} and \ref{sec:result_ssl_car} we used 18 annotated and 68 unannounced images, i.e., the annotation ratio was around 20\%. In Section~\ref{sec:result_ssl_ablation}, we experimented with different annotation ratios including 10\%, 20\% and 50\%.
\subsubsection{Hyper-Parameter Setting}\label{sec:icrawl_hyper-para}
First, to investigate the best value of hyper parameter $\gamma$ in Eq.~\eqref{eq:weight} that controls the weight of unannotated images, we measured the model's performance on the validation set at the first round of I-CRAWL with $\gamma$ ranging from 1.0 to 5.0. They are compared with ``Initial" that refers to model pre-trained on the annotated images and ``No weighting" that denotes treating all unannotated images equally without considering the quality of pseudo labels. Quantitative measurements shown in Table~\ref{tab:valid_gamma} demonstrate that the best $\gamma$ value was 3.0, and its corresponding Dice score was 68.75\%, which was better than 66.62\% obtained by ``Initial" and 67.83\% obtained by ``No weighting". Therefore, we set $\gamma=3.0$ in the following experiments.
\subsubsection{Uncertainty and Confidence-Aware Refinement}\label{sec:result_ssl_car}
Fig.~\ref{fig:crf_compare} shows a visual comparison between a standard fully connected CRF~\cite{Krahenbuhl2011} and our proposed CAR that leverages confidence (uncertainty) for refinement of pseudo labels. The first row of each subfigure shows an unannotated image and the pseudo label with uncertainty obtained by the CNN, and the second row shows the updated pseudo labels. In Fig.~\ref{fig:crf_compare}(a), the initial pseudo label has a large under-segmentated region, which is associated with high values in the uncertainty map, i.e., low confidence. The standard CRF only fixed the pseudo label moderately. In contrast, with the help of confidence, our CAR largely improved the pseudo label's quality by recovering the under-segmented region. In Fig.~\ref{fig:crf_compare}(b), the initial pseudo label has some over-segmentation in airways, and the uncertainty map indicates the potentially wrong segmentation in the corresponding regions well. We can observe that CAR outperformed CRF in removing the over-segmented airways.
\begin{table}
\centering
\caption{\textcolor{black}{Comparison of different uncertainty estimation methods used in CAR for the validation set in the first round. $^*$ denotes significant improvement from Initial.}}
\label{tab:unc_compare}
\begin{tabular
{l|c|c|c}
\hline
& Dice (\%) & RVE (\%) & HD$_{95}$ (mm) \\ \hline
Initial & 66.62$\pm$8.53 & 23.91$\pm$15.14 & 11.11$\pm$6.00
\\
CAR (entropy min) & 67.68$\pm$10.64 & 22.73$\pm$17.38 & 10.54$\pm$6.39
\\
CAR (Bayesian) & 68.31$\pm$9.52 & 21.80$\pm$16.53 & 10.39$\pm$6.03
\\
CAR (MC Dropout)& \textbf{69.05$\pm$9.41$^*$} & \textbf{21.31$\pm$16.10} & \textbf{9.94$\pm$6.16$^*$} \\
\hline
\end{tabular}
\end{table}
\textcolor{black}{We also compared MC Dropout~\cite{Gal2016} with two other uncertainty estimation methods: entropy minimization and Bayesian network~\cite{Jena2019}. A visual comparison of them is shown in Fig.~\ref{fig:uncertain_compare}. We found that in spite of longer time required than the other methods, MC Dropout could generate more calibrated uncertainty estimation. As shown in Fig.~\ref{fig:uncertain_compare}, the uncertain region obtained by entropy minimization and Bayesian network~\cite{Jena2019} are mostly located around the border of the segmentation output, and uncertainty map obtained by MC dropout can better indicate under- and over-segmentation regions, which is highlighted by red and yellow arrows in Fig.~\ref{fig:uncertain_compare}. For quantitative comparison, we applied CAR as a post processing method to the validation set in the first round of I-CRAWL, where these uncertainty estimation methods were used respectively. Results in Table~\ref{tab:unc_compare} show that using MC dropout for CAR improved the prediction accuracy from 66.62\% to 69.05\% in terms of Dice, which outperformed using the other two uncertainty estimation methods. }
\textcolor{black}{Fig.~\ref{fig:uncertain_rounds} shows a visualization of pseudo labels and uncertainty as the round increases. It can be observed that the initial pseudo label at round 0 has a large under-segmented region with high uncertainty. The pseudo label becomes more accurate and confident as the round increases. }
\subsubsection{Ablation Study of I-CRAWL}\label{sec:result_ssl_ablation}
For ablation study of I-CRAWL, we set the PF-Net trained only with the annotated images as a baseline, and it was compared with: 1) IT that refers to naive iterative training, where in each round pseudo label of an unannotated image is reset to the prediction given by the network without refinement, 2) IT + CRF that uses standard fully connected CRF~\cite{Krahenbuhl2011} to refine pseudo labels, 3) IT + CAR denoting that our confidence-aware refinement is used to update pseudo labels in each round, and 4) our I-CRAWL (IT + CAR + IW) where IW denotes our confidence-based image weighting of pseudo labels.
The performance of these methods at different rounds are shown in Fig.~\ref{fig:iteration_num}. Note that round 0 is the baseline, and all the methods based on iterative training performed better than the baseline. However, the improvement obtained by only using IT is slight. Using CRF or CAR to refine the pseudo labels at different rounds achieved a large improvement of Dice, and our CAR considering the voxel-level confidence of predictions outperformed the naive CRF. Weighting of pseudo labels based on image-level confidence helped to obtain more accurate result, and our I-CRAWL outperformed the other variants. Fig.~\ref{fig:iteration_num} also shows that the improvement from round 0 to round 1 of I-CRAWL is large, but the model's performance does not change much at round 2 and 3.
Table~\ref{tab:ssl_compare1} shows quantitative comparison between the baseline and variants of I-CRAWL at the end of training (round 3). It can be observed that IT's performance was not far from the baseline, with an average Dice of 70.87\% compared with 70.36\%. Using CRF and CAR improved the average Dice to 72.13\% and 72.71\%, respectively, showing the superiority of CAR compared with CRF. I-CRAWL improved the average Dice to 73.04\%, with HD$_{95}$ value of 7.92 mm in average, which was better than the other variants.
\begin{figure}
\centering
\subfloat[Dice \label{fig:1a}]
{\includegraphics[width=0.49\linewidth]{figures/iteration_dice.pdf}}
\hfill
\subfloat[HD$_{95}$ \label{fig:1b}] {\includegraphics[width=0.49\linewidth]{figures/iteration_hd95.pdf}}
\hfill
\caption{Comparison between variants of I-CRAWL at different training rounds for semi-supervised learning. Round 0 means the baseline that only learns from annotated images.}
\label{fig:iteration_num}
\end{figure}
\begin{table}
\centering
\caption{Ablation study of our semi-supervised method I-CRAWL for PF lesion segmentation. IT: iterative training. CAR: Confidence-aware refinement. IW: Image weighting of pseudo labels. \textcolor{black}{$^*$ denotes significant improvement from the baseline ($p$-value $<$ 0.05).}
}
\label{tab:ssl_compare1}
\begin{tabular
{l|c|c|c}
\hline
& Dice (\%) & RVE (\%) & HD$_{95}$ (mm) \\ \hline
Baseline & 70.36$\pm$10.14 & 27.96$\pm$21.72 & 10.87$\pm$9.28
\\
IT & 70.87$\pm$8.77 & 28.37$\pm$17.38 & 9.75$\pm$6.36
\\
IT + CRF & 72.13$\pm$8.97 & 27.38$\pm$20.39 & 8.38$\pm$4.75
\\
IT + CAR & 72.71$\pm$9.50$^*$ & \textbf{26.05$\pm$18.79} & 8.30$\pm$5.17
\\
IT + CAR + IW & \textbf{73.04$\pm$10.14$^*$} & 26.12$\pm$20.37 & \textbf{7.92$\pm$4.87$^*$} \\
\hline
\textcolor{black}{Full Supervision} &
74.58$\pm$10.99 & 23.07$\pm$15.79 & 7.03$\pm$4.09 \\
\hline
\end{tabular}
\end{table}
\subsubsection{Comparison with Existing Methods}
I-CRAWL was compared with several sate-of-the-art semi-supervised methods for medical image segmentation: 1) Fan et al.~\cite{Fan2020} that uses a randomly selected propagation strategy for semi-supervised COVID-19 lung infection segmentation, 2) Bai et al.~\cite{Bai2017a} that uses CRF to refine pseudo labels in an iterative training framework, corresponding to ``IT + CRF" described previously, 3) Cui et al.~\cite{Cui2019} that is an adapted mean teacher method, and 4) UA-MT~\cite{Yu2019} that is uncertainty-aware mean teacher. For all these methods, we used our PF-Net as the backbone network.
\begin{table}
\centering
\caption{Quantitative comparison of different semi-supervised methods for PF lesion segmentation. \textcolor{black}{$R_{an}$: Ratio of annotated images in the training set. $^\dagger$ denotes there is no significant difference from full supervision ($p$-value $>$ 0.05).}
}
\label{tab:ssl_compare2}
\begin{tabular
{l|l|c|c|c}
\hline
$R_{an}$ &Method & Dice (\%) & RVE (\%) & HD$_{95}$ (mm) \\ \hline
\multirow{6}{*}{\textcolor{black}{10\%}} & Baseline & 54.73$\pm$16.34 & 45.83$\pm$24.99 & 25.79$\pm$18.71
\\
& Fan et al.~\cite{Fan2020} & 60.44$\pm$15.86 & 42.40$\pm$21.06 & 17.44$\pm$16.00
\\
& Bai et al.~\cite{Bai2017a} & 65.03$\pm$18.01 & 32.98$\pm$23.16 & 16.82$\pm$14.82
\\
& Cui et al.~\cite{Cui2019} & 61.86$\pm$16.20 & 41.16$\pm$20.78 & 16.72$\pm$16.18
\\
& UA-MT~\cite{Yu2019} & 63.94$\pm$15.72 & 38.47$\pm$21.22 & 15.14$\pm$15.34
\\
& I-CRAWL & \textbf{67.18$\pm$14.02} & \textbf{32.12$\pm$18.50} & \textbf{11.85$\pm$9.82} \\ \hline
\multirow{6}{*}{20\%}&Baseline & 70.36$\pm$10.14 & 27.96$\pm$21.72 & 10.87$\pm$9.28
\\
& Fan et al.~\cite{Fan2020} & 71.62$\pm$10.32 & 27.57$\pm$19.09 & 10.54$\pm$8.41
\\
& Bai et al.~\cite{Bai2017a} & 72.13$\pm$8.97 & 27.38$\pm$20.39 & 8.38$\pm$4.75
\\
& Cui et al.~\cite{Cui2019} & 70.39$\pm$10.42 & 26.63$\pm$24.85 & 9.48$\pm$8.82
\\
& UA-MT~\cite{Yu2019} & 70.46$\pm$9.53 & 27.14$\pm$23.17 & 8.25$\pm$4.93
\\
& I-CRAWL & \textbf{73.04$\pm$10.14} & \textbf{26.12$\pm$20.37} & \textbf{7.92$\pm$4.87$^\dagger$} \\ \hline
\multirow{6}{*}{\textcolor{black}{50\%}}&Baseline & 72.09$\pm$12.54 & 26.49$\pm$23.19 & 9.38$\pm$6.43
\\
& Fan et al.~\cite{Fan2020} & 72.69$\pm$11.41 & 26.46$\pm$17.71 & 9.19$\pm$6.45
\\
& Bai et al.~\cite{Bai2017a} & 72.92$\pm$10.41 & 25.71$\pm$19.97$^\dagger$ & 7.45$\pm$4.98$^\dagger$
\\
& Cui et al.~\cite{Cui2019} & 72.72$\pm$12.44 & 24.57$\pm$19.96$^\dagger$ & 8.11$\pm$5.76$^\dagger$
\\
& UA-MT~\cite{Yu2019} & 73.40$\pm$11.48 & 24.67$\pm$19.08$^\dagger$ & 7.86$\pm$4.56$^\dagger$
\\
& I-CRAWL & \textbf{74.02$\pm$13.77$^\dagger$} & \textbf{23.93$\pm$16.75$^\dagger$} & \textbf{7.24$\pm$5.03}$^\dagger$ \\ \hline
\multicolumn{2}{l|}{\textcolor{black}{Full Supervision}} &
74.58$\pm$10.99 & 23.07$\pm$ 15.79 & 7.03$\pm$4.09 \\
\hline
\end{tabular}
\end{table}
\textcolor{black}{We investigated the performance of these methods with different ratios of labeled data: 10\%, 20\% and 50\%.} For each setting, the baseline was learning only from the labeled images, \textcolor{black}{and the upper bound was ``full supervision" where 100\% training images were labeled.} The results are shown in Table~\ref{tab:ssl_compare2}. It can be observed that with only 10\% images annotated, the baseline performed poorly with an average Dice of 54.73\%. Our I-CRAWL improved it to 67.18\%, which largely outperformed the other methods. When 20\% images were annotated, our method improved the average Dice from 70.36\% to 73.04\% compared with the baseline, while the mean teacher-based methods did not bring much performance gain. When 50\% images were annotated, our method also outperformed the others, and it was comparable with full supervision (74.02$\pm$13.77\% versus 74.58$\pm$10.99\% in terms of Dice, with $p$-value $>$ 0.05).
\section{Discussion and Conclusion}
\textcolor{black}{Our PF-Net combines 2D and 3D convolutions to deal with anisotropic resolutions, and we set $M=2$ as the in-plane resolution is four times of the through-plane resolution in our dataset. It may be set to other values according to the spacing information of different datasets. Multi-scale guided dense attention in PF-Net is important for dealing with PF lesions with various positions, shapes and scales. We noticed that Sinha et al.~\cite{Sinha2021} also proposed a multi-scale attention, but it has key differences from ours. First, Sinha et al.~\cite{Sinha2021} concatenated feature at different levels of the encoder to obtain a multi-scale feature, which is used as input for parallel attention modules at different scales. While PF-Net learns multi-scale attentions sequentially, where attention at a lower resolution level is used as input for all the higher resolution levels with dense connections. Second, Sinha et al.~\cite{Sinha2021} used self-attention inspired by non-local block~\cite{Wang2017c} that is computationally expensive with large memory consumption, while PF-Net uses convolution to obtain the attention coefficients at different spatial positions, which has higher memory and computational efficiency. In addition, for each attention module, the attention maps are calculated in two steps in~\cite{Sinha2021}, and a consistency between the two steps is imposed via an L2 distance of their encoded representations, which is called guided attention by the authors. In contrast, guided attention in PF-Net refers to supervising attention maps directly by the resampled segmentation ground truth. }
\textcolor{black}{Our I-CRAWL is a pseudo label-based method for semi-supervised learning. Despite that pseudo label has been previously investigated~\cite{Bai2017a,Fan2020}, I-CRAWL is superior to these works mainly for the following reasons. First, the pseudo labels generated by a model trained with a small set of annotated images inevitably contain a lot of inaccurate predictions. Improving the quality of pseudo labels would benefit the final segmentation model. However, the method in~\cite{Fan2020} does not refine pseudo labels, and Bai et al.~\cite{Bai2017a} refines pseudo labels by CRF without considering their confidence. Our method employs uncertainty estimation to find uncertain regions that are likely to be mis-segmented, and the confidence-aware refinement is more effective to refine these mis-segmented regions, leading to improved accuracy of pseudo labels. Second, the quality of pseudo labels of different images varies a lot, and it is important to exclude low-quality pseudo labels that may corrupt the segmentation model. However, Bai et al.~\cite{Bai2017a} and~\cite{Fan2020} ignored this point and treated all the pseudo labels equally. In contrast, I-CRAWL uses image-level uncertainty information to highlight more confident pseudo labels and down-weight uncertain ones that are unreliable. Thus, the model is less affected by low-quality pseudo labels.}
\textcolor{black}{Uncertainty estimation plays an important role in our I-CRAWL framework. We found that the simple yet effective MC dropout performed better than alternatives including entropy minimization and Bayesian networks~\cite{Jena2019}. Despite that MC Dropout is slow for uncertainty estimation, it is used offline at the beginning of each round of our method and takes a short time compared with the batch training step. In the scenario of $20\%$ annotated data, our uncertainty estimation takes 8.72 minutes (7.69s per 3D image), and CAR and image weighting take 12.98 minutes (11.45s per 3D image). In contrast, the model update with batch training takes around 4 hours for each round, i.e., the first three steps of I-CRAWL account for 8.29\% of the entire runtime of each round, which could be further accelerated by multi-thread parallel computing. Therefore, our uncertainty estimation and CAR require little extra time compared with naive iterative training and Bai et al.~\cite{Bai2017a}. }
In conclusion, we present a novel 2.5D network structure and an uncertainty-based semi-supervised learning method for automatic segmentation of pulmonary fibrosis from anisotropic CT scans. To deal with complex PF lesions with irregular structures and appearance in CT volumes with anisotropic 3D resolution, we propose PF-Net that combines a 2.5D network baseline with multi-scale guided dense attention. To leverage unannotated images for learning, we propose I-CRAWL that is an iterative training framework, where a confidence-aware refinement process is introduced to update pseudo labels and a confidence-based image weighting is proposed to suppress images with low-quality pseudo labels. Experimental results with lung CT scans from Rhesus Macaques showed that our PF-Net outperformed existing 2D, 3D and 2.5D networks for PF lesion segmentation, and our I-CRAWL could better leverage unannotated images for training than state-of-the-art semi-supervised methods. Our methods can be extended to deal with other structures and human CT scans in the future.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
\section{Introduction}\label{sec:introduction}}
\else
\section{Introduction}
\label{sec:introduction}
\fi
\IEEEPARstart{T}{his} demo file is intended to serve as a ``starter file''
for IEEE Computer Society journal papers produced under \LaTeX\ using
IEEEtran.cls version 1.8b and later.
I wish you the best of success.
\hfill mds
\hfill August 26, 2015
\subsection{Subsection Heading Here}
Subsection text here.
\subsubsection{Subsubsection Heading Here}
Subsubsection text here.
\section{Conclusion}
The conclusion goes here.
\appendices
\section{Proof of the First Zonklar Equation}
Appendix one text goes here.
\section{}
Appendix two text goes here.
\ifCLASSOPTIONcompsoc
\section*{Acknowledgments}
\else
\section*{Acknowledgment}
\fi
The authors would like to thank...
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\section{Introduction}
This demo file is intended to serve as a ``starter file''
for IEEE conference papers produced under \LaTeX\ using
IEEEtran.cls version 1.8b and later.
I wish you the best of success.
\hfill mds
\hfill August 26, 2015
\subsection{Subsection Heading Here}
Subsection text here.
\subsubsection{Subsubsection Heading Here}
Subsubsection text here.
\section{Conclusion}
The conclusion goes here.
\section*{Acknowledgment}
The authors would like to thank...
\section{Introduction}
This demo file is intended to serve as a ``starter file''
for IEEE Computer Society conference papers produced under \LaTeX\ using
IEEEtran.cls version 1.8b and later.
I wish you the best of success.
\hfill mds
\hfill August 26, 2015
\subsection{Subsection Heading Here}
Subsection text here.
\subsubsection{Subsubsection Heading Here}
Subsubsection text here.
\section{Conclusion}
The conclusion goes here.
\ifCLASSOPTIONcompsoc
\section*{Acknowledgments}
\else
\section*{Acknowledgment}
\fi
The authors would like to thank...
|
\section{Introduction}
Let $R$ be a commutative integral domain of characteristic 0, and let $R\la X,Y\ra$ be
the free associative algebra of two (non-commutative)
variables $X,Y$ over $R$.
For $u,v\in R\la X,Y\ra$, we shall write $[u,v]$
to denote the Lie bracket $uv-vu$.
In \cite{M37}, W.Magnus introduced the associative subalgebra
$S_X\subset R\la X,Y\ra$ generated by
(what are called) the elements arising by elimination of $X$:
\begin{equation}
Y^{(0)}:=Y, \quad
Y^{(k+1)}:=[X,Y^{(k)}] \quad
(k=0,1,2,\dots),
\end{equation}
and showed that $S_X$ is freely generated by the $Y^{(k)}$
$(k=0,1,2,\dots)$.
Moreover, he derived that every element $Z$ of $R\la X,Y\ra$ can be
written uniquely in the form
\begin{equation} \label{1st-elimination}
Z=\alpha_0 X^m +s_1 X^{m-1}+\cdots+s_m,
\end{equation}
where $\alpha_0\in R$, $s_1,\dots, s_m\in S_X$
(see \cite[Hilfssatz 2]{M37}, \cite[Lemma 5.6]{MKS}).
This observation is the first step preceding to repeated elimination
for the construction
of the basic Lie elements (an ordered basis of free Lie algebra)
whose powered products in decreasing orders give Poincare-Birkoff-Witt
basis of the enveloping algebra $R\la X,Y\ra$ (\cite[Theorem 5.8]{MKS}).
Apparently, this theory was historically a starting point
toward subsequent developments of
finer constructions of free Lie algebra bases
due to Lazard, Hall, Lyndon, Viennot, and so on
(cf. e.g., \cite[Notes 4.5, 5.7]{Reu}).
In this note, we however stay on the first step of elimination
(\ref{1st-elimination}) and attempt to watch a combinatorial aspect
with respect to a certain explicit $R$-linear basis
``$\{\MM^{(\bk)}\}_{\bk\in \N_0^{(\infty)}}$'' of $R\la X,Y\ra$
(to be called the Magnus polynomials below)
designed as follows:
\begin{Notation}
Let $\N_0$ denote the set of non-negative integers, and
let
$$
\N_0^{(\infty)}:=\bigcup_{d=0}^\infty
\left(
\prod_{k=1}^{d}\N_0\right)
\times \N_0
$$
be the collection of finite sequences
$
\bk=(k_1,\dots,k_d; k_\infty)
$
of non-negative integers equipped
with a special last entry $k_\infty\in \N_0$.
Here, we suppose $(;k_\infty)$ also as elements of
$\N_0^{(\infty)}$ coming from $d=0$.
For $\bk\in \N_0^{(\infty)}$, define
$|\bk|:=\sum_{k=i}^\infty k_i=k_1+\cdots+k_d+k_\infty$
(resp. $\dep(k):=d$),
and call it the size (resp. depth) of $\bk$.
\end{Notation}
\begin{Definition}[Magnus polynomial]
\label{defMagnus}
For $\bk=(k_1,\dots,k_d; k_\infty)\in\N_0^{(\infty)}$, define
$$
\MM^{(\bk)}:=Y^{(k_1)}\cdots Y^{(k_d)} \cdot X^{k_\infty} \in R\la X,Y\ra.
$$
We also set $\MM^{(;0)}=1$, $\MM^{(;k)}=X^k$ ($k=1,2,\dots)$.
Note that $\MM^{(k;0)}=Y^{(k)}$ for $k\ge 0$.
\end{Definition}
\begin{Example}
$\MM^{(1,0;2)}=Y^{(1)}Y^{(0)}X^2=(XY-YX)YX^2=XY^2X^2-YXYX^2$.
\end{Example}
It is not difficult to see that the Magnus polynomial $\MM^{(\bk)}\in R\la X,Y\ra$
is homogeneous
of bidegree $(|\bk|,\dep(\bk))$ in $X$ and $Y$.
The above mentioned Magnus expression (\ref{1st-elimination})
can then be rephrased as
\begin{equation} \label{Magnus}
Z=\sum_{\bk\in \N_0^{(\infty)}}
\alpha_{\bk}\, \MM^{(\bk)}
\end{equation}
with uniquely determined coefficients $\alpha_{\bk}\in R$ for any given
$Z\in R\la X,Y\ra$.
In other words, the collection $\{\MM^{(\bk)}\mid \bk\in \N_0^{(\infty)}\}$
forms an $R$-linear basis of $R\la X,Y\ra$.
Below in \S 2, we will construct another $R$-linear basis
$\{\SS^{(\bk)}\mid \bk\in \N_0^{(\infty)}\}$
(formed by what we call the `demi-shuffle' polynomials)
and show that
$\{\MM^{(\bk)}\}_\bk$
and $\{\SS^{(\bk)}\}_\bk$
are dual to each other under the standard pairing
with respect to the monomials of $R\la X,Y\ra$
(Theorem \ref{orthogonality}).
We then in \S 3 shortly generalize the duality to the case of
free associative algebras of more variables (Theorem
\ref{orthogonality2}).
In \S 4, we apply the formation of dual basis
to derive a formula of Le-Murakami, Furusho type
that expresses arbitrary coefficients of a group-like series
$J\in R\lala X,Y\rara$ by the `regular' coefficients of $J$
(Theorem \ref{LeMuFu}).
\section{Demi-shuffle duals and array binomial coefficients}
Let $W$ be the subset of $R\la X,Y\ra$ formed by the monomials in $X,Y$
together with $1$, and call any element of $W$ a word.
It is clear that $W$ forms a free monoid by the concatenation product
that restricts the multiplication of $R\la X,Y\ra$.
Each element of $R\la X,Y\ra$ is an $R$-linear combination of words in $W$.
For two elements $u,v\in R\la X,Y\ra$, define the standard pairing $\lp u,v\rp\in R$
so as to extend $R$-linearly the Kronecker symbol $\lp w,w'\rp:=\delta_{w}^{w'}\in\{0,1\}$
for words $w,w'\in W$.
\begin{Notation} We use the notation $w_\bk:=X^{k_1}Y\cdots X^{k_d}YX^{k_\infty}$
and call it the word associated to $\bk=(k_1,\dots,k_d; k_\infty)\in\N_0^{(\infty)}$.
The mapping $\bk\mapsto w_\bk$ gives a bijection between $\N_0^{(\infty)}$ onto $W$.
(Note that $w_{(;0)}=1$.)
The standard pairing $\lp w_\bk,w_{\bk'}\rp$ is equal to $0$ or $1$
according to whether $\bk\ne \bk'$ or $\bk=\bk'$.
\end{Notation}
The purpose of this section is to describe the duals of Magnus basis
$\{\MM^{(\bk)}\}_{\bk\in \N_0^{(\infty)}}$ with respect to the standard
pairing.
\begin{Definition}[Demi-shuffle polynomial]
\label{demi-shuffle}
For $\bk=(k_1,\dots,k_d; k_\infty)\in\N_0^{(\infty)}$, define
$$
\SS^{(\bk)}:=
(\cdots((X^{k_1}Y)\sha X^{k_2})Y)\sha\cdots\sha ) X^{k_d})Y)\sha
X^{k_\infty} \in R\la X,Y\ra,
$$
where $\sha$ denotes the usual shuffle product.
We also set $\SS^{(;0)}=1$, $\SS^{(;k)}=X^k$ ($k=1,2,\dots)$.
Note that $\SS^{(k;0)}=X^kY$ for $k\ge 0$.
\end{Definition}
The construction of $\SS^{(\bk)}$ can be interpreted
as forming the linear sum
of all words obtained from
the word
$w_\bk=X^{k_1}Y\cdots X^{k_d}YX^{k_\infty}$
by consecutively applying `left shuffles' of letters $X$'s
and `concatenations' of letters $Y$'s in $w_\bk$.
\begin{Example}
Here are a few examples:
$\SS^{(0,1;0)}=(Y\sha X)Y=YXY+XYY$;
$\SS^{(1,1;0)}=((XY)\sha X)Y=XYXY+2XXYY$;
$\SS^{(1,0,1;0)}=(((XY)Y)\sha X)Y=XYYXY+XYXY^2+2X^2Y^3$.
Using the first identity, one can also compute
\begin{align*}
&\SS^{(0,1;1)}=((Y\sha X)Y)\sha X=(YXY+XYY)\sha X \\
&=(YXYX+2YXXY+XYXY)+(XYYX+XYXY+2XXYY) \\
&=2XXYY+2XYXY+XYYX+2YXXY+YXYX.
\end{align*}
\end{Example}
\begin{Theorem}[Duality] \label{orthogonality}
For $\bt, \bk\in \N_0^{(\infty)}$, we have
$$
\lp \SS^{(\bt)},\MM^{(\bk)}\rp=\delta_\bt^{\bk}.
$$
Here $\delta_\bt^{\bk}$ is the Kronecker symbol, i.e., designating $0$ or $1$
according to whether $\bt\ne\bk$ or $\bt=\bk$ respectively.
\end{Theorem}
Before going to the proof of the above theorem, we introduce the following
notation.
\begin{Definition}[Array binomial coefficient]
For $\bt,\bk\in \N_0^{(\infty)}$ with $\dep(\bt)=\dep(\bk)$,
$|\bt|=|\bk|$, define
\begin{equation} \label{ArrayBino}
\binom{\bt}{\bk}:=
\binom{t_1}{k_1}
\binom{t_1+t_2-k_1}{k_2}
\cdots
\binom{t_1+\cdots t_d-k_1-\cdots-k_{d-1}}{k_d},
\end{equation}
where $\bt=(t_1,\dots,t_d,t_\infty)$,
$\bk=(k_1,\dots,k_d,k_\infty)$.
We set $\binom{\bt}{\bk}:=0$ if either $\dep(\bt)\ne\dep(\bk)$ or
$|\bt|\ne|\bk|$ holds.
\end{Definition}
\begin{Remark}
\label{rem2.6}
The special case $\binom{N,0,\dots,0;0}{k_1,k_2,\dots,k_d;k_\infty}$
is the same as the usual multinomial coefficient
$\binom{N}{k_1,k_2,\dots,k_d,k_\infty}$ in combinatorics.
Note also that $\binom{\bt}{\bk}\ne 0$ implies $t_\infty\le k_\infty$,
as the last factor of $\binom{\bt}{\bk}$ could survive only when
$(t_1+\cdots t_d-k_1-\cdots-k_{d-1})-k_d=k_\infty-t_\infty\ge 0$.
\end{Remark}
It turns out that the array binomial coefficients give
the expansion of $\SS^{(\bt)}$ as a linear sum of the monomials in $W$.
Recall that, for $\bt=(t_1,\dots,t_d,t_\infty)\in \N_0^{(\infty)}$,
$w_\bt$ denotes the word $X^{t_1}YX^{t_2}Y\cdots X^{t_d}YX^{t_\infty}\in W$.
\begin{Lemma}[Monomial expansion]
\label{dS-mono}
$$
\SS^{(\bk)}=\sum_{\bt \in\N_0^{(\infty)}}
\binom{\bt}{\bk} w_{\bt}.
$$
\end{Lemma}
\begin{proof}
This follows from elementary combinatorics of counting: Consider letters $Y$
as partitions of $X$-boxes. The $i$-th binomial
factor of (\ref{ArrayBino})
represents the number of ways to choose those places of $X$ in $w_\bt$
originated from the old box $YX^{k_i}Y$ among those $t_i$ pieces of $X$
in the box $YX^{t_i}Y$ or
$(t_1-k_1)+\cdots+(t_{i-1}-k_{i-1})$ pieces of transported $X$'s
in the other left boxes $X^{t_1}Y$, $YX^{t_j}Y$ $(j=2,\dots,i-1)$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{orthogonality}]
It is not difficult to see from the formula
$Y^{(k)}=\sum_{i=0}^k (-1)^i\binom{k}{i}X^{k-i}YX^i$ (\cite[(4)]{M37})
that the expansion of the Magnus polynomial in monomials
is given by
\begin{equation} \label{Mag-Mono}
\MM^{(\bk)}=\sum_{\bt \in\N_0^{(\infty)}}
\magnus{\bk}{\bt} w_{\bt}
\end{equation}
with
\begin{equation}
\label{MagnusBino}
\magnus{\bk}{\bt}:=(-1)^{\sum_{i=1}^d (d-i+1)(k_i-t_i)}
\binom{k_1}{k_1-t_1}
\binom{k_2}{k_1+k_2-t_1-t_2} \cdots
\binom{k_d}{\sum_{i=1}^d(k_i-t_i)}
\end{equation}
for $\bt:=(t_1,\dots, t_d;t_\infty)$, $\bk:=(k_1,\dots,k_d;k_\infty)$.
Since $\lp\SS^{(\bt)},\MM^{(\bk)}\rp=\sum_{\bu\in\N_0^{(\infty)}}
\lp\SS^{(\bt)},w_\bu\rp \lp\MM^{(\bk)},w_\bu\rp$, it suffices to show
\begin{equation} \label{target}
\sum_\bu \magnus{\bk}{\bu} \binom{\bu}{\bt}=\delta_\bk^\bt.
\end{equation}
Noting that non-zero pairing $\lp\SS^{(\bt)},\MM^{(\bk)}\rp$ occurs only when
$|\bt|=|\bk|$, $\dep(\bt)=\dep(\bk)$, without loss of generality, we may
assume that $\bu$ in the above summation also runs over those
with the fixed size $N:=|\bt|=|\bk|$ and depth $d:=\dep(\bt)=\dep(\bk)$.
Then, the summation $\sum_\bu$ with $\bu=(u_1,\dots,u_d;u_\infty)$
has $d$ independent parameters $u_1,\dots, u_d$ that determine
$u_\infty=N-\sum_{i=1}^d u_i$. We may also regard each $u_i$ running over $\Z$,
as the coefficients $\magnus{\bk}{\bu}$, $\binom{\bu}{\bt}$ vanishes when
combinatorial meaning is lost.
Then, in the summation $\sum_{(u_1,\dots,u_d)\in\Z^d}$ in (\ref{target}),
the partial factor of summation
involved with the last parameter $u_d$ can be factored out in the form:
\begin{align*}
&
\sum_{u_d\in\Z}(-1)^{-u_d}
\binom{k_d}{u_d+\sum_{i=1}^{d-1}(u_i-k_i)}
\binom{u_d+\sum_{i=1}^{d-1}(u_i-t_i)}{t_d} \\
&=
(-1)^{\sum_{i=1}^{d-1}(u_i-k_i)-k_d}
\binom{\sum_{i=1}^{d-1}(k_i-t_i)}{t_d-k_d}.
\end{align*}
(Use \cite[(5.2.4)]{GKP}.)
Repeating this process inductively on $d$, we eventually find
$$
\lp\SS^{(\bt)},\MM^{(\bk)}\rp=
\binom{0}{t_1-k_1}
\binom{k_1-t_1}{t_2-k_2}
\binom{k_1+k_2-t_1-t_2}{t_3-k_3}
\cdots
\binom{\sum_{i=1}^{d-1}(k_i-t_i)}{t_d-k_d}
$$
which is equal to $\delta^\bk_\bt$ as desired.
\end{proof}
\begin{Corollary} \label{coro}
Each element $u\in R\la X,Y\ra$ can be written as
$$
u=\sum_{\bk \in\N_0^{(\infty)}}
\lp\SS^{(\bk)}, u\rp \,\MM^{(\bk)}
=\sum_{\bk \in\N_0^{(\infty)}}\lp\MM^{(\bk)}, u\rp \,\SS^{(\bk)}.
$$
\end{Corollary}
Note that the summation in RHS is essentially a finite sum.
\section{Generalization to the case $R\la X,Y_1,Y_2,\cdots \ra$}
It is not difficult to generalize the above duality in $R\la X,Y\ra$
(Theorem \ref{orthogonality})
to similar duality in $R\la X,Y_\lambda \ra_{\lambda\in \Lambda}$
($\Lambda$: a nonempty index set), viz. in
the associative algebra freely generated by
the symbols $X,Y_\lambda$ $(\lambda\in\Lambda)$ over $R$.
In fact, introducing
\begin{equation}
Y_\lambda^{(0)}:=Y_\lambda, \quad
Y_\lambda^{(k+1)}:=[X,Y_\lambda^{(k)}] \quad
(\lambda\in\Lambda, k=0,1,2,\dots)
\end{equation}
that are called the elements arising by elimination of $X$,
Magnus (\cite[Hilfssatz 2]{M37}, \cite[Lemma 5.6]{MKS})
showed that every element $Z$ of $R\la X,Y_\lambda \ra_{\lambda\in \Lambda}$
has the unique expression (\ref{1st-elimination})
with
$S_X$
the subalgebra
freely generated by the
$Y_\lambda^{(k)}$ ($k\in\N_0,\lambda\in\Lambda$).
\begin{Definition}[Depth-varied Magnus/demi-shuffle polynomials and monomials]
\label{def3.1}
Let $d$ be a positive integer.
For $\bk=(k_1,\dots,k_d; k_\infty)\in\N_0^{(\infty)}$ and a
finite sequence $\blambda=(\lambda_1,\dots,\lambda_d)\in\Lambda^d$,
define
\begin{align*}
&\MM^{(\bk,\blambda)}:=Y_{\lambda_1}^{(k_1)}\cdots Y_{\lambda_d}^{(k_d)} \cdot X^{k_\infty};\\
&\SS^{(\bk,\blambda)}:=(\cdots((X^{k_1}Y_{\lambda_1})\sha X^{k_2})Y_{\lambda_2})\sha\cdots\sha ) X^{k_d})Y_{\lambda_d})\sha X^{k_\infty}; \\
&w_{\bk,\blambda}:=X^{k_1}Y_{\lambda_1}\cdots X^{k_d}Y_{\lambda_d}X^{k_\infty}.
\end{align*}
For $d=0$ with $\bk=(;k)$, $\blambda=()$,
we simply set $w_{(;k),()}=\MM^{((;k),())}=\SS^{((;k),())}=X^k$.
\end{Definition}
Note that the monomials $w_{\bk,\blambda}$ ($\bk\in\N_0^{(\infty)}$,
$\blambda\in\Lambda^{\dep(\bk)})$
form an $R$-linear basis of $R\la X,Y_\lambda \ra_{\lambda\in \Lambda}$.
Let us write $\la\ ,\ \ra$ for the standard pairing defined by the Kronecker
symbol with respect to these monomials.
\begin{Theorem}[Duality] \label{orthogonality2}
For $\bt, \bk\in \N_0^{(\infty)}$ and $\blambda\in\Lambda^{\dep(\bt)}$,
$\bmu\in\Lambda^{\dep(\bk)}$, we have
$$
\lp \SS^{(\bt,\blambda)},\MM^{(\bk,\bmu)}\rp=\delta_{(\bt,\blambda)}^{(\bk,\bmu)}.
$$
Here $\delta_{(\bt,\blambda)}^{(\bk,\bmu)}$ is the Kronecker symbol,
i.e., designating $1$ or $0$
according to whether the pairs $(\bt,\blambda)$ and $(\bk,\bmu)$
coincide or not respectively.
\end{Theorem}
\begin{proof}
Given a fixed $\blambda=(\lambda_1,\dots,\lambda_d)\in\Lambda^d$,
let $V_\blambda$ be the $R$-linear
subspace of $R\la X,Y_\lambda \ra_{\lambda\in \Lambda}$ generated by
the monomials $\{w_{\bk,\blambda}\mid\bk\in\N_0^{(\infty)}, \dep(\bk)=d\}$.
It is obvious that if $\blambda\ne\bmu$ then
$V_\blambda$ and $V_\bmu$ are mutually orthogonal under the standard pairing $\la\ ,\ \ra$.
Since $\MM^{(\bk,\bmu)}\in V_\bmu$, $\SS^{(\bt,\blambda)}\in V_\blambda$,
we only need to look at the case $\bmu=\blambda\in \Lambda^d$.
Consider the $R$-linear subspace $V_d$ of $R\la X,Y\ra$ generated by
$\{w_\bk\mid \bk\in \N_0^{(\infty)}, \dep(\bk)=d\}$. Then, the mapping
$w_\bk\mapsto w_{\bk,\blambda}$ defines an isometry, i.e., an $R$-linear
isomorphism $\phi_\blambda:V_d \isom V_\blambda$ preserving $\la\ ,\ \ra$.
The assertion then follows at once from Theorem \ref{orthogonality}
after observing $\phi_\blambda( \SS^{(\bt)})=\SS^{(\bt,\blambda)}$ and
$\phi_\blambda(\MM^{(\bk)})=\MM^{(\bk,\blambda)}$.
\end{proof}
\section{Application to a formula of Le-Murakami and Furusho type}
In this section, we assume that $R$ is a field and consider
$R\la X,Y\ra$ as a subalgebra of the ring of non-commutative
formal power series $R\lala X,Y\rara$, where a standard
comultiplication $\Delta$ is defined by setting
$\Delta(a)=1\otimes a+a\otimes 1$ for $a\in\{X,Y\}$.
An element $J\in R\lala X,Y\rara$ is called group-like if
it has constant term 1 and satisfies $\Delta(J)=J\otimes J$.
There are many group-like elements;
for example, the subgroup multiplicatively generated by $\exp(X)$ and $\exp(Y)$
in $R\lala X,Y\rara^\times$ consists of group-like elements and forms a free group of rank 2.
\begin{Theorem}[Le-Murakami, Furusho type formula]
\label{LeMuFu}
Let $J\in R\lala X,Y\rara$ be a group-like element in the form
$$
J=\sum_{\bk\in\N_0^{(\infty)}} c_\bk w_\bk,
$$
and write $c_X$ for the coefficient $c_{(;1)}$ of $X$ in $J$.
Then,
$$
c_{(k_1,\dots,k_d;k_\infty)}=\sum_{\substack{s,t\ge 0 \\ s+t=k_\infty}}
(-1)^s
\frac{{(c_X)}^t}{t!}
\sum_{\substack{ s_1,\dots,s_d\ge 0 \\ s=s_1+\dots+s_d}}
\binom{k_1+s_1}{k_1}\cdots \binom{k_d+s_d}{k_d}
c_{(k_1+s_1,\dots,k_d+s_d;0)}
\ .
$$
\end{Theorem}
We first prove an elementary identity that will be used for the proof of the above formula.
\begin{Lemma} \label{lem2}
Let $\bkappa=(k_1,\dots,k_d)\in\N_0^d$ and $\bs=(s_1,\dots,s_d)\in \Z^d$
satisfy $s=s_1+\cdots+s_d\ge 0$ and $k_i+s_i\ge 0$ $(i=1,\dots,d)$.
Then, we have
$$
\sum_{\btau\in\N_0^d}
\lp\SS^{(\btau;0)},w_{(\bkappa+\bs;0)}\rp
\cdot \lp\MM^{(\btau;0)}, w_{(\bkappa;s)}\rp
=(-1)^s
\binom{k_1+s_1}{k_1}\cdots \binom{k_d+s_d}{k_d}.
$$
\end{Lemma}
\begin{proof}
We shall compute the LHS explicitly with
$$
\la\SS^{(\btau;0)},w_{(\bkappa+\bs;0)}\ra=\binom{(\bkappa+\bs;0)}{(\btau;0)}
=\binom{k_1+s_1}{t_1}
\cdots
\binom{\sum_{i=1}^{d-1}(k_i+s_i)-\sum_{i=1}^{d-2}t_i}{t_{d-1}}
\binom{t_d}{t_d}
$$
by Lemma \ref{dS-mono} and with
$$
\la\MM^{(\btau;0)}, w_{(\bkappa;s)}\ra
=\magnus{(\btau;0)}{(\bkappa;s)}
=(-1)^{s+\sum_{i=1}^{d-1}(d-i)(t_i-k_i)}
\binom{t_1}{t_1-k_1}
\cdots
\binom{t_{d-1}}{\sum_{i=1}^{d-1}(t_i-k_i)}
\binom{t_d}{s}
$$
by (\ref{Mag-Mono}).
As $\binom{(\bkappa+\bs;0)}{(\btau;0)}\magnus{(\btau;0)}{(\bkappa;s)}\ne 0$
only when all entries of $\btau=(t_1,\dots,t_d)$ are nonnegative
and $t_1+\cdots+t_d=\sum_{i=1}^d (k_i+s_i)$ (constant), the summation
$\sum_\btau \binom{(\bkappa+\bs;0)}{(\btau;0)}\magnus{(\btau;0)}{(\bkappa;s)}$
can be taken over the tuples $(t_1,\dots,t_{d-1})\in\Z^{d-1}$
with entries running as independent variables.
As $s=\sum_{i=1}^{d}(t_i-k_i)$, the partial sum involved with the last variable
$t_{d-1}$ may be factored out as
\begin{align*}
&\sum_{t_{d-1}}(-1)^{t_{d-1}}
\binom{\sum_{i=1}^{d-1}(k_i+s_i)-\sum_{i=1}^{d-2}t_i}{t_{d-1}}
\binom{t_{d-1}}{\sum_{i=1}^{d-1}(t_i-k_i)}
\binom{t_d}{s} \\
&=\sum_{t_{d-1}}(-1)^{t_{d-1}}
\binom{\sum_{i=1}^{d-1}(k_i+s_i)-\sum_{i=1}^{d-2}t_i}{ \sum_{i=1}^{d-1}k_i-\sum_{i=1}^{d-2}t_i }
\binom{\sum_{i=1}^{d-1} s_i}{\sum_{i=1}^{d-1}(t_i-k_i)}
\binom{\sum_{i=1}^d(k_i+s_i)-\sum_{i=1}^{d-1}t_i}{s} \\
&=\binom{\sum_{i=1}^{d-1}(k_i+s_i)-\sum_{i=1}^{d-2}t_i}{ \sum_{i=1}^{d-1}k_i-\sum_{i=1}^{d-2}t_i }
(-1)^{\sum_{i=1}^{d-1}k_i -\sum_{i=1}^{d-2} t_i}
\binom{k_d+s_d}{s_d},
\end{align*}
where \cite[(5.21)]{GKP} is applied for the first equality and
\cite[(5.24)]{GKP} for the second.
After factoring out the constant $\binom{k_d+s_d}{s_d}$ and repeating
the similar process with the other variables $t_{d-2},\dots, t_1$ consecutively,
we eventually obtain the asserted formula.
Below in Note \ref{Quest4.3}, we also provide an alternative proof of the lemma
free from intricate use of
\cite[(5.21),(5.24)]{GKP}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{LeMuFu}]
We argue in the beautiful framework exploited in Reutenauer's book \cite[1.5]{Reu}
using the complete tensor product
$$
\mathscr{A}=R\lala X,Y\rara \bar\otimes R\lala X,Y\rara
$$
equipped with a product induced from
the shuffle product (resp. the concatenation product)
on the left (resp. right) of $\bar\otimes$.
Recall that the ring of
$R$-linear endomorphism $\mathrm{End}_RR\lala X,Y\rara$
can be embedded into $\mathscr{A}$ by
$f\mapsto \sum_{w\in W} w\otimes f(w)$, and that the product of
$\mathscr{A}$ restricts to the convolution product of $\mathrm{End}_RR\lala X,Y\rara$
defined by $f\ast g:=\mathrm{conc}\circ(f\otimes g)\circ \Delta$
(`$\mathrm{conc}$' means concatenation of left and right sides of $\otimes$).
Note that, for $f\in\mathrm{End}_R R\lala X,Y\rara$ and $J\in R\lala X,Y\rara$,
we have $f(J)=\sum_{w\in W} \lp w,J\rp f(w)$.
Since, by Corollary \ref{coro}, every word $w$ can be written as
$\sum_{\bt\in \N_0^{(\infty)}}
\lp \SS^{(\bt)},w \rp \MM^{(\bt)}$,
the element of $\mathscr{A}$ corresponding to
the identity $\mathrm{id}\in \mathrm{End}_RR\lala X,Y\rara$ is:
\begin{align*}
\sum_{w\in W} w\otimes w &=
\sum_w w\otimes \sum_\bt \lp \SS^{(\bt)},w \rp\MM^{(\bt)}
=
\sum_\bt (\sum_w \lp\SS^{(\bt)},w\rp w) \otimes \MM^{(\bt)} \\
&=
\sum_\bt \SS^{(\bt)} \otimes \MM^{(\bt)} \\
&=
\left(\sum_{d=0}^\infty
\sum_{\btau\in\N_0^d} \SS^{(\btau;0)}\otimes \MM^{(\btau;0)}\right)
\cdot
\left(\sum_{t=0}^\infty X^t\otimes X^t\right),
\end{align*}
where used are
$\SS^{(\bt)}=\SS^{(\btau;t)}=\SS^{(\btau;0)}\sha X^t$ and
$\MM^{(\bt)}=\MM^{(\btau;t)}=\MM^{(\btau;0)}\cdot X^t$.
Observing that both factors of the above last side correspond to specific
$R$-linear endomorphisms,
we can apply $\mathrm{id}$ to $J$ as the convolution product of them and find from
$\Delta(J)=J\otimes J$ that
\begin{equation} \label{J-expansion}
J=\mathrm{id}(J)=
\left(\sum_{d=0}^\infty
\sum_{\btau\in\N_0^d} \lp \SS^{(\btau;0)},J \rp \MM^{(\btau;0)}\right)
\left(\sum_{t=0}^\infty \frac{(c_X)^t}{t !} X^t\right).
\end{equation}
Note here that the pairing of $J$ with
$X^t=X^{\sha\, t}/t!$ is equal to $(c_X)^t/t !$,
as easily seen from the fact that the specialization
$J(X,0)\in R\lala X\rara$ at $Y=0$ is a group like
element $\exp(c_X\cdot X)$.
To settle the proof of Theorem \ref{LeMuFu},
given a fixed $\bk=(\bkappa;k_\infty)=(k_1,\dots,k_d; k_\infty)\in\N_0^{(\infty)}$
and $0\le s\le k_\infty$,
we compute
the coefficient of $w_{(\kappa;s)}=X^{k_1}Y\cdots X^{k_d}YX^s$ in the expansion
of the first factor of the above right hand side as follows:
\begin{align*}
&\sum_{d=0}^\infty
\sum_{\btau\in\N_0^d}
\lp \SS^{(\btau;0)},J \rp
\lp \MM^{(\btau;0)},w_{(\bkappa;s)} \rp
=\left\lp\sum_{d=0}^\infty
\sum_{\btau\in\N_0^d} \lp\SS^{(\btau;0)},J \rp \MM^{(\btau;0)},
w_{(\bkappa;s)} \right\rp \\
=&\left\lp\sum_{d=0}^\infty
\sum_{\btau\in\N_0^d}
\biggl\lp\SS^{(\btau;0)},\sum_{\bu\in \N_0^{(\infty)}}(J,w_\bu)w_\bu\biggr\rp
\MM^{(\btau;0)},
w_{(\bkappa;s)}\right\rp \\
=&\sum_{\bu} \lp J,w_\bu \rp \sum_{d=0}^\infty
\sum_{\btau\in\N_0^d} \lp \SS^{(\btau;0)},w_\bu \rp
\lp \MM^{(\btau;0)}, w_{(\bkappa;s)} \rp .
\end{align*}
But since $\lp\SS^{(\btau;0)},w_\bu\rp \lp \MM^{(\btau;0)}, w_{(\bkappa;s)} \rp$
survives only when
$\dep(\btau;0)=\dep(\bkappa;s)=\dep(\bu)$ and $|(\btau;0)|=|(\bkappa;s)|=|\bu|$,
the summation $\sum_\bu$ in the above last side occurs only for
those $\bu$ of the form
$(\bkappa+\bs;0)\in \N_0^{(\infty)}$
with $\bs=(s_1,\dots,s_d)\in\Z^d$,
$s=s_1+\cdots+s_d\ge 0$
(cf. also Remark \ref{rem2.6}).
Then, it follows from Lemma \ref{lem2} that the above last side is equal to
$$
\sum_{d=0}^\infty \sum_{\substack{\bs\in\N_0^d \\ |(\bs;0)|=s}}
\lp J, w_{(\bkappa+\bs;0)} \rp
(-1)^s
\binom{k_1+s_1}{k_1}\cdots \binom{k_d+s_d}{k_d}.
$$
(Note: The prescribed condition $\bs\in\Z^d$ has been replaced
with $\bs\in\N_0^d$ for a posteriori survivals of binomial factors).
{}From this and (\ref{J-expansion}) together with
$\lp J, w_{(\bkappa+\bs;0)} \rp=c_{(k_1+s_1,\dots,k_d+s_d;0)}$,
we conclude the assertion.
\end{proof}
\begin{Note}[{\it Alternative proof of Lemma} \ref{lem2}] \label{Quest4.3}
In the right hand side of Lemma \ref{lem2}, the quantity
$\binom{k_1+s_1}{k_1}\cdots \binom{k_d+s_d}{k_d}$ can be interpreted as
the pairing
$\lp w_{(k_1,\dots,k_d;0)}\sha X^s, w_{(k_1+s_1,\dots,k_d+s_d;0)} \rp$.
Therefore, the assertion of Lemma is equivalent to the identity
\begin{equation}
\sum_{\btau\in\N_0^d}
\lp\SS^{(\btau;0)},w_{(\bkappa+\bs;0)} \rp \cdot \lp\MM^{(\btau;0)}, w_{(\bkappa;0)}\!\cdot\!\! X^s\rp
=(-1)^s
\lp w_{(\bkappa;0)}\sha X^s, w_{(\bkappa+\bs;0)} \rp
\label{alterLemma}
\end{equation}
for $\bkappa=(k_1,\dots,k_d)\in\N_0^d$, $\bs=(s_1,\dots,s_d)\in\Z^d$
satisfying $s=s_1+\cdots+s_d\ge 0$ and $\bkappa+\bs\in\N_0^d$.
We now give an alternative proof for it using the Magnus/demi-shuffle duality:
First, by Corollary \ref{coro}, we have
$w_{(\bkappa;0)}=\sum_\bbr\la \MM^{(\bbr)},w_{(\bkappa;0)}\ra \SS^{(\bbr)}$
and $w_{(\bkappa+\bs;0)}=\sum_\bt\la \SS^{(\bt)},w_{(\bkappa+\bs;0)}\ra \MM^{(\bt)}$
so that the RHS of (\ref{alterLemma}) can be written as
\begin{align}
\label{RHS4.2}
&(-1)^s
\lp w_{(\bkappa;0)}\sha X^s, w_{(\bkappa+\bs;0)} \rp \\
&=(-1)^s \sum_{\bbr,\bt\in\N_0^{(\infty)}}
\lp \SS^{(\bbr)}\sha X^s, \MM^{(\bt)} \rp
\lp \MM^{(\bbr)} ,w_{(\bkappa;0)} \rp
\lp \SS^{(\bt)} ,w_{(\bkappa+\bs;0)}\rp
\notag \\
&=(-1)^s \sum_{\brho\in\N_0^d}
\lp \MM^{(\brho;0)} ,w_{(\bkappa;0)}\rp
\lp \SS^{(\brho;s)}, w_{(\bkappa+\bs;0)} \rp.
\notag
\end{align}
Here in the second equality, we use the fact that
$\lp \MM^{(\bbr)}, w_{(\bkappa;0)} \rp$
survives only if $\bbr=(\brho;0)\in\N_0^{(\infty)}$ for some $\brho\in\N_0^d$
and then apply the duality (Theorem \ref{orthogonality})
to $\lp \SS^{(\bbr)}\sha X^s, \MM^{(\bt)} \rp$
with $\SS^{(\brho;0)}\sha X^s=\SS^{(\brho;s)}$
(cf. Definitions \ref{defMagnus} and \ref{demi-shuffle}).
On the other hand, in the LHS of (\ref{alterLemma}),
one observes that nontrivial terms of the summation arise
only from those $\btau=(\tau_1,\dots,\tau_d)\in\N_0^d$ subject to
$\sum_{i=1}^d \tau_i=s+\sum_{i=1}^d \kappa_i$ (constant).
But then, the last binomial factor in (\ref{MagnusBino}) for
$ \lp\MM^{(\btau;0)}, w_{(\bkappa;0)}\!\cdot\!\! X^s\rp
=\magnus{(\tau_1,\dots,\tau_d;0)}{(\kappa_1,\dots,\kappa_d;s)}
$
equals $\binom{\tau_d}{s}$ which is non-zero only if
$\tau_d\ge s$.
Therefore,
the summation $\sum_\btau$ may be replaced by
$\sum_\brho$ with $\brho=\btau-(\mathbf{0},s)$
in $\N_0^d$ (where $\mathbf{0}\in\N_0^{d-1}$: the zero vector).
Thus,
the LHS of (\ref{alterLemma}) can be written as
\begin{align}
\label{LHS4.2}
&\sum_{\btau\in\N_0^d}
\lp\SS^{(\btau;0)},w_{(\bkappa+\bs;0)} \rp \cdot
\lp\MM^{(\btau;0)}, w_{(\bkappa;0)}\!\cdot\!\! X^s\rp \\
&=
\sum_{\brho\in\N_0^d}
\lp\SS^{(\brho+(\mathbf{0},s);0)},w_{(\bkappa+\bs;0)} \rp \cdot
\lp\MM^{(\brho+(\mathbf{0},s);0)}, w_{(\bkappa;s)} \rp.
\notag
\end{align}
Comparing summands of the above (\ref{RHS4.2}) and (\ref{LHS4.2}) for
individual $\brho\in\N_0^d$ in view of coefficients of
monomial expansions of demi-shuffle/Magnus polynomials
(Lemma \ref{dS-mono} and (\ref{Mag-Mono})), we reduce
the formula (\ref{alterLemma}) to the following elementary identity
for $\bkappa=(k_i),\brho=(r_i)\in\N_0^d$ and $\bs=(s_i)\in\Z^d$
satisfying
$\sum_{i=1}^d k_i=\sum_{i=1}^d r_i$, $\bs+\bkappa\in \N_0^d$
and $s:=\sum_{i=1}^d s_i\ge 0$:
\begin{equation}
\binom{(\bkappa+\bs;0)}{(\brho;s)}
\magnus{(\brho;0)}{(\bkappa;0)}
=
(-1)^s
\binom{(\bkappa+\bs;0)}{(\brho+(\mathbf{0},s);0)}
\magnus{(\brho+(\mathbf{0},s);0)}{(\bkappa,s)}
\end{equation}
that is an immediate consequence of definitions
of these symbols $\{^*_* \}$, $(^*_*)$.
(Observe that only difference between the corresponding symbols
occurs from the last binomial coefficient in (\ref{ArrayBino})
and (\ref{MagnusBino}).)
\qed
\end{Note}
\begin{Example}
The following shows an output of a group-like element
$J=\sum_{w\in W}c_w w$ of $R\lala X,Y\rara$ with the
shuffle relation (which is necessary and
sufficient for group-likeness due to Ree \cite{Ree})
counted from a computation using software \cite{Maple}
up to total degree 4.
$ \displaystyle J=
1+c_{X} X +c_{Y} Y +\frac{c_{X}^{2} \mathit{XX}}{2}+c_{\mathit{XY}} \mathit{XY} +\left(c_{X} c_{Y}-c_{\mathit{XY}}\right) \mathit{YX}
+\frac{c_{Y}^{2} \mathit{YY}}{2}+\frac{c_{X}^{3} \mathit{XXX}}{6} \\
+c_{\mathit{XXY}} \mathit{XXY} +\left(c_{X} c_{\mathit{XY}}-2 c_{\mathit{XXY}}\right) \mathit{XYX} +c_{\mathit{XYY}} \mathit{XYY} +\left(\frac{1}{2} c_{X}^{2} c_{Y}-c_{X} c_{\mathit{XY}}+c_{\mathit{XXY}}\right) \mathit{YXX}
\\
+\left(c_{\mathit{XY}} c_{Y}-2 c_{\mathit{XYY}}\right) \mathit{YXY}
+\left(\frac{1}{2} c_{X} c_{Y}^{2}-c_{\mathit{XY}} c_{Y}+c_{\mathit{XYY}}\right) \mathit{YYX} +\frac{c_{Y}^{3} \mathit{YYY}}{6}
\\
+\frac{c_{X}^{4} \mathit{XXXX}}{24}+c_{\mathit{XXXY}} \mathit{XXXY} +\left(c_{X} c_{\mathit{XXY}}-3 c_{\mathit{XXXY}}\right) \mathit{XXYX} +c_{\mathit{XXYY}} \mathit{XXYY}
\\
+\left(\frac{1}{2} c_{X}^{2} c_{\mathit{XY}}-2 c_{X} c_{\mathit{XXY}}+3 c_{\mathit{XXXY}}\right) \mathit{XYXX} +\left(\frac{c_{\mathit{XY}}^{2}}{2}-2 c_{\mathit{XXYY}}\right) \mathit{XYXY}
\\
+\left(c_{X} c_{\mathit{XYY}}-\frac{c_{\mathit{XY}}^{2}}{2}\right) \mathit{XYYX}
+c_{\mathit{XYYY}} \mathit{XYYY} +\left(\frac{1}{6} c_{X}^{3} c_{Y}-\frac{1}{2} c_{X}^{2} c_{\mathit{XY}}+c_{X} c_{\mathit{XXY}}-c_{\mathit{XXXY}}\right) \mathit{YXXX}
\\
+\left(c_{\mathit{XXY}} c_{Y}-\frac{c_{\mathit{XY}}^{2}}{2}\right) \mathit{YXXY}
+\left(c_{X} c_{\mathit{XY}} c_{Y}-2 c_{X} c_{\mathit{XYY}}-2 c_{\mathit{XXY}} c_{Y}+\frac{1}{2} c_{\mathit{XY}}^{2}+2 c_{\mathit{XXYY}}\right) \mathit{YXYX}
\\
+\left(c_{\mathit{XYY}} c_{Y}-3 c_{\mathit{XYYY}}\right) \mathit{YXYY}
+\left(\frac{1}{4} c_{X}^{2} c_{Y}^{2}-c_{X} c_{\mathit{XY}} c_{Y}+c_{X} c_{\mathit{XYY}}+c_{\mathit{XXY}} c_{Y}-c_{\mathit{XXYY}}\right) \mathit{YYXX}
\\
+\left(\frac{1}{2} c_{\mathit{XY}} c_{Y}^{2}-2 c_{\mathit{XYY}} c_{Y}+3 c_{\mathit{XYYY}}\right) \mathit{YYXY}
+\left(\frac{1}{6} c_{X} c_{Y}^{3}-\frac{1}{2} c_{\mathit{XY}} c_{Y}^{2}+c_{\mathit{XYY}} c_{Y}-c_{\mathit{XYYY}}\right) \mathit{YYYX}
\\
+\frac{c_{Y}^{4} \mathit{YYYY}}{24}
\quad+\quad $(terms of degree $\ge 5$).
\medskip
In the above computation, one observes that the coefficient $c_{XYXY}$
is expressed by lower simpler coefficients of $J$. This does not follow
from Theorem \ref{LeMuFu}, however, does reflect the fact
that $XYXY$ is not a Lyndon word.
Discussions on the most economical expression using only the
coefficients of Lyndon words can be found in \cite{MPH}.
\end{Example}
\begin{Note}
In the modern theory of multiple zeta values, a certain standard solution
$G_0^z(X,Y)\in \C\lala X,Y\rara$ to the KZ-equation on $z\in \C-\{0,1\}$
is known as the generating function for the multiple polylogarithms (MPL).
It is also used to define the Drinfeld associator $\Phi(X,Y) \in\C\lala X,Y\rara$.
The coefficients of $w_{(k_1,\dots,k_d;0)}$ in $\Phi(X,Y)$ (resp. in $G_0^z(X,Y)$)
are regular multiple zeta values (resp. regular MPL) of multi-index $(k_1,\dots,k_d)$,
but the other coefficients are in general not.
Le-Murakami \cite{LM}, Furusho \cite{F} derived formulas that express
all coefficients of $\Phi(X,Y)$ and $G_0^z(X,Y)$ by those `regular' coefficients
explicitly.
In \cite[Remark 2]{N21}, the author posed a question if it could be a similar
case for `$\ell$-adic Galois associator $f_\sigma^z(X,Y)\in\Q_\ell\lala X,Y\rara$',
in which context analytic theory of KZ-equation is unavailable yet.
Since $f_\sigma^z(X,Y)$ is by definition a group-like element,
the above Theorem \ref{LeMuFu} answers the question affirmatively.
\end{Note}
\begin{Note}
\label{zinbiel}
A noteworthy notion closely related to our $\SS^{(\bk)}$,
$\SS^{(\bk,\blambda)}$ is
the free Zinbiel (or, dual Leibniz) algebra
studied by J.-L Loday \cite{Lo95}, I.Dokas \cite{Do10}, F.Chapoton \cite{Ch21} et.al.
Let $V$ be a vector space with a basis $\mathfrak{B}=\{X_0,X_1,\dots\}$ and
$T(V)$ be the tensor algebra (free associative algebra) generated by the
letters in $\mathfrak{B}$. Loday introduced the ``half-shuffle'' product $\prec$
in $T(V)$ as the linear extension of the binary product on words
given by:
$$
(x_0x_1\cdots x_p)\prec (x_{p+1}\cdots x_{p+q})
:=
x_0\cdot \bigl( (x_1\cdots x_p)\sha (x_{p+1}\cdots x_{p+q})\bigr),
$$
where $x_i$ are letters in $\mathfrak{B}$ ($i=0,\dots, p+q$).
It is remarkable that, while the usual shuffle product $w\,\sha\, w'=w\prec w'+w'\prec w$
is associative (and commutative), the half-shuffle product
$\prec$ is not even associative --- however satisfying
$(w_1\prec w_2)\prec w_3=w_1\prec (w_2\prec w_3)+w_1\prec (w_3\prec w_2)$.
We may relate the `Zinbiel monomials'
with our demi-shuffle polynomials
$\SS^{(\bk,\blambda)}$ in Definition \ref{def3.1} as follows:
Write $\ast\mapsto \overline{\ast}$ for the anti-automorphism of
$R\la X,Y_\lambda\ra_{\lambda\in\Lambda}$
reversing the order of letters in each word, e.g.,
$\overline{XXY_\lambda}=Y_\lambda XX$.
Then,
\begin{equation}
\label{semi-shuffle}
\overline{\SS^{(\bk,\blambda)}}=
X^{k_\infty}\sha \bigl(...
\bigl(Y_{\lambda_d}X^{k_d} \prec\bigl( Y_{\lambda_{d-1}}X^{k_{d-1}} \prec\bigl(
\cdots
\prec \bigl( Y_{\lambda_2}X^{k_2} \prec Y_{\lambda_1}X^{k_1} \bigr)\bigr)...\bigr)
\end{equation}
for $\bk=(k_1,\dots,k_d; k_\infty)\in\N_0^{(\infty)}$,
$\blambda=(\lambda_1,\dots,\lambda_d)\in\Lambda^d$.
These polynomials also appeared in \cite[Proposition 5.10]{Minh2019} to
illustrate the coefficients (of the main factor) of a solution of the
KZ-equation expanded in
$(\mathrm{ad}_{-X}^{k_1}Y)\cdots(\mathrm{ad}_{-X}^{k_d}Y)$.
Calling $\SS^{(\bk)}$, $\SS^{(\bk,\blambda)}$
`demi-shuffle' in Definitions \ref{demi-shuffle}, \ref{def3.1}
or reserving `semi-shuffle' for
names of anything else
might keep a moderate distance from the already overwhelming naming
`half-shuffle' of the operation $\prec$ in literature.
\end{Note}
\medskip
{\it Acknowledgement}:
The author is grateful to Hidekazu Furusho for hinting a positive answer to the
question posed in \cite[Remark 2]{N21} toward the form of
Theorem \ref{LeMuFu}
of the present paper, and for valuable comments and information
including what are mentioned in Note \ref{zinbiel}.
The author also thanks Densuke Shiraishi for stimulating discussions that
share awareness of various open problems around $\ell$-adic Galois multiple polylogarithms.
This work was supported by JSPS KAKENHI Grant Numbers JP20H00115.
\ifx\undefined\bysame
\newcommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\,}
\fi
|
\section{Introduction}
Humans interactively converse every day with other humans, and more recently with machines. Humans constantly converse with each other, online or offline, for example, when attending meetings or using customer service. Therefore, textual dialogues are an essential component of the interactions between users and user agents.
This abundance of personal and public conversations represents a valuable source of information, but analyzing such an immense amount of data to meet specific information needs can lead to information overload problems~\cite{jones2004information}.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=1\linewidth]{./figures/pos_example.png}
\end{center}
\caption{Example utterance of formal and informal sentences of the same meaning from different speakers, with the different parts-of-speech labeled. The histogram shows the different individual textual styles.}
\label{fig:pos_example}
\end{figure}
Recently, dialogue summarization has emerged as a means to resolve this issue. Dialogue summarization is the task of distilling the highlights of an instance of a dialogue and rewriting them into an abridged version. This task is closely associated with abstractive text summarization~\cite{shi2021neural,gehrmann-etal-2018-bottom,nallapati2016abstractive}.
However, most of the existing investigation efforts on abstractive text summarization have been concentrated on single-speaker documents, such as news articles~\cite{paulus2017deep,see2017get} and scientific documents~\cite{cachola-etal-2020-tldr, erera-etal-2019-summarization}. Research has largely focused on the composition of strictly formatted paragraphs ({\em{e.g.,}} introduction-method-conclusion) or is simply dependent on locational biases ({\em{i.e.,}} the tendency towards a particular location such as the lead or tail of the text)~\cite{kim2019abstractive}.
Unlike these strictly structured formats, the format of dialogues between multiple interlocutors is often informal. Dialogues are represented through a variety of utterances, including the expression of personal perspectives, opinions, markers of certainty and doubt, speaker interruptions~\cite{precht2008sex,sacks1978simplest}, colloquial representations, and even the use of emojis in a textual way, and important information is scattered throughout dialogue, not necessarily restricted to predictable areas. These features lead to a focus on informative utterances. Figure~\ref{fig:pos_example} presents the representation of different speaker styles for informal and formal utterances that we attempt to address in this paper.
Thus, the critical challenges of dialogue summarization task are: 1) \textit{\textbf{Multiple speakers and the different textual styles;}} the essential pieces of information in a conversation are scattered across the utterances of interlocutors through their different textual styles~\cite{koppel2009computational}. 2) \textit{\textbf{Informal structure;}} dialogues consist of an informal structure, including slang and colloquial language, free of the locational biases found in formal structures~\cite{chen2020multi}.
To address these challenges, we investigated the relationship between textual styles and representative attributes of utterances. \citet{kubler2010adding} proposed that the types ({\em{e.g.,}} intent or role of a speaker) of sentences from speakers are associated with different syntactic structures ({\em{i.e.,}} linguistic information), such as part-of-speech (POS) tagging. This is derived from the fact that different speaker roles are characterized by different syntactic structures.
Research in dialogue summarization is benefited from research in related fields, such as speaker recognition. Speaker recognition is the process of identity recognition, and specifically uses identity information ({\em{i.e.,}} voiceprint) from the human voice. The speaker's speech signal is considered as the hidden features that can be used to distinguish the identity of that individual and is commonly used in the speaker recognition research field~\cite{guo2021speaker,liu2018gmm}. In essence, the uttered text has a unique representation from each speaker, like a voiceprint. Based on this prior research, we began our study with the assumption that because syntactic structures tend to be associated with a representative of a sentence uttered from speakers, these structures would help distinguish the different styles of utterances. This assumption was also maintained in previous research. \citet{zhu2020hierarchical} proposed a hierarchical structure to handle the transcripts of long meetings and adopted a role vector to represent the individual speakers in conversation summarizations.
Inspired by the previous works, we propose a novel abstractive dialogue summarization model for use in a daily conversation setting, which is characterized by an informal style of text, including emoticons and abbreviations of chat terms. Furthermore, we explore the locational biases in dialogue structures. Although dialogues show independent locational biases different from that of formal documentation, we evaluated different simple baselines based on locational biases motivated by~\citet{gliwa2019samsum,kim2019abstractive} (see details in Section~\ref{subsec:implementation}). The main contributions of this paper are fourfold.
\begin{itemize}
\item First, we propose a novel approach for the abstractive dialogue summarization task. Specifically, the multi-task learning model is proposed to learn abstractive dialogue summarization and perform sequence labeling tasks simultaneously, to reflect the syntactic features on the dialogue summarization model.
\item Second, to the best of our knowledge, this is the first study to perform multi-task learning on the dialogue summarization task using the SAMSum corpus and, specifically, to integrate these tasks using linguistic information.
\item Third, we propose a novel input type training method, rather than using the traditional method of truncating the input, to investigate locational biases.
\item Finally, the proposed method outperformed the base models for all ROUGE scores~\cite{lin2004rouge}.
\end{itemize}
\section{Proposed Method}
\subsection{Why Part-of-speech-tagging?}
Part-of-speech-tagging is a valuable resource for analysis in the syntax-aware approach. \citet{arifin2018sentence} conducted multi-document summarization to find representative sentences, not only by sentence distribution to select the most important sentences but also by how informative a term is in a sentence. They used part-of-speech (POS) tagging information to resolve this and obtain improved performance. This approach is characterized by its use of grammatical information, which is carried by POS labels, and the presence or absence of informative content in a sentence.
We further considered that incorporating the linguistic information from POS tagging could help alleviate structure/context ({\em{i.e.,}} formal and informal) dependency issues for text summarization, and also in dialogue summarization. Simultaneously learning the syntax-aware approach using linguistic information and language generation allows the sharing of grammatical information that constrains next word generation.
Also, it is possible to deal with the first challenge by applying syntax-awareness to the entirety of the utterances from the dialogue because this will recognize the linguistic information from the speakers and also intrinsically represent the textual styles. Therefore, the model obtains the built-in ability to distinguish text styles.
\subsection{Problem Formulation}
We formalize the problem of dialogue summarization as follows. The input consists of dialogues $\chi$ and dialogue speakers $S$. Assume there are $d$ dialogues in total. The dialogues are $\chi = \{X_1,...,X_d\}$. Each dialogue consists of multiple turns, where each turn is the utterance of a speaker. Therefore, $X_i = \{(s_1, u_1), (s_2, u_2), ..., (s_{L_i}, u_{L_i})\}$, where $s_j \in S$, $1\leq j \leq L_i$, is a speaker and $u_j = (w_1,...,w_{l_j})$ is the tokenized utterance from $s_j$. The human-annotated summary for dialogue $X_i$, denoted by $Y_i$, is also a sequence of tokens. In the end, the aim of the task is to generate a dialogue summary $\hat{Y}=(\hat{y}_1, ..., \hat{y}_l)$ given the dialogues $X=\{(s_1, u_1), (s_2, u_2), ..., (s_m, u_m)\}$ and the reference summaries $Y=(y_1, ..., y_k)$.
The purpose of the sequence labeling task is to predict the sequence label $Y_{POS}=(pos_1, ..., pos_n)$, where $pos_i=\{tag_1,...,tag_{l_j}\}$, and $|pos|$ is the number of tags $tag$ in an utterance.
To summarize, the final goal of dialogue summarization is to maximize the conditional probability of the dialogue summary $Y$, given dialogues $X$ and model parameters $\theta: P(Y|X;\theta)$.
\subsection{Preprocessing for Syntax-Aware SAMSum}
We automatically annotated sequence labels for all the utterances as these are not included in the SAMSum corpus. We used these labels for training syntax-aware information using the steps below.
\paragraph{Tokenization for data labeling} To better recognize syntax-aware information, we used Twokenizer\footnote{\url{https://github.com/myleott/ark-twokenize-py} Note that Twokenizer is used for data labeling not for model training.} \cite{owoputi2013improved} before annotating the sequence labeling. The Twokenizer was revised for tweet text to conduct part-of-speech (POS) tagging. Tweet text consists of online conversational text that also contains many nonstandard lexical items and syntactic patterns, such as emojis and emoticons. Also, the daily chat includes those like tweet text.
Emoticons ({\em{e.g.,}} :), XD) refer to the generation of a face or icon using traditional alphabetic or punctuation symbols, whereas emojis ({\em{e.g.,}} \cChangey{2}, \cChangey{-1}) refer to when small pictures used as symbols. The Twokenizer accurately recognizes emoticons as, for example, ``:)'' not as ``: ).''
\paragraph{Part-of-speech tagging} We obtained tokenized utterances using the above method. This process improves the model's ability to recognize each token to use the POS tagger. For the sequence labeling, we adopted the CMU-Twitter-POS-tagger\footnote{\url{http://www.cs.cmu.edu/~ark/TweetNLP/}}~\cite{owoputi2013improved,gimpel2010part}, which addresses the problem of POS tagging for English data from the popular micro-blogging service Twitter.
\begin{figure*}[h!]
\begin{center}
\includegraphics[width=0.9\linewidth]{./figures/syntax_bart_overview.png}
\end{center}
\caption{Overview of the model architecture. The syntax-aware encoder with a task-specific linear head learns the sequence labeling task given the dialogue utterances in a bidirectional encoder setting from the BART encoder. The conversation decoder ({\em{i.e.,}} autoregressive decoder from the BART decoder) learns the dialogue summarization task through the linear head.}
\label{fig:overall}
\end{figure*}
\subsection{Model Overview}
Regarding the multi-task learning for BART backbone, we address two different tasks simultaneously: token classification ({\em{i.e.,}} sequence labeling) and language modeling ({\em{i.e.,}} generation). BART consists of a bidirectional encoder and an autoregressive decoder. Therefore, we conducted the token classification task in the encoder ({\em{i.e.,}} syntax-aware encoder) and the language model task in the decoder ({\em{i.e.,}} conversational decoder). As illustrated in Figure~\ref{fig:overall}, task-specific linear heads were trained through multi-task learning, which performs the main task as a dialogue summarization task and the POS sequence labeling task as an auxiliary task.
\subsection{Syntax-Aware Encoder} \label{sec:syntax-aware-encoder}
We sought to address the application of syntax-aware information to a dialogue summarization model through the sharable encoder.
Each utterance $u_i$ was composed of a special [EOU] token, as was done in previous work~\cite{gliwa2019samsum}, considering each utterance separately. In general, the input sequence,
\begin{equation}
\mathrm{{X_i}^{'}}=\{s_1;u_1;[EOU] ... s_{L_i};u_{L_i};[EOU]\},
\label{eq:equ-1}
\end{equation}
is fed into the bottom encoder $\mathrm{\bf E}$ of BART. Given the hidden outputs of the encoder's last layer $\{{h_1}^{L}, ..., {h_n}^{L}\}$, the output layer for the sequence labeling task was a linear classifier $f : \mathbb{R}^{d} \to \mathcal{Y}$, where $\mathbb{R}^{d}$ denotes the dimension of the hidden layer and $\mathcal{Y}$ is a $(k-1)$ simplex, where $k$ is the number of POS tags.
\begin{equation}
\begin{aligned}
\{{h_1}^{L}, ..., {h_n}^{L}\} = \mathrm{\bf E}(\{\mathrm{{X^{'}_1}, ..., {X^{'}_n}\}}),
\label{eq:equ-2}
\end{aligned}
\end{equation}
, where $L$ is the last layer. The probability that the word $w_{l_j}$ aligns with the $m$-th POS tag is computed using softmax:
\begin{equation}
\begin{aligned}
P(m|w_{l_j}) = \mathrm{Softmax}(W_m{h_n}^{L}),
\end{aligned}
\end{equation}
, where $W_m$ is a parameter to be learned.
\subsection{Conversational Decoder}
To integrate the syntax-aware encoder with the decoder, the dialogue summarization model consists of combining the shared syntax-aware encoder using the Equation~\ref{eq:equ-2} and the conversational decoder.
\paragraph{Shared Syntax-Aware Encoder} We used the same encoder over all-around layers from the Section~\ref{sec:syntax-aware-encoder} to apply the syntax-aware information to the dialogue summarization model. Both encoders from each task were shared.
The syntactic information could provide different conversational aspects for the models to learn and further determine which set of utterances deserve more attention to generate better dialogue summaries.
The input to the decoder included $l-1$ previously generated tokens $\{t_1, ..., t_{l-1}\}$. We fed the tokens to the conversation decoder $\mathrm{\bf D}$, and the $l$-th token ($\hat{t}_l$) was predicted as follows:
\begin{equation}
\begin{aligned}
\tilde{\mathrm{\bf E}} = \mathrm{\bf E}(\{\mathrm{{X^{'}_1}, ..., {X^{'}_n}\}}),
\label{eq:equ-3}
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
\{\hat{y}_1, ..., \hat{y}_{l-1}\} = \mathrm{\bf D}(\{{t_1}, ..., {t_{l-1}}\}, \tilde{\mathrm{\bf E}}),
\label{eq:equ-4}
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
P(\hat{t}_l|t_{<l}, X^{'}) = \mathrm{Softmax}(W\hat{y}_{l-1}),
\label{eq:equ-5}
\end{aligned}
\end{equation}
, where $W$ is a parameter to be learned.
\subsection{Syntax-Aware Multi-task Learning}
We trained the two tasks jointly using multi-task learning. We considered the dialogue summarization task as the main task and the sequence labeling task as an auxiliary task.
\paragraph{Joint training} During training, the two self-learning objectives were combined with the cross-entropy loss for each task, and we sought to minimize the loss as follows:
\begin{equation}
\begin{aligned}
\mathcal{L_{DS}} = -\sum\mathrm{log}P({y_l}|\hat{t}_{<l}, X^{'}),
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
\mathcal{L_{POS}} = -\sum\mathrm{log}P(pos_{l_j}|m),
\end{aligned}
\end{equation}
Thus, the final loss $\mathcal{L}$ of our model is:
\begin{equation}
\begin{aligned}
\mathcal{L} = \lambda\mathcal{L_{POS}} + (1-\lambda)\mathcal{L_{DS}}.
\end{aligned}
\end{equation}
, where $\mathcal{L_{DS}}$ and $\mathcal{L_{POS}}$ are the loss of the dialogue summarization model and the sequence labeling model, respectively, and $\lambda$ denotes the parameter of strength in each task.
Finally, the model can activate linguistic information to enhance its ability to distinguish a speaker's utterance style.
\subsection{Speaker Styles of Utterance for Ad-hoc Analysis}
\label{sec:speaker_style}
In order to represent the uttering styles of speakers, we considered a list of POS tags extracted from utterances of each speaker as a style of the speaker as Equation \ref{eq:style_is_a_set_of_tags} where $i$ denotes the index of speaker, and $T$ is the number of tags. It has the same form of a document made up of words.
\begin{equation}
\label{eq:style_is_a_set_of_tags}
{style_{i}}=\{tag_1, ..., tag_T\},
\end{equation}
With the style documents, we conducted tf-idf, a commonly used method to weight the importance of each term in a document. The tf-idf formula is as follows:
\begin{equation}
\label{eq:tfidf}
W_{ij}=tf_{ij}*idf_{ij}=tf*log(\frac{n}{1+df_{ij}}),
\end{equation}
In Equation \ref{eq:tfidf}, $tf_{ij}$ represents a term frequency of the $j$-th tag in the $i$-th speaker style, and $df_{ij}$ denotes the number of $i$-th speaker styles in which the $j$-th tag appears, and $n$ is the total number of speaker styles. Then we employed K-means clustering for grouping the speaker styles. Consequently, this ad-hoc analysis was used to show whether our proposed strategies have been worked as intended in the trained models.
\section{Experimental Setup}
\subsection{Dataset and Baselines}
We trained and evaluated our model on a large-scale dialogue summary dataset SAMSum~\cite{gliwa2019samsum}. SAMSum is the first daily chat corpus for use in dialogue summarization and truly informal conversation. The subject of each conversation is open domain, and the conversation type is informal.
The details on data statistics in SAMSum corpus are shown below, including in Table~\ref{tab:data_statistics}.
\begin{table}[tbh!]
\centering
\scalebox{0.95}{
\resizebox{\columnwidth}{!}{
\begin{tabular}{rl|cc|cc|cc}
\hline
\multicolumn{2}{r|}{\multirow{2}{*}{\textbf{\# Conv}}} & \multicolumn{2}{c|}{\textbf{S.L}} & \multicolumn{2}{c|}{\textbf{\# Speakers}} & \multicolumn{2}{c}{\textbf{\# Turns}} \\
& & Mean & Range & Mean & Range & Mean & Range \\ \hline
Train & 14732 & 23.44 & [2, 73] & 2.40 & [1, 14] & 11.17 & [1, 46] \\
Dev & 818 & 23.42 & [4, 68] & 2.39 & [2, 12] & 10.83 & [3, 30] \\
Test & 819 & 23.12 & [4, 71] & 2.36 & [2, 11] & 11.25 & [3, 30] \\ \hline
\end{tabular}}}
\caption{Data statistics of the SAMSum corpus. S.L denotes summary length. Range indicates the minimum and maximum values.} \label{tab:data_statistics}
\end{table}
\paragraph{Data Statistics} To better understand the characteristic of this corpus, we explored the density of the number of utterances.
As shown in Figure~\ref{fig:utter-statistic} (see Appendix~\ref{appendix:density}), the density of the number of utterances is consistently below ten. This result indicates that the utterances are generally fewer than ten in the training set. This result allows us to choose the locational biases, like input sequence types such as LEAD-3~\cite{see2017get}, which takes the three leading sentences of the source text as the summary (we discuss this in Section~\ref{subsec:implementation}).
\paragraph{Baselines} We evaluated the model's performance with the following summarization models, based upon previous works~\cite{gliwa2019samsum}.
\begin{itemize}
\item \textbf{Pointer Generator}~\cite{see2017get} This model followed~\citet{gliwa2019samsum}, wherein separators are added between each utterance and utilized as input for the pointer generator model.
\item \textbf{DynamicConv + GPT-2}~\cite{wu2019pay} Based on ~\citet{gliwa2019samsum}, this model uses GPT-2 to initialize token embeddings~\cite{radford2019language}.
\item \textbf{Fast Abs RL Enhanced}~\cite{chen2018fast} adopts a hybrid method that selects salient sentences and then paraphrases them as abstractive sentences through sentence-level policy gradient methods.
\item \textbf{BART}~\cite{lewis-etal-2020-bart}
We used BART as the vanilla in the following setting and added a separator in each utterance. The default parameter setting was \texttt{BART-base}\footnote{\url{https://huggingface.co/ transformers/model_doc/bart.html}}. Additionally, we fed the input type with the LONGEST-10 settings, as described in Section~\ref{subsec:implementation}.
\end{itemize}
\subsection{Evaluation Metrics}
We utilized different evaluation metrics, including several recently introduced methods used in text summarization and generation tasks.
\paragraph{ROUGE-N\footnote{\url{https://github.com/pltrdy/rouge} Note that different packages may generate different ROUGE scores.}}~\cite{lin2004rouge} mostly used evaluation metrics for the text summarization task. We calculated ROUGE-1, ROUGE-2, and ROUGE-L.
\paragraph{BertScore\footnote{\url{https://pypi.org/project/bert-score/}}}~\cite{zhang2019bertscore} calculates the aligning similarity scores between the generated and reference summaries on a token level using BERT.
\subsection{Implementation Details}\label{subsec:implementation}
We tested different input type modes in our proposed model. The traditional method for handling long sequences in a pretrained language model is to truncate the sequence in an uncompleted format, not in the true utterance format, owing to limitations in system memory. To alleviate this issue, we propose a novel input type method to retain the utterance format.
Inspired by previous work~\cite{gliwa2019samsum,see2017get}, we defined the input types as LEAD-n, MIDDLE-n, and LONGEST-n. The underlying assumption of these input types is that the locational biases~\cite{kim2019abstractive} contain the essential information at the head ({\em{i.e.,}} the beginning of the lead) or middle of lengthy conversations. Additionally, this method preserves the entire utterance sequences without breaking up sentence information. To support the above assumptions, we performed data statistics as described in the above section. We trained the model according to the different settings described as follows: \textbf{LEAD-n} - takes $n$ leading utterances of the dialogues, \textbf{MIDDLE-n} - takes $n$ utterances from the middle of the dialogue, and \textbf{LONGEST-n} - takes $n$ utterances from the longest of the dialogue.
We used the \texttt{BART-base} model to initialize the backbone of the encoder/decoder frame and followed the default settings. The learning rate was set to 3e-4. We trained the model for 20 epochs. Also, we set $\lambda$ as 0.1 in the final model. The training was conducted on a single RTX 8000 GPU with 48 GB memory. We trained the model with the Adam optimizer~\cite{kingma2014adam} and an early stop on validation set ROUGE-1. During the inference, the beam size was 4, including for the baseline models.
\section{Main Results}
\subsection{Quantitative Results}
We evaluated the models across the different settings with ROUGE-1, ROUGE-2 and ROUGE-L, and BertScore on the SAMSum test set. The experimental results are shown in Tables~\ref{tab:result-type} and ~\ref{tab:result}.
\begin{table}[h!]
\centering
\resizebox{\columnwidth}{!}{
\begin{tabular}{l|c|l|cccc}
\hline
\textbf{Type} & $\lambda$ & \textbf{n} & \textbf{ROUGE-1} & \textbf{ROUGE-2} & \textbf{ROUGE-L} & \textbf{BertScore} \\ \hline
\textbf{LEAD}& \multirow{3}{*}{0.5} &\multirow{3}{*}{10} &0.409&0.170&0.390& 0.909 \\
\textbf{MIDDLE} & &&0.403&0.167&0.382&0.908 \\
\textbf{LONGEST} &&&\textbf{0.425}&0.183&0.405&0.909 \\ \hline
\textbf{LEAD}&\multirow{2}{*}{0.5}& \multirow{2}{*}{20}&\textbf{0.425}&\textbf{0.188}&\textbf{0.409}& \textbf{0.910} \\
\textbf{MIDDLE}&&& 0.414 &0.181&0.404&0.908 \\\hline \hline
\textbf{LEAD}& \multirow{3}{*}{0.1} &\multirow{3}{*}{10} &0.426&0.188&0.414& 0.910 \\
\textbf{MIDDLE} & &&0.428&0.192&0.414&0.910 \\
\textbf{LONGEST}&&&\textbf{0.431}&\textbf{0.189}&\textbf{0.420}&\textbf{0.910} \\ \hline
\textbf{LEAD}&\multirow{2}{*}{0.1}& \multirow{2}{*}{20}&0.424&0.187&0.415& 0.909 \\
\textbf{MIDDLE}&&&0.425&0.189&0.416&0.909 \\\hline
\end{tabular}} \caption{Performance comparison according to the different input type settings for training. $\lambda$ and n indicate the strength of the task ability and the number of utterances in dialogue, respectively.} \label{tab:result-type}
\end{table}
\begin{table}[tbh!]
\centering
\scalebox{0.7}{
\begin{tabular}{c|c|c}
\toprule
\textbf{Model} & \textbf{Type} & \textbf{avg \# words} \\ \hline
Ground summary & - & 23.12 \\ \hline
BART & LONG-10 & 22.25 \\ \hline
\multirow{3}{*}{\shortstack[c]{Syntax-aware BART\\($\lambda=0.1$)}} & LEAD-10 & 19.95 \\
& MIDDLE-10 & 18.25 \\
& LONG-10 & 21.95 \\
\bottomrule
\end{tabular}}
\caption{The average number of words of the generated summaries at an inference.}
\label{tab:quantitative-2}
\end{table}
\paragraph{Internal model verification} In Table~\ref{tab:result-type}, we explore the influence locational biases in dialogue have on performance. The input type settings are depicted according to the different measures, based on the F1 scores of both ROUGE and BertScore. We set $\lambda$ as 0.5 and 0.1 and compared $n$ of 10 and 20 for each setting\footnote{Note that we did not set the LONGEST-20 due to limitations in computing power.}. With $\lambda$ set as 0.1, the LONGEST-10 model showed the best performance across every measure. However, the performance when $\lambda$ was set as 0.5 was lower than when $\lambda$ was 0.1, in general. Although the BertScore showed a subtle difference, it also showed the highest performance in this result. We interpret this result as indicating that lengthy utterances are valuable when generating summaries, and the key topics are located at the length of 10 in a dialogue.
\paragraph{Comparison of the generated length} We investigated the length of the generative summary, which varies with the input type. As shown in Table~\ref{tab:quantitative-2}, we examined the average of words according to the generated summaries at each inference step. The BART base model ({\em{i.e.,}} BART$\dagger$) used 22.25 words when it generated the summaries. Moreover, the average words generated by our proposed model according to the input type shows that our best performance model ({\em{i.e.,}} Syntax-aware BART (LONG-10)) used only 21.95 words, thus requiring fewer words than the baseline model. This observation reveals that our proposed model often favors generating slightly shorter summaries than does the BART baseline model, which leads to more concise summaries while still capturing the important information. According to input types, input length influences the average number of words.
\paragraph{External model verification} In Table~\ref{tab:result}, we present the ROUGE-1, ROUGE-2, and ROUGE-L scores between our model and other, comparison models. First, our proposed model outperformed the other baselines with respect to F1 for all ROUGE scores. \textit{As hypothesized previously, our experiments demonstrate that the usage of linguistic information is worthwhile to enhance the model performance.} Fast Abs RL Enhanced achieved slightly better scores than Pointer Generator and DynamicConv+GPT2. This indicates that the use of reinforcement learning to first select important sentences is beneficial. The key factor related to the overall lower performance of the baseline models seems to be that the baseline models fundamentally are not based on the language model; however, the DynamicConv model with the GPT-2 embeddings is based on the usage of pretrained embeddings from the language model GPT-2, which is trained on a large corpus.
\begin{table*}[ht!]
\centering
\resizebox{15cm}{!}{%
\begin{tabular}{c|c|ccc|ccc|ccc}
\hline
\multirow{2}{*}{\textbf{Model}} & \multirow{2}{*}{\textbf{Type}} & \multicolumn{3}{c|}{\textbf{ ROUGE-1}} & \multicolumn{3}{c|}{\textbf{ ROUGE-2}} & \multicolumn{3}{c}{\textbf{ ROUGE-L}}\\ \cline{3-11}
&{-} & F & P & R & F & P & R & F & P & R \\ \hline
Pointer Generator~\cite{see2017get}* & - & 0.401 & - & - & 0.153 & - & - & 0.366 & - & - \\
DynamicConv + GPT-2~\cite{wu2019pay}* & - & 0.418 & - & - & 0.164 & - & - & 0.376 & - & - \\
Fast Abs RL Enhanced~\cite{chen2018fast}* & - & 0.420 & - & - & 0.181 & - & - & 0.392 & - & - \\ \hline \hline
\multirow{1}{*}{BART $\dagger$} & LONG-10 &0.426&0.488&0.419&0.188&0.220&0.184&0.419&0.464&0.415 \\
\multirow{1}{*}{Syntax-aware BART \(\dagger\) (\(\lambda\) =0.1)} & LONG-10 &\textbf{0.431}&0.486&0.426&\textbf{0.189}&0.216&0.186&\textbf{0.420}&0.460&0.418 \\\hline
\end{tabular}%
}
\caption{Performance comparison of the proposed method with different models on the test set. * denotes the results from~\cite{chen2020multi}, and $\dagger$ corresponds to our proposed method model, which shows the best performance (LONGEST-10). Note that F, P, and R indicate F1, precision, and recall scores, respectively.}
\label{tab:result}
\end{table*}
\begin{table*}[h!]
\centering
\resizebox{16cm}{!}{%
\begin{tabular}{|l|l|l|}
\toprule
\bf Dialogue 1 & \bf Dialogue 2 & \bf Speaker Style Utterance (abbreviated) \\\hline
1. \textcolor{ku}{lilly:} sorry, I'm gonna be late &
1. \textcolor{ku}{randolph:} honey & \\
2. \textcolor{ku}{lilly:} don't wait for me and order the food &
2. \textcolor{ku}{randolph:} are you still in the pharmacy? & \\
3. \textcolor{jhu}{gabriel:} no problem, shall we also order &
3. \textcolor{jhu}{maya:} yes & (1) Stlye A vs B\\
something for you? &
4. \textcolor{ku}{randolph:} buy me some earplugs please & ...Robert: ...The Swedes didn't even bother to \textcolor{blue}{find out}...\\
4. \textcolor{jhu}{gabriel:} so that you get it as soon as you get &
5. \textcolor{jhu}{maya:} how many pairs? & they started \textcolor{blue}{laying them off}... (B) \\
to us? &
6. \textcolor{ku}{randolph:} 4 or 5 packs & Cynthia: ...i'd like us to go to this new bistro i \textcolor{blue}{discovered}... (A)\\
5. \textcolor{ku}{lilly:} good idea &
7. \textcolor{jhu}{maya:} i'll get you 5 & \\
6. \textcolor{ku}{lilly:} \colorbox{lime!}{pasta with salmon and basil} as always &
8. \textcolor{ku}{randolph:} thanks darling & \\
very tasty there & & \\ \cline{3-3}
\cline{1-2}
{\bf REF:} lilly will be late. gabriel will order \colorbox{lime!}{pasta with salmon and basil} for her.&
{\bf REF:} maya will buy 5 packs of earplugs for & \\
& randolph at the pharmacy. & (2) Style C\\ \cline{1-2} \cline{1-2}
& & {\shortstack[l]{...Iris : \textcolor{blue}{<file other>} My husband is famous...\\Haha. You don't even realize what this...}}\\
{\shortstack[l]{{\bf FE:} lilly will be late. lilly and gabriel are going \\ to pasta with salmon and basil is always tasty.\textcolor{orange}{[62/46/68]}}} &
{\shortstack[l]{{\bf FE:} randolph is in the pharmacy. \colorbox{magenta!}{randolph} will buy some earplugs \\ for randolph. maya will get 5.\textcolor{orange}{[64/38/71]}}} & ...Dan : \textcolor{blue}{<photo file>}...But its not working any more and it hurts \textcolor{blue}{:(}...\\
{\shortstack[l]{{\bf B:} lilly will be late. \colorbox{pink!}{gabriel and lilly} will order food for \colorbox{pink!}{lilly and gabriel}.\\ \textcolor{orange}{[72/39/63]}}} &
{\shortstack[l]{{\bf B:} \colorbox{magenta!}{laurie} is in the pharmacy. maya will buy 4 or 5 pairs of earphones\\ for him. \textcolor{orange}{[51/23/51]}}} &{\shortstack[l]{...Simon : \textcolor{blue}{BTW} it's so annoying that people can't see that \\such immigration policy reduces...}}\\ \cline{1-2}
{\shortstack[l]{{\bf SB:} lilly is going to be late. gabriel will order food for her. \\ \colorbox{lime!}{lilly will get pasta with salmon and basil} and \\ \colorbox{cyan}{she will get it as soon as she arrives at them.}\textcolor{orange}{[68/23/68]}}} &
{\shortstack[l]{{\bf SB:} maya will buy \colorbox{cyan!}{4} or 5 pairs of earplugs \\ for \colorbox{magenta!}{raymond} at the pharmacy. \textcolor{orange}{[63/34/63]}}} & \\
\bottomrule
\end{tabular}%
}
\caption{Examples of dialogues from each model. REF -- reference summary, FE -- Fast Abs RL Enhanced, B -- vanilla BART, and SB -- Syntax-aware BART (Ours). \textcolor{orange}{[R-1/R-2/R-L]} indicates F1 score from ROUGE-n. The error consists of the following factors \colorbox{pink!}{(i)}, \colorbox{magenta!}{(ii)}, and \colorbox{cyan}{(iii)}; otherwise, the accurate case is colored \colorbox{lime!}{lime}.}
\label{tab:qualitative-1}
\end{table*}
\subsection{Qualitative Results}
We compared the generated examples from several baselines including our proposed model in terms of their ROUGE scores. We present the qualitative results in Table~\ref{tab:qualitative-1}. We observed the error analysis through the following major error types -- \textit{(i) Incorrect reasoning}: indicates that the model came to the incorrect conclusion, which occurred when the generated summaries reasoned relations in the dialogue incorrectly. \textit{(ii) Incorrect reference}: indicates the association of one's locations or actions with an incorrect speaker, regardless of the original context in the generated summaries. \textit{(iii) Redundancy}: is the case wherein the content of the generated summaries was not mentioned in a reference. \textit{(iv) Missing information}: content existing in the reference is absent in generated summaries.
\paragraph{Error anaylsis}
In dialogue 1, FE ({\em{i.e.,}} Fast Abs RL Enhanced) and SB ({\em{i.e.,}} Syntax-aware BART) performed well at capturing the meaning of the reference summary, despite it being slightly lengthy. The B model ({\em{i.e.,}}vanilla BART model) showed the highest score but contained instances of (i) incorrect reasoning (\textit{gabriel and lilly}). However, our SB model highlighted ({\em{i.e.,}} lime-colored) the content influenced by the lengthy utterance at line 6. It appears that the model was affected by the lengthy input type. Alternatively, there is also the (iii) Redundancy case, as shown in SB. The related content appeared in dialogue but was absent from the reference.
\begin{figure}[hbt!]
\begin{center}
\includegraphics[width=0.5\linewidth]{./figures/speaker.png}
\end{center}
\caption{Two-dimensional PCA projection of each speaker style - A (margenta), B (blue), and C (purple). The legend indicates the center point of each cluster. }
\label{fig:speaker_pca}
\end{figure}
\begin{figure}[hbt!]
\begin{center}
\includegraphics[width=0.8\linewidth]{./figures/style_histogram.png}
\end{center}
\caption{Average tf-idf score on top six ranked POS features by standard deviation (std) according to the speaker styles (A, B, and C).}
\label{fig:group_analysis}
\end{figure}
In dialogue 2, the SB model showed the highest performance. FE mismatched (ii) the information of who was acting ({\em{i.e.,}} the subject) and who received the action ({\em{i.e.,}} the object). This observation also true of the B (\textit{laurie}) and SB (\textit{raymond}) models. Additionally, there was a case of (iii) redundancy in the SB (\textit{4}) model, despite existing content being present in the dialogue.
In sum, according to the observations listed in Table~\ref{tab:qualitative-1}, our proposed model captured the lengthy utterance well in terms of our objective to use the locational information. Nonetheless, there are limitations to this study, such as the inclusion of incorrect references. This is an area that we aim to improve in future work.
\begin{table}[H]
\centering
\resizebox{\columnwidth}{!}{
\begin{tabular}{c|cccccc}
\hline
& \textbf{$\sim$} & \textbf{T} & \textbf{E} & \textbf{G} & \textbf{\& } & \textbf{X} \\ \hline
Description & discourse/file marker & verb particle & emoticon & \begin{tabular}[c]{@{}c@{}}abbreviation,\\ foreign words\end{tabular} & \begin{tabular}[c]{@{}c@{}}coordinating \\ conjunction\end{tabular} & \begin{tabular}[c]{@{}c@{}}existential \textbackslash{}textit\{there\}, \\ predeterminers\end{tabular} \\ \hline
Example & @user:hello & up, out & ;-), :b & btw (by the way) & and, but & both, all, half \\ \hline
\end{tabular}}
\caption{POS tag description for the top-6 ranked.}
\label{tab:tag_info}
\end{table}
\paragraph{Speaker utterance style (Vitamin)}
We discovered the speaker styles from the test set by following the Section~\ref{sec:speaker_style} for ad-hoc analysis. In details, our research question\footnote{We carried out the group characteristics in Appendix~\ref{appendix:speaker_style}.} was \textit{``what are the differences between the speakers?''}.
Figure~\ref{fig:speaker_pca} depicts the distributional characteristics through PCA (principal components analysis) projection for each speaker style and K-means clustering (K=3). The style `A' and `B' have some intersection what is even shown in Figure~\ref{fig:group_analysis}, and `C' is relatively distant from other groups.
In Figure~\ref{fig:group_analysis}, we illustrate the top-6 ranked POS features to distinguish the groups (see the detail values in Appendix~\ref{appendix:speaker_style}). Style `B' specifies \textup{T} that represent the verb particle, and style `C' mainly consists of \textup{$\sim$, E, G} as different factors than other groups. However, style `A' shows the relatively flatten performance below 1.0 as Figure~\ref{fig:group_analysis} and Appendix~\ref{appendix:speaker_style}. In Table~\ref{tab:qualitative-1}'s right table, we compared the speaker styles: (1) Style `B' used verb particle (\textit{find out}), but `A' used the same representation in a different way(\textit{discovered}) and (2) Style `C' mostly tend to represent their intention using picture or contents. In the end, we found that the ability to distinguish those speaker styles was reflected in our model.
\section{Conclusion}
In this study, we proposed a novel syntax-aware sequence-to-sequence model that leverages syntactic information (i.e., POS tagging), considering the informal daily chat structure constraints, and distinguishes the different textual styles from multiple speakers for abstractive dialogue summarization. To strategically combine syntactic information to the dialogue summarization task, we adopted multi-task learning to reproduce both syntactic information and dialogue summarization. Furthermore, we presented a novel input type to train the model to explore locational biases in dialogue structures.
We benchmarked the experiments using the SAMSum corpus, and the experimental results demonstrate that the proposed method improves comparison models for all ROUGE scores.
There are promising future directions regarding this research. It would be worthwhile to apply the traditional truncation method with our proposed model to deeply compare performance differences.
|
\section{Introduction}
\label{section:Introduction}
We consider operators with $\lambda \geq 1$,
\begin{equation}
\label{eq:OscillatoryIntegralOperators}
T^\lambda f(x) = \int e^{i \phi^\lambda(x;\omega)} a^\lambda(x;\omega) f(\omega) d\omega
\end{equation}
and $a \in C_c^\infty(\mathbb{R}^{n} \times \mathbb{R}^{n-1},\mathbb{R})$, $\phi \in C^\infty(\mathbb{R}^n \times \mathbb{R}^{n-1} \backslash 0;\mathbb{R})$, $\phi^\lambda(x;\omega) = \lambda \phi(x/\lambda;\omega)$, $a^\lambda(x;\omega) = a(x/\lambda;\omega)$. We suppose that $\phi$ is $1$-homogeneous in $\omega$, i.e.,
\begin{equation}
\label{eq:HomogeneityPhaseFunction}
\phi(x;\mu \omega) = \mu \phi(x;\omega)
\end{equation}
for $\mu > 0$. For the support of $a$ we suppose that
\begin{equation*}
\text{supp}(a) \subseteq A^{n-1} = B_{n-1}(0,2) \backslash B_{n-1}(0,1/2).
\end{equation*}
We write $x=(x',x_n) \in \mathbb{R}^{n-1} \times \mathbb{R}$ and impose the following conditions on $\phi$ in $\text{supp}(a)$:
\begin{align*}
&C1) \quad \text{rank}(\partial^2_{x \omega} \phi) = n-1, \\
&C2^+) \quad \partial^2_{\omega \omega} \langle \partial_x \phi, G(x;\omega_0) \rangle \big|_{\omega = \omega_0} \text{ has } n-2 \text{ non-vanishing eigenvalues of the same sign},
\end{align*}
where $G$ denotes the Gauss map
\begin{equation}
\label{eq:GaussMap}
G_0(x;\omega) = \bigwedge_{j=1}^{n-1} \partial^2_{x \omega_j} \phi(x;\omega), \qquad G = G_0 / |G_0|
\end{equation}
of the embedded surface $\omega \mapsto \partial_x \phi(x;\omega)$. We identify $\bigwedge^{n-1} \mathbb{R}^n \simeq \mathbb{R}^n$.
The operators defined in \eqref{eq:OscillatoryIntegralOperators} naturally extend the adjoint Fourier restriction operator for the cone
\begin{equation}
\label{eq:ConeExtensionOperator}
\mathcal{E} f(x) = \int_{A^{n-1}} e^{i (x'.\omega + x_n |\omega|)} f(\omega) d\omega.
\end{equation}
In this note we prove new estimates
\begin{equation}
\label{eq:LpLpBoundVariableCoefficient}
\| T^\lambda f \|_{L^p(\mathbb{R}^n)} \lesssim_{\varepsilon,\phi,a} \lambda^\varepsilon \Vert f \Vert_{L^p(A^{n-1})}
\end{equation}
for operators \eqref{eq:OscillatoryIntegralOperators} like described above. Firstly, we recall the conjectured range of $L^p$-estimates
\begin{equation}
\label{eq:FourierExtensionCone}
\| \mathcal{E} f \|_{L^p(\mathbb{R}^n)} \lesssim \| f \|_{L^p(A^{n-1})}.
\end{equation}
is given by $p>\frac{2(n-1)}{n-2}$. This prominent open problem is known as \emph{restriction conjecture for the cone} and goes back to Stein.
The conjecture was solved for $n=3$ by Barcelo \cite{Taberner1985}, for $n=4$ by Wolff \cite{Wolff2001} via bilinear estimates, and for $n=5$ by Ou--Wang \cite{OuWang2021} via polynomial partitioning. Let
\begin{equation}
\label{eq:PolynomialPartitioningRange}
p_n = \begin{cases}
&4, \quad n=3, \\
&2 \cdot \frac{3n+1}{3n-3}, \quad n > 3 \text{ odd}, \\
&2 \cdot \frac{3n}{3n-4}, \quad n>3 \text{ even}.
\end{cases}
\end{equation}
Ou--Wang showed \eqref{eq:FourierExtensionCone} for $p > p_n$, which is also currently the widest range in higher dimensions to the best of the author's knowledge. Notably, in the case of Carleson-Sj\"olin phase functions (cf. \cite{CarlesonSjoelin1972,Hoermander1973}), which are not $1$-homogeneous anymore, where $C2^+)$ is replaced with
\begin{equation*}
H2^+) \quad \partial^2_{\omega \omega} \langle \partial_x \phi(x;\omega) , G(x;\omega_0) \rangle \big|_{\omega = \omega_0} \text{ has } n-1 \text{ eigenvalues of the same sign},
\end{equation*}
Guth--Hickman--Iliopoulou \cite{GuthHickmanIliopoulou2019} showed the sharp range of $L^p - L^p$ estimates, in the sense that there are phase functions for which the estimate fails for lower values of $p$. The deviation from the corresponding generalized restriction conjecture for the paraboloid occurs due to Kakeya compression. This was initially observed by Bourgain \cite{Bourgain1991}, see also Wisewell \cite{Wisewell2005} and Bourgain--Guth \cite{BourgainGuth2011}. Related phenomena were discussed by Minicozzi--Sogge \cite{MinicozziSogge1997} and Sogge \cite{Sogge1999}. In this note we point out Kakeya compression for $1$-homogeneous phases with variable coefficients, which shows that the following $L^p$-estimates are sharp up to endpoints:
\begin{theorem}
\label{thm:LpLpEstimatesVariableCoefficients}
Let $\phi:\mathbb{R}^n \times \mathbb{R}^{n-1} \backslash 0 \rightarrow \mathbb{R}$ be a $1$-homogeneous phase satisfying $C1)$ and $C2^+)$ and $a \in C^\infty_c(A^{n-1})$ be an amplitude. Then, we find the estimate \eqref{eq:LpLpBoundVariableCoefficient} to hold for $p \geq p_n$ with $p_n$ as in \eqref{eq:PolynomialPartitioningRange}.
\end{theorem}
We remark that for $p>p_n$ the $\lambda^\varepsilon$-factor can be dropped. Guth--Hickman--Iliopoulou showed the $\varepsilon$-removal lemma for oscillatory integral operators in \cite[Section~12]{GuthHickmanIliopoulou2019}, albeit with a stronger non-degeneracy hypothesis than presently considered. The idea goes back to Tao \cite{Tao1998,Tao1999}. In Section \ref{section:epsRemoval} we prove the following global estimates for $p > p_n$ by a small variation of the argument in \cite{GuthHickmanIliopoulou2019}:
\begin{equation*}
\| T^\lambda f \|_{L^p(\mathbb{R}^n)} \lesssim_{\phi,a} \| f \|_{L^p(A^{n-1})}.
\end{equation*}
\bigskip
The proof of Theorem \ref{thm:LpLpEstimatesVariableCoefficients} combines ideas from the case of constant-coefficient homogeneous phases due to Ou--Wang \cite{OuWang2021} and Gao--Liu--Miao--Xi \cite{GaoLiuMiaoXi2021} and variable-coefficient non-homogeneous phases due to Guth--Hickman--Iliopoulou \cite{GuthHickmanIliopoulou2019}. We digress for a moment to describe the tools we will use and put them into context. Bennett--Carbery--Tao \cite{BennettCarberyTao2006} delivered an important contribution with sharp $n$-multilinear restriction estimates. We note that the multilinear estimates were shown as well for constant-coefficient phase functions as smooth perturbations thereof. Bourgain--Guth \cite{BourgainGuth2011} devised an iteration to deduce linear estimates from multilinear estimates. Guth \cite{Guth2016} observed that the full strength of $k$-multilinear estimates is not required, but a slightly weaker variant given by $k$-broad norms suffices to run the iteration. He used polynomial partitioning to improve on the previous results in \cite{Guth2016,Guth2018}. The idea is to equipartition the broad norm with polynomials of controlled degree: After wave packet decomposition, one finds that either the broad norm is concentrated on ``cells" or on the ``wall", which is a neighbourhood of a variety. To oversimplify matters for a moment, if the broad norm is concentrated on the cells, then sharp bounds follow from induction on scales. If the broad norm is concentrated along the wall, then we are morally dealing with a restriction problem in lower dimensions, which is amenable to another induction hypothesis.
\medskip
We introduce the $k$-broad norms in the present context: For its definition decompose $A^{n-1}$ into finitely overlapping sectors $\tau$ of aperture $\sim K^{-1}$ and length $\sim 1$, where $K$ is a large constant. Given $f: A^{n-1} \to \mathbb{C}$, write $f = \sum f_\tau$, where $f_\tau$ is supported in $\tau$. In view of the rescaling $\phi^\lambda$ of the phase, we define the rescaled Gauss map
\begin{equation*}
G^\lambda(x;\omega) = G(\frac{x}{\lambda};\omega) \text{ for } (x;\omega) \in \text{supp } (a^\lambda).
\end{equation*}
For each $x \in B(0,\lambda)$
\begin{equation*}
G^\lambda(x;\tau) = \{ G^\lambda(x;\omega) \; : \; \omega \in \tau \text{ and } (x;\omega) \in \text{supp} a^\lambda \}.
\end{equation*}
For $V \subseteq \mathbb{R}^n$ a linear subspace, let $\angle(G^\lambda(x;\tau),V)$ denote the smallest angle between any non-zero vector $v \in V$ and $G^\lambda(x;\tau)$.
The spatial ball $B(0,\lambda)$ is decomposed into relatively small balls $B_{K^2}$ of radius $K^2$. We fix $\mathcal{B}_{K^2}$ a collection of finitely-overlapping $K^2$-balls, which are centred in and cover $B(0,\lambda)$. For $B_{K^2} \in \mathcal{B}_{K^2}$ centred at $\bar{x} \in B(0,\lambda)$, define
\begin{equation}
\label{eq:BroadNorm}
\mu_{T^\lambda f}(B_{K^2}) = \min_{V_1,\ldots,V_A \in Gr(k-1,n)} \big( \max_{\tau: \angle(G^\lambda(\bar{x};\tau),V_a) > K^{-2} \; \forall a} \| T^\lambda f_{\tau} \|^p_{L^p(B_{K^2})} \big),
\end{equation}
where $Gr(k-1,n)$ denotes the Grassmannian manifold of $(k-1)$-dimensional subspaces in $\mathbb{R}^n$. We stress the deviation from \cite{GuthHickmanIliopoulou2019}, in which the angle threshold $K^{-1}$ was considered. In case of the Fourier extension operator associated with the cone, we have to strengthen the angle condition to $K^{-2}$ to further confine the narrow part.
We write $\tau \not \in V_a$ as shorthand for $\angle(G^\lambda(\bar{x};\tau),V_a) > K^{-2}$ provided that $\bar{x}$ is clear from context. Thus, we can write as well
\begin{equation*}
\mu_{T^\lambda f}(B_{K^2}) = \min_{V_1,\ldots,V_A \in Gr(k-1,n)} \big( \max_{\substack{\tau: \tau \not \in V_a, \\ \text{ for } 1 \leq a \leq A}} \| T^\lambda f_\tau \|^p_{L^p(B_{K^2})} \big).
\end{equation*}
For $U \subseteq \mathbb{R}^n$ the $k$-broad norm is defined as
\begin{equation*}
\| T^\lambda f \|_{BL^p_{k,A}}(U) = \big( \sum_{\substack{B_{K^2} \in \mathcal{B}_{K^2}, \\ B_{K^2} \cap U \neq \emptyset}} \mu_{T^\lambda f}(B_{K^2}) \big)^{1/p}.
\end{equation*}
A key step in the proof of the $L^p$-$L^p$-estimate is the proof of $k$-broad estimates:
\begin{theorem}
\label{thm:kBroadEstimate}
For $2 \leq k \leq n$ and all $\varepsilon > 0$, there exists a constant $C_\varepsilon > 1$ and an integer $A$ such that, whenever $T^\lambda$ is an oscillatory integral operator with reduced $1$-homogeneous phase satisfying $C1)$ and $C2^+)$, the estimate
\begin{equation}
\| T^\lambda f \|_{BL^p_{k,A}(\mathbb{R}^n)} \lesssim_\varepsilon K^{C_\varepsilon} \lambda^\varepsilon \| f \|_{L^2(A^{n-1})}
\end{equation}
holds for all $\lambda \geq 1$ and $K \geq 1$ whenever
\begin{equation}
p \geq \bar{p}(k,n) = \frac{2(n+k)}{n+k-2}.
\end{equation}
\end{theorem}
Reduced phase functions are introduced in Section \ref{subsection:BasicReductions}. These phases are basically small $C^N$ perturbations of $1$-homogeneous phases with constant coefficients. These reduced phase functions were previously used by Beltran--Hickman--Sogge \cite{BeltranHickmanSogge2020} to derive decoupling estimates. As in \cite{BeltranHickmanSogge2020}, general phase functions satisfying $C1)$ and $C2^+)$ are transformed by partitioning the support of the amplitude and Lorentz rescaling to reduced phases.
The arguments to deduce Theorem \ref{thm:LpLpEstimatesVariableCoefficients} from Theorem \ref{thm:kBroadEstimate} are essentially due to Bourgain--Guth and we give a sketch in the following: As mentioned previously, it is enough to work with reduced phases and amplitudes.
Firstly, we write
\begin{equation*}
\| T^\lambda f \|^p_{L^p(B(0,\lambda))} \lesssim \sum_{B_{K^2} \in \mathcal{B}_{K^2}} \| T^\lambda f \|^p_{L^p(B_{K^2})}.
\end{equation*}
Fixing one $K^2$-ball, there is a collection of $(k-1)$-dimensional subspaces $V_1,\ldots,V_A$ such that
\begin{equation*}
\mu_{T^\lambda f}(B_{K^2}) = \max_{\substack{\tau \not \in V_a \\ \text{for } 1 \leq a \leq A}} \| T^\lambda f_\tau \|^p_{L^p(B_{K^2})}.
\end{equation*}
Note that there are only $K^{O(1)}$ $K^{-1}$-sectors. Writing $T^\lambda f = \sum_\tau T^\lambda f_\tau$, sectors $\tau \not \in V_a$ for $1 \leq a \leq A$ are estimated by $\mu_{T^\lambda f}(B_{K^2})$. The remaining sectors are isolated. This yields
\begin{equation}
\label{eq:BroadNarrowEstimate}
\| T^\lambda f\|^p_{L^p(B_{K^2})} \lesssim_A K^{O(1)} \mu_{T^\lambda f}(B_{K^2}) + \sum_{a=1}^A \| \sum_{\tau \in V_a} T^\lambda f_\tau \|^p_{L^p(B_{K^2})}.
\end{equation}
The first term is captured by the broad estimate; the second term is estimated by $\ell^p$-decoupling (cf. \cite{BourgainDemeter2015,BeltranHickmanSogge2020}) and induction on scales \cite{OuWang2021,GaoLiuMiaoXi2021}.
\medskip
Very recently, Gao--Liu--Miao--Xi \cite{GaoLiuMiaoXi2021} proved an extension of Ou--Wang's result for the circular cone $\phi(x,\omega) = x'.\omega+x_n|\omega|$ for more general conic surfaces, but still with constant coefficients. For these constant coefficient phase functions, the Kakeya compression described in the present work cannot happen. Gao \emph{et al.} \cite{GaoLiuMiaoXi2021} used $k$-broad estimates to derive new local smoothing estimates for the wave equation in Euclidean space. At small spatial scales, the variable coefficient phases are approximated with extension operators for conic surfaces. Then we can use arguments from \cite{GaoLiuMiaoXi2021}. Furthermore, in \cite{HickmanIliopoulou2020} Hickman and Iliopoulou showed sharp $L^p$-estimates for non-homogeneous phases with indefinite signature. This suggests to study also homogeneous phase functions with indefinite signature with the methods of the paper.
\medskip
Notably, we do not use the usual wave packet decomposition for the cone as e.g. in \cite{OuWang2021} or \cite{GaoLiuMiaoXi2021} to prove the broad estimate. Instead, we stick to the wave packet decomposition commonly used for the Fourier extension operator of the paraboloid or its variable coefficient counterpart \cite{GuthHickmanIliopoulou2019}. This allows to use many arguments from \cite{GuthHickmanIliopoulou2019} without change and hints at the possibility of a unified approach. A major change happens for the transverse equidistribution estimates, to be analyzed in Section \ref{section:TransverseEquidistribution}. Secondly, the narrow decoupling requires additional considerations, see Section \ref{section:LinearEstimates}. Since the polynomial partitioning approach is involved, we elected to elaborate on the argument in Sections \ref{section:Preliminaries} - \ref{section:LinearEstimates}.
We remark that the idea to use the same wave packet decomposition for homogeneous and inhomogeneous phase functions in the variable coefficient context is not new: In \cite{Lee2006} Lee considered linear and bilinear estimates for oscillatory integral operators and could treat variable coefficient versions of the Fourier extension operator of the paraboloid and the cone with the same wave packet decomposition. He generalized bilinear estimates due to Tao \cite{Tao2003} and Wolff \cite{Wolff2001} to variable coefficient phases. Notably, in \cite{Lee2006} was pointed out for the first time that a convexity condition as $H2^+)$ or $C2^+)$ allows to go beyond Tomas--Stein $L^2-L^p$-estimates, which are sharp for phases without convexity condition. Bourgain \cite{Bourgain1991} showed in the context of non-homogeneous phases without convexity conditions that the Tomas-Stein range is sharp (see also \cite{BourgainGuth2011}). In the present work, the $L^p$-$L^p$-estimates for general oscillatory integral operators with phase satisfying $C1)$ and $C2^+)$ due to Lee \cite{Lee2006} are improved to the sharp range up to the endpoint for $n \geq 5$.
\medskip
In Section \ref{section:LocalSmoothing} we apply the new estimates for oscillatory integral operators to prove new local smoothing estimates for solutions to wave equations on compact Riemannian manifolds with $\dim(M) \geq 3$.
Local parametrices are given by
\begin{equation*}
\mathcal{F} f(x',x_n) = \int_{\mathbb{R}^{n-1}} e^{i \phi(x',x_n;\omega)} a(x;\omega) \hat{f}(\omega) d\omega
\end{equation*}
with $\phi \in C^\infty(\mathbb{R}^n \times \mathbb{R}^{n-1} \backslash 0)$ a phase function satisfying $C1)$ and $C2^+)$ and $a \in S^0(\mathbb{R}^{2d+1})$ with compact support in $x$. Hence, it suffices to prove local smoothing estimates of rescaled Fourier integral operators $\mathcal{F}^\lambda$. In Theorem \ref{thm:ImprovedLocalSmoothing} and Corollary \ref{cor:InterpolatedSmoothing} we extend the recent results due to Gao \emph{et al.} \cite{GaoLiuMiaoXi2021} for the Euclidean wave equation to wave equations on compact Riemannian manifolds. This improves on the previously best local smoothing estimates due to Beltran--Hickman--Sogge \cite{BeltranHickmanSogge2020} for wave equations on compact manifolds.
\medskip
\emph{Outline of the paper:} In Section \ref{section:NecessaryConditions} we show the necessary conditions for $L^p$-estimates for variable-coefficient $1$-homogeneous phases. Preliminaries for the polynomial partitioning argument to show Theorem \ref{thm:kBroadEstimate} are given in Section \ref{section:Preliminaries}. In this section we introduce the notion of a reduced homogeneous phase function and collect geometric consequences. This will simplify the proof of Theorem \ref{thm:kBroadEstimate}. We recall the wave packet analysis in the context of variable coefficients \cite{Lee2006,GuthHickmanIliopoulou2019} and collect facts on the $k$-broad norms. In Section \ref{section:PolynomialPartitioning} we recall the polynomial partitioning tools. In Section \ref{section:TransverseEquidistribution} transverse equidistribution estimates are proved. These differ from the transverse equidistribution estimates shown in \cite{GuthHickmanIliopoulou2019} for Carleson--Sj\"olin phase functions. In Section \ref{section:ComparingWavePackets} we compare wave packet decompositions at different scales, which is necessary to run the induction on scales in Section \ref{section:MainInductiveArgument}. In this section we deduce Theorem \ref{thm:kBroadEstimate} from Theorem \ref{thm:MainInductionTheorem}, which is suitable for induction on dimension and radius. In Section \ref{section:LinearEstimates} we show how Theorem \ref{thm:kBroadEstimate} implies Theorem \ref{thm:LpLpEstimatesVariableCoefficients}. In Section \ref{section:epsRemoval} we show how the $\lambda^\varepsilon$-factor can be removed away from the endpoint.In Section \ref{section:LocalSmoothing} we apply the oscillatory integral estimates to show new local smoothing estimates for solutions to wave equations on compact Riemannian manifolds.
\section{Kakeya compression}
\label{section:NecessaryConditions}
In the following we modify the example due to Guth--Hickman--Iliopoulou \cite[Section~2]{GuthHickmanIliopoulou2019} (see also \cite{BourgainGuth2011}) for homogeneous phase functions. This yields the necessary conditions:
\begin{proposition}
\label{prop:Necessary}
Necessary for the estimate \eqref{eq:LpLpBoundVariableCoefficient} to hold for $n \geq 5$ is
\begin{equation*}
p \geq
\begin{cases}
&2 \cdot \frac{3n}{3n-4}, \quad n \text{ even}, \\
&2 \cdot \frac{3n+1}{3n-3}, \quad n \text{ odd}.
\end{cases}
\end{equation*}
\end{proposition}
We only consider $n \geq 5$ because the bilinear estimates due to Wolff \cite{Wolff2001} and Lee \cite{Lee2006} solve the cone restriction conjecture for $3 \leq n \leq 4$. Let
\begin{equation*}
x=(\underbrace{x^{\prime \prime},x_{n-1}}_{x^\prime},x_n) \in \mathbb{R}^{n-2} \times \mathbb{R} \times \mathbb{R} \text{ and } \omega = (\omega^\prime,\omega_{n-1}) \in \mathbb{R}^{n-2} \times \mathbb{R}.
\end{equation*}
We consider the phase functions
\begin{equation}
\label{eq:PhaseFunction}
\phi(x;\omega) = x^\prime.\omega + \frac{\langle A(x_n) \omega^\prime, \omega^\prime \rangle}{2 \omega_{n-1}}, \quad \omega_{n-1} \in (1/2,1).
\end{equation}
$A(x_n)$ denotes the $(n-2) \times (n-2)$--positive definite matrix
\begin{equation*}
A(x_n) =
\begin{cases}
\bigoplus_{i=1}^{\frac{n-2}{2}}
\begin{pmatrix}
x_n & x_n^2 \\
x_n^2 & x_n +x_n^3
\end{pmatrix}, \quad n-2 \text{ even}, \\
\bigoplus_{i=1}^{\frac{n-3}{2}}
\begin{pmatrix}
x_n & x_n^2 \\
x_n^2 & x_n +x_n^3
\end{pmatrix}
\oplus (x_n), \quad n-2 \text{ odd}.
\end{cases}
\end{equation*}
The main idea is to construct many wave packets which are concentrated in the neighbourhood of a lower dimensional algebraic variety. Whereas the direction governed by the frequency $\omega_\theta$ below varies, for fixed $\omega_\theta$ we consider precisely one starting position $v_\theta$. This concentration in a low dimensional algebraic variety does not happen in the linear case \eqref{eq:ConeExtensionOperator}.
We consider wave packets adapted to $\phi$ as follows: $\Xi = (B^{n-2}(0,c) \times (1/2,1)$ is covered by essentially disjoint elongated caps
\begin{equation*}
\Xi_{\theta} = \{ (\omega',\omega_{n-1}) \in \Xi : | \omega'/\omega_{n-1} - \omega_{\theta}| \lesssim \lambda^{-\frac{1}{2}} \}
\end{equation*}
with $\omega_{\theta} \in B^{n-2}(0,c)$ for $|c| \ll 1$. Apparently, $\Xi$ can be covered by $\sim \lambda^{\frac{ n-2}{2}}$ finitely overlapping sets $\Xi_{\theta}$. We consider a corresponding smooth partition of unity $(\psi_\theta)_{\omega_\theta \in \Xi}$ and wave packets
\begin{equation*}
f_{\theta,v}(\omega) = e^{-i \lambda \langle v, \omega^\prime \rangle} \psi_\theta(\omega), \quad v = (v_1,\ldots,v_{n-2}) \in \mathbb{R}^{n-2}.
\end{equation*}
We have by non-oscillation of the phase
\begin{equation*}
| T^\lambda f_{\theta,v}(x^{\prime \prime},x_{n-1},x_n) | \gtrsim \lambda^{- \frac{n-2}{2}} \chi_{T_{\theta,v}}(x).
\end{equation*}
$\chi_{T_{\theta,v}}$ denotes the characteristic function of $T_{\theta,v}$. The $T_{\theta,v}$ are curved slabs of size $( 1 \times \underbrace{\lambda^{1/2} \times \ldots \times \lambda^{1/2}}_{ n-2 \text{ times}} \times \lambda )$ with
\begin{equation*}
T_{\theta,v} \subseteq \{ x \in B(0,\lambda) \, : \, |x^{\prime \prime} - \lambda \gamma_{\theta,v} \left( \frac{x_n}{\lambda} \right) | < c \lambda^{\frac{1}{2}+\varepsilon} \text{ and } |x_{n-1} - \lambda \gamma^\prime_{\theta}(x_n / \lambda) | < c \lambda^{\frac{1}{2}+\varepsilon} \},
\end{equation*}
for any $\varepsilon >0$, which follows from non-stationary phase; $c$ denotes a small constant and $\gamma_{\theta,v}$, $\gamma_{\theta}'$ denote curves:
\begin{equation*}
\gamma_{\theta,v}(x_n) = v - A(x_n) \omega_\theta, \quad \gamma^\prime_\theta (x_n) = \frac{1}{2} \langle A(x_n) \omega_\theta, \omega_\theta \rangle.
\end{equation*}
Furthermore, note that the condition
\begin{equation*}
\big| \frac{\omega'}{\omega_{n-1}} - \omega_\theta \big| \lesssim \lambda^{-\frac{1}{2}}, \; \omega_{n-1} \in (1,2), \; \omega' \in B(\omega,c \lambda^{-\frac{1}{2}})
\end{equation*}
corresponds to considering $\lambda^{-\frac{1}{2}}$-sectors into direction $(\omega_\theta,1)$. The degeneracy of $\partial^2 \phi$ into radial direction gives the localization of tubes to size $\lambda^\varepsilon$ into this direction: We have
\begin{equation*}
\partial_\omega \phi(x;\omega).\frac{(\omega_\theta,1)}{|(\omega_\theta,1)|} = \partial_\omega \phi(x;(\omega_\theta,1)).\frac{(\omega_\theta,1)}{|(\omega_\theta,1)|} + O(\lambda^{-1}) \text{ for } | \frac{\omega}{|\omega|} - \frac{(\omega_\theta,1)}{|(\omega_\theta,1)|} | \lesssim \lambda^{-\frac{1}{2}}.
\end{equation*}
The non-degeneracy of $\partial^2 \phi$ gives localization to size $\lambda^{1/2+\varepsilon}$ into the remaining directions. We argue in the following why the curved tubes $\chi_{T_{\theta,v}}$ are in fact of size $1 \times \lambda^{\frac{1}{2}} \times \ldots \times \lambda^{\frac{1}{2}} \times \lambda$: Consider the oscillatory integral
\begin{equation*}
F(x') = \int e^{i(x'.\omega + \lambda \tilde{\phi}(x_n/\lambda,\omega))} \psi_{\theta}(\omega) d\omega
\end{equation*}
with $\psi_{\theta} \in C^\infty_c(A^{n-1})$ localizing to a slab into direction $\theta \in \mathbb{S}^{n-2}$ and
\begin{equation*}
\tilde{\phi}(x_n,\mu \omega)= \mu \tilde{\phi}(x_n,\omega) \text{ for } \mu > 0.
\end{equation*}
We use Taylor expansion in $\omega$ to write
\begin{equation*}
\begin{split}
\lambda \tilde{\phi}(x_n/\lambda,\omega) &= |\omega| (\lambda \tilde{\phi}(x_n/\lambda,\omega/|\omega|)) \\
&= |\omega| (\lambda \tilde{\phi}(x_n/\lambda,\theta) + \lambda \nabla_{\omega} \tilde{\phi}(x_n/\lambda,\theta) ( \frac{\omega}{|\omega|} - \theta) + O(c)).
\end{split}
\end{equation*}
For $\omega \in \text{supp}( \psi_{\theta})$ we have
\begin{equation*}
|\omega| = \omega.\theta + O(c \lambda^{-1}).
\end{equation*}
Hence, we can write
\begin{equation*}
\lambda \tilde{\phi}(x_n/\lambda,\omega) = \lambda \tilde{\phi}(x_n/\lambda,\theta)(\omega.\theta) + \lambda \nabla_{\omega} \tilde{\phi}(x_n/\lambda,\theta)(\omega - (\omega.\theta) \theta) + O(c).
\end{equation*}
Let $\{\theta^1_{\perp},\ldots,\theta_{\perp}^{n-2},\theta \}$ be an orthonormal basis of $\mathbb{R}^{n-1}$. Then,
\begin{equation*}
\lambda \tilde{\phi}(x_n/\lambda,\omega) = \lambda \tilde{\phi}(x_n/\lambda,\theta)(\omega.\theta) + \lambda \sum_{i=1}^{n-2} (\nabla_\omega \tilde{\phi}(x_n/\lambda,\theta).\theta_{\perp}^i ) (\omega.\theta_{\perp}^i) + O(c).
\end{equation*}
Consequently,
\begin{equation*}
\begin{split}
x'.\omega + \lambda \tilde{\phi}(x_n/\lambda,\omega) &= (x'.\theta + \lambda \tilde{\phi}(x_n/\lambda,\theta))(\omega.\theta) \\
&+ \sum_{i=1}^{n-2} (x'.\theta_{\perp}^i + \lambda \nabla_{\omega} \tilde{\phi}(x_n/\lambda,\theta).\theta_{\perp}^i )(\omega.\theta_{\perp}^i) + O(c).
\end{split}
\end{equation*}
And for $|x'.\theta + \lambda \tilde{\phi}(x_n/\lambda,\theta)| \ll c$ and $|x'.\theta_{\perp}^i + \lambda \nabla_{\omega} \tilde{\phi}(x_n/\lambda,\theta).\theta_{\perp}^i| \ll c \lambda^{1/2}$, we see that the whole phase is $O(c)$. Hence, there is no oscillation within $\text{supp}(\psi_{\theta})$ and for fixed $x_n$ this defines a region $A_{x_n}$ for $x'$ of size $1 \times \lambda^{1/2} \times \ldots \times \lambda^{1/2}$. Taking $T_{\theta} = \bigcup_{x_n} A_{x_n}$ yields the $1 \times \lambda^{1/2} \times \ldots \times \lambda^{1/2} \times \lambda$-tube. Note that the factor $e^{-i \lambda \langle v,\omega' \rangle}$ amounts to a shift in $x'$ by $\lambda v$, but does not change the size of the tube.
\medskip
\noindent We prepare the initial data with randomized signs:
\begin{equation*}
f = \sum_{\theta} \varepsilon_\theta f_{\theta,v}.
\end{equation*}
By Khintchine's theorem, the expected value of $|T^\lambda f(x)|$ is given by the square sum:
\begin{equation*}
\mathbf{E}[|T^\lambda f(x)|] \sim \big( \sum_\theta |T^\lambda f_{\theta,v_\theta} |^2 \big)^{1/2} \gtrsim \lambda^{-\frac{n-2}{2}} \big( \sum_\theta \chi_{T_{\theta,v_\theta}}(x) \big)^{1/2}.
\end{equation*}
Taking $L^p$-norms yields by Minkowski's inequality
\begin{equation*}
\lambda^{- \frac{n-2}{2}} \big( \int \big( \sum_{\theta} \chi_{T_{\theta,v_{\theta}}} \big)^{\frac{p}{2}} \big)^{\frac{1}{p}} \lesssim \mathbf{E}[ \| T^\lambda f \|_{L^p} ].
\end{equation*}
Next, we find by applying H\"older's inequality
\begin{align*}
\lambda^{-\frac{n-2}{2}} \big( \int \sum_\theta \chi_{T_{\theta,v_\theta}} \big)^{1/2} &\lesssim \left| \bigcup_{\theta} T_{\theta,v_\theta} \right|^{1/2 - 1/p} \mathbf{E}[ \| T^\lambda f \|_{L^p} ] \\
&\lesssim \left| \bigcup_{\theta} T_{\theta,v_\theta} \right|^{1/2-1/p} \| f \|_p \lesssim \left| \bigcup_{\theta} T_{\theta,v_\theta} \right|^{1/2 - 1/p}.
\end{align*}
The penultimate estimate is by hypothesis, and the final estimate follows from $|f| = 1$ and $|\text{supp } f| \sim 1$. Since the tubes $T_{\theta,v_\theta}$ are $( 1 \times \lambda^{1/2} \times \ldots \times \lambda^{1/2} \times \lambda)$-slabs, $\int \chi_{T_{\theta, v_\theta}} \sim \lambda^{\frac{n}{2}}$. Moreover, there are about $\lambda^{\frac{n-2}{2}}$ slabs. Hence,
\begin{equation*}
\lambda^{-\frac{n-2}{2}} \big( \int \sum \chi_{T_{\theta, v_\theta}} \big)^{1/2} \sim \lambda^{\frac{1}{2}}.
\end{equation*}
Thus, we arrive at
\begin{equation}
\label{eq:EstimateLp}
1 \sim \| f \|_{L^p(B^{n-1})} \lesssim \left| \bigcup_{\theta} T_{\theta,v_{\theta}} \right|^{1/2 - 1/p} \lambda^{-\frac{1}{2} } \| f \|_{L^p(B^{n-1})}.
\end{equation}
Next, we shall see how to choose $v_\theta$ such that the curved slabs are concentrated in a neighbourhood of a low-dimensional algebraic variety inspired by \cite{GuthHickmanIliopoulou2019}.\\
For $\Xi_{\theta}$, we set
\begin{equation}
\label{eq:StartingPosition}
v_{\theta,2j-1} = - (\omega_{\theta})_{2j} \text{ and } v_{\theta,2j} = v_{\theta,n-2} = 0 \text{ for } 1 \leq j \leq \lfloor \frac{n-2}{2} \rfloor.
\end{equation}
Let $d = n-1 - \lfloor \frac{n-2}{2} \rfloor$ and $Z = Z(P_1,\ldots,P_{n-1-d})$ be the common zero set of the polynomials
\begin{equation}
\label{eq:Polynomials}
P_j(x_1,\ldots,x_{n-2},x_n) = \lambda x_{2j} - x_{2j-1} x_n \text{ for } 1 \leq j \leq \lfloor \frac{n-2}{2} \rfloor.
\end{equation}
It is straight-forward to show that $x_n \mapsto (\lambda \gamma_{\theta,v_\theta}(x_n/\lambda),x_n)$ is contained in $Z(P_1,\ldots,P_{n-1-d})$. $Z$ is an algebraic variety of dimension
\begin{equation}
\label{eq:DimensionVariety}
d=(n-1) - \lfloor \frac{n-2}{2} \rfloor
\end{equation}
in $\mathbb{R}^{n-1}$ and of degree $O_n(1)$. Thus, Wongkew's theorem (cf. Theorem \ref{thm:WongkewCovering}) on the size of neighbourhoods of algebraic varieties applies, and we find
\begin{equation}
\label{eq:SizeEstimateNeighbourhood}
|N_{\lambda^{1/2}}(Z) \cap B^{n-1}(0,\lambda) | \lesssim \lambda^{d+\frac{n-1-d}{2}}.
\end{equation}
We find by \eqref{eq:DimensionVariety} and \eqref{eq:SizeEstimateNeighbourhood}
\begin{equation}
\label{eq:EstimateVarietyII}
|N_{\lambda^{\frac{1}{2}}}(Z)| \lesssim
\begin{cases}
\lambda^{\frac{3n-2}{4}}, \quad n \text{ even}, \\
\lambda^{\frac{3n-1}{4}}, \quad n \text{ odd.}
\end{cases}
\end{equation}
Moreover, for $(x_1,\ldots,x_n) \in T_{\theta,v_\theta}$ we have $x_{n-1} \in B(\lambda \gamma'_\theta(x_n/\lambda),\lambda^{\frac{1}{2}+\varepsilon})$.
This yields
\begin{equation}
\label{eq:EstimateTubeUnion}
\left| \bigcup_{\theta} T_{\theta,v_\theta} \right|^{1/2 - 1/p} \lesssim |N_{\lambda^{1/2}}(Z) \cdot \lambda^{\frac{1}{2}}|^{1/2- 1/p}.
\end{equation}
Plugging \eqref{eq:EstimateTubeUnion} into \eqref{eq:EstimateLp} with the estimate from \eqref{eq:EstimateVarietyII}, we find
\begin{equation*}
p \geq
\begin{cases}
&2 \cdot \frac{3n}{3n-4}, \quad n \text{ even}, \\
&2 \cdot \frac{3n+1}{3n-3}, \quad n \text{ odd}.
\end{cases}
\end{equation*}
This finishes the proof of Proposition \ref{prop:Necessary}. $\hfill \Box$
\section{Preliminaries}
\label{section:Preliminaries}
\subsection{Basic reductions of the phase function}
\label{subsection:BasicReductions}
In this paragraph we shall reduce $1$-homogeneous phase functions satisfying the above assumptions to a form, which highlights that the class of considered phase functions are indeed smooth perturbations of the translation-invariant case
\begin{equation*}
\phi_*(x;\omega) = \langle x', \omega \rangle + \frac{t (\omega')^2}{\omega_{n-1}}, \quad \omega' \in B(0,c), \; \omega_{n-1} \in (1,2).
\end{equation*}
Constant-coefficient perturbations were analyzed in \cite{GaoLiuMiaoXi2021}.
The arguments were provided in \cite[Section~2]{BeltranHickmanSogge2020} and details are omitted here (see also \cite{Lee2006}). After localisation and translation, we may assume that $a$ is supported inside $X \times \Xi$, where $X = X' \times T$ for $X' \subseteq B(0,1) \subseteq \mathbb{R}^{n-1}$ and $T \subseteq (-1,1) \subseteq \mathbb{R}$ are small open neighbourhoods of the origin and $\Xi \subseteq A^{n-1}$ is a small sector around $e_{n-1} = (0,\ldots,0,1) \in \mathbb{R}^{n-1}$.\\
Firstly, we can suppose that
\begin{align*}
&C1^\prime) \quad \det \partial^2_{\omega x'} \phi(x;\omega) \neq 0 \text{ for all } (x,\omega) \in X \times \Xi, \\
&C2^\prime) \quad \partial^2_{\omega^\prime \omega^\prime} \partial_{x_n} \phi(x,\omega) \text{ has eigenvalues of the same sign for all } (x,\omega) \in X \times \Xi.
\end{align*}
This follows as in \cite{BeltranHickmanSogge2020}. By rotation in the $x$-variables, we can also suppose that
\begin{equation*}
G(0;e_{n-1}) = e_n \text{ and } \partial^2_{x_n \omega} \phi(0;e_{n-1}) = 0.
\end{equation*}
Hence, by making $\Xi$ small enough, we find that
\begin{equation}
\label{eq:XiXnDerivativeBound}
|\partial^2_{x_n \omega} \phi(x;\omega)| \leq c_{par} \text{ for } (x,\omega) \in X \times \Xi.
\end{equation}
By non-degeneracy $C1')$, we find a smooth locally inverse mapping $\Phi_{x_n,\omega} : X' \to \mathbb{R}^{n-1}$ such that
\begin{equation*}
\partial_{\omega} \phi(\Phi_{x_n,\omega}(x'),x_n;\omega) = x'.
\end{equation*}
We shall also write $\Phi_{x_n,\omega}(x') = \Phi(x',x_n;\omega)$. There is also a smooth mapping $\Psi(x,\cdot)$ with
\begin{equation*}
\partial_{x^\prime} \phi(x;\Psi(x;\omega)) = \omega.
\end{equation*}
For $\lambda \geq 1$, we consider the rescaled versions $\Phi^\lambda(x;\omega) = \lambda \Phi(x/\lambda;\omega)$ and $\Psi^\lambda(x;\omega) = \Psi(x/\lambda;\omega)$. We assume that $X$ and $\Xi$ are such that the above mappings are defined on the whole support of $a$.
In the following we shall quantify the deviation from $\phi_*$ further, by restricting the values of second and third derivatives and bounding higher derivatives:
Let $c_{par} > 0$ denote a small constant. Firstly note that there are (possibly large) constants $A_1,A_2,A_3 \geq 1$ such that
\begin{align*}
&C1^{\prime \prime}) \quad |\partial^2_{\omega x^\prime} \phi(x;\omega) - I_{n-1}| \leq c_{par} A_1 \text{ for } (x;\omega) \in X \times \Xi,\\
&C2^{\prime \prime}) \quad |\partial^2_{\omega^\prime \omega^\prime} \partial_{x_n} \phi(x;\omega) - \frac{I_{n-1}}{\omega_{n-1}} | \leq c_{par} A_2 \text{ for } (x;\omega) \in X \times \Xi.
\end{align*}
For the higher derivatives, we suppose that
\begin{align*}
&D1) \; \| \partial^{\beta}_\omega \partial_{x_k} \phi \|_{L^\infty(Z \times \Xi)} \leq c_{par} A_1 \text{ for } 1 \leq k \leq n-1 \text{ and } \beta \in \mathbb{N}_0^{n-1} \text{ with } 2 \leq |\beta| \leq 3 \\
&\qquad \text{ such that } |\beta^\prime| \geq 2; \\
&\qquad \| \partial^{\beta^\prime}_{\omega^\prime} \partial_{x_n} \phi \|_{L^\infty(Z \times \Xi)} \leq \frac{c_{par}}{2n} A_1 \text{ for all } \beta^\prime \in \mathbb{N}_0^{n-2} \text{ with } |\beta^\prime| = 3, \\
&D2) \; \text{For some large integer } N \in \mathbb{N}, \text{ one has} \\
&\qquad \| \partial_\omega^\beta \partial_x^\alpha \phi \|_{L^\infty(X \times \Xi)} \leq \frac{c_{par}}{2n} A_3 \text{ for all } (\alpha,\beta) \in \mathbb{N}_0^{n} \times \mathbb{N}_0^{n-1} \text{ with } 2 \leq |\alpha| \leq 4N \\
&\qquad \text{and } 1 \leq |\beta| \leq 2N+2 \text{ satisfying } 1 \leq |\beta| \leq 2N \text{ or } |\beta^\prime| \geq 2.
\end{align*}
A phase $\phi$ satisfying the above four conditions with constants $A_1$, $A_2$, $A_3$ is said to be of type $(A_1,A_2,A_3)$.
By parabolic rescaling (cf. Lemma \ref{lem:ParabolicRescaling}), we see that we can reduce to initial data with $A_i = 1$; these phases are said to be reduced.
\subsection{Geometric consequences}
Let $\phi$ be a reduced phase function in the above sense. We shall see how the corresponding hypersurfaces $\Sigma_x$ parametrized by $\omega \mapsto \partial_x \phi(x;\omega)$ resemble the ones from $\phi_*$. To see this, recall that $\Psi: U \rightarrow \Xi$ satisfies $\partial_{x^\prime} \phi(x;\Psi(x;\omega)) = \omega$. Hence, $\Sigma_x$ is the graph of the function $h_x(\omega) = \partial_{x_n} \phi(x;\Psi(x;\omega))$ over the fibre $U_x$.\\
Each $h_x$ is a perturbation of the translation invariant case in the following sense:
\begin{lemma}
The following estimate holds true for all $\omega \in U_x$:
\begin{equation}
\label{eq:HPerturbation}
\| \partial^2_{\omega^\prime \omega^\prime} h_x(\omega) - I_{n-1}/ \omega_{n-1} \|_{L^\infty} \lesssim c_{par}.
\end{equation}
Here $c_{par}>0$ denotes the constant from the definition of a reduced phase function.
\end{lemma}
\begin{proof}
This is a consequence of properties of $\Psi$. Firstly, we record that $\Psi(x;e_{n-1}) = 1$. By the implicit function theorem and non-degeneracy of $\phi$, we find
\begin{equation*}
\partial_\omega \Psi(x;\omega) = \partial^2_{x^\prime \omega} \phi(x;\Psi(x;\omega))^{-1}.
\end{equation*}
Hence,
\begin{equation*}
\| \partial_\omega \Psi(x;\omega) - I_{n-1} \|_{L^\infty} = O(c_{par}).
\end{equation*}
As a consequence of this identity (and choosing $c_{par}$ to be sufficiently small),
\begin{equation*}
|\Psi(x;\omega) - \Psi(x;\omega^\prime)| \sim |\omega - \omega^\prime| \text{ for all } \omega, \omega^\prime \in U_x
\end{equation*}
with implicit constant only depending on $n$.\\
Additionally, if $1 \leq k \leq n-1$, then by twice differentiating the identity
\begin{equation*}
\partial_{x_k} \phi(x;\Psi(x;\omega)) = \omega_k
\end{equation*}
in the $\omega$-variables, it follows that
\begin{equation*}
\| \partial^2_{\omega \omega} \Psi_k(x;\omega) \|_{L^\infty} = O(c_{par}).
\end{equation*}
By the previous estimate, \eqref{eq:HPerturbation} follows from $C2'')$.
\end{proof}
By similar means, we infer estimates for the generalized Gauss map associated with $T^\lambda$. To give the results, let
\begin{equation*}
X^\lambda = \{ x \in \mathbb{R}^n \; | \; \frac{x}{\lambda} \in X \}
\end{equation*}
denote the $\lambda$-dilate of $X$, so that $a^\lambda$ is supported in $X^\lambda \times \Xi$.
\begin{lemma}
\label{lem:ConsequencesReductionGaussMap}
For all $x, \bar{x} \in X^\lambda$ and $\omega, \bar{\omega} \in \Xi$, the estimates
\begin{equation}
\label{eq:EstimatesReducedGaussMap}
\begin{split}
\angle(G^\lambda(x;\omega),G^\lambda(x;\bar{\omega})) &\sim |\frac{\omega}{|\omega|} - \frac{\bar{\omega}}{|\bar{\omega}|}| \sim \angle (\omega, \bar{\omega}), \\
\angle(G^\lambda(x;\omega),G^\lambda(\bar{x};\omega)) &\lesssim \lambda^{-1} |x-\bar{x}|
\end{split}
\end{equation}
hold true.
\end{lemma}
This will be helpful to understand the wave packet analysis in the following sections.
\subsection{Wave packet decomposition}
\label{subsection:WavePacket}
We carry out the wave packet decomposition with respect to some spatial parameter $1 \ll R \ll \lambda$. For this purpose, we follow \cite{GuthHickmanIliopoulou2019} and use that the construction only depends on the non-degeneracy condition $C1)$. We do not use the usual wave packet decomposition for the cone as e.g. in \cite{OuWang2021}, but adapt the parabolic case, as previously done by Lee \cite{Lee2006}. The reason is that in Section \ref{section:ComparingWavePackets} we would sort the smaller cone tubes into larger tubes anyway. It appears that the present choice of wave packet decomposition allows to transfer arguments from \cite{GuthHickmanIliopoulou2019} to the homogeneous setting more directly. In the following we introduce notations from \cite{GuthHickmanIliopoulou2019}.
Cover $A^{n-1}$ by finitely overlapping balls of radius $R^{-1/2}$, and let $\psi_\theta$ be a smooth partition of unity adapted to this cover. These $\theta$ will frequently be referred to as $R^{-1/2}$-balls. For a ball $\theta$, cover $\mathbb{R}^{n-1}$ with finitely overlapping balls of size $R^{\frac{1+\delta}{2}} \times \ldots \times R^{\frac{1+\delta}{2}}$ with center $v \in R^{\frac{1+\delta}{2}} \mathbb{Z}^{n-1}$. Let $\eta_v=\eta(\cdot - v)$ denote a bump function adapted to $B(v,R^{\frac{1+\delta}{2}})$ such that
\begin{equation*}
\sum_{v \in \mathbb{Z}^{n-1}} \eta_v = 1
\end{equation*}
with $\hat{\eta}_v$ essentially supported in $B(0,C R^{-\frac{1+\delta}{2}})$. This is possible by the Poisson summation formula.
Let $\mathbb{T}$ denote the collection of all pairs $(\theta,v)$. Then, for $f:\mathbb{R}^{n-1} \to \mathbb{C}$ with support in $A^{n-1}$ and sufficiently regular, we find
\begin{equation*}
f = \sum_{(\theta,v) \in \mathbb{T}} (\eta_v(\psi_\theta f)\check{\;})\hat{\;} = \sum_{(\theta,v) \in \mathbb{T}} \hat{\eta}_v * (\psi_\theta f).
\end{equation*}
For each $R^{-1/2}$-ball $\theta$, let $\omega_\theta$ denote its centre. Choose a real-valued smooth function $\tilde{\psi}$ so that $\tilde{\psi}_\theta$ is supported in $\theta$, and $\tilde{\psi}_\theta(\omega) = 1$ whenever $\omega$ belongs to a $cR^{-1/2}$-neighbourhood of the support of $\psi_\theta$ for some small $c>0$. Finally, define
\begin{equation*}
f_{\theta,v} = \tilde{\psi}_\theta \cdot [ \hat{\eta}_v * (\psi_\theta f)].
\end{equation*}
The function $\hat{\eta}_v$ is rapidly decaying outside $B(0,CR^{-\frac{1+\delta}{2}})$ and, consequently,
\begin{equation*}
\| f_{\theta,v} - (\hat{\eta}_v * (\psi_\theta f)) \|_{L^\infty(\mathbb{R}^{n-1})} \leq \text{RapDec}(R) \| f \|_{L^2(A^{n-1})}.
\end{equation*}
The functions $f_{\theta,v}$ are almost orthogonal: if $\mathbb{S} \subseteq \mathbb{T}$, then
\begin{equation*}
\Vert \sum_{(\theta,v) \in \mathbb{S}} f_{\theta,v} \|^2_{L^2(\mathbb{R}^{n-1})} \sim \sum_{(\theta,v) \in \mathbb{S}} \| f_{\theta,v} \|^2_{L^2(\mathbb{R}^{n-1})}.
\end{equation*}
Let $T^\lambda$ be an oscillatory integral operator with reduced phase $\phi$ satisfying $C1')$ and amplitude $a$ supported in $X \times \Xi$. For $(\theta,v) \in \mathbb{T}$ define the curve $\gamma^1_{\theta,v}: I^1_{\theta,v} \to \mathbb{R}^{n-1}$ by setting $\gamma^1_{\theta,v}(t)=\Phi(v,t;\omega_\theta)$, where $\Phi$ is the function introduced above and
\begin{equation*}
I^1_{\theta,v} = \{ x_n \in T \, | \, \partial_\omega \phi(x^\prime,x_n;\omega_\theta) = v \text{ for some } x^\prime \in X^\prime \}.
\end{equation*}
Hence, $\partial_\omega \phi(\gamma^1_{\theta,v}(x_n),x_n;\omega_\theta) = v$ for all $x_n \in I^1_{\theta,v}$. For the rescaled curve
\begin{equation*}
\gamma^\lambda_{\theta,v}(t) = \lambda \gamma^1_{\theta,v/\lambda}(t/\lambda),
\end{equation*}
we find
\begin{equation*}
\partial_\omega \phi^\lambda(\gamma^\lambda_{\theta,v}(x_n),x_n;\omega_\theta) = v \text{ for all } t \in I^\lambda_{\theta,v} = \{t \in \mathbb{R} \, : \, \frac{t}{\lambda} \in I^1_{\theta,v} \}.
\end{equation*}
Let $\Gamma^\lambda_{\theta,v}: I^\lambda_{\theta,v} \to \mathbb{R}^n$ denote the graph mapping $\Gamma^\lambda_{\theta,v}(x_n) = (\gamma^\lambda_{\theta,v}(x_n),x_n)$; for the sake of brevity, the image of this mapping is denoted by $\Gamma^\lambda_{\theta,v}$, too.
\begin{lemma}[{\cite[Lemma~5.2]{GuthHickmanIliopoulou2019}}]
The tangent space $T_{\Gamma^\lambda_{\theta,v}(x_n)} \Gamma^\lambda_{\theta,v}$ lies in the direction of the unit vector $G^\lambda(\Gamma^\lambda_{\theta,v}(x_n);\omega_\theta)$ for all $x_n \in I^\lambda_{\theta,v}$.
\end{lemma}
We consider curved tubes
\begin{equation*}
T_{\theta,v} = \{(x^\prime,x_n) \in B(0,R) \, : \, x_n \in I^\lambda_{\theta,v} \text{ and } x^\prime \in B(\gamma^\lambda_{\theta,v}(x_n),R^{\frac{1}{2}+\delta}) \}.
\end{equation*}
We refer to the curve $\Gamma^\lambda_{\theta,v}$ as the core of $T_{\theta,v}$. Since $\phi$ is of reduced form, we find by the diffeomorphism property of $\Phi$ (writing $x' = \Phi^{-1}_{x_n,\omega_\theta} \circ \Phi_{x_n,\omega_\theta} (x')$)
\begin{equation*}
|x^\prime - \gamma^\lambda_{\theta,v}| \sim |\partial_\omega \phi^\lambda(x;\omega_\theta) - v|,
\end{equation*}
for all $x = (x^\prime,x_n) \in X_\lambda$ with $x_n \in I^\lambda_{\theta,v}$ uniformly in $\lambda$. This has the following consequence:
\begin{lemma}[{\cite[Lemma~5.4]{GuthHickmanIliopoulou2019}}]
If $1 \ll R \ll \lambda$ and $x \in B(0,R) \backslash T_{\theta,v}$, then
\begin{equation*}
|T^\lambda f_{\theta,v}(x)| \leq (1+ R^{-1/2} |\partial_\omega \phi^\lambda(x;\omega_\theta) -v |)^{-(n+1)} \text{RapDec}(R) \| f \|_{L^2(A^{n-1})}.
\end{equation*}
\end{lemma}
\subsection{$L^2$-$L^2$-estimate}
We recall the following generalization of Parseval's theorem, only depending on non-degeneracy $C1')$ of the phase function (cf. \cite[Section~2.1]{Sogge2017}):
\begin{lemma}[{\cite[Lemma~5.5]{GuthHickmanIliopoulou2019}}]
If $1 \leq R \leq \lambda$ and $B_R$ is any ball of radius $R$, then
\begin{equation}
\label{eq:L2Estimate}
\| T^\lambda f \|_{L^2(B_R)} \lesssim R^{1/2} \| f \|_{L^2(A^{n-1})}.
\end{equation}
\end{lemma}
This follows from the following estimate:
\begin{lemma}[{\cite[Lemma~5.6]{GuthHickmanIliopoulou2019}}]
For any fixed $x_n \in \mathbb{R}$, we find the estimate
\begin{equation}
\label{eq:FixedTimeL2Estimate}
\| T^\lambda f \|_{L^2(\mathbb{R}^{n-1} \times \{ x_n \})} \lesssim \| f \|_{L^2(A^{n-1})} .
\end{equation}
\end{lemma}
\subsection{$k$-broad norms}
\label{subsection:kBroadNorms}
Here we recall basic properties of the $k$-broad norms. Although the naming is misleading as $k$-broad norms are strictly speaking no norms, the properties are similar enough to make the following arguments work. We shall also see that $U \mapsto \| T^\lambda f \|^p_{BL^p_{k,A}(U)}$ behaves as a measure.
\begin{lemma}[Finite (sub-)additivity, {\cite[Lemma~6.1]{GuthHickmanIliopoulou2019}}]
Let $U_1,U_2 \subseteq \mathbb{R}^n$ and $U = U_1 \cup U_2$. If $1 \leq p < \infty$ and $A$ is a non-negative integer, then
\begin{equation}
\label{eq:SubadditivityBroadNorm}
\| T^\lambda f \|^p_{BL^p_{k,A}(U)} \leq \| T^\lambda f \|^p_{BL^p_{k,A}(U_1)} + \| T^\lambda f \|^p_{BL^p_{k,A}(U_2)}
\end{equation}
holds for all integrable $f: A^{n-1} \to \mathbb{C}$.
\end{lemma}
Secondly, we have the following variant of the triangle inequality:
\begin{lemma}[Triangle inequality, {\cite[Lemma~6.2]{GuthHickmanIliopoulou2019}}]
If $U \subseteq \mathbb{R}^n$, $1 \leq p < \infty$ and $A = A_1 + A_2$ for $A_1$ and $A_2$ non-negative integers, then
\begin{equation}
\| T^\lambda(f_1 + f_2) \|_{BL^p_{k,A}(U)} \lesssim \| T^\lambda f_1 \|_{BL^p_{k,A_1}(U)} + \| T^\lambda f_2 \|_{BL^p_{k,A_2}(U)}
\end{equation}
holds for all integrable $f_1,f_2: A^{n-1} \to \mathbb{C}$.
\end{lemma}
We further have the following variant of H\"older's inequality:
\begin{lemma}[Logarithmic convexity, {\cite[Lemma~6.3]{GuthHickmanIliopoulou2019}}]
Suppose that $U \subseteq \mathbb{R}^n$, $1 \leq p,p_1,p_2 < \infty$ and $0 \leq \alpha_1,\alpha_2 \leq 1$ satisfy $\alpha_1 + \alpha_2 = 1$ and
\begin{equation*}
\frac{1}{p} = \frac{\alpha_1}{p_1} + \frac{\alpha_2}{p_2}.
\end{equation*}
If $A = A_1 + A_2$ for $A_1$, $A_2$ non-negative integers, then
\begin{equation*}
\| T^\lambda f \|_{BL^p_{k,A}(U)} \lesssim \| T^\lambda f \|^{\alpha_1}_{BL^{p_1}_{k,A_1}(U)} \| T^\lambda f \|^{\alpha_2}_{BL^{p_2}_{k,A_2}(U)}.
\end{equation*}
\end{lemma}
Later on, we shall only consider $A \gg 1$, which will allow for the use of H\"older's and Minkowski's inequality for $k$-broad norms.
\section{Polynomial partitioning}
\label{section:PolynomialPartitioning}
A key tool in the proof will be polynomial partitioning following previous work by Guth \cite{Guth2016,Guth2018} (see also Guth--Katz \cite{GuthKatz2015}) and in the variable coefficient case Guth--Hickman--Iliopoulou \cite{GuthHickmanIliopoulou2019}. The idea is to divide the ball $B_R$ through the zero set of a polynomial into cells, which equidistribute the broad norm. Either $\mu_{T^\lambda f}$ will be concentrated in the cells or at the wall, i.e., neighbourhood of the zero locus of the polynomial. Both cases will be handled by induction. We recall some facts from \cite{GuthHickmanIliopoulou2019}, which we will rely on in the following.
\subsection{Tools from algebraic geometry}
Given a polynomial $P$ in $\mathbb{R}^n$, its zero set is denoted by $Z(P)$. To make the varieties $Z(P_1,\ldots,P_{n-m})$ smooth $m$-dimensional manifolds, we consider transverse complete intersections:
\begin{definition}
Let $m \in \mathbb{N}$, $ m \leq n$, and let $P_1,\ldots,P_{n-m}$ be polynomials on $\mathbb{R}^n$ whose common zero set is denoted by $Z(P_1,\ldots,P_{n-m})$. The variety $Z(P_1,\ldots,P_{n-m})$ is called a transverse complete intersection if
\begin{equation*}
\nabla P_1(x) \wedge \ldots \wedge \nabla P_{n-m}(x) \neq 0 \qquad \forall x \in Z(P_1,\ldots,P_{n-m}).
\end{equation*}
The degree of the transverse complete intersection $\overline{\deg} Z$ is defined as \\
$\max_{j=1,\ldots,n-m} \deg P_j$.
\end{definition}
We have the following partitioning argument:
\begin{theorem}[{\cite[Theorem~7.3]{GuthHickmanIliopoulou2019}}]
\label{thm:EquidistributionPolynomialPartitioning}
Suppose that $W \geq 0$ is a non-zero $L^1$-function on $\mathbb{R}^n$. Then, for any degree $D \in \mathbb{N}$, there exists a non-zero polynomial $P$ of degree $\deg P \lesssim D$ such that the following holds:
\begin{itemize}
\item The set $Z(P)$ is a finite union of $\log D$ transverse complete intersection.
\item If $(O_i)_{i \in \mathcal{I}}$ denotes the set of connected components of $\mathbb{R}^n \backslash Z(P)$, then $\# \mathcal{I} \lesssim D^n$ and
\begin{equation}
\label{eq:EquidistributionPartitioning}
\int_{O_i} W \sim D^{-n} \int_{\mathbb{R}^n} W \text{ for all } i \in \mathcal{I}.
\end{equation}
\end{itemize}
\end{theorem}
The connected components are called \emph{cells}.
We further need the following lemma on transverse intersections of tubes with varieties:
\begin{lemma}[{\cite[Lemma~5.7]{Guth2018}}]
\label{lem:TransverseIntersectionsStraightLines}
Let $T$ be a cylinder of radius $r$ with central line $\ell$ and suppose that $Z=Z(P_1,\ldots,P_{n-m}) \subseteq \mathbb{R}^n$ is a transverse complete intersection, where the polynomials $P_j$ have degree at most $D$. For $\alpha > 0$, let
\begin{equation*}
Z_{> \alpha} = \{ z \in Z: \text{Angle}(T_z Z, \ell) > \alpha \}.
\end{equation*}
Then $Z_{> \alpha} \cap T$ is contained in a union of $\lesssim D^n$ balls of radius $\lesssim r \alpha^{-1}$.
\end{lemma}
For the application, we are interested in $r = R^{(1+\delta)/2}$, as this will be the radius of the (thin) tubes and $\alpha = R^{-\frac{1}{2}+\delta}$.
\subsection{Polynomial approximation}
\label{subsection:PolynomialApproximation}
However, with smooth core curves, Lemma \ref{lem:TransverseIntersectionsStraightLines} is not applicable directly. We approximate the core curves by polynomials such that algebraic methods can still be applied to the curved tubes. We follow \cite[Section~7.2]{GuthHickmanIliopoulou2019}. Let $\varepsilon > 0$ be a small parameter and let $N=N_\varepsilon:= \lceil 1/(2 \varepsilon) \rceil \in \mathbb{N}$. Suppose that $\Gamma:(-1,1) \to \mathbb{R}^n$ is a smooth curve with
\begin{equation*}
\| \Gamma \|_{C^{N+1}(-1,1)} = \max_{0 \leq k \leq N+1} \sup_{|t| < 1} | \Gamma^{(k)}(t)| \lesssim 1.
\end{equation*}
After reductions of Section \ref{subsection:BasicReductions}, we find the following estimates:
\begin{lemma}
\label{lem:PolynomialApproximation}
The curves $\Gamma^1_{\theta,v}$ satisfy
\begin{equation*}
|(\Gamma^1_{\theta,v})'(t)| \sim 1 \text{ for all } t \in I^1_{\theta,v},
\end{equation*}
and
\begin{equation*}
\sup_{t \in I^1_{\theta,v}} | ( \Gamma^1_{\theta,v})^{(k)}(t)| \lesssim c_{par} \text{ for } 2 \leq k \leq N.
\end{equation*}
\end{lemma}
The proof from \cite[Lemma~7.4]{GuthHickmanIliopoulou2019} applies verbatim although the phase functions are from different classes, but because of bounds \eqref{eq:XiXnDerivativeBound} and $D2)$ from Subsection \ref{subsection:BasicReductions}.
We denote by $[\Gamma]_\varepsilon: \mathbb{R} \to \mathbb{R}^n$ the polynomial curve given by the degree-$N$ Taylor approximation of $\Gamma$ around zero. Observe that
\begin{equation*}
\| [ \Gamma ]_\varepsilon \|_{C^\infty(-2,2)} \leq e^2 \| \Gamma \|_{C^N(-1,1)} \lesssim 1.
\end{equation*}
Furthermore, for $\lambda \gg 1$, noting that $\lambda^{-\varepsilon N} \leq \lambda^{-1/2}$, Taylor's theorem yields
\begin{equation*}
| \Gamma^{(i)}(t) - [ \Gamma ]^{(i)}_\varepsilon(t) | \lesssim_\varepsilon \lambda^{-\frac{1}{2}} |t|^{1-i} \text{ for all } |t| \lesssim_\varepsilon \lambda^{-\varepsilon} \text{ and } i = 0,1.
\end{equation*}
Letting $\Gamma^\lambda:(-\lambda,\lambda) \to \mathbb{R}^n$ denote the rescaled curve $\Gamma^\lambda(t) = \lambda \Gamma(t/\lambda)$, the above inequalities imply that
\begin{equation}
\label{eq:RescaledBoundsPolynomialApproximation}
\| [ \Gamma^\lambda]'_\varepsilon \|_{C^\infty(-2\lambda,2\lambda)} \lesssim 1 \text{ and } \| [ \Gamma^\lambda]''_\varepsilon \|_{C^\infty(-2\lambda,2\lambda)} \lesssim \lambda^{-1},
\end{equation}
and
\begin{equation*}
| (\Gamma^\lambda)^{(i)}(t) - ([ \Gamma^\lambda]_\varepsilon )^{(i)}(t) | \lesssim_\varepsilon \lambda^{-\frac{1}{2}} |t|^{1-i} \text{ for all } |t| \lesssim_\varepsilon \lambda^{1-\varepsilon} \text{ and } i=0,1.
\end{equation*}
As a consequence of $|(\Gamma^\lambda)'(t)| \sim |[\Gamma^\lambda]'_\varepsilon(t) | \sim 1$, the tangent spaces to the curves $\Gamma^\lambda$ and $[\Gamma^\lambda]_\varepsilon$ have a small angular separation, i.e.,
\begin{equation}
\label{eq:TangentSpaceApproximant}
\angle (T_{\Gamma^\lambda(t)} \Gamma^\lambda, T_{[\Gamma^\lambda]_\varepsilon(t)} [ \Gamma^\lambda]_\varepsilon ) \lesssim_\varepsilon \lambda^{-\frac{1}{2}} \text{ for all } |t| \lesssim_\varepsilon \lambda^{1-\varepsilon}.
\end{equation}
\subsection{Transverse interactions between curved tubes and varieties}
Next, we generalize the transverse interaction of straight lines and varieties as in Lemma \ref{lem:TransverseIntersectionsStraightLines} to curved tubes, which are approximated by polynomials.
Let $Z=Z(P_1,\ldots,P_{n-m})$ be a transverse complete intersection and $\Gamma: \mathbb{R} \to \mathbb{R}^n$ be a polynomial curve. Given $\alpha,r>0$, the problem is to estimate the size of the set
\begin{equation*}
Z_{>\alpha,r,\Gamma} := \{ z \in Z : \text{there exists } x \in \Gamma \text{ with } |x-z|<r \text{ and } \angle(T_z Z, T_x \Gamma) > \alpha \}.
\end{equation*}
We further assume that $\Gamma$ is a \emph{polynomial graph}, which means it can be rotated so that it is given by $\Gamma(t) = (\gamma(t),t)$ for some polynomial mapping $\gamma: \mathbb{R} \to \mathbb{R}^{n-1}$. This is the case considered in the present context.
We have the following generalization of Lemma \ref{lem:TransverseIntersectionsStraightLines}:
\begin{lemma}[{\cite[Lemma~7.5]{GuthHickmanIliopoulou2019}}]
\label{lem:TransverseIntersectionPolynomialApproximant}
Let $n \geq 2$, $1 \leq m \leq n$ and $Z=Z(P_1,\ldots,P_{n-m}) \subseteq \mathbb{R}^n$ be a transverse complete intersection. Suppose that $\Gamma:\mathbb{R} \to \mathbb{R}^n$ is a polynomial graph satisfying
\begin{equation}
\label{eq:DerivativeConditionsPolynomialGraph}
\| \Gamma' \|_{L^\infty(-2\lambda,2\lambda)} \lesssim 1 \text{ and } \| \Gamma'' \|_{L^\infty(-2\lambda,2\lambda)} \leq \delta
\end{equation}
for some $\lambda,\delta>0$. There exists a dimensional constant $\bar{C} > 0$ such that, for all $\alpha > 0$ and $0<r<\lambda$ satisfying $\alpha \geq \bar{C} \delta r$, the set $Z_{>\alpha,r,\Gamma} \cap B(0,\lambda)$ is contained in a union of
\begin{equation*}
O((\overline{\deg} Z \cdot \deg \Gamma)^n)
\end{equation*}
balls of radius $r/\alpha$.
\end{lemma}
In the present context, we consider the polynomial approximant $\Gamma = [\Gamma^\lambda_{\theta,v}]_\varepsilon$ of the curve $\Gamma^\lambda_{\theta,v}$ as defined in Subsection \ref{subsection:PolynomialApproximation}. Then, $\deg \Gamma \lesssim_\varepsilon 1$ and by \eqref{eq:RescaledBoundsPolynomialApproximation} we find \eqref{eq:DerivativeConditionsPolynomialGraph} to hold with $\delta \sim_\varepsilon 1/\lambda$. Consequently, for $\alpha > 0$ and $0<r<\lambda$ with $\alpha \gtrsim r/\lambda$, the set $Z_{>\alpha,r,\Gamma} \cap B(0,\lambda)$ can be covered by $O_\varepsilon((\overline{\deg} Z)^n)$ balls of radius $r/\alpha$.
\section{Transverse equidistribution estimates}
\label{section:TransverseEquidistribution}
\subsection{Linearizing the phase function}
\label{subsection:Linearizing}
In this section transverse equidistribution estimates for wave packets tangential to varieties will be examined. This is a key point in the main induction argument. Contrary to \cite{OuWang2021} or \cite{GaoLiuMiaoXi2021}, however, we stick to the wave packet decomposition used in \cite{GuthHickmanIliopoulou2019}.
\begin{definition}
Let $Z = Z(P_1,\ldots,P_{n-m})$ be a transverse complete intersection. A wave packet $(\theta,v)$ is said to be $R^{-\frac{1}{2}+\delta_m}$-tangent to $Z$ in $B(0,R)$ if
\begin{equation}
\label{eq:TangentialConditionI}
T_{\theta,v} \cap B_R \subseteq N_{R^{\frac{1}{2} + \delta_m}}(Z)
\end{equation}
and
\begin{equation}
\label{eq:TangentialConditionII}
\angle(G^\lambda(x;\omega_\theta), T_z Z) \leq \bar{c}_{tang} R^{-\frac{1}{2}+\delta_m}
\end{equation}
for any $x \in T_{\theta,v}$ and $z \in Z \cap B(0,2R)$ with $|x-z| \leq \bar{C}_{tang} R^{\frac{1}{2}+\delta_m}$.
\end{definition}
We want to study functions concentrated on the collection of wave packets
\begin{equation*}
\mathbb{T}_Z = \{ (\theta,v) \in \mathbb{T} : T_{\theta,v} \text{ in } R^{-\frac{1}{2}+\delta_m}-\text{tangent to } Z \text{ in } B(0,R)\}.
\end{equation*}
Precisely, we make the following definition:
\begin{definition}
If $\mathbb{S} \subseteq \mathbb{T}$, then $f$ is said to be concentrated on wave packets from $\mathbb{S}$ if
\begin{equation*}
f = \sum_{(\theta,v) \in \mathbb{S}} f_{\theta,v} + \text{RapDec}(R) \| f \|_{L^2}.
\end{equation*}
\end{definition}
Let $B \subseteq \mathbb{R}^n$ be a ball of radius $R^{\frac{1}{2}+\delta_m}$ with centre $\bar{x} \in B(0,R)$. We study $\eta_B \cdot T^\lambda g$, where $\eta_B$ is a suitable choice of Schwartz function adapted to $B$. A stationary phase argument yields that $\eta_B \cdot T^\lambda g_{\theta,v}$ is concentrated near the surface
$\Sigma = \{ \Sigma(\omega) \, : \omega \in A^{n-1} \}$, where $\Sigma(\omega) = \partial_x \phi^\lambda(\bar{x};\omega)$. This leads to the refined set of wave packets
\begin{equation*}
\mathbb{T}_{Z,B} = \{(\theta,v) \in \mathbb{T}_Z : T_{\theta,v} \cap B \neq \emptyset \}.
\end{equation*}
For $(\theta,v) \in \mathbb{T}_{Z,B}$, the direction $G^\lambda(\bar{x};\omega_\theta)$ of $T_{\theta,v}$ on the ball $B$ must make a small angle with each of the tangent spaces $T_z Z$ for all $z \in Z \cap B$. This constrains $\Sigma(\omega_\theta)$ to lie in a small neighbourhood of some typically $m$-dimensional manifold $S_\xi$. But in the homogeneous case, $S_\xi$ might only be one-dimensional, or ``close" to a one-dimensional manifold. This will be quantified below. This case does not contribute essentially in the broad norm. To linearize $S_\xi$, if it is not a ``thin", essentially one-dimensional set, let $R^{\frac{1}{2}} < \rho \ll R$ and for the remainder of this section, let $\tau \subseteq A^{n-1}$ denote a sector of radius $O(\rho^{-\frac{1}{2}+\delta_m})$. We define
\begin{equation*}
\mathbb{T}_{Z,B,\tau} = \{ (\theta,v) \in \mathbb{T}_Z : \theta \cap \tau \neq \emptyset \text{ and } T_{\theta,v} \cap B \neq \emptyset \}.
\end{equation*}
\medskip
We recall the constant-coefficient examples. Suppose that $Z$ is an $m$-dimensional affine plane so that $T_z Z = V$ for all $z \in Z$, where $V \parallel Z$. The extension operator for the paraboloid has the unnormalized Gauss map $G_0(\omega) = (-\omega,1)$. Consequently,
\begin{equation*}
A_\omega = \{ \omega \in \mathbb{R}^{n-1} : G_0(\omega) \in V \}
\end{equation*}
is an affine subspace of $\mathbb{R}^{n-1}$ of dimension $m-1$. If $G_0(\omega) \in V$, then $\Sigma(\omega) \in A_\xi = A_u \times \mathbb{R}$. Due to the frequency localization and the uncertainty principle, $\eta_B \cdot T^\lambda g$ will decay little in directions transverse to $A_\xi$. This was exploited in \cite{Guth2018}.\\
The situation for the cone is a little different. Here, $G_0(\omega) = \big( - \frac{\omega}{|\omega|}, 1 \big)$ is not an affine map. Let $V^+ = \{ \omega : G_0(\omega) \in V \}$. By a crucial observation due to Ou--Wang \cite{OuWang2021}, if $V^+$ is tangent to $\mathcal{C} = \{ (\omega, \frac{\omega}{|\omega|}) : \omega \in A^{n-1} \}$ up to an angle $R^{-\delta_m}$, then $N_{R^{-\frac{1}{2}+\delta_m}} V^+ \cap \mathcal{C}$ is a $O(R^{-\delta_m})$-neighbourhood of $O(1)$ radial lines.
In the variable coefficient case, we see that if $V^+$ is tangent to $\mathcal{C}_x = \{ \partial_x \phi^\lambda(\bar{x};\omega) \}$
up to an angle $R^{-\delta_m}$, then $N_{R^{-\frac{1}{2}+\delta_m}} V^+ \cap C$ is a $O(R^{-\delta_m})$-neighbourhood of $O(1)$ radial lines. A contribution like this can be neglected in the $K$-broad norm. Otherwise, we shall see that we find quantitative transversality to hold and can deduce transverse equidistribution estimates similar to the paraboloid case (or its variable coefficient counterpart). In the constant coefficient case, but for arbitrary cones, this was recently investigated in \cite{GaoLiuMiaoXi2021}. We shall see how the arguments adapt to the variable coefficient case.
Consider an $m$-dimensional linear subspace $V= \{ \sum_{j=1}^n a_{i,j} x_j =0, \; i =1,\ldots,n-m \}$ and let $V^- = \{ \sum_{j=1}^{n-1} a_{i,j} x_j = 0 \}$. We change to $u$-frequencies via $\Psi^\lambda$, which recall is defined by
\begin{equation*}
\partial_{x'} \phi^\lambda(\bar{x};\Psi^\lambda(\bar{x};u)) = u.
\end{equation*}
We use short-hand notation $\Psi(u) := \Psi^\lambda(\bar{x};u)$. It is easy to see that $\Psi$ like $\partial_{x'} \phi^\lambda$ and the identity mapping is $1$-homogeneous because
\begin{equation*}
\partial_{x'} \phi^\lambda(\bar{x};\Psi(\mu u)) = \mu u = \mu \partial_{x'} \phi^\lambda(\bar{x};\Psi(u)) = \partial_{x'} \phi^\lambda(\bar{x};\mu \Psi(u)).
\end{equation*}
By substituting $\tilde{\phi}(u) = h_{\bar{x}}(u) = \partial_{x_n} \phi^\lambda(\bar{x};\Psi(u))$ the arguments from \cite{GaoLiuMiaoXi2021} apply. We define a set
\begin{equation*}
L = \{ u \in A^{n-1}: \sum_{j=1}^{n-1} a_{i,j} \partial_{j} \tilde{\phi}(u) - a_{i,n} = 0; \quad i =1,\ldots,n-m \}.
\end{equation*}
The set $\{ (u,\tilde{\phi}(u)) : u \in L \}$ describes the points on the generalized cone, which have a normal in $V$. The \emph{tangential case} gives a negligible contribution to the broad norm:
\begin{lemma}[{\cite[Lemma~4.5]{GaoLiuMiaoXi2021}}]
\label{lem:NegligibleContributionBall}
Let $\eta \in \mathbb{S}^{n-2} \subseteq \mathbb{R}^{n-1}$. If $\eta \in L$ and $\angle(\eta,V^-) > \frac{\pi}{2} - K^{-2}$, then $L$ is contained in the set $\{ \xi \in \mathbb{R}^{n-1} \backslash 0: \angle(\xi,\eta) \lesssim K^{-2} \}$.
\end{lemma}
It is important to note that, contrary to the transverse case analyzed below, the lemma does not hinge on a stronger localization of $\eta$. For later purposes, note that for the suitably defined $k$-broad norm balls $B(\bar{x};R^{\frac{1}{2}+\delta_m})$, for which Lemma \ref{lem:NegligibleContributionBall} applies, make a negligible contribution.
We turn to the more involved \emph{transverse case}: In general $\{(u,\tilde{\phi}(u)) : u \in L\}$ may not lie in an affine subspace because $L$ may have curvature. We start by linearizing $L$. By taking the orthogonal complement of a suitable extension of the tangent space we shall construct $W$, which is quantitatively transverse to $V$. Let $\eta \in L \cap \mathbb{S}^{n-2}$ with $\text{Ang}(\eta,V^-) < \frac{\pi}{2} - K^{-2}$. Define $\tilde{V}$ to be the $n-m$-dimensional linear subspace spanned by $\gamma_1,\ldots,\gamma_{n-m}$ given by
\begin{equation*}
\gamma_i = \partial^2 \tilde{\phi} (\eta) \alpha_i, \quad \alpha_i = (a_{i,1},\ldots,a_{i,n}), \quad i=1,\ldots,n-m.
\end{equation*}
The angle condition $\angle(\eta,V^-) < \frac{\pi}{2} - K^{-2}$ ensures that $\gamma_i$, $i=1,\ldots,n-m$ are linearly independent. Indeed, the Hessian is degenerate in the direction of $\eta$, but $\alpha_i$ is orthogonal to $V^-$. Let $\bar{V}^-$ be the orthogonal complement of $\tilde{V}$ in $\mathbb{R}^{n-1}$, i.e.,
\begin{equation*}
\mathbb{R}^{n-1} = \tilde{V} \oplus \bar{V}^-.
\end{equation*}
Note that $\bar{V}^-$ denotes the tangent space of $L$: Starting from the equations
\begin{equation*}
\sum_{j=1}^{n-1} a_{i,j} \partial_{j} \tilde{\phi}(\eta + \xi') - a_{i,n} = 0,
\end{equation*}
linearizing yields for $\xi' \in T_\eta L$
\begin{equation*}
\sum_{j=1}^{n-1} a_{i,j} \sum_{k=1}^{n-1} \partial^2_{j k} \tilde{\phi}(\eta) \xi_k' = (\partial^2 \tilde{\phi}(\eta) \alpha_i, \xi') = 0.
\end{equation*}
Let $\bar{V}$ be the linear subspace spanned by $\bar{V}^-$ and $e_{n}$. Define $W$ to be the orthogonal complement of $\bar{V}$ in $\mathbb{R}^{n}$, i.e.,
\begin{equation*}
\mathbb{R}^{n} = \bar{V} \oplus W.
\end{equation*}
As pointed out in \cite{GaoLiuMiaoXi2021}, all the linear subspaces depend on the choice of $\eta$. We have the following quantitative transversality:
\begin{lemma}[{\cite[Lemma~4.6]{GaoLiuMiaoXi2021}}]
\label{lem:QuantitativeTransverse}
Let $\eta \in \mathbb{S}^{n-2} \cap L$. If $\angle(\eta,V^-) \leq \frac{\pi}{2} - K^{-2}$, then $W$ and $V$ are transverse in the sense that $\angle(V,W) \gtrsim K^{-4}$.
\end{lemma}
\subsection{Verifying the transverse equidistribution estimate}
We turn to the key equidistribution estimate. In the following let $\tau \subseteq A^{n-1}$ denote a sector of aperture $O(\rho^{-\frac{1}{2}+\delta_m})$ and $B$ a ball of radius $R^{\frac{1}{2}+\delta_m}$. Moreover, we suppose that $Z$ is $K$-flat, i.e., for any $z,z' \in Z \cap 2B$ we have
\begin{equation*}
\angle(T_z Z, T_{z'} Z) \lesssim K^{-5}
\end{equation*}
with $K \lesssim R^\delta \ll \rho^{\delta_m}$.
\begin{lemma}
\label{lem:TransverseEquidistributionEstimate}
Let $Z$ be a $K$-flat, transverse complete intersection with $\dim Z = m$, $\overline{\text{deg}} \; Z \lesssim_\varepsilon 1$, $B=B(\bar{x},R^{\frac{1}{2}+\delta_m})$ a ball of radius $R^{\frac{1}{2}+\delta_m}$, and let $g$ be concentrated on wave packets in $\mathbb{T}_{Z,B,\tau}$. Suppose that with the notations of Subsection \ref{subsection:Linearizing}, with $\tilde{\phi} = h_{\bar{x}}$, and for some $\eta \in \Psi^{-1}(\tau) \cap \mathbb{S}^{n-2}$ we are in the situation of Lemma \ref{lem:QuantitativeTransverse}. Then, for any $\rho \leq R$,
\begin{equation}
\label{eq:TransverseEquidistributionEstimateI}
\int_{B \cap N_{\rho^{\frac{1}{2}+\delta_m}}(Z)} |T^\lambda g|^2 \lesssim R^{\frac{1}{2}+O(\delta_m)} \big( \frac{\rho}{R} \big)^{\frac{n-m}{2}} \| g \|^2_{L^2}.
\end{equation}
\end{lemma}
For the proof, we recall the following quantifications of the uncertainty principle from \cite[Subsection~8.2]{GuthHickmanIliopoulou2019}: Let $G: \mathbb{R}^n \to \mathbb{C}$ be frequency supported on a ball of radius $r >0$. Then we have the moral estimate due to lack of $L^2$-norm concentration:
\begin{equation*}
\Xint-_{B(x_0,\rho)} |G|^2 \lesssim \Xint-_{B(x_0,r^{-1})} |G|^2.
\end{equation*}
We use the below manifestation with $\hat{G}$ essentially supported in a ball of radius $r>0$.
\begin{lemma}
\label{lem:UncertaintyQuantification}
If $r^{-\frac{1}{2}} \leq \rho \leq r^{-1}$, then for any $B(x_0,\rho)$, $\xi_0 \in \mathbb{R}^n$ and $\rho >0 $ one has
\begin{equation*}
\Xint-_{B(x_0,\rho)} |G|^2 \lesssim_\delta \| \hat{G} w^{-1}_{B(\xi_0,r)} \|_\infty^{\frac{2\delta}{1+\delta}} \frac{1}{|B(r)|^{-1}} \big( \int_{\mathbb{R}^n} |G|^2 \big)^{\frac{1}{1+\delta}}.
\end{equation*}
Above, $w_{B(\xi_0,r)}$ is a weight concentrated on $B(\xi_0,r)$ given by
\begin{equation*}
w_{B(\xi_0,r)}(\xi) = (1+r^{-1} |\xi-\xi_0|)^{-N} \text{ for some large } N = N_\delta \in \mathbb{N}.
\end{equation*}
\end{lemma}
As first step in the proof of Lemma \ref{lem:TransverseEquidistributionEstimate}, we consider wave packets tangential to linear subspaces: In the following transverse equidistribution estimates are considered with respect to some fixed linear subspace $V \subseteq \mathbb{R}^n$. Let $B$ be a ball of radius $R^{\frac{1}{2}+\delta_m}$ with centre $\bar{x} \in \mathbb{R}^n$, and define
\begin{equation*}
\mathbb{T}_{V,B} = \{(\theta,v) : \angle(G^\lambda(\bar{x};\omega_\theta),V) \lesssim R^{-\frac{1}{2}+\delta_m} \text{ and } T_{\theta,v} \cap B \neq \emptyset \}.
\end{equation*}
Let $R^{\frac{1}{2}} < \rho < R$ and, for $\tau \subseteq \mathbb{R}^{n-1}$ be a sector of aperture $O(\rho^{-\frac{1}{2}+\delta_m})$ centred around a point in $A^{n-1}$, define
\begin{equation*}
\mathbb{T}_{V,B,\tau} = \{ (\theta,v) \in \mathbb{T}_{V,B} : \theta \cap \big( \frac{\tau}{10} \big) \neq \emptyset \}.
\end{equation*}
\begin{lemma}
\label{lem:TransverseEquidistributionEstimateLinearSubspace}
If $V \subseteq \mathbb{R}^n$ is a linear subspace, then there exists a linear subspace $W$ with the following properties:
\begin{itemize}
\item[(1)] $\dim V + \dim W = n$;
\item[(2)] $V$ and $W$ are quantitatively transverse with $\angle(v,w) \gtrsim K^{-4}$ for any $v \in V$, $w \in W$, $v,w \neq 0$;
\item[(3)] if $g$ is concentrated on wave packets from $\mathbb{T}_{V,B,\tau}$ and there is $\eta \in \Psi^{-1}(\tau)$ such that for $\tilde{\phi} = h_{\bar{x}}$ the assumptions of Lemma \ref{lem:QuantitativeTransverse} are valid, $\Pi$ is any plane parallel to $W$ and $x_0 \in \Pi \cap B$, then the inequality
\begin{equation*}
\int_{\Pi \cap B(x_0,\rho^{\frac{1}{2}+\delta_m})} |T^\lambda g|^2 \lesssim_\delta R^{O(\delta_m)} \big( \frac{\rho}{R} \big)^{\frac{\dim W}{2}} \| g \|^{2\delta / (1+\delta)}_{L^2} \big( \int_{\Pi \cap 2B} |T^\lambda g|^2 \big)^{\frac{1}{1+\delta}}
\end{equation*}
holds, up to inclusion of $\text{RapDec}(R) \| g \|_{L^2}$ on the right-hand side.
\end{itemize}
\end{lemma}
\begin{proof}
\textbf{Constructing the subspace $W$:} We choose $W$ after linearizing as in Lemma \ref{lem:QuantitativeTransverse}. Recall that $h_{\bar{x}}(u) = \partial_{x_n} \phi^\lambda(\bar{x};\Psi(u))$ with $\partial_{x'} \phi^\lambda(\bar{x};\Psi(u)) = u$ such that $(u,h_{\bar{x}}(u))$ is a graph parametrization of $\partial_{x} \phi^\lambda(\bar{x}; \cdot)$ in $u$-frequencies with $\omega = \Psi(u)$. Fix some $\eta \in L \cap \Psi^{-1}(\tau)$ and construct $W$ as in Lemma \ref{lem:QuantitativeTransverse}. Note that
\begin{equation*}
\partial^2_{x u_1} \phi^\lambda(\bar{x};\Psi(u)) \wedge \ldots \wedge \partial^2_{x u_{n-1}} \phi^\lambda(\bar{x};\Psi(u)) = G_0(\bar{x};\omega) \cdot \det J \Psi(u).
\end{equation*}
If $L \cap \Psi^{-1}(\tau) = \emptyset$, then $\Psi(L) \cap \tau = \emptyset$, but then, by Lemma \ref{lem:ConsequencesReductionGaussMap} we had $\mathbb{T}_{V,B,\tau} = \emptyset$ and there is nothing to show. Hence, we can construct $W$ around $\eta \in L \cap \Psi^{-1}(\tau)$ as in Subsection \ref{subsection:Linearizing}. $W$ and $V$ are quantitatively transverse as in (2) by Lemma \ref{lem:QuantitativeTransverse}.
\medskip
\textbf{Verifying the transverse equidistribution estimate:} Recall that $g$ is concentrated on wave packets $\mathbb{T}_{V,B,\tau}$, $B$ is a $R^{\frac{1}{2}+\delta_m}$-ball, and $\tau$ is a $O(\rho^{-\frac{1}{2}+\delta_m})$-sector. $W$ is constructed as above. Let $\eta_B(x) = \eta((x-\bar{x})/R^{\frac{1}{2}+\delta_m})$ denote a Schwartz cutoff, which satisfies $\eta(x) = 1$ for $x \in B(0,2)$. We have
\begin{equation}
\label{eq:FourierTransformLocalization}
(\eta_B T^\lambda g_{\theta,v} \vert_{\Pi} ) \widehat (\xi ) = e^{-2 \pi i x_0.\xi} R^{\dim W \big( \frac{1}{2} + \delta_m \big)} \int_{A^{n-1}} K^{\lambda,R}(\xi;\omega) g_{\theta,v}(\omega) d \omega
\end{equation}
with $K^{\lambda,R}$ given by
\begin{equation*}
K^{\lambda,R}(\xi;\omega) = \int_{W} e^{2 \pi i \phi^{\lambda,R}_\omega(z)} a_\omega^{\lambda,R}(z) dz
\end{equation*}
for the phase and amplitude function
\begin{align*}
\phi_\omega^{\lambda,R}(z) &= \phi^\lambda(x_0+R^{\frac{1}{2}+\delta_m} z; \omega) - R^{\frac{1}{2}+\delta_m} \langle z;\xi \rangle, \\
a_\omega^{\lambda,R}(z) &= a^\lambda(x_0+R^{\frac{1}{2}+\delta_m} z; \omega) \tilde{\eta}(z),
\end{align*}
and
\begin{equation*}
\tilde{\eta}(z) = \eta(z + \frac{x_0-\bar{x}}{R^{\frac{1}{2}+\delta_m}}).
\end{equation*}
Let $\Sigma(\omega) = \partial_x \phi^\lambda(\bar{x};\omega)$.
Fixing $\omega \in \Xi$, $\xi \in \hat{\mathbb{R}}^n$ such that $|\xi- \text{proj}_{W} \Sigma(\omega)| \gtrsim R^{-\frac{1}{2}+\delta_m}$ and $R \gg 1$, the following estimates hold on $\text{supp}(a_z^{\lambda,R})$:
\begin{itemize}
\item[(i)] $|\partial_z \phi^{\lambda,R}_\omega(z)| \sim R^{\frac{1}{2}+\delta_m} | \xi - \text{proj}_{W} \Sigma(\omega)| \gtrsim R^{2\delta_m}$,
\item[(ii)] $|\partial_z^\alpha \phi^{\lambda,R}_\omega(z)| \lesssim |\partial_z \phi^{\lambda,R}_\omega|$ for $2\leq |\alpha| \leq N_{par}$,
\item[(iii)] $|\partial_z^\alpha a^{\lambda,R}_\omega| \lesssim_\varepsilon 1$.
\end{itemize}
This is verified as in \cite[Claim~2,~p.~308]{GuthHickmanIliopoulou2019}.
Furthermore, $\omega \in \text{supp} (g_{\theta,v})$, then $|\omega- \omega_\theta| < R^{-\frac{1}{2}}$, and so $|\Sigma(\omega) - \xi_\theta| \lesssim R^{-\frac{1}{2}}$, where $\xi_\theta = \Sigma(\omega_\theta)$. Consequently, by non-stationary phase,
\begin{equation*}
|(\eta_B \cdot T^\lambda g_{\theta,v}) \vert_{\Pi} \widehat (\xi) | \lesssim_N R^{O(1)} w_{B(\text{proj}_{W} \xi_\theta, R^{-\frac{1}{2}})}(\xi) \| g_{\theta,v} \|_{L^2}.
\end{equation*}
Let
\begin{align*}
L = \{ u \in A^{n-1} : (-\nabla h_{\bar{x}}(u),1) \in V \}, \\
S_\omega = \{ \omega \in A^{n-1} : G^\lambda(\bar{x};\omega) \in V \}.
\end{align*}
Let $A_u = T_\eta L^{\text{aff}}$ denote the affine variant of the linear subspace $T_\eta L$, and $A_\xi = A_u \times \mathbb{R}$. With $V_\xi$ denoting the linear subspace associated with $A_\xi$, we have $V_\xi^\perp = W$.
Next, we shall show that
\begin{equation}
\label{eq:DistanceFrequencySubspace}
\text{dist}(\xi_\theta, A_\xi) \lesssim R^{-\frac{1}{2}+\delta_m}.
\end{equation}
Then it follows
\begin{equation*}
w_{B(\text{proj}_{W} \xi_\theta, R^{-\frac{1}{2}})} \lesssim_\delta w_{B(\xi_*, R^{-\frac{1}{2}+\delta_m})}.
\end{equation*}
Here $\xi_*$ denotes the centre of a ball of radius $O(R^{-\frac{1}{2}+\delta_m})$ containing $\text{proj}_W(\xi_\theta)$ for $(\theta,v) \in \mathbb{T}_{B,\tau,V}$. We can again refer to \cite{GuthHickmanIliopoulou2019} for details. The claim follows then by estimating
\begin{equation*}
\| (\eta_B \cdot T^\lambda g \vert_\Pi) \widehat{\,} \; w^{-1}_{B(\xi_*,R^{-\frac{1}{2}+\delta_m})} \|_\infty
\end{equation*}
via Lemma \ref{lem:UncertaintyQuantification}.
We turn to the proof of \eqref{eq:DistanceFrequencySubspace}: Fix $(\theta,v) \in \mathbb{T}_{B,\tau,V}$ and let
\begin{equation*}
u_\theta = \text{proj}_{x_n^\perp} (\Sigma(\omega_\theta)).
\end{equation*}
We compute by triangle inequality
\begin{equation*}
\text{dist}(\xi_\theta,A_\xi) = \text{dist}(u_\theta,A_u) \leq \text{dist}(u_\theta, L \cap \Psi^{-1}(\tau)) + \sup_{u_* \in L \cap \Psi^{-1}(\tau)} \text{dist}(u_*,A_u).
\end{equation*}
Furthermore, by Lemma \ref{lem:ConsequencesReductionGaussMap},
\begin{equation*}
\text{dist}(u_\theta, L \cap \Psi^{-1}(\tau)) \sim \text{dist}(\omega_\theta,S_\omega \cap \tau) \lesssim \angle(G^\lambda(\bar{x};\omega_\theta),V) \lesssim R^{-\frac{1}{2}+\delta_m},
\end{equation*}
where the last inequality is by the definition of $\mathbb{T}_{V,B,\tau}$.
We turn to the estimate of the second term: Fix $u_* \in L \cap \Psi^{-1}(\tau)$. We note that $\text{dist}(u_*,A_u) = \text{dist}(u_*,A_{\bar{u}})$ for $\bar{u} = \frac{\| u_* \|}{\| u \|} u$ by null direction. Let $A_{\bar{u}} = u_0 + V_u$ for some linear subspace $V_u$. Now we note that the surface $L \cap \Psi^{-1}(\tau) \cap \{ \| u_* \| \mathbb{S}^{n-2} \}$, provided $\rho$ is large enough, can be written as subset of graph of a function $\psi: \mathcal{W} \to V_u^{\perp}$, where $\mathcal{W} \subseteq V_u$ is a subset around the origin of size $O(\rho^{-\frac{1}{2}+\delta_m})$. More precisely, we may write
\begin{equation*}
L \cap \Psi^{-1}(\tau) \cap \{ \| u_* \| \mathbb{S}^{n-2} \} \subseteq \{ w +\psi(w) : w \in \mathcal{W} \} + u_0
\end{equation*}
with $\psi(0) = 0$ and $\nabla \psi(0) = 0$. The estimate now follows from Taylor expansion as in \cite[p.~310]{GuthHickmanIliopoulou2019}.
\end{proof}
For proof of the transverse equidistribution estimate in Lemma \ref{lem:TransverseEquidistributionEstimate} requires we have to extend the estimate from fixed vector space to variety. The argument follows \cite[Section~8.4]{GuthHickmanIliopoulou2019} with the difference that the quantitative transversality mildly depends on the scale. We use the following result of Wongkew \cite{Wongkew1993} to control the size of neighbourhoods of varieties.
\begin{theorem}[{\cite{Wongkew1993}}]
\label{thm:WongkewCovering}
Suppose $Z=Z(P_1,\ldots,P_{n-m})$ is an $m$-dimensional transverse complete intersection in $\mathbb{R}^n$ with $\overline{\deg} Z \leq D$. For any $0 < \rho \leq R$ and $R$-ball $B_R$, $N_\rho(Z \cap B_R)$ can be covered by $O_D((R/\rho)^m)$ balls of radius $\rho$.
\end{theorem}
Next, we consider planar slices of neighbourhoods of varieties. We recall the following from \cite{GuthHickmanIliopoulou2019}: Any $m$-dimensional linear subspace $V$ can be expressed as a transverse complete intersection $V = Z(P_{N_1},\ldots,P_{N_{n-m}})$ with $\{ N_1,\ldots, N_{n-m} \}$ an orthonormal basis of $V^\perp$ and $P_{N_j}(x) = \langle x, N_j \rangle$. Let $V_1$, $V_2$ be linear subspaces in $\mathbb{R}^n$ and suppose that
\begin{equation}
\label{eq:DimensionConditionI}
\dim V_1 + \dim V_2 \geq n.
\end{equation}
$V_1 \cap V_2$ is a transverse complete intersection if and only if
\begin{equation*}
\dim(V_1 \cap V_2) = \dim V_1 + \dim V_2 - n.
\end{equation*}
This means that $V_1 \cap V_2$ is as small as possible.
\begin{definition}
A pair $(V_1,V_2)$ of linear subspaces $\mathbb{R}^n$ satisfying \eqref{eq:DimensionConditionI} is said to be \emph{quantitatively transverse} if the following hold:
\begin{itemize}
\item $\dim (V_1 \cap V_2) = \dim V_1 + \dim V_2 -n$;
\item $\angle(v_1,v_2)\geq c_{\text{trans}}$ for all non-zero vectors $v_j \in (V_1 \cap V_2)^\perp \cap V_j$, $j=1,2$.
\end{itemize}
\end{definition}
In the main argument the constant $c_{\text{trans}}$ will not be fixed, contrary to \cite{GuthHickmanIliopoulou2019}. We need to quantify the dependence on $c_{\text{trans}}$ in \cite[Lemma~8.13]{GuthHickmanIliopoulou2019}:
\begin{lemma}
\label{lem:PlanarSlicesLemma}
There exists some dimensional constant $C>0$ such that the following holds. Let $B_r \subseteq \mathbb{R}^n$ be an $r$-ball, $V \subseteq \mathbb{R}^n$ be a linear subspace, $Z$ be a transverse complete intersection and suppose that $\dim Z + \dim V \geq n$ and $(T_z Z, V)$ is a quantitatively transverse pair for all $z \in Z \cap 2 B_r$. Then, the following inclusion holds
\begin{equation*}
V \cap B_r \cap N_{\rho}(Z) \subseteq N_{2 \bar{C} \rho}(V \cap Z).
\end{equation*}
for all $0< \rho \ll r$ with $\bar{C} = \sin (c_{\text{trans}})^{-1}$.
\end{lemma}
\begin{proof}
This follows from the proof of \cite[Lemma~8.13]{GuthHickmanIliopoulou2019}.
\end{proof}
We are ready for the proof of Lemma \ref{lem:TransverseEquidistributionEstimate}:
\begin{proof}[Proof of Lemma \ref{lem:TransverseEquidistributionEstimate}]
First, by
\begin{equation*}
|G^\lambda(\bar{x};\theta) - G^\lambda(x;\theta)| \lesssim |x-\bar{x}|/\lambda \lesssim R^{-\frac{1}{2}+\delta_m},
\end{equation*}
we infer that
\begin{equation*}
\angle(G^\lambda(\bar{x};\theta), T_z Z) \lesssim R^{-\frac{1}{2}+\delta_m} \text{ for all } z \in Z \cap 2B.
\end{equation*}
Letting $V= T_z Z$, we have
\begin{equation*}
\mathbb{T}_{Z,B,\tau} \subseteq \mathbb{T}_{V,B,\tau}.
\end{equation*}
We can apply Lemma \ref{lem:TransverseEquidistributionEstimateLinearSubspace} to find a subspace $W$ such that
\begin{equation}
\label{eq:TransversalityVW}
\angle (V,W) \gtrsim K^{-4}
\end{equation}
and
\begin{equation}
\label{eq:TransverseEquidistributionSubspace}
\int_{\Pi \cap B(x_0,\rho^{\frac{1}{2}+\delta_m})} |T^\lambda g|^2 \lesssim_\delta R^{O(\delta_m)} \big( \frac{\rho}{R} \big)^{\frac{\dim W}{2}} \| g \|^{\frac{2 \delta}{1+\delta}}_{L^2} \big( \int_{\Pi \cap 2B} |T^\lambda g|^2 \big)^{\frac{1}{1+\delta}}
\end{equation}
for every affine subspace $\Pi$ parallel to $W$. By $K$-flatness of $Z$, we have that $(T_z Z, W)$ is a quantitatively transverse pair for all $z \in Z \cap 2B$. By Lemma \ref{lem:PlanarSlicesLemma}, we have
\begin{equation*}
\Pi \cap N_{\rho^{\frac{1}{2}+\delta_m}}(Z) \cap B \subseteq N_{C K^4 \rho^{\frac{1}{2}+\delta_m}}(\Pi \cap Z) \cap 2B.
\end{equation*}
Since $\Pi \cap Z$ is a transverse complete intersection of dimension $\dim V' + \dim Z - n$, Theorem \ref{thm:WongkewCovering} yields that $\Pi \cap N_{\rho^{\frac{1}{2}+\delta_m}}(Z) \cap B$ can be covered by
\begin{equation*}
O \big( R^{O(\delta_m)} \big( \frac{R}{\rho}\big)^{(\dim V + \dim Z - n)/2} \big) = O(R^{O(\delta_m)})
\end{equation*}
balls of radius $\rho^{\frac{1}{2}+\delta_m}$ because $K \lesssim R^\delta \ll \rho^{\delta_m}$. Applying \eqref{eq:TransverseEquidistributionSubspace} to each of these balls and summing, one deduces that
\begin{equation*}
\int_{\Pi \cap N_{\rho^{\frac{1}{2}+\delta_m}}(Z) \cap B} |T^\lambda g|^2 \lesssim_\delta R^{O(\delta_m)} \big( \frac{\rho}{R} \big)^{\frac{n-m}{2}} \| g \|^{2\delta/(1+\delta)}_{L^2} \big( \int_{\Pi \cap 2B} |T^\lambda g |^2 \big)^{\frac{1}{1+\delta}}.
\end{equation*}
Following the steps from \cite[p.~318]{GuthHickmanIliopoulou2019} completes the proof.
\end{proof}
\section{Comparing wave packets at different spatial scales}
\label{section:ComparingWavePackets}
For the induction on scales, we shall compare wave packet decompositions at different radii. Let $1 \ll R \ll \lambda$, and
\begin{equation*}
T^\lambda f(x) = \sum_{(\theta,v) \in \mathbb{T}} T^\lambda f_{\theta,v}(x) + \text{RapDec}(R) \| f \|_{L^2(A^{n-1})}.
\end{equation*}
In this section we recall the results from \cite[Section~9]{GuthHickmanIliopoulou2019}, which again did not hinge on $H2^+)$, but on non-degeneracy.
\subsection{Wave packets at smaller scale}
Let $R^{\frac{1}{2}} \leq \rho \leq R$ and fix $B(y,\rho) \subseteq B(0,R)$. $T^\lambda f_{\theta,v}$ can be decomposed into wave packets at scale $\rho$ over $B(y,\rho)$. For $g:A^{n-1} \to \mathbb{C}$, define $\tilde{g} = e^{i \phi^\lambda(y;\cdot)} g$, so that
\begin{equation*}
T^\lambda g(x) = \tilde{T}^\lambda \tilde{g}(\tilde{x}) \text{ for } \tilde{x} = x-y,
\end{equation*}
where $\tilde{T}^\lambda$ is the oscillatory integral operator with phase $\tilde{\phi}^\lambda$ and amplitude $\tilde{a}^\lambda$ given by
\begin{equation}
\label{eq:TranslationPhase}
\tilde{\phi}(x;\omega) = \phi(x + \frac{y}{\lambda};\omega) - \phi(\frac{y}{\lambda};\omega) \text{ and } \tilde{a}(x;\omega) = a(x+\frac{y}{\lambda};\omega).
\end{equation}
This yields by linearity
\begin{equation*}
T^\lambda f(x) = \sum_{(\theta,v) \in \mathbb{T}} \tilde{T}^\lambda(( f_{\theta,v}) \tilde{\,} )(\tilde{x}) + \text{RapDec}(R) \| f \|_{L^2(A^{n-1})}.
\end{equation*}
Each $T^\lambda f_{\theta,v}$ is (spatially) concentrated on the curve $R^{\frac{1}{2}+\delta}$-tube $T_{\theta,v}$ and, consequently, each $\tilde{T}^\lambda (f_{\theta,v}) \tilde{\,}$ is concentrated on $T_{\theta,v} - y$. Since
\begin{equation}
\label{eq:RelationDirectionSmallLargeTube}
\partial_\omega \tilde{\phi}^\lambda((\gamma^\lambda_{\omega,v}(t),t)-y;\omega) = v - \partial_\omega \phi^\lambda(y;\omega),
\end{equation}
the core curve $\Gamma^\lambda_{\theta,v} - y$ of $T_{\theta,v} - y$ is equal to $\Gamma^\lambda_{\theta,v - \bar{v}(y;\omega_\theta)}$, where
\begin{equation*}
\bar{v}(y;\omega) = \partial_\omega \phi^\lambda(y;\omega).
\end{equation*}
We repeat the construction of wave packets for each $\tilde{T}^\lambda (f_{\theta,v}) \tilde{\,}$ at scale $\rho$. Cover $A^{n-1}$ by finitely overlapping balls $\tilde{\theta}$ of radius $\rho^{-\frac{1}{2}}$, and $\mathbb{R}^{n-1}$ by finitely-overlapping balls of radius $\rho^{\frac{1+\delta}{2}}$ centered at vectors $\tilde{v} \in \rho^{\frac{1+\delta}{2}} \mathbb{Z}^{n-1}$. Let $\tilde{\mathbb{T}}$ denote the set of all pairs $(\tilde{\theta},\tilde{v})$. For each $(\theta,v) \in \mathbb{T}$ one may decompose
\begin{equation*}
( f_{\theta,v}) \tilde{\,} = \sum_{(\tilde{\theta},\tilde{v}) \in \tilde{\mathbb{T}}} (f_{\theta,v})^{\tilde{\,}}_{\tilde{\theta},\tilde{v}} + \text{RapDec}(R) \| f \|_{L^2(A^{n-1})}.
\end{equation*}
The significant contributions to this sum arise from pairs $(\tilde{\theta},\tilde{v})$ belonging to
\begin{equation*}
\tilde{\mathbb{T}}_{\theta,v} = \{(\tilde{\theta},\tilde{v}) \in \tilde{\mathbb{T}} : \, \text{dist}(\theta,\tilde{\theta}) \lesssim \rho^{-\frac{1}{2}} \text{ and } |v-\bar{v}(y;\omega_\theta)- \tilde{v} | \lesssim R^{\frac{1+\delta}{2}} \}.
\end{equation*}
\begin{lemma}[{\cite[Lemma~9.1]{GuthHickmanIliopoulou2019}}]
\label{lem:SmallWavePacketsFromLarge}
The following holds:
\begin{equation*}
(f_{\theta,v}) \tilde{\,} = \sum_{(\tilde{\theta},\tilde{v}) \in \tilde{\mathbb{T}}_{\theta,v}} (f_{\theta,v})_{\tilde{\theta},\tilde{v}}^{\tilde{\,}} + \text{RapDec}(R) \| f \|_{L^2}.
\end{equation*}
\end{lemma}
\subsection{Tangency properties}
In this subsection we recall how tangency properties of the large wave packets are inherited by the small wave packets (cf. \cite[Section~9.2]{GuthHickmanIliopoulou2019}). On the other hand, recall that a small wave packet coming from a large packet, which is tangential to $Z$, need not be contained in a neighbourhood of $Z$ on a small scale. The small wave packet is located too far away from $Z$. However, with the angle condition inherited, we shall see that the small wave packet is contained in a small neighbourhood of a translate of the variety.
We analyze functions $h$ concentrated on wave packets from
\begin{equation*}
\mathbb{T}_{Z,B(y,\rho)} = \{ (\theta,v) \in \mathbb{T}_Z : T_{\theta,v} \cap B(y,\rho) \neq \emptyset \}.
\end{equation*}
For this purpose, we consider the core of a small tube:
\begin{equation*}
\partial_\omega \tilde{\phi}^\lambda(\tilde{\gamma}^\lambda_{\omega,v}(t),t;\omega) = v
\end{equation*}
for $t \in (-\rho,\rho)$. By \eqref{eq:RelationDirectionSmallLargeTube}, we have the identity:
\begin{equation}
\label{eq:RelationCoresSmallTubesLargeTubes}
\gamma^\lambda_{\omega,v}(t) = \tilde{\gamma}^\lambda_{\omega,v-\bar{v}(y;\omega)}(t-y_n)+y'.
\end{equation}
Let $\tilde{T}_{\omega,v}$ be the $\rho^{\frac{1}{2}+\delta}$-tube with core curve $\tilde{\Gamma}^\lambda_{\omega,v} = (\tilde{\gamma}^\lambda_{\omega,v}(t),t)$.
We have the following:
\begin{lemma}[{\cite[Lemma~9.3]{GuthHickmanIliopoulou2019}}]
\label{lem:DistanceSmallLargeTubes}
If $(\theta,v) \in \mathbb{T}$ and $(\tilde{\theta},\tilde{v}) \in \tilde{\mathbb{T}}_{\theta,v}$, then
\begin{equation*}
| \tilde{\Gamma}^\lambda_{\tilde{\theta},\tilde{v}}(t) - ( \Gamma^\lambda_{\theta,v}(t+y_n) - y) | \lesssim R^{\frac{1+\delta}{2}}
\end{equation*}
for all $t \in (-\rho,\rho)$.
\end{lemma}
Fix $(\theta,v) \in \mathbb{T}_Z$ and $(\tilde{\theta},\tilde{v}) \in \tilde{\mathbb{T}}_{\theta,v}$. Next, we show that for $x \in \tilde{T}_{\tilde{\theta},\tilde{v}}$ and $z \in Z$ and $b \in B(0,2 R^{\frac{1}{2}+\delta_m})$ are such that
\begin{equation*}
z-y+b \in B(0,4 \rho) \text{ and } |x-(z-y+b)| \leq \bar{C}_{tang} \rho^{\frac{1}{2}+\delta_m},
\end{equation*}
then we find the following estimate to hold:
\begin{equation}
\label{eq:SmallAngleSmallTube}
\angle (\tilde{G}^\lambda(x;\omega_{\tilde{\theta}}), T_{z-y+b} (Z-y+b)) \leq \bar{c}_{tang} \rho^{-\frac{1}{2}+\delta_m},
\end{equation}
where $\tilde{G}^\lambda$ is the generalized Gauss map associated with $\tilde{\phi}^\lambda$. We have $\tilde{G}^\lambda(x;\omega) = G^\lambda(x+y;\omega)$ and $T_{z-y+b}(Z-y+b) = T_z Z$, so it is equivalent to check that
\begin{equation*}
\angle(G^\lambda(x+y;\omega_{\tilde{\theta}}), T_z Z) \leq \bar{c}_{tang} \rho^{-\frac{1}{2}+\delta_m}.
\end{equation*}
By Lemma \ref{lem:DistanceSmallLargeTubes}, the definition of $\tilde{T}_{\tilde{\theta},\tilde{v}}$, and assuming that $\rho \leq R^{1-\delta}$, it follows
\begin{equation*}
|x+y - \Gamma^\lambda_{\theta,v}(x_n+y_n)| \lesssim R^{\frac{1+\delta}{2}}.
\end{equation*}
By expanding the Gauss map, we find
\begin{equation*}
\angle(G^\lambda(x+y;\omega_{\tilde{\theta}}), T_z Z) \lesssim \angle (G^\lambda(\Gamma^\lambda_{\theta,v}(x_n+y_n);\omega_\theta),T_z Z) + \rho^{-\frac{1}{2}}.
\end{equation*}
Finally, $\Gamma^\lambda_{\theta,v}(x_n+y_n) \in T_{\theta,v}$, which is $R^{-\frac{1}{2}+\delta_m}$-tangent to $Z$. Hence,
\begin{equation*}
\angle (G^\lambda(\Gamma^\lambda_{\theta,v}(x_n+y_n);\omega_\theta), T_z Z) \leq \bar{c}_{tang} R^{-\frac{1}{2}+\delta_m}.
\end{equation*}
Likewise the argument in \cite[pp.~326f]{GuthHickmanIliopoulou2019} shows that a $\rho^{\frac{1}{2}+\delta}$-tube, which intersects $N_{\rho^{\frac{1}{2} + \delta_m}/2}(Z-y+b) \cap B(0,\rho)$, is actually contained in $N_{\rho^{\frac{1}{2}+\delta_m}}(Z-y+b)$ by virtue of \eqref{eq:SmallAngleSmallTube}.
We arrive at the following proposition (cf. \cite[Proposition~9.2]{GuthHickmanIliopoulou2019}):
\begin{proposition}
\label{prop:InheritedTangencyProperties}
Let $R^{\frac{1}{2}} \leq \rho \leq R^{1-\delta}$ and $Z \subseteq \mathbb{R}^n$ be a transverse complete intersection.
\begin{itemize}
\item[(1)] Let $(\theta,v) \in \mathbb{T}_Z$ and $b \in B(0,2 R^{\frac{1}{2}+\delta_m})$. If $(\tilde{\theta},\tilde{v}) \in \tilde{\mathbb{T}}_{\theta,v}$ satisfies
\begin{equation*}
\tilde{T}_{\tilde{\theta},\tilde{v}} \cap N_{\rho^{\frac{1}{2}+\delta_m}/2}(Z-y+b) \neq \emptyset,
\end{equation*}
then $(\tilde{\theta},\tilde{v}) \in \tilde{\mathbb{T}}_{Z-y+b}$.
\item[(2)] If $h$ is concentrated on wave packets in $\mathbb{T}_{Z,B(y,\rho)}$, then $\tilde{h}$ is concentrated on wave packets in
\begin{equation*}
\bigcup_{|b| \lesssim R^{\frac{1}{2}+\delta_m}} \tilde{\mathbb{T}}_{Z-y+b}.
\end{equation*}
\end{itemize}
\end{proposition}
We also make the following definition:
\begin{equation*}
\tilde{\mathbb{T}}_b = \{ (\tilde{\theta},\tilde{v}) : (\tilde{\theta},\tilde{v}) \in \bigcup_{(\theta,v) \in \mathbb{T}_{Z,B(y,\rho)}} \tilde{\mathbb{T}}_{\theta,v} : \tilde{T}_{\tilde{\theta},\tilde{v}} \cap N_{\rho^{\frac{1}{2}+\delta_m}/2}(Z-y+b) \neq \emptyset \}.
\end{equation*}
By the above, we have $\tilde{\mathbb{T}}_b \subseteq \tilde{\mathbb{T}}_{Z-y+b}$. For a function $h$ concentrated on wave packets in $\mathbb{T}_{Z,B(y,\rho)}$, we consider a function of the form
\begin{equation*}
\tilde{h}_b = \sum_{(\tilde{\theta},\tilde{v}) \in \tilde{\mathbb{T}}_b} \tilde{h}_{\tilde{\theta},\tilde{v}}.
\end{equation*}
Since $\tilde{\mathbb{T}}_b \subseteq \tilde{\mathbb{T}}_{Z-y+b}$ by Proposition \ref{prop:InheritedTangencyProperties}, we have
\begin{equation*}
\tilde{T}^\lambda \tilde{h}_b(\tilde{x}) = T^\lambda h_b(x) \chi_{N_{\rho^{\frac{1}{2}+\delta_m}}(Z+b)}(x) + \text{RapDec}(R) \| h \|_{L^2}
\end{equation*}
for all $x=\tilde{x}+y \in B(y,\rho)$.
\subsection{Sorting wave packets}
In this section we recall sorting the large wave packets by ``medium tubes". This was carried out in\cite[Section~9.3]{GuthHickmanIliopoulou2019}. Given a ball $B(y,\rho)$, many large wave packets $(\theta,v)$ might give rise to essentially the same set $\tilde{\mathbb{T}}_{\theta,v}$. Medium tube segments of $T_\rho$ of length $\rho$ and radius $R^{\frac{1}{2}+\delta}$ allow for a grouping into large and small wave packets.
We give the details: Let $\mathcal{T}$ denote the collection of all pairs $(\tilde{\theta},w)$ formed by a $\rho^{-\frac{1}{2}}$-ball $\tilde{\theta}$ and $w \in R^{(1+\delta)/2} \mathbb{Z}^{n-1}$. For each $(\tilde{\theta},w) \in \mathcal{T}$, choose some
\begin{equation*}
\mathcal{T}_{\tilde{\theta},w} \subseteq \{ (\theta,v) \in \mathbb{T}: \text{dist}(\theta,\tilde{\theta}) \lesssim \rho^{-\frac{1}{2}} \text{ and } |v-\bar{v}(y;\omega_\theta) - w | \lesssim R^{(1+\delta)/2} \}
\end{equation*}
so that the family $\{ \mathcal{T}_{\tilde{\theta},w}: (\tilde{\theta},w) \in \mathcal{T} \}$ forms a covering of $\mathbb{T}$ by disjoint sets. The medium tubes are given by
\begin{equation*}
T_{\tilde{\theta},w} = \bigcup_{(\theta,v) \in \mathcal{T}_{\tilde{\theta},w}} T_{\theta,v} \cap B(y,\rho).
\end{equation*}
If $(\tilde{\theta},w) \in \mathcal{T}$ and $(\theta,v) \in \mathcal{T}_{\tilde{\theta},w}$, then (cf. \cite[Cor.~9.4]{GuthHickmanIliopoulou2019})
\begin{equation*}
\text{dist}_H(T_{\theta,v} \cap B(y,\rho), T_{\tilde{\theta},w}) \lesssim R^{\frac{1}{2}+\delta}.
\end{equation*}
Let $g:A^{n-1} \to \mathbb{C}$ be integrable and define
\begin{equation*}
g_{\tilde{\theta},w} = \sum_{(\theta,v) \in \mathcal{T}_{\tilde{\theta},w}} g_{\theta,v}
\end{equation*}
for all $(\tilde{\theta},w) \in \mathcal{T}$. Since $\mathcal{T}_{\tilde{\theta},w}$ cover $\mathbb{T}$ and are disjoint, it follows that
\begin{equation*}
g = \sum_{(\tilde{\theta},w) \in \mathcal{T}} g_{\tilde{\theta},w} + \text{RapDec}(R) \| g \|_{L^2}.
\end{equation*}
The functions $g_{\tilde{\theta},w}$ are almost orthogonal and, consequently,
\begin{equation*}
\| g \|^2_{L^2} \sim \sum_{(\tilde{\theta},w) \in \mathcal{T}} \| g_{\tilde{\theta},w} \|_{L^2}^2.
\end{equation*}
$(g_{\tilde{\theta},w}) \tilde{\,}$ is concentrated on scale $\rho$ wave packets belonging to $\bigcup_{(\theta,v) \in \mathcal{T}_{\tilde{\theta},w}} \tilde{\mathbb{T}}_{\theta,v}$. This union is contained in
\begin{equation*}
\tilde{\mathcal{T}}_{\tilde{\theta},w} = \{ (\tilde{\theta}',\tilde{v}) \in \tilde{\mathbb{T}} : \text{dist}(\tilde{\theta}',\tilde{\theta}) \lesssim \rho^{-1/2} \text{ and } |\tilde{v}-w| \lesssim R^{\frac{1+\delta}{2}} \}.
\end{equation*}
The family $\{ \tilde{\mathcal{T}}_{\tilde{\theta},w} : (\tilde{\theta},w) \in \mathcal{T} \}$ forms a covering of $\tilde{\mathbb{T}}$ by almost disjoint sets. Hence, we have almost orthogonality between the scale $\rho$ wave packets of the different functions $(g_{\tilde{\theta},w}) \tilde{\,}$ (cf. \cite[Eq.~(9.17)]{GuthHickmanIliopoulou2019}):
\begin{equation*}
\big\| \sum_{(\tilde{\theta},w) \in \mathcal{T}} (g_{\tilde{\theta},w})_b \tilde{\,} \big\|^2_{L^2} \sim \sum_{(\tilde{\theta},w) \in \mathcal{T}} \| (g_{\tilde{\theta},w})_b \tilde{\,} \|_{L^2}^2.
\end{equation*}
\subsection{Reverse H\"ormander $L^2$-estimate}
In the following we record a reverse H\"ormander $L^2$-estimate. This will imply transverse equidistribution estimates for functions concentrated on wave packets, which are sorted as above. This was previously done in \cite[Subsection~9.4]{GuthHickmanIliopoulou2019}, whose statements carry over. Thus, the proofs are omitted. We collect the relevant estimates here for future reference.
Let $Z$ be an $m$-dimensional transverse complete intersection, $(\tilde{\theta},w) \in \mathcal{T}$ and $h$ be a function concentrated on $\mathbb{T}_{Z \cap B(y,\rho)} \cap \mathcal{T}_{\tilde{\theta},w}$. By the above, every scale $R$ wave packet of $h$ intersects $B(y,\rho)$ on the set $T_{\tilde{\theta},w}$, which has a Hausdorff distance $\lesssim R^{\frac{1}{2}+\delta_m}$ to $T_{\theta,v} \cap B(y,\rho)$ for any $(\theta,v) \in \mathcal{T}_{\tilde{\theta},w}$. Moreover, the scale $\rho$ wave packets of $\tilde{h}$ will intersect $B(x_0-y,CR^{\frac{1}{2}+\delta_m})$. In this case, the following reverse of H\"ormander's $L^2$-estimate holds:
\begin{lemma}[{\cite[Lemma~9.5,~p.~329]{GuthHickmanIliopoulou2019}}]
\label{lem:ReverseL2Bound}
Let $T^\lambda$ be an oscillatory integral operator with phase $\phi^\lambda$ given by a translate of a reduced phase in the sense of \eqref{eq:TranslationPhase} and $1 \leq R^{\frac{1}{2}+\delta} \leq r \lesssim \lambda^{\frac{1}{2}}$. There exists a family of oscillatory integral operators $\mathbf{T}^\lambda$ all with phase $\phi^\lambda$ such that the following hold:
\begin{itemize}
\item[(i)] each $T^\lambda \in \mathbf{T}^\lambda$ is again an operator with phase given by a translate of a reduced phase as in \eqref{eq:TranslationPhase},
\item[(ii)] $\# \mathbf{T}^\lambda = O(1)$;
\item[(iii)] if $f$ is concentrated on wave packets to (with respect to $T^\lambda$) which intersect some $B(\bar{x},r) \subseteq B(0,R)$, then
\begin{equation*}
\| f \|^2_{L^2} \lesssim r^{-1} \| T^\lambda_* f \|^2_{L^2(B(\bar{x};Cr))}
\end{equation*}
holds for some $T^\lambda_* \in \mathbf{T}^\lambda$.
\end{itemize}
\end{lemma}
Lemma \ref{lem:ReverseL2Bound} is proved in \cite{GuthHickmanIliopoulou2019} via Fourier series expansion and Plancherel's theorem. The proof hinges only on the non-degeneracy $C1)$ of the phase function. Hence, it applies to the homogeneous phase functions presently considered as well. For $h$ as above, $x_0 \in T_{\tilde{\theta},w}$ and $|b| \lesssim R^{\frac{1}{2}+\delta_m}$, this implies that $\tilde{h}_b$, as defined above, is a sum of wave packets which intersect
\begin{equation*}
B(x_0-y,CR^{\frac{1}{2}+\delta_m}).
\end{equation*}
Applying Lemma \ref{lem:ReverseL2Bound} at scale $\rho$ with $r \sim R^{\frac{1}{2}+\delta_m}$ to $\tilde{h}_b$ yields
\begin{equation*}
\| \tilde{h}_b \|^2_{L^2} \lesssim R^{-\frac{1}{2}-\delta_m} \| \tilde{T}^\lambda_* \tilde{h}_b \|^2_{L^2(B(x_0-y,CR^{\frac{1}{2}+\delta_m})}.
\end{equation*}
Since the tangency properties of $T^\lambda$ are inherited by $T^\lambda_*$, we infer
\begin{equation*}
\| \tilde{h}_b \|^2_{L^2} \lesssim R^{-\frac{1}{2}-\delta_m} \| T^\lambda_* h_b \|^2_{L^2(N_{\rho^{\frac{1}{2}+\delta_m}}(Z+b) \cap B(x_0,CR^{\frac{1}{2}+\delta_m}))}.
\end{equation*}
Applying H\"ormander's $L^2$-bound yields the following lemma:
\begin{lemma}
{\cite[Lemma~9.6]{GuthHickmanIliopoulou2019}}
\label{lem:Orthogonalityhb}
Let $h$ be concentrated on wave packets from $\mathbb{T}_{Z\cap B(y,\rho)} \cap \mathcal{T}_{\tilde{\theta},w}$ for some $(\tilde{\theta},w) \in \mathcal{T}$. Let $\mathfrak{B} \subseteq B(0,CR^{\frac{1}{2}+\delta_m})$ be such that the sets
\begin{equation*}
N_{\rho^{\frac{1}{2}+\delta_m}}(Z+b) \cap B(x_0,CR^{\frac{1}{2}+\delta_m})
\end{equation*}
are essentially disjoint over $b \in \mathfrak{B}$. Then,
\begin{equation*}
\sum_{b \in \mathfrak{B}} \| \tilde{h}_b \|^2_{L^2} \lesssim \| h \|^2_{L^2}.
\end{equation*}
\end{lemma}
\section{Main inductive argument}
\label{section:MainInductiveArgument}
The $k$-broad estimate is a consequence of the following claim, which is suitable for induction. Let
\begin{equation*}
\bar{p}(k,n) = 2 \cdot \frac{n+k}{n+k-2}.
\end{equation*}
\begin{theorem}
\label{thm:MainInductionTheorem}
For $\varepsilon > 0$, sufficiently small, there are
\begin{equation*}
0<\delta \ll \delta_{n-1} \ll \ldots \ll \delta_1 \ll \delta_0 \ll \varepsilon
\end{equation*}
and large dyadic parameters $\bar{A}_\varepsilon$, $\bar{C}_\varepsilon$, $D_{m,\varepsilon} \lesssim_\varepsilon 1$ and $\theta_m < \varepsilon$ such that the following holds. Suppose $Z=Z(P_1,\ldots,P_{n-m})$ is a transverse complete intersection with $\overline{\deg} Z \leq D_{m,\varepsilon}$. For all $2 \leq k \leq n$, $1 \leq A \leq \bar{A}_\varepsilon$ dyadic and $1 \leq K \leq R \leq \lambda$, the inequality
\begin{equation}
\label{eq:MainInductionEstimate}
\| T^\lambda f \|_{BL^p_{k,A}(B(0,R))} \lesssim_\varepsilon K^{\bar{C}_\varepsilon} R^{\theta_m + \delta(\log \bar{A}_\varepsilon - \log A) - e_{k,n}(p) + \frac{1}{2}} \| f \|_{L^2(A^{n-1})}
\end{equation}
holds whenever $f$ is concentrated on wave packets from $\mathbb{T}_Z$ and
\begin{equation}
\label{eq:Inductionp}
2 \leq p \leq \bar{p}_0(k,m) =
\begin{cases}
\bar{p}(k,m), \quad \text{if } k <m, \\
\bar{p}(m,m) + \delta, \quad \text{if } k=m.
\end{cases}
\end{equation}
Above,
\begin{equation*}
e_{k,n}(p) = \frac{1}{2} \big( \frac{1}{2} - \frac{1}{p} \big) (n+k).
\end{equation*}
\end{theorem}
In the first step we reduce to $R \lesssim_\varepsilon \lambda^{1-\varepsilon}$ by covering $B(0,\lambda)$ with balls of radius $\lambda^{1-\varepsilon}$. The technical details are provided in \cite[Lemma~10.2]{GuthHickmanIliopoulou2019}. This reduction is necessary to allow for polynomial approximation of the core curve $\gamma^\lambda_{\omega,v}$ uniformly in $R$.
Next, we set up the induction argument for $1 \leq R \lesssim_\varepsilon \lambda^{1-\varepsilon}$. For $\varepsilon >0$ sufficiently small, it is enough to consider $K \lesssim R^\delta$ by choosing $\bar{C}_\varepsilon$ sufficiently large (as the claim then follows from the trivial $L^1$-$L^\infty$-estimate and crude summation). We let furthermore
\begin{equation}
\label{eq:Parameters}
\begin{split}
D_{m,\varepsilon} = \varepsilon^{-\delta^{-(2n-m)}}, \quad \theta(\varepsilon) = \varepsilon - c_n \delta_m, \quad \bar{A}_\varepsilon = \lceil e^{\frac{10n}{\delta}} \rceil, \\
\delta_i = \delta_i(\varepsilon) = \varepsilon^{2i + 1} \text{ for all } i=1, \ldots, n-1, \text{ and } \delta = \delta(\varepsilon) \ll \delta_{n-1}.
\end{split}
\end{equation}
The base case is given by $m \leq k-1$, and $A \geq 2^{10}$. For details we refer to \cite[Subsection~10.3]{GuthHickmanIliopoulou2019}.
\subsection{Inductive step} Let $2 \leq k \leq n-1$, $k \leq m \leq n$, and $K \lesssim_\varepsilon R^\delta$. Assume, by way of induction hypothesis, that \eqref{eq:MainInductionEstimate} holds whenever $\text{dim} Z \leq m-1$ or the radial parameter is at most $\frac{R}{2}$. Fix $\varepsilon >0$, $1 < A \leq \bar{A}_\varepsilon$ and a transverse complete intersection $Z = Z(P_1,\ldots,P_{n-m})$ with $\overline{\deg} Z \leq D_{m,\varepsilon}$, where $\bar{A}_\varepsilon$ and $D_{m,\varepsilon}$ are as in \eqref{eq:Parameters}. Let $f$ be concentrated on wave packets from $\mathbb{T}_Z$. It suffices to show \eqref{eq:MainInductionEstimate} for $p = \bar{p}_0(k,m)$ by interpolation with the trivial $L^2$-bound. We recall the two cases to be analyzed:
\medskip
\textbf{The algebraic case:} There exists a transverse complete intersection $Y^l \subseteq Z$ of dimension $1 \leq l \leq m-1$ of maximum degree at most $(D_{m,\varepsilon})^n$ such that
\begin{equation}
\label{eq:AlgebraicCase}
\| T^\lambda f \|^p_{BL^p_{k,A}(N_{R^{\frac{1}{2}+\delta_m}/4}(Y^l) \cap B(0,R))} \geq c_{alg} \| T^\lambda f \|^p_{BL^p_{k,A}(B(0,R))}.
\end{equation}
Here $c_{alg}>0$ depends on $n$ and $\varepsilon$.
\textbf{The cellular case:} For any transverse complete intersection $Y^{l} \subseteq Z$ of dimension $1 \leq l \leq m-1$ and maximum degree at most $(D_{m,\varepsilon})^n$, the inequality
\begin{equation}
\label{eq:CellularCase}
\| T^\lambda f \|^p_{BL^p_{k,A}(N_{R^{\frac{1}{2} + \delta_m}/4}(Y^l)) \cap B(0,R)} < c_{alg} \| T^\lambda f \|^p_{BL^p_{k,A}(B(0,R))}
\end{equation}
holds.
The cellular case is as usually treated by induction on the radius. Via polynomial partitioning the $BL^p_{k,A}$-norm is equidistributed among the cells and the induction closes. The algebraic case is more involved: $T^\lambda f$ can be regarded as concentrated near a low-dimensional and low degree variety $Y^l$ (for an oversimplification, think of a hyperplane). If the wave packets from $f$ are also tangential to this variety, then we can use induction on the dimension to conclude. If this is not the case and many wave packets are transverse to $Y^l$, we conclude via transverse equidistribution estimates.
\subsubsection{Cellular case}
This case is handled as in \cite[Section~10.5]{GuthHickmanIliopoulou2019}. We omit the details.
\subsubsection{Algebraic case}
In this case transverse equidistribution estimates become important at one step. This is different than \cite[Section~10.6]{GuthHickmanIliopoulou2019}, and we turn to the details. Fix a transverse complete intersection $Y^l$ of dimension $1 \leq l \leq m-1$ of maximum degree $\overline{\deg} Y^l \leq (D_{m,\varepsilon})^n$, which satisfies \eqref{eq:MainInductionEstimate}. Let $R^{\frac{1}{2}} \ll \rho \ll R$ be such that $\rho^{\frac{1}{2}+\delta_l} = R^{\frac{1}{2}+\delta_m}$, and note that
\begin{equation*}
R \leq R^{2 \delta_l} \rho \text{ and } \rho \leq R^{-\delta_l/2} R.
\end{equation*}
Let $\mathcal{B}_\rho$ be a finitely overlapping cover of $B(0,R)$ by $\rho$-balls, and for each $B \in \mathcal{B}_\rho$ define
\begin{equation*}
\mathbb{T}_B = \{(\theta,v) \in \mathbb{T}: T_{\theta,v} \cap N_{R^{\frac{1}{2} + \delta_m}/4}(Y^l) \cap B \neq \emptyset \}
\end{equation*}
and
\begin{equation*}
f_B:= \sum_{(\theta,v) \in \mathbb{T}_B} f_{\theta,v}.
\end{equation*}
We have by the triangle inequality for broad norms
\begin{equation*}
\| T^\lambda f \|^p_{BL^p_{k,A}(B(0,R))} \lesssim \sum_{B \in \mathcal{B}_\rho} \| T^\lambda f_B \|^p_{BL^p_{k,A}(N_{R^{\frac{1}{2}+\delta_m}/4}(Y^l) \cap B))}
\end{equation*}
up to $\text{RapDec}(R) \| f \|^p_{L^2}$ on the right-hand side by the rapid decay off the wave packets.
For $B=B(y,\rho) \in \mathcal{B}_\rho$, let $\mathbb{T}_{B,tang}$ denote the set of all $(\theta,v) \in \mathbb{T}_B$ with the property that, whenever $x \in T_{\theta,v}$ and $z \in Y^l \cap B(y,2 \rho)$ satisfy $|x-z| \leq 2 \bar{C}_{tang} \rho^{\frac{1}{2}+\delta_l}$, it follows that
\begin{equation*}
\angle (G^\lambda(x;\omega_\theta), T_z Y^l) \leq \frac{1}{2} \bar{c}_{tang} \rho^{-\frac{1}{2}+\delta_l},
\end{equation*}
where $\bar{C}_{tang}$ and $\bar{c}_{tang}$ are the constants appearing in the definition of tangency. Furthermore, let $\mathbb{T}_{B,trans} = \mathbb{T}_B \backslash \mathbb{T}_{B,tang}$ and define
\begin{equation*}
f_{B,tang} = \sum_{(\theta,v) \in \mathbb{T}_{B,tang}} f_{\theta,v} \text{ and } f_{B,trans} = \sum_{(\theta,v) \in \mathbb{T}_{B,trans}} f_{\theta,v}.
\end{equation*}
It follows that $f_B = f_{B,tang} + f_{B,trans}$ and, by the triangle inequality for broad norms, one concludes that
\begin{equation*}
\| T^\lambda f \|^p_{BL^p_{k,A}(B(0,R))} \lesssim \sum_{B \in \mathcal{B}_\rho} \| T^\lambda f_{B,tang} \|^p_{BL^p_{k,A/2}(B)} + \sum_{B \in \mathcal{B}_\rho} \| T^\lambda f_{B,trans} \|^p_{BL^p_{k,A/2}(B)}.
\end{equation*}
Either the tangential or transverse contribution to the above sum dominates, and each case is treated separately.
\textbf{Tangential subcase:} Suppose that the tangential term dominates and we have
\begin{equation}
\| T^\lambda f \|^p_{BL^p_{k,A}(B(0,R))} \lesssim \sum_{B \in \mathcal{B}_\rho} \| T^\lambda f_{B,tang} \|^p_{BL^p_{k,A/2}(B)}.
\end{equation}
This case can be handled as in \cite[pp.345-346]{GuthHickmanIliopoulou2019}, and we skip the details.
\medskip
\textbf{Transverse sub-case:} In this case, we have
\begin{equation*}
\| T^\lambda f \|^p_{BL^p_{k,A}(B(0,R))} \lesssim \sum_{B \in \mathcal{B}_\rho} \| T^\lambda f_{B,trans} \|^p_{BL^p_{k,A/2}(B)}.
\end{equation*}
Following \cite{GuthHickmanIliopoulou2019}, we use an argument similar to the cellular case. In the transverse case the number of cells a given tube can enter is controlled by transversality as follows:
\begin{lemma}[{\cite[Eq.~(10.23)]{GuthHickmanIliopoulou2019}}]
We find the following estimate to hold:
\begin{equation}
\label{eq:AlmostOrthogonalityRhoBallsTransverseCase}
\sum_{B \in \mathcal{B}_\rho} \| f_{B,trans} \|^2_{L^2(A^{n-1})} \lesssim_\varepsilon \| f \|^2_{L^2(A^{n-1})}.
\end{equation}
\end{lemma}
The strategy in the transverse case is to use induction on radius to show that for some $\overline{c}_\varepsilon > 0$ one has (redenoting $f_j$ for $f_{B_j,trans}$)
\begin{equation}
\label{eq:TransverseSubcaseKeyEstimate}
\| T^\lambda f_{j} \|_{BL^p_{k,A/2}(B)} \leq \overline{c}_\varepsilon E_{m,A}(R) \| f_{j} \|_{L^2(A^{n-1})}
\end{equation}
for all $B_j \in \mathcal{B}_\rho$.
Provided $\overline{c}_\varepsilon > 0$ is chosen sufficiently small, depending only on $n$ and $\varepsilon$,
\eqref{eq:TransverseSubcaseKeyEstimate} can be combined with \eqref{eq:AlmostOrthogonalityRhoBallsTransverseCase} and the estimate
\begin{equation*}
\| f_{B,trans} \|_{L^2(A^{n-1})} \lesssim \| f \|_{L^2}
\end{equation*}
to yield
\begin{equation*}
\| T^\lambda f \|_{BL^p_{k,A/2}(B(0,R))} \lesssim_\varepsilon \bar{c}_\varepsilon E_{m,A}(R) \| f \|^{1-\frac{2}{p}}_{L^2} \big( \sum_{B \in \mathcal{B}_\rho} \| f_{B,trans} \|^2_{L^2} \big)^{\frac{1}{p}} \leq E_{m,A}(R) \| f \|_{L^2}.
\end{equation*}
The main obstacle is that $f_{j}$ do not, in general, satisfy the hypothesis of Theorem \ref{thm:MainInductionTheorem} at scale $\rho$. The remedy is to break $f_{j}$ into pieces $f_{j,b}$, which are $\rho^{\frac{1}{2}+\delta_m}$-tangent to a translated variety of $Z+b$.
\medskip
\textbf{Flattening the variety:}
In the first step we flatten the variety up to $K^{-5}$ (cf. \cite[p.~25]{OuWang2021}). This requires to estimate $K^{5(n-1)}$-expressions of the form $\| T^\lambda f_j \|_{BL^p_{k,A/2}(B)}$ with $Z$ flat up to angles $K^{-5}$. The factor of $K^{O(n)}$ is admissible (see \eqref{eq:ImprovedTransverseSummation} below).
\medskip
\textbf{Separating essentially and non-essentially contributing $R^{\frac{1}{2}+\delta_m}$-balls}: Cover $B_j$ by finitely overlapping $R^{\frac{1}{2}+\delta_m}$-balls $B_{j,k}$. Let $(\theta,v) \in \mathbb{T}_{Z,B_{j,k}}$ and $x \in T_{\theta,v} \cap N_{R^{\frac{1}{2}+\delta_m}}(Z) \cap B_{j,k}$, $z \in Z$ with $|x-z| \leq \bar{C}_{tang} R^{\frac{1}{2}+\delta_m}$. By definition of tangency, we have
\begin{equation*}
\angle (G^\lambda(x;\omega_\theta),T_{z} Z) \lesssim R^{-\frac{1}{2}+\delta_m}.
\end{equation*}
Let $V= T_z Z$. By Lemma \ref{lem:ConsequencesReductionGaussMap}, we have
\begin{equation*}
\angle (G^\lambda(\bar{x};\omega_\theta),V) \lesssim R^{-\frac{1}{2}+\delta_m}.
\end{equation*}
Now we consider the linearization $\tilde{\phi}_{\bar{x}}(u) = \partial_{x_n} \phi^\lambda (\bar{x};\Psi^\lambda(u))$ around $\bar{x}$, the centre of $B_{j,k}$. We consider as in Section \ref{subsection:Linearizing}
\begin{equation*}
V= \{ \sum_{j=1}^n a_{i,j} x_j = 0, \quad i =1,\ldots,n-m \}, \quad L_{\bar{x}} = \{ u \in A^{n-1} : \sum_{j=1}^{n-1} a_{i,j} \partial_j \tilde{\phi}_{\bar{x}}(u) - a_{i,n} = 0 \}
\end{equation*}
such that $L_{\bar{x}}$ denotes the $u$-frequencies with normal in $V$.\\
We apply the dichotomy of Section \ref{subsection:Linearizing}: Either $L_{\bar{x}}$ is contained in $O(1)$ slabs of size $1 \times K^{-2} \times \ldots \times K^{-2}$ by Lemma \ref{lem:NegligibleContributionBall}. This is referred to as Case I. Note that if $L_{\bar{x}}$ is contained in $O(1)$ $1 \times K^{-2} \times \ldots \times K^{-2}$-slabs, then so is $\bigcup \theta$ with $\angle (G^\lambda(\bar{x};\omega_\theta),V) \lesssim R^{-\frac{1}{2}+\delta_m}$ by Lemma \ref{lem:ConsequencesReductionGaussMap}. Consequently, Case I-balls can be neglected in the $k$-broad norm (see \eqref{eq:EssentialContribution} below).
Otherwise, we consider the further refinement $\mathbb{T}_{V,B_{j,k},\tau}$ with $\tau$ a $\rho^{-\frac{1}{2}}$-sector. By Lemma \ref{lem:QuantitativeTransverse}, there is a quantitatively transverse subspace $W$ with $\bar{V} \oplus W = \mathbb{R}^n$ and
\begin{equation*}
\angle (V,W) \gtrsim K^{-4}.
\end{equation*}
$\bar{V}$ denotes a suitable extension of a tangent space of $L_{\bar{x}}$ from Subsection \ref{subsection:Linearizing} (Case II). We let $X_I$ and $X_{II}$ denote the union of balls $B_{j,k}$ from Cases I and II.
Next, we use the sorting into medium tubes as in Section \ref{subsection:Linearizing}. Recall notations $\mathcal{T}_{\tilde{\theta},w}$ and $\tilde{\mathcal{T}}_{\tilde{\theta},w}$ with $\tilde{\theta}$ a $\rho^{-\frac{1}{2}}$-cap and $w \in R^{\frac{1+\delta}{2}} \mathbb{Z}^{n-1}$ for sortings, which relate $\rho$-wave packets on $B_j$ with the large $R$-wave packets. For the sake of brevity let $g = f_{j,trans}$. We define as in \cite[p.~25]{OuWang2021}:
\begin{equation*}
g_{ess} = \sum_{(\tilde{\theta},w) \in \mathcal{T}_{ess}} g_{\tilde{\theta},w} = g - \sum_{(\tilde{\theta},w) \in \mathcal{T}_{tail}} g_{\tilde{\theta},w},
\end{equation*}
where
\begin{equation*}
\begin{split}
\mathcal{T}_{ess} &= \{ (\tilde{\theta},w) : \, \exists (\theta,v) \in \mathcal{T}_{\tilde{\theta},w} \text{ so that } T_{\theta,v} \cap X_{II} \neq \emptyset \}, \\
\mathcal{T}_{tail} &= \{ (\tilde{\theta},w) : \, \forall (\theta,v) \in \mathcal{T}_{\tilde{\theta},w} : T_{\theta,v} \cap X_{II} = \emptyset \}.
\end{split}
\end{equation*}
Like in \cite{OuWang2021}, we infer that
\begin{equation}
\label{eq:EssentialContribution}
\| T^\lambda g \|_{BL^p_{k,A}(B_j)} \leq \| T^\lambda g_{ess} \|_{BL^p_{k,A/2}(B_j)} + \text{RapDec}(R) \| f \|_{L^2}.
\end{equation}
As in \cite{GuthHickmanIliopoulou2019}, we choose a set of translates $\mathcal{B}$, so that we can write
\begin{equation}
\label{eq:TransverseSubcaseKeyEstimateII}
\| T^\lambda g_{ess} \|_{BL^p_{k,A/2}(B_j)} \lesssim \big( \sum_{b \in \mathcal{B}} \| T^\lambda g_{ess,b} \|^p_{BL^p_{k,A/2}(B_j)} \big)^{\frac{1}{p}},
\end{equation}
where each piece $g_{ess,b}$ is defined so that it is concentrated on scale $\rho$ wave packets, which are tangential to some translate $Z-y+b$ of $Z$. At this point, we can use transverse equidistribution and infer that $g_{ess,b}$ satisfy favorable $L^2$-estimates. Moreover, the radial induction hypothesis is applied to each of the $T^\lambda g_{ess,b}$. To close the induction, one must estimate
\begin{equation*}
\big( \sum_{b \in \mathcal{B}} \| g_{ess,b} \|^p_{L^2} \big)^{\frac{1}{p}}
\end{equation*}
in terms of $\| g_{ess} \|_{L^2}$. The gain in $(\rho/R)$ stemming from transverse equidistribution is crucial. We can sum the contributions from the individual pieces $g_{ess,b}$ without any (significant) loss in $R$.
To ensure that $g_{ess,b}$ form a reasonable decomposition of $g_{ess}$ so that \eqref{eq:TransverseSubcaseKeyEstimateII} holds up to logarithmic factors, the set of translates $\mathcal{B}$ must be chosen so that $\bigcup_{b \in \mathcal{B}} N_{\rho^{\frac{1}{2} + \delta_m}}(Z-y+b)$ covers $N_{R^{\frac{1}{2}+\delta_m}}(Z)$ (where the mass of $T^\lambda g_{ess,b}$ is concentrated) and so that the $N_{\rho^{\frac{1}{2}+\delta_m}}(Z-y+b)$ are essentially disjoint. This was achieved in \cite{GuthHickmanIliopoulou2019} using a probabilistic construction: Fix $B=B(y,\rho) \in \mathcal{B}_\rho$, one may show the following:
\begin{lemma}[{\cite[Lemma~10.5]{GuthHickmanIliopoulou2019}}]
\label{lem:ProbabilisticLemma}
There exists a finite set $\mathcal{B} \subseteq B(0,2 R^{\frac{1}{2}+\delta_m})$ and a collection
\begin{equation*}
\mathcal{B}' \subseteq \{ B_{K^2} \in \mathcal{B}_{K^2} : B_{K^2} \cap B(y,\rho) \neq \emptyset \}
\end{equation*}
such that, up to inclusion of a rapidly decreasing error term,
\begin{equation}
\label{eq:TranslateEstimate}
\| T^\lambda f_{B,trans} \|_{BL^p_{k,A/2}(B)} \lesssim (\log R)^2 \big( \sum_{B_{K^2} \in \mathcal{B}'} \mu_{T^\lambda f_{B,trans}}(B_{K^2}) \big)^{\frac{1}{p}}
\end{equation}
and for each $B_{K^2} \in \mathcal{B}'$ the following holds:
\begin{itemize}
\item[(i)] there exists some $b \in \mathcal{B}$ such that
\begin{equation}
\label{eq:ContainmentKBall}
B_{K^2} \subseteq N_{\rho^{\frac{1}{2}+\delta_m}/2}(Z+b);
\end{equation}
\item[(ii)] there exist at most $O(1)$ vectors $b \in \mathcal{B}$ for which
\begin{equation*}
B_{K^2} \cap N_{\rho^{\frac{1}{2} + \delta_m}}(Z+b) \neq \emptyset.
\end{equation*}
\end{itemize}
\end{lemma}
By the lemma, we may argue as follows: For each $b \in \mathcal{B}$, let $\mathcal{B}_b'$ denote the collection of all $B_{K^2} \in \mathcal{B}'$ for which \eqref{eq:ContainmentKBall} holds. Then, by \eqref{eq:TranslateEstimate} and property (i) in the lemma,
\begin{equation*}
\| T^\lambda f_{B,trans} \|_{BL^p_{k,A/2}(B)} \lesssim (\log R)^2 \big( \sum_{b \in \mathcal{B}} \sum_{B_{K^2} \in \mathcal{B}_b'} \mu_{\tilde{T}^\lambda (f_{B,trans}) \tilde{\,}}(B_{K^2} - y) \big)^{\frac{1}{p}}.
\end{equation*}
Define the collection of wave packets
\begin{equation*}
\tilde{\mathbb{T}}_b' = \{ (\tilde{\theta},\tilde{v} ) \in \bigcup_{(\theta,v) \in \mathbb{T}_{ess}} \tilde{\mathbb{T}}_{\theta,v} : \tilde{T}_{\tilde{\theta},\tilde{v}} \cap \big( \bigcup_{B_{K^2} \in \mathcal{B}_b'} (B_{K^2} - y) \big) \neq \emptyset \}.
\end{equation*}
If $g_{ess,b}$ is defined by
\begin{equation*}
(g_{ess,b}) \tilde{\,} = \sum_{(\tilde{\theta},\tilde{v}) \in \tilde{\mathbb{T}}_{b}'} (g_{ess})_{\tilde{\theta},\tilde{v}}^{\tilde{\,}},
\end{equation*}
then $(g_{ess,b}) \tilde{\,}$ is concentrated on wave packets that are $\rho^{-\frac{1}{2}+\delta_m}$-tangent to $Z-y+b$. Furthermore, again up to a rapidly decreasing error term, one has
\begin{equation*}
\| T^\lambda g_{ess,b} \|_{BL^p_{k,A/4}(B_j)} \lesssim (\log R)^2 \big( \sum_{b \in \mathcal{B}} \| \tilde{T}^\lambda (g_{ess,b}) \tilde{\,} \|^p_{BL^p_{k,A/4}(B(0,\rho))} \big)^{\frac{1}{p}}.
\end{equation*}
The function $(g_{ess,b}) \tilde{\,}$ satisfies the hypotheses of Theorem \ref{thm:MainInductionTheorem} at scale $\rho$ and therefore the radial induction hypothesis yields
\begin{equation*}
\big( \sum_{b \in \mathcal{B}} \| \tilde{T}^\lambda (g_{ess,b}) \tilde \|^p_{BL^p_{k,A/4}(B(0,\rho))} \big)^{\frac{1}{p}} \leq E_{m,A/4}(\rho) \big( \sum_{b \in \mathcal{B}} \| g_{ess,b} \|^p_{L^2} \big)^{\frac{1}{p}}.
\end{equation*}
We claim that
\begin{equation}
\label{eq:ImprovedTransverseSummation}
\big( \sum_{b \in \mathcal{B}} \| g_{ess,b} \|^p_{L^2} \big)^{\frac{1}{p}} \lesssim R^{O(\delta_m)} \big( \frac{\rho}{R} \big)^{(n-m) \big( \frac{1}{4} - \frac{1}{2p} \big)} \| g_{ess} \|_{L^2}.
\end{equation}
We show this via interpolation between $p=2$ and $p=\infty$.
For $p=2$ this follows from orthogonality of the wave packets and property (ii) of Lemma \ref{lem:ProbabilisticLemma}.
\medskip
For $p=\infty$ we use transverse equidistribution. By almost orthogonality of $(\tilde{\theta},w) \in \mathcal{T}$ and the definition of $\mathcal{T}_{ess}$ we have
\begin{equation*}
\| \tilde{g}_{ess,b} \|^2_{L^2} \sim \sum_{(\tilde{\theta},w) \in \mathcal{T}_{ess}} \| \tilde{g}_{ess,b,\tilde{\theta},w} \|^2_{L^2}.
\end{equation*}
By construction of $g_{ess,b,\tilde{\theta},w}$ there is $(\theta,v) \in \mathcal{T}_{\tilde{\theta},w}$ such that $T_{\theta,v}$ for $(\theta,v)$ intersects $X_{II}$. Let $B= B(\bar{x};R^{\frac{1}{2}+\delta_m})$ denote the corresponding ball in $X_{II}$. Since the Hausdorff distance between $T_{\theta_1,v_1}$ for further $(\theta_1,v_1) \in \mathcal{T}_{\tilde{\theta},w}$ is $\lesssim R^{\frac{1}{2}+\delta}$, we can apply Lemma \ref{lem:ReverseL2Bound} at scale $\rho$ with $r \sim R^{\frac{1}{2}+\delta_m}$ to find that
\begin{equation}
\label{eq:InverseHoermanderApplication}
\| \tilde{g}_{ess,b,\tilde{\theta},w} \|_{L^2}^2 \lesssim R^{-\frac{1}{2}-\delta_m} \| \tilde{T}_*^\lambda \tilde{g}_{ess,b,\tilde{\theta},w} \|_{L^2(10B)}^2.
\end{equation}
Next, we can apply Lemma \ref{lem:TransverseEquidistributionEstimate} to find that
\begin{equation}
\label{eq:TransverseEquidistributionApplication}
\begin{split}
\| T^\lambda g_{ess,b,\tilde{\theta},w} \|^2_{L^2(10B \cap N_{\rho^{\frac{1}{2}+\delta_m}}(Z+b))} &\lesssim R^{\frac{1}{2}+O(\delta_m)} \big( \frac{\rho}{R} \big)^{\frac{n-m}{2}} \| g_{ess,b,\tilde{\theta},w} \|^2_{L^2} \\
&\lesssim R^{\frac{1}{2}+O(\delta_m)} \big( \frac{\rho}{R} \big)^{\frac{n-m}{2}} \| g_{ess,b} \|^2_{L^2}.
\end{split}
\end{equation}
Taking \eqref{eq:InverseHoermanderApplication} and \eqref{eq:TransverseEquidistributionApplication} together, we find
\begin{equation*}
\| \tilde{g}_{ess,b,\tilde{\theta},w} \|_{L^2}^2 \lesssim R^{O(\delta_m)} \big( \frac{\rho}{R} \big)^{\frac{n-m}{2}} \| g_{ess,b} \|^2_{L^2},
\end{equation*}
which is the claimed $p=\infty$ estimate for \eqref{eq:ImprovedTransverseSummation}.
\medskip
At this point, the computation to close the induction follows \cite[p.~351]{GuthHickmanIliopoulou2019} verbatim. The proof of Theorem \ref{thm:MainInductionTheorem} is complete.
$\hfill \Box$
\section{From $k$-broad to linear estimates}
\label{section:LinearEstimates}
In this section we deduce the linear estimates from the $k$-broad estimates by applying the Bourgain--Guth argument \cite{BourgainGuth2011}. We show the following proposition:
\begin{proposition}
\label{prop:LinearFromBroadEstimates}
Suppose that for all $K \geq 1$ and all $\varepsilon > 0$ any oscillatory integral operator $T^\lambda$ with reduced $1$-homogeneous phase satisfying $C1)$ and $C2^+)$ obeys the $k$-broad inequality
\begin{equation}
\label{eq:HypothesisKBroad}
\| T^\lambda f \|_{BL^p_{k,A}(B(0,R))} \lesssim_\varepsilon K^{C_\varepsilon} R^\varepsilon \| f \|_{L^p(A^{n-1})}
\end{equation}
for some fixed $k$, $A$, $p$, $C_\varepsilon$, and all $R \geq 1$. If
\begin{equation}
\label{eq:ConditionsP}
p(k,n) \leq p \leq \frac{2n}{n-2}, \qquad p(k,n) =
\begin{cases}
2 \cdot \frac{n-1}{n-2} \quad \text{ if } 2 \leq k \leq 3, \\
2 \cdot \frac{2n-k+1}{2n-k-1} \quad \text{ if } k > 3,
\end{cases}
\end{equation}
then any oscillatory integral operator with $C1)$ and $C2^+)$ phase $\phi$ and amplitude $a$ satisfies
\begin{equation}
\label{eq:LinearEstimateFromBroad}
\| T^\lambda f \|_{L^p(\mathbb{R}^n)} \lesssim_{\phi,\varepsilon,a} \lambda^\varepsilon \| f \|_{L^p(A^{n-1})}.
\end{equation}
\end{proposition}
From this proposition Theorem \ref{thm:LpLpEstimatesVariableCoefficients} is immediate by choosing $k = \frac{n+1}{2}$ for $n$ odd and $k = \frac{n}{2} + 1$ for $n$ even as $\max(p(k,n),\bar{p}(k,n))$ gives the lower bound for $p$ in Theorem \ref{thm:LpLpEstimatesVariableCoefficients}. For the proof we use induction on scales: $Q_{p,\delta}(R)$ will denote the infimum over all constants $C$ for which the estimate
\begin{equation*}
\| T^\lambda f \|_{L^p(B(0,r))} \leq C \| f \|_{L^p(A^{n-1})}
\end{equation*}
holds for $1 \leq r \leq R$ and all oscillatory integral operators built from a suitable class of phase functions, which is invariant under rescaling and amenable to narrow decoupling, which is explained below.
With this definition, it remains to prove that for $p$ as in Proposition \ref{prop:LinearFromBroadEstimates}
\begin{equation*}
Q_{p,\delta}(R) \lesssim_\varepsilon R^\varepsilon
\end{equation*}
for all $\varepsilon >0$ and $1 \leq R\leq \lambda$.\\
For this purpose, we decompose $B(0,R)$ into finitely overlapping balls $B_{K^2}$ of radius $K^2$ and estimate $\| T^\lambda f \|_{L^p(B_{K^2})}$. $f$ is decomposed into ``broad" and ``narrow" term. The narrow term is of the form
\begin{equation}
\label{eq:NarrowTerm}
\sum_{\substack{\tau \in V_a \\ \text{for some } a}} f_\tau,
\end{equation}
consisting of contributions to $f$ from sectors for which $G^\lambda(\bar{x};\tau)$ makes a small angle with some member of a family of $(k-1)$-planes. Here $\bar{x}$ denotes the centre of $B_{K^2}$. The broad term consists of contributions to $f$ from the remaining sectors. One may choose the planes $V_1, \ldots , V_A$ so that the broad term can be bounded by the $k$-broad inequality. Thus, $f$ of the form \eqref{eq:NarrowTerm} has to be analyzed. This is accomplished by narrow $\ell^p$-decoupling and rescaling. We use the following decoupling result:
\begin{proposition}[Variable coefficient decoupling]
\label{prop:DecouplingNarrowTerm}
Suppose that $T^\lambda$ is an oscillatory integral operator with reduced $C1)$ and $C2^+)$ phase, which is $K$-flat and let $B_{K^2} \subseteq \lambda^{1-\delta}$ with $1 \leq K^2 \leq \lambda^{\frac{1}{2}-\delta}$, $0<\delta \leq 1/2$. If $V \subseteq \mathbb{R}^n$ is an $m$-dimensional linear subspace, then for $2 \leq p \leq \frac{2n}{n-2}$ and any $\delta>0$ one has
\begin{equation*}
\begin{split}
\big\| \sum_{\tau \in V} T^\lambda g_\tau \big\|_{L^p(B_{K^2})} &\lesssim_\delta \max(1, K^{(m-2)\big( \frac{1}{2}- \frac{1}{p} \big)}) K^\delta \big( \sum_{\tau \in V} \| T^\lambda g_\tau \|_{L^p(w_{B_{K^2}})}^p \big)^{\frac{1}{p}} \\
&\quad + \lambda^{-\frac{\delta N}{2}} \| g \|_{L^2}.
\end{split}
\end{equation*}
Here, the sum ranges over sectors $\tau$ for which $\angle(G^\lambda(\bar{x};\tau),V) \leq K^{-2}$, where $\bar{x}$ is the centre of $B_{K^2}$ and $w_{B_{K^2}} = (1+|x-\bar{x}|)^{-N}$ is a rapidly decaying weight off $B_{K^2}$ with $N$ the same as in the notion of $K$-flatness.
\end{proposition}
We remark that on the right-hand side there are strictly speaking slightly different amplitude functions involved. If we choose $B_{K^2} \subseteq B(0,\lambda^{1-\delta})$ for $\lambda$ large enough however, the amplitude functions satisfy the uniform bounds
\begin{equation*}
|\partial_\omega^\alpha a(x;\omega)| \lesssim_N 1
\end{equation*}
for $0 \leq |\alpha| \leq N$, $N$ being the parameter from $K$-flatness. This technicality of dealing with different amplitude functions is handled by appropriate definition of the induction quantity.
In the translation-invariant case, e.g., with $\mathcal{E}$ as in \eqref{eq:ConeExtensionOperator}, this follows from the $\ell^2$-decoupling
\begin{equation*}
\big\| \sum_\tau \mathcal{E} g_\tau \big\|_{L^p(B_{K^2})} \lesssim_\delta K^\delta \big( \sum_{\tau \in V} \| \mathcal{E} g_\tau \|^2_{L^p(w_{B_{K^2}})} \big)^{\frac{1}{2}}
\end{equation*}
for $2 \leq p \leq \frac{2n}{n-2}$ and by counting the sectors $\tau$ such that $\angle(G(\tau),V) \leq K^{-2}$. This is carried out in \cite{OuWang2021}; see also \cite[Lemma~2.2]{Harris2019}. The error term $\lambda^{-\frac{\delta N}{2}} \| f \|_{L^2}$ comes from approximation with constant coefficient operators. Gao \emph{et al.} \cite{GaoLiuMiaoXi2021} used $K$-flatness to prove narrow decoupling of general homogeneous phases in the constant coefficient case.
\begin{definition}
We say that a $1$-homogeneous smooth $\phi:\mathbb{R}^{n-1} \backslash 0 \to \mathbb{R}$ supported in $\Xi$ is $K$-flat if
\begin{equation*}
\begin{split}
\phi(\omega',\omega_{n-1}) &= \omega_{n-1} \phi(\omega'/\omega_{n-1},1) \\
&= \omega_{n-1} \phi(e_{n-1}) + \partial_{\omega'} \phi(e_{n-1}) \omega' + \frac{\langle \partial^2_{\omega' \omega'} \phi(e_{n-1}) \omega', \omega' \rangle}{2 \omega_{n-1}} + K^{-4} E(\omega)
\end{split}
\end{equation*}
with $E(\omega)$ $1$-homogeneous, satisfying $|\partial^\alpha E_R| \lesssim_\alpha 1$ for $0 \leq |\alpha| \leq N$.
\end{definition}
In the course of the argument, we will need to consider higher derivatives; above unspecified as $N$. These are needed for approximation with constant-coefficient operators. In the end, we choose $N=N(\varepsilon)$ (since $\delta = \delta(\varepsilon)$) large enough such that the error term $\lambda^{- \frac{\delta N}{2}} \| f \|_{L^2}$ propagates through the argument. Note that by comparison with Taylor's formula we have
\begin{equation*}
K^{-4} E(\omega) = \sum_{|\alpha| = 3} \frac{3}{\alpha !} \int_0^1 (1-s)^2 (\partial_{\omega'}^\alpha \phi)(\frac{s \omega'}{\omega_{n-1}},1) ds \frac{(\omega')^\alpha}{\omega_{n-1}^2}.
\end{equation*}
For these constant-coefficient operators, Harris's argument \cite[Lemma~2.2]{Harris2019} of sector counting applies. To apply narrow decoupling for the variable coefficient operator on a small $K^2$-ball with $K^2 \lesssim \lambda^{\frac{1}{2}-\varepsilon}$, we approximate the variable coefficient phase with a constant coefficient phase. Beltran--Hickman--Sogge \cite{BeltranHickmanSogge2020} worked out that this is possible by Taylor expansion.\\
We need the following notations: Let $\phi$ be a reduced phase and $\bar{x} \in \mathbb{R}^n$, which will be the centre of the small ball on which we want to apply decoupling. Recall that $u \mapsto \partial_x \phi^\lambda(\bar{x};\Psi^\lambda(\bar{x};u))$ is a graph parametrization of the hypersurface $\Sigma_{\bar{u}}$. We have
\begin{equation*}
\langle x, (\partial_x \phi^\lambda) (\bar{x};\Psi^\lambda(\bar{x};u)) \rangle = \langle x',u \rangle + x_n h_{\bar{x}}(u)
\end{equation*}
for all $x = (x',x_n) \in \mathbb{R}^n$ with $h_{\bar{z}}(u) = (\partial_{x_n} \phi^\lambda) (\bar{z};\Psi^\lambda(\bar{z};u))$. We suppose for technical reasons that $a(x;\omega) = a_1(x) a_2(\omega)$; the general case is reduced to this by Fourier series expansion. Let $E_{\bar{x}}$ denote the extension operator associated to $\Sigma_{\bar{x}}$, given by
\begin{equation*}
E_{\bar{x}} g(x) = \int_{\mathbb{R}^{n-1}} e^{i(\langle x',u \rangle + x_n h_{\bar{x}}(u))} a_{\bar{x}}(u) g(u) du \text{ for all } x \in \mathbb{R}^n,
\end{equation*}
where $a_{\bar{x}}(u) = a_2 \circ \Psi^\lambda(\bar{x};u) |\det \partial_u \Psi^\lambda(\bar{x};u)|$. We recall how $T^\lambda$ is approximated by $E_{\bar{x}}$: Let $x \in B(\bar{x};K^2) \subseteq B(0,3\lambda/4)$. By change of variables $\omega = \Psi^\lambda(\bar{x};u)$ and a Taylor expansion of $\phi^\lambda$ around $\bar{x}$, we have
\begin{equation*}
T^\lambda f(x) = \int_{\mathbb{R}^{n-1}} e^{i( \langle x-\bar{x}, (\partial_x \phi^\lambda)(\bar{x};\Psi^\lambda(\bar{z};u)) \rangle + \mathcal{E}^\lambda_{\bar{x}}(x-\bar{x};u))} a_1^\lambda(x) a_{\bar{z}}(u) f_{\bar{z}}(u) du
\end{equation*}
with $f_{\bar{x}} = e^{i \phi^\lambda(\bar{x};\Psi^\lambda(\bar{x};\cdot))} f \circ \Psi^\lambda(\bar{x};\cdot)$ and by Taylor expansion
\begin{equation*}
\mathcal{E}^\lambda_{\bar{x}}(v;u) = \frac{1}{\lambda} \int_0^1 (1-r) \langle (\partial_{xx} \phi)((\bar{x}+rv)/\lambda;\Psi^\lambda(\bar{x};u)) v, v \rangle dr.
\end{equation*}
By the derivative bounds
\begin{equation*}
\sup_{(v;u) \in B(0,K^2) \times \text{supp} a_{\bar{x}}} | \partial^\beta_\omega \mathcal{E}_{\bar{x}}^\lambda(v;u) | \lesssim_N 1
\end{equation*}
and Fourier series expansion, the oscillation of $\mathcal{E}^\lambda_{\bar{x}}$ can be neglected. This yields the following lemma:
\begin{lemma}[{\cite[Lemma~2.6]{BeltranHickmanSogge2020}}]
Let $T^\lambda$ be an oscillatory integral operator with reduced $C1)$ and $C2^+)$ phase. Let $0 < \delta \leq 1/2$, $1 \leq K^2 \leq \lambda^{\frac{1}{2}-\delta}$ and $\bar{x}/\lambda \in X$ so that $B(\bar{x};K^2) \subseteq B(0,3\lambda/4)$.
\begin{itemize}
\item Then
\begin{equation}
\label{eq:LinearApproximationI}
\| T^\lambda f \|_{L^p(w_{B(\bar{x};K^2)})} \lesssim_N \| E_{\bar{x}} f_{\bar{x}} \|_{L^p(w_{B(0;K^2)})} + \lambda^{-\frac{\delta N}{2}} \| f \|_{L^2}
\end{equation}
holds provided that $N$ is sufficiently large depending on $n$, $\delta$, and $p$.
\item Suppose that $|\bar{x}| \leq \lambda^{1-\delta'}$. There exists a family of operators $\mathbf{T}^\lambda$ all with phase $\phi$ and of type $(1,1,C)$ data such that
\begin{equation}
\label{eq:LinearApproximationII}
\| E_{\bar{x}} f_{\bar{x}} \|_{L^p(w_{B(0,K^2)})} \lesssim_N \| T^\lambda_* f \|_{L^p(w_{B(\bar{x};K^2)})} + \lambda^{- \frac{N \min(\delta,\delta')}{2}} \| f \|_2
\end{equation}
holds for some $T^\lambda_* \in \mathbf{T}^\lambda$. The family $\mathbf{T}^\lambda$ has cardinality $O_N(1)$ and is independent of $B(\bar{x};K^2)$.
\end{itemize}
\end{lemma}
To apply the narrow decoupling to $E_{\bar{x}} f_{\bar{x}}$, we need that the constant coefficient phase
\begin{equation*}
h_{\bar{x}}(u) = \partial_{x_n} \phi^\lambda(\bar{x};\Psi^\lambda(\bar{x};u))
\end{equation*}
is $K$-flat.
\begin{definition}
Let $K \gg 1$. We say that a reduced homogeneous phase $\phi: \mathbb{R}^n \times \mathbb{R}^{n-1} \backslash 0 \to \mathbb{R}$ is $K$-flat, if all its constant-coefficient approximations $h_{\bar{x}}$ are $K$-flat and
\begin{equation*}
\begin{split}
|\partial_{x'} \partial^\alpha_{\xi} \phi | &\lesssim K^{-4} \qquad 2 \leq |\alpha'| \leq N, \\
|\partial_{x_n} \partial_\xi^\alpha \phi | &\lesssim K^{-4} \qquad 3 \leq |\alpha'| \leq N.
\end{split}
\end{equation*}
\end{definition}
The derivative bounds are required to control a change of variables in frequencies. We remark that with this definition, Proposition \ref{prop:DecouplingNarrowTerm} now follows from the constant-coefficient decoupling and the approximation by constant-coefficient operators provided by the previous lemma.
We can give the definition of the inductive quantity now:
\begin{definition}
For $1 \leq p \leq \infty$ and $R \geq 1$ let $Q_{p,\delta}(R)$ denote the infimum over all constants $C$ for which the estimate
\begin{equation*}
\| T^\lambda f \|_{L^p(B(0,r))} \leq C \| f \|_{L^p(A^{n-1})}
\end{equation*}
holds for $1 \leq r \leq R$ and all oscillatory integral operators $T^\lambda$ with reduced $C1)$ and $C2^+)$ $1$-homogeneous phase, which is $\lambda^{\delta}$-flat, and all $\lambda \geq R$. Furthermore, we require estimates
\begin{equation*}
|\partial_\omega^\alpha a(x;\omega)| \lesssim_N 1
\end{equation*}
for the amplitude function with $0 \leq |\alpha| \leq N$.
\end{definition}
Before we turn to the parabolic rescaling, note that by homogeneity,
\begin{equation}
\partial_{x'} \phi(x,\omega) = \sum_{j=1}^{n-1} \omega_j \cdot \partial_{\omega_j} \partial_{x^\prime} \phi(x,\omega).
\end{equation}
Thus, for each $t \in (-1,1)$ and $\omega \in \mathbb{R}^{n-1}$ the Jacobian determinant of the map $x' \mapsto ((\partial_{\omega^\prime} \phi)(x;\omega), \phi(x;\omega))$ is given by $\omega_{n-1} \cdot \det \partial^2_{\omega x^\prime} \phi(x;\omega)$ and hence, non-vanishing. Let $x = (x'',x_{n-1},x_n) \in \mathbb{R}^{n-2} \times \mathbb{R} \times \mathbb{R}$. The implicit function theorem yields a smooth local inverse mapping $\Upsilon(\cdot,x_n;\omega)$, which satisfies
\begin{equation*}
\partial_\omega \phi(\Upsilon(x^\prime,x_n;\omega),x_n;\omega) = x^{\prime \prime} \text{ and } \phi(\Upsilon(x;\omega),x_n;\omega) = x_{n-1}.
\end{equation*}
\begin{lemma}[Parabolic rescaling]
\label{lem:ParabolicRescaling}
Let $\text{supp} (f) \subseteq \Xi$ be supported in a $\rho^{-1}$-plate and $\phi$ be a reduced phase, that is $\lambda^{\delta}$-flat. Then, for any $1 \leq \rho \leq R \leq \lambda$:
\begin{equation}
\label{eq:LorentzRescaling}
\| T^\lambda f \|_{L^p(B(0,R))} \lesssim_{\delta'} R^{\delta'} Q_{p,\delta}(R/\rho) \rho^{\frac{2(n-1)}{p}-(n-2)} \| f \|_{L^p}.
\end{equation}
\end{lemma}
The proof combines arguments from \cite{BeltranHickmanSogge2020} and \cite{GuthHickmanIliopoulou2019}. In \cite{BeltranHickmanSogge2020} the phase after parabolic rescaling was computed, and it was shown how after rescaling we find the bounds for higher derivatives introduced in Section \ref{subsection:BasicReductions} to hold, even for arbitrary phases. We shall also see that these phases are $\lambda^\delta$-flat. Since we need expressions from the computations in \cite{BeltranHickmanSogge2020}, some details are repeated.
\begin{proof}
Let $\omega \in B_{n-2}(0,1)$ with $(\omega,1)$ the centre of the $\rho^{-1}$-plate encasing the support of $g$:
\begin{equation*}
\text{supp} (g) \subseteq \{ (\xi',\xi_{n-1}) \in \mathbb{R}^{n-1} : \, 1/2 \leq \xi_{n-1} \leq 2 \text{ and } \big| \frac{\xi'}{\xi_{n-1}} - \omega \big| \leq \rho^{-1} \}.
\end{equation*}
We perform the change of variables
\begin{equation*}
(\xi',\xi_{n-1}) = (\eta_{n-1} \omega + \rho^{-1} \eta', \eta_{n-1}),
\end{equation*}
after which follows
\begin{equation*}
T^\lambda g(x) = \int_{\mathbb{R}^{n-1}} e^{i \phi^\lambda(x;\eta_{n-1} \omega + \rho^{-1} \eta', \eta_{n-1})} a^\lambda(x;\eta_{n-1} \omega + \rho^{-1} \eta', \eta_{n-1}) \tilde{g}(\eta) d\eta,
\end{equation*}
where $\tilde{g}(\eta) = \rho^{-(n-2)} g(\eta_{n-1} \omega + \rho^{-1} \eta',\eta_{n-1})$ and $\text{supp} (\tilde{g}) \subseteq \Xi$. By Taylor expansion and homogeneity of the phase, we find
\begin{equation*}
\begin{split}
\phi(x;\eta_{n-1} \omega + \rho^{-1} \eta', \eta_{n-1}) &= \phi(x;\omega,1) \eta_{n-1} + \rho^{-1} \langle \partial_{\omega'} \phi (x;\omega,1), \eta' \rangle \\
&\quad + \rho^{-2} \int_0^1 (1-r) \langle \partial^2_{\omega' \omega'} \phi(x;\eta_{n-1} \omega + r \rho^{-1} \eta', \eta_{n-1}) \eta', \eta' \rangle dr.
\end{split}
\end{equation*}
Let $\Upsilon_\omega(y',y_n) = (\Upsilon(y',y_n;\omega,1),y_{n-1})$ and $\Upsilon^\lambda_\omega(y',y_n) = \lambda \Upsilon_\omega(y'/\lambda,y_n/\lambda)$ and consider anisotropic dilations
\begin{equation*}
D_\rho( y'', y_{n-1}, y_n) = (\rho y'',y_{n-1}, \rho^2 y_n) \text{ and } D'_{\rho^{-1}}(y'',y_{n-1}) = (\rho^{-1} y'', \rho^{-2} y_{n-1})
\end{equation*}
on $\mathbb{R}^n$ and $\mathbb{R}^{n-1}$, respectively. By definition of $\Upsilon$, we find
\begin{equation*}
T^\lambda g \circ \Upsilon^\lambda_\omega \circ D_\rho = \tilde{T}^{\lambda/\rho^2} \tilde{g}
\end{equation*}
where
\begin{equation*}
\tilde{T}^{\lambda/\rho^2} \tilde{g}(y) = \int_{\mathbb{R}^{n-1}} e^{i \tilde{\phi}^{\lambda/\rho^2}(y;\eta)} \tilde{a}^\lambda(y;\eta) \tilde{g}(\eta) d\eta
\end{equation*}
for the phase $\tilde{\phi}(y;\eta)$ given by
\begin{equation*}
\langle y',\eta \rangle + \int_0^1 (1-r) \langle \partial^2_{\xi' \xi'} \phi(\Upsilon_\omega(D'_{\rho^{-1}} y',y_{n}); \eta_{n-1} \omega + r \rho^{-1} \eta', \eta_{n-1}) \eta', \eta' \rangle dr
\end{equation*}
and the amplitude $\tilde{a}(y;\eta) = a(\Upsilon_\omega(D'_{\rho^{-1}}y',y_n);\eta_{n-1} \omega + \rho^{-1} \eta', \eta_{n-1})$. By a change of spatial variables, we find
\begin{equation*}
\| T^\lambda g \|_{L^p(B_R)} \lesssim \rho^{\frac{n}{p}} \| \tilde{T}^{\lambda/\rho^2} \tilde{g} \|_{L^p((\Upsilon^\lambda_\omega \circ D_\rho)^{-1}(B_R))}.
\end{equation*}
We want to apply the induction hypothesis at scale $R/\rho^2$. First, we make a harmless linear change of variables: Let $L \in GL(n-1;\mathbb{R})$ be such that $L e_{n-1} = e_{n-1}$ and
\begin{equation*}
\partial^2_{\eta' \eta'} \partial_{y_n} \tilde{\phi}_L (0,0;e_{n-1}) = I_{n-1},
\end{equation*}
where
\begin{equation*}
\tilde{\phi}_L(y;\eta) = \tilde{\phi}(L^{-1} y',y_n; L \eta).
\end{equation*}
It suffices to analyze $\tilde{T}^{\lambda/\rho^2}_L \tilde{g}_L$ with $\tilde{T}^{\lambda/\rho^2}_L$ defined with respect to the datum $(\tilde{\phi}_L,\tilde{a}_L)$ for $\tilde{\phi}_L$ as above, $\tilde{a}_L(y;\eta) = \tilde{a}(L^{-1} y',y_n;L \eta)$ and $\tilde{g}_L = | \det L| \tilde{g} \circ L$.\\
To see that $\tilde{\phi}_L$ is still a reduced phase, note the representations
\begin{equation*}
\tilde{\phi}_L(y;\eta) = \rho^2 \phi(\Upsilon_\omega(D'_{\rho^{-1}} \circ L^{-1} y', y_n), y_n; \eta_n \omega + \rho^{-1} L' \eta', \eta_n)
\end{equation*}
and
\begin{equation*}
\langle y',\eta \rangle + \int_0^1 (1-r) \langle \partial^2_{\omega' \omega'} \phi(\Upsilon_\omega(D'_{\rho^{-1}} \circ L^{-1} y',y_n);\eta_{n-1} \omega + r \rho^{-1} L' \eta', \eta_{n-1}) L' \eta', L' \eta' \rangle dr,
\end{equation*}
where $L'$ denotes the $(n-2) \times (n-2)$-submatrix of $L$, containing the first $n-2$ rows and columns. In \cite{BeltranHickmanSogge2020} was then shown that, starting with a reduced phase $\phi$, that $\tilde{\phi}_L$ is again a reduced phase. For sake of simplicity, suppose that $L=1$ in the following as taking derivatives only gives additional components of $L$. For reduced phase functions the components are bounded. We still have to show that it is still $\lambda^{\delta}$-flat:
Consider the formula
\begin{equation*}
\tilde{\phi} = \langle y',\eta \rangle + \int_0^1 (1-r) \langle \partial^2_{\omega' \omega'} \phi(\Upsilon_\omega(y',y_n); \eta_{n-1} \omega + r \rho^{-1} \eta', \eta_{n-1}) \eta', \eta' \rangle dr
\end{equation*}
Hence, we find
\begin{equation}
\label{eq:ExtensionOperatorI}
\partial_{x_n} \tilde{\phi} = \int_0^1 (1-r) \sum_{i,j,k} \partial^2_{\omega_j' \omega'_k} \partial_{x_i} \phi(\Upsilon_\omega(D'_{\rho^{-1}} y',y_n); \eta_{n-1} \omega + r \rho^{-1} \eta', \eta_{n-1}) \cdot \frac{\partial \Upsilon_\omega^i}{\partial x_n} (\eta'_j) (\eta'_k) dr
\end{equation}
and for $\Psi(\bar{x};u) = u$, it is straight-forward from taking additional derivatives in $\eta'$ that the resulting extension operator is indeed $\lambda^\delta$-flat. Next, we consider
\begin{equation*}
h_{\bar{x}}(u) = \partial_{x_n} \tilde{\phi}(\bar{x};\Psi(\bar{x};u)).
\end{equation*}
By definition of $\Psi(\bar{x};u)=(\Psi'(\bar{x};u),\Psi_{n-1}(\bar{x};u))$, we find
\begin{equation}
\label{eq:ErrorChangeFrequencies}
\begin{split}
u_i &= \partial_{x_i'} \tilde{\phi}(\bar{x};\Psi(\bar{x};u)) \\
&= \Psi'_i(\bar{x};u) + \rho^{-1} \int_0^1 (1-r) \sum_{\ell,j,k} \partial_{x_{\ell}'} \partial^2_{\omega'_j \omega_k'} \phi(\Upsilon_\omega(D'_{\rho^{-1}} y',y_n); \\
&\quad \Psi_{n-1} \omega + r \rho^{-1} \Psi'(\bar{x};u) \Psi_{n-1}(\bar{x};u)) \frac{\partial \Upsilon_\omega^\ell}{\partial_{x_i'}}(D'_{\rho^{-1}} y',y_n) \Psi_j'(\bar{x};u) \Psi_k'(\bar{x};u) dr.
\end{split}
\end{equation}
We find
\begin{equation*}
\Psi(\bar{x};u) = u + f(u), \quad |f(u)| \lesssim \rho^{-1}.
\end{equation*}
This also yields bounds for the derivatives of $f$ in \eqref{eq:ErrorChangeFrequencies} and taking the bounds of $\Psi$ and we see that $h_{\bar{x}}(u)$ is $\rho^{-1}$-flat. The argument also shows that, if $\phi$ was already $\tilde{\rho}^{-1}$-flat, in particular, $|\partial_{x'} \partial^2_{\omega' \omega'} \phi| \lesssim \rho^{-1}$ and bounds for higher derivatives, then $\tilde{\phi}$ is $\rho^{-1} \tilde{\rho}^{-1}$-flat. This matches the heuristic that rescaling makes the phase more resemble the translation-invariant case.
Hence, it suffices to show that
\begin{equation*}
\| T^\lambda f \|_{L^p(D_{\textbf{R}})} \lesssim_{\delta'} Q_{p,\delta}(R) R^{\delta'} \| f \|_{L^p}
\end{equation*}
for an ellipse
\begin{equation*}
D_{\textbf{R}} = \{ x \in \mathbb{R}^n : \big( \frac{|x'|}{R'} \big)^2 + \big( \frac{|x_n|}{R} \big)^2 \leq 1 \}
\end{equation*}
and an oscillatory integral operator with $\lambda^\delta$-flat phase. This can be argued as in \cite[Section~11.2]{GuthHickmanIliopoulou2019}.
\end{proof}
The narrow decoupling allows to separate the contribution of $T^\lambda f_\tau$ and it remains to estimate $\| T^\lambda f_\tau \|_{L^p(B_R)}$.
We are ready for the proof of Proposition \ref{prop:LinearFromBroadEstimates}:
\begin{proof}[Proof of Proposition \ref{prop:LinearFromBroadEstimates}]
It suffices to prove the linear estimate for $p$ satisfying the additional constraint
\begin{equation*}
p(k,n) < p
\end{equation*}
by interpolation. In the first step, for $\lambda \gg 1$, we carry out a parabolic rescaling depending on the phase such that it is enough to consider $\lambda^{\tilde{\delta}}$-flat phase functions. This loses a factor $C_\phi \lambda^{O(n) \tilde{\delta} }$ by partitioning $\Xi$ into sectors, which will be admissible provided that
\begin{equation}
\label{eq:InitialRescaling}
\lambda^{O(n) \tilde{\delta}} \leq \lambda^\varepsilon.
\end{equation}
\medskip
\noindent Let $\varepsilon > 0$. By the assumed $k$-broad estimate, we find
\begin{equation}
\label{eq:kBroadEstimateI}
\sum_{\substack{B_{K^2} \in \mathcal{B}_{K^2}, \\ B_{K^2} \cap B(0,R) \neq \emptyset}} \min_{V_1,\ldots,V_A} \max_{\tau \notin V_a} \int_{B_{K^2}} |T^\lambda f_\tau |^p \leq \tilde{C}_\varepsilon K^{C_\varepsilon} R^{\frac{p \varepsilon}{2}} \| f \|_{L^p(A^{n-1})}^p,
\end{equation}
where $V_1$,...,$V_A$ are $(k-1)$-planes and $\tau \notin V_a$ is short-hand for
\begin{equation*}
\angle (G^\lambda(\bar{x};\tau),V_a) > K^{-2},
\end{equation*}
with $\bar{x}$ being centre of $B_{K^2}$.\\
We choose $V_1$,...,$V_A$ for each $B_{K^2}$, which attains the minimum in \eqref{eq:kBroadEstimateI}. By this, we may write
\begin{equation*}
\int_{B_{K^2}} |T^\lambda f |^p \lesssim K^{O(n)} \max_{\tau \notin V_a} \int_{B_{K^2}} |T^\lambda f_\tau|^p + \sum_{a=1}^A \int_{B_{K^2}} \big| \sum_{\tau \in V_a} T^\lambda f_\tau \big|^p.
\end{equation*}
By summing over $B_{K^2}$ and using \eqref{eq:kBroadEstimateI}, we find
\begin{equation*}
\int_{B(0,R)} |T^\lambda f|^p \lesssim K^{O(n)} \tilde{C}_\varepsilon K^{C_\varepsilon} R^{p \varepsilon/2} \| f \|^p_{L^p} + \sum_{\substack{B_{K^2} \in \mathcal{B}_{K^2}, \\ B_{K^2} \cap B(0,R) \neq \emptyset}} \sum_{a=1}^A \int_{B_{K^2}} \big| \sum_{\tau \in V_a} T^\lambda f_\tau \big|^p.
\end{equation*}
By the decoupling result Proposition \ref{prop:DecouplingNarrowTerm}, we find for any $\delta' > 0$, provided that $K \leq \lambda^{\tilde{\delta}}$,
\begin{equation*}
\int_{B_{K^2}} \big| \sum_{\tau \in V_a} T^\lambda f_\tau \big|^p \lesssim_{\delta'} K^{\max((k-3)(\frac{p}{2}-1),0) + \delta'} \sum_{\tau \in V_a} \int_{\mathbb{R}^n} |T^\lambda f_\tau|^p w_{B_{K^2}}
\end{equation*}
and summing over $a$ and $B_{K^2}$, we find
\begin{equation}
\label{eq:NarrowAfterDecoupling}
\sum_{B_{K^2} \in \mathcal{B}_{K^2}} \sum_{a=1}^A \int_{B_{K^2}} \big| \sum_{\tau \in V_a} T^\lambda f_\tau \big|^p \lesssim_{\delta'} K^{\max((k-3)(p/2-1),0)+ \delta'} \sum_{\tau: K^{-1}} \int_{B(0,2R)} |T^\lambda f_\tau |^p.
\end{equation}
The separated expressions $T^\lambda f_\tau$ are amenable to Lemma \ref{lem:ParabolicRescaling} and an application gives
\begin{equation}
\label{eq:RescaledVersions}
\int_{B(0,2R)} |T^\lambda f_\tau|^p \lesssim_\delta (Q_{p,\tilde{\delta}}(R))^p R^\delta K^{2(n-1)-(n-2)p} \| f_\tau \|^p_{L^p}.
\end{equation}
Plugging \eqref{eq:RescaledVersions} into \eqref{eq:NarrowAfterDecoupling}, we find
\begin{equation*}
\int_{B(0,R)} |T^\lambda f|^p \leq (K^{O(n)} \tilde{C}_\varepsilon K^{C_\varepsilon} R^{p \varepsilon/2} + C_{\delta,\delta'} (Q_{p,\tilde{\delta}}(R))^p R^\delta K^{-e(k,p)+\delta'} ) \| f \|^p_{L^p(A^{n-1})}.
\end{equation*}
This yields
\begin{equation*}
(Q_{p,\tilde{\delta}}(R))^p \leq K^{O(n)} \tilde{C}_\varepsilon K^{C_\varepsilon} R^{p \varepsilon/2} + C_{\delta,\delta'} (Q_{p,\tilde{\delta}}(R))^p R^\delta K^{-e(k,p)+\delta'}.
\end{equation*}
Since $p$ is as in \eqref{eq:ConditionsP}, we find $e(k,p)>0$, and may choose $\delta' = e(k,p)/2$, so that the $K$ exponent in the second term on the right-hand side is negative. Moreover, we can choose $\delta$ small enough such that $\frac{2 \delta}{e(k,p)} C_\varepsilon \leq \frac{p \varepsilon}{8}$ and $\frac{O(n) \delta}{2 e(k,p)} \leq \frac{\varepsilon}{8}$. This ensures for the first term on the right-hand side:
\begin{equation*}
K^{O(n)} K^{C_\varepsilon} \leq \tilde{D}_\varepsilon R^{\frac{ 3p \varepsilon}{4}}.
\end{equation*}
Thus, if $K=K_0 R^{\frac{2 \delta}{e(k,p)}}$ for a sufficiently large $K_0$, depending on $\varepsilon$, $\delta=\delta(\varepsilon)$, $p$ and $n$, it follows that
\begin{equation*}
(Q_{p,\tilde{\delta}}(R))^p \leq \tilde{D}_\varepsilon R^{\frac{ 3p \varepsilon}{4}} + \frac{1}{2} (Q_{p,\tilde{\delta}}(R))^p.
\end{equation*}
By choosing $\tilde{\delta}= \frac{3 \delta}{e(k,p)}$ and $\lambda \geq E(\varepsilon)$ such that $\lambda^{\tilde{\delta}} \geq K_0 \lambda^{\frac{2 \delta}{e(k,p)}}$, the proof is complete because \eqref{eq:InitialRescaling} is ensured by $\frac{O(n) \delta}{2 e(k,p)} \leq \frac{\varepsilon}{8}$.
\end{proof}
\section{$\varepsilon$-removal away from the endpoint}
\label{section:epsRemoval}
In the following we prove the estimate
\begin{equation}
\label{eq:GlobalEstimate}
\| T^\lambda f \|_{L^p(\mathbb{R}^n)} \lesssim_{\phi,a} \| f \|_{L^p(A^{n-1})}
\end{equation}
for $p > p_n$ with $p_n$ defined in \eqref{eq:PolynomialPartitioningRange}. The argument is essentially well-known in the literature \cite{Tao1998,Tao1999,GuthHickmanIliopoulou2019} and we shall be brief. The detailed argument from \cite{GuthHickmanIliopoulou2019} cannot be applied directly because it relies on non-degenerate curvature properties $H2)$ of the phase function. However, we shall see that the partial non-degeneracy
\begin{equation}
\label{eq:OneNondegenerateCurvature}
\exists \text{non-vanishing eigenvalue of } \partial^2_{\omega \omega} \langle \partial_x \phi^\lambda(x;\omega), G^\lambda(x;\omega_0) \rangle \vert_{\omega = \omega_0}
\end{equation}
suffices for the argument. In the following we suppose that the phase $\phi$ satisfies the non-degeneracy $C1)$ and \eqref{eq:OneNondegenerateCurvature}. We shall prove that, if for $\bar{p} \geq 2$ and for all $\varepsilon > 0$ the estimate
\begin{equation}
\label{eq:LocalEstimate}
\| T^\lambda f \|_{L^p(B_R)} \lesssim_{\varepsilon,\phi,a} R^\varepsilon \| f \|_{L^p(A^{n-1})}
\end{equation}
holds for all $p \geq \bar{p}$, all $R$-balls $B_R$, and any amplitude, then we find the global estimate \eqref{eq:GlobalEstimate} to hold for all $p > \bar{p}$. The following notion plays an important role in the argument:
\begin{definition}[Tao \cite{Tao1999}]
Let $R \geq 1$. A collection $\{B(x_j,R)\}_{j=1}^N$ of $R$-balls in $\mathbb{R}^d$ is sparse if $\{x_1,\ldots,x_N\}$ are $(RN)^{\bar{C}}$-separated. Here $\bar{C} \geq 1$ is a fixed constant, chosen large enough to satisfy the requirements of the forthcoming argument.
\end{definition}
Like in previous instances of the argument, we are reduced to the analysis of sparse families of balls.
\begin{lemma}[{\cite[Lemma~12.2]{GuthHickmanIliopoulou2019}}]
\label{lem:SparseReduction}
To prove \eqref{eq:GlobalEstimate} for all $p > \bar{p}$, it suffices to show that for all $\varepsilon>0$ the estimate
\begin{equation}
\label{eq:SparseEstimate}
\| T^\lambda f \|_{L^{\bar{p}}(S)} \lesssim_{\varepsilon,\phi,a} R^\varepsilon \| f \|_{L^{\bar{p}}(A^{n-1})}
\end{equation}
holds whenever $R \geq 1$ and $S \subseteq \mathbb{R}^n$ is a union of $R$-balls belonging to a sparse collection, for any choice of amplitude function.
\end{lemma}
The key ingredient in the proof of Lemma \ref{lem:SparseReduction} is the following covering lemma due to Tao \cite{Tao1998}:
\begin{lemma}[Covering~lemma,~\cite{Tao1998,Tao1999}]
\label{lem:CoveringLemma}
Suppose that $E \subseteq \mathbb{R}^n$ is a finite union of $1$-cubes and $N \geq 1$. Define the radii $R_j$ inductively by $R_0 =1$ and $R_j=(R_{j-1} |E|)^{\bar{C}}$ for $1 \leq j \leq N-1$. Then, for each $0 \leq j \leq N-1$, there exists a family of sparse collections $(\mathcal{B}_{j,\alpha})_{\alpha \in A_j}$ of balls of radius $R_j$ such that the index sets $A_k$ have cardinality $O(|E|^{1/N})$ and
\begin{equation*}
E \subseteq \bigcup_{j=0}^{N-1} \bigcup_{\alpha \in A_j} S_{j,\alpha},
\end{equation*}
where $S_{j,\alpha}$ is the union of all the balls belonging to the family $\mathcal{B}_{j,\alpha}$.
\end{lemma}
With Lemma \ref{lem:CoveringLemma} at hand, the proof of Lemma \ref{lem:SparseReduction} from \cite{GuthHickmanIliopoulou2019} applies. It remains to establish the estimates for $T^\lambda$ over sparse collections of $R$-balls.
\begin{lemma}
Under the above hypotheses, if $p \geq \bar{p}$, then the estimate
\begin{equation*}
\| T^\lambda f \|_{L^p(S)} \lesssim_{\varepsilon,\phi,a} R^\varepsilon \| f \|_{L^p}
\end{equation*}
holds for all $ \varepsilon > 0$ whenever $S \subseteq \mathbb{R}^n$ is a union of $R$-balls belonging to a sparse collection.
\end{lemma}
\begin{proof}
Let $(B(x_j,R))_{j=1}^N$ be a sparse collection of balls. We can suppose that $R \ll \lambda$ and that all $B(x_j,R)$ intersect the $x$-support of $a^\lambda$. Furthermore, letting $c_{diam} > 0$ be a small constant $\text{diam} X < c_{diam}$ so that
\begin{equation*}
\frac{|x_{j_1} - x_{j_2}|}{\lambda} \lesssim c_{diam} \text{ for all } 1 \leq j_1,j_2 \leq N.
\end{equation*}
Fix $\eta \in C^\infty_c(\mathbb{R}^{n-1})$ satisfying $0 \leq \eta \leq 1$, $\text{supp}(\eta) \subseteq B^{n-1}$ and $\eta(z) = 1$ for all $z \in B(0,1/2)$. For $R_1 := CNR$, where $C \geq 1$ is a large constant, define $\eta_{R}(z) = \eta(z/R_1)$. Let $\psi \in C^\infty_c(\mathbb{R}^{n-1})$ satisfy $0 \leq \psi \leq 1$, $\text{supp}(\psi) \subseteq \Omega$ and $\psi(\omega) = 1$ for $\omega$ belonging to the $\omega$-support of $a^\lambda$. Fix $1 \leq j \leq N$ and write
\begin{equation*}
e^{i \phi^\lambda(x_j;\cdot)}\psi f = P_j f + (e^{i \phi^\lambda(x_j;\cdot)} \psi f - P_j f) =: P_j f + f_{j,\infty},
\end{equation*}
where $P_j f = \hat{\eta}_{R_1} * [e^{i \phi^\lambda(x_j;\cdot)} \psi f]$. If one defines
\begin{equation*}
\text{Err}(x) = \int_{\mathbb{R}^{n-1}} e^{i(\phi^\lambda(x;\omega) - \phi^\lambda(x_j;\omega))} a^\lambda(x;\omega) f_{j,\infty}(\omega) d\omega,
\end{equation*}
then it follows that
\begin{equation*}
T^\lambda f(x) = T^\lambda [ e^{-i \phi^\lambda(x_j;\cdot)} P_j f](x) + \text{Err}(x).
\end{equation*}
For $x \in B(x_j;R)$, the term $\text{Err}(x)$ is negligible. By Plancherel's theorem,
\begin{equation*}
\text{Err}(x) = \int \overline{G_x^{\vee} (z)} \cdot(1 - \eta_{R_1}(z)) [ e^{i \phi^\lambda(x_j;\cdot)} \psi f]^{\vee} (z) dz,
\end{equation*}
where
\begin{equation*}
G_x^{\vee} (z) = \frac{1}{(2 \pi)^{n-1}} \int_{\mathbb{R}^{n-1}} e^{i( \langle z, \omega \rangle - \phi^\lambda(x;\omega) + \phi^\lambda(x_j;\omega))} a^\lambda(x;\omega) d\omega.
\end{equation*}
Taking the $\omega$-derivatives of the phase of $G_x^{\vee} (z)$, one obtains
\begin{equation*}
\begin{split}
z - \lambda \big( \partial_\omega \phi \big( \frac{x}{\lambda}; \omega \big) - \partial_\omega \phi \big( \frac{x_j}{\lambda};\omega \big) \big) &= z + O(R) \\
\Rightarrow - \lambda \big( \partial_\omega^\alpha \phi \big( \frac{x}{\lambda};\omega \big) - \partial_\omega^\alpha \phi \big( \frac{x_j}{\lambda};\omega \big) \big) &= O(R) \text{ for } |\alpha| \geq 2.
\end{split}
\end{equation*}
Hence, if $z$ belongs to the support of $1-\eta_{R_1}$, then integration by parts shows that $G_x(z)$ is rapidly decaying in $R_1$, and we find
\begin{equation*}
|\text{Err}(x)| \leq \text{RapDec}(R_1) \| f \|_{L^p}.
\end{equation*}
By applying the estimate for $T^\lambda$ with $R^\varepsilon$-loss over each ball $B(x_j;R)$, one obtains
\begin{equation*}
\begin{split}
\| T^\lambda f \|_{L^p(S)} &\leq \big( \sum_{j=1}^N \| T^\lambda [ e^{-i \phi^\lambda(x_j;\cdot)} P_j f ] \|^p_{L^p(B(x_j;R))} \big)^{\frac{1}{p}} \\
&\lesssim_{\varepsilon, a, \phi} R^\varepsilon \big( \sum_{j=1}^N \| P_j f \|^p_{L^p(A^{n-1})} \big)^{\frac{1}{p}} + \| f \|_{L^p}.
\end{split}
\end{equation*}
Thus, it suffices to show that
\begin{equation*}
\big( \sum_{j=1}^N \| P_j f \|^p_{L^p} \big)^{\frac{1}{p}} \lesssim \| f \|_{L^p}.
\end{equation*}
This follows via interpolation between $p=2$ and $p= \infty$. For $p= \infty$, this is a consequence of Young's inequality. The estimate for $p=2$ is by duality equivalent to
\begin{equation*}
\big\| \sum_{j=1}^N e^{-2 \pi i \phi^\lambda(x_j;\cdot)} \psi \cdot [\hat{\eta}_{R_1} * g_j] \big\|_{L^2(\mathbb{R}^{d-1})} \lesssim \big( \sum_{j=1}^N \| g_j \|^2_{L^2} \big)^{\frac{1}{2}}.
\end{equation*}
By squaring the left-hand side, we find
\begin{equation*}
\sum_{j_1,j_2=1}^N \int_{\mathbb{R}^{d-1}} \overline{G_{j_1,j_2}(\omega)} \hat{\eta}_{R_1} * g_{j_1}(\omega) \overline{\hat{\eta}_{R_1} * g_{j_2}(\omega)} d\omega,
\end{equation*}
where
\begin{equation*}
G_{j_1,j_2}(\omega) = e^{i (\phi^\lambda(x_{j_1};\omega) - \phi^\lambda(x_{j_2};\omega))} \psi^2(\omega).
\end{equation*}
Plancherel's theorem yields
\begin{equation}
\label{eq:SquaredKernel}
\int_{\mathbb{R}^{n-1}} \overline{G_{j_1,j_2}(\omega)} \hat{\eta}_{R_1} * g_{j_1}(\omega) \overline{\hat{\eta}_{R_1} * g_2(\omega)} d\omega = \int_{\mathbb{R}^{n-1}} \overline{G_{j_1,j_2}^{\vee} (z)} (\eta_{R_1} \check{g}_{j_1}) * (\eta_{R_1} \check{g}_{j_2})^{\tilde{\;}} (z) dz.
\end{equation}
Here $(\eta_{R_1} \check{g}_{j_2})^{\tilde{\;}}(z) = \overline{\eta_{R_1} \check{g}_{j_2}(-z)}$. Fix $1 \leq j_1,j_2 \leq N$ with $j_1 \neq j_2$, let $z \in \mathbb{R}^{n-1}$ with $|z| \lesssim R_1 < |x_{j_2} - x_{j_1}|$ and consider
\begin{equation*}
G_{j_1,j_2}^{\vee}(z) = \frac{1}{(2 \pi)^{n-1}} \int_{\mathbb{R}^{n-1}} e^{i (\langle z,\omega \rangle + \phi^\lambda(x_{j_1};\omega) - \phi^\lambda(x_{j_2};\omega))} \psi^2(\omega) d\omega.
\end{equation*}
For $\alpha \in \mathbb{N}^{n-1}$ with $|\alpha| \leq 2$, consider
\begin{equation*}
\partial_\omega^\alpha [\phi^\lambda(x_{j_1};\omega) - \phi^\lambda(x_{j_2};\omega)] = \partial_\omega^\alpha \langle \partial_x \phi^\lambda(x_{j_1};\omega), x_{j_2} - x_{j_1} \rangle + O(c_{\text{diam}} |x_{j_2}-x_{j_1}|).
\end{equation*}
Let $c_{crit}>0$ be a small constant, chosen to satisfy the further needs of the arguments, and $\omega_0 \in \Omega$. Suppose that
\begin{equation}
\label{eq:KernelDirection}
\big| \pm \frac{x_{j_2}-x_{j_1}}{|x_{j_2}-x_{j_1}|} - G^\lambda(x_{j_1};\omega_0) \big| \geq c_{crit}.
\end{equation}
The non-degeneracy $C1)$ implies that the vector $G^\lambda(x;\omega_0)$ spans the kernel of $\partial^2_{\omega x} \phi^\lambda(x;\omega_0)$. This yields, in case of \eqref{eq:KernelDirection},
\begin{equation*}
| \partial_\omega [\langle \partial_x \phi^\lambda(x_{j_1};\omega),x_{j_2}-x_{j_1} \rangle \vert_{\omega = \omega_0} | \gtrsim |x_{j_2} - x_{j_1}|
\end{equation*}
and consequently,
\begin{equation*}
|\partial_\omega [ \phi^\lambda(x_{j_1};\omega) - \phi^\lambda(x_{j_2};\omega)] \vert_{\omega = \omega_0} | \gtrsim |x_{j_2} - x_{j_1}|.
\end{equation*}
Then, rapid decay of $\check{G}_{j_1,j_2}$ follows by integration by parts. If \eqref{eq:KernelDirection} fails, then
\begin{equation*}
\partial_\omega^\alpha \langle \partial_x \phi^\lambda(x_{j_1};\omega), \frac{x_{j_2}-x_{j_1}}{|x_{j_2}-x_{j_1}|} \rangle \vert_{\omega = \omega_0} = \partial_\omega^\alpha \langle \partial_x \phi^\lambda(x_{j_1};\omega), G^\lambda(x_{j_1};\omega_0) \rangle \vert_{\omega = \omega_0} + O(c_{crit}).
\end{equation*}
Hence, by a Van der Corput-argument \cite[Proposition~5,~p.~342]{Stein1993} we still find due to \eqref{eq:OneNondegenerateCurvature}
\begin{equation*}
| \check{G}_{j_1,j_2}(z)| \lesssim |x_{j_2}-x_{j_1}|^{-\frac{1}{2}}.
\end{equation*}
This yields the estimate for the absolute value of \eqref{eq:SquaredKernel}:
\begin{equation*}
\begin{split}
|\eqref{eq:SquaredKernel}| \lesssim R_1^{- \bar{C}/2} \| (\eta_{R_1} \check{g}_{j_1}) * (\eta_{R_1} \check{g}_{j_2})^{\tilde{\;}} \|_{L^1(\mathbb{R}^{n-1})} &\lesssim R_1^{- \bar{C}/2} \prod_{j=1}^2 \| \eta_{R_1} \check{g}_{j_i} \|_{L^1(\mathbb{R}^{n-1})} \\
&\lesssim R_1^{-\bar{C}/2 + n-1} \prod_{j=1}^2 \| g_{j_i} \|_{L^2(\mathbb{R}^{n-1})}.
\end{split}
\end{equation*}
Since there are only $O(N^2)$ choices of indices $j_1$ and $j_2$ and $R_1 = CRN$, the trivial estimate
\begin{equation*}
\prod_{i=1}^2 \| g_{j_i} \|_{L^2} \lesssim \sum_{j=1}^N \| g_{j} \|^2_{L^2}
\end{equation*}
suffices to sum the off-diagonal terms. The diagonal contribution is estimated by
\begin{equation*}
\big( \sum_{j=1}^N \| \hat{\eta}_{R_1} * g_j \|^2_{L^2(A^{n-1})} \big)^{1/2} \lesssim \big( \sum_{j=1}^N \| g_j \|^2_{L^2(A^{n-1})} \big)^{1/2}.
\end{equation*}
The proof is complete.
\end{proof}
\section{Improved local smoothing for Fourier integral operators}
\label{section:LocalSmoothing}
In this section we improve $L^p$-smoothing estimates for solutions to wave equations on compact Riemannian manifolds $(M,g)$ with $\dim (M) \geq 3$: We consider
\begin{equation}
\label{eq:WaveEquationCompactManifold}
\left\{ \begin{array}{cl}
\partial_t^2 u - \Delta_g u &= 0, \quad (t,x) \in \mathbb{R} \times M, \\
u(\cdot,0) &= f_0, \qquad \dot{u}(\cdot,0) = f_1
\end{array} \right.
\end{equation}
with the solution $u$ to \eqref{eq:WaveEquationCompactManifold} given by
\begin{equation*}
u(t) = \cos(t \sqrt{- \Delta_g}) f_0 + \frac{\sin(t \sqrt{- \Delta_g})}{\sqrt{- \Delta_g}} f_1.
\end{equation*}
Parametrices for the half-wave equation are provided by Fourier integral operators (FIOs); see below.
By results due to Seeger--Sogge--Stein \cite{SeegerSoggeStein1991} relying on the parametrix representation (see also \cite{Peral1980,Miyachi1980} in the Euclidean case), it is known that the fixed-time estimate
\begin{equation*}
\| u(\cdot,t) \|_{L^p(\mathbb{R}^d)} \lesssim \| f_0 \|_{L^p_{\bar{s}_p}(\mathbb{R}^d)} + \| f_1 \|_{L^p_{\bar{s}_p-1}(\mathbb{R}^d)}
\end{equation*}
with
\begin{equation}
\label{eq:SharpFixedTimeRegularity}
\bar{s}_p = (d-1) \big| \frac{1}{2} - \frac{1}{p} \big|
\end{equation}
is sharp for all $1<p<\infty$ provided that $t$ avoids a discrete set. The local smoothing conjecture due to C. Sogge \cite{Sogge1991} for the Euclidean wave equation, i.e., $(M,g) = (\mathbb{R}^d,(\delta^{ij}))$ in \eqref{eq:WaveEquationCompactManifold}, states that
\begin{equation}
\label{eq:LocalSmoothingEstimate}
\big( \int_1^2 \| u(\cdot,t) \|_{L^p(\mathbb{R}^d)}^p \big)^{\frac{1}{p}} \lesssim \| f_0 \|_{L^p_{\bar{s}_p-\sigma}(\mathbb{R}^d)} + \| f_1 \|_{L^p_{\bar{s}_p-1-\sigma}(\mathbb{R}^d)}
\end{equation}
for $\sigma < \frac{1}{p}$ and $\frac{2d}{d-1} \leq p < \infty$. (Note that $\bar{s}_p - \frac{1}{p} = 0$ for $p = \frac{2d}{d-1}$.) This conjecture stands on top of prominent open problems in Harmonic Analysis as it implies as well the restriction conjecture as the Bochner--Riesz conjecture. Initial progress was due to Sogge \cite{Sogge1991} and Mockenhaupt--Seeger--Sogge \cite{MockenhauptSeegerSogge1993}. Wolff identified decoupling inequalities \cite{Wolff2000} to yield sharp local smoothing estimates. Further progress in this direction was made in \cite{GarrigosSeeger2009,Lee2020,LeeVargas2012}. Bourgain--Demeter \cite{BourgainDemeter2015} covered the sharp range for decoupling inequalities, which implies sharp local smoothing estimates for $p \geq \frac{2(d+1)}{d-1}$. We refer to the survey by Beltran--Hickman--Sogge \cite{BeltranHickmanSogge2018} for local smoothing estimates for FIOs. Guth--Wang--Zhang \cite{GuthWangZhang2020} verified the Euclidean local smoothing conjecture for $d=2$ by a sharp $L^4$-square function estimate. Gao \emph{et al.} \cite{GaoLiuMiaoXi2020} extended this to compact Riemannian surfaces. We remark that for $d \geq 3$, counterexamples due to Minicozzi--Sogge \cite{MinicozziSogge1997} show that \eqref{eq:LocalSmoothingEstimate} fails if one replaces $\mathbb{R}^d$ with general compact Riemannian manifolds for $\sigma < 1/p$, if $p < p_{d,+}$ with
\begin{equation}
\label{eq:LpCompactManifold}
p_{d,+} =
\begin{cases}
\frac{2 \cdot (3d+1)}{3d-3}, \text{ if } d \text{ is odd}, \\
\frac{2 \cdot (3d+2)}{3d-2}, \text{ if } d \text{ is even}.
\end{cases}
\end{equation}
Hence, local smoothing estimates for solutions to wave equations on compact Riemannian manifolds are only conjectured for $p \geq p_{d,+}$ with $\sigma < 1/p$.
\medskip
Gao \emph{et al.} \cite{GaoLiuMiaoXi2021} also improved the Euclidean local smoothing estimates for $d \geq 3$ and $2 \leq p \leq \frac{2(d+1)}{d-1}$ due to Bourgain--Demeter by a broad--narrow iteration. Presently, we extend their arguments to the variable coefficient case.
Let $d \geq 3$ and
\begin{equation}
\label{eq:RangeSmoothingFIO}
p_d =
\begin{cases}
2 \cdot \frac{3d+5}{3d+1} \text{ for } d \text{ odd}, \\
2 \cdot \frac{3d+6}{3d+2} \text{ for } d \text{ even}.
\end{cases}
\end{equation}
We show the following:
\begin{theorem}[Improved local smoothing on compact manifolds]
\label{thm:ImprovedLocalSmoothing}
Let $(M,g)$ be a compact Riemannian manifold with $\dim(M) \geq 3$. Let $\bar{s}_p$ be as in \eqref{eq:SharpFixedTimeRegularity}, $p_d \leq p < \infty$ with $p_d$ as in \eqref{eq:RangeSmoothingFIO} and $\sigma < \frac{2}{p} - \frac{1}{2}$. Let $u$ be a solution to \eqref{eq:WaveEquationCompactManifold}.
Then, we find the following estimate to hold:
\begin{equation}
\label{eq:LocalSmoothingCompactManifold}
\| u \|_{L_t^p([1,2], L_x^p(M))} \lesssim_{M,g,p,\sigma} \| f_0 \|_{L^p_{\bar{s}_p-\sigma}(M)} + \| f_1 \|_{L^p_{\bar{s}_p-\sigma+1}(M)}.
\end{equation}
\end{theorem}
We can interpolate with the trivial $L^2$-estimate and the sharp local smoothing estimates for $p \geq \frac{2(d+1)}{d-1}$ due to Beltran--Hickman--Sogge \cite{BeltranHickmanSogge2020} to find a broader range of estimates (cf. \cite[Corollary~1.3]{GaoLiuMiaoXi2021}):
\begin{corollary}
\label{cor:InterpolatedSmoothing}
Let $d \geq 3$ and $u$ be a solution to \eqref{eq:WaveEquationCompactManifold}. Then, \eqref{eq:LocalSmoothingCompactManifold} holds true for $\sigma < \sigma_p$, where, if $d \geq 3$ is odd,
\begin{equation*}
\sigma_p =
\begin{cases}
\frac{3d-3}{4} \big( \frac{1}{2} - \frac{1}{p} \big), \quad 2 <p \leq p_d, \\
\frac{d-1}{d+3} \big( \frac{3d+1}{6d+10} - \frac{1}{p} \big) + \frac{3d-3}{6d+10}, \quad p_d < p \leq \frac{2(d+1)}{d-1},
\end{cases}
\end{equation*}
and, if $d \geq 3$ is even,
\begin{equation*}
\sigma_p =
\begin{cases}
\frac{3d-2}{4} \big( \frac{1}{2} - \frac{1}{p} \big), \quad 2 < p \leq p_d, \\
\frac{d-2}{d+4} \big( \frac{3d+2}{6d+12} - \frac{1}{p} \big) + \frac{3d-2}{6d+12}, \quad p_d < p \leq \frac{2(d+1)}{d-1}.
\end{cases}
\end{equation*}
\end{corollary}
It is well-known (cf. \cite[Chapter~4]{Sogge2017}, \cite[p.~224]{MinicozziSogge1997}) that local parametrices for \eqref{eq:WaveEquationCompactManifold} take the form of FIOs
\begin{equation}
\label{eq:FIO}
(\mathcal{F} f)(x,t) = \int_{\mathbb{R}^d} e^{i \phi(x,t;\xi)} a(x,t;\xi) \hat{f}(\xi) d\xi
\end{equation}
with phase functions $\phi \in C^\infty(\mathbb{R}^{d+1} \times \mathbb{R}^d \backslash 0)$, which are $1$-homogeneous in $\xi$ and satisfy $C1)$ and $C2^+)$. $a \in S^0(\mathbb{R}^{2d+1})$ is a symbol of order zero, compactly supported in $(x,t)$.
It turns out that for the proof of Theorem \ref{thm:ImprovedLocalSmoothing}, it suffices to prove bounds for rescaled operators
\begin{equation}
\label{eq:RescaledFIO}
(\mathcal{F}^\lambda f)(x,t) = \int_{\mathbb{R}^d} e^{i \phi^\lambda(x,t;\xi)} a^\lambda(x,t;\xi) \hat{f}(\xi) d\xi
\end{equation}
with $a^\lambda$ and $\phi^\lambda$ defined like in previous sections. Theorem \ref{thm:ImprovedLocalSmoothing} is a consequence of the following (cf. \cite[Section~3]{BeltranHickmanSogge2020}):
\begin{proposition}
\label{prop:SmoothingFIO}
Let $\mathcal{F}$ be an FIO as in \eqref{eq:FIO} and $p_d$ as in \eqref{eq:RangeSmoothingFIO}. Then, we find the following local smoothing estimate to hold for $p_d \leq p < \infty$:
\begin{equation}
\label{eq:SmoothingFIOs}
\| \mathcal{F}^\lambda f \|_{L^p_{t,x}(\mathbb{R}^{d+1})} \lesssim_{\varepsilon,\phi,a} \lambda^{d \big( \frac{1}{2} - \frac{1}{p} \big) + \varepsilon} \| f \|_{L^p(\mathbb{R}^d)}.
\end{equation}
\end{proposition}
Proposition \ref{prop:SmoothingFIO} improves on the previously best estimates due to Beltran--Hickman--Sogge \cite{BeltranHickmanSogge2020}, which read
\begin{equation*}
\| \mathcal{F}^\lambda f \|_{L^p_{t,x}(\mathbb{R}^{d+1})} \lesssim_{\varepsilon, \phi, a} \lambda^{\frac{(d-1)}{2} \big( \frac{1}{2} - \frac{1}{p} \big)+ \frac{1}{p} + \varepsilon} \| f \|_{L^p(\mathbb{R}^d)}
\end{equation*}
for $2 \leq p \leq \frac{2(d+1)}{d-1}$. Beltran--Hickman--Sogge \cite{BeltranHickmanSogge2020} extended the decoupling inequalities in the constant coefficient case \cite{BourgainDemeter2015} to variable coefficients. This argument also yields local smoothing estimates for FIOs, which do not satisfy the convexity condition $C2^+)$. Indeed, the FIOs, for which decoupling yields the sharp smoothing estimates (cf. \cite[Section~4]{BeltranHickmanSogge2020}), are the ones with $d$ odd, and
\begin{equation*}
\partial^2_{\xi \xi} \langle \partial_x \phi(x,t;\xi), G_0(x,t;\xi) \rangle
\end{equation*}
having $\frac{d-1}{2}$ positive and $\frac{d-1}{2}$ negative eigenvalues.
For the proof of Proposition \ref{prop:SmoothingFIO}, we run almost the same iteration as in the proof of Theorem \ref{thm:LpLpEstimatesVariableCoefficients}. The following lemma based on finite speed of propagation allows to convert $L^2$-estimates for $T^\lambda$ into $L^p$-estimates for $\mathcal{F}^\lambda$:
\begin{lemma}
\label{lem:FiniteSpeedPropagation}
Let $(\phi,a)$ be reduced data and $\psi \in \mathcal{S}(\mathbb{R}^d)$ such that $\text{supp} (\hat{\psi}) \subseteq B(0,1)$, $\sum_{\ell \in \mathbb{Z}^d} \psi(x-\ell) \equiv 1$ for any $x \in \mathbb{R}^d$. Assume $\text{supp} (\hat{f}) \subseteq A^d$. Then, for any $\varepsilon > 0$, the following estimate holds true:
\begin{equation}
\label{eq:FiniteSpeedPropagation}
\begin{split}
|\mathcal{F}^\lambda f(x,t)| &\lesssim_\varepsilon |\mathcal{F}^\lambda (\psi_{R^{1+\varepsilon}(x_0)} f)(x,t)| \\
&\quad + \text{RapDec}(R) \sum_{|\ell| > R^\varepsilon} (1+|\ell|)^{-M} \| f |\psi_{\ell}(\cdot-x_0)|^{\frac{1}{2}} \|_{L^p(w_{B_R^d(x_0)})}
\end{split}
\end{equation}
for $(x,t) \in B(x_0,R) \times [-R,R]$, $1<p<\infty$, where
\begin{equation*}
\psi_{R^{1+\varepsilon}(x_0)}(x) = \sum_{|\ell| < R^\varepsilon} \psi(R^{-1}(x-x_0) - \ell).
\end{equation*}
\end{lemma}
\begin{proof}
The claim follows from a kernel estimate. We have
\begin{equation*}
\begin{split}
\mathcal{F}^\lambda(x,t) &= \int e^{i \phi^\lambda(x,t;\xi)} a^\lambda(x,t;\xi) \hat{f}(\xi) d\xi \\
&= \frac{1}{(2 \pi)^d} \int e^{i(\phi^\lambda(x,t;\xi) - \langle y,\xi \rangle)} a^\lambda(x,t;\xi) f(y) dy d\xi.
\end{split}
\end{equation*}
We set $K^\lambda(x,t;y) = \int e^{i(\phi^\lambda(x,t;\xi) - \langle y, \xi \rangle)} a^\lambda(x,t;\xi) d\xi$. Let $\Phi^\lambda(x,y,\xi,t) = \phi^\lambda(x,t;\xi) - \langle y, \xi \rangle$. We have
\begin{equation*}
\nabla_\xi \Phi^\lambda(x,y,\xi,t) = \nabla_\xi \phi^\lambda(x,t;\xi) - y.
\end{equation*}
For a reduced phase function, we have
\begin{equation*}
\nabla_\xi \phi^\lambda(x,0;\xi) = x, \quad \nabla_\xi \phi^\lambda(x,t;\xi) = \frac{1}{\lambda} \nabla_\xi \int_0^{\lambda t} \partial_t \phi^\lambda(x,s;\xi) ds + x.
\end{equation*}
By $|\nabla_\xi \partial_t \phi(x,t;\xi)| \lesssim 1$ for a reduced phase function, we find for $|t| \leq R$ and $|x-y| \geq R^{1+\varepsilon}$ rapid decay by non-stationary phase. We have the estimate
\begin{equation*}
|K^\lambda (x,t;y)| \leq C_N (1+R|x-y|)^{-N}.
\end{equation*}
Provided that $|\partial_\xi^\alpha a| \leq c_{par}$ for $0 \leq |\alpha| \leq N$ and reduced phase functions, $C_N$ can be chosen uniformly. By this, we find \eqref{eq:FiniteSpeedPropagation} to hold.
\end{proof}
By the same arguments as in Section \ref{section:LinearEstimates}, we can show the following narrow decoupling:
\begin{proposition}
\label{prop:NarrowDecoupling}
Let $B_{K^2} \subseteq B(0,\lambda^{1-\delta'})$ be a $K^2$-ball. Let $(\phi,a)$ be a $K$-flat datum. Let $k \geq 3$ and $V$ be a $(k-1)$-dimensional vector space. Suppose that $\text{supp}(f) \subseteq \bigcup_\nu S_{\nu}$ be a union of $K^{-1}$-slabs such that $\angle(G^\lambda(\bar{x};\tau),V) \leq K^{-2}$. Then, we find the following estimate to hold:
\begin{equation}
\label{eq:NarrowDecoupling}
\| T^\lambda f \|_{L^p(B^{d+1}_{K^2})} \lesssim_\delta K^\delta \big( \sum_\nu \| T^\lambda f_{\nu} \|^2_{L^p(w_{B^{d+1}_{K^2}})} \big)^{1/2} + \lambda^{- \frac{\min( \delta, \delta') N}{2}} \| f \|_{L^2}
\end{equation}
for $ 2 \leq p \leq \frac{2(k-1)}{k-3}$.
\end{proposition}
Like in Proposition \ref{prop:DecouplingNarrowTerm}, the amplitude functions on the right-hand side are slightly different, but satisfy uniform bounds. The discrepancy will be hidden in the induction hypothesis again.
As further ingredient we use the following Lorentz rescaling for FIOs. Let $Q_{p,\delta}(R)$ be the infimum over all constants such that
\begin{equation*}
\| \mathcal{F}^\lambda f \|_{L^p(B(0,R))} \leq Q_{p,\delta}(R) R^{d \big( \frac{1}{2} - \frac{1}{p} \big)} \| f \|_{L^p}
\end{equation*}
for $1 \leq R \leq \lambda$ and all FIOs $\mathcal{F}$ with $\lambda^\delta$-flat phase functions and amplitude functions, which satisfy
\begin{equation*}
| \partial_{\xi}^\alpha a(x,t;\xi)| \leq c_{par}
\end{equation*}
for $0 \leq |\alpha| \leq N$.
\begin{lemma}[Lorentz rescaling for FIOs]
\label{lem:RescalingFIOs}
Let $(\phi,a)$ be reduced data with $\phi$ a $\lambda^\delta$-flat phase function and $\hat{f}$ supported in a $\rho^{-1}$-slab in $A^d$. Then, for any $1 \leq \rho \leq R \leq \lambda$:
\begin{equation}
\label{eq:ParabolicRescalingFIO}
\| \mathcal{F}^\lambda g \|_{L^p(B(0,R))} \lesssim_{\delta'} R^{\delta'} Q_{p,\delta}(R/\rho^2) R^{d \big( \frac{1}{2} - \frac{1}{p} \big)} \rho^{\frac{2(d+1)}{p} - d} \| g \|_{L^p}.
\end{equation}
\end{lemma}
\begin{proof}
The proof has much in common with the proof of Lemma \ref{lem:ParabolicRescaling}. However, after rescaling, we use almost orthogonality in space-time, which comes from finite speed of propagation (cf. Lemma \ref{lem:FiniteSpeedPropagation}). Let $\omega \in B_{d-1}(0,1)$ with $(\omega,1)$ the centre of the $\rho^{-1}$-slab encasing the support of $\hat{g}$:
\begin{equation*}
\text{supp}(\hat{g}) \subseteq \{ (\xi',\xi_d) \in \mathbb{R}^d \, : \, 1/2 \leq \xi_d \leq 2 \text{ and } \big| \frac{\xi'}{\xi_d} - \omega \big| \leq \rho^{-1} \}.
\end{equation*}
We perform the change of variables:
\begin{equation*}
(\xi',\xi_d) = (\eta_d \omega + \rho^{-1} \eta', \eta_d),
\end{equation*}
after which follows
\begin{equation*}
(\mathcal{F}^\lambda g)(x,t) = \int_{\mathbb{R}^d} e^{i \phi^\lambda(x,t;\eta_d \omega + \rho^{-1} \eta', \eta_d)} a^\lambda(x,t;\eta_d \omega + \rho^{-1} \eta', \eta_d) \hat{\tilde{g}}(\eta) d\eta,
\end{equation*}
where $\hat{\tilde{g}}(\eta) = \rho^{-(d-1)} \hat{g}(\eta_d \omega + \rho^{-1} \eta', \eta_d)$ and $\text{supp}(\hat{\tilde{f}}) \subseteq \Xi$. By Taylor expansion and homogeneity of the phase, we find
\begin{equation*}
\begin{split}
\phi(x,t;\eta_d \omega + \rho^{-1} \eta', \eta_d) &= \phi(x,t;\omega,1) \eta_d + \rho^{-1} \langle \partial_{\xi'} \phi(x,t;\omega,1), \eta' \rangle \\
&\quad + \rho^{-2} \int_0^1 (1-r) \langle \partial^2_{\xi' \xi'} \phi(x,t;\eta_d \omega + r \rho^{-1} \eta', \eta_d) \eta', \eta' \rangle dr.
\end{split}
\end{equation*}
Let $\Upsilon_\omega(x,t) = (\Upsilon(x,t;\omega,1),x_d)$ and $\Upsilon^\lambda_\omega(x,t) = \lambda \Upsilon_\omega(x/\lambda,t/\lambda)$ and consider anisotropic dilations
\begin{equation*}
D_\rho (x',x_d,t) = (\rho x', x_d, \rho^2 t) \text{ and } D'_{\rho^{-1}} (x',x_d) = (\rho^{-1} x', \rho^{-2} x_d)
\end{equation*}
on $\mathbb{R}^{d+1}$ and $\mathbb{R}^d$, respectively. By definition of $\Upsilon$, we find
\begin{equation*}
\mathcal{F}^\lambda g \circ \Upsilon^\lambda_\omega \circ D_\rho = \tilde{\mathcal{F}}^{\lambda /\rho^2} \tilde{g},
\end{equation*}
where
\begin{equation*}
\tilde{\mathcal{F}}^{\lambda / \rho^2} \tilde{g}(y,\tau) = \int_{\mathbb{R}^d} e^{i \tilde{\phi}^{\lambda/\rho^2}(y,\tau;\eta)} \tilde{a}^{\lambda/\rho^2}(y,\tau;\eta) \hat{\tilde{g}}(\eta) d\eta
\end{equation*}
for the phase $\tilde{\phi}(y,\tau;\eta)$ given by
\begin{equation*}
\langle y, \eta \rangle + \int_0^1 (1-r) \langle \partial^2_{\xi' \xi'} \phi(\Upsilon_\omega(D'_{\rho^{-1}} y,y_d); \eta_d \omega + r \rho^{-1} \eta', \eta_d) \eta', \eta' \rangle dr
\end{equation*}
and the amplitude
\begin{equation*}
\tilde{a}(y,\tau;\eta) = a(\Upsilon_\omega (D'_{\rho^{-1}} y;\tau);\eta_d \omega + \rho^{-1} \eta', \eta_d).
\end{equation*}
By change of space-time variables, we find
\begin{equation}
\label{eq:RescalingFIOI}
\| \mathcal{F}^\lambda g \|_{L^p(B_R)} \lesssim \rho^{\frac{d+1}{p}} \| \tilde{\mathcal{F}}^{\lambda/\rho^2} \tilde{g} \|_{L^p((\Upsilon^\lambda_\omega \circ D_\rho)^{-1}(B_R)}.
\end{equation}
Note that $(\Upsilon^\lambda_\omega \circ D_\rho)^{-1}(B_R)= D_R$ is roughly a set of size $R/\rho \times \ldots \times R/\rho \times R \times R/\rho^2$. We want to apply the induction hypothesis, to which end we use finite speed of propagation: Since the time-scale is $R/\rho^2$, the localization by Lemma \ref{lem:FiniteSpeedPropagation} yields
\begin{equation}
\label{eq:RescalingFIOII}
\| \tilde{\mathcal{F}}^{\lambda/\rho^2} \tilde{g} \|_{L^p(D_R)} \lesssim_{\delta'} R^{\delta'} Q_{p,\delta}(R/\rho^2) (R/\rho^2)^{d \big( \frac{1}{2} - \frac{1}{p} \big) } \| \tilde{g} \|_{L^p}.
\end{equation}
Since $\tilde{g}(x) = g(\rho x', x_d - \omega x')$, we find
\begin{equation}
\label{eq:RescalingFIOIII}
\| \tilde{g} \|_{L^p} = \rho^{- \frac{d-1}{p}} \| g \|_{L^p}.
\end{equation}
Taking \eqref{eq:RescalingFIOI}, \eqref{eq:RescalingFIOII}, and \eqref{eq:RescalingFIOIII} together, we find
\eqref{eq:ParabolicRescalingFIO} to hold.
\end{proof}
We are ready for the proof of the following proposition:
\begin{proposition}
\label{prop:NarrowIterationSmoothingFIO}
Let $d \geq 3$, $2 \leq k \leq d$, and $\lambda \geq 1$. If for all $\varepsilon > 0$ and
\begin{equation*}
\bar{p}(k,d) \leq p \leq
\begin{cases}
\infty, \quad 2 \leq k \leq 3, \\
2 \frac{k-1}{k-3}, \quad k \geq 4,
\end{cases}
\text{ with }
\bar{p}(k,d) =
\begin{cases}
\frac{2(d+1)}{d}, \quad k=2, \\
2 \cdot \frac{2d-k+5}{2d-k+3}, \quad k \geq 3,
\end{cases}
\end{equation*}
FIOs with reduced data $(\phi,a)$ obey the $k$-broad estimate for all $1 \leq K \leq R \leq \lambda$ and some fixed choice of $A$
\begin{equation*}
\| \mathcal{F}^\lambda f \|_{BL^p_{k,A}(B^{d+1}_R)} \leq \bar{C}_\varepsilon K^{C_\varepsilon} R^\varepsilon R^{d \big( \frac{1}{2} - \frac{1}{p} \big) } \| f \|_{L^p},
\end{equation*}
then we have for some $D_{\varepsilon,\phi,a}$
\begin{equation}
\label{eq:SmoothingFIONarrow}
\| \mathcal{F}^\lambda f \|_{L^p(\mathbb{R}^{d+1})} \leq D_{\varepsilon,\phi,a} \lambda^{d \big( \frac{1}{2} - \frac{1}{p} \big) + \varepsilon} \| f \|_{L^p(\mathbb{R}^d)}.
\end{equation}
\end{proposition}
Proposition \ref{prop:SmoothingFIO} follows from Proposition \ref{prop:NarrowIterationSmoothingFIO} by choosing $k= \frac{d+5}{2}$ for $d$ odd and $k= \frac{d+4}{2}$ for $d$ even. This will complete the proof of Theorem \ref{thm:ImprovedLocalSmoothing}.
\begin{proof}[Proof of Proposition \ref{prop:NarrowIterationSmoothingFIO}]
The proof has many similarities with the proof of Proposition \ref{prop:LinearFromBroadEstimates}, and we shall be brief. By one parabolic rescaling depending on the phase as in the beginning of the proof of Proposition \ref{prop:LinearFromBroadEstimates}, we can suppose that $(\phi,a)$ is $\lambda^{\delta}$-flat, and $R \leq \lambda^{1-\frac{\varepsilon}{10d}}$.
In the following $Q_{p,\tilde{\delta}}(R)$ denotes the smallest constant such that for all $\lambda^{\tilde{\delta}}$-flat phase functions and normalized amplitude functions, we have
\begin{equation*}
\| \mathcal{F}^\lambda f \|_{L^p(B(0,R))} \leq R^{d \big( \frac{1}{2} - \frac{1}{p} \big)} Q_{p,\tilde{\delta}}(R) \| f \|_{L^p}.
\end{equation*}
It suffices to prove that for any $\varepsilon > 0$ there is $C_\varepsilon > 0$ such that $Q_{p,\tilde{\delta}}(R) \leq C_\varepsilon R^\varepsilon$ for any $R \leq \lambda^{1-\frac{\varepsilon}{10 d}}$.
For a given ball $B_{K^2} \subseteq B(0,R)$, let $V_1,\ldots,V_A$ be $(k-1)$-dimensional linear subspaces, which achieve the minimum in the definition of the $k$-broad norm, such that
\begin{equation*}
\begin{split}
\int_{B_{K^2}^{d+1}} | \mathcal{F}^\lambda f(x,t)|^p dx dt &\lesssim K^{O(1)} \max_{\tau \notin V_{\ell}} \int_{B_{K^2}^{d+1}} | \mathcal{F}^\lambda f^\tau(x,t)|^p dx dt \\
&\quad + \sum_{\ell=1}^A \int_{B_{K^2}^{d+1}} \big| \sum_{\tau \in V_{\ell}} \mathcal{F}^\lambda f^\tau(x,t) \big|^p dx dt.
\end{split}
\end{equation*}
Summing over a finitely overlapping family $\big( B_{K^2} \big) = \mathcal{B}_{K^2}$ covering $B(0,R)$ yields
\begin{equation*}
\begin{split}
\int_{B(0,R)} | \mathcal{F}^\lambda f(x,t)|^p dx dt &\lesssim K^{O(1)} \sum_{B_{K^2} \in \mathcal{B}_{K^2}} \min_{V_1,\ldots,V_A} \max_{\tau \notin V_{\ell}} \int_{B_{K^2}} | \mathcal{F}^\lambda f^\tau(x,t)|^p dx dt \\
&\quad + \sum_{B_{K^2} \in \mathcal{B}_{K^2}} \sum_{\ell=1}^A \int_{B_{K^2}} \big| \sum_{\tau \in V_{\ell}} \mathcal{F}^\lambda f^\tau(x,t)|^p dx dt.
\end{split}
\end{equation*}
By the broad norm estimate, we find
\begin{equation*}
\sum_{B_{K^2} \in \mathcal{B}_{K^2}} \min_{V_1,\ldots,V_A} \max_{\tau \notin V_{\ell}} \int_{B_{K^2}} | \mathcal{F}^\lambda f^\tau(x,t)|^p dx dt \leq \bar{C}_\varepsilon K^{C_{\varepsilon}} R^{\frac{\varepsilon p}{2}} R^{dp \big( \frac{1}{2} - \frac{1}{p} \big)} \| f \|_{L^p}^p.
\end{equation*}
The narrow contribution is estimated by Proposition \ref{prop:NarrowDecoupling}:
\begin{equation}
\label{eq:NarrowEstimateI}
\begin{split}
\sum_{\ell=1}^A \int_{B_{K^2} \in \mathcal{B}_{K^2} } \big| \sum_{\tau \in V_{\ell}} \mathcal{F}^\lambda f^\tau(x,t) \big|^p dx dt &\leq C_{\delta} K^{\delta} K^{\max( (k-3)\big( \frac{1}{2} - \frac{1}{p} \big) p, 0)} \\
&\qquad \times \sum_{\tau} \int_{\mathbb{R}^{d+1}} w_{B_{K^2}} |\mathcal{F}^\lambda f^\tau(x,t)|^p dx dt,
\end{split}
\end{equation}
where we have used the sector counting estimate
\begin{equation*}
\# \{ \tau : \tau \in V_{\ell} \} \lesssim \max(1,K^{k-3}).
\end{equation*}
Summing over $B_{K^2} \in \mathcal{B}_{K^2}$ in \eqref{eq:NarrowEstimateI}, we find
\begin{equation*}
\begin{split}
\sum_{B_{K^2} \in \mathcal{B}_{K^2}} \sum_{\ell=1}^A \int_{B_{K^2}^{d+1}} \big| \sum_{\tau \in V_{\ell}} \mathcal{F}^\lambda f^\tau(x,t) \big|^p dx dt &\leq C_{\delta} K^\delta K^{\max( (k-3) \big( \frac{1}{2} - \frac{1}{p} \big) p, 0)} \\
&\quad \times \sum_\tau \int_{\mathbb{R}^{d+1}} w_{B(0,2R)} |\mathcal{F}^\lambda f^\tau(x,t) |^p dx dt.
\end{split}
\end{equation*}
By Lemma \ref{lem:ParabolicRescaling}, we find
\begin{equation}
\label{eq:NarrowEstimateII}
\begin{split}
&\quad \int_{B(0,R)} |\mathcal{F}^\lambda f^\tau(x,t) |^p dx dt \\
&\lesssim_{\delta_1} K^{-2d \big(\frac{1}{2} - \frac{1}{p} \big) p + 2 } Q_{p,\tilde{\delta}}^p (R/K^2) R^{dp \big( \frac{1}{2} - \frac{1}{p} \big) + \delta_1} \| f^\tau \|^p_p + \text{RapDec}(R) \| f \|_p^p.
\end{split}
\end{equation}
Note the following by Plancherel's theorem for $p=2$, the kernel estimate for $p=\infty$, and interpolation:
\begin{equation}
\label{eq:SectorSummation}
\big( \sum_\tau \| f^\tau \|^p_p \big)^{\frac{1}{p}} \lesssim \| f \|_p.
\end{equation}
Hence, summing \eqref{eq:NarrowEstimateII} over $\tau$ yields by \eqref{eq:SectorSummation}
\begin{equation*}
\begin{split}
\int_{B_R^{d+1}} |\mathcal{F}^\lambda f(x,t)|^p dx dt &\leq C_\varepsilon R^{dp \big( \frac{1}{2}- \frac{1}{p} \big) + \frac{\varepsilon p}{2}} \| f \|_{L^p}^p \\
&\quad + C_{\delta,\delta_1} K^\delta R^{dp \big( \frac{1}{2} - \frac{1}{p} \big) + \delta_1} K^{-e(p,k,d)} Q_{p,\tilde{\delta}}^p(R/K^2) \| f \|_{L^p}^p
\end{split}
\end{equation*}
with
\begin{equation*}
e(p,k,d) = \max \{ 2d \big( \frac{1}{2} - \frac{1}{p} \big) p - 2, 2d \big( \frac{1}{2} - \frac{1}{p} \big) p - 2 - (k-3) \big( \frac{1}{2} - \frac{1}{p} \big) p \}.
\end{equation*}
We find $e(p,k,d) \geq 0$, if
\begin{equation*}
p \geq
\begin{cases}
&\frac{2(d+1)}{d}, \quad k=2, \\
&2 \cdot \frac{2d-k+5}{2d-k+3}, \quad k \geq 3.
\end{cases}
\end{equation*}
By the induction hypothesis, we have
\begin{equation*}
Q_{p,\tilde{\delta}}^p(R) \leq K^{O(1)} \tilde{C}_{\varepsilon} K^{\bar{C}_\varepsilon} R^{\frac{\varepsilon p}{2}} + C_{\delta,\delta_1} R^{\delta_1 } Q_{p,\tilde{\delta}}^p(R) K^{-e(p,k,d) + \delta}.
\end{equation*}
We can choose $\delta(\varepsilon)$, $\delta_1(\varepsilon)$, and $K=K_0 R^{\tilde{\delta}}$ similarly as at the end of the proof of Proposition \ref{prop:LinearFromBroadEstimates} to close the induction.
\end{proof}
\section*{Acknowledgements}
Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) -- Project-ID 258734477 -- SFB 1173.
\bibliographystyle{plain}
|
\section{Introduction}
The connection between long gamma-ray bursts (those GRBs with prompt gamma-ray durations lasting longer than a few seconds) and the deaths of massive stars is strongly supported, both theoretically and observationally, including the direct association of supernova events coincident with many long gamma-ray bursts \citep{Woos93,MW99, BKD02, Hjorth03,WB06,WH06,KNJ08a, KNJ08b, HB12, Ly17}. It is also clear that long gamma-ray bursts (lGRBs) are rare events - not every massive stellar death produces a highly relativistic jet that leads to a gamma-ray burst and its subsequent afterglow. The conditions required to produce an lGRB from a collapsing star \citep{MW99,YL05,HMM05,Yoon06,WH06}, and indeed the jet launching mechanism itself \cite[e.g.][]{BK08, KB09, LR19bz,KP21}, are still open questions. It is an ongoing pursuit to determine and understand {\em which} stars make gamma-ray bursts, and {\em why}.
We can hope to unravel this mystery through a careful examination of the emission we observe from these objects. For example, the $10$'s to $100$'s seconds of highly variable prompt gamma-ray emission offers insight into the central engine of these objects \citep{KPS97, NP02, LR16,LR18}. The duration and strength of this prompt emission, produced by internal dissipation processes in the jet, reflect the amount of angular momentum and the overall energy budget of the black hole-accretion disk system.\footnote{Here, we do not consider magnetar central engines, although our arguments can be generalized to that case.} Meanwhile, the long-lived (days, weeks, months, years) afterglow of a gamma-ray burst, produced by the deceleration of the jet as it sweeps up the external medium, carries information about the circumburst environment since it is the cumulative mass of this environment that serves to decelerate the jet.
Unfortunately, unraveling the properties of a GRB progenitor from the prompt and afterglow light curves and spectra is difficult, while the analysis is plagued with degeneracies among the model parameters; that is, several equally reasonable models may fit the data equally well and it becomes difficult to distinguish among them \cite[e.g.][]{Li07,Via18,TLR21}. As more and better broadband data arrive, and statistical fitting and analysis techniques progress, our understanding is improving (see, e.g., the paper by \cite{BGG20}, who discuss ways to break the degeneracies by considering the shape or curvature of the lightcurve). However, we still have much to learn about teasing out the underlying physics of a GRB.
Another approach is to search for correlations between observed variables, which can offer clues into physical processes playing a role in the GRB event. For example, \cite{LPM00} found that the well known gamma-ray flux-spectral peak energy correlation seen in BATSE GRB data \citep{Mal95} is due to an intrinsic correlation between isotropic radiated energy $E_{iso}$ and the peak of the intrinsic spectrum $E_{p}$; they predicted a relationship $E_{iso} \sim E_{p}^{0.55 \pm 0.15}$, which was indeed later found in the data for those lGRBs with measured redshifts \citep{AM02}. \cite{LPM00} attributed this correlation to synchrotron radiation as the dominant emission mechanism producing the prompt gamma-rays. Many other significant correlations have been seen in the GRB data, all of which offer clues into the physics at play, and which can potentially be linked to the underlying progenitor system (for an extensive and detailed summary of relations seen in GRB data and their physical implications, see the book by \cite{Dai19} and the following reviews: \cite{Dai17, Dai18, DA18}).
One can also examine whether the distributions of different observed variables suggest different classes, or different potential progenitors, of GRBs.\footnote{A classic example of this is the distribution of observed prompt gamma-ray duration \citep{Kouv93}. This distribution is highly bimodal and it was suggested early on that long GRBs ($>2s$) come from different progenitors than short GRBs ($<2s$). This indeed has been confirmed thanks to coincident gravitational wave emission indicating that short GRBs result from double neutron star mergers \citep{Ab17}.} Along these lines, inspired by the presence of radio-loud and radio-quiet AGN, \cite{HGM13} first suggested that there may be a population of intrinsically radio quiet GRBs - GRBs for which very little radio emission is intrinsically present. \cite{LR17} and \cite{LR19} tested this idea and, looking in detail at the data and accounting for observational selection effects, found distinct differences between GRBs with and without radio afterglows. Those with radio afterglows are significantly longer in their prompt duration and more energetic than those without, potentially suggesting two different populations.
In this paper, we consider the possibility that radio loud lGRBs come from the collapse of a massive star that is in an interacting binary system. The binary interaction allows for: 1) the high angular momentum needed to create a long-lived accretion disk and launch a jet, 2) a potentially higher energy budget both from the jet power's dependence on angular momentum and mass gained from a companion, and 3) a dense and extended circumburst environment, in which the GRB jet is more likely to produce the bright radio afterglow. Additionally, we consider the fraction of stars thought to be in binary systems and compare this to the fraction of radio loud GRBs.
An lGRB occurring from the collapse of a massive star in a binary system has been considered in the past \cite[e.g][]{Iv02,vdHy07,BK10,Pod10, CSE20}. In particular, \cite{vdHy07} and \cite{BK10} consider a massive star collapse in a close binary system with a compact object companion. The latter show that the GRB progenitor can have significantly higher mass compared to a single star system, sufficient spin to launch a GRB jet, and in some cases a very long lived accretion disk. \cite{CSE20} consider binary progenitors for long gamma-ray bursts and show that not only can binary systems provide the necessary stripping of the stellar envelope (see also \cite{Pod10, Lap20, FW22}) and sufficient angular momentum to launch a jet, but also, from their population synthesis models, can broadly accommodate the rates of lGRBs.
Recently, \cite{Bav21} examined massive stars collapsing in binary systems where both stars end their lives as black holes (and therefore are the progenitors of binary black hole mergers observed by LIGO and Virgo). Modelling three types of binaries - a) close binaries with a common envelope phase, b) wider binaries with stable mass transfer, and c) close binaries with chemically homogeneous evolution - they found that a star collapsing in systems corresponding to cases a) and c) can produce a highly spinning black hole that launches a GRB jet. Furthermore, their rate estimates, based on LIGO/Virgo's binary black hole gravitational wave observations, are consistent with these systems producing a large fraction (the majority) of lGRBs.
\begin{figure*}[!t]
\includegraphics[width=2.3in]{CumdistTint2.png}\includegraphics[width=2.3in]{CumdistEiso2.png}\includegraphics[width=2.3in]{CumdistRed2.png}\\
\includegraphics[width=2.3in]{Tinthisto2.png}\includegraphics[width=2.3in]{Eisohisto2.png}
\includegraphics[width=2.3in]{Redhisto2.png}
\caption{Cumulative (upper row) and differential (lower row) distributions of intrinsic GRB duration $T_{int}$ (left panel), isotropic equivalent energy $E_{iso}$ (middle panel), and redshift $(1+z)$ (right panel) for radio bright (dotted blue line) and radio quiet (solid green line) GRBs. A KS test indicates the radio loud and quiet samples have significantly different distributions in both $T_{int}$ and $E_{iso}$, but not $(1+z)$. From \cite{LR19}}
\label{fig:cumdist}
\end{figure*}
In this paper, we argue that radio loud lGRBs are a result of the collapse of a massive star in an interacting binary system, while the radio quiet GRBs have single star progenitors (or exist in non-interacting binaries or multiples). Our paper is organized as follows: In \S 2, we summarize the observational results that suggest that radio loud and radio quiet lGRBs come from distinct populations. In \S 3, we discuss how binary interaction can affect the duration, energy budget and circumburst medium of an lGRB, and accommodate the trends we see in the radio loud and radio quiet data. In \S 4, we discuss the rates of binary systems and show that lGRBs with radio afterglows are consistent with these rates. In \S 5, we present our conclusions.
\begin{deluxetable}{lccc}
\tablecaption{Radio Bright vs Radio Dark Samples}
\tablecolumns{4}
\tablewidth{\linewidth}
\tablehead{Sample & $\bar{z}$ & $\bar{T}_{int}/s$ & $\bar{E}_{iso}/10^{52}$ erg}
\startdata
Radio Bright (78 bursts) | & 2.8 & 39. & 51. \\
Radio Dark (41 bursts) | & 2.6 & 16. & 9. \\
\enddata
\tablecomments{ Average values of the redshift $z$, intrinsic prompt duration $T_{int}$, and isotropic emitted energy $E_{iso}$ for the sample of energetic ($E_{iso} > 10^{52}$ erg) bursts with and without detected radio afterglows. A Student's $t-$test gives a probability of less than $.001$ that these samples are from the same distribution, based on the average values of $T_{int}$ and $E_{iso}$.}
\label{}
\end{deluxetable}
\section{Properties of Radio Loud versus Radio Quiet GRBs}
Based on the suggestion by \cite{HGM13} that there may exist a truly radio dark population of GRBs, \cite{LR17} and \cite{LR19} analyzed a sample of energetic lGRBs for which radio follow-up was attempted. Their data are primarily taken from \cite{CF12}, although their second paper adds to the sample utilizing data found on Jochen Greiner's GRB page\footnote{\url{2 http://www.mpe.mpg.de/ jcg/grbgen.html}}. After taking steps to account for detector selection effects, they found somewhat suprising results: radio bright GRBs tend to be significantly longer (by more than a factor of 2) in their prompt gamma-ray duration and have significantly higher isotropic energy than radio dark GRBs (we note they found no correlation between isotropic energy and measured radio flux in the radio bright subsample, suggesting that this is not simply a flux-limit effect). Importantly, the redshift distributions of the radio bright and radio dark samples are {\em not} statistically different. These results were recently confirmed by \cite{Zhang21}, and summarized in Table 1.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{Tintz.png}
\caption{Intrinsic duration versus redshift for the radio bright sample. Even when accounting for selection effects that artificially truncate the upper right and lower left of this plane, we find a statistically significant ($> 4 \sigma$) anti-correlation. The correlation is not present in the radio quiet sample.}
\label{fig:my_label}
\end{figure}
\cite{LR19} also found that the prompt duration for the radio bright sample is anti-correlated with redshift in only the radio bright sample (see Figure 2; this was confirmed by \cite{Dain21} in a sub-sample of radio bright GRBs with plateau-like features). Additionally, they found that {\em extended} very high energy emission (VHE; above 100MeV) is present only in the radio bright sample. These results require further confirmation, but are also suggestive of two different sources/environments for the radio bright and dark sub-samples. For example, \cite{Huang21} showed that strong wind environment suppresses VHE emission (and/or it decays faster) and the presence of VHE is more likely in a ISM-like medium.
We note they found no apparent difference between X-ray and optical emission between the radio bright and dark samples, nor in their jet opening angles; a more detailed analysis of the differences between radio bright and dark samples is in progress (Chakraborty et al., in prep).
Therefore, because observables related to both the inner engine (duration and isotropic energy) and environment (radio flux) of the GRB appear to be significantly different between the radio quiet and radio loud samples, these results suggest that these populations could be coming from different progenitor systems. Here, we consider in broad terms whether the radio loud sample may be a result of a GRB occurring in an interacting binary system.\footnote{We emphasize that we are considering the collapse of a star in a binary system in which interaction with the companion is significant, but not necessarily a merger as the catalyst for the GRB.} \\ \\
\section{General Characteristics of GRBs in Binary vs Single Star Systems}
We consider the expected prompt duration, energy budget and circumburst medium profile of a massive star in an interacting binary system. \\
\subsection{Angular Momentum and GRB Duration}
The details of the particle acceleration and radiation processes that lead to the highly variable prompt gamma-ray emission of a GRB are still a matter of investigation, but it is believed this emission results from {\em internal} dissipation processes, like internal shocks or magnetic reconnection, in the jet as opposed to an external shock \cite[for a comprehensive review summarizing the arguments for this, see][]{Pir04}. That means the duration of the gamma-ray emission is constrained by the lifetime of the jet, which itself depends on the amount of time the accretion disk sufficiently feeds the black hole and/or the jet launching mechanism is active. One can estimate this timescale by considering the amount of material supplied to the central engine (i.e. the amount of mass in the disk) divided by the average rate at which that mass is accreted on to the central black hole.
\begin{equation}
T_{int} \approx M_{{\rm disk}}/\dot{M}_{\rm disk}
\end{equation}
\noindent where $T_{int}$ is the intrinsic prompt gamma-ray duration, $M_{{\rm disk}}$ is the mass in the disk and $\dot{M}_{\rm disk}$ is the average accretion rate.
These variables - the amount of mass in the disk and the (complicated, ever-changing) accretion rate are difficult to determine. A disk will form only if there is enough specific angular momentum, $l$, in the gas (angular momentum per unit mass), $l > l_{critical} \sim 2R_{g}c$, where $R_{g}$ is the gravitational radius of the central black hole \cite[see, e.g., the descriptions in][]{PB03,Pr06,JP08}, and a system with more angular momentum has the ability to create a more massive disk\footnote{Again, here we consider a a black hole central engine but see, e.g., \cite{KB07} and \cite{Zou19} for discussion of a magnetar central engine.}. That is, more of the collapsing stellar mass can be centrifugally supported in a disk as opposed to collapsing directly to the central black hole\footnote{ \cite{JP08} showed this mass can be estimated as $M_{disk} \approx l_{max}c/4G - M_{core}$, where $l_{max}$ is the maximum angular momentum in the collapsing gas and $M_{core}$ is the core mass of the collapsing massive star.}.
Additionally, the accretion rate is in general lower for higher angular momentum systems - a more highly spinning disk will accrete at a lower rate than a lower angular momentum disk \cite[for a look at this effect in both stellar mass and supermassive black hole-disk systems, see, e.g.,][ respectively]{PB03, RG15}. As is clear from equation 1 above, both of these effects of a higher angular momentum system - that is, a more massive $M_{disk}$ and lower accretion rate $\dot{M}_{disk}$ - serve to increase the duration, $T_{int}$, of lGRB.\\
We can make an order of magnitude estimate of how duration depends on the amount of angular momentum in our GRB system as follows: We assume the GRB central engine is a black hole-accretion disk system, which serves to launch the GRB jet.
For a solid rotating disk, we can write the mass of the disk as:
\begin{equation}
M_{\rm disk} \sim 2 J R_{disk}^{-2} \omega^{-1}
\end{equation}
\noindent where $J$ is the angular momentum, $R_{disk}$ is the radius of the disk and $\omega$ is its average rotational velocity. A rough but informative estimate of the accretion rate of the disk is given by the mass in the disk divided by a dynamical timescale, $\dot{M} \sim M_{\rm disk}/t_{\rm dyn}$. For a disk with angular momentum $J$ and mass $M_{\rm disk}$ rotating around a black hole of mass $M_{BH}$ at an average rotational velocity of $\omega$, the dynamical time can be written as:
\begin{equation}
t_{dyn} \sim J^{3/4}M_{\rm disk}^{-3/4}\omega^{-3/4}G^{-1/2}M_{BH}^{-1/2}
\end{equation}
\noindent where $G$ is Newton's gravitational constant. Then the accretion rate can be estimated as:
\begin{equation}
\dot{M} \sim J^{-3/4}\omega^{3/4}M_{\rm disk}^{7/4}M_{BH}^{1/2}G^{1/2}
\end{equation}
\noindent so that the intrinsic duration of the gamma-ray burst, which - again - is given by the lifetime of the disk that sustains the GRB jet, is related to angular momentum as:
\begin{equation}
T_{int} \approx M_{{\rm disk}}/\dot{M}_{\rm disk} \propto J^{7/4}
\end{equation}
This simplistic order-of-magnitude estimate aligns with the basic intuition that a system with more angular momentum should indeed lead to a longer duration GRB. We now discuss angular momentum in single star vs. binary systems. \\
\subsubsection{Prompt Duration in Single Star Systems}
Because the angular momentum of single massive stars at the end of their lives is a huge unknown \citep{Heg05,Heg06,Ros12,Sund13}, there are a wide range of possibilities for what the expected value of the intrinsic duration might be in a single massive star model. \\
\cite{JP08} looked at a number of models for the distribution of angular momentum, as well as three distinct accretion scenarios (depending on how gas in the polar region is treated), for the gas in a single massive star collapse. They show, in these single star systems, that the total mass that forms a disk is usually only a small fraction of the stellar envelope. From the amount of material that forms a disk, and using two different accretion rate assumptions (a constant between $0.1$ - $1.0 M_{\odot}$, and a variable accretion rate based on the free fall velocity of the gas in the torus), they estimate the duration of a single star collapsar producing a GRB, and show that most collapsars can produce GRBs with durations {\em at most} of about $50s$. Even this duration required somewhat extreme conditions, indicating single star collapsars producing GRBs are not only indeed very unique, but can only produce lGRBs {\em on the lower end of the prompt duration distribution}, consistent with the average prompt durations ($\sim 16s$) that \cite{LR19} found for radio quiet GRBs. \\
\subsubsection{Prompt Duration in Binary Systems}
In a binary system, on the other hand, we
may have significant angular momentum transfer from the companion star that can in principle spin up the GRB collapsar and create a higher angular momentum system. Stars in tidally locked binaries systems will be spun up (or down) until the spin angular momentum is roughly the same as the orbital velocity \citep{Izz04}. For the spin values needed for a successful GRB (black hole spin parameters $a \gtrsim 0.5$), this requires relatively short orbital periods ($<$ 10hr), which in turn implies a compact object companion \citep{Pod04, Jan13}. Such systems have been shown to lead to GRBs that can last hundreds to even thousands of seconds \citep{BK10}. \cite{Bav21} recently examined binary systems that are progenitors of lGRBs, and showed that two of the three formation channels they consider (common envelope and chemically homogeneous evolution) lead to systems with significant tidal interaction and angular momentum increase. The high angular momenta of the cores of the progenitor stars in their systems are preserved until the collapse of the second massive star, even when angular momentum transport is efficient. \\
As pointed out above and discussed further below, we consider interacting binary systems in cases where the companion star is a compact object in close orbit; this companion imparts significant angular momentum to the primary star (the massive star that will collapse and form a GRB). For our analysis below, we consider the case where the compact object companion is a black hole; however, our analysis is readily extended to a neutron star or white dwarf companion as well.\\
The total angular momentum in massive star-black hole binary system is given by:
\begin{equation}
J = \left(\frac{Gr(M_{BH}M_{star})^{2}}{M_{tot}} \right)^{1/2} + I_{BH}\omega_{BH} + I_{star}\omega_{star}
\end{equation}
\noindent where $M_{tot} = (M_{BH} + M_{star})$ is the total mass of the binary system, $r$ is the semi-major axis, $I_{BH}$ and $I_{star}$ are the moments of inertia of the black hole and star respectively, and $\omega_{BH}$ and $\omega_{star}$ are the rotational velocities of the black hole and star respectively.\\
We consider close, interacting binary systems in which tidal locking will occur on short timescales compared to the lifetime of the massive star that collapses to create the GRB. The tidal locking timescale is given by \citep{GLAD96}:
\begin{equation}
t_{lock} \approx \frac{16 r^{6} \rho_{star} \omega_{o} Q}{45 G M_{BH}^{2} k_{2}}
\end{equation}
\noindent where $\rho_{star}$ is the average density of the star, $\omega_{o}$ is its initial rotational velocity, $Q$ is the so-called dissipation function of the star, and $k_{2}$ is the tidal love number. Quantities like $Q$ and $k_{2}$ are highly unknown \citep{GLAD96} and usually taken to be around $\sim 100$ and of order unity, respectively. This expression, however, is most sensitive to the orbital separation $r$ of our binary system and it becomes clear why close binary systems are necessary for tidal locking to occur on a timescale less than the lifetime of the massive star. Taking high mass X-ray binary systems as a guiding point for the values in the expression above, we find:
\begin{equation}
\begin{aligned}
t_{lock} \approx & 10^{3} \ {\rm yr} \ (r/10^{12}cm)^{6} (\rho_{star}/10^{3}g cm^{-3}) \\
& (\omega_{o}/10^{-5}s^{-1}) (Q/100) (M_{BH}/10M_{\odot})^{2} (k_{2}/0.1)
\end{aligned}
\end{equation} \\
\noindent indicating that for closely separated binaries, tidal locking can occur on short timescales compared to the life of the massive star.
In this case, when the system becomes tidally locked, the angular momentum of the massive star in the binary system that will create the GRB upon collapse is given by:
\begin{equation}
J_{star, binary} = \alpha M_{star} R_{star}^{2} \left(\frac{G M_{tot}}{r^{3}} \right)^{1/2}
\end{equation}
\noindent where we have estimated the moment of inertia of the star to be $I_{star} \sim \alpha M_{star} R_{star}^{2}$. Ultimately, what we are doing here is comparing the rotation rate, $\omega_{star} = \sqrt{G M_{tot}/r^{3}}$, of a massive star in a tidally locked close binary system to that of a single massive star. Once again, the rotation rates of single massive stars can take on a range of values and are very unknown \citep{Heg05,Heg06}, but several works \citep{Ros12,Sund13} suggest periods on the order of days to tens or even hundreds of days. \\
Using the fiducial values for our binary system as above, and conservatively taking the rotation rate of a single massive star to be $\sim days$, we find
\begin{equation}
\begin{aligned}
& J_{binary}/J_{single} = \omega_{binary}/\omega_{single} \\
& \approx 5 \ (M_{tot}/20M_{\odot})^{1/2} (r/10^{12}cm)^{-3/2}
\end{aligned}
\end{equation}
Because a binary system has in principle much more angular momentum available and will, in cases of close binaries with a compact companions, spin up the massive star that will make a GRB on a timescale shorter than the lifetime of the massive star, it can sustain a disk (and therefore a jet) much longer compared to a single star system. This allows for a longer duration gamma-ray burst. \\
To summarize, the amount of mass present in an accretion disk of a collapsing star and rate of accretion onto the black hole both fundamentally depend on the angular momentum in the system. Although the details are complicated (and depend on the specifics of stellar structure and gas microphysical processes), in general we expect stars with larger angular momentum to be able to support more massive disks a lower accretion rate. Because massive stars in interacting binary systems can have significantly higher angular momentum than in single star systems, they can form a more massive, longer-lived accretion disk upon collapse, leading to a longer duration lGRB.\\
\subsection{Jet Power and Energy Budget}
A star in an interacting binary system can gain mass from its companion throughout its lifetime. Indeed, there is evidence that Type II SNe in binary systems end up with larger core mass than their single-star cousins \citep{BK10,Zap21}.
The right panel of Figure 1 in \cite{Zap21} shows that that core mass of a star in a binary system that produces a TypeII SNe is roughly one and half to two times larger than that of a single star.
Gain in mass from a companion provides an overall larger (gravitational potential energy) budget in these systems, which may explain the higher values of inferred higher isotropic energies in radio loud lGRBs. However, these apparently larger isotropic energies may also result from the larger angular momentum in a binary system (as discussed in the previous section), leading to a more powerful jet. Let us consider the jet power's dependence on angular momentum. If we assume that the GRB jet is magnetically driven (launched by, for example, the Blandford-Znajek (BZ) mechanism \citep{BZ77}, which has been shown to be a more efficient launch mechanism for an lGRB, compared to, say, neutrino annihilation), the luminosity of the jet is given by $L_{BZ} = (kf/4\pi c) \phi^{2} \omega^{2}$, where $k$ is a geometrical factor related to the magnetic field geometry (of order $\sim 0.05$), $\omega$ is the angular velocity of the spinning black hole, $f$ is a factor of order unity, and $\phi$ is the magnetic flux on the black hole. We can re-write this luminosity in terms of the the black hole dimensionless spin parameter $a = Jc/GM_{BH}^{2}$ ($J$ being the black hole angular momentum) and black hole mass $M_{BH}$:
\begin{equation}
\centering
\begin{aligned}
L_{BZ} & = (kfc^{5}/64\pi G^{2})a^{2}\phi^{2}M_{BH}^{-2}
\end{aligned}
\end{equation}
The jet power depends strongly on the spin of the black hole, the magnetic flux, and the mass of the black hole, although we note the complicated interplay between all three of these variables (for example, as discussed in the previous section, the angular momentum of the system will affect how much mass initially collapses into the black hole vs forming a centrifugally supported disk). We can simplistically write down, under the assumption of flux conservation, the magnetic flux in terms of magnetic field strength $B$ and mass of the black hole\footnote{ $\phi \approx 4 \pi B R^{2}$, where $B$ is the magnetic field, and $R$ is the Kerr radius given by $R = GM/c^{2} + \sqrt{(GM/c^{2})^{2} - a^{2}}$}, which leads us to the following expression for the observed jet luminosity \citep{mck05, TNM10, TG15, LR19bz}:
\begin{equation}
\centering
\begin{aligned}
& L_{GRB} \approx \\
& 10^{50} {\rm erg} (\eta/0.1) (a/0.9)^{2}(B/10^{16}{\rm G})^{2}(M_{BH}/5M_{\odot})^{2}
\end{aligned}
\end{equation}
\noindent and where we have used an efficiency factor $\eta$ between the BZ jet power and the observed jet luminosity, $L_{GRB} = \eta L_{BZ}$. The measured isotropic energy is a measure of the integral of this luminosity over the duration of the gamma-ray emission (the frequency at which an lGRB in general emits most of its energy). Therefore, more highly spinning, massive systems can lead to higher measured isotropic energies, as seen in the radio loud lGRB sample\footnote{The relationship between a massive star's angular momentum and the final resultant spin of the black hole is certainly not straightforward \citep{Heg06}. We are operating under the reasonable assumption, however, that stars with higher angular momentum will on average lead to black holes with higher spin values.}.\\
Because the true energy emitted in the jet, $E_{\gamma}$, is related to the isotropic energy as $E_{\gamma} = \frac{1}{2}(1-{\rm{cos}}(\theta_{j}))E_{iso}$, we also consider the role of the jet beaming angle, $\theta_{j}$, plays in our understanding of higher isotropic energies for radio loud lGRBs. A larger $E_{iso}$ may not necessarily indicate a higher energy budget but, instead, a narrower beaming angle. A narrower jet may indeed occur in systems with more angular momentum (Hurtado et al. in prep) and in systems with denser stellar envelopes \cite[i.e. more massive, lower metallicity stars, as shown in][]{LR20}.
However, in our data sample, we do {\em not} find a difference between the distribution of the beaming angles between the radio loud and quiet samples \cite[the measured beaming angles of GRB jets, when available, can be found in the metadata table compiled by][]{Wang20}, although we note the numbers of lGRBs with jet opening angle measurements in our samples are small and accounting for the complicated selection effects when measuring beaming angle is difficult \cite[a more detailed description of can be found in][]{LR20}. Nonetheless, the data currently suggest that the higher isotropic energy measurement in radio loud lGRBs is not a beaming angle effect, but an indication of a truly higher energy/power budget in these systems.
\subsection{Circumburst Media and Radio Light Curves}
\begin{figure*}
\centering
\includegraphics[width=0.46\textwidth]{RadISMWind1to50GHz.png} \includegraphics[width=0.46\textwidth]{RadISMWind6to9GHz.png} \\
\includegraphics[width=0.46\textwidth]{RadISMWind400to500GHz.png} \includegraphics[width=0.46\textwidth]{RadISMWind.6to600GHz.png}
\caption{Representative radio light curves in different radio frequency ranges, assuming synchrotron radiation, for jets traversing wind (green dash and dash-dot lines) and ISM-like (blue solid and dotted lines) media. In both cases (ISM and wind) we assume an isotropic energy of $10^{52}$ ergs, an electron energy distribution power law index $p=2.5$, a magnetic energy fraction of $\epsilon_{B} = 0.01$, an electron energy fraction of $\epsilon_{e} = 0.1$, and a redshift $z=1$. For the ISM curves, we assume a constant density of $n_{o} = 50 \ cm^{-3}$ (solid blue line) and $n_{o}=10 \ cm^{-3}$ (dotted blue line). For the wind curves, we assume a wind normalization $A_{*}$ (where $\rho = A_{*}r^{-2}$) of $A_{*} = 5 \times 10^{12} g/cm$ (green dashed-dot curve) and $A_{*} = 5 \times 10^{11} g/cm$ (green dash curve). }
\label{fig:my_label}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.46\textwidth]{1.5GHzFluxDensity2.png} \includegraphics[width=0.46\textwidth]{3GHzFluxDensity2.png} \\
\includegraphics[width=0.46\textwidth]{6GHzFluxDensity.png} \includegraphics[width=0.46\textwidth]{200GHzFluxDensity2.png}
\caption{Flux density (mJy) as a function of time for typical radio observational frequencies, assuming synchrotron radiation, for jets traversing wind (green dash and dash-dot lines) and ISM-like (blue solid and dotted lines) media. In both cases (ISM and wind) we assume an isotropic energy of $10^{52}$ ergs, an electron energy distribution power law index $p=2.5$, a magnetic energy fraction of $\epsilon_{B} = 0.01$, an electron energy fraction of $\epsilon_{e} = 0.1$, and a redshift $z=1$. For the ISM curves, we assume a constant density of $n_{o} = 50 \ cm^{-3}$ (solid blue line) and $n_{o}=10 \ cm^{-3}$ (dotted blue line). For the wind curves, we assume a wind normalization $A_{*}$ (where $\rho = A_{*}r^{-2}$) of $A_{*} = 5 \times 10^{12} g/cm$ (green dashed-dot curve) and $A_{*} = 5 \times 10^{11} g/cm$ (green dash curve). }
\label{fig:my_label}
\end{figure*}
Interacting binaries can go through a phase in their evolution that can create a complicated, enhanced environment around the system \citep{Bate00, Olof15, Lines15, Lap20, Schr21}, compared to single star systems in which a massive star's wind is expected to produce a roughly $1/r^{2}$ density profile around the star \citep{CH00}. \cite{Cast21} show, both analytically and with the code {\em Guacho}, that extended, dense shells can form as a result of the interacting winds around two giants in a binary system. Similarly, \cite{Schr21} show how winds from a star in an interacting binary can produce a a dense, extensive gas structure around the system, to large radii. \cite{Bate00} show that the density profiles in interacting binaries can take on a wide range of profiles, lying between single star wind profiles and uniform density profiles, but more often than not resembling the latter (we note, however, they are examining the circumbinary medium at the formation of the binary system). \\
Further detailed modelling of density profiles around stars in interacting binary systems with at least one massive star companion (i.e. relevant to GRB progenitors) is needed, as there are many system variables to examine \citep{Olof15}. However, it is clear that deviations from a simple $1/r^{2}$ density profile are not unexpected in interacting binaries, and indeed a profile that may appear like a relatively dense uniform density profile (in terms of inferring this profile from the observed GRB emission) is reasonable (see the papers by, e.g., \cite{NG07} and \cite{Gat13}, who show that modest circumburst density fluctuations will not create large fluctuations in the observed GRB light curve; this can lead to spectral and light curve fits that give an apparent constant density profile for the circumburst medium). \\
Now, consider a gamma-ray burst occurring in a constant density medium vs. a wind-like medium. The {\em peak} of the radio spectrum at the characteristic synchrotron frequencies depends of course on the circumburst density profile. The peak flux at the synchrotron self-absorption frequency $\nu_{sa}$ (the frequency at which the system is optically thick to synchrotron photons) and the so-called the minimum electron frequency $\nu_{m}$(the frequency that corresponds to the peak of the electron energy distribution) are proportional to the values of the density (or wind normalization density factor) and time as follows \citep{Pach70,SPN98,GS02}: \\
\begin{equation}
\Bigg\{
\begin{aligned}
& F_{p,\nu_{sa}, ISM} \propto n^{7/10} t_{day}^{1/2} \\
& F_{p,\nu_{m}, ISM} \propto n^{1/2}t_{day}^{0}
\end{aligned}
\end{equation}
\begin{equation}
\Bigg\{
\begin{aligned}
& F_{p,\nu_{sa}, wind} \propto A_{*}^{7/5}t_{day}^{-1/2} \\
& F_{p,\nu_{m}, wind} \propto A_{*}t_{day}^{-1/2}
\end{aligned}
\end{equation}
\noindent where $n_{o}$ is a constant density in units of $cm^{-3}$ and $A_{*}$ is the normalization in units of $g/cm$ for a wind like density profile, $\rho(r) = A_{*}r^{-2}$. \\
These proportionalities illustrate the more rapid decay of the peak of the lGRB radio spectrum in a wind-like medium, compared to an ISM-like medium. However, to get a better sense of the relative strength (and detectability) of radio emission from a wind or ISM-like medium, we need to compute the full light curves over the relevant observational frequency range. In Figure 3, we do just that: we have computed the flux in the radio band as a function of time, over a range of frequencies (that span ranges over which most afterglow searches are performed) for both ISM and wind-like media. We keep careful track of the relative values of the synchrotron self-absorption frequency and minimum electron frequency, throughout the evolution of the light curve \cite[see Figure 1 and Table 2 of][]{GS02}. \\
We use the following expressions for the approximate synchrotron spectrum across different frequency ranges in the ISM and wind cases \citep{GS02}:
\begin{equation}
\begin{aligned}
A_{ISM} = & 1.18 \times 10^{8} mJy \ (4.59 - p) (1+z)^{9/4} \\
& \epsilon_{B}^{-1/4} n^{-1/2} E_{52}^{1/4} t_{days}^{5/4}d_{L, 28}^{-2} \nu_{14}^{5/2} \\
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
B_{ISM} = & 4.20 \times 10^{9} mJy \ \frac{3p+2}{3p-1} (1+z)^{5/2} \\
& \bar{\epsilon_{e}} n^{-1/2} E_{52}^{1/2} t_{days}^{1/2}d_{L, 28}^{-2} \nu_{14}^{2} \\
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
D_{ISM} = & 27.9 mJy \ \frac{p-1}{3p-1} (1+z)^{5/6} \\
& \bar{\epsilon_{e}}^{-2/3} \epsilon_{B}^{1/3} n^{-1/2} E_{52}^{5/6} t_{days}^{1/2}d_{L, 28}^{-2} \nu_{14}^{1/3} \\
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
G_{ISM} = & 0.461 mJy \ (p-.04)e^{2.53p} (1+z)^{(3+p)/4} \\
& \bar{\epsilon_{e}}^{p-1} \epsilon_{B}^{(1+p)/4} n^{-1/2} E_{52}^{(3+p)/4} t_{days}^{3(1-p)/4} \\
& d_{L, 28}^{-2} \nu_{14}^{(1-p)/2} \\
\end{aligned}
\end{equation} \\
\noindent where $z$ is redshift, $p$ is the index of a power-law energy distribution for the radiating electrons, $n$ is density in units of $cm^{-3}$, $\epsilon_{B}$ is the fraction of energy in the magnetic field, $\bar{\epsilon_{e}}$ is the fraction of energy in the electrons, $\epsilon_{e}$, times the factor $(p-2)/(p-1)$, $E_{52}$ is the energy in units of $10^{52}$ ergs, $t_{day}$ is observation time in units of days, $d_{L, 28}$ is the luminosity distance in units of $10^{28}cm$, $\nu_{14}$ is the frequency in units of $10^{14}Hz$. \\ \\
In the wind circumburst medium case we have:
\begin{equation}
\begin{aligned}
A_{wind} = & 2.96 \times 10^{7} mJy \ (4.59 - p) (1+z)^{7/4} \\
& \epsilon_{B}^{-1/4} A_{*}^{-1} E_{52}^{3/4} t_{days}^{7/4}d_{L, 28}^{-2} \nu_{14}^{5/2} \\
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
B_{wind} = & 1.33 \times 10^{9} mJy \ \frac{3p+2}{3p-1} (1+z)^{2} \\
& \bar{\epsilon_{e}} A_{*}^{-1} E_{52} t_{days}d_{L, 28}^{-2} \nu_{14}^{2} \\
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
D_{wind} = & 211 mJy \ \frac{p-1}{3p-1} (1+z)^{4/3} \\
& \bar{\epsilon_{e}}^{-2/3} \epsilon_{B}^{1/3} A_{*} E_{52}^{1/3} t_{days}^{0}d_{L, 28}^{-2} \nu_{14}^{1/3} \\
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
G_{wind} = & 3.82 mJy \ (p-.18)e^{2.54p} (1+z)^{(5+p)/4} \\
& \bar{\epsilon_{e}}^{p-1} \epsilon_{B}^{(1+p)/4} A_{*} E_{52}^{(1+p)/4} t_{days}^{(1-3p)/4} \\
& d_{L, 28}^{-2} \nu_{14}^{(1-p)/2} \\
\end{aligned}
\end{equation} \\
\noindent where, again, $A_{*}$ is the wind density profile normalization factor in units of $g/cm$.\\ \\
To compute the radio light curve , then, our prescription is as follows: In the case when the self-absorption frequency is greater than the minimum electron frequency, the flux between observed frequencies $\nu_{1}$ and $\nu_{2}$ for an ISM/wind medium is given by:
\begin{equation}
F = \int_{\nu_{1}}^{\nu_{sa}} B_{k} d\nu + \int_{\nu_{sa}}^{\nu_{m}} D_{k} d\nu + \int_{\nu_{m}}^{\nu_{2}} G_{k} d\nu
\end{equation}
\noindent where $k$ indicates whether we are in an ISM ($k=0$) or wind ($k=2$) medium (this is based on how density varies with radius, $\rho \propto r^{-k}$), in the respective media.\\
When the self-absorption frequency $\nu_{sa}$ is less than the minimum electron frequency, we have:
\begin{equation}
F = \int_{\nu_{1}}^{\nu_{sa}} B_{k} d\nu + \int_{\nu_{sa}}^{\nu_{m}} A_{k} d\nu + \int_{\nu_{m}}^{\nu_{2}} G_{k} d\nu
\end{equation}.
\noindent To compute the flux densities (flux per unit frequency) shown in Figure 4, we omit the integration and evaluate the $A, B, C$, and $D$ expressions at a particular relevant observational frequency.
The above expressions assume the physical break frequencies are within the observed band; i.e. $\nu_{1}< \nu_{sa} < \nu_{m} < \nu_{2}$ for equation 21 and $\nu_{1}< \nu_{m} < \nu_{sa} < \nu_{2}$ for equation 22. Of course if the physical break frequencies fall outside the observed band, we adjust the limits of our integral accordingly. \\
In Figure 3, we show light curves for systems with an isotropic energy of $10^{52}$ ergs, an electron energy distribution power law index $p=2.5$, a magnetic energy fraction of $\epsilon_{B} = 0.01$, an electron energy fraction of $\epsilon_{e} = 0.1$, and a redshift $z=1$. For the ISM curves, we assume a constant density of $n_{o}=10 \ cm^{-3}$ (dotted blue line) and $n_{o} = 50 \ cm^{-3}$ (solid blue line). For the wind curves, we assume a wind normalization $A_{*} = 5 \times 10^{12} g/cm$ (green dashed-dot curve) and $A_{*} = 5 \times 10^{11} g/cm$ (green dash curve). Figure 4 shows the radio flux densities in at specific relevant observational frequencies. Certainly we have the freedom to adjust the normalization of these curves based on the values of the physical parameters we choose; however, we have chosen relatively conservative and typical values expected for GRBs, based on both theory and fits to their afterglow light curves \citep{vE13,Gomp18,KF21}.\\
It is clear that the constant density medium case, even for relatively modest densities, is brighter and longer-lasting that the wind density case, indicating a higher likelihood of detecting a radio afterglow in such an environment.\footnote{We note that fits to broadband light curves and spectra give inconclusive results for the density profiles. \cite{GFP18} fit a sample of lGRBs to both wind and ISM medium density models and did not find that radio loud GRBs were better fit by one density profile over the other. However, because of the simplicity of the density models and the degeneracy of density with other model parameters (as well as small number statistics; there were only 15 GRBs from our radio loud GRBs in their sample), it is difficult to draw definitive conclusions about the external density profiles from these fits. We consider this an open issue to be further explored. } \\
Our main point here is that, for fiducial values of intrinsic GRB properties, we expect brighter radio emission in systems with an extended ISM-like environment, which we suggest is more likely in an interacting binary system compared to a single massive star system.
\section{Rates}
In our sample of energetic lGRBs with radio follow-up, we find that roughly 60 \% of GRBs appear to be radio loud. Accordingly, from their radio image stacking analysis of lGRBs, \cite{HGM13} suggested that roughly 30 to 40 \% are truly radio quiet. In Figure 5, we plot the radio loud fraction of lGRBs as a function of redshift for five redshift bins. Although it appears to evolve somewhat (peaking at around $(1+z) \approx 2.5$), we caution that within the large error bars, this ratio is consistent with staying roughly constant over redshift. We would like to compare this - the fraction of lGRBs that are radio loud - to the binary star fraction (more specifically, the fraction of binary systems that can produce an lGRB).
Unfortunately, this question is plagued with many complicating factors. First of all, getting an accurate handle on overall binary rates, through both observations and binary population synthesis studies, is notoriously difficult \citep{Han20, DW20, Mazz20}. Furthermore, determining those massive stars in close, interacting binaries that would produce an lGRB has additional complexities associated with understanding the conditions necessary for a successful GRB.
Several studies \cite[e.g.][]{Sana12,Sana13} have put the binary star fraction at around 60 \%, with some studies \citep{MG09} showing rates closer to 75 \%. Although this percentage is interestingly similar to what we find for the fraction of radio loud GRBs, we again caution that directly comparing these percentages ignores the details involved in understanding the rates of those binary star systems that could create lGRBs. Nonetheless, if we can make the simplistic but reasonable assumption that the fraction of single stars that create lGRBs is roughly equivalent to the fraction of binary stars that create lGRBs (more on this below), then, combined with the other factors described in \S 3 for why interacting binary progenitors are consistent with the data from radio loud lGRBS, we find the similarity of the ratios suggestive enough to warrant further investigation of this issue.
Consider the rate of lGRBs ($\dot{dN}/dz)$ as it relates to the star formation rate $\dot{\rho}_{\rm SFR}(z)$ \citep{Kist08,Yuk08,Kist09, LR20}:
\begin{equation}
(\dot{dN}/dz)= \dot{\rho}_{\rm SFR}(z)/({\rm f_{beam}(z)}) \left(\frac{dV/dz}{(1+z)} \right) \epsilon(z)
\end{equation}
\noindent where $f_{beam}(z)$ is a factor ($>1$) that accounts for the number of GRBs missed due to GRB jets that are pointed away from our detectors, and $\epsilon(z)$ parameterizes the fraction of stars that make GRBs. Both of these functions can in principle depend on redshift (\cite{LR20} show that indeed $\rm f_{beam}(z)$ does). The factor $dV/dz$ is the cosmological volume element given by:
\begin{equation}
\begin{split}
dV/dz = & 4 \pi (\frac{c}{H_{o}})^{3} \bigg[\int_{1}^{1+z} \frac{d(1+z)}{\sqrt{\Omega_{\Lambda} + \Omega_{m}(1+z)^{3}}}\bigg]^{2} \\
& \times \frac{1}{\sqrt{\Omega_{\Lambda} + \Omega_{m}(1+z)^{3}}}
\end{split}
\end{equation}
\noindent where $\Omega_{m}$, $\Omega_{\Lambda}$ are the matter and cosmological constant density parameters, and the $H_{o}$ is the Hubble constant.
If we can assume that a function $\epsilon(z)$ is a sufficient way to account for how each population (i.e. radio loud or radio quiet) independently traces the star formation rate, and if we assume that the jet beaming angle and its evolution over cosmic time are roughly similar in both radio loud and quiet populations, then the fraction of radio loud lGRBs is the ratio of the integral of $\epsilon(z)$ for the radio bright sample relative to the whole population:
\begin{equation}
\frac{\int (\dot{dN}/dz_{RL})}{\int (\dot{dN}/dz)} = \frac{\int \epsilon_{RL}(z)}{\int\epsilon_{RQ}(z) + \int \epsilon_{RL}(z)} \approx 0.6 \pm 0.2
\end{equation}
The functions $\epsilon(z)$ are as yet undetermined for the progenitor models we suggest here (i.e. collapsars in interacting binaries and single star/non-interacting binary collapsars; again, this function is parameterizing the fundamental, underlying question of which stars produce GRBs and why). As we described above, to fairly compare the binary star fraction to the radio loud lGRB fraction (or, alternatively, the single star fraction to the radio quiet lGRB fraction), we need to assume that $\epsilon(z)$ is roughly similar for both radio loud and radio quiet progenitors. \\
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{FracRB.png}
\caption{Fraction of radio loud GRBs (relative to the total population), for those with attempted radio follow-up, as a function of redshift. Data are taken from the sample described in \cite{LR19}, who use a subset of the data published in \cite{CF12}.}
\label{fig:my_label}
\end{figure}
We again mention the simulations of \cite{Bav21}, who indicate that common envelope (CE) and chemically homogeneous evolution (CHE) interacting binaries (ones that would apply to systems we are discussing here) add up to $= 71 \%$ ($57 \%$ (CE) $+ 14 \%$ (CHE)) of the binary systems they consider (note they argue binary systems can explain the rates of the entire population of lGRBs; this could suggest, as mentioned above, that the radio loud vs radio quiet is more a reflection of the type of binary interaction rather than binary vs. single star systems).
Finally, we comment that metallicity plays an important role in all of the considerations above. It has long been suggested that low metallicity is a necessary condition to produce a successful GRB jet from a collapsing star because it mitigates mass and angular momentum loss \citep{MW99,YL05,HMM05,Yoon06,WH06}, and indeed lGRBs tend to be found in lower metallicity galaxies (for a recent review summarizing these observational results, see \cite{Per16}). Meanwhile, the close binary fraction may be higher for massive and lower metallicity stars \citep{DM91, Bav21, Klen21}. \cite{Mazz20} show, using APOGEE data, an anti-correlation between the close binary fraction and [Fe/H] ratio, highlighting the necessity of low metallicity in producing these systems (see their Figure 5; similarly \cite{Hwang21} present recent results showing the wide binary fraction is positively correlated with metallicity). On the other hand, \cite{Liu21} show that low metallicity (specifically Pop III) binary systems are born wider. However, \cite{Det08} show that massive star winds widen binary systems so low metallicity is needed to achieve close binary systems. \cite{Schr21} show that either widen or tighten with metallicity depending on the velocity of the wind (relative to the orbital velocity) and the mass ratio of the binary. The detailed connection between metallicity, GRBs, and interacting binary systems is beyond the scope of this paper but is important factor in future work investigating these systems.
\section{Conclusions}
The radio emission from a long GRB is a reflection of a GRB jet's interaction with its surrounding external medium, while the duration of the prompt gamma-ray emission comes from internal dissipation processes in the jet and reflect (along with the energy budget) the system's inner engine. That lGRBs with radio afterglows appear have significantly longer gamma-ray durations and higher isotropic energies than those with no radio afterglows, suggests that radio loud and radio quiet lGRBs may originate from different progenitor systems. \\
We have argued broadly in this paper that these observational results can be explained if radio loud GRBs emanate from interacting binary systems, while radio quiet GRBs emerge from the collapse of a single star (or a massive star collapsing in a non-interacting binary, where we define interacting binary as one where the companion has significant influence on the primary star in terms of mass and angular momentum transfer). Our main arguments are:
\begin{itemize}
\item A massive star in a closely interacting binary can have significant angular momentum, due to spin up from its companion. This can produce both a more massive disk and lower accretion rate in the system, leading to a longer duration gamma-ray burst.
\item A massive star in a closely interacting binary may gain mass from its companion which can lead to a higher isotropic energy when the star collapses. The higher isotropic energies seen in the radio loud sample of lGRBs may also be a reflection of the system's higher angular momentum and its ability to provide more power to the jet, given a Blandford-Znajek scenario for the jet launch mechanism.
\item Radio emission is brighter in an extended, uniform medium. Closely interacting binary systems have the ability to create a more extended, dense circumbinary medium that may lead to brighter, longer-lived radio emission compared to a single star systems with a wind-like medium.
\item Although we acknowledge the complicating factors that prohibit a direct comparison, the ratio of binary systems to single star systems aligns with the ratio of radio loud to radio dark GRBs.
\end{itemize}
Additional clues, including the presence or absence of extended very high energy emission (shown to be more likely in extended constant density media and not wind-like media) as well as correlations specific to either one of the radio loud or quiet populations, such as the duration redshift anti-correlation shown in Figure 2, can further test our hypothesis. Correlations of the rates of radio loud lGRBs with that of high mass X-ray binaries may also offer additional clues (Upton-Sanderbeck, private communication). Nonetheless, as we accumulate and analyze more lGRB afterglow data, it is becoming clear that the radio emission in particular shows unusual behavior and has challenged our standard model of lGRBs. For example, recently, \cite{KF21,Mar21} and Dvornak et al (in prep) show that the standard synchrotron shock emission models do not fit the radio light curves well. \cite{Lev21} show that the break times and plateau-like behavior seen in radio afterglow light curves are unique compared to optical and X-ray plateaus.
As such, radio observations of lGRBs may offer us a unique and deeper insight into the underlying physics of these objects. Continued radio follow-up, more observational and theoretical work on the formation of binary and single stars, as well as a deeper investigation into the conditions necessary to produce a relativistic GRB jet, will allow us to better test the idea presented in this paper and further understand the fascinating processes behind lGRBs and their role in our universe.
\section{Acknowledgements}
We are very grateful to the referee for thoughtful and helpful commments that improved this manuscript. We thank Olivia Cantrell Rodriguez, Jarrett Johnson, Angana Chakraborty, Ken Luu, Roseanne Cheng, Maria Dainotti, Valeria Hurtado, Celia Tandon, Aycin Aykutalp, Phoebe Upton-Sanderbeck, Delina Levine and Kevin Zvonarek for lively and interesting discussions related to this work. Los Alamos National Laboratory is operated by Triad National Security, LLC, for the National Nuclear Security Administration of U.S. Department of Energy (Contract No. 89233218CNA000001). LA-UR-22-20471
|
\section{Introduction}
Strong interactions between charge and spin degrees of freedom are responsible for numerous
novel phases in strongly correlated electron materials. One particularly attractive class
of materials for investigating such effects are organic charge transfer salts \cite{Powell2011,Hotta2012,Dressel2020}.
These materials display phenomena such as unconventional superconductivity and spin liquid behaviour \cite{Shimizu2003,Powell2006}.
In addition to low temperature phenomena, at temperatures on the orders of tens of kelvins, slow and
glassy charge dynamics have been observed, particularly in the dielectric relaxation of the $\kappa$-(BEDT-TTF)$_2$X family of organic charge transfer
salts \cite{Lunkenheimer2015a,Tomic2015}. In the organic charge transfer salt $\kappa$-(BEDT-TTF)$_2$Cu$_2$(CN)$_3$, which shows spin liquid
behaviour at low temperatures, there is broad in-plane dielectric relaxation below $\sim 60$ K, and
out of plane relaxor-like dielectric response below about 60 K \cite{Abdel-Jawad2010}.
Glassy response has also been observed in the dielectric function of $\kappa$-(BEDT-TTF)$_2$Ag$_2$(CN)$_3$ \cite{Pinteric2018}
and glassy freezing of electrons at low temperatures has been suggested for $\kappa$-(BEDT-TTF)$_2$Hg(SCN)$_2$Br \cite{Hemmida2018}.
The origin of the electric dipoles that give rise to the relaxor ferroelectric behaviour in $\kappa$-(BEDT-TTF)$_2$X salts
is still an active area of investigation.
There is evidence for charge disproportionation in dimers in $\kappa$-(BEDT-TTF)$_2$Hg(SCN)$_2$Cl \cite{Drichko2014}
but Sedlmeier {\it et al.} \cite{Sedlmeier2012} did not find evidence for charge disproportionation in
$\kappa$-(BEDT-TTF)$_2$Cu$_2$(CN)$_3$, $\kappa$-(BEDT-TTF)$_2$Cu[N(CN)$_2$]Cl
or $\kappa$-(BEDT-TTF)$_2$Cu[N(CN)$_2$]Br. Pinteri\'{c} {\it et al.} \cite{Pinteric2014} have argued that optical
measurements of $\kappa$-(BEDT-TTF)$_2$Cu$_2$(CN)$_3$
are not consistent with local dipoles. Suggested mechanisms for glassy dynamics
that do not involve dipoles in dimers include domain walls between dimer Mott and charge ordered phases \cite{Fukuyama2017}
and a ``dielectric catastrophe'' \cite{Pustogow2021}.
At higher temperatures disorder in the conformational orientation of ethylene groups in BEDT-TTF \cite{Guterding2015}
has been implicated in glassy dynamics in $\kappa$-(BEDT-TTF)$_2$X materials \cite{Hartmann2015,Muller2015}.
Relaxor ferroelectric or glassy behaviour has also been seen in several
other families of organic charge transfer salts, such as $\beta^\prime$-Pd(dmit)$_2$ salts \cite{Abdel-Jawad2013,Fujiyama2018},
$\beta^\prime$-(BEDT-TTF)$_2$ICl$_2$ \cite{Iguchi2013,Muller2020}, $\theta$-(BEDT-TTF)$_2$RbZn(SCN)$_4$ \cite{Kagawa2013}
and $\theta$-(BEDT-TTF)$_2$CsZn(SCN)$_4$ \cite{Sato2014,Sato2016}, usually in the tens of kelvin temperature range.
A dielectric peak similar to those seen in relaxor ferroelectrics is also seen in $\alpha$-(BEDT-TTF)$_2$I$_3$ \cite{Lunkenheimer2015b,Ivek2017}.
Many of the $\kappa$-(BEDT-TTF)$_2$X salts such as
$\kappa$-(BEDT-TTF)$_2$Cu$_2$(CN)$_3$ can be described by a triangular lattice of dimers forming
a quarter-filled extended two dimensional Hubbard model with both intra- and inter-dimer interactions \cite{Hotta2010,Naka2010,Gomi2013}.
Quarter filled extended Hubbard models have been identified as enhancing
geometric frustration of charge degrees of freedom \cite{Merino2005} and
geometric frustration in charge ordering has been emphasized as a factor in leading to a
charge cluster glass with no long range order in $\theta$-(BEDT-TTF)$_2$RbZn(SCN)$_4$ \cite{Kagawa2013}.
In this paper we study a quarter-filled extended dimer Hubbard model with the same form as that proposed to
describe $\kappa$-(BEDT-TTF)$_2$Cu$_2$(CN)$_3$. Following Ref.~\cite{Hotta2010} we obtain the low energy theory in
the one electron per dimer limit, which can be written in terms of spin and dipole degrees of freedom.
Going beyond Ref.~\cite{Hotta2010}, we take into account the occupation of
next-nearest neighbour sites on the couplings in the effective Hamiltonian as was
done for an extended Hubbard model on a square lattice \cite{Farrell2013,Farrell2014}.
This leads to a distribution of couplings between spin and dipole degrees of freedom in the low-energy theory.
Rather than simulate the resulting model directly, which would be prohibitively computationally expensive,
we make a classical approximation, which gives a model of vector spins coupled to Ising dipoles and study this model with classical Monte Carlo
simulations. Our main result is that in this approximate model we find that glassiness in the charge degrees of freedom
occurs over a range of intra-dimer and inter-dimer nearest neighbour interaction strengths.
This paper is structured as follows: in Sec.~\ref{sec:model} we introduce the model we study and describe how we obtain
the low energy theory. In Sec.~\ref{sec:MC} we discuss our Monte Carlo simulations of the simplified model and the
results we obtain from our simulations. Finally, in Sec.~\ref{sec:disc} we discuss our results and conclude.
\section{Model}
\label{sec:model}
We consider a two dimensional extended Hubbard model of dimers on a triangular lattice, introduced by Hotta \cite{Hotta2010} for
$\kappa$-(BEDT-TTF)$_2$Cu$_2$(CN)$_3$.
We allow for both intra-dimer and inter-dimer hopping, on-site interactions and both intra-dimer and inter-dimer
nearest neighbour interactions. All hops and interactions are illustrated in Fig.~\ref{fig:model}.
We write the Hamiltonian as
\begin{equation}
H = H_T + H_U + H_V,
\label{eq:Ham}
\end{equation}
where $H_T$ is the hopping part of the Hamiltonian, $H_U$ is the on-site Hubbard interaction and $H_V$ is the
nearest neighbour interaction term. We now consider each term in detail.
Let $(x,y)$ label a dimer on the triangular lattice. The hopping part of the Hamiltonian is then
\begin{eqnarray}
H_T &=\sum_{\alpha}t_\alpha \sum_{(x,y), i}\sum_{(x',y'),j}^\prime \sum_{\sigma}c^{\dagger}_{(x',y'), j, \sigma} c^{\vphantom{\dagger}}_{(x,y), i, \sigma}, \nonumber \\
&
\label{eq:hopping}
\end{eqnarray}
where $i = 1,2$ labels the lattice sites on dimer $(x,y)$, $j = 1,2$ labels the lattice sites on dimer $(x',y')$, spin is
labelled by $\sigma = \uparrow, \downarrow$ and $\alpha = d, b, p ,q$ specifies the type of hopping, following the notation of
Ref.~\cite{Hotta2010}.
The restriction that $(x',y'),j$ are limited to nearest neighbour sites and dimers is denoted by the prime on the sum.
The operator $c^{\dagger}_{(x',y'), j, \sigma}$ creates an electron of spin $\sigma$ on lattice site $j$ of dimer $(x',y')$.
Similarly $c_{(x,y), i, \sigma}$ destroys an electron of spin $\sigma$ on lattice site $i$ of dimer $(x,y)$.
\begin{figure}[ht]
\includegraphics[width=8cm]{Figure1.pdf}
\caption{Triangular lattice of dimers used for the extended Hubbard model. The position of the central dimer is $(x,y)$. The $t_i$ denote different types of
hopping and the $V_i$ denote nearest neighbour interactions.}
\label{fig:model}
\end{figure}
The on-site interaction term is given by
\begin{equation}\label{eq:Hubbard_U}
H_U =U\sum_{(x,y)}\sum_{i}n_{(x,y), i, \uparrow} n_{(x,y), i, \downarrow},
\end{equation}
with the number operator on site $i$ of dimer $(x, y)$ for an electron of spin $\sigma$ given by $n_{(x,y), i, \sigma} = c^{\dagger}_{(x,y), i, \sigma}c_{(x,y), i, \sigma} $ and
the nearest neighbour interaction Hamiltonian is
\begin{equation}\label{eq:Hubbard_V}
H_V =\sum_{\alpha}V_\alpha \sum_{(x,y),i}\hspace{2mm}\sum_{(x',y'),j}{\vphantom{\sum}}'\sum_{\sigma, \sigma'}n_{(x',y'), j, \sigma} n_{(x,y), i, \sigma'}.
\end{equation}
Density functional theory calculations \cite{Kandpal2009,Nakamura2009,Nakamura2012}
have predicted that the number of free electrons in $\kappa$-(BEDT-TTF)$_2$X salts is equal to the number of
dimers. Measurements of the out-of-plane optical conductivity
have also been performed for various $\kappa\text{-(BEDT-TTF)}_2$X salts and the the locations
of vibrational modes were found to be consistent with one electron per dimer \cite{Sedlmeier2012}.
In the limit that $U$ and $V_d$ are much larger than other microscopic energy scales, these
terms will strongly penalize double occupancy of dimers, and we can expect the low energy physics
to be dominated by singly occupied dimers for
\begin{equation}\label{Eq:Low-Energy}
\frac{t_{\alpha}}{U}, \frac{t_{\alpha}}{V_\alpha}, \frac{V_{\alpha}}{U} \ll 1.
\end{equation}
In the low energy limit, with single occupancy of dimers there are four possible states per
dimer, which we may write as
$\ket{\uparrow, 0}, \ket{\downarrow, 0},\ket{0,\uparrow}, \ket{0, \downarrow}$
where we list the occupation of site 1 before that of site 2. We can view the dipole moment
of the dimer as a pseudospin with site 1 corresponding to $P^z = 1/2$ and site 2 corresponding
to $P^z = -1/2$ and switch to the $\ket{P^z, S^z}$ basis to represent the state of each
dimer \cite{Hotta2010}.
We use a strong coupling expansion for the extended Hubbard model, Eq.~(\ref{eq:Ham}), to derive a low-energy effective model.
We project to the one electron per dimer limit and make use of the method set out in Ref.~\cite{MacDonald1988}. The effective model we derive follows the approach of Hotta
\cite{Hotta2010} to write the Hamiltonian in the dipole-spin basis. However, we go beyond
the expansion considered by Hotta by including all nearest neighbour interactions, leading to modified couplings
in the low energy effective theory.
\subsection{Low energy theory}
In the standard strong-coupling expansion of the Hubbard model, e.g. \cite{MacDonald1988},
one projects on to the single electron per site limit. The situation we consider is slightly more
complicated in that we consider the one-electron per dimer limit. An additional complication is the
presence of nearest neighbour interactions, since the occupation of sites on adjacent dimers will
affect the allowed terms in the expansion. To include these terms, we follow closely the approach used
in Refs.~\cite{Farrell2013,Farrell2014} for the extended Hubbard model on a square lattice, modifying
their approach for a triangular lattice of dimers.
We write the Hamiltonian in the form
\begin{equation}
H = H_0 + H_T,
\end{equation}
where $H_0 = H_U + H_V$.
$H_0$ does not change the number of electrons per dimer, whereas $H_T$
includes hops which may change the dimer occupation (the intra-dimer hopping does not change the dimer occupation,
but since it can change the nearest neighbour interaction energy, we do not include it in $H_0$). We introduce
a canonical unitary transformation, $S$, so that $H_0$ remains a constant of motion to a desired order in
$1/U$. Let $H^\prime$ be the transformed Hamiltonian, then
\begin{equation}
H^\prime = H_0 + H_T^\prime,
\end{equation}
where
\begin{equation}
H_T^\prime = e^{iS}H_Te^{-iS},
\label{eq:htprime}
\end{equation}
and to ensure that $H_0$ remains a constant of motion to order $1/U$, we must have
\begin{equation}
\left[H_0,H_T^\prime\right] = 0,
\label{eq:EoMconstraint}
\end{equation}
and expanding $S$ in powers of $1/U$, we get
\begin{equation} \label{eq:s-series}
S = -i\sum_{n=1}^{\infty}\frac{S_n}{U^n}.
\end{equation}
Since we consider the expansion to second order in perturbation theory, we must
determine both $S_1$ and $S_2$. Expanding the hopping part of the Hamiltonian in a power
series in $1/U$ also,
\begin{equation} \label{eq:t-series}
H_T^\prime = \sum_{n=1}^{\infty}\frac{H_{T,n}^\prime}{U^{n-1}},
\end{equation}
and using Eqs.~(\ref{eq:htprime}), (\ref{eq:s-series}) and (\ref{eq:t-series}), we obtain the following
expressions for the first and second order corrections:
\begin{equation}\label{eq:1st-order}
H_{T,1}^\prime = H_T + [S_1,\tilde{H}_0],
\end{equation}
and
\begin{equation}\label{eq:2nd-order}
H_{T,2}^\prime = [S_1,H_T]+\frac{1}{2}\big[S_1,[S_1,\tilde{H}_0]\big]+[S_2,\tilde{H}_0],
\end{equation}
where $\tilde{H}_0 = H_0/U$.
These equations do not give a clear path for obtaining $S_1$ or $S_2$, but applying the
equation of motion requirement Eq.~(\ref{eq:EoMconstraint}), and demanding that it apply
at each order sequentially gives
\begin{equation} \label{eq:s1}
\big[H_0,[S_1,\tilde{H}_0]+H_T\big] = 0,
\end{equation}
and
\begin{equation} \label{eq:s2}
\big[H_0,[S_1,[S_1,H_T]+\frac{1}{2}\big[S_1,[S_1,\tilde{H}_0]\big]+[S_2,\tilde{H}_0]\big] = 0.
\end{equation}
To determine the solutions of Eqs.~(\ref{eq:s1}) and (\ref{eq:s2}) it is helpful to decompose
the hopping term in the Hamiltonian into channels that are differentiated by whether they change the occupation of
a dimer, similar to the procedure used at the site level in Ref.~\cite{MacDonald1988}.
Define the hole occupancy $h_{(x,y),i,\sigma} = 1 - n_{(x,y),i,\sigma}$ which is 0 if site $(x,y),i$ is occupied by an electron with spin $\sigma$ and 1 otherwise. Using the identity $ h_{(x,y),i,\sigma} + n_{(x,y),i,\sigma} = 1$, and acting from both the right and the left, we may write $H_T = \sum_\alpha (T_\alpha^1 + T_\alpha^0 + T_\alpha^{-1})$, where (with $\bar{\sigma}$ indicating the opposite spin to $\sigma$)
\begin{widetext}
\begin{eqnarray}
T_\alpha^1 & = & t_\alpha \sum_{(x,y),i}\hspace{2mm}\sum_{(x',y'),j}\hspace*{-2mm}{\vphantom{\sum}}^\prime \hspace{2mm}\sum_{\sigma}n^{\phantom{\dagger}}_{(x',y'),j,\bar{\sigma}}c^{\dagger}_{(x',y'), j, \sigma} c^{\vphantom{\dagger}}_{(x,y), i, \sigma} h^{\vphantom{\dagger}}_{(x,y),i,\bar{\sigma}},
\\
T_\alpha^0 & = & t_\alpha \sum_{(x,y),i}\hspace{2mm}\sum_{(x',y'),j}\hspace*{-2mm}{\vphantom{\sum}}^\prime \hspace{2mm}\sum_{\sigma}\left\{n^{\phantom{\dagger}}_{(x',y'),j,\bar{\sigma}}c^{\dagger}_{(x',y'), j, \sigma} c^{\vphantom{\dagger}}_{(x,y), i, \sigma} n^{\vphantom{\dagger}}_{(x,y),i,\bar{\sigma}}
+h^{\phantom{\dagger}}_{(x',y'),j,\bar{\sigma}}c^{\dagger}_{(x',y'), j, \sigma} c^{\vphantom{\dagger}}_{(x,y), i, \sigma} h^{\vphantom{\dagger}}_{(x,y),i,\bar{\sigma}}\right\},
\\
T_\alpha^{-1} & = & t_\alpha \sum_{(x,y),i}\hspace{2mm}\sum_{(x',y'),j}\hspace*{-2mm}{\vphantom{\sum}}^\prime \hspace{2mm}\sum_{\sigma}h^{\phantom{\dagger}}_{(x',y'),j,\bar{\sigma}}c^{\dagger}_{(x',y'), j, \sigma} c^{\phantom{\dagger}}_{(x,y), i, \sigma} n^{\phantom{\dagger}}_{(x,y),i,\bar{\sigma}}.
\end{eqnarray}
Writing the summand as $(T_\alpha^m)_{(x,y;\, x',y'),i,j,\sigma}$, with $m \in \{1,0,-1\}$, the hopping term may be written as
\begin{equation} \label{eq:hopping_hub_decomp}
H_T =\sum_{\alpha} \sum_{(x,y),i}\hspace{2mm}\sum_{(x',y'),j}\hspace*{-2mm}{\vphantom{\sum}}^\prime \hspace{2mm}
\sum_{\sigma,m}(T_\alpha^m)_{(x,y;\, x',y'),i,j,\sigma}.
\end{equation}
Note that each $T_\alpha^m$ channel changes the number of doubly occupied sites by $m$ and so
the interaction energy interaction energy in $H_0$ changes by an amount
$mU$. However, nearest neighbour interactions mean that hops that change site occupation also
change nearest neighbour interaction energies. In order to deal with this, we follow Ref.~\cite{Farrell2014}
and introduce a nearest neighbour projection operator, which projects out all states except those which have an electronic configuration $\tilde n^\beta$ neighbouring site $(x,y),i$. Formally it is defined as
\begin{eqnarray}
O^\beta_{(x,y),i}[\tilde n^\beta] & = & \prod_{(\delta_{\beta x}, \delta_{\beta y},\delta_{\beta}), \sigma}
\{\tilde{n}_{(\delta_{\beta x}, \delta_{\beta y}), \delta_{\beta},\sigma}n_{(x+\delta_{\beta x}, y+\delta_{\beta y}), i+\delta_{\beta},\sigma}
+ (1-\tilde{n}_{(\delta_{\beta x}, \delta_{\beta y}), \delta_{\beta},\sigma})h_{(x+\delta_{\beta x}, y+\delta_{\beta y}), i+\delta_{\beta},\sigma}\},
\end{eqnarray}
where the $(\delta_{\beta x}, \delta_{\beta y},\delta_{\beta})$ connect site $(x,y),i$ to the neighbouring sites via $\beta = d, b, p, q$ and
$\tilde{n}_{(\delta_{\beta x}, \delta_{\beta y}), \delta_{\beta},\sigma}$ is 1 when the site at
$(x+\delta_{\beta x}, y+\delta_{\beta y}), i+\delta_{\beta}$ is occupied and zero otherwise.
Inserting forms of the identity $$1 = \prod_{(\delta_{\beta x}, \delta_{\beta y},\delta_{\beta}), \sigma}\left(n_{(x+\delta_{\beta x}, y+\delta_{\beta y}), i+\delta_{\beta},\sigma}+h_{(x+\delta_{\beta x}, y+\delta_{\beta y}), i+\delta_{\beta},\sigma}\right),$$ on either side of Eq.~(\ref{eq:hopping_hub_decomp}) one can find
(similarly to Ref.~\cite{Farrell2014})
\begin{equation}
H_T = \sum_{\alpha, m, \{M_1\},\{M_2\}} T_{\alpha}^{m, \{M_2\}, \{M_1\}},
\label{eq:hop_chan}
\end{equation}
where $\{M_1\} = \{M_1^d, M_1^b, M_1^p, M_1^q\}$ indicates the number of occupied neighbouring $d$, $b$, $p$, and $q$ sites before the hop
and $\{M_2\}$ indicates the neighbours after the hop. Hence
\begin{eqnarray}
T_{\alpha}^{m, \{M_2\}, \{M_1\}} & = & t_\alpha \sum_{(x,y),i}\hspace{2mm}\sum_{(x',y'),j}\hspace*{-2mm}{\vphantom{\sum}}^\prime \hspace{2mm}\sum_{\sigma}\left\{\prod_\beta \sum_{S[n_2^\beta] = M_2^\beta}\sum_{S[n_1^\beta] = M_1^\beta}\right\} \nonumber \\
& &\times \left\{\prod_{\gamma} O_{(x',y'),j}^\gamma[n_2^\gamma]\right\}(T_\alpha^m)_{(x,y;\, x',y'),i,j,\sigma}\left\{\prod_{\eta} O_{(x,y),i}^\eta[n_1^\eta]\right\},
\end{eqnarray}
and $S[n^\beta] = \sum_{(\delta_{\beta x}, \delta_{\beta y},\delta_{\beta}),\sigma} {n}_{(\delta_{\beta x}, \delta_{\beta y}), \delta_{\beta},\sigma}$ is the total number of occupied
neighbouring sites of type $\beta$.
The different hopping channel operators have the commutators (see Appendix C for a derivation)
\begin{equation}\label{eq:decomp_com}
\left[H_U + H_V, T_{\alpha}^{m, \{M_2\}, \{M_1\}}\right] = \left[mU + \sum_{\beta}V_\beta\left(M_2^\beta - M_1^\beta\right)\right] T_{\alpha}^{m, \{M_2\}, \{M_1\}},
\end{equation}
and using the decomposition in Eq.~(\ref{eq:hop_chan}) with the commutator Eq.~(\ref{eq:decomp_com}) allows one to obtain $S_1$ and $S_2$ as outlined
in Appendix \ref{app:strong_coup}.
The first order correction is
\begin{equation}
H_{T,1}^\prime = \sum_{\alpha}\sum_{\{M\}}T_{\alpha}^{0, \{M\}, \{M\}},
\end{equation}
and the second order correction is
\begin{eqnarray}\label{eq.prime.example}
H_{T,2}^\prime & = & U \sum_{\alpha,\nu}\hspace{2mm}\tilde{\sum_{\{M_{2}\}, \{M_{1}\}, \{N\},m}}
\frac{ T_{\alpha}^{m, \{M_{2}\}, \{M_1\}} T_{\nu}^{-m, \{N\}, \{M_{2}\} - \{M_{1}\}
+ \{N\}}}{mU+\sum_{\gamma}V_\gamma(M_2^\gamma - M_1^\gamma)} ,
\end{eqnarray}
where the tilde on the sum over occupation numbers indicates that the sum excludes values of $m$, $\{M_1\}$ and $\{M_2\}$ for which the denominator vanishes.
The first order correction corresponds to a process with a single hop, which can only maintain the one
electron per dimer limit if it is an intra-dimer hop, i.e. $\alpha = d$, no other hops are allowed. It
also requires that the number of nearest neighbour electrons is the same before and after the hop.
In summary, the full second order effective Hamiltonian is
\begin{eqnarray}\label{eq:exp-ham-1}
H^\prime & = & H_U + H_V + \sum_{\{M\}}T_{d}^{0, \{M\}, \{M\}}
+ \sum_{\alpha,\nu}\hspace{2mm}\tilde{\sum_{\{M_{2}\}, \{M_{1}\}, \{N\},m}} \frac{ T_{\alpha}^{m, \{M_{2}\}, \{M_1\}} T_{\nu}^{-m, \{N\}, \{M_{2}\} - \{M_{1}\} + \{N\}}}{mU+\sum_{\gamma}V_\gamma(M_2^\gamma - M_1^\gamma)} .
\end{eqnarray}
We now proceed to write the low energy model in terms of dipole and spin operators, following Hotta \cite{Hotta2010}.
The full form of the effective low energy model in the dipole-spin basis is
\begin{eqnarray}\label{eq:Ham_C_ij}
H^\prime & =& V_q\sum_{i}\sum_{j}{\vphantom{\sum}}' P_i^z P_j^z - V_p\sum_{i}\sum_{j}{\vphantom{\sum}}' P_i^z P_j^z - V_b \sum_{i}\sum_{j}{\vphantom{\sum}}' P_i^z P_j^z + t_d \sum_{i}(P_i^+ + P_i^-) \nonumber \\
& & +\sum_{i}\sum_{j}{\vphantom{\sum}}'\bigg\{ C_{i,j}^0 P_i^z + C_{i,j}^1 P_j^z + C_{i,j}^2 P_i^zP_j^z + C_{i,j}^3 \vec{S_i}\cdot\vec{S_j}
+ C_{i,j}^4 P_i^z\vec{S_i}\cdot\vec{S_j} + C_{i,j}^5 P_j^z\vec{S_i}\cdot\vec{S_j} \nonumber \\
& & \hspace*{1cm} + C_{i,j}^6 P_i^zP_j^z\vec{S_i}\cdot\vec{S_j} + C_{i,j}^7 \left( P_i^+ + P_i^-\right) + C_{i,j}^8 \left( P_i^+P_j^z + P_i^-P_j^z\right) + C_{i,j}^{9} \left(P_i^+\vec{S_i}\cdot\vec{S_j} + P_i^-\vec{S_i}\cdot\vec{S_j}\right) \nonumber \\
& & \hspace*{1cm} + C_{i,j}^{10} \left( P_i^+P_j^z\vec{S_i}\cdot\vec{S_j} + P_i^-P_j^z\vec{S_i}\cdot\vec{S_j} \right) + C_{i,j}^{11} \left( P_j^+ +
P_j^- \right) + C_{i,j}^{12} \left( P_j^+P_i^z + P_j^-P_i^z \right) \nonumber \\
& & \hspace*{1cm} + C_{i,j}^{13} \left( P_j^+\vec{S_i}\cdot\vec{S_j} + P_j^-\vec{S_i}\cdot\vec{S_j}\right) + C_{i,j}^{14} \left( P_j^+P_i^z\vec{S_i}\cdot\vec{S_j}
+ P_j^-P_i^z\vec{S_i}\cdot\vec{S_j}\right)\bigg\},
\end{eqnarray}
where the various $C_{i,j}^n$ are listed in Appendix \ref{App.CoefTable}.
When viewed as a spin model for the dipole pseudospins, the model has ferromagnetic interactions ($V_p$ and $V_b$), antiferromagnetic
interactions ($V_q$) and interactions between dipoles that depend on the states of the physical spins of the dimers (the $C^n_{ij}$). There are also
random field-like terms for the dipoles that depend on the occupations of nearest neighbours. Given that the model
sits on a triangular lattice, this suggests that there may be frustrating interactions between the dipoles that could possibly lead to
slow dynamics and/or glassy behaviour. Performing a dynamical simulation of this model for more than a small number of spins and dipoles
is not practical due to the quantum terms involving dipole raising and lowering operators $P^+$ and $P^-$. Hence, in an effort to learn
whether this low energy model contains glassy physics, we drop the ``quantum'' terms, and focus on a simplified classical spin model
of dipoles coupled to spins on a triangular lattice
\begin{eqnarray}
H_{\rm classical} & =& V_q\sum_{i}\sum_{j}{\vphantom{\sum}}' P_i^z P_j^z - V_p\sum_{i}\sum_{j}{\vphantom{\sum}}' P_i^z P_j^z - V_b \sum_{i}\sum_{j}{\vphantom{\sum}}' P_i^z P_j^z
\label{eq:eff_model}
\\
& & +\sum_{i}\sum_{j}{\vphantom{\sum}}'\bigg( C_{i,j}^0 P_i^z + C_{i,j}^1 P_j^z + C_{i,j}^2 P_i^zP_j^z + C_{i,j}^3 \vec{S_i}\cdot\vec{S_j}
+ C_{i,j}^4 P_i^z\vec{S_i}\cdot\vec{S_j} + C_{i,j}^5 P_j^z\vec{S_i}\cdot\vec{S_j} + C_{i,j}^6 P_i^zP_j^z\vec{S_i}\cdot\vec{S_j}\bigg) . \nonumber
\end{eqnarray}
We treat the dipoles $P_i^z$ as Ising variables and the spins $\bvec{S}_i$ as classical vectors. This allows the model to be
simulated using classical Monte Carlo techniques.
\end{widetext}
\section{Monte Carlo simulations}
\label{sec:MC}
In this section we present and discuss results from equilibrium Monte Carlo simulations of the model defined in Eq.~(\ref{eq:eff_model}).
We calculate the polarization, magnetization, electric and magnetic susceptibility and Edwards-Anderson order parameters
for both spin and charge degrees of freedom. The couplings in $H_{\rm classical}$ depend not only on the parameters of the extended Hubbard
model but also on the occupation of the neighbouring sites in the lattice. This means that the couplings in a Monte Carlo simulation depend
on the state of the system and will evolve as dipoles are flipped. In order to obtain a more computationally tractable problem, we instead
use the following procedure to specify the model: i) calculate the distribution of all possible values of couplings for each term in the
Hamiltonian, ii) fit the distribution to a simplified form, iii) draw couplings between dipoles and spins randomly from the distributions
calculated in ii), iv) average over multiple sets of couplings.
\begin{figure}[h]
\includegraphics[width=8cm]{Figure2.pdf}
\caption{Dimers that must be included to calculate couplings for the dimer $i$. Each black dot represents a lattice site for which the nearest neighbour occupancy
needs to be known to properly calculate each $C_n$, meaning that a minimum of 15 dimers are needed to calculate the full set of couplings.}
\label{fig:15_con}
\end{figure}
Each of the coefficients, $C^n_{ij}$, in Eq.~(\ref{eq:eff_model}) is determined by the electronic
configurations of the neighbouring dimers (see Tables I-\ref{tab:four} in Appendix \ref{app:couplings}). Therefore for each distinct electronic
configuration of nearest neighbours about some central dimer $i$ there will be a different set of $C_{i,j}^n$ for that dimer.
For large lattices,
calculating the couplings for each dimer is a very computationally expensive task. To run Monte Carlo simulations one must first calculate
the set of couplings $C_{i,j}^n$ from the set of $C_n$ as set out in Appendix \ref{app:couplings}. In order to simplify the calculations we find the distributions of the various
$C_{i,j}^n$ and then sample from these distributions to set the couplings for each simulation.
In this way we aim to capture the effect of the distribution of couplings that arise from the different occupations without recalculating all of
the couplings at each step of the calculation as dipoles flip and the charge distribution evolves with time.
We now discuss the calculation of the couplings. The denominator of each $C_n$ relating to the hopping between dimers $i$ and $j$ depends on the occupancy
of the nearest neighbours of a lattice site on dimer $j$ as well as the occupancy of the nearest neighbours of a lattice site on the dimer $i$.
As can be seen in Fig.~\ref{fig:15_con} this implies that to calculate a set of $C_n$ the electronic configuration of a dimer $i$ as well as its
14 nearest and next-nearest neighbors must be known.
We calculated the distributions of the $C_{i,j}^n$ as follows: first, for a single 15 dimer electronic configuration we calculated the set of $C_n$.
Second, we used the set of $C_n$ to calculate the corresponding set of $C_{i,j}^n$ (Appendix \ref{App.CoefTable}). We repeated this process until the electronic
configuration space was properly sampled and binned the various $C_{i,j}^n$ to create histograms.
There are $2^{15} = 32768$ possible electronic configurations.
\begin{widetext}
\begin{figure}[ht]
\centering
\includegraphics[scale=0.31]{Figure3a.pdf}
\includegraphics[scale=0.31]{Figure3b.pdf}
\caption{Normalized probability distributions for $C_{i,j}^2$ couplings to the two types of dimers neighbouring a central dimer located at $i=(x,y)$. a)
($x \pm \frac{1}{2}, y \pm \frac{1}{2}$) neighbours; b) ($x\pm 1$, $y$) neighbours. Each bar has a width of $1.4 \times 10^{-5}$. Parameters used are $U/t = 15.000113$,
$V_d/t = 10.000219$, $V_b/t = 1.000547$, $V_p/t = 1.000623$, and $V_q/t = 1.000412$.}
\label{fig:non_disorder_individual}
\end{figure}
\begin{figure}[ht]
\includegraphics[width=8cm]{Figure4a.pdf}
\includegraphics[width=8cm]{Figure4b.pdf}
\includegraphics[width=8cm]{Figure4c.pdf}
\includegraphics[width=8cm]{Figure4d.pdf}
\caption{Histograms combining couplings for all neighbouring dimers for various $C_{i,j}^n$ couplings: a) $C_{i,j}^0$, b) $C_{i,j}^2$, c) $C_{i,j}^3$, d) $C_{i,j}^4$.
Parameters used are the same as Fig.~\ref{fig:non_disorder_individual}.}
\label{fig:non_disorder_mult_com}
\end{figure}
\end{widetext}
While motivated by the $\kappa$-(BEDT-TTF)$_2$X family of salts, we do not choose parameter values for any specific material, but are somewhat
guided by the parameter choices in Ref.~\cite{Hotta2010}. In all calculations, we take all of the hopping parameters, $t_d = t_b = t_p = t_q = t$,
and measuring other parameters in terms of $t$, we take $U = 15 t$, with $V_d$ taking values between $4t$ and $12 t$ and $V_b, V_p$ and $V_q$
taking values between $t$ and $3t$. For reference, Pinteri\'{c} {\it et al.} estimate $U/t = 7.3$ in $\kappa$-(BEDT-TTF)$_2$Cu$_2$(CN)$_3$ \cite{Pinteric2014}.
The choice of all hopping parameters being equal leads to some coupling coefficients being equal that would otherwise take on different values.
There is a subtlety involved when calculating the denominators of the set of $C_n$. As discussed in Appendix \ref{app:couplings}, the strong coupling
expansion excludes all terms which have a zero denominator. The denominator takes the form $mU + \sum_\gamma
V_\gamma (M_2^\gamma - M_1^\gamma)$ where $m$ and $(M_2^\gamma - M_1^\gamma)$ are integers, implying that if the $U$ and
the $V_\gamma$ are factors of each other it is possible to have a vanishing denominator.
According to the strong coupling expansion if a process leads to a vanishing denominator it does not contribute to the second order Hamiltonian.
To avoid such terms we selected parameters $U$ and $V_\gamma$ such that the denominator does not vanish.
\begin{figure}[h]
\centering
\includegraphics[width=8cm]{Figure5.pdf}
\caption{An example of the approximate probability density function shown in Eq. (\ref{eq:prob}) for $C^0_{ij}$:
here $w_1 = -0.0068$, $w_2 = -0.0023$, $w_3 = 0.0023$, $w_4 = 0.0068$, and $b_1 = b_2 = 0.2$.}
\label{fig:hist_fit}
\end{figure}
We determined histograms for all of the couplings $C^n_{ij}$ that enter into the effective model. Examples of these histograms for
$C_{i,j}^1$ (which give the strengths of the $P_i^z P_j^z$ interaction) are shown in Fig.~\ref{fig:non_disorder_individual}.
The couplings to dimers $\left(x \pm \frac{1}{2}, y \pm \frac{1}{2}\right)$ are identical due to symmetry for the chosen hopping parameters as are those
to dimers located at $\left(x \pm 1, y\right)$, so there are only two types of $C^1_{ij}$ couplings for our choice of parameters. The combined
distributions of all $C^0_{ij}$, $C^2_{ij}$, $C^3_{ij}$, $C^4_{ij}$ couplings are shown in Fig.~\ref{fig:non_disorder_mult_com}.
As a further simplification we find approximate analytical forms for the distributions and draw the couplings from these approximate distributions.
Many parameter choices lead to histograms which can be modeled adequately by a probability density function of the form
\begin{widetext}
\begin{eqnarray} \label{eq:prob}
P(x) =
\begin{cases}
0, & x<w_1, \\
b_1, & w_1 \leq x < w_2, \\
0, & w_2 \leq x < 0, \\
\big(1-b_2(w_4 - w_3) - b_1(w_2-w_1)\big)\delta(x), & x=0, \\
0, & 0 < x \leq w_3, \\
b_2, & w_3 < x \leq w_4, \\
0, &w_4 < x,
\end{cases}
\end{eqnarray}
with $w_4 > w_3 > w_2 > w_1$ and $b_1 \geq 0$, $b_2 \geq 0$. A general graphical representation of this distribution is shown in Fig. \ref{fig:hist_fit} using $C^0_{ij}$ as an example.
\end{widetext}
\begin{figure}[hbt]
\includegraphics[width=8cm]{Figure6a.pdf}
\includegraphics[width=8cm]{Figure6b.pdf}
\caption{Cumulative probability functions for the combined histograms of a) $C^{0}_{i,j}$ and b) $C^{4}_{i,j}$.
The parameters used are $U/t=15.0113$, $V_{d/t}=8.0219$, $V_{b}/t=1.0547$, $V_{p}/t=1.0623$, and $V_{q}/t=1.0412$.}
\label{fig:coef_cum}
\end{figure}
We select the $w_n$ and $b_n$ parameters so that the
corresponding cumulative probability function obtained from Eq. (\ref{eq:prob}) matches the cumulative probability function of the various histograms.
This ensures that the behaviour of the distribution is properly captured.
We first choose the parameters $w_n$ based on the cumulative probability function of the various histograms and then perform a least squares analysis to calculate the $b_n$ fitting parameters. This process ensures that all the desired peaks are captured and is performed for every fit. Some results of this fitting procedure are shown in Fig.~\ref{fig:coef_cum} for $C^0_{ij}$ and $C^4_{ij}$.
In the Monte Carlo simulations we create a triangular lattice of $N$ dimers and then assign couplings randomly from the calculated distributions.
As there are many more coupling configurations than $N$, we average over couplings to obtain a representative sample of the configuration space.
In one Monte Carlo step, we attempt to flip $N$ spins and $N$ dipoles, choosing spins and dipoles at random. A spin flip corresponds to assigning
a new direction in space for the vector spin and flipping a dipole corresponds to changing the sign of the Ising degree of freedom.
We use the Metropolis algorithm to determine the probability of whether a move that leads to an energy change $\Delta E$ is accepted:
\begin{equation}
P(\Delta E) =
\begin{cases}
1, & \Delta E \leq 0,\\
e^{-\beta \Delta E}, & \Delta E > 0,\\
\end{cases}
\end{equation}
\noindent where $\beta = 1/(k_{B}T)$ with
$k_{B}=1.38 \times 10^{-23}$m$^{2}$ JK$^{-1}$.
We consider lattices of up to $N = 14^2 = 196$ dimers and establish equilibration by considering two replicas of the system, prepared with the
same set of couplings but different thermal histories (i.e. different random numbers in the Metropolis algorithm). The system is taken to be
equilibrated when chosen thermodynamic variables (the electric and magnetic susceptibilities) calculated in the two replicas agree to within a
specific tolerance. After the system has equilibrated we sample to obtain averages of the quantities that we discuss below.
First, we calculate the average polarization per dipole at time step $t_{j}$ for a lattice size of $N$ dimers, given by
\begin{equation}
P(t_{j}) = \dfrac{1}{N}\left| \sum_{i=1}^{N} P_{i}(t_{j}) \right| ,
\end{equation}
with $P_{i}(t_{j})$ the polarization of dimer $i$ at time step $t_{j}$. Second, we calculate the average magnetization per spin at time step $t_{j}$
\begin{equation}
M(t_{j}) = \dfrac{1}{N}\left| \sum_{i=1}^{N} \vec{S}_{i}(t_{j}) \right|,
\end{equation}
where $\vec{S}_{i}(t_{j})$ is the spin of dimer $i$ at time step $t_{j}$.
We then calculate the average polarization $P = \left<P(t_j)\right>$ and average magnetization $M = \left<M(t_j)\right>$ where
the notation $<\ldots>$ indicates a time average over $N_{t}$ Monte Carlo time steps. We
also calculate the electric susceptibility and magnetic susceptibility, defined respectively as
\begin{equation}
\chi _{P} = \beta \lim_{t_j \to \infty} \left( \big\langle P(t_{j})^{2} \big\rangle - \big\langle P(t_{j}) \big\rangle ^{2} \right),
\end{equation}
and
\begin{equation}
\chi _{M} = \beta \lim_{t_j \to \infty} \left( \big\langle M(t_{j})^{2} \big\rangle - \big\langle M(t_{j}) \big\rangle ^{2} \right) .
\end{equation}
Our equilibration criterion is that
the electric and magnetic susceptibilities of the two replicas agree to $2\%$ tolerance or completion of
$2 \times 10^{7}$ Monte Carlo time steps, whichever comes first. We were unable to reach equilibrium at temperatures lower than $0.0065\, U/$k$_{B}$.
\begin{widetext}
\begin{figure}[hbt]
\includegraphics[width=8cm]{Figure7a.pdf}
\includegraphics[width=8cm]{Figure7b.pdf}
\caption{Polarization averaged over 18 bond configurations as a function of temperature for a lattice size of $10^2 = 100$.
a) Fixed inter-dimer interactions: $U/t=15.0113$, $V_{b}/t=1.0547$, $V_{p}/t=1.0623$, and $V_{q}/t=1.0412$ with varying $V_d/t$; b)
Fixed intra-dimer coupling: $U/t=15.0113$ and $V_{d}/t=10.0219$ with varying $V_\alpha/t$ with $\alpha = b$, $p$ and $q$.}
\label{fig:Polarization}
\end{figure}
\begin{figure}
\includegraphics[width=8cm]{Figure8a.pdf}
\includegraphics[width=8cm]{Figure8b.pdf}
\caption{Magnetization averaged over 18 bond configurations as a function of temperature for a lattice size of $10^2 = 100$.
a) Fixed inter-dimer interactions: $U/t=15.0113$, $V_{b}/t=1.0547$, $V_{p}/t=1.0623$, and $V_{q}/t=1.0412$ with varying $V_d/t$; b)
Fixed intra-dimer coupling: $U/t=15.0113$ and $V_{d}/t=10.0219$ with varying $V_\alpha/t$ with $\alpha = b$, $p$ and $q$.}
\label{fig:Magnetization}
\end{figure}
\begin{figure}
\includegraphics[width=8cm]{Figure9a.pdf}
\includegraphics[width=8cm]{Figure9b.pdf}
\caption{Electric susceptibility averaged over 18 bond configurations as a function of temperature for a lattice size of $10^2 = 100$.
a) Fixed inter-dimer interactions: $U/t=15.0113$, $V_{b}/t=1.0547$, $V_{p}/t=1.0623$, and $V_{q}/t=1.0412$ with varying $V_d/t$; b)
Fixed intra-dimer coupling: $U/t=15.0113$ and $V_{d}/t=10.0219$ with varying $V_\alpha/t$ with $\alpha = b$, $p$ and $q$.}
\label{fig:El_Sus}
\end{figure}
\begin{figure}
\includegraphics[width=8cm]{Figure10a.pdf}
\includegraphics[width=8cm]{Figure10b.pdf}
\caption{Magnetic susceptibility averaged over 18 bond configurations as a function of temperature for a lattice size of $10^2 = 100$.
a) Fixed inter-dimer interactions: $U/t=15.0113$, $V_{b}/t=1.0547$, $V_{p}/t=1.0623$, and $V_{q}/t=1.0412$ with varying $V_d/t$; b)
Fixed intra-dimer coupling: $U/t=15.0113$ and $V_{d}/t=10.0219$ with varying $V_\alpha/t$ with $\alpha = b$, $p$ and $q$.}
\label{fig:Mag_Sus}
\end{figure}
To check for glassy behaviour we calculate the Edwards-Anderson order parameters for both charge and spin degrees of freedom.
For the Ising like dipole degrees of freedom, the Edwards-Anderson order parameter is
\begin{equation}
Q_{\rm EA}^{\rm Pol} = \lim_{t_{j}\to\infty} \left| \overline{\left(\left\langle\dfrac{1}{N}\sum_{i=1}^{N}P^{z,A}_{i}(t_{j})P^{z,B}_{i}(t_{j})\right\rangle - \left\langle \sum_{i=1}^{N}P^{z,A}_{i}(t_{j})\right\rangle \left\langle \sum_{i=1}^{N}P^{z,B}_{i}(t_{j})\right\rangle \right)}\right|.
\label{Pol_EA}
\end{equation}
Here $P^{z,A}_{i}(t_{j})$ and $P^{z,B}_{i}(t_{j})$ are the polarization on dimer $i$ at time step $t_{j}$ for replicas $A$ and $B$ respectively. The over-bar indicates an average over the bond configurations. For a lattice of $N$ dimers, the Edwards-Anderson order parameter for the spin degrees of freedom takes the form
\begin{equation}
Q_{\rm EA}^{\rm Mag} = \lim_{t_{j}\to\infty} \left|\overline{\left(\left\langle\dfrac{1}{N}\sum_{i=1}^{N}\vec{S}^{A}_{i}(t_{j})\cdot \vec{S}^{B}_{i}(t_{j})\right\rangle - \left\langle \sum_{i=1}^{N}\vec{S}^{A}_{i}(t_{j})\right\rangle \cdot \left\langle \sum_{i=1}^{N}\vec{S}^{B}_{i}(t_{j})\right\rangle \right)} \right| ,
\label{Mag_EA}
\end{equation}
where $\vec{S}^{A}_{i}(t_{j})$ and $\vec{S}^{B}_{i}(t_{j})$ are the spins at dimer $i$ at time step $t_{j}$ for replicas $A$ and $B$ respectively.
\end{widetext}
In order to establish whether there is a phase transition, and if so, to
estimate of the critical temperature, $T_{c}$, we use the Binder cumulant statistic. For the Edwards-Anderson order parameter, $Q$ the statistic is given by
\begin{equation}
U_{L}(Q) = \dfrac{3}{2}\left[1 - \left(\dfrac{\langle Q^{4} \big\rangle _{L}}{3\langle Q^{2} \big\rangle ^{2} _{L}} \right) \right],
\end{equation}
where $L$ is the size of the system. We plot $U_{L}$ as a function of temperature for different system sizes. If there is a transition to an ordered
phase at a temperature $T_c$ then in the large $L$ limit, $U_{L}$ tends to a finite value for $T<T_{c}$ and 0 for $T> T_{c} $.
Up to small finite size corrections, the Binder cumulant curves should intersect at the critical temperature, where in the $L \to \infty$ limit, the
Binder cumulant is independent of system size.
\begin{figure}[thb]
\includegraphics[width=1\linewidth]{Figure11a.pdf}
\includegraphics[width=1\linewidth]{Figure11b.pdf}
\caption{a) $Q^{\rm Pol}_{\rm EA}$; b) $Q^{\rm Mag}_{\rm EA}$ for fixed inter-dimer interaction strengths on a lattice of 10$^2$ = 100 dimers,
averaged over 18 bond configurations. The parameters used were $U/t=15.0113$, $V_{b}/t=1.0547$, $V_{p}/t =1.0623$, $V_{q}/t=1.0412$ for varying $V_{d}/t$.}
\label{fig:EA_Vd}
\end{figure}
\begin{figure}[thb]
\includegraphics[width=1\linewidth]{Figure12a.pdf}
\includegraphics[width=1\linewidth]{Figure12b.pdf}
\caption{a) $Q^{\rm Pol}_{\rm EA}$; b) $Q^{\rm Mag}_{\rm EA}$ for fixed intra-dimer interaction strengths on a lattice of 10$^2$ = 100 dimers,
averaged over 18 bond configurations. The parameters used were $U/t=15.0113$, $V_{d}/t=10.0219$ for varying $V_{\alpha }$ with $\alpha = b$, $p$ and $q$.}
\label{fig:EA_Va}
\end{figure}
\begin{figure}[hbt]
\includegraphics[width=8cm]{Figure13.pdf}
\caption{Binder cumulant plot of $V_{d}/t = 10.0219$ for lattice sizes $6\times 6$, $8\times 8$ and $12\times 12$. Other parameters used are $U/t=15.0113$, $V_{b}/t=1.0547$,
$V_{p}/t=1.0623$, $V_{q}/t=1.0412$.}
\label{fig:Binder_10}
\end{figure}
Figure \ref{fig:Polarization} shows the polarization as a function of temperature for various intra- and inter-dimer interaction energies respectively.
There is no evidence for ordering in the charge degrees of freedom for the parameters considered. We also see no ordering in the spin degrees of freedom shown in
Fig.~\ref{fig:Magnetization}.
Figures \ref{fig:El_Sus} and \ref{fig:Mag_Sus} show the electric and magnetic susceptibilities as a function of temperature
respectively. We see similar behaviour for a wide range of $V_{d}$ and $V_{\alpha }$ for different lattice sizes of $N$ dimers.
In Fig.~\ref{fig:EA_Vd} we show $Q^{\rm Pol}_{\rm EA}$ and $Q^{\rm Mag}_{\rm EA}$
as a function of temperature. In the low temperature regime ($\lesssim 0.02\, U/k_{B}$) $Q^{\rm Pol}_{\rm EA}$
grows with decreasing temperature for $V_{d}/t \approx $ 10, 11, and 12. For intra-dimer interaction energies; $V_{d}/t \lesssim 10$, $Q^{\rm Pol}_{\rm EA}$
becomes appreciable for temperatures $\lesssim 0.03\, U/k_{B}$.
Smaller intra-dimer interactions ($V_{d}/t \lesssim 10$) show stronger evidence
of a non-zero Edwards-Anderson order parameter in the charge degrees of freedom at the chosen temperatures.
Even though we see evidence of glassy ordering in charge degrees of freedom at low temperatures in Fig.~\ref{fig:EA_Vd}, we do not see any
evidence of glassy ordering in the spin degrees of freedom. $Q^{\rm Mag}_{\rm EA}$ takes on values indistinguishable
from zero for all temperatures and $V_{d}/t$ values considered.
Similar results to those seen in Fig.~\ref{fig:EA_Vd} are shown in Fig.~\ref{fig:EA_Va} where we plot $Q^{\rm Pol}_{\rm EA}$ and $Q^{\rm Mag}_{\rm EA}$
as a function of temperature for different inter-dimer interaction energies. Here, we observe that small
$V_{\alpha}/t$ values lead to a larger $Q^{\rm Pol}_{\rm EA}$ at low temperatures than large $V_{\alpha}/t$.
A simple estimate of the phase transition is provided by the temperature at which the Edwards-Anderson order parameter becomes clearly distinct
from zero, which can be determined from Fig.~\ref{fig:EA_Vd} and \ref{fig:EA_Va} to be $k_{B}T \approx 0.025\, U/k_{B})$.
In the following discussion we present a quantitative estimate of the phase transition in our model obtained from the Binder cumulant statistic.
Taking $V_d/t \simeq 10$, the intersection of the curves in Fig.~\ref{fig:Binder_10} determines an estimate for the critical temperature for the phase
transition of $\approx 0.022\, U/k_{B}$, as compared to the naive estimate of $\approx 0.025\, U/k_{B}$ obtained from the $Q^{\rm Pol}_{EA}$ curve in Fig.~\ref{fig:EA_Vd}.
We observed similar behaviour for other parameter values $(e.g. V_{d}/t = 8$ and 9), but this is the clearest example of a transition.
\section{Discussion and Conclusions}
\label{sec:disc}
In this paper we have considered an extended Hubbard model of dimers on a triangular lattice, and obtained the low energy effective theory
in terms of separate spin and charge degrees of freedom in the one electron per dimer limit, taking into account the occupation of
nearest neighbour sites, which extends a previous perturbative expansion of the model \cite{Hotta2010}. In order to study the tendency
towards glassiness of charge degrees of freedom in the model, motivated by evidence of relaxor ferroelectric behaviour in
the $\kappa -$(BEDT-TTF$)_{2}X$ family of salts, we made a classical approximation to the effective model, so that we could study
it with Monte Carlo simulations. The couplings in the effective model depend on the occupations of the neighbouring sites,
so in order to avoid recalculating the couplings at each step of the Monte Carlo simulation we calculated the full distribution of
couplings and then drew couplings randomly from this distribution. Under these conditions we have shown that a non-zero Edwards-Anderson order parameter
for charge degrees of freedom develops below a critical temperature $T_c$.
In order to make some connection to experiment, we take $U \sim 0.7$ eV \cite{Nakamura2012}, which would place $T_c = 0.022\, U/k_B \sim 180$ K, which is much
higher than the temperatures at which glassy dynamics is seen in $\kappa -$(BEDT-TTF$)_{2}X$ salts, which is usually on the order of tens of kelvins.
However, several of the approximations we have made are likely to enhance ordering, so it is not surprising that our temperature estimate is considerably
higher than experimental observations. There are two main approximations we have made that likely enhance glassiness in the model.
First, we drop ``quantum'' terms in Eq.~(\ref{eq:Ham_C_ij}) to obtain our effective model, which corresponds to ignoring quantum fluctuations, that can be expected to depress $T_c$. Second, we draw the couplings we use in the Monte Carlo simulation from a fixed distribution that is independent of time, rather than
working with couplings that fluctuate with time. This has the effect of generating quenched rather than time-dependent disorder, which will also enhance
glassy tendencies in the model. However, despite these approximations, glassiness appears quite readily in the charge degrees of freedom, so even if
the two approximations we have made are relaxed, it seems likely that there will remain strong tendencies towards glassiness, that may well manifest
themselves at lower temperature scales in the true system. Previous work on models of classical models of spins coupled to charge degrees of freedom has
also demonstrated glassy dynamics, even in the absence of disorder \cite{Kennett2005}.
Another observation of note is that we find no ordering in the spin degrees of freedom and no development of charge polarization
at the temperatures we can access, consistent with the experimental observations in $\kappa -$(BEDT-TTF$)_{2}X$ salts. In addition, the observation of a
non-zero Edwards-Anderson transition temperature in a two dimensional model might be surprising until one considers that it is on a triangular lattice, for
which each site has six neighbours, unlike the two dimensional Edwards-Anderson model on a square lattice with $\pm J$ couplings, for which there are
only four neighbours per site, that is a spin glass only at zero temperature \cite{Singh1986,Bhatt1988}.
We regard this work as a proof of principle that glassiness can arise in models that are relevant for $\kappa -$(BEDT-TTF$)_{2}X$ salts and see future
avenues for exploration through i) studying a wider range of extended Hubbard model parameters, ii) allowing for time dependent couplings, and iii) investigating the effects of quantum terms in the effective model. In addition to the equilibrium calculations considered here, the study of out of equilibrium dynamics in
larger systems may help to determine the functional time dependence of aging dynamics that might also be accessible in experiments.
\section{Acknowledgements}
The authors acknowledge Compute Canada resources that were used to obtain the numerical
results in this work. M. B. D. and M. P. K. were supported by NSERC.
\begin{appendix}
\begin{widetext}
\section{Commutators Used in the Strong Coupling Expansion} \label{App.Commutator}
In this Appendix we provide derivations of a number of commutators that prove useful in constructing the strong coupling expansion of the
Hamiltonian. Many of the commutators derived here have corresponding simpler versions in Ref.~\cite{Farrell2013}
in which the strong coupling expansion was performed on a square lattice of sites rather than a triangular lattice of dimers.
\subsection{Useful Commutators}
Before deriving the commutator $\left[H_0, T_{\alpha}^{m, \{M_2\}, \{M_1\}}\right]$ it is helpful to first derive a few basic commutators that are used in the derivation of the commutator of $H_0$ with the hopping operator. We begin with
\begin{align}\label{eq:numbercom}
[n^{\vphantom{\dagger}}_{(x,y),i,\sigma},c^{\dagger}_{(x',y'), j, \sigma'} c^{\vphantom{\dagger}}_{(x'',y''), k, \sigma'}]
=\, &\delta_{\sigma, \sigma'}\left[\delta_{(x,y),(x',y')} \delta_{i,j} - \delta_{(x,y),(x'',y'')} \delta_{i,k}\right]c^{\dagger}_{(x',y'), j, \sigma'}
c^{\vphantom{\dagger}}_{(x'',y''), k, \sigma'}.
\end{align}
Setting $(x'',y'') = (x',y'), k=j$ in this result it follows directly that
\begin{align}
[n_{(x,y),i,\sigma},n_{(x',y'),j,\sigma'}] &= 0,
\end{align}
which implies
\begin{align}
[n_{(x,y),i,\sigma},h_{(x',y'),j,\sigma'}] = \left[n_{(x,y),i,\sigma}, O_{(x',y'),j}^\beta[\tilde{n}^\beta]\right] = \left[h_{(x,y),i,\sigma},
O_{(x',y'),j}^\beta[\tilde{n}^\beta]\right]= 0.
\end{align}
The next set of commutators needed are those of the form $\big[n_{(x,y),i,\sigma'},(T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma} \big]$ which must be calculated independently for each $m = -1,0,1$.
In general they can be written as
\begin{align} \label{eq:n-hop-com}
\Big[n_{(x,y),i,\sigma'},&(T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma} \Big]
= (T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma}\delta_{\sigma,\sigma'}\left[\delta_{(x,y),(x'',y'')}\delta_{i,k} - \delta_{(x,y),(x',y')}\delta_{i,j}\right] .
\end{align}
\subsection{On-site Interaction Commutator}
Using Eq.~(\ref{eq:n-hop-com}) one can show straightforwardly that
in general
\begin{align}\label{eq:U_comm}
\left[H_U, T_{\alpha}^{m, \{M_2\}, \{M_1\}}\right] = m U T_{\alpha}^{m, \{M_2\}, \{M_1\}}.
\end{align}
\subsection{Nearest Neighbour Interaction Commutator}
Before deriving the next commutator it is useful to first rewrite the nearest neighbour interaction term as
\begin{align}
H_{V} = \frac{1}{2}\sum_{\gamma}V_\gamma\sum_{\mathclap{\substack{(x,y) \\ i}}}\hspace{4mm} \sum_{\mathclap{\substack{(\delta_{x_\gamma},\delta_{y_\gamma}) \\ \delta_\gamma}}}\hspace{3.5mm}\sum_{\sigma, \sigma'}n_{(x,y),i,\sigma}n_{(x+\delta_{x_\gamma},y+\delta_{y_\gamma}),i+\delta_\gamma,\sigma'},
\end{align}
where $\gamma = d, q, b, p$ and the factor of $\frac{1}{2}$ is to prevent double counting. It is possible to further rewrite this as
\begin{align}
H_{V} = \frac{1}{2}\sum_{\gamma}V_\gamma \sum_{\mathclap{\substack{(x,y) \\ i}}}\sum_{\sigma}n_{(x,y),i,\sigma}\tilde{n}^\gamma_{(x,y),i},
\end{align}
where
\begin{align}
\tilde{n}^\gamma_{(x,y),i}=\hspace{3.5mm}\sum_{\mathclap{\substack{(\delta_{x_\gamma},\delta_{y_\gamma}) \\ \delta_\gamma}}}\hspace{3.5mm}\sum_{\tilde{\sigma}}n_{(x+\delta_{x_\gamma},y+\delta_{y_\gamma}),i+\delta_\gamma,\tilde{\sigma}}.
\end{align}
The desired commutator now takes the form
\begin{align} \label{eq:Hv-com-step1}
\left[H_{V}, T_{\alpha}^{m, \{M_2\}, \{M_1\}}\right] =& \frac{1}{2}\sum_{\gamma}V_\gamma\sum_{\mathclap{\substack{(x,y) \\\ i}}}\hspace{2mm}\sum_{\mathclap{\substack{(x',y') \\ j}}}\hspace{3.2mm}\sum_{\mathclap{\substack{(x'',y'') \\ k}}}{\vphantom{\sum}}'\left\{\prod_{\beta}\sum_{S[n_2^\beta] = M_2^\beta}\sum_{S[n_1^\beta] = M_1^\beta}\right\}
\left\{\prod_{\eta} O_{(x'',y''),k}^\eta[n_2^\eta]\right\} \nonumber \\
&\times \sum_{\sigma,\sigma'} \left[n_{(x,y),i,\sigma'}\tilde{n}^\gamma_{(x,y),i},(T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma}\right]
\left\{\prod_{\xi} O_{(x',y'),j}^\xi[n_1^\xi]\right\} \nonumber \\
=& \frac{1}{2}\sum_{\gamma}V_\gamma\sum_{\mathclap{\substack{(x,y) \\\ i}}}\hspace{2mm}\sum_{\mathclap{\substack{(x',y') \\ j}}}\hspace{3.2mm}\sum_{\mathclap{\substack{(x'',y'') \\ k}}}{\vphantom{\sum}}'\left\{\prod_{\beta}\sum_{S[n_2^\beta] = M_2^\beta}\sum_{S[n_1^\beta] = M_1^\beta}\right\}
\left\{\prod_{\eta} O_{(x'',y''),k}^\eta[n_2^\eta]\right\} \nonumber \\
&\times \sum_{\sigma,\sigma'} \Bigg\{ n_{(x,y),i,\sigma'}\left[\tilde{n}^\gamma_{(x,y),i},(T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma}\right] \nonumber \\
&+ \left[n_{(x,y),i,\sigma'},(T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma}\right]\tilde{n}^\gamma_{(x,y),i}\Bigg\}
\left\{\prod_{\xi} O_{(x',y'),j}^\xi[n_1^\xi]\right\}.
\end{align}
Equation (\ref{eq:n-hop-com}) can then be used to obtain
\begin{align}\label{eq:nt-hop-com}
&\left[\tilde{n}^\gamma_{(x,y),i},(T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma}\right] \nonumber \\ &=\hspace{3.5mm} \sum_{\mathclap{\substack{(\delta_{x_\gamma},\delta_{y_\gamma}) \\\delta_\gamma}}}\hspace{3.5mm}\sum_{\tilde{\sigma}}(T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma}\delta_{\sigma,\tilde{\sigma}}\left[\delta_{(x+\delta_{x_\gamma},y+\delta_{y_\gamma}),(x'',y'')}\delta_{i+\delta_\gamma,k}-\delta_{(x+\delta_{x_\gamma},y+\delta_{y_\gamma}),(x',y')}\delta_{i+\delta_\gamma,j}\right].
\end{align}
Applying Eq.~(\ref{eq:n-hop-com}) and relabeling some indicies leads to
\begin{eqnarray} \label{eq:Hv-com-step2}
\left[H_{V}, T_{\alpha}^{m, \{M_2\}, \{M_1\}}\right] &=& \frac{1}{2}\sum_{\gamma}V_\gamma \sum_{\mathclap{\substack{(x',y') \\ j}}}
\hspace{3.2mm}\sum_{\mathclap{\substack{(x'',y'') \\ k}}}
{\vphantom{\sum}}'\sum_{\sigma}\left\{\prod_{\beta}\sum_{S[n_2^\beta] = M_2^\beta}\sum_{S[n_1^\beta] = M_1^\beta}\right\}
\left\{\prod_{\eta} O_{(x'',y''),k}^\eta[n_2^\eta]\right\} \nonumber \\
& & \times \Bigg\{ \left[\tilde{n}^\gamma_{(x'',y''),k}-\tilde{n}^\gamma_{(x',y'),j}\right] (T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma} \nonumber \\
& & \hspace*{0.5cm} +(T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma}\left[\tilde{n}^\gamma_{(x'',y''),k}-\tilde{n}^\gamma_{(x',y'),j}\right]\Bigg\}
\left\{\prod_{\xi} O_{(x',y'),j}^\xi[n_1^\xi]\right\}.
\end{eqnarray}
The central part of this equation can be rewritten as
\begin{align} \label{eq:Hv-com-step2.5}
\tilde{n}^\gamma_{(x'',y''),k}&(T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma}-\tilde{n}^\gamma_{(x',y'),j}(T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma} \nonumber \\
&+(T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma}\tilde{n}^\gamma_{(x'',y''),k}-(T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma}\tilde{n}^\gamma_{(x',y'),j} \nonumber \\
=& 2\left[\tilde{n}^\gamma_{(x'',y''),k}(T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma}-(T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma}\tilde{n}^\gamma_{(x',y'),j}\right] \nonumber \\
& + \left[(T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma},\tilde{n}^\gamma_{(x'',y''),k}\right] - \left[\tilde{n}^\gamma_{(x',y'),j},(T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma}\right] .
\end{align}
When Eq.~(\ref{eq:Hv-com-step2.5}) is inserted into Eq.~(\ref{eq:Hv-com-step2}) and Eq.~(\ref{eq:nt-hop-com}) is also applied this leads to the result
\begin{align} \label{eq:Hv-com-step3}
\left[H_{V}, T_{\alpha}^{m, \{M_2\}, \{M_1\}}\right] =& \sum_{\gamma}V_\gamma\sum_{\mathclap{\substack{(x',y') \\ j}}}\hspace{3.2mm}\sum_{\mathclap{\substack{(x'',y'') \\ k}}}{\vphantom{\sum}}'\sum_{\sigma}\left\{\prod_{\beta}\sum_{S[n_2^\beta] = M_2^\beta}\sum_{S[n_1^\beta] = M_1^\beta}\right\}
\left\{\prod_{\eta} O_{(x'',y''),k}^\eta[n_2^\eta]\right\} \nonumber \\
&\times \left[\tilde{n}^\gamma_{(x'',y''),k}(T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma} - (T_\alpha^m)_{(x',y';\, x'',y''),j,k,\sigma}\tilde{n}^\gamma_{(x',y'),j} \right]
\left\{\prod_{\xi} O_{(x',y'),j}^\xi[n_1^\xi]\right\}.
\end{align}
and this becomes
\begin{align}
\left[H_{V}, T_{\alpha}^{m, \{M_2\}, \{M_1\}}\right] = \sum_{\gamma}V_\gamma(M_2^\gamma - M_1^\gamma)\, T_{\alpha}^{m, \{M_2\}, \{M_1\}},
\end{align}
which gives a commutator analogous to Eq.~(\ref{eq:U_comm}).
\section{Details of the Strong Coupling Expansion} \label{app:strong_coup}
In this Appendix we write the decomposed hopping operator as
\begin{align}
T_{\alpha}^{m, \{M_{2}\}, \{M_1\}} = Y_{m,\alpha}^{\{M_{2}\},\{M_1\}},
\end{align}
which satisfies the commutator
\begin{align}
\left[H_0,Y_{m,\alpha}^{ \{M_{2}\},\{M_1\}}\right] = \epsilon_{m}^{ \{M_{2}\},\{M_1\}}Y_{m,\alpha}^{\{M_{2}\},\{M_1\}},
\end{align}
with $\epsilon_{m}^{\{M_{2}\},\{M_1\}} = mU+\sum_{\gamma}V_\gamma(M_2^\gamma - M_1^\gamma)$ and $H_0 = H_U + H_V$.
Recall that solutions for $S_n$ and $H_{T,n}'$ are required such that Eq.~(\ref{eq:EoMconstraint}) is satisfied to some desired order in $1/U$. For $S_1$ we must solve
\begin{align} \label{eq:s1_solve_1}
\big[H_0,[S_1,\tilde{H}_0]+H_T\big] = 0,
\end{align}
and remembering that $\tilde{H}_0 = H_0/U$ it is possible to rearrange this as
\begin{align} \label{eq:s1_solve_2}
\big[\tilde{H}_0,[\tilde{H}_0,S_1]]&= [\tilde{H}_0,H_T]
=\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}\left[\tilde{H}_0,Y_{m,\alpha}^{ \{M_{2}\},\{M_1\}}\right]
= \sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}\frac{\epsilon_{m}^{\{M_{2}\},\{M_1\}}}{U} Y_{m,\alpha}^{ \{M_{2}\},\{M_1\}}.
\end{align}
We can verify that $S_1$ takes the form:
\begin{align} \label{eq:s1_gen}
S_1 = \tilde{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}}\hspace{3mm}
\frac{U}{\epsilon_{m}^{\{M_{2}\},\{M_1\}}} Y_{m,\alpha}^{ \{M_{2}\},\{M_1\}},
\end{align}
where the tilde over the sum indicates that the all terms in the sum which have $\epsilon_{m}^{ \{M_{2}\},\{M_1\}} = 0$ are excluded.
Substituting this into Eq.~(\ref{eq:s1_solve_2}) gives
\begin{align}
\tilde{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}}\hspace{3mm} \frac{U}{\epsilon_{m}^{ \{M_{2}\},\{M_1\}}}
\left(\frac{\epsilon_{m}^{\{M_{2}\},\{M_1\}}}{U}\right)^2 Y_{m,\alpha}^{ \{M_{2}\},\{M_{1}\}}
= \sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}\frac{\epsilon_{m}^{\{M_{2}\},\{M_1\}}}{U} Y_{m,\alpha}^{ \{M_{2}\},\{M_1\}},
\end{align}
and since it is possible to drop the $\epsilon_{m}^{\{M_{2}\},\{M_1\}} = 0$ terms from the sum on the right hand side with no consequence this becomes
\begin{align}
\tilde{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}}\frac{\epsilon_{m}^{ \{M_{2}\},\{M_1\}}}{U} Y_{m,\alpha}^{ \{M_{2}\},\{M_1\}}
= \tilde{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}}\frac{\epsilon_{m}^{\{M_{2}\},\{M_1\}}}{U} Y_{m,\alpha}^{\{M_{2}\},\{M_1\}},
\end{align}
verifying that the form of $S_1$ in Eq.~(\ref{eq:s1_gen}) solves Eq.~(\ref{eq:s1_solve_2}). Now that the form of $S_1$ is known it is possible to calculate the form of the first order correction
\begin{eqnarray}
H_{T,1}' = H_T - [\tilde{H}_0,S_1]
&= &\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}Y_{m,\alpha}^{\{M_{2}\},\{M_1\}} -\tilde{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}}\hspace{3mm}
\frac{U}{\epsilon_{m}^{\{M_{2}\},\{M_1\}}}\frac{\epsilon_{m}^{ \{M_{2}\},\{M_1\}}}{U} Y_{m,\alpha}^{ \{M_{2}\},\{M_1\}} \nonumber \\
&= &\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}Y_{m,\alpha}^{\{M_{2}\},\{M_1\}} - \tilde{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}}
Y_{m,\alpha}^{ \{M_{2}\},\{M_1\}} \nonumber \\
&= & \sum^*_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}Y_{m,\alpha}^{ \{M_{2}\},\{M_1\}}.
\end{eqnarray}
where the star above the sum denotes that the sum only includes terms in which $\epsilon_{m}^{ \{M_{2}\},\{M_1\}} = 0$. In the hopping operator notation this correction takes the form
\begin{align}
H_{T,1}' &=\sum_{\alpha}\hspace{5mm}\sum^*_{\mathclap{\substack{m \\ \{M_{2}\},\{M_{1}\} }}}\hspace{2mm}T_{\alpha}^{m, \{M_{2}\}, \{M_1\}}.
\end{align}
Strictly, it is possible that the starred restriction is satisfied by having the $mU+\sum_{\gamma}V_\gamma(M_2^\gamma - M_1^\gamma)$ perfectly cancel for non-zero $m$, $\{M_1\}$
and $\{M_2\}$, however, if $U$ and the $V_\gamma$ are chosen so that this is never the case we will have $m=0, \{M_2\} = \{M_1\}$, allowing the first order correction to be written as
\begin{align}
H_{T,1}' &= \sum_{\alpha}\sum_{\{M\}}T_{\alpha}^{0, \{M\}, \{M\}},
\end{align}
with no additional conditions, completing the first order perturbation theory.
Before finding $S_2$ and the second order correction it is worthwhile to recall the the Jacobi identity
\begin{align}
\big[A,[B,C]\big] + \big[C,[A,B]\big] + \big[B,[C,A]\big] = 0,
\end{align}
which we can use to write
\begin{align}\label{eq:jacobi}
\left[\tilde{H}_0,\left[Y_{m,\alpha}^{\{M_{2}\},\{M_1\}}, Y_{n,\nu}^{\{N_{2}\},\{N_1\}}\right]\right]
= \frac{\epsilon_{m}^{ \{M_{2}\},\{M_1\}}+\epsilon_{n}^{ \{N_{2}\},\{N_1\}}}{U}
\left[Y_{m,\alpha}^{\{M_{2}\},\{M_1\}}, Y_{n,\nu}^{\{N_{2}\},\{N_1\}}\right].
\end{align}
For $S_2$ we must solve
\begin{align}\label{eq:s_2_int}
\left[\tilde{H}_0,[\tilde{H}_0,S_2]\right] =& \tfrac{1}{2}\left[\tilde{H}_0,\left[S_1,[S_1,\tilde{H}_0]\right]\right] + \left[\tilde{H}_0,\left[S_1,H_T\right]\right] \nonumber \\
=& -\tilde{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}}\hspace{10mm}\tilde{\sum_{\mathclap{\substack{n,\nu \\ \{N_{2}\},\{N_1\} }}}}
\frac{\epsilon_{m}^{ \{M_{2}\},\{M_1\}}+\epsilon_{n}^{ \{N_{2}\},\{N_1\}}}{2\epsilon_{m}^{\{M_{2}\},\{M_1\}}} \left[ Y_{m,\alpha}^{\{M_{2}\},\{M_1\}},
Y_{n,\nu}^{\{N_{2}\},\{N_1\}}\right] \nonumber \\
&+\tilde{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}}\hspace{10mm}\sum_{\mathclap{\substack{n,\nu \\ \{N_{2}\},\{N_1\} }}}
\frac{\epsilon_{m}^{ \{M_{2}\},\{M_1\}}
+\epsilon_{n}^{ \{N_{2}\},\{N_1\}}}{\epsilon_{m}^{\{M_{2}\},\{M_1\}}} \left[ Y_{m,\alpha}^{\{M_{2}\},\{M_1\}},
Y_{n,\nu}^{\{N_{2}\},\{N_1\}} \right] ,
\end{align}
where we used Eq.~(\ref{eq:jacobi}) in Eq.~(\ref{eq:s_2_int}).
It is possible to rewrite
\begin{equation}
\sum_{\mathclap{\substack{n,\nu \\ \{N_{2}\},\{N_1\} }}} \hspace*{3mm} = \hspace*{3mm}
\tilde{\sum_{\mathclap{\substack{n,\nu \\ \{N_{2}\},\{N_1\} }}}} \hspace*{3mm} + \hspace*{3mm}
\sum^*_{\mathclap{\substack{n,\nu \\ \{N_{2}\},\{N_1\} }}} ,
\label{eq:break_up}
\end{equation}
by breaking up the sum into parts which have $\epsilon_{n}^{ \{N_{2}\},\{N_1\}} \neq 0$ and parts which have $\epsilon_{n}^{ \{N_{2}\},\{N_1\}} = 0$.
The commutator equation then becomes
\begin{align}
\left[\tilde{H}_0,[\tilde{H}_0,S_2]\right] =& \tilde{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}}\hspace{10mm}
\tilde{\sum_{\mathclap{\substack{n,\nu \\ \{N_{2}\},\{N_1\} }}}}
\frac{\epsilon_{m}^{ \{M_{2}\},\{M_1\}}+\epsilon_{n}^{ \{N_{2}\},\{N_1\}}}{2\epsilon_{m}^{\{M_{2}\},\{M_1\}}}
\left[ Y_{m,\alpha}^{\{M_{2}\},\{M_1\}}, Y_{n,\nu}^{\{N_{2}\},\{N_1\}}\right] \nonumber \\
&+\tilde{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}}\hspace{10mm}\sum^*_{\mathclap{\substack{n,\nu \\ \{N_{2}\},\{N_1\} }}}
\frac{\epsilon_{m}^{ \{M_{2}\},\{M_1\}}+0}{\epsilon_{m}^{\{M_{2}\},\{M_1\}}}
\left[ Y_{m,\alpha}^{\{M_{2}\},\{M_1\}}, Y_{n,\nu}^{\{N_{2}\},\{N_1\}}\right].
\end{align}
Dropping the $\epsilon_{m}^{ \{M_{2}\},\{M_1\}}+\epsilon_{n}^{ \{N_{2}\},\{N_1\}} = 0$ terms from the first summation has no consequence allowing this to be written as
\begin{align} \label{eq:S2.LHS}
\left[\tilde{H}_0,[\tilde{H}_0,S_2]\right] =& \overline{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}\hspace{10mm}
\sum_{\mathclap{\substack{n,\nu \\ \{N_{2}\},\{N_1\} }}}}\hspace{3mm}
\frac{\epsilon_{m}^{ \{M_{2}\},\{M_1\}}+\epsilon_{n}^{ \{N_{2}\},\{N_1\}}}{2\epsilon_{m}^{\{M_{2}\},\{M_1\}}}
\left[ Y_{m,\alpha}^{\{M_{2}\},\{M_1\}}, Y_{n,\nu}^{\{N_{2}\},\{N_1\}}\right] \nonumber \\
&+\tilde{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}}\hspace{10mm}\sum^*_{\mathclap{\substack{n,\nu \\ \{N_{2}\},\{N_1\} }}} \,\,
\left[ Y_{m,\alpha}^{\{M_{2}\},\{M_1\}}, Y_{n,\nu}^{\{N_{2}\},\{N_1\}} \right] ,
\end{align}
where the over-bar indicates that the sums exclude all terms which have $\epsilon_{m}^{ \{M_{2}\},\{M_1\}} = 0$, $\epsilon_{n}^{ \{N_{2}\},\{N_1\}} = 0$,
and $\epsilon_{m}^{ \{M_{2}\},\{M_1\}}+\epsilon_{n}^{ \{N_{2}\},\{N_1\}} = 0$. Again, rather than solving this equation explicitly for $S_2$ a solution of the form
\begin{align}
S_2 =& \overline{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}\hspace{10mm}
\sum_{\mathclap{\substack{n,\nu \nonumber \\ \{N_{2}\},\{N_1\} }}}}\hspace{3mm} \frac{U^2}{2\epsilon_{m}^{\{M_{2}\},\{M_1\}}
\left(\epsilon_{m}^{ \{M_{2}\},\{M_1\}}+\epsilon_{n}^{ \{N_{2}\},\{N_1\}}\right)} \left[ Y_{m,\alpha}^{\{M_{2}\},\{M_1\}},
Y_{n,\nu}^{\{N_{2}\},\{N_1\}}\right] \\
&+\tilde{ \sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}}\hspace{10mm}\sum^*_{\mathclap{\substack{n,\nu \\ \{N_{2}\},\{N_1\} }}}\hspace{3mm}
\left(\frac{U}{\epsilon_{m}^{ \{M_{2}\},\{M_1\}}}\right)^2
\left[Y_{m,\alpha}^{\{M_{2}\},\{M_1\}}, Y_{n,\nu}^{\{N_{2}\},\{N_1\}}\right],
\end{align}
can be guessed and easily verified by substitution into Eq.~(\ref{eq:S2.LHS}). Now that both $S_1$ and $S_2$ are known it is possible to calculate the second order correction
\begin{align}
H_{T,2}' =& \left[S_1,H_T\right]+\tfrac{1}{2}\left[S_1,[S_1,\tilde{H}_0]\right]+[S_2,\tilde{H}_0] \nonumber \\
=&\tilde{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}}\hspace{10mm}\sum_{\mathclap{\substack{n,\nu \\ \{N_{2}\},\{N_1\} }}}\hspace{3mm}
\frac{U}{\epsilon_{m}^{\{M_{2}\},\{M_1\}}}
\left[Y_{m,\alpha}^{\{M_{2}\},\{M_1\}}, Y_{n,\nu}^{\{N_{2}\},\{N_1\}}\right] \nonumber \\
&-\tilde{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}}\hspace{10mm}\tilde{\sum_{\mathclap{\substack{n,\nu \\ \{N_{2}\},\{N_1\} }}}}\hspace{3mm}
\frac{U}{2\epsilon_{m}^{\{M_{2}\},\{M_1\}}}
\left[Y_{m,\alpha}^{\{M_{2}\},\{M_1\}}, Y_{n,\nu}^{\{N_{2}\},\{N_1\}}\right] \nonumber \\
&-\overline{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}\hspace{10mm}\sum_{\mathclap{\substack{n,\nu \\ \{N_{2}\},\{N_1\} }}}}\hspace{3mm}
\frac{U}{2\epsilon_{m}^{\{M_{2}\},\{M_1\}}}\left[Y_{m,\alpha}^{\{M_{2}\},\{M_1\}}, Y_{n,\nu}^{\{N_{2}\},\{N_1\}}\right] \nonumber \\
&- \tilde{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}}\hspace{10mm}\sum^*_{\mathclap{\substack{n,\nu \\ \{N_{2}\},\{N_1}\} }}\hspace{3mm}
\frac{U}{\epsilon_{m}^{ \{M_{2}\},\{M_1\}}}\left[ Y_{m,\alpha}^{\{M_{2}\},\{M_1\}}, Y_{n,\nu}^{\{N_{2}\},\{N_1\}}\right] ,
\end{align}
and using Eq.~(\ref{eq:break_up}) to break up the sum leads directly to
\begin{align}
H_{T,2}' =& \tilde{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}}\hspace{10mm}\tilde{\sum_{\mathclap{\substack{n,\nu \\ \{N_{2}\},\{N_1\} }}}}\hspace{3mm}
\frac{U}{2\epsilon_{m}^{\{M_{2}\},\{M_1\}}}\left[Y_{m,\alpha}^{\{M_{2}\},\{M_1\}}, Y_{n,\nu}^{\{N_{2}\},\{N_1\}}\right] \nonumber \\
&-\overline{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\} }}}\hspace{10mm}\sum_{\mathclap{\substack{n,\nu \\ \{N_{2}\},\{N_1\} }}}}\hspace{3mm}
\frac{U}{2\epsilon_{m}^{\{M_{2}\},\{M_1\}}}\left[Y_{m,\alpha}^{\{M_{2}\},\{M_1\}}, Y_{n,\nu}^{\{N_{2}\},\{N_1\}}\right] ,
\end{align}
which can be simplified to
\begin{align}
H_{T,2}' &= \overline{\overline{\sum_{\mathclap{\substack{m,\alpha \\ \{M_{2}\},\{M_1\}, }}}\hspace{10mm}
\sum_{\mathclap{\substack{n,\nu \\ \{N_{2}\},\{N_1\} }}}}}\hspace{3mm}
\frac{U}{\epsilon_{m}^{\{M_{2}\},\{M_1\}}}Y_{m,\alpha}^{\{M_{2}\},\{M_1\}}Y_{n,\nu}^{\{N_{2}\},\{N_1\}},
\end{align}
where the double over-bar indicates that the sums include only terms which satisfy $\epsilon_{m}^{ \{M_{2}\},\{M_1\}}+\epsilon_{n}^{ \{N_{2}\},\{N_1\}} = 0,
\epsilon_{m}^{ \{M_{2}\},\{M_1\}} \neq 0, \epsilon_{n}^{ \{N_{2}\},\{N_1\}} \neq 0$.
To simplify this further it is necessary to go back to the full hopping operator notation in which case
\begin{align}
H_{T,2}' =\, & U \sum_{\alpha,\nu}\hspace{5mm}\overline{\overline{\sum_{\mathclap{\substack{ \{M_{2}\}, \{M_{1}\} \\ \{N_{2}\}, \{N_{1}\}}}}
\hspace{5.7mm}\sum_{\mathclap{\substack{ m, n }}}}} \hspace{1mm}
\frac{ T_{\alpha}^{m, \{M_{2}\}, \{M_1\}} T_{\nu}^{n, \{N_{2}\}, \{N_1\}}}{mU+
\sum_{\gamma}V_\gamma(M_2^\gamma - M_1^\gamma)}.
\end{align}
Now, for the double over-bar restriction to be satisfied for all $U$ and $V_\gamma$ the only
terms that will be in the sum are those that have $n=-m, \{N_{1}\} = \{M_{2}\} - \{M_{1}\} + \{N_{2}\}$,
which allows $H_{T,2}'$ to be reduced to
\begin{align}
H_{T,2}' &= U \sum_{\alpha,\nu}\hspace{5mm}\tilde{\sum_{\mathclap{\substack{ \{M_{2}\}, \{M_{1}\} \\ \{N\},m}}}}
\hspace{3mm}\frac{ T_{\alpha}^{m, \{M_{2}\}, \{M_1\}} T_{\nu}^{-m, \{N\}, \{M_{2}\} - \{M_{1}\} + \{N\}}}{mU+\sum_{\gamma}V_\gamma(M_2^\gamma - M_1^\gamma)},
\end{align}
with no additional conditions, which completes the second order perturbation theory.
\section{Second Order Processes and Contributions}
\label{app:couplings}
There are eight different types of second order processes, each of which is associated with a particular coupling.
We list each of the eight couplings, $C_n$, with $n = 1$ to $8$, along with an illustrative example for each coupling in Table~\ref{tab:one},
where
\begin{eqnarray}
C_1 = \frac{{t_b}^2}{\sum_\gamma V_\gamma (M_2^\gamma - M_1^\gamma) } , \quad C_2 = \frac{{t_b}^2}{-U + \sum_\gamma V_\gamma (M_2^\gamma - M_1^\gamma) } , \nonumber \\
C_3 = \frac{{t_p t_q}}{\sum_\gamma V_\gamma (M_2^\gamma - M_1^\gamma) } , \quad C_4 = \frac{{t_q t_p}}{-U + \sum_\gamma V_\gamma (M_2^\gamma - M_1^\gamma) } , \nonumber \\
C_5 = \frac{{t_q}^2}{\sum_\gamma V_\gamma (M_2^\gamma - M_1^\gamma) } , \quad C_6 = \frac{{t_p}^2}{\sum_\gamma V_\gamma (M_2^\gamma - M_1^\gamma) } , \nonumber \\
C_7 = \frac{{t_q}^2}{-U + \sum_\gamma V_\gamma (M_2^\gamma - M_1^\gamma) } , \quad C_8 = \frac{{t_p}^2}{-U + \sum_\gamma V_\gamma (M_2^\gamma - M_1^\gamma) } . \nonumber
\end{eqnarray}
\begin{center}
\begin{longtable}{|c|c|}
\caption{Examples of Second Order Processes for each of the couplings $C_1$ to $C_8$.}\\
\hline
Process & Contribution to Hamiltonian \\ \hhline{|=|=|}
\shortstack{\includegraphics[scale=0.2]{Proc1.pdf} \\ $i=(x,y)$ and $j=(x+1,y)$ \\ \, } &
\shortstack{ $C_1(\tfrac{1}{2}+P_i^z)(\tfrac{1}{2}+P_j^z)$ \\ \, \\ \, \\ \, \\ \, \\ \, \\ \, \\ \, }\\ \hline
\shortstack{\includegraphics[scale=0.2]{Proc2.pdf}\\ $i=(x,y)$ and $j=(x+1,y)$ \\ \, } &
\shortstack{$C_2(\tfrac{1}{2}+P_i^z)(\tfrac{1}{2}-P_j^z)(\tfrac{1}{2}-2\vec{S_i}\cdot\vec{S_j})$
\\ \, \\ \, \\ \, \\ \, } \\ \hline
\shortstack{\includegraphics[scale=0.2]{Proc3.pdf} \\ $i=(x,y)$ and $j=(x+\tfrac{1}{2},y - \tfrac{1}{2})$ \\ \, } &
\shortstack{$C_3(\tfrac{1}{2}-P_j^z)P_i^-$ \\ \, \\ \, \\ \,\\ \, \\ \, \\ \,}\\ \hline
\shortstack{\includegraphics[scale=0.2]{Proc4.pdf} \\ $i=(x,y)$ and $j=(x+\tfrac{1}{2},y - \tfrac{1}{2})$ \\ \,} &
\shortstack{$C_4(\tfrac{1}{2}+P_j^z)P_i^-(\tfrac{1}{2}-2\vec{S_i}\cdot\vec{S_j})$\\ \, \\ \, \\ \, \\ \, } \\ \hline
\shortstack{\includegraphics[scale=0.2]{Proc5.pdf} \\ $i=(x,y)$ and $j=(x+\tfrac{1}{2},y - \tfrac{1}{2})$ \\ \,} &
\shortstack{$C_{5}(\tfrac{1}{2}+P_i^z)(\tfrac{1}{2}-P_j^z)$ \\ \, \\ \, \\ \,\\ \, \\ \, \\ \,}\\ \hline
\shortstack{\includegraphics[scale=0.2]{Proc6.pdf} \\ $i=(x,y)$ and $j=(x-\tfrac{1}{2},y + \tfrac{1}{2})$ \\ \,} &
\shortstack{$C_6(\tfrac{1}{2}+P_i^z)(\tfrac{1}{2}+P_j^z)$ \\ \, \\ \, \\ \,\\ \, \\ \, \\ \,}\\ \hline
\shortstack{\includegraphics[scale=0.2]{Proc7.pdf} \\ $i=(x,y)$ and $j=(x+\tfrac{1}{2},y + \tfrac{1}{2})$ \\ \, } &
\shortstack{$C_{7}(\tfrac{1}{2}+P_i^z)(\tfrac{1}{2}+P_j^z)(\tfrac{1}{2}-2\vec{S_i}\cdot\vec{S_j})$\\ \, \\ \, \\ \, \\ \, \\ \,} \\ \hline
\shortstack{\includegraphics[scale=0.2]{Proc8.pdf} \\ $i=(x,y)$ and $j=(x-\tfrac{1}{2},y - \tfrac{1}{2})$ \\ \,} &
\shortstack{$C_8(\tfrac{1}{2}-P_i^z)(\tfrac{1}{2}+P_j^z)(\tfrac{1}{2}-2\vec{S_i}\cdot\vec{S_j})$\\ \, \\ \, \\ \, \\ \, \\ \, } \\ \hline
\label{tab:one}
\end{longtable}
\end{center}
\subsection{Tables of Coefficients} \label{App.CoefTable}
In tables \ref{tab:two} to \ref{tab:four} we set out the relationships between the coefficients $C_1$ to $C_8$ listed above and the coefficients that enter the
effective model Eq.~(\ref{eq:Ham_C_ij}). The coefficients vary for each type of pair of dimers. We only list non-zero
coefficients
\renewcommand{\arraystretch}{1.3}
\begin{table}[ht]
\caption{Coefficients for $i = (x,y)$ and $j = (x\pm 1,y)$} \label{tab:coef_1}
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
Coupling & $i = (x,y)$ and $j = (x+ 1,y)$ & $i = (x,y)$ and $j = (x-1,y)$ \\
\hline
$C_{i,j}^0$ & $\frac{C_2}{2}$ & $-\frac{C_2}{2}$ \\
$C_{i,j}^1$ & $-\frac{C_2}{2}$ & $\frac{C_2}{2}$ \\
$C_{i,j}^2$ & $2C_1 - C_2$ & $2C_1 - C_2$ \\
$C_{i,j}^3$ & $-C_2$ & $-C_2$ \\
$C_{i,j}^4$ & $-2C_2$ & $2C_2$ \\
$C_{i,j}^5$ & $2C_2$ & $-2C_2$ \\
$C_{i,j}^6$ & $4C_2$ & $4C_2$ \\
\hline
\end{tabular}
\label{tab:two}
\end{center}
\end{table}
\renewcommand{\arraystretch}{1}
\renewcommand{\arraystretch}{1.3}
\begin{table}[ht]
\caption{Coefficients for $i = (x,y)$ and $j = (x+\tfrac{1}{2},y-\tfrac{1}{2})$ or $j = (x-\tfrac{1}{2},y+\tfrac{1}{2})$} \label{tab:coef_2}
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
Coupling & $i = (x,y)$ and $j = (x+\tfrac{1}{2},y - \tfrac{1}{2})$ & $i = (x,y)$ and $j = (x-\tfrac{1}{2},y+\tfrac{1}{2})$ \\
\hline
$C_{i,j}^0$ & $\frac{1}{2}\left(C_7 - C_8\right)$ & $-\frac{1}{2}\left(C_7 - C_8\right)$ \\
$C_{i,j}^1$ & $\frac{1}{2}\left(C_7 + C_8\right)$ & $-\frac{1}{2}\left(C_7 + C_8\right)$ \\
$C_{i,j}^2$ & $2\left(C_6 - C_5\right) + \left(C_7 - C_8\right)$ & $2\left(C_6 - C_5\right) - \left(C_7 - C_8\right)$ \\
$C_{i,j}^3$ & $-\left(C_7 + C_8\right)$ & $-\left(C_7 + C_8\right)$ \\
$C_{i,j}^4$ & $2\left(C_8 - C_7\right)$ & $2\left(C_7 - C_8\right)$ \\
$C_{i,j}^5$ & $-2\left(C_7 + C_8\right)$ & $2\left(C_7 + C_8\right)$ \\
$C_{i,j}^6$ & $4\left(C_8 - C_7\right)$ & $-4\left(C_7 - C_8\right)$ \\
$C_{i,j}^7$ & $\frac{1}{4}\left(3C_3 + C_4\right)$ & $\frac{1}{4}\left(3C_3 + C_4\right)$ \\
$C_{i,j}^8$ & $-\frac{1}{2}\left(C_3 - C_4\right)$ & $\frac{1}{2}\left(C_3 - C_4\right)$ \\
$C_{i,j}^{9}$ & $C_3 - C_4$ & $C_3 - C_4$ \\
$C_{i,j}^{10}$ & $2\left(C_3 - C_4\right)$ & $-2\left(C_3 - C_4\right)$ \\
\hline
\end{tabular}
\label{tab:three}
\end{center}
\end{table}
\renewcommand{\arraystretch}{1.3}
\begin{table}[ht]
\caption{Coefficients for $i = (x,y)$ and $j = (x+\tfrac{1}{2},y+\tfrac{1}{2})$ or $j = (x-\tfrac{1}{2},y-\tfrac{1}{2})$} \label{tab:coef_2}
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
Coupling & $i = (x,y)$ and $j = (x+\tfrac{1}{2},y + \tfrac{1}{2})$ & $i = (x,y)$ and $j = (x-\tfrac{1}{2},y-\tfrac{1}{2})$ \\
\hline
$C_{i,j}^0$ & $\frac{1}{2}\left(C_7 + C_8\right)$ & $-\frac{1}{2}\left(C_7 + C_8\right)$ \\
$C_{i,j}^1$ & $\frac{1}{2}\left(C_7 - C_8\right)$ & $-\frac{1}{2}\left(C_7 - C_8\right)$ \\
$C_{i,j}^2$ & $2\left(C_6 - C_5\right) + \left(C_7 - C_8\right)$ & $2\left(C_6 - C_5\right) + \left(C_7 - C_8\right)$ \\
$C_{i,j}^3$ & $-\left(C_7 + C_8\right)$ & $-\left(C_7 + C_8\right)$ \\
$C_{i,j}^4$ & $-2\left(C_7 + C_8\right)$ & $2\left(C_7 + C_8\right)$ \\
$C_{i,j}^5$ & $-2\left(C_7 - C_8\right)$ & $2\left(C_7 - C_8\right)$ \\
$C_{i,j}^6$ & $-4\left(C_7 + C_8\right)$ & $-4\left(C_7 - C_8\right)$ \\
$C_{i,j}^{11}$ & $\frac{1}{4}\left(3C_3 + C_4\right)$ & $\frac{1}{4}\left(3C_3 + C_4\right)$ \\
$C_{i,j}^{12}$ & $\frac{1}{2}\left(C_2 - C_3\right)$ & $\frac{1}{2}\left(C_3 - C_4\right)$ \\
$C_{i,j}^{13}$ & $C_3 - C_2$ & $C_3 - C_4$ \\
$C_{i,j}^{14}$ & $2\left(C_3 - C_2\right)$ & $-2\left(C_3 - C_4\right)$ \\
\hline
\end{tabular}
\label{tab:four}
\end{center}
\end{table}
\end{widetext}
\end{appendix}
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
Learning from a few labeled examples to acquire skills for new task is essential to achieve machine intelligence. Take object recognition in personalized self-driving system as an example \cite{bae2020selfdriving}. Learning each user's personal driving preference model forms one task. The system is first deployed in the small city, Rochester. The company later extends its market to New York. The user base of New York is much larger than that of Rochester, causing domain imbalance. Also, after adapting to New York users, the learned user behavior from Rochester will be easily forgotten. Similar scenarios occur when learning to solve NLP tasks on a sequence of different languages \cite{incrementNLP} with imbalanced resources of different languages.
\begin{figure}
\centering
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{figure1.png}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{figure1mechanism.png}
\end{subfigure}
\caption{Illustration of meta learning for few shot object recognition on a sequence of imbalanced domains. Our focused problems including domain change detection, how to manage memory and sample memory tasks for joint training with streaming tasks.}
\label{fig:overview}
\end{figure}
Meta learning is a promising approach for solving such few-shot learning problems. One common assumption of current models is that the task distribution is stationary during meta training.
However, real world scenarios (such as the above self-driving system) are more complex and often involve learning across different domains (environments), with challenges such as: (1) task distributions change among different domains; (2) tasks from previous domains are usually unavailable when training on a new domain; (3) the number of tasks from each domain could be highly imbalanced; (4) domain difficulty could vary significantly in nature across the domain sequence. An example is shown in Figure \ref{fig:overview}. Directly applying current meta learning models to such scenarios is not suitable to tackle these challenges, e.g., the object recognition accuracy of meta learned neural networks generally deteriorates significantly on previous context after adapting to a new environment \cite{gidaris2018dynamic, ren2018incremental, yoon2020xtarnet}.
In this work, we cope with such challenges by considering a more realistic problem setting that (1) learning on a sequence of domains; (2) task stream contains significant domain size imbalance; (3) domain labels and boundaries remain unavailable during both training and testing; (4) domain difficulty is non-uniform across the domain sequence. We term such problem setup as \textit{Meta Learning on a Sequence of Imbalanced Domains with Varying Difficulty} (MLSID).
MLSID requires the meta learning model both adapting to a new domain and retaining the ability to recognize objects from previous domains. To tackle this challenging problem, we adopt replay-based approaches, i.e., a small number of tasks from previous domains are maintained in a memory buffer. Accordingly, there are two main problems that need to be solved: (1) how to determine which task should be stored into the memory buffer and which to be moved out. To address this problem, we propose an adaptive memory management mechanism based on the domain distribution and difficulty, so that the tasks in memory buffer could maximize the retained knowledge of previous domains; (2) how to determine which tasks to sample from memory during meta training. We propose an efficient adaptive task sampling approach to accelerate meta training and reduce gradient estimation variance according to our derived optimal task sampling distribution. Our intuition is that not all tasks are equally important for joint training at different iterations. It is thus desirable to dynamically determine which tasks to sample and to be jointly trained with current tasks to mitigate catastrophic forgetting at each training iteration.
Our contributions are summarized as following:
\begin{itemize}
\item To our best knowledge, this is the first work of meta learning on a sequence of imbalanced domains. For convenient evaluation of different models, we propose a new challenging benchmark consisting of imbalanced domain sequences.
\item We propose a novel mechanism, \textit{``Memory Management with Domain Distribution and Difficulty Awareness''}, to maximize the retained knowledge of previous domains in the memory buffer.
\item We propose an efficient adaptive task sampling method during meta training, which significantly reduces gradient estimation variance with theoretical guarantees, making the meta training process more stable and boosting the model performance.
\item Our method is orthogonal to specific meta learning methods and can be integrated with them seamlessly. Extensive experiments with gradient-based and metric-based meta learning methods on the proposed benchmark demonstrate the effectiveness of our method.
\end{itemize}
\section{Problem Setting}
A series of mini-batch training tasks ${\mathcal{T}}_{1}, {\mathcal{T}}_{2}, \dots, {\mathcal{T}}_{N}$ arrive sequentially, with possible domain shift occurring in the stream, i.e., the task stream can be segmented by continual latent domains, ${\mathcal{D}}_{1}, {\mathcal{D}}_{2}, \dots, {\mathcal{D}}_{L}$. $\mathcal{T}_{t}$ denotes the mini-batch of tasks arrived at time $t$. The domain identity associated with each task remains unavailable during both meta training and testing. Domain boundaries, i.e., indicating current domain has finished
and the next domain is about to start, are unknown. This is a more practical and general setup. Each task $\mathcal{T}$ is divided into training and testing data $\{\mathcal{T}^{train}, \mathcal{T}^{test} \}$. Suppose $\mathcal{T}_t^{train}$ consists of $K$ examples, $\{(\bm{x}^k, \bm{y}^k)\}_{k=1}^{K}$, where in object recognition, $\bm{x}^k$ is the image data and $\bm{y}^k$ is the corresponding object label.
We assume the agent stays within each domain for some consecutive time. Also, we consider a simplified setting where the agent will not return back to previous domains and put the contrary case into future work. Our proposed learning system maintains a memory buffer $\mathcal{M}$ to store a small number of training tasks from previous domains for replay to avoid forgetting of previous knowledge. Old tasks are not revisited during training unless they are stored in the memory $\mathcal{M}$. The total number of tasks processed is much larger than memory capacity.
At the end of meta training, we randomly sample a large number of \textit{unseen few-shot tasks} from each latent domain, ${\mathcal{D}}_{1}, {\mathcal{D}}_{2}, \dots, {\mathcal{D}}_{L}$ for meta testing. The model performance is the average accuracy on all the sampled tasks.
\section{Methodology}
\subsection{Conventional Reservoir Sampling and Its Limitations}
Reservoir sampling (RS) \cite{reservoir1985, ERRing19} is a random sampling method for choosing $k$ samples from a data stream in a single pass without knowing the actual value of total number of items in advance.
Straightforward adoption of RS here is to maintain a fixed memory and
uniformly sample tasks from the task stream. Each task in the stream is assigned equal probability $\frac{n}{N}$ of being moved into the memory buffer, where $n$ is the memory capacity size and $N$ is the total number of tasks seen so far. However, it is not suitable for the practical scenarios previously described, with two major shortcomings: (1) the task distribution in memory can be skewed when the input task stream is highly imbalanced in our setting. This leads to under-representation of the minority domains; (2) the importance of each task varies as some domains are more difficult to learn than others. This factor is also not taken into account with RS.
\begin{figure}
\centering
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{balancememory2.png}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.21\textwidth}
\centering
\includegraphics[width=\textwidth]{piechart3.png}
\caption{reservoir sampling}
\end{subfigure}
\begin{subfigure}[b]{0.26\textwidth}
\centering
\includegraphics[width=\textwidth]{piechart4.png}
\caption{proposed memory management}
\end{subfigure}
\caption{An example of (a) reservoir sampling and (b) proposed memory management method jointly considering domain distribution and difficulty when meta learning on task stream from three latent domains.}
\label{fig:memorymanage}
\end{figure}
To address the above issues, we propose to first detect domain change in the input task stream to associate each task with a latent domain label. We then present a new mechanism, called \textit{Memory Management with Domain Distribution and Difficulty Awareness} by utilizing the associated latent domain label with each task. For simple illustration, we construct an imbalanced input task stream from Miniimagenet, Omniglot and Aircraft as shown in Figure \ref{fig:memorymanage}.
Evidently, the resulting distribution of stored tasks with RS is highly imbalanced and dramatically influenced
by the input task stream distribution. In contrast, our memory management mechanism balances the three domain proportions by jointly considering domain distribution and difficulty.
\textbf{Model Summary}: We first illustrate on our domain change detection component in Section \ref{sec:domaincluster}, which is used for (1) managing and balancing tasks in the memory buffer by incorporating the task difficulty (defined in Section \ref{sec:importancesampling}) to determine whether the new incoming task should be moved into the memory and which old task should be moved out of memory in Section \ref{sec:memorysampling}; (2) adaptive sampling tasks from memory buffer during meta training by dynamically adjusting the sampling probability of each task in the memory according to the task gradient for mitigating catastrophic forgetting in Section \ref{sec:importancesampling}.
\subsection{Online Domain Change Detection} \label{sec:domaincluster}
Online domain change detection is a difficult problem due to: (1) few shot tasks are highly diverse within a single domain; (2) there are varying degree of variations at domain boundaries across the sequence. In our initial study, we found that it is inadequate to set a threshold on the change of mini-batch task loss value for detecting domain change. We thus construct a low dimensional projected space and perform online domain change detection on this space.
\paragraph{Projected space} Tasks $\mathcal{T}_t$ are mapped into a common space $\bm{o}_t = f_{{\bm{\theta}}_t}({\mathcal{T}_t^{train}}) = \frac{1}{K} \sum_{k=1}^{K} f_{{\bm{\theta}}_t}(\{\bm{x}^k \})$ where $K$ is the number of training data and $f_{{\bm{\theta}}_t}$ is the CNN embedding network. The task embedding could be further refined by incorporating the image labels, e.g., concatenating the word embedding of the image categories with image embedding. We leave this direction as interesting future work. To reduce the variance across different few shot tasks and capture the general domain information, we compute the exponential moving average of task embedding $ \bm{O}_{t}$ as $ \bm{O}_{t} = \alpha \bm{o}_t + (1-\alpha)\bm{O}_{t-1}$, where the constant $\alpha$ is the weighting multiplier which encodes the relative importance between current task embedding and past moving average. A sliding window stores the past $m$ ($m$ is a small number) steps moving average, $\bm{O}_{t-1}, \bm{O}_{t-2}, \cdots, \bm{O}_{t-m}$, which are used to form the low dimensional projection vector $\zb_t$, where the $i$-th dimensional element of $\zb_t$ is the distance between $\bm{o}_t$ and $\bm{O}_{t-i}$, $d(\bm{o}_t, \bm{O}_{t-i})$. The projected $m$ dimensional vector $\zb_t$ captures longer context similarity information spanning across multiple consecutive tasks.
\textbf{Online domain change detection} At each time $t$, we utilize the above constructed projected space for online domain change detection. Assume we have two windows of projected embedding of previous tasks $\mathcal{U}^B = \{\zb_{t-2B}, \zb_{t-2B+1}, \cdots, \zb_{t-B-1} \}$ with distribution $Q$ and $\mathcal{V}^{B} = \{\zb_{t-B}, \zb_{t-B+1}, \cdots, \zb_t \}$ with distribution $R$, where $B$ is the window size. In other words, $\mathcal{V}^{B}$ represents the most recent window of projection space (test window) and $\mathcal{U}^B$ represents the projection space of previous window (reference window). $\mathcal{U}^B$ and $\mathcal{V}^B$ are non-overlapping windows. For notation clarity and presentation convenience, we use another notation to denote the $\mathcal{U}^B = \{\ub_1, \ub_2, \cdots, \ub_B \}$ and $\mathcal{V}^B = \{V_B_1, V_B_2, \cdots, V_B_B \}$, i.e., $\ub_i = \zb_{t-2B+i-1}$ and $V_B_i = \zb_{t-B+i-1}$. Our general framework is to first measure the distance between the two distributions $Q$ and $R$, $d(Q, R)$; then, by setting a threshold $b$, the domain change is detected when $d(Q, R)>b$. Here, we use Maximum Mean Discrepancy (MMD) to measure the distribution distance. Following \cite{li2018scan}, the MMD distance between $Q$ and $R$ is defined as:
\begin{equation}
\text{MMD}[\mathcal{F}, Q, R] := \sup\limits_{f\in \mathcal{F}} \{\mathbb{E}_{\ub \sim Q} [f(\ub)] - \mathbb{E}_{V_B \sim R} [f(V_B)] \}
\end{equation}
U-statistics \cite{JMLRgretton12a} can be used for estimating ${\text{MMD}}^2$:
\begin{equation}
W^B_t = \text{MMD}^2[\mathcal{U}^{B}, \mathcal{V}^{B}] = \frac{1}{B(B-1)} \sum_{i\neq j}^{B} h(\ub_i, \ub_j, V_B_i, V_B_j)
\end{equation}
and $h(\cdot)$ is defined as:
\begin{equation}
h(\ub_i, \ub_j, V_B_i, V_B_j) = k(\ub_i, \ub_j) + k(V_B_i, V_B_j) - k(\ub_i, V_B_j) - k(\ub_j, V_B_i)
\end{equation}
where $k(\cdot, \cdot)$ is RKHS kernel. In this paper, we assume
RBF kernel $k(x, x^{\prime}) = exp(-||x- x^{\prime}||^2/2 \sigma^2) $ is used.
The detection statistics at time $t$ is $W^B_t$.
If $Q$ and $R$ are close, $W^B_t$ is expected to be small, implying small probability of existence of domain change. If $Q$ and $R$ are significantly different distributions, $W^B_t$ is expected to be large, implying higher chance of domain shift. Thus, $W^B_t$ characterizes the chance of domain shift at time $t$.
We then test on the condition of $W^B_t>b$ to determine whether domain change occurs, where $b$ is a threshold. Each task $\mathcal{T}_t$ is associated with a latent domain label $L_t$, $L_0=0$. If $W^B_t>b$, $L_t = L_{t-1}+1$, i.e., a new domain arrives (Note that the actual domain changes could happen a few steps ago, but for simplicity, we could assume domain changes occur at time $t$); otherwise, $L_t = L_{t-1}$, i.e., the current domain continues. We leave the more general case with domain revisiting as future work. How to set the threshold is a non-trivial task and is described in the following.
\textbf{Setting the threshold} Clearly, setting the threshold $b$ involves a trade-off between two aspects: (1) the probability of $W^B_t>b$
when there is no domain change; (2) the probability of
$W^B_t>b$ when there is domain change. As a result, if the domain similarity and difficulty vary significantly, simply setting a fixed threshold across the entire training process is highly insufficient. In other words, adaptive threshold of $b$ is necessary. Before we present the adaptive threshold method, we first show the theorem which characterizes the property of detection statistics $W^B_t$ in the following.
\begin{algorithm}[H]
\small
\caption{Online Domain Change Detection (ODCD).}
\label{alg:adaptthresh}
\begin{algorithmic}[1]
\REQUIRE stream of detection statistics $W^B_t$; constant $\rho$; desired quantile (significance level) $\delta$; Initialize $\mu_0 = 0$ and $\mu_0^{(2)} = 0$
\STATE \textbf{Function} ODCD ($W^B_t$, $\rho$, $\delta$)
\STATE $\db = False$; // indicator of domain shift
\STATE $\mu_t = (1 - \rho)\mu_{t-1} + \rho (W^B_t)^2$
\STATE $\mu_t^{(2)} = (1 - \rho)\mu_{t-1}^{(2)} + \rho (W^B_t)^4$
\STATE $\sigma_t = \sqrt{\mu_t^{(2)} - \mu_t^2}$
\IF{ $W^B_t> \mu_t + \delta\sigma_t$}
\STATE $\db = True$; //there is domain shift at time $t$
\ENDIF
\RETURN $\db$
\STATE \textbf{EndFunction}
\end{algorithmic}
\end{algorithm}
\begin{theorem}
Assume $\zb_i$ are drawn i.i.d. from $Q$. Suppose that $\mathbb{E}_{Q}||k(\zb, \cdot) ||^4 < \infty$. Set $\mu \overset{def}{=} \mathbb{E}_{Q} k(\zb, \cdot)$ and $K(\zb, \zb^{\prime}) \overset{def}{=} \langle k(\zb, \cdot) -\mu, k(\zb^{\prime}, \cdot) -\mu \rangle$. Suppose the eigenvalue $\xi_l$ and eigenvectors $\phi_l^2$ of $K$ satisfy $\xi_l \geq 0$ and $\mathbb{E}_Q \phi_l^2 < \infty$ such that $K(\zb, \zb^{\prime}) = \sum_{l\geq 1} \xi_l \phi_l(\zb) \phi_l(\zb^{\prime})$ and $\langle \phi_l, \phi_{l^{\prime}} \rangle = \textbf{1}_{l=l^{\prime}}$. Then,
\begin{equation}
W^B_t \overset{d}{\to} \beta \sum_{l\geq 1} \xi_l Z_l^2
\end{equation}
\end{theorem}
Where $\overset{d}{\to}$ means converge in distribution and $(Z_l)_{l\geq 1}$ is a collection of infinite independent standard normal random variables and $\beta$ is a constant. The theorem and proof follow from \cite{theoremkernel, newCP}. We can observe that $W^B_t$ asymptotically follows a distribution formed by a weighted linear combination of independent normal distribution. By Lindeberg’s central limit theorem \cite{vaart_1998}, it is reasonable to assume $W^B_t$ is approximately Gaussian distribution. The problem is thus reduced to estimate its mean $\mu_t$ and $\sigma_t$. The adaptive threshold $b$, following from \cite{newCP}, can be estimated by online approximation, $b = \mu_t + \delta\sigma_t$, where $\delta$ is a constant and set to be the desired quantile of the normal distribution. This adaptive method for online domain change detection is shown in Algorithm \ref{alg:adaptthresh}.
\subsection{Memory Management with Domain Distribution and Difficulty Awareness}
\label{sec:memorysampling}
In this section, we design the memory management mechanism for determining which task to be stored in the memory and which task to be moved out. The mechanism, named \textit{Memory Management with Domain Distribution and Difficulty Awareness} (M2D3), jointly considers the difficulty and distribution of few shot tasks in our setting.
M2D3 first estimates the probability of the current task $\mathcal{T}_t$ to be moved into the memory.
The model will then determine the task to be moved out in the event that a new task move-in happens.
To improve efficiency, we utilize the obtained latent domain information associated with each task (as described in previous section) to first estimate this move-out probability at cluster-level before sampling single task, as in Figure \ref{fig:memory}.
\begin{figure}[h]
\centering
\includegraphics[width=7.6cm]{memory.png}
\caption{Illustration on the memory management process. Each colored circle represents one cluster in the buffer and each dot denotes one task.}
\label{fig:memory}
\end{figure}
Here we define the notations involved in the following method description. Each task $\mathcal{T}_t$ in the memory is associated with a latent domain label $L_t$ and all the tasks with the same latent domain label form one cluster. $\mathcal{M}_{i}$ denotes the cluster formed by all the tasks with latent domain label $i$ in memory $\mathcal{M}$, $n_{i} = |\mathcal{M}_{i}|$ denotes the number of tasks in $\mathcal{M}_{i}$ and $n = |\mathcal{M}|$ denotes the total number of tasks in memory, and $\mathcal{I}_i$ denotes the importance score of cluster $\mathcal{M}_i$.
\textbf{Probability of new task moving into memory}
When the new task $\mathcal{T}_t$ arrives, the chance of $\mathcal{T}_t$ being stored in memory is estimated, with the basic principle being the more incremental knowledge is brought by $\mathcal{T}_t$, the higher the probability of $\mathcal{T}_t$ being stored. This depends on the difficulty and prevalence of current latent domain. We propose an approach on top of this principle to estimate this probability. The score function of $\mathcal{T}_t$ is defined as:
\begin{equation}
S_{new} = (1-\frac{n_{L_{t}}}{n})\mathcal{I}_{t}^T
\end{equation}
Where $\mathcal{I}_{t}^{T}$ represents the importance for $\mathcal{T}_t$ , which is defined as the task-specific gradient norm in Section \ref{sec:importancesampling}. $n_{L_{t}}$ denotes the number of tasks of current latent domain cluster in memory buffer. $\frac{n_{L_{t}}}{n}$ denotes the prevalence of current latent domain in memory. $\mathcal{I}_{i}$ represents the importance for cluster $\mathcal{M}_i$, which is defined as the cluster-specific gradient norm $G_i$ in Section \ref{sec:importancesampling} (The computation is shared and corresponding terms are computed only once.). The importance of in-memory tasks is defined as $ M_{s} = \frac{1}{L_t - 1} \sum_{i=1}^{L_t - 1} \frac{n_{i}}{n}\mathcal{I}_{i}$. The score function of in-memory tasks is defined as:
\begin{equation}
S_{mem} =\frac{n_{L_{t}}}{n} M_{s}
\end{equation}
The probability of moving $\mathcal{T}_t$ into the memory is:
\begin{equation}\label{eq:in}
P_{in} = \frac{e^{S_{new}}}{e^{S_{new}} + e^{S_{mem}}}
\end{equation}
This task selection mechanism maximizes the incremental knowledge of each task added into memory.
\begin{algorithm}[H]
\small
\caption{Memory Management with Domain Distribution and Difficulty Awareness (M2D3).}
\label{alg:memory-manage}
\begin{algorithmic}[1]
\REQUIRE mini-batch training tasks ${\mathcal{T}}_{t}$; memory tasks $\mathcal{M}$; domain label $L_{t-1}$
\STATE \textbf{Function} M2D3 $(\mathcal{M}, \mathcal{T}_t)$
\STATE calculate the probability $\mathcal{P}_{in}$ to move $\mathcal{T}_t$ into memory as Eq. \ref{eq:in}.
calculate detection statistics of $W^B_t$
\STATE $\db = $ ODCD($W^B_t$, $\rho$, $\delta$); detect domain change by Alg. \ref{alg:adaptthresh}.
\IF {$\db$}
\STATE $L_{t} = L_{t-1} +1$
\STATE $\mathcal{M}_{L_{t}} = \{\}$
\ENDIF
\IF{memory $\mathcal{M}$ is not full}
\STATE $\mathcal{M}_{L_{t}} \leftarrow \mathcal{M}_{L_{t}} \cup \mathcal{T}_t$
\ELSE
\IF {${\mathcal{T}}_{t}$ is moved into memory by Eq. \ref{eq:in}}
\STATE calculate the move-out probability for each cluster $\mathcal{P}_t^i$ and sample cluster $j$ according to Eq. \ref{eq:score} and \ref{eq:sampleout}.
\STATE sample task from $\mathcal{M}_j$ to move out of memory.
\STATE move $\mathcal{T}_t$ into memory $\mathcal{M}_{L_{t}} \leftarrow \mathcal{M}_{L_{t}} \cup {\mathcal{T}}_{t}$
\ENDIF
\ENDIF
\RETURN updated memory buffer $\mathcal{M}$
\STATE \textbf{EndFunction}
\end{algorithmic}
\end{algorithm}
\textbf{Probability of existing tasks moving out of memory}
To improve the efficiency of removing the tasks currently in memory, we perform a hierarchical sampling approach. We perform sampling first at cluster level before focusing on individual tasks, as shown in Figure \ref{fig:memory}.
The estimated probability is correlated with both the size of each cluster in memory and its importance. The factor for each cluster $\mathcal{M}_i$ is defined as:
\begin{equation}\label{eq:score}
\mathcal{A}_i \propto -(1-\frac{n_i}{n})\mathcal{I}_i
\end{equation}
The moving out probability for each cluster $\mathcal{M}_i$ at time $t$ is then defined as
\begin{equation}\label{eq:sampleout}
\mathcal{P}_t^i = \frac{e^{\mathcal{A}_i}}{\sum_{i=1}^{i=L_t - 1} e^{\mathcal{A}_i}}
\end{equation}
The complete mechanism is summarized in Algorithm \ref{alg:memory-manage}.
\subsection{Adaptive Memory Task Sampling for Training} \label{sec:importancesampling}
During meta training, a mini-batch of tasks are sampled from the memory and are jointly trained with current tasks to mitigate catastrophic forgetting.
Direct uniform sampling tasks from memory incurs high variance, and results in unstable training \cite{khodak2019provable, NEURIPS2019_1dba5eed}.
On the other hand, our intuition for non-uniform task sampling mechanism is that the tasks are not equally important for retaining the knowledge from previous domains. The tasks that carry more information are more beneficial for the model to remember previous domains, and should be sampled more frequently. To achieve this goal, we propose an efficient adaptive task sampling scheme in memory that accelerates training and reduces gradient estimation variance. As shown in Figure \ref{fig:tasksampling}, the sampling probability of Miniimagenet and Aircraft are adjusted and increased based on the scheme suggesting the importance of these domains are higher than that of Omniglot for retaining knowledge.
\begin{figure}[h]
\centering
\includegraphics[width=8.0cm]{hierarchicalsampling.png}
\caption{A simple example of uniform task sampling and
our adaptive memory task sampling method for sampling tasks from memory buffer during meta training.}
\label{fig:tasksampling}
\end{figure}
With the task specific loss function $\mathcal{L}_{{\bm{\theta}}}(\mathcal{T}_i) = P(\mathcal{T}^{test}_i|{\bm{\theta}}, \mathcal{T}^{train}_i)$. The optimization objective at time $t$ is defined as minimizing on the loss function of both the new tasks and memory tasks $\mathcal{H}({\bm{\theta}}) = \mathcal{L}_{{\bm{\theta}}}(\mathcal{T}_t) + \sum \limits_{\mathcal{T}_i\in \mathcal{M}} \mathcal{L}_{{\bm{\theta}}}(\mathcal{T}_i)$.
At time $t$, our proposed adaptive sampling mechanism assigns each task $\mathcal{T}_i \in \mathcal{M}$ a probability $q_i^t$ such that $\sum_{i=1} ^{i=n} q_i^t = 1$, we then sample $\mathcal{T}_{i_t}$ based on the distribution $\qb_t$ $= (q_1^t, q_2^t, \cdots, q_n^t)$. We temporally omit the subscript (superscript) $t$ for the following theorem for notation clarity.
\begin{theorem}\label{theorem:tasksample}
Let $\pb(\mathcal{T})$ be the distribution of the tasks in memory $\mathcal{M}$. Then,
\begin{equation}
\mathbb{E}_{\pb(\mathcal{T})} \nabla_{{\bm{\theta}}} \mathcal{L}_{{\bm{\theta}}}(\mathcal{T}) = \mathbb{E}_{\qb(\mathcal{T})} [\frac{\pb(\mathcal{T})}{\qb(\mathcal{T})} \nabla_{{\bm{\theta}}} \mathcal{L}_{{\bm{\theta}}}(\mathcal{T}) ] = \Omega
\end{equation}
Let $\mathbb{V}_{\qb}[\Omega]$ denotes the covariance of the above estimator associated with $\qb$. Then, the trace of $\mathbb{V}_{\qb}[\Omega]$ is minimized by the following optimal $\qb^{*}$
\begin{equation}
\qb^{*}(\mathcal{T}) = \frac{\pb (\mathcal{T}) ||\nabla_{{\bm{\theta}}} \mathcal{L}_{{\bm{\theta}}}(\mathcal{T}) ||_2}{\int \pb (\mathcal{T}) ||\nabla_{{\bm{\theta}}} \mathcal{L}_{{\bm{\theta}}}(\mathcal{T}) ||_2}.
\end{equation}
In particular,
if no prior information is available on task distribution, uniform sampling of tasks from memory is adopted and $\pb (\mathcal{T}) = \frac{1}{n}$, $\qb^{*}(\mathcal{T}_i) = \frac{ ||\nabla_{{\bm{\theta}}} \mathcal{L}_{{\bm{\theta}}}(\mathcal{T}_i) ||_2}{ \sum_{j=1}^n ||\nabla_{{\bm{\theta}}} \mathcal{L}_{{\bm{\theta}}}(\mathcal{T}_j) ||_2}. $
Thus, $w(\mathcal{T}_i) = \frac{\pb(\mathcal{T}_i)}{\qb(\mathcal{T}_i)} = \frac{1}{n \qb(\mathcal{T}_i)}$
\end{theorem}
Proof is provided in Appendix \ref{app:proof}.
The parameters are updated as:
\begin{equation}
{\bm{\theta}}_{t+1} = {\bm{\theta}}_{t} - \eta w_i^{t} \nabla_{{\bm{\theta}}_{t}} \mathcal{L}_{{\bm{\theta}}}(\mathcal{T}_{i_t})
\end{equation}
Where $\eta$ is the learning rate, $w_i^t = \frac{1}{nq_i^t}$.
Similar to standard SGD analysis \cite{pmlr-v97-qian19b}, we define the convergence speed of meta training as the shrinkage of distance to optimal parameters ${\bm{\theta}}^{*}$ between two consecutive iterations $C = - \mathbb{E}_{\qb_t} [||{\bm{\theta}}_{t+1} -{\bm{\theta}}^{*} ||_2^2 - ||{\bm{\theta}}_{t} -{\bm{\theta}}^{*} ||_2^2]$.
Following \cite{pmlr-v80-katharopoulos18a, alain2016variance}, it can be expressed as:
\begin{equation}\label{eq:variance}
C = 2\eta ({\bm{\theta}}_{t} -{\bm{\theta}}^{*}) \Omega - \eta^2 \Omega^{T}\Omega - \eta^2 \text{Tr}(\mathbb{V}_{\qb_t}[\Omega] )
\end{equation}
Theorem \ref{theorem:tasksample} illustrates the optimal task sampling distribution for reducing the gradient variance is proportional to the per-task gradient norm. Minimizing the gradient variance (last term of RHS in Eq.\ref{eq:variance}) as in Theorem \ref{theorem:tasksample} also speeds up the convergence (maximize $C$) as a byproduct. However, it is computationally prohibitive to compute this distribution. We therefore propose efficient approximation to it.
By Section \ref{sec:domaincluster}, each memory task is associated with a latent cluster label. Utilizing this property, we can first sample $R$ (small) tasks from each cluster, then calculate the gradient norm for each cluster as $G_i$.
By doing so, the computational efficiency of the optimal task sampling distribution will be significantly improved. The sampling probability for each cluster is calculated as:
\begin{equation}\label{eq:outprob}
\mathcal{Z}_t^i = \frac{n_i G_i}{\sum_{j=1}^{j=L_t}n_jG_j}
\end{equation}
The sampling scheme is to first sample cluster indexes from memory according to Eq. \ref{eq:outprob}, then randomly sample tasks from the specified clusters. We name this task sampling scheme as \textit{ada\textbf{P}tive m\textbf{E}mory \textbf{T}ask \textbf{S}ampling} (PETS).
Eq. \ref{eq:outprob} illustrates that the original sampling distribution of each cluster (measured by the frequency of each cluster in the memory buffer) is weighted by the corresponding importance of each cluster measured by the gradient norm $G_i$. In practice, the computational efficiency can be further improved by computing the sampling distribution every $s$ steps with the same distribution during each time interval.
PETS is summarized in Algorithm \ref{alg:importance-sampling}.
\begin{algorithm}[H]
\small
\caption{Adaptive Memory Task Sampling (PETS).}
\label{alg:importance-sampling}
\begin{algorithmic}[1]
\REQUIRE A sequence of mini-batch training tasks ${\mathcal{T}}_{1}, {\mathcal{T}}_{2}, \dots, {\mathcal{T}}_{N}$; memory buffer $\mathcal{M}$; model parameters ${\bm{\theta}}$;
\FOR{$t = 1$ to $N$}
\FOR {each cluster $\mathcal{M}_j$ in $\mathcal{M}$}
\STATE sample mini-batch tasks from cluster $\mathcal{M}_j$ and calculate gradient norm $G_j$ for $\mathcal{M}_j$.
\ENDFOR
\STATE calculate the task sampling distribution from each cluster $\mathcal{M}_j$ as in Eq. \ref{eq:outprob}.
\STATE sample tasks $\mathcal{B}$ from $\mathcal{M}$ according to the distribution $\mathcal{Z}_t$ as in Eq. \ref{eq:outprob}.
\STATE update ${\bm{\theta}}$ by meta training on $\mathcal{T}_t \cup \mathcal{B}$
\STATE Memory tasks update
$\mathcal{M} = \text{M2D3} (\mathcal{M}, \mathcal{T}_t)$
\ENDFOR
\end{algorithmic}
\end{algorithm}
\section{Related Work}
\paragraph{Meta Learning:}
Meta learning \cite{metanet} focuses on rapidly adapting to unseen tasks by learning on a large number of similar tasks. Representative works include \cite{matching16, protonet17, Edwards2017TowardsAN, finn17a, PMAML2018, pmlr-v48-santoro16, antoniou2018train, munkhdalai2017meta, learn2learn2016, SNAILICLR18, lee2019meta, Wang2020Bayesian, zhou2021metalearning, ravichandran2020fewshot, tokmakov2018learning, Zhang_2019_ICCV}, etc. All of these methods work on the simplified setting where task distributions are stationary during meta training. Completely different from these works, we focus on the more challenging setting where task distributions are non-stationary and imbalanced.
Online meta learning \cite{finn2019online} stores all previous tasks in online setting to avoid forgetting with small number of tasks. \cite{jerfel2018reconciling} use Dirichlet process mixtures (DPM) to model the latent tasks structure and expand network. By contrast, ours focuses on mitigating catastrophic forgetting with single model when meta learning on imbalanced domain sequences with only limited access to previous domains.
Multi-domain meta learning \cite{triantafillou2019metadataset, tseng2020crossdomain, vuorio2019multimodal} assume tasks from all domains are available during meta training. We focus on the case that each domain in an imbalanced domain sequence sequentially arrives.
\paragraph{Continual Learning:}
Continual learning (CL) aims to maintain previous knowledge when learning on sequentially arriving data with distribution shift. Many works focus on mitigating catastrophic forgetting during the learning process. Representative works include \cite{lopezpaz2017gradient, AGEM19, riemer2018learning, yoon2018lifelong, EWC16, nguyen2017variational, ebrahimi2020adversarial,aljundi2018memory, IL2MCV, aljundi2019taskfree}, etc. \textit{Continual few-shot learning} \cite{antoniou2020defining} (CFSL) focuses on remembering previously learned few-shot tasks in a single domain.
To our best knowledge, the replay-based approach to imbalanced streaming setting of continual learning has been only considered in \cite{aljundi2019gradient, imbalance2020, kim2020imbalanced}. Different from these works, which focus on learning on \textit{a small number of tasks} and aim to generalize to previous tasks, our work focuses on the setting where the model learns on \textit{a large number of tasks} with domain shift and imbalance, and aims to generalize to the \textit{unseen tasks} from previous domains without catastrophic forgetting instead of remembering on a specific task.
\paragraph{Incremental and Continual Few-shot Learning:}
Incremental few-shot learning \cite{gidaris2018dynamic, ren2018incremental, yoon2020xtarnet} aim to learn new categories while retaining knowledge on old categories within a single domain and assume access to the base categories is unlimited. This paper, by contrast, requires good generalization to \textit{unseen} categories in previous domains and access to previous domains is limited.
Continual-MAML \cite{Caccia2020} aims for online fast adaptation to new tasks while accumulating knowledge on old tasks and assume previous tasks can be unlimited revisited. MOCA \cite{harrison2020continuous} works in online learning and learns the experiences from previous data to improve sequential prediction. In contrast, ours focuses on generalizing to previous domain when learning on a large number of tasks with sequential domain shift and limited access to previous domains.
\section{Experiments}
Our method is orthogonal to specific meta learning models and can be integrated into them seamlessly.
For illustration, we evaluate our method on representative meta learning models including (1) gradient-based meta learning \textbf{ANIL} \cite{raghu2020rapid}, which is a simplified model of MAML \cite{finn17a}; (2) metric-based meta learning \textbf{Prototypical Network} (\textbf{PNet}) \cite{protonet17}. Extension to other meta learning models is straightforward.
\textbf{Baselines}:
(1) \textbf{sequential training}, which learns the latent domains sequentially without any external mechanism and demonstrates the model forgetting behavior; (2) \textbf{reservoir sampling (RS)} \cite{reservoir1985}; (3) \textbf{joint offline training}, which learns all the domains jointly in a multi-domain meta-learning setting; (4) \textbf{independent training}, which trains each domain independently. Among them, \textbf{joint offline training} and \textbf{independent training} serve as the performance upper bound.
In addition, since continual learning (CL) methods only
apply to a small number of tasks, directly applying CL methods to our setting with large number of tasks (more than 40K) is infeasible. Instead, we combine several representative CL methods with meta learning base model. We modify and adapt \textbf{GSS} \cite{aljundi2019gradient}, \textbf{MIR} \cite {aljundi2019online}, \textbf{AGEM} \cite{AGEM19} and \textbf{MER} \cite{riemer2018learning} to our setting and combine them with meta learning base models to serve as strong baselines. We denote these baselines as PNet-GSS, ANIL-GSS, etc.
\textbf{Proposed benchmark}
To simulate realistic imbalanced domain sequences, we construct a new benchmark and collect 6 domains with varying degree of similarity and difficulty, including \textbf{Quickdraw} \cite{quickdraw}, \textbf{AIRCRAFT} \cite{maji2013finegrained}, \textbf{CUB} \cite{WelinderEtal2010}, \textbf{Miniimagenet} \cite{matching16}, \textbf{Omniglot} \cite{Omniglot2011}, \textbf{Necessities} from Logo-2K+ \cite{wang2019logo2k}. We resize all images into the same size of $84\times84$. All the methods are compared for 5-way 1-shot and 5-way 5-shot learning. All the datasets are publicly available with more details provided in Appendix \ref{appendix:dataset}. We calculate the average accuracy on unseen testing tasks from all the domains for evaluation purpose.
\textbf{Implementation details}
For ANIL-based \cite{raghu2020rapid} baselines, following \cite{antoniou2018train}, we use a four-layer CNN with 48 filters and one fully-connected layer as the meta learner. For PNet-based \cite{protonet17} baselines, we use a five-layer CNN with 64 filters of kernel size being 3 for meta learning. Following \cite{protonet17}, we do not use any fully connected layers for PNet-based models. Similar architecture is commonly used in existing meta learning literature. We do not use any pre-trained network feature extractors which may contain prior knowledge on many pre-trained image classes, as this violates our problem setting that future domain knowledge is completely unknown. We perform experiments on different domain orderings, with the default ordering being Quickdraw, MiniImagenet, Omniglot, CUB, Aircraft and Necessities. To simulate imbalanced domains in streaming setting, each domain on this sequence is trained on 5000, 2000, 6000, 2000, 2000, 24000 steps respectively. In this setup, reservoir sampling will underrepresent
most domains. All experiments are averaged over three independent runs. More implementation details are given in Appendix \ref{app:detail}.
\subsection{Comparison to Baselines}
We compare our methods to the baselines. The memory maintains 300 batches (2) tasks. Results are shown in Table \ref{tab:protonetbaseline} and \ref{tab:ANILbaseline}. We can observe that our method significantly outperforms baselines by a large margin of $5.21 \%$ for 5-shot learning and $4.95\%$ for 1-shot learning with PNet-based model. For ANIL-based baselines, our method outperforms baselines by $4.60 \%$ for 5-shot learning and $2.19\%$ for 1-shot learning. This shows the effectiveness of our method.
\subsection{Effect of Memory Capacity}
We explore the effect of memory capacity for the performance of baselines and our method.
Table \ref{tab:ablation-memory1} and \ref{tab:ablation-memory2} show the results with memory capacity (batches) of 200, 300 and 500 respectively. Our method significantly outperforms all the baselines in each capacity case.
\begin{table}
\centering
\caption{Comparisons with PNet-based baselines}
\begin{adjustbox}{scale=0.75,tabular= lccc,center}
\begin{tabular}{lrrrrrrr}
\toprule
&&\multicolumn{2}{c}{5-Way 1-Shot}& \multicolumn{2}{c}{5-Way 5-Shots}\\
Algorithm& & ACC & &ACC\\
\midrule
PNet-Sequential && $31.82\pm 0.56$ && $48.21\pm 0.50$ \\
\midrule
PNet-RS && $34.68\pm 1.96$ && $53.69\pm 0.76$ \\
\midrule
PNet-GSS && $36.15\pm 1.59$ && $55.16\pm 0.72$\\
\midrule
PNet-AGEM && $34.07\pm 1.71$ && $52.61\pm 0.68$\\
\midrule
PNet-MIR &&$34.53\pm 1.45$ && $53.91\pm 0.56$\\
\midrule
PNet-MER &&$35.82\pm 1.69$ &&$54.28\pm 0.61$\\
\midrule
PNet-Ours && $\mathbf{41.10\pm 0.42}$ && $\mathbf{60.37\pm 0.32}$ \\
\midrule
Joint-training &&$52.96\pm 0.45$ && $68.56\pm 0.37$ \\
\midrule
Independent-training && $58.25\pm 0.36$ && $72.23\pm 0.29$ \\
\bottomrule
\end{tabular}
\label{tab:protonetbaseline}
\end{adjustbox}
\end{table}
\begin{table}
\centering
\caption{Comparisons with ANIL-based baselines}
\begin{adjustbox}{scale=0.75,tabular= lccc,center}
\begin{tabular}{lrrrrrrr}
\toprule
&&\multicolumn{2}{c}{5-Way 1-Shot}& \multicolumn{2}{c}{5-Way 5-Shots}\\
Algorithm& & ACC & &ACC\\
\midrule
ANIL-Sequential && $30.68\pm 0.67$ && $41.39\pm 0.37$ \\
\midrule
ANIL-RS && $32.11\pm 0.90$ && $48.72\pm 0.79$ \\
\midrule
ANIL-GSS && $31.78\pm 1.08$ && $48.93\pm 0.83$ \\
\midrule
ANIL-AGEM && $32.23\pm 1.21$ && $48.56\pm 0.91$ \\
\midrule
ANIL-MIR && $31.85\pm 0.97$ && $48.34\pm 0.72$ \\
\midrule
ANIL-MER &&$32.72\pm 1.06$ && $49.05\pm 0.96$ \\
\midrule
ANIL-Ours && $\mathbf{34.91\pm 0.73}$ && $\mathbf{53.65\pm 0.56}$\\
\midrule
Joint-training && $52.37\pm 0.72$ && $66.21\pm 0.61$ \\
\midrule
Independent-training && $56.52\pm 0.57$ && $69.67\pm 0.53$ \\
\bottomrule
\end{tabular}
\label{tab:ANILbaseline}
\end{adjustbox}
\end{table}
\begin{table}
\centering
\caption{Effect of memory size for PNet-based baselines}
\begin{adjustbox}{scale=0.75,tabular= lccc,center}
\begin{tabular}{lrrrrrrr}
\toprule
&&\multicolumn{2}{c}{5-Way 1-Shot}& \multicolumn{2}{c}{5-Way 5-Shots}\\
Algorithm& & ACC & &ACC\\
\midrule
PNet-RS ($n= 200$) && $34.12\pm 1.12$ && $53.29\pm 0.42$ \\
PNet-Ours ($n= 200$) && $40.11\pm 0.73$ && $59.86\pm 0.27$ \\
\midrule
PNet-RS ($n= 300$) && $34.68\pm 1.96$ && $53.69\pm 0.76$ \\
PNet-Ours ($n= 300$) && $41.10\pm 0.42$ && $60.37\pm 0.32$ \\
\midrule
PNet-RS ($n= 500$) && $35.67\pm 0.82$ && $55.95\pm 0.79$\\
PNet-Ours ($n= 500$) && $41.82\pm 0.90$ && $61.05\pm 0.60$\\
\bottomrule
\end{tabular}
\label{tab:ablation-memory1}
\end{adjustbox}
\end{table}
\begin{table}
\centering
\caption{Effect of memory size for ANIL-based baselines}
\begin{adjustbox}{scale=0.75,tabular= lccc,center}
\begin{tabular}{lrrrrrrr}
\toprule
&&\multicolumn{2}{c}{5-Way 1-Shot}& \multicolumn{2}{c}{5-Way 5-Shots}\\
Algorithm& & ACC & &ACC\\
\midrule
ANIL-RS ($n= 200$) && $31.03\pm 0.97$ && $45.96\pm 0.81$ \\
ANIL-Ours ($n= 200$) && $32.83\pm 0.71$ && $48.21\pm 0.61$ \\
\midrule
ANIL-RS ($n= 300$) && $32.11\pm 0.90$ && $48.72\pm 0.79$ \\
ANIL-Ours ($n= 300$) && $34.91\pm 0.73$ && $53.65\pm 0.56$\\
\midrule
ANIL-RS ($n= 500$) && $39.35\pm 0.76$ && $53.86\pm 0.68$ \\
ANIL-Ours ($n= 500$) && $42.79\pm 0.67$ && $59.23\pm 0.49$\\
\bottomrule
\end{tabular}
\label{tab:ablation-memory2}
\end{adjustbox}
\end{table}
\subsection{Effect of Domain Ordering}
We also compare to other two orderings: \textbf{Necessities, CUB, Omniglot, Aircraft, MiniImagenet, Quickdraw}; and \textbf{Omniglot, Aircraft, Necessities, CUB, Quickdraw, MiniImagenet}. The results are shown in Appendix \ref{app:moreresults}. In all cases, our method substantially outperforms the baselines.
\subsection{Effect of Different Ratios of Domains}
To explore how the different domain ratios affect the model performance, we did another set of experiments with iterations of 4K, 4K, 3K, 4K, 4K, 22K steps on each domain respectively. The results are shown in Table \ref{tab:differentratio} in Appendix.
\subsection{Effect of Domain Revisiting}
To investigate the effect of domain revisiting on baselines and our method, we perform experiment on the domain sequence with domain revising of Quickdraw. The details and results are shown in Table \ref{tab:domainrevisit} in Appendix. We currently assume that there is no domain-revising, properly handling domain-revisiting is left as interesting future work.
\subsection{Ablation Study}
\textbf{Effect of memory management mechanism}
To verify the effectiveness of M2D3 proposed in section \ref{sec:memorysampling}, Table \ref{tab:gradientvariance} in Appendix shows the experiments with simple reservoir sampling without M2D3 (PNet-RS) and with M2D3 (PNet-Ours (without PETS)) respectively. Our method with M2D3 significantly outperforms baseline by $4.1\%$ and $4.2\%$ respectively. The memory proportion for each latent domain is shown in Figure \ref{fig:memorybalance} in Appendix. For RS baseline, the memory proportion for each domain is highly imbalanced. On the contrary, our memory management mechanism enables the memory proportion for each domain is relatively balanced, demonstrating the effectiveness of our method.
\textbf{Effect of PETS}
To verify the effectiveness of PETS proposed in section \ref{sec:importancesampling}, we compare the gradient variance with uniform sampling and our adaptive task sampling method, the gradient variance during training is shown in Figure \ref{fig:gradientvariance} in Appendix. We can see that our adaptive task sampling achieves much less gradient variance especially when training for longer iterations. Table \ref{tab:gradientvariance} in Appendix shows that with PETS, the performance is improved by more than $2.2\%$ and $2.4\%$ for 1-shot and 5-shot learning respectively.
\section{Conclusion}
This paper addresses the forgetting problem when meta learning on non-stationary and imbalanced task distributions. To address this problem, we propose a new memory management mechanism to balance the proportion of each domain in the memory buffer. Also, we introduce an efficient adaptive memory task sampling method to reduce the task gradient variance. Experiments demonstrate the effectiveness of our proposed methods.
For future work, it would be interesting to meta learn the proportional of each domain automatically.
\textbf{Acknowledgements}
This research was supported in part by NSF through grants IIS-1910492.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
The {bittide}\ system is designed to enable synchronous execution
at large scale without the need for a global clock. Synchronous
communication and processing offers significant benefits for
determinism, performance, utilization, and
robustness. Synchronization in {bittide}\ is decentralized -- every
node in the system adjusts its frequency based on the observed
communication with its neighbors. This mechanism, first
proposed in~\cite{spalink_2006}, defines a synchronous logical clock
that is resilient to variations in physical clock frequencies.
The design objective is for {bittide}\ systems to possess shared
logical time. It is \emph{not} a requirement for this logical time to
perfectly match physical time. All machines on the network share a
logical discrete clock that ticks in lockstep. This idea is called
\emph{logical synchronization}, to distinguish it from \emph{physical
synchronization}. Viewed from the inside, the behavior of a logically
synchronized system is identical to that of a system with a single
shared physical clock. Viewed from the outside, the logical time is
fully disconnected from physical wall-clock time, meaning that logical
time steps can vary in physical duration, both over time and between nodes.
Applications running on the system use the logical time to
coordinate their actions, which replaces the need to reference
physical time. Thus, {bittide}\ enables perfect \emph{logical}
synchronization, using imperfect \emph{physical} synchronization.
The decentralized nature of the {bittide}\ synchronization mechanism
enables building large-scale systems that are synchronized with an
accuracy that is otherwise hard or prohibitively expensive to achieve.
Simply overlaying synchronization information onto asynchronous
communication layers is possible, but in practice has led to large
communication requirements and limited
accuracy~\cite{ntp,corbett_spanner_2013,li_sundial_2020}. The
{bittide}\ system instead achieves synchronization using
\emph{low-level} data flows inherent to serial data links. There is no
communication overhead (in-band signaling) required by the
synchronization mechanism, as the continuous data (meaningful or not)
exchanged at the physical layer is sufficient to provide the necessary
input to our control system. As we will demonstrate in this paper,
the logical synchronization is accurate, even though the underlying
substrate is only approximately synchronized. This enables building a
much wider class of synchronized systems.
\paragraph{Prior work.}
This particular scheme for synchronization using low-level network
mechanisms originates in~\cite{spalink_2006}. However, other
synchronous network protocols exist, including the heavily-used
SONET~\cite{sonet}. High-level protocols for clock-synchronization
such as NTP are also widely used~\cite{ntp}. Large systems such as
Spanner use proprietary hardware and protocols to keep clocks
as close to each other as
possible~\cite{corbett_spanner_2013,li_sundial_2020}.
The literature discussing synchronization dynamics is large, and we
can only touch upon it briefly here. The behavior of networked
systems that achieve synchronization via feedback control mechanisms
has been widely studied, and the clock frequency behavior here is
similar to that of several other systems which have been analyzed in
the literature. These include flocking models~\cite{boyds}, Markov
chain averaging models~\cite{hastings,boyd2004}, congestion control
protocols~\cite{kelly}, power networks~\cite{strogatz}, vehicle
platooning~\cite{hedrick}, and flocking~\cite{jadbabaie}. The earliest
work to study such coupled oscillator models of synchronization is
Winfree~\cite{winfree1967}.
Several aspects of the {bittide}\ control system are still open
challenges, and we discuss them in Section~\ref{sec:controlobjectives}.
\section{The {bittide}\ synchronization mechanism}
We now describe in more detail the structure of
a {bittide}\ system. We give here for the sake of clarity the simplest
implementation and omit several possible variations and extensions.
We have a network of computers, represented as an undirected graph,
each node being a computer with a single processor. Each edge
connects a pair of computers, and corresponds to a pair of
links, one in each direction.
These connections are direct; in this simple case there are no
switches or routers between neighboring nodes. Bits are sent across
these links grouped into frames. In the simplest case, all frames are
the same (fixed) size. Implementations may choose the frame size,
which is unrestricted by the {bittide}\ system definition. Because the
frames are of fixed size, determining the boundaries between frames is
straightforward and has low overhead.
Consider two neighboring nodes. At each node there is a queue (called
the \emph{elastic buffer}). Frames are added to the tail of the
elastic buffer as they arrive. At the head of the elastic buffer,
frames are removed from the buffer and read by the processor. Whenever
a frame is removed from the buffer, a new frame is sent on the
outgoing link back to the sender. Thus each edge between two nodes on
the graph corresponds to \emph{four} objects; a link in each
direction, and an elastic buffer at each node. Because removing a
frame from the head of the queue is always consequent with sending a
new frame on the outgoing link, if we are interested solely in the
network dynamics (and not the actual data in the frames) we can
conceptually view these frames as identical; it would be the same if
each node simply sent back the frames it received, after they
propagated through the elastic buffer. In this sense, the two links
and two buffers form a closed cycle, with frames flowing around
perpetually, as illustrated in Figure~\ref{fig:twonodes}.
\begin{figure}[ht]
\centerline{\begin{overpic}[width=0.21\textwidth]{buffers2}
\put(50,15){\clap{$\to$ link}}
\put(12,-4){\clap{$i$}}
\put(90,-4){\clap{$j$}}
\put(50,80){\clap{link $\leftarrow$}}
\put(100,50){\parbox[t]{1in}{elastic\\ buffer}}
\put(0,50){\llap{\parbox[t]{0.4in}{elastic\\ buffer}}}
\end{overpic}}
\caption{Two adjacent nodes.}
\label{fig:twonodes}
\end{figure}
There is one more wrinkle to this behavior. If a node $i$ has, say
$d_i$ neighbors, then it has $d_i$ incoming links (with $d_i$ elastic
buffers), and $d_i$ outgoing links. The critical feature here is the
timing; frames are sent simultaneously on all outgoing links. It is
worth pointing out that exact simultaneity is not necessary. Instead
we simply need the different transmissions to remain in discrete
lockstep with each other, so that transmissions occur in stages of
$d_i$ frames, one on each link. The next stage does not start until
the previous stage has finished.
We can view each edge as a ring with two buffers and two links, and
frames moving around the ring. Because of the above timing behavior,
each ring turns in lockstep with all other rings. This holds even
though the links may have different physical latency and the buffer
occupancies may vary. The rings behave somewhat analogously to
mechanical gears; even if gears have different sizes and hence
different angular rates, at any point where two gears meet, the teeth
of both gears pass the point at the same rate.
At each node there is an oscillator, which drives the processor
clock, which on a modern system might run at a few
GHz. On a {bittide}\ system, the oscillator also drives the network
(possibly via frequency multiplication or division), so that
with each clock tick a set of $d_i$ frames is sent, one to each
neighbor. Each network clock tick corresponds to a stage above.
Because the same oscillator is used for the transmission of
frames as for the processor, the lockstep behavior of the network
induces an identical lockstep behavior for all processors. All nodes
coordinate to synchronize these clocks, as will be discussed below. At
a node, with each tick of the clock, a frame is read from each elastic
buffer, a frame is sent on each outgoing link, and a processor
instruction is executed.
The {bittide}\ system operates at Layer 1 (physical level) in the OSI
network layer model. This is a significant advantage for
synchronization, as it ensures that there is minimal buffering. This
is in contrast with higher-level protocols such as NTP~\cite{ntp},
where synchronization is at a comparatively coarser level.
The frames that are being transmitted around the network have a dual
purpose; they are used for communication, and they are used for
synchronization. For the latter, the content of the frames is
irrelevant to the synchronization, and a processor must simply send a
frame whenever it reads one. If there is no data to be communicated,
it must therefore send a frame containing garbage data. (This is not
generally wasteful, because most standard wired network links operate
this way transparently to the user by default.) Additionally, since
each link is between directly-connected neighbors, it is the
responsibility of higher-level network protocols to perform multi-hop
communication and routing.
The {bittide}\ system enables an entire datacenter, and possibly even
larger networks, to operate in logical synchrony. Such behavior is a
significant departure from how current datacenters work, where large
applications run via networks of asynchronous processes. For many
applications, one of the disadvantages of implementing systems in this
way is \emph{tail-latency}~\cite{dean2013tail}, where delays caused by
a few occasionally-slow processes compound to limit overall
performance. The {bittide}\ system offers a way to eliminate sources of
latency variability via deterministic ahead-of-time scheduling of
processes.
\section{Using feedback control to maintain synchrony}
Consider two neighboring nodes $i$ and $j$. Each node transmits a
frame on an outgoing link whenever it removes a frame from the
corresponding elastic buffer, thus conserving the total number of
frames in the cycle of two links and two buffers. If the oscillator
of $j$ is faster than that of $i$, then $j$ will pull frames from its
buffer more frequently than they are being supplied, and consequently
its elastic buffer will become emptier. Conversely, $i$'s receive
buffer will become fuller. The number of frames in each buffer is
called \emph{buffer occupancy}. Each node has a controller process that
measures buffer occupancy, which provides a feedback signal indirectly
capturing the relative speed of its oscillator compared to the
other. It will then adjust its oscillator speed appropriately.
A node which has $d_i$ neighbors has $d_i$ elastic buffers, which are
all emptied according to the local clock frequency, and filled
according to the frequency of the corresponding neighbor. The
collection of occupancies of these buffers is therefore a feedback
signal, providing information regarding the relative frequencies of
all of the neighboring nodes. The question that remains is how best
to adjust the local oscillator frequency in response to these
signals. This is the responsibility of the control algorithm.
We also need to address the challenge that the
information that each node has about its neighbor is delayed by the
corresponding latency of the physical link; that is, the amount of
time between a frame being sent from the head of the sender's elastic
buffer and it being received into the tail of the recipient's elastic
buffer. When a node changes its frequency, the effects are felt by
all of its neighbors delayed by the corresponding latencies, and this
in turn leads to the neighbors adjusting their frequencies, and so
on. The effects of frequency adjustment therefore propagate through
the network. As a result of these dynamics, both the latencies and the
graph topology can have a significant effect on the overall behavior
of our decentralized synchronization mechanism.
\section{Modeling the dynamics of frames}
Let the set of all nodes in the graph be
$\mathcal{V} = \{1,2,\dots,n\}$. Each node has a clock, whose value at
any time $t$ is a real number $\theta_i(t)$, called the \emph{phase}
of the clock. It is driven by an oscillator whose frequency is
$\omega_i(t)$, and so satisfies the dynamics
\[
\dot\theta_i(t) = \omega_i(t) \quad \text{for all } i \in \mathcal V
\]
The value of $\theta_i(t)$ is called the \emph{local time} at node
$i$, or the time in \emph{local ticks}. The frequency $\omega_i(t)$
may vary over time, both as a result of physical variations such as
temperature, and as a result of adjustments by the controller at node
$i$. This clock drives the frame reading and transmission processes
in the following way. Every time~$t$ at which the clock $\theta_i(t)$
is an integer, node $i$ removes a frame from the head of the
corresponding elastic buffer and sends a frame on the link from $i$ to
$j$. Therefore, if~$\omega_i$ does not change over time, node $i$
sends $\omega_i$ frames per second.
This simple model is enough to determine the location of all of the
frames within the system, as follows. Suppose $s < t$, then the number
of frames that have been sent on the time interval $(s,t]$ by node $i$
is
\[
\sigma_i\interval{s}{t} = \floor{\theta_i(t)} - \floor{\theta_i(s)}
\]
Between any two nodes $i$ and $j$, there are two links, one in each
direction. The link from $i$ to $j$ has latency $l_{ij}$. Note that
the latency includes the time to serialize a frame, transmit it across
the physical link, and deserialize the frame into the elastic buffer.
It does not include the time for the frame to propagate from the tail
to the head of the elastic buffer, and so is not a measure of
communication delay between nodes.
On this link, the number of frames at time~$t$ is therefore
\begin{align*}
\gamma_{ij}(t) &= \sigma_i\interval{t-l_{ij}}{t} \\
&= \floor{\theta_i(t)} - \floor{\theta_i(t-l_{ij})}
\end{align*}
Note that this implies a specific interpretation of boundary points,
so that frames which at time $t$ are exactly at the start of the link
are considered as on the link, and frames that are at the end of the
link are considered as no longer on the link.
Define the number of frames received into the elastic buffer at $j$
from $i$ over the interval $(s,t]$ to be $\rho_{ij}\interval{s}{t}$.
Since the links do not drop frames, the number received is simply
the number sent, delayed by the latency. We have
\[
\rho_{ij}\interval{s}{t} = \sigma_i\interval{s-l_{ij}}{t-l_{ij}}
\]
To determine the occupancy of the elastic buffer, we need to specify
additional information, because while knowledge of $\theta$ informs us
of how many frames arrive and leave the buffer within a particular
time interval, the total number of frames in the buffer also depends
on how many it contained beforehand. So we define $\beta_{ij}(t)$ to
be the occupancy of the elastic buffer at node $j$ associated with the
link from node $i$ at time $t$, and $\beta^0_{ij}$ to be the
occupancy at time $t=0$. The following gives a mathematical definition
of $\beta_{ij}$. We construct $\beta_{ij}$ as the unique function
for which
\[
\beta_{ij}(0) = \beta_{ij}^0
\]
and for which the following difference relationship holds for all $t,s
\geq 0$. The difference between the occupancy at time $t$ and the
occupancy at time $s$ is simply the number of frames received minus
the number sent over that interval.
That is,
\begin{align}
\beta_{ij}(t) - \beta_{ij}(s) \hskip -40pt & \nonumber \\
&= \underbrace{\rho_{ij}\interval{s}{t}}_\text{received} -
\underbrace{\sigma_j\interval{s}{t}\vphantom{\rho_i}}_\text{sent\vphantom{d}} \nonumber \\
&=
\label{eqn:betadiff}
\sigma_i\interval{s-l_{ij}}{t-l_{ij}} -\sigma_j\interval{s}{t} \\
&=
\floor{\theta_i(t-l_{ij})} - \floor{\theta_i(s-l_{ij})}
- \floor{\theta_j(t)} + \floor{\theta_j(s)} \nonumber
\end{align}
Now define
\begin{equation}
\label{eqn:ugn}
\lambda_{ij}(t) = \beta_{ij}(t) -\floor{\theta_i(t-l_{ij})}
+ \floor{\theta_j(t)}
\end{equation}
and notice that equation~\eqref{eqn:betadiff} implies that
\[
\lambda_{ij}(t) = \lambda_{ij}(s) \qquad \text{for all }s,t
\]
Hence $\lambda_{ij}$ is constant. Evaluating equation~\eqref{eqn:ugn} at
$t=0$ shows that it may be determined from $\beta_{ij}^0$ and the
initial conditions for $\theta$, which specify the value of
$\theta(t)$ for all $t \leq 0$. Note that the initial conditions are
not simply given by $\theta(0)$ since the system contains delays, and
is therefore infinite-dimensional. Then we have
\[
\beta_{ij}(t)
= \floor{\theta_i(t-l_{ij})} - \floor{\theta_j(t)} + \lambda_{ij}
\]
Note in particular that $\lambda_{ij}$ and $\lambda_{ji}$ may differ.
Using the above definition of $\gamma_{ij}$, it is convenient to
write $\lambda_{ij}$ as
\[
\lambda_{ij} = \beta_{ij}(t) + \gamma_{ij}(t) + \floor{\theta_j(t)} - \floor{\theta_i(t)}
\]
This number is the buffer occupancy plus the link occupancy, plus the
clock offset between the nodes. We can interpret its invariance as
follows. It is constant because the first two terms sum to the number
of frames on the path from the head of the buffer at $i$ to the head
of the buffer at $j$. The only way this can change is via one or
other clock increasing by one, and those two actions correspond to a
frame being added or removed from this path. Furthermore, we have
\[
\lambda_{ji} + \lambda_{ij} = \beta_{ji}(t) + \gamma_{ji}(t)
+ \beta_{ij}(t) + \gamma_{ij}(t)
\]
which means that the total number of frames on the two links plus two
buffers is conserved.
In the above description we have for simplicity only discussed the
case where all links transmit at the same rate. However, it is
straightforward to extend this model to include links which send
at different rates. To do this, one adds \emph{gearboxes} $g_{ij}$
so that node $i$ sends $g_{ij}$ frames onto link $i\to j$ for every tick
of $\theta_i$. Then the buffer occupancies become
\[
\beta_{ij}(t) = \floor{g_{ij}\theta_i(t-l_{ij})}
- \floor{g_{ij} \theta_j(t)} + \lambda_{ij}
\]
Since this additional complexity does not affect the control mechanism
responsible for synchronization we do not discuss it further here.
In a hardware implementation, each node has memory dedicated to the
elastic buffer for each network interface. Two pointers are stored
which keep track of each end of the buffer, so that adding or removing
a frame does not require data to be moved in memory. However, the
buffer has a fixed size, and so can overflow. Both overflow and
underflow at any node are fatal errors for the {bittide}\ system.
The requirement that the elastic buffers neither overflow nor
underflow means that the difference in clock frequencies at the two
ends of a link cannot stay too large for too long; if it does, either
the buffer at the low-frequency end will overflow or the buffer at the
high-frequency end will underflow, or both.
Define the elastic buffer length to be $\beta^\text{max}$. Then we
must ensure that the frequencies $\omega$ are such that the
occupancies $\beta_{ij}$ satisfy
\[
0 \leq \beta_{ij}(t) \leq \beta^\text{max}
\text{ for all } t\geq 0 \text{ and } i,j \in \mathcal V
\]
This is the fundamental performance requirement that the control
system must enforce. Additionally, it is preferable that
$\beta_{ij}(t)$ be small, since smaller buffers mean smaller
communication latency. Here, by communication latency we mean the
amount of real time (wall-clock time) that it takes for a frame to
leave the head of the source elastic buffer and arrive at the head of
the destination elastic buffer.
Achieving this requirement means we must have clocks that are
operating (on average) at the same frequency at all nodes. In
practice, left alone, no two clocks will remain perfectly
synchronized, and over time their counters will diverge. Some of the
most stable clocks are atomic clocks, which offer a relative error of
about 1 part in $10^{11}$. This much error means that the buffer will
accumulate about 1 bit every 100 seconds on a gigabit link. To avoid
buffer overflow and underflow, we therefore need to use feedback
control to stabilize the buffer occupancies.
\section{Connecting a controller}
The basic idea of the control system is that it can measure the buffer
occupancies $\beta$ and set the frequencies $\omega$. However, there
are several complicating factors. The most fundamental is that the
controller cannot actually set the exact frequency $\omega$. The
oscillator at node $i$ has a frequency at which it will operate
when it is not corrected, which is called the \emph{uncorrected frequency},
denoted by $\omega^\text{u}_i$. When the oscillator is controlled, the
frequency $\omega_i$ is
\[
\omega_i(t) = c_i(t) + \omega^\text{u}_i(t)
\]
where $c_i$ is the \emph{correction} set by the controller. The
uncorrected frequency is typically not known exactly, and is subject
to both manufacturing tolerances and the effects of aging,
temperature, and other physical effects.
It changes with time, and is not
measurable while the system is running. Models for the change in $\omega^\text{u}$
over time include phenomena such as drift and
jitter~\cite{allan1987time}.
At each node, by design, there is no way to measure wall-clock time $t$; the best
the processor can do is observe the local clock $\theta_i$. It is this
notion of time that determines when the occupancy of the elastic
buffers is sampled and when the frequency is updated. Therefore, the
sample-rate at each node varies, depending on the state of the
system. This is one of the sources of nonlinear behavior in the
system.
The model of the system is written using wall-clock time $t$ as
independent variable. However, we do not assume that the controller
can observe $t$. Node $i$ measures the buffer occupancies
at times $t=t_i^0, t_i^1, t_i^2, \dots$. These sample times are
defined by
\[
\theta_i(t_i^k) = \theta_i^0 + kp
\]
Here $p\in\mathbb{Z}_+$ is the sample period, in local ticks. Notice that,
since the initial conditions specify that $\theta_i(0)= \theta_i^0$,
the first sample time is $t_i^0 = 0$. After sampling at time
$t=t^k_i$, node $i$ sets the frequency correction $c_i$ at a
time~$d$ local ticks later. Specifically, the correction is set at times
$t=s_i^0, s_i^1, s_i^2, \dots$, defined by
\[
\theta_i(s_i^k) = \theta_i^0 + kp + d
\]
It is important that the initial phase $\theta_i^0$ is not an
integer. Otherwise, the buffer occupancy is measured at exactly the
times when $\theta_i$ is integral, and those are precisely the times
at which a frame is removed from the elastic buffer. While this is
mathematically well defined, in practice we cannot measure buffer
occupancy exactly at this time. The interpretation of the fractional
part $\theta_i^0 - \floor{\theta_i^0}$ is that it specifies when the
samples are made, relative to the removal of frames from the buffer.
At each time $t_i^k$, the controller at node $i$ measures $y_k^i$, the set
of buffer occupancies at that node,
that is
\[
y^k_i = \{ (j, \beta_{ji}(t_i^k)) \mid j \in \mathop{\mathopfont{neighbors}}(i)\}
\]
Each buffer occupancy is labeled with the neighboring node $j$ that
supplies it. The controller is a function $\chi_i^k$ which
maps the history of these measurements to the correction $c_i^k\in\mathbb{R}$,
according to
\begin{equation}
\label{eqn:ioctrl}
c_i^k = \chi_i^k(y_i^0, \dots, y_i^k)
\end{equation}
This correction is applied on the interval $[s_i^k,
s_i^{k+1})$, and
as a result the frequency is \emph{piecewise constant}, with
\begin{equation}
\dot\theta_i(t) = c_i^k + \omega^\text{u}_i \quad \text{for } t \in[s_i^k,s_i^{k+1})
\end{equation}
\subsection{The abstract frame model}
The model defined by the above equations is called the \emph{abstract
frame model} for {bittide}. It defines the connection between the
controller and the clock, using an ideal abstraction for the frames,
by which one can determine the location of all of the frames using
only the history of~$\theta$. Since this is a transport model, one
might also model it using an advection partial differential equation,
but here we use the delay-differential equation form. We summarize
the model here. For all $t \geq0$, $i \in \mathcal V$, and~$k\in\mathbb{Z}_+$,
\begin{equation}
\label{eqn:afm}
\begin{aligned}
\dot\theta_i(t) &= c_i^k + \omega^\text{u}_i \quad \text{for } t \in[s_i^k,s_i^{k+1}) \\
\beta_{ji}(t) &= \floor{\theta_j(t-l_{ji})} - \floor{\theta_i(t)} + \lambda_{ji}\\
\theta_i(t_i^k) &= \theta_i^0 + kp \\
\theta_i(s_i^k) &= \theta_i^0 + kp + d \\
y^k_i &= \{ (j, \beta_{ji}(t_i^k)) \mid j \in \mathop{\mathopfont{neighbors}}(i)\} \\
c_i^k &= \chi_i^k(y_i^0, \dots, y_i^k)
\end{aligned}
\end{equation}
The initial conditions of the model are
\begin{equation}
\label{eqn:ics}
\theta_i(t) =
\begin{cases}
\theta_i^0 + \omega_i^{(-2)}t & \text{for } t \in [t^\text{e}, 0] \\
\theta_i^0 + \omega_i^{(-1)}t & \text{for } t \in [0, d/\omega_i^{(-1)}]
\end{cases}
\end{equation}
The first of these equations ensures that the dynamics are
well-defined, by specifying the initial value of $\theta$ on a
sufficiently large interval of time for the delay dynamics. The second
equation defines $\theta$ on the period between time zero and the time
at which the first controller action takes effect.
A controller is called \emph{admissible} if
\[
\chi_i^k(y_i^0, \dots, y_i^k) + \omega^\text{u}_i > \omega^{\text{min}}
\]
for all $i,k,y_i^0,\dots,y_i^k$. This ensures that
\begin{equation}
\label{eqn:freqbound}
\omega_i(t) > \omega^{\text{min}}
\end{equation}
The parameters of the model are as follows. The minimum frequency is
$\omega^{\text{min}}>0$. The \emph{epoch} is $t^\text{e}<0$, and it
must satisfy
\[
t^\text{e} \leq -(l_{ij} + d/\omega^\text{min}) \text{ for all } i,j \in \mathcal V
\]
The initial buffer occupancies are $\beta_{ij}^0 \in\mathbb{Z}_+$. The
sampling period is $p \in\mathbb{Z}_+$, the number of clock cycles between
controller updates. The delay $d\in\mathbb{Z}_+$ models the time (in local
ticks) taken to compute the controller update and for the oscillator
to respond to a frequency change. We assume $d<p$. The initial
frequencies are
\[
\omega_i^{(-1)} > \omega^\text{min} \text{ and }
\omega_i^{(-2)}> \omega^\text{min}
\]
The initial clock phases are $\theta^0_i \in\mathbb{R}^+ \backslash \mathbb{Z}$. The uncorrected
oscillator frequencies are $\omega^\text{u}_i\in\mathbb{R}_+$.
The constants
$\lambda_{ij}$ are computed according to
\[
\lambda_{ij} = \beta_{ij}^0 -\floor{\theta_i(-l_{ij})} + \floor{\theta_j^0}
\]
With a state-space decentralized controller, we have
\begin{align*}
\xi_i^{k+1} &= f_i(\xi_i^k, y_i^k) \quad \\
c_i^k &= g_i(\xi_i^k, y_i^k)
\end{align*}
where $\xi_i^k$ is state of the controller at node $i$ and step~$k$. This results
in an input-output controller map of the form~\eqref{eqn:ioctrl}.
\section{Existence and uniqueness of solutions}
It is important to establish the existence and uniqueness of solutions
for the abstract frame model. The model's dynamics are a type of
hybrid system, with nonlinearities, state-dependent multi-rate
sampling and delays, making the analysis challenging. In addition, the
buffer occupancies are discrete, and this allows for the possibility
that the dynamics will exhibit Zeno behavior. We will establish that
admissibility~\eqref{eqn:freqbound} is a critical condition to avoid
Zeno behavior when frames enter or leave the elastic buffer. As such,
we can ensure the existence and uniqueness of solutions for any
controller that satisfies the assumptions above.
We summarize the controller and sampling behavior of the model by writing it as
\begin{align*}
\dot\theta_i(t) &= c_i^k + \omega^\text{u}_i && \text{for } t \in[s_i^k,s_i^{k+1}) \\
\theta_i(s_i^k) &= \theta_i^0 + kp + d \\
c_i^k &= G_i(\theta, s_i^k)
\end{align*}
Here the function $G_i$ contains the construction of the sample times
$t_i^k$, the measurement, and the controller. Notice that the first
argument of $G_i$ is the entire history of $\theta$, not just its
value at a particular time. To state this precisely, define the
(non-minimal) state space for the system as follows. Consider
functions $f:[t^\text{e}, b] \to \mathbb{R}$ with $b>0$ or $f:[t^\text{e}, \infty) \to
\mathbb{R}$, which are piecewise linear, continuous, and satisfy
\[
f'(t) > \omega^\text{min} \text{ for almost all } t\in\mathop{\mathopfont{dom}} f
\]
Let $\mathcal P$ denote the set of all such functions, and $\mathcal X
= \mathcal P^n$. The function $G_i$ has domain $\mathcal D \subset
\mathcal X \times \mathbb{R}$, and $G_i:\mathcal D \to \mathbb{R}$. Specifically, for
$\theta \in \mathcal X$, $i\in\mathcal V$, and $s\in\mathbb{R}$, we have
$\theta,s \in \mathcal D$ if $s\in\mathop{\mathopfont{dom}} \theta$ and $\theta_i(s) =
\theta_i(0) + kp + d$ for some $k\in\mathbb{Z}_+$. Under these conditions,
for $l=0,\dots,k$, let $t^l$ be such that
\[
\theta_i(t^l) = \theta_i(0) +lp
\]
which must exist, since $\theta_i$ is strictly increasing. Evaluate
\begin{align*}
y^l &= \{ (j, \beta_{ji}(t^l)) \mid j \in \mathop{\mathopfont{neighbors}}(i)\} \\
c &= \chi_i^k(y^0, \dots, y^k)
\end{align*}
and define $G_i(\theta,s) = c$. Notice that evaluating $\beta_{ji}$
requires evaluating $\theta$ at times $-l_{ji}$, which lie within the
domain of $\theta$ by the requirements on $t^\text{e}$.
We can now prove the following result.
\begin{theorem}
There exists a unique $\theta\in\mathcal{X}$ satisfying the abstract
frame model~\eqref{eqn:afm} and initial conditions~\eqref{eqn:ics}.
\end{theorem}
\begin{proof}
First let $\theta\in\mathcal X$ be defined by initial
conditions~\eqref{eqn:ics}. We can now define a map $F:\mathcal X
\to \mathcal X$ as follows. Given $\theta \in \mathcal X$, define
$i$ according to
\begin{equation}
\label{eqn:choice}
i = \min \arg \min_i \max \mathop{\mathopfont{dom}} \theta_i
\end{equation}
which simply finds the component of $\theta$ which has the smallest
domain, breaking ties by choosing the smallest index. Let $s=
\max\mathop{\mathopfont{dom}} \theta_i$. We have $\theta,s$ lies in $\mathop{\mathopfont{dom}} G_i$, and so
let $c=G_i(\theta,s)$. We now extend the piecewise linear function
$\theta_i$ by adding a point so that
\[
\theta_i(s + p/(c+\omega^\text{u}_i)) = \theta_i(s)+p
\]
and extending $\theta_i$ to this point via linear interpolation.
Let this newly extended function be $\theta^+$, whose derivative is
by construction bounded below by $\omega^\text{min}$. We define~$F$
by $\theta^+ = F(\theta)$. The proof now proceeds by induction,
constructing a sequence of functions by repeatedly applying~$F$ to
$\theta$. At each step, the domain of one component of $\theta$ is
extended by at least $p/\omega^\text{min}$. This is always the
component with the smallest domain, and so in the limit the domain
of $\theta$ extends to $[t^\text{e},\infty)$ as desired. Uniqueness
follows by observing that the order in which updates are
performed, as determined by the breaking of ties
in~\eqref{eqn:choice}, does not affect the resulting solution.
\end{proof}
\section{Simulation}
The above proof of existence also leads immediately to the following algorithm
for simulating the system.
\begin{tabbing}
\hskip 20pt \= $s \leftarrow 0$ \\[1mm]
\> $\xi_i \leftarrow \xi_i^0 \text{ for all } i \in \mathcal V$ \\[1mm]
\> $\theta_i \leftarrow \text{initial conditions of~\eqref{eqn:ics}} \text{ for all } i \in \mathcal V$ \\[1mm]
\> while $s < t_\text{max}$ \\[1mm]
\> \hskip 20pt \=
$i \leftarrow \min \arg \min_i \max \mathop{\mathopfont{dom}} \theta_i$\\[1mm]
\> \> $s \leftarrow \max \mathop{\mathopfont{dom}} \theta_i$ \\[1mm]
\> \> $t \leftarrow \theta_i^{-1}(\theta_i(s) - d)$ \\[1mm]
\> \> $y \leftarrow \{ (j, \beta_{ji}(t) \mid j \in \mathop{\mathopfont{neighbors}}(i) \}$ \\[1mm]
\> \> $\xi_i \leftarrow f_i (\xi_i, y)$ \\[1mm]
\> \> $c \leftarrow g_i (\xi_i, y)$ \\[1mm]
\> \> $\mathop{\mathopfont{append}}(\theta_i, (s+p/(c+\omega^\text{u}_i), \theta_i(s) + p) )$
\end{tabbing}
Here $\theta_i$ is a piecewise linear increasing function, stored as a
list of knots, {\it i.e.}, pairs of real numbers. It can be evaluated via
linear interpolation, as can the inverse function $\theta_i^{-1}$. The
$\mathop{\mathopfont{append}}$ function simply adds a knot to the end of the list. Also
$\max \mathop{\mathopfont{dom}} \theta_i$ is given by the independent variable of the last
knot.
Code to simulate this system is available at \linebreak \url
{https://bittide.googlesource.com/callisto}.
\subsection{Limitations of the model}
The abstract frame model allows one to determine the location of all
of the individual frames on the network. For control systems, one
often uses models at a coarser level of precision. The AFM does,
however, omit certain phenomena. In particular, the model assumes that
a frame is inserted into the elastic buffer as soon as it has
traversed the link.
The model also does not include limitations which may be imposed by
the particular physical oscillator used. The oscillator frequency is
set by writing the desired frequency offset into a register, the
number of bits of which determines the number of quantization levels
available for control. Depending on the specific system parameters,
this quantization may have a significant effect on the control system.
\section{Control objectives and requirements} \label{sec:controlobjectives}
The objective of {bittide}\ is the maintenance of logical
synchronization. No matter what frequencies the nodes run at, the
logical synchronization happens as a consequence of the lockstep
behavior of the system --- until, that is, the buffer overflows, or
underflows. Thus, the primary objective of the control system is to
manage the buffer occupancy --- keep it within limits and make it as
small as possible. Buffer occupancy translates directly into
communication latency, therefore smaller is better, as long it it does
not underflow. Keeping the buffer within limits also requires
frequencies to not deviate too much, or for too long from each other.
Minimizing the frequency deviation is not a goal. However,
the physical oscillator may drift, and the controller should tolerate
this drift.
A secondary objective is to keep the frequency as large as possible;
if there were no frequency limits, a trivial control strategy for
managing buffer occupancy would simply be to reduce all nodes to close
to zero frequency. Since the processor cores are clocked exactly in
lockstep with frame transmission, higher frequencies are better
because they result in faster computation. The controller should
behave well when encountering constraints imposed by frequency limits.
All of the above objectives must be achieved with a
\emph{decentralized controller.} This has profound consequences for
the application layer, enabling strong security guarantees. For the
control algorithm, it means that the individual controllers cannot
communicate directly with each other to share measurement information
and coordinate their actions.
We impose specific design choices on the controllers for a
{bittide}\ system, motivated by the desire for strong isolation. They
cannot use Paxos-style consensus algorithms, or elect a leader. They
cannot mark or inspect frames. They cannot observe the real time
$t$. They cannot broadcast information on the network. Each node must
update its frequency using only observations of its elastic buffers.
There is however an additional freedom available to
the {bittide}\ controller. After booting, once the buffer occupancies and frequencies
have reached equilibrium, each node may adjust its buffer, discarding
frames or adding frames, and thereby reset its buffer occupancy to a
desired value. This can happen only at bootup, because at this stage
the frames do not yet contain application data. Therefore the
objective that buffer occupancies should be regulated applies only
after the initial transients of the boot phase.
The {bittide}\ control system presented here measures buffer occupancy
and uses this directly to choose the frequency correction. More
sophisticated schemes which estimate frequencies at neighbors are possible
and may offer performance benefits.
There are further practical considerations for the {bittide}\ system
controller. These include allowing nodes to leave or join the network
gracefully, and detection of node and link failures. It must also
work over a wide range of network topologies, and it is preferable
that minimal configuration should be necessary to inform the
controller of the topology and link latencies. The controller should
handle a broad range of frequencies. There are many ways to formalize
these objectives, and the design of a controller that achieves all of
these objectives remains a subject for research.
\section{Example}
\begin{figure}[t]
\centerline{\begin{overpic}[width=1\linewidth]{bms_both2}
\put(0,17){\llap{\small$\omega$}}
\put(0,68){\llap{\small$\beta$}}
\put(38,-1){\clap{\small$t$}}
\put(65,66.3){\tiny$\beta_{31}$}
\put(65,62.5){\tiny$\beta_{32}$}
\put(65,58.5){\tiny$\beta_{21}$}
\put(65,50){\tiny$\beta_{12}$}
\put(65,46.5){\tiny$\beta_{23}$}
\put(65,42.5){\tiny$\beta_{13}$}
\put(8,31.5){\tiny$\omega_3$}
\put(8,14.5){\tiny$\omega_2$}
\put(8,7.3){\tiny$\omega_1$}
\end{overpic}}
\vspace*{3mm}
\caption{Occupancy and frequency for a system with three nodes.}
\label{fig:threenodes}
\end{figure}
Figure~\ref{fig:threenodes} shows an example simulation for a graph
with three nodes and three edges in a triangular topology. The parameters
for this system are
\[
p = 10, \quad d = 2, \quad l_{ji}=1, \quad \theta^0_{i} = 0.1,
\quad \beta_{ji}^0 = 50
\]
for all $i,j$, and the uncorrected frequencies are
\[
\omega^\text{u} = (1.1, 1.4, 2.0)
\]
These parameters are chosen for illustrative purposes only (so that
the variations in occupancy and frequency are visible in the figure).
For example, on a modern system typically the uncorrected frequencies would
be much closer together. The controller used here is a simple
proportional controller, given by
\[
c_i = k_P \sum_{j \in \mathop{\mathopfont{neighbors}}(i)} \beta_{ji}
\]
and the gain $k_P=0.01$. A stability analysis of this system can be
found in~\cite{ecc}.
Some features of this mechanism are apparent. The frequency at a node
is proportional to the sum of the buffer lengths, and so short buffer
lengths at a node will cause that node to have a low rate of
transmission, and so its buffer lengths will increase. Consequently we
expect that frequencies and buffer lengths will tend to equilibrate,
as seen in these plots, and that at equilibrium all frequencies will
be close to each other. We can also observe that the buffer
occupancies at each end of a link almost sum to a constant. This
mirroring of the buffer occupancies is not perfect, due to the effects
of latency.
\section{Conclusions}
In this paper we have presented a model for the {bittide}\ synchronization
mechanism. We have discussed the unique features and
requirements of the control design problem. As these systems are
deployed, we anticipate further research will be developed for these
systems.
\section{Acknowledgments}
We thank Sam Grayson, Sahil Hasan, Sarah Aguasvivas Manzano, Jean-Jacques
Slotine, and Tong Shen for many fruitful
discussions. We particularly thank Sarah Aguasvivas Manzano for
carefully reading the manuscript.
\input{bms_arxiv2.bbl}
\end{document}
|
\section{Introduction}
The development of low-dimensional topology over the last four decades has been greatly influenced by ideas and methods of gauge theory and quantum topology, dating back to the work of Donaldson \cite{Donaldson} and Jones \cite{Jones}. There are many formulations of invariants originating in these theories, but categorically and structurally the two frameworks are quite different: the former is analytic in nature and gives rise to $(3+1)$-dimensional topological quantum field theories, associating to a closed $3$-manifold versions of Floer homology (originally defined in the instanton context in \cite{Floer}). Starting with a quantum group, the latter gives a family of $(2+1)$-dimensional TQFTs, associating to a closed $3$-manifold a collection of numerical Witten-Reshetikhin-Turaev invariants \cite{Witten, RT} at roots of unity.
Our work builds on two theories that are known to recover, for a certain class of $3$-manifolds, Floer homology and ${\rm SU}(2)$ quantum invariants respectively:
lattice cohomology defined by N\'{e}methi \cite{Nem_Lattice_cohomology} and the $\widehat Z$ invariant of Gukov-Pei-Putrov-Vafa \cite{GPPV}.
We show that for negative definite plumbed $3$-manifolds, equipped with a spin$^c$-structure, there is a natural construction giving a common refinement of these two theories. As we discuss below, our construction has novel properties that are not satisfied by either lattice cohomology or the $\widehat Z$ invariant. To explain this in more detail, we first summarize the context considered in this paper.
Motivated by the study of normal surface singularities, \cite{Nem_Lattice_cohomology}
introduced lattice cohomology $\Hbb^*(Y, \mathfrak s)$ of negative definite plumbed $3$-manifolds $Y$ with a spin$^c$ structure $\mathfrak s$. For a subclass of negative definite plumbings, $\Hbb^0(Y, \mathfrak s)$ is isomorphic to Heegaard Floer homology $HF^+(-Y, \mathfrak s)$ defined by Ozsv\'{a}th-Szab\'{o} \cite{MR2113019}, as modules over $\Z[U]$, see Section \ref{sec:lattice cohomology} for a more detailed discussion. The generators of $\Hbb^0(Y, \mathfrak s)$ and the action of $U$ are encoded by the {\em graded root}, a certain infinite tree associated to $(Y, \mathfrak s)$, first defined in \cite{Nem_On_the}.
We construct a refinement, an invariant of $(Y,\mathfrak s)$ which takes the form of a graded root labelled by a collection of $2$-variable Laurent polynomials $P^{}_F$, see Figure \ref{Weighted_graded_root} for an example. As in \cite{Nem_On_the}, the graded root $(R,\chi)$ is defined starting from a negative definite plumbing representing $Y$ and a particular representative of the spin$^c$ structure $\mathfrak s$. The Laurent polynomials $P^{}_F$ labelling the vertices of the graded root in our construction depend on a choice of {\em admissible functions} $F=\{F_n : \Z \to \mathcal{R}\}_{n\geq 0}$ where $\mathcal{R}$ is a commutative ring, see Definition \ref{def:admissible} for details.
We work in the setting of $3$-manifolds which are negative definite plumbing trees, as in \cite{Nem_Lattice_cohomology}.
Our main theorem, proved in Section \ref{sec: The invariant}, is that the result is a topological invariant:
\begin{thm} \label{main thm}
For any admissible family of functions $F$, the weighted graded root $(R, \chi, P^{}_F)$ is an invariant of the $3$-manifold $Y$ equipped with the $\spinc$ structure $\mathfrak s$.
\end{thm}
\begin{figure}[t]
\centering
\includegraphics[height=7.7cm]{GradedRoot2.pdf}
{\scriptsize
\put(-88,-7){$0$}
\put(-128,-7){$0$}
\put(-164,20){$0$}
\put(-55,20){$0$}
\put(-130,20){$0$}
\put(-86,20){$0$}
\put(-113,57){$0$}
\put(-113,87){$0$}
\put(-55,75){$0$}
}
{\scriptsize
\put(-180,82){$\frac{1}{2}q^{\frac{13}{2}}$}
\put(-180,110){$\frac{1}{2}q^{\frac{13}{2}}t^{-2}+\frac{1}{2}q^{\frac{13}{2}}$}
\put(-180,135){$\frac{1}{2}q^{\frac{13}{2}}t^{-2}+\frac{1}{2}q^{\frac{13}{2}}$}
\put(-215,161){$\frac{1}{2}q^{\frac{13}{2}}t^{-2}+\frac{1}{2}q^{\frac{13}{2}}- q^{\frac{23}{2}}t^{-1}$}
\put(-215,188){$\frac{1}{2}q^{\frac{13}{2}}t^{-2}+\frac{1}{2}q^{\frac{13}{2}}- q^{\frac{23}{2}}t^{-1}$}
}
{\small
\put(5,0){$0$}
\put(5,27){$2$}
\put(5,54){$4$}
\put(5,81){$6$}
\put(5,108){$8$}
\put(5,135){$10$}
\put(5,162){$12$}
\put(5,189){$14$}
}
\caption{The weighted graded root for the Brieskorn sphere $\Sigma(2,7,15)$ corresponding to the admissible family $\widehat F$. See grading conventions \ref{grading conv} for an explanation of the numbers in the right column. A more detailed discussion of this example is given in Section \ref{section:conj}.}
\label{Weighted_graded_root}
\end{figure}
In Section \ref{sec:two variable series} we show that the sequence of Laurent polynomials $P^n_F$, obtained by summing the labels over the vertices of the graded root $R$ in grading $\chi=n$, stabilizes to a $2$-variable series. Up to an overall normalization, this limit is a Laurent series in $q$ whose coefficients are Laurent polynomials in $t$. Theorem \ref{thm:convergence} shows that it is an invariant of $(Y,\mathfrak s).$
As discussed in Section \ref{sec:weighted graded root}, there is considerable flexibility in the choice of an admissible family of functions $F$. Each such family gives a weighted graded root and a $2$-variable series which are invariants of the $3$-manifold with a spin$^c$ structure. A particular choice, denoted $\widehat F$ in Section \ref{sec:Zhat_a}, gives rise to a $2$-variable series $\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!_{Y,\mathfrak{s}}(q,t)$. To state its properties, we recall the context of the GPPV invariant.
Based on the study of BPS states and certain supersymmetric $6$-dimensional quantum field theories, \cite{GPV, GPPV} formulated a physical definition of homological invariants of $3$-manifolds, denoted $\mathcal{H}^{i,j}_{\text{BPS}} (Y, \mathfrak s)$.
When the underlying Lie group is ${\rm SU}(2)$, the Euler characterisic of this homology theory is expected to be a $q$-series of the form
\begin{equation} \label{eq: Zhat}
\widehat{Z}_{Y, \mathfrak s}(q) = \sum_{i,j} (-1)^i q^j \, {\rm rk}\, \mathcal{H}^{i,j}_{\rm{BPS}} (Y, {\mathfrak s}) \; \in \; 2^{-c} q^{\Delta_{\mathfrak s}} \Z[[q]],
\end{equation}
for some $c \in \Z_+$ and $\Delta_{\mathfrak s} \in \Q$ depending on $(Y, {\mathfrak s})$. A mathematically rigorous definition of $\widehat{Z}_{Y,\mathfrak s}(q)$ in general is not yet available. A concrete mathematical formulation for negative definite plumbed $3$-manifolds was given in \cite[Appendix A]{GPPV}; also see Section \ref{sec:Zhat_a} below for a more detailed discussion. An earlier instance of these $q$-series, motivated by the study of WRT invariants and of modular forms, was considered in the case of the Poincar\'{e} homology sphere, and more generally Seifert fibered integer homology spheres with three singular fibers in \cite{Lawrence-Zagier}.
For certain classes of negative definite plumbed $3$-manifolds, the $\widehat Z$ $q$-series are known to satisfy (quantum) modularity properties, cf. \cite{Lawrence-Zagier, 3d_modularity, Quantum_modular_forms}. It is not yet known what kinds of modular forms arise as the $q$-series of other $3$-manifolds including more general negative definite plumbings, and other examples such as Dehn surgeries on hyperbolic knots considered in \cite{GM}.
For Seifert fibered integer homology spheres with three singular fibers, $\widehat Z_{Y,{\mathfrak s}_0}(q)$ is a holomorphic function in the unit disk $|q|<1$, and, up to a normalization, radial limits to roots of unity give ${\rm SU}(2)$ WRT invariants \cite[Theorem 3]{Lawrence-Zagier}, see also \cite[Remark 4.5]{GM}. More generally, for rational homology spheres it is conjectured \cite{GPV} that radial limits of a certain linear combination of $\widehat Z_{Y,\mathfrak s}$ over $\spinc$ structures recovers the WRT invariant of $Y$; a precise statement is also given in \cite[Conjecture 3.1]{GM}.
Our next result, established in Sections \ref{sec:two variable series}, \ref{sec:Zhat_a}, relates the $2$-variable series that may be read off from the weighted graded root $(R, \chi, P^{}_{\widehat F})$, as discussed above, to the $\widehat Z$ $q$-series.
\begin{thm} \label{thm: Zhat intro}
The $2$-variable series $\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!_{Y,\mathfrak{s}}(q,t)$ is an invariant of the $3$-manifold $Y$ with a spin$^c$ structure $\mathfrak s$, and its specialization at $t=1$ equals $\widehat{Z}_{Y,\mathfrak s}(q)$.
\end{thm}
The series $\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!_{Y,\mathfrak{s}}(q,t)$ for the Brieskorn sphere $\Sigma(2,7,15)$ in Figure \ref{Weighted_graded_root} is considered in Example \ref{2 7 15 ex involution0}. It is an interesting question whether there are analogues for the $2$-variable series of the properties of $\widehat Z$ discussed above, in particular the limiting behavior of $\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!(q,t)$ along radial limits of the $q$ variable to roots of unity, as well as modularity of other specializations of $\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!(q,t)$.
Some common features of the $\widehat Z$ invariant with the gauge theory setting were apparent in \cite{GPV, GPPV}; indeed bridging the gap between gauge theoretic and quantum invariants was mentioned as a motivation in \cite{GPPV}. Crucially, the $\widehat Z$ $q$-series depends not just on a $3$-manifold, but also on a spin$^c$ structure. Further, it is shown in \cite{GPPcobordism} that certain numerical gauge-theoretic invariants can be recovered from the $\widehat Z$ series, and a physical discussion of a relation with Heegaard Floer homology is given in \cite{GPV}. Our contribution, as stated in Theorems \ref{main thm}, \ref{thm: Zhat intro} is a new structure that is a common refinement of both perspectives; moreover the weighted graded root $(R, \chi, P^{}_F)$ has new features that are not present in either of them.
Lattice cohomology and the $\widehat Z$ $q$-series are known to be invariant under conjugation of the $\spinc$ structure, see Section \ref{section:conj} for further details. Corollary \ref{prop: Zhathat conjugation} states a more subtle transformation of the $2$-variable series $\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!$ under this conjugation. Moreover, Example \ref{prop: Zhathat conjugation} gives a plumbing where conjugate $\spinc$ structures have different weighted graded roots. This example also shows that the Laurent polynomial weights of the graded root carry more information than the limiting series.
A version of the theory developed here is likely to have an analogue for knot lattice homology of \cite{knot} and the invariant of plumbed knot complements introduced in \cite{GM}. This extension is outside the scope of the present paper; we plan to pursue this in future work.
We conclude by recalling the problem of categorifying WRT invariants of $3$-manifolds, which remains a central open question in quantum topology. The $\widehat Z$ $q$-series provide a very promising approach to this problem. Indeed, as discussed above there is a physical prediction $\mathcal{H}^{i,j}_{\text{BPS}} (Y, \mathfrak s)$ for a homology based on the theory of BPS states \cite{GPV, GPPV}. It is interesting to note that the $2$-variable series $\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!_{Y,\mathfrak{s}}$ constructed in this paper is {\em different} from the expected Poincar\'{e} series of the BPS homology, see Section \ref{sec: Recovering q series}, thus indicating the possibility of a different (or more refined) categorification.
{\bf Acknowledgements.} PJ thanks his advisor, Tom Mark, for his continued support and introducing PJ to lattice cohomology. VK is grateful to Sergei Gukov for discussions on the GPPV invariant.
RA was supported by NSF RTG grant DMS-1839968 and the Jefferson Scholars Foundation. PJ was supported by NSF RTG grant DMS-1839968.
VK was supported in part by Simons Foundation fellowship 608604, and NSF Grant DMS-2105467.
\tableofcontents
\newpage
\section{Negative definite plumbed 3-manifolds}
This section summarizes background material and fixes notational conventions on plumbed $3$-manifolds and $\spinc$ structures; the reader is referred to \cite{Neumann, GompfStipsicz} for more details.
\subsection{Plumbings} \label{subsec: negative definite plumbings}
A \emph{plumbing graph} is a finite graph $\Gamma$ equipped with extra data.
For the purposes of this paper, we restrict to plumbing graphs which are trees equipped with a weight function $m : \mathcal{V}(\Gamma) \to \Z$, where $\mathcal{V}(\Gamma)$ is the set of vertices of $\Gamma$. Let $s= \vert\mathcal{V}(\Gamma)\vert$ be the number of vertices of $\Gamma$ and let $\delta_{v}$ be the degree of $v\in \mathcal{V}(\Gamma)$. We will often implicitly choose an ordering on $\mathcal{V}(\Gamma)$, so that $\mathcal{V}(\Gamma) = \{ v_1, \ldots, v_s\}$, and write quantities associated to $v_i\in \mathcal{V}(\Gamma)$ according to the subscript $i$. For example, $m_i = m(v_i)$, $\delta_i = \delta_{v_i}$, etc. Denote by $m, \delta \in \Z^s$ the weight and degree vectors, respectively:
\begin{equation*}
m = (m_1, \ldots, m_s), \; \;
\delta = (\delta_1, \ldots, \delta_s).
\end{equation*}
Assign to $\Gamma$ the symmetric $s\times s$ matrix $M= M(\Gamma)$ with entries:
\[
M_{i,j} =
\begin{cases}
m_{i} & \text{ if } i= j,\\
1 & \text{ if } i\neq j, \text{ and } v_{i} \text{ and } v_{j} \text{ are connected by an edge},\\
0 & {\text{otherwise.} }
\end{cases}
\]
We say $\Gamma$ is \textit{negative definite} if $M$ is negative definite.
From $\Gamma$ we obtain the following manifolds. Consider the framed link $\mathcal{L}(\Gamma)\subset S^3$ given by taking an unknot at each vertex $v$ with framing $m(v)$, and Hopf linking these unknots when the corresponding vertices are adjacent; see Figure \ref{fig:plumbing and link ex} for an example. Let $X=X(\Gamma)$ denote the $4$-manifold obtained by attaching $2$-handles to $D^4$ along $\mathcal{L}(\Gamma)$. Equivalently, $X$ is obtained by plumbing disk bundles over $S^2$ with Euler numbers $m(v)$. Let $Y=Y(\Gamma)$ denote the boundary of $X$, that is the $3$-manifold obtained by Dehn surgery on $\mathcal{L}(\Gamma)$.
\begin{figure}
\centering
\subcaptionbox{A plumbing tree $\Gamma$.
\label{fig:plumbing ex}}[.45\linewidth]
{\includestandalone[scale=.9]{plumbing_ex}
}
\subcaptionbox{The framed link $\mathcal{L}(\Gamma)$.
\label{fig:link ex}
}[.5\linewidth]
{\includestandalone[scale=.7]{link_ex}
}
\caption{A negative definite plumbing $\Gamma$ and its associated framed link $\mathcal{L}(\Gamma)$. The $3$-manifold $Y(\Gamma)$ is the Brieskorn sphere $\Sigma(2,7,15)$.}\label{fig:plumbing and link ex}
\end{figure}
A \emph{negative definite plumbed 3-manifold} is a 3-manifold that is homeomorphic to $Y(\Gamma)$ for some negative definite plumbing tree $\Gamma$. In \cite{Neumann}, the relationship between different representations of a given negative definite plumbed 3-manifold is studied. In particular, from the results in \cite{Neumann}, one can deduce the following theorem which is used in \cite[Proposition 3.4.2]{Nem_Lattice_cohomology} to prove the topological invariance of lattice cohomology:
\begin{thm}[\cite{Neumann}]
Two negative definite plumbing trees $\Gamma$ and $\Gamma'$ represent the same 3-manifold if and only if they can be related by a finite sequence of type (a) and (b) Neumann moves shown in Figure \ref{fig:Neumann moves}.
\end{thm}
\begin{notn}
\label{notation type a and b}
For future reference, we establish notation associated to type (a) and (b) Neumann moves.
\begin{itemize}
\item We use primes to distinguish quantities associated with $\Gamma'$ from those associated with $\Gamma$. For example, $\delta'$ denotes the degree vector for $\Gamma'$, whereas $\delta$ denotes the degree vector for $\Gamma$.
\item For a type (a) move, we will always order vertices so that the $-1$ weighted vertex in $\Gamma'$ which is blown down is labeled by $v_{0}'$, and the two adjacent vertices with weights $m_{1}-1$ and $m_{2}-1$ are labeled by $v_{1}'$ and $v_{2}'$ respectively. In the $\Gamma$ graph of a type (a) move, there is no vertex $v_{0}$ and the two vertices with weights $m_{1}$ and $m_{2}$ are labeled by $v_{1}$ and $v_{2}$ respectively.
\item For a type (b) move, the $-1$ weighted vertex on $\Gamma'$ is labeled by $v_{0}'$ and its adjacent vertex is labeled by $v_{1}'$. In the $\Gamma$ graph, there is no $v_{0}$ vertex and the vertex with weight $m_{1}$ is labeled by $v_{1}$.
\end{itemize}
\end{notn}
\begin{figure}
\centering
\subcaptionbox{Type (a) move.
\label{fig:type a}}[.5\linewidth]
{\includestandalone{type_a}
}
\subcaptionbox{Type (b) move.
\label{fig:type b}
}[.4\linewidth]
{\includestandalone{type_b}
}
\caption{Type (a) and (b) Neumann moves relating negative definite plumbing trees for homeomorphic $3$-manifolds. }\label{fig:Neumann moves}
\end{figure}
\subsection{Identification of \texorpdfstring{$\spinc$}{} structures}\label{sec:identification of spinc structures}
$\Spinc$ structures are important ingredients to both lattice cohomology and the $\widehat{Z}$-invariant. The two theories use different identifications of $\spinc$ structures in terms of plumbing data. We recall a translation between the two identifications, following \cite[Section 4.2]{GM}.
To begin, we describe the relationships between various (co)homology groups of $X$ and $Y$. First, note that $\Gamma$ gives a convenient choice of basis for $H_2(X;\Z)$ in the following way. For $v\in \mathcal{V}(\Gamma)$, let $[v]\in H_2(X;\Z)$ be the class of the $2$-sphere obtained by capping off the core of the $2$-handle corresponding to $v$. Then $H_2(X;\Z) \cong \Z^s$, with a basis given by $\{[v_{1}], \ldots, [v_{s}]\}$. With respect to this basis, the intersection form on $H_2(X;\Z)$
is the bilinear form associated with $M$, $\langle x,y\rangle = x^t M y$. We will also write
\[
\langle-,-\rangle : \Z^s \times \Z^s \to \Z
\]
to denote this bilinear form when $H_2(X;\Z)$ is identified with $\Z^s$ as above.
\begin{rem}
In some of the lattice cohomology literature the intersection form is denoted by $(-,-)$. However, in \cite{GM} the intersection form is denoted using angled brackets $\langle-,-\rangle$, as we do above, and $(-,-)$ instead refers to the usual dot product. To minimize confusion, we will use $\cdot$ for the dot product.
\end{rem}
Since $X$ is a 2-handlebody with no 1-handles, $H_1(X;\Z) = 0$, and we can identify $H^2(X;\Z)$ with $\Hom_{\Z}(H_2(X;\Z),\Z)$. Furthermore, using the above basis of $H_2(X;\Z)$, we have a distinguished isomorphism $\Hom_{\Z}(H_2(X;\Z),\Z) \cong \Z^s$, so that a vector $k\in \Hom_{\Z}(H_2(X;\Z),\Z)$ is represented as $k = (k([v_1]), \ldots, k([v_s])) \in \Z^s$. Combining these two identifications, we get an identification of $H^2(X;\Z)$ with $\Z^s$ such that for $k\in H^2(X,\Z)$ and $x\in H_2(X;\Z)$, we have $k(x) = k \cdot x$. The identifications described above are used throughout the paper.
\begin{defn}
An element $k\in H^2(X;\Z)$ is called \textit{characteristic} if $k\cdot x + \langle x, x\rangle \equiv 0 \bmod 2$ for all $x\in H_{2}(X;\Z)$. We denote the set of characteristic vectors of $X$ by $\Char(X)$.
\end{defn}
In terms of our identification of $H^2(X;\Z)$ with $\Z^s$, it follows that:
\begin{align*}
\Char(X) = m + 2\Z^s.
\end{align*}
It is a standard fact in 4-manifold topology that for simply connected $X$ the map which takes a $\spinc$ structure on $X$ to the first Chern class of its determinant line bundle is a bijection from $\spinc(X)$ to $\Char(X)$, cf. \cite[Proposition 2.4.16]{GompfStipsicz}. Moreover, by restricting $\spinc$ structures to the boundary 3-manifold $Y$, we get the following identification:
\begin{equation}
\label{eq:LC spinc structures}
\spinc(Y) \cong \frac{m+2\Z^s}{2M\Z^s}.
\end{equation}
The right side of the above identification is to be interpreted as the set $m+2\Z^s$ up to the equivalence relation defined by $k\sim k'$ if $k-k'\in 2M\Z^s$. If $k\in m +2\Z^s$, we let $[k]\in \spinc(Y)$ denote the equivalence class containing $k$.
The identification of $\spinc$ structures just described is the one used in the lattice cohomology and Heegaard Floer homology literature (see for example \cite[Section 2.2.2]{Nem_Lattice_cohomology} and \cite[Section 1]{O-S_Floer_Homology_Plumbed}). The identification used in the $\widehat{Z}$ literature is given as follows:
\begin{equation}
\label{eq:GM spinc structures}
\spinc(Y) \cong \frac{\delta+2\Z^s}{2M\Z^s}.
\end{equation}
As in the first identification, for $a\in \delta + 2\Z^s$, we let $[a]\in \spinc(Y)$ denote the equivalence class containing $a$. To avoid confusion between the two $\spinc$ identifications, throughout the paper we will use the letter $k$ for vectors in $m + 2\Z^s$ and the letter $a$ for vectors in $\delta + 2\Z^s$.
The lattice cohomology and $\widehat{Z}$ identifications of $\spinc$ structures are related to each other in the following way. For a fixed plumbing graph $\Gamma$, there is a bijection:
\begin{equation}
\label{eq:translating between spinc identifications}
\frac{m+2\Z^s}{2M\Z^s} \overset{\sim}{\longrightarrow} \frac{\delta+2\Z^s}{2M\Z^s}, \hspace{ 1cm}
[k] \longmapsto [k-(m+\delta)].
\end{equation}
\begin{rem}
If we let $u = (1, 1,\ldots, 1)\in \Z^s$, then $Mu = m+\delta$ and we can alternatively express the above bijection as $[k] \mapsto [k-Mu]$.
\end{rem}
The above bijection is natural with respect to type (a) and type (b) Neumann moves in the sense that if $\Gamma$ and $\Gamma'$ are two plumbing graphs related by either one of these two moves, we get the following commutative diagram:
\begin{center}
\begin{tikzpicture}[baseline= (a).base]
\node[scale=1.2] (a) at (0,0){
\begin{tikzcd}
\large
\frac{m+2\Z^s}{2M\Z^s}
\ar[r] \ar{d}{\alpha} & \frac{\delta+2\Z^s}{2M\Z^s} \ar{d}{\beta}\\
\frac{m'+2\Z^{s+1}}{2M'\Z^{s+1}} \ar[r] & \frac{\delta'+2\Z^{s+1}}{2M'\Z^{s+1}}
\end{tikzcd}
};
\end{tikzpicture}
\end{center}
Here $\alpha, \beta$ are also bijections, which at the level of representatives are defined as follows.
Type (a) move:
\begin{equation}
\label{eq:k' for type a}
\alpha(k) =k':= (0, k) + (1, -1, -1, 0, \ldots, 0), \hspace{.5cm}
\beta(a)=a':=(0, a).
\end{equation}
Type (b) move:
\begin{equation}
\label{eq:k' for type b}
\alpha(k) =k':=(0, k) + (-1, 1, 0,\ldots, 0), \hspace{.5cm}
\beta(a)=a':=(0, a) + (-1, 1, 0,\ldots, 0).
\end{equation}
\section{Lattice cohomology}
\label{sec:lattice cohomology}
Lattice cohomology was introduced by Némethi in \cite{Nem_Lattice_cohomology}, building on earlier work of Ozsváth-Szabó in \cite{O-S_Floer_Homology_Plumbed}. It is a theory which assigns to a given plumbing graph $\Gamma$ and $\spinc$ structure $[k]$, a $\Z[U]$-module:
\begin{align*}
\Hbb^*(\Gamma, [k]) = \bigoplus_{i=0}^{\infty}\Hbb^i(\Gamma, [k])
\end{align*}
Each $\Hbb^i(\Gamma, [k])$ is a $(2\Z)$-graded $\Z[U]$-module. Hence, $\Hbb^*$ is bigraded; it carries a homological grading given by the superscript $i$ as well as an internal $(2\Z)$-grading.
It was shown in \cite{Nem_Lattice_cohomology} that for negative definite plumbings $\Hbb^*$ is invariant under Neumann moves and therefore is a topological invariant. Furthermore, extending results from \cite{O-S_Floer_Homology_Plumbed}, it was shown that for a subset of negative definite plumbed 3-manifolds, namely \textit{almost rational} plumbings, there exists an isomorphism between lattice cohomology and Heegaard Floer homology:
\begin{thm}[{\cite[Theorems 4.3.3 and 5.2.2] {Nem_Lattice_cohomology}}]\label{thm: LC-HF iso}
If $\Gamma$ is almost rational, then as graded $\Z[U]$-modules,
\begin{align}
\begin{aligned}
\label{eq:LC-HF iso}
\Hbb^i(\Gamma, [k])\left[-\max\limits_{k'\in [k]}\frac{(k')^2+s}{4}\right]&\cong
\begin{cases}
HF^+(-Y(\Gamma), [k])&\text{if }i =0\\
0&\text{otherwise}.
\end{cases}
\end{aligned}
\end{align}
Here $-Y$ is $Y$ with the opposite orientation, $(k')^2:=(k')^tM^{-1}k'$, and the left side of the equation is $\Hbb^i(\Gamma, [k])$ with its internal grading shifted up by $-\max\limits_{k'\in [k]}\frac{(k')^2+s}{4}$.
\end{thm}
\begin{rem}\label{rem: k^2 meaning}
The quantity $(k')^2$ has the following geometric meaning: it is the square of the first Chern class of the $\spinc$ structure on $X(\Gamma)$ corresponding to the characteristic vector $k'$. Even if a vector $x$ is not characteristic, we will still define $x^2 := x^tM^{-1}x\in \Q$.
\end{rem}
\begin{rem}\label{rem: d inv}
The minimal internal grading of $\Hbb^0(\Gamma, [k])$ is always equal to zero. Hence, after the grading shift, the minimal grading of the left side of equation (\ref{eq:LC-HF iso}) is equal to $-\max\limits_{k'\in [k]}\frac{(k')^2+s}{4}$. In particular, Theorem \ref{thm: LC-HF iso} implies the $d$-invariant of $-Y$ at the $\spinc$ structure $[k]$ is equal to $-\max\limits_{k'\in [k]}\frac{(k')^2+s}{4}$, or, equivalently, the $d$-invariant of $Y$ at $[k]$ is $\max\limits_{k'\in [k]}\frac{(k')^2+s}{4}$.
\end{rem}
In light of Theorem \ref{thm: LC-HF iso}, we focus our attention on $\Hbb^0$ rather than recalling the full definition of lattice cohomology. However, it is worth noting that there do exist negative definite plumbings with non-trivial $\Hbb^{i}$, $i\geq 1$ (see for example \cite[Example 4.4.1]{Nem_Lattice_cohomology}) and there also exist other classes of plumbed manifolds in which the lattice cohomology - Heegaard Floer homology isomorphism holds, cf. \cite{OSS-spectral-sequence}.
\subsection{Graded roots}
The $(2\Z)$-graded $\Z[U]$-module $\Hbb^0(\Gamma, [k])$ has the nice feature that it can be encoded by a graph, called a \textit{graded root}, proven in \cite[Proposition 4.6]{Nem_On_the} to itself be a topological invariant. We now recall the notion of a graded root $(R, \chi)$ as an abstract object and show how to associate to it a $(2\Z)$-graded $\Z[U]$-module $\Hbb(R,\chi)$. In the next subsection, we show how to obtain a graded root $\groot{[k]}$ from a pair $(\Gamma, [k])$ and we define $\Hbb^0(\Gamma, [k]) = \Hbb\groot{[k]}$. For complete details, see \cite[Section 3]{Nem_On_the}.
\begin{defn}
\label{def:graded root} A \emph{graded root} is an infinite tree $R$, with vertices and edges denoted $\mathcal{V}$ and $\Edges$ respectively, together with a grading function $\chi : \mathcal{V} \to \Z$ satisfying the properties listed below. We write an edge with endpoints $u$ and $v$ as $[u,v]\in \Edges$.
\begin{enumerate}
\item $\chi(u) - \chi(v) = \pm 1$ for any $[u,v] \in \Edges$.
\item $\chi(u) > \min \{ \chi(v), \chi(w) \}$ for any $[u,v], [u,w] \in \Edges$ with $v\neq w$.
\item $\chi$ is bounded below, each preimage $\chi^{-1}(n)$ is finite, and $\vert \chi^{-1}(n) \vert =1$ for sufficiently large $n$.
\end{enumerate}
An isomorphism of graded roots is an isomorphism of the underlying graphs which commutes with the grading functions. For $r\in \Z$, let $(R,\chi)[r] = (R,\chi[r])$ denote the graded root with the same underlying tree and the grading shifted up by $r$, so that $\chi[r](v) = \chi(v) + r$.
\end{defn}
We visualize a graded root by embedding it into the plane such that vertices of the same grading are placed at the same horizontal level, see Figure \ref{fig:graded root example} for an example.
\begin{figure}
\centering
\includestandalone{graded_root_example}
\caption{An example of a graded root.}
\label{fig:graded root example}
\end{figure}
The $(2\Z)$-graded $\Z[U]$-module $\Hbb(R,\chi)$ associated to a graded root $(R, \chi)$ is defined as follows:
\begin{itemize}
\item To each $v\in\mathcal{V}$, we associate a copy of $\Z$, which we denote $\Z_v$. By an abuse of notation, we let $v$ also denote a distinguished generator of $\Z_{v}$.
\item As a graded $\Z$-module, we let $\Hbb(R,\chi):=\bigoplus\limits_{v\in \mathcal{V}}\Z_{v}$, where $\Z_{v}$ has grading equal to $2\chi(v)$.
\item For each generator $v$, we let $Uv = v_{1}+\cdots + v_{n}$ where
\begin{align*}
\{v_{1},\ldots, v_{n}\} = \{w\in \mathcal{V}\ |\ \chi(w) = \chi(v)-1\text{ and } w\text{ is connected to }v\text{ by an edge}\}.
\end{align*}
In particular, if the above set is empty, then $Uv =0$.
\item We then extend the $U$-action by $\Z$-linearity. Note $U$ decreases grading by $2$.
\end{itemize}
\subsection{Graded roots for negative definite plumbings}
\label{sec:graded roots for neg def plumbings}
We now recall from \cite[Section 4]{Nem_On_the} the graded root $\groot{k}$ associated to a negative definite plumbing $\Gamma$ and a $\spinc$ representative $k\in m + 2\Z^s$. We then show how to obtain a graded root $\groot{[k]}$ corresponding to $(\Gamma, [k])$, independent of the choice of $\spinc$ representative $k$.
Define a function $\chi^{}_k : \Z^s \to \Z$ by
\begin{equation}
\label{eq:chi}
\chi^{}_k(x) = - (k\cdot x + \langle x,x\rangle)/2.
\end{equation}
Note that $\chi^{}_k(x)\in \Z$ since $k$ is characteristic.
Consider the standard cubical complex structure on $\R^s$, with $0$-dimensional cells located at the lattice points $\Z^s \subset \R^s$. We can extend $\chi^{}_k$ to a function on closed cells $\square$ (of any dimension), by defining
\begin{align*}
\chi^{}_{k}(\square) = \max\{\chi^{}_{k}(v) \mid v\text{ is a 0-cell of }\square\}
\end{align*}
Let $S_j \subset \R^s$ denote the subcomplex consisting of cells $\square$ such that $\chi^{}_k(\square) \leq j$. We call $S_j$ a \emph{sublevel set}. Note that each $S_{j}$ is compact since the intersection form $\langle-,-\rangle$ is assumed to be negative definite. More precisely, if one considers $\chi^{}_{k}$ as a function from $\R^s\to \R$, then it is bounded below and its level sets are $(s-1)$-dimensional ellipsoids.
Write each sublevel set as a disjoint union over its connected components,
\begin{align*}
S_j = C_{j,1} \sqcup \cdots \sqcup C_{j,n_j}
\end{align*}
The vertices of $R_k$ consist of connected components among all the $S_j$,
\[
\mathcal{V} := \{C_{j,\ell} \mid j\in \Z, 1\leq \ell \leq n_j\},
\]
and the grading is given by $\chi^{}_k(C_{j,\ell}) = j$, where as in \cite{Nem_On_the} we use $\chi_{k}$ to denote both a function on closed cells of our cellular decomposition as well as a grading on the vertices of $R_k$. Edges of $R_k$ correspond to inclusions of connected components: there is an edge connecting $C_{j,\ell}$ and $C_{j+1, \ell'}$ if $C_{j,\ell} \subseteq C_{j+1,\ell'}$. By \cite[Proposition 4.3]{Nem_On_the}, $\groot{k}$ is a graded root.
Let us now recall from \cite{Nem_On_the} how $\groot{k}$ depends on the choice of a $\spinc$ representative. Let $k\in m+2\Z^s$ and let $k'=k + 2My$ be another representative for $[k]\in \spinc(Y)$. One readily checks that:
\begin{equation}
\label{eq:chi for different reps}
\chi^{}_{k'}(x) = \chi^{}_{k}(x+y) - \chi^{}_{k}(y)
\end{equation}
for all $x\in \Z^s$. As stated in \cite[Proposition 4.4]{Nem_On_the}, there is an isomorphism of graded roots:
\begin{equation}\label{eq:GR for different reps}
\groot{k'} \cong \groot{k}[-\chi^{}_k(y)],
\end{equation}
given by applying the translation $x\mapsto x+y$, to each connected component $C$ in each sublevel set of $\chi^{}_{k'}$. Since the collection of graded roots $\{\groot{k'}\}_{k'\in [k]}$ are all isomorphic up to an overall grading shift, we normalize gradings in the following way to obtain a graded root independent of the choice of $\spinc$ representative. This is the same normalization as \cite[Section 4.5.1]{Nem_On_the}.
\begin{defn}
\label{def:normalizating graded root}
Let $[k]\in \spinc$ and define $\groot{[k]}$ by taking any representative $k'\in [k]$ and shifting the $\chi^{}_{k'}$-grading on $(R_{k'}, \chi^{}_{k'})$ so that its minimal grading is zero.
\end{defn}
\begin{gconvs}\label{grading conv}
When drawing the graded root $\groot{[k]}$ associated to a plumbing $\Gamma$ and $\spinc$ structure $[k]$, the numbers we list in the vertical column to the right are the gradings of the corresponding generators of $\Hbb^0(\Gamma, [k])\left[-\max\limits_{k'\in [k]}\frac{(k')^2+s}{4}\right]$. The reason we do this is so that when the isomorphism (\ref{eq:LC-HF iso}) holds, the gradings one sees are the $HF^+$ gradings. See for example Figure \ref{Weighted_graded_root}, where $d(-\Sigma(2,7,15))=0$.
\end{gconvs}
\section{Admissible functions and weighted graded roots}
\label{sec:weighted graded root}
This section illustrates the main construction of the paper in a preliminary context.
Let $\Gamma$ be a negative definite plumbing and $k\in m + 2\Z^s$ a $\spinc$ representative. Given a function
\[
F_{\Gamma,k} : \Z^s \to \mathcal{R}
\]
valued in some ring $\mathcal{R}$, each vertex $v$ in the graded root $\groot{k}$ can be given a {\em weight} by taking the sum of $F_{\Gamma,k}$ over lattice points in the connected component representing $v$. Precisely, for a connected component $C$ in some sublevel set of $\chi^{}_k$, let $L(C) = C \cap \Z^s$ denote its lattice points, and define its weight to be
\begin{equation}
\label{eq:sum over connected component}
F_{\Gamma,k}(C) := \sum_{x\in L(C)} F_{\Gamma,k}(x).
\end{equation}
In this section we explain a way to obtain $F_{\Gamma,k}$ from an \emph{admissible} family of functions $F = \{F_n\}_{n\geq 0}$. Theorem \ref{thm:sum of weights is invt} shows that the graded root $\groot{k}$ with these weights is an invariant of $(Y(\Gamma), [k])$.
This result follows from the more general Theorem \ref{thm:invariance}.
\begin{defn} \label{def:admissible} Fix a commutative ring $\mathcal{R}$.
A family of functions $F = \{F_n : \Z \to \mathcal{R}\}_{n\geq 0}$ is \emph{admissible} if
\begin{enumerate}[label= (A\arabic*)]
\item\label{item:A1} $F_2(0) = 1$ and $F_2(r) = 0$ for all $r\neq 0$.
\item \label{item:A2} For all $n\geq 1$ and $r\in \Z$,
\[ F_n(r+1) - F_n(r-1) = F_{n-1}(r).\]
\end{enumerate}
\end{defn}
For an admissible family $F = \{F_n\}_{n\geq 0}$, define $F_{\Gamma,k} :\Z^s \to \mathcal{R}$ by
\begin{equation}
\label{eq:F_Gamma,k}
F_{\Gamma,k}(x) = \prod_{v\in \mathcal{V}(\Gamma)} F_{\delta_v}\left( (2Mx + k - Mu)_v \right),
\end{equation}
where $u=(1,1,\ldots, 1)\in \Z^s$ and $(-)_v$ denotes the component corresponding to $v$.
\begin{rem} \label{Gamma remark}
We stress that $F_n$ and $F_{\Gamma,k}$ are only set-theoretic functions rather than homomorphisms. The definition \eqref{eq:F_Gamma,k} of $F_{\Gamma,k}(x)$ is motivated by the $\widehat Z$-invariant, see Section \ref{sec:Zhat_a} and in particular Proposition \ref{prop:Zhat_k in terms of F}.
\end{rem}
Note that if $k'=k+2My$ is another representative for $[k]\in \spinc(Y(\Gamma))$, then
\begin{equation}
\label{eq:F different representatives}
F_{\Gamma,k'}(x) = F_{\Gamma,k}(x+y),
\end{equation}
so the weights in equation \eqref{eq:sum over connected component} are compatible with the isomorphism $\groot{k'} \cong \groot{k}[-\chi^{}_k(y)]$ from Section \ref{sec:graded roots for neg def plumbings}. Denote by $(R_{[k]}, \chi^{}_{[k]}, F_{[k]})$ the graded root $\groot{[k]}$ equipped with these weights.
\begin{thm}
\label{thm:sum of weights is invt}
For any admissible family of functions $F$, the weighted graded root $(R_{[k]}, \chi^{}_{[k]}, F_{[k]})$ is an invariant of the $3$-manifold $Y(\Gamma)$ endowed with the $\spinc$ structure $[k]$.
\end{thm}
\noindent The proof of this result follows from Theorem \ref{thm:invariance} upon specializing $q=t=1$.
We end this section with some remarks about admissible families of functions. Explicit $\Z$- and $\Z[\frac{1}{2}]$-valued examples motivated by the $\widehat{Z}$ invariant are given in Definitions \ref{def:Fhat} and \ref{def:Fhatpm}. Note that not only $F_2$, but also $F_0$ and $F_1$ are uniquely determined by conditions \ref{item:A1} and \ref{item:A2}:
\begin{equation}
\label{eq:F_1 and F_0}
F_1(r) =
\begin{cases}
1 & \text{ if } r=-1, \\
-1 & \text{ if } r=1, \\
0 & \text{ otherwise.}
\end{cases}
\hskip3em
F_0(r) =
\begin{cases}
1 & \text{ if } r=\pm 2, \\
-2 & \text{ if } r=0, \\
0 & \text{ otherwise.}
\end{cases}
\end{equation}
We can characterize admissible families in the following way. Let $\operatorname{Adm}(\mathcal{R})$ denote the set of all admissible $\mathcal{R}$-valued families of functions, and let $(\mathcal{R} \times \mathcal{R})^{\mathbb{N}}$ be the set of all sequences with entries in $\mathcal{R} \times \mathcal{R}$.
\begin{prop}\label{prop: adm identification}
There is a bijection $\operatorname{Adm}(\mathcal{R}) \cong (\mathcal{R} \times \mathcal{R})^{\mathbb{N}}$.
\end{prop}
\begin{proof}
We show that $\Psi: \operatorname{Adm}(\mathcal{R})\to (\mathcal{R} \times \mathcal{R})^{\mathbb{N}}$ defined by $\Psi(F) = (F_{n+2}(0), F_{n+2}(1))_{n\geq 1}$ is a bijection. Suppose $F = \{F_{n}\}_{n\geq 0}$ is an admissible family. Fix $n\geq 1$. By applying the recursive relation \ref{item:A2} inductively, we see that $F_n(r)$ is determined by $F_{n-1}, F_{n}(0)$ if $r$ is even and $F_{n-1},F_{n}(1)$ if $r$ is odd, so $\Psi$ is injective. Likewise, given $(a_{n}, b_{n})_{n\geq 1}\in (\mathcal{R} \times \mathcal{R})^{\mathbb{N}}$, we set $(F_{n+2}(0),F_{n+2}(1))=(a_n, b_n)$ and use \ref{item:A2} to inductively construct an admissible family $F$ such that $\Psi(F) = (a_{n}, b_{n})_{n\geq 1}$.
\end{proof}
\section{The invariant} \label{sec: The invariant}
This section introduces the main construction of this paper, a refinement of the weights from equation \eqref{eq:sum over connected component} in the form of a collection of two-variable Laurent polynomials.
Section \ref{sec: invariance} shows that the resulting weighted graded root is a $3$-manifold invariant.
\subsection{Refined weights}
We start by establishing the following notation.
\begin{notn} \label{notn}
For $k,x \in \Z^s$, define
\begin{align*}
\Delta_k = -\frac{(k-Mu)^2 + 3s + \sum_v m_v}{4},\hspace{2em}\varepsilon_k(x) = \Delta_k + 2\chi^{}_k(x) + \langle x,u\rangle
\end{align*}
where the notation $(-)^2$ is the same as in Remark \ref{rem: k^2 meaning}. Note, the term $\Delta_k$ is an overall normalization used to eliminate dependence on the choice of $\spinc$ representative and is similar in form to the $d$-invariant from Heegaard Floer homology (see Remark \ref{rem: d inv}). Also, recall that $Mu = m+\delta$, so $\langle x,u\rangle = x\cdot (m+ \delta)$.
\end{notn}
\begin{defn}
\label{def:weighted graded root k}
Let $k\in m + 2\Z^s$ be a $\spinc$ representative and let $F=\{F_n\}_{n\geq 0}$ be an admissible family valued in a commutative ring $\mathcal{R}$. To each vertex of the graded root $\groot{k}$ we assign a weight valued in $q^{\Delta_k}\cdot\mathcal{R}[q^{\pm 1}, t^{\pm 1}]$ as follows. For a vertex represented by a connected component $C$ in some sublevel set, let $L(C) = C \cap \Z^s$ denote its lattice points. Set
\begin{equation}
\label{eq:laurent poly enhancement}
P_{F,k}(C) = \sum\limits_{x\in L(C)}F_{\Gamma, k}(x)q^{\varepsilon_k(x)}t^{\langle x,u\rangle},
\end{equation}
and let $\wgrootF{k}$ denote the graded root $\groot{k}$ with these weights. We will often omit the reference to $F$ by writing $P_k$ instead of $P_{F,k}$. Note that specializing $q=t=1$ recovers the weights in equation \eqref{eq:sum over connected component}.
\end{defn}
The above weights can be interpreted geometrically as follows. For $n\in \Z$, the coefficient of $t^n$ in $P_{k}(C)$ is given by summing $F_{\Gamma,k}(x)q^{\Delta_k + 2\chi^{}_k(x) + n}$ over all $x\in \Z^s$ which lie on the intersection of $C$ with the hyperplane $\{y\in \R^s \mid \langle y,u\rangle = n \}$.
Let us verify that the weights $P_{k}(C)$ are compatible with the isomorphisms (\ref{eq:GR for different reps}) relating graded roots for different representatives of $[k]$.
\begin{lem}
\label{lem:equations for different spinc reps}
Let $k, k'=k+2My \in m+2\Z^s$ be two representatives for $[k]\in \spinc (Y)$. Then $\varepsilon_{k'}(x) = \varepsilon_k(x+y)$ for all $x\in \Z^s$.
\end{lem}
\begin{proof}
First note that
\[
(k'-Mu)^2 = (k-Mu)^2 +4 y\cdot k - 4 \langle y,u\rangle + 4 \langle y,y\rangle,
\]
which implies
\begin{align*}
\Delta_{k'} &= \Delta_k - y\cdot k +\langle y,u\rangle - \langle y,y\rangle \\
&= \Delta_k + 2\chi^{}_{k}(y) + \langle y,u\rangle.
\end{align*}
The desired equality now follows from the above equality and equation \eqref{eq:chi for different reps}.
\end{proof}
For $n\in \Z$, let $\wgroot{k}\{n\}$ denote the weighted graded root with each Laurent polynomial weight multiplied by $t^n$.
\begin{prop}
\label{prop:different reps for weighted graded root}
Let $k\in m + 2\Z^s$ and let $k' = k + 2My$ for some $y\in \Z^s$. Then
\[
\wgroot{k'} \cong \wgroot{k}[-\chi^{}_{k}(y)]\{-\langle y,u\rangle\}.
\]
\end{prop}
\begin{proof}
Recall that the isomorphism $\groot{k} \cong \groot{k'}[-\chi^{}_{k}(y)]$ of graded roots in equation \eqref{eq:GR for different reps} is induced by the translation $T(x) = x+y$. A lattice point $x\in L(C)$ contributes the summand
\[
F_{\Gamma, k'}(x)q^{\varepsilon_{k'}(x)}t^{\langle x,u\rangle }
\]
to $P_{F,k'}(C)$, while $T(x)$ contributes
\[
t^{\langle y,u\rangle} \cdot F_{\Gamma, k}(x+y)q^{\varepsilon_k(x+y)}t^{\langle x,u\rangle}
\]
to $P_{F,k}(T(C))$. Lemma \ref{lem:equations for different spinc reps} and equation \eqref{eq:F different representatives} imply that
\[
F_{\Gamma, k'}(x)q^{\varepsilon_{k'}(x)}t^{\langle x,u\rangle } = F_{\Gamma, k}(x+y)q^{\varepsilon_k(x+y)}t^{\langle x,u\rangle},
\]
which completes the proof.
\end{proof}
In light of the above proposition, we consider weighted graded roots up to simultaneous multiplication of all Laurent polynomial weights by some overall power of $t$; that is, $\wgroot{k}$ is equivalent to $\wgroot{k}\{n\}$ for all $n\in \Z$.
\begin{defn}
\label{def:weighted graded root [k]}
Set $\wgroot{[k]}$ to be $\groot{[k]}$, as in Definition \ref{def:normalizating graded root}, equipped with the weights $P_k$ for some $k\in [k]$.
\end{defn}
\noindent Lemma \ref{lem:equations for different spinc reps} and Proposition \ref{prop:different reps for weighted graded root} guarantee that, up to the above equivalence, $\wgroot{[k]}$ does not depend on $k\in [k]$.
\begin{rem}
\label{rmk:normalizing t}
The $t$-ambiguity could be removed by fixing the following normalization, similar to the normalization in Definition \ref{def:normalizating graded root}. First, if for some $k\in [k]$ all $P_k$ weights are zero then, by Proposition \ref{prop:different reps for weighted graded root}, for any other $k'\in[k]$ all $P_{k'}$ weights are zero. Hence, there is no ambiguity. Otherwise, let $n'\in \Z$ be the minimal $t$-power among all nonzero $P_k$ weights of the vertices in $\chi^{}_k$-grading $n$, and set
\[
\wgroot{[k]} = (R_{[k]}, \chi^{}_{[k]}, P_{k})\{-n'\}.
\]
Proposition \ref{prop:different reps for weighted graded root} implies $\wgroot{[k]}$, as defined in the previous equation, is independent of $k\in [k]$.
\end{rem}
\subsection{Invariance} \label{sec: invariance}
In this section we prove invariance of $\wgroot{[k]}$ under the two Neumann moves shown in Figure \ref{fig:Neumann moves}. This establishes Theorem \ref{main thm}, which is restated as Theorem \ref{thm:invariance} below using a more detailed notation. In what follows, $\Gamma$ is a negative definite plumbing tree with $s$ vertices, and $\Gamma'$ is a plumbing tree with $s'=s+1$ vertices obtained from $\Gamma$ by one of the type (a) or (b) moves. We will use the conventions established in Notation \ref{notation type a and b}, as well as the following additional notation for the two moves.
{\bf Type (a):} The intersection form for $\Gamma'$ is given by $M' = \til{M} + A$ where
\begin{align}
\label{eq:M' for type a}
\til{M} =
\begin{pmatrix}
0 & 0 \\
0 & M
\end{pmatrix}
\hskip2em
A =
{\scriptsize
\begin{pmatrix}
-1 & \phantom{-}1 & \phantom{-}1 & 0 & \cdots & 0 \\
\phantom{-}1 & -1 & -1 & 0 & \cdots & 0 \\
\phantom{-}1 & -1 & -1 & 0 & \cdots & 0 \\
\phantom{-}0 & \phantom{-}0 & \phantom{-}0 & 0 & \cdots & 0 \\
\phantom{-}\vdots & & & & & \vdots \\
\phantom{-}0 & \phantom{-}\cdots & \phantom{-}\cdots & \cdots & \cdots & 0
\end{pmatrix}
}
\end{align}
As in \cite[Proposition 3.4.2]{Nem_Lattice_cohomology}, define the projection $\pi_{*}:\Z^{s+1}\to \Z^s$ by
\begin{align}
\label{eq:pi_* map}
\pi_{*}(x_{0}, x_{1}, \ldots, x_{s}) = (x_{1}, \ldots, x_{s})
\end{align}
and the inclusion $\pi^*:\Z^s\to\Z^{s+1}$ by
\begin{align}
\label{eq:pi^* map type a}
\pi^*(x_{1}, x_2, \ldots, x_{s}) = (x_{1}+x_{2}, x_{1}, x_{2}, \ldots, x_{s}).
\end{align}
{\bf Type (b):} The intersection form for $\Gamma'$ is given by $M' = \til{M} + A$ where
\begin{align} \label{M prime type b}
\til{M}=\begin{pmatrix}
0 & 0 \\
0 & M
\end{pmatrix}
\hskip2em
A =
{\scriptsize
\begin{pmatrix}
-1 & \phantom{-}1 & 0 & \cdots & 0 \\
\phantom{-}1 & -1 & 0 & \cdots & 0 \\
\phantom{-}0 & \phantom{-}0 & 0 & \cdots & 0 \\
\phantom{-}\vdots & & & & \vdots \\
\phantom{-}0 & \phantom{-}\cdots & \cdots & \cdots & 0
\end{pmatrix}
}
\end{align}
As in the type (a) case, let $\pi_{*}:\Z^{s+1}\to\Z^s$ denote the projection $\pi_*(x_0,x_{1}, \ldots, x_{s}) = (x_{1}, x_{2}, \ldots, x_{s})$, and define two inclusions $\pi^*, \rho^* : \Z^s \to \Z^{s+1}$ by
\begin{align}
\pi^*(x_{1}, \ldots, x_{s}) &= (x_{1}, x_{1}, \ldots, x_{s}), \label{eq:pi^* map type b} \\
\rho^* (x_{1}, \ldots, x_{s}) &= (x_{1}-1, x_{1}, \ldots, x_{s}). \label{eq:p map type b}
\end{align}
With the notation in place, we now record several results regarding various contributions to the Laurent polynomial weights.
\begin{lem} \label{lem: exponent of normalization term}
Let $\Gamma$, $\Gamma'$ be negative definite plumbings related by a Neumann move as above. Let $k\in m+2\Z^s$ be a $\spinc$ representative, and let $k'\in m'+2\Z^{s+1} $ denote the associated representative, as in equations \eqref{eq:k' for type a} and \eqref{eq:k' for type b}. Then $\Delta_k = \Delta_{k'}$.
\end{lem}
\begin{proof}
We first address the type (a) move. Let
$u'=(1,u)$. Observe that $3s + \sum_v m_v = 3s' + \sum_v m_v'$, so it remains to verify that
\[
(k'-M'u')^2 = (k-Mu)^2.
\]
Expressions for $M'$ and $k'$ are given in equations \eqref{eq:M' for type a} and \eqref{eq:k' for type a}. Note that
\[
k' - M'u' = (0,k-Mu).
\]
Let $y=(y_1, y_2, \ldots, y_{s})\in \Q^{s}$ be such that $k-Mu = My$. Then $M'(y_1 + y_2, y) = (0,k-Mu)$, and it follows that $(k'-M'u')^2 = (k-Mu)^2$ which completes the proof of the type (a) case.
For the type (b) move, $M'$ and $k'$ are given in equations \eqref{M prime type b}, \eqref{eq:k' for type b}. Note
$$k'-M'u'= (-1,1, 0,\ldots,0)+(0, k-Mu).
$$
Denote these two summands by $w:=(-1,1, 0,\ldots,0)$ and $\widetilde k:= (0, k-Mu)$.
We claim that
\begin{equation} \label{eq: claim}
(\widetilde k)^t(M')^{-1}\widetilde k = (k-Mu)^t M^{-1}(k-Mu).
\end{equation}
To prove the claim, let $y=M^{-1}(k-Mu)$ or equivalently $k-Mu=My$. Then one checks that $\widetilde k=M'(y_1, y)$, where $y_1$ is the first coordinate of $y$. Thus the left-hand side of (\ref{eq: claim}) equals $(0, k-Mu)^t(y_1, y)$ which equals the right-hand side of (\ref{eq: claim}), verifying the claim.
Expanding linearly, consider
$$(k'-M'u')^t(M')^{-1}(k'-M'u') - (k-Mu)^t M^{-1}(k-Mu)=
$$
\begin{equation} \label{eq: three terms}
w^t(M')^{-1}w + 2(\widetilde k)^t(M')^{-1} w.
\end{equation}
It follows from equations (\ref{M prime type b}) that $(M')^{-1} w = (1,0,\ldots,0),$ thus the first term in (\ref{eq: three terms}) equals $-1$ and the second term is zero. Under the (b) move, we have $3s+\sum_{v}m_{v} = 3s' + \sum_v m_v' -1$, precisely offsetting the change in $(k-Mu)^2$ computed above.
\end{proof}
For the type (a) move in the following lemma, recall the function $\pi^*$ from equation \eqref{eq:pi^* map type a}.
For the type (b) move, recall the functions $\pi^*$, and $\rho^*$ from equations \eqref{eq:pi^* map type b} and \eqref{eq:p map type b}.
\begin{lem}
\label{lem:pi* maps and u}
Let $\Gamma, \Gamma', k$, and $k'$ be as in the statement of Lemma \ref{lem: exponent of normalization term}.
\begin{enumerate}
\item In the type (a) case, for any $x\in \Z^{s}$,
\[
\langle x , u \rangle =\langle \pi^*(x) , u'\rangle.
\]
\item In the type (b) case, for any $x\in \Z^s$,
\[
\langle x ,u\rangle =\langle \pi^*(x) , u'\rangle = \langle \rho^*(x),u'\rangle.
\]
\end{enumerate}
\end{lem}
\begin{proof}
Recall that $\langle x,u\rangle = x\cdot (m+\delta)$. For item (1), $m' + \delta' = (1,-1,-1, 0, \ldots, 0) + (0, m+\delta)$. Then
\begin{align*}
\pi^*(x)\cdot (m' +\delta') &= (x_{1}+x_{2}, x_{1}, x_{2}, \ldots, x_{s})\cdot [(1,-1,-1,0, \ldots, 0) + (0, m+\delta)]\\
&= x\cdot (m+\delta).
\end{align*}
For item (2), $m' + \delta' = (0,m + \delta)$, and the desired equality follows.
\end{proof}
\begin{lem}
\label{lem:F functions and pi maps}
Let $\Gamma, \Gamma', k$, and $k'$ be as in the statement of Lemma \ref{lem: exponent of normalization term}.
\begin{enumerate}
\item For the type (a) move, $F_{\Gamma, k}(x) = F_{\Gamma', k'}(\pi^*(x))
$ for all $x\in \Z^s$,
\item For the type (b) move, $F_{\Gamma, k}(x) = F_{\Gamma',k'}(\pi^*(x)) + F_{\Gamma',k'}(\rho^*(x))
$ for all $x\in \Z^s$.
\end{enumerate}
\end{lem}
\begin{proof}
To show item (1), first note that
\[
2M'\pi^*(x) + k' - M'u' =(0, 2Mx+ k -Mu),
\]
so, using property \ref{item:A1}, $F_{\Gamma', k'}(\pi^*(x)) = F_{2}(0)\prod\limits_{i=1}^{s} F_{\delta_{i}}((2Mx + k -Mu)_{v_{i}})=F_{\Gamma, k}(x)$.
We now verify item (2). Observe that
\begin{align*}
2M'\pi^*(x) + k' - M'u' &= (0,2Mx+k-Mu) + (-1,1,0,\ldots, 0), \\
2M'\rho^*(x) + k' - M'u' &= (0, 2Mx +k-Mu)+ (1,-1, 0, \ldots, 0).
\end{align*}
Introduce the notation
\begin{align*}
r = (2Mx + k - Mu)_{v_1},\hspace{2em} \overline{r} = \prod_{i=2}^s F_{\delta_i}(2Mx + k - Mu)_{v_i},
\end{align*}
and recall from equation \eqref{eq:F_1 and F_0} that $F_1(\pm 1) = \mp 1$. Then we have
\begin{align*}
F_{\Gamma',k'}(\pi^*(x)) &= F_{\delta_1 +1}(r + 1) \cdot \overline{r}, \\
F_{\Gamma',k'}(\rho^*(x)) &= -F_{\delta_1 +1 }(r-1) \cdot \overline{r}, \\
F_{\Gamma,k}(x) &= F_{\delta_1}(r) \cdot \overline{r},
\end{align*}
and the desired equality follows from property \ref{item:A2}.
\end{proof}
We are in a position to prove our main result:
\begin{thm}
\label{thm:invariance}
For any admissible family of functions $F$, the weighted graded root $\wgroot{[k]}$ is an invariant of the $3$-manifold $Y(\Gamma)$ equipped with the $\spinc$ structure $[k]$.
\end{thm}
\begin{proof}
We will demonstrate an isomorphism $
\wgroot{k} \cong \wgroot{k'}$ of weighted graded roots when $\Gamma, \Gamma', k$, and $k'$ are as in the statement of Lemma \ref{lem: exponent of normalization term}. Note, here the symbol $\cong$ means isomorphism with no $t$-ambiguity; the $t$-ambiguity only becomes relevant when one changes $\spinc$ representative.
For each of the two moves we first give an explicit isomorphism of graded roots $\groot{k}\cong \groot{k'}$, following the proofs of \cite[Proposition 4.6]{Nem_On_the} and \cite[Proposition 3.4.2]{Nem_Lattice_cohomology}. We then show that this isomorphism respects our Laurent polynomial weights.
We begin with the type (a) move. Recall the functions $\pi_*$ and $\pi^*$ from \eqref{eq:pi_* map} and \eqref{eq:pi^* map type a}, and that $M' = \widetilde{M} + A$, as in equation \eqref{eq:M' for type a}.
For $x'= (x_0, x_1, x_2, \ldots, x_s) \in \Z^{s+1}$, we have
\[
(x')^t A x' = - (x_0 - x_1 - x_2)^2,
\]
so $(x')^t M' x' = \pi_*(x')^t M \pi_*(x') - (x_0 - x_1 - x_2)^2$. It is then straightforward to verify that
\[
\chi^{}_{k'}(x') = \chi^{}_{k}(\pi_{*}(x')) + \frac{1}{2}[(x_{0}-x_{1}-x_{2})(x_{0}-x_{1}-x_{2}-1)].
\]
In particular, substituting $x'=\pi^*(x)$ for $x\in \Z^s$, this implies
\begin{equation}
\label{eq:pi^* commutes with chi type a}
\chi^{}_{k'} \circ \pi^* = \chi^{}_{k},
\end{equation}
so $\pi^*$ induces an inclusion $ \chi_{k}^{-1}((-\infty, j])\hookrightarrow \chi_{k'}^{-1}((-\infty, j])$ of sublevel sets.
As in the proof of \cite[Proposition 3.4.2]{Nem_Lattice_cohomology}, $\pi^*$ also induces a bijection, denoted $\tilde{\pi}^*$, on connected components in these sublevel sets. The isomorphism $\groot{k} \cong \groot{k'}$ of graded roots is given by $\tilde{\pi}^*$, sending a connected component $C$ to the connected component $\til{\pi}^*(C)$ that contains $\pi^*(C)$.
Fix a connected component $C$ in some sublevel set of $\chi^{}_k$. We will now show that
\[
P_k(C) = P_{k'}(\tilde{\pi}^*(C)).
\]
The term on the right-hand side above is a sum over contributions from lattice points in the component $\tilde{\pi}^*(C)$, which contains all the lattice points in $\pi^*(C)$, but is in general strictly bigger. As we shall now see, only lattice points in $\pi^*(C)$ contribute. We have
\[
(2M'x'+ k'-M'u')_{v_0'} = -2(x_0 - x_1 - x_2).
\]
Since $\delta_{v_0'}= 2$, property \ref{item:A1} implies that $F_{\Gamma',k'}(x') = 0$ unless $x_0 = x_1 + x_2$, so
\[
P_{k'}(\tilde{\pi}^*(C)) = \sum_{x'\in \pi^*(L(C))} F_{\Gamma', k'}(x') q^{\varepsilon_{k'}(x')} t^{\langle x',u'\rangle}.
\]
Therefore, it suffices to show
\[
F_{\Gamma, k}(x) q^{\varepsilon_k(x)} t^{\langle x,u\rangle}
=
F_{\Gamma', k'}(\pi^*(x)) q^{\varepsilon_{k'}(\pi^*(x))} t^{\langle \pi^*(x),u'\rangle}.
\]
for all $x\in C$. Equation \eqref{eq:pi^* commutes with chi type a}, Lemma \ref{lem: exponent of normalization term}, and item (1) of Lemma \ref{lem:pi* maps and u} guarantee that the powers of $q$ and $t$ are equal, and $F_{\Gamma, k}(x) = F_{\Gamma', k'}(\pi^*(x))$ by Lemma \ref{lem:F functions and pi maps} (1). This concludes the proof of the type (a) move.
We now address the type (b) move. Recall the functions $\pi_*, \pi^*$, and $\rho^*$ from equations \eqref{eq:pi_* map}, \eqref{eq:pi^* map type b}, \eqref{eq:p map type b}, and that $M'= \widetilde{M} + A$ as in equation \eqref{M prime type b}. For $x'\in \Z^{s+1}$, we have
\[
(x')^t A x' = - (x_0 - x_1)^2,
\]
so $(x')^t M' x' = \pi_*(x')^t M \pi_*(x') - (x_0 - x_1)^2$. It is then easy to see that
\[
\chi^{}_{k'}(x') = \chi^{}_{k}(\pi_*(x')) +\frac{1}{2}(x_{0}-x_{1})(x_{0}-x_{1}+ 1),
\]
which implies
\begin{equation}
\label{eq:pi^* commutes with chi type b}
\chi^{}_{k'} \circ \pi^* = \chi^{}_k = \chi^{}_{k'} \circ \rho^*.
\end{equation}
Thus both $\pi^*$ and $\rho^*$ induce inclusions $ \chi_{k}^{-1}((-\infty, j])\hookrightarrow \chi_{k'}^{-1}((-\infty, j])$ of sublevel sets. As in the type (a) case above, $\pi^*$ also induces a bijection $\tilde{\pi}^*$ between connected components of each sublevel set, and the isomorphism of graded roots $\groot{k} \cong \groot{k'}$ is given by $\tilde{\pi}^*$.
To complete the proof we check that
\[
P_k(C) = P_{k'}(\tilde{\pi}^*(C))
\]
for every connected component $C$ of each sublevel set of $\chi^{}_k$. As in the type (a) case, we will now see that only a particular subset of lattice points in $\tilde{\pi}^*(C)$ contribute to $P_{k'}(\tilde{\pi}^*(C))$. To begin, note
\[
(2M'x' + k' - M'u')_{v_0'} = -2(x_0 - x_1) - 1.
\]
Since $\delta_{v_0'} =1$, the formula for $F_1$ from equation \eqref{eq:F_1 and F_0} implies that $F_{\Gamma',k'}(x') = 0$ unless $-2(x_0 - x_1) - 1 = \pm 1$, or, equivalently, unless $x' = \pi^*(x)$ or $x' = \rho^*(x)$ for some $x\in \Z^s$. Observe that $\pi^*(x) - \rho^* (x) = (1,0, \ldots, 0)$, so $\pi^*(x)$ and $\rho^*(x)$ are in the same component of $\chi_{k'}^{-1}(j)$. It follows that
\[
P_{k'}(\tilde{\pi}^*(C)) =
\sum_{x'\in \pi^*(L(C))} F_{\Gamma', k'}(x') q^{\varepsilon_{k'}(x')} t^{\langle x',u'\rangle}
+
\sum_{x'\in \rho^*(L(C))} F_{\Gamma', k'}(x') q^{\varepsilon_{k'}(x')} t^{\langle x',u'\rangle}.
\]
To complete the proof, we have
\[
F_{\Gamma, k}(x) q^{\varepsilon_k(x)} t^{\langle x,u\rangle }
=
F_{\Gamma', k'}(\pi^*(x)) q^{\varepsilon_{k'}(\pi^*(x))} t^{\langle \pi^*(x),u'\rangle}
+ F_{\Gamma', k'}(\rho^*(x)) q^{\varepsilon_{k'}(\rho^*(x))} t^{\langle \rho^*(x),u'\rangle}
\]
by combining Lemma \ref{lem: exponent of normalization term}, equation \eqref{eq:pi^* commutes with chi type b}, item (2) of Lemma \ref{lem:pi* maps and u}, and item (2) of Lemma \ref{lem:F functions and pi maps}.
\end{proof}
\section{The two-variable series}
\label{sec:two variable series}
In this section we extract a two-variable series from $\wgroot{[k]}$ by taking a limit (in an appropriate sense) of the weights $P_k(C)$. Theorem \ref{thm:convergence} shows that this limiting procedure yields a well-defined invariant of $(Y(\Gamma), [k])$. Throughout this section some $\mathcal{R}$-valued admissible family $F$ will be fixed, and references to it will be omitted for brevity of notation.
We first establish some preliminary notions. For a commutative ring $\mathcal{R}$, denote by $\mathcal{R}[q^{-1},q]]$ the ring of Laurent series in $q$ and by $\widetilde{\mathcal{R}}$ the set of Laurent series in $q$ whose coefficients are Laurent polynomials in $t$,
\[
\widetilde{\mathcal{R}} = \left(\mathcal{R}[t^{\pm 1}] \right)[q^{-1},q]].
\]
Given $\Delta\in \Q$, $f\in q^\Delta \cdot \widetilde{\mathcal{R}} $, and $i,j \in \Z$, let $[f]_{i,j} \in \mathcal{R}$
be the coefficient of $q^{\Delta+i} t^{j}$ in $f$.
\begin{defn}
\label{def:stabilization}
We say a sequence $f_1, f_2, \ldots \in q^\Delta \cdot \mathcal{R} [q^{\pm 1}, t^{\pm 1}]$ \emph{stabilizes} if for all $i,j\in \Z$, the sequence of coefficients $([f_1]_{i,j}, ([f_2]_{i,j}, \ldots )$ is eventually constant. For such a sequence, the \emph{limit} $f$ is the bi-infinite series in $q,t$ defined
by setting $[f]_{i,j}$ to be the limit of $[f_n]_{i,j}$ as $n\to \infty$.
\end{defn}
As stated in the definition, the limit of a stabilizing sequence in general is a bi-infinite series in $q,t$. In Theorem \ref{thm:convergence} below, the limit is claimed to be an element of $q^\Delta \cdot \widetilde{\mathcal{R}}$. In addition to proving that the sequence $(f_n)$ stabilizes, this will be shown by establishing that \begin{enumerate}[label=(\roman*)]
\item
there exists $i_0\in \Z$ such that $[f_n]_{i,j} = 0$ for all $n\geq 0$, $j\in \Z$, and $i\leq i_0$, and
\item for any fixed $i$, the set of $j$ such that $[f_n]_{i,j} \neq 0$ is bounded.
\end{enumerate}
Returning to weighted graded roots, fix a negative definite plumbing tree $\Gamma$ and $\spinc$ representative $k\in m+2\Z^s$. Consider the weighted graded root $\wgroot{k}$, as given in Definition \ref{def:weighted graded root k}. For $n\in \Z$, let
\begin{equation}
\label{eq:P_k^n}
P_k^n := \sum_{C\in \chi^{-1}_{[k]}(n)} P_k(C) \in q^{\Delta_k} \cdot \mathcal{R}[q^{\pm 1}, t^{\pm 1}]
\end{equation}
denote the sum of the Laurent polynomial weights over vertices $C$ of $R_k$ in $\chi^{}_{k}$-grading $n$. Recall that $\chi^{}_k$ is bounded below by some $n_0\in \Z$, and consider the sequence $(P_k^{n_0}, P_k^{n_0 + 1}, P_k^{n_0 + 2}, \ldots )$.
\begin{rem}
Note that $P_k^n$ is the sum of $F_{\Gamma,k}(x)q^{\varepsilon_k(x)}t^{\langle x,u\rangle}$ over all lattice points $x$ in the entire $n$-sublevel set of $\chi^{}_{k}$. Moreover, since there is only one connected component for large enough $n$, one may just as well start the sequence at a sufficiently high $\chi^{}_{k}$-grading, making the sum in \eqref{eq:P_k^n} be given by a single $P_k(C)$.
\end{rem}
\begin{thm}
\label{thm:convergence}
The sequence $\left(P_k^{n_0}, P_k^{n_0 + 1}, P_k^{n_0 + 2}, \ldots\right)$ stabilizes to an element of $q^{\Delta_k} \cdot \widetilde{\mathcal{R}}$. Up to multiplication by a power of $t$, its limit $P_k^\infty$ is an invariant of the $3$-manifold $Y(\Gamma)$ equipped with the $\spinc$ structure $[k]$.
\end{thm}
\begin{proof}
For $n\in \Z$, define
\begin{align*}
\mathcal{S}_n &= \{ x\in \Z^s \mid \chi^{}_k(x) \leq n \}, &
\til{\mathcal{S}}_n &= \{ x\in \Z^s \mid 2\chi^{}_k(x) + \langle x, u\rangle \leq n \}, \\
\partial \til{\mathcal{S}}_n &= \{ x\in \Z^s \mid 2\chi^{}_k(x) + \langle x, u\rangle =n \},
& \mathcal{A}_n &= \{ x\in \Z^s \mid \langle x,u\rangle = n\}.
\end{align*}
By definition,
\[
P_k^n = \sum_{x\in \mathcal{S}_n} F_{\Gamma,k}(x)q^{\varepsilon_k(x)} t^{\langle x,u\rangle}.
\]
It follows from Notation \ref{notn} that for fixed $i,j\in \Z$, the coefficient of $q^{\Delta_k + i} t^{j}$ in $P_k^n$ is equal to
\begin{equation}
\label{eq:coefficient fixed n i and j}
\sum_{x} F_{\Gamma,k}(x),
\end{equation}
where the sum is over $x\in \mathcal{S}_n \cap \partial\til{\mathcal{S}}_i \cap \mathcal{A}_j$. Both $(\mathcal{S}_n)$, $(\til{\mathcal{S}}_n)$ are sequences of nested finite sets whose union is $\Z^s$. Hence for a fixed $i$ there exists $N$ such that $\mathcal{S}_i \subset \mathcal{S}_n$ for all $n\geq N$. Then for $n\geq N$, we have
\[
\mathcal{S}_n \cap \partial\til{\mathcal{S}}_i \cap \mathcal{A}_j = \partial\til{\mathcal{S}}_i \cap \mathcal{A}_j,
\]
so that the sum in equation \eqref{eq:coefficient fixed n i and j} is independent of $n$ for $n$ sufficiently large. See Figure \ref{fig:sublevel sets} for an illustration when $s=2$. This verifies stabilization of the sequence.
As discussed after Definition \ref{def:stabilization}, we will check two conditions (i), (ii) ensuring that the limit is an element of $q^\Delta \cdot \widetilde{\mathcal{R}}$.
The condition (i) follows from the fact that the $q$-powers in the $P_k^{n}$ are given by $\varepsilon_k$, which is bounded below. To check (ii), observe that for a fixed $i$, the exponent of $t$ is given by $\langle x, u\rangle$ which is bounded on the set $\til{\mathcal{S}}_i$.
\begin{figure}
\centering
\includestandalone{sublevel_sets}
\caption{A schematic depiction of stabilization. Left: $\mathcal{S}_n$ is represented by the shaded ellipse, $\partial\til{\mathcal{S}}_i$ is the unshaded ellipse, and the hyperplane $\mathcal{A}_j$ is the straight line (the actual sets are discrete subsets of the illustration). The intersection $\mathcal{S}_n \cap \partial\til{\mathcal{S}}_i \cap \mathcal{A}_j$ is marked by a dot. Right: taking $n>>0$ ensures $\partial\til{\mathcal{S}}_i \subset \mathcal{S}_n$, and the intersection (two dots) is the same for all sufficiently large $n$.}
\label{fig:sublevel sets}
\end{figure}
That $P_k^\infty$ is an invariant of $(Y(\Gamma),[k])$ up to multiplication by $t$ follows from Proposition \ref{prop:different reps for weighted graded root} and Theorem \ref{thm:invariance}.
\end{proof}
For $f,g\in q^{\Delta}\cdot \widetilde{\mathcal{R}}$, we write $f \doteq g$ if $f= t^n g$ for some $n\in \Z$. As a result of the above theorem, we can now make the following definition.
\begin{defn}\label{defn:P infinity}
Set $P_{[k]}^{\infty} = P_{k}^{\infty}$ for any choice of $k\in [k]$, with the understanding that $P_{[k]}^{\infty}$ is well-defined up to multiplication by a power of $t$. More explicitly, it follows from the definition \eqref{eq:laurent poly enhancement} of the Laurent polynomial weights and from Theorem \ref{thm:convergence} that
\begin{align} \label{P [k] infty}
P_{[k]}^{\infty} \doteq \sum\limits_{x\in \Z^s}F_{\Gamma,k}(x)q^{\varepsilon_{k}(x)}t^{\langle x, u\rangle}
\end{align}
for any choice of $k\in [k]$. To specify the underlying admissible family $F$, the notation $P_{F, [k]}^{\infty}$ will be used.
\end{defn}
\begin{rem}
For any $F$, setting $t=1$ in $P_{[k]}^\infty$ gives a well-defined Laurent $q$-series invariant of $(Y(\Gamma,[k])$. Moreover, if the $t$-powers in $\wgroot{[k]}$ are normalized as in Remark \ref{rmk:normalizing t}, then the weights stabilize to a well-defined element of $q^{\Delta_k} \cdot \widetilde{\mathcal{R}}$.
\end{rem}
\section{The \texorpdfstring{$\widehat{Z}_a(q)$}{} power series}
\label{sec:Zhat_a}
This section starts with a review of the GPPV invariant of negative definite plumbed manifolds, motivating the definition of the admissible family of functions $\widehat F$. In fact, three closely related admissible families are discussed in Section \ref{subsec: Three admissible families}, $\widehat F$, $\widehat F^+$, and $\widehat F^-$. Section \ref{sec: The lattice and Zhat} reformulates the $\widehat Z$ invariant using the lattice cohomology convention for $\spinc$ structures.
Theorem \ref{thm: Zhathat Zhat} shows that the GPPV invariant is a specialization of the $2$-variable series $P_{\widehat F, [k]}^{\infty}$ at $t=1$, thus establishing Theorem \ref{thm: Zhat intro} stated in the introduction. Additionally, Section \ref{sec: Recovering q series} gives calculations in specific examples.
Let $a\in \delta + 2\Z^s$ be a representative of a $\spinc$ structure $[a]$ on $Y$, using the convention \eqref{eq:GM spinc structures}. Following \cite{GPPV} (see also \cite[Section 4.3]{GM}), consider
\begin{equation}
\label{eq:Zhat_a}
\widehat{Z}_a(q) := q^{- \frac{3s + \sum_v m_v}{4}} \cdot
v.p.
\oint\limits_{\vert z_v \vert =1} \prod_{v\in \mathcal{V}(\Gamma)} \frac{dz_{v}}{2\pi i z_{v}}\left(z_{v}-\frac{1}{z_{v}}\right)^{2-\delta_v} \cdot \Theta_a^{-M}(z),
\end{equation}
where
\begin{equation}
\label{eq:theta function}
\Theta_a^{-M}(z) := \sum_{\ell \in a + 2M\Z^s} q^{-\frac{\ell^t M^{-1}\ell}{4}} \prod_{v\in \mathcal{V}(\Gamma)} z_v^{\ell_v}.
\end{equation}
In \eqref{eq:Zhat_a}, $v.p.$ denotes the principal value, that is the average of the integrals over $\vert z_v \vert = 1 +\varepsilon$ and over $\vert z_v \vert = 1 -\varepsilon$, for small $\varepsilon>0$. Note that, since $M$ is negative definite, for each power of $q$ the expression (\ref{eq:theta function}) for
$\Theta_a^{-M}(z)$ is a Laurent polynomial in the variables $\{z_v\}_{v\in \mathcal{V}}$, and the exponent $\left(-\ell^t M^{-1} \ell\right)/4$ as $\ell$ varies is bounded below. It is clear from the definition that $\widehat{Z}_a(q)$ is independent of the choice of representative $a\in [a]$.
We begin by rewriting \eqref{eq:Zhat_a} as
\begin{align}\label{eq: rewrite}
\widehat{Z}_a(q)
=
q^{- \frac{3s+ \sum_{v} m_v}{4}}\sum\limits_{\ell\in a+2M\Z^s}
\left[
\prod\limits_{v}v.p. \frac{1}{2\pi i}\oint\limits_{|z_{v}|=1}\frac{1}{z_{v}}\left(z_{v}-\frac{1}{z_{v}}\right)^{2-\delta_v}z_{v}^{\ell_{v}}dz_{v}
\right]
q^{-\frac{\ell^t M^{-1}\ell}{4}}.
\end{align}
Our analysis of the $\widehat Z$ invariant, in particular Proposition \ref{prop:Zhat_k in terms of F} and Theorem \ref{thm: Zhathat Zhat} below, depends on the properties of the coefficient given by the expression in the square brackets in equation \eqref{eq: rewrite}. Thus we start by rewriting it in more concrete terms. We note that this preliminary analysis, leading to Definition \ref{def:Fhat}, amounts to taking a detailed look at the coefficients denoted $F_{\vec{\ell}}$ in \cite[equation (43)]{GM}.
To compute the integral in equation \eqref{eq: rewrite}, write $\left(z_{v}-z_{v}^{-1}\right)^{2-\delta_v}$ as a Laurent series $E_v^-$ in $z_v$ for the integral over $\lr{z_v} = 1 - \varepsilon$, and as a Laurent series $E_v^+$ in $z_v^{-1}$ for the integral over $\lr{z_v} = 1+ \varepsilon$. Then for $\ell_v\in \Z$,
\begin{equation*}
\label{eq:principal value with residues}
v.p. \frac{1}{2\pi i}\oint\limits_{|z_{v}|=1}\frac{1}{z_{v}}\left(z_{v}-\frac{1}{z_{v}}\right)^{2-\delta_v}z_{v}^{\ell_{v}}dz_{v}
=
\frac{1}{2}\left[
\operatorname{Res}\left(z_v^{\ell_v - 1} E_v^-,0\right)
+
\operatorname{Res}\left(z_v^{-\ell_v - 1}E_v^+(z_v^{-1}),0\right)
\right].
\end{equation*}
Note that $\operatorname{Res}\left(z_v^{\ell_v - 1} E_v^-,0\right) $ and $\operatorname{Res}\left(z_v^{-\ell_v - 1}E_v^+(z_v^{-1}),0\right) $ equal the coefficient of $z^{-\ell_v}$ in $E_v^-$ and $E_v^+$, respectively. We will now identify the Laurent series $E_v^-$ and $E_v^+$ more explicitly.
When $\delta_v\leq 2$, the exponent in $(z_v- z_v^{-1})^{2-\delta_v}$ is non-negative and $E_v^+ =E_v^-$ = $(z_v- z_v^{-1})^{2-\delta_v}$ is a Laurent polynomial. In particular, if $\delta_v\leq 2$ for all vertices $v$, then $\widehat{Z}_a(q)$ is a Laurent polynomial with integer coefficients. More generally, coefficients of $\widehat{Z}_a(q)$ are in $2^{-c}\cdot \Z$ where $c$ is the number of vertices of degree at least 3.
We now describe the Laurent series expansions $E_v^{\pm}$ of $(z_v - z_v^{-1})^{2-\delta_v}$ when $\delta_v\geq 3$. Fix $n\geq 3$. For $\lr{z}<1$, using the expansion $
\left( z - \frac{1}{z} \right)^{-1}
=
\frac{-z}{1-z^2}
=
-\sum\limits_{i\geq 0}z^{2i+1} $,
we can write
\begin{equation}
\label{eq:expansion <1}
\left(z - \frac{1}{z}\right)^{2-n} = \left( -\sum\limits_{i\geq 0} z^{2i+1}\right)^{n -2}.
\end{equation}
For $\lr{z}>1$, the expansion $\left( z - \frac{1}{z} \right)^{-1}
=
\frac{z^{-1}}{1-z^{-2}}
=
\sum\limits_{i\geq 0}z_v^{-(2i+1)}$ gives
\begin{equation}
\label{eq:expansion >1}
\left(z - \frac{1}{z}\right)^{2-n} = \left( \sum_{i\geq 0} z^{-(2i+1)} \right)^{n-2}.
\end{equation}
Then $E_v^-$ and $E_v^+$ are given by substituting $z=z_v$, $n=\delta_v$ into the right-hand side of \eqref{eq:expansion <1} and \eqref{eq:expansion >1}, respectively. We summarize the discussion so far: the expression in square brackets in equation \eqref{eq: rewrite}
equals the product over $v\in \mathcal{V}(\Gamma)$ of the average of the coefficients of $z^{-\ell_v}$ in
\eqref{eq:expansion <1}, \eqref{eq:expansion >1}.
We now define a family of functions $\widehat{F} = \{\widehat{F}_n\}_{n\geq 0}$ which record the coefficients in the average of the two expansions. In Proposition \ref{prop:Fhats are admissible} we show this family is admissible.
\begin{defn}
\label{def:Fhat}
Consider the following family of functions $\{ \widehat{F}_n : \Z \to \Q \}_{n\in \Z_+}$.
For $0\leq n \leq 2$, set $\widehat{F}_n(r)$ to be the coefficient of $z^{-r}$ in $(z-z^{-1})^{2-n}$. For $n\geq 3$, $\widehat{F}_n(r)$ is defined to be the average of the coefficients of $z^{-r}$ in equations \eqref{eq:expansion <1} and \eqref{eq:expansion >1}.
\end{defn}
Note that $\widehat{F}_n$ takes values in $\Z$ for $0\leq n\leq 2$ and in $\frac{1}{2}\Z$ for $n \geq 3$. Thus $\widehat{F}_{\Gamma,k}$, defined in equation \eqref{eq:F_Gamma,k}, takes values in $2^{-c} \Z$ where $c$ is the number of vertices of degree at least $3$. Although an explicit formula for $\widehat{F}_n$, $n\geq 3$, will not be used in this paper, for the reader's convenience we record it in equation \eqref{eq:Fhat formula}.
\begin{align}
\label{eq:Fhat formula}
\widehat{F}_{n}(r) =
\begin{cases}
\frac{1}{2}\operatorname{sgn}(r)^{n}\begin{pmatrix}
\frac{n+|r|}{2}-2\\
n-3
\end{pmatrix}&\text{ if }|r|\geq n-2\text{ and } r\equiv n\bmod 2\\
0&\text{ otherwise.}
\end{cases}
\end{align}
Here $\operatorname{sgn}(r) \in \{-1, 1\}$ denotes the sign of $r$.
\subsection{Three admissible families} \label{subsec: Three admissible families}
In this section we introduce families $\widehat{F}^{+}$, $\widehat{F}^{-}$, closely related to $\widehat{F}$, and show that they are all admissible.
\begin{defn}
\label{def:Fhatpm}
For $r\in \Z$ and $0\leq n \leq 2$, set $\widehat{F}^+_n(r) = \widehat{F}^-_n(r) = \widehat{F}_n(r)$ to be the coefficient of $z^{-r}$ in $(z-z^{-1})^{2-n}$. For $n\geq 3$, $\widehat{F}_n^{-}(r)$ and $\widehat{F}_n^+(r)$ are defined to be the coefficient of $z^{-r}$ in \eqref{eq:expansion <1} and \eqref{eq:expansion >1}, respectively.
\end{defn}
The following general observation is used in the proof of the proposition below. If $F^1, F^2, \ldots, F^m$ are admissible families valued in a field of characteristic zero, then the family $\operatorname{av}(F^1, \ldots, F^m)$ given by the average
\begin{equation}
\label{eq:averaging}
\left(\operatorname{av}(F^1, \ldots, F^m) \right)_n = \frac{1}{m} \sum_{i=1}^m F^1_n + \cdots + F^m_n
\end{equation}
is again admissible.
\begin{prop}
\label{prop:Fhats are admissible}
The families $\widehat{F}^+$, $\widehat{F}^-$, and $\widehat{F}$ are admissible.
\end{prop}
\begin{proof}
Property \ref{item:A1} is straightforward to verify. To show property \ref{item:A2} for $\widehat{F}^+$, note that
\[
(z-z^{-1}) \sum_{i\geq 0} z^{-(2i+1)} = 1.
\]
Therefore
\[
z \left( \sum_{i\geq 0} z^{-(2i+1)} \right)^{n-2} - z^{-1} \left( \sum_{i\geq 0} z^{-(2i+1)} \right)^{n-2} = \left(\sum_{i\geq 0} z^{-(2i+1)} \right)^{n-3},
\]
which demonstrates \ref{item:A2}. Alternatively, \ref{item:A2} may also be seen from a binomial coefficient identity, using an explicit formula for $\widehat F^{\pm}$, analogous to \eqref{eq:Fhat formula}. The calculation for $\widehat{F}^-$ is similar. Finally, note that $\widehat{F}$ is the average of $\widehat{F}^+$ and $\widehat{F}^-$ and is therefore admissible.
\end{proof}
\subsection{The lattice and \texorpdfstring{$\widehat{Z}$}{}} \label{sec: The lattice and Zhat}
In this section we reformulate $\widehat{Z}$ as a sum of contributions of the associated function $\widehat{F}_{\Gamma,k}$ (see equation \eqref{eq:F_Gamma,k} and Remark \ref{Gamma remark}) over lattice points, using the lattice cohomology identification of $\spinc$ structures.
As a first step, we reparameterize definition \eqref{eq:Zhat_a} in the following way. Every $\ell \in a+2M\Z^s$ can be written in the form $\ell = a + 2M x$ for a unique $x\in \Z^s$. Then
\[
\frac{\ell^t M^{-1} \ell}{4} = \frac{a^2}{4} + a\cdot x + \langle x,x\rangle
= \frac{a^2}{4} - 2\chi^{}_{a}(x),
\]
using the notation of Remark \ref{rem: k^2 meaning} and \eqref{eq:chi}. Compare with \cite[Equation (46)]{GM}. Thus we get
\begin{equation}
\label{eq:Zhat_a reformulation}
\widehat{Z}_a(q) =
q^{- \frac{a^2 + 3s + \sum_v m_v}{4}} v.p.
\oint\limits_{|z_{v}|=1}\prod_{v}\frac{dz_{v}}{2\pi i z_{v}}\left(z_{v}-\frac{1}{z_{v}}\right)^{2-\delta_v}
\left(\sum\limits_{x\in \mathbb{Z}^{s}}q^{2\chi^{}_{ a}(x)}\prod\limits_{v}z_{v}^{(a+2Mx)_{v}}\right).
\end{equation}
We now move on to the main goal of this section. Recall from Section \ref{sec:identification of spinc structures} that graded roots and lattice cohomology use a different identification of $\spinc$ structures than the $\widehat{Z}$ invariant. The translation between these two identifications is given in equation \eqref{eq:translating between spinc identifications}. Given $k\in m+2\Z^s$, let $a=k-Mu \in \delta + 2\Z^s$ denote the corresponding $\spinc$ representative, and set
\begin{equation*}
\Zhat^{\lattice}_k(q) := \widehat{Z}_a(q).
\end{equation*}
Recall Notation \ref{notn} for ${\Delta_k}$ and $\varepsilon_k(x)$ in the following statement.
\begin{prop}
\label{prop:Zhat_k in terms of F}
For $k\in m + 2\Z^s$, we have
\[
\Zhat^{\lattice}_k(q) = \sum_{x\in \Z^s} \widehat{F}_{\Gamma,k}(x) q^{\varepsilon_k(x)}.
\]
\end{prop}
\begin{proof}
Note that $2\chi^{}_{a}(x) = 2\chi^{}_{k}(x) + \langle x,u\rangle$ for all $x\in \Z^s$, so equation \eqref{eq:Zhat_a reformulation} with $a=k-Mu$ can be written as
\[
\widehat{Z}_a(q) =
q^{\Delta_k} v.p.
\oint\limits_{|z_{v}|=1}\prod_{v}\frac{dz_{v}}{2\pi i z_{v}}\left(z_{v}-\frac{1}{z_{v}}\right)^{2-\delta_v}
\left(\sum\limits_{x\in \mathbb{Z}^{s}}q^{2\chi^{}_{ k}(x) + \langle x,u\rangle}\prod\limits_{v}z_{v}^{(2Mx+k-Mu)_{v}}\right).
\]
From the above equation and the discussion preceding Section \ref{subsec: Three admissible families}, we see that for every $j\in \Z$, the coefficient of $q^{\Delta_k + j}$ in $\widehat{Z}_k^{\lattice}$ is equal to
\[
\sum\limits_{\substack{x\in \Z^s \\ 2\chi^{}_{k}(x)+\langle x,u\rangle=j}}\widehat{F}_{\Gamma, k}(x),
\]
which verifies the desired equality.
\end{proof}
\subsection{Recovering the $q$-series} \label{sec: Recovering q series}
In this section we show that, when the admissible family is $\widehat{F}$, the two-variable series specializes to $\widehat{Z}_k^{\lattice}(q)$ by setting $t=1$. Calculations for $S^3$ and $\Sigma(2,7,15)$ are presented further below.
\begin{defn}
Fix a negative definite plumbing $\Gamma$ and a $\spinc$ structure $[k]$. Define
\begin{align}
\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!_{[k]}(q,t) := P^{\infty}_{\widehat{F},[k]}
\end{align}
which, as we recall from Definition \ref{defn:P infinity}, is an invariant of $(Y(\Gamma), [k])$ up to multiplication by a power of $t$.
\end{defn}
\begin{thm} \label{thm: Zhathat Zhat}
With the above notation,
\[
\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!_{[k]}(q,1) = \widehat{Z}_k^{\lattice}(q).
\]
\end{thm}
\begin{proof}
Fix $j\in \Z$. Using the notation in the proof of Theorem \ref{thm:convergence}, the coefficient of $q^{\Delta_k + j}$ in $\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!_{[k]}(q,1)$ is equal to
\[
\sum_{x\in \partial\til{\mathcal{S}}_j} \widehat{F}_{\Gamma,k}(x),
\]
which by Proposition \ref{prop:Zhat_k in terms of F} equals the coefficient of $q^{\Delta_k + j}$ in $\widehat{Z}_k^{\lattice}(q)$.
\end{proof}
\begin{exmp}
Consider $Y=S^3$ represented as a plumbing $\Gamma$ consisting of a single vertex $v$ with weight $m_v=-1$. Let $k=-1\in m + 2M\Z$ be a representative of the unique $\spinc$ structure. We have
\[
\chi^{}_k(x) = \frac{1}{2}(x^2+x),\ \ \varepsilon_k(x) = -\frac{1}{2} + x^2,\ \ \widehat{F}_{\Gamma,k}(x) = \widehat{F}_0(-2x).
\]
The formula in \eqref{eq:F_1 and F_0} implies $\widehat{F}_{\Gamma,k}(\pm 1) = 1$, $\widehat{F}_{\Gamma,k}(0) = -2$, and $\widehat{F}_{\Gamma,k}(x) = 0$ for all other values of $x$. Since $\chi^{-1}_k(1)$ contains the lattice points $-1,0,1$, we see that the weights stabilize at the sublevel set $\chi^{-1}_k((-\infty, 1])$. The weighted graded root and $\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!_{[k]}$ are given in Figure \ref{fig:wgroot S^3}.
\begin{figure}
\centering
{ \begin{tikzpicture}
\node (0) at (0,0) {$\bullet$};
\node (1) at (0,1) {$\bullet$};
\node (2) at (0,2) {$\bullet$};
\node at (0,2.75) {$\vdots$};
\node[right] at (2.05,2.75) {$\vdots$};
\draw (0,0) -- (0,2);
\draw[dashed] (0,0) --++ (2,0);
\draw[dashed] (0,1) --++ (2,0);
\draw[dashed] (0,2) --++ (2,0);
\node[right] at (2,0) {$0$};
\node[right] at (2,1) {$2$};
\node[right] at (2,2) {$4$};
\node[left] at (-.2,0) {$-2q^{-\frac{1}{2}} + tq^{\frac{1}{2}}$};
\node[left] at (-.2,1) {$-2q^{-\frac{1}{2}} + tq^{\frac{1}{2}} + t^{-1}q^{\frac{1}{2}}$};
\node[left] at (-.2,2) {$-2q^{-\frac{1}{2}} + tq^{\frac{1}{2}} + t^{-1}q^{\frac{1}{2}}$};
\node at (7,1) {$\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!_{[k]}(q,t) = -2q^{-\frac{1}{2}} + (t + t^{-1})q^{\frac{1}{2}}$};
\end{tikzpicture}
}
\caption{The weighted graded root and two-variable series $\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!_{[k]}(q,t)$ for $S^3$ corresponding to the admissible family $\widehat{F}$. Recall from grading conventions \ref{grading conv} that the numbers $0,2,4$ to the right of the graded root denote the Heegaard Floer gradings. }\label{fig:wgroot S^3}
\end{figure}
\end{exmp}
In particular, specializing $\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!_{[k]}(q,t)$ at $t=1$ recovers the $\widehat{Z}$ invariant for $S^3$. The calculation above shows that the $2$-variable series $\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!$ introduced in this paper is different from the conjectured Poincar\'{e} series of the BPS homology \cite[Equation (6.80)]{GPV}, \cite[Equation (18)]{GM}.
\begin{exmp} \label{2 7 15 ex involution0}
Consider the Brieskorn sphere $\Sigma(2,7,15)$, which can be represented as a negative definite plumbing, as shown in Figure \ref{fig:plumbing and link ex}. Since $\Sigma(2,7,15)$ is an integer homology sphere, we denote its unique $\spinc$ structure by $\mathfrak{s}_0$.
One can compute, cf. \cite[Section 4.6]{GM}:
\begin{align*}
\widehat{Z}_{\mathfrak{s}_{0}}(q) &=
q^{1739/840}\sum\limits_{n=0}^{\infty}\left[q^{(61+420n)^2/840}+q^{(149+420n)^2/840}+q^{(299+420n)^2/840}+q^{(331+420n)^2/840}\right]\\
&-q^{1739/840}\sum_{n=0}^{\infty}\left[q^{(89+420n)^2/840}+q^{(121+420n)^2/840}+q^{(271+420n)^2/840}+q^{(359+420n)^2/840}\right]\\
&=q^{13/2} - q^{23/2} - q^{39/2}+q^{57/2} - q^{179/2} + q^{217/2} + q^{265/2} - q^{311/2} +\cdots
\end{align*}
The beginning of the weighted graded root (a result of a computer calculation) for $\Sigma(2,7,15)$ is shown in Figure \ref{Weighted_graded_root}; additional weights are given in the table below.
\begin{table}[H]
\begin{tabular}{|c|c|}
\hline
Weight & Grading \\[.2cm]\hline
\rule{0pt}{3ex}$\frac{1}{2}(t^{-2} + 1)q^{\frac{13}{2}} - t^{-1}q^{\frac{23}{2}} - t^{-1}q^{\frac{39}{2}}$ & 20 \\[.2cm] \hline
\rule{0pt}{3ex} $\frac{1}{2}(t^{-2} + 1)q^{\frac{13}{2}} - t^{-1}q^{\frac{23}{2}} - t^{-1}q^{\frac{39}{2}} + \frac{1}{2}q^{\frac{57}{2}}$ & 28 \\[.2cm] \hline
\end{tabular}
\end{table}
\noindent In particular, setting $t=1$ in the weight at grading $28$, one can see the first few terms of $\widehat{Z}_{\mathfrak{s}_{0}}(q)$ as result of stabilization, which is a consequence of Theorems \ref{thm:convergence}, \ref{thm: Zhathat Zhat}.
\end{exmp}
\section{$\Spinc$ conjugation}\label{section:conj}
In this section we study the behavior of $\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!$ and weighted graded roots under $\spinc$ conjugation. Under the identification (\ref{eq:LC spinc structures}), conjugation is given by the map $[k]\to [-k]$. Both $\widehat{Z}$ and lattice cohomology, in particular graded roots, are invariant under conjugation. However, when considering our new theory of weighted graded roots, a different, more refined, story emerges which we now describe.
Let $F$ be an $\mathcal{R}$-valued admissible family. Consider the following property.
\begin{equation}
\label{eq:A3}
F_n(-r) = (-1)^n F_n(r) \text{ for all } n\geq 0 \text{ and } r\in \Z. \tag{A3}
\end{equation}
\begin{prop}\label{t inversion}
If $F$ is an admissible family which satisfies property \eqref{eq:A3}, then $P_{F,[k]}^{\infty}(q,t) \doteq P_{F,[-k]}^{\infty}(q,t^{-1})$ for all $k\in m+2\Z^s$.
\end{prop}
\begin{proof}
Note that $k':= -k + 2Mu$ is a representative for $[-k]$. We will show that
\begin{equation} \label{eq: Zhathat conj proof}
F_{\Gamma,k}(x) q^{\varepsilon_k(x)}t^{\langle x, u\rangle} = F_{\Gamma,k'}(-x)q^{\varepsilon_{k'}(-x)}(t^{-1})^{\langle -x, u\rangle}
\end{equation}
for all $x\in \Z^s$, and the claim follows. First,
\[
\Delta_{k'} = \Delta_{k},\hspace{2em} 2\chi_k(x) = 2\chi^{}_{k'}(-x) - 2 \langle x,u\rangle,
\]
so $\varepsilon_{k}(x) = \varepsilon_{k'}(-x)$ for all $x\in \Z^s$. Next, $2Mx + k - Mu = -\left(2M(-x) +k' - Mu \right)$, so
\begin{align*}
F_{\Gamma,k}(x) = (-1)^{\sum_v \delta_v} F_{\Gamma,k'}(-x) = F_{\Gamma,k'}(-x),
\end{align*}
where the first equality follows from property \eqref{eq:A3} and the second is due to the sum of degrees in any graph being even. Lastly, $t^{\langle x, u\rangle} = (t^{-1})^{\langle -x, u\rangle}$ is automatic.
\end{proof}
Now note that $\widehat{F}$, introduced in Definition \ref{def:Fhat}, satisfies property \eqref{eq:A3}.\footnote{Although it will not be used, we note that $\widehat{F}^{\pm}$ from Definition \ref{def:Fhatpm} do not satisfy \eqref{eq:A3}.}
\begin{cor} \label{prop: Zhathat conjugation}
$\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!_{[k]}(q, t) \doteq \widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!_{[-k]}(q, t^{-1})$. In particular, setting $t = 1$ recovers the conjugation invariance of $\widehat{Z}$.
\end{cor}
We now turn to weighted graded roots and illustrate, via two examples, some interesting behavior under $\spinc$ conjugation. First, we briefly recall how graded roots transform under conjugation and describe the corresponding story in Heegaard Floer homology.
Given a negative definite plumbing $\Gamma$ and a $\spinc$ structure $[k]$, the map $\Z^s\to \Z^s$, sending $x$ to $-x$
induces an isomorphism $(R_{k}, \chi^{}_{k})\cong (R_{-k}, \chi^{}_{-k})$ since $\chi^{}_{k}(x) = \chi^{}_{-k}(-x)$. Similarly, in Heegaard Floer homology, for any closed oriented 3-manifold $Y$ and $\spinc$ structure $\mathfrak{s}$, there is an isomorphism $HF^+(Y,\mathfrak{s})\cong HF^+(Y, \bar{\mathfrak{s}})$, where $\bar{\mathfrak{s}}$ is the conjugate of $\mathfrak{s}$; see \cite[Theorem 2.4]{OS-applications}.
Moreover, for a self-conjugate $\spinc$ structure we get an involution on the graded root and on Heegaard Floer homology. The involution on the graded root is induced by the map
\begin{align*}
\Z^s\to \Z^s,\hspace{1em} x\mapsto -x-M^{-1}k.
\end{align*}
Note here $M^{-1}k\in \Z^s$ since $[k]=[-k]$. For Heegaard Floer homology, the involution
\begin{align*}
\iota: HF^+(Y, \mathfrak{s}) \to HF^+(Y, \mathfrak{s})
\end{align*}
comes from a chain map obtained by considering what happens when a pointed Heegaard diagram $(\Sigma, \alpha, \beta, z)$ representing $Y$ is replaced with $(-\Sigma, \beta, \alpha, z)$. The involution $\iota$ is at the foundation of involutive Heegaard Floer homology, an extension of Heegaard Floer homology due to Hendricks-Manolescu \cite{involutive}.
For $\Gamma$ an almost rational plumbing, Dai-Manolescu show that the two involutions described above are identified under the isomorphism given in Theorem \ref{thm: LC-HF iso} (see \cite[Theorem 3.1]{DM}). Furthermore, they show that the graded root is symmetric about the infinite stem and the involution is the reflection about the infinite stem.
\begin{exmp} \label{2 7 15 ex involution}
Consider again the Brieskorn sphere $\Sigma(2,7,15)$. Note, the plumbing given in Figure \ref{fig:plumbing and link ex} describing $\Sigma(2,7, 15)$ is almost rational. Also, since $\Sigma(2, 7, 15)$ only has one $\spinc$ structure, $\mathfrak{s}_{0}$, it is self-conjugate by default. Hence, the corresponding graded root is symmetric about the infinite stem and the involution is the reflection. However, as seen in Figure \ref{Weighted_graded_root}, the weighted graded root is no longer symmetric and the involution does not preserve all of the weights. There is a node at grading level $6$ which has weight $\frac{1}{2}q^{13/2}$, whereas the node on the opposite side of the infinite stem has weight $0$. The reason for this symmetry breaking is a result of the failure of $\widehat{F}_{\Gamma, k}(-x-M^{-1}k)q^{\varepsilon_{k}(-x-M^{-1}k)}t^{\langle-x-M^{-1}k, u\rangle}$ to equal $\widehat{F}_{\Gamma, k}(x)q^{\varepsilon_{k}(x)}t^{\langle x, u\rangle}$.
\end{exmp}
The following example shows that, unlike $\widehat{Z}$ and graded roots, the weighted graded root can distinguish conjugate $\spinc$ structures. Moreover, it exhibits a new phenomenon different from that in Corollary \ref{prop: Zhathat conjugation}.
\begin{exmp} \label{ex: asymmetric conjugation}
Let $\Gamma$ be the plumbing pictured below:
\begin{center}
\begin{tikzpicture}[scale=.7]
\node (m) at (0,0) {$\bullet$};
\node (r) at (2,0) {$\bullet$};
\node (l) at (-2,0) {$\bullet$};
\node (b) at (0,-2) {$\bullet$};
\node (t) at (0,2) {$\bullet$};
\draw (-2,0) --++ (4,0);
\draw (0,2) --++ (0,-4);
\node[above] at (-.5,0) {$-1$};
\node[above] at (r) {$-7$};
\node[above] at (l) {$-11$};
\node[left] at (b) {$-10$};
\node[left] at (t) {$-3$};
\end{tikzpicture}
\end{center}
Order the vertices so that $v_{1}, v_{2}, v_{3}, v_{4}, v_{5}$ correspond to the vertices with weights $-1$, $-7$,$-10$,\\$-11$,$-3$. Let $k=(-5,5,8,9, 1)$. Consider the $\spinc$ structure $[k]$ and its conjugate $[-k]$.
\begin{figure}
\centering
{\begin{tikzpicture}
\node at (-3,0) {$\bullet$};
\node at (-3,1) {$\bullet$};
\node at (-3,2) {$\bullet$};
\node at (-3,3) {$\bullet$};
\node at (-3,4) {$\bullet$};
\node at (-3,5) {$\bullet$};
\node at (-3,6) {$\bullet$};
\node at (-3,7) {$\bullet$};
\node at (-3,8) {$\bullet$};
\node at (-3,9) {$\bullet$};
\node at (3,0) {$\bullet$};
\node at (3,1) {$\bullet$};
\node at (3,2) {$\bullet$};
\node at (3,3) {$\bullet$};
\node at (3,4) {$\bullet$};
\node at (3,5) {$\bullet$};
\node at (3,6) {$\bullet$};
\node at (3,7) {$\bullet$};
\node at (3,8) {$\bullet$};
\node at (3,9) {$\bullet$};
\draw (-3,0)--(-3, 9);
\draw (3,0)--(3, 9);
\node[right] at (7,0) {$\frac{570}{769}$};
\node[right] at (7,1) {$\frac{2108}{769}$};
\node[right] at (7,2) {$\frac{3646}{769}$};
\node[right] at (7,3) {$\frac{5184}{769}$};
\node[right] at (7,4) {$\frac{6722}{769}$};
\node[right] at (7,5) {$\frac{8260}{769}$};
\node[right] at (7,6) {$\frac{9798}{769}$};
\node[right] at (7,7) {$\frac{11336}{769}$};
\node[right] at (7,8) {$\frac{12874}{769}$};
\node[right] at (7,9) {$\frac{14412}{769}$};
\draw[dashed] (-3,0)--(7,0);
\draw[dashed] (-3,1)--(7,1);
\draw[dashed] (-3,2)--(7,2);
\draw[dashed] (-3,3)--(7,3);
\draw[dashed] (-3,4)--(7,4);
\draw[dashed] (-3,5)--(7,5);
\draw[dashed] (-3,6)--(7,6);
\draw[dashed] (-3,7)--(7,7);
\draw[dashed] (-3,8)--(7,8);
\draw[dashed] (-3,9)--(7,9);
\node[left] at (3,.2) {$0$};
\node[left] at (3,1.2) {$0$};
\node[left] at (3,2.2) {$0$};
\node[left] at (3,3.2) {$0$};
\node[left] at (3,4.3) {$\frac{1}{2}q^{\frac{15009}{1538}}$};
\node[left] at (3,5.3) {$\frac{1}{2}q^{\frac{15009}{1538}}$};
\node[left] at (3,6.3) {$\frac{1}{2}q^{\frac{15009}{1538}}$};
\node[left] at (3,7.3) {$q^{\frac{19623}{1538}}t^{-3} + \frac{1}{2}q^{\frac{15009}{1538}}$};
\node[left] at (3,8.3) {$q^{\frac{19623}{1538}}t^{-3} + \frac{1}{2}q^{\frac{15009}{1538}}$};
\node[left] at (3,9.3) {$q^{\frac{19623}{1538}}t^{-3} + \frac{1}{2}q^{\frac{15009}{1538}}$};
\node[left] at (-3,.2) {$0$};
\node[left] at (-3,1.2) {$0$};
\node[left] at (-3,2.2) {$0$};
\node[left] at (-3,3.2) {$0$};
\node[left] at (-3,4.2) {$0$};
\node[left] at (-3,5.2) {$0$};
\node[left] at (-3,6.2) {$0$};
\node[left] at (-3,7.3) {$q^{\frac{19623}{1538}}
t^{3}+\frac{1}{2}q^{\frac{15009}{1538}}$};
\node[left] at (-3,8.3) {$q^{\frac{19623}{1538}}
t^{3}+\frac{1}{2}q^{\frac{15009}{1538}}$};
\node[left] at (-3,9.3) {$q^{\frac{19623}{1538}}
t^{3}+\frac{1}{2}q^{\frac{15009}{1538}}$};
\node at (-3,10) {$\vdots$};
\node at (3,10) {$\vdots$};
\node at (7.5,10) {$\vdots$};
\node at (-3,-1) {$(R_k, \chi^{}_{k}, P_{\widehat{F},k})$};
\node at (3,-1) {$(R_{-k}, \chi^{}_{-k}, P_{\widehat{F},-k})$};
\end{tikzpicture}}
\caption{Note, $d(-Y(\Gamma), [(-5,5,8,9,1)]) = d(-Y(\Gamma), [(5,-5,-8,-9,-1)]) = \frac{570}{769}$ \label{fig: cross graded root}. We use the normalization discussed in Remark \ref{rmk:normalizing t}.}
\end{figure}
Initial segments of the weighted graded roots (a result of a computer calculation) corresponding to $[k]$ and $[-k]$ are pictured in Figure \ref{fig: cross graded root}. As discussed in the beginning of this section, the $\widehat{Z}$ invariant and graded roots are invariant under $\spinc$ conjugation. In this example the weighted graded roots not only distinguish $[k]$ and $[-k]$, they do this by more than just inversion of $t$ in all the weights, compare with Corollary \ref{prop: Zhathat conjugation}. (Note that the weights of graded roots are well defined up to an overall normalization by a power of $t$.)
For example, the node at grading level $\frac{6722}{769}$ for $(R_{[k]}, \chi^{}_{[k]}, P_{\widehat{F},[k]})$ is $0$, while the corresponding node for $(R_{[-k]}, \chi^{}_{[-k]}, P_{\widehat{F},[-k]})$ is $\frac{1}{2}q^{\frac{15009}{1538}}$.
Note that $\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!(q,t)$ is the limit of the Laurent polynomial weights, whose coefficients stabilize in every bidegree according to Theorem \ref{thm:convergence}. The weighted graded roots carry the unstable information as well; this explains the discrepancy between this example and Corollary \ref{prop: Zhathat conjugation}.
On a more detailed level, the reason for the discrepancy by more than just inversion of $t$ is due to the failure of $\widehat{F}_{\Gamma, -k}(-x)q^{\varepsilon_{-k}(-x)}(t^{-1})^{\langle -x,u\rangle}$ to equal $\widehat{F}_{\Gamma, k}(x)q^{\varepsilon_{k}(x)}t^{\langle x,u\rangle}$. Equation
\eqref{eq: Zhathat conj proof} in the proof of Proposition \ref{t inversion}, where $k'= -k + 2Mu$,
was sufficient for showing $\widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!_{[k]}(q, t)\doteq \widehat{\vphantom{\rule{5pt}{10pt}}\smash{\widehat{Z}}\,}\!_{[-k]}(q, t^{-1})$ because the sum is taken over all lattice points $x\in \Z^s$. However, the weights on the nodes of the graded root are sums over lattice points in some connected component of a sublevel set of $\chi^{}_{k}$ for $(R_k, \chi^{}_{k}, P_{\widehat{F},k})$, and of $\chi^{}_{-k}$ for $(R_{-k}, \chi^{}_{-k}, P_{\widehat{F},-k})$. But the map $x\mapsto -x$ takes the connected components of $\chi^{}_{k}$ sublevel sets to connected components of $\chi^{}_{-k}$ sublevel sets, not connected components of $\chi^{}_{-k + 2Mu}$ sublevel sets.
\end{exmp}
\bibliographystyle{amsalpha}
|
\section{Acknowledgements}
\textbf{\textit{Acknowledgements}} \, We thank Brian Batell for useful discussions on leptophobic DM models. This work was funded by the U.S. Department of Energy, the U.S. National Science Foundation, Los Alamos National Laboratory LDRD program, and PAPIIT-UNAM grant No.\,\,IT100420. We also wish to acknowledge support from the LANSCE Lujan Center and LANL's Accelerator Operations and Technology (AOT) division. This research used resources provided by the Los Alamos National Laboratory Institutional Computing Program, which is supported by the U.S. Department of Energy National Nuclear Security Administration under Contract No.\,89233218CNA000001.
\section{Acknowledgements}
\textbf{\textit{Acknowledgements}} \, We thank Brian Batell for useful discussions on leptophobic DM models. This work was funded by the U.S. Department of Energy, the U.S. National Science Foundation, Los Alamos National Laboratory LDRD program, and PAPIIT-UNAM grant No.\,\,IT100420. We also wish to acknowledge support from the LANSCE Lujan Center and LANL's Accelerator Operations and Technology (AOT) division. This research used resources provided by the Los Alamos National Laboratory Institutional Computing Program, which is supported by the U.S. Department of Energy National Nuclear Security Administration under Contract No.\,89233218CNA000001.
|
\section{Introduction}
\label{sec:intro}
\label{sec:intro}
\subsection{Dynamical sampling problems over graphs} Signals that arise from social, biological, and sensor networks are typically structured and interconnected, which are modelled as residing on weighted
graphs \cite{ceccon2017graph}. The demand for large-scale data processing tasks over graphs such as graph-based filtering \cite{chen2013adaptive,loukas2016frequency} and semi-supervised classifications \cite{ekambaram2013wavelet,chen2014semi} has inspired an emerging field of graph signal processing \cite{shuman2013emerging,sandryhaila2014big}, where a cornerstone problem is the robust recovery of graph signals from its sampled values on a proper subset of nodes \cite{pesenson2008sampling,pesenson2009variational,chen2015discrete,segarra2016reconstruction,puy2018random}.
In general, this type of inverse problem is ill-posed and the exact reconstruction is impossible. Furthermore, the task of selecting a subset of nodes whose values enable reconstruction of the values in the entire graph with minimal loss is known to be NP- hard \cite{davis1997adaptive}.
However, many signals in real world applications exhibit structural features such as smoothness. In classical signal processing, the smoothness assumptions are often defined in terms of the signal’s Fourier transform. An important class of smooth signals is that of bandlimited signals, i.e., signals whose Fourier transforms are compactly supported. Bandlimited graph signals are defined in a similar way using the graph Fourier transform. That is to say, the signal is spanned by the eigenvectors corresponding to the small eigenvalues of the graph Laplacian. Such signals have appeared in various applications such as
temperature readings across U.S., wind speed across Minnesota, and natural image reconstruction \cite{chen2016signal}. Motivated by these applications, the sampling and reconstruction of bandlimited graph signals have drawn considerable attention in recent years.
Most previous work in this area considered the static setting: various sampling and reconstruction algorithms for bandlimited graph signals, inspired from classical Shannon-sampling \cite{chen2015discrete,pesenson2008sampling,pesenson2010sampling} and compressed sensing \cite{puy2018random}, have been proposed. For $k$-bandlimited graph signals, one has to select at least $k$ nodes to ensure exact reconstruction in the noiseless case, and it has been proved that $\mathcal{O}(k \log k)$ random samples are sufficient to recover all $k$-bandlimited signals with high probability \cite{puy2018random}.
In many cases, the bandlimited signals over graphs are evolving and the underlying dynamical process can be modelled as a heat diffusion process over graphs, such as rumor propagation over social networks \cite{xiao2019rumor}, traffic movement over transportation networks \cite{yu2017spatiotemporal}, spatial temperature profiles measured by a wireless sensor network \cite{thanou2017learning}, and neural activities at different regions of the brain \cite{sporns2010networks}. The existing sampling and reconstruction algorithms, however, do not take dynamical process into account. In practice, we can obtain the space-time samples of the diffusive field by placing the sensors at various spatial locations and activating them at different time instances. On one hand, due to application-specific restrictions, it is often the case that one can not take a sufficient number of spatial samples at a single time instance to recover a bandlimited signal. On the other hand, the acquisition cost of space-time samples
is much cheaper than the same amount of spatial samples, since the latter requires the physical presence of more sensors in the network, whereas the former is, in theory, only constrained by the communication capacity and energy budget of the sensor network.
The above mentioned practical concerns motivated a new sampling approach termed ``dynamical sampling" \cite{aldroubi2013dynamical,aldroubi2017dynamical}. Dynamical sampling deals with processing a time series of evolving signals. The goal is to recover the dynamical systems from a union of coarse spatial samples at multiple time instances. A core problem to address is where to put sensors, when to take samples and how to reconstruct the dynamical system from measured samples.
In this paper, we consider the dynamical sampling of a bandlimited diffusion field over the graph, where we assume the physical law is known. Specifically, the homogeneous heat diffusion process over $\mathcal{G}$ is governed by
\begin{equation}\label{linearsystem}
\frac{\partial }{\partial t}\mathbf{x}_t =-L\mathbf{x}_t, t\geq 0,
\end{equation} where $L$ is the normalized graph Laplacian of $\mathcal{G}$. The diffusive field $\mathbf{x}_t$ is completely determined by the initial condition $\mathbf{x}_0$:
$\mathbf{x}_t =e^{-tL}\mathbf{x}_0=:A^{t/\Delta t}\mathbf{x}_0, t\geq 0$
where $A \triangleq e^{-\Delta t}$ and $\Delta t>0$. In other words, $\mathbf{x}_t$ is just a filtered version of the initial state $\mathbf{x}_0$, with the filter being a time varying Laplacian kernel.
We observe the dynamics up to $T$ time instances. For simplicity, let $\Omega \subset [n] \times \{0, \Delta t,\cdots, (T-1) \Delta t \}$ denote the space-time locations where the entries of the diffusive field $\{\mathbf{x}_t: t\geq 0\}$ are observed. Now we decompose $\Omega=\bigcup_{\ell=0}^{T-1}\Omega_{\ell} \times \{\ell\}$ where $\Omega_{\ell}\subset [n]$ is the set of observational spatial locations at $t_{\ell}=\ell\Delta t$. We use $S_{\Omega_{\ell}}$ to denote the observation matrix that
$S_{\Omega_{\ell}}\mathbf{x}=\sum_{i\in \Omega_{\ell}}\mathbf{x}(i)\mathbf{e}_i$ where $\{\mathbf{e}_i\}_{i=1}^{n}$ is the standard orthonormal basis in $\mathbb{C}^n$. Motivated by the way we gather the information from the social networks, we will investigate three random space-time sampling regimes with an example shown in Fig.~\ref{fig:sampling}: { \ding{172} Random spatial locations at initial time, \ding{173} Random spatial locations at consecutive times and \ding{174} Random space-time locations}. In the first regime, the sampling locations are the same for all time instances
and are randomly selected according to a predefined probability distribution $\mathbf{p}^{(1)}$ on $[n]$. For the second regime, the sampling locations at different time instances are randomly selected according to their own predefined probability distributions $\mathbf{p}^{(2)}(t)$. Different from the first two sampling regimes, the evolved signals are treated as a signal in $\mathbb{R}^{nT}$ in the third sampling regime and the locations are randomly selected according to a probability distribution $\mathbf{p}^{(3)}$ in $[nT]$
\begin{figure}
\centering
\begin{minipage}{.3\linewidth} \centering \small Regime 1 \end{minipage}
\begin{minipage}{.28\linewidth} \centering \small Regime 2 \end{minipage}
\begin{minipage}{.28\linewidth} \centering \small Regime 3 \end{minipage}\\
\includegraphics[width=.9\linewidth, keepaspectratio]{figs/sampling.pdf}
\caption{\footnotesize\label{fig:sampling} An example of three space-time sampling regimes using the same number of samples. Here the $x$ axis denotes the time instances and the $y$ axis denotes the spatial values of evolving signals with colors representing the magnitude. Only values at colored spots are observed. }
\end{figure}
The goal is to investigate the conditions on $\Omega$ to achieve stable recovery of the initial signal $\mathbf{x}_0$ from (possibly noisy) partial observations of states from a single trajectory $\{S_{\Omega_\ell}\mathbf{x}_{t_\ell}: \ell=0,1,\cdots, T-1\}$ and propose robust reconstruction strategies.
\subsection {Summary of contributions}
In this paper, we propose three random space-time sampling regimes to sample {bandlimited graph diffusive fields}.
For each regime, we introduce a quantity called the \textit{spectral graph weighted coherence of order $(k,T)$}. Such a quantity depends on the interplay between the sampling distribution and the localization of the bandlimited diffusion fields over the space-time locations. We provide sampling complexity bounds to ensure robust recovery of initial signal with high probability and show that they scale linearly with the square of spectral graph weighted coherence of order $(k,T)$. We show that temporal information can reduce the spatial sampling density and one can achieve comparable (even better) results given the same amount of space-time samples with the static case. We further provide a detailed comparison of three regimes in terms of their sample complexity and the influence of sampling probability distributions on the performance. More details are discussed in Section \ref{subs:properties}.
Finally, we propose a robust method to reconstruct any $k$-bandlimited signal from its noisy samples, with a detailed error analysis. The idea is to use an $\ell_2$ regularization term that utilizes the graph Laplacian and bandlimited property of the graph signals. The results show that our methods can recover $k$-bandlimited signals exactly in the absence of noise and the method is also robust to measurement acquisition noise and model errors. In addition, extensive numerical simulations are conducted on various graphs including an example of deblurring a real image from the space-time samples of its diffusion process, which further support our theoretical results and confirm that random dynamical sampling offers a robust alternative for the standard spatial sampling.
Note that our space-time sampling theorems are applicable to any dynamical system over the graph of the form $x_t=Ax_{t-1}$ with the evolution operator $A$ being a diagonalizable operator. For example, $A$ can be a polynomial function of any symmetrical Laplacian or adjacency matrix of any weighted undirected graph. It is also possible to generalize the current framework to bandlimited limited graph processes governed by a linear time-invariant system \cite{isufi2020observing}.
Nevertheless, the reconstruction methods proposed here are specifically designed to take advantage of the semi-definite positivity of the evolution operator. For simplicity and conciseness, we therefore concentrate on the normalized Laplacians.
Our work is built upon recent progress on sampling of bandlimited graph signals and dynamical sampling over regular domains. In particular, we generalize the work in \cite{puy2018random} to bandlimited diffusion fields with three random-space sampling regimes. In Section \ref{relatedwork}, we give more detailed comparisons with these relevant works.
\section{Related work} \label{relatedwork}
\subsection{Dynamical sampling over regular domain} Previous work on dynamical sampling has focused on deterministic sampling over regular domains. Dynamical sampling was first proposed in \cite{aldroubi2013dynamical,aldroubi2017dynamical} by Aldroubi et al for linear time invariant systems, motivated by the pioneering work of \cite{lu2009spatial} that considered the space-time sampling of bandlimited diffusion fields over the real line, see also \cite{ranieri2011sampling,dokmanic2011sensor, murray2015estimating,murray2016physics} for various extensions. The majority of the work \cite{aceska2013dynamical,cabrelli2017dynamical,aldroubi2021sampling,aldroubi2020phaseless,tang2017universal,ulanovskii2021reconstruction,aldroubi2019frames,lai2017undersampled,christensen2019frame} has focused on developing deterministic sampling theorems to ensure exact reconstruction, where the number and the location of sensors are fixed for successive time instances. Our paper is the first one to extend the dynamical sampling framework to graphs and consider the random space-time sampling regimes: the locations and numbers of sensors can change over time, and the observation time period is not necessarily successive. We will show that this allows us to adapt the graph topology to the sampling such that one needs much fewer number of space-time samples yet yields a more robust reconstruction.
\subsection{Sampling and reconstruction of bandlimited graph signals}
Sampling and reconstruction of bandlimited graph signals (corresponding to $T=1$ in our case) have been extensively studied and achieved remarkable progress. The existing approaches can be mainly categorized into selection sampling and aggregation sampling. In the first category, deterministic sampling (see e.g.~\cite{chen2015discrete,chamon2017greedy, pesenson2008sampling,pesenson2009variational, anis2016efficient, huang2020reconstruction}) and random sampling scheme (see e.g.~\cite{chen2015signal,chen2016signal,hashemi2018accelerated,puy2018random}) were proposed. Using ideas of variable density sampling from compressed sensing, Vandergheynst et al. \cite{puy2018random} derive random sampling schemes based on probability distributions over the graph vertexes. In this paper, we generalize their work to the bandlimited diffusion field and propose three different random space-time sampling schemes so that the work \cite{puy2018random} becomes a special case of our regimes with $T=1$. In the second category, aggregation sampling aims to recover signals from (random weighted) linear combinations of signal values on a neighborhood of selected nodes. It corresponds to sampling a time-varying signal driven by a graph shift operator, whose matrix form has the same sparsity pattern with the adjacent matrix. The deterministic scheme proposed in \cite{marques2015sampling} (see \cite{wang2016local,yang2021orthogonal} for extensions) can be viewed a special case of our regime 1 with a pre-selected set of nodes; the random scheme \cite{valsesia2018sampling} used a graph shift operator with i.i.d.~Gaussian entries, corresponding to the case of $T=2$ and observations are only made at $T=2$.
Compared to the static setting, there are relatively fewer work considering the sampling and reconstruction problem in dynamical system setting. Different models for the time series over graph have been proposed. The examples include distributed reconstruction algorithms for graph low-pass filtering processes \cite{wang2015distributed} and batch reconstruction methods \cite{qiu2017time} for graph signals whose temporal difference is bandlimited.
\cite{isufi2020observing} considered
both deterministic and Bernoulli random space-time sampling for bandlimited graph processes. Necessary conditions on probabilities are derived to ensure the exact reconstruction of the process and a convex optimization approach was proposed to choose the optimal sampling design. The proposed regime is most related to our regime 3. However, our work is different: we also consider regimes 1 and 2, and we use a different observation model by extending the idea of variable density sampling.
\section{Preliminaries and Notation} \label{sec: NP}
\subsection{Operators associated to graphs} We consider an undirected weighted graph $\mathcal{G}=(\mathcal{V},\mathcal{E},W)$ where $\mathcal{V}=\{v_1,v_2,\cdots, v_n\}$ is set of $n$ vertices and $ \mathcal{E} \subset\mathcal{ V\times V}$ is a set of edges. {Let $\{v_i,v_j\}$ denote the edge if $v_i$ and $v_j$ are connected}.
The weighted adjacent matrix $W$ is defined as
\[
W(i,j) \triangleq
\begin{cases}
\alpha_{ij}&\mbox{if $\{v_i,v_j\} \in \mathcal{E}$}\\
0 &\mbox{otherwise}
\end{cases}; \alpha_{ij} \in \mathbb{R}_{+}; \text{ for all } v_i,v_j \in \mathcal{V}.\]
The weight associated with each edge in the graph often {indicates similarity or dependency between the corresponding vertices}. The connectivity and edge weights are dictated by the physics of the problem at hand, for instance, the weights are inversely proportional to the physical distance between nodes in the network. In particular, $W$ is symmetric. The \textit{degree} $\deg(v_i)$ of a vertex $v_i \in V$ is defined as $\deg(v_i)=\sum_{j=1}^n W(i,j)$. ${D} :={\mathrm{diag}}(\deg(v_i)_{i=1}^{n})$ denotes the diagonal degree matrix. In the following, we introduce important operators associated with the graph $\mathcal{G}$.
\begin{definition}The normalized diffusion operator of a graph $\mathcal{G}$ with weighted adjacent matrix ${W} \in \mathbb{R}^{n \times n}$ is defined by
$N = {D}^{-\frac{1}{2}}W {D}^{-\frac{1}{2}}$. We call $N$ the \textit{diffusion matrix} of $\mathcal{G}$. The normalized graph Laplacian operator is $L =I-N$.
\end{definition}
\subsection{Bandlimited graph signal} The matrix $L$ is real symmetric, and so it admits an eigen-decomposition ${L}=U\Sigma U^{\top}$ where the columns of $U\in \mathbb{R}^{n\times n}$ are orthonormal eigenvectors and the diagonal entries of $\Sigma$ consist of $n$ eigenvalues $\sigma_1,\cdots, \sigma_n$. Furthermore, semi-definite positivity of $L$ implies that all eigenvalues are non-negative. Without loss of generality, we assume that $\sigma_1\leq\cdots\leq\sigma_n$. {Note that $A=\exp(-\Delta t L)$, we have $A= U\Lambda U^\top $ with $\Lambda:=\exp(-\Delta t \Sigma)$. Let's denote the diagonal entries of $\Lambda$ as $\lambda_1=\exp(-\Delta t \sigma_1), \cdots,\lambda_n=\exp(-\Delta t\sigma_n)$. Thus we have $\lambda_1\geq \cdots\geq \lambda_n$. }
In graph signal processing, the matrix $U$ is often viewed as the graph Fourier transform. For any signal $\mathbf{x}\in \mathbb{R}^n$ defined on the nodes of the graph $\mathcal{G}$, $\widehat{\mathbf{x}}=U^\top\mathbf{x}$ {represents} the Fourier coefficients of $\mathbf{x}$ ordered in increasing frequencies. The smooth graph signals are modelled
as $k$ bandlimited signals: for $\mathbf{x}\in \mathbb{R}^n$ on $\mathcal{G}$ with bandwidth $k>0$ it satisfies
$\mathbf{x}=U_k\widehat{\mathbf{x}}^k$
where $\widehat{\mathbf{x}}^k:=\widehat{\mathbf{x}}(1:k) \in \mathbb{R}^k$ and
$U_k:=(\mathbf{u}_1
\cdots,\mathbf{u}_k) \in \mathbb{R}^{n\times k}$,
i.e., $U_k$ is the restriction of $U$ to its first $k$ vectors. This yields the following formal definition of a $k-$bandlimited signal.
\begin{definition}[Bandlimited signal on $\mathcal{G}$] A signal $\mathbf{x} \in \mathbb{R}^n$ defined on the nodes of the graph $\mathcal{G}$ is
$k$-bandlimited with $k\in \mathbb{N}\setminus\{0\}$ if $\mathbf{x}\in \mathrm{span}(U_k)$.\end{definition}
Note that we use $U_k$ in our definition of $k$-bandlimited signals to handle the case where the eigen-decomposition
is not unique.
\subsection{Notation}
In this work, capital letters are used for matrices, lower boldface letters for vectors, and regular letters for scalars. The set of the first $d$ natural numbers is denoted by $[d]$. For integers $i,n, T$, the notation $ i:n:Tn$ stands for the set $\{k=i+jn\in[i, Tn]: \text{for some }j\in\mathbb{Z}\}$ and $I_d$ stands for the $d\times d$ identity matrix.
We let ${\mathrm{diag}}(\mathbf{x})$ denote the diagonal matrix with diagonal entries given by vector $\mathbf{x}$ and ${\mathrm{diag}}(X)={\mathrm{diag}}(X(1,1),\cdots,X(d,d))$ for $X\in\mathbb{R}^{d\times d}$.
Let $\mathbf{x}_1, \cdots,\mathbf{x}_t$ be vectors and $X_1,\cdots,X_t$ be matrices, we define ${\mathrm{diag}}(\mathbf{x}_1;\cdots;\mathbf{x}_{t})$ and ${\mathrm{diag}}(X_1;\cdots;X_{t})$ as the block diagonal matrix with
the $(i,i)$-th block of ${\mathrm{diag}}(\mathbf{x}_1; \cdots;\mathbf{x}_{t})$ being ${\mathrm{diag}}(\mathbf{x}_i)$ and the $(i,i)$-th block of ${\mathrm{diag}}(X_1;\cdots;X_{t})$ being $X_i$ respectively.
Let $T$ be a positive integer and we define
\begin{align}\label{map:f_T}
f_{T}(\lambda)=\sqrt{\sum_{t=0}^{T-1}\lambda^{2t}},\;
F_{k,T}= {\mathrm{diag}}\left(\frac{1}{f_{T}(\lambda_1)},\cdots,\frac{1}{f_{T}(\lambda_k)}\right),\; \Lambda_k={\mathrm{diag}}(\lambda_1,\cdots,\lambda_k).
\end{align}
\section{Sampling of $k$-bandlimited diffusion fields} In this section, we describe three regimes that select a subset of space-time locations to sample a $k$-bandlimited diffusion field. Then, we prove that the corresponding space-time sampling procedure stably embeds the set of initial signals that consists of $k$-bandlimited signals.
We describe how to reconstruct initial signals from the space-time samples in Section 5.
\subsection{The space-time sampling procedure}
To select the space-time sampling set $\Omega$, we introduce various probability distributions $\mathbf{p}$ serving as sampling distributions for three random regimes.
\begin{definition}\label{def:distributions}
The probability distributions can be defined as follows:
\begin{itemize}
\item[\ding{172}] \textbf{Random spatial locations at initial time:} let $\mathbf{p}^{(1)}$ be a probability distribution on $[n]$. We define $P_1 \in\mathbb{R}^{Tn\times Tn}$ as $P_1={\mathrm{diag}}({\mathbf{p}^{(1)};\cdots;\mathbf{p}^{(1)}})$.
\item[\ding{173}] {\bf Random spatial locations at consecutive times:} let $\mathbf{p}_t^{(2)}$ be a probability distribution on $[n]$ describing sampling at time $t$, for all $t=0,1,\cdots,T-1$. The associated matrix $P_2\in\mathbb{R}^{Tn\times Tn}$ is defined as $P_2={\mathrm{diag}}(\mathbf{p}_0^{(2)}; \cdots;\mathbf{p}_{T-1}^{(2)} )$.
\item[\ding{174}] {\bf Random space-time locations:} we let $\mathbf{p}^{(3)}$ be a probability distribution on $[Tn]$ and define $P_3\in\mathbb{R}^{Tn\times Tn}$ as $\mathrm{diag}(\mathbf{p}^{(3)})$.
\end{itemize}
\end{definition}
We assume that the probability distributions above are all non-degenerate, i.e., entry-wise nonzero. Once the probability distributions are assigned, we obtain the nodes in the space-time sampling sets $\Omega$
independently (with replacement) according to the probability distributions. Note that $\Omega$ only needs to be selected once to sample all $k$-bandlimited diffusion fields over $\mathcal{G}$. After the sampling distributions and sampling set $\Omega$ are determined, we define the associated variables as follows.
\begin{definition}\label{def_MPW}
For three regimes, we define the associated sampling matrices and weighting matrices as follows.
\begin{itemize}
\item[\ding{172}] \textbf{Random spatial locations at initial time:} $\Omega^{(1)}:= \Omega^{(1)}_0 \times [T]$ where $\Omega^{(1)}_0 =\{\omega_1^{(1)},\cdots,\omega_{m}^{(1)}\}$ is constructed by drawing $m$ indices independently (with replacement) from set $[n]$ according to probability $\mathbf{p}^{(1)}$, i.e.,
$\mathbb{P}(\omega_j^{(1)}=i)=\mathbf{p}^{(1)}(i), \text{ for all } j\in[m] \text{ and }i\in[n]$.
We define the sampling matrix
\begin{equation}\label{eqn:samp_mat_fixed}
S_1={\mathrm{diag}}( \underbrace{S_1^{(0)};\cdots; S_1^{(0)}}_{T}),
\end{equation}
where $ S_1^{(0)}\in\mathbb{R}^{m\times n}$ is defined as
$ S_1^{(0)}(i,j):=\begin{cases}
1, &\text{ if }j=\omega_i\\
0,& \text{otherwise}
\end{cases}
$. Thus, we can define the probability matrix on the sampling set $P_{\Omega^{(1)}}=S_1P_1S_1^\top$.
To average the samples' information at each time instance, we also introduce $M=Lm$ and the weighting matrix $W_1\in\mathbb{R}^{M\times M}$ as
\begin{equation}\label{eqn:wgt_mat_fixed}
W_1=\frac{1}{m}{\mathrm{diag}}(\underbrace{I_{m};\cdots;I_{m}}_{T}).
\end{equation}
\item[\ding{173}] \textbf{Random spatial locations at consecutive times}: let $\Omega^{(2)}= \cup_{t=0}^{T-1}(\Omega_{t}^{(2)}\times \{t\})$ with $\Omega_t^{(2)}:=\{\omega_{t,1}^{(2)},\cdots,\omega_{t,m_t}^{(2)}\}$, where $\Omega_t^{(2)}$ is constructed by drawing $m_t$ indices independently (with replacement) from the set $[n]$ according to the probability distribution $\mathbf{p}_t^{(2)}$, i.e.,
$\mathbb{P}(\omega_{t,j}^{(2)}=i)=\mathbf{p}^{(2)}_t(i)$, for all $j\in[m_t] \text{ and }i\in[n]$.
We set $M=\sum_{t=0}^{T-1}m_t$ and define the sampling matrix $S_2\in\mathbb{R}^{M\times Tn}$ as
\begin{equation}\label{eqn:samp_mat_rand_space}
S_2={\mathrm{diag}}( S_2^{(0)}; S_2^{(1)};\cdots; S_2^{(T-1)})
\end{equation}
where $S_2^{(t)}\in\mathbb{R}^{m_t\times n}$ is defined as $S_2^{(t)}(i,j):=\begin{cases}
1, &j=\omega_{t,i}^{(2)}\\
0,& \text{otherwise},
\end{cases}
$ for $t=0,\cdots,T-1$. Thus, we could define the probability matrix on the sampling set $P_{\Omega^{(2)}}=S_2P_2S_2^\top$.
To average the samples at each time instance, we define the weighting matrix $W_2\in\mathbb{R}^{M\times M}$ as
\begin{equation}\label{eqn:wgt_mat_fixed}
W_2={\mathrm{diag}}(\frac{1}{m_0}I_{m_0}; \frac{1}{m_1}I_{m_1};\cdots; \frac{1}{m_{T-1}}I_{m_{T-1}}).
\end{equation}
\item[\ding{174}] \textbf{Random space-time locations}: $\Omega^{(3)}:=\{\omega_1^{(3)},\cdots,\omega_{m}^{(3)}\}$ is constructed by drawing $m$ indices independently (with replacement) from $[Tn]$ according to probability distribution $\mathbf{p}^{(3)}$, i.e.,
$\mathbb{P}(\omega_j=i)=\mathbf{p}^{(3)}(i), \forall j\in[m] \text{ and }i\in[Tn]$.
For this sampling regime, we define the sampling matrix $S_3\in\mathbb{R}^{m\times Tn}$ as
\begin{equation}\label{eqn:samp_mat_rand_tsp}
S_3(i,j):=\begin{cases}
1, &\text{ if }j=\omega_i^{(3}\\
0,& \text{otherwise}
\end{cases},
\end{equation}
the probability matrix on the sampling set $P_{\Omega^{(3)}}=S_3P_3S_3^\top$, and weighting matrix $W_3=\frac{1}{M}I_{M}$ with $M=m$.
\end{itemize}
\end{definition}
\subsection{Stable embedding of $k$-bandlimited signals via space-time samples}
First let's introduce a map
$\pi_{A,T}:\mathbb{R}^n \rightarrow \mathbb{R}^{Tn}$ that maps $k$-bandlimited signals to $k$-bandlimited diffusion fields and is defined by $\pi_{A,T}(\mathbf{x})=[\mathbf{x}^\top,(A\mathbf{x})^\top, \cdots, (A^{T-1}\mathbf{x})^\top]^\top$.
Notice that if $\mathbf{x}\in \mathrm{span}(U_k)$, then $\pi_{A,T}(\mathbf{x})\in \mathrm{span}(\widetilde{U}_{k,T})$, where $\widetilde{U}_{k,T} \in \mathbb{R}^{Tn\times k}$ consists of orthonormal columns vectors and is defined below
\begin{equation}\label{eqn:tilde_UkL}
\widetilde{U}_{k,T}=\begin{bmatrix}\frac{1}{f_{T}(\lambda_1)}\mathbf{u}_1& \frac{1}{f_{T}(\lambda_2)} \mathbf{u}_2&\cdots&\frac{1}{f_{T}(\lambda_k)} \mathbf{u}_k\\\frac{\lambda_1}{f_{T}(\lambda_1)} \mathbf{u}_1& \frac{\lambda_2}{f_{T}(\lambda_2)} \mathbf{u}_2&\cdots & \frac{\lambda_k}{f_{T}(\lambda_k)} \mathbf{u}_k\\ \vdots&\vdots&\cdots&\vdots\\ \frac{\lambda_1^{T-1}}{f_{T}(\lambda_1)} \mathbf{u}_1 &\frac{\lambda_2^{T-1}}{f_{T}(\lambda_2)}\mathbf{u}_2&\cdots&\frac{\lambda_k^{T-1}}{f_{T}(\lambda_k)} \mathbf{u}_k\end{bmatrix}.
\end{equation}
This observation implies that $\pi_{ A,T}:\mathrm{span}(U_k) \rightarrow \mathrm{span}(\widetilde{U}_{k,T})$. The following lemma shows that $\pi_{A,T}$ is a stable embedding and specifies lower and upper embedding constants.
\begin{lemma}\label{lmm: embedding}
For any $\mathbf{x}\in \mathrm{span}(U_k)$, we have that
\begin{equation*}
f_{T}(\lambda_k)\|\mathbf{x}\|_2 \leq \|\pi_{{A},T}(\mathbf{x})\|_2 \leq f_{T}(\lambda_1)\|\mathbf{x}\|_2,
\end{equation*}
where the map $f_T$ is defined in \eqref{map:f_T}.
\end{lemma}
The key observation that $\pi_{ A,T}(\mathrm{span}(U_k))= \mathrm{span}(\widetilde{U}_{k,T})$ inspires us to interpret $k$-bandlimited diffusion fields as $k$-bandlimited signals over an extended space-time graph $\mathcal{G}_T$, whose top $k$ Laplacian eigenvectors are given by $ \widetilde{U}_{k,T}$. Leveraging this interpretation, we connect the space-time sampling of bandlimited diffusion fields over $\mathcal{G}$ with the classical spatial sampling of $k$-bandlimited graph signals over $\mathcal{G}_T$.
Similar to compressed sensing, the number of space-time samples needed to guarantee stable reconstruction of $k$-bandlimited diffusion fields will depend on a quantity, the spectral graph weighted coherence of order $(k,T)$, that represents how the energy of bandlimited diffusion fields spreads over the space-time nodes. We will derive and understand this quantity via sampling on $\mathcal{G}_T$.
Let $\delta_i \in \mathbb{R}^{Tn}$ denote the Dirac at the space-time node $i$.
The quantity
\begin{equation}\label{gcoherence}
\|\widetilde{U}_{k,T}(i,:)\|_2 = \|\widetilde{U}_{k,T}^{\top}\delta_i\|_2
\end{equation} describes how much the energy of $\delta_i$ is concentrated on the first $k$ Fourier modes of $\mathcal{G}_T$, or equivalently, the distribution of the energy of the $k$-bandlimited diffusion fields over the space-time nodes.
When the quantity \eqref{gcoherence} is one, then there exist $k$-bandlimited diffusion fields whose energy is solely concentrated on the $i$th space-time node, and not sampling the $i$th node jeopardises the chance of reconstructions. When it is zero, then no $k$-bandlimited diffusion fields over $\mathcal{G}$ has a part of its energy on the $i$th node; one can safely remove this node from the sampling set. We thus see that the quality of our sampling method will depend on the interplay between the sampling distribution and the quantities $\|\widetilde{U}_{k,T}^{\top}\delta_i\|_2$ for $i \in [nT]$. Ideally, the larger $\|\widetilde{U}_{k,T}^{\top}\delta_i\|_2$ is, the higher probability of the node $i$ is observed. Below, we introduce the spectral graph weighted coherence of order $(k,T)$ to characterize the interplay between the sampling distribution for each regime and space-time nodes, where the case of $T=1$ corresponds to the graph weighted coherence introduced in \cite[Definition 2.1]{puy2018random} for the static case.
\begin{definition}[Graph spectral weighted coherence of order $(k,T)$ $\mathbf{\nu}_{j,\mathbf{p}^{(j)}}^{k,T}$ ]
\label{def:gswc}
Let $\mathbf{p}^{(j)}$ represent a sampling distribution for $j=1,2,3$. The graph spectral weighted coherence $\mathbf{\nu}_{j,\mathbf{p}^{(j)}}^{k,T}$ of order $(k,T)$ for the triple $(\mathcal {G},T,\mathbf{p}^{(j)})$ is defined as following
\begin{itemize}
\item[\ding{172}] \textbf{Random spatial locations at the initial time:} \vspace{-1.5mm}
\begin{equation}\label{eqn:co_fixed}
\begin{aligned}
\mathbf{\nu}_{1,\mathbf{p}^{(1)}}^{k,T}:=
\max_{1\leq i\leq n}\left\{\frac{1}{\sqrt{\mathbf{p}^{(1)}(i)}}\left\| \widetilde{U}_{k,T}((i:n:Tn),:) \right\|_2\right\},
\end{aligned}
\end{equation}
where $\widetilde{U}_{k,T}((i:n:Tn),:)$ is the row submatrix of $\widetilde{U}_{k,T}$ with row indices in $(i:n:Tn)$
\item[\ding{173}] \textbf{Random spatial locations at consecutive times:}
\vspace{-1.5mm}
\begin{equation}\label{eqn:co_rand_space}
\mathbf{\nu}_{2,\mathbf{p}^{(2)}}^{k,T}:=\begin{bmatrix}
\mathbf{\nu}_{\mathbf{p}_0^{(2)}}\\
\vdots\\
\mathbf{\nu}_{\mathbf{p}_{T-1}^{(2)}}
\end{bmatrix}=\begin{bmatrix}\max_{1\leq i\leq n} \frac{1}{\sqrt{\mathbf{p}_0^{(2)}(i)}}\|\widetilde{U}_{k,T}^\top\delta_{i}\|_2\\
\vdots\\
\max_{1\leq i\leq n} \frac{1}{\sqrt{\mathbf{p}_{T-1}^{(2)}(i)}}\|\widetilde{U}_{k,T}^\top\delta_{i+(T-1)n}\|_2\\
\end{bmatrix},
\end{equation}
\item[\ding{174}] \textbf{Random space-time locations: }
\begin{equation}\label{eqn:co_tsp}
\mathbf{\nu}_{3,\mathbf{p}^{(3)}}^{k,T}:=\max_{1\leq i\leq Tn}\left\{\frac{1}{\sqrt{\mathbf{p}_{i}^{(3)}}}\|\widetilde{U}_{k,T}^{\top}\delta_i\|_2 \right\},
\end{equation}
\end{itemize}
\end{definition}
From the definition, the spectral graph weighted coherence of order $(k,T)$ for each regime is thus a characterization of the interaction among the $k$-bandlimited signals, spectrum of the graph Laplacian, and the respective sampling distributions. For the ease of the readers, we summarize three sampling regimes in Table \ref{tab:var_SR}.
\begin{table}[t]\footnotesize
\centering
\begin{tabular}{|l|l|l|l| }
\hline
Sampling &\multirow{2}{*}{ \ding{172}} &\multirow{2}{*}{ \ding{173}} &\multirow{2}{*}{ \ding{174} } \\
Regimes &&&\\\hline
$\mathbf{p}^{(j)}$& $\mathbf{p}^{(1)}\in\mathbb{R}^{n}$ & $\mathbf{p}^{(2)}=[\mathbf{p}_0^{(2)};
\cdots;
\mathbf{p}_{T-1}^{(2)}]\in\mathbb{R}^{n\times T}$& $\mathbf{p}^{(3)}\in\mathbb{R}^{Tn}$ \\
\hline
$P_j$ &$P_1={\mathrm{diag}}(\underbrace{\mathbf{p}^{(1)};\cdots;\mathbf{p}^{(1)}}_{T})$&$P_2={\mathrm{diag}}\left(\mathbf{p}_0^{(2)};\mathbf{p}_1^{(2)};\cdots;\mathbf{p}_{T-1}^{(2)}\right) $ &$P_3={\mathrm{diag}}(\mathbf{p}^{(3)})$\\
\hline
\multirow{2}{*}{ $\Omega$}& { $\Omega^{(1)}=\Omega^{(1)}_0 \times [T]$} & $\Omega^{(2)}= \bigcup_{t=0}^{T-1}\Omega_{t}^{(2)} \times \{t\}$, & \multirow{2}{*}{ $\Omega^{(3)}=\{\omega_1^{(3)},\cdots,\omega_{m}^{(3)}\}$}\\
& $\Omega^{(1)}_0 =\{\omega_1^{(1)},\cdots,\omega_{m}^{(1)}\}$& $\Omega_t^{(2)}=\{\omega_{t,1}^{(2)},\cdots,\omega_{t,m_t}^{(2)}\}$.&\\
\hline
\multirow{4}{*}{ $S_j$}& $S_{1}={\mathrm{diag}}( \underbrace{S_1^{(0)}; \cdots; S_1^{(0)}}_{T})$,& $S_2={\mathrm{diag}}( S_2^{(0)}; \cdots; S_2^{(T-1)})$ &\multirow{2}{*}{ $S_3\in\mathbb{R}^{m\times Tn}$ with}\\
&$ S_1^{(0)}\in\mathbb{R}^{m\times n}$ with & $S_2^{(t)}\in\mathbb{R}^{m_t\times n}$ with &\\
& $ S_1^{(0)}(i,j)=\begin{cases}
1, & j=\omega_i\\
0,& \text{otherwise}.
\end{cases}$ & $S_2^{(t)}(i,j)=\begin{cases}
1, & j=\omega_{t,i}\\
0,& \text{otherwise}
\end{cases}$
& $S_3(i,j)=\begin{cases}
1, & j=\omega_i\\
0,& \text{otherwise}.\\
\end{cases}$\\
&&for $t=0,\cdots,T-1$. &\\
\hline
$P_{\Omega^{(j)}}$& $P_{\Omega^{(1)}}=
{\mathrm{diag}}(S_1 P_1 S_1^{\top})$& $P_{\Omega^{(2)}}={\mathrm{diag}}(S_2P_2S_2^{\top})$ & $P_{\Omega^{(3)}}={\mathrm{diag}}(S_3P_3S_3^{\top})$\\
\hline
$W_j$ & $W_1=\frac{1}{m}I_{mT}$&$W_2={\mathrm{diag}}(\frac{1}{m_0}I_{m_0}; \cdots; \frac{1}{m_{T-1}}I_{m_{T-1}})$& $W_3=\frac{1}{m}I_m $\\%\frac{1}{m}{\mathrm{diag}}(\underbrace{I_m;\cdots;I_m}_{T})$&$W_2={\mathrm{diag}}(\frac{1}{m_0}I_{m_0}; \cdots; \frac{1}{m_{T-1}}I_{m_{T-1}})$& $W_3=\frac{1}{m}I_m$\\
\hline
$\mathbf{\nu}_{j,\mathbf{p}^{(j)}}^{k,T}$&$\mathbf{\nu}_{1,\mathbf{p}^{(1)}}^{k,T}$ in \eqref{eqn:co_fixed} & $\mathbf{\nu}_{2,\mathbf{p}^{(2)}}^{k,T}$ in \eqref{eqn:co_rand_space}& $\mathbf{\nu}_{3,\mathbf{p}^{(3)}}^{k,T}$ in \eqref{eqn:co_tsp}\\
\hline
\end{tabular}
\caption{\footnotesize Relevant notations and matrices for different random space-time sampling regimes.}\label{tab:var_SR}
\end{table}
\subsection{Properties of spectral graph weighted coherences}\label{subs:properties} As shown later, the sampling size required to achieve stable embedding essentially depends on the spectral graph weighted coherence. Next we analyze its properties in Proposition \ref{prop:gcoherence}.
\begin{proposition}\label{prop:gcoherence} We list several properties of the spectral graph weighted coherences of order $(k,T)$ that are useful for later analysis.
\begin{itemize}
\item[(a)]For any probability distribution $ \mathbf{p}^{(j)}$ for $j=1,2,3$, we have that
\begin{itemize}
\item [(a.1)] $ \left(\mathbf{\nu}_{1,\mathbf{p}^{(1)}}^{k,T}\right)^2\geq \sum_{i=1}^{n}\left\| \widetilde{U}_{k,T}((i:n:Tn),:) \right\|_2^2 \geq k$ and the first inequality becomes an equality when
\begin{equation}\label{opt1}
\mathbf{p_{opt}}^{(1)}(i)=\frac{ \left\|
\widetilde{U}_{k,T}((i:n:Tn),:)) \right\|_2^2}{\sum_{\ell=1}^{n}\left\|
\widetilde{U}_{k,T}((\ell:n:Tn),:) \right\|_2^2} \quad \text{ for } i=1,\cdots,n.
\end{equation}
\item[(a.2)] $(\nu_{2,\mathbf{p}^{(2)}}^{k,T}(t))^2 \geq \sum\limits_{\ell=1}^{k}\frac{\lambda_{\ell}^{2t}}{f_{T}^2(\lambda_\ell)} $ for $t=0,\cdots,T-1$ and the equality holds when
\begin{equation}\label{opt2}
\mathbf{p_{t,opt}}^{(2)}(i)= \frac{\|\widetilde{U}_{k,T}^{\top}\delta_{i+tn}\|_2^2}{\sum_{\ell=1}^{k}\lambda_{\ell}^{2t}/f_{T}^2(\lambda_\ell) }.
\end{equation}
As a result,
$\sum_{t=1}^{T} \left(\nu_{2,\mathbf{p}^{(2)}}^{k,T}(t)\right)^2 \geq k$.
\item[(a.3)] $\left(\mathbf{\nu}_{3,\mathbf{p}^{(3)}}^{k,T}\right)^2\geq k$ and the equality holds when
\begin{equation}\label{opt3}
\mathbf{p_{opt}}^{(3)}(i)=\frac{ \|\widetilde{U}_{k,T}^{\top}\delta_i\|_2^2}{k}
\end{equation}
\end{itemize}
\item[(b)] For any sampling distributions, we have that
\begin{equation*}
T\left(\nu_{1,\mathbf{p}^{(1)}}^{k,T}\right)^2 \geq \left(\nu_{3,\mathbf{p}^{(3)}}^{k,T}\right)^2 \geq \sum_{t=0}^{T-1}\left(\nu_{2,\mathbf{p}^{(2)}}^{k,T}(t)\right)^2. \end{equation*}
In particular, we call the probability defined in \eqref{opt1}, \eqref{opt2} and \eqref{opt3} the optimal distribution for the regime 1,2 and 3. In this case, we have
\begin{equation*} T\left(\nu_{1,\mathbf{p}^{(1)}}^{k,T}\right)^2 \geq \left(\nu_{3,\mathbf{p}^{(3)}}^{k,T}\right)^2 = \sum_{t=0}^{T-1}\left(\nu_{2,\mathbf{p}^{(2)}}^{k,T}(t)\right)^2=k. \end{equation*}
\end{itemize}
\end{proposition}
We are now ready to present the main theorem in this section. We represent three random space-time sampling procedures of $k$-bandlimited diffusion fields as three operators $P_{\Omega^{(j)}}^{-\frac{1}{2}}W_j^{\frac{1}{2}}S_j$ for $j=1,2,3$. We provided a unified theory to show that their composition with the map $\pi_{ A,T}$ yields stable embeddings of $k$-bandlimited signals with high probability, as long as the number of space-time samples taken is above a threshold depending on the respective spectral graph coherence.
\begin{theorem}\label{thm:embedding_all}
Let $P_{\Omega^{(j)}},S_j,W_j$ be defined in Table \ref{tab:var_SR} with the sampling distribution $\mathbf{p}^{(j)}$.
For any $\delta,\epsilon \in (0,1)$, the following inequality
\begin{equation}\label{eqn:frame}
(1-\delta)f_{T}^2(\lambda_k)\|\mathbf{x}\|^2\leq \| P_{\Omega^{(j)}}^{-\frac{1}{2}}W_j^{\frac{1}{2}}S_j\pi_{ A,T}(\mathbf{x})\|_2^2 \leq (1+\delta)f_{T}^2(\lambda_{1})\|\mathbf{x}\|_2^2,
\end{equation}
holds with probability at least $1-\epsilon$,
provided that $m\geq \frac{3}{\delta^2}\left(\mathbf{\nu}_{j,\mathbf{p}^{(j)}}^{k,T}\right)^2\log\left(\frac{2k}{\epsilon} \right)$ for $j=1,3$, and $m_t\geq \frac{3}{\delta^2}\left(\mathbf{\nu}_{2,\mathbf{p}^{(2)}}^{k,T}(t) \right)^2\log\left(\frac{2k}{\epsilon} \right)$ for $j=2$, where $m_t$ is the number of space-time samples taken at time $t$.
\end{theorem}
\eqref{eqn:frame} represents the stable embedding of $k$-bandlimited signals into $\mathbb{R}^m$. Let us take two different vectors $\mathbf{x}_1 \neq \mathbf{x}_2$ in $\mathrm{span} (U_k)$. Notice that $\mathbf{x}_1-\mathbf{x}_2 \in \mathrm{span}(U_k)$, then $\| P_{\Omega^{(j)}}^{-\frac{1}{2}}W_j^{\frac{1}{2}}S_j\pi_{ A,T}(\mathbf{x}_1-\mathbf{x}_2)\|_2^2\geq (1-\delta)f_{T}^2(\lambda_k)\|\mathbf{x}_1-\mathbf{x}_2\|^2>0.$
\begin{remark}We make several important comments on Theorem \ref{thm:embedding_all}.
\begin{itemize}
\item In the case of $T=1$, $f_{T}^2(\lambda_{1})=f_{T}^2(\lambda_{k})=1$ and all the spectral graph weighted coherences equal to $(\nu_{\mathbf{p}}^{k})^2=\max_{1\leq i\leq n} \frac{\|U_k^{\top}\delta_i\|^2_2}{\mathbf{p}(i)}$, which coincides with \cite[ Theorem 2.2]{puy2018random}, that the embedding operator satisfies the Restricted Isometry Property (RIP). Note that for regime 1, the total number of space-time samples is $Tm$ and our analysis in Proposition \ref{prop:gcoherence} shows that
$T(\nu_{1,\mathbf{p}^{(1)}}^{k,T})^2$ is the largest among three regimes. Even in this case, due to the temporal dynamics, one would expect fewer requirements of spatial sensors than the static case: Indeed, one can verify that $ \left(\nu_{1,\mathbf{p}^{(1)}}^{k,T}\right)^2 \leq \left(\nu_{\mathbf{p}}^{k}\right)^2$.
We also refer to the numerical section (Section \ref{sec:experiments}) for more empirical evidences.
\item In practical situations, one just needs to measure the bandlimited diffusion fields by $S_j\pi_{ A,T}(\mathbf{x})$. The re-weighting by
$P_{\Omega^{(j)}}^{-\frac{1}{2}}W_j^{\frac{1}{2}}$ can be done offline.
\item The spectral graph weighted coherences $(\mathbf{\nu}_{j,\mathbf{p}^{(j)}}^{k,T})^2\geq k$ for $j=1,3$ and $\sum_{t=1}^{T}(\mathbf{\nu}_{j,\mathbf{p}^{(2)}}^{k,T})^2(t)\geq k$ (see Proposition \ref{prop:gcoherence}). It indicates that, we need to have at least $k$ space-time samples in all three regimes. Note that $k$ is also the minimum number of measurements that one must take to ensure the reconstruction of the $k$ bandlimited signals. {Furthermore, the optimal spectral graph weighted coherence for regimes 2 and 3 is $k$ and therefore is independent of the total sampling time. This indicates that, one can expect to use fewer number of spatial sensors at each time instance on average as $T$ increase without hurting too much reconstruction accuracy, we refer to Theorem \ref{thm: err_dirct}.}
\item By part~(b) of Proposition \ref{prop:gcoherence}, the sampling complexity bounds satisfy that regime 1 $\geq$ regime 3 $\geq$ regime 2. This is consistent with our numerical results in Section 6.
\end{itemize}
\end{remark}
Theorem \ref{thm:embedding_all} generalizes the well-known compressed sensing results in bounded orthonormal systems to the sampling of bandlimited diffusion fields with various space-time sampling regimes. Here we are in a different setting where the signal model is the bandlimited diffusion field. We leverage this temporal structure to connect with the orthonormal matrix $\widetilde U_{k,T}$. This connection allows us to utilize the same concentration inequalities used in compressed sensing yet refine and tighten the space-time sampling conditions.
In part $(a)$ of Proposition \ref{prop:gcoherence}, we derive $\mathbf{p}_{\textbf{opt}}^{(j)}
for regime $j$, that minimizes the corresponding $\nu_{j,\mathbf{p}^{(j)}}^{k,T}$.
Combining Theorem \ref{thm:embedding_all} and part $(a.2)$, part $(a.3)$ of Proposition \ref{prop:gcoherence}, we can see that there always exists a sampling distribution for regime 2 and regime 3 such that $\mathcal{O}(k\log k)$ space-time samples are enough to capture all $k$-bandlimited signals. For regime 1, due to the inequality in part $(a.1)$, one would expect more space-time samples than other two regimes.
\subsection{Intuitive links between graph diffusion process and the spectral graph weighted coherence}\label{graphdiffusion} We will interpret $i\in [nT]$
as a space-time node in $[n]\times [T]$. We give here some examples showing how the spectral graph weighted coherence changes for diffusion processes over different graphs.
Consider circulant graphs with periodic boundary conditions, whose adjacent matrices are real and symmetric circulant matrices. In this case, the eigenvectors of their Laplacian are the $n$ dimensional classical Fourier modes. For simplicity, assume that the top $k$ eigenvalues are distinct. Let $\lambda_1,\lambda_2,\cdots,\lambda_k$ denote the eigenvalues of evolution operator $A$ for the heat diffusion process. { In this case, one can show that $\mathbf{p_{opt}}^{(1)}$ is the uniform distribution, and $\mathbf{p_{opt}}^{(2)}$ and $\mathbf{p_{opt}}^{(3)}$ have the identical components for a fixed time instance. These conclusions come from the following two facts:}\\
(i)For regime 1,
$\left\| \widetilde{U}_{k,T}((i:n:Tn),:) \right\|_2 $ is the same for all $1\leq i\leq n$ and it equals to the operator norm of the matrix
\begin{equation*}\frac{1}{\sqrt{n}}\begin{bmatrix}1&1&\cdots&1\\ \lambda_1&\lambda_2&\cdots&\lambda_k\\ \vdots&\vdots&\cdots&\vdots\\ \lambda_1^{T-1}&\lambda_2^{T-1}&\cdots&\lambda_k^{T-1} \end{bmatrix} \begin{bmatrix} {1}/{f_T(\lambda_1)}&0&\cdots&0\\ 0& {1}/{f_T(\lambda_2)} &\cdots&0\\ \vdots&\vdots&\cdots&\vdots\\ 0 & 0 &\cdots& {1}/{f_T(\lambda_k)}\end{bmatrix}.
\end{equation*}
(ii) For regime 2 and regime 3,
$\left\| \widetilde{U}_{k,T}^\top \delta_i \right\|_2$ is the same for all node $i$ at the time instance $t$ and it equals to
$\frac{1}{\sqrt{n}}\sqrt{\sum_{j=1}^{k} {\lambda_j^{2(t-1)}}/{f_T(\lambda_j)^2}}$.
Let us consider the graph made of $k$ disconnected components of size $n_1,\cdots, n_k$ such that $\sum_{i=1}^{k}n_i=n$. Let the Laplacian $L$ be the graph Laplacian with edge weights all equal to one. Then a basis of $\mathrm{span}(U_k)$ is the concatenation of the square root of the degree operator times the indicator vectors of each component. Moreover, $\mathrm{span}(U_k)$ is the eigenspace associated to the eigenvalue 0 of $L$. So the heat diffusion process over this eigenspace is the trivial identity. So the temporal dynamics play no role in sampling and it will be the same with the static case. By calculation, for the node $i$ at component $j$, we have
\begin{equation*}
\left\| \widetilde{U}_{k,T}((i:n:Tn),:) \right\|_2 =\frac{\sqrt{d_{ij}}}{\sqrt{d_j}} ~\text{ and }~
\left\| \widetilde{U}_{k,T}^\top \delta_i \right\|_2 =\frac{\sqrt{d_{ij}}}{\sqrt{d_jT}},
\end{equation*}where $d_{ij}$ is the degree of node $i$ and $d_j$ is the total degree of nodes in component $j$.
To keep things simplified, let us assume each component is regular, i.e., each node has the same degree. In this case, in $\mathbf{p_{opt}}^{(1)}$, the probability of choosing the spatial node $1\leq i \leq n$ at component $j$ is $\frac{1}{kn_j}$, coinciding with the case $T=1$. For regime 2 and 3, the optimal probability of choosing the spatial node $i$ at time $t$ is $\frac{1}{k{n_j}}$ and $\frac{1}{k{n_jT}}$ respectively. Thus, if all components have the same size, then uniform sampling is the optimal sampling. If the components have different sizes, {the smaller a component is, the larger the probability of sampling one of its nodes is}. We can see that optimal sampling requires that each component is sampled with probability $\frac{1}{k}$, independent of its size. The probability that each component is sampled at least once, a necessary condition for perfect recovery, is thus higher than that using uniform sampling. In the numerical section (Section \ref{sec:experiments}), we consider the loosely defined
community-structured graph, which is a relaxation of the strictly disconnected component example. In this case, one also expects that $\mathbf{p}_{\textbf{opt}}^{(j)}$ to sample a node is inversely proportional to the size of its community.
\section{Reconstruction of $k$-bandlimited signals from space-time samples}
\label {Main}
In this section, we are interested in designing a robust procedure to recover the initial signal in $\mathrm{span}(U_k)$ from $M$ space-time samples. Specifically,
the space-time samples $\mathbf{y}\in\mathbb{R}^{M}$ are defined as
\[\mathbf{y}=S_j(\pi_{ A,T}(\mathbf{x}))+\mathbf{e}
\]
where $\mathbf{e}\in\mathbb{R}^{M}$ models the observational noise and $S_j$ is defined in Definition \ref{def_MPW}.
When one knows a basis for $\mathrm{span}(U_k)$, we adopt the standard least squares method to estimate $\mathbf{x}\in\text{span}(U_k)$ from $\mathbf{y}$. This is done by solving
\begin{equation}\label{eqn:obj1}
{\mathbf{x}}^*=\arg\min_{\widetilde{\mathbf{x}}\in\text{span}({U}_{k})}\|P_{\Omega^{(j)}}^{-1/2}W_{j}^{\frac{1}{2}}(S_j\pi_{ A,T}(\widetilde{\mathbf{x}})-\mathbf{y})\|_2.
\end{equation}
Here we introduce the weighting matrix $P_{\Omega^{(j)}}^{-1/2}W_{j}^{\frac{1}{2}}$ in \eqref{eqn:obj1} to ensure the stable embedding property in Theorem \ref{thm:embedding_all}. The solution to \eqref{eqn:obj1} is obtained by solving
\[\mathrm{z}=\arg\min_{\widetilde{\mathrm{z}}\in\mathbb{R}^{k}}\|P_{\Omega^{(j)}}^{-1/2}W_{j}^{\frac{1}{2}}(S_j\pi_{ A,T}U_{k}(\widetilde{\mathbf{z}})-\mathbf{y})\|_2
\]
and setting ${\mathbf{x}}^*=U_k \mathbf{z}$. We generalize the idea of \cite[Theorem 3.1]{puy2018random} to prove that ${\mathbf{x}}^*$ is a faithful estimation of $\mathbf{x}$ so that \cite[Theorem 3.1]{puy2018random} becomes a special case of the following theorem:
\begin{theorem}\label{thm: err_dirct}
Let $\Omega^{(j)}$ be a sampling set according to a sampling distribution $\mathbf{p}^{(j)}$, and $S_j$ be the sampling matrix associated to $\Omega^{(j)}$. Let $\varepsilon ,\delta\in(0,1)$ and suppose that $m\geq \frac{3}{\delta^2}(\nu_{j,\mathbf{p}^{(j)}}^{k,T})^2\log(\frac{2k}{\varepsilon})$ for $j= 1, 3$ or $m_t\geq \frac{3}{\delta^2}(\nu_{2,\mathbf{p}^{(2)}}^{k,T}(t))^2\log(\frac{2k}{\varepsilon})$ for $j=2$. With probability at least $1-\varepsilon$, the following holds for all $\mathbf{x}\in\text{span}( {U}_{k})$ and all $\mathbf e\in\mathbb{R}^{M}$.
\begin{enumerate}[label=\roman*)]
\item Let ${\mathbf{x}}^*$ be the solution of Problem \eqref{eqn:obj1} with $\mathbf{y}=S_j(\pi_{ A,T}(\mathbf{x}))+\mathbf{e}$. Then
\begin{equation}\label{err:1}
\|\mathbf{x}^*-\mathbf{x}\|_2\leq \frac{2}{\sqrt{1-\delta}f_{T}(\lambda_k)}\|P_{\Omega^{(j)}}^{-1/2}W_{j}^{\frac{1}{2}}\mathbf{e}\|_2,
\end{equation}
where $f_T$ is defined in \eqref{map:f_T}
\item There exist particular vectors $\mathbf e_0\in\mathbb{R}^{M}$ such that the solution ${\mathbf{x}}^*$ of Problem \eqref{eqn:obj1} with $\mathbf{y}=S_j\pi_{ A,T}(\mathbf{x})+\mathbf e_0$ satisfies
\begin{equation}\label{err:2}
\|\mathbf{x}^*- {\mathbf{x}}\|_2\geq \frac{1}{\sqrt{1+\delta} f_{T}(\lambda_1)}\|P_{\Omega^{(j)}}^{-1/2}W_{j}^{\frac{1}{2}}\mathbf e_0\|_2.
\end{equation}
\end{enumerate}
\end{theorem}
The estimate \eqref{err:1} tells us that if $\mathbf{e}=0$, then ${\mathbf{x}}^*=\mathbf{x}$ with high probability. In the presence of noise, \eqref{err:1} shows that $\|\mathbf{x}^*- {\mathbf{x}}\|_2 $ scales linearly with $\|P_{\Omega^{(j)}}^{-1/2}W_{j}^{\frac{1}{2}}\mathbf{e}\|_2$, and the estimate \eqref{err:2} shows that the bound on $\|\mathbf{x}^*- {\mathbf{x}}\|_2 $ is sharp up to a constant. In the case of uniform sampling, we have
\begin{eqnarray}
\|P_{\Omega^{(j)}}^{-1/2}W_{j}^{\frac{1}{2}}\mathbf{e}\|_2^2&=&{\frac{nT}{M}}\|\mathbf{e}\|_2^2, j=1,3 \label{bd1}\\
\|P_{\Omega^{(2)}}^{-1/2}W_{2}^{\frac{1}{2}}\mathbf{e}\|_2^2&=& \sum_{t=0}^{T-1}\frac{n}{m_t}\|\mathbf{e}_t\|_2^2\label{bd2},
\end{eqnarray} where $\mathrm{e}_t$ denotes the projection of $\mathrm{e}$ into the components at time instance $t$.
In the case of non-uniform sampling, the noise may be amplified a lot at some particular draws of $\Omega$ due to some probability weights could be very close to 0. Non-uniform sampling could thus be very sensitive to noise unlike uniform sampling. Fortunately, this is a worst case scenario and it is unlikely to draw samples from locations where the probability is small by our sampling procedures. For each regime, the expectation $\mathbb{E}\|P_{\Omega^{(j)}}^{-1/2}W_{j}^{\frac{1}{2}}\mathbf{e}\|_2^2$ over the choice of $\Omega^{(j)}$
is as in the right hand side of \eqref{bd1} and \eqref{bd2} respectively. Therefore the noise term is not too large on average over the draw of $\Omega$.
Notice that to solve \eqref{eqn:obj1}, we need to know $U_k$ in advance, which could be computationally expensive for large scale graphs. When $U_k$ is not given, we propose to estimate $\mathbf{x}$ by solving the following {regularized least square} problem
\begin{equation}\label{eqn:obj2}
\min_{\widetilde{\mathbf{ x}}\in\mathbb{R}^{n}}\|P_{\Omega^{(j)}}^{-1/2}W_{j}^{\frac{1}{2}}(S_j\pi_{ A,T}(\widetilde{\mathbf {x}})-\mathbf{y})\|_2^2+\gamma \widetilde{\mathbf{ x}}^\top g(L) \widetilde{{\mathbf{x}}},
\end{equation}
where $\gamma>0$ and $g:\mathbb{R}\rightarrow\mathbb{R}$ is a nonnegative and nondecreasing polynomial function. One can find a solution to \eqref{eqn:obj2} by solving the following equation
\[ (\pi_{ A,T}^\top S_j^\top W_j^{\frac{1}{2}}P_{\Omega^{(j)}}^{-1}W_{j}^{\frac{1}{2}}S_j\pi_{ A,T}+\gamma g(L) ){\mathbf{x}}=\pi_{ A,T}^\top S_j^\top W_j^{\frac{1}{2}}P_{\Omega^{(j)}}^{-1}W_{j}^{\frac{1}{2}} \mathbf{y}.
\]
The idea of this relaxation comes from graph-filtering techniques which have found connections with decoders used in the semi-supervised learning on graphs \cite{puy2018random}. In our problem, we use the penalty term $ \gamma \widetilde{\mathbf {x}}^\top g(L)\widetilde{\mathbf{x}}$ to incorporate the smoothness of initial signal. The next theorem provides bounds for the error between the original signal $\mathbf{x}$ and the solution of \eqref{eqn:obj2}, where \cite[Theorem 3.2]{puy2018random} becomes a special case for $T=1$.
\begin{theorem}\label{thm:main_regular}
Let $\Omega^{(j)}$ be a sampling set which obtained according to the sampling distribution $\mathbf{p}^{(j)}$. $P_{\Omega^{(j)}}$, $S_j$, $W_j$ are provided in Table \ref{tab:var_SR} associated with $\mathbf{p}^{(j)}$.
$M_{\max}>0$ is a constant such that $\|P_{\Omega^{(j)}}^{-1/2} W_{j}^{\frac{1}{2}}S_j\|_2\leq M_{\max}$. Let $\varepsilon ,\delta\in(0,1)$ and suppose that $m\geq \frac{3}{\delta^2}(\nu_{j,\mathbf{p}^{(j)}}^{k,T})^2\log(\frac{2k}{\varepsilon})$ for $j= 1, 3$ (and $m_t\geq \frac{3}{\delta^2}\left(\mathbf{\nu}_{2,\mathbf{p}^{(2)}}^{k,T}(t) \right)^2\log(\frac{2k}{\varepsilon})$ for $j=2$). With probability at least $1-\varepsilon$, the following holds for all $\mathbf{x}\in\text{span}({U}_{k})$ and all $\mathbf e\in\mathbb{R}^{M}$, all $\gamma>0$, and all nonnegative and nondecrasing polynomial functions $g$ such that $g(\sigma_{k+1})>0$, where $\sigma_{k+1}$ is the ($k+1$)th singular value of $L$.
Let $\mathbf{x}^*$ be the solution of Problem \eqref{eqn:obj2} with $\mathbf{y}=S_j(\pi_{ A,T}(\mathbf{x}))+\mathbf e$. Then
\begin{equation}\label{eqn:thm:main_regular1}
\begin{aligned}
&\|\alpha^*-\mathbf{x}\|_2\leq \frac{1}{\sqrt{1-\delta}f_{T}(\lambda_{k})}\left(2+\frac{M_{\max}f_{T}(\lambda_{k+1}) }{\sqrt{\gamma g(\sigma_{k+1}) }}\right)\|P_{\Omega^{(j)}}^{-1/2}W_{j}^{\frac{1}{2}}\mathbf{e}\|_2+\\
&\frac{1}{\sqrt{1-\delta}f_{T}(\lambda_{k})}\left(M_{\max}f_{T}(\lambda_{k+1})\sqrt{\frac{g(\sigma_{k})}{g(\sigma_{k+1})}}+\sqrt{\gamma g(\sigma_{k})} \right)\|\mathbf{x}\|_2
\end{aligned}
\end{equation}
and
\begin{equation}\label{eqn:thm:main_regular2}
\|\beta^*\|_2\leq\frac{1}{\sqrt{\gamma g(\sigma_{k+1})}}\|P_{\Omega^{(j)}}^{-1/2}W_{j}\mathbf{e}\|_2+\left( \sqrt{\frac{g(\sigma_{k})}{g(\sigma_{k+1})}}\|\mathbf{x}\|_2 \right),
\end{equation}
where $\alpha^*:=U_kU_k^\top\mathbf{x}^*$, $\beta^*:=(I_n-U_kU_k^\top)\mathbf{x}^*$ and $f_T$ is defined in \eqref{map:f_T}.
\end{theorem}
Notice that when $T=1$, the bounds in Theorem \ref{thm:main_regular} are exactly the bounds in \cite[Theorem 3.2]{puy2018random}, due to that $f_{T}(\lambda_k)=f_{T}(\lambda_{k+1})=1$ and $W_{j}=\frac{1}{m}I_{n}$.
In the estimates, $x^*=\alpha^*+\beta^*$.
To find a bound for $\|x^*-x\|_2$, one could simply use the triangular inequality and the bounds in \eqref{eqn:thm:main_regular1}, \eqref{eqn:thm:main_regular2}.
In the absence of noise, we thus have
\begin{equation*}
\|x^*-x\|_2\leq \left(\left(\frac{M_{\max}f_{T}(\lambda_{k+1})}{\sqrt{1-\delta}f_{T}(\lambda_{k})}+1\right)\sqrt{\frac{g(\sigma_{k})}{g(\sigma_{k+1})}}+\frac{\sqrt{\gamma g(\sigma_{k})}}{\sqrt{1-\delta}f_{T}(\lambda_{k})} \right)\|\mathbf{x}\|_2.
\end{equation*}
If $g(\sigma_{k})=0$, we notice that we obtain a perfect reconstruction. Note that $g$ is supposed to be nondecreasing and nonnegative. In addition $\sigma_1\leq\cdots\leq\sigma_{n}$. Thus $g(\sigma_{k})=0$ implies that we also have $g(\sigma_{1})=\cdots=g(\sigma_{k-1})=0$. If $g(\sigma_{k})\neq 0$, the above bound shows that we should choose $\gamma$ as close as possible to $0$ and seek to minimize the ratio $\frac{g(\sigma_{k})}{g(\sigma_{k+1})}$ to minimize the upper bound on the reconstruction error. Notice that if $g(L)=L^{\ell}$ for $\ell\in\mathbb{N}^*$, then the ratio $g(\sigma_{k})/g(\sigma_{k+1})$ decreases as $\ell$ increases. Thus increasing the power of $L$ and taking $\gamma$ sufficiently small to compensate the potential growth of $g(\sigma_{k})$ is a simple solution to improve the reconstruction quality in the absence of noise. {In addition, notice that $f_T$ is an increasing function with respect to $T$ (see \eqref{map:f_T}). Thus $1/f_T(\lambda_k)$ will be decreasing as $T$ is increasing. Additionally, $\lambda_k\geq \lambda_{k+1}$, we have that $f_T(\lambda_{k+1})/f_T(\lambda_k)$ is a nonincreasing function with respect to $T$. Therefore, for fixed $g$, the error bound will be a nonincreasing function with respect to $T$. }
In the presence of noise, for a fixed function $g$, the upper bound on the reconstruction error is minimized for a value of $\gamma$ proportional to $\|P_{\Omega^{(j)}}W_j\mathbf{e}\|_2/\|\mathbf{x}\|_2$. To optimize the result further, one should seek to have $g(\sigma_{k})$ as small as possible and $g(\sigma_{k+1})$ as large as possible.
Therefore, the reconstruction of the original bandlimited graph signal can be summarized in Algorithm \ref{ALGO:reconstruction}.
\begin{algorithm}[h!]
\caption{ Procedure to reconstruct the original signal}\label{ALGO:reconstruction}
\begin{algorithmic}[1]
\STATE\textbf{Input:} { The signal evolution operator ${A}\in\mathbb{R}^{n\times n}$, the samples $\{\mathbf{y}_{t}^{(j)}\}_{t=0}^{T-1}$, sampling sets $\Omega^{(j)}$, and the probability distributions $\mathbf{p}^{(j)}$.}
\STATE Generate the sampling matrix $S_j$, the weighting matrix $W_{j}$ from the sampling set $\Omega^{(j)}$ and matrix $P_j$ from the probability distribution $\mathbf{p}^{(j)}$.
\STATE Stack the samples together to a column vector and denote it as $\mathbf{y}^{(j)}$
\IF{The support $U_k$ of $\mathbf{x}$ is known}
\STATE Set ${\mathbf{x}}^*=\arg\min\limits_{\widetilde{\mathbf{ x}}\in\text{span}({U}_{k})}\|P_{\Omega^{(j)}}^{-1/2}W_{j}^{\frac{1}{2}}(S_j\pi_{ A,T}(\widetilde{\mathbf{x}})-\mathbf{y})\|_2$ as estimate of $\mathbf{x}$.
\ELSE
\STATE Choose some constant $\gamma$ and regularizization function $g(z)$.
\STATE Set ${\mathbf{x}}^*=\arg\min\limits_{\widetilde{\mathbf{ x}}\in\mathbb{R}^{n}}\|P_{\Omega^{(j)}}^{-1/2}W_{j}^{\frac{1}{2}}(S_j\pi_{ A,T}(\widetilde{\mathbf{ x}})-\mathbf{y})\|_2^2+\gamma \widetilde{\mathbf{x}}^{\top} g(L)\widetilde{\mathbf{ x}}$ as estimate of $\mathbf{x}$.
\ENDIF
\STATE\textbf{Output: $\mathbf{x}^*$. }
\end{algorithmic}
\end{algorithm}
\section{Numerical Experiments}
\label{sec:experiments}
\begin{figure}[h]
\centering
\includegraphics[width=.18\linewidth, keepaspectratio]{figs/community.pdf}
\includegraphics[width=.18\linewidth, keepaspectratio]{figs/minnesota.pdf}
\includegraphics[width=.18\linewidth, keepaspectratio]{figs/bunny.pdf}
\includegraphics[width=.18\linewidth, keepaspectratio]{figs/sensor.pdf}
\includegraphics[width=.18\linewidth, keepaspectratio]{figs/path_graph.pdf}
\caption{\footnotesize\label{fig:graphs} The five different graphs used in the simulations: the figures are community graph, Minnesota graph, bunny graph, sensor graph and ring graph from left to right.}
\vspace{-0.1in}
\end{figure}
In this section, we empirically illustrate the theoretical findings and demonstrate the performance of the proposed algorithms.
\paragraph{Graphs} We use almost the same set of graphs in \cite{puy2018random}. It consists of {five} different types of graphs, and is all available in the GSP toolbox \cite{perraudin2014} which is presented in Fig.~\ref{fig:graphs}: a) different community-type graphs of size $V = 1000$; b)Minnesota road graph of size $V = 2642$; c) Stanford bunny graph of size $V = 2503$; d) an unweighted sensor graph of size $V = 1000$ \footnote{We replace the binary tree graph in \cite{puy2018random} with the sensor graph, since the latter is more relevant to applications considered in our paper.}; e) a path graph of size $V = 1000$. Each node
is connected to its left and right neighbours except for the
boundary nodes which only have one neighbour.
\paragraph{Heat diffusion processes} We use the normalized Laplacian $L$ and consider the continuous heat diffusion processes over the graphs in all experiments with the evolution operator $A=\exp(-\Delta t L)$. Note that in \cite{puy2018random}, the combinatorial Laplacian was used, so we also run the experiments for $T=1$ for comparison with static case, instead of using the numerical results in \cite{puy2018random}. The specific parameters are summarized in Table~\ref{tab:table-diffusion}.
\begin{table}[th]
\footnotesize
\centering
\begin{tabular}{|l|l|l|l|l|l|l|}
\hline
Type & Community & Minesota & Bunny & Sensor & Path & Real images \\ \hline
$\Delta t$ & 4 & 30 & 30 & 10 & 4 & 1 \\ \hline
$T$ & 20 & 20 & 20 & 10 & 20 & 6 \\ \hline
\end{tabular}
\caption{\label{tab:table-diffusion} Parameters in heat diffusion processes }
\vspace{-0.2in}
\end{table}
\paragraph{Space-time sampling} All space-time
samples are taken with replacement as stated in Theorem~\ref{thm:embedding_all}. The total number of space-time samples is denoted by $M$ and $m_t$ for $t=0,\cdots,T-1$ are set to be equal for regime 2.
\subsection{Effect of the space-time sampling distributions on $M$}
First we show how $\mathbf{p}^{(j)}$ on the proposed sampling regimes
affect the required total number of samples $M$ to ensure stable embedding \eqref{eqn:frame}.
We compute the lower embedding constant
\begin{equation*}
\begin{aligned}
\small
\underline{\delta}_{k,\mathbf{p}^{(j)}} :&= \min_{x\in \mathrm{span}(U_k),\|x\|=1} \| P_{\Omega^{(j)}}^{-\frac{1}{2}}W_j^{\frac{1}{2}}S_j\pi_{ A,T}(\mathbf{x})\|_2^2\\
&= \lambda_{\text{min}}(U_k^\top\pi_{ A,T}^\top S_j^\top W^{\frac{1}{2}}P_{\Omega^{(j)}}^{-1}W^{\frac{1}{2}}S_j\pi_{ A,T}U_k)
\vspace{-1mm}
\end{aligned}
\end{equation*}
for {different $M$}.
We compute $\underline{\delta}_{k,\mathbf{p}^{(j)}}$ for 250 independent draws of the matrices $S_j$s.
\subsubsection{Using community graphs}
Inspired by \cite{puy2018random}, we use three types of community graphs\footnote{The community graphs $C_1,C_2,C_3$ used in our paper correspond to $C_1,C_3,C_5$ in \cite{puy2018random}}, denoted by $C_1, C_2, C_3$, to study the effect of the size of the communities on sampling distributions. All these graphs have $10$ communities with $9$ of them of approximately equal size and the last community being reduced size. Specifically: (i) the graphs of type $C_1$ have $10$ communities of size $100$; (ii) the graphs of type $C_2$ have $1$ community of size $25$, $8$ communities of size $108$, and $1$ community of size $111$;
(iii) the graphs of type $C_3$ have $1$ community of size $13$, $8$ communities of size $109$, and $1$ community of size $115$.
\begin{figure}[th]
\centering
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} $\underline{\delta}_{10,\pi^{(1)}}$ \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} $\underline{\delta}_{10,\pi^{(2)}}$ \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} $\underline{\delta}_{10,\pi^{(3)}}$ \end{minipage}\\
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/eigs_community_uniform_p1-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/eigs_community_uniform_p2-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/eigs_community_uniform_p3-eps-converted-to.pdf}\\
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} $\underline{\delta}_{10, \mathbf{p}_{\mathrm{opt}}^{(1)}}$ \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} $\underline{\delta}_{10, \mathbf{p}_{\mathrm{opt}}^{(2)}}$ \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} $\underline{\delta}_{10, \mathbf{p}_{\mathrm{opt}}^{(3)}}$ \end{minipage}\\
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/eigs_community_optimal_p1-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/eigs_community_optimal_p2-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/eigs_community_optimal_p3-eps-converted-to.pdf}\\
\caption{\footnotesize\label{fig:eig_community} Probability that $\underline{\delta}_{10}$ is great than $0.005$ as a function of $M$ for $5$ different types of community graphs: $C_1$ in black plus, $C_2$ in blue diamond, $C_3$ in orange hexagon. We choose $ M=20:20:500$.
}
\vspace{-0.2in}
\end{figure}
\begin{figure}[th]
\centering
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Regime 1 \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Regime 2 \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Regime 3 \end{minipage}\\
\includegraphics[width=.30\linewidth,height=0.25\linewidt
]{figs/Opt_plus_Unf_coherence1_vs_time-eps-converted-to.pdf}
\includegraphics[width=.30\linewidth, height=0.25\linewidt
]{figs/Opt_plus_Unf_coherence2_vs_time-eps-converted-to.pdf}
\includegraphics[width=.30\linewidth,height=0.25\linewidt
]{figs/Opt_plus_Unf_coherence3_vs_time-eps-converted-to.pdf}\\
\caption{\footnotesize\label{fig:coherence_vs_Time} The relation between spectral graph weighted coherence of order $(k,T)$ and $T$ with $k=10$ and $T$ ranging from 1 to 20 for community graphs.
We present results for $C_1$ in black plus, $C_2$ in blue diamond, and $C_3$ in orange hexagon; uniform sampling distribution in solid curve and optimal sampling distribution in dashed curve.
For regime 1 ,2 and 3, we present the values of $(\nu_{1,\bf{p}^{(1)}}^{k,T})^2$, $\sum_{t=0}^{T-1}(\nu_{\bf{p}_t^{(2)}})^2/T$ and $(\nu_{3,\bf{p}^{(3)}}^{k,T})^2/T$, respectively. All figures show that the value are decreasing as $T$ increases, with very slow rate in regime 1 (see two representative data pairs). For the case of optimal sampling distribution, the curves in regime 2 and regime 3 are in fact $10/T$ (see the representative data pair (10,1)). }
\vspace{-0.3in}
\end{figure}
We present in Fig.~\ref{fig:eig_community} the probabilities that $\underline{\delta}_{10, \mathbf{p}^{(j)}}$ is greater than 0.005, i.e, $\mathrm{Prob}(\underline{\delta}_{10,\mathbf{p}^{(j)}} \geq 0.005)$, estimated over $250$ trials.
Let $M_{i,\mathbf{p}^{(j)}}^*$ be the required number of space-time samples to ensure $\mathrm{Prob}(\underline{\delta}_{10,\mathbf{p}^{(j)}} \geq 0.005) = 0.9$ for $i$-th graph-type and sampling distribution $\mathbf{p}^{(j)}$ of each sampling regime. Theorem~\ref{thm:embedding_all} predicts that $M_{i, \mathbf{p}^{(j)}}^*$ scales linearly with $T(\mathbf{\nu}_{1,\mathbf{p}^{(1)}}^{k,T})^2$, $\sum_{t=0}^{T-1}(\mathbf{\nu}_{2,\mathbf{p}^{(2)}}^{k,T}(t))^2$ and $(\mathbf{\nu}_{3,\mathbf{p}^{(3)}}^{k,T})^2$ for $j=1,2,3$ respectively of community graph $i$.
For the uniform distribution $\mathbf{p}:=\pi$, the first figure from the top panel in Fig.~\ref{fig:eig_community} shows that $M_{1, \mathbf{p}^{(j)}}^*\geq M_{2, \mathbf{p}^{(j)}}^* \geq M_{3, \mathbf{p}^{(j)}}^*$. This matches their ordering of $\mathbf{\nu}_{j,\mathbf{p}^{(j)}}^{k,T}$ presented in Fig.~\ref{fig:coherence_vs_Time}, and is in accordance with Theorem \ref{thm:embedding_all}. For the optimal sampling distribution $\mathbf{p}_{\mathrm{opt}}$, we have $\mathbf{\nu}_{j,\mathbf{p}_{\textbf{opt}}^{(j)}}^{k,T}$ approximately the same for all types of graphs and $j=1,2,3$. Therefore $M_{i,\mathbf{p}^{(j)}}^*$ must be nearly identical for all graph-types, as observed in the second panel of Fig.~\ref{fig:eig_community}. For each type of graph, in Proposition \ref{gcoherence}, we showed that for all distributions,
\vspace{-1.5mm}
\[\sum_{t=0}^{T-1}(\mathbf{\nu}_{2,\mathbf{p}_t^{(2)}}^{k,T})^2\leq (\mathbf{\nu}_{3,\mathbf{p}^{(3)}}^{k,T})^2 \leq T(\mathbf{\nu}_{1,\mathbf{p}^{(1)}}^{k,T})^2,\] this matches the numerical results in Fig.~\ref{fig:coherence_vs_Time}.
In addition, we also computed the spectral graph weighted coherence for $C_1$, $C_2$, $C_3$ with respect to different sampling distributions and total sampling time $T$ in Fig.~\ref{fig:coherence_vs_Time}. The results show that the number of total space-time samples required in regime 2 and regime 3 are the same as $T$ varies and thus one can expect use fewer spatial samples as $T$ increase with approximately 1-1 trade-off. While in regime 1, the temporal information only brings a tiny ratio of reduction in spatial samples.
\subsubsection{The results for other graphs} We also perform the same experiments for four other graphs that have wide applications: Minnesota, bunny, path, and sensor graphs (cf.~\cite{perraudin2014}). For the first three graphs, the experiments are performed for bandlimited signals with bandwidth $100$. For the sensor graph, we set the bandwidth $50$. Here our goal is to test the performance with respect to signals with different bandwidth. For the path graph, some eigenvalues have multiplicities larger than 1. We choose the bandwidth to ensure that $\lambda_k <\lambda_{k+1}$ as required in our assumptions. The results for $\text{Prob}(\underline{\delta}_{k,\mathbf{p}^{(i)}}>0.005)$ \textit{v.s.} $M$ are summarized in Fig.~\ref{fig:eig_other}.
\subsection{Comparisons}
\subsubsection{Optimal distribution versus uniform distribution} Compared with the uniform distribution $\pi^{(j)}$, the optimal distribution $\mathbf{p}_{\textbf{opt}}^{(j)}$ for each regime takes the graph topology and temporal information into account so that one can use fewer space-time samples yet yield robust reconstruction. In the case $k=100$, the advantage of using $\mathbf{p}_{\textbf{opt}}^{(j)}$ is most obvious for the Minnesota and bunny graphs in regime 3. For the Minnesota graph, we reach only a probability of $0.008$ at $M = 800$ with $\pi^{(j)}$, whereas $M = 400$ are sufficient to reach a probability $1$ with $\mathbf{p}_{\textbf{opt}}^{(j)}$. $\pi^{(j)}$ exhibits a poorer performance for the bunny graph: we reach only a probability of $0.016$ at $M = 1400$ with $\pi^{(j)}$, whereas $M = 400$ are sufficient to reach a probability 1 with $\mathbf{p}_{\textbf{opt}}^{(j)}$.
Uniform sampling is not working well in regime 3 for the Minnesota and bunny graphs, since there exist few eigenmodes of the bandlimited diffusion field whose energy is highly concentrated on few space-time nodes. In other words, we have $ \|\widetilde{U}_{100,T}\delta_i\| \approx 1$ for some space-time node $i$. In this case, the corresponding $\nu_{j,\mathbf{\pi}^{(3)}}^{k,T}$ is approximately $\sqrt{Tn}$. In contrast, $\nu_{3,\mathbf{p}_{\textbf{opt}}^{(3)}}^{k,T}$ is $\sqrt{k}$ for $\mathbf{p}_{\textbf{opt}}^{(j)}$. In regime 2, thanks to the the same number of spatial samples taken at consecutive times, the gap between the uniform sampling and optimal sampling is much smaller.
For more regular graphs like the sensor graph and path graph, as our analysis in Section \ref{graphdiffusion} shows, the optimal distributions are close to the uniform distributions, and therefore they have comparable performances.
\textbf{Dynamical sampling versus static sampling.} The static sampling ($T=1$) results using uniform distributions are presented in Table \ref{tab:table-uniformdistribution}. For path graph and community graph, we observed that the embedding produced by space-time samples chosen according to regime 2 and 3 are as good as the one produced by the same number of spatial samples, we even do better in community graph. For the bunny graph and Minnesota graph, if we count the average spatial samples at each time instance, then we use much fewer spatial samples than the static setting to achieve a comparable performance. From the perspective of sampling, this means that one can use fewer number of sensors and dynamical sampling provides a much cheaper alternative for data acquisition. We also refer to the captions of Figure \ref{fig:rec_nonoise} and Figure \ref{fig:rec_noisy} for the comparison
of reconstruction with the static case.
\begin{table}[h]
\footnotesize
\centering
\begin{tabular}{|l|l|l|l|l|l|l|l|l|l|l|l|l|}
\hline
Type & \multicolumn{3}{l|}{Community} & \multicolumn{3}{l|}{Minnesota} & \multicolumn{3}{l|}{Bunny} & \multicolumn{3}{l|}{Path} \\ \hline
Regime & 1 & 2 & 3 & 1 & 2 & 3 & 1 & 2 & 3 & 1 & 2 & 3 \\ \hline
$M$ & 400 & 80 & 80 & 3000 & 1600 & 1600 & 3200 & 2400 & 2600 & 3200 & 400 & 400 \\ \hline
$M/T$ & 20 & 4 & 4 & 150 & 80 & 80 & 200 & 120 & 120 & 200 & 20 & 20 \\ \hline
Static & \multicolumn{3}{l|}{130} & \multicolumn{3}{l|}{450} & \multicolumn{3}{l|}{$320$} & \multicolumn{3}{l|}{360} \\ \hline
\end{tabular}
\caption{\label{tab:table-uniformdistribution}For the uniform distributions, the number of samples that required to reach a probability at least 0.95.}
\vspace{-0.2in}
\end{table}
\begin{figure}[th]
\centering
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} $\text{Minnesota, } \underline{\delta}_{100,\mathbf{p}^{(1)}}$ \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} $\text{Minnesota, } \underline{\delta}_{100,\mathbf{p}^{(2)}}$ \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} $\text{Minnesota, } \underline{\delta}_{100,\mathbf{p}^{(3)}}$ \end{minipage}\\
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/eigs_minesota_optimal_p1-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/eigs_minesota_optimal_p2-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/eigs_minesota_optimal_p3-eps-converted-to.pdf}\\
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} $\text{Bunny, } \underline{\delta}_{100,\mathbf{p}^{(1)}}$ \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} $\text{Bunny, } \underline{\delta}_{100,\mathbf{p}^{(2)}}$ \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} $\text{Bunny, } \underline{\delta}_{100,\mathbf{p}^{(3)}}$ \end{minipage}\\
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/eigs_bunny_optimal_p1-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/eigs_bunny_optimal_p2-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/eigs_bunny_optimal_p3-eps-converted-to.pdf}\\
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} $\text{Sensor, } \underline{\delta}_{50,\mathbf{p}^{(1)}}$ \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} $\text{Sensor, } \underline{\delta}_{50,\mathbf{p}^{(2)}}$ \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} $\text{Sensor, } \underline{\delta}_{50,\mathbf{p}^{(3)}}$ \end{minipage}\\
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/eigs_sensor_optimal_p1-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/eigs_sensor_optimal_p2-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/eigs_sensor_optimal_p3-eps-converted-to.pdf}\\
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} $\text{Path, } \underline{\delta}_{100,\mathbf{p}^{(1)}}$ \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} $\text{Path, } \underline{\delta}_{100,\mathbf{p}^{(2)}}$ \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} $\text{Path, } \underline{\delta}_{100,\mathbf{p}^{(3)}}$ \end{minipage}\\
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/eigs_path_optimal_p1-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/eigs_path_optimal_p2-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/eigs_path_optimal_p3-eps-converted-to.pdf}\\
\caption{\footnotesize\label{fig:eig_other} Probability that $\underline{\delta}_{10}$ is great than $0.005$ as a function of $M$.
The panels show the results at $k=100$ for the Minnesota, bunny and path graphs, and at $k=50$ for the sensor graph using optimal and uniform distributions. The gap for $M$ is 200. }
\vspace{-0.2in}
\end{figure}
\textbf{Comparison of sampling regimes.} Let $M_{\mathbf{p}^{(j)}}^*$ denote the required number of space-time samples in regime $j$ such that $\underline{\delta}_{k,\mathbf{p}^{(j)}}> 0.005$. We have found that $M_{\mathbf{p}^{(1)}}^*$ is largest for all graphs used in the experiments.
And regime 2 has the best performance in terms of the sample complexity. That is to say, we have $ M_{\mathbf{p}^{(2)}}^* \leq M_{ \mathbf{p}^{(3)}}^*\leq M_{ \mathbf{p}^{(1)}}^*$,
which indicates the relationship
$T(\nu_{1,\mathbf{p}^{(1)}}^{K,T})^2 \geq (\nu_{3,\mathbf{p}^{(3)}}^{K,T})^2\geq \sum_{t=0}^{T-1}(\nu_{2,\mathbf{p}^{(2,t)}}^{K,T})^2 $ for uniform and optimal distributions. These observations coincide with the respective coherence and are consistent with Theorem~\ref{thm:embedding_all} and the analysis in Proposition \ref{prop:gcoherence}. %
\subsection{Reconstruction of bandlimited signals}
In this part, we examine the performance of the reconstruction of initial signals from space-time samples on community graphs of type $C_3$, the Minnesota graphs and bunny graphs. We consider the recovery of $k$-bandlimited signals with $k=10$, in contrast with reconstruction results for static sampling setting in \cite{puy2018random}.
We choose $M = 200$ for $\mathbf{p}_{\textbf{opt}}^{(j)}$ with $j=1,2,3$.
The experiments are performed with and without noise on the sampled data. In the presence of noise, the noise vector $\mathbf{e}$ is random with i.i.d. mean 0 and variance $\sigma^2$ Gaussian entries \footnote{For nodes sampled multiple times, the realization of the noise is thus different each time for the same sampled node. Thus, the noise vector contains no duplicated entry.}.
$\sigma$ are chosen from $\{0, 1.5 \times 10^{-3}, 3.7 \times 10^{-3}, 8.8 \times 10^{-3}, 2.1 \times 10^{-2}, 5.0 \times 10^{-2} \}$. The signals are reconstructed by solving \eqref{eqn:obj2} for different values of the
parameter $\gamma$ and different functions $g$. For the community graph and the bunny graph, the regularisation parameter $\gamma$ varies between $10^{-3}$ and $10^2$. For the Minnesota graph, it varies between $10^{-1}$ and $10^{5}$. {For each $\sigma$, $10$ independent random signals of unit norm are drawn, sampled and reconstructed with all
possible $(\gamma, g)$}. Then the mean errors\footnote{See Theorem \ref{thm:main_regular} for the definition of ${\alpha}^*$ and ${\beta}^*$.} $\|x^* - x\|_2$, $\|{\alpha}^* - x\|_2$ and $\|{\beta}^*\|_2$ over these $10$ signals are computed.
\begin{figure}[]
\vspace{-0.1in}
\centering
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Regime 1 \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Regime 2 \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Regime3 \end{minipage}\\
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Community \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Community \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Community \end{minipage}\\
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/rec_community_uniform_p1-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/rec_community_uniform_p4-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/rec_community_uniform_p7-eps-converted-to.pdf}\\
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Bunny \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Bunny \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Bunny \end{minipage}\\
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/rec_bunny_uniform_p1-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/rec_bunny_uniform_p4-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/rec_bunny_uniform_p7-eps-converted-to.pdf}\\
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Minesota \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Minesota \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Minesota \end{minipage}\\
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/rec_minesota_uniform_p1-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/rec_minesota_uniform_p4-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth, height=0.25\linewidt
]{figs/rec_minesota_uniform_p7-eps-converted-to.pdf}\\
\caption{\footnotesize Mean reconstruction errors $\log_{10}\|x-x^*\|$ of 10 bandlimited signals as a function of $\gamma$ for the community graph $C_3$, Bunny graph and Minnesota graph. We take $M=200$ noise free space-time samples for three regimes. The black dash curve represents results with $g(L)=L$. The blue dotted curves indicate the results with $g(L)=L^2$. The red dashed-dot curves indicate the results with $g(L)=L^4$. The green solid curve indicate the results with $g(L)=\exp(I-L)$. The first, second and third columns show the results for regime 1, regime 2 and regime 3 using the optimal sampling distribution respectively. We refer to the reconstruction errors of $\|x-\alpha^*\|$ and $\|\beta^*\|$ in log scale to {the supplementary information (SI)} section \ref{addnum} (see their definitions in Theorem \ref{thm:main_regular}). Compared to the static case (Figure \ref{fig:rec_nonoise_static} in SI), our regime 1 achieved comparable (cf. community graph) or slightly better (cf. bunny and Minesota graphs) performance, and we observed significantly improvement in regime 2 and regime 3.
}\label{fig:rec_nonoise}
\vspace{-0.3in}
\end{figure}
The mean reconstruction errors are presented in Fig.~\ref{fig:rec_nonoise} when the measurements are noise-free.
In these experiments, the signals are reconstructed by setting $g(L) = L, L^2, L^4, \exp(I-L)$.
Let's recall that the ratio $g(\sigma_{k})/g(\sigma_{k+1})$ decreases as the power of $L$ increases. We observe that all reconstruction errors, $\|x^* - x\|_2$, $\|{\alpha}^* - x\|_2$ and $ \|{{\beta}^*}\|_2$ decrease when the ratio $g(\sigma_{k})/g(\sigma_{k+1})$ {decreases} in the range of small $\gamma$ (cf.~plots in SI section~\ref{addnum}), as predicted by the upper bounds in Theorem~\ref{thm:main_regular}. Additionally, Compared to the static case (Figure \ref{fig:rec_nonoise_static} in SI), our regime 1 achieved comparable (even better) performance, and we observed significantly improvement in regime 2 and 3.
We present the mean reconstruction errors when the measurements are noisy
in Fig.~\ref{fig:rec_noisy}. In these experiments, we reconstruct the signals using $g(L) = L^4$. As expected the best regularisation parameter $\gamma$ increases with the noise level. Compared with static sampling setting (cf. Fig~\ref{fig:rec_static} in SI section \ref{addnum}), our estimates in regime 2 and regime 3 achieve comparable accuracy. The estimates obtained in Regime 1 did not perform as well as other regimes, as indicated by our analysis: one needs to have more space-time samples due to the large spectral graph weighted coherence.
\begin{figure}[!t]
\centering
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Regime 1 \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Regime 2 \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Regime3 \end{minipage}\\
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Community \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Community \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Community \end{minipage}\\
\includegraphics[width=.3\linewidth,height=0.25\linewidth]{figs/noise_rec_community_uniform_p1-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth,height=0.25\linewidth]{figs/noise_rec_community_uniform_p4-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth,height=0.25\linewidth]{figs/noise_rec_community_uniform_p7-eps-converted-to.pdf}\\
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Bunny \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Bunny \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Bunny \end{minipage}\\
\includegraphics[width=.3\linewidth,height=0.25\linewidth]{figs/noise_bunny_uniform_p1-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth,height=0.25\linewidth]{figs/noise_bunny_uniform_p4-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth,height=0.25\linewidth]{figs/noise_bunny_uniform_p7-eps-converted-to.pdf}\\
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Minesota \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Minesota \end{minipage}
\begin{minipage}{.3\linewidth} \centering \small \hspace{2mm} Minesota \end{minipage}\\
\includegraphics[width=.3\linewidth,height=0.25\linewidth]{figs/noise_rec_minesota_uniform_p1-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth,height=0.25\linewidth]{figs/noise_rec_minesota_uniform_p4-eps-converted-to.pdf}
\includegraphics[width=.3\linewidth,height=0.25\linewidth]{figs/noise_rec_minesota_uniform_p7-eps-converted-to.pdf}\\
\caption{\footnotesize Mean reconstruction error $\log_{10}\|x^* - x\|_2$ of $10$-bandlimited signals as a function of $\gamma$ with $g(L) = L^4$. The simulations are performed in presence of noise. The standard deviation of the noise is $0.0015$ (blue plus), $0.0037$ (red square), $0.0088$ (black left-pointing triangle), $0.0210$ (green right-pointing triangle), $0.0500$ (cyan pentagram). The best reconstruction errors are indicated by orange circles. The first, second and third columns show the results for a community graph of type $C_5$, and the Minnesota graph, respectively. The reconstruction accuracy achieved in regime 2 and regime 3 are satisfactory, compared with the static sampling results summarized in Fig.~\ref{fig:rec_static} in SI, since we only use 10 spatial samples on average. For bunny graph, we may need to choose bigger $\gamma$, and we expect that the optimal error will also be as comparable with the static case in Fig.~\ref{fig:rec_static}. }\label{fig:rec_noisy}
\vspace{-0.2in}
\end{figure}
\subsection{Illustration: space-time sampling of a series of blurring real images}
\begin{figure}[h]
\vspace{-0.1in}
\centering
\subfloat{\includegraphics[width=.6\linewidth]{figs/evolved_images_jiuzhaigou_tau6_sample7200.pdf}}
\caption{\footnotesize The evolved images of the Jiuzhaigou over the first 6 time instances.
} \label{fig:evolved_images}
\vspace{-0.2in}
\end{figure}
The experimental section is finished with an example of de-blurring an image from the space-time samples. We use the photo of Jiuzhaigou park in Sichuan, China. The original photo is (approximately) bandlimited in a nearest neighborhood graph. We consider the heat diffusion process over the graph so we have a series of blurring images. We illustrate the dynamics in Fig.~\ref{fig:evolved_images}.
This RGB image contains $640\times 640$ pixels. We divide this image into patches of $8 \times 8$ pixels, thus obtaining $80\times 80$ patches of $64$ pixels per RGB channel. Each patch is denoted by $\mathbf{q}_{i,j,\ell} \in\mathbb{R} ^{64}$
with $i \in\{1,\cdots,80\}$, $j \in\{ 1,\cdots,80\}$, and $\ell \in \{1,2,3\}$. The pair of indices $(i, j)$ encodes the spatial location of the patch and $\ell$ encodes the color channel. Using these patches, we obtain the following matrix
\[X:=\begin{bmatrix}
\mathbf{q}_{1,1,1} & \mathbf{q}_{2,1,1} &\cdots&\mathbf{q}_{2,1,1} &\cdots &\mathbf{q}_{80,80,1}\\
\mathbf{q}_{1,1,2} & \mathbf{q}_{2,1,2} &\cdots&\mathbf{q}_{2,1,2} &\cdots &\mathbf{q}_{80,80,2}\\
\mathbf{q}_{1,1,3} & \mathbf{q}_{2,1,3} &\cdots&\mathbf{q}_{2,1,3} &\cdots &\mathbf{q}_{80,80,3}
\end{bmatrix}\in\mathbb{R}^{192\times n}, n=6400.
\]
Each column of $X$ represents a color patch of the original image at a given position.
Motivated by the simulations in \cite{puy2018random}, we build a graph by modelling the similarity between the columns of $X$. Let $\mathbf{x}_i:=X(:,i) \in\mathbb{R}^{192}$.
For each $i$, we search for the $20$ nearest
neighbours of $\mathbf{x}_i$ among all other columns of $X$. Let $\mathbf{x}_j \in\mathbb{R}^{192}$ be a vector connected to $\mathbf{x}_i$. The weight $W_{ij}$ of the weighted adjacency matrix $W \in\mathbb{R}^{n\times n}$ satisfies
\begin{equation*}
W_{ij}:=\exp\left(-\frac{\|\mathbf{x}_i-\mathbf{x}_j\|_2^2}{2\sigma^2} \right),
\end{equation*}
where $\sigma> 0$ is the standard deviation of all Euclidean distances between pairs of connected columns/patches. We then symmetrize $W$. Each column of $X$ is thus connected to at least $20$ other columns after symmetrization. The construction of the graph is finished by computing the normalized Laplacian $L \in\mathbb{R}^{n\times n}$ associated to $W$. Set the evolution operator $A=\exp(-L)$. We let $A$ act on $X^\top$ iteratively with $T=6$ and we treat each column of $X^\top$ a bandlimited graph signal with bandwidth 300. For this experiment, we take $M=7200$ measurements using the uniform and optimal distributions
for these three sampling regimes and denote the samples as matrix $Y$ which means the samples at each time instance is less than $20\%$ for regime 1 and 2.
The sampled images are presented in Fig.~\ref{fig:samples}, where all non-sampled pixels are in black.
The image is reconstructed by solving \eqref{eqn:obj2} for each column of Y with $\gamma = 10^{-5}$ and $g(L) = L^4$. Fig.~\ref{fig:real_image} shows that a very accurate reconstruction of the original image is obtained.
The SNRs between the original and the reconstructed images are presented in Table \ref{tab:SNRs}. The optimal sampling distribution allows a better image reconstruction quality than the results of the uniform sampling distribution. Additionally, the reconstructions from sampling regimes 2 and 3 have no big difference but both are better than the results from sampling regime 1. The reconstruction results from space-time samples show that incorporating the temporal dynamics is very effective in reducing the spatial sampling density, and one achieved much better performance than the static setting with 4 times higher spatial sampling densities.
\begin{figure}[h]
\vspace{-0.08in}
\centering
\subfloat{\includegraphics[width=.3\linewidth]{figs/uniform_regime1_samples_jiuzhaigou_tau6_sample7200.pdf}}\hspace{0.005in}
\subfloat{\includegraphics[width=.3\linewidth]{figs/uniform_regime2_samples_jiuzhaigou_tau6_sample7200.pdf}}\hspace{0.005in}
\subfloat{\includegraphics[width=.3\linewidth]{figs/uniform_regime3_samples_jiuzhaigou_tau6_sample7200.pdf}}\\
\vspace{-0.1in}
\subfloat[Regime 1]{\includegraphics[width=.3\linewidth]{figs/opt_regime1_samples_jiuzhaigou_tau6_sample7200.pdf}}\hspace{0.005in}
\subfloat[Regime 2]{\includegraphics[width=.3\linewidth]{figs/opt_regime2_samples_jiuzhaigou_tau6_sample7200.pdf}}\hspace{0.005in}
\subfloat[Regime 3]{\includegraphics[width=.3\linewidth]{figs/opt_regime3_samples_jiuzhaigou_tau6_sample7200.pdf}}
\caption{\footnotesize
Samples of the real image by using these three sampling regimes:
\textbf{Top row:} uniform sampling. \textbf{Bottom row:} optimal sampling.
} \label{fig:samples}
\vspace{-0.1in}
\end{figure}
\begin{figure}[h]
\centering
\subfloat{\includegraphics[width=.2\linewidth]{figs/Jiuzhaigou.pdf}} \hspace{0.005in}
\subfloat{\includegraphics[width=.2\linewidth]{figs/uniform_regime1_rec_jiuzhaigou_tau6_sample7200_1e-05.pdf}}\hspace{0.005in}
\subfloat{\includegraphics[width=.2\linewidth]{figs/uniform_regime2_rec_jiuzhaigou_tau6_sample7200_1e-05.pdf}}\hspace{0.005in}
\subfloat{\includegraphics[width=.2\linewidth]{figs/uniform_regime3_rec_jiuzhaigou_tau6_sample7200_1e-05.pdf}}
\vspace{-.1in}
\subfloat[Original]{\includegraphics[width=.2\linewidth]{figs/Jiuzhaigou.pdf}} \hspace{0.005in}
\subfloat[Regime 1]{\includegraphics[width=.2\linewidth]{figs/opt_regime1_rec_jiuzhaigou_tau6_sample7200_1e-05.pdf}}\hspace{0.005in}
\subfloat[Regime 2]{\includegraphics[width=.2\linewidth]{figs/opt_regime2_rec_jiuzhaigou_tau6_sample7200_1e-05.pdf}}\hspace{0.005in}
\subfloat[Regime 3]{\includegraphics[width=.2\linewidth]{figs/opt_regime3_rec_jiuzhaigou_tau6_sample7200_1e-05.pdf}}
\caption{\footnotesize
Comparisons of the reconstruction of the real image by using these three sampling regimes with regularizer $\gamma=10^{-5}$:
\textbf{Top row:} Uniform sampling. \textbf{Bottom row:} Optimal sampling.
} \label{fig:real_image}
\vspace{-0.1in}
\end{figure}
\begin{table}[!h]
\caption{\footnotesize
The SNR between the original and the reconstructed image: The number of samples per time instance for regime 1,2,3 is 1200 and the total number of samples for static setting is 4800. }\label{tab:SNRs}
\centering
\begin{tabular}{ |c||c|c|c|c|}
\hline
~ & \multicolumn{4}{c|}{ $\gamma =10^{-5} $} \cr
\cline{2-5}
~ & Regime 1 & Regime 2 & Regime 3 & Static \cr
\hhline {|=||=|=|=|=|}
\textbf{Uniform} &$22.59$ & $ 27.92$ & $28.14$ & $21.47$ \cr
\hline
\textbf{Optimal} &$22.93$ & $30.18$ & $30.87$ & $21.37$ \cr
\hline
\end{tabular}
\end{table}
\bibliographystyle{siamplain}
|
\section{Introduction}
\label{sec:intro}
\subsection{Probability Measures}
This paper studies probability measures using Bessel generating functions. However, an important approach used is to view the Bessel generating functions as formal power series. We are given a positive real number $\theta$ where $\beta=2\theta$. Here, $\beta=1$, $2$, and $4$ correspond to the GUE, GOE, and GSE, respectively. The multivariate Bessel function $B_{(a_1,\ldots,a_N)}(x_1,\ldots,x_N;\theta)$ for $a_1<\cdots<a_N$ is based on the $\theta$-corners process, also known as the Gaussian $\beta$-corners process, with fixed top row $(a_1,\ldots,a_N)$ (\cite{crystalrandommat}, \cite{matrix}), see \Cref{def:multivariate}. For positive integers $N$, let $\mathcal{M}_N$ denote the set of Borel probability measures on ordered $N$-tuples $(a_1, \ldots, a_N)$ in $\mathbb{R}^N$ such that $a_1\leq \cdots\leq a_N$. With $\theta$, for $\mu\in\mathcal{M}_N$, the Bessel generating function $G_\theta(x_1,\ldots,x_N;\mu)$ is
\[
G_\theta(x_1,\ldots,x_N;\mu)=\int_{a_1\leq\cdots\leq a_N} B_{(a_1,\ldots,a_N)}(x_1,\ldots,x_N;\theta) \mu(da_1,\ldots,da_N),
\]
see \Cref{def:BGF}.
For a sequence $\{\mu_N\}_{N\geq 1}$ of probability measures such that $\mu_N\in\mathcal{M}_N$ for $N\geq 1$, for $k\geq 1$, we let the random variable $p_k^N$ be
\begin{equation}
\label{eq:moments}
p_k^N = \frac{1}{N}\sum_{i=1}^N \left(\frac{a_i}{N}\right)^k,
\end{equation}
where $(a_1\leq \cdots\leq a_N)$ are distributed according to $\mu_N$. Note that in $p_k^N$, the $a_i$ are scaled by $\frac{1}{N}$. With this, the moments $\{m_k\}_{k\geq 1}$ of $\{\mu_N\}_{N\geq 1}$ are given by $m_k=\lim_{N\rightarrow\infty}\mathbb{E}_{\mu_N}(p_k^N)$ for $k\geq 1$. In this paper, we look at the Law of Large Numbers considering the $p_k^N$ and moments $\{m_k\}_{k\geq 1}$. Also, note that in later results in the paper we look at sequences $\{\mu_N\}_{N\geq 1}$ of probability measures satisfying certain conditions, which we discuss later.
\begin{definition}
\label{def:LLN}
A sequence $\{\mu_N\}_{N\geq 1}$ of probability measures such that $\mu_N\in\mathcal{M}_N$ for $N\geq 1$ \textit{satisfies a Law of Large Numbers} if there exists a sequence $\{m_k\}_{k\geq 1}$ of real numbers such that
\begin{equation}
\label{eq:llnsatisfactory}
\lim_{N\rightarrow\infty} \mathbb{E}_{\mu_N}\left(\prod_{i=1}^s p_{k_i}^N\right) = \prod_{i=1}^s m_{k_i}
\end{equation}
for any positive integer $s$ and positive integers $k_i$, $1\leq i\leq s$.
\end{definition}
In the above definition, $\{m_k\}_{k\geq 1}$ are the moments. Later on in this paper, we consider $\{\mu_N\}_{N\geq 1}$ such that $\mu_N\in\mathcal{M}_N$ is exponentially decaying. The condition for $\mu\in\mathcal{M}_N$ being exponentially decaying is below.
\begin{definition}
\label{def:decay}
A probability measure $\mu$ in $\mathcal{M}_N$ is \textit{exponentially decaying} with exponent $R>0$ if
\[\int_{a_1\leq a_2\leq \cdots\leq a_N} e^{NR\max_{1\leq i\leq N}|a_i|} \mu(da_1, \ldots, da_N)\]
is finite.
\end{definition}
\begin{lemma}[\cite{matrix}*{Lemma 2.9}]
\label{lemma:convergence}
If $\mu\in\mathcal{M}_N$ is exponentially decaying with exponent $R>0$, then $G_\theta(x_1, \ldots, x_N; \mu)$ converges for all $(x_1, \ldots, x_N)$ in the domain
\[\Omega_R = \{(x_1, \ldots, x_N)\in\mathbb{C}^N | |\text{Re}(x_i)|<R, 1\leq i\leq N\},\]
and is holomorphic in the domain.
\end{lemma}
A reason we consider $\mu\in\mathcal{M}_N$ which are exponentially decaying is that $G_\theta(x_1,\ldots,x_N;\mu)$ is convergent and holomorphic in a neighborhood of the origin, see \Cref{lemma:convergence}. Soon, we will be able to state the main result of this paper, \Cref{lln}.
\begin{definition}
\label{def:orderedtuple} For an ordered $N$-tuple $a=(a_1, a_2, \ldots, a_N)$ of nonnegative integers, let the $\textit{equivalent partition}$ of $a$, denoted by $\pi(a)$, be the partition obtained from removing the $a_i$ which are $0$ and ordering the remaining $a_i$ in nonincreasing order from left to right.
\end{definition}
For a partition $\nu=(a_1\geq a_2\geq\cdots\geq a_m)$, let $P(\nu)$ denote the number of distinct permutations of $(a_1, a_2, \ldots, a_m)$. Also, for a finite ordered list $S$ of positive integers with maximum element $M$, suppose that $n_i$ of the elements of $S$ are $i$ for $1\leq i\leq M$. Then, if $n=(n_1, n_2, \ldots, n_M)$, let $\sigma(S)$ be $\pi(n)$.
\begin{theorem}
\label{lln}
Suppose $\theta$ is a positive real number. Let $\{\mu_N\}_{N\geq 1}$ be a sequence of probability measures such that for all $N\geq 1$, $\mu_N$ is in $\mathcal{M}_N$ and is exponentially decaying. Assume that for all partitions $\nu$ with $|\nu|\geq 1$, there exists a real number $c_\nu$ such that
\begin{equation}
\label{eq:llnconditions}
\displaystyle\lim_{N\rightarrow\infty}\frac{1}{N}\cdot\frac{\partial}{\partial x_{i_1}}\cdots\frac{\partial}{\partial x_{i_r}} \ln(G_\theta(x_1,\ldots,x_N;\mu_N))\bigg|_{x_i=0,1\leq i\leq N} = \frac{\ell(\nu)(|\nu|-1)!c_\nu}{P(\nu)}
\end{equation}
for all positive integers $i_1, \ldots, i_r$ such that $\sigma((i_1,\ldots,i_r))=\nu$. Then, $\{\mu_N\}_{N\geq 1}$ satisfy a LLN, and for positive integers $k$,
\begin{equation}
\label{eq:llnmoments}
m_k = \sum_{\pi\in NC(k)} \prod_{B\in \pi}\theta^{|B|-1}\left(\sum_{\nu\in P, |\nu|=|B|}(-1)^{\ell(\nu)-1} P(\nu) c_\nu\right).
\end{equation}
\end{theorem}
\Cref{lln} generalizes \cite{matrix}*{Claim 9.1} and is similar to, but differs in important places with, \cite{matrix}*{Theorem 3.8}, where the reverse direction is also shown. Particularly, in \cite{matrix}*{Theorem 3.8}, $\theta N\rightarrow\gamma$ as $N\rightarrow\infty$ and the partial derivatives are not scaled; but, in this paper, $\theta$ is constant and the partial derivatives are scaled by $\frac{1}{N}$. However, similarly, the moments and LLN satisfaction are obtained from the value at $x_i=0,1\leq i\leq N$ after Dunkl operators are used on the Bessel generating functions, see \Cref{sec:measures}.
Additionally, this paper considers when the limit of any partial derivative can be nonzero. On the other hand, in \cite{matrix}, the limits of partial derivatives with two or more distinct indices are $0$. This is a reason why formal power series are needed to prove \Cref{lln}. Note that in \Cref{lln}, the free cumulant of order $k$ for $k\geq 1$ would be
\[
c_k = \theta^{k-1}\sum_{\nu\in P, |\nu|=k}(-1)^{\ell(\nu)-1} P(\nu) c_\nu,
\]
where free cumulants are discussed in \cite{freeprobability}. We see that $c_k$ is a linear combination of the limits of the order $k$ partial derivatives.
\subsection{Formal Power Series}
Suppose we have variables $x_1,\ldots,x_N$ and $\vec{x}=(x_1,\ldots,x_N)$. A formal power series $F(x_1,\ldots,x_N)$ has form
\begin{equation}
\label{eq:formalseries}
F(x_1,\ldots,x_N)=\sum_{\alpha=(\alpha_1,\ldots,\alpha_N)\in\mathbb{Z}_{\geq 0}^N} c_F^\alpha \prod_{i=1}^N x_i^{\alpha_i},
\end{equation}
where the $c_F^\alpha$ are constants. In the paper, we consider formal power series $F(x_1,\ldots,x_N)$ which are symmetrical, or symmetrical in $N-1$ variables. Also, for simplicity, we usually call a formal power series a formal series instead.
In \Cref{sec:measures}, we see that we can view $\ln(G_\theta(x_1,\ldots,x_N;\mu))$ as a symmetrical polynomial, and afterwards, we can use formal series. In order to find the value at $x_i=0$, $1\leq i\leq N$, after using Dunkl operators on the Bessel generating function, we look at the constant terms of formal series. Particularly, in \pref{eq:formalseries}, we have that the constant term of $F(x_1,\ldots,x_N)$ is $[1] F(x_1,\ldots,x_N) = c_F^0$, where $0$ is $\alpha$ with $\alpha_i=0$, $1\leq i\leq N$.
\subsection{Paper Organization}
The organization of this paper is as follows. In \Cref{sec:operators}, we introduce the notation for formal power series and Dunkl operators. Next, in \Cref{sec:formal}, we go over results on the limits of operators on formal series with $N$ variables which are symmetric in $N-1$ variables as $N$ increases to $\infty$. Afterwards, in \Cref{sec:main}, we prove results important to proving \Cref{lln}, and in \Cref{sec:measures}, we prove \Cref{lln}. Following this, in \Cref{sec:eigenvalue},
we apply \Cref{lln} to the $\beta$-Hermite ensemble. Next, in \Cref{appendix:a}, we state and prove several combinatorial formulas which are used throughout the paper. Finally, in \Cref{appendix:b}, we state additional results for formal series.
\textbf{Acknowledgements.} This is a continuation of research done in MIT UROP+. The author would like to thank Matthew Nicoletti for being the mentor for the research and suggesting the problem. Also, the author would like to thank Alexei Borodin and Vadim Gorin for giving comments on the paper.
\section{Formal Power Series and Operators}
\label{sec:operators}
\subsection{Formal Power Series}
Later on in the paper, we look at formal power series in $x_1,\ldots,x_N$. For $\vec{x}=(x_1,\ldots, x_N)$ and a partition $\nu$, let
\[
M_{\nu}(\vec{x})=\sum_{\substack{a=(a_1, \ldots, a_N), \\ a_i\in \mathbb{Z}_{\geq 0}, 1\leq i\leq N, \pi(a)=\nu}} \prod_{i=1}^N x_i^{a_i}.
\]
Above, $N=\infty$ is possible. Also, suppose that the set of all partitions is $P$.
For a symmetric formal series $F(x_1,\ldots, x_N)$, we let
\[
F(x_1,\ldots,x_N) = \sum_{\nu\in P, \ell(\nu)\leq N} c_F^\nu M_\nu(\vec{x}),
\]
where the $c_F^\nu\in\mathbb{C}$ are constants. Later on, in \Cref{sec:main}, when we view the $c_F^\nu$ as variables, $c_F^\nu$ has degree $|\nu|$ for $\nu\in P$.
For positive integers $N\geq i$, let $\mathcal{F}_i^N$ denote the set of formal series $F(x_1,\ldots,x_N)$ which are symmetrical in $x_j$ for $1\leq j\leq N$, $j\not=i$. Suppose that $F(x_1,\ldots,x_N)\in \mathcal{F}_i^N$. Then, where $\vec{x}_{i}=(x_j)_{\substack{1\leq j\leq N, j\not= i}}$ is $\vec{x}$ with $x_i$ removed, we have
\[
F(x_1,\ldots,x_N) = \sum_{d=0}^\infty\left(\sum_{\nu\in P} c_F^{d,\nu} M_\nu(\vec{x}_{i})\right)x_i^d,
\]
where the $c_F^{d,\nu}\in\mathbb{C}$ are constants.
We can let $c_F^\nu=0$ for $\nu$ such that $\ell(\nu)\geq N+1$, since $\vec{x}$ has $N$ variables, and $c_F^{d,\nu}=0$ for $\nu$ such that $\ell(\nu)\geq N$, since $\vec{x}_i$ has $N-1$ variables. Particularly, for $F(x_1,\ldots,x_N)\in\mathcal{F}_i^N$,
\begin{equation}
\label{eq:formalseriescoeff}
F(x_1,\ldots,x_N) = \sum_{d=0}^\infty\left(\sum_{\nu\in P, \ell(\nu)\leq N-1} c_F^{d,\nu} M_\nu(\vec{x}_{i})\right)x_i^d.
\end{equation}
\begin{definition}
For a sequence of constants $s=\{c^{d,\nu}\}_{d\geq 0, \nu\in P}$ and a positive integer $i$, let the $\textit{formal series}$ of $s$ in $\mathcal{F}_i^N$ for all $N\geq i$ be
\[
F_i(s)(x_1,\ldots,x_N) = \sum_{d=0}^\infty\sum_{\nu\in P, \ell(\nu)\leq N-1} c^{d,\nu} M_\nu(\vec{x}_i)x_i^d.
\]
Also, where $\vec{x}_i=(x_j)_{j\geq 1, j\not=i}$, let
\[
F_i(s) = \sum_{d=0}^\infty\sum_{\nu\in P} c^{d,\nu} M_\nu(\vec{x}_i)x_i^d
\]
as a formal series in $(x_j)_{j\geq 1}$.
\end{definition}
\subsection{Dunkl Operator}
For $N$ variables $x_1, \ldots, x_N$, let $s_{i,j}$ be the operator which switches $x_i$ and $x_j$.
\begin{definition}
\label{def:operators}
The \textit{Dunkl operators} are, for $1\leq i\leq N$,
\[
\mathcal{D}_i:=\frac{\partial}{\partial x_i}+\theta\sum_{\substack{1\leq j\leq N, \\ j\not=i}}\frac{1}{x_i-x_j}(1-s_{i,j}).
\]
\end{definition}
For positive integers $k$, let
\begin{equation}
\label{eq:operator1}
\mathcal{P}_k=\sum_{i=1}^N\mathcal{D}_i^k.
\end{equation}
Also, let $\frac{\partial}{\partial x_i}$ be denoted by $\partial_i$ for $1\leq i\leq N$. Moreover, for $i\not= j$, let
\[
\frac{\theta}{x_i-x_j}(1-s_{i,j})
\]
be called the \textit{switch} from $i$ to $j$. We know that the Dunkl operators are commutative, which is in \Cref{prop:commutative} below. Due to this, the $\mathcal{P}_i$ are commutative as well.
\begin{proposition}[\cite{dunkloperators}*{Theorem 1.9}]
\label{prop:commutative}
For positive integers $i,j$, $1\leq i, j\leq N$, $\mathcal{D}_i\mathcal{D}_j=\mathcal{D}_j\mathcal{D}_i$.
\end{proposition}
\section{Limits of Operators On Formal Power Series}
\label{sec:formal}
In this section, we look at operators on formal power series in $x_j$, $1\leq j\leq N$ such that there exists $i$, $1\leq i\leq N$, for which the formal power series is symmetrical for $j\not=i$. The main results are \Cref{value} and \Cref{finalvalue1}, with \Cref{value} being a special case of \Cref{finalvalue1}.
\subsection{Basic Results}
For a partition $\nu=(a_1\geq a_2\geq\cdots\geq a_m$, let
\[
S(\nu)=\{(p_1, p_2)|p_i=(b_{i,1}, \ldots, b_{i,m}), 1\leq i\leq 2, b_{1,j}+b_{2,j}=a_j, 1\leq j\leq m\}.
\]
If we have partitions $\nu_1, \nu_2, \ldots, \nu_k$, we let $\nu=\nu_1+\cdots+\nu_k$ be the partition formed when $\nu_1, \ldots,\nu_k$ are joined to each other.
For $1\leq i,j\leq N$, $i\not=j$, let $C_{i,j}$ be the operator such that for nonnegative integers $a_k$, $1\leq k\leq N$,
\begin{equation}
\label{eq:change}
C_{i,j}\left(\prod_{k=1}^N x_k^{a_k}\right) = \begin{cases}
x_i^{a_j-1}\displaystyle\prod_{1\leq k\leq N, k\not= i,j} x_k^{a_k} & \text{if $a_i=0$ and $a_j\geq 1$}, \\
0 & \text{ if $a_i\geq 1$ or $a_i=a_j=0$}.
\end{cases}
\end{equation}
With this, for positive integers $N\geq i$ and a formal series $f(x_1,\ldots,x_N)\in \mathcal{F}_i^N$, let the operator $\mathcal{Q}_i^N(f(x_1,\ldots,x_N))$ be
\begin{equation}
\label{eq:operator2}
\mathcal{Q}_i^N(f(x_1,\ldots,x_N)) =\theta\sum_{\substack{1\leq j\leq N, \\ j\not= i}} \frac{d_i-C_{i,j}}{N}+f(x_1, \ldots, x_N).
\end{equation}
In \Cref{operatorpoly}, we show that $\mathcal{Q}_i^N(f(x_1,\ldots,x_N))$ is an operator from $\mathcal{F}_i^N$ to $\mathcal{F}_i^N$. Sometimes, for $f(x_1,\ldots,x_N)\in\mathcal{F}_i^N$, we denote $\mathcal{Q}_i^N(f(x_1,\ldots,x_N))$ by $\mathcal{Q}_i^N(f)$, and for $s=\{c^{d,\nu}\}_{d\geq 0,\nu\in P}$, we denote $\mathcal{Q}_i^N(F_i(s)(x_1,\ldots,x_N))$ by $\mathcal{Q}_i^N(F_i(s))$. In this paper, for operators $\mathcal{T}_i$, $1\leq i\leq m$, the product
\[\prod_{i=1}^m \mathcal{T}_i \]
denotes the $\mathcal{T}_i$ being applied from $i=1$ to $m$.
\begin{definition}
For a positive integer $i$, a sequence of formal series $\{f_N(x_1,\ldots,x_N)\}_{N\geq i}$ is symmetric $\textit{outside}$ of $i$ if $f_N(x_1,\ldots,x_N)\in \mathcal{F}_i^N$ for all $N\geq i$.
\end{definition}
\begin{definition}
\label{def:limitingsequence}
Suppose that $i$ is a positive integer and a sequence of formal series $f=\{f_N(x_1,\ldots,x_N)\}_{N\geq i}$ is symmetrical outside of $i$. Then, the $\textit{limit}$ of $f$ outside of $i$ exists if $\lim_{N\rightarrow\infty} c_{f_N}^{d,\nu}$ exists for all $d, \nu$. If the limit outside of $f$ outside of $i$ exists, the $\textit{limiting sequence}$ of $f$ outside of $i$ is $s=\{\lim_{N\rightarrow\infty} c_{f_N}^{d,\nu}\}_{d\geq 0, \nu\in P}$, and the limit of $f$ as $N\rightarrow\infty$ outside of $i$ is
\[\lim_{N\rightarrow\infty} f_N(x_1,\ldots,x_N) = F_i(s) = \sum_{d=0}^\infty\left(\sum_{\nu\in P} \left(\lim_{N\rightarrow\infty} c_{f_N}^{d,\nu}\right)\cdot M_\nu(\vec{x}_{i})\right)x_i^d.\]
\end{definition}
\begin{remark}
For a sequence $\{f_N(x_1,\ldots,x_N)\}_{N\geq i}$ such that $f_N(x_1,\ldots,x_N)$ is in $\mathcal{F}_i^N$ for all $N\geq i$, the limiting sequence of $f_N(x_1,\ldots,x_N)$ outside of $i$ being $s$ is equivalent to the limit of $f_N(x_1,\ldots,x_N)$ as $N\rightarrow\infty$ outside of $i$ being $F_i(s)$.
\end{remark}
\begin{proposition}
\label{operatorpoly}
For $f(x_1,\ldots, x_N)$ and $g(x_1,\ldots, x_N)$ in $\mathcal{F}_i^N$, $\mathcal{Q}_i^N(f)g(x_1,\ldots,x_N)$ is in $\mathcal{F}_i^N$, and is
\begin{align*}
& \sum_{d=0}^\infty \sum_{\nu\in P, \ell(\nu)\leq N-1}\left(\theta c_g^{d+1, \nu}-\theta c_g^{0, \nu + (d+1)} + \sum_{\substack{a+b=d, \\ (p_1, p_2)\in S(\nu)}} c_f^{a,\pi(p_1)}c_g^{b,\pi(p_2)}\right)M_{\nu}(\vec{x}_i)x_i^d \\
& +\frac{\theta}{N}\cdot\sum_{d=0}^\infty \sum_{\nu\in P, \ell(\nu)\leq N-1}(-c_g^{d+1,\nu}+(\ell(\nu)+1)c_g^{0,\nu+(d+1)})M_\nu(\vec{x}_i)x_i^d.
\end{align*}
\end{proposition}
\begin{proof}
We see that
\[
g_1(x_1, \ldots, x_N)=d_ig(x_1, \ldots, x_N) = \sum_{d=1}^\infty \sum_{\nu\in P} c_{g}^{d,\nu}M_{\nu}(\vec{x}_{i})x_i^{d-1}=\sum_{d=0}^\infty \sum_{\nu\in P} c_{g}^{d+1,\nu}M_{\nu}(\vec{x}_{i})x_i^d.
\]
Also, let
\[
g_2(x_1, \ldots, x_N)=\sum_{1\leq j\leq N, j\not= i} C_{i,j} g(x_1, \ldots, x_N) = \left(\sum_{1\leq j\leq N, j\not= i} C_{i,j}\right)\left(\sum_{\nu\in P} c_{g}^{0, \nu} M_\nu(\vec{x}_i)\right).
\]
Observe that $g_2(x_1, \ldots, x_N)$ is a formal power series which is symmetric in $\vec{x}_i$. We find the coefficient of $M_\nu(\vec{x}_i)x_i^d$. If $\ell(\nu)\geq N$, the coefficient is $0$ in $g_2(x_1,\ldots,x_N)$. Then, suppose $\ell(\nu)\leq N-1$. For a monomial $p$ in $M_\nu(\vec{x}_i)$ with variables $(\vec{x}_i)_j$ with $1\leq j\leq \ell(\nu)$, there are $N-1-\ell(\nu)$ $j$ such that $x_j$ is not in $p$, and these are the $j$ such that $C_{i,j}$ applied to a monomial will give $px_i^d$. For a monomial $q$, if
\[C_{i,j}q=px_i^d,\]
then $q=px_j^{d+1}$ and has coefficient $c_{g}^{0, \nu+(d+1)}$ in $g$. Therefore, the coefficient of $px_i^d$, and thus $M_\nu(\vec{x}_i)x_i^d$, in $g_2(x_1, \ldots, x_N)$ is $(N-\ell(\nu)-1)c_g^{0, \nu+(d+1)}$. From this,
\[g_2(x_1, \ldots, x_N)=\sum_{d=0}^\infty\sum_{\nu\in P,\ell(\nu)\leq N-1} (N-\ell(\nu)-1) c_g^{0, \nu+(d+1)}M_\nu(\vec{x}_i)x_i^d.\]
Next, let
\begin{align*}
g_3(x_1, \ldots, x_N) & = f(x_1, \ldots, x_N)g(x_1, \ldots, x_N) \\
& = \sum_{d_1, d_2=0}^\infty\sum_{\nu_1, \nu_2\in P} c_f^{d_1, \nu_1}c_g^{d_2, \nu_2}M_{\nu_1}(\vec{x}_i)M_{\nu_2}(\vec{x}_i)x_i^{d_1+d_2}.
\end{align*}
For $\nu\in P$ where $\nu$ can be $0$, we find the coefficient of $M_\nu(\vec{x}_i)x_i^d$ in $g_3(x_1,\ldots,x_N)$. If $\ell(\nu)\geq N$, this coefficient will be $0$. Suppose $\ell(\nu)\leq N-1$, and let
\[
q = \prod_{j=1}^{\ell(\nu)} x_{(\vec{x}_i)_j}^{\nu_j}.
\]
Note that if $q_1$ and $q_2$ are monic monomials such that $q_1q_2=q$, then for $1\leq s\leq 2$, if $b_{s,j}$ is the degree of $x_{(\vec{x}_i)_j}$ in $q_s$ for $1\leq j\leq \ell(\nu)$ and $p(q_s)=(b_{s,1}, \ldots, b_{s, \ell(\nu)})$, $(p(q_1), p(q_2))\in S(\nu)$. We see that if $S$ is the set of $(q_1,q_2)$ such that $q_1$ and $q_2$ are monic monomials with $q_1q_2=q$,
\[p:S\rightarrow S(\nu), (q_1,q_2)\mapsto (p(q_1),p(q_2))\]
is injective as well as surjective, and therefore is a bijection. Also, the coefficient of $q_1x_1^a$ and $q_2x_1^b$ is $c_f^{a, \pi(p(q_1))}$ in $f$ and $c_g^{b, \pi(p(q_2))}$ in $g$, respectively. Then, the coefficient of $qx_i^d$ is
\[
\sum_{\substack{a+b = d, \\ q_1q_2=q}} c_f^{a, \pi(p(q_1))}c_g^{b, \pi(p(q_2))}=\sum_{\substack{a+b=d, \\ (p_1, p_2)\in S(\nu)}} c_f^{a, \pi(p_1)}c_g^{b, \pi(p_2)},
\]
which is also the coefficient of $M_\nu(\vec{x}_i)x_i^d$ in $g_3(x_1,\ldots,x_N)$. From this,
\[
g_3(x_1,\ldots,x_N) = \sum_{d=0}^\infty\sum_{\nu\in P, \ell(\nu)\leq N-1}\left(\sum_{\substack{a+b=d, \\ (p_1,p_2)\in S(\nu)}} c_f^{a,\pi(p_1)}c_g^{b,\pi(p_2)}\right)M_\nu(\vec{x}_i)x_i^d.
\]
Since $g_1(x_1,\ldots, x_N)$, $g_2(x_1,\ldots,x_N)$, and $g_3(x_1,\ldots,x_N)$ are in $\mathcal{F}_i^N$, $\mathcal{Q}_i(f)g(x_1,\ldots,x_N)$ is in $\mathcal{F}_i^N$, with
\[
\mathcal{Q}_i(f)g(x_1,\ldots,x_N) = \frac{(N-1)\theta g_1(x_1,\ldots x_N)}{N}-\frac{\theta g_2(x_1,\ldots,x_N)}{N}+g_3(x_1,\ldots,x_N).
\]
With the expressions we obtained, this completes the proof.
\end{proof}
\begin{corollary}
\label{limitexists}
Let $m\geq 0$ be an integer. Suppose $\{f_{j,N}(x_1,\ldots,x_N)\}_{N\geq i}$ for $1\leq j\leq m$ and $\{g_N(x_1,\ldots,x_N)\}_{N\geq i}$ are symmetrical outside of $i$. Also, assume that the limit of $f_{j,N}(x_1, \ldots, x_N)$ for $1\leq j\leq m$ and $g_N(x_1, \ldots, x_N)$ as $N\rightarrow\infty$ exist outside of $i$. Then,
\[
\lim_{N\rightarrow\infty} \left(\prod_{j=1}^m\mathcal{Q}_i^N(f_{j,N})\right) g_N(x_1, \ldots, x_N)
\]
exists outside of $i$. Also, the limit of $\mathcal{Q}_i^N(f_{1,N})g_N(x_1,\ldots,x_N)$ as $N\rightarrow\infty$ outside of $i$ is equal to
\begin{equation}
\label{eq:limitoperator}
\sum_{d=0}^\infty \sum_{\nu\in P} \lim_{N\rightarrow\infty}\left(\theta c_{g_N}^{d+1, \nu}-\theta c_{g_N}^{0, \nu + (d+1)} + \sum_{\substack{a+b=d, \\ (p_1, p_2)\in S(\nu)}} c_{f_{1,N}}^{a,\pi(p_1)}c_{g_N}^{b,\pi(p_2)}\right)M_\nu(\vec{x}_{i})x_i^d.
\end{equation}
\end{corollary}
\begin{proof}
The $m=0$ case is clear, and for $m=1$, the limit of $\mathcal{Q}_i^N(f_{1,N})g_N(x_1,\ldots,x_N)$ as $N\rightarrow\infty$ outside of $i$ existing and being equal to \pref{eq:limitoperator} follows from taking the $N\rightarrow\infty$ limit of \Cref{operatorpoly}. Then, the rest of \Cref{limitexists} follows from $m=1$ and induction.
\end{proof}
\begin{proposition}
\label{limit}
Let $i\geq 1$ and $m\geq 0$ be integers. Suppose $\{f_{j,N}(x_1,\ldots,x_N)\}_{N\geq 1}$ for $1\leq j\leq m$ and $\{g_N(x_1,\ldots,x_N)\}_{N\geq 1}$ are symmetrical outside of $i$. Assume that for $1\leq j\leq m$, $f_{j,N}(x_1, \ldots, x_N)$ has a limit as $N\rightarrow\infty$ with limiting sequence $f_j$ outside of $i$. Also, assume that $g_N(x_1, \ldots, x_N)$ has a limit as $N\rightarrow\infty$ with limiting sequence $g$ outside of $i$. Then, outside of $i$,
\begin{equation}
\label{eq:limitsequal}
\lim_{N\rightarrow\infty} \left(\prod_{j=1}^m\mathcal{Q}_i^N(f_{j,N})\right) g_N(x_1, \ldots, x_N) = \lim_{N\rightarrow\infty} \left(\prod_{j=1}^m\mathcal{Q}_i^N(F_i(f_j))\right) F_i(g)(x_1, \ldots, x_N).
\end{equation}
\end{proposition}
\begin{proof}
From \Cref{limitexists}, the limits in both sides exist. Suppose that for $1\leq j\leq m$, $f_j=\{c_{f_j}^{d,\nu}\}_{d\geq 0, \nu\in P}$, and $g=\{c_g^{d,\nu}\}_{d\geq 0, \nu\in P}$. We use induction on $m$, where the base case $m=0$ is clear. For $m=1$, we can use \pref{eq:limitoperator} to show that $\lim_{N\rightarrow\infty} \mathcal{Q}_i^N(f_{1,N})g_N(x_1,\ldots,x_N)=\lim_{N\rightarrow\infty} \mathcal{Q}_i^N(F(f_1))F(g)(x_1,\ldots,x_N)$.
Assume the statement holds for $m\geq 1$. We want to show the statement holds for $m+1$. Let
\[r_N(x_1, \ldots, x_N) = \mathcal{Q}_{i}^N(f_{1,N}) g_N(x_1, \ldots, x_N),\]
and $r$ be the limiting sequence of $\mathcal{Q}_{i}^N(F_i(f_1)) F_i(g)(x_1, \ldots, x_N)$ as $N\rightarrow\infty$ outside of $i$. From $m=1$, $\lim_{N\rightarrow\infty} r_N(x_1, \ldots, x_N) = \lim_{N\rightarrow\infty} \mathcal{Q}_{i}^N(F_i(f_1)) F_i(g)(x_1, \ldots, x_N) = F_i(r)$, or $r$ is the limiting sequence of $r_N(x_1,\ldots,x_N)$ as $N\rightarrow\infty$ outside of $i$. Then, by the inductive hypothesis,
\begin{align*}
& \lim_{N\rightarrow\infty} \left(\prod_{j=1}^{m+1}\mathcal{Q}_i^N(f_{j,N})\right) g_N(x_1, \ldots, x_N) = \lim_{N\rightarrow\infty} \left(\prod_{j=2}^{m+1}\mathcal{Q}_i^N(f_{j,N})\right) r_N(x_1, \ldots, x_N) \\
& = \lim_{N\rightarrow\infty} \left(\prod_{j=2}^{m+1}\mathcal{Q}_i^N(F_i(f_j))\right) F_i(r)(x_1, \ldots, x_N)
\\
& = \lim_{N\rightarrow\infty} \left(\prod_{j=2}^{m+1}\mathcal{Q}_i^N(F_i(f_j))\right) \mathcal{Q}_i^N(F_i(f_1))F_i(g)(x_1, \ldots, x_N) \\
& = \lim_{N\rightarrow\infty} \left(\prod_{j=1}^{m+1}\mathcal{Q}_i^N(F_i(f_j))\right) F_i(g)(x_1, \ldots, x_N).
\end{align*}
This completes the proof.
\end{proof}
\begin{proposition}
\label{prop:commutativeswitch}
Suppose that $i, j$ are integers such that $1\leq i,j\leq N$ and $i\not= j$. Then, for integers $k$ with $1\leq k \leq N$ and $k\not= i,j$, if $f\in\mathcal{F}_k^N$,
\[
s_{i,j}\mathcal{Q}_k^N(f) = \mathcal{Q}_k^N(f) s_{i,j}
\]
as operators from $\mathcal{F}_k^N$ to $\mathcal{F}_k^N$. Also, if $f\in\mathcal{F}_i^N$,
\[s_{i,j}\mathcal{Q}_i^N(f) = \mathcal{Q}_j^N(s_{i,j}f) s_{i,j}\]
as operators from $\mathcal{F}_i^N$ to $\mathcal{F}_j^N$.
\end{proposition}
\begin{proof} This is clear from expanding the operators.
\end{proof}
\subsection{Constant Term}
The most important part of the formal series after the $\mathcal{Q}_i^N$ operators is the constant term, as will be seen in \Cref{sec:main}. For a sequence of coefficients $s=\{c^{d,\nu}\}_{d\geq 0, \nu\in P}$, this corresponds to $c^{0,0}$. In this subsection, we look at \Cref{value} to compute the constant term following $\mathcal{Q}_i^N$ operators where there is only one $i$. Later on, in \Cref{finalvalue1}, we look at the constant term following $\mathcal{Q}_i^N$ operators where there are any number of distinct $i$. These values are computed with free cumulants, which are in \Cref{def:cumulants}. However, we must go over noncrossing partitions first.
Suppose that we have a partition $\pi$ of a finite, nonempty set $S$ of real numbers. We let
\begin{equation}
\label{eq:partitionblocks}
\pi = B_1\sqcup B_2\sqcup\cdots\sqcup B_m,
\end{equation}
where $B_i$, $1\leq i\leq m$, are the blocks of $\pi$, such that the smallest element of $B_{i+1}$ is greater than the smallest element of $B_i$ for $1\leq i\leq m-1$. Also, the length of $\pi$ is $\ell(\pi)=m$.
A partition $\pi$ is $\textit{noncrossing}$ if for any distinct blocks $B_1$ and $B_2$ of $\pi$, there do not exist $a,b\in B_1$ and $c,d\in B_2$ such that $a<c<b<d$. Let the set of noncrossing partitions of a finite, nonempty set $S$ of real numbers be $NC(S)$, and for $k\geq 1$, let $NC(k)=NC(\{1,\ldots,k\})$. A way to represent a partition $\pi$ is the $\textit{circular representation}$, where the elements are spaced around a circle in order and the convex hulls of the elements of each block of $\pi$ are added. If $\pi$ is noncrossing, the convex hulls will be disjoint.
\begin{figure}[!h]
\centering
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=1.1cm,y=1.1cm]
\clip(-2,-1.1) rectangle (10,3.3);
\fill[line width=1.5pt,fill=black,fill opacity=0.3] (-0.43969262078590776,2.493620766483186) -- (0.5,2.8356409098088537) -- (1.,0.) -- (0.,0.) -- cycle;
\fill[line width=1.5pt,fill=black,fill opacity=0.3] (1.4396926207859084,2.4936207664831853) -- (1.939692620785908,1.6275953626987465) -- (1.7660444431189777,0.642787609686539) -- cycle;
\node at (0.5,-0.9) {\normalsize Noncrossing partition};
\draw [line width=1.5pt] (0.5,1.3737387097273115) circle (1.4619022000815423*1.1cm);
\draw [line width=1.5pt] (-0.939692620785908,1.6275953626987478)-- (-0.7660444431189781,0.6427876096865397);
\draw [line width=1.5pt] (-0.43969262078590776,2.493620766483186)-- (0.5,2.8356409098088537);
\draw [line width=1.5pt] (0.5,2.8356409098088537)-- (1.,0.);
\draw [line width=1.5pt] (1.,0.)-- (0.,0.);
\draw [line width=1.5pt] (0.,0.)-- (-0.43969262078590776,2.493620766483186);
\draw [line width=1.5pt] (1.4396926207859084,2.4936207664831853)-- (1.939692620785908,1.6275953626987465);
\draw [line width=1.5pt] (1.939692620785908,1.6275953626987465)-- (1.7660444431189777,0.642787609686539);
\draw [line width=1.5pt] (1.7660444431189777,0.642787609686539)-- (1.4396926207859084,2.4936207664831853);
\begin{scriptsize}
\draw [fill=black] (0.,0.) circle (2.pt);
\draw[color=black] (-0.11563625798620362,-0.3177080076873236) node {1};
\draw [fill=black] (1.,0.) circle (2.pt);
\draw[color=black] (1.1156362579862036,-0.3177080076873239) node {2};
\draw [fill=black] (1.7660444431189777,0.642787609686539) circle (2.pt);
\draw[color=black] (2.058845726811989,0.47373870972731047) node {3};
\draw [fill=black] (1.939692620785908,1.6275953626987465) circle (2.pt);
\draw[color=black] (2.272653955421975,1.6863054295277848) node {4};
\draw [fill=black] (1.4396926207859084,2.4936207664831853) circle (2.pt);
\draw[color=black] (1.657017697435771,2.752618707341471) node {5};
\draw [fill=black] (0.5,2.8356409098088537) circle (2.pt);
\draw[color=black] (0.5,3.173738709727311) node {6};
\draw [fill=black] (-0.43969262078590776,2.493620766483186) circle (2.pt);
\draw[color=black] (-0.6570176974357707,2.7526187073414725) node {7};
\draw [fill=black] (-0.939692620785908,1.6275953626987478) circle (2.pt);
\draw[color=black] (-1.2726539554219745,1.6863054295277862) node {8};
\draw [fill=black] (-0.7660444431189781,0.6427876096865397) circle (2.pt);
\draw[color=black] (-1.0588457268119897,0.4737387097273116) node {9};
\fill[line width=1.5pt,fill=black,fill opacity=0.3] (8.439692620785909,2.4936207664831853) -- (8.766044443118978,0.642787609686539) -- (8.,0.) -- cycle;
\fill[line width=1.5pt,fill=black,fill opacity=0.3] (6.560307379214092,2.493620766483186)-- (6.233955556881022,0.6427876096865397) -- (6.060307379214092,1.6275953626987478) -- cycle;
\node at (7.5,-0.9) {\normalsize Crossing partition};
\draw [line width=1.5pt] (7.5,1.373738709727307) circle (1.4619022000815467*1.1cm);
\draw [line width=1.5pt] (7.,0.)-- (8.939692620785908,1.6275953626987465);
\draw [line width=1.5pt] (6.560307379214092,2.493620766483186)-- (6.233955556881022,0.6427876096865397);
\draw [line width=1.5pt] (6.233955556881022,0.6427876096865397)--(6.060307379214092,1.6275953626987478);
\draw [line width=1.5pt] (6.560307379214092,2.493620766483186)--(6.060307379214092,1.6275953626987478);
\draw [line width=1.5pt] (8.439692620785909,2.4936207664831853)-- (8.766044443118978,0.642787609686539);
\draw [line width=1.5pt] (8.766044443118978,0.642787609686539)-- (8.,0.);
\draw [line width=1.5pt] (8.,0.)-- (8.439692620785909,2.4936207664831853);
\draw [fill=black] (7.,0.) circle (2.pt);
\draw[color=black] (7-0.11563625798620362,-0.3177080076873236) node {1};
\draw [fill=black] (8.,0.) circle (2.pt);
\draw[color=black] (8.1156362579862036,-0.3177080076873239) node {2};
\draw [fill=black] (8.7660444431189777,0.642787609686539) circle (2.pt);
\draw[color=black] (9.058845726811989,0.47373870972731047) node {3};
\draw [fill=black] (8.939692620785908,1.6275953626987465) circle (2.pt);
\draw[color=black] (9.272653955421975,1.6863054295277848) node {4};
\draw [fill=black] (8.4396926207859084,2.4936207664831853) circle (2.pt);
\draw[color=black] (8.657017697435771,2.752618707341471) node {5};
\draw [fill=black] (7.5,2.8356409098088537) circle (2.pt);
\draw[color=black] (7.5,3.173738709727311) node {6};
\draw [fill=black] (7-0.43969262078590776,2.493620766483186) circle (2.pt);
\draw[color=black] (7-0.6570176974357707,2.7526187073414725) node {7};
\draw [fill=black] (7-0.939692620785908,1.6275953626987478) circle (2.pt);
\draw[color=black] (7-1.2726539554219745,1.6863054295277862) node {8};
\draw [fill=black] (7-0.7660444431189781,0.6427876096865397) circle (2.pt);
\draw[color=black] (7-1.0588457268119897,0.4737387097273116) node {9};
\end{scriptsize}
\end{tikzpicture}
\caption{Circular representations of a noncrossing and crossing partition of $\{1,2, \ldots, 9\}.$}
\end{figure}
\begin{definition}
\label{def:cumulants}
For a sequence of coefficients $s=\{c^{d,\nu}\}_{d\geq 0, \nu\in P}$ and a positive integer $k$, let the $\textit{free cumulant}$ of order $k$ for $s$ be
\begin{equation}
\label{eq:cumulants}
c_k(s) = \theta^{k-1}\sum_{\substack{\nu\in P, d\geq 0,\\ |\nu|+d=k-1}} (-1)^{\ell(\nu)}P(\nu)c^{d, \nu}.
\end{equation}
\end{definition}
A way the free cumulants appear is in \pref{eq:limitcumulants}. Note that these free cumulants will be used in \Cref{sec:main} and \Cref{sec:measures} to obtain the free cumulants mentioned previously for the moments $m_k$ in \Cref{lln}.
\begin{theorem}
\label{value}
Let $i$ and $k$ be positive integers. Suppose $\{f_{j,N}(x_1, \ldots, x_N)\}_{N\geq i}$ for $1\leq j\leq k-1$ and $\{g_N(x_1,\ldots, x_N)\}_{N\geq i}$ are sequences of formal series which are symmetric outside of $i$. For $1\leq j\leq k$, assume that $f_{j,N}(x_1,\ldots,x_N)$ has a limit as $N\rightarrow\infty$ with limiting sequence $f$ outside of $i$. Also, suppose $g_N(x_1,\ldots, x_N)$ has a limit as $N\rightarrow\infty$ with limiting sequence $g$ outside of $i$. Then,
\begin{equation}
\label{eq:limitcumulants}
\lim_{N\rightarrow\infty}\left([1]\prod_{j=1}^{k-1}\mathcal{Q}_i^N(f_{j,N})g_N(x_1, \ldots, x_N)\right) = \sum_{\substack{\pi\in NC(k), \\ \pi = B_1\sqcup\cdots\sqcup B_m}} c_{|B_1|}(g)\prod_{i=2}^m c_{|B_i|}(f).
\end{equation}
\end{theorem}
\begin{proof}
From \Cref{limitexists}, the limit of $\prod_{j=1}^{k-1}\mathcal{Q}_i^N(f_{j,N})g_N(x_1, \ldots, x_N)$ as $N\rightarrow\infty$ outside of $i$ exists, and the left hand side of \pref{eq:limitcumulants} is
\[
[1]\lim_{N\rightarrow\infty}\prod_{j=1}^{k-1}\mathcal{Q}_i^N(f_{j,N})g_N(x_1, \ldots, x_N),
\]
We can prove the Theorem with induction. For the base case $k=1$,
\[
\lim_{N\rightarrow\infty} [1] g_N(x_1, \ldots, x_N)= c_g^{0,0} = c_1(g).
\]
Next, assume that the statement holds for a positve integer $k$. We want to show the statement holds for $k+1$. From \Cref{limit}, outside of $i$,
\[\lim_{N\rightarrow\infty}\prod_{j=1}^k\mathcal{Q}_i^N(f_{j,N})g_N(x_1, \ldots, x_N)=\lim_{N\rightarrow\infty}\mathcal{Q}_i^N(F_i(f))^kF_i(g)(x_1, \ldots, x_N).\]
Using \Cref{limitexists}, let $g'$ be the limiting sequence of $\mathcal{Q}_i^N(F_i(f)) F_i(g)(x_1, \ldots, x_N)$ outside of $i$, with
\[
c_{g'}^{d, \nu} = \theta c_g^{d+1, \nu}-\theta c_g^{0, \nu + (d+1)}+\sum_{\substack{a+b=d, \\ (p_1, p_2)\in S(\nu)}} c_f^{a, \pi(p_1)}c_g^{b, \pi(p_2)}
\]
for $d\geq 0$ and $\nu\in P$. By \Cref{limit},
\[
\lim_{N\rightarrow\infty}\mathcal{Q}_i^N(F_i(f))^kF_i(g)(x_1, \ldots, x_N) = \lim_{N\rightarrow\infty}\mathcal{Q}_i^N(F_i(f))^{k-1}F_i(g')(x_1, \ldots, x_N).
\]
However, by the inductive hypothesis,
\begin{align*}
& [1]\lim_{N\rightarrow\infty}\mathcal{Q}_i^N(F_i(f))^{k-1}F_i(g')(x_1, \ldots, x_N) = \sum_{\substack{\pi\in NC(k), \\ \pi = B_1\sqcup\cdots\sqcup B_m}} c_{|B_1|}(g')\prod_{i=2}^m c_{|B_i|}(f) \\
& = \sum_{\substack{\pi\in NC(k), \\ \pi = B_1\sqcup\cdots\sqcup B_m}} \theta^{|B_1|-1}\left(\sum_{\substack{\nu\in P, d\geq 0 \\ |\nu|+d=|B_1|-1}} (-1)^{\ell(\nu)}P(\nu)c_{g'}^{d, \nu}\right)\prod_{i=2}^m c_{|B_i|}(f).
\end{align*}
Also, from \Cref{noncrossingsum},
\begin{align*}
&\sum_{\substack{\pi\in NC(k+1), \\ \pi = B_1\sqcup\cdots\sqcup B_m}} c_{|B_1|}(g)\prod_{i=2}^m c_{|B_i|}(f) \\
& = \sum_{\substack{\pi\in NC(k), \\ \pi = B_1\sqcup\cdots\sqcup B_m}} \left(c_{|B_1|+1}(g)+\sum_{j=1}^{|B_1|} c_j(f)c_{|B_1|+1-j}(g)\right) \prod_{i=2}^m c_{|B_i|}(f).
\end{align*}
Then, for $k+1$, it suffices to show that
\[
\theta^{|B_1|-1}\sum_{\substack{\nu\in P, d\geq 0 \\ |\nu|+d=|B_1|-1}} (-1)^{\ell(\nu)}c_{g'}^{d, \nu}P(\nu) = c_{|B_1|+1}(g)+\sum_{j=1}^{|B_1|} c_j(f)c_{|B_1|+1-j}(g).
\]
The left hand side is
\begin{align*}
& \theta^{|B_1|}\sum_{\substack{\nu\in P, d\geq 0 \\ |\nu|+d=|B_1|-1}}(-1)^{\ell(\nu)}P(\nu)c_g^{d+1, \nu}-\theta^{|B_1|}\sum_{\substack{\nu\in P, d\geq 0 \\ |\nu|+d=|B_1|-1}}(-1)^{\ell(\nu)}P(\nu)c_g^{0, \nu + (d+1)} \\
& + \theta^{|B_1|-1}\sum_{\substack{\nu\in P, d\geq 0 \\ |\nu|+d=|B_1|-1}}\sum_{\substack{a+b=d, \\ (p_1, p_2)\in S(\nu)}} (-1)^{\ell(\nu)}P(\nu)c_f^{a, \pi(p_1)}c_g^{b, \pi(p_2)}.
\end{align*}
However, the right hand side is
\begin{align*}
& \theta^{|B_1|}\sum_{\substack{\nu\in P, d\geq 0,\\ |\nu|+d=|B_1|}} (-1)^{\ell(\nu)}P(\nu)c_g^{d, \nu}\\
& +\theta^{|B_1|-1}\sum_{j=1}^{|B_1|}\left(\sum_{\substack{\nu_1\in P, d_1\geq 0,\\ |\nu_1|+d_1=j-1}} (-1)^{\ell(\nu_1)}P(\nu_1)c_f^{d_1, \nu_1}\right)\left(\sum_{\substack{\nu_2\in P, d_2\geq 0,\\ |\nu_2|+d_2=|B_1|-j}} (-1)^{\ell(\nu_2)}P(\nu_2)c_g^{d_2, \nu_2}\right).
\end{align*}
Observe that
\[
\sum_{\substack{\nu\in P, d\geq 0,\\ |\nu|+d=|B_1|}} (-1)^{\ell(\nu)}P(\nu)c_g^{d, \nu}
= \sum_{\substack{\nu\in P, d\geq 0, \\ |\nu|+d=|B_1|-1}}(-1)^{\ell(\nu)}P(\nu)c_g^{d+1, \nu} + \sum_{\nu\in P, |\nu|=|B_1|} (-1)^{\ell(\nu)}P(\nu)c_g^{0, \nu}.
\]
Here, we want to show that
\[
\sum_{\nu\in P,|\nu|=|B_1|} (-1)^{\ell(\nu)}P(\nu)c_g^{0, \nu} = \sum_{\substack{\nu\in P, d\geq 0, \\ |\nu|+d=|B_1|-1}}(-1)^{\ell(\nu)+1}P(\nu)c_g^{0, \nu + (d+1)}.
\]
We look at the coefficient of $c_g^{0, \nu}$ with $|\nu|=|B_1|\geq 1$. If $\nu=(a_1\geq\cdots\geq a_m)$ and $R(\nu)$ is the set $\{a_i| 1\leq i\leq m\}$, the coefficient on the right hand side is
\[(-1)^{\ell(\nu)}\sum_{i\in R(\nu)} P(\nu_i),\]
where $\nu_i$ is $\nu$ with $i$ removed. However, in a permutation of $\nu$, the first integer must be an element $i$ of $R(\nu)$. Then, there are $P(\nu_i)$ permutations of the remaining components of $\nu$, and therefore, for $\nu$ with $|\nu|\geq 1$,
\begin{equation}
\label{eq:partitionsplit}
P(\nu)=\sum_{i\in R(\nu)} P(\nu_i).
\end{equation}
With this, the coefficient on the left and right hand side are equal.
In order to complete the proof, it suffices to show that
\begin{equation}
\label{eq:prodseries}
\begin{split}
& \sum_{\substack{\nu\in P, d\geq 0, \\ |\nu|+d=|B_1|-1}}\sum_{\substack{a+b=d, \\ (p_1, p_2)\in S(\nu)}} (-1)^{\ell(\nu)}P(\nu)c_f^{a, \pi(p_1)}c_g^{b, \pi(p_2)} \\
& = \sum_{j=1}^{|B_1|}\left(\sum_{\substack{\nu_1\in P, d_1\geq 0,\\ |\nu_1|+d_1=j-1}} (-1)^{\ell(\nu_1)}P(\nu_1)c_f^{d_1, \nu_1}\right)\left(\sum_{\substack{\nu_2\in P, d_2\geq 0,\\ |\nu_2|+d_2=|B_1|-j}} (-1)^{\ell(\nu_2)}P(\nu_2)c_g^{d_2, \nu_2}\right).
\end{split}
\end{equation}
We find the coefficient of $c_f^{d_1,\nu_1}c_g^{d_2,\nu_2}$ on both sides, where $d_1+d_2+|\nu_1|+|\nu_2|=|B_1|-1$. On the right hand side, the coefficient is
$(-1)^{\ell(\nu_1)+\ell(\nu_2)}P(\nu_1)P(\nu_2)$. For $\nu$ with $|\nu|=|B_1|-d-1$, let $T(\nu)=\{(p_1, p_2)\in S(\nu) | \pi(p_1)=\nu_1, \pi(p_2)=\nu_2\}$. Then, on the left hand side, the coefficient is, where $d=d_1+d_2$,
\begin{align*}
\sum_{\substack{\nu\in P,\\ |\nu|=|B_1|-d-1}} \sum_{\substack{(p_1, p_2)\in S(\nu), \\ \pi(p_1)=\nu_1, \pi(p_2)=\nu_2}} (-1)^{\ell(\nu)}P(\nu) & = \sum_{\substack{\nu\in P,\\ |\nu|=|B_1|-d-1}} (-1)^{\ell(\nu)}|T(\nu)|\cdot P(\nu)
\end{align*}
Suppose $|T(\nu)|\not=0$. Then, if $\ell(\nu)=\ell(\nu_1)+\ell(\nu_2)-k$, $0\leq k\leq \min(\ell(\nu_1), \ell(\nu_2))$.
\begin{claim}
\label{claim:coeffsum}
For $0\leq k\leq \min(\ell(\nu_1), \ell(\nu_2))$,
\[
\sum_{\substack{\nu\in P, |\nu|=|B_1|-d-1, \\ \ell(\nu)=\ell(\nu_1)+\ell(\nu_2)-k}} |T(\nu)|\cdot P(\nu)=\frac{\binom{\ell(\nu_1)}{k}\binom{\ell(\nu_2)}{k}}{\binom{\ell(\nu_1)+\ell(\nu_2)}{k}}\cdot\frac{(\ell(\nu_1)+\ell(\nu_2))!}{\ell(\nu_1)!\ell(\nu_2)!}P(\nu_1)P(\nu_2).
\]
\end{claim}
\begin{proof}
Let $l=\ell(\nu_1)+\ell(\nu_2)-k$ and $\vec{x}=(x_1,\ldots,x_l)$. Suppose partition $\nu$ has $|\nu|=|B_1|-d-1$ and $\ell(\nu)=l$. We see that the coefficient of any term of $M_\nu(\vec{x})$ in $M_{\nu_1}(\vec{x})M_{\nu_2}(\vec{x})$ is $|T(\nu)|$. Therefore, the sum of the coefficients in $M_{\nu_1}(\vec{x})M_{\nu_2}(\vec{x})$ of terms of $M_\nu(\vec{x})$ is $|T(\nu)|\cdot P(\nu)$, where there are $P(\nu)$ terms of $M_\nu(\vec{x})$.
Suppose that $S$ is the sum of the terms of $M_{\nu_1}(\vec{x})M_{\nu_2}(\vec{x})$ which contain all of the $x_i$, $x_1,\ldots,x_l$. If a term $p$ of $S$ contains all of the $x_i$, $p$ without its coefficient must be a term of $M_\nu(\vec{x})$ for some $\nu\in P$ with $|\nu|=|B_1|-d-1$ and $\ell(\nu)=l$. But, the sum of the coefficients in $M_{\nu_1}(\vec{x})M_{\nu_2}(\vec{x})$ of terms of $M_\nu(\vec{x})$ is $|T(\nu)|\cdot P(\nu)$. Therefore, the sum of the coefficients of $S$ is
\[
C=\sum_{\substack{\nu\in P,|\nu|=|B_1|-d-1, \\ \ell(\nu)=\ell(\nu_1)+\ell(\nu_2)-k}} |T(\nu)|\cdot P(\nu).
\]
However, for terms $p_i$ in $M_{\nu_i}(\vec{x})$, $1\leq i\leq 2$, if $p_1p_2$ contains all of the $x_i$, then exactly $k$ of the $x_i$ must be in both $p_1$ and $p_2$. Note that $S$ is the sum of $p_1p_2$ for such $(p_1,p_2)$, and $C$ is the number of $(p_1,p_2)$; we count $C$. There are
\[
\frac{(\ell(\nu_1)+\ell(\nu_2)-k)!}{k!(\ell(\nu_1)-k)!(\ell(\nu_2)-k)!}=\frac{\binom{\ell(\nu_1)}{k}\binom{\ell(\nu_2)}{k}}{\binom{\ell(\nu_1)+\ell(\nu_2)}{k}}\cdot\frac{(\ell(\nu_1)+\ell(\nu_2))!}{\ell(\nu_1)!\ell(\nu_2)!}
\]
ways to choose the $x_i$ for $p_1$ and $p_2$. For each choice of the $x_i$, there are $P(\nu_1)$ and $P(\nu_2)$ possibilities for $p_1$ and $p_2$, respectively. Thus,
\[
\sum_{\substack{\nu\in P, |\nu|=|B_1|-d-1, \\ \ell(\nu)=\ell(\nu_1)+\ell(\nu_2)-k}} |T(\nu)|\cdot P(\nu)=C=\frac{\binom{\ell(\nu_1)}{k}\binom{\ell(\nu_2)}{k}}{\binom{\ell(\nu_1)+\ell(\nu_2)}{k}}\cdot\frac{(\ell(\nu_1)+\ell(\nu_2))!}{\ell(\nu_1)!\ell(\nu_2)!}\cdot P(\nu_1)P(\nu_2).
\]
\end{proof}
Afterwards, the coefficient of $c_f^{d_1,\nu_1}c_g^{d_2,\nu_2}$ on the left hand side of \pref{eq:prodseries} is, with \Cref{claim:coeffsum} and \Cref{binomsum},
\begin{align*}
\sum_{\substack{\nu\in P, \\ |\nu|=|B_1|-d-1}}(-1)^{\ell(\nu)}|T(\nu)|\cdot P(\nu)
& = \sum_{k=0}^{\min(\ell(\nu_1), \ell(\nu_2))}(-1)^{\ell(\nu_1)+\ell(\nu_2)-k}\sum_{\substack{\nu\in P, |\nu|=|B_1|-d-1, \\ \ell(\nu)=\ell(\nu_1)+\ell(\nu_2)-k}} |T(\nu)|\cdot P(\nu)\\
& = (-1)^{\ell(\nu_1)+\ell(\nu_2)}P(\nu_1)P(\nu_2).
\end{align*}
We are done.
\end{proof}
\subsection{Distinct Indices}
Also, we look at the constant term following combinations of $\mathcal{Q}_i^j$ with distinct indices $i$ in \Cref{finalvalue1}. This is important for the proof of \Cref{lln}, as will be seen in \Cref{sec:main} and \Cref{sec:measures}, and for this, we use the operator in \pref{eq:multoperators}. But, it is necessary to consider the free cumulants of $s=\{c^{d,\nu}\}_{d\geq 0, \nu\in P}$ which are symmetrical in the $x_i$ beforehand.
\begin{proposition}
\label{symmetricalcumulants}
Suppose that $i$ is a positive integer, and $s = \{c^{d,\nu}\}_{d\geq 0, \nu\in P}$ is a sequence such that $F_i(s)$ is symmetrical with respect to $(x_j)_{j\geq 1}$. Then, $c_k(s)=0$ for all integers $k\geq 2$.
\begin{proof} Since $F_i(s)$ is symmetrical, we have that for all $\lambda\in P$, there exists a constant $c^{\lambda}$ such that $c^{d,\nu}=c^{\lambda}$ for all $d, \nu$ such that $\nu+(d)=\lambda$. The free cumulant of $s$ of order $k$ is
\[
c_k(s) = \theta^{k-1}\sum_{\substack{\nu\in P, d\geq 0,\\ |\nu|+d=k-1}} (-1)^{\ell(\nu)}P(\nu)c^{d, \nu}
= \theta^{k-1}\sum_{\lambda\in P, |\lambda| = k-1}c^\lambda\sum_{\substack{\nu\in P, d\geq 0,\\ \nu+(d)=\lambda}} (-1)^{\ell(\nu)}P(\nu),
\]
and $c_k(s)=0$ for $k\geq 2$ follows from \pref{eq:partitionsplit}.
\end{proof}
\end{proposition}
Suppose we have positive integers $i\leq j$, and $k$. For $F(x_1,\ldots,x_j), G(x_1,\ldots,x_j)\in\mathcal{F}_i^j$, let
\begin{equation}
\label{eq:multoperators}
\mathcal{R}_{i,j}^k(F(x_1,\ldots,x_j)) G(x_1,\ldots,x_j) = \mathcal{Q}_i^j\left(F(x_1,\ldots,x_j)\right)^k G(x_1,\ldots,x_j)|_{x_i=0}.
\end{equation}
\begin{lemma}
\label{symmetric}
Suppose $\lambda$ is a partition of length $m\geq 1$, $N\geq m$ is a positive integer, $f_i\in\mathcal{F}_{N-i+1}^{N-i+1}$ for $1\leq i\leq m$, and $g\in\mathcal{F}_N^N$. Then,
\[
\left(\prod_{i=1}^m \mathcal{R}_{N-i+1,N-i+1}^{\lambda_i}(f_i)\right)(g)
\]
is a symmetric formal series in $x_1, \ldots, x_{N-m}$.
\end{lemma}
\begin{proof}
This follows from induction on $m$ from $m=1$ to $N$.
\end{proof}
\begin{lemma}
\label{limit2}
Suppose $\lambda$ is a partition of length $m\geq 1$. Let $\{f_{i,N}\}_{N\geq m}$ for $1\leq i\leq m$ and $\{g_N(x_1,\ldots,x_N)\}_{N\geq m}$ be sequences of formal series such that $f_{i,N}\in\mathcal{F}_{N-i+1}^{N-i+1}$ and $g_N\in\mathcal{F}_N^N$ for $N\geq m$. Also, suppose $\{s_{1,N-i+1}f_{i,N}\}_{N\geq m}$ for $1\leq i\leq m$ and $\{s_{1,N}g_N\}_{N\geq m}$, which are symmetrical outside of $1$, have limits as $N\rightarrow\infty$ outside of $1$. Then, for $i\geq 1$, the limit from $N=i+m$ to $\infty$ of the formal series
\[\left(\prod_{i=1}^m \mathcal{R}_{N-i+1,N-i+1}^{\lambda_i}(f_{i,N})\right)(g_N)
\]
in $x_1,\ldots,x_{N-m}$ exists outside of $i$.
\end{lemma}
\begin{proof}
We show this with induction on $m$. For the base case $m=1$,
\[
\lim_{N\rightarrow\infty}\mathcal{Q}_1^N(s_{1,N}f_{1,N})^{\lambda_1}(s_{1,N}g_N)
\]
exists outside of $1$ from \Cref{limitexists} because $s_{1,N}f_{1,N}$ and $s_{1,N}g_N$ have limits outside of $1$. For some $i\geq 1$, consider $N\geq i+1$. From \Cref{symmetric}, $\mathcal{R}_{N,N}^{\lambda_1}(f_{1,N})(g_N)\in\mathcal{F}_i^{N-1}$.
From \Cref{prop:commutativeswitch}, $s_{1,N}\mathcal{Q}_N^N(f_{1,N})^{\lambda_1}(g_N)= \mathcal{Q}_1^N(s_{1,N}f_{1,N})^{\lambda_1}(s_{1,N}g_N)$. With this, the coefficient $c^{d, \nu}$ of $\mathcal{Q}_N^N(f_{1,N})^{\lambda_1}(g_N)|_{x_N=0}$ as a formal series in $x_1,\ldots,x_{N-1}$ outside of $i$ will be the coefficient $c^{0,\nu+(d)}$ of $\mathcal{Q}_N^N(f_{1,N})^{\lambda_1}(g_N)$ outside of $N$, and thus the coefficient $c^{0,\nu+(d)}$ of $\mathcal{Q}_1^N(s_{1,N}f_{1,N})^{\lambda_1}(s_{1,N}g_N)$ outside of $1$. Since the limit of $\mathcal{Q}_1^N(s_{1,N}f_{1,N})^{\lambda_1}(s_{1,N}g_N)$ exists, the limit of this coefficient exists from $N=i+1$ to $\infty$. As this holds for all $d, \nu$, the base case is complete.
Next, assume that the result holds for $m\geq 1$. We want to show that the result holds for $m+1$. Let
\[
h_{N-m}=\left(\prod_{i=1}^m \mathcal{R}_{N-i+1,N-i+1}^{\lambda_i}(f_{i,N})\right)(g_N).
\]
By the inductive hypothesis and because $h_{N-m}$ is symmetric, the limit of $s_{1,N-m}h_{N-m}=h_{N-m}$ from $N-m=1$ to $\infty$ exists outside of $1$. Then,
\[
\left(\prod_{i=1}^{m+1} \mathcal{R}_{N-i+1,N-i+1}^{\lambda_i}(f_{i,N})\right)(g_N)=\mathcal{R}_{N-m, N-m}^{\lambda_{m+1}}(f_{m+1,N})(h_{N-m}).
\]
is the base case $m=1$ with $f_{m+1, N}$ as $f_{1,N-m}$, $h_{N-m}$ as $g_{N-m}$, and $N-m$ as $N$, where the conditions are satisfied. Therefore, from $m=1$, the limit of the above expression outside of $i$ for $i\geq 1$ exists from $N-m=i+1$ to $\infty$, or from $N=i+m+1$ to $\infty$, completing the proof.
\end{proof}
\begin{theorem}
\label{finalvalue1}
Suppose $\lambda$ is a partition of length $m\geq 1$. Let $\{f_{i,N}(x_1,\ldots,x_{N-i+1})\}_{N\geq m}$ for $1\leq i\leq m$ and $\{g_N(x_1,\ldots,x_N)\}_{N\geq m}$ be sequences of formal series such that $f_{i,N}\in\mathcal{F}_{N-i+1}^{N-i+1}$ and $g_N\in\mathcal{F}_N^N$ for $N\geq m$. Moreover, for $1\leq i\leq m$, assume that $\{s_{1,N-i+1}f_{i,N}\}_{N\geq m}$, which is symmetrical outside of $1$, has a limit as $N\rightarrow\infty$ with limiting sequence $f_i$ outside of $1$. Also, assume that $\{s_{1,N}g_N\}_{N\geq m}$, which is symmetrical outside of $1$, has a limit as $N\rightarrow\infty$ with limiting sequence $g$ outside of $1$. Then,
\begin{equation}
\label{eq:valuemultiply1}
\begin{split}
& \lim_{N\rightarrow\infty}[1]\left(\prod_{i=1}^m \mathcal{R}_{N-i+1,N-i+1}^{\lambda_i}(f_{i,N})\right)(g_N)\\
& = \left(\sum_{\substack{\pi\in NC(\lambda_1+1), \\ \pi = B_1\sqcup\cdots\sqcup B_{\ell(\pi)}}}c_{|B_1|}(g)\prod_{i=2}^{\ell(\pi)} c_{|B_i|}(f_1)\right) \prod_{i=2}^m\left(\sum_{\substack{\pi\in NC(\lambda_i), \\ \pi = B_1\sqcup\cdots\sqcup B_{\ell(\pi)}}}\prod_{i=1}^{\ell(\pi)} c_{|B_i|}(f_i)\right).
\end{split}
\end{equation}
\end{theorem}
\begin{proof} We prove this with induction on $m$. For the base case $m=1$, use \Cref{value}. Suppose the result holds for $m\geq 1$. We want to show it holds for $m+1$.
From \Cref{symmetric},
\[
h_{N-m}(x_1,\ldots,x_{N-m})=\left(\prod_{i=1}^m\mathcal{R}_{N-i+1,N-i+1}^{\lambda_i}(f_{i,N})\right)(g_N)
\]
is symmetric in $x_1, \ldots, x_{N-m}$ for all $N\geq m$. We denote $h_N(x_1,\ldots,x_N)$ by $h_N$. Then,
\[
s_{1,N-m}\mathcal{Q}_{N-m}^{N-m}\left(f_{m+1,N}\right)^{\lambda_{m+1}}h_{N-m} =\mathcal{Q}_1^{N-m}\left(s_{1,N-m}f_{m+1,N}\right)^{\lambda_{m+1}}h_{N-m}.
\]
from \Cref{prop:commutativeswitch}. Since the switch does not change the constant,
\begin{align*}
& [1]\mathcal{R}_{N-m,N-m}^{N, \lambda_{m+1}}(f_{m+1,N})\left(\prod_{i=1}^m \mathcal{R}_{N-i+1,N-i+1}^{\lambda_i}(f_{i,N})\right)(g_N) = [1]\mathcal{Q}_{N-m}^{N-m}\left(f_{m+1,N}\right)^{\lambda_{m+1}}h_{N-m} \\& = [1]\mathcal{Q}_1^{N-m}\left(s_{1,N-m}f_{m+1,N}\right)^{\lambda_{m+1}}h_{N-m}.
\end{align*}
We want to compute the limit of this as $N\rightarrow\infty$. With $f_{N-m}=s_{1,N-m}f_{m+1,N}$ for $N\geq m+1$, outside of $1$, as $N\rightarrow\infty$, $f_N$ has limit $f_{m+1}$, and from \Cref{limit2}, $h_N$ has a limit; let $h$ be the limiting sequence of $h_N$ as $N\rightarrow\infty$. Then, from \Cref{value} with $N-m$ for $N$,
\[
\lim_{N\rightarrow\infty} [1]\mathcal{Q}_1^{N-m}(f_{N-m})^{\lambda_{m+1}} h_{N-m}= \sum_{\substack{\pi\in NC(\lambda_{m+1}+1), \\ \pi = B_1\sqcup\cdots\sqcup B_{\ell(\pi)}}} c_{|B_1|}(h)\prod_{i=2}^{\ell(\pi)} c_{|B_i|}(f_{m+1}).
\]
But, as $h_N$ is symmetric in $x_i$, $1\leq i\leq N$, for $N\geq 1$, $h$ is symmetric. Then, by \Cref{symmetricalcumulants}, $c_k(h)=0$, $k\geq 2$, and $c_1(h)=c_h^{0,0}$. From this,
\begin{align*}
& \lim_{N\rightarrow\infty} [1]\mathcal{Q}_1^{N-m}(f_{N-m})^{\lambda_{m+1}} h_{N-m} = c_h^{0,0}\left(\sum_{\substack{\pi\in NC(\lambda_{m+1}),\\ \pi = B_1\sqcup\cdots\sqcup B_{\ell(\pi)}}} \prod_{i=1}^{\ell(\pi)} c_{|B_i|}(f_{m+1})\right) \\
& = \left(\sum_{\substack{\pi\in NC(\lambda_1+1), \\ \pi = B_1\sqcup\cdots\sqcup B_{\ell(\pi)}}}c_{|B_1|}(g)\prod_{i=2}^{\ell(\pi)} c_{|B_i|}(f_1)\right) \prod_{i=2}^{m+1}\left(\sum_{\substack{\pi\in NC(\lambda_i), \\ \pi = B_1\sqcup\cdots\sqcup B_{\ell(\pi)}}} \prod_{i=1}^{\ell(\pi)} c_{|B_i|}(f_i)\right)
\end{align*}
using the inductive hypothesis, as needed.
\end{proof}
\section{Sequences of Operators}
\label{sec:main}
\subsection{Setup}
Suppose indices $\mathfrak{r}=\{i_j\}_{1\leq j\leq k}$ and a positive integer $N\geq\max(\mathfrak{r})$ are given. For a symmetric formal series $F(x_1,\ldots,x_N)$, let
\begin{equation}
\label{eq:sequenceoperators}
\mathcal{D}_\mathfrak{r}(F(x_1,\ldots,x_N)) = \left(\prod_{j=1}^k\left(\mathcal{D}_{i_j}+\frac{\partial}{\partial x_{i_j}} F(x_1,\ldots,x_N)\right)\right)(1).
\end{equation}
Above, the operators are for formal series. Also, suppose the set $I_N(\lambda)$ consists of indices $l$ of length $|\lambda|$ such that the first $\lambda_1$ indices are $i_1$, the next $\lambda_2$ indices are $i_2$, and so forth, until the last $\lambda_m$ indices are $i_m$, where $1\leq i_j\leq N$ for $1\leq j\leq m$.
The main result of \Cref{sec:main} is \Cref{lln:generalization} below. This is used to prove \Cref{lln}, see \Cref{subsec:mainproof}. Note that \Cref{lln} is a generalization of Claim 9.1 of \cite{matrix}, and some techniques from Section 5 of that paper are used in the proof of \Cref{lln:generalization}.
\begin{theorem}
\label{lln:generalization}
Suppose $\{F_N(x_1,\ldots,x_N)\}_{N\geq 1}$ is a sequence of symmetric formal series. Assume that for all partitions $\nu$ with $|\nu|\geq 1$, there exists a real number $c_\nu$ such that
\[
\displaystyle\lim_{N\rightarrow\infty}\frac{1}{N}\cdot[1]\frac{\partial}{\partial x_{i_1}}\cdots\frac{\partial}{\partial x_{i_r}} F_N(x_1,\ldots,x_N) = \frac{\ell(\nu)(|\nu|-1)!c_\nu}{P(\nu)}
\]
for all positive integers $i_1, \ldots, i_r$ such that $\sigma((i_1,\ldots,i_r))=\nu$. Then, for a partition $\lambda$ with $|\lambda|\geq 1$,
\begin{align*}
&\lim_{N\rightarrow\infty} \left(\frac{1}{N^{\ell(\lambda)+|\lambda|}}\sum_{l\in I_N(\lambda)}[1]\mathcal{D}_l(F_N(x_1,\ldots,x_N))\right) \\
& = \prod_{i=1}^{\ell(\lambda)}\left(\sum_{\pi\in NC(\lambda_i)} \prod_{B\in\pi}\theta^{|B|-1}\left(\sum_{\nu\in P, |\nu|=|B|}(-1)^{\ell(\nu)-1} P(\nu) c_{\nu}\right)\right).
\end{align*}
\end{theorem}
Note that for $\nu\in P$ with $|\nu|\geq 1$, if $i_1, \ldots, i_r$ are positive integers such that $\sigma((i_1,\ldots,i_r))=\nu$, then
\[
[1]\frac{\partial}{\partial x_{i_1}}\cdots\frac{\partial}{\partial x_{i_r}} F_N(x_1,\ldots,x_N) = \frac{|\nu|!}{P(\nu)}\cdot c_{F_N}^\nu.
\]
After this, the condition of \Cref{lln:generalization} gives
\begin{equation}
\label{eq:limcoeff}
\lim_{N\rightarrow\infty}\frac{|\nu|}{\ell(\nu)}\cdot\frac{c_{F_N}^\nu}{N}=c_\nu.
\end{equation}
\subsection{Sequences}
We consider sequences $s$ of $k$ operators over $x_1,\ldots,x_N$, where $s_j$ is the operator at step $j$, $1\leq j\leq k$. Each $s_j$ has an associated index $i_j$, where $1\leq i_j\leq N$. We say that $\mathfrak{r}=\{i_j\}_{1\leq j\leq k}$ are the $\textit{indices}$ of $s$. Here are the possibilities for $s_j$:
\begin{enumerate}
\item (\textit{Derivative}) $\frac{\partial}{\partial x_{i_j}},$ denoted by $\partial_{i_j}$.
\item (\textit{Switch}) $\frac{\theta}{x_{i_j}-x_i}(1-s_{i_j, i})$, where $1\leq i\leq N$, $i\not= i_j$. This is the $\textit{switch}$ from $i_j$ to $i$.
\item (\textit{Term Multiplication}) Multiplication by $c_F^\nu\partial_{i_j}(x_1^{a_1}\cdots x_N^{a_N})$ for $a_i$, $1\leq i\leq N$, such that $\nu\in P$ with $|\nu|\geq 1$ and $\pi((a_1,\ldots,a_N))=\nu$. We say that $c_F^\nu$ is the $\textit{constant}$ of the term multiplication.
\item (\textit{Change}) $\theta (d_{i_j}-C_{{i_j},i})$, where $1\leq i\leq N$, $i\not= i_j$. This is the $\textit{change}$ from $i_j$ to $i$.
\end{enumerate}
Observe that the term multiplications are equivalent to multiplying by $\partial_{i_j} F(x_1,\ldots,x_N)$. Also, if $s_j$ is a term multiplication by
\[p=c_F^\nu\frac{\partial}{\partial x_{i_j}}x_1^{a_1}\cdots x_N^{a_N},\]
we say that $x_i$ is $\textit{in}$ $s_j$ or $s_j$ $\textit{contains}$ $x_i$ if $p\not=0$ and the degree of $x_i$ in $p$ is at least $1$.
Next, for $0\leq j\leq k$, where $r(s)_0=1$, let
\begin{equation}
\label{eq:seqval}
r(s)_j = s_j\circ s_{j-1}\circ\cdots\circ s_1(1).
\end{equation}
Also, suppose we have indices $\mathfrak{r}=\{i_j\}_{1\leq j\leq k}$ and $N\geq\max(\mathfrak{r})$. For sequences $s$ with indices $\mathfrak{r}$ over $x_1,\ldots,x_N$ such that $r(s)_k$ is nonzero and has degree $0$ in the $x_i$, let $D_{N,\mathfrak{r}}(s)$ be the set of $x_i$ such that $i\notin\mathfrak{r}$ which are in term multiplications in $s$.
From now on, view the coefficients $c_F^\nu$ of $F(x_1,\ldots,x_N)$ as variables rather than constants. We view $c_F^\nu$ as having degree $|\nu|$, and from \pref{eq:limcoeff}, $c_F^\nu$ has order $N$. Additionally, the constant term $[1]F(x_1,\ldots,x_N)=c_F^0$ does not matter since we are taking the derivative. Therefore, we set $c_F^0=0$, and when we refer to a coefficient $c_F^\nu$ with $\nu\in P$, assume $|\nu|\geq 1$.
\begin{lemma}
\label{sequences}
Suppose $s$ is a sequence of length $k$. For $0\leq j\leq k$,
\[
r(s)_j = \prod_{i=1}^m c_F^{\nu_i} \cdot P(x_1,\ldots,x_N),
\]
where $s_j$ is term multiplication for $j=j_i$, $1\leq i\leq m$, with $j_i<j_{i+1}$ for $1\leq i\leq m-1$, the term multiplication $s_{j_i}$ has constant $c_F^{\nu_i}$, and $P(x_1,\ldots,x_N)$ is some homogeneous integer polynomial which is $0$ or has degree $\sum_{i=1}^m |\nu_i|-j.$
\end{lemma}
\begin{proof}
We can prove this with induction on $j$. The base case $j=0$ is clear. Assume the result holds for $j$, $0\leq j\leq k-1$. We want to show the result holds for $j+1$. Suppose $P(x_1,\ldots,x_N)$ and $P'(x_1,\ldots,x_N)$ are the polynomials for $r(s)_j$ and $r(s)_{j+1}$, respectively.
If $r(s)_j=0$, then $r(s)_{j+1}=0$, and the statement holds. Assume $r(s)_j\not=0$. If $s_{j+1}$ is a derivative, switch, or change, we see that in $r(s)_{j+1}$, $P'(x_1,\ldots,x_N)=0$ or $P'(x_1,\ldots,x_N)$ is a homogeneous integer polynomial with degree
$\sum_{i=1}^m |\nu_i|-j-1$, and the $c_F^\nu$ remain the same. However, if $s_{j+1}$ is term multiplication by $c_F^{\nu_{m+1}}\partial_{i_j}(x_1^{a_1}\cdots x_N^{a_N})$,
\[
r(s)_{j+1} = \prod_{i=1}^{m+1} c_F^{\nu_i}\cdot\left(\frac{\partial}{\partial x_{i_j}}x_1^{a_1}\cdots x_N^{a_N}\right)P(x_1,\ldots,x_N).
\]
There are now $m+1$ term multiplications, where $j_{m+1}=j+1$ and $s_{j_{m+1}}=s_{j+1}$ has constant $c_F^{\nu_{m+1}}$. Moreover, because $\pi((a_1,\ldots,a_N))=\nu_{m+1}$, we have that $P'(x_1,\ldots,x_N)=(\partial_{i_j}x_1^{a_1}\cdots x_N^{a_N})P(x_1,\ldots,x_N)$
is $0$ or is a homogeneous integer polynomial with degree $\sum_{i=1}^{m+1} |\nu_i|-j-1$. The induction is complete.
\end{proof}
\begin{corollary}
\label{polynomial}
For indices $\mathfrak{r}=\{i_j\}_{1\leq j\leq k}$ and $N\geq\max(\mathfrak{r})$, $[1]\mathcal{D}_\mathfrak{r}(F(x_1,\ldots,x_N))$ is a polynomial in the $c_F^\nu$ for $\nu\in P$ with $|\nu|\geq 1$ which is homogeneous of degree $k$, where $c_F^\nu$ has degree $|\nu|$ for $\nu\in P$ with $|\nu|\geq 1$.
\end{corollary}
\begin{proof}
This follows from \Cref{sequences}.
\end{proof}
From \Cref{sequences}, if a sequence $s$ of length $k$ has $r(s)_k$ have degree $0$ in the $x_i$, we know that $r(s)_k=d\cdot\prod_{i=1}^m c_F^{\nu_i}$ for nonzero $d$ with the sum of the $|\nu_i|$ equal to $k$, where the $c_F^{\nu_i}$ are from \Cref{sequences}. Then, for such $s$, where $k$ can be any positive integer, let
\begin{equation}
\label{eq:endconstant}
C(s)=d.
\end{equation}
\begin{definition}
\label{def:setseq}
For a positive integer $k$, suppose we have a sequence of indices $\mathfrak{r}=\{i_j\}_{1\leq j\leq k}$. For $N\geq\max(\mathfrak{r})$, let $T_N(\mathfrak{r})$, $T_N^1(\mathfrak{r})$, and $T_N^2(\mathfrak{r})$ denote set of sequences $s$ over $x_1,\ldots,x_N$ with indices $\mathfrak{r}$ such that $r(s)_k$ is nonzero and has degree $0$ in the $x_i$ such that:
\begin{itemize}
\item The operators of $s$ in $T_N(\mathfrak{r})$ can be any operator (derivatives, term multiplications, switches, and changes).
\item The operators of $s$ in $T_N^1(\mathfrak{r})$ are only derivatives, term multiplications, and switches.
\item The operators of $s$ in $T_N^2(\mathfrak{r})$ are only term multiplications and changes.
\end{itemize}
\end{definition}
Throughout \Cref{sec:main}, $T_N(\mathfrak{r})$, $T_N^1(\mathfrak{r})$, and $T_N^2(\mathfrak{r})$ are referred to. Particularly, we have that
\[
[1]\mathcal{D}_{\mathfrak{r}}(F(x_1,\ldots,x_N))= [1]\left(\prod_{j=1}^k\left(\mathcal{D}_{i_j} + \frac{\partial}{\partial x_{i_j}}F(x_1,\ldots,x_N)\right)\right)(1)= \sum_{s\in T_N^1(\mathfrak{r})} r(s)_k.
\]
\subsection{Orders}
\begin{definition}
\label{def:order}
Suppose $Q(c^\nu; N)$ is a polynomial in the $c^\nu$, $\nu\in P$, $|\nu|\geq 1$, with finite degree and finitely many terms, and coefficients which are functions of $N$. Then, $Q(c^\nu; N)$ has an $\textit{order}$ of $N^k$ if for a term
\[
\prod_{i=1}^m c^{\nu_i},
\]
the coefficient $f(N)$ in $Q(c^\nu; N)$ has $|f(N)|=O(N^{k-m})$ for positive integers $N$.
\end{definition}
If $Q(c^\nu; N)$ has an order of $N^k$, we say that $Q(c^\nu; N)$ is $O(N^k)$. Later, we use $c_F^\nu$ for $c^\nu$, and generally, for $k$ indices $\mathfrak{r}$, if we have sets $S_N$ in $T_N(\mathfrak{r})$ for $N\geq 1$, we let
\begin{equation}
\label{eq:order}
Q(c_F^\nu; N) = \sum_{s\in S_N} r(s)_k.
\end{equation}
We get $Q(c_F^\nu;N)=\mathcal{D}_{\mathfrak{r}}(F(x_1,\ldots,x_N))|_{x_i=0, 1\leq i\leq N}$ with $S_N=T_N^1(\mathfrak{r})$ for $N\geq 1$. Later on, \Cref{def:order} and \pref{eq:order} are used with $\mathcal{D}_{\mathfrak{r}}$ as well as other operators.
\begin{proposition}
\label{prop:boundedconst}
Suppose $\mathfrak{r}=\{i_j\}_{1\leq j\leq k}$. There exists a constant $C>0$ such that for all positive integers $N$, for any sequence $s\in T_N(\mathfrak{r})$, $|C(s)|\leq C$.
\end{proposition}
\begin{proof}
For such a sequence $s$, for $0\leq j\leq k$, suppose $P_j(x_1,\ldots,x_N)$ is the polynomial for $r(s)_j$ in \pref{eq:seqval}. Suppose $s$ has $m$ term multiplications. Since each other operator decreases the degree in the $x_i$ by $1$, the degree in the $x_i$ of $P_j(x_1,\ldots,x_N)$ for $0\leq j\leq k$ is at most $k-m$, which is less than $k$. Using this, \Cref{prop:boundedconst} can be seen by, for $0\leq j\leq k-1$, bounding the factor $s_{j+1}$ changes the sum of the absolute values of the coefficients from $P_j(x_1,\ldots,x_N)$ to $P_{j+1}(x_1,\ldots,x_N)$. Note that $P_k(x_1,\ldots,x_N)=C(s)$.
\end{proof}
\begin{proposition}
\label{order}
For a positive integer $k$ and indices $\mathfrak{r}=\{i_j\}_{1\leq j\leq k}$, for $N\geq\max(\mathfrak{r})$, $[1]\mathcal{D}_\mathfrak{r}(F(x_1,\ldots,x_N))$ has order $N^k$.
\end{proposition}
\begin{proof}
We know from \Cref{polynomial} that the expression will be a polynomial in the $c_{F}^\nu$ which is homogeneous of order $k$. Suppose that for a positive integer $m$, for $1\leq i\leq m$, $\nu_i\in P$ such that $|\nu_i|\geq 1$, and the sum of the $|\nu_i|$ is $k$. Note that the number of possible $\nu_1, \ldots, \nu_m$, $1\leq m\leq k$, is finite. Then, if we show that for $N\geq\max(\mathfrak{r})$, the coefficient of
\[
p=\prod_{i=1}^m c_{F}^{\nu_i}
\]
is of order $N^{k-m}$, we will be done by \Cref{def:order}. For $N\geq\max(\mathfrak{r})$, suppose the coefficient of $p$ is $r$; we want to show that $|r|=O(N^{k-m}).$ Let $T$ be the set of sequences $s$ such that $r(s)_k=dp$, where $d\not=0$ is a real number. From \Cref{prop:boundedconst}, there exists $C>0$ such that for all $N$, for all $s\in T$, $|C(s)|\leq C$. Then,
\[
r=\sum_{s\in T} C(s),
|r|=\Bigg|\sum_{s\in T} C(s)\Bigg| \leq \sum_{s\in T} |C(s)| \leq C\cdot |T|
\]
with the triangle inequality, and it suffices to show that $|T|=O(N^{k-m})$.
Suppose $s\in T$. We know that exactly $m$ of the $s_i$ must be term multiplications, and as seen in the proof of \Cref{prop:boundedconst}, the degree in the $x_i$ of $r(s)_j$ for $0\leq j\leq k$ is at most $k-m$.
\begin{claim}
\label{claim:switches}
Suppose $s\in T$. If $x_i\in D_{N,\mathfrak{r}}(s)$, suppose the first term multiplication $x_i$ is in is $s_{j'}$. Then, $s_j$ must be a switch from $i_j$ to $i$ for some $j$, $j'<j\leq k$.
\end{claim}
\begin{proof}
Observe that, in $r_{j'}(s)$, $x_i$ is in all of the terms, and must be removed because $r_k(s)$ cannot contain $x_i$. For the sake of contradiction, assume the statement does not hold. If $j'=k$, then $r_k(s)=r_{j'}(s)$ contains $x_i$, a contradiction since $r_k(s)=dp$ for a nonzero real number $d$. Therefore, $j'<k$. Below, we view $r_j(s)$ for $1\leq j\leq k$ as a polynomial in the $x_i$, and a term in $r_j(s)$ is a product of the $x_i$ with its coefficient.
Consider $s_j$, $j'<j\leq k$. We see that if $s_j$ is a derivative, term multiplication, or switch, all nonzero terms of $r_j(s)$ will have $x_i$. Then, in $r_k(s)\not= 0$, the terms will contain $x_i$, a contradiction.
\end{proof}
Let $S=\{x_{i_j}|1\leq j\leq k\}$. If $|D_{N,\mathfrak{r}}(s)|=d$, from \Cref{claim:switches}, we see that we can find one switch from $i_j$ to $i$ for each $x_i\in D_{N,\mathfrak{r}}(s)$, giving $d$ switches in total, where $0\leq d\leq k-m$. Also, the number possibilities for $D_{N,\mathfrak{r}}(s)$, the number subsets of
$\{x_1, \ldots, x_N\}\backslash S$ with size $d$, is at most $N^d$. Next, suppose $X\subset\{x_1, \ldots, x_N\}\backslash S$, $|X|=d$. Because the variables outside of $\mathfrak{r}$ are symmetrical, the number of $s\in T$ such that $D_{N,\mathfrak{r}}(s)=X$ is the same for all such $X$; let this number be $T_d$.
In $s$, there are $m$ term multiplications, known as $\alpha(s)$ operators. Also, there are $d\leq k-m$ operators
which are switches from $i_j$ to $i$, where $x_i\in D_{N,\mathfrak{r}}(s)$ and the $i$ are distinct, known as $\beta(s)$ operators. The other $k-m-d$ operators, known as $\gamma(s)$ operators, can be derivatives or any switch. In this proof, there can be overcounting of $N_d$, with $s\in T$ counted multiple times.
For $d$, $0\leq d\leq k-m$, there are
\[
\frac{k!}{m!d!(k-m-d)!}
\]
possible groupings of the $s_j$ into $\alpha(s), \beta(s)$, and $\gamma(s)$ operators. For a term multiplication, or a $\alpha(s)$ operator, suppose the term before the derivative is $q$. We know that all $x_i$ in $q$ must be in $S\cup D_{N,\mathfrak{r}}(s)$. Also, the total degree of $q$ after the derivative is at most $k-m$. We look at degrees of the $x_i$ in $q$. Note that for $x_{i_j}$, where $\partial_{i_j}q\not=0$, the degree must be at least $1$ and at most $k-m+1$, giving $k-m+1$ possibilities. For the other $x_i$, the degree must be at least $0$ and at most $k-m$, also giving $k-m+1$ possibilities. With this, the number of possibilities for the $\alpha(s)$ operators is at most
\[
(k-m+1)^{(k+|D_{N,\mathfrak{r}}(s)|)m}=(k-m+1)^{(k+d)m}.
\]
On the other hand, for the $\beta(s)$ operators, there being $d$ in total, there are $d!$ possible orderings. Finally, each of the other $\gamma(s)$ operators, there being $k-m-d$ in total, have $N$ possibilities, giving a total of $N^{k-m-d}$. Then, for a constant $C_d$ not depending on $N$,
\[
T_d\leq \frac{k!}{m!d!(k-m-d)!}\cdot (k-m+1)^{(k+d)m}\cdot d!\cdot N^{k-m-d}= C_d\cdot N^{k-m-d}.
\]
From this, for $0\leq d\leq k-m$, the number of $s\in T$ such that $|D_{N,\mathfrak{r}}(s)|=d$ is at most $N^d\cdot T_d\leq C_d\cdot N^{k-m}$, where there are at most $N^d$ choices for $D_{N,\mathfrak{r}}(s)$. Therefore,
\[
|T| \leq \left(\sum_{d=0}^{k-m} C_d\right)N^{k-m},
\]
and $|T|=O(N^{k-m})$, as desired.
\end{proof}
\subsection{Remainders}
\label{subsec:remainders}
\begin{proposition}
\label{prop:limitzero}
Suppose $s=\{a_\nu(N)\}_{\nu\in P}$ is a sequence of variables over positive integers $N$ such that for each $\nu\in P$, there exists a constant $C>0$ such that $|a_\nu(N)|\leq CN$ for all positive integers $N$. Then, if $Q(c^\nu; N)$ is order $N^k$,
\[
\lim_{N\rightarrow\infty}\frac{Q(a_\nu(N); N)}{N^{k+1}}=0,
\]
where we let $c^\nu=a_\nu(N)$ for all $\nu\in P$.
\end{proposition}
\begin{proof}
This is clear from considering each term of $Q(a_\nu(N);N)$, with there being a finite number of terms.
\end{proof}
Suppose we have a polynomial $R$ in the $c_F^\nu$ which is homogeneous of degree $k$ with order $N^k$. For $N\geq 1$, we can let $c_F^\nu=c_{F_N}^\nu$, and from the conditions of \Cref{lln:generalization}, we see that for each $\nu\in P$, if $a_\nu(N)=c_{F_N}^\nu$, there exists a constant $C>0$ such that $|a_\nu(N)|\leq CN$ for all $N\geq 1$. Afterwards, we use \Cref{prop:limitzero} on $R$ with $c_F^\nu=a_\nu(N)$ for $N\geq 1$ to get that
\begin{equation}
\label{eq:remainderto0}
\lim_{N\rightarrow\infty} \frac{R(a_\nu(N); N)}{N^{k+1}} = \lim_{N\rightarrow\infty} \frac{R(c_{F_N}^\nu; N)}{N^{k+1}} = 0.
\end{equation}
We use \pref{eq:remainderto0} in the proof of \Cref{lln:generalization} in Subsection \ref{subsec:llngeneralization} to show that as $N\rightarrow\infty$, such remainders $R$ go to $0$. Particularly, \pref{eq:remainderto0} can be used with various results from Subsections \ref{subsec:remainders} and \ref{subsec:differentop}.
\begin{proposition}
\label{remainder1}
Suppose that $\lambda$ is a partition with $\ell(\lambda)=m$ and $|\lambda|=k$. Also, suppose $\mathfrak{r}=\{i_j\}_{1\leq j\leq k}$, where the first $\lambda_1$ indices are $1$, the next $\lambda_2$ indices are $2$, and so forth, until the last $\lambda_m$ indices are $m$. Then, for $N\geq m$,
\[
\frac{1}{N^{m}} \sum_{l\in I_N(\lambda)} [1]\mathcal{D}_{l}(F(x_1,\ldots,x_N)) = [1]D_\mathfrak{r}(F(x_1,\ldots,x_N)) + R,
\]
where $R$ is a homogeneous polynomial in the $c_F^\nu$ with degree $k$ and order $N^{k-1}$.
\end{proposition}
\begin{proof}
Suppose that in $l\in I_N(\lambda)$, the first $\lambda_1$ indices are $i_1$, the next $\lambda_2$ indices are $i_2$, and so forth, until the last $\lambda_m$ indices are $i_m$, where $1\leq i_j\leq N$ for $1\leq j\leq m$. From \Cref{order}, for all $l\in I_N(\lambda)$, $\mathcal{D}_{l}(F(x_1,\ldots,x_N))$ has order $N^k$. Note that if $l$ has all $i_j$ distinct, $[1]\mathcal{D}_{l}(F(x_1,\ldots,x_N))$ is equal, by symmetry, to $[1]\mathcal{D}_{\mathfrak{r}}(F(x_1,\ldots,x_N))$. Consider the set $S$ of other $l$, where some of the $i_j$ are equal. Let
\[
R'=\frac{1}{N^m}\sum_{l\in S} [1]\mathcal{D}_l(F(x_1,\ldots,x_N)).
\]
From \Cref{polynomial}, $R'$ is a polynomial in the $c_F^\nu$ which is homogeneous of degree $k$.
Note that there are $N^{m}-O(N^{m-1})$ $l$ where all of the $i_j$ are distinct, and dividing by $N^{m}$ will give
\[
\left(1-O\left(\frac{1}{N}\right)\right)[1]\mathcal{D}_{\mathfrak{r}}(F(x_1,\ldots,x_N)).
\]
Also, $|S|=O(N^{m-1})$ and each $\mathcal{D}_{\mathfrak{r}}(F(x_1,\ldots,x_N))$ is order $N^k$, so $R'$ is $\frac{1}{N^m}\cdot O(N^{m-1})\cdot N^k=O(N^{k-1})$.
Then, we get
\[
R=R'-O\left(\frac{1}{N}\right)[1]\mathcal{D}_{\mathfrak{r}}(F(x_1,\ldots,x_N))
\]
is a homogeneous polynomial in the $c_F^\nu$ with degree $k$ and order $N^{k-1}$, as desired.
\end{proof}
\begin{lemma}
\label{remainder2}
Suppose that $\mathfrak{r}=\{i_j\}_{1\leq j\leq k}$ are indices. For $N\geq \max(\mathfrak{r})$, let $H$ be the set of sequences $s$ in $T_N^1(\mathfrak{r})$ which satisfy the following conditions:
\begin{itemize}
\item For $1\leq j\leq k$, if $s_j$ is the switch from $i_j$ to $i$, $i\notin \mathfrak{r}$.
\item There do not exist integers $i$, $j_1$, and $j_2$, $1\leq i\leq N$, $1\leq j_1, j_2\leq k$, $j_1\not=j_2$, such that $s_{j_1}$ is the switch from $i_{j_1}$ to $i$ and $s_{j_2}$ is the switch from $i_{j_2}$ to $i$.
\end{itemize}
Then,
\[[1]\mathcal{D}_{\mathfrak{r}}(F(x_1,\ldots,x_N)) = \sum_{s\in H} r(s)_k + R,\]
where $R$ is a homogeneous polynomial in the $c_F^\nu$ with degree $k$ and order $N^{k-1}$.
\end{lemma}
\begin{proof}
Suppose that $Q$ is the set $s\in T_N^1(\mathfrak{r})$ such that at least one of the conditions is not followed. Then,
\[
R = \sum_{s\in Q} r(s)_k.
\]
Since $Q\subset T_N^1(\mathfrak{r})$, $R$ is a polynomial in the $c_F^{\nu}$ which is homogeneous of degree $k$. It suffices to show that the coefficient in $R$ of each term
\[
p=\prod_{i=1}^m c_{F}^{\nu_i}
\]
is order $N^{k-m-1}$ from \Cref{def:order}. We consider the set $T$ of $s\in Q$ which end with $p$. Following the proof of \Cref{order}, let the coefficient of $p$ be $r$. We want to show that $r$ is $O(N^{k-m-1})$, and we know that $|C(s)|\leq C$ for a $C>0$ independent of $N$. Therefore, it suffices to show that $|T|=O(N^{k-m-1})$.
In $s\in T$, there are $m$ term multiplications, or $\alpha(s)$ operators. Also, there are $d$, $d\leq k-m$, $\beta(s)$ operators
which are switches from $i_j$ to $i$ where $x_i\in D_{N,\mathfrak{r}}(s)$, such that each $x_i\in D_{N,\mathfrak{r}}(s)$ is in exactly one $\beta(s)$ operator. The other $k-m-d$ operators are $\gamma(s)$ operators, and are derivatives or switches.
Consider when one of the switches is from $i_j$ to $i\in \mathfrak{r}$. Since $\mathfrak{r}$ and $D_{N,\mathfrak{r}}(s)$ are disjoint, the switch must be a $\gamma(s)$ operator, and the number of possibilities for the $\gamma(s)$ operators is $O(N^{k-m-d-1})$. With the number of possibilities for $D_{N,\mathfrak{r}}(s)$ being at most $N^d$ and for the other choices being bounded, the total number of possible $s$ is then $O(N^{k-m-1})$.
For the next case, suppose there are two switches with the same $i$. Then, we could have a $\beta(s)$ and $\gamma(s)$ switch or two $\gamma(s)$ switches with the same $i$. For the first case, there are at most $N$ ways to pick $i$. Afterwards, the number of possibilities for $D_{N,\mathfrak{r}}(s)$ is at most $N^{d-1}$ and the $\gamma(s)$ switches is $O(N^{k-m-d-1})$. Where there are $O(1)$ possibilities for the other choices, this gives $O(N^{k-m-1})$ possibilities for the first case. Next, for the second case, there are at most $N$ ways to choose $i$. Afterwards, the number of possibilities for $D_{N,\mathfrak{r}}(s)$ is at most $N^d$ and the $\gamma(s)$ switches is $O(N^{k-m-d-2})$. Where there are $O(1)$ possibilities for the other choices, this gives
$O(N^{k-m-1})$ possibilities for the second case.
Therefore, $|T|=O(N^{k-m-1})$, as desired.
\end{proof}
\subsection{Sequences For Different Operators}
\label{subsec:differentop}
Next, we consider the operators $\mathcal{Q}_i^N$ from \Cref{sec:formal}. Particularly,
\[
N\mathcal{Q}_i^N\left(\frac{\partial_i F}{N}\right) = \theta\sum_{\substack{1\leq j\leq N, \\ j\not= i}} (d_i-C_{i,j}) + \frac{\partial}{\partial x_i} F(x_1,\ldots,x_N)
\]
is used. In \Cref{remainder3}, we show that we can replace each $\mathcal{D}_i$ with $N\mathcal{Q}_i^N\left(\frac{\partial_{i_j}F}{N}\right)$. Afterwards, in \Cref{remainder4}, we use the $\mathcal{R}_{i,j}^k$ operators to evaluate the resulting expression. Similarly to before, for $\mathfrak{r}=\{i_j\}_{1\leq j\leq k}$ and $N\geq\max(\mathfrak{r})$, we let
\begin{equation}
\label{eq:sequenceoperators2}
\mathcal{Q}_{\mathfrak{r}}^N(F(x_1,\ldots,x_N)) = N^k\left(\prod_{j=1}^k \mathcal{Q}_{i_j}^N\left(\frac{\partial_{i_j} F}{N}\right)\right)(1) = \sum_{s\in T_2^k(\mathfrak{r})} r(s)_k.
\end{equation}
\begin{proposition}
\label{order2}
Let $\mathfrak{r}=\{i_j\}_{1\leq j\leq k}$. For $N\geq\max(\mathfrak{r})$, $[1]\mathcal{Q}^N_{\mathfrak{r}}(F(x_1,\ldots,x_N))$ has order $N^k$.
\end{proposition}
\begin{proof}
We can use the same proof as \Cref{order}, but with no derivatives and each switch replaced by a change, with the switch from $i$ to $j$ replaced by the change from $i$ to $j$.
\end{proof}
\begin{lemma}
\label{newremainder2}
Suppose that $\mathfrak{r}=\{i_j\}_{1\leq j\leq k}$. For $N\geq\max(\mathfrak{r})$, suppose $H$ is the set of sequences $s$ in $T_N^2(\mathfrak{r})$ which satisfy the following conditions:
\begin{itemize}
\item For $1\leq j\leq k$, if $s_j$ is the change from $i_j$ to $i$, $i\notin\mathfrak{r}$.
\item There do not exist integers $i$, $j_1$, and $j_2$, $1\leq i\leq N$, $1\leq j_1, j_2\leq k$, $j_1\not=j_2$, such that $s_{j_1}$ is the change from $i_{j_1}$ to $i$ and $s_{j_2}$ is the change from $i_{j_2}$ to $i$.
\end{itemize}
Then,
\[
[1]\mathcal{Q}_{\mathfrak{r}}^N(F(x_1,\ldots,x_N)) = \sum_{s\in H} r(s)_k+ R,
\]
where $R$ is a homogeneous polynomial in the $c_F^\nu$ with degree $k$ and order $N^{k-1}$.
\end{lemma}
\begin{proof}
We can use the same proof as \Cref{remainder2}, but with no derivatives and each switch replaced by a change, with the switch from $i$ to $j$ replaced by the change from $i$ to $j$.
\end{proof}
\begin{theorem}
\label{remainder3}
Suppose that $k$ is a positive integer and $\mathfrak{r}=\{i_j\}_{1\leq j\leq k}$. For $N\geq \max(\mathfrak{r})$,
\begin{equation}
\label{eq:diffoperators}
[1]\mathcal{D}_\mathfrak{r}(F(x_1,\ldots,x_N))= [1]\mathcal{Q}_\mathfrak{r}^N(F(x_1,\ldots,x_N)) + R,
\end{equation}
where $R$ is a homogeneous polynomial in the $c_F^\nu$ with degree $k$ and order $N^{k-1}$.
\end{theorem}
\begin{proof}
We note that in $T_N^1(\mathfrak{r})$, the sequences can have derivatives, but cannot in $T_N^2(\mathfrak{r})$, where $T_N^1(\mathfrak{r})$ and $T_N^2(\mathfrak{r})$ correspond to the left and right hand side of \pref{eq:diffoperators}, respectively. Then, let $T_N^3(\mathfrak{r})$ be the set of $s\in T_N^1(\mathfrak{r})$ which do not contain derivatives.
Suppose that $H$ is the set of $s$ in $T_N^1(\mathfrak{r})$ satisfying the \Cref{remainder2} conditions. From the Lemma,
\[
[1]\mathcal{D}_{\mathfrak{r}}(F(x_1,\ldots,x_N))=\sum_{s\in H} r(s)_k + R_1,
\]
where $R_1$ is a homogeneous polynomial in the $c_F^\nu$ with degree $k$ and order $N^{k-1}$. Consider $H_1=H\cap T_N^3(\mathfrak{r})$. We see that $H_1$ is the set of $s\in T_N^1(\mathfrak{r})$ satisfying the conditions of \Cref{remainder2} with only switches and term multiplications, and
\[
\sum_{s\in H} r(s)_k = \sum_{s\in H_1} r(s)_k + R_2, R_2= \sum_{s\in H\cap (T_N^1(\mathfrak{r})\backslash T_N^3(\mathfrak{r}))} r(s)_k.
\]
Also, let $H_2$ be the set of $s$ in $T_N^2(\mathfrak{r})$ satisfying the \Cref{newremainder2} conditions. By the Lemma,
\[
[1]\mathcal{Q}_{\mathfrak{r}}^N(F(x_1,\ldots,x_N))=\sum_{s\in H_2} r(s)_k+R_3,
\]
where $R_3$ is a homogeneous polynomial in the $c_F^\nu$ with degree $k$ and order $N^{k-1}$.
\begin{claim}
$R_2$ is a homogeneous polynomial in the $c_F^\nu$ with degree $k$ and order $N^{k-1}$
\end{claim}
\begin{proof}
We bound the number of $s\in T_N^1(\mathfrak{r})\backslash T_N^3(\mathfrak{r})$, the $s\in T_N^1(\mathfrak{r})$ which have a derivative, such that $r(s)_k=dp$, where
\[
p=\prod_{i=1}^m c_F^{\nu_i}.
\]
Looking at the proof of \Cref{order}, the derivatives of $s$ only appear in $\gamma(s)$ operators. Where $|D_{N,\mathfrak{r}}(s)|=d$, the number of ways to choose the $\gamma(s)$ operators such that there is at least $1$ derivative is $O(N^{k-m-d-1})$. Then, we see that the number of $s$ in $T_N^1(\mathfrak{r})\backslash T_N^3(\mathfrak{r})$, and therefore $H\cap (T_N^1(\mathfrak{r})\backslash T_N^3(\mathfrak{r}))$, ending with $p$ is $O(N^{k-m-1})$, as needed.
\end{proof}
For a sequence $s$ in $H_1$, let $c(s)$ be the sequence in $H_2$ where if $s_{j}$ is the switch from $i_j$ to $i$, then $c(s)_j$ is the change from $i_j$ to $i$, and all other operators are the same. We have that $s\in H_1$ have switches and term multiplications, while $s\in H_2$ have changes and term multiplications. Clearly, $c$ is a bijection from $H_1$ to $H_2$.
\begin{claim}
\label{claim:bijval}
For all $s\in H_1$, $r(s)_k = r(c(s))_k$.
\end{claim}
\begin{proof}
Consider $s\in T_N(\mathfrak{r})$ with term multiplications, switches, and changes such that:
\begin{itemize}
\item For $1\leq j\leq k$, if $s_j$ is the switch or change from $i_j$ to $i$, $i\notin \mathfrak{r}$.
\item There do not exist integers $i$, $j_1$, and $j_2$, $1\leq i\leq N$, $1\leq j_1, j_2\leq k$, $j_1\not=j_2$, such that $s_{j_1}$ is the switch or change from $i_{j_1}$ to $i$ and $s_{j_2}$ is the switch or change from $i_{j_2}$ to $i$.
\end{itemize}
Suppose that $s_j$, $1\leq j\leq k$, is a switch from $i_j$ to $i$ and $s_{j'}$, $j<j'\leq k$, are not changes. It suffices to show that if we make $s_j$ the change from $i_j$ to $i$, $r(s)_k$ is the same. Then, for $s\in H_1$, if the switches of $s$ are $s_{j_1}, s_{j_2}, \ldots, s_{j_r}$, $1\leq j_1<\cdots<j_r\leq k$, we could make $s_{j_u}$ into a change from $u=1$ to $r$ and get $c(s)$. This would give $r(c(s))_k=r(s)_k$.
Suppose $q$ is a term of $r(s)_{j-1}$. We look at $s_j(q)$, which is a sum of terms or $0$. Suppose $q'$ is a term of $s_j(q)$ and the degree of $x_i$ in $q'$ is at least $1$. Afterwards, where $j< j'\leq k$, using that $s\in H_1$ is the switch from $i_{j'}$ to $i'\not=i$ or a term multiplication. In $r(s)_k$, the degree of $x_i$ must be $0$, but to decrease the degree of $x_i$, $s_{j'}$, $j'>j$, will make all terms $0$. Therefore, $q'$ will not contribute to $r(s)_k$. So, the only term of $s_j(q)$ which will contribute to $r(s)_k$ will be $q'$ with degree of $x_i$ equal to $0$, and since $s_j$ is the switch from $i_j$ to $i$, in $s_j(q)$, there will be at most one such $q'$. With this, we can set
\[
s_j(q)=q'
\]
if such $q'$ exists, and $s_j(q)=0$ if not.
Assume the degree of $x_i$ and $x_{i_j}$ are both at least $1$ in $q$. Then, in $s_j(q)$, the degree in $x_i$ of all terms will be at least $1$, so $q$ does not contribute to $r(s)_k$. Now, replace $s_j$ with the change from $i_j$ to $i$. Note that in $\theta(d_{i_j}-C_{i_j,i})(q)=\theta d_{i_j}(q)$, the degree of $x_i$ is at least $1$, and because of this, using the logic from above, after replacing $s_j$ there are also no contributions to $r(s)_k$. Therefore, in this case, $r(s)_k$ is unchanged.
Next, suppose the degree of $x_i$ is $0$ and $x_{i_j}$ is at least $1$ in $q$. In this case, $q'$ exists and will be $\theta d_{i_j}(q)$. But, since the degree of $x_{i_j}$ is at least $1$,
\[
\theta(d_{i_j}-C_{i_j,i})(q) = \theta d_{i_j}(q) = q' = s_j(q).
\]
Therefore, we can replace $s_j$ with $\theta(d_{i_j}-C_{i_j,i})$ and $r(s)_k$ will be the same.
Consider when the degree of $x_{i_j}$ in $q$ is $0$. If the degree of $x_i$ is at least $1$ in $q$, then $q'$ will exist, and be $-\theta C_{i_j,i}(q)$. Otherwise, if the degree of $x_i$ is $0$, $s_j(q)=0=-\theta C_{i_j,i}(q)$. However, since the degree of $x_{i_j}$ in $q$ is $0$,
\[
\theta(d_{i_j}-C_{i_j,i})(q)=-\theta C_{{i_j},i}(q)=s_j(q).
\]
Therefore, after replacing $s_j$ with $\theta(d_{i_j}-C_{i_j,i})$, $r(s)_k$ is unchanged here as well.
\end{proof}
Since $c:H_1\rightarrow H_2$ is a bijection, with \Cref{claim:bijval},
\[
\sum_{s\in H_1} r(s)_k = \sum_{s\in H_2} r(s)_k,
\]
and we are done.
\end{proof}
\begin{theorem}
\label{remainder4}
Suppose that $\lambda$ is a partition with $\ell(\lambda)=m$ and $|\lambda|=k$. Then, for $N\geq m$, if, for $1\leq i\leq N$, $\vec{X}_i=(x_1,\ldots,x_i, 0,\ldots,0)$ is $\vec{x}$ with $x_j=0$ for $i+1\leq j\leq N$,
\begin{align*}
& [1]N^k\prod_{i=1}^m \mathcal{Q}_{i}^N\left(\frac{\partial_i F}{N}\right)^{\lambda_i}(1) \\
& = [1]\prod_{i=1}^m\left((N-i+1)^{\lambda_i}\mathcal{R}_{N-i+1,N-i+1}^{\lambda_i}\left(\frac{\partial_{N-i+1} F(\vec{X}_{N-i+1})}{N-i+1}\right)\right)(1)+R,
\end{align*}
where $R$ is a homogeneous polynomial in the $c_F^\nu$ with degree $k$ and order $N^{k-1}$.
\end{theorem}
\begin{proof}
By symmetry,
\[
[1]N^{k}\prod_{i=1}^m \mathcal{Q}_{i}^N\left(\frac{\partial_i F}{N}\right)^{\lambda_i}(1)
=[1]N^{k}\prod_{i=1}^m \mathcal{Q}_{N-i+1}^N\left(\frac{\partial_{N-i+1} F}{N}\right)^{\lambda_i}(1).
\]
Here, $\mathfrak{r}=\{i_j\}_{1\leq j\leq k}$ consists of $k$ indices, the first $\lambda_1$ being $N$, the next $\lambda_2$ being $N-1$, and so forth, until the last $\lambda_m$ are $N-m+1$. Let $T$ be the set of sequences $s$ with indices $\mathfrak{r}$ over $x_1,\ldots,x_N$ consisting of term multiplications and changes such that:
\begin{itemize}
\item For $1\leq j\leq m-1$, after $\lambda_1+\cdots+\lambda_j$ operators, the next $\lambda_{j+1}$ operators cannot have changes from $N-j$ to $i$ and term multiplications which contain $x_i$, $N-j+1\leq i\leq N$.
\item $r(s)_k$ has degree $0$ in the $x_i$.
\end{itemize}
Note that it is possible for $s\in T$ to have $r(s)_k=0$.
Suppose $0\leq d\leq m$. For a sequence $s$ of term multiplications and changes, let $r_d(s)_i$, $0\leq i\leq k$, be as follows. Set $r_d(s)_0=1$, and for $1\leq i\leq k$, $i\not=\lambda_1+\cdots+\lambda_j$, $1\leq j\leq d$, let $r_d(s)_i=s_i(r_d(s)_{i-1})$. Also, for $1\leq j\leq d$, let \[r_d(s)_{\lambda_1+\cdots+\lambda_j}=(s_{\lambda_1+\cdots+\lambda_j}(r_d(s)_{\lambda_1+\cdots+\lambda_j-1}))|_{x_{N-j+1}=0}.\]
Note that $r_0(s)_i=r(s)_i$, $0\leq i\leq k$.
\begin{claim}
\label{claim:samedeg}
For $0\leq i\leq k$ and any sequence $s$ of term multiplications and changes, there exists a positive integer $n$ such that for $0\leq d\leq m$, $r_d(s)_i$ is homogeneous in $x_1, \ldots, x_N$ with degree $n$ or $r_d(s)_i=0$.
\end{claim}
\begin{proof}
This is clear by induction from $i=0$ to $k$.
\end{proof}
\begin{claim}
\label{claim:equalend}
For all $s\in T$, $r_d(s)_k$ is the same for $0\leq d\leq m$.
\end{claim}
\begin{proof}
To show this, we prove that $r_d(s)_k=r_{d-1}(s)_k$, $1\leq d\leq m$. Observe that $r_d(s)_i=r_{d-1}(s)_i$ for $1\leq i\leq \lambda_1+\cdots+\lambda_d-1$, and
\[
r_{d-1}(s)_k-r_d(s)_k = s_k\circ\cdots\circ s_{\lambda_1+\cdots+\lambda_d+1}(r_{d-1}(s)_{\lambda_1+\cdots+\lambda_d}-r_{d-1}(s)_{\lambda_1+\cdots+\lambda_d}|_{x_{N-d+1}=0}),
\]
where each term of $r_{d-1}(s)_{\lambda_1+\cdots+\lambda_d}-r_{d-1}(s)_{\lambda_1+\cdots+\lambda_d}|_{x_{N-d+1}=0}$ contains $x_{N-d+1}$. But, if $\lambda_1+\cdots+\lambda_d+1\leq i\leq k$, $s_i$ does not have a $x_{N-d+1}$ since $s\in T$, and $s_i$ cannot remove $x_{N-d+1}$, unless all terms become $0$. Then, all nonzero terms of $r_d(s)_k-r_{d-1}(s)_k$ contain $x_{N-d+1}$. Since $s\in T$, $r_0(s)_k=r(s)_k$ has degree $0$ in the $x_i$, so $r_{d-1}(s)_k$ and $r_d(s)_k$, and hence $r_d(s)_k-r_{d-1}(s)_k$, have degree $0$ in the $x_i$ from \Cref{claim:samedeg}. But, all nonzero terms of $r_d(s)_k-r_{d-1}(s)_k$ contain $x_i$, and therefore, $r_d(s)_k-r_{d-1}(s)_k=0$.
\end{proof}
\begin{claim}
\[
[1]\prod_{i=1}^m \left((N-i+1)^{\lambda_i}\mathcal{R}_{N-i+1,N-i+1}^{\lambda_i}\left(\frac{\partial_{N-i+1} F(\vec{X}_{N-i+1})}{N-i+1}\right)\right)(1) = \sum_{s\in T\cap T_N^2(\mathfrak{r})} r(s)_k.
\]
\end{claim}
\begin{proof}
The left hand side of the above expression is
\[
\sum_{s\in T} r_m(s)_k = \sum_{s\in T} r(s)_k = \sum_{s\in T\cap T_N^2(\mathfrak{r})} r(s)_k,
\]
since, by \Cref{claim:equalend}, $r_m(s)_k=r(s)_k$ for $s\in T$, and $r(s)_k\not=0$ if and only if $s\in T\cap T_N^2(\mathfrak{r})$.
\end{proof}
Suppose that $T'$ is the set of $s\in T_N^2(\mathfrak{r})$ satisfying the \Cref{newremainder2} conditions. From the Lemma,
\[
[1]N^k\prod_{i=1}^m \mathcal{Q}_{N-i+1}^N\left(\frac{\partial_{N-i+1} F}{N}\right)^{\lambda_i}(1) = \sum_{s\in T'} r(s)_k + R_1,
\]
where $R_1$ is a homogeneous polynomial in the $c_F^\nu$ with degree $k$ and order $N^{k-1}$.
\begin{claim}
The set $T'$ is a subset of $T$.
\end{claim}
\begin{proof}
It suffices to show that if $s\in T'$, for $d>\lambda_1+\cdots+\lambda_j$, $s_d$ does not contain $x_i$, $N-j+1\leq i\leq N$. For the sake of contradiction, assume $s\in T'$ and $s_d$ contains $x_i$, $N-j+1\leq i\leq N$, for some $d>\lambda_1+\cdots+\lambda_j$. From the conditions of \Cref{newremainder2}, $x_i$ must be in a term multiplication. After, $x_i$ will be in all terms of $r(s)_d$. But, $x_i$ cannot be removed by an operator unless all the terms become $0$, a contradiction to $s\in T_N^2(\mathfrak{r})$.
\end{proof}
Since $T'\subset T_N^2(\mathfrak{r})$ as well, $T'\subset T\cap T_N^2(\mathfrak{r})$. Again, we consider the coefficient of
\[
p=\prod_{i=1}^m c_{F}^{\nu_i}.
\]
Let $U$ and $U'$ be the set of $s$ in $T_N^2(\mathfrak{r})$ and $T'$, respectively, such that $r(s)_k=dp$, $d\not=0$. Then, $U\cap T$ will be the set of $s$ in $T\cap T_N^2(\mathfrak{r})$ such that $r(s)_k=dp$, $d\not=0$. We show that $|U\backslash U'|$ is $O(N^{k-m-1})$; where $U'\subset U\cap T$, $|U\cap T\backslash U'|$ will be $O(N^{k-m-1})$ as well. But, as in \Cref{newremainder2}, this follows from the proof of \Cref{remainder2} without derivates and using changes instead of switches. After, with \Cref{prop:boundedconst},
\[
\sum_{s\in T\cap T_N^2(\mathfrak{r})} r(s)_k = \sum_{s\in T'} r(s)_k + R_2,
\]
where $R_2$ is a homogeneous polynomial in the $c_F^\nu$ with degree $k$ and order $N^{k-1}$, completing the proof.
\end{proof}
\subsection{Proof of \Cref{lln:generalization}}
\label{subsec:llngeneralization}
By \Cref{remainder1}, as $N\rightarrow\infty$,
\[
\lim_{N\rightarrow\infty} \left(\frac{1}{N^{m+|\lambda|}}\sum_{l\in I_N(\lambda)}[1]\mathcal{D}_{l}(F_N(x_1,\ldots,x_N))\right) = \lim_{N\rightarrow\infty}\frac{1}{N^{|\lambda|}}[1]\mathcal{D}_\mathfrak{r}(x_1,\ldots,x_N),
\]
using \pref{eq:remainderto0} with $k=|\lambda|-1$. Afterwards, from \Cref{remainder3} and \Cref{remainder4},
\[
\lim_{N\rightarrow\infty}\frac{1}{N^{|\lambda|}}[1]\mathcal{D}_\mathfrak{r}(x_1,\ldots,x_N) = \lim_{N\rightarrow\infty} [1]\prod_{i=1}^m\left( \mathcal{R}_{N-i+1,N-i+1}^{\lambda_i}\left(\frac{\partial_{N-i+1} F_N(\vec{X}_{N-i+1})}{N-i+1}\right)\right)(1),
\]
where \pref{eq:remainderto0} is used again.
To evaluate this, note that
\begin{equation}
\label{eq:switchderiv}
s_{1,N-i+1}\frac{\partial_{N-i+1} F_N(\vec{X}_{N-i+1})}{N-i+1} = \frac{\partial_1 F_N(\vec{X}_{N-i+1})}{N-i+1},
\end{equation}
and the coefficient of $x_1^dM_\nu(x_2,\ldots,x_{N-i+1})$ in \pref{eq:switchderiv} is $\frac{(d+1)c_{F_N}^{\nu+(d+1)}}{N-i+1}$ for $\nu\in P$ with $\ell(\nu)\leq N-i$. Here, we can have $\nu=0$. But, from \pref{eq:limcoeff},
\[
\lim_{N\rightarrow\infty} \frac{(d+1)c_{F_N}^{\nu+(d+1)}}{N-i+1} = \frac{(d+1)(\ell(\nu)+1)c_{\nu+(d+1)}}{|\nu|+d+1},
\]
and the limiting sequence of \pref{eq:switchderiv} outside of $1$ for $1\leq i\leq m$ is $f=\{c^{d,\nu}\}_{d\geq 0, \nu\in P}$, where $c^{d,\nu}=\frac{(d+1)(\ell(\nu)+1)c_{\nu+(d+1)}}{|\nu|+d+1}$. From \Cref{finalvalue1},
\begin{align*}
\lim_{N\rightarrow\infty} [1]\prod_{i=1}^m\left( \mathcal{R}_{N-i+1,N-i+1}^{\lambda_i}\left(\frac{\partial_{N-i+1} F_N(\vec{X}_{N-i+1})}{N-i+1}\right)\right)(1) =\prod_{i=1}^m\left(\sum_{\substack{\pi\in NC(\lambda_i)}} \prod_{B\in\pi} c_{|B|}(f)\right).
\end{align*}
With this, $\{\mu_N\}_{N\geq 1}$ satisfy a LLN, where
\[
\lim_{N\rightarrow\infty} \mathbb{E}_{\mu_N}\left(\prod_{i=1}^m p_{\lambda_i}^N\right) = \prod_{i=1}^m m_{\lambda_i}, m_k = \sum_{\pi\in NC(k)}\prod_{B\in\pi} c_{|B|}(f)
\]
for $k\geq 1$. Suppose that for $\nu\in P$, for $d\geq 1$, $N_d(\nu)$ of the components of $\nu$ are $d$. Then, if $\nu'+(d)=\nu$, $\ell(\nu)P(\nu')=N_d(\nu)P(\nu)$, and for $l\geq 1$,
\begin{align*}
c_l(f) & = \theta^{l-1}\sum_{\substack{\nu'\in P, d\geq 0,\\ |\nu'|+d=l-1}} (-1)^{\ell(\nu')}P(\nu')\frac{(d+1)(\ell(\nu')+1)c_{\nu'+(d+1)}}{|\nu'|+d+1} \\
& = \theta^{l-1} \sum_{\nu\in P, |\nu|=l} (-1)^{\ell(\nu)-1}\frac{\ell(\nu)c_\nu}{|\nu|}\sum_{\substack{\nu'\in P, d\geq 1,\\ \nu'+(d)=\nu}}dP(\nu')\\
& = \theta^{l-1} \sum_{\nu\in P, |\nu|=l} (-1)^{\ell(\nu)-1}\frac{P(\nu)c_\nu}{|\nu|}\sum_{\substack{\nu'\in P, d\geq 1,\\ \nu'+(d)=\nu}}dN_d(\nu) \\
& = \theta^{l-1} \sum_{\nu\in P, |\nu|=l} (-1)^{\ell(\nu)-1} P(\nu)c_\nu,
\end{align*}
which gives \pref{eq:llnmoments}, as required.
\section{Probability Measures}
\label{sec:measures}
\subsection{Bessel Generating Functions}
A $\textit{Gelfand-Tsetlin pattern}$ with $N$ rows is a set of real numbers $\{\lambda_i^j\}_{1\leq i\leq j\leq N}$ such that $\lambda_i^{j+1}\leq \lambda_i^j\leq \lambda_{i+1}^{j+1}$ for $1\leq i\leq j\leq N-1$. We let $\mathcal{G}_N$ denote the set of Gelfand-Tsetlin patterns with $N$ rows.
\begin{definition}
\label{def:multivariate}
The $\textit{multivariate Bessel function}$ for $a_1<\cdots<a_N$ is
\begin{align*}
B_{(a_1, \ldots, a_N)}(x_1, \ldots, x_N; \theta) = &
\int_{\substack{d=\{y_i^k\}\in\mathcal{G}_N, \\ y_i^N = a_i, 1\leq i\leq N}}\left[\exp\left(\sum_{k=1}^N x_k\cdot\left(\sum_{i=1}^k y_i^k - \sum_{j=1}^{k-1} y_j^{k-1}\right)\right)\right]\cdot \\ & f_{(a_1, \ldots, a_N)}(d)\prod_{1\leq i\leq k\leq N-1} dy_i^k,
\end{align*}
where for $d= \{y_i^k\}\in\mathcal{G}_N$ such that $y_i^N = a_i, 1\leq i\leq N$,
\begin{align*}
f_{(a_1, \ldots, a_N)}(d)= & \prod_{k=1}^N\frac{\Gamma(k\theta)}{\Gamma(\theta)^k}\prod_{1\leq i<j\leq N} (a_j-a_i)^{1-2\theta}\cdot \\
& \prod_{k=1}^{N-1}\left[\prod_{1\leq i<j\leq k}(y_j^k-y_i^k)^{2-2\theta}\prod_{a=1}^k\prod_{b=1}^{k+1}|y_a^k-y_b^{k+1}|^{\theta-1}\right].
\end{align*}
\end{definition}
\begin{definition}
\label{def:BGF}
The \textit{Bessel generating function} of $\mu\in\mathcal{M}_N$ is defined as
\begin{equation}
\label{eq:BGF}
G_\theta(x_1, \ldots, x_N; \mu):= \int_{a_1\leq\cdots\leq a_N} B_{(a_1,\ldots, a_N)}(x_1, \ldots, x_N; \theta)\mu(da_1, \ldots, da_N).
\end{equation}
\end{definition}
Observe that $f_{(a_1, \ldots, a_N)}(d)$ is the probability density of the $\theta$-corners process or $\beta$-corners process with top row $a_1<\cdots<a_N$ in \cite{crystalrandommat}, \cite{matrix}. Note that this is also referred to as the orbital $\beta$ process in \cite{betaprocess}. A scaled multivariate Bessel function and the Bessel generating function for $\theta=1$ are studied in \cite{Airy}. The following result is well known.
\begin{theorem}
\label{thm:operator}
For any symmetric polynomial $P(x_1, \ldots, x_N)$ and each $N$-tuple of reals $(a_1< a_2<\cdots< a_N)$,
\begin{equation}
\label{eq:symmetricpoly}
P(\mathcal{D}_1, \ldots, \mathcal{D}_N)B_{(a_1, \ldots, a_N)} = P(a_1, \ldots, a_N)B_{(a_1, \ldots, a_N)}.
\end{equation}
\end{theorem}
From \Cref{lemma:convergence}, if $\mu\in\mathcal{M}_N$ is exponentially decaying, then the BGF of $\mu$ converges in an open neighborhood of the origin in $\mathbb{C}^N$. Additionally, from \pref{eq:symmetricpoly},
\[
\mathcal{P}_kB_{(a_1, \ldots, a_N)} = \left(\sum_{i=1}^Na_i^k\right)B_{(a_1, \ldots, a_N)},
\]
which can be used to prove the following Proposition.
\begin{proposition}[\cite{matrix}*{Proposition 2.11}]
\label{prop:expectedvalue}
For a positive integer $s$, let $k_1, \ldots, k_s$ be positive integers. Suppose $\mu\in\mathcal{M}_N$ is exponentially decaying. Then,
\begin{equation}
\label{eq:expected}
\left(\prod_{i=1}^s \mathcal{P}_{k_i}\right) G_\theta(x_1,\ldots, x_N; \mu)\bigg|_{x_i=0, 1\leq i\leq N} = \mathbb{E}_{\mu}\left(\prod_{i=1}^s\left(\sum_{j=1}^N a_j^{k_i}\right)\right).
\end{equation}
\end{proposition}
\subsection{Setup}
\label{subsec:setup}
In \Cref{lln}, let $F_N(x_1,\ldots,x_N)= \ln(G_N(x_1,\ldots,x_N;\theta))$ for $N\geq 1$. Note that for $\nu\in P$ with $|\nu|\geq 1$, if $i_1, \ldots, i_r$ are positive integers such that $\sigma((i_1,\ldots,i_r))=\nu$, then
\[
\frac{\partial}{\partial x_{i_1}}\cdots\frac{\partial}{\partial x_{i_r}} \ln(G_N^\theta)\bigg|_{x_i=0,1\leq i\leq N} = \frac{|\nu|!}{P(\nu)}\cdot c_{F_N}^\nu.
\]
Also, from \Cref{prop:expectedvalue}, for exponentially decaying $\mu\in\mathcal{M}_N$, for a partition $\lambda$ with $|\lambda|\geq 1$,
\begin{equation}
\label{eq:llnev}
\begin{split}
\mathbb{E}_{\mu} \left(\prod_{i=1}^{\ell(\lambda)} p_{\lambda_i}^N\right) & =\mathbb{E}_{\mu}\left(\prod_{i=1}^{\ell(\lambda)}\left(\frac{1}{N}\sum_{j=1}^N \left(\frac{a_j}{N}\right)^{\lambda_i}\right)\right) \\
& = \frac{1}{N^{\ell(\lambda)+|\lambda|}}\left(\prod_{i=1}^{\ell(\lambda)} \mathcal{P}_{\lambda_i}\right) G_\theta(x_1,\ldots, x_N; \mu)\bigg|_{x_i=0, 1\leq i\leq N}.
\end{split}
\end{equation}
To get that $\{\mu_N\}_{N\geq 1}$ satisfy a LLN, we use \pref{eq:llnev} with $\mu=\mu_N$ and take the $N\rightarrow\infty$ limit.
\begin{proposition}[\cite{matrix}*{Lemma 5.2}]
\label{prop:equalformal}
Suppose that $F$ is a $(k+1)$-times continuously differentiable function on a neighborhood of $(0,\ldots,0)\in\mathbb{C}^N$, with Taylor series expansion
\[
F(x_1,\ldots,x_N)=\sum_{\nu, |\nu|\leq k, \ell(\nu)\leq N} c_F^\nu M_{\nu}(\vec{x}) + O(\norm{x}^{k+1}).
\]
Then, for any $\lambda=(\lambda_1, \ldots, \lambda_m)$ with $|\lambda|=k$,
\[\left(\prod_{i=1}^m\mathcal{P}_{\lambda_i}\right)\exp(F(x_1,\ldots,x_N))\bigg|_{x_i=0, 1\leq i\leq N} = \left(\prod_{i=1}^m\mathcal{P}_{\lambda_i}\right)\exp(\tilde{F}(x_1,\ldots,x_N))\bigg|_{x_i=0, 1\leq i\leq N},\]
where
\[
\tilde{F}(x_1,\ldots,x_N)=\sum_{\nu, |\nu|\leq k, \ell(\nu)\leq N} c_F^\nu M_{\nu}(\vec{x}).
\]
\end{proposition}
Suppose $\mu\in\mathcal{M}_N$ is exponentially decaying. Then, from \Cref{lemma:convergence}, $G_\theta(x_1,\ldots,x_N;\mu)$ and therefore $F(x_1,\ldots,x_N)=\ln(G_\theta(x_1,\ldots,x_N;\mu))$ are symmetric as and holomorphic in a neighborhood of the origin. Moreover, $F(0,\ldots,0)=\ln(G_\theta(0,\ldots,0;\mu))=0$.
\subsection{Proof of \Cref{lln}}
\label{subsec:mainproof}
Suppose we have a symmetric polynomial $\tilde{F}(x_1,\ldots,x_N)$ with $\tilde{F}(0,\ldots,0)=0$. We have that
\begin{align*}
& \frac{\partial}{\partial x_i}R(x_1,\ldots,x_N)\exp(\tilde{F}(x_1,\ldots,x_N)) \\
& = \left(\frac{\partial}{\partial x_i}R(x_1,\ldots,x_N)+R(x_1,\ldots,x_N)\frac{\partial}{\partial x_i}\tilde{F}(x_1,\ldots,x_N)\right)\exp(\tilde{F}(x_1,\ldots,x_N)).
\end{align*}
Moreover, because $\tilde{F}(x_1,\ldots,x_N)$ is symmetric,
\[
s_{i,j}(R(x_1,\ldots,x_N)\exp(\tilde{F}(x_1,\ldots,x_N))) = (s_{i,j}R(x_1,\ldots,x_N))\exp(\tilde{F}(x_1,\ldots,x_N)).
\]
Also, since $\tilde{F}(0,\ldots,0)=0$, $R(x_1,\ldots,x_N)\exp(\tilde{F}(x_1,\ldots,x_N))$ at $x_i=0$, $1\leq i\leq N$ is $R(0,\ldots,0)$. From above, where $\mathfrak{r}=\{i_j\}_{1\leq j\leq k}$
\begin{align*}
\left(\prod_{j=1}^k\mathcal{D}_{i_j}\right)\exp(\tilde{F}(x_1,\ldots,x_N)) & = \left(\prod_{j=1}^k \left(\mathcal{D}_{i_j} + \frac{\partial}{\partial x_{i_j}}\tilde{F}(x_1,\ldots,x_N)\right)\right)(1)\cdot\exp(\tilde{F}(x_1,\ldots,x_N))\\
& = \mathcal{D}_{\mathfrak{r}}(\tilde{F}(x_1,\ldots,x_N))\cdot\exp(\tilde{F}(x_1,\ldots,x_N)),
\end{align*}
and
\[
\left(\prod_{j=1}^k\mathcal{D}_{i_j}\right)\exp(\tilde{F}(x_1,\ldots,x_N))\bigg|_{x_i=0,1\leq i\leq N} = [1]\mathcal{D}_{\mathfrak{r}}(\tilde{F}(x_1,\ldots,x_N)).
\]
Note that since $\tilde{F}(x_1,\ldots,x_N)$ is a polynomial, $\mathcal{D}_{\mathfrak{r}}(\tilde{F}(x_1,\ldots,x_N))$ is also a polynomial. Also, for a partition $\lambda$ with $\ell(\lambda)=m$, we have that
\begin{equation}
\label{eq:conversion}
\left(\prod_{i=1}^m \mathcal{P}_{\lambda_i}\right)\exp(\tilde{F}(x_1,\ldots,x_N))\bigg|_{x_i=0, 1\leq i\leq N} = \sum_{l\in I_N(\lambda)} [1]\mathcal{D}_l(\tilde{F}(x_1,\ldots,x_N)).
\end{equation}
Next, let $\lambda$ be a partition of length $m$ with $|\lambda|\geq 1$. Suppose $\mathfrak{r}=\{i_j\}_{1\leq j\leq |\lambda|}$, where the first $\lambda_1$ indices are $1$, the next $\lambda_2$ are $2$, and so forth, until the last $\lambda_m$ are $m$. For \Cref{lln}, we look at the limit of \pref{eq:llnev} for $\mu=\mu_N$ as $N\rightarrow\infty$. For $N\geq 1$, we let $F_N(x_1,\ldots,x_N)=\ln(G_\theta(x_1,\ldots,x_N; \mu_N)$. For $N\geq 1$, because $\mu_N$ is exponentially decaying, we see that $F_N(x_1,\ldots,x_N)$ is symmetric and holomorphic in a neighborhood of the origin. Then, we can use \Cref{prop:equalformal} on $F_N(x_1,\ldots,x_N)$.
Suppose that $\tilde{F}_N(x_1,\ldots,x_N)$ is the polynomial for $F_N(x_1,\ldots,x_N)$ from \Cref{prop:equalformal} with $k=|\lambda|$. By \Cref{prop:equalformal}, \pref{eq:llnev}, and \pref{eq:conversion},
\begin{align*}
\mathbb{E}_{\mu_N}\left(\prod_{i=1}^m p_{\lambda_i}^N\right) & = \frac{1}{N^{m+|\lambda|}}\left(\prod_{i=1}^m\mathcal{P}_{\lambda_i}\right)\exp(F_N(x_1,\ldots,x_N))\bigg|_{x_i=0, 1\leq i\leq N} \\
& = \frac{1}{N^{m+|\lambda|}}\left(\prod_{i=1}^m\mathcal{P}_{\lambda_i}\right)\exp(\tilde{F}_N(x_1,\ldots,x_N))\bigg|_{x_i=0, 1\leq i\leq N} \\
& = \frac{1}{N^{m+|\lambda|}} \sum_{l\in I_N(\lambda)}[1]\mathcal{D}_{l}(\tilde{F}_N(x_1,\ldots,x_N)).
\end{align*}
Where $\tilde{F}_N(x_1,\ldots,x_N)$, a symmetric polynomial, is a symmetric formal series for $N\geq 1$, with the conditions in \Cref{lln}, we can use \Cref{lln:generalization} on $\{\tilde{F}_N(x_1,\ldots,x_N)\}_{N\geq 1}$. Then, with the above expression, we are done.
\section{Eigenvalue Distribution of Random Matrices}
\label{sec:eigenvalue}
From \cite{AGZ}, for all $\beta>0$, consider the measure, known as the $\beta$-Hermite ensemble, on $\mathcal{M}_N$ with probability density
\begin{equation}
\label{eq:betahermite}
d_{N, \beta}(x_1, \ldots, x_N) = C_{N, \beta}\prod_{1\leq i<j\leq N} |x_i-x_j|^\beta\prod_{i=1}^N e^{-\frac{\beta x_i^2}{4}},
\end{equation}
where
\[
C_{N,\beta} = (2\pi)^{-\frac{N}{2}}\left(\frac{\beta}{2}\right)^{\frac{\beta N(N-1)}{4}+\frac{N}{2}}\Gamma\left(\frac{\beta}{2}\right)^N\left(\prod_{i=1}^N\Gamma\left(\frac{i\beta}{2}\right)\right)^{-1}.
\]
In the paper, we scale $d_{N,\beta}(x_1,\ldots,x_N)$ by $\sqrt{N}$ to get the probability density
\begin{equation}
\label{eq:betahermitescaled}
d^s_{N,\beta}(x_1,\ldots,x_N) = \frac{C_{N, \beta}}{N^{\frac{\beta N(N-1)}{4}+\frac{N}{2}}}\prod_{1\leq i<j\leq N} |x_i-x_j|^\beta\prod_{i=1}^N e^{-\frac{\beta x_i^2}{4N}}
\end{equation}
on $\mathcal{M}_N$. For $\beta=1,2,$ and $4$, $d_{N,\beta}$ is the probability density of the ordered eigenvalues of the GUE, GOE, and GSE, respectively. Also, from \cite{AGZ}, there exist random tridiagonal matrices with eigenvalue distribution being the $\beta$-Hermitian ensemble for $\beta>0$.
Suppose $\{\mu_N\}_{N\geq 1}$ are a sequence of probability measures such that $\mu_N$ is in $\mathcal{M}_N$ and has density $d^s_{N,\beta}$. In \Cref{scaledlln}, we see that if $\beta>0$ is fixed, \Cref{lln} on $\{\mu_N\}_{N\geq 1}$ gives the moments as well as LLN satisfaction. This is an example of how the results of this paper can be used on probability measures.
\begin{proposition}[\cite{betaprocess}*{Corollary 3.7}]
\label{prop:scaledbessel}
Let $a_1,\ldots, a_N, y_1, \ldots, y_N, c$ be $2N+1$ arbitrary complex numbers. Then,
\[
B_{(ca_1,\ldots,ca_N)}(y_1,\ldots,y_N)=B_{(a_1,\ldots,a_N)}(cy_1,\ldots,cy_N).
\]
\end{proposition}
\begin{proposition}
\label{prop:scaledbgf}
Suppose $\mu\in\mathcal{M}_N$ is exponentially decaying. For $c>0$, let $c\mu$ be $\mu$ scaled by $c$. Then,
\[
G_\theta(x_1,\ldots,x_N; c\mu) = G_\theta(cx_1,\ldots,cx_N;\mu).
\]
\begin{proof}
The probability density of $c\mu$ is $c\mu(a_1,\ldots,a_N)=\frac{1}{c^N}\mu\left(\frac{a_1}{c},\ldots,\frac{a_N}{c}\right)$. With \Cref{prop:scaledbessel}, where $a'=(a_1'\leq\cdots\leq a_N')$, $a_i'=\frac{a_i}{c}$, $1\leq i\leq N$,
\begin{align*}
G_\theta(x_1,\ldots,x_N; c\mu) & = \int_{a=(a_1\leq\cdots\leq a_N)} B_a(x_1,\ldots,x_N;\theta) c\mu(a_1,\ldots,a_N)da_1\cdots da_N \\
& = \int_{a'=(a_1'\leq\cdots\leq a_N')} B_{ca'}(x_1,\ldots,x_N;\theta) c^Nc\mu(ca_1',\ldots, ca_N') da_1'\cdots da_N' \\
& = \int_{a'=(a_1'\leq\cdots\leq a_N')} B_{a'}(cx_1,\ldots,cx_N;\theta) \mu(a_1',\ldots,a_N')da_1'\cdots da_N' \\
& = G_\theta(cx_1,\ldots,cx_N;\mu).
\end{align*}
\end{proof}
\end{proposition}
\begin{proposition}
\label{scaledlln}
Suppose $\theta>0$ is fixed. Consider the sequence $\{\mu_N\}_{N\geq 1}$ of probability measures such that for all positive integers $N$, $\mu_N\in\mathcal{M}_N$ and $\mu_N$ has probability density $d_{N,2\theta}^s$ from \pref{eq:betahermitescaled}. Then, $\{\mu_N\}_{N\geq 1}$ satisfy a LLN, with $m_{2k-1}=0$ and $m_{2k}=\frac{1}{k+1}\binom{2k}{k}$ for $k\geq 1$.
\end{proposition}
\begin{proof}
For $N\geq 1$, consider $\frac{\mu_N}{\sqrt{N}}$, which has density $d_{N,2\theta}$ in \pref{eq:betahermite}. It is easy to see that $\mu_N$ and $\frac{\mu_N}{\sqrt{N}}$ are exponentially decaying. From \cite{betaprocess}*{Proposition 4.2},
\begin{equation}
\label{eq:bgfbeta}
G_\theta\left(x_1,\ldots,x_N;\frac{\mu_N}{\sqrt{N}}\right)=\exp\left(\frac{1}{2\theta}\sum_{i=1}^N x_i^2\right).
\end{equation}
Then, from \Cref{prop:scaledbgf} and \pref{eq:bgfbeta},
\[
G_\theta(x_1,\ldots,x_N;\mu_N)=G_\theta\left(\sqrt{N}x_1, \ldots, \sqrt{N}x_N;\frac{\mu_N}{\sqrt{N}}\right)=\exp\left(\frac{N}{2\theta}\sum_{i=1}^N x_i^2\right).
\]
Afterwards, by \Cref{lln}, $\{\mu_N\}_{N\geq 1}$ satisfy a LLN with $c_{(2)}=\frac{1}{\theta}$ and $c_{\nu}=0$ for other $\nu\in P$, $|\nu|\geq 1$. For $k\geq 1$, $m_{2k-1}=0$ and $m_{2k}=|T(k)|$, where $T(k)$ is the set of $\pi\in NC(2k)$ which have all blocks of size $2$. But, $|T(k)|=\frac{1}{k+1}\binom{2k}{k}$, the $k$th Catalan number.
\end{proof}
From \Cref{scaledlln}, if $(a_1\leq\cdots\leq a_N)$ are distributed with density $d_{N,\beta}(x_1,\ldots,x_N)$, for positive integer $s$ and positive integers $k_i$, $1\leq i\leq s$,
\[
\lim_{N\rightarrow\infty} \mathbb{E}_{\mu_N}\left(\prod_{i=1}^s \left(\frac{1}{N}\sum_{i=1}^N\left(\frac{a_i}{\sqrt{N}}\right)^{k_i}\right)\right) = \prod_{i=1}^s m_{k_i},
\]
with the moments $m_k$ given in \Cref{scaledlln}. Sometimes, as seen above, for sequences $\{\mu_N\}_{N\geq 1}$ of exponentially decaying probability measures, for $N\geq 1$, we can scale $\mu_N$ by an appropriate power of $N$ and use \Cref{prop:scaledbgf} to satisfy the conditions of \Cref{lln}.
|
\section{Introduction}
The relativistic plasmas is found in the astrophysical objects and in the laboratory experiments on the laser-surface interaction and thermonuclear devises.
The role of the relativistic effects is also important for the quantum plasmas
\cite{Hakim book Rel Stat Phys},
\cite{Mahajan PoP 2002},
\cite{Shatashvili ASS 97},
\cite{Shatashvili PoP 99},
\cite{Shatashvili PoP 20},
\cite{Weinberg Gr 72},
\cite{Mahajan PoP 2011},
\cite{Comisso PRL 14},
\cite{Mahajan PRL 03},
\cite{Mahajan PRL 08}.
The degenerate electrons demonstrates the noticeable quantum effects in the relativistic regime
\cite{Andreev PRE 15 SEAW}, \cite{Zhu PRE 10},
\cite{Andreev JPP 21}, \cite{Shukla UFN 10}.
Therefore, the full relativistic model requires the quantum effects as well.
However, this paper is focused on the classical plasmas and the relativistic temperature effects.
The hydrodynamic description of relativistic plasmas with the relativistic temperatures
(so at least one species has relativistic temperature)
is under consideration,
so the temperature (in the energy units) of electrons (like the lightest particle in the plasma)
is comparable with the rest energy of the electron $m_{e}c^{2}$,
where $m_{e}$ is the mass of electron, and $c$ is the speed of light.
One of simplest derivations of the relativistic hydrodynamics of the ideal liquid starts from the energy-momentum tensor in the rest frame
\cite{Landau v2}, \cite{Landau v6}:
$T^{\alpha\beta}=(\epsilon+\tilde{p})u^{\alpha}u^{\beta}-\tilde{p} g^{\alpha\beta}$,
where the energy density $\epsilon$,
the isotropic pressure $\tilde{p}$ are combined into the enthalpy $w=\epsilon+\tilde{p}$,
and the following components of the four-velocity field in the rest frame $u^{0}=1$, $u^{i}=0$.
Here and below the Greek letters denote the four dimensions $\alpha=0$, $1$, $2$, $3$,
and the Latin letters denote three dimensional (spatial) indexes $i=1$, $2$, $3$.
Knowledge of the energy-momentum tensor leads to the following equation of motion \cite{Landau v6}
\begin{equation}\label{RHD2018Cl d T a b 0} \partial_{\beta}T^{\alpha\beta}=0.\end{equation}
If we consider the charged particles interacting with the electromagnetic field
we need to extend equation (\ref{RHD2018Cl d T a b 0})
\begin{equation}\label{RHD2018Cl d T a b F} \partial_{\beta}T^{\alpha\beta}=q n F^{\alpha\beta}u^{\beta},\end{equation}
where $q$ is the charge of particle,
$n$ is the concentration,
and
$F^{\alpha\beta}$ is the tensor of electromagnetic field.
Here the rest frame is defined for the macroscopically motionless fluid.
However, the rest frame can be defined in the presence of the local macroscopic flows
or/and turbulence.
Next step is generalization of the presented in the rest frame hydrodynamic equations to the arbitrary inertial frame.
It includes the Lorentz transformation for the representation of hydrodynamic equations
(or the energy-momentum tensor) in the arbitrary frame,
where the arbitrary frame moves with the velocity $u^{i}$ relatively the rest frame.
Apparently, the velocity of the frame $u^{i}$ is not related to the local flows
which can exist in the fluid and described by the velocity field.
Moreover, it is essential to point out that
velocity $u^{i}$ is a constant.
It does not depend on the position
while the velocity field of the fluid is a function of space.
The following remark is in order.
The application of the described method locally cannot be done,
since the Lorentz transformation is a global transformation.
Described transformation is useful for the study of beams or global flows \emph{with nonrelativistic temperatures}.
It can be assumed that
the energy density and the pressure in the energy-momentum tensor correspond to the relativistic temperatures via suitable equations of state.
Since the Lorentz transformation makes the motion of the whole system or the motion of the observer relatively medium and does not related to the appearance of the local flows.
The energy-momentum tensor $T^{\alpha\beta}$ is constructed of all possible tensors and vectors in the isotropic fluid
which are the Kronecker symbol $\delta^{\alpha\beta}$ and the velocity vector $v^{\alpha}$.
However, the presented below analysis of the relativistically hot plasmas demonstrates presence of another nontrivial four vector.
Hence, the energy-momentum tensor will be reconsidered after the introduction of our model.
The comments described above show some problems of the hydrodynamic model constructed phenomenologically on the macroscopic scale.
Anyway, present level of knowledge requires some microscopic justification of the macroscopic models.
Sometimes authors refer to the kinetic theory as the microscopic model.
Here, we mean that
the microscopic level is the scale,
where motion of each particle is distinguishable
(the scale of electrons and protons for the hydrogen plasmas).
From this point of view the kinetics is the macroscopic method of description formulated in the six dimensional space of coordinates and momentums.
In our derivation we avoid derivation of the kinetic model as an intermediate stage.
We directly derive the hydrodynamics from the microscopic motion of particles.
This paper is organized as follows.
In Sec. II derivation of the relativistic hydrodynamic model based on exact microscopic motion of particles is shown.
The basic definitions and method of derivation are demonstrated.
In Sec. III the suggested relativistic hydrodynamic model is represented in terms of the velocity field.
In Sec. IV method of derivation of equations of state is described to make truncation of the set of equations.
In Sec. V necessity of the covariant structure of hydrodynamic equations.
In Sec. VI the microscopic structure of the energy-momentum tensor appearing via the microscopic derivation is demonstrated and discussed.
In Sec. VII the spectrum of high-frequency longitudinal plane waves in the relativistically hot isotropic plasmas.
In Sec. VIII the zero-temperature limit of the model is found to demonstrate its agreement with well-known results
for the relativistic beams in the low temperature plasmas.
In Sec. IX brief discussion of the plane waves in the relativistically hot magnetized plasmas is given.
In Sec. X a brief summary of obtained results is presented.
\section{Model}
Our goal is the derivation of the macroscopic hydrodynamic equations for the relativistic plasmas
with the relativistically large temperatures.
We are going to trace exact microscopic evolution of particles obeying the relativistic modification of the Newton equations of motion.
\subsection{Basic definitions}
We start our derivation with the definition of the concentration of particles.
In classical physics the particle is modeled as the point-like object.
Hence, its mathematical representation is the delta-function
(one particle in the zero volume space gives infinite particle number density in one point and zero density in other point).
Therefore, the microscopic concentration of the classical system of particles is the sum of delta functions:
\begin{equation}\label{RHD2018Cl concentration mic}
n_{mic}(\textbf{r},t)=\sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i}(t)),\end{equation}
where the subindex \textit{mic} refers to the microscopic definition of the concentration of particles.
Function $\textbf{r}_{i}(t)$ is the radius vector of $i$-th particle.
Its evolution happens in accordance with the Newton equation of motion,
where the interaction with all surrounding particles is included.
Presented here microscopic definition of concentration evolve with accordance with exact microscopic evolution of particles.
So, if one wants to find its evolution one need to solve set of Newton equations of motion.
Our goal is to construct approximate macroscopic model to capture main features of the relativistic plasmas.
However, if we consider the macroscopic theory
we need to introduce the macroscopically infinitesimal element of volume $\Delta$
and present the number of particles in this element of volume
\begin{equation}\label{RHD2018Cl concentration definition}
n(\textbf{r},t)\equiv n_{mac}(\textbf{r},t)=\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}
\sum_{i=1}^{N}\delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t)). \end{equation}
The presented integral does not directly related to the simplified point-like structure of particles.
It counts number of particles in the chosen vicinity $\Delta$ of the chosen point $\textbf{r}$
since the integral of the delta function is equal to one if the particle is in the vicinity
and the integral of the delta function is equal to zero if the particle is outside of the vicinity.
Equation (\ref{RHD2018Cl concentration definition}) can be used to the center of mass of finite size objects.
However, for the point-like objects,
equation (\ref{RHD2018Cl concentration definition}) is equivalent to microscopic concentration (\ref{RHD2018Cl concentration mic}).
The macroscopic concentration can be rewritten in different form
which can provide more physical inside of the presented definition (\ref{RHD2018Cl concentration definition}):
\begin{equation}\label{RHD2018Cl concentration REdefinition}
n(\textbf{r},t)=\frac{1}{m\Delta} \sum_{i=1}^{N(\textbf{r},t)}m_{i}, \end{equation}
where all $m_{i}$ are equal to each other
(the concentration is defined for each species in plasmas).
Function $N(\textbf{r},t)$ is the number of particles in the $\Delta$-vicinity of point $\textbf{r}$ at the fixed moment in time $t$.
These definitions (\ref{RHD2018Cl concentration definition}) and (\ref{RHD2018Cl concentration REdefinition}) are equivalent to each other,
but (\ref{RHD2018Cl concentration REdefinition}) less useful for the derivation of the continuity equation
since the unknown function $N(\textbf{r},t)$ is present in the upper limit of summation.
The integral operator in equation (\ref{RHD2018Cl concentration definition})
\begin{equation}\label{RHD2018Cl concentration definition}
\int_{\Delta}d\mbox{\boldmath $\xi$}
\sum_{i=1}^{N}\delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t))=N(\textbf{r},t) \end{equation}
counts the number of particles $N(\textbf{r},t)$ following their exact motion in accordance with the trajectory of each particle $\textbf{r}_{i}(t)$.
Equation (\ref{RHD2018Cl concentration REdefinition}) can be rewritten with no application of mass of particles $m_{i}$
via the count of units:
$n(\textbf{r},t)=\sum_{i=1}^{N(\textbf{r},t)}1_{i}/\Delta=N(\textbf{r},t)/\Delta$.
At initial step we have no conditions on volume $\Delta$.
If we need to present a macroscopic theory
we should have $\Delta$ large enough
so $\Delta$-vicinity of each point of space contains macroscopically large number of particles.
So, $\Delta$-volume can play the role of the macroscopically infinitesimal element of space.
On the other hand we can show transition from equation (\ref{RHD2018Cl concentration definition})
to equation (\ref{RHD2018Cl concentration mic}) at $\Delta\rightarrow0$.
If the point of space $\textbf{r}$ contains a particle (or several point-like particles)
we have $n(\textbf{r},t)=1/(\Delta\rightarrow0)\rightarrow\infty$.
If the point of space $\textbf{r}$ does not contain any particle
we find $n(\textbf{r},t)=lim_{\Delta\rightarrow0}(0/\Delta)=0$.
So, we have $N$ delta functions giving $N$ infinite values at points $\textbf{r}=\textbf{r}_{i}(t)$ for $i\in[1,N]$
and zero values in other points.
So, we have same distribution as in equation (\ref{RHD2018Cl concentration mic}).
A possible candidate for the concentration definition is
\begin{equation}\label{RHD2018Cl concentration a possible definition} \tilde{n}(\textbf{r},t)=\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N}\gamma_{i}\delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t)), \end{equation}
where $\gamma_{i}=(1-\textbf{v}_{i}^{2}(t)/c^{2})^{-1/2}$
with $\textbf{v}_{i}(t)=d\textbf{r}_{i}(t)/dt$ the velocity of $i$th particle.
However, function $\tilde{n}$ is proportional to the energy density $\tilde{n}=\varepsilon(\textbf{r},t)/mc^{2}$,
hence it obviously does not satisfy the continuity equation.
Presented method is the three-dimensional reduction of the method of microscopic derivation of the relativistic kinetics \cite{Kuz'menkov 91}.
\subsection{Continuity equation}
Analysis of the concentration dynamics can be obtained without the equation of motion
since it is defined by the kinematic effects
$$\partial_{t}n(\textbf{r},t)
=\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} \partial_{t}\delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t))$$
\begin{equation}\label{RHD2018Cl concentration derivative}
=\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} (-\textbf{v}_{i}(t))
\cdot\partial_{\textbf{r}}\delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t)).
\end{equation}
The derivative on the space variable $\partial_{\textbf{r}}$ can be placed outside of the integral in the last term.
Hence, the time evolution of the concentration leads to the appearance of the particles current:
\begin{equation}\label{RHD2018Cl particle current definition}
\textbf{j}(\textbf{r},t)=\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} \textbf{v}_{i}(t)
\delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t)).
\end{equation}
Summing up presented results as the well-known continuity equation:
\begin{equation}\label{RHD2018Cl continuity equation} \partial_{t}n+\nabla\cdot\textbf{j}=0. \end{equation}
There is no statistical averaging in the described prescription.
Moreover, it is not necessary to use the statistics
while we trace the microscopic motion itself.
However, some short notations (which can remind statistical physics) are in order.
For example, the particle current appears as
the action of operator
\begin{equation}\label{RHD2018Cl brackets}\langle ... \rangle\equiv
\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} ... \delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t))\end{equation}
on the velocities of particles.
Therefore, the particle current can be written in a short form: $\textbf{j}=\langle \textbf{v}_{i}\rangle$.
The nonrelativistic hydrodynamics is considered within this method in Ref. \cite{Drofa TMP 96}.
It is also briefly discussed in Ref. \cite{Andreev PIERS 2012}.
\subsection{Equation for the current evolution}
Derivation of the other equations including the current evolution equation
requires equation of motion for each particle
to trace their exact motion in terms of collective (macroscopic) variables.
we use the relativistic Newton equations in terms of the velocity evolution
(or the acceleration caused by the interaction) \cite{Landau v2} (see section 17):
\begin{equation}\label{RHD2018Cl Newton eq} \dot{\textbf{v}}_{i}=\frac{e_{i}}{m_{i}}\sqrt{1-\frac{\textbf{v}_{i}^{2}}{c^{2}}} \biggl[\textbf{E}_{i}+\frac{1}{c}[\textbf{v}_{i}\times\textbf{B}_{i}]-\frac{1}{c^{2}}\textbf{v}_{i}(\textbf{v}_{i}\cdot\textbf{E}_{i})\biggr], \end{equation}
where $\textbf{v}_{i}=\textbf{v}_{i}(t)$,
$\textbf{E}_{i}=\textbf{E}(\textbf{r}_{i}(t),t)$ and
$\textbf{B}_{i}=\textbf{B}(\textbf{r}_{i}(t),t)$ are the electric and magnetic fields
in the point $\textbf{r}_{i}(t)$ and at time $t$ acting on $i$th particle.
We consider the evolution of the particle current $\textbf{j}(\textbf{r},t)$.
To this end we calculate the time derivative of the current:
$$\partial_{t}j^{a}(\textbf{r},t)
=\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} v_{i}^{a}(t)\partial_{t}\delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t))$$
\begin{equation}\label{RHD2018Cl current derivative}
+\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} \dot{v}_{i}^{a}(t)\delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t)).
\end{equation}
We present the presented equation introducing the flux of the particle current $\Pi^{ab}$ and the force field $F^{a}$
\begin{equation}\label{RHD2018Cl current derivative Euler equation}
\partial_{t}j^{a}+\partial_{b}\Pi^{ab}=F^{a},
\end{equation}
where
\begin{equation}\label{RHD2018Cl def Pi}
\Pi^{ab}=\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} v_{i}^{a}(t)v_{i}^{b}(t)\delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t)),
\end{equation}
and
\begin{equation}\label{RHD2018Cl force field via acceleration}
F^{a}=\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} \dot{v}_{i}^{a}(t)\delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t)).
\end{equation}
Strictly speaking the introduced force field $F^{a}$ is not exactly the force field,
since the force causes the change of momentum $\dot{\textbf{p}}_{i}=\mathcal{F}$.
While $F^{a}$ is the acceleration of particles given by equation (\ref{RHD2018Cl Newton eq})
in terms of the electromagnetic field acting on the particle.
Next, we substitute the acceleration from equation (\ref{RHD2018Cl Newton eq}) in the force field $F^{a}$ (\ref{RHD2018Cl force field via acceleration})
and find the expression containing new hydrodynamic functions
$$F^{a}=\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} \frac{e_{i}}{m_{i}\gamma_{i}}E^{a}(\textbf{r}+\mbox{\boldmath $\xi$},t) \delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t))$$
$$+\frac{1}{c}\varepsilon^{abc}\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} \frac{e_{i}}{m_{i}\gamma_{i}}v_{i}^{b}B^{c}(\textbf{r}+\mbox{\boldmath $\xi$},t) \delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t))$$
\begin{equation}\label{RHD2018Cl force field via acceleration}
-\frac{1}{c^{2}}\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} \frac{e_{i}}{m_{i}\gamma_{i}}v_{i}^{a}v_{i}^{b}E^{b}(\textbf{r}+\mbox{\boldmath $\xi$},t) \delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t)).
\end{equation}
In equation (\ref{RHD2018Cl force field via acceleration})
we apply the replacement of coordinate of particles in the argument of the electromagnetic field $\textbf{r}_{i}(t)$
on $\textbf{r}+\mbox{\boldmath $\xi$}$ using the $\delta$-function $\delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t))$.
We consider the monopole approximation of the electric and magnetic fields
$E^{a}(\textbf{r}+\mbox{\boldmath $\xi$},t)\approx E^{a}(\textbf{r},t)$,
and
$B^{a}(\textbf{r}+\mbox{\boldmath $\xi$},t)\approx B^{a}(\textbf{r},t)$.
It corresponds to the mean-field (self-consistent field) approximation traditionally applied at the study of the plasmas.
This approximation is possible if the electric and magnetic fields have small change on the scale of $\Delta$-vicinity.
For instance, if the particle is under the action of the electromagnetic wave
its wavelength $\lambda$ should be large in compare with the radius of the vicinity:
$\lambda\gg\sqrt[3]{\Delta}$.
The mean-field approximation simplify the force field to the following form
$$F^{a}=E^{a}(\textbf{r},t)\cdot\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} \frac{e_{i}}{m_{i}\gamma_{i}} \delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t))$$
$$+\frac{1}{c}\varepsilon^{abc}B^{c}(\textbf{r},t)\cdot\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} \frac{e_{i}}{m_{i}\gamma_{i}}v_{i}^{b} \delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t))$$
\begin{equation}\label{RHD2018Cl force field expanded}
-\frac{1}{c^{2}}E^{b}(\textbf{r},t)\cdot\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} \frac{e_{i}}{m_{i}\gamma_{i}}v_{i}^{a}v_{i}^{b} \delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t)).
\end{equation}
This presigere gives the mean-field (the self-consistent) approximation.
Evolution of the particle current leads to three new hydrodynamic variables
\begin{equation}\label{RHD2018Cl Gamma function definition} \Gamma=\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} \frac{1}{\gamma_{i}} \delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t)),\end{equation}
\begin{equation}\label{RHD2018Cl Theta function definition} \Theta^{a}=\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} \frac{1}{\gamma_{i}}v_{i}^{a} \delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t)),\end{equation}
and
\begin{equation}\label{RHD2018Cl Xi function definition}
\Xi^{ab}=\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} \frac{1}{\gamma_{i}}v_{i}^{a}v_{i}^{b} \delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t)).
\end{equation}
Overall, we find the following equation
\begin{equation}\label{RHD2018Cl current derivative Euler equation explicit}
\partial_{t}j^{a}+\partial_{b}\Pi^{ab}
=\frac{e}{m}\biggl(\Gamma E^{a} +\frac{1}{c}\varepsilon^{abc}\Theta^{a}B^{c} -\frac{1}{c^{2}}\Xi^{ab} E^{b}\biggr),
\end{equation}
which is the particle current evolution equation.
It is one of the generalizations of the nonrelativistic Euler equation for the velocity field on the relativistic regime.
A brief presentation of this model is given in short report \cite{Andreev 2021 05}
Some steps towards present model are made in Ref. \cite{Andreev arXiv 12 relativistic}.
\subsection{New relativistic hydrodynamic variables}
Relativistic equation for the particle current evolution
gives three new hydrodynamic functions $\Gamma$, $\Theta^{a}$, and $\Xi^{ab}$.
These hydrodynamic functions do not exist in the nonrelativistic limit.
The particle current $j^{a}$ coincides with the momentum density in the nonrelativistic limit.
Moreover, the hydrodynamic Gamma function $\Gamma$ tends to the concentration $n$ in the nonrelativistic limit.
The hydrodynamic Theta function $\Theta^{a}$ has same nonrelativistic limit as the current of particles $j^{a}$ and the momentum density.
The hydrodynamic Xi function has same form in the nonrelativistic regime as the flux of the particle current $\Pi^{ab}$ (\ref{RHD2018Cl def Pi})
and the momentum flux (containing the pressure).
We have two different functions in the relativistic hydrodynamics:
the current of particles and the momentum density.
Hence, we need to choose
which of them should be included in the model.
The answer on this question follows from the Maxwell equations,
where the electromagnetic field is caused by the concentration and the current of particles.
Therefore, we consider the momentum density as a function
which is unnecessary for application in our model.
Hydrodynamic Gamma function $\Gamma$ is
the "average" reverse relativistic gamma factor $\Gamma=\langle\frac{1}{\gamma_{i}}\rangle$
(see equation (\ref{RHD2018Cl Gamma function definition})).
Hydrodynamic Theta vector function $\Theta^{a}$ is the current of the reverse gamma factor.
So, vector field $\Theta^{a}$ is the current of the scalar function $\Gamma$:
$\Theta^{a}=\langle v_{i}^{a}/\gamma_{i}\rangle$ (see equation (\ref{RHD2018Cl Theta function definition})).
Next, the tensor field $\Xi^{ab}$ is the flux or current of the vector field $\Theta^{a}$:
$\Xi^{ab}=\langle v_{i}^{a}v_{i}^{b}/\gamma_{i}\rangle$
(see equation \ref{RHD2018Cl Xi function definition}).
\subsection{Equations evolution for the new relativistic hydrodynamic variables}
For the further development of the relativistic hydrodynamic
equations consider the evolution of the hydrodynamic Gamma and Theta functions.
It is well-known that
the application of nonrelativistic hydrodynamic composed of the continuity and Euler equations
gives incorrect coefficient for the thermal contributions in the spectrum
in compare with the kinetic results \cite{Tokatly PRB 99}, \cite{Tokatly PRB 00}.
However, the extension of the hydrodynamic model including the pressure evolution equation allows to improve the results
\cite{Andreev JPP 21}, \cite{Tokatly PRB 99}, \cite{Tokatly PRB 00}, \cite{Miller PoP 16}.
Here, we try to create a relativistic minimal coupling hydrodynamic model,
which is based on the evolution of the scalar and vector functions
$n$, $j^{a}$, $\Gamma$, $\Theta^{a}$,
where the tensor fields of the higher tensor rank
(like $\Pi^{ab}$ and $\Xi^{ab}$)
are expressed using equations of state.
\subsubsection{Equation of the Gamma function evolution}
We consider the temporal evolution of the Gamma function according to the described method
(by the calculation of the time derivative of the definition of required function)
\begin{equation}\label{RHD2018Cl equation of Gamma evolution 1}
\partial_{t}\Gamma+\partial_{a}\Theta^{a}= -\frac{1}{c^{2}}\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} \frac{1}{\gamma_{i}}v_{i}^{a}\dot{v}_{i}^{a} \delta_{i},
\end{equation}
where
the second term in equation comes from differentiating of the delta function,
the last term appears from differentiating of the reverse gamma factor,
$\delta_{i}\equiv\delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t))$
is the short representation of the delta function.
Moreover, let us mention a technical detail
that the gamma factor is in the denominator in the last term of equation (\ref{RHD2018Cl equation of Gamma evolution 1})
We consider the right-hand side of equation (\ref{RHD2018Cl equation of Gamma evolution 1}) using the relativistic Newton equation (\ref{RHD2018Cl Newton eq})
\begin{equation}\label{RHD2018Cl} \biggl\langle\frac{1}{\gamma_{i}}v_{i}^{a}\dot{v}_{i}^{a}\biggr\rangle=\frac{e}{m}\langle v_{i}^{a}E_{i}^{a}\rangle-\frac{1}{c^{2}}\frac{e}{m}\langle v_{i}^{a}\textbf{v}_{i}^{2}E_{i}^{a}\rangle, \end{equation}
where we use that
$\gamma_{i}v_{i}^{a}\dot{v}_{i}^{a}
=\frac{e}{m}v_{i}^{a}(E^{a}-\frac{1}{c^{2}}v_{i}^{a}(\textbf{v}_{i}\cdot\textbf{E}))$
$=\frac{e}{m}(\textbf{v}_{i}\cdot\textbf{E}))(1-\frac{1}{c^{2}}\textbf{v}_{i}^{2})$.
It can be rewritten via the gamma factor
$\frac{e}{m}(\textbf{v}_{i}\cdot\textbf{E}))\frac{1}{\gamma_{i}^{2}}$,
but it gives no further usage.
Ones again, we consider the monopole approximation of the electric field
$E^{a}(\textbf{r}+\mbox{\boldmath $\xi$},t)\approx E^{a}(\textbf{r},t)$,
which basically gives the mean-field approximation,
and find
\begin{equation}\label{RHD2018Cl} \langle v_{i}^{a}E_{i}^{a}\rangle-\frac{1}{c^{2}}\langle v_{i}^{a}\textbf{v}_{i}^{2}E_{i}^{a}\rangle
=E^{a}(\textbf{r},t)\biggl(j^{a}-\frac{1}{c^{2}}Q^{a}\biggr), \end{equation}
where
\begin{equation}\label{RHD2018Cl Q a def} Q^{a}=\langle v_{i}^{a}\textbf{v}_{i}^{2}\rangle\end{equation}
is the flux vector of the velocity square which coincides (proportional) with the kinetic energy current in the nonrelativistic regime.
Finally, we present the evolution equation for the Gamma function
\begin{equation}\label{RHD2018Cl equation of Gamma evolution 2 monopole}
\partial_{t}\Gamma+\partial_{a}\Theta^{a}= -\frac{1}{c^{2}}\textbf{E}\biggl(\textbf{j}-\frac{1}{c^{2}}\textbf{Q}\biggr),
\end{equation}
where
we see that the Theta function is the current of the Gamma function
(as it is mentioned above judging on the structure of the definitions).
Hence, the set of equation partially closes itself.
So, evolution of Gamma function leads to functions $j^{a}$, $\Theta^{a}$,
which are introduced above, and single new function $Q^{a}$.
\subsubsection{Equation for the Theta function evolution}
Next equation which appears in the developing model is the evolution equation for the hydrodynamic Theta function.
Differentiating function (\ref{RHD2018Cl Theta function definition}) with respect to time we find the required equation.
Derivative of function (\ref{RHD2018Cl Theta function definition}) leads to three terms:
the derivative of the velocity,
the derivative of the delta function,
and the derivative of the reverse gamma factor.
The derivative of the delta function leads to the flux of the hydrodynamic Theta function
which is presented by function $\Xi^{ab}$ (\ref{RHD2018Cl Xi function definition}).
The derivative of the reverse gamma factor is considered at the derivation of the hydrodynamic Gamma function evolution:
$\partial_{t}(\gamma_{i}^{-1})$$=-\frac{e}{mc^{2}}\gamma_{i}^{-2}(\textbf{v}_{i}\cdot\textbf{E}_{i})$.
The acceleration provides the following term
$\gamma_{i}^{-1}\dot{v}_{i}^{a}=$
$(e/m)\gamma_{i}^{-2}[E_{i}^{a}+\varepsilon^{abc}v_{i}^{b}B_{i}^{c}/c-v_{i}^{a}v_{i}^{b}E_{i}^{b}/c^{2}]$.
The coefficient $\gamma_{i}^{-2}$ splits on two terms.
Hence, the term caused by the acceleration $\gamma_{i}^{-1}\dot{v}_{i}^{a}$ splits on six terms.
We present the evolution equation for the hydrodynamic Theta function
in the monopole (mean-field) approximation of the electromagnetic field
$$\partial_{t}\Theta^{a} +\partial_{b}\Xi^{ab}=\frac{e}{m}E^{a}\biggl[n-\frac{\Pi^{bb}}{c^{2}}\biggr]$$
\begin{equation}\label{RHD2018Cl equation of Theta evolution}
+\frac{e}{mc}\varepsilon^{abc}\biggl[j^{b}-\frac{Q^{b}}{c^{2}}\biggr]B^{c}-\frac{2e}{mc^{2}}E^{b}\biggl[\Pi^{ab}-\frac{L^{abcc}}{c^{2}}\biggr],\end{equation}
where
\begin{equation}\label{RHD2018Cl L fourth order definition} L^{abcd}=\langle v_{i}^{a}v_{i}^{b}v_{i}^{c}v_{i}^{d}\rangle. \end{equation}
It is essential to point out that
the evolution of $\Theta^{a}$ is mostly expressed via the concentration $n$, the current $j^{a}$, the flux of current $\Pi^{ab}$.
It means that the set equations is almost closing itself one again.
One additional function appears at this step $L^{abcd}$ (\ref{RHD2018Cl L fourth order definition})
along with function $Q^{a}$ (\ref{RHD2018Cl Q a def}) obtained at the derivation of equation
(\ref{RHD2018Cl equation of Gamma evolution 2 monopole}).
Therefore, we can try to stop at this step and truncate the set of equations.
\subsubsection{Xi function evolution}
The particle current evolution equation (\ref{RHD2018Cl current derivative Euler equation explicit})
suggests that we need evolution equation for the Xi function $\Xi^{ab}$.
However, it is necessary to create a limited set of equations.
Hence, at this stage of the model development find an equation of state for the Xi function $\Xi^{ab}$
(\ref{RHD2018Cl Xi function definition}).
\subsection{Equations of electromagnetic field}
The mean-field electromagnetic field $\textbf{E}$ and $\textbf{B}$
appearing in equations (\ref{RHD2018Cl current derivative Euler equation explicit}),
(\ref{RHD2018Cl equation of Gamma evolution 2 monopole}),
and (\ref{RHD2018Cl equation of Theta evolution})
satisfies the Maxwell equations
\begin{equation}\label{RHD2018Cl div B} \nabla \cdot\textbf{B}=0,\end{equation}
\begin{equation}\label{RHD2018Cl rot E}
\nabla\times \textbf{E}=-\frac{1}{c}\partial_{t}\textbf{B}, \end{equation}
\begin{equation}\label{RHD2018Cl div E} \nabla \cdot\textbf{E}=4\pi\sum_{s=e,i}e_{s}n_{s},\end{equation}
and
\begin{equation}\label{RHD2018Cl rot B with time}
\nabla\times \textbf{B}=\frac{1}{c}\partial_{t}\textbf{E}+\frac{4\pi}{c}\sum_{s=e,i}e_{s}\textbf{j}_{s}.\end{equation}
Full set of hydrodynamic equations is developed for each species,
while the superposition of all particles appears as the source of the electromagnetic field.
Hence, the sum of species "s" is presented in equations (\ref{RHD2018Cl div E}) and (\ref{RHD2018Cl rot B with time}).
as an example we show the summation on two species:
the electrons and ions.
Quantum analog of this concept is suggested in Ref. \cite{Maksimov QHM 99}.
The many-particle quantum hydrodynamics is developed for the number of physical systems
\cite{Andreev JPP 21}, \cite{Andreev Ch 21}, \cite{Andreev PoF 21}.
\section{Velocity field in equations for the relativistic hydrodynamics}
Introduce the velocity field in the traditional way as the ratio between the particle current and the concentration $\textbf{v}=\textbf{j}/n$.
Next, we need to recognize the contribution of the velocity field in other hydrodynamic functions.
The velocity field is the local average velocity or in other words it is the average velocity of all particle in the delta vicinity $\Delta$ of point $\textbf{r}$.
Therefore, we can split the velocity of each particle on the average velocity (the velocity field) $\textbf{v}$
and the deviation from the velocity field $\textbf{u}_{i}$ caused by the difference of velocities of particles related to the thermal effects: $\textbf{v}_{i}=\textbf{v}+\textbf{u}_{i}$.
Hence, function $\textbf{u}_{i}$ can be interpreted as the local thermal velocity of particles.
The definitions of the current (\ref{RHD2018Cl particle current definition}) and the velocity field show that the average of the thermal velocity is equal to zero $\langle\textbf{u}_{i}\rangle=0$.
We substitute the decomposition of the velocities in the definition of the hydrodynamic functions.
We start with the current flux
\begin{equation}\label{RHD2018Cl} \Pi^{ab}=\langle v_{i}^{a}v_{i}^{b}\rangle=nv^{a}v^{b}+p^{ab},\end{equation}
which gives the flux of the current of particles on the thermal velocities $p^{ab}=\langle u_{i}^{a}u_{i}^{b}\rangle$
which is an analog of pressure,
but it is a different function.
The terms linear on the thermal velocity go to zero.
Next, we consider the hydrodynamic Theta function
\begin{equation}\label{RHD2018Cl}
\Theta^{a}=\biggl\langle \frac{v_{i}^{a}}{\gamma_{i}}\biggr\rangle=\Gamma v^{a}+t^{a},\end{equation}
where
$t^{a}=\langle\frac{u_{i}^{a}}{\gamma_{i}}\rangle$,
with
\begin{equation}\label{RHD2018Cl}
\gamma_{i}=\frac{1}{\sqrt{1-\frac{[\textbf{v}^{2}+2\textbf{v}\cdot\textbf{u}_{i}+\textbf{u}_{i}^{2}]}{c^{2}}}}.\end{equation}
Function $t^{a}$ can be called the thermal part of the hydrodynamic Theta function.
However, function $t^{a}$ also contains the velocity field in nonadditive form.
Same is true for the hydrodynamic Gamma function $\Gamma$
since "averaged" reverse gamma factor contains both the velocity field and the thermal velocity.
we also describe the structure of the hydrodynamic Xi function $\Xi^{ab}$ in terms of the velocity field:
\begin{equation}\label{RHD2018Cl}
\Xi^{ab}=\biggl\langle \frac{v_{i}^{a}v_{i}^{b}}{\gamma_{i}}\biggr\rangle
=\Gamma v^{a}v^{b}+ v^{a}t^{b} +t^{a}v^{b} +t^{ab},\end{equation}
where
$t^{ab}=\langle\frac{u_{i}^{a}u_{i}^{b}}{\gamma_{i}}\rangle$.
We ready to represent continuity equation and the particle current evolution equation,
but the hydrodynamic Gamma function evolution equation contains vector $Q^{a}$
and the hydrodynamic Theta function evolution equation includes vector $Q^{a}$ and the partial trace of tensor $L^{abcd}$.
here, we study the structure of presented functions:
$$Q^{a}=\langle v_{i}^{a}v_{i}^{b}v_{i}^{b}\rangle$$
\begin{equation}\label{RHD2018Cl}=nv^{a}v^{b}v^{b}+v^{a}p^{bb}+2v^{b}p^{ab}+q^{a},\end{equation}
where
\begin{equation}\label{RHD2018Cl}q^{a}=\langle u_{i}^{a}u_{i}^{b}u_{i}^{b}\rangle,\end{equation}
and
$$L^{abcd}=\langle v_{i}^{a}v_{i}^{b}v_{i}^{c}v_{i}^{d}\rangle
=nv^{a}v^{b}v^{c}v^{d}$$
\begin{equation}\label{RHD2018Cl}+[v,v,p]^{a,b,cd} +[v,q]^{a,bcd} +M^{abcd}.\end{equation}
We have rather huge expression for tensor $L^{abcd}$ in terms of the velocity field.
Therefore, we introduce the following notations including permutations of similar terms.
\begin{equation}\label{RHD2018Cl}[v,q]^{a,bcd}\equiv v^{a}q^{bcd}+v^{b}q^{acd}+v^{c}q^{abd}+v^{d}q^{abc},\end{equation}
and
$$[v,v,p]^{a,b,cd}\equiv v^{a}v^{b}p^{cd}+v^{a}v^{c}p^{bd}$$
\begin{equation}\label{RHD2018Cl}+v^{a}v^{d}p^{bc} +v^{b}v^{c}p^{ad}+v^{b}v^{d}p^{ac}
+v^{c}v^{d}p^{ab}.\end{equation}
Above, we also introduce the third and fourth rank tensors composed of the thermal velocities:
\begin{equation}\label{RHD2018Cl}q^{abc}=\langle u_{i}^{a}u_{i}^{b}u_{i}^{c}\rangle,\end{equation}
and
\begin{equation}\label{RHD2018Cl}M^{abcd}=\langle u_{i}^{a}u_{i}^{b}u_{i}^{c}u_{i}^{d}\rangle.\end{equation}
Equation for the hydrodynamic Theta function evolution contains the partial trace of tensor $L^{abcc}$,
which has the following form:
$$L^{abcc}=nv^{a}v^{b}\textbf{v}^{2} +v^{a}v^{b}p^{cc} +2v^{a}v^{c}p^{bc}+2v^{b}v^{c}p^{ac}$$
\begin{equation}\label{RHD2018Cl}+v^{2}p^{ab} +v^{a}q^{b}+v^{b}q^{a}+2v^{c}q^{abc} +M^{abcc}.\end{equation}
\subsection{Intermediate form of the suggested hydrodynamic model}
The representation of the hydrodynamic functions via the velocity is given.
Therefore, we can represent the hydrodynamic equations (\ref{RHD2018Cl continuity equation}), (\ref{RHD2018Cl current derivative}),
(\ref{RHD2018Cl equation of Gamma evolution 2 monopole}), and (\ref{RHD2018Cl equation of Theta evolution})
via the velocity field.
The continuity equation has the traditional form:
\begin{equation}\label{RHD2018Cl continuity equation via velocity}
\partial_{t}n+\nabla\cdot (n\textbf{v})=0. \end{equation}
Equation of the velocity field evolution:
$$n(\partial_{t}+(\textbf{v}\cdot\nabla))v^{a}+\partial_{b}p^{ab}
=\frac{e}{m}\Gamma E^{a}+\frac{e}{mc}\Gamma\varepsilon^{abc}v^{b}B^{\gamma}$$
\begin{equation}\label{RHD2018Cl velocity field evolution equation}
+\frac{e}{mc}\varepsilon^{abc}t^{b}B^{\gamma}-\frac{e}{mc^{2}}(\Gamma v^{a}v^{b}+v^{a}t^{b}+t^{a}v^{b}+t^{ab})E^{b}
\end{equation}
has familiar left-hand side
while the right-hand side (at leats first three terms) has an intuitively understandable structure,
but new (untraditional) functions define the right-hand side.
Tensor $p^{ab}$ looks like the pressure.
However, the traditional pressure is the momentum transmitted through the element of surface during the unit of time.
While $p^{ab}$ is the current of particles transmitted through the element of surface during the unit of time.
Next, we find equation for the hydrodynamic Gamma function
$$\partial_{t}\Gamma +\partial_{b}(v^{b}\Gamma)+\partial_{b}t^{b} =-\frac{e}{mc^{2}}\biggl[nv^{b}E_{b}$$
\begin{equation}\label{RHD2018Cl evolution of Gamma via velocity}
-\frac{1}{c^{2}}E^{b}(nv_{b}\textbf{v}^{2} +v_{b}p^{cc}+2p^{bc}v_{c}+q_{b})\biggr]. \end{equation}
Finally, we present the hydrodynamic Theta function evolution equation
$$\partial_{t}t^{a} +\partial_{b}(t^{a}v^{b})+t^{b}\partial_{b}v^{a}+\partial_{b}t^{ab}$$
$$-\frac{e}{mc^{2}}v^{a}\biggl[nv^{b}E_{b}
-\frac{1}{c^{2}}E^{b}(nv_{b}\textbf{v}^{2} +v_{b}p^{cc}+2p^{bc}v_{c}+q_{b})\biggr]$$
$$-\frac{\Gamma}{n}\partial_{b}p^{ab}
+\frac{e}{m}\frac{\Gamma}{n}\biggl[\Gamma E^{a} +\frac{1}{c}\Gamma\varepsilon^{abc}v_{b}B_{c}-\frac{1}{c^{2}}\Gamma v^{a}v^{b}E_{b}$$
$$+\frac{1}{c}\varepsilon^{abc}t_{b}B_{c} -\frac{1}{c^{2}}(v^{a}t^{b}+t^{a}v^{b}+t^{ab})E_{b}\biggr]
=\frac{e}{m}nE^{a}$$
$$+\frac{e}{mc}n\varepsilon^{abc}v_{b}B_{c} -\frac{2e}{mc^{2}}(nv^{a}v^{b}+p^{ab})E_{b}
-\frac{e}{mc^{2}}E^{a}(n\textbf{v}^{2}+p^{bb})$$
$$-\frac{e}{mc^{2}}\frac{1}{c}\varepsilon^{abc}(nv_{b}\textbf{v}^{2}+v_{b}p^{dd}+2v^{d}p_{bd}+q_{b})B_{c}$$
$$+\frac{2e}{mc^{2}}\frac{1}{c^{2}}[nv^{a}v^{b}\textbf{v}^{2}+v^{a}v^{b}p^{cc}+p^{ab}\textbf{v}^{2}+2p^{ac}v^{b}v_{c}+2v^{a}p^{bc}v_{c}$$
\begin{equation}\label{RHD2018Cl evolution of Theta via velocity}
+v^{a}q^{b}+v^{b}q^{a}+2v^{c}q^{abc}+M^{abcc}]E_{b}. \end{equation}
Equations (\ref{RHD2018Cl continuity equation via velocity})-(\ref{RHD2018Cl evolution of Gamma via velocity})
are the basic set of relativistic hydrodynamic equations of the ideal plasmas suggested in this paper.
However, equations (\ref{RHD2018Cl continuity equation via velocity})-(\ref{RHD2018Cl evolution of Gamma via velocity})
should be truncated to obtain a closed set of equations applicable for the study of plasmas.
This paper is focused on the method of derivation of hydrodynamic equations for the relativistic plasmas.
However, these equations have been presented in Refs.
\cite{Andreev 2021 05}, \cite{Andreev 2021 06}, \cite{Andreev 2021 07}, \cite{Andreev 2021 08},
including waves in the magnetized plasmas \cite{Andreev 2021 06}, \cite{Andreev 2021 07}, \cite{Andreev 2021 08}.
\section{Truncation}
The truncation requires equations of state for several functions including
$\Pi^{ab}$ (reducing to $p^{ab}$), $\Xi^{ab}$ (reducing to $t^{ab}$),
$Q^{a}$ (reducing to $q^{a}$), $L^{abcd}$ (reducing to $M^{abcd}$).
They should be presented via functions $n$, $v^{a}$, $\Gamma$, $t^{a}$.
Explicit microscopic definitions like (\ref{RHD2018Cl concentration definition}), (\ref{RHD2018Cl Gamma function definition}),
(\ref{RHD2018Cl Theta function definition}), (\ref{RHD2018Cl Xi function definition}) are not useful for calculation of equilibrium values of considering functions or derivations of equations of state.
These definitions are based on the exact microscopic motion of particles
which is unknown.
However, the approximate application of the equations for the higher rank tensors can give this information.
However, we have not developed extended model,
it is planed as the future work.
Here we go another way and neglect the physical picture demonstrated during derivation.
So, some formal technic is applied to get the required equations of state.
This situation is usual for the hydrodynamic models of classical fluids and plasmas.
Required values can be obtained via the equilibrium distribution functions $f_{0}(p)$,
where $p$ is the module of momentum.
For instance the Gamma function is the average of the reverse of the relativistic factor $\gamma_{i}$.
Hence, the equilibrium Gamma unction can be calculated as
$\Gamma_{0}=\int\gamma^{-1}(p)f_{0}(p)d^{3}p$,
where
$\gamma(p)=\sqrt{p^{2}+m^{2}c^{2}}/mc$.
Here we make substitution of concepts.
We use the distribution function instead of the arithmetic average on the $\Delta$-vicinity.
\subsection{Equations of state: rest frame}
Calculate required expressions using relativistic equilibrium distribution function presented in the rest frame
\begin{equation}\label{RHD2018Cl eq rest frame distribution}
f_{0r}(p)=\textrm{Z} e^{-\epsilon/T},
\end{equation}
where
the subindex $0$ refers to the equilibrium state,
the subindex $r$ refers to the rest frame,
\begin{equation}\label{RHD2018Cl}
\textrm{Z}=\frac{n}{4\pi m^{2}cTK_{2}(\frac{mc^{2}}{T})},
\end{equation}
with $T$ is the equilibrium temperature in the energy units,
$p$ is the momentum,
$K_{2}(b)$ is the second order Macdonald function,
and $\epsilon=\sqrt{m^{2}c^{4}+p^{2}c^{2}}$.
The rest frame is the inertial frame,
where the fluid is macroscopically motionless or it has minimal energy.
The energy momentum tensor presented in the Introductions is constructed in the first regime for the macroscopically motionless fluid.
Same assumption is made for the distribution function (\ref{RHD2018Cl eq rest frame distribution}).
The useful application of the rest frame includes the assumption
that
the deviations from the macroscopically motionless state are relatively small.
So, we can consider the medium as the average background,
where all structures and flows can be viewed as processes happening on this background.
Hence, we introduce some equilibrium-like background for the large amplitude nonlinear processes.
The rest frame can be easily associated with this background.
If we have infinite number of the large amplitude structures overlapping each over we have no distinguishable background.
In this case the rest frame can be found as the inertial frame,
where system have minimal energy.
However, this formal step does not give simple physical picture to include in calculations of the distribution function or the energy-momentum tensor.
The Macdonald functions $K_{\mu}(b)$ have the following definition (the integral form):
\begin{equation}\label{RHD2018Cl}
K_{\mu}(b)=\frac{\sqrt{\pi}b^{\mu}}{\Gamma(\mu+\frac{1}{2})}\int_{1}^{+\infty}(t^{2}-1)^{\mu-1/2}e^{-bt}dt,
\end{equation}
where $Re\mu>-1/2$ and $Re b>0$.
For instance, the Macdonald functions $K_{\mu}(b)$ allow to calculate analytically the equilibrium value of the average reverse $\gamma$-factor:
$\Gamma_{0}=m^{2}cT\textrm{Z}K_{1}(b)$,
where $b=mc^{2}/T$.
Presented analysis provides the following equations of state
$p^{ab}=U_{p}^{2}n\delta^{ab}$,
$t^{ab}=U_{t}^{2}n\delta^{ab}$,
$q^{a}=0$,
$M^{abcd}=(U_{M}^{4}/3) nI_{0}^{abcd}$,
where $I_{0}^{abcd}=\delta^{ab}\delta^{cd}+\delta^{ac}\delta^{bd}+\delta^{ad}\delta^{bc}$,
and $U_{p}$, $U_{t}$, $U_{M}$ are constants.
After the application of the equations of state for the high-rank tensors introduced in the model
we obtain the following set of truncated hydrodynamic equations.
The continuity equation:
\begin{equation}\label{RHD2018Cl continuity equation via velocity Truncated}
\partial_{t}n+\nabla\cdot (n\textbf{v})=0. \end{equation}
Equation of the velocity field evolution:
$$n(\partial_{t}+(\textbf{v}\cdot\nabla))\textbf{v}+U_{p}^{2}\nabla n
=\frac{e}{m}\Gamma \textbf{E}$$
$$+\frac{e}{mc}\Gamma [\textbf{v}\times \textbf{B}]
+\frac{e}{mc} [\textbf{t}\times \textbf{B}]
-\frac{e}{mc^{2}}\biggl(
\Gamma \textbf{v} (\textbf{v}\cdot\textbf{E})$$
\begin{equation}\label{RHD2018Cl velocity field evolution equation Truncated}
+\textbf{v}(\textbf{t}\cdot\textbf{E})
+\textbf{t}(\textbf{v}\cdot\textbf{E})
+U_{t}^{2} n \textbf{E}\biggr).
\end{equation}
The Gamma function evolution equation
$$\partial_{t}\Gamma +\nabla\cdot(\textbf{v}\Gamma)+\nabla\cdot \textbf{t} =-\frac{e}{mc^{2}}\biggl[n\textbf{v}\cdot\textbf{E}$$
\begin{equation}\label{RHD2018Cl evolution of Gamma via velocity Truncated}
-\frac{1}{c^{2}}(n(\textbf{v}\cdot\textbf{E}) \textbf{v}^{2} +5U_{p}^{2}n \textbf{v}^{2})\biggr]. \end{equation}
The Theta function evolution equation
$$\partial_{t}\textbf{t} +(\textbf{v}\cdot\nabla)\textbf{t}+\textbf{t}(\nabla\cdot\textbf{v})+(\textbf{t}\cdot\nabla)\textbf{v}+U_{t}^{2}\nabla n$$
$$-\frac{e}{mc^{2}}v^{a}\biggl[n\textbf{v}\cdot \textbf{E}
-\frac{1}{c^{2}}\biggl(n(\textbf{v}\cdot\textbf{E}) \textbf{v}^{2} +(\textbf{v}\cdot\textbf{E}) 5U_{p}^{2}n\biggr)\biggr]$$
$$-\frac{\Gamma}{n}U_{p}^{2}\nabla n
+\frac{e}{m}\frac{\Gamma}{n}\biggl[\Gamma \textbf{E} +\frac{1}{c}\Gamma [\textbf{v}\times\textbf{B}]
-\frac{1}{c^{2}}\Gamma \textbf{v}(\textbf{v}\cdot\textbf{E})$$
$$+\frac{1}{c} [\textbf{t}\times \textbf{B}]
-\frac{1}{c^{2}}(\textbf{v}(\textbf{t}\times\textbf{E})+\textbf{t}(\textbf{v}\times\textbf{E})+U_{t}^{2}n\textbf{E})\biggr]
$$
$$=\frac{e}{m}n\textbf{E}
+\frac{e}{mc}n[\textbf{v}\times \textbf{B}]
-\frac{2e}{mc^{2}}(n\textbf{v}(\textbf{v}\times\textbf{E})+U_{p}^{2}n\textbf{E})$$
$$-\frac{e}{mc^{2}}\textbf{E}(n\textbf{v}^{2}+3U_{p}^{2}n)
-\frac{e}{mc^{2}}\frac{1}{c}(n\textbf{v}^{2}+5U_{p}^{2}n)[\textbf{v}\times\textbf{B}]$$
$$+\frac{2e}{mc^{2}}\frac{1}{c^{2}}
[n\textbf{v}(\textbf{v}\cdot\textbf{E}) \textbf{v}^{2}
+5U_{p}^{2}n\textbf{v}(\textbf{v}\cdot\textbf{E})
$$
\begin{equation}\label{RHD2018Cl evolution of Theta via velocity Truncated}
+U_{p}^{2}n\textbf{E} \textbf{v}^{2}
+2\textbf{v} U_{p}^{2}n(\textbf{v}\cdot \textbf{E})
+\frac{5}{3}U_{M}^{4}\textbf{E}]. \end{equation}
Equations (\ref{RHD2018Cl continuity equation via velocity Truncated})
-(\ref{RHD2018Cl evolution of Theta via velocity Truncated})
are coupled to the Maxwell equations (\ref{RHD2018Cl div B})-(\ref{RHD2018Cl rot B with time}).
Three different characteristic velocities $U_{p}$, $U_{t}$, and $U_{M}$ are calculated in the following way.
The application of the isotropic distribution function leads to diagonal form of tensors $p^{ab}$ and $t^{ab}$:
$p^{ab}=p\delta^{ab}$ and $t^{ab}=\tilde{t}\delta^{ab}$.
The "diagonal" form is obtained for tensor $M^{abcd}$ as well:
$M^{abcd}=M_{0}(\delta^{ab}\delta^{cd}+\delta^{ac}\delta^{bd}+\delta^{ad}\delta^{bc})/3$.
The equilibrium expressions for functions
$p$, $\tilde{t}$, $\textbf{q}$, $M$ are used as the equations of state for the nonequilibrium functions.
Approximate calculation of functions $p^{ab}$, $t^{ab}$, $\textbf{q}$, and $M^{abcd}$ gives to the following representations
$p^{ab}=c^{2}\delta^{ab}\tilde{Z}f_{1}(\beta)/3$,
$t^{ab}=c^{2}\delta^{ab}\tilde{Z}f_{2}(\beta)/3$
$M^{abcd}=c^{4}(\delta^{ab}\delta^{cd}+\delta^{ac}\delta^{bd}+\delta^{ad}\delta^{bc})\tilde{Z}f_{3}(\beta)/15$,
and
$\textbf{q}=0$,
where
$\beta=mc^{2}/T$,
$\tilde{Z}=4\pi Z (mc)^{3}=n\beta K_{2}^{-1}(\beta)$,
\begin{equation}\label{RHD2021ClLM f 1} f_{1}(\beta)=\int_{1}^{+\infty}\frac{d x}{x}(x^{2}-1)^{3/2}e^{-\beta x}, \end{equation}
\begin{equation}\label{RHD2021ClLM f 2} f_{2}(\beta)=\int_{1}^{+\infty}\frac{d x}{x^{2}}(x^{2}-1)^{3/2}e^{-\beta x}, \end{equation}
and
\begin{equation}\label{RHD2021ClLM f 3} f_{3}(\beta)=\int_{1}^{+\infty}\frac{d x}{x^{3}}(x^{2}-1)^{5/2}e^{-\beta x}. \end{equation}
Functions $f_{1}(\beta)$, $f_{2}(\beta)$ and $f_{3}(\beta)$ are calculated numerically below for the chosen values of temperatures.
For each function describing the thermal we introduce corresponding velocity
$\delta p=U_{p}^{2} \delta n$,
$\delta \tilde{t}=U_{t}^{2} \delta n$,
$\delta M=U_{M}^{4} \delta n$.
\subsection{Equations of state: arbitrary inertial frame}
For the calculation of equations of state
we have used the relativistic Maxwellian distribution function in the rest frame.
This approach narrows down the variety of physical scenario,
where the suggested model can be applied.
Let us describe some backgrounds for the further generalizations.
First, we can use the relativistic Maxwellian distribution in the arbitrary frame:
\begin{equation}\label{RHD2018Cl distribution p A}
f_{0}(p)=\hat{Z} e^{U_{\alpha}P^{\alpha}/T},
\end{equation}
where
$P^{\alpha}=p^{\alpha}+eA^{\alpha}$ is the canonical momentum,
$p^{\alpha}=\{\varepsilon/c,\textbf{p}\}$ is the four-momentum,
$A^{\alpha}=\{\phi,\textbf{A}\}$ is the four-potential of the electromagnetic field,
$U_{\alpha}=\{-c,\textbf{v}\}$
(while $U^{\alpha}=\{c,\textbf{v}\}$)
is the hydrodynamic four-velocity field giving the four-current $j^{\alpha}=nU^{\alpha}$,
$\textbf{v}$ is the hydrodynamic velocity field,
and
\begin{equation}\label{RHD2018Cl Z hat}
\hat{Z}=\frac{n}{4\pi m^{2}cTK_{2}(\frac{mc^{2}}{T})} e^{e\phi/T}.
\end{equation}
with $A^{0}=\Phi$
The hydrodynamic four-velocity field entering the distribution function along with the concentration presented in $\hat{Z}$.
More general distribution functions are used in literature for the derivation of hydrodynamics from the kinetic model.
For instance, generalized prefactor in front of exponential function is discussed in Ref. \cite{Hazeltine APJ 2002},
where the prefactor is constructed of several hydrodynamic functions in the covariant form.
This function is aimed to cover the gyrokinetic effects in the relativistic plasmas.
\section{On the covariant structure of hydrodynamic equations}
At the phenomenological derivations of the macroscopic equations of motion
we need to use some fundamental principles
which allows to avoid some misinterpretations.
One of such fundamental principles in relativistic physics is the covariant form of equations.
However, sometimes one needs to neglect the covariant form of equations in order to find an appropriate macroscopic in a particular frame.
One of the most important examples is the quantum field theory at the finite temperatures,
where there is chosen reference frame bound to the thermostat.
In our derivation
we are not bound to particular inertial frame.
we start our derivation in the arbitrary inertial frame.
However, afterwords we work in this microscopically fixed frame in order to make transition to the macroscopic scale.
Systematic application of the equations correctly describing the microscopic motion allows us to obtain the correct macroscopic equations.
We use microscopic equation of motion presented in noncovariant form (\ref{RHD2018Cl Newton eq}).
Consequently, the obtained macroscopic equations appear in noncovariant notations.
Moreover, presented equations are found in the three-vector notations.
Nevertheless, we can use these equations for the analysis of relativistic effects working in the single inertial frame.
Introduced above hydrodynamic functions are presented in the three-vector notations.
However, they can be combined as components of the four-tensors:
\begin{equation}\label{RHD2018Cl four tensor Pi} \Pi^{\alpha\beta}=\left(
\begin{array}{cc}
n c^{2} & n v^{b} c \\
nv^{a} c & \Pi^{ab} \\
\end{array}
\right),
\end{equation}
and
\begin{equation}\label{RHD2018Cl four tensor Gamma} \Gamma^{\alpha\beta}=\left(
\begin{array}{cc}
\Gamma c^{2} & t^{b} c \\
t^{a} c & t^{ab} \\
\end{array}
\right).
\end{equation}
These are analogs of the energy-momentum tensor
which is describe below at the analysis of the four-momentum density evolution.
\section{On a possible structure of the energy-momentum tensor}
Presented method of the microscopic derivation allows to derive the energy-momentum four-vector
and derive equation for its evolution providing the energy-momentum tensor.
This equation is not a part of suggested hydrodynamic equations
because the found structure of hydrodynamic equations is a consequence revealed by the concentration evolution.
Relativistic hydrodynamic allows several generalization of concentration and velocity field.
Such as the particle current and the momentum density coincide in nonrelativistic limit,
but they are differen functions in general case.
Equations (\ref{RHD2018Cl continuity equation via velocity})-(\ref{RHD2018Cl evolution of Gamma via velocity}) show
that the energy and momentum do not appear as a part of evolution of concentration and current.
While the concentration and current are relevant since
they exist as sources of field in the Maxwell equations.
However, we show the energy-momentum $\{\epsilon/c, p^{a}\}$ evolution
to demonstrate agreement between presented method and relativistic hydrodynamics presented in other works.
Let us to point out the general matrix structure of the energy-momentum tensor together with its microscopic definitions:
\begin{equation}\label{RHD2018Cl T a b matrix structure} T^{\alpha\beta}=\left(
\begin{array}{cc}
\epsilon & p^{b}c \\
p^{a}c & T^{ab} \\
\end{array}
\right)
=\langle v_{i}^{\alpha}p_{i}^{\beta}\rangle.
\end{equation}
to the best of my knowledge all relativistic hydrodynamic models are based on the momentum balance equation.
Therefore, we present the microscopic definition of the momentum density
\begin{equation}\label{RHD2018Cl momentum density definition}
p^{\alpha}(\textbf{r},t)=\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N} p^{\alpha}_{i}(t)
\delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t)),
\end{equation}
where $p^{\alpha}_{i}(t)=\{\epsilon_{i}/c,\textbf{p}_{i}\}=\{mc\gamma_{i},m\gamma_{i}\textbf{v}_{i}\}$.
Next, we consider the time evolution of the momentum density via the calculation of the time derivative of the presented function
$$\partial_{t}p^{a}=\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N}
\biggl(q_{i}E_{i}^{a}+\frac{q_{i}}{c}\varepsilon^{abc}v_{i}^{b}B_{i}^{c}\delta_{i}\biggr)$$
\begin{equation}\label{RHD2018Cl momentum evolution time derivative}
-\partial_{b}\frac{1}{\Delta}\int_{\Delta}d\mbox{\boldmath $\xi$}\sum_{i=1}^{N}p_{i}^{a}v_{i}^{b}\delta_{i},\end{equation}
where $\delta_{i}=\delta(\textbf{r}+\mbox{\boldmath $\xi$}-\textbf{r}_{i}(t))$.
In the meanfield approximation equation (\ref{RHD2018Cl momentum evolution time derivative}) reduces to the following equation:
\begin{equation}\label{RHD2018Cl momentum evolution meanfield}
\partial_{t}p^{a}+\partial_{b}\langle mv_{i}^{a}v_{i}^{b}\gamma_{i}\rangle
=qnE^{a}+\frac{q}{c}n\varepsilon^{abc}v^{b}B^{c}. \end{equation}
the left-hand side of equation (\ref{RHD2018Cl momentum evolution meanfield}) can be rewritten via the energy-momentum tensor in the four-vector.
Therefore, the right-hand side should be rewritten in some notations via the tensor of electromagnetic field.
Equation (\ref{RHD2018Cl momentum evolution meanfield}) requires two equations of state.
First, one needs to get rid of the momentum density $p^{a}$ presenting it in terms of the velocity field
to make it in agreement with the right-hand side of momentum evolution equation (\ref{RHD2018Cl momentum evolution meanfield})
(which is represented via velocity),
the Maxwell equations and the continuity equation.
Moreover, the momentum flux $\langle v_{i}^{a}v_{i}^{b}\gamma_{i}\rangle$ (containing the pressure tensor) requires
an equation of state as the function of concentration and the velocity field.
Any hydrodynamic model requires one or several equations of state.
Our model shows that in the relativistic case there are three generalizations of the Euler equation.
they are equations for particle current (velocity field) $\textbf{j}\sim \textbf{v}$, t-vector function $\textbf{t}$, and the momentum $\textbf{p}$.
All of them have same nonrelativistic limit.
The microscopic energy momentum tensor (\ref{RHD2018Cl T a b matrix structure}) is derived
as a part of the momentum balance equation (\ref{RHD2018Cl momentum evolution meanfield}).
However, we show above that
the relativistic hydrodynamic can be contracted of several scalars $n$, $\Gamma$, $\tilde{t}$, $p$, $M_{0}$
and two vector fields $\textbf{v}$ and $\textbf{t}$.
It gives the following four-vector notations $u^{\alpha}=\{c,\textbf{v}\}$ and $\Gamma^{\alpha}=\{\Gamma, \textbf{t}\}$.
Hence, an attempt of the microscopic contraction of the energy-momentum tensor $T^{\alpha\beta}$ discussed in the Introduction
can be generalized up to account of term proportional to $\Gamma^{\alpha}\Gamma^{\beta}$
along with $u^{\alpha}u^{\beta}$ and $\delta^{\alpha\beta}$.
However, our further discussion demonstrate no necessity to extend the energy momentum tensor,
but the introduction of novel tensors $\Pi^{\alpha\beta}$ (\ref{RHD2018Cl four tensor Pi})
and $\Gamma^{\alpha\beta}$ (\ref{RHD2018Cl four tensor Gamma}).
\section{Longitudinal one-dimensional waves in the relativistic isotropic plasmas}
Let us demonstrate the application of the developed relativistic hydrodynamic model
(\ref{RHD2018Cl continuity equation via velocity Truncated})-(\ref{RHD2018Cl evolution of Theta via velocity Truncated})
on the relatively simple (but fundamentally important) examples.
We start our illustration with the high-frequency longitudinal waves in the isotropic plasmas.
We focus on the regime,
where plasma is assumed macroscopically motionless,
but this plasmas has the relativistic temperature $T\sim m_{e}c^{2}$.
The temperature is proportional to the trace of pressure,
which is a part of the momentum flux tensor $T^{ab}=\langle v_{i}^{a}p_{i}^{b}\rangle$.
hence, the temperature can be larger then $m_{e}c^{2}$ due to the relativistic $\gamma$-factor i the definition of $T^{ab}$.
Considering system is described by nonzero equilibrium concentration $n_{0}$ and the equilibrium average $\gamma$-factor $\Gamma_{0}=n_{0}K_{1}(b)/K_{2}(b)$.
The equilibrium velocity field $\textbf{v}_{0}$, vector $\textbf{t}_{0}$, and the electric field $\textbf{E}_{0}$ are equal to zero.
The equilibrium magnetic field is equal to zero and perturbations of the magnetic field are absent.
We present the corresponding linearized equations set
which follows from equations
(\ref{RHD2018Cl continuity equation via velocity Truncated})-(\ref{RHD2018Cl evolution of Theta via velocity Truncated})
for the linear approximation on the small amplitude perturbations:
\begin{equation}\label{RHD2018Cl continuity equation lin 1D}
\partial_{t}\delta n+n_{0}\partial_{x} \delta v_{x}=0, \end{equation}
\begin{equation}\label{RHD2018Cl velocity field evolution equation lin 1D}
n_{0}\partial_{t}\delta v_{x}+\partial_{x}\delta p
=\frac{e}{m}\Gamma_{0} \delta E_{x}-\frac{e}{mc^{2}}\tilde{t}_{0}\delta E^{x},
\end{equation}
\begin{equation}\label{RHD2018Cl evolution of Gamma lin 1D}
\partial_{t}\delta\Gamma +\Gamma_{0}\partial_{x}\delta v_{x}+\partial_{x}\delta t_{x} =\frac{e}{mc^{2}}\frac{q_{0}^{x}}{c^{2}}\delta E^{x}, \end{equation}
and
$$\partial_{t}\delta t_{x} +\partial_{x}\delta \tilde{t}-\frac{\Gamma_{0}}{n_{0}}\partial_{x}\delta p +\frac{e}{m}\frac{\Gamma_{0}^{2}}{n_{0}}\delta E_{x}$$
\begin{equation}\label{RHD2018Cl evolution of Theta lin 1D}
=\frac{e}{m}n_{0}\delta E_{x} -5\frac{e}{mc^{2}}p_{0}\delta E_{x} +2\frac{e}{mc^{2}}M_{0}^{xxcc}\delta E_{x}. \end{equation}
Solve equations (\ref{RHD2018Cl continuity equation lin 1D})
and (\ref{RHD2018Cl velocity field evolution equation lin 1D})
to get $\delta n$ and $\delta v_{x}$ as functions of the electric field perturbations $\delta E_{x}$
and find the following expressions:
\begin{equation}\label{RHD2018Cl v x via E}
\delta v_{x}=\imath\omega\frac{e}{m}\frac{\Gamma_{0}-\frac{1}{c^{2}}\tilde{t}_{0}}{n_{0}(\omega^{2}-U_{p}^{2}k^{2})}\delta E_{x}, \end{equation}
\begin{equation}\label{RHD2018Cl concentration lin via E}
\delta n=\imath k\frac{e}{m}\frac{\Gamma_{0}-\frac{1}{c^{2}}\tilde{t}_{0}}{\omega^{2}-U_{p}^{2}k^{2}}\delta E_{x}. \end{equation}
Functions $\delta \Gamma$ and $\delta t_{x}$ can be also found as functions of the electric field perturbations $\delta E_{x}$ from equations
(\ref{RHD2018Cl evolution of Gamma lin 1D}) and (\ref{RHD2018Cl evolution of Theta lin 1D}),
but
these functions go nowhere.
They do not affect the evolution of the concentration $\delta n$ and the velocity field $\delta v_{x}$ in the linear approximation.
They do not contribute in the equations of field either.
The nonlinear evolution should be affected by these functions.
This conclusion is correct for the isotropic medium.
More complex picture would appear at the presence of the flows and/or the magnetic field.
We substitute the concentration (\ref{RHD2018Cl concentration lin via E}) in the Poisson equation $\nabla\cdot\textbf{E}=4\pi\rho$,
where $\rho$ is the charge density of the plasmas.
In our case it goes to
$\partial_{x}E_{x}=4\pi e\delta n$.
It gives the spectrum of the relativistic Langmuir waves:
\begin{equation}\label{RHD2018Cl Langm wave spectrum}
\omega^{2}=\omega_{Le}^{2}\frac{\Gamma_{0}-\frac{1}{c^{2}}\tilde{t}_{0}}{n_{0}} +U_{p}^{2}k^{2}.
\end{equation}
The Langmuir wave spectrum (\ref{RHD2018Cl Langm wave spectrum}) requires three equations of state,
while whole set of linearized equations includes more unknown functions approximately calculated for the truncation.
Necessary equations of state are demonstrated above,
where it is shown
that function $\tilde{t}_{0}$ is represented via the characteristic velocity $U_{t}$.
However, we want to discuss briefly its more general structure for different values of the velocity field $\textbf{v}$ contributing in $\tilde{t}_{0}$ in the general case.
Function $\tilde{t}$ has the following structure
in the arbitrary regime, where
the thermal velocity $\textbf{u}_{i}$ and the velocity field $\textbf{v}$ are relativistic and comparable to each other
\begin{equation}\label{RHD2018Cl tilde t arbitrary v and u}\tilde{t}=\biggl\langle\frac{u_{i}^{x}u_{i}^{x}}{\gamma_{i}}\biggr\rangle
=\biggl\langle u_{i}^{x}u_{i}^{x}\sqrt{1-\frac{\textbf{v}^{2}+2\textbf{u}_{i}\cdot\textbf{v}+\textbf{u}_{i}^{2}}{c^{2}}}\biggr\rangle.\end{equation}
Function $\tilde{t}$ reduces to the flux of particle current
If we dial with the relativistic beam (beam of electrons for instance),
where the thermal effects are neglegible
\begin{equation}\label{RHD2018Cl t via p cold}\tilde{t}=\sqrt{1-\frac{\textbf{v}^{2}}{c^{2}}}\cdot p,\end{equation}
where
the velocity field of the beam $\textbf{v}$ is also incorporated.
However, the opposite limit is more interesting.
If the thermal effects dominate over the velocity field existing in the relativistically hot beamless plasmas
function $\tilde{t}$ combines of fluxless part of $\tilde{t}$ and corrections caused by the velocity field:
\begin{equation}\label{RHD2018Cl expansion of t}\tilde{t}\approx\biggl\langle u_{i}^{x}u_{i}^{x} \sqrt{1-\frac{\textbf{u}_{i}^{2}}{c^{2}}}\biggr\rangle
-\frac{1}{c^{2}}v^{a}\biggl\langle u_{i}^{x}u_{i}^{x}u_{i}^{a} \frac{1}{\sqrt{1-\frac{\textbf{u}_{i}^{2}}{c^{2}}}}\biggr\rangle.\end{equation}
For the further analysis of spectrum (\ref{RHD2018Cl Langm wave spectrum})
we consider the main term in the expansion (\ref{RHD2018Cl expansion of t}).
Hence, we completely neglect the contribution of the velocity field in this function.
Finally, we have the following spectrum for the longitudinal waves in the macroscopically motionless isotropic plasmas:
\begin{equation}\label{RHD2018Cl Langmuir wave spectrum}
\omega^{2}=\omega_{Le}^{2}\biggl(\frac{K_{1}}{K_{2}}-\frac{U_{t}^{2}}{c^{2}}\biggr) +U_{p}^2 k^{2}.
\end{equation}
Fraction $\frac{K_{1}(b)}{K_{2}(b)}$ goes to 1 at the small temperatures (so the argument $b=mc^{2}/T$ goes to infinity),
since the argument of the Macdonald functions goes to infinity.
Numerical analysis of spectrum (\ref{RHD2018Cl Langmuir wave spectrum}) can be found in Ref. \cite{Andreev 2021 05}.
Nevertheless, we want to point out that additional multiplayer in front of the square of the Langmuir frequency is below unit.
Hence, the relativistic temperature effects reduces the minimal frequency of the minimal frequency of the longitudinal wave below $\omega_{Le}$.
This decrease can be of several orders.
\section{Relativistic beams in the nonrelativistic plasmas}
Major goal for the development of the presented hydrodynamic model is the study of plasmas with the large temperatures of order of $m_{e}c^{2}$ and above.
However, for justification of the model,
it is useful to consider relativistic effects in absence of the temperature effects.
This regime corresponds to the propagation of the relativistic beam through the plasmas.
Let us consider the high-frequency perturbations in the macroscopically motionless plasmas through which the electron beam propagates.
The beam is characterized by the equilibrium concentration $n_{0b}\ll n_{0e}$.
The equilibrium velocity of beam is $\textbf{v}_{0}=\{v_{0},0,0\}$.
We consider propagation of excitations parallel to the direction of the beam propagation $\textbf{k}=\{k,0,0\}$.
Under described conditions
the set of hydrodynamic equations
(\ref{RHD2018Cl continuity equation via velocity Truncated})-(\ref{RHD2018Cl evolution of Theta via velocity Truncated})
used for each species gives the following dispersion equation:
\begin{equation}\label{RHD2018Cl dispersion eq with beam}
1-\frac{\omega_{Le}^{2}\biggl(\frac{K_{1}}{K_{2}}-\frac{U_{t}^{2}}{c^{2}}\biggr)}{\omega^{2}-U_{p}^2 k^{2}} -\frac{\omega_{Lb}^{2}}{\gamma_{0}^{3}(\omega-kv_{0})^{2}}=0, \end{equation}
where $\gamma_{0}=\sqrt{1-v_{0}^{2}/c^{2}}^{-1}$,
and
$\omega_{Lb}^{2}=4\pi e^{2}n_{0b}/m_{e}$.
Presented cold regime corresponds to equation found in Refs.
\cite{Bret PoP 2006},
\cite{Bret PPCF 06},
\cite{Bret ApJ 09},
\cite{Bret PoP 2016},
\cite{Akhiezer book 74}.
\section{Plane waves in the relativistically hot magnetized plasmas}
\subsection{Linearly polarized waves propagating parallel to the magnetic field}
Similarly to the Sec. VII
we consider the plasmas
which are macroscopically motionless in the equilibrium state.
However, here we consider the plasmas placed in the uniform constant magnetic field.
The equilibrium state is characterized by the constant nonzero values for the concentration $n_{0}$,
the equilibrium average $\gamma$-factor $\Gamma_{0}$,
and the magnetic field $\textbf{B}_{0}=B_{0}\textbf{e}_{z}$.
We consider waves propagating parallel to the external magnetic field $\textbf{k}=\{0,0,k\}\parallel \textbf{B}_{0}$
with the electric field perturbations parallel to the external magnetic field $\delta \textbf{E}=\{0,0,E_{z}\}\parallel \textbf{B}_{0}$ as well.
We focus on the small amplitude excitations and consider they linear dynamic.
In this regime the set of hydrodynamic equations
(\ref{RHD2018Cl continuity equation via velocity Truncated})-(\ref{RHD2018Cl evolution of Theta via velocity Truncated})
reduces to the following couple of equations:
\begin{equation}\label{RHD2018Cl eq v z lin with B} -\imath\omega n_{0}\delta v_{ez}
=\frac{q_{e}}{m}\Gamma_{0}\delta E_{z}-\frac{q_{e}}{mc^{2}}U_{t}^{2} n_{0}\delta E_{z}, \end{equation}
and
\begin{equation}\label{RHD2018Cl eq E z lin with B} (\omega^{2}-k^{2}c^{2})\delta E_{z}+4\pi\imath\omega q_{e}n_{0e}\delta v_{ez}=0 \end{equation}
Equations (\ref{RHD2018Cl eq v z lin with B}) and (\ref{RHD2018Cl eq E z lin with B}) give the following spectrum:
\begin{equation}\label{RHD2018Cl spectrum transverse} \omega^{2}=k^{2}c^2+\omega_{Le}^{2}\biggl(\frac{K_{1}}{K_{2}}-\frac{U_{t}^{2}}{c^{2}}\biggr) \end{equation}
It gives same result as the electromagnetic one-dimensional waves in the relativistic isotropic plasmas.
Equation (\ref{RHD2018Cl spectrum transverse}) shows that
the transverse waves includes same reduction of the Langmuir frequency
as it happens for the longitudinal waves (\ref{RHD2018Cl Langmuir wave spectrum}).
\section{Conclusion}
Hydrodynamic model for the relativistically hot ideal plasmas has been developed as a chain of equations.
Derivation of equations is performed for functions
which appear in previous equations.
For example, derivation of the continuity equation leads to the current of particles.
Next, the equation for evolution of the current has been derived.
The particle current evolution contains four novel functions.
Two of them are the three-scalar and three-vector.
Therefore, equations for the evolution of these functions have been found.
So, no functions is introduced \textit{ad hoc}.
Therefore, no energy-momentum tensor is involved.
In stead of it two four-tensors relatively to equations for two four-vectors have been found.
They are the four-particle current,
which also exists in the Maxwell equations as the source of field,
and the four-Gamma vector,
which is the average of the reverse gamma factor,
and the average of the thermal velocity divided by the relativistic gamma factor.
These variables are chosen for a simple reason.
The evolution of four-particle current leads to appearance of four-Gamma vector.
While the evolution of four-Gamma vector produces the four-particle current.
Therefore, the set of equations tries to close itself.
Few new functions similar to the pressure appear either.
Therefore, the set of equations is not completely closed.
Consequently, the chain of equations can be continued.
Or, the chain can be truncated at this stage forming an example of minimal coupling model for relativistic plasmas.
Relativistic plasmas show particular interest in the regime of the relativistically hot plasmas.
Consequently, details of evolution in the momentum space are essential.
However, the application of kinetic equations can be rather complicate.
Therefore, we need as much of higher moments as it is possible to cover the reach kinetic behavior.
Presented model does not cover high tensor dimensional hydrodynamic functions,
but consider two variations of four-vectors.
There is the third variation of four-vector.
It is the four-momentum.
However, it is not appear in obtained equations.
Derivation of hydrodynamic equations starts with the definition of concentration,
where the integral operator explicitly averages the microscopic dynamics over the macroscopically small volume.
The truncation presiger is discussed after derivation of basic equations.
The closed set of equations is applied for study of fundamental collective phenomena in relativistic plasmas.
The spectrum of Langmuir waves is obtained.
The dispersion equation for the cold relativistic beam propagating through macroscopically motionless relativistically hot plasmas is demonstrated.
The cold regimes, particularly, the beam propagation, shows agrement with the well-known models of relativistic plasmas motion.
While, the thermal effects, particularly in the magnetic field, gives new information about plasma dynamics in hydrodynamic regime
(these effects are discussed in earlier papers).
\section{Acknowledgements}
Work is supported by the Russian Foundation for Basic Research (grant no. 20-02-00476).
This paper has been supported by the RUDN University Strategic Academic Leadership Program.
\section{DATA AVAILABILITY}
Data sharing is not applicable to this article as no new data were
created or analyzed in this study, which is a purely theoretical one.
|
\subsection{Proof of Lemma \ref{lemma:distribution}}
\begin{proof}
Notice that $\sum_{q_1=0}^{\infty}\sum_{q_2=-d}^{\infty}\pi_{q_1,q_2} = 1$. We can then prove the result by checking that $\pi_{q_1,q_2}$ satisfies the balance equation of the Markov chain because the steady state distribution is the unique solution to the balance equation given that it is normalized\cite{harchol2013performance}. For any $x$, denote $x^+=\max(0,x)$. In the considered Markov chain, for a state $(q_1,q_2)$, its transitions are given as follows:
\[
(q_1,q_2) \rightarrow \left \{
\begin{aligned}
(q_1+1,q_2) &,~\text{of rate }1; \\
(q_1-1,q_2+1) &,~\text{of rate }\mu~\text{if }q_1 > 0; \\
(q_1,q_2-1) &,~\text{of rate }\lambda_{(q_2-M)^+}~\text{if }q_2 > -d.
\end{aligned}
\right.
\]
For any two state $Q,P$, denote the rate from $Q$ to $P$ by $r_{Q,P}$. Fix a state $Q=(q_1,q_2)$. Here we only check the balance equation for state $(q_1,q_2)$ satisfying $q_1 > 0,q_2 > -d$. The same strategy can verify other cases similarly. The balance equation of this state is that
\[
\text{inflow}\colon = \sum_{P} \pi_{P}r_{P,Q} = \sum_{P} \pi_Q r_{Q,P} =\colon \text{outflow}.
\]
We need to consider three cases. First, suppose that $q_2 < M$. Then by the form of $\pi$, it holds
\[
\begin{aligned}
\text{inflow}&=\pi_{q_1-1,q_2}+\pi_{q_1,q_2+1}\lambda_0+\pi_{q_1+1,q_2-1}\mu \\
&= C\left(\frac{1}{\mu^{q_1-1}\lambda_0^{q_2}}+\frac{1}{\mu^{q_1}\lambda_0^{q_2}}+\frac{1}{\mu^{q_1}\lambda_0^{q_2-1}}\right) \\
&= \pi_{q_1,q_2}\left(1+\lambda_0+\mu\right)=\text{outflow}.
\end{aligned}
\]
Then if $q_2 = M$, it satisfies that
\[
\begin{aligned}
\text{inflow}&=\pi_{q_1-1,q_2}+\pi_{q_1,q_2+1}\lambda_1+\pi_{q_1+1,q_2-1}\mu \\
&= C\left(\frac{1}{\mu^{q_1-1}\lambda_0^{q_2}}+\frac{1}{\mu^{q_1}\lambda_0^{q_2}}+\frac{1}{\mu^{q_1}\lambda_0^{q_2-1}}\right) \\
&= \pi_{q_1,q_2}\left(1+\lambda_0+\mu\right)=\text{outflow}.
\end{aligned}
\]
Finally, for $q_2 > M$,
\[
\begin{aligned}
&\mspace{23mu}\text{inflow} \\
&=\pi_{q_1-1,q_2}+\pi_{q_1,q_2+1}\lambda_{q_2+1-M}+\pi_{q_1+1,q_2-1}\mu \\
&= C\left(\frac{1}{\mu^{q_1-1}\lambda_0^{q_2}}+\frac{1}{\mu^{q_1}\lambda_0^{M}}\left(\prod_{i=1}^{q_2-M}\frac{1}{\lambda_i}+\prod_{i=1}^{q_2-M-1}\frac{1}{\lambda_i}\right)\right) \\
&= \pi_{q_1,q_2}\left(1+\lambda_{q_2-M}+\mu\right)=\text{outflow},
\end{aligned}
\]
which completes the proof.
\end{proof}
\subsection{Proof of Lemma \ref{lemma:meanValue}}
\begin{proof}
Define $\expect{N_o}$ be the expected number of unsent orders. By Little's Law, $\expect{T_o} = \expect{N_o}$ because orders' arrival rate is $1$. It holds that
\[
\expect{N_o} = \sum_{q_1=0}^{\infty}\sum_{q_2=0}^{\infty} (q_1+q_2)\pi_{q_1,q_2}.
\]
Simplifying Lemma \ref{lemma:distribution} by setting $M = \infty$. We have $\pi_{q_1,q_2} = (1-\rho)\rho^d(1-\frac{1}{\mu})\frac{1}{\mu^q_1}\rho^{q_2}$ for $q_1 \geq 0,q_2 \geq -d$.
Then
\[
\begin{aligned}
\expect{N_o} &= \sum_{q_1,q_2=0}^{\infty} q_1 \pi_{q_1,q_2} + \sum_{q_1,q_2=0}^{\infty} q_2 \pi_{q_1,q_2} \\
&= \frac{1}{\mu - 1} + \sum_{q_2=0}^{\infty} (1-\rho)\rho^{d+j}j\\
&= \frac{1}{\mu - 1} + \frac{\rho^{d+1}}{1-\rho}.
\end{aligned}
\]
On the other hand, define $\expect{N_r}$ be the expected waiting riders. Again, by Little's Law, $\expect{N_r} = \hat{\lambda}\expect{T_r}$, where $\hat{\lambda}$ is the "true" dispatch rate of riders. That is,
\[
\hat{\lambda} = \lambda_0\left(1 - \sum_{q_1=0}^{\infty}\pi_{q_1,-d}\right)=\lambda_0\rho = 1.
\]
Then for $\expect{N_r}$, it holds
\[
\begin{aligned}
\expect{N_r} = \sum_{q_1=0}^{\infty}\sum_{q_2=-d}^{-1} -q_2 \pi_{q_1,q_2} &= (1-\rho)\rho^d\sum_{j=1}^d j\rho^j \\
&= d - \frac{1 - \rho^d}{\lambda_0 - 1},
\end{aligned}
\]
and thus $\expect{T_r} = d - \frac{1 - \rho^d}{\lambda_0 - 1}$.
\end{proof}
\subsection{Formal Proof of Theorem \ref{thm:optimal-algo}}
We first have the following lemma demonstrating the optimality of choosing smallest feasible $d$.
Recall the function $D(\rho)$ and the function $\gamma(n)$ from Algorithm \ref{algo:optimal-parameter} defined for a fixed $T^*$. Let $d_l =max(1,D(\frac{1}{\Lambda}))$. For every $d \geq d_l$, define $h(d) = \expectsub{\frac{1}{\gamma(d)},d}{T_r}$.
\begin{lemma}\label{lemma:monotonicity}
It holds that $h(d)$ is a strictly increasing function when $d \geq d_l$.
\end{lemma}
\begin{proof}
Let $\rho_l$ be the smallest real number such that $\frac{\rho_l^{2}}{1-\rho_l} \geq T^*$. And if $\rho_l < \frac{1}{\Lambda}$, set it to be $\frac{1}{\Lambda}$. We can then see that $d_l = D(\rho_l)$. Now for every $\rho \geq \rho_l$, define $x(\rho)$ such that $\frac{\rho^{x(\rho)+1}}{1-\rho} = T^*$. Then $x(\rho)+1 = \frac{\ln T^* + \ln(1-\rho)}{\ln \rho} \geq 1$. Note that $x(\rho)$ may not be integer.
We claim that for every $d \geq d_l$, it holds $x(\gamma(d)) = d$. That is, $x$ is like a reverse function of $\gamma$, but it extends the definitional domain. We prove this claim by contradiction. Suppose it is not. It must hold that $x(\gamma(d)) < d$, and $\frac{(\gamma(d))^{d+1}}{1-\gamma(d)} < T^*$. However, in this case, we can find a $\rho > \gamma(d)$ such that $\frac{\rho^{d+1}}{1-\rho)} \leq T^*$, which contradicts the fact that $\gamma(d)$ is the largest such possible $\rho$. As a result, we must have $x(\gamma(d))=d$.
For $\rho \geq \rho_l$, define $g(\rho)\colon= x(\rho)-\frac{\rho - \rho^{x(\rho)+1}}{1-\rho}$. Then to prove that $h(d)$ is strictly increasing, it is sufficient to prove that $g(\rho)$ is strictly increasing since $h(d) = g(\gamma(d))$. Furthermore, observe that for $\rho \geq \rho_l$, $\frac{\rho^{x(\rho)+1}}{1-\rho} = T^*$, which is a constant. It is enough to prove the function $G(\rho)\colon=\frac{\ln T^* + \ln(1-\rho)}{\ln \rho}-\frac{\rho}{1-\rho}-1$ is strictly increasing.
To prove it, take derivative of $G$, and we can get
\begin{equation}
\begin{aligned}
G'(\rho) &= \frac{-1}{(1-\rho)\ln \rho}-\frac{\ln T^*+\ln(1-\rho)}{\rho(\ln \rho)^2}-\frac{1}{(1-\rho)^2} \\
&=\frac{-1}{(1-\rho)\ln \rho}-\frac{(x(\rho)+1)\ln \rho}{\rho(\ln \rho)^2}-\frac{1}{(1-\rho)^2} \\
&\geq \frac{-1}{(1-\rho)\ln \rho}-\frac{1}{\rho\ln \rho}-\frac{1}{(1-\rho)^2},
\end{aligned}
\end{equation}
where the last inequality is because $x(\rho)+1\geq 1$ and $\ln(\rho) < 0.$
It remains to show that $G_1(\rho)=\frac{-1}{(1-\rho)\ln \rho}-\frac{1}{\rho\ln \rho}-\frac{1}{(1-\rho)^2}$ is positive in $(0,1)$. Simplifying the terms gives $G_1(\rho)=\frac{-\rho+\rho\ln \rho + 1}{-(1-\rho)^2\rho \ln \rho}$. For the function $G_2(\rho)=-\rho + \rho \ln \rho + 1$, its derivative is $\ln \rho$, which is negative in $(0,1)$. Therefore, the function $G_2(\rho)$ is a strictly decreasing function in $(0,1)$. Since $G_2(1) = 0$, we know both $G_2(\rho)$ is always positive in $(0,1)$, and so is $G_1(\rho)$. Consequently, $G(\rho)$ is a strictly increasing function, which completes the proof.
\end{proof}
We can now finish the proof of Theorem \ref{thm:optimal-algo}.
\begin{proof}
In Algorithm \ref{algo:optimal-parameter}, we first check whether $\frac{\Lambda}{1-\Lambda} \leq T^*$. If it is, it means it is feasible to set $d = 0$, and $\lambda_0 = \Lambda$. It is optimal because $\expectsub{\Lambda,0}{T_r} = 0$.
Otherwise, we have to set $d \geq 1$. Notice that when $d$ is fixed, $\expectsub{\Lambda,0}{T_o}$ is a decreasing function, while $\expectsub{\Lambda,0}{T_r}$ is an increasing function. Therefore, suppose we want to select $d$ one by one. We can first check whether the pair of $(\lambda_0 = \Lambda,d)$ can satisfy the constraint in (\ref{eq:optimization}). If it does, then we decrease the value of $\lambda_0$ to obtain lower averaged rider waiting time. The above discussion claims that feasible $d$ must be at least $D(\frac{1}{\Lambda})$. And for each feasible $d$, the optimal choice is to choose $\lambda_0 = \frac{1}{\gamma(d)}$. The averaged rider waiting time given by this parameter setting is exactly $h(d)$. Now according to Lemma \ref{lemma:monotonicity}, $h(d)$ is an increasing function. Therefore, the optimal solution is to set $d = D(\frac{1}{\Lambda})$, and $\lambda_0 = \frac{1}{\gamma(d)}$.
\end{proof}
\subsection{Proof of Theorem \ref{thm:data-value}}
We now provide the proof of Theorem \ref{thm:data-value}. The intuition is that,
by Little's Law, it suffices to prove that after using the new policy $(\vec{\tau},d,m)$, the mean number of waiting riders remains the same, but the mean number of waiting orders decreases.
Then the proof starts by coupling the Markov chain of the threshold $m$ and that of the threshold $M$, $M > m$. After that, we propose an optimization problem to minimize the mean number of waiting orders, and use a greedy policy to solve it. Algorithm \ref{algo:policy-improve} corresponds to the final outcome of this policy, and thus makes an improvement.
\subsubsection{System Coupling}
To enable analysis, recall the discussion in Section \ref{sec:queue-decoupling}, we only need to consider the equivalent Markov chain of $\mathcal{Q}_2$ to calculate the mean numbers of waiting orders and riders. Denote the new Markov chain of $(\vec{\lambda},d,M)$ as $A$, and that of $(\vec{\tau},d,m)$ as $B$. Denote their steady-state distributions by $\pi_A(q), \pi_B(q)$. We can now couple the two Markov chains on state $m$. Recall that $\rho = \frac{1}{\lambda_0}$. By solving balance equations with state $m$ as the initial start, we could notice that
\[
\pi_A(m)\left(\sum_{i=-d}^{M} \rho^{i-m}+\rho^{M-m}\sum_{i=M+1}^{\infty}\prod_{j=1}^{i-M} \frac{1}{\lambda_j}\right) = 1,
\]
and
\[
\pi_B(m)\left(\sum_{i=-d}^{m} \rho^{i-m} + \tau_1\sum_{i=m+1}^{\infty} \left(\frac{1}{\Lambda}\right)^{i-m-1}\right)=1.
\]
Then by the definition of $\tau_1$, it holds that $\pi_A(m) = \pi_B(m)$. As a result, by Lemma \ref{lemma:equivalence}, it leads to the fact that
\[
\expectsub{\vec{\tau},d,m}{T_r} =\sum_{i=-d}^{-1}(-i)\rho^{i-m}\pi_A(m)= \expectsub{\vec{\lambda},d,M}{T_r}.
\]
\subsubsection{Optimization Framework and Greedy Improvement}
It remains to show that the mean number of waiting orders decreases. For Markov chain $A$, define $a_q = \frac{\pi_A(q)}{\pi_A(m)}$ for $q > 0$. Similarly, define $b_q$ for Markov chain $B$. Indeed, $a_q = \frac{a_{q-1}}{\lambda_{\max(q-M,0)}}$. Then by Lemma \ref{lemma:equivalence}, we have
\[
\expectsub{\vec{\lambda},d,M}{T_o} - \expectsub{\vec{\tau},d,m}{T_o} = \pi_A(m)\sum_{q=1}^{\infty} q\left(a_q - b_q\right).
\]
Furthermore, we can observe that by the definition of $\tau_1$, it holds $\sum_{q=1}^{\infty} a_q = \sum_{q=1}^{\infty} b_q < \infty$. Denote this constant by $H$. To show that $\expectsub{\vec{\lambda},d,M}{T_o} \geq \expectsub{\vec{\tau},d,m}{T_o}$, it suffices to prove $\{b_q\}$ is the solution to the following optimization problem:
\begin{equation}\label{eq:optimize-delay}
\begin{aligned}
& \underset{\vec{x}=x_1,\cdots}{\text{minimize}}
& & \sum_{i=1}^{\infty} i\cdot x_i \\
& \text{subject to}
& & \sum_{i=1}^{\infty} x_i = H \\
& & & \frac{x_i}{x_{i - 1}} \geq \frac{1}{\Lambda}, \forall i > 1 \\
& & & x_1 \geq \frac{1}{\Lambda}.
\end{aligned}
\end{equation}
The requirement of $\frac{x_i}{x_{i - 1}} \geq \frac{1}{\Lambda}$ comes from the fact that the dispatch rate cannot exceed $\Lambda$.
Greedily, given a feasible setting of $\vec{x}$ in \ref{eq:optimize-delay}, the best strategy is to keep $x_i$ large for small $i$, and keep $x_i$ small for large $i$. Particularly, we claim that the solution must satisfy $\frac{x_{i+1}}{x_i} = \frac{1}{\Lambda}$ for all $i \geq 1$. Otherwise, we can find the smallest $i$ with $\frac{x_{i+1}}{x_i} > \frac{1}{\Lambda}$. It always improves the objective by decreasing $x_{i+1}$ and increasing $x_i$. Then we can solve out $x_1$, and it turns out that the produced solution is indeed $\{b_q\}$, which completes the proof.
\hspace*{\fill}~\QED\par\endtrivlist\unskip
\subsection{Proof of Theorem \ref{thm:tradeoff}}
Fix $\lambda_0$ and $d$ satisfying the constraint in (\ref{eq:optimization}). By the assumption in Theorem \ref{thm:tradeoff}, it holds that $\frac{1}{\lambda_0} \geq 0.3$ and $d \geq 3$. Then to show that there is a trade-off between customer experience and labor right, one key observation is that, by Lemma \ref{lemma:meanValue},
\begin{equation}
\expect{T_r} + \expect{T_o}-\frac{1}{\mu-1} = d + \frac{2\rho^{d+1}-\rho}{1-\rho},
\end{equation}
where $\rho = \frac{1}{\lambda_0}$. Let $f(\rho) = \frac{2\rho^{d+1}-\rho}{1-\rho}$. We then consider two cases.
First, when $\rho^{d+1} \geq \frac{3}{4}$, it holds $f(\rho) \geq \frac{1}{2(1-\rho)}$ because $\rho < 1$. Therefore, $\expect{T_r} + \expect{T_o}-\frac{1}{\mu-1} \geq \frac{1}{2(1-\rho)}.$
Second, when $\rho^{d+1} < \frac{3}{4}$, we claim that $\expect{T_r} + \expect{T_o}-\frac{1}{\mu-1} \geq \frac{d}{2}$. To see why it is true, we only need to show that $f(\rho)=\rho \frac{2\rho^d-1}{1-\rho} \geq -\frac{\rho d}{2}$. For $x \in [0,1)$, Define $u(x) = \frac{2x-1}{1-x^{1/d}}.$ We claim that $u(x) \geq -\frac{d}{2}$.
\begin{proof}
If not, then it holds that
\[
\frac{2x-1}{1-x^{1/d}} \leq -\frac{d}{2} \Longleftrightarrow 0 \leq -\frac{d}{2}+\frac{d}{2}x^{1/d}-2x+1.
\]
Define $v(x) = -\frac{d}{2}+\frac{d}{2}x^{1/d}-2x+1$. Take derivative of $v(x)$, we know $v'(x) = \frac{1}{2}x^{1/d-1}-2$. Since $d \geq 3$, $v'(x)$ is a decreasing function. In addition, $\lim_{x \to 0^+} v'(x) > 0$, and $v'(1) < 0$. Therefore, the value $\bar{x}$ such that $v'(\bar{x}) = 0$ is the maximum point of $v(x)$. Set $v'(\bar{x}) = 0$. It holds $\bar{x} = 4^{\frac{-d}{d-1}}$. Therefore, the maximum value of $h(x)$ for $x \in (0,1)$ would be $h(\bar{x}) = -\frac{d}{2}+\frac{d}{2}4^{\frac{1}{d-1}}-2\bar{x}+1 \leq -2\cdot 4^{-1.5}+1<0$. It thus contradicts the assumption that there is some $x$ satisfying $h(x) \geq 0$, which completes the proof.
\end{proof}
As a result, it holds $\expect{T_r}+\expect{T_o}-\frac{1}{\mu - 1} \geq \frac{d}{2} \geq \frac{\ln(4/3)}{-2\ln \rho}$ since $\rho^d \leq \frac{3}{4}$. Furthermore, by assumption, $\rho \geq 0.3$, meaning that $-\ln \rho \leq 2(1-\rho)$. Therefore, summarizing the above discussions, we have
\begin{equation}
\expect{T_r} \geq \frac{\ln(4/3)}{4(1-\rho)}-\left(\expect{T_o}-\frac{1}{\mu-1})\right) \geq \frac{\ln(4/3)}{4(1-\hat{\rho})}-T^*,
\end{equation}
where $\hat{\rho} = \frac{1}{\Lambda}$.\hspace*{\fill}~\QED\par\endtrivlist\unskip
\section{INTRODUCTION}
\subsection{Background}
Labor rights in meal delivery have become one of the top public concerns in platform economy. During the last decade, the meal-delivery industry has been flourishing due to the penetration of smart mobile phones. However, a growing number of traffic violations and accidents due to risky driving of riders have attracted people to discuss the relationship between platform algorithms and labor rights. Current statistics reveal that meal delivery riders contribute to a significant share of traffic violations. In 2017, approximately one food delivery rider was reported to be injured or killed every 2.5 days in Shanghai, China\cite{ChinaNews}. The emerging food-delivery platform also contributes to a nonnegligible fraction of traffic accidents in Korea \cite{byun2017characteristics}. In the U.S., food delivery riders are viewed as one of the most dangerous jobs\cite{DangerousJob}.
It is necessary to develop and analyze dispatch algorithms of meal-delivery platforms from the labor-right protecting perspective. However, the current algorithmic models for meal delivery are either user-centric or platform profit-driven and ignore the perspective of protecting riders from traffic risks. The absence of the perspective of labor rights of riders causes two problems. First, the current dispatch models simplify the procedure of the working flow of riders by merging riders’ picking-up order period and delivery period into one period. This oversimplification causes the current models to ignore the potential of improving rider traffic safety by managing the rider waiting time during the period of picking up orders. Consequently, user-centric or profit-driven dispatch results in overwaiting riders. Second, the user-centric dispatch focuses mainly on timely delivery for attracting consumers while profit-driven dispatch pursues minimizing the population of riders for cost saving\cite{DriverRisk, christie2019health}. Consequently, the current dispatch algorithms limit the number of riders and continuously shorten the time window of deliveries to attract more consumers by quickly delivering foods. Both mechanisms, user-centric and profit-driven dispatch, contribute to overload of riders. The above two problems of the current food-delivery models together drive riders to overspeed.
In this research, we propose a queuing-model-based framework of designing dispatch policies from the perspective of protecting the traffic safty of riders in meal delivery platforms. In contrast to user-centric dispatch or profit-driven dispatch, in this research, we develop a framework that carefully models the working flow of riders and seeks to mitigate the stress of the delivery-time limit that forces riders to drive at excessive speeds. Our rider-centric framework enables us to analyze the pathway of protecting rider safety. According to our framework, we develop a labor-rights-protection dispatch algorithm to control the waiting time of riders in restaurants. Our algorithm can achieve the optimal dispatch strategy that minimizes the waiting time of riders with guaranteed consumer-acceptable order waiting time. We also apply our framework to explore how restaurant information can further improve the labor rights of riders with a refined platform dispatch policy.
Our rider-centric framework further allows us to comprehensively capture the impacts from all stakeholders, including restaurants, consumers, platforms, and riders themselves. We found that information about the number of restaurant offline orders is critical for managing rider waiting time for orders. Therefore, our framework decomposes the meal-delivery task into two processes: an information disclosure process and a dispatch process. The information disclosure process captures the impact of restaurants sharing their information, while the dispatch process captures the impact of the patience of consumers and decisions of platforms. We found that restaurants can help with labor rights protection by sharing their private information about the number of their offline orders. This conclusion reveals the tight connection between private information and labor rights in platform economy.
\subsection{Related work}
The growth of meal delivery platforms opens a wide range of research areas. Due to the unique process structure of meal deliveries, queuing models are frequently used to understand profit brought by meal delivery platforms, or broadly on-demand platforms \cite{bai2019coordinating,taylor2018demand,feldman2019can,chen2019food,chen2020courier}. The current user-centric or profit-driven dispatch adopts mainly a single-sided queueing model focusing on timely service of the randomly arriving orders by the least size of rider population. In particular, \cite{feldman2019can} studies the benefit of delivery platforms to restaurants by analyzing an associated queuing model and proposing new contracts. Then, in \cite{chen2019food}, the authors propose a queuing model with a pricing scheme to study interactions between platforms and customers. Their model also discusses the impact of a finite rider pool on the profit of a platform. In addition, the model of \cite{chen2020courier} analyzes the benefit of order batching via queues with batch arrivals.
From a technical perspective, our queueing model belongs to the class of two-sided queues, whose study can be dated back to the 1990s \cite{gelenbe1991queues}. In the original model, there are arrivals of positive jobs to one single queue. There are also negative jobs that can cancel one positive job waiting in the queue. However, when there is no positive job, the negative job will directly leave. The sojourn time of positive jobs has been derived \cite{harrison1993sojourn}, and the model has also been extended to more general service time \cite{harrison1996m} and arrival processes \cite{li2004map}. The stationary distribution is known to have a product form in queueing networks with negative arrivals \cite{henderson1993queueing, gelenbe1991product}. Our model, however, is different from previous studies in the sense that negative arrivals, i.e., riders, can wait in another queue for future orders. The arrival process of riders can depend on the queue length of orders instead of a time-homogeneous process.
In addition to queuing models, there are optimization and reinforcement learning studies aiming at developing online optimization strategies for efficient real-time meal deliveries. A stream of work models rider dispatch problems by optimizations with the objective of maximizing the number of timely orders or minimizing the averaged delays \cite{yildiz2019provably, reyes2018meal,joshi2020batching}. In addition, studies of meal deliveries when a customer's order can include food from multiple restaurants is studied in \cite{steever2019dynamic}. Randomness in meal delivery problems is considered in the optimization framework in \cite{ulmer2020restaurant,liu2020time}. The model in \cite{liu2020integrating} uses deep inverse reinforcement learning to learn a better dispatch policy. In addition to providing efficient dispatch, customer satisfaction is another core concern of current dispatch studies. Some recent studies discussed reliable estimation of arrival time by using supervised learning on deliveries \cite{zhu2020order,hildebrandt2020supervised}. There also exist works pursuing improving customer's satisfaction by designing reasonable range of customers to whom each restaurant can expose \cite{ding2020delivery}.
Although the labor rights of riders have attracted sufficient social attention, there are only limited studies on the topics, most of which focus on the impacts of wage design on welfare distribution. For instance, \cite{benjaafar2020labor,wu2020two} examine how wages in platforms impact customers and laborers but have not discussed how the process of dispatching may impact labor welfare. To the best of our knowledge, there is still a lack of systematic discussion about improving the dispatch for mitigating the time-limit stress on riders, which forces riders to drive riskily during delivery orders. We also found that discussion about the impacts of information on labor rights is rare. \cite{mao2019faster,liu2020time} integrate the estimation of travel times, attitudes of customers toward late orders and familiarity of riders with roads into optimization models and show by simulations that data can improve dispatch efficiency. Nevertheless, to the best of our knowledge, there is yet no work discussing the impact of restaurant data and providing theoretical insights into how restaurant data can help improve dispatch efficiency and labor welfare.
\subsection{Labor rights of riders in food-delivery platforms}
Labor rights can be divided into two categories: 1) working benefits such as minimum wages and reasonable working hours; 2) the safety of working conditions\cite{laborright}. Most current relative discussions in the context of platform economy focus on the first economic aspect of labor rights \cite{de2015rise, benjaafar2020labor, parrott2018earnings}, while the working safety of riders is absent from the current literature about platform economy.
Here, our work provides a first step to consider the driving safety of riders in platform economy by minimizing the waiting time of riders. We focus on the waiting time of riders in restaurants for two reasons. First, minimizing the waste of time during the order-picking-up process allows riders to have sufficient time to deliver the order. Second, in contrast to user-centric or profit-driven dispatch, minimizing the rider's waiting time in restaurants can prevent the algorithm from continuously shortening the available time window of deliveries. We believe that our work could offer a new angle for the control community to think about platform economy.
\subsection{Contributions}
We summarize our three contributions below.
\begin{itemize}
\item \textbf{Framework of modeling labor-right protecting problems in meal-delivery platforms.} In contrast to the literature, we developed a model framework addressing labor rights and specifically labor safety in platform economy. As riders play the roles linking all other players on the platform, the labor-right protecting model has to comprehensively model interactions between consumers, restaurants, riders, and the platform.
\item \textbf{Optimal dispatch algorithm for labor rights protection}. Based on the above framework, we propose an efficient dispatch algorithm and rigorously prove that it minimizes the time waste of riders with satisfactory user experience.
\item \textbf{Value of restaurant data on labor rights protection.} Finally, our analysis discovers the sharp value of the private information of restaurants in improving waiting time of riders. Such a result reveals a new perspective on the role of restaurants in protecting labor welfare.
\end{itemize}
\section{Model setting}
In this section, we provide the detailed setting of an online meal delivery platform. Particularly, the environment enrolls four elements: a platform, a restaurant, arrivals of food orders and delivery riders.
\subsection{Model of a Meal-Delivery Process}
In online meal delivery service, an order needs to go through the following process until its corresponding customer can receive the food. First, immediately after being placed online (via apps or websites), the order joins into the processing queue at the restaurant. We assume that the restaurant prepares orders in the processing queue one by one in a first come first serve manner, and it can only prepare one order at one time. After an order is done, the order enters a new waiting queue and waits there until a rider picks it up. In the meantime, the platform will consistently dispatch riders to the restaurant at possibly time-varying rates. In our setting, we assume that one arriving rider will take the earliest prepared order from the waiting queue. If there are no prepared orders, riders will also wait in a queue. When one order is ready, the earliest arriving rider will take this order and leave.
We assume that orders arrive in a Poisson process with rate $1$, and its service time is an exponential distribution of rate $\mu$ independent of the arrival process. The assumption of a Poisson arrival process and exponential service time follows the vast study of restaurant service systems \cite{chen2019food, hwang2010joint, chen2020courier}.
Note that the assumption of a unit-rate arrival process is without loss of generality because we can always normalize the time scale. Denote the food processing queue as $\mathcal{Q}_1$, and the waiting queue of prepared food as $\mathcal{Q}_2$. Then, $\mathcal{Q}_1$ itself is an M/M/1 queue. When one order leaves $\mathcal{Q}_1$, it immediately joins $\mathcal{Q}_2$ to wait for riders. Let $Q_1(t), Q_2(t)$ be the number of orders in $\mathcal{Q}_1, \mathcal{Q}_2$ at time $t$, respectively. When there is no waiting order but $a$ waiting riders, we would write $Q_2(t) = -a.$ In this sense, riders are like negative arrivals that can cancel an order from the queue. Figure \ref{fig:order_procedure} provides an illustration of the whole process.
\begin{figure}[bp]
\centering
\includegraphics[scale=0.25]{figures/food_process.pdf}
\caption{Illustration of Meal Delivery Services}
\label{fig:order_procedure}
\end{figure}
\subsection{Information Disclosure Dispatch Process}
Based on knowledge of the order arrival rate, the platform must design corresponding dispatch algorithms to send riders. In addition, the platform has two sources of information, one from users and the other from the restaurant, both of which may play a role in the dispatch process. Specifically, we divide the decision process into two parts: 1) users' and restaurant's information disclosure scheme and 2) dispatch process of the platform.
\subsubsection{\textbf{Information Disclosure}}
When a user submits an order to the platform, he/she can reveal his/her patience time $T^*$, that is, how much time he/she can endure for extra order delays due to deliveries. Users' patience time is crucial for riders since in practice, riders would get low tips or even be disqualified for late deliveries \cite{doordash}. We assume that all orders will fully disclose the same fixed value $T^*$. The platform has to design algorithms satisfying this constraint. We formalize this decision problem in Section \ref{sec:labor-problem}.
From the view of the restaurant, when $Q_2(t) \geq 0$, its value is only visible to the restaurant, not the platform, since the platform has no information of whether one order is finished or not. Therefore, the restaurant could decide whether to reveal such information to the platform. To model the restaurant's willingness to share its private data with the platform, we design the following information disclosure mechanism. In particular, the restaurant sets a nonnegative integer $M$ as a threshold that is also known by the platform. For every time $t$, the restaurant maintains a signal $\theta(t)$ visible to the platform, given by
\begin{equation}
\theta(t) = \left\{
\begin{aligned}
0, & ~\text{if }Q_2(t) \leq M;\\
Q_2(t)-M, & ~\text{if }Q_2(t) > M.
\end{aligned}
\right.
\end{equation}
We allow $\theta(t) \equiv 0$, and write it as $M = \infty$. Such a screening policy is motivated by the fact that restaurants would like to hide the number of waiting prepared orders when it is small. In this way, the platform may send excessive riders, which can help reduce the waiting time of orders. The threshold $M$ measures how much information is being shared. When $M = 0$, the restaurant shares all of its private information. However, if $M = \infty$, the restaurant would provide no data to the platform.
\subsubsection{\textbf{Dispatch process}} Finally, we define the dispatch process of the platform. The dispatch policy is assumed to be a time-varying Poisson arrival process of riders whose rates change according to $\theta(t)$ provided by the restaurant. Extending the dispatch policy to more general processes may entail more technical challenges, which exceeds the focus of this paper.
In particular, the platform first sets a list of dispatch rate parameters $\vec{\lambda}=(\lambda_0,\lambda_1,\cdots)$ where $\lambda_i \leq \Lambda$ and the upper bound $\Lambda$ measures the maximal availability of riders near the restaurant. After choosing $\vec{\lambda}$, the platform will send riders to the restaurant in a Poisson process whose rate is given by $\lambda_{\theta(t)}.$ That is, conditioned on the event that the restaurant signal is equal to a value $i$, riders will arrive to the restaurant in a Poisson process of rate $\lambda_{i}$. Furthermore, if there are already $d$ riders waiting at the restaurant, i.e., $Q_2(t) = -d$, the platform will not send any riders even when it is going to do so. We call $d$ the buffer length.
For ease of exposition, we assume a rider will arrive at the restaurant immediately when he/she is summoned by the platform. In addition, we assume that the arrival process of riders is independent of the order arrival process and service time of orders.
We note here that $\vec{\lambda}, d$ and $\Lambda$ are tightly connected to two crucial parts in today's meal deliveries: postponed dispatch \cite{ulmer2020restaurant} and delivery scope design \cite{ding2020delivery}. Specifically, in postponed dispatch, the platform will not immediately dispatch a rider to restaurants when an order arrives so that they will not arrive with orders being unprepared. Our arrival rates $\vec{\lambda}$ can thus be seen as how fast the platform will send out a rider. Furthermore, when there are already too many riders ($d$ in our model) waiting for food, the platform won't send any rider. In addition, the platform in general will divide a whole city into multiple regions and deploy different numbers of riders in each area. The number of riders in one region will then decide the magnitude of the parameter $\Lambda$.
To ensure that the platform's capacity is sufficient to serve incoming orders and that the system is stable, we impose the following assumptions on dispatch rates of riders.
\begin{assumption}\label{as:stable}
Rates $\mu, \Lambda, \lambda_0$ are larger than $1$, and $\lambda_i > 0$ for all $i > 0$. In addition, it satisfies that there exists a constant $C<\infty$, such that
$
\sum_{i=1}^{\infty}\prod_{j=1}^i \frac{1}{\lambda_i} = C.
$
\end{assumption}
Under the above assumption, the system state can be the pair of $(Q_1(t), Q_2(t))$, which forms a continuous-time Markov chain. Fig. \ref{fig:transition} illustrates a snapshot of this Markov chain. Furthermore, the system is positively recurrent and enjoys a unique stationary distribution based on Assumption \ref{as:stable} \cite{harchol2013performance}. To measure customer experience and labor rights, this paper focuses on the mean order waiting time $\expect{T_o}$ and the mean rider waiting time $\expect{T_r}$ in the system. To be more specific, the waiting time of an order is measured as the time difference between its arrival at the restaurant and its time fetched by a rider. The mean order waiting time is the time-averaged waiting time among all orders. Similarly, the waiting time of one rider is the time between his/her arrival at the restaurant and the time he/she fetches one order. The mean rider waiting time is then the time-averaged waiting time of all riders.
Note that the two expectations are determined by the platform's setting of $\vec{\lambda}, d, M$. Therefore, when we want to stress such dependence, they are written as $\expectsub{\vec{\lambda}, d, M}{T_o}$ and $ \expectsub{\vec{\lambda}, d, M}{T_r}$.
\begin{figure}
\begin{center}
\includegraphics[width=3.5in]{figures/transition.pdf}
\caption{The Markov chain with state $(q_1,q_2)$}
\label{fig:transition}
\end{center}
\end{figure}
\section{Labor-Right Protecting Dispatch Problem and Analysis}\label{sec:labor-problem}
Based on the mathematical model of a meal delivery platform, we propose the labor-right protecting dispatch problem in this section. The goal is to protect laborers while guaranteeing customer experience. In particular, this section includes two parts. In the first part, we formulate the labor rights-protecting dispatch problem. In the second part, we analyze the characteristics of the model, which enables us to design algorithms for the dispatch problem in Section \ref{sec:algorithm}.
\subsection{Problem Formulation}
To formalize the experience guarantee, we first notice that the expected order waiting time is at least the mean response time in an $M/M/1$ queue of arrival rate $1$ and service rate $\mu$. Then, we know $\expect{T_o}$ is at least $\frac{1}{\mu - 1}$ \cite{harchol2013performance}.
\begin{proposition}
The averaged order waiting time $\expect{T_o}$ is at least $\frac{1}{\mu - 1}$.
\end{proposition}
Based on this lower bound on expected order waiting time, we can formalize the meaning of patient time $T^*$ in the sense that we require $\expect{T_o} \leq \frac{1}{\mu - 1} + T^*$ since $\expect{T_o} - \frac{1}{\mu - 1}$ measures the extra delays brought by deliveries. Then, after the restaurant reveals its threshold $M$, the platform needs to find the best $\vec{\lambda},d$ to minimize the averaged rider waiting time while satisfying requirements of users.
Based on the above discussion, we propose the following optimization framework for the labor-right protecting dispatch problem.
\begin{equation}\label{eq:optimization}
\begin{aligned}
& \underset{\vec{\lambda},d}{\text{minimize}}
& & \expectsub{\vec{\lambda},d,M}{\text{rider waiting time }T_r} \\
& \text{subject to}
& & \expectsub{\vec{\lambda},d,M}{\text{order waiting time }T_o} \\
& & &\leq \frac{1}{\mu - 1} + \text{user patience time }T^* .
\end{aligned}
\end{equation}
Note that our optimization problem also belongs to the vast domain of constrained Markov decision processes (CMDPs) \cite{altman1999constrained}. The goal of CMDP is to obtain the optimal state-dependent policy of a certain objective while satisfying the given constraints. The optimal policy can be complicated and even randomized. However, in the context of our paper, we not only want to solve the optimization problem (\ref{eq:optimization}) but also seek to explicitly understand the trade-off between the objective "rider waiting time" and the constraint "order waiting time" and how the decisions and the trade-off could change in response to the volume of information provided by the restaurant. Therefore, the policy space in our optimization problems is kept as simple as possible to discuss the above information-related decisions and does not involve the general space of randomized policies. We believe a new framework is needed for such discussions and thus leave it as a future study.
\subsection{Model Analysis}
To obtain a deeper understanding of the problem, we first study the stationary distribution of the system in the general setting where the restaurant allows a threshold $M$, and the platform sets the associated dispatch rate $\vec{\lambda}$ and the buffer length $d$. Based on the distribution, we observe a decoupling phenomenon between $\mathcal{Q}_1$ and $\mathcal{Q}_2$. Finally, we restrict our scope to the case of $M=\infty$ and obtain close-form formulas for $\expect{T_o}$ and $\expect{T_r}$.
\subsubsection{\textbf{Stationary Distribution of the General Model}}
Recall that the state of the system is given by $(Q_1(t), Q_2(t))$, namely, the number of orders waiting for preparation and the number of prepared orders waiting for riders. If $Q_2(t) < 0$, it is the negative of waiting riders. Let $(\mathbf{Q}_1,\mathbf{Q}_2)$ be the corresponding random variables distributed with the stationary distribution of $(Q_1(t), Q_2(t))$. Let $\pi_{q_1,q_2}$ be the stationary distribution of the event $(\mathbf{Q_1} = q_1, \mathbf{Q_2} = q_2)$. Fix the threshold $M$, the dispatch rates $\vec{\lambda} = (\lambda_0,\lambda_1,\cdots)$, and the buffer level $d$, and assume they satisfy the assumption \ref{as:stable}. The state space of the system would be $\mathbb{N} \times \left(\{-d, d+1,\cdots, 1\}\cup \mathbb{N}\right)$.
Define $\rho = \frac{1}{\lambda_0}$. The following lemma shows that$\pi_{q_1,q_2}$ exhibits a product-form property similar to the property in queueing networks with negative arrivals \cite{henderson1993queueing}. The proof of this lemma can be found in the appendix.
\begin{lemma}\label{lemma:distribution}
It holds that for $q_1 \geq 0, q_2 \geq -d$,
\[
\pi_{q_1,q_2} = \left\{
\begin{aligned}
&C_1\left(1-\frac{1}{\mu}\right)\left(\frac{1}{\mu}\right)^{q_1}\rho^{q_2},~\text{if }q_2 \leq M; \\
&C_1\left(1-\frac{1}{\mu}\right)\left(\frac{1}{\mu}\right)^{q_1}\rho^{M}\prod_{i=1}^{q_2-M}\frac{1}{\lambda_i},~\text{otherwise,}
\end{aligned}
\right.
\]
where
\begin{equation}
C_1 = \frac{1-\rho}{\rho^{-d}-\rho^{M+1}+(1-\rho)\rho^M\sum_{i=1}^{\infty}\prod_{j=1}^i \frac{1}{\lambda_i}}.
\end{equation}
\end{lemma}
\subsubsection{Queue Decoupling}\label{sec:queue-decoupling}
Observe that by Lemma \ref{lemma:distribution}, we can indeed decouple $\mathcal{Q}_2$ from $\mathcal{Q}_1$. If we consider only the distribution of events $\mathbf{Q}_2 = q$ for $q \geq -d$, the stationary distribution of $\mathcal{Q}_2$ is equivalent to the stationary distribution of the following Markov chain. In the equivalent Markov chain, the set of the state is $\{-d,\cdots,-1\} \cup \mathbb{N}$. For each state $q \geq -d$, it transitions to state $q + 1$ with rate $1$ and to $q - 1$ with rate $\lambda_{\max(q-M,0)}$ if $q > -d$. The stationary distribution of this Markov chain, denoted by $\nu_q$, is exactly given by
\begin{equation}
\nu_q = \left\{
\begin{aligned}
&C_1\rho^{q_2},~\text{if }q_2 \leq M; \\
&C_1\rho^{M}\prod_{i=1}^{q_2-M}\frac{1}{\lambda_i},~\text{otherwise,}
\end{aligned}
\right.
\end{equation}
where $\rho,C_1$ are the same as those in Lemma \ref{lemma:distribution}.
In addition, let $\mathbf{Q}^*$ be the random variable corresponding to the state in this Markov chain. Then, by equivalence, we have the following property connecting $\expect{T_o},\expect{T_r}$ with $\mathbf{Q}^*$. Due to space limit, we omit its proof.
\begin{lemma}\label{lemma:equivalence}
It holds that
\begin{equation}
\begin{aligned}
\expect{T_o} &= \frac{1}{\mu - 1} + \expect{\max(\mathbf{Q}^*,0)} \\
\expect{T_r} &= -\expect{\min(\mathbf{Q}^*,0)}.
\end{aligned}
\end{equation}
\end{lemma}
\subsubsection{\textbf{Analysis for $M = \infty$}}
Restricting our scope to $M = \infty$, based on Lemma \ref{lemma:distribution}, we can obtain expressions of $\expect{T_o}$ and $\expect{T_r}$ via Little's Law\cite{harchol2013performance}. Formally, we have this lemma.
\begin{lemma}\label{lemma:meanValue}
For fixed $\lambda_0 > 1$ and $d$, let $\rho = \frac{1}{\lambda_0}$. It satisfies that
\begin{subequations}
\begin{equation}
\expectsub{\lambda_0,d}{T_o} = \frac{1}{\mu - 1} + \frac{\rho^{d+1}}{1-\rho},
\end{equation}
\begin{equation}
\expectsub{\lambda_0,d}{T_r} = d - \frac{\rho-\rho^{d+1}}{1-\rho}.
\end{equation}
\end{subequations}
\end{lemma}
\section{Labor-right protecting algorithm} \label{sec:algorithm}
Since it is generally nontrivial to solve (\ref{eq:optimization}), in the first part of this section, we start our analysis by assuming that there is no restaurant information, i.e., $M = \infty$. Under this assumption, we propose Algorithm \ref{algo:optimal-parameter} that provides the procedure to find the optimal $\lambda_0, d$. Second, when there are restaurant data, we design Algorithm \ref{algo:policy-improve} to improve policies based on new information.
\subsection{Algorithm and its Optimality when No Restaurant Data}
When there is no restaurant information, the platform needs only to seek the optimal design of $\lambda_0, d$. The closed-form expressions of $\expect{T_o}$ and $\expect{T_r}$ shown in the Lemma \ref{lemma:meanValue} provides the basis of the following algorithm to solve (\ref{eq:optimization}).
\begin{algorithm}[Optimal Dispatch Algorithm]\label{algo:optimal-parameter}
The algorithm takes $T^*$ and $\Lambda$ as inputs and outputs two parameters $\lambda_0, d$ as follows.
\begin{itemize}
\item If $\frac{\Lambda}{1-\Lambda} \leq T$, simply returns $d = 0, \lambda_0 = \Lambda$;
\item Otherwise, for every $\rho \in [0,1)$, define $D(\rho)$ as the smallest nonnegative integer such that $\frac{\rho^{D(\rho)+1}}{1-\rho} \leq T^*$; and for every nonnegative integer $n$, define $\gamma(n)$ as the maximum real number in $(0,1)$ such that $\frac{\gamma(n)^{n+1}}{1-\gamma(n)} \leq T^*$.
\item Let $d_1 = D(\frac{1}{\Lambda}).$ The algorithm returns $d_1, \frac{1}{\gamma(d_1)}$.
\end{itemize}
\end{algorithm}
The next theorem supports the optimality of Algorithm \ref{algo:optimal-parameter}.
\begin{theorem}\label{thm:optimal-algo}
Under the two parameters $\lambda_0, d$ found by Algorithm \ref{algo:optimal-parameter}, and assuming $M = \infty$, it satisfies $\expect{T_o} = \frac{1}{\mu - 1} + T^*$, and $\expect{T_r}$ is minimized.
\end{theorem}
Let's look at a proof sketch of Theorem \ref{thm:optimal-algo}. A formal proof is provided in the appendix. The goal is to minimize $\expect{T_r}$ while satisfying the delay constraint in (\ref{eq:optimization}). By Lemma \ref{lemma:meanValue}, we want to find $\lambda_0, d$ such that $d-\frac{\rho-\rho^{d+1}}{1-\rho}$ is minimized and $\frac{\rho^{d+1}}{1-\rho} \leq T^*$. The intuition of Algorithm \ref{algo:optimal-parameter} is to choose the smallest $\lambda_0$ satisfying the constraint in (\ref{eq:optimization}). Note that $\expect{T_r}$ is a decreasing function of $\rho$ when $d$ is fixed. Therefore, after finding the smallest $d$, we would increase the value of $\rho$ (or equivalently decrease $\lambda_0$) until the constraint in (\ref{eq:optimization}) is violated.
\subsection{Policy Improvement using Restaurant Information}
In this section, we explore how to improve platform policies using restaurant data. Particularly, the restaurant provides more information to the platform by reducing its threshold from $M$ to $m$, where $m < M$. Our results highlight that restaurant data are of particular importance to help the platform make better decisions.
Formally, let us assume that the platform has a policy given by when the threshold is $M$. Then, algorithm \ref{algo:policy-improve} outputs a new policy for a given threshold $m, m < M$ that is arguably better than the original policy. Algorithm \ref{algo:policy-improve} is given below.
\begin{algorithm}[Policy Improvement with Restaurant's Data]\label{algo:policy-improve}
The algorithm takes $\vec{\lambda}=(\lambda_0,\cdots),d,M,m$ and $\Lambda$ as inputs and outputs the dispatch rates and buffer level, $\vec{\tau}=(\tau_0,\cdots), d_1$, of the new policy as follows. Define $\rho = 1/\lambda_0$.
\begin{itemize}
\item Set $C = \frac{\rho-\rho^{M-m+1}}{1-\rho}+\rho^{M-m}\sum_{i=M+1}^{\infty}\prod_{j=1}^{i-M} \frac{1}{\lambda_j}.$
\item Set $\tau_0 = \lambda_0, \tau_1 = \frac{1}{(1-\frac{1}{\Lambda})C}$, and $\tau_i = \Lambda$ for all $i > 1$.
\item Let $d_1 = d$. Output $(\vec{\tau}, d_1)$.
\end{itemize}
\end{algorithm}
Our next result shows that using Algorithm \ref{algo:policy-improve}, the platform can always find a policy with better averaged order waiting time and the same averaged rider waiting time based on any policy of a larger threshold.
\begin{theorem}\label{thm:data-value}
Suppose that the platform has a policy with parameters $\vec{\lambda}=(\lambda_0,\cdots),d$ and the restaurant's threshold is $M$, $M > 0$. If $M$ reduces to $m$, $m < M$, then using Algorithm \ref{algo:policy-improve}, the platform can construct a new set of dispatch rates $\vec{\tau}$, such that
\begin{equation}\label{eq:policy-improve}
\expectsub{\vec{\tau},d,m}{T_o} \leq \expectsub{\vec{\lambda},d,M}{T_o}, ~\expectsub{\vec{\tau},d,m}{T_r} = \expectsub{\vec{\lambda},d,M}{T_r},
\end{equation}
where the inequality becomes equal only if
\[
\rho=\frac{\Lambda-1}{\Lambda}\left(\frac{\rho-\rho^{M-m+1}}{1-\rho}+\rho^{M-m}\sum_{i=M+1}^{\infty}\prod_{j=1}^{i-M} \frac{1}{\lambda_j}\right).
\]
\end{theorem}
\begin{remark}
Note that if $M = \infty$ in Theorem \ref{thm:data-value}, the inequality in (\ref{eq:policy-improve}) becomes equal if and only if $\lambda_0 = \Lambda$, that is, the platform is already dispatching riders at its full capacity.
\end{remark}
The proof of Theorem \ref{thm:data-value} can be found in the appendix.
\section{Obligation of different stakeholders}
In this section, we discuss obligations in protecting the labor rights of the three stakeholders: the platform, users, and the restaurant.
First, for restaurant information, Theorem \ref{thm:data-value} tells us that whenever there are data from the restaurant side, the platform can improve the experience of customers while maintaining the same averaged rider waiting time. In this sense, restaurant data help to improve labor rights by making the labor more efficient. We will further quantify the effect of queueing data of the restaurant in Section \ref{sec:simulation}.
Second, we desire to understand how the customer's requirement of experience and the platform's ability to recruit riders may affect the averaged rider waiting time. For ease of exposition, we assume that $M = \infty$ in this part. Furthermore, suppose that the platform is still trying to solve the labor-right protecting dispatch problem introduced in (\ref{eq:optimization}).
In this case, we show that there is a fundamental lower bound of $\expect{T_r}$ given that $T^*$ is low. The proof is deferred to the appendix.
\begin{theorem}[Dilemma between customers and riders]\label{thm:tradeoff}
Suppose that $\hat{\rho}\colon=\frac{1}{\Lambda} \geq 0.3$ and $\frac{\hat{\rho}^4}{1-\hat{\rho}} \geq T^*$ and $M = \infty$. Then, it holds
$
\expect{T_r} \geq \frac{\ln(4/3)}{4}\frac{1}{1-\hat{\rho}}-T^*,
$
for all feasible $d$ and $\lambda_0$ satisfying the constraint in (\ref{eq:optimization}).
\end{theorem}
This result shows that the sum of $T^*$ and $\expect{T_r}$ is lower bounded by a value determined by the platform capacity $\Lambda$. As a result, when $\Lambda$ is fixed, no matter how the platform acts, it is impossible for both $T^*$ and $\expect{T_r}$ to become negligible simultaneously. Nevertheless, the lower bound decreases when $T^*$ or $\Lambda$ increases, thus implying that if customers allow a longer delivery time or if the platform is able to recruit more riders, then laborers can enjoy a lower average waiting time.
\section{Numerical Case} \label{sec:simulation}
In this section, we aim to illustrate the importance of the labor rights protection problem, the dilemma between customer experience and rider waiting time, and the value of restaurant information using simulations. Particularly, we assume that $\lambda_0 = 1$, and $\mu = \Lambda = 1.5$. It is a reasonable assumption because in practice, the restaurant and the platform usually only recruit nearly sufficient numbers of riders to reduce cost.
\subsection{Importance of labor rights protection and Trade-offs}
To study the impact of the labor rights protection problem, we consider three settings of patient time, $T^*=0.01, 0.05, 0.1$. In addition, we assume that there is no restaurant information, so $M = \infty$. Then, for $\lambda_0 \in (1,\Lambda)$, we plot the curve of $\expectsub{\lambda_0,d}{T_r}$, where $d$ is the minimum integer such that $\expectsub{\lambda_0,d}{T_r} - \frac{1}{\mu - 1} \leq T^*$. The result is shown in Fig. \ref{fig:right-protect}.
\begin{figure}[!hbtp]
\centering
\includegraphics[scale=0.3]{deliver-time.jpg}
\caption{Averaged Rider waiting time under Different Dispatch Rates Given that Customer Experience is satisfied}
\label{fig:right-protect}
\end{figure}
As we can see, for a fixed $T^*$, although all the patience time is satisfied, the averaged rider waiting time varies greatly. A clever setting of $\lambda_0$ can reduce the waiting time by a factor of $4$. In addition, when $T^*$ increases, the averaged rider waiting time decreases. It characterizes the trade-off between customer experience and labor rights.
\subsection{Value of Restaurant's Data}
To highlight the value of the data from the restaurant, we assume that $M$ could be $0,10,\infty$. In this case, we are able to differentiate different levels of restaurant information. For the platform's policy, we fix $d = 0$ and then allow $\lambda_0$ to vary from $1$ to $\Lambda$ to be the policy for $M = \infty$. Then, for other $M$, we employ Algorithm \ref{algo:policy-improve} to generate new policies. We plot the value of $\expectsub{\vec{\lambda},d,M}{T_o}$ after the policy improvement as shown in Fig. \ref{fig:policy-improve}.
\begin{figure}[!hbtp]
\centering
\includegraphics[scale=0.3]{order-time.jpg}
\caption{Averaged Order Waiting Time After Using Restaurant Data}
\label{fig:policy-improve}
\end{figure}
As we can tell from Fig. \ref{fig:policy-improve}, restaurant data and algorithm \ref{algo:policy-improve} can indeed greatly improve the customer experience, decreasing $\expect{T_o}$ by a factor of two when $M = 0$ and $\lambda_0 = 1.2$. Specifically, we can observe that the improvement is greater when $\lambda_0$ or $M$ is smaller.
\section{CONCLUSIONS}
In this paper, we formulated a queueing model for meal delivery platforms that incorporated four key stakeholders: the platform, the restaurant, riders, and customers. In particular, we proposed the labor rights protecting problem, whose goal is to reduce waiting time of riders while maintaining satisfactory customer experience. When there is no restaurant information, we designed an efficient algorithm to find the optimal dispatch parameters for the platform. When the restaurant is willing to share its private data, we further characterized a policy improvement method that takes any policy as an input and returns a refined policy based on restaurant data. The refined policy provably improves customer experience, while labor rights did not get hurt. Our results also indicated that there was a trade-off between customer experience and labor rights in the sense that they cannot be satisfied simultaneously. Our simulations confirmed the theoretical results. Specifically, they revealed that an optimal labor-rights protecting algorithm could dramatically reduce the waiting time of riders, and restaurant data could considerably reduce the order waiting time, especially when the restaurant would like to share more information.
\bibliographystyle{IEEEtran}
\subsection{Proof of Lemma \ref{lemma:distribution}}
\begin{proof}
Notice that $\sum_{q_1=0}^{\infty}\sum_{q_2=-d}^{\infty}\pi_{q_1,q_2} = 1$. We can then prove the result by checking that $\pi_{q_1,q_2}$ satisfies the balance equation of the Markov chain because the steady state distribution is the unique solution to the balance equation given that it is normalized\cite{harchol2013performance}. For any $x$, denote $x^+=\max(0,x)$. In the considered Markov chain, for a state $(q_1,q_2)$, its transitions are given as follows:
\[
(q_1,q_2) \rightarrow \left \{
\begin{aligned}
(q_1+1,q_2) &,~\text{of rate }1; \\
(q_1-1,q_2+1) &,~\text{of rate }\mu~\text{if }q_1 > 0; \\
(q_1,q_2-1) &,~\text{of rate }\lambda_{(q_2-M)^+}~\text{if }q_2 > -d.
\end{aligned}
\right.
\]
For any two state $Q,P$, denote the rate from $Q$ to $P$ by $r_{Q,P}$. Fix a state $Q=(q_1,q_2)$. Here we only check the balance equation for state $(q_1,q_2)$ satisfying $q_1 > 0,q_2 > -d$. The same strategy can verify other cases similarly. The balance equation of this state is that
\[
\text{inflow}\colon = \sum_{P} \pi_{P}r_{P,Q} = \sum_{P} \pi_Q r_{Q,P} =\colon \text{outflow}.
\]
We need to consider three cases. First, suppose that $q_2 < M$. Then by the form of $\pi$, it holds
\[
\begin{aligned}
\text{inflow}&=\pi_{q_1-1,q_2}+\pi_{q_1,q_2+1}\lambda_0+\pi_{q_1+1,q_2-1}\mu \\
&= C\left(\frac{1}{\mu^{q_1-1}\lambda_0^{q_2}}+\frac{1}{\mu^{q_1}\lambda_0^{q_2}}+\frac{1}{\mu^{q_1}\lambda_0^{q_2-1}}\right) \\
&= \pi_{q_1,q_2}\left(1+\lambda_0+\mu\right)=\text{outflow}.
\end{aligned}
\]
Then if $q_2 = M$, it satisfies that
\[
\begin{aligned}
\text{inflow}&=\pi_{q_1-1,q_2}+\pi_{q_1,q_2+1}\lambda_1+\pi_{q_1+1,q_2-1}\mu \\
&= C\left(\frac{1}{\mu^{q_1-1}\lambda_0^{q_2}}+\frac{1}{\mu^{q_1}\lambda_0^{q_2}}+\frac{1}{\mu^{q_1}\lambda_0^{q_2-1}}\right) \\
&= \pi_{q_1,q_2}\left(1+\lambda_0+\mu\right)=\text{outflow}.
\end{aligned}
\]
Finally, for $q_2 > M$,
\[
\begin{aligned}
&\mspace{23mu}\text{inflow} \\
&=\pi_{q_1-1,q_2}+\pi_{q_1,q_2+1}\lambda_{q_2+1-M}+\pi_{q_1+1,q_2-1}\mu \\
&= C\left(\frac{1}{\mu^{q_1-1}\lambda_0^{q_2}}+\frac{1}{\mu^{q_1}\lambda_0^{M}}\left(\prod_{i=1}^{q_2-M}\frac{1}{\lambda_i}+\prod_{i=1}^{q_2-M-1}\frac{1}{\lambda_i}\right)\right) \\
&= \pi_{q_1,q_2}\left(1+\lambda_{q_2-M}+\mu\right)=\text{outflow},
\end{aligned}
\]
which completes the proof.
\end{proof}
\subsection{Proof of Lemma \ref{lemma:meanValue}}
\begin{proof}
Define $\expect{N_o}$ be the expected number of unsent orders. By Little's Law, $\expect{T_o} = \expect{N_o}$ because orders' arrival rate is $1$. It holds that
\[
\expect{N_o} = \sum_{q_1=0}^{\infty}\sum_{q_2=0}^{\infty} (q_1+q_2)\pi_{q_1,q_2}.
\]
Simplifying Lemma \ref{lemma:distribution} by setting $M = \infty$. We have $\pi_{q_1,q_2} = (1-\rho)\rho^d(1-\frac{1}{\mu})\frac{1}{\mu^q_1}\rho^{q_2}$ for $q_1 \geq 0,q_2 \geq -d$.
Then
\[
\begin{aligned}
\expect{N_o} &= \sum_{q_1,q_2=0}^{\infty} q_1 \pi_{q_1,q_2} + \sum_{q_1,q_2=0}^{\infty} q_2 \pi_{q_1,q_2} \\
&= \frac{1}{\mu - 1} + \sum_{q_2=0}^{\infty} (1-\rho)\rho^{d+j}j\\
&= \frac{1}{\mu - 1} + \frac{\rho^{d+1}}{1-\rho}.
\end{aligned}
\]
On the other hand, define $\expect{N_r}$ be the expected waiting riders. Again, by Little's Law, $\expect{N_r} = \hat{\lambda}\expect{T_r}$, where $\hat{\lambda}$ is the "true" dispatch rate of riders. That is,
\[
\hat{\lambda} = \lambda_0\left(1 - \sum_{q_1=0}^{\infty}\pi_{q_1,-d}\right)=\lambda_0\rho = 1.
\]
Then for $\expect{N_r}$, it holds
\[
\begin{aligned}
\expect{N_r} = \sum_{q_1=0}^{\infty}\sum_{q_2=-d}^{-1} -q_2 \pi_{q_1,q_2} &= (1-\rho)\rho^d\sum_{j=1}^d j\rho^j \\
&= d - \frac{1 - \rho^d}{\lambda_0 - 1},
\end{aligned}
\]
and thus $\expect{T_r} = d - \frac{1 - \rho^d}{\lambda_0 - 1}$.
\end{proof}
\subsection{Formal Proof of Theorem \ref{thm:optimal-algo}}
We first have the following lemma demonstrating the optimality of choosing smallest feasible $d$.
Recall the function $D(\rho)$ and the function $\gamma(n)$ from Algorithm \ref{algo:optimal-parameter} defined for a fixed $T^*$. Let $d_l =max(1,D(\frac{1}{\Lambda}))$. For every $d \geq d_l$, define $h(d) = \expectsub{\frac{1}{\gamma(d)},d}{T_r}$.
\begin{lemma}\label{lemma:monotonicity}
It holds that $h(d)$ is a strictly increasing function when $d \geq d_l$.
\end{lemma}
\begin{proof}
Let $\rho_l$ be the smallest real number such that $\frac{\rho_l^{2}}{1-\rho_l} \geq T^*$. And if $\rho_l < \frac{1}{\Lambda}$, set it to be $\frac{1}{\Lambda}$. We can then see that $d_l = D(\rho_l)$. Now for every $\rho \geq \rho_l$, define $x(\rho)$ such that $\frac{\rho^{x(\rho)+1}}{1-\rho} = T^*$. Then $x(\rho)+1 = \frac{\ln T^* + \ln(1-\rho)}{\ln \rho} \geq 1$. Note that $x(\rho)$ may not be integer.
We claim that for every $d \geq d_l$, it holds $x(\gamma(d)) = d$. That is, $x$ is like a reverse function of $\gamma$, but it extends the definitional domain. We prove this claim by contradiction. Suppose it is not. It must hold that $x(\gamma(d)) < d$, and $\frac{(\gamma(d))^{d+1}}{1-\gamma(d)} < T^*$. However, in this case, we can find a $\rho > \gamma(d)$ such that $\frac{\rho^{d+1}}{1-\rho)} \leq T^*$, which contradicts the fact that $\gamma(d)$ is the largest such possible $\rho$. As a result, we must have $x(\gamma(d))=d$.
For $\rho \geq \rho_l$, define $g(\rho)\colon= x(\rho)-\frac{\rho - \rho^{x(\rho)+1}}{1-\rho}$. Then to prove that $h(d)$ is strictly increasing, it is sufficient to prove that $g(\rho)$ is strictly increasing since $h(d) = g(\gamma(d))$. Furthermore, observe that for $\rho \geq \rho_l$, $\frac{\rho^{x(\rho)+1}}{1-\rho} = T^*$, which is a constant. It is enough to prove the function $G(\rho)\colon=\frac{\ln T^* + \ln(1-\rho)}{\ln \rho}-\frac{\rho}{1-\rho}-1$ is strictly increasing.
To prove it, take derivative of $G$, and we can get
\begin{equation}
\begin{aligned}
G'(\rho) &= \frac{-1}{(1-\rho)\ln \rho}-\frac{\ln T^*+\ln(1-\rho)}{\rho(\ln \rho)^2}-\frac{1}{(1-\rho)^2} \\
&=\frac{-1}{(1-\rho)\ln \rho}-\frac{(x(\rho)+1)\ln \rho}{\rho(\ln \rho)^2}-\frac{1}{(1-\rho)^2} \\
&\geq \frac{-1}{(1-\rho)\ln \rho}-\frac{1}{\rho\ln \rho}-\frac{1}{(1-\rho)^2},
\end{aligned}
\end{equation}
where the last inequality is because $x(\rho)+1\geq 1$ and $\ln(\rho) < 0.$
It remains to show that $G_1(\rho)=\frac{-1}{(1-\rho)\ln \rho}-\frac{1}{\rho\ln \rho}-\frac{1}{(1-\rho)^2}$ is positive in $(0,1)$. Simplifying the terms gives $G_1(\rho)=\frac{-\rho+\rho\ln \rho + 1}{-(1-\rho)^2\rho \ln \rho}$. For the function $G_2(\rho)=-\rho + \rho \ln \rho + 1$, its derivative is $\ln \rho$, which is negative in $(0,1)$. Therefore, the function $G_2(\rho)$ is a strictly decreasing function in $(0,1)$. Since $G_2(1) = 0$, we know both $G_2(\rho)$ is always positive in $(0,1)$, and so is $G_1(\rho)$. Consequently, $G(\rho)$ is a strictly increasing function, which completes the proof.
\end{proof}
We can now finish the proof of Theorem \ref{thm:optimal-algo}.
\begin{proof}
In Algorithm \ref{algo:optimal-parameter}, we first check whether $\frac{\Lambda}{1-\Lambda} \leq T^*$. If it is, it means it is feasible to set $d = 0$, and $\lambda_0 = \Lambda$. It is optimal because $\expectsub{\Lambda,0}{T_r} = 0$.
Otherwise, we have to set $d \geq 1$. Notice that when $d$ is fixed, $\expectsub{\Lambda,0}{T_o}$ is a decreasing function, while $\expectsub{\Lambda,0}{T_r}$ is an increasing function. Therefore, suppose we want to select $d$ one by one. We can first check whether the pair of $(\lambda_0 = \Lambda,d)$ can satisfy the constraint in (\ref{eq:optimization}). If it does, then we decrease the value of $\lambda_0$ to obtain lower averaged rider waiting time. The above discussion claims that feasible $d$ must be at least $D(\frac{1}{\Lambda})$. And for each feasible $d$, the optimal choice is to choose $\lambda_0 = \frac{1}{\gamma(d)}$. The averaged rider waiting time given by this parameter setting is exactly $h(d)$. Now according to Lemma \ref{lemma:monotonicity}, $h(d)$ is an increasing function. Therefore, the optimal solution is to set $d = D(\frac{1}{\Lambda})$, and $\lambda_0 = \frac{1}{\gamma(d)}$.
\end{proof}
\subsection{Proof of Theorem \ref{thm:data-value}}
We now provide the proof of Theorem \ref{thm:data-value}. The intuition is that,
by Little's Law, it suffices to prove that after using the new policy $(\vec{\tau},d,m)$, the mean number of waiting riders remains the same, but the mean number of waiting orders decreases.
Then the proof starts by coupling the Markov chain of the threshold $m$ and that of the threshold $M$, $M > m$. After that, we propose an optimization problem to minimize the mean number of waiting orders, and use a greedy policy to solve it. Algorithm \ref{algo:policy-improve} corresponds to the final outcome of this policy, and thus makes an improvement.
\subsubsection{System Coupling}
To enable analysis, recall the discussion in Section \ref{sec:queue-decoupling}, we only need to consider the equivalent Markov chain of $\mathcal{Q}_2$ to calculate the mean numbers of waiting orders and riders. Denote the new Markov chain of $(\vec{\lambda},d,M)$ as $A$, and that of $(\vec{\tau},d,m)$ as $B$. Denote their steady-state distributions by $\pi_A(q), \pi_B(q)$. We can now couple the two Markov chains on state $m$. Recall that $\rho = \frac{1}{\lambda_0}$. By solving balance equations with state $m$ as the initial start, we could notice that
\[
\pi_A(m)\left(\sum_{i=-d}^{M} \rho^{i-m}+\rho^{M-m}\sum_{i=M+1}^{\infty}\prod_{j=1}^{i-M} \frac{1}{\lambda_j}\right) = 1,
\]
and
\[
\pi_B(m)\left(\sum_{i=-d}^{m} \rho^{i-m} + \tau_1\sum_{i=m+1}^{\infty} \left(\frac{1}{\Lambda}\right)^{i-m-1}\right)=1.
\]
Then by the definition of $\tau_1$, it holds that $\pi_A(m) = \pi_B(m)$. As a result, by Lemma \ref{lemma:equivalence}, it leads to the fact that
\[
\expectsub{\vec{\tau},d,m}{T_r} =\sum_{i=-d}^{-1}(-i)\rho^{i-m}\pi_A(m)= \expectsub{\vec{\lambda},d,M}{T_r}.
\]
\subsubsection{Optimization Framework and Greedy Improvement}
It remains to show that the mean number of waiting orders decreases. For Markov chain $A$, define $a_q = \frac{\pi_A(q)}{\pi_A(m)}$ for $q > 0$. Similarly, define $b_q$ for Markov chain $B$. Indeed, $a_q = \frac{a_{q-1}}{\lambda_{\max(q-M,0)}}$. Then by Lemma \ref{lemma:equivalence}, we have
\[
\expectsub{\vec{\lambda},d,M}{T_o} - \expectsub{\vec{\tau},d,m}{T_o} = \pi_A(m)\sum_{q=1}^{\infty} q\left(a_q - b_q\right).
\]
Furthermore, we can observe that by the definition of $\tau_1$, it holds $\sum_{q=1}^{\infty} a_q = \sum_{q=1}^{\infty} b_q < \infty$. Denote this constant by $H$. To show that $\expectsub{\vec{\lambda},d,M}{T_o} \geq \expectsub{\vec{\tau},d,m}{T_o}$, it suffices to prove $\{b_q\}$ is the solution to the following optimization problem:
\begin{equation}\label{eq:optimize-delay}
\begin{aligned}
& \underset{\vec{x}=x_1,\cdots}{\text{minimize}}
& & \sum_{i=1}^{\infty} i\cdot x_i \\
& \text{subject to}
& & \sum_{i=1}^{\infty} x_i = H \\
& & & \frac{x_i}{x_{i - 1}} \geq \frac{1}{\Lambda}, \forall i > 1 \\
& & & x_1 \geq \frac{1}{\Lambda}.
\end{aligned}
\end{equation}
The requirement of $\frac{x_i}{x_{i - 1}} \geq \frac{1}{\Lambda}$ comes from the fact that the dispatch rate cannot exceed $\Lambda$.
Greedily, given a feasible setting of $\vec{x}$ in \ref{eq:optimize-delay}, the best strategy is to keep $x_i$ large for small $i$, and keep $x_i$ small for large $i$. Particularly, we claim that the solution must satisfy $\frac{x_{i+1}}{x_i} = \frac{1}{\Lambda}$ for all $i \geq 1$. Otherwise, we can find the smallest $i$ with $\frac{x_{i+1}}{x_i} > \frac{1}{\Lambda}$. It always improves the objective by decreasing $x_{i+1}$ and increasing $x_i$. Then we can solve out $x_1$, and it turns out that the produced solution is indeed $\{b_q\}$, which completes the proof.
\hspace*{\fill}~\QED\par\endtrivlist\unskip
\subsection{Proof of Theorem \ref{thm:tradeoff}}
Fix $\lambda_0$ and $d$ satisfying the constraint in (\ref{eq:optimization}). By the assumption in Theorem \ref{thm:tradeoff}, it holds that $\frac{1}{\lambda_0} \geq 0.3$ and $d \geq 3$. Then to show that there is a trade-off between customer experience and labor right, one key observation is that, by Lemma \ref{lemma:meanValue},
\begin{equation}
\expect{T_r} + \expect{T_o}-\frac{1}{\mu-1} = d + \frac{2\rho^{d+1}-\rho}{1-\rho},
\end{equation}
where $\rho = \frac{1}{\lambda_0}$. Let $f(\rho) = \frac{2\rho^{d+1}-\rho}{1-\rho}$. We then consider two cases.
First, when $\rho^{d+1} \geq \frac{3}{4}$, it holds $f(\rho) \geq \frac{1}{2(1-\rho)}$ because $\rho < 1$. Therefore, $\expect{T_r} + \expect{T_o}-\frac{1}{\mu-1} \geq \frac{1}{2(1-\rho)}.$
Second, when $\rho^{d+1} < \frac{3}{4}$, we claim that $\expect{T_r} + \expect{T_o}-\frac{1}{\mu-1} \geq \frac{d}{2}$. To see why it is true, we only need to show that $f(\rho)=\rho \frac{2\rho^d-1}{1-\rho} \geq -\frac{\rho d}{2}$. For $x \in [0,1)$, Define $u(x) = \frac{2x-1}{1-x^{1/d}}.$ We claim that $u(x) \geq -\frac{d}{2}$.
\begin{proof}
If not, then it holds that
\[
\frac{2x-1}{1-x^{1/d}} \leq -\frac{d}{2} \Longleftrightarrow 0 \leq -\frac{d}{2}+\frac{d}{2}x^{1/d}-2x+1.
\]
Define $v(x) = -\frac{d}{2}+\frac{d}{2}x^{1/d}-2x+1$. Take derivative of $v(x)$, we know $v'(x) = \frac{1}{2}x^{1/d-1}-2$. Since $d \geq 3$, $v'(x)$ is a decreasing function. In addition, $\lim_{x \to 0^+} v'(x) > 0$, and $v'(1) < 0$. Therefore, the value $\bar{x}$ such that $v'(\bar{x}) = 0$ is the maximum point of $v(x)$. Set $v'(\bar{x}) = 0$. It holds $\bar{x} = 4^{\frac{-d}{d-1}}$. Therefore, the maximum value of $h(x)$ for $x \in (0,1)$ would be $h(\bar{x}) = -\frac{d}{2}+\frac{d}{2}4^{\frac{1}{d-1}}-2\bar{x}+1 \leq -2\cdot 4^{-1.5}+1<0$. It thus contradicts the assumption that there is some $x$ satisfying $h(x) \geq 0$, which completes the proof.
\end{proof}
As a result, it holds $\expect{T_r}+\expect{T_o}-\frac{1}{\mu - 1} \geq \frac{d}{2} \geq \frac{\ln(4/3)}{-2\ln \rho}$ since $\rho^d \leq \frac{3}{4}$. Furthermore, by assumption, $\rho \geq 0.3$, meaning that $-\ln \rho \leq 2(1-\rho)$. Therefore, summarizing the above discussions, we have
\begin{equation}
\expect{T_r} \geq \frac{\ln(4/3)}{4(1-\rho)}-\left(\expect{T_o}-\frac{1}{\mu-1})\right) \geq \frac{\ln(4/3)}{4(1-\hat{\rho})}-T^*,
\end{equation}
where $\hat{\rho} = \frac{1}{\Lambda}$.\hspace*{\fill}~\QED\par\endtrivlist\unskip
\section{INTRODUCTION}
\subsection{Background}
Labor rights in meal delivery have become one of the top public concerns in platform economy. During the last decade, the meal-delivery industry has been flourishing due to the penetration of smart mobile phones. However, a growing number of traffic violations and accidents due to risky driving of riders have attracted people to discuss the relationship between platform algorithms and labor rights. Current statistics reveal that meal delivery riders contribute to a significant share of traffic violations. In 2017, approximately one food delivery rider was reported to be injured or killed every 2.5 days in Shanghai, China\cite{ChinaNews}. The emerging food-delivery platform also contributes to a nonnegligible fraction of traffic accidents in Korea \cite{byun2017characteristics}. In the U.S., food delivery riders are viewed as one of the most dangerous jobs\cite{DangerousJob}.
It is necessary to develop and analyze dispatch algorithms of meal-delivery platforms from the labor-right protecting perspective. However, the current algorithmic models for meal delivery are either user-centric or platform profit-driven and ignore the perspective of protecting riders from traffic risks. The absence of the perspective of labor rights of riders causes two problems. First, the current dispatch models simplify the procedure of the working flow of riders by merging riders’ picking-up order period and delivery period into one period. This oversimplification causes the current models to ignore the potential of improving rider traffic safety by managing the rider waiting time during the period of picking up orders. Consequently, user-centric or profit-driven dispatch results in overwaiting riders. Second, the user-centric dispatch focuses mainly on timely delivery for attracting consumers while profit-driven dispatch pursues minimizing the population of riders for cost saving\cite{DriverRisk, christie2019health}. Consequently, the current dispatch algorithms limit the number of riders and continuously shorten the time window of deliveries to attract more consumers by quickly delivering foods. Both mechanisms, user-centric and profit-driven dispatch, contribute to overload of riders. The above two problems of the current food-delivery models together drive riders to overspeed.
In this research, we propose a queuing-model-based framework of designing dispatch policies from the perspective of protecting the traffic safty of riders in meal delivery platforms. In contrast to user-centric dispatch or profit-driven dispatch, in this research, we develop a framework that carefully models the working flow of riders and seeks to mitigate the stress of the delivery-time limit that forces riders to drive at excessive speeds. Our rider-centric framework enables us to analyze the pathway of protecting rider safety. According to our framework, we develop a labor-rights-protection dispatch algorithm to control the waiting time of riders in restaurants. Our algorithm can achieve the optimal dispatch strategy that minimizes the waiting time of riders with guaranteed consumer-acceptable order waiting time. We also apply our framework to explore how restaurant information can further improve the labor rights of riders with a refined platform dispatch policy.
Our rider-centric framework further allows us to comprehensively capture the impacts from all stakeholders, including restaurants, consumers, platforms, and riders themselves. We found that information about the number of restaurant offline orders is critical for managing rider waiting time for orders. Therefore, our framework decomposes the meal-delivery task into two processes: an information disclosure process and a dispatch process. The information disclosure process captures the impact of restaurants sharing their information, while the dispatch process captures the impact of the patience of consumers and decisions of platforms. We found that restaurants can help with labor rights protection by sharing their private information about the number of their offline orders. This conclusion reveals the tight connection between private information and labor rights in platform economy.
\subsection{Related work}
The growth of meal delivery platforms opens a wide range of research areas. Due to the unique process structure of meal deliveries, queuing models are frequently used to understand profit brought by meal delivery platforms, or broadly on-demand platforms \cite{bai2019coordinating,taylor2018demand,feldman2019can,chen2019food,chen2020courier}. The current user-centric or profit-driven dispatch adopts mainly a single-sided queueing model focusing on timely service of the randomly arriving orders by the least size of rider population. In particular, \cite{feldman2019can} studies the benefit of delivery platforms to restaurants by analyzing an associated queuing model and proposing new contracts. Then, in \cite{chen2019food}, the authors propose a queuing model with a pricing scheme to study interactions between platforms and customers. Their model also discusses the impact of a finite rider pool on the profit of a platform. In addition, the model of \cite{chen2020courier} analyzes the benefit of order batching via queues with batch arrivals.
From a technical perspective, our queueing model belongs to the class of two-sided queues, whose study can be dated back to the 1990s \cite{gelenbe1991queues}. In the original model, there are arrivals of positive jobs to one single queue. There are also negative jobs that can cancel one positive job waiting in the queue. However, when there is no positive job, the negative job will directly leave. The sojourn time of positive jobs has been derived \cite{harrison1993sojourn}, and the model has also been extended to more general service time \cite{harrison1996m} and arrival processes \cite{li2004map}. The stationary distribution is known to have a product form in queueing networks with negative arrivals \cite{henderson1993queueing, gelenbe1991product}. Our model, however, is different from previous studies in the sense that negative arrivals, i.e., riders, can wait in another queue for future orders. The arrival process of riders can depend on the queue length of orders instead of a time-homogeneous process.
In addition to queuing models, there are optimization and reinforcement learning studies aiming at developing online optimization strategies for efficient real-time meal deliveries. A stream of work models rider dispatch problems by optimizations with the objective of maximizing the number of timely orders or minimizing the averaged delays \cite{yildiz2019provably, reyes2018meal,joshi2020batching}. In addition, studies of meal deliveries when a customer's order can include food from multiple restaurants is studied in \cite{steever2019dynamic}. Randomness in meal delivery problems is considered in the optimization framework in \cite{ulmer2020restaurant,liu2020time}. The model in \cite{liu2020integrating} uses deep inverse reinforcement learning to learn a better dispatch policy. In addition to providing efficient dispatch, customer satisfaction is another core concern of current dispatch studies. Some recent studies discussed reliable estimation of arrival time by using supervised learning on deliveries \cite{zhu2020order,hildebrandt2020supervised}. There also exist works pursuing improving customer's satisfaction by designing reasonable range of customers to whom each restaurant can expose \cite{ding2020delivery}.
Although the labor rights of riders have attracted sufficient social attention, there are only limited studies on the topics, most of which focus on the impacts of wage design on welfare distribution. For instance, \cite{benjaafar2020labor,wu2020two} examine how wages in platforms impact customers and laborers but have not discussed how the process of dispatching may impact labor welfare. To the best of our knowledge, there is still a lack of systematic discussion about improving the dispatch for mitigating the time-limit stress on riders, which forces riders to drive riskily during delivery orders. We also found that discussion about the impacts of information on labor rights is rare. \cite{mao2019faster,liu2020time} integrate the estimation of travel times, attitudes of customers toward late orders and familiarity of riders with roads into optimization models and show by simulations that data can improve dispatch efficiency. Nevertheless, to the best of our knowledge, there is yet no work discussing the impact of restaurant data and providing theoretical insights into how restaurant data can help improve dispatch efficiency and labor welfare.
\subsection{Labor rights of riders in food-delivery platforms}
Labor rights can be divided into two categories: 1) working benefits such as minimum wages and reasonable working hours; 2) the safety of working conditions\cite{laborright}. Most current relative discussions in the context of platform economy focus on the first economic aspect of labor rights \cite{de2015rise, benjaafar2020labor, parrott2018earnings}, while the working safety of riders is absent from the current literature about platform economy.
Here, our work provides a first step to consider the driving safety of riders in platform economy by minimizing the waiting time of riders. We focus on the waiting time of riders in restaurants for two reasons. First, minimizing the waste of time during the order-picking-up process allows riders to have sufficient time to deliver the order. Second, in contrast to user-centric or profit-driven dispatch, minimizing the rider's waiting time in restaurants can prevent the algorithm from continuously shortening the available time window of deliveries. We believe that our work could offer a new angle for the control community to think about platform economy.
\subsection{Contributions}
We summarize our three contributions below.
\begin{itemize}
\item \textbf{Framework of modeling labor-right protecting problems in meal-delivery platforms.} In contrast to the literature, we developed a model framework addressing labor rights and specifically labor safety in platform economy. As riders play the roles linking all other players on the platform, the labor-right protecting model has to comprehensively model interactions between consumers, restaurants, riders, and the platform.
\item \textbf{Optimal dispatch algorithm for labor rights protection}. Based on the above framework, we propose an efficient dispatch algorithm and rigorously prove that it minimizes the time waste of riders with satisfactory user experience.
\item \textbf{Value of restaurant data on labor rights protection.} Finally, our analysis discovers the sharp value of the private information of restaurants in improving waiting time of riders. Such a result reveals a new perspective on the role of restaurants in protecting labor welfare.
\end{itemize}
\section{Model setting}
In this section, we provide the detailed setting of an online meal delivery platform. Particularly, the environment enrolls four elements: a platform, a restaurant, arrivals of food orders and delivery riders.
\subsection{Model of a Meal-Delivery Process}
In online meal delivery service, an order needs to go through the following process until its corresponding customer can receive the food. First, immediately after being placed online (via apps or websites), the order joins into the processing queue at the restaurant. We assume that the restaurant prepares orders in the processing queue one by one in a first come first serve manner, and it can only prepare one order at one time. After an order is done, the order enters a new waiting queue and waits there until a rider picks it up. In the meantime, the platform will consistently dispatch riders to the restaurant at possibly time-varying rates. In our setting, we assume that one arriving rider will take the earliest prepared order from the waiting queue. If there are no prepared orders, riders will also wait in a queue. When one order is ready, the earliest arriving rider will take this order and leave.
We assume that orders arrive in a Poisson process with rate $1$, and its service time is an exponential distribution of rate $\mu$ independent of the arrival process. The assumption of a Poisson arrival process and exponential service time follows the vast study of restaurant service systems \cite{chen2019food, hwang2010joint, chen2020courier}.
Note that the assumption of a unit-rate arrival process is without loss of generality because we can always normalize the time scale. Denote the food processing queue as $\mathcal{Q}_1$, and the waiting queue of prepared food as $\mathcal{Q}_2$. Then, $\mathcal{Q}_1$ itself is an M/M/1 queue. When one order leaves $\mathcal{Q}_1$, it immediately joins $\mathcal{Q}_2$ to wait for riders. Let $Q_1(t), Q_2(t)$ be the number of orders in $\mathcal{Q}_1, \mathcal{Q}_2$ at time $t$, respectively. When there is no waiting order but $a$ waiting riders, we would write $Q_2(t) = -a.$ In this sense, riders are like negative arrivals that can cancel an order from the queue. Figure \ref{fig:order_procedure} provides an illustration of the whole process.
\begin{figure}[bp]
\centering
\includegraphics[scale=0.25]{figures/food_process.pdf}
\caption{Illustration of Meal Delivery Services}
\label{fig:order_procedure}
\end{figure}
\subsection{Information Disclosure Dispatch Process}
Based on knowledge of the order arrival rate, the platform must design corresponding dispatch algorithms to send riders. In addition, the platform has two sources of information, one from users and the other from the restaurant, both of which may play a role in the dispatch process. Specifically, we divide the decision process into two parts: 1) users' and restaurant's information disclosure scheme and 2) dispatch process of the platform.
\subsubsection{\textbf{Information Disclosure}}
When a user submits an order to the platform, he/she can reveal his/her patience time $T^*$, that is, how much time he/she can endure for extra order delays due to deliveries. Users' patience time is crucial for riders since in practice, riders would get low tips or even be disqualified for late deliveries \cite{doordash}. We assume that all orders will fully disclose the same fixed value $T^*$. The platform has to design algorithms satisfying this constraint. We formalize this decision problem in Section \ref{sec:labor-problem}.
From the view of the restaurant, when $Q_2(t) \geq 0$, its value is only visible to the restaurant, not the platform, since the platform has no information of whether one order is finished or not. Therefore, the restaurant could decide whether to reveal such information to the platform. To model the restaurant's willingness to share its private data with the platform, we design the following information disclosure mechanism. In particular, the restaurant sets a nonnegative integer $M$ as a threshold that is also known by the platform. For every time $t$, the restaurant maintains a signal $\theta(t)$ visible to the platform, given by
\begin{equation}
\theta(t) = \left\{
\begin{aligned}
0, & ~\text{if }Q_2(t) \leq M;\\
Q_2(t)-M, & ~\text{if }Q_2(t) > M.
\end{aligned}
\right.
\end{equation}
We allow $\theta(t) \equiv 0$, and write it as $M = \infty$. Such a screening policy is motivated by the fact that restaurants would like to hide the number of waiting prepared orders when it is small. In this way, the platform may send excessive riders, which can help reduce the waiting time of orders. The threshold $M$ measures how much information is being shared. When $M = 0$, the restaurant shares all of its private information. However, if $M = \infty$, the restaurant would provide no data to the platform.
\subsubsection{\textbf{Dispatch process}} Finally, we define the dispatch process of the platform. The dispatch policy is assumed to be a time-varying Poisson arrival process of riders whose rates change according to $\theta(t)$ provided by the restaurant. Extending the dispatch policy to more general processes may entail more technical challenges, which exceeds the focus of this paper.
In particular, the platform first sets a list of dispatch rate parameters $\vec{\lambda}=(\lambda_0,\lambda_1,\cdots)$ where $\lambda_i \leq \Lambda$ and the upper bound $\Lambda$ measures the maximal availability of riders near the restaurant. After choosing $\vec{\lambda}$, the platform will send riders to the restaurant in a Poisson process whose rate is given by $\lambda_{\theta(t)}.$ That is, conditioned on the event that the restaurant signal is equal to a value $i$, riders will arrive to the restaurant in a Poisson process of rate $\lambda_{i}$. Furthermore, if there are already $d$ riders waiting at the restaurant, i.e., $Q_2(t) = -d$, the platform will not send any riders even when it is going to do so. We call $d$ the buffer length.
For ease of exposition, we assume a rider will arrive at the restaurant immediately when he/she is summoned by the platform. In addition, we assume that the arrival process of riders is independent of the order arrival process and service time of orders.
We note here that $\vec{\lambda}, d$ and $\Lambda$ are tightly connected to two crucial parts in today's meal deliveries: postponed dispatch \cite{ulmer2020restaurant} and delivery scope design \cite{ding2020delivery}. Specifically, in postponed dispatch, the platform will not immediately dispatch a rider to restaurants when an order arrives so that they will not arrive with orders being unprepared. Our arrival rates $\vec{\lambda}$ can thus be seen as how fast the platform will send out a rider. Furthermore, when there are already too many riders ($d$ in our model) waiting for food, the platform won't send any rider. In addition, the platform in general will divide a whole city into multiple regions and deploy different numbers of riders in each area. The number of riders in one region will then decide the magnitude of the parameter $\Lambda$.
To ensure that the platform's capacity is sufficient to serve incoming orders and that the system is stable, we impose the following assumptions on dispatch rates of riders.
\begin{assumption}\label{as:stable}
Rates $\mu, \Lambda, \lambda_0$ are larger than $1$, and $\lambda_i > 0$ for all $i > 0$. In addition, it satisfies that there exists a constant $C<\infty$, such that
$
\sum_{i=1}^{\infty}\prod_{j=1}^i \frac{1}{\lambda_i} = C.
$
\end{assumption}
Under the above assumption, the system state can be the pair of $(Q_1(t), Q_2(t))$, which forms a continuous-time Markov chain. Fig. \ref{fig:transition} illustrates a snapshot of this Markov chain. Furthermore, the system is positively recurrent and enjoys a unique stationary distribution based on Assumption \ref{as:stable} \cite{harchol2013performance}. To measure customer experience and labor rights, this paper focuses on the mean order waiting time $\expect{T_o}$ and the mean rider waiting time $\expect{T_r}$ in the system. To be more specific, the waiting time of an order is measured as the time difference between its arrival at the restaurant and its time fetched by a rider. The mean order waiting time is the time-averaged waiting time among all orders. Similarly, the waiting time of one rider is the time between his/her arrival at the restaurant and the time he/she fetches one order. The mean rider waiting time is then the time-averaged waiting time of all riders.
Note that the two expectations are determined by the platform's setting of $\vec{\lambda}, d, M$. Therefore, when we want to stress such dependence, they are written as $\expectsub{\vec{\lambda}, d, M}{T_o}$ and $ \expectsub{\vec{\lambda}, d, M}{T_r}$.
\begin{figure}
\begin{center}
\includegraphics[width=3.5in]{figures/transition.pdf}
\caption{The Markov chain with state $(q_1,q_2)$}
\label{fig:transition}
\end{center}
\end{figure}
\section{Labor-Right Protecting Dispatch Problem and Analysis}\label{sec:labor-problem}
Based on the mathematical model of a meal delivery platform, we propose the labor-right protecting dispatch problem in this section. The goal is to protect laborers while guaranteeing customer experience. In particular, this section includes two parts. In the first part, we formulate the labor rights-protecting dispatch problem. In the second part, we analyze the characteristics of the model, which enables us to design algorithms for the dispatch problem in Section \ref{sec:algorithm}.
\subsection{Problem Formulation}
To formalize the experience guarantee, we first notice that the expected order waiting time is at least the mean response time in an $M/M/1$ queue of arrival rate $1$ and service rate $\mu$. Then, we know $\expect{T_o}$ is at least $\frac{1}{\mu - 1}$ \cite{harchol2013performance}.
\begin{proposition}
The averaged order waiting time $\expect{T_o}$ is at least $\frac{1}{\mu - 1}$.
\end{proposition}
Based on this lower bound on expected order waiting time, we can formalize the meaning of patient time $T^*$ in the sense that we require $\expect{T_o} \leq \frac{1}{\mu - 1} + T^*$ since $\expect{T_o} - \frac{1}{\mu - 1}$ measures the extra delays brought by deliveries. Then, after the restaurant reveals its threshold $M$, the platform needs to find the best $\vec{\lambda},d$ to minimize the averaged rider waiting time while satisfying requirements of users.
Based on the above discussion, we propose the following optimization framework for the labor-right protecting dispatch problem.
\begin{equation}\label{eq:optimization}
\begin{aligned}
& \underset{\vec{\lambda},d}{\text{minimize}}
& & \expectsub{\vec{\lambda},d,M}{\text{rider waiting time }T_r} \\
& \text{subject to}
& & \expectsub{\vec{\lambda},d,M}{\text{order waiting time }T_o} \\
& & &\leq \frac{1}{\mu - 1} + \text{user patience time }T^* .
\end{aligned}
\end{equation}
Note that our optimization problem also belongs to the vast domain of constrained Markov decision processes (CMDPs) \cite{altman1999constrained}. The goal of CMDP is to obtain the optimal state-dependent policy of a certain objective while satisfying the given constraints. The optimal policy can be complicated and even randomized. However, in the context of our paper, we not only want to solve the optimization problem (\ref{eq:optimization}) but also seek to explicitly understand the trade-off between the objective "rider waiting time" and the constraint "order waiting time" and how the decisions and the trade-off could change in response to the volume of information provided by the restaurant. Therefore, the policy space in our optimization problems is kept as simple as possible to discuss the above information-related decisions and does not involve the general space of randomized policies. We believe a new framework is needed for such discussions and thus leave it as a future study.
\subsection{Model Analysis}
To obtain a deeper understanding of the problem, we first study the stationary distribution of the system in the general setting where the restaurant allows a threshold $M$, and the platform sets the associated dispatch rate $\vec{\lambda}$ and the buffer length $d$. Based on the distribution, we observe a decoupling phenomenon between $\mathcal{Q}_1$ and $\mathcal{Q}_2$. Finally, we restrict our scope to the case of $M=\infty$ and obtain close-form formulas for $\expect{T_o}$ and $\expect{T_r}$.
\subsubsection{\textbf{Stationary Distribution of the General Model}}
Recall that the state of the system is given by $(Q_1(t), Q_2(t))$, namely, the number of orders waiting for preparation and the number of prepared orders waiting for riders. If $Q_2(t) < 0$, it is the negative of waiting riders. Let $(\mathbf{Q}_1,\mathbf{Q}_2)$ be the corresponding random variables distributed with the stationary distribution of $(Q_1(t), Q_2(t))$. Let $\pi_{q_1,q_2}$ be the stationary distribution of the event $(\mathbf{Q_1} = q_1, \mathbf{Q_2} = q_2)$. Fix the threshold $M$, the dispatch rates $\vec{\lambda} = (\lambda_0,\lambda_1,\cdots)$, and the buffer level $d$, and assume they satisfy the assumption \ref{as:stable}. The state space of the system would be $\mathbb{N} \times \left(\{-d, d+1,\cdots, 1\}\cup \mathbb{N}\right)$.
Define $\rho = \frac{1}{\lambda_0}$. The following lemma shows that$\pi_{q_1,q_2}$ exhibits a product-form property similar to the property in queueing networks with negative arrivals \cite{henderson1993queueing}. The proof of this lemma can be found in the appendix.
\begin{lemma}\label{lemma:distribution}
It holds that for $q_1 \geq 0, q_2 \geq -d$,
\[
\pi_{q_1,q_2} = \left\{
\begin{aligned}
&C_1\left(1-\frac{1}{\mu}\right)\left(\frac{1}{\mu}\right)^{q_1}\rho^{q_2},~\text{if }q_2 \leq M; \\
&C_1\left(1-\frac{1}{\mu}\right)\left(\frac{1}{\mu}\right)^{q_1}\rho^{M}\prod_{i=1}^{q_2-M}\frac{1}{\lambda_i},~\text{otherwise,}
\end{aligned}
\right.
\]
where
\begin{equation}
C_1 = \frac{1-\rho}{\rho^{-d}-\rho^{M+1}+(1-\rho)\rho^M\sum_{i=1}^{\infty}\prod_{j=1}^i \frac{1}{\lambda_i}}.
\end{equation}
\end{lemma}
\subsubsection{Queue Decoupling}\label{sec:queue-decoupling}
Observe that by Lemma \ref{lemma:distribution}, we can indeed decouple $\mathcal{Q}_2$ from $\mathcal{Q}_1$. If we consider only the distribution of events $\mathbf{Q}_2 = q$ for $q \geq -d$, the stationary distribution of $\mathcal{Q}_2$ is equivalent to the stationary distribution of the following Markov chain. In the equivalent Markov chain, the set of the state is $\{-d,\cdots,-1\} \cup \mathbb{N}$. For each state $q \geq -d$, it transitions to state $q + 1$ with rate $1$ and to $q - 1$ with rate $\lambda_{\max(q-M,0)}$ if $q > -d$. The stationary distribution of this Markov chain, denoted by $\nu_q$, is exactly given by
\begin{equation}
\nu_q = \left\{
\begin{aligned}
&C_1\rho^{q_2},~\text{if }q_2 \leq M; \\
&C_1\rho^{M}\prod_{i=1}^{q_2-M}\frac{1}{\lambda_i},~\text{otherwise,}
\end{aligned}
\right.
\end{equation}
where $\rho,C_1$ are the same as those in Lemma \ref{lemma:distribution}.
In addition, let $\mathbf{Q}^*$ be the random variable corresponding to the state in this Markov chain. Then, by equivalence, we have the following property connecting $\expect{T_o},\expect{T_r}$ with $\mathbf{Q}^*$. Due to space limit, we omit its proof.
\begin{lemma}\label{lemma:equivalence}
It holds that
\begin{equation}
\begin{aligned}
\expect{T_o} &= \frac{1}{\mu - 1} + \expect{\max(\mathbf{Q}^*,0)} \\
\expect{T_r} &= -\expect{\min(\mathbf{Q}^*,0)}.
\end{aligned}
\end{equation}
\end{lemma}
\subsubsection{\textbf{Analysis for $M = \infty$}}
Restricting our scope to $M = \infty$, based on Lemma \ref{lemma:distribution}, we can obtain expressions of $\expect{T_o}$ and $\expect{T_r}$ via Little's Law\cite{harchol2013performance}. Formally, we have this lemma.
\begin{lemma}\label{lemma:meanValue}
For fixed $\lambda_0 > 1$ and $d$, let $\rho = \frac{1}{\lambda_0}$. It satisfies that
\begin{subequations}
\begin{equation}
\expectsub{\lambda_0,d}{T_o} = \frac{1}{\mu - 1} + \frac{\rho^{d+1}}{1-\rho},
\end{equation}
\begin{equation}
\expectsub{\lambda_0,d}{T_r} = d - \frac{\rho-\rho^{d+1}}{1-\rho}.
\end{equation}
\end{subequations}
\end{lemma}
\section{Labor-right protecting algorithm} \label{sec:algorithm}
Since it is generally nontrivial to solve (\ref{eq:optimization}), in the first part of this section, we start our analysis by assuming that there is no restaurant information, i.e., $M = \infty$. Under this assumption, we propose Algorithm \ref{algo:optimal-parameter} that provides the procedure to find the optimal $\lambda_0, d$. Second, when there are restaurant data, we design Algorithm \ref{algo:policy-improve} to improve policies based on new information.
\subsection{Algorithm and its Optimality when No Restaurant Data}
When there is no restaurant information, the platform needs only to seek the optimal design of $\lambda_0, d$. The closed-form expressions of $\expect{T_o}$ and $\expect{T_r}$ shown in the Lemma \ref{lemma:meanValue} provides the basis of the following algorithm to solve (\ref{eq:optimization}).
\begin{algorithm}[Optimal Dispatch Algorithm]\label{algo:optimal-parameter}
The algorithm takes $T^*$ and $\Lambda$ as inputs and outputs two parameters $\lambda_0, d$ as follows.
\begin{itemize}
\item If $\frac{\Lambda}{1-\Lambda} \leq T$, simply returns $d = 0, \lambda_0 = \Lambda$;
\item Otherwise, for every $\rho \in [0,1)$, define $D(\rho)$ as the smallest nonnegative integer such that $\frac{\rho^{D(\rho)+1}}{1-\rho} \leq T^*$; and for every nonnegative integer $n$, define $\gamma(n)$ as the maximum real number in $(0,1)$ such that $\frac{\gamma(n)^{n+1}}{1-\gamma(n)} \leq T^*$.
\item Let $d_1 = D(\frac{1}{\Lambda}).$ The algorithm returns $d_1, \frac{1}{\gamma(d_1)}$.
\end{itemize}
\end{algorithm}
The next theorem supports the optimality of Algorithm \ref{algo:optimal-parameter}.
\begin{theorem}\label{thm:optimal-algo}
Under the two parameters $\lambda_0, d$ found by Algorithm \ref{algo:optimal-parameter}, and assuming $M = \infty$, it satisfies $\expect{T_o} = \frac{1}{\mu - 1} + T^*$, and $\expect{T_r}$ is minimized.
\end{theorem}
Let's look at a proof sketch of Theorem \ref{thm:optimal-algo}. A formal proof is provided in the appendix. The goal is to minimize $\expect{T_r}$ while satisfying the delay constraint in (\ref{eq:optimization}). By Lemma \ref{lemma:meanValue}, we want to find $\lambda_0, d$ such that $d-\frac{\rho-\rho^{d+1}}{1-\rho}$ is minimized and $\frac{\rho^{d+1}}{1-\rho} \leq T^*$. The intuition of Algorithm \ref{algo:optimal-parameter} is to choose the smallest $\lambda_0$ satisfying the constraint in (\ref{eq:optimization}). Note that $\expect{T_r}$ is a decreasing function of $\rho$ when $d$ is fixed. Therefore, after finding the smallest $d$, we would increase the value of $\rho$ (or equivalently decrease $\lambda_0$) until the constraint in (\ref{eq:optimization}) is violated.
\subsection{Policy Improvement using Restaurant Information}
In this section, we explore how to improve platform policies using restaurant data. Particularly, the restaurant provides more information to the platform by reducing its threshold from $M$ to $m$, where $m < M$. Our results highlight that restaurant data are of particular importance to help the platform make better decisions.
Formally, let us assume that the platform has a policy given by when the threshold is $M$. Then, algorithm \ref{algo:policy-improve} outputs a new policy for a given threshold $m, m < M$ that is arguably better than the original policy. Algorithm \ref{algo:policy-improve} is given below.
\begin{algorithm}[Policy Improvement with Restaurant's Data]\label{algo:policy-improve}
The algorithm takes $\vec{\lambda}=(\lambda_0,\cdots),d,M,m$ and $\Lambda$ as inputs and outputs the dispatch rates and buffer level, $\vec{\tau}=(\tau_0,\cdots), d_1$, of the new policy as follows. Define $\rho = 1/\lambda_0$.
\begin{itemize}
\item Set $C = \frac{\rho-\rho^{M-m+1}}{1-\rho}+\rho^{M-m}\sum_{i=M+1}^{\infty}\prod_{j=1}^{i-M} \frac{1}{\lambda_j}.$
\item Set $\tau_0 = \lambda_0, \tau_1 = \frac{1}{(1-\frac{1}{\Lambda})C}$, and $\tau_i = \Lambda$ for all $i > 1$.
\item Let $d_1 = d$. Output $(\vec{\tau}, d_1)$.
\end{itemize}
\end{algorithm}
Our next result shows that using Algorithm \ref{algo:policy-improve}, the platform can always find a policy with better averaged order waiting time and the same averaged rider waiting time based on any policy of a larger threshold.
\begin{theorem}\label{thm:data-value}
Suppose that the platform has a policy with parameters $\vec{\lambda}=(\lambda_0,\cdots),d$ and the restaurant's threshold is $M$, $M > 0$. If $M$ reduces to $m$, $m < M$, then using Algorithm \ref{algo:policy-improve}, the platform can construct a new set of dispatch rates $\vec{\tau}$, such that
\begin{equation}\label{eq:policy-improve}
\expectsub{\vec{\tau},d,m}{T_o} \leq \expectsub{\vec{\lambda},d,M}{T_o}, ~\expectsub{\vec{\tau},d,m}{T_r} = \expectsub{\vec{\lambda},d,M}{T_r},
\end{equation}
where the inequality becomes equal only if
\[
\rho=\frac{\Lambda-1}{\Lambda}\left(\frac{\rho-\rho^{M-m+1}}{1-\rho}+\rho^{M-m}\sum_{i=M+1}^{\infty}\prod_{j=1}^{i-M} \frac{1}{\lambda_j}\right).
\]
\end{theorem}
\begin{remark}
Note that if $M = \infty$ in Theorem \ref{thm:data-value}, the inequality in (\ref{eq:policy-improve}) becomes equal if and only if $\lambda_0 = \Lambda$, that is, the platform is already dispatching riders at its full capacity.
\end{remark}
The proof of Theorem \ref{thm:data-value} can be found in the appendix.
\section{Obligation of different stakeholders}
In this section, we discuss obligations in protecting the labor rights of the three stakeholders: the platform, users, and the restaurant.
First, for restaurant information, Theorem \ref{thm:data-value} tells us that whenever there are data from the restaurant side, the platform can improve the experience of customers while maintaining the same averaged rider waiting time. In this sense, restaurant data help to improve labor rights by making the labor more efficient. We will further quantify the effect of queueing data of the restaurant in Section \ref{sec:simulation}.
Second, we desire to understand how the customer's requirement of experience and the platform's ability to recruit riders may affect the averaged rider waiting time. For ease of exposition, we assume that $M = \infty$ in this part. Furthermore, suppose that the platform is still trying to solve the labor-right protecting dispatch problem introduced in (\ref{eq:optimization}).
In this case, we show that there is a fundamental lower bound of $\expect{T_r}$ given that $T^*$ is low. The proof is deferred to the appendix.
\begin{theorem}[Dilemma between customers and riders]\label{thm:tradeoff}
Suppose that $\hat{\rho}\colon=\frac{1}{\Lambda} \geq 0.3$ and $\frac{\hat{\rho}^4}{1-\hat{\rho}} \geq T^*$ and $M = \infty$. Then, it holds
$
\expect{T_r} \geq \frac{\ln(4/3)}{4}\frac{1}{1-\hat{\rho}}-T^*,
$
for all feasible $d$ and $\lambda_0$ satisfying the constraint in (\ref{eq:optimization}).
\end{theorem}
This result shows that the sum of $T^*$ and $\expect{T_r}$ is lower bounded by a value determined by the platform capacity $\Lambda$. As a result, when $\Lambda$ is fixed, no matter how the platform acts, it is impossible for both $T^*$ and $\expect{T_r}$ to become negligible simultaneously. Nevertheless, the lower bound decreases when $T^*$ or $\Lambda$ increases, thus implying that if customers allow a longer delivery time or if the platform is able to recruit more riders, then laborers can enjoy a lower average waiting time.
\section{Numerical Case} \label{sec:simulation}
In this section, we aim to illustrate the importance of the labor rights protection problem, the dilemma between customer experience and rider waiting time, and the value of restaurant information using simulations. Particularly, we assume that $\lambda_0 = 1$, and $\mu = \Lambda = 1.5$. It is a reasonable assumption because in practice, the restaurant and the platform usually only recruit nearly sufficient numbers of riders to reduce cost.
\subsection{Importance of labor rights protection and Trade-offs}
To study the impact of the labor rights protection problem, we consider three settings of patient time, $T^*=0.01, 0.05, 0.1$. In addition, we assume that there is no restaurant information, so $M = \infty$. Then, for $\lambda_0 \in (1,\Lambda)$, we plot the curve of $\expectsub{\lambda_0,d}{T_r}$, where $d$ is the minimum integer such that $\expectsub{\lambda_0,d}{T_r} - \frac{1}{\mu - 1} \leq T^*$. The result is shown in Fig. \ref{fig:right-protect}.
\begin{figure}[!hbtp]
\centering
\includegraphics[scale=0.3]{deliver-time.jpg}
\caption{Averaged Rider waiting time under Different Dispatch Rates Given that Customer Experience is satisfied}
\label{fig:right-protect}
\end{figure}
As we can see, for a fixed $T^*$, although all the patience time is satisfied, the averaged rider waiting time varies greatly. A clever setting of $\lambda_0$ can reduce the waiting time by a factor of $4$. In addition, when $T^*$ increases, the averaged rider waiting time decreases. It characterizes the trade-off between customer experience and labor rights.
\subsection{Value of Restaurant's Data}
To highlight the value of the data from the restaurant, we assume that $M$ could be $0,10,\infty$. In this case, we are able to differentiate different levels of restaurant information. For the platform's policy, we fix $d = 0$ and then allow $\lambda_0$ to vary from $1$ to $\Lambda$ to be the policy for $M = \infty$. Then, for other $M$, we employ Algorithm \ref{algo:policy-improve} to generate new policies. We plot the value of $\expectsub{\vec{\lambda},d,M}{T_o}$ after the policy improvement as shown in Fig. \ref{fig:policy-improve}.
\begin{figure}[!hbtp]
\centering
\includegraphics[scale=0.3]{order-time.jpg}
\caption{Averaged Order Waiting Time After Using Restaurant Data}
\label{fig:policy-improve}
\end{figure}
As we can tell from Fig. \ref{fig:policy-improve}, restaurant data and algorithm \ref{algo:policy-improve} can indeed greatly improve the customer experience, decreasing $\expect{T_o}$ by a factor of two when $M = 0$ and $\lambda_0 = 1.2$. Specifically, we can observe that the improvement is greater when $\lambda_0$ or $M$ is smaller.
\section{CONCLUSIONS}
In this paper, we formulated a queueing model for meal delivery platforms that incorporated four key stakeholders: the platform, the restaurant, riders, and customers. In particular, we proposed the labor rights protecting problem, whose goal is to reduce waiting time of riders while maintaining satisfactory customer experience. When there is no restaurant information, we designed an efficient algorithm to find the optimal dispatch parameters for the platform. When the restaurant is willing to share its private data, we further characterized a policy improvement method that takes any policy as an input and returns a refined policy based on restaurant data. The refined policy provably improves customer experience, while labor rights did not get hurt. Our results also indicated that there was a trade-off between customer experience and labor rights in the sense that they cannot be satisfied simultaneously. Our simulations confirmed the theoretical results. Specifically, they revealed that an optimal labor-rights protecting algorithm could dramatically reduce the waiting time of riders, and restaurant data could considerably reduce the order waiting time, especially when the restaurant would like to share more information.
\bibliographystyle{IEEEtran}
|
\section{Introduction} \label{sec:introduction}
Video-language segmentation (VLS) aims to predict a set of segmentation masks for the referent in a video under the guidance of a natural language description. This practical problem is fundamental to many applications such as interactive video editing, object tracking and human-machine interaction~\cite{wang2021survey}. Unlike the general semantic segmentation tasks that is operated over a fixed set of pre-defined categories, VLS requires fine-grained joint modeling for vision and language, yielding a more challenging scenario due to the great diversity of linguistic concepts and extra semantics in temporal domain.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.98\linewidth]{images/Moti.pdf}
\end{center}
\captionsetup{font=small}
\caption{\small
\textbf{Illustration of the video-language segmentation task.}
Given a linguistic expression, the predictions of state-of-the-art method \cite{wang2019asymmetric} and CVLS are shown in (a) and (b) respectively.
Previous methods (a) mainly rely on the implicit fusion to formulate cross-modal interaction and fail to distinguish semantically similar objects while CVLS (b) does the opposite.
}\label{fig:motivation}
\end{figure}
A commonplace solution in prior work is first extracting visual and linguistic features respectively, followed by a fusion process to implicitly interwind cross-modal features, et al.}{\emph{et al}\onedot{e.g.}, straightforward concatenation \cite{shi2018key,hu2016segmentation}, cross-modal attention-mechanism \cite{wang2019asymmetric,ningpolar,hui2021collaborative} or dynamic filter \cite{wang2020context,gavrilyuk2018actor}, and then conducting segmentation based on the fused feature. However, these solutions rarely consider the crucial fine-grained alignment of the cross-modal information hence prone to fail in distinguishing the semantically similar objects. As in Fig.~\ref{fig:motivation}, the model with implicit fusion strategy fails to capture the contextual difference among multiple similar regions, et al.}{\emph{et al}\onedot{i.e.}. person conducting different actions.
In this paper, we focus on explicitly modeling the vision-language interactions for the VLS problem. Our approach is inspired by the contrastive representation learning theory~\cite{wu2018unsupervised,chen2020simple,wang2021exploring} whose fundamental purpose is to capture similarities between examples of the same class and contrast them with examples from other classes. Dubbed Contrastive Video-Language Segmentation (CVLS), our method presents a cross-modal contrastive objective to align the referred visual object with the linguistic description in representation space while contracting the unreferred video information with the description at the same time. In this way, fine-grained vision-language interactions can be explicitly learned thus boosting the representation discriminability among visually similar objects based on a linguistic description.
Although vanilla contrastive representation learning provides a potentially feasible solution for the vision-language interactions, we find it problematic to apply the Na\"ive paradigm mechanically on the raw cross-modal representations.
Upon reviewing the VLS task, we notice that the background areas occupy a large proportion across frames while the visually similar candidates do the opposite.
Since the background areas can be easily excluded from the candidate referent just using the visual features while the visually similar objects can be only distinguished on condition of a linguistic feature\footnote{For convenience, we use ``visual-distinguishable region'' to refer the former while use ``linguistic-distinguishable object'' to refer the latter hereafter.}.
For example, as shown in Fig.~\ref{fig:motivation}, models can easily distinguish foreground objects (person) apart from a large area of background even without a linguistic guidance.
A trivial construction of instance-level positive-negative pairs would introduce excessive uninformative visual samples that are over-easy for the current linguistic embedding to discriminate, which might directly degrade the cross-modal contrastive objective to a single-modal feature classifier, et al.}{\emph{et al}\onedot{i.e.}, a visual saliency detector.
To address this issue, we present two simple yet effective hard instance mining strategies, et al.}{\emph{et al}\onedot{i.e.}, Language-relevant Channel Filter (LCF) and Relative Hard Instance Construction (RHIC), which are specially designed for the cross-modal scenario.
They encourage the network to exclude uninformative visual-distinguishable features and to focus on linguistic-distinguishable objects during the contrastive training.
On one hand, with the guidance of linguistic description, LCF module generates channel-wise filters for visual features to highlight the linguistic-distinguishable features.
On the other hand, in RHIC module, only those misclassified pixels with higher confidence are considered while positive-negative pair sampling, to avoid gradient from being dominated by samples from visual-distinguishable regions.
Main contributions of our work are as follows:
\begin{itemize}
\item We propose CVLS, a cross-modal contrastive learning-based method for VLS task, which explicitly captures the commondality of vision-language representations thus boosting representation discriminability among visually similar objects on condition of linguistic descriptions.
\item We propose two simple yet efective hard mining strategies tailored for cross-modal contrastive objective. These strategies are able to filter out over-easy video content from visual features, which enriches the informative gains for contrastive learning.
\item The proposed method significantly outperforms SOTA methods in most metrics on two popular video-language segmentation benchmarks, et al.}{\emph{et al}\onedot{i.e.}, A2D, and J-HMDB.
\end{itemize}
\section{Related Work}
\subsection{Video-Language Segmentation}
Towards understanding fine-grained video representation, Xu et al.}{\emph{et al}\onedot{et al.}~\cite{xu2015can} introduce a task of actor and action video segmentation with the Actor-Action Dataset (A2D) containing a fixed set of actor-action pairs and pixel-level annotations.
Later,
Gavrilyuk et al.}{\emph{et al}\onedot{et al.}~\cite{gavrilyuk2018actor} extend Actor-Action Dataset (A2D) with human-annotated sentences and introduce the challenging text-based video segmentation task that requires further comprehension of both vision and language modalities.
They adopt dynamic convolution filters generated by language to align the multi-modal feature.
Then, Wang et al.}{\emph{et al}\onedot{et al.}~\cite{wang2020context} extend vanilla dynamic convolution with vision context to generate spatial-relevant kernels. Ning et al.}{\emph{et al}\onedot{et al.}~\cite{ningpolar} embed spatial expressions into attention module in terms of direction and range for better linguistic spatial formulation. Wang et al.}{\emph{et al}\onedot{et al.}~\cite{wang2019asymmetric} utilize asymmetric attention mechanisms to facilitate visual guided linguistic feature learning. Hui et al.}{\emph{et al}\onedot{et al.}~\cite{hui2021collaborative} further introduce an additional 2D spatial encoder to alleviate the spatial information misaligned problem brought by 3D CNNs.
Mcintosh et al.}{\emph{et al}\onedot{et al.}~\cite{mcintosh2020visual} introduce a capsule-based approach for better capturing the relationship between multi-modal features. To further mining continuous temporal information, they extend the A2D dataset with annotations for all frames.
Even with certain achievement, existing methods mainly suffers from coarse cross-modal interaction, et al.}{\emph{et al}\onedot{i.e.}, implicit multi-modal fusion, which directly leads to ambiguous boundaries for semantically similar objects.
In this paper, we address this limitation by constructing instance-level multi-modal interaction in an explicitly contrastive manner.
\subsection{Multi-modal Contrastive Learning}
Contrastive Learning techniques learn representations in a discriminative manner by pulling positive pairs closer against negative pairs.
Recently, self-supervised visual representation learning surges and earns a great success either by partial data prediction \cite{liu2020pic,doersch2015unsupervised,dosovitskiy2015discriminative}, or instance-level training image discrimination \cite{chen2020simple,he2020momentum,tian2019contrastive,wu2018unsupervised}. These methods mainly rely on instance augmentation to form positive pairs.
Beyond augmentation-based approach, multi-modal contrastive learning leverages multiple modalities of the same instance to construct positive pairs. And cross-modal representations are learned during the procedure.
Examples of commonly used data modalities include appearance of image, motion of video \cite{harley2019learning}, depth \cite{tian2019contrastive}, luminance and chrominance \cite{zhang2017split}, audio \cite{wu2019DAM,wu2021explore}, and text \cite{zhang2020contrastive,gupta2020contrastive} as in our work.
Zhang et al.}{\emph{et al}\onedot{et al.}~\cite{zhang2020contrastive} and Gupta et al.}{\emph{et al}\onedot{et al.}~\cite{gupta2020contrastive} both design sophisticated strategies to align vision region with expression and maximize the mutual information between the modalities. In this work, inspired by such great progress, we propose a specially-designed cross-modal contrastive learning objective for the first attempt of introducing explicit cross-modal interaction into Video-Language Segmentation.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.9\linewidth]{images/pipeline.pdf}
\end{center}
\captionsetup{font=small}
\caption{\small
\textbf{Illustration of Contrastive Cross-modal Learning Module (CCLM)}. Visual encoder and linguistic encoder are applied to extract object-level visual features $\{r_p, r_{n_1}^1, \cdots, r_{n_2}^3\}$ and reference feature $r_l$ respectively, which are then projected to a $128$-dimensional space and $L2$ normalized.
$r_p$ is the corresponding referent feature.
$r_{n_i}^t$ denotes negative instances in the video with time step $t$, including background (grey region).
Optimal representations are learned via cross-modal discrimination, where the referred object and the language description are pulled closer and the distracting contents are pushed apart across frames.
Language-relevant Channel Filter and Relative Hard Instance Construction are abbreviated as ``LCF" and ``RHIC" (Sec.~\ref{section:LCF}) respectively.
}\label{fig:pipeline}
\end{figure*}
\section{Method}
In this section, we first present the basic Cross-modal Conrastive Learning procedure in Sec.~\ref{section:ObjCons}. To further alleviate degradation problem, we then introduce two specially-designed cross-modal hard mining approaches in Sec.~\ref{section:HM}. Even not being restricted to any certain feature extractor, in Sec.~\ref{section:overall}, we finally describe the overall architecture implemented in this work.
\subsection{Na\"ive Instance-level Cross-modal Contrastive Representation Learning (CCLM)} \label{section:ObjCons}
As illustrated in Fig.~\ref{fig:pipeline}, with a pair of reference (natural language description) and corresponding referent (visual object), our goal is to learn a good representation space that the positive pairs would be pulled together while other objects in a video would be pushed away at mean time.
\subsubsection{Within-modal representation construction.}
Different from the prevailing segmentation models that formulate visual embeddings in grid-level as in \cite{zhao2020contrastive}, in VLS task we argue that it is more reasonable and efficient to formulate within-modal embeddings and construct cross-modal commonalities in instance-level concerning that natural language expressions are mostly annotated in an object view.
Now the problem lies in how to construct within-modal representations for the reference and objects.
For linguistic modality, reference feature $r_l$ can be directly extracted by a sentence-level linguistic encoder (Detailed later in Sec.~\ref{section:LE}).
For visual modality, instance-level feature is obtained by conducting average-pooling on mask-cropped visual feature map extracted from visual encoder (Sec.~\ref{section:VE}).
Formally, the naive process for obtaining instance representation $r_v^t$ at time step $t$ could be defined as:
\begin{equation}
r_v^t = AVG(X^t_v \odot o^t). \label{equation:OT}
\end{equation}
Here we denote $X^t_v$ and $o^t$ as the extracted visual feature and a binary object mask.
$\odot$ and $AVG$ correspond to element-wise multiplication and average pooling respectively.
We treat the entire video clip as a pool of object features $\{r_{v_1}^1,r_{v_1}^2,\cdots,r_{v_{N_v}}^T,\}$, where $N_v$ is the total number of instances in a video clip\footnote{Following definitions in literature~\cite{lin2021video,wang2021end}, we use the term et al.}{\emph{et al}\onedot{instance} for video-level identification and use et al.}{\emph{et al}\onedot{object} for frame-level, et al.}{\emph{et al}\onedot{e.g.}, there are $2$ instances and $2T$ objects in the example video clip (Shown in Fig.~\ref{fig:pipeline})} with $T$ frames.
For ease of notation, we define the positive referent object corresponding to each linguistic reference $r_l$ as $r_p$ and all other negative visual objects as $r_{n_i}^t$, where $i \in \{1,\cdots,N_v-1\}$ and $t \in \{1,\cdots,T\}$.
Correspondingly, $o_p$ is the referent region and $o_{n_i}^t$ is the negative object region. Both of them can be obtained from ground-truth masks during training. In later sections, we omit time stamp $t$ for ease of notation.
It should be noted that, background region (grey region in Fig.~\ref{fig:pipeline}) is also considered as a negative instance and counted in $N_v$.
Within-modal features in the pool are further fed into a projection network, $Proj(\cdot)$, mapped to a same cross-modal representation space as in Fig.~\ref{fig:pipeline}, et al.}{\emph{et al}\onedot{i.e.}, $z_p = Proj(r_p)$, $z_{n_i} = Proj(r_{n_i})$, $z_l = Proj(r_l)$.
We instantiate $Proj(\cdot)$ as a multi-layer perceptron as in~\cite{chen2020simple} with a single hidden layer of size $512$ and output vector of size $128$.
The outputs are then normalized to lie on an unit hypersphere, which enables using an inner product to measure distances in the projection space, et al.}{\emph{et al}\onedot{i.e.}, $||z_p||=||z_{n_i}||=||z_l||=1$.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.98\linewidth]{new_images/pipelinelatest.pdf}
\end{center}
\captionsetup{font=small}
\caption{\small
\textbf{Detailed architecture of CVLS pipeline.}
Visual feature and linguistic feature are extracted by a FPN-like encoder and a bi-LSTM \cite{huang2015bidirectional} respectively \textbf{(c)}.
Language-relevant Channel Filter (\textbf{LCF}) (Sec.~\ref{section:LCF}) is integrated in visual encoder to exclude the over-simple visual-distinguishable feature.
Multi-scale output of the visual encoder is then concatenated as $X_v$ and fed into the Relative Hard Instance Construction (\textbf{RHIC}) (Sec.~\ref{section:LCF}) to aggregate instance-level visual features.
``\textbf{CCLM}" denotes the Cross-modal Contrastive Learning Module (Sec.~\ref{section:ObjCons}) as in Fig.~\ref{fig:pipeline}.
Detailed process of \textbf{(a)} Spatial Encoding Module (\textbf{SEM}) and \textbf{(b)} Language-relevant Channel Filter \textbf{(LCF)}.
}\label{fig:overall}
\end{figure*}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.95\linewidth]{images/RHIC.pdf}
\end{center}
\captionsetup{font=small}
\caption{\small
\textbf{Illustration of region-of-interests} while aggregating visual instance representation between Naive Instance Construction (Naive IC) and Relative Hard Instance Construction (RHIC).
Instead of pooling all the pixels, RHIC only considers relatively harder region of each instance, based on the confidence score of each pixel.
The positive referent and negative candidate are painted in green and red respectively.
Background region is also considered as a negative instance and is painted in blue in both figures.
}\label{fig:rhic}
\end{figure}
\subsubsection{ Instance-level Cross-Modal Contrast.}
With constructed instance-level features, the representation space is then shaped with the instance-level cross-modal contrastive objective $\mathcal{L}_c$, which optimizes the triple representations of the positive referent, the negative objects and the language description in the latent space:
\begin{equation}
\mathcal{L}_c = - \log \frac{\exp({z_p} \cdot z_l/\tau)}{\sum_{j} \exp({z_{n_j}} \cdot z_l/\tau) + \exp({z_p} \cdot z_l/\tau)},
\end{equation}
\noindent
where $\tau$ is a scalar temperature parameter, and a lower temperature value would increase the influence of harder negatives \cite{wu2018unsupervised,gunel2020supervised}.
The cross-modal contrastive objective regularizes the embedding space and force the network to explore explicit cross-modal discrimination among semantically similar objects.
\subsubsection{Overall Training Objective.}
During contrastive training, our method is supervised under both task-specific segmentation loss $\mathcal{L}_s$ and instance-level cross-modal contrastive loss $\mathcal{L}_c$ defined in previous section.
$\mathcal{L}_s$ optimizes the segmentation mask based on the
extracted multi-modal tensor:
\begin{equation}
\mathcal{L}_s = - s \odot \log (\sigma(e)) - (1-s) \odot \log(1-\sigma(e)),
\end{equation}
where the $\cdot$ symbol denotes the inner (dot) product, $\odot$ is element-wise multiplication. And $e$ is the output of mask decoder and $s$ denotes the ground truth.
And overall training target of our method is a weighted average of $\mathcal{L}_s$ and $\mathcal{L}_c$:
\begin{equation}
\mathcal{L} = \mathcal{L}_s + \lambda \mathcal{L}_c,
\end{equation}
\noindent
where $\lambda$ is a weighting hyperparameter, and $\lambda \in [0.1,1]$.
\subsection{Cross-Modal Hard Example Mining}\label{section:HM}
Though vanilla contrastive learning provides an explicit standpoint for modeling the fine-grained vision-language interaction, as we described in Sec.\ref{sec:introduction}, equally treat the visual-distinguishable regions and the linguistic-distinguishable objects (et al.}{\emph{et al}\onedot{e.g.}, average pooling) might directly degrade the cross-modal contrastive objective to a single-modal feature classifier.
We thus propose two hard mining strategies tailored for cross-modal scenario, aiming to reduce the impact of excessive uninformative visual samples and enlarge the informative gains from both implicit and explicit aspects.
\subsubsection{Language-relevant Channel Filter (LCF). } \label{section:LCF}
As stated in~\cite{hu2018squeeze,chen2017sca,zeiler2014visualizing}, channel-wise features intrinsically encode the responses of different convolutional filters, and the visual concepts can be integral with different combination of channel responses.
Following this insight, we employ natural language as a filter to enhance language-relevant visual features and depress the language-irrelevant feature before contrastive training by applying dynamically conditioned weights on channels.
Specifically, we design a simple gating mechanism with a sigmoid activation as a filter on each channel.
With the linguistic feature $r_l$ and visual feature map $X$, the filter can be employed through:
\begin{equation}
\begin{aligned}
g &= \sigma(W_2\ \delta ({W_1}{r_l} + b_1)+b_2), \\
X_v &= g \odot X,
\end{aligned}
\end{equation}
where $\delta(\cdot)$ refers to the LeakyRELU operation \cite{xu2015empirical}. $W_i$ and $b_i$ are learnable matrix and bias.
Rather than enforcing the one-hot activation, we employ sigmod function $\sigma(\cdot)$ to normalize the excitation range to $(0,1)$, which enables a non-mutually-exclusive relationship among all channels. In this way, the impact of the visual-distinguishable feature declines, which implicitly forces the network to focus on linguistic-distinguishable content during the contrastive training.
\subsubsection{Relative Hard Instance Construction (RHIC). } \label{section:rhic}
An explicit hard mining strategy is introduced to further reduce the impact of uninformative visual regions. Dubbed Relative Hard Instance Construction (RHIC), it explicitly selects a portion of the entire object as et al.}{\emph{et al}\onedot{hard} regions for visual object representation construction.
Specifically, during instance construction, only the misclassified pixels with high confidence scores are considered as et al.}{\emph{et al}\onedot{hard} regions, which are sampled with a relative ratio (1:3) of hard to easy pixels.
Formally, given the model’s estimated probability $c$ predicted by segmentation decoder, we define misclassified degree $c_p$:
\begin{equation}
\begin{aligned}
c_p =
\begin{cases}
1-c & \text{if $y=1$} \\
c & \text{otherwise}.
\end{cases}
\end{aligned}
\end{equation}
In the above $y\!\in\!\{0,1\}$ specifies ground-truth label of each pixel.
The instance region in Equation \ref{equation:OT} can be updated by:
\begin{equation}
o^t = Top\_K(c_p) \odot o^t,
\end{equation}
where $Top\_K$ denotes the relative selecting process.
The visualization of the selected et al.}{\emph{et al}\onedot{hard} regions by RHIC is shown in Fig.~\ref{fig:rhic}(b).
It is worth-noted that we employ a et al.}{\emph{et al}\onedot{relative} selecting strategy here which introduces a flexible value of negative samples and $K$ corresponding to the total
pixel number of the object region and the prediction will be et al.}{\emph{et al}\onedot{progressively} improved by et al.}{\emph{et al}\onedot{online} selecting the relative-hard regions in each training step.
In practise, as occupation of different objects vary dramatically, we find a fixed threshold, et al.}{\emph{et al}\onedot{i.e.}, a fixed number of $K$ is hard to tune. And we conduct additional ablations on relative ratio in Sec.\ref{sec:exp}.
\subsection{Overall Architecture}\label{section:overall}
In this section we instantiate the detailed architecture design as shown in Fig.~\ref{fig:overall}. Nevertheless, the idea of CVLS is general in the sense that basic encoder and decoder can be easily replaced with other stronger baselines.
\subsubsection{Linguistic Feature Encoder. } \label{section:LE}
Given a linguistic sentence $S = \{s_i\}_{N_l}$ with $N_l$ words,
linguistic feature is encoded with a bi-LSTM \cite{huang2015bidirectional} module followed by a self-guided attention fusion.
Specifically, linguistic feature $r_l \in \mathbb{R}^{C_v}$ can be calculated from:
\begin{equation}
r_l = W_l(\sum_{i=1}^{N_l}{\alpha_{i}h_i}) + b_l,
\end{equation}
where $h_i$ is the corresponding hidden state of $i$-th word, $\alpha$ is the word-level attention weights derived from: $\alpha_i = softmax(fc(h_i))$. $W_l$ and $b_l$ are learnable parameters. $r_l$ is projected to the feature visual feature space with $C_v$ dimensions.
he self-guided attention fusion introduces a flexible way for highlighting keywords in a sentence and reduce the negative impact caused by sentence truncation or padding.
\subsubsection{Visual Feature Encoder. }\label{section:VE}
As illustrated in Fig.~\ref{fig:overall}(c), our network takes a video clip $\mathcal{V}=\{f_i\}_{T}$ as input.
For each frame, the multi-level visual features $\{x_i\}$ are extracted with a CNN backbone (et al.}{\emph{et al}\onedot{e.g.}, ResNet-50 \cite{he2016deep}) and respectively fused with an 8-D spatial coordinate feature $p_i \in \mathbb{R}^{H_i \times W_i \times 8}$ ($i \in \{2,3,4,5\}$), where $H_i$ and $W_i$ are the height and width of the visual features.
We collect the spatial augmenting operations as a Spatial Encoding Module (SEM) and illustrates it in Fig.~\ref{fig:overall}(a). It is designed for remedying the weak positional sensation of spatial-agnostic CNN-extracted visual feature.
Formally, the spatial enhanced visual feature $X_i \in \mathbb{R}^{H_i \times W_i \times C_v^i}$ at level $i$ can be calculated from:
\begin{equation}
X_i = x_i + W_p(p_i),
\end{equation}
where $W_p$ is a learnable matrix and $C_v^i$ is the channel dimension.
$\{X_i\}$ are then enhanced with multi-scale information by merging with features from a top-down pathway via simple upsampling and addition, like in \cite{kirillov2019panoptic}.
Multi-level features are then concatenated to obtain final feature map, which would be used for task-specific mask prediction and contrastive learning ($X_v$ in Sec.~\ref{section:ObjCons}).
\setlength{\tabcolsep}{5pt}
\begin{table*}
\small
\begin{center}
\resizebox{1.\textwidth}{!}{
\setlength\tabcolsep{8pt}
\renewcommand\arraystretch{1.0}
\begin{tabular}{l|c|ccccc|c|cc}
\hline
\multirow{2}{*}{\textbf{Methods}} & \multirow{2}{*}{\textbf{Venue}} & \multicolumn{5}{c|}{\textbf{Overlap}} & \textbf{mAP} & \multicolumn{2}{|c}{\textbf{IoU}} \\
& & P@0.5 & P@0.6 & P@0.7 & P@0.8 & P@0.9 & 0.5:0.95 & Overall & Mean \\
\shline
\cite{hu2016segmentation} & ECCV16 & 34.8 & 23.6 & 13.3 & 3.3 & 0.1 & 13.2 & 47.4 & 35.0 \\
\cite{li2017tracking} & CVPR17 & 38.7 & 29.0 & 17.5 & 6.6 & 0.1 & 16.3 & 51.5 & 35.4 \\
\cite{gavrilyuk2018actor} & CVPR18 & 53.8 & 43.7 & 31.8 & 17.1 & 2.1 & 26.9 & 57.4 & 48.1 \\
\cite{wang2019asymmetric} & ICCV19 & 55.7 & 45.9 & 31.9 & 16.0 & 2.0 & 27.4 & 60.1 & 49.0 \\
\cite{mcintosh2020visual} & CVPR20 & 52.6 & 45.0 & 34.5 & 20.7 & 3.6 & 30.3 & 56.8 & 46.0 \\
\cite{ningpolar} & IJCAI20 & 63.4 & 57.9 & 48.3 & 32.2 & 8.3 & 38.8 & 66.1 & 52.9 \\
\cite{wang2020context} & AAAI20 & 60.7 & 52.5 & 40.5 & 23.5 & 4.5 & 33.3 & 62.3 & 53.1 \\
\cite{hui2021collaborative} & CVPR21 & \textbf{65.4} & \textbf{58.9} & 49.7 & 33.3 & 9.1 & 39.9 & 66.2 & 56.1 \\
\hline
\textbf{Ours} & - & 64.2 & 58.7 & \textbf{50.5} & \textbf{35.8} & \textbf{11.6} & \textbf{40.8} & \textbf{68.3} & \textbf{57.1} \\
\hline
\end{tabular}
}
\end{center}
\captionsetup{font=small}
\caption{\small
Comparison with the state-of-the-art methods on the A2D Sentences \texttt{valid} using et al.}{\emph{et al}\onedot{IoU} and et al.}{\emph{et al}\onedot{Precision@K} as metrics.
}
\label{table:a2dsota}
\end{table*}
\setlength{\tabcolsep}{5pt}
\begin{table*}
\small
\begin{center}
\resizebox{1.\textwidth}{!}{
\setlength\tabcolsep{8pt}
\renewcommand\arraystretch{1.0}
\begin{tabular}{l|c|ccccc|c|cc}
\hline
\multirow{2}{*}{\textbf{Methods}} & \multirow{2}{*}{\textbf{Venue}} & \multicolumn{5}{c|}{\textbf{Overlap}} & \textbf{mAP} & \multicolumn{2}{|c}{\textbf{IoU}} \\
& & P@0.5 & P@0.6 & P@0.7 & P@0.8 & P@0.9 & 0.5:0.95 & Overall & Mean \\
\shline
\cite{hu2016segmentation} & ECCV16 & 63.3 & 35.0 & 8.5 & 0.2 & 0.0 & 17.8 & 54.6 & 52.8 \\
\cite{li2017tracking} & CVPR17 & 57.8 & 33.5 & 10.3 & 0.6 & 0.0 & 17.3 & 52.9 & 49.1 \\
\cite{gavrilyuk2018actor} & CVPR18 & 71.2 & 51.8 & 26.4 & 3.0 & 0.0 & 26.7 & 55.5 & 57.0 \\
\cite{wang2019asymmetric} & ICCV19 & 75.6 & 56.4 & 28.7 & 3.4 & 0.0 & 28.9 & 57.6 & 58.4 \\
\cite{mcintosh2020visual} & CVPR20 & 67.7 & 51.3 & 28.3 & 5.1 & 0.0 & 26.1 & 53.5 & 55.0 \\
\cite{ningpolar} & IJCAI20 & 69.1 & 57.2 & 31.9 & 6.0 & 0.1 & 29.4 & - & - \\
\cite{wang2020context} & AAAI20 & 74.2 & 58.7 & 31.6 & 4.7 & 0.0 & 30.1 & 55.4 & 57.6 \\
\cite{hui2021collaborative} & CVPR21 & 78.3 & 63.9 & 37.8 & 7.6 & 0.0 & 33.5 & 59.8 & 60.4 \\
\hline
\textbf{Ours} & - & \textbf{86.0} & \textbf{72.7} & \textbf{44.8} & \textbf{9.7} & \textbf{0.1} & \textbf{39.2} & \textbf{66.1} & \textbf{64.3} \\
\hline
\end{tabular}
}
\end{center}
\captionsetup{font=small}
\caption{\small
Comparison with the state-of-the-art methods on J-HMDB Sentences \texttt{test} using et al.}{\emph{et al}\onedot{IoU} and et al.}{\emph{et al}\onedot{Precision@K} as metrics.
}
\label{table:jhmdbsota}
\end{table*}
\begin{table}
\small
\begin{center}
\resizebox{0.46\textwidth}{!}{
\setlength\tabcolsep{8pt}
\renewcommand\arraystretch{1.1}
\begin{tabular}{ccc|c|cc}
\hline
\multicolumn{3}{c|}{\textbf{Modules}} & \textbf{mAP} & \multicolumn{2}{|c}{\textbf{IoU}} \\
CCL & RHIC & LCF & 0.5:0.95 & Overall & Mean \\
\shline
& & & 26.6 & 56.9 & 47.1 \\
$\surd$ & & & 34.4 & 64.2 & 52.7 \\
$\surd$ & $\surd$ & & 36.2 & 65.6 & 53.3 \\
$\surd$ & & $\surd$ & 39.7 & 67.5 & 56.8 \\
$\surd$ & $\surd$ & $\surd$ & \textbf{40.8} & \textbf{68.3} & \textbf{57.1} \\
\hline
\end{tabular}
}
\end{center}
\captionsetup{font=small}
\caption{\small
\textbf{Component analysis} on A2D Sentences \texttt{valid}.
}
\label{table:ablation}
\end{table}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=.86\linewidth]{images/quantight.pdf}
\end{center}
\captionsetup{font=small}
\caption{\small
\textbf{Qualitative results} of VLS on A2D Sentences.
Predicted result is represented by the same color with the corresponding sentence.
Cross-modal Hard Example Mining is abbreviated as ``CHEM".
(a) Original image. (b)-(d) are predicted results by:
(b) Baseline model, et al.}{\emph{et al}\onedot{i.e.}, implicit fusion (row 1 in Table \ref{table:ablation});
(c) CVLS et al.}{\emph{et al}\onedot{w/o} CHEM (row 2 in Table \ref{table:ablation});
(d) Full CVLS (row 5 in Table \ref{table:ablation}).
(e) Ground truth.
}\label{fig:qualiti}
\end{figure*}
\section{Experiments}\label{sec:exp}
\subsection{Datasets and Evaluation Criteria}
All experiments in this paper are conducted on two extended datasets: \textbf{A2D Sentences} and \textbf{J-HMDB Sentences}. These datasets are released in \cite{gavrilyuk2018actor} by additionally providing natural descriptions on original A2D \cite{xu2015can} and J-HMDB \cite{jhuang2013towards} respectively.
\noindent\textbf{A2D Sentences} contains 3782 videos in total with 8 action classes performed by 7 actor classes. Each video in A2D has 3 to 5 frames annotated with pixel-level actor-action segmentation masks. Besides, it contains 6,655 sentences corresponding to actors and their actions. Following the settings in \cite{wang2019asymmetric}, we split the whole dataset into 3017 training videos, 737 testing videos, and 28 unlabeled videos.
\noindent\textbf{J-HMDB Sentences} contains 928 short videos with 928 corresponding sentences describing 21 different action classes. Pixel-wise 2D articulated human puppet masks are provided for evaluating segmentation performance.
We evaluate our proposed method with the criteria of Intersection-over-Union (IoU) and precision. Overall IoU is the ratio of the total intersection area divided by the total union area over testing samples. The mean IoU is the averaged IoU over all samples, which may not be affected by the size of samples. We also measure precision@K which considers the percentage of testing samples whose IoU scores are higher than threshold K and calculate mean average precision over 0.50:0.05:0.95 \cite{gavrilyuk2018actor}.
\subsection{Implementation Details}
We adopt ResNet50 \cite{he2016deep} model pre-trained on ImageNet \cite{deng2009imagenet} as visual backbone and use the output of Res2, Res3, Res4 and Res5 for multi-level feature construction. And we instantiate mask decoder with three stacked $3\times3$ convolution layers for decoding followed by one $1\times1$ convolutional layer for outputting the final segmentation mask.
A bi-LSTM~\cite{huang2015bidirectional} module is utilized as text encoder. All input frames are resized to $320 \times 320$.
Following the settings in \cite{gavrilyuk2018actor}, the maximum length of sentences is set to 20 and the dimension of word vector is 1000. The word vectors are initialized with one-hot vectors without any pre-trained weights applied. The hidden states of bi-LSTM \cite{huang2015bidirectional} are encoded as sentence features with a dimension of 2000.
Training is done with Adam optimizer \cite{kingma2014adam} with an initial learning rate of $0.0002$. We employ a scheduler that waits for $2$ epochs after loss stagnation to reduce the learning rate by a factor of $10$ and the batch size is set to $8$ by default.
During inference, following the setting in \cite{wang2020context}, we take a pixel as foreground when its value is higher than $\beta$ of the maximum value in probability map. And $\beta$ is set to $0.8$ in our implementation.
\subsection{Comparison with State-of-the-Art Methods}
Following the standard settings, we compare our CVLS with other state-of-the-art video-language segmentation models published in recent years on two datasets: A2D Sentences and J-HMDB Sentences. The comparison results are demonstrated in Table \ref{table:a2dsota} and Table \ref{table:jhmdbsota}.
First on A2D Sentences, baseline methods \cite{hu2016segmentation,li2017tracking} are pre-trained on ReferIt dataset \cite{kazemzadeh2014referitgame} and then fine-tuned on A2D sentences. Other methods including ours are trained on A2D Sentences exclusively.
As shown in Table \ref{table:a2dsota}, we bring 1.2$\%$ improvement on Mean IoU and 2.1$\%$ on Overall IoU over SOTA respectively, which directly proves the effectiveness of our method.
The state-of-the-art performance has been achieved on most metrics especially at higher IoU thresholds, et al.}{\emph{et al}\onedot{i.e.}, $P@0.8$ and $P@0.9$.
On $P@0.8$, our method outperforms the SOTA by a margin of 2.5$\%$. On $P@0.9$, 2.5$\%$ absolute improvement is achieved with a 27$\%$ relative improvement.
Then, we conduct experiments on the whole J-HMDB Sentences to evaluate the generalization ability of our CVLS. As a default setting, the best model trained on A2D sentences is employed without any additional fine-tuning.
With finer cross-modal interaction, our CVLS significantly outperforms previous state-of-the-art methods on all metrics considered.
\subsection{Ablation Study}
\subsubsection{Component Analysis. }
\noindentet al.}{\emph{et al}\onedot{Settings:}
We conduct extensive component analysis to verify the effectiveness and reveal the internal mechanism of each component in CVLS. We summarize ablation results of each proposed module in Table \ref{table:ablation}.
Baseline, in line1, ignores the fine-grained cross-modal interaction and directly predicts segmentation mask with implicit concatenation-convolution~\cite{shi2018key,hu2016segmentation} fusion.
In other lines, the explicitly cross-modal interaction is involved and the network with different components is supervised under a combination of cross-modal contrastive loss and segmentation loss.
\noindentet al.}{\emph{et al}\onedot{Observations:}
All ablations are conducted on the validation set of A2D Sentences.
We could obtain the following observations:
\textbf{(1)} Due to the redundant uninformative visual-distinguishable feature, the vanilla CCL merely brings limited improvement compared with baseline.
\textbf{(2)} Both of the hard mining strategies can significantly improve the vanilla CCL, which validates the effects of the design.
\textbf{(3)} Conclusively, these results confirm the merits of the explicitly cross-modal interaction formulation again.
\begin{figure}[t]
\begin{center}
\includegraphics[width=.98\linewidth]{new_images/sensitive.pdf}
\end{center}
\captionsetup{font=small}
\caption{\small
\textbf{Variation of model performance} with different selecting ratios in RHIC (Sec.~\ref{section:rhic}) and different objective weighting $\lambda$.
}\label{fig:sensitive}
\end{figure}
\subsubsection{Relative ratio of et al.}{\emph{et al}\onedot{hard} samples in RHIC. }
We supply an additional experiment on ranges of relative ratio of RHIC in Fig.\ref{fig:sensitive}~(a). We modify relative ratio of hard to easy pixels from 1:2 to 1:5, and final performance of CVLS is stable ($\pm$0.15$\%$ in terms of mAP) but extra epochs are needed for convergence. While the ratio is ranged from 1:1 to 4:1, final performance progressively drops $\sim$0.9$\%$, as RHIC is progressively losing efficacy which causes model degradation. We empirically set the ratio to 1:3 for balancing the performance and convergence speed.
\subsubsection{Parameter sensitivity analysis of $\lambda$. } As we employ a joint loss of segmentation term and contrastive term, we further evaluate different values of the weighting parameter $\lambda$ (Sec.~\ref{section:ObjCons}) and summarize the results
in Fig.~\ref{fig:sensitive}~(b). In practice, we find the proposed approach is robust to parameter variation and the performance is slightly jittered among different settings. We finally set $\lambda$ to 0.8 empirically.
\subsection{Qualitative Analysis}
In this section, we investigate the internal mechanism in CVLS by analyzing qualitative results.
The comparisons between the full model and alternative structures are illustrated in Fig.~\ref{fig:qualiti}.
Compared with the implicit fusion baseline (Fig.~\ref{fig:qualiti}(b)), even without hard mining strategies, the vanilla cross-modal contrastive learning pipeline could still capture finer cross-modal interaction and shows reasonable segmentation results.
Due to the model degradation caused by extra visual-distinguishable feature, the model still messes up the semantically similar region, et al.}{\emph{et al}\onedot{i.e.}, two guys with red and green masks respectively (the third line in Fig.~\ref{fig:qualiti}(c)). While in Fig.~\ref{fig:qualiti}(d), the problem is greatly alleviated.
Conclusively, these visualized results reconfirm the effectiveness of CVLS and the hard mining strategies which facilitates the entire method in both implicit (LCF) and explicit (RHIC) aspects.
\section{Conclusion}
In this paper, we propose a highly effective pipeline with a novel cross-modal contrastive learning objective
for the first attempt of introducing the explicit cross-modal interaction into video-language segmentation field.
To prevent the model from degradation to single-modal, we further employ two cross-modal hard mining strategies to exclude the over-simple visual-distinguishable feature from both implicit and explicit aspects, et al.}{\emph{et al}\onedot{i.e.}, Language-relevant Channel Filter and Relative Hard Instance Construction.
Evaluations on commonly used benchmarks demonstrate that CVLS surpasses all the state-of-the-art methods by large margins.
\section{Introduction} \label{sec:introduction}
Video-language segmentation (VLS) aims to predict a set of segmentation masks for the referent in a video under the guidance of a natural language description. This practical problem is fundamental to many applications such as interactive video editing, object tracking and human-machine interaction~\cite{wang2021survey}. Unlike the general semantic segmentation tasks that is operated over a fixed set of pre-defined categories, VLS requires fine-grained joint modeling for vision and language, yielding a more challenging scenario due to the great diversity of linguistic concepts and extra semantics in temporal domain.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.98\linewidth]{images/Moti.pdf}
\end{center}
\captionsetup{font=small}
\caption{\small
\textbf{Illustration of the video-language segmentation task.}
Given a linguistic expression, the predictions of state-of-the-art method \cite{wang2019asymmetric} and CVLS are shown in (a) and (b) respectively.
Previous methods (a) mainly rely on the implicit fusion to formulate cross-modal interaction and fail to distinguish semantically similar objects while CVLS (b) does the opposite.
}\label{fig:motivation}
\end{figure}
A commonplace solution in prior work is first extracting visual and linguistic features respectively, followed by a fusion process to implicitly interwind cross-modal features, et al.}{\emph{et al}\onedot{e.g.}, straightforward concatenation \cite{shi2018key,hu2016segmentation}, cross-modal attention-mechanism \cite{wang2019asymmetric,ningpolar,hui2021collaborative} or dynamic filter \cite{wang2020context,gavrilyuk2018actor}, and then conducting segmentation based on the fused feature. However, these solutions rarely consider the crucial fine-grained alignment of the cross-modal information hence prone to fail in distinguishing the semantically similar objects. As in Fig.~\ref{fig:motivation}, the model with implicit fusion strategy fails to capture the contextual difference among multiple similar regions, et al.}{\emph{et al}\onedot{i.e.}. person conducting different actions.
In this paper, we focus on explicitly modeling the vision-language interactions for the VLS problem. Our approach is inspired by the contrastive representation learning theory~\cite{wu2018unsupervised,chen2020simple,wang2021exploring} whose fundamental purpose is to capture similarities between examples of the same class and contrast them with examples from other classes. Dubbed Contrastive Video-Language Segmentation (CVLS), our method presents a cross-modal contrastive objective to align the referred visual object with the linguistic description in representation space while contracting the unreferred video information with the description at the same time. In this way, fine-grained vision-language interactions can be explicitly learned thus boosting the representation discriminability among visually similar objects based on a linguistic description.
Although vanilla contrastive representation learning provides a potentially feasible solution for the vision-language interactions, we find it problematic to apply the Na\"ive paradigm mechanically on the raw cross-modal representations.
Upon reviewing the VLS task, we notice that the background areas occupy a large proportion across frames while the visually similar candidates do the opposite.
Since the background areas can be easily excluded from the candidate referent just using the visual features while the visually similar objects can be only distinguished on condition of a linguistic feature\footnote{For convenience, we use ``visual-distinguishable region'' to refer the former while use ``linguistic-distinguishable object'' to refer the latter hereafter.}.
For example, as shown in Fig.~\ref{fig:motivation}, models can easily distinguish foreground objects (person) apart from a large area of background even without a linguistic guidance.
A trivial construction of instance-level positive-negative pairs would introduce excessive uninformative visual samples that are over-easy for the current linguistic embedding to discriminate, which might directly degrade the cross-modal contrastive objective to a single-modal feature classifier, et al.}{\emph{et al}\onedot{i.e.}, a visual saliency detector.
To address this issue, we present two simple yet effective hard instance mining strategies, et al.}{\emph{et al}\onedot{i.e.}, Language-relevant Channel Filter (LCF) and Relative Hard Instance Construction (RHIC), which are specially designed for the cross-modal scenario.
They encourage the network to exclude uninformative visual-distinguishable features and to focus on linguistic-distinguishable objects during the contrastive training.
On one hand, with the guidance of linguistic description, LCF module generates channel-wise filters for visual features to highlight the linguistic-distinguishable features.
On the other hand, in RHIC module, only those misclassified pixels with higher confidence are considered while positive-negative pair sampling, to avoid gradient from being dominated by samples from visual-distinguishable regions.
Main contributions of our work are as follows:
\begin{itemize}
\item We propose CVLS, a cross-modal contrastive learning-based method for VLS task, which explicitly captures the commondality of vision-language representations thus boosting representation discriminability among visually similar objects on condition of linguistic descriptions.
\item We propose two simple yet efective hard mining strategies tailored for cross-modal contrastive objective. These strategies are able to filter out over-easy video content from visual features, which enriches the informative gains for contrastive learning.
\item The proposed method significantly outperforms SOTA methods in most metrics on two popular video-language segmentation benchmarks, et al.}{\emph{et al}\onedot{i.e.}, A2D, and J-HMDB.
\end{itemize}
\section{Related Work}
\subsection{Video-Language Segmentation}
Towards understanding fine-grained video representation, Xu et al.}{\emph{et al}\onedot{et al.}~\cite{xu2015can} introduce a task of actor and action video segmentation with the Actor-Action Dataset (A2D) containing a fixed set of actor-action pairs and pixel-level annotations.
Later,
Gavrilyuk et al.}{\emph{et al}\onedot{et al.}~\cite{gavrilyuk2018actor} extend Actor-Action Dataset (A2D) with human-annotated sentences and introduce the challenging text-based video segmentation task that requires further comprehension of both vision and language modalities.
They adopt dynamic convolution filters generated by language to align the multi-modal feature.
Then, Wang et al.}{\emph{et al}\onedot{et al.}~\cite{wang2020context} extend vanilla dynamic convolution with vision context to generate spatial-relevant kernels. Ning et al.}{\emph{et al}\onedot{et al.}~\cite{ningpolar} embed spatial expressions into attention module in terms of direction and range for better linguistic spatial formulation. Wang et al.}{\emph{et al}\onedot{et al.}~\cite{wang2019asymmetric} utilize asymmetric attention mechanisms to facilitate visual guided linguistic feature learning. Hui et al.}{\emph{et al}\onedot{et al.}~\cite{hui2021collaborative} further introduce an additional 2D spatial encoder to alleviate the spatial information misaligned problem brought by 3D CNNs.
Mcintosh et al.}{\emph{et al}\onedot{et al.}~\cite{mcintosh2020visual} introduce a capsule-based approach for better capturing the relationship between multi-modal features. To further mining continuous temporal information, they extend the A2D dataset with annotations for all frames.
Even with certain achievement, existing methods mainly suffers from coarse cross-modal interaction, et al.}{\emph{et al}\onedot{i.e.}, implicit multi-modal fusion, which directly leads to ambiguous boundaries for semantically similar objects.
In this paper, we address this limitation by constructing instance-level multi-modal interaction in an explicitly contrastive manner.
\subsection{Multi-modal Contrastive Learning}
Contrastive Learning techniques learn representations in a discriminative manner by pulling positive pairs closer against negative pairs.
Recently, self-supervised visual representation learning surges and earns a great success either by partial data prediction \cite{liu2020pic,doersch2015unsupervised,dosovitskiy2015discriminative}, or instance-level training image discrimination \cite{chen2020simple,he2020momentum,tian2019contrastive,wu2018unsupervised}. These methods mainly rely on instance augmentation to form positive pairs.
Beyond augmentation-based approach, multi-modal contrastive learning leverages multiple modalities of the same instance to construct positive pairs. And cross-modal representations are learned during the procedure.
Examples of commonly used data modalities include appearance of image, motion of video \cite{harley2019learning}, depth \cite{tian2019contrastive}, luminance and chrominance \cite{zhang2017split}, audio \cite{wu2019DAM,wu2021explore}, and text \cite{zhang2020contrastive,gupta2020contrastive} as in our work.
Zhang et al.}{\emph{et al}\onedot{et al.}~\cite{zhang2020contrastive} and Gupta et al.}{\emph{et al}\onedot{et al.}~\cite{gupta2020contrastive} both design sophisticated strategies to align vision region with expression and maximize the mutual information between the modalities. In this work, inspired by such great progress, we propose a specially-designed cross-modal contrastive learning objective for the first attempt of introducing explicit cross-modal interaction into Video-Language Segmentation.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.9\linewidth]{images/pipeline.pdf}
\end{center}
\captionsetup{font=small}
\caption{\small
\textbf{Illustration of Contrastive Cross-modal Learning Module (CCLM)}. Visual encoder and linguistic encoder are applied to extract object-level visual features $\{r_p, r_{n_1}^1, \cdots, r_{n_2}^3\}$ and reference feature $r_l$ respectively, which are then projected to a $128$-dimensional space and $L2$ normalized.
$r_p$ is the corresponding referent feature.
$r_{n_i}^t$ denotes negative instances in the video with time step $t$, including background (grey region).
Optimal representations are learned via cross-modal discrimination, where the referred object and the language description are pulled closer and the distracting contents are pushed apart across frames.
Language-relevant Channel Filter and Relative Hard Instance Construction are abbreviated as ``LCF" and ``RHIC" (Sec.~\ref{section:LCF}) respectively.
}\label{fig:pipeline}
\end{figure*}
\section{Method}
In this section, we first present the basic Cross-modal Conrastive Learning procedure in Sec.~\ref{section:ObjCons}. To further alleviate degradation problem, we then introduce two specially-designed cross-modal hard mining approaches in Sec.~\ref{section:HM}. Even not being restricted to any certain feature extractor, in Sec.~\ref{section:overall}, we finally describe the overall architecture implemented in this work.
\subsection{Na\"ive Instance-level Cross-modal Contrastive Representation Learning (CCLM)} \label{section:ObjCons}
As illustrated in Fig.~\ref{fig:pipeline}, with a pair of reference (natural language description) and corresponding referent (visual object), our goal is to learn a good representation space that the positive pairs would be pulled together while other objects in a video would be pushed away at mean time.
\subsubsection{Within-modal representation construction.}
Different from the prevailing segmentation models that formulate visual embeddings in grid-level as in \cite{zhao2020contrastive}, in VLS task we argue that it is more reasonable and efficient to formulate within-modal embeddings and construct cross-modal commonalities in instance-level concerning that natural language expressions are mostly annotated in an object view.
Now the problem lies in how to construct within-modal representations for the reference and objects.
For linguistic modality, reference feature $r_l$ can be directly extracted by a sentence-level linguistic encoder (Detailed later in Sec.~\ref{section:LE}).
For visual modality, instance-level feature is obtained by conducting average-pooling on mask-cropped visual feature map extracted from visual encoder (Sec.~\ref{section:VE}).
Formally, the naive process for obtaining instance representation $r_v^t$ at time step $t$ could be defined as:
\begin{equation}
r_v^t = AVG(X^t_v \odot o^t). \label{equation:OT}
\end{equation}
Here we denote $X^t_v$ and $o^t$ as the extracted visual feature and a binary object mask.
$\odot$ and $AVG$ correspond to element-wise multiplication and average pooling respectively.
We treat the entire video clip as a pool of object features $\{r_{v_1}^1,r_{v_1}^2,\cdots,r_{v_{N_v}}^T,\}$, where $N_v$ is the total number of instances in a video clip\footnote{Following definitions in literature~\cite{lin2021video,wang2021end}, we use the term et al.}{\emph{et al}\onedot{instance} for video-level identification and use et al.}{\emph{et al}\onedot{object} for frame-level, et al.}{\emph{et al}\onedot{e.g.}, there are $2$ instances and $2T$ objects in the example video clip (Shown in Fig.~\ref{fig:pipeline})} with $T$ frames.
For ease of notation, we define the positive referent object corresponding to each linguistic reference $r_l$ as $r_p$ and all other negative visual objects as $r_{n_i}^t$, where $i \in \{1,\cdots,N_v-1\}$ and $t \in \{1,\cdots,T\}$.
Correspondingly, $o_p$ is the referent region and $o_{n_i}^t$ is the negative object region. Both of them can be obtained from ground-truth masks during training. In later sections, we omit time stamp $t$ for ease of notation.
It should be noted that, background region (grey region in Fig.~\ref{fig:pipeline}) is also considered as a negative instance and counted in $N_v$.
Within-modal features in the pool are further fed into a projection network, $Proj(\cdot)$, mapped to a same cross-modal representation space as in Fig.~\ref{fig:pipeline}, et al.}{\emph{et al}\onedot{i.e.}, $z_p = Proj(r_p)$, $z_{n_i} = Proj(r_{n_i})$, $z_l = Proj(r_l)$.
We instantiate $Proj(\cdot)$ as a multi-layer perceptron as in~\cite{chen2020simple} with a single hidden layer of size $512$ and output vector of size $128$.
The outputs are then normalized to lie on an unit hypersphere, which enables using an inner product to measure distances in the projection space, et al.}{\emph{et al}\onedot{i.e.}, $||z_p||=||z_{n_i}||=||z_l||=1$.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.98\linewidth]{new_images/pipelinelatest.pdf}
\end{center}
\captionsetup{font=small}
\caption{\small
\textbf{Detailed architecture of CVLS pipeline.}
Visual feature and linguistic feature are extracted by a FPN-like encoder and a bi-LSTM \cite{huang2015bidirectional} respectively \textbf{(c)}.
Language-relevant Channel Filter (\textbf{LCF}) (Sec.~\ref{section:LCF}) is integrated in visual encoder to exclude the over-simple visual-distinguishable feature.
Multi-scale output of the visual encoder is then concatenated as $X_v$ and fed into the Relative Hard Instance Construction (\textbf{RHIC}) (Sec.~\ref{section:LCF}) to aggregate instance-level visual features.
``\textbf{CCLM}" denotes the Cross-modal Contrastive Learning Module (Sec.~\ref{section:ObjCons}) as in Fig.~\ref{fig:pipeline}.
Detailed process of \textbf{(a)} Spatial Encoding Module (\textbf{SEM}) and \textbf{(b)} Language-relevant Channel Filter \textbf{(LCF)}.
}\label{fig:overall}
\end{figure*}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.95\linewidth]{images/RHIC.pdf}
\end{center}
\captionsetup{font=small}
\caption{\small
\textbf{Illustration of region-of-interests} while aggregating visual instance representation between Naive Instance Construction (Naive IC) and Relative Hard Instance Construction (RHIC).
Instead of pooling all the pixels, RHIC only considers relatively harder region of each instance, based on the confidence score of each pixel.
The positive referent and negative candidate are painted in green and red respectively.
Background region is also considered as a negative instance and is painted in blue in both figures.
}\label{fig:rhic}
\end{figure}
\subsubsection{ Instance-level Cross-Modal Contrast.}
With constructed instance-level features, the representation space is then shaped with the instance-level cross-modal contrastive objective $\mathcal{L}_c$, which optimizes the triple representations of the positive referent, the negative objects and the language description in the latent space:
\begin{equation}
\mathcal{L}_c = - \log \frac{\exp({z_p} \cdot z_l/\tau)}{\sum_{j} \exp({z_{n_j}} \cdot z_l/\tau) + \exp({z_p} \cdot z_l/\tau)},
\end{equation}
\noindent
where $\tau$ is a scalar temperature parameter, and a lower temperature value would increase the influence of harder negatives \cite{wu2018unsupervised,gunel2020supervised}.
The cross-modal contrastive objective regularizes the embedding space and force the network to explore explicit cross-modal discrimination among semantically similar objects.
\subsubsection{Overall Training Objective.}
During contrastive training, our method is supervised under both task-specific segmentation loss $\mathcal{L}_s$ and instance-level cross-modal contrastive loss $\mathcal{L}_c$ defined in previous section.
$\mathcal{L}_s$ optimizes the segmentation mask based on the
extracted multi-modal tensor:
\begin{equation}
\mathcal{L}_s = - s \odot \log (\sigma(e)) - (1-s) \odot \log(1-\sigma(e)),
\end{equation}
where the $\cdot$ symbol denotes the inner (dot) product, $\odot$ is element-wise multiplication. And $e$ is the output of mask decoder and $s$ denotes the ground truth.
And overall training target of our method is a weighted average of $\mathcal{L}_s$ and $\mathcal{L}_c$:
\begin{equation}
\mathcal{L} = \mathcal{L}_s + \lambda \mathcal{L}_c,
\end{equation}
\noindent
where $\lambda$ is a weighting hyperparameter, and $\lambda \in [0.1,1]$.
\subsection{Cross-Modal Hard Example Mining}\label{section:HM}
Though vanilla contrastive learning provides an explicit standpoint for modeling the fine-grained vision-language interaction, as we described in Sec.\ref{sec:introduction}, equally treat the visual-distinguishable regions and the linguistic-distinguishable objects (et al.}{\emph{et al}\onedot{e.g.}, average pooling) might directly degrade the cross-modal contrastive objective to a single-modal feature classifier.
We thus propose two hard mining strategies tailored for cross-modal scenario, aiming to reduce the impact of excessive uninformative visual samples and enlarge the informative gains from both implicit and explicit aspects.
\subsubsection{Language-relevant Channel Filter (LCF). } \label{section:LCF}
As stated in~\cite{hu2018squeeze,chen2017sca,zeiler2014visualizing}, channel-wise features intrinsically encode the responses of different convolutional filters, and the visual concepts can be integral with different combination of channel responses.
Following this insight, we employ natural language as a filter to enhance language-relevant visual features and depress the language-irrelevant feature before contrastive training by applying dynamically conditioned weights on channels.
Specifically, we design a simple gating mechanism with a sigmoid activation as a filter on each channel.
With the linguistic feature $r_l$ and visual feature map $X$, the filter can be employed through:
\begin{equation}
\begin{aligned}
g &= \sigma(W_2\ \delta ({W_1}{r_l} + b_1)+b_2), \\
X_v &= g \odot X,
\end{aligned}
\end{equation}
where $\delta(\cdot)$ refers to the LeakyRELU operation \cite{xu2015empirical}. $W_i$ and $b_i$ are learnable matrix and bias.
Rather than enforcing the one-hot activation, we employ sigmod function $\sigma(\cdot)$ to normalize the excitation range to $(0,1)$, which enables a non-mutually-exclusive relationship among all channels. In this way, the impact of the visual-distinguishable feature declines, which implicitly forces the network to focus on linguistic-distinguishable content during the contrastive training.
\subsubsection{Relative Hard Instance Construction (RHIC). } \label{section:rhic}
An explicit hard mining strategy is introduced to further reduce the impact of uninformative visual regions. Dubbed Relative Hard Instance Construction (RHIC), it explicitly selects a portion of the entire object as et al.}{\emph{et al}\onedot{hard} regions for visual object representation construction.
Specifically, during instance construction, only the misclassified pixels with high confidence scores are considered as et al.}{\emph{et al}\onedot{hard} regions, which are sampled with a relative ratio (1:3) of hard to easy pixels.
Formally, given the model’s estimated probability $c$ predicted by segmentation decoder, we define misclassified degree $c_p$:
\begin{equation}
\begin{aligned}
c_p =
\begin{cases}
1-c & \text{if $y=1$} \\
c & \text{otherwise}.
\end{cases}
\end{aligned}
\end{equation}
In the above $y\!\in\!\{0,1\}$ specifies ground-truth label of each pixel.
The instance region in Equation \ref{equation:OT} can be updated by:
\begin{equation}
o^t = Top\_K(c_p) \odot o^t,
\end{equation}
where $Top\_K$ denotes the relative selecting process.
The visualization of the selected et al.}{\emph{et al}\onedot{hard} regions by RHIC is shown in Fig.~\ref{fig:rhic}(b).
It is worth-noted that we employ a et al.}{\emph{et al}\onedot{relative} selecting strategy here which introduces a flexible value of negative samples and $K$ corresponding to the total
pixel number of the object region and the prediction will be et al.}{\emph{et al}\onedot{progressively} improved by et al.}{\emph{et al}\onedot{online} selecting the relative-hard regions in each training step.
In practise, as occupation of different objects vary dramatically, we find a fixed threshold, et al.}{\emph{et al}\onedot{i.e.}, a fixed number of $K$ is hard to tune. And we conduct additional ablations on relative ratio in Sec.\ref{sec:exp}.
\subsection{Overall Architecture}\label{section:overall}
In this section we instantiate the detailed architecture design as shown in Fig.~\ref{fig:overall}. Nevertheless, the idea of CVLS is general in the sense that basic encoder and decoder can be easily replaced with other stronger baselines.
\subsubsection{Linguistic Feature Encoder. } \label{section:LE}
Given a linguistic sentence $S = \{s_i\}_{N_l}$ with $N_l$ words,
linguistic feature is encoded with a bi-LSTM \cite{huang2015bidirectional} module followed by a self-guided attention fusion.
Specifically, linguistic feature $r_l \in \mathbb{R}^{C_v}$ can be calculated from:
\begin{equation}
r_l = W_l(\sum_{i=1}^{N_l}{\alpha_{i}h_i}) + b_l,
\end{equation}
where $h_i$ is the corresponding hidden state of $i$-th word, $\alpha$ is the word-level attention weights derived from: $\alpha_i = softmax(fc(h_i))$. $W_l$ and $b_l$ are learnable parameters. $r_l$ is projected to the feature visual feature space with $C_v$ dimensions.
he self-guided attention fusion introduces a flexible way for highlighting keywords in a sentence and reduce the negative impact caused by sentence truncation or padding.
\subsubsection{Visual Feature Encoder. }\label{section:VE}
As illustrated in Fig.~\ref{fig:overall}(c), our network takes a video clip $\mathcal{V}=\{f_i\}_{T}$ as input.
For each frame, the multi-level visual features $\{x_i\}$ are extracted with a CNN backbone (et al.}{\emph{et al}\onedot{e.g.}, ResNet-50 \cite{he2016deep}) and respectively fused with an 8-D spatial coordinate feature $p_i \in \mathbb{R}^{H_i \times W_i \times 8}$ ($i \in \{2,3,4,5\}$), where $H_i$ and $W_i$ are the height and width of the visual features.
We collect the spatial augmenting operations as a Spatial Encoding Module (SEM) and illustrates it in Fig.~\ref{fig:overall}(a). It is designed for remedying the weak positional sensation of spatial-agnostic CNN-extracted visual feature.
Formally, the spatial enhanced visual feature $X_i \in \mathbb{R}^{H_i \times W_i \times C_v^i}$ at level $i$ can be calculated from:
\begin{equation}
X_i = x_i + W_p(p_i),
\end{equation}
where $W_p$ is a learnable matrix and $C_v^i$ is the channel dimension.
$\{X_i\}$ are then enhanced with multi-scale information by merging with features from a top-down pathway via simple upsampling and addition, like in \cite{kirillov2019panoptic}.
Multi-level features are then concatenated to obtain final feature map, which would be used for task-specific mask prediction and contrastive learning ($X_v$ in Sec.~\ref{section:ObjCons}).
\setlength{\tabcolsep}{5pt}
\begin{table*}
\small
\begin{center}
\resizebox{1.\textwidth}{!}{
\setlength\tabcolsep{8pt}
\renewcommand\arraystretch{1.0}
\begin{tabular}{l|c|ccccc|c|cc}
\hline
\multirow{2}{*}{\textbf{Methods}} & \multirow{2}{*}{\textbf{Venue}} & \multicolumn{5}{c|}{\textbf{Overlap}} & \textbf{mAP} & \multicolumn{2}{|c}{\textbf{IoU}} \\
& & P@0.5 & P@0.6 & P@0.7 & P@0.8 & P@0.9 & 0.5:0.95 & Overall & Mean \\
\shline
\cite{hu2016segmentation} & ECCV16 & 34.8 & 23.6 & 13.3 & 3.3 & 0.1 & 13.2 & 47.4 & 35.0 \\
\cite{li2017tracking} & CVPR17 & 38.7 & 29.0 & 17.5 & 6.6 & 0.1 & 16.3 & 51.5 & 35.4 \\
\cite{gavrilyuk2018actor} & CVPR18 & 53.8 & 43.7 & 31.8 & 17.1 & 2.1 & 26.9 & 57.4 & 48.1 \\
\cite{wang2019asymmetric} & ICCV19 & 55.7 & 45.9 & 31.9 & 16.0 & 2.0 & 27.4 & 60.1 & 49.0 \\
\cite{mcintosh2020visual} & CVPR20 & 52.6 & 45.0 & 34.5 & 20.7 & 3.6 & 30.3 & 56.8 & 46.0 \\
\cite{ningpolar} & IJCAI20 & 63.4 & 57.9 & 48.3 & 32.2 & 8.3 & 38.8 & 66.1 & 52.9 \\
\cite{wang2020context} & AAAI20 & 60.7 & 52.5 & 40.5 & 23.5 & 4.5 & 33.3 & 62.3 & 53.1 \\
\cite{hui2021collaborative} & CVPR21 & \textbf{65.4} & \textbf{58.9} & 49.7 & 33.3 & 9.1 & 39.9 & 66.2 & 56.1 \\
\hline
\textbf{Ours} & - & 64.2 & 58.7 & \textbf{50.5} & \textbf{35.8} & \textbf{11.6} & \textbf{40.8} & \textbf{68.3} & \textbf{57.1} \\
\hline
\end{tabular}
}
\end{center}
\captionsetup{font=small}
\caption{\small
Comparison with the state-of-the-art methods on the A2D Sentences \texttt{valid} using et al.}{\emph{et al}\onedot{IoU} and et al.}{\emph{et al}\onedot{Precision@K} as metrics.
}
\label{table:a2dsota}
\end{table*}
\setlength{\tabcolsep}{5pt}
\begin{table*}
\small
\begin{center}
\resizebox{1.\textwidth}{!}{
\setlength\tabcolsep{8pt}
\renewcommand\arraystretch{1.0}
\begin{tabular}{l|c|ccccc|c|cc}
\hline
\multirow{2}{*}{\textbf{Methods}} & \multirow{2}{*}{\textbf{Venue}} & \multicolumn{5}{c|}{\textbf{Overlap}} & \textbf{mAP} & \multicolumn{2}{|c}{\textbf{IoU}} \\
& & P@0.5 & P@0.6 & P@0.7 & P@0.8 & P@0.9 & 0.5:0.95 & Overall & Mean \\
\shline
\cite{hu2016segmentation} & ECCV16 & 63.3 & 35.0 & 8.5 & 0.2 & 0.0 & 17.8 & 54.6 & 52.8 \\
\cite{li2017tracking} & CVPR17 & 57.8 & 33.5 & 10.3 & 0.6 & 0.0 & 17.3 & 52.9 & 49.1 \\
\cite{gavrilyuk2018actor} & CVPR18 & 71.2 & 51.8 & 26.4 & 3.0 & 0.0 & 26.7 & 55.5 & 57.0 \\
\cite{wang2019asymmetric} & ICCV19 & 75.6 & 56.4 & 28.7 & 3.4 & 0.0 & 28.9 & 57.6 & 58.4 \\
\cite{mcintosh2020visual} & CVPR20 & 67.7 & 51.3 & 28.3 & 5.1 & 0.0 & 26.1 & 53.5 & 55.0 \\
\cite{ningpolar} & IJCAI20 & 69.1 & 57.2 & 31.9 & 6.0 & 0.1 & 29.4 & - & - \\
\cite{wang2020context} & AAAI20 & 74.2 & 58.7 & 31.6 & 4.7 & 0.0 & 30.1 & 55.4 & 57.6 \\
\cite{hui2021collaborative} & CVPR21 & 78.3 & 63.9 & 37.8 & 7.6 & 0.0 & 33.5 & 59.8 & 60.4 \\
\hline
\textbf{Ours} & - & \textbf{86.0} & \textbf{72.7} & \textbf{44.8} & \textbf{9.7} & \textbf{0.1} & \textbf{39.2} & \textbf{66.1} & \textbf{64.3} \\
\hline
\end{tabular}
}
\end{center}
\captionsetup{font=small}
\caption{\small
Comparison with the state-of-the-art methods on J-HMDB Sentences \texttt{test} using et al.}{\emph{et al}\onedot{IoU} and et al.}{\emph{et al}\onedot{Precision@K} as metrics.
}
\label{table:jhmdbsota}
\end{table*}
\begin{table}
\small
\begin{center}
\resizebox{0.46\textwidth}{!}{
\setlength\tabcolsep{8pt}
\renewcommand\arraystretch{1.1}
\begin{tabular}{ccc|c|cc}
\hline
\multicolumn{3}{c|}{\textbf{Modules}} & \textbf{mAP} & \multicolumn{2}{|c}{\textbf{IoU}} \\
CCL & RHIC & LCF & 0.5:0.95 & Overall & Mean \\
\shline
& & & 26.6 & 56.9 & 47.1 \\
$\surd$ & & & 34.4 & 64.2 & 52.7 \\
$\surd$ & $\surd$ & & 36.2 & 65.6 & 53.3 \\
$\surd$ & & $\surd$ & 39.7 & 67.5 & 56.8 \\
$\surd$ & $\surd$ & $\surd$ & \textbf{40.8} & \textbf{68.3} & \textbf{57.1} \\
\hline
\end{tabular}
}
\end{center}
\captionsetup{font=small}
\caption{\small
\textbf{Component analysis} on A2D Sentences \texttt{valid}.
}
\label{table:ablation}
\end{table}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=.86\linewidth]{images/quantight.pdf}
\end{center}
\captionsetup{font=small}
\caption{\small
\textbf{Qualitative results} of VLS on A2D Sentences.
Predicted result is represented by the same color with the corresponding sentence.
Cross-modal Hard Example Mining is abbreviated as ``CHEM".
(a) Original image. (b)-(d) are predicted results by:
(b) Baseline model, et al.}{\emph{et al}\onedot{i.e.}, implicit fusion (row 1 in Table \ref{table:ablation});
(c) CVLS et al.}{\emph{et al}\onedot{w/o} CHEM (row 2 in Table \ref{table:ablation});
(d) Full CVLS (row 5 in Table \ref{table:ablation}).
(e) Ground truth.
}\label{fig:qualiti}
\end{figure*}
\section{Experiments}\label{sec:exp}
\subsection{Datasets and Evaluation Criteria}
All experiments in this paper are conducted on two extended datasets: \textbf{A2D Sentences} and \textbf{J-HMDB Sentences}. These datasets are released in \cite{gavrilyuk2018actor} by additionally providing natural descriptions on original A2D \cite{xu2015can} and J-HMDB \cite{jhuang2013towards} respectively.
\noindent\textbf{A2D Sentences} contains 3782 videos in total with 8 action classes performed by 7 actor classes. Each video in A2D has 3 to 5 frames annotated with pixel-level actor-action segmentation masks. Besides, it contains 6,655 sentences corresponding to actors and their actions. Following the settings in \cite{wang2019asymmetric}, we split the whole dataset into 3017 training videos, 737 testing videos, and 28 unlabeled videos.
\noindent\textbf{J-HMDB Sentences} contains 928 short videos with 928 corresponding sentences describing 21 different action classes. Pixel-wise 2D articulated human puppet masks are provided for evaluating segmentation performance.
We evaluate our proposed method with the criteria of Intersection-over-Union (IoU) and precision. Overall IoU is the ratio of the total intersection area divided by the total union area over testing samples. The mean IoU is the averaged IoU over all samples, which may not be affected by the size of samples. We also measure precision@K which considers the percentage of testing samples whose IoU scores are higher than threshold K and calculate mean average precision over 0.50:0.05:0.95 \cite{gavrilyuk2018actor}.
\subsection{Implementation Details}
We adopt ResNet50 \cite{he2016deep} model pre-trained on ImageNet \cite{deng2009imagenet} as visual backbone and use the output of Res2, Res3, Res4 and Res5 for multi-level feature construction. And we instantiate mask decoder with three stacked $3\times3$ convolution layers for decoding followed by one $1\times1$ convolutional layer for outputting the final segmentation mask.
A bi-LSTM~\cite{huang2015bidirectional} module is utilized as text encoder. All input frames are resized to $320 \times 320$.
Following the settings in \cite{gavrilyuk2018actor}, the maximum length of sentences is set to 20 and the dimension of word vector is 1000. The word vectors are initialized with one-hot vectors without any pre-trained weights applied. The hidden states of bi-LSTM \cite{huang2015bidirectional} are encoded as sentence features with a dimension of 2000.
Training is done with Adam optimizer \cite{kingma2014adam} with an initial learning rate of $0.0002$. We employ a scheduler that waits for $2$ epochs after loss stagnation to reduce the learning rate by a factor of $10$ and the batch size is set to $8$ by default.
During inference, following the setting in \cite{wang2020context}, we take a pixel as foreground when its value is higher than $\beta$ of the maximum value in probability map. And $\beta$ is set to $0.8$ in our implementation.
\subsection{Comparison with State-of-the-Art Methods}
Following the standard settings, we compare our CVLS with other state-of-the-art video-language segmentation models published in recent years on two datasets: A2D Sentences and J-HMDB Sentences. The comparison results are demonstrated in Table \ref{table:a2dsota} and Table \ref{table:jhmdbsota}.
First on A2D Sentences, baseline methods \cite{hu2016segmentation,li2017tracking} are pre-trained on ReferIt dataset \cite{kazemzadeh2014referitgame} and then fine-tuned on A2D sentences. Other methods including ours are trained on A2D Sentences exclusively.
As shown in Table \ref{table:a2dsota}, we bring 1.2$\%$ improvement on Mean IoU and 2.1$\%$ on Overall IoU over SOTA respectively, which directly proves the effectiveness of our method.
The state-of-the-art performance has been achieved on most metrics especially at higher IoU thresholds, et al.}{\emph{et al}\onedot{i.e.}, $P@0.8$ and $P@0.9$.
On $P@0.8$, our method outperforms the SOTA by a margin of 2.5$\%$. On $P@0.9$, 2.5$\%$ absolute improvement is achieved with a 27$\%$ relative improvement.
Then, we conduct experiments on the whole J-HMDB Sentences to evaluate the generalization ability of our CVLS. As a default setting, the best model trained on A2D sentences is employed without any additional fine-tuning.
With finer cross-modal interaction, our CVLS significantly outperforms previous state-of-the-art methods on all metrics considered.
\subsection{Ablation Study}
\subsubsection{Component Analysis. }
\noindentet al.}{\emph{et al}\onedot{Settings:}
We conduct extensive component analysis to verify the effectiveness and reveal the internal mechanism of each component in CVLS. We summarize ablation results of each proposed module in Table \ref{table:ablation}.
Baseline, in line1, ignores the fine-grained cross-modal interaction and directly predicts segmentation mask with implicit concatenation-convolution~\cite{shi2018key,hu2016segmentation} fusion.
In other lines, the explicitly cross-modal interaction is involved and the network with different components is supervised under a combination of cross-modal contrastive loss and segmentation loss.
\noindentet al.}{\emph{et al}\onedot{Observations:}
All ablations are conducted on the validation set of A2D Sentences.
We could obtain the following observations:
\textbf{(1)} Due to the redundant uninformative visual-distinguishable feature, the vanilla CCL merely brings limited improvement compared with baseline.
\textbf{(2)} Both of the hard mining strategies can significantly improve the vanilla CCL, which validates the effects of the design.
\textbf{(3)} Conclusively, these results confirm the merits of the explicitly cross-modal interaction formulation again.
\begin{figure}[t]
\begin{center}
\includegraphics[width=.98\linewidth]{new_images/sensitive.pdf}
\end{center}
\captionsetup{font=small}
\caption{\small
\textbf{Variation of model performance} with different selecting ratios in RHIC (Sec.~\ref{section:rhic}) and different objective weighting $\lambda$.
}\label{fig:sensitive}
\end{figure}
\subsubsection{Relative ratio of et al.}{\emph{et al}\onedot{hard} samples in RHIC. }
We supply an additional experiment on ranges of relative ratio of RHIC in Fig.\ref{fig:sensitive}~(a). We modify relative ratio of hard to easy pixels from 1:2 to 1:5, and final performance of CVLS is stable ($\pm$0.15$\%$ in terms of mAP) but extra epochs are needed for convergence. While the ratio is ranged from 1:1 to 4:1, final performance progressively drops $\sim$0.9$\%$, as RHIC is progressively losing efficacy which causes model degradation. We empirically set the ratio to 1:3 for balancing the performance and convergence speed.
\subsubsection{Parameter sensitivity analysis of $\lambda$. } As we employ a joint loss of segmentation term and contrastive term, we further evaluate different values of the weighting parameter $\lambda$ (Sec.~\ref{section:ObjCons}) and summarize the results
in Fig.~\ref{fig:sensitive}~(b). In practice, we find the proposed approach is robust to parameter variation and the performance is slightly jittered among different settings. We finally set $\lambda$ to 0.8 empirically.
\subsection{Qualitative Analysis}
In this section, we investigate the internal mechanism in CVLS by analyzing qualitative results.
The comparisons between the full model and alternative structures are illustrated in Fig.~\ref{fig:qualiti}.
Compared with the implicit fusion baseline (Fig.~\ref{fig:qualiti}(b)), even without hard mining strategies, the vanilla cross-modal contrastive learning pipeline could still capture finer cross-modal interaction and shows reasonable segmentation results.
Due to the model degradation caused by extra visual-distinguishable feature, the model still messes up the semantically similar region, et al.}{\emph{et al}\onedot{i.e.}, two guys with red and green masks respectively (the third line in Fig.~\ref{fig:qualiti}(c)). While in Fig.~\ref{fig:qualiti}(d), the problem is greatly alleviated.
Conclusively, these visualized results reconfirm the effectiveness of CVLS and the hard mining strategies which facilitates the entire method in both implicit (LCF) and explicit (RHIC) aspects.
\section{Conclusion}
In this paper, we propose a highly effective pipeline with a novel cross-modal contrastive learning objective
for the first attempt of introducing the explicit cross-modal interaction into video-language segmentation field.
To prevent the model from degradation to single-modal, we further employ two cross-modal hard mining strategies to exclude the over-simple visual-distinguishable feature from both implicit and explicit aspects, et al.}{\emph{et al}\onedot{i.e.}, Language-relevant Channel Filter and Relative Hard Instance Construction.
Evaluations on commonly used benchmarks demonstrate that CVLS surpasses all the state-of-the-art methods by large margins.
|
\subsection{Experimental Setup}
In our experiments, we consider contact sequences of length $N=10$, trajectories of length $M=10.8 \text{ seconds}$, and 0.1 second optimization time steps. For these results, we fix the initial pose of the object and randomly sample 60 object goal orientations between -$\pi$ and $\pi$ radians. In our experiments, we obtain the object pose trajectory for the contact sequence planner by linearly interpolating between initial and final object pose. We use the same initial contact configuration for all trials. The object we consider is a 20cm$\times$10cm rectangular object with coefficient of friction $\mu = 0.7$ and mass \mbox{$m = 50$ grams}.
\subsection{Baselines}
We compare TrajectoTree to these CITO baselines:
1) ``CITO'': General CITO formulation from Section \ref{ssec:cio}, initializing decision variables such that the system is in static equilibrium throughout the entire trajectory. We hypothesize that this formulation will produce highly dynamic trajectories prone to dropping the object during execution.
2) ``CITO, warm-start'': General formulation from Section \ref{ssec:cio}, initializing object pose and joint angles to those given by the high-level plan. We hypothesize that compared to the first baseline, this initialization should move solutions closer to a non-dynamic manipulation sequence.
For all three methods, we use cost function (\ref{eqn:cost2}) with the same values for $L$, $Q$, and $R$
\subsection{Planning Speed and Solution Quality}
We show that TrajectoTree achieves faster total planning times than the baseline methods. For methods that use the contact sequence planner, we consider total planning time to be the sum of the search time and the trajectory optimization solve time. Fig. \ref{fig:time_and_cost} (left) reports the total planning times across 60 trials for each method. TrajectoTree achieves an average planning time of 14 seconds. ``CITO'' and ``CITO, warm-start'' achieve average planning times of 98 seconds and 76 seconds, respectively. The planning times of TrajectoTree have a significantly narrower interquartile range compared to those of the baselines, suggesting that our method also performs more consistently across various goal poses.
This improved planning speed comes at the expense of finding trajectories that have a higher cost in comparison with the baselines. This is expected, as TrajectoTree solves an optimization problem that is the same as for the baselines, but with additional constraints corresponding to the desired contact sequence. Specifically, TrajectoTree constrains particular fingers to remain in contact with the object via typical equality constraints, rather than allowing them to make and break contact via complementarity constraints. Because TrajectoTree is more constrained than the baselines, the optimal objective value is guaranteed to be no lower than the baselines, and will be higher than the baselines if these additional constraints are active, as evidenced in Fig.~\ref{fig:time_and_cost} (right). What is crucial to note, however, is that the trajectories found by TrajectoTree are considerably more robust to dropping the object when executed in a physics simulator, as shown in Fig.~\ref{fig:tracking} and detailed in sec~\ref{sec:controller}.
A comparison of the two baseline methods also demonstrates that the choice of the initial guess impacts solutions, as other works \cite{PosaCIO, Dai2014} have also found; however, our results suggest that imposing additional constraints, as done in TrajectoTree, can significantly influence the solution cost and planning time. While is it clear that these additional constraints will increase the trajectory cost, it may be surprising that these extra constraints can also significantly reduce solution time and lead to more robust plans in practice.
\subsection{Executing trajectories in simulation}
We show in simulation that controllers which track trajectories found with TrajectoTree are significantly less likely to drop the object than when tracking trajectories planned with the baseline methods. We execute trajectories in a planar PyBullet environment by planning a trajectory once and tracking the reference fingertip positions with the low-level controller described in Section \ref{sec:controller}. We use the same hand-tuned controller gains for all methods.
We compare executing trajectories planned by all three methods, across the same randomly chosen goal poses, and show the mean absolute tracking errors for object position and angle in Fig. \ref{fig:tracking}.
Although we only perform closed-loop tracking of fingertip positions, the controller is able to maintain relatively small tracking errors on object pose when tracking trajectories planned by TrajectoTree. In contrast, tracking the trajectories planned by the baseline methods results in much larger errors and frequently dropping the object, as shown by the dramatically increasing errors in the $y$ dimension (gravity points in the negative $y$ direction in our simulation). This reinforces that imposing the additional constraints given by the high-level plan, which only considers kinematic feasibility and grasp stability, moves the CITO towards solutions that are of higher cost, but also more robust during execution.
\subsection{Contact sequence planner}
\input{Figures/graph}
To find a feasible finger contact sequence for a given object trajectory, we use depth-first search with a heuristic to construct and search through a tree. The planner takes as input an $N$-length object pose trajectory $(x^*_0, ..., x^*_{N-1})$ and initial contact point locations for each finger, $p^*_0$. A node $z$ in the tree consists of an object pose, finger joint configurations, and finger contact point locations on the object. To construct a tree, we expand nodes that are kinematically feasible and have a stable grasp. To expand a node, we find its neighbors by exploring a discrete set of contact switches for each finger not needed to maintain grasp stability. We outline our search algorithm with pseudo-code in Algorithm \ref{alg:search}, and Fig. \ref{fig:graph} illustrates each of its components with a simple example. Below, we describe each of the functions in Algorithm \ref{alg:search}.
\begin{algorithm}
\newcommand\mycommfont[1]{\footnotesize\ttfamily\textcolor{gray}{#1}}
\SetCommentSty{mycommfont}
\DontPrintSemicolon
\SetAlgoLined
\SetKwInOut{Input}{input}
\SetKwInOut{Output}{output}
\Input{$(x^*_0, ..., x^*_{N-1})$, $p_0^*$}
\Output{$\mathcal{T} = (V,E)$, Planner success status}
$V_{\text{open}} \leftarrow \{z_{\text{start}}\}$,
$V_{\text{closed}} \leftarrow \emptyset$ \\
$\mathcal{T} \leftarrow (V \{z_{\text{start}}\}, E \{\text{none}\})$ \tcp{Initialize tree}
\While{$V_{\text{open}} \neq \emptyset$}{
\tcp{Get deepest node from open set}
$z_{\text{current}} \leftarrow \textsc{SelectNode}(V_{\text{open}})$ \\
$V_{\text{open}} \leftarrow V_{\text{open}} \setminus z_{\text{current}}$ \tcp{Remove from open set}
$V_{\text{closed}} \leftarrow V_{\text{closed}} \cup z_{\text{current}}$ \tcp{Add to closed set}
\If{$\textsc{IsFeasible}(z_{\text{current}})$}{
\If{$x^*_{\text{current}}$ equals $x^*_{N}$}{
\tcp{Exit if object goal pose reached}
\Return{$\mathcal{T}$, True}
}
\tcp{Add the neighbors to tree}
$Z_{\text{near}} \leftarrow \textsc{GetNeighbors}(z_{\text{current}})$ \\
\For{$z \in Z_{\text{near}}$}{
$\mathcal{T} \leftarrow \textsc{Insert}(z, z_{\text{current}})$\\
$V_{\text{open}} \leftarrow V_{\text{open}} \cup z$
}
}
}
\Return{$\mathcal{T}$, False}
\caption{Contact sequence planner}
\label{alg:search}
\end{algorithm}
\textsc{SelectNode}($V_{\text{open}}$):
From the nodes in the open set $V_{\text{open}}$ deepest in the tree, return the node with the lowest heuristic value. As a heuristic for choosing feasible nodes, we use the deviation of the second finger joint from its nominal joint angle of 45 degrees, guided by the intuition that the more extended fingers are, the more likely a contact configuration will be infeasible. While we found that this heuristic helped improve search times marginally, it can be swapped with other heuristics or removed entirely.
\textsc{IsFeasible}($z$): Determine if node $z$ is valid, and should be expanded, or a dead-end. It is a dead end if the object is not in frictional form closure \cite{SpringerGrasping}, or the contact points are not reachable by the fingers. We check for frictional form closure by solving the linear program formulated in \cite{SpringerGrasping} and check for kinematic feasibility with inverse kinematics. Return \textit{False} if $z$ is a dead end and \textit{True} otherwise.
\textsc{GetNeighbors}($z$): Return all neighbor nodes of a feasible node $z$. Given the current node's object pose, $x^*_n$, the object pose of all neighbor nodes is $x^*_{n+1}$. To get a neighbor node $z_{\text{near}}$ of $z$, first determine all free fingers in $z$. A finger is ``free'' if it is not needed to maintain frictional form closure on the object if the other fingers remain fixed on the object. Next, choose a new contact point location for that free finger by displacing its current contact location along the object's surface by some displacement chosen from a pre-defined, fixed set of displacements $D_{\text{cp}}$. The set $D_{\text{cp}}$ includes the zero displacement, meaning that the free finger will remain at its current contact location. Then, create a new neighbor for each displacement in $D_{\text{cp}}$. Only one finger can perform a contact switch between two nodes, so each neighbor node will have the same set of contact points as the parent node, $z_{\text{current}}$, except for one contact point which may have changed.
\textsc{Insert}($z_{\text{new}}$, $z_{\text{current}}$): Add $z_{\text{new}}$ to $V$, and add the edge between $z_{\text{new}}$ and its parent $z_{\text{current}}$ to $E$.
Algorithm \ref{alg:search} outputs path of $N$ nodes from the initial to final object pose, where each node contains the object pose $x^*_{n}$, joint angles $q^*_{n}$, and contact point locations in the object frame $p_{n}^*$ for \mbox{$n=0,...,N-1$}. At most one finger can perform a contact switch at any given time, meaning that with four fingers in total, any two consecutive nodes in the path must have three common contact points. We call a pair of consecutive nodes in a path a transition segment. In TrajectoTree, we enforce these transition segments in CITO by constraining contact point locations.
\subsection{General contact-implicit trajectory optimization} \label{ssec:cio}
We formulate the general CITO problem for planar dexterous manipulation. A discussion of extending this formulation to 3D can be found in \cite{PosaCIO}. In Section \ref{sec:contact-transition-to}, we augment this formulation with constraints from the contact sequence plan to arrive at the contact transition trajectory optimization used in TrajectoTree. Table \ref{table:notation} contains the primary notation.
We find trajectories for the object pose, finger joint configurations and torques, and corresponding contact forces by solving a CITO problem of the form
\begin{align}\label{eqn:opt}
&\underset{\{\mathbf{x},\mathbf{q},\dot{\mathbf{x}},\dot{\mathbf{q}},\boldsymbol{\tau},\boldsymbol{\lambda},\boldsymbol{\gamma}\}}{\text{minimize}}\;\;\;
F(\mathbf{x},\boldsymbol{\tau}, \boldsymbol{\lambda}, \boldsymbol{\gamma})\\
& \quad \text{s.t.} \quad \;\;\; \text{dynamics constraints (\ref{eq:obj_dyn}), (\ref{eq:hand_dyn})}\nonumber\\[-3pt]
& \quad \quad \quad \quad \text{friction cone constraint (\ref{eq:con_fc})}\nonumber\\[-3pt]
& \quad \quad \quad \quad \text{complementarity constraints (\ref{eq:con_comp}), (\ref{eq:con_fgp1})}\nonumber\\[-3pt]
& \quad \quad \quad \quad \text{path constraints for each decision variable,}\nonumber
\end{align}
where $F(\mathbf{x}, \boldsymbol{\tau}, \boldsymbol{\lambda}, \boldsymbol{\gamma})$ is a quadratic tracking objective function of the form
\begin{equation} \label{eqn:cost}
\begin{split}
F(\mathbf{x},\boldsymbol{\tau},\boldsymbol{\lambda}, \boldsymbol{\gamma}) &= \sum_{k=0}^{M} (x_k - x_{\text{goal}})^T Q (x_k - x_{\text{goal}})\\
& \quad + \tau_k^T R \tau_k + \lambda_k^T L \lambda_k + ||\boldsymbol{\gamma}||_1,
\end{split}
\end{equation}
where $\mathbf{x} = (x_0, \ldots, x_M)$ is the object pose trajectory, and $\mathbf{q}$, $\boldsymbol{\tau}$, and $\boldsymbol{\lambda}$ are defined similarly as the joint angle trajectory, joint torque trajectory, and contact force trajectory, respectively. $L$, $Q$, and $R$ are weight matrices, and $\boldsymbol{\gamma}$ is a vector of slack variables, which we use to relax several equality constraints to improve convergence of the optimization problem \cite{Manchester2017}.
The constraints in (\ref{eqn:opt}) are as follows.
We use trapezoidal collocation \cite{Kelly2017AnIT} to discretize the finger and object dynamics,
\begin{alignat}{2}
M_{\text{obj}}(x_k)\Ddot{x_k} &= G(x_k, q_k) \lambda_k + g_{\text{obj}}(x_k), \label{eq:obj_dyn}\\
M_{\text{hand}}(q_k)\Ddot{q_k} &= \tau_k - J^T(x_k,q_k)\lambda_k. \label{eq:hand_dyn}
\end{alignat}
In these constraints, $g_{\text{obj}}$ is the vector of gravitational forces on the object, and $M_{\text{obj}}$ and $M_{\text{hand}}$ are the object and hand mass matrices, respectively. We omit the Coriolis and centrifugal terms in (\ref{eq:hand_dyn}) because these are negligible in a quasi-static setting with low-mass manipulators. Higher-order collocation methods, like Hermite-Simpson or orthogonal collocation, can be used instead of trapezoidal collocation to improve solution accuracy, as \cite{Patel2019ContactImplicitTO}.
We model contact in the optimization problem with friction cone constraints and complementarity constraints:
\begin{alignat}{2}
\mu \lambda_{n,k} - |\lambda_{t,k}| & \geq 0 \label{eq:con_fc}\\
\gamma_i - \phi(q_{k})\lambda_{n,k} &\geq 0 \label{eq:con_comp}\\
-\gamma \leq (J\dot{q}_k - G^T \dot{x}_k)_{\{x,y\}} \lambda_{n,k} &\leq \gamma \label{eq:con_fgp1}\\
\lambda_{n,k} & \geq 0\\
\phi(q_k) & \geq 0 \label{eq:con_phi}\\
\boldsymbol{\gamma} &\geq \boldsymbol{0} \label{eq:con_slack_var}.
\end{alignat}
This ensures that contact forces are zero when the fingertips are not in contact with the object. Equations (\ref{eq:con_fc}) -- (\ref{eq:con_slack_var}) apply for each finger $f = 1,...,n_f$ separately. To minimize notation clutter, we drop the additional subscript $f$ and use $q$ and $\lambda$ to denote joint angles and contact forces for a single finger $f$.
The vector of contact forces $\lambda$ is comprised of normal and tangential components $\lambda_n$ and $\lambda_t$, respectively. Equation (\ref{eq:con_fc}) constrains the contact forces to lie within the planar friction cones of an object with coefficient of friction $\mu$. The function $\phi(q)$ can be thought of as a signed-distance field of the object, which enforces a non-penetration constraint with (\ref{eq:con_phi}) and only equals zero when the fingertip is in contact with the object. Equation (\ref{eq:con_comp}) is the complementarity constraint on normal contact forces, relaxed with a slack variable. We include another relaxed complementarity constraint, (\ref{eq:con_fgp1}), to constrain each fingertip to remain fixed at their contact points when on the object; these equations constrain the $\{x,y\}$ components of a fingertip's velocity, in the world frame, to be equal to the $\{x,y\}$ components of the contact point velocity, also in the world frame. We solve the problem in (\ref{eqn:opt}) with the constraints in (\ref{eq:obj_dyn}) -- (\ref{eq:con_slack_var}) using IPOPT \cite{ipopt}.
\subsection{Constraining CITO to the high-level contact sequence} \label{sec:contact-transition-to}
Finally, we formulate the contact transition trajectory optimization by combining the contact sequence plan with CITO. Given a high-level contact sequence plan, we modify the cost function and introduce additional task-specific constraints that specify which fingers are free and fixed, as well as the contact point locations of each finger. For a high-level plan with $N-1$ transition segments, we fix each segment to be $\hat{M}$ time-steps long, making the full trajectory $M = (N-1)\hat{M}$ time-steps long. Each of the $s=0,...,(N-2)$ transition segments consists of two consecutive nodes.
First, we introduce a cost function that uses object poses $(x^*_1, ..., x^*_{N-1})$ given by the high-level plan as incremental goal object poses in the running state cost, resulting in a cost function of the form
\begin{equation} \label{eqn:cost2}
\begin{split}
\hat{F}(\mathbf{x},\boldsymbol{\tau},\boldsymbol{\lambda}, \boldsymbol{\gamma}) = \sum_{j=0}^{M} \tau_j^T R \tau_j + \lambda_j^T L \lambda_j + ||\boldsymbol{\gamma}||_1 + \\%[-5pt]
\sum_{n=1}^{N-1} \sum_{k=0}^{\hat{M}} (x_{(n-1)\hat{M} + k} - x_{n}^*)^T Q (x_{(n-1)\hat{M} + k} - x_{n}^*).
\end{split}
\end{equation}
If segment $s$ has a free finger $f'$, we constrain the contact points to be fixed at locations specified by the high-level plan at the first and last time-step of the segment, and keep the complementarity constraints for the other time-steps with constraints of the form
\vspace*{-\baselineskip}
\begin{alignat}{2}
\text{FK}(q_{s\hat{M},f'}) &= p_{s,f'}^* \label{eq:con_free_cp1}\\
\text{FK}(q_{(s+1)\hat{M}-1,f'}) &= p_{(s+1),f'}^*\\
\text{(\ref{eq:con_comp}), (\ref{eq:con_fgp1})} \text{ with } k &\in \{s\hat{M}+1,...,(s+1)\hat{M}-2\}.
\end{alignat}
The fixed fingers $f$ are constrained to be fixed at locations specified by the high-level plan for the entire segment $s$ with equality constraints of the form
\vspace*{-\baselineskip}
\begin{alignat}{2}
\text{FK}(q_{k,f}) &= p_{s,f}^* \label{eq:con_fixed_cp} \quad \text{for} \quad k &\in \{s\hat{M},...,(s+1)\hat{M}-1\}
\end{alignat}
where $\text{FK}(q_{k,f})$ is the forward kinematics that computes the fingertip position of finger $f$ at time-step $k$. In our implementation, we relax the contact point equality constraints with slack variables, but omit them here for brevity.
\input{Figures/solve_time_cost}
\input{Figures/obj_tracking}
Solving (\ref{eqn:opt}) with (\ref{eqn:cost2}) as the cost function and (\ref{eq:con_free_cp1}) - (\ref{eq:con_fixed_cp}) in place of (\ref{eq:con_comp}) and (\ref{eq:con_fgp1}) results in trajectories that adhere to the contact sequence in the high-level plan. With this formulation, we solve one optimization problem for the entire $M = (N-1)\hat{M}$ length trajectory. Additionally, we include the boundary constraints
\begin{alignat}{2}
x_0 &= x_0^* \label{eq:con_x0}\\
q_0 &= q_0^* \label{eq:con_q0}\\
\dot{x}_{s\hat{M}} , \dot{x}_{(s+1)\hat{M}-1} & = 0 \;\;\forall\;\; s \in \{0,...,N-2\} \label{eq:con_dx}\\
\dot{q}_{s\hat{M}} , \dot{q}_{(s+1)\hat{M}-1} & = 0 \;\;\forall\;\; s \in \{0,...,N-2\},
\label{eq:con_dq}
\end{alignat}
where (\ref{eq:con_x0}) and (\ref{eq:con_q0}) constrain the initial object pose and joint angles of the entire trajectory to those specified by the first node in the high-level plan. Equations (\ref{eq:con_dx}) and (\ref{eq:con_dq}) constrain the object and joints to be stationary at the first and last time-step of each segment. We do not constrain the final object pose or joint angles of each segment to those specified in the high-level plan, thereby allowing for some deviation
\subsection{Controller} \label{sec:controller}
To execute trajectories planned by TrajectoTree in simulation, we track the fingertip trajectories in Cartesian space using the following simplified impedance controller \cite{wuthrich2020trifinger} with additional gravity compensation for the fingers (time index omitted for brevity)
\begin{equation} \label{eq:ctr2}
\tau = J^T\Big(k_p(p_{\text{ref}}-p)+k_v(\dot{p}_{\text{ref}}-\dot{p})+\lambda_{\text{ref}}\Big) + g_{\text{hand}},
\end{equation}
where $p_{\text{ref}}$, $\dot{p}_{\text{ref}}$, and $\lambda_{\text{ref}}$ are the reference fingertip positions, velocities, and contact forces from the trajectory optimization solution, $g_{\text{hand}}$ is the gravity compensation vector, and $k_p$ and $k_v$ are hand-tuned controller gains.
\subsection{Contact-implicit trajectory optimization}
Trajectory optimization is a key method for planning robot motion, generating motion plans that locally optimize a given cost functional, subject to a set of constraints. Although trajectory optimization generalizes well to different systems, traditional methods can only handle smooth dynamics, and not the discontinuous dynamics introduced by contact switching. In recent years, several works have introduced a class of trajectory optimization methods that address this limitation. These methods, termed CITO, simultaneously plan state, control input, and contact force trajectories without needing a pre-specified contact mode schedule. They handle the hybrid dynamics of contact with either complementarity constraints \cite{Stewart2000, Yunt2006,PosaCIO,Manchester2017,Landry2019BilevelOF} or with soft constraints implemented as a penalty term in the cost function \cite{Mordatch2012DiscoveryOC, Mordatch2012ContactinvariantOF}. In this work, we follow the formulation introduced in \cite{PosaCIO}.
Since the introduction of CITO methods, many works have applied them to whole-body dynamic motion planning \cite{Dai2014, Marcucci2017ATT}, quadruped locomotion \cite{Landry2019BilevelOF}, and single-leg jumping \cite{Mastalli2016HierarchicalPO}.
Of these works, \cite{Marcucci2017ATT} and \cite{Mastalli2016HierarchicalPO} use hierarchical planning schemes, first planning a trajectory with a simplified robot model and then using this to warm-start the full trajectory optimization. TrajectoTree adopts a similar hierarchical design strategy, first considering a simpler kinematics problem before introducing full dynamics.
More recently, CITO methods have also been used for planning non-prehensile object manipulation \cite{Sleiman2019, Onol2019ContactImplicitTO}. There have also been a few works that use CITO for planning dexterous manipulation tasks \cite{Mordatch2012ContactinvariantOF, Gabiccini2018CIODexMan}.
In \cite{Mordatch2012ContactinvariantOF}, the authors impose both dynamics and contact constraints using a soft penalty formulation to synthesize manipulation motions for computer graphics. The method proposed in \cite{Gabiccini2018CIODexMan} is more similar to ours, in that it uses a complementarity constraint formulation. While both methods are able to plan dexterous manipulation tasks that involve contact switching, results are only shown as animated visualizations of the trajectories, and do not include discussions of the planning times.
While one might argue that the versatility of CITO makes it trivial to apply to planning dexterous manipulation tasks, in reality, obtaining good solutions with reasonable solve times still depends on providing an appropriate initialization and task-specific constraints. TrajectoTree accomplishes this by augmenting CITO with an additional high-level planner.
\subsection{Planning dexterous manipulation}
Contact switching is an integral part of dexterous manipulation, but reasoning about how to make and break contact with an object is difficult due to the combinatorics and hybrid dynamics associated with the problem. To reduce the complexity associated with reasoning about contact switching, many dexterous manipulation planning methods rely on breaking down the task by choosing hand-designed manipulation primitives \cite{Omata1996,Cruciani2018DMG,Li2020LearningHC}. Often, primitives are designed for specific tasks or objects, thus requiring significant implementation effort to generalize to different scenarios.
Finger gaiting is another common method for reducing the complexity of dexterous manipulation planning by allowing only a single finger to switch contacts at any given time. Finger gaiting has most often been used for in-hand re-grasping \cite{Sundaralingam2018GeometricIR,Fan2017RealTimeFG}. Although finger gaiting can be considered a primitive, it is more general than higher-level manipulation primitives such as sliding, pushing, or pivoting.
In contrast to breaking down a task by imposing structure on motion, data-driven approaches have also been used to plan dexterous manipulation tasks \cite{nagabandi2019deep,openai}, and while the results have been impressive, they require large amounts of training data and time to generalize to different scenarios.
Of these three strategies, our proposed approach is most similar to finger gaiting, and does not rely on choosing hand-designed primitives. Unlike \cite{Sundaralingam2018GeometricIR}, we consider the task of moving an object to a goal pose and do not assume a given order in which fingers move. \cite{Fan2017RealTimeFG} assumes that the object is smooth to enable fingers to slide along its surface.
\subsection{Contact planning for locomotion and manipulation}
Finally, we address the contact planning works most directly related to this work.
First, we discuss a recent work that combines search-based planning and dynamics optimization for humanoid contact planning \cite{LinBerenson2018}, as well as an extension to that work that considers the robustness of the method to disturbances \cite{Lin2020RobustHC}. The core structure of the methods in these works is similar to TrajectoTree, first constructing and searching through a graph to find a contact sequence, and then using a dynamics optimization step to generate full trajectories given this contact sequence. However, their graph search finds contact sequences with low dynamics costs, which traditionally would involve solving a slow dynamics optimization problem to obtain the cost of every contact transition, whereas our discrete planner only cares about finding feasible solutions. Consequently, the goal of their work is to speed up the graph search by training a network to predict the optimal objective value of the dynamics optimization for a contact transition. In contrast, this work focuses on speeding up trajectory optimization by using a search-based planner.
Next, we discuss several hybrid feedback control works that reason about contact mode scheduling. These works employ a variety of strategies, including mixed-integer quadratic programming (MIQP) \cite{Marcucci2017ApproximateHM, Hogan2016FeedbackCO}, sequencing high-level primitives \cite{Woodruff2017PlanningAC}, enumerating all possible contact sequences \cite{Doshi2019}, and learning contact schedule selection with Bayesian optimization \cite{Seyde2019LocomotionPT}. Of these, only \cite{Doshi2019} considers multi-contact dexterous manipulation tasks. The authors present a hybrid differential dynamic programming (DDP) algorithm for closed-loop execution of planar pushing and pivoting primitives with frictional contact switches. Their method uses input-constrained DDP to explore and rank all feasible contact mode sequences. Unlike CITO methods, DDP requires a fixed contact mode sequence. Consequently, their approach requires enumerating all possible contact sequences, limiting the number of contact switches they can consider while still obtaining reasonable planning times. The maximum number of contact switches they consider is two.
Finally, we highlight two recent papers that both reason about how to plan contact switches for planar dexterous manipulation tasks. The first, \cite{AceitunoCabezas2020AGQ}, reformulates the problem with a quasi-dynamic relaxation and constructs the planning problem as mixed-integer program. The second, \cite{cheng2020contact}, uses rapidly-exploring random tree guided by contact modes. Both methods plan a diverse set of dexterous manipulation tasks, but neither consider full finger kinematics, and instead assume that the contact points will always be reachable.
\section{Introduction}
\input{intro}
\section{Related Work}
\input{related_work}
\section{Method}
\input{methods}
\section{Experiments \& Results}
\input{experiments}
\section{Conclusion}
\input{conclusion}
\bibliographystyle{unsrt}
|
\section{Introduction}
For a left Artinian ring $R$, it is known that the prime radical of $R$ coincides with its Jacobson radical. Therefore, in any left Artinian ring the prime radical is always nilpotent. For an arbitrary ring $R$, it is no longer true that the prime radical equals the Jacobson radical. In general, the prime radical of $R$ is contained properly in its Jacobson radical. It can be seen that the prime radical of a ring $R$ consists of nilpotent elements, that is, the prime radical is a nil ideal. It has been a recurrent question in ring theory whether the nil (one-sided) ideals are nilpotent. Levitzki, in 1939 (see \cite[10.30]{lamfirst} and the paragraph below the proof) proved that every nil one-sided ideal of a left Noetherian ring is nilpotent. In particular, the prime radical of a left Noetherian ring is nilpotent. This result was extended for a left Goldie ring by Lanski \cite{lanski1969nil} in 1969.
In \cite{BicanPr}, a product of modules generalizing the product of a left ideal with a left module was introduced. Hence, this product gives a product of submodules of a given module and it coincides with the usual product of ideals when the module is the base ring. In 2002, the first author proved that this product in a left $R$-module $M$ is associative provided that $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$ \cite[Proposition 5.6]{beachy2002m}. Later, in \cite{raggiprime} (resp. \cite{raggisemiprime}), it is introduced the concept of prime (resp. semiprime) submodule of a module $M$ using the product mentioned above. As part of the second author's dissertation, in \cite{maugoldie} semiprime Goldie modules were presented as a generalization of semiprime left Goldie rings. In this paper, for a left $R$-module $M$ which is $M^{(\Lambda)}$-projective for every index set $\Lambda$, we consider the intersection of all prime submodules of $M$ and we prove that this intersection is nilpotent when $M$ is Goldie, extending Lanski's result \cite[Theorem 1]{lanski1969nil}. The nilpotency of the prime radical of a Goldie module will be used in a subsequent paper when the finite reduced rank is studied in categories of type $\sigma[M]$ \cite{beachyreduced}.
In order to achieve our goal, this paper is divided in five sections. The first section is this introduction and the next one compiles the background needed to make this work as self-contained as possible. In Section \ref{LocNilSub}, nil-submodules and locally nilpotent submodules are defined. It is proved that the sum of all locally nilpotent submodules of a module inherits the property (Lemma \ref{sumlocnil}). For a module $M$, this maximal locally nilpotent submodule is denoted by $\mathfrak{L}(M)$ and it happens that $\mathfrak{L}\left(M/\mathfrak{L}(M) \right)=0$ provided that $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$ (Proposition \ref{Lfiyrad}). Moreover, the submodule $\mathfrak{L}(M)$ coincides with the intersection of all prime submodules of $M$ (Corollary \ref{nesL}). At the end of this section, using the operator $\mathfrak{L}(\_)$, a characterization of semiprime rings in terms of free modules is given (Corollary \ref{rsp}). In Section \ref{AccandNil}, we introduce the notion of ascending chain condition (acc) on annihilators on a module $M$. We study the behavior of nil submodules of a module satisfying acc on annihilators. We show that fully invariant nil submodules are locally nilpotent when $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$ and satisfies acc on annihilators (Proposition \ref{accnillocnil}). Also in this section, a fundamental result of the paper is proved (Proposition \ref{subm}). Finally in Section \ref{Nilpotency}, the main result is proved. It is shown that a fully invariant nil submodule of a Goldie module $M$ is nilpotent, under the hypotheses that $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$ and retractable (Theorem \ref{main}). As under the previous hypothesis on $M$, when $M$ is a Goldie module its endomorphism ring is a right Goldie ring (Lemma \ref{mgolsgol}), the proof of Theorem \ref{main} is focused on finding elements in the endomorphism ring of a Goldie module like those in \cite[Lemma 8]{lanski1969nil}.
Throughout this paper $R$ will denote an associative ring with unit and all the $R$-modules will be unitary left modules. The notation $N\leq M$ and $N<M$ will mean $N$ is a submodule of $M$ and $N$ is a proper submodule of $M$ respectively. Given an $R$-module $M$, the endomorphism ring of $M$ usually will be denoted by $S=\End_R(M)$. A direct sum of copies of a module $M$ will be denoted by $M^{(\Lambda)}$ where $\Lambda$ is a set.
\section{Preliminaries}
Let $M$ and $N$ be left $R$-modules. It is said that $M$ is \emph{$N$-projective} if for any epimorphism $\rho:N\to L$ and any homomorphism $\alpha:M\to L$, there exists $\overline{\alpha}:M\to N$ such that $\rho\overline{\alpha}=\alpha$. It is true that if $M$ is $N_i$-projective for a finite family of modules $\{N_1,...,N_\ell\}$, then $M$ is $\bigoplus_{i=1}^\ell N_i$-projective \cite[18.2]{wisbauerfoundations}. In general, this is not true for arbitrary families. A module $M$ is called \emph{quasi-projective} if, $M$ is $M$-projective. In most of the results in this paper, we will assume that $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$, this hypothesis is satisfied by every finitely generated quasi-projective module for example \cite[18.2]{wisbauerfoundations}. The condition that $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$ is equivalent to say that $M$ is projective in the category $\sigma[M]$ \cite[18.3]{wisbauerfoundations}, where $\sigma[M]$ is the category consisting of all $M$-subgenerated modules. In many of our references, this last equivalence is used. It will also be used, inside the proofs, that if $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$ and $N$ is a fully invariant submodule of $M$ then, $M/N$ is $(M/N)^{(\Lambda)}$-projective for every index set $\Lambda$ \cite[Lemma 9]{vanmodules}.
Given a left $R$-module $M$ and $N,K\leq M$, \emph{the product of $N$ with $K$ in $M$} is defined as
\[N_MK=\sum\{f(N)\mid f\in\Hom_R(M,K)\}.\]
If $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$, this product gives an associative operation in the lattice of submodules of $M$. Moreover, if $N$ and $K$ are fully invariant submodules of $M$, then $N_MK$ is fully invariant. For $N\leq M$, \emph{the powers of $N$} are defined recursively as follows: $N^1=N$ and $N^{\ell+1}={N^\ell}_MN$. \footnote{Do not confuse the notation with the direct product of $\ell+1$ copies of $N$ which will be denoted by $N^{(\ell+1)}$.} It is said that a submodule $N\leq M$ is \emph{nilpotent} if $N^\ell=0$ for some $\ell>0$. Some general properties of this product are listed in \cite[Proposition 1.3]{PepeGab}. Also, we give the following useful lemmas.
\begin{lemma}\label{proddirsumm}
Let $M$ be a module and $N\leq M$. Then $K_ML\subseteq K_NL$ for all $K,L\leq N$. In addition, if $N\leq^\oplus M$ then $K_ML=K_NL$.
\end{lemma}
\begin{proof}
Let $K,L$ be submodules of $N$. Using the restriction of a homomorphism, if $f:M\to L$ is any homomorphism then $f|_N(K)=f(K)$ since $K\leq N$. Thus $K_ML\subseteq K_NL$. Now suppose $N$ is a direct summand of $M$. If $f:N\to L$ then $f\oplus 0:M\to L$. Hence $K_NL\subseteq K_ML$. Thus $K_NL=K_ML$.
\end{proof}
\begin{lemma}\label{fprod}
Let $M$ be a quasi-projective module. If $A$ and $B$ are submodules of $M$ and $f\in\End_R(M)$, then $f(A_MB)=A_Mf(B)$ and $f(A)_MB\subseteq A_MB$.
\end{lemma}
\begin{proof}
Let $f\left( \sum_{i=1}^\ell g_i(a_i)\right)\in f(A_MB)$. Then,
\[f\left( \sum_{i=1}^\ell g_i(a_i)\right)=\sum_{i=1}^\ell fg_i(a)\in A_Mf(B).\]
On the other hand, let $\sum_{i=1}^\ell h_i(a_i)\in A_Mf(B)$. The restriction $f|_B:B\to f(B)$ is an epimorphism. Since $M$ is quasi-projective, there exists $g_i:M\to B$ such that $fg_i=h_i$ for all $i$. Hence,
\[\sum_{i=1}^\ell h_i(a_i)=\sum_{i=1}^\ell fg_i(a_i)=f\left( \sum_{i=1}^\ell g_i(a_i)\right)\in f(A_MB).\]
For the other assertion, let $\sum_{i=1}^\ell g_i(f(a_i))\in f(A)_MB$. Then $g_if:M\to B$. Thus, $\sum_{i=1}^\ell g_if(a_i)\in A_MB$.
\end{proof}
\begin{lemma}\label{epiproduct}
Let $M$ be quasi-projective and let $K,N\leq M$ be two submodules with $K$ fully invariant in $M$. If $\pi:M\to M/K$ is the canonical projection, then $\pi(N_MN)=\pi(N)_{M/K}\pi(N)$.
\end{lemma}
\begin{proof}
Since $K$ is fully invariant in $M$, given $f\in\Hom_R(M,N)$ there is an $\bar{f}\in\Hom_R(M/K,\pi(N))$ such that $\pi f=\bar{f}\pi$. Now, since $M$ is quasi-projective, if $\bar{f}\in\Hom_R(M/K,\pi(N))$ then there exists $f\in\Hom_R(M,N)$ such that $\pi f=\bar{f}\pi$. Thus, there is an epimorphism $\Hom_R(M,N)\to \Hom_R(M/K,\pi(N))\to 0$ induced by $\pi$.
It follows that:
\begin{equation*}
\begin{split}
\pi(N_MN) & =\pi\left( \sum\{f(N)\mid f\in\Hom_R(M,N)\}\right) \\
& =\sum\{\pi f(N)\mid f\in \Hom_R(M,N)\} \\
& =\sum\{\bar{f}\pi(N)\mid \bar{f}\in\Hom_R(M/K,\pi(N))\}\\
& =\pi(N)_{M/K}\pi(N).
\end{split}
\end{equation*}
\end{proof}
In a natural way, a fully invariant submodule $P$ of a module $M$ is said to be a \emph{prime submodule} (resp. \emph{semiprime submodule}) if $N_MK\leq P$ (resp. $N_MN\leq P$) implies $N\leq P$ or $K\leq P$ (resp. $N\leq P$) for every fully invariant submodules $N$ and $K$. It is said that $M$ is a \emph{prime module} (resp. \emph{semiprime module}) if $0$ is a prime (resp. semiprime) submodule. Many properties known for prime or semiprime ideals can be extended to prime or semiprime sumodules as it has been done in \cite{beachy2002m,beachy2020fully,PepeGab,mauozcan,raggiprime,raggisemiprime}. Recall that a ring $R$ is \emph{left Goldie}, if $R$ satisfies the ascending chain condition (acc) on left annihilators and has finite uniform dimension. This notion can be extended to left modules as follows: a left module $M$ is called \emph{Goldie} if the set $\{\bigcap_{f\in X}\Ker f\mid X\subseteq \End_R(M)\}$ satisfies acc and $M$ has finite uniform dimension. It is clear that a ring $R$ is a left Goldie ring if and only if $_RR$ is a Goldie module \cite{maugoldie}.
\section{Locally nilpotent submodules}\label{LocNilSub}
\begin{defn}
Let $M$ be an $R$-module and $S=\End_R(M)$. A submodule $N$ of $M$ is called a \emph{nil-submodule} if for all $n\in N$, the right ideal $\Hom_R(M,Rn)$ of $S$ is nil. That is, if each $f:M\to Rn$ is nilpotent.
\end{defn}
\begin{lemma}\label{factornil}
Let $M$ be a quasi-projective module and let $N$ be a nil-submodule of $M$. Then, $\frac{N+K}{K}$ is a nil-submodule of $M/K$ for all $K\leq M$.
\end{lemma}
\begin{proof}
Let $n+K\in \frac{N+K}{K}$ and $f\in\Hom_R(M/K,R(n+K))$. Let $\pi:M\to M/K$ be the canonical projection and consider $\pi|:Rn\to R(n+K)$ the restriction to $Rn$. Since $M$ is quasi-projective, there exists $g:M\to Rn$ such that $(\pi|)g=f\pi$. By hypothesis, there exists $k>1$ such that $g^k=0$. Therefore,
\[0=(\pi|)g^k=f\pi g^{k-1}=f(\pi g)g^{k-2}=f(f\pi)g^{k-2}=\cdots=f^k\pi.\]
This implies that $f^k=0$ and hence $\frac{N+K}{K}$ is a nil-submodule of $M/K$.
\end{proof}
Recall that a subset $T$ of a ring $R$ is called \emph{locally nilpotent} if, for any finite subset $\{t_1,...,t_n\}\subseteq T$ there exists an integer $\ell$ such that any product of $\ell$ elements from $\{t_1,...,t_n\}$ is zero \cite[pp. 166]{lamfirst}.
\begin{defn}
Let $N$ be a submodule of a module $M$. It is said that $N$ is \emph{locally nilpotent} if for any subset $\{n_1,...,n_k\}\subseteq N$ there exists an integer $\ell>0$ such that any combination of $\ell$ elements of $\{n_1,...,n_k\}$ satisfies that
\[{Rn_{i_1}}_M{Rn_{i_2}}_M\cdots{_MRn_{i_\ell}}=0.\]
\end{defn}
Note that if $N$ is a locally nilpotent submodule of $M$, then $Rn$ is a nilpotent submodule for every $n\in N$, but $N$ might not be nilpotent as the following example shows.
\begin{example}
Consider the $\mathbb{Z}$-module $\mathbb{Q}$. Every nonzero cyclic submodule of $_\mathbb{Z}\mathbb{Q}$ is isomorphic to $\mathbb{Z}$ and $\mathbb{Z}_{\mathbb{Q}}\mathbb{Z}=0$. This implies that $_\mathbb{Z}\mathbb{Q}$ is locally nilpotent. But $\mathbb{Q}^\ell=\mathbb{Q}$ for all $\ell>0$.
\end{example}
\begin{rem}
It is clear that if a left ideal $I$ of a ring $R$ is a locally nilpotent submodule, then $I$ is a locally nilpotent subset. However, the converse might not be true. For, consider the ring $R$ given in \cite[pp. 167]{lamfirst}. The ring $R$ is prime, so has no nonzero nilpotent submodules. But it is proved that $R$ has a cyclic left ideal $Rx$ which is locally nilpotent. Therefore, $Rx$ cannot be a locally nilpotent submodule of $_RR$.
\end{rem}
\begin{rem}
If $N\leq M$ is locally nilpotent, then $N$ is a nil-submodule. For, let $n\in N$ and $f\in\Hom_R(M,Rn)$. Consider $\{n\}\subseteq N$. Then, there is an $\ell>0$ such that $(Rn)^\ell=0$. Since $f(n)\in Rn_MRn$, $f^{\ell-1}(n)\in(Rn)^{\ell}=0$. Thus, $f^\ell=0$. On the other hand, if $N\leq M$ is a nilpotent submodule, then $N$ is locally nilpotent.
\end{rem}
\begin{lemma}\label{fgnilp}
Suppose $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$ and let $N$ be a finitely generated submodule of $M$. If $N$ is locally nilpotent, then $N$ is nilpotent.
\end{lemma}
\begin{proof}
We proceed by induction on the number of generators of $N$. Suppose $N=Rn$. Since $N$ is locally nilpotent, there exists $\ell>0$ such that the product $\ell$ times of $Rn$ is zero, that is, $N^\ell=(Rn)^\ell=0$. Now suppose that $N=Rn_1+\cdots+Rn_k$ and that the result is valid for any locally nilpotent finitely generated submodule with less than $k$ generators. Put $K=Rn_1+\cdots+Rn_{k-1}$. By induction hypothesis there exist $\ell_1,\ell_2>0$ such that $K^{\ell_1}=0$ and $(Rn_k)^{\ell_2}=0$. Let $\ell>\ell_1+\ell_2$. Then,
\begin{equation}\label{a}
\begin{split}
N^\ell & =(K+Rn_k)^\ell\\
& =K^\ell+\cdots+{K^{\alpha_1}}_M{(Rn_k)^{\beta_1}}_M\cdots _M{K^{\alpha_h}}_M(Rn_k)^{\beta_h}+\cdots +(Rn_k)^\ell
\end{split}
\end{equation}
where $\alpha_i,\beta_i\geq 0$, $\sum_{i=1}^h\alpha_i+\beta_i=\ell$ and with the convention ${A^0}_MB=B$ and $A_MB^0=A$. Then $\sum_{i=1}^h\alpha_i>\ell_1$ or $\sum_{i=1}^h\beta_i>\ell_2$ since $\ell>\ell_1+\ell_2$. Suppose $\beta=\sum_{i=1}^h\beta_i>\ell_2$. Consider the summand $X={K^{\alpha_1}}_M{(Rn_k)^{\beta_1}}_M\cdots _M{K^{\alpha_h}}_M(Rn_k)^{\beta_h}$ in the expression (\ref{a}). If $\beta_h\neq 0$, then ${K^{\alpha_h}}_M(Rn_k)^{\beta_h}\subseteq (Rn_k)^{\beta_h}$. Hence
\begin{equation*}
\begin{split}
{K^{\alpha_1}}_M{(Rn_k)^{\beta_1}}_M\cdots _M{K^{\alpha_h}}_M(Rn_k)^{\beta_h} & \subseteq {K^{\alpha_1}}_M{(Rn_k)^{\beta_1}}_M\cdots _M{K^{\alpha_{h-1}}}_M(Rn_k)^{\beta_{h-1}+\beta_h} \\
& \cdots \\
& \subseteq {K^{\alpha_1}}_M(Rn_k)^\beta.
\end{split}
\end{equation*}
Hence $X=0$. If $\beta_h=0$, we have $X={K^{\alpha_1}}_M{(Rn_k)^{\beta_1}}_M\cdots _M{(Rn_k)^{\beta_{h-1}}}_M{K^{\alpha_h}}$. Proceeding as above, $X\subseteq {K^{\alpha_1}}_M{(Rn_k)^\beta}_MK^{\alpha_h}$. Thus, $X=0$. Analogously if $\alpha=\sum_{i=1}^h\alpha_i>\ell_1$. Hence $N^\ell=0$.
\end{proof}
\begin{rem}
Notice that the proof of Lemma \ref{fgnilp} implies that in a module $M$ which is $M^{(\Lambda)}$-projective for every index set $\Lambda$, the finite sum of nilpotent submodules is nilpotent.
\end{rem}
\begin{lemma}\label{sumlocnil}
Suppose $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$. If $N$ and $L$ are locally nilpotent submodules of $M$, then $N+L$ is locally nilpotent.
\end{lemma}
\begin{proof}
Let $\{n_1+l_1,..,n_k+l_k\}\subseteq N+L$. For $\{n_1,...,n_k\}\subseteq N$ there exists $\ell_N>0$ such that any combination of $\ell_N$ elements of $\{n_1,...,n_k\}$ satisfies that ${Rn_{i_1}}_M{Rn_{i_2}}_M\cdots{_MRn_{i_{\ell_N}}}=0$. Analogously, for the subset $\{l_1,...,l_k\}\subseteq L$, there exists $\ell_L>0$ such that any combination of $\ell_L$ elements of $\{l_1,...,l_k\}$ satisfies that ${Rl_{i_1}}_M{Rl_{i_2}}_M\cdots{_MRl_{i_{\ell_L}}}=0$. Take $\ell>\ell_N+\ell_L$. Then
\[{R(n_{i_1}+l_{i_1})}_M\cdots{_MR(n_{i_\ell}+l_{i_\ell})}\subseteq\left( {Rn_{i_1}+Rl_{i_1}}\right) _M\cdots _M\left(Rn_{i_\ell}+Rl_{i_\ell}\right)\]
\[= \cdots +{(Rn_{i_1})^{\alpha_1}}_M{(Rl_{i_1})^{\beta_1}}_M{(Rn_{i_2})^{\alpha_2}}_M{(Rl_{i_2})^{\beta_2}}_M\cdots_M{(Rn_{i_\ell})^{\alpha_\ell}}_M{(Rl_{i_\ell})^{\beta_\ell}}+\cdots\]
where $\alpha_j,\beta_j\in\{0,1\}$ with the convention that ${(Rn)^0}_MRl=Rl$ and $Rn_M(Rl)^0=Rn$. Also, these exponents satisfy $\sum_{j=1}^\ell \alpha_j+\beta_j=\ell$.
Let us focus on the summand
\[X={(Rn_{i_1})^{\alpha_1}}_M{(Rl_{i_1})^{\beta_1}}_M{(Rn_{i_2})^{\alpha_2}}_M{(Rl_{i_2})^{\beta_2}}_M\cdots_M{(Rn_{i_\ell})^{\alpha_\ell}}_M{(Rl_{i_\ell})^{\beta_\ell}}\]
Suppose that $\sum_{j=1}^\ell\alpha_j\geq \ell_N$. Since the product ${(Rn_{k-1})^{\beta_{k-1}}}_M(Rl_{k})^{\alpha_{k}}\subseteq (Rl_{k})^{\alpha_{k}}$,
\[X\subseteq\left( {(Rn_{i_1})^{\alpha_1}}_M{(Rn_{i_2})^{\alpha_2}}_M\cdots_M{(Rn_{i_{\ell_N}})^{\alpha_{\ell_N}}}\right) _M{(Rl_{i_{\ell_N}})^{\beta_{\ell_N}}}_M\cdots_M{(Rn_{i_\ell})^{\alpha_\ell}}_M{(Rl_{i_\ell})^{\beta_\ell}}.\]
Then $X=0$. Analogously, if $\sum_{j=1}^\ell\beta_j\geq \ell_L$, then $X=0$. Because of the choice of $\ell$ it cannot be possible that $\sum_{j=1}^\ell\alpha_j< \ell_N$ and $\sum_{j=1}^\ell\beta_j< \ell_L$. Thus, $N+L$ is locally nilpotent.
\end{proof}
\begin{prop}\label{Lfiyrad}
Suppose $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$ and let $\mathfrak{L}(M)$ be the sum of all locally nilpotent submodules of $M$. Then $\mathfrak{L}(M)$ is a fully invariant locally nilpotent submodule. Moreover $\mathfrak{L}(M/\mathfrak{L}(M))=0$.
\end{prop}
\begin{proof}
It follows from Lemma \ref{sumlocnil} that $\mathfrak{L}(M)$ is locally nilpotent. Let $f\in\End_R(M)$ be any endomorphism and let $\{f(m_1),...,f(m_k)\}$ be any subset of $f(\mathfrak{L}(M))$. Since $\mathfrak{L}(M)$ is locally nilpotent and $\{m_1,...,m_k\}\subseteq \mathfrak{L}(M)$, there exists $\ell>0$ such that any combination of $\ell$ elements of $\{m_1,...,m_k\}$ satisfies that ${Rm_{i_1}}_M{Rm_{i_2}}_M\cdots{_MRm_{i_{\ell}}}=0$. Then, by Lemma \ref{fprod},
\[{Rf(m_{i_1})}_M\left( {Rf(m_{i_2})} _M\cdots{_MRf(m_{i_{\ell}})}_MRf(m_{i_{\ell+1}})\right) \]
\[\subseteq {Rm_{i_1}}_M\left( {Rf(m_{i_2})} _M\cdots{_MRf(m_{i_{\ell}})}_MRf(m_{i_{\ell+1}})\right)\]
\[\subseteq {Rm_{i_1}}_M {Rm_{i_2}} _M\cdots{_MRf(m_{i_{\ell}})}_MRf(m_{i_{\ell+1}})\]
\[\subseteq\cdots\subseteq {Rm_{i_1}}_M{Rm_{i_2}} _M\cdots{_MRm_{i_{\ell}}} _MRf(m_{i_{\ell+1}})=0.\]
Therefore, $f(\mathfrak{L}(M))\subseteq \mathfrak{L}(M)$. Thus, $\mathfrak{L}(M)$ is fully invariant. Now, let $\pi:M\to M/\mathfrak{L}(M)$ be the canonical projection and $x+\mathfrak{L}(M)\in M/\mathfrak{L}(M)$. Consider $R(x+\mathfrak{L}(M))=\pi(Rx)$. It follows from Lemma \ref{epiproduct}, that $\mathfrak{L}(M/\mathfrak{L}(M))=0$.
\end{proof}
\begin{cor}\label{Lsp}
Suppose $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$. Then, $\mathfrak{L}(M)$ is a semiprime submodule of $M$.
\end{cor}
\begin{proof}
Let $K/\mathfrak{L}(M)$ be a submodule of $M/\mathfrak{L}(M)$ such that $\left(K/\mathfrak{L}(M) \right) ^2=0$. Hence $K/\mathfrak{L}(M)$ is nilpotent and so $K/\mathfrak{L}(M)\subseteq \mathfrak{L}(M/\mathfrak{L}(M))=0$. Thus, $M/\mathfrak{L}(M)$ is a semiprime module which implies that $\mathfrak{L}(M)$ is a semiprime submodule of $M$ by \cite[Proposition 13]{raggisemiprime}.
\end{proof}
\begin{defn}
Let $M$ be a module. The \emph{prime radical} of $M$ is defined as the intersection of all prime submodules of $M$.
\end{defn}
If $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$, in \cite[22.3]{wisbauerfoundations} it is shown that $\Rad(M)\neq M$, that is, $M$ has maximal submodules. By \cite[Corollary 4.11 and Example 4.14]{medinageneralization} each maximal submodule of $M$ contains a prime submodule, hence $Spec(M)\neq\emptyset$ and so the prime radical of $M$ is a proper submodule.
\begin{cor}\label{nesL}
Suppose $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$. Let $N$ be the prime radical of $M$. Then, $N=\mathfrak{L}(M)$.
\end{cor}
\begin{proof}
It follows from Corollary \ref{Lsp}, that $\mathfrak{L}(M)$ is a semiprime submodule. Hence, $N\subseteq \mathfrak{L}(M)$ by \cite[Proposition 1.11]{maugoldie}. Now, let $l\in \mathfrak{L}(M)$. Since $\mathfrak{L}(M)$ is a locally nilpotent submodule, $Rl$ is nilpotent. Since $N$ is a semiprime submodule, $l\in N$. Thus, $\mathfrak{L}(M)\subseteq N$.
\end{proof}
\begin{cor}\label{prnilnet}
Let $M$ be a quasi-projective module. If $M$ is Noetherian, then the prime radical of $M$ is nilpotent.
\end{cor}
\begin{proof}
It follows from Lemma \ref{fgnilp} and Corollary \ref{nesL}.
\end{proof}
\begin{example}
Let $n,p$ be integers with $p$ prime. Then, the $\mathbb{Z}$-module $\mathbb{Z}/p^n\mathbb{Z}$ is quasi-projective. It follows from Corollary \ref{prnilnet} that $\mathfrak{L}(\mathbb{Z}/p^n\mathbb{Z})=\mathbb{Z}/p^{n-1}\mathbb{Z}$ is a nilpotent submodule.
\end{example}
\begin{lemma}\label{Lsumas}
Suppose $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$. If $M=\bigoplus_{i\in I} M_i$, then $\mathfrak{L}(M)=\bigoplus_{i\in I}\mathfrak{L}(M_i)$.
\end{lemma}
\begin{proof}
It follows from Lemma \ref{proddirsumm} that $\mathfrak{L}(M_i)\subseteq \mathfrak{L}(M)$ for all $i\in I$. Let $\pi_j:M\to M_j$ be the canonical projection. Since $M_j\leq M$, we can see $\pi$ as an endomorphism of $M$. Hence $\pi(\mathfrak{L}(M))\subseteq M_j\cap \mathfrak{L}(M)$ by Proposition \ref{Lfiyrad}. Now, let $X\leq M$ be a locally nilpotent submodule such that $X\leq M_j$. Since $M_j$ is a direct summand of $M$, $X\leq \mathfrak{L}(M_j)$ by Lemma \ref{proddirsumm}. Therefore $\pi(\mathfrak{L}(M))\leq \mathfrak{L}(M_j)$. Hence $\pi(\mathfrak{L}(M))\leq \mathfrak{L}(M_i)$ for all $i\in I$. It follows that $\mathfrak{L}(M)\subseteq \bigoplus_{i\in I} \mathfrak{L}(M_i)$. Thus, $\mathfrak{L}(M)=\bigoplus_{i\in I} \mathfrak{L}(M_i)$.
\end{proof}
\begin{prop}
Suppose $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$ and that $M=\bigoplus_{i\in I}M_i$. Then, $M$ is semiprime if and only if $M_i$ is a semiprime module for all $i\in I$.
\end{prop}
\begin{proof}
It follows from Lemma \ref{Lsumas} that
\[\mathfrak{L}(M)=\mathfrak{L}\left(\bigoplus_{i\in I} M_i \right)=\bigoplus_{i\in I} \mathfrak{L}(M_i).\]
Thus, $\mathfrak{L}(M)=0$ if and only if $\mathfrak{L}(M_i)=0$ for all $i\in I$.
\end{proof}
\begin{cor}\label{rsp}
The following conditions are equivalent for a ring $R$.
\begin{enumerate}[label=\emph{(\alph*)}]
\item $R$ is a semiprime ring.
\item Every free left (right) $R$-module is semiprime.
\item Every projective left (right) $R$-module is semiprime.
\end{enumerate}
\end{cor}
\begin{cor}
The following conditions are equivalent for a ring $R$:
\begin{enumerate}[label=\emph{(\alph*)}]
\item The prime radical of $R$ is nilpotent.
\item The prime radical of every left (right) free $R$-module is nilpotent.
\item The prime radical of every left (right) projective $R$-module is nilpotent.
\end{enumerate}
\end{cor}
\begin{proof}
(a)$\Rightarrow$(b) Consider a free $R$-module $R^{(\Lambda)}$. Then $\mathfrak{L}\left(R^{(\Lambda)} \right)=\left(\mathfrak{L}(R) \right)^{(\Lambda)}$ by Lemma \ref{Lsumas}. By hypothesis, there exists $n>0$ such that $\mathfrak{L}(R)^n=0$. Suppose $n=2$. Using Lemma \ref{proddirsumm} and the properties of the product we get
\[{\left(\mathfrak{L}(R) \right)^{(\Lambda)}}_{R^{(\Lambda)}}\left(\mathfrak{L}(R) \right)^{(\Lambda)}=\left[ {\left(\mathfrak{L}(R) \right)^{(\Lambda)}}_{R^{(\Lambda)}}\mathfrak{L}(R)\right] ^{(\Lambda)}=\left[\sum_{\Lambda}\left(\mathfrak{L}(R) _{R^{(\Lambda)}}\mathfrak{L}(R) \right)\right]^{(\Lambda)}\]
\[=\left[\sum_{\Lambda}\left(\mathfrak{L}(R) _{R}\mathfrak{L}(R) \right)\right]^{(\Lambda)}=\left[\sum_{\Lambda}\left(\mathfrak{L}(R)^2\right)\right]^{(\Lambda)}=0.\]
Thus,
\[\left(\mathfrak{L}\left(R^{(\Lambda)}\right) \right)^n=\left[\sum_{\Lambda^{n-1}}\left(\mathfrak{L}(R)^n\right)\right]^{(\Lambda)}=0.\]
Therefore, the prime radical of $R^{(\Lambda)}$ is nilpotent.
(b)$\Rightarrow$(c) Let $P$ be a projective $R$-module. Then there exists a free $R$-module $R^{(\Lambda)}$ such that $R^{(\Lambda)}=P\oplus P'$. Hence $\mathfrak{L}\left( R^{(\Lambda)}\right) =\mathfrak{L}(P)\oplus\mathfrak{L}(P')$. By hypothesis, there exists $n>0$ such that $\mathfrak{L}\left( R^{(\Lambda)}\right)^n=0$. Therefore $\left(\mathfrak{L}(P)\oplus\mathfrak{L}(P') \right)^n=0$. Suppose $n=2$. Then
\[0=\left(\mathfrak{L}(P)\oplus\mathfrak{L}(P') \right)^2=\left[\mathfrak{L}(P)\oplus\mathfrak{L}(P')\right]_{R^{(\Lambda)}}\left[\mathfrak{L}(P)\oplus\mathfrak{L}(P')\right]\]
\[=\left( \left[\mathfrak{L}(P)\oplus\mathfrak{L}(P')\right]_{R^{(\Lambda)}}\mathfrak{L}(P)\right) \oplus\left( \left[\mathfrak{L}(P) \oplus \mathfrak{L}(P')\right]_{R^{(\Lambda)}}\mathfrak{L}(P')\right).\]
On the other hand,
\[\mathfrak{L}(P)_P\mathfrak{L}(P)=\mathfrak{L}(P)_{R^{(\Lambda)}}\mathfrak{L}(P)\subseteq\left[\mathfrak{L}(P)\oplus\mathfrak{L}(P')\right]_{R^{(\Lambda)}}\mathfrak{L}(P)=0.\]
This implies that $\mathfrak{L}(P)$ is a nilpotent submodule of $P$.
(c)$\Rightarrow$(a) is trivial.
\end{proof}
\section{Nil-submodules and acc on annihilators}\label{AccandNil}
For a ring $R$ and $Y\subseteq R$, the left (resp. right) annihilator of $Y$ in $R$ is denoted by $\ann_R^\ell(Y)$ (resp. $\ann_R^r(Y)$). For a subset $X\subseteq M$ of an $R$-module $M$ with $S=\End_R(M)$, the left annihilator of $X$ in $S$ is denoted by $\mathbf{l}_S(X)=\{f\in S\mid f(X)=0\}$. If $\left\langle X\right\rangle $ is the $R$-submodule generated by $X\subseteq M$ then $\mathbf{l}_S(X)=\mathbf{l}_S\left( \left\langle X\right\rangle \right)$. On the other hand, an $R$-module $M$ is said to have the \emph{acc on annihilators} if any chain
\[\bigcap_{f\in Y_1}\Ker f\subseteq\bigcap_{f\in Y_2}\Ker f\subseteq\cdots\]
with $Y_i\subseteq \End_R(M)$ ($i>0$) becomes stationary in finitely many steps.
\begin{prop}\label{maccsacc}
If $M$ satisfies acc on annihilators, then $S=\End_R(M)$ satisfies acc on right annihilators.
\end{prop}
\begin{proof}
We will show that $S$ satisfies the descending chain condition (dcc) on left annihilators. Let
\[\ann_S^\ell(X_1)\supseteq \ann_S^\ell(X_2)\supseteq\cdots\]
be a descending chain of left annihilators in $S$. Then, there is an ascending chain
\[\bigcap_{f\in\ann_S^\ell(X_1)}\Ker f\subseteq \bigcap_{f\in\ann_S^\ell(X_2)}\Ker f\subseteq\cdots\]
in $M$. By hypothesis, there exists a positive integer $k$ such that $\bigcap_{f\in\ann_S^\ell(X_k)}\Ker f= \bigcap_{f\in\ann_S^\ell(X_{k+i})}\Ker f$ for all $i>0$. We claim that
\[\mathbf{l}_S\left( \bigcap_{f\in\ann_S^\ell(X_k)}\Ker f\right) =\ann_S^\ell(X_1).\]
Let $f\in\ann_S^\ell(X_k)$. It follows that $f\left( \bigcap_{f\in\ann_S^\ell(X_k)}\Ker f\right)=0$. Hence, $\ann_S^\ell(X_k)\subseteq\mathbf{l}_S\left( \bigcap_{f\in\ann_S^\ell(X_k)}\Ker f\right)$. Now, let $g\in\mathbf{l}_S\left( \bigcap_{f\in\ann_S^\ell(X_k)}\Ker f\right)$ and $h\in X_k$. It follows that $fh=0$ for all $f\in\ann_S^\ell(X_k)$, i.e., $h(M)\subseteq\bigcap_{f\in\ann_S^\ell(X_k)}\Ker f$. Therefore, $gh=0$. Thus, $g\in\ann_S^\ell(X_k)$. Then,
\[\ann_S^\ell(X_k)=\mathbf{l}_S\left( \bigcap_{f\in\ann_S^\ell(X_k)}\Ker f\right) = \mathbf{l}_S\left( \bigcap_{f\in\ann_S^\ell(X_{k+i})}\Ker f\right) =\ann_S^\ell(X_{k+i})\]
for all $i>0$. This implies that $S$ satisfies acc on right annihilators.
\end{proof}
Notice that the converse of Proposition \ref{maccsacc} is not true in general, as the next example shows.
\begin{example}
Consider the $\mathbb{Z}$-module $M=\mathbb{Z}_{p^\infty}$ with $p$ a prime number. It is known that $S=\End_\mathbb{Z}(M)$ is isomorphic to the ring of $p$-adic numbers. Thus, $S$ is a commutative Noetherian ring. On the other hand, each submodule of $M$ is the kernel of an endomorphism of $M$, hence $M$ does not satisfy the acc on annihilators.
\end{example}
\begin{lemma}\label{nilpsubnil}
Let $M$ be $M^{(\Lambda)}$-projective for every index set $\Lambda$ and let $K<N$ fully invariant submodules of $M$ with $N$ a nil-submodule.
If $M$ satisfies acc on annihilators, then $N/K$ contains a nonzero nilpotent submodule.
\end{lemma}
\begin{proof}
If $\Hom_R(M/K,N/K)=0$, then $N/K$ is nilpotent. Also, if $\Hom(M,N)=0$, then $\Hom_R(M/K,N/K)=0$ by the projectivity condition. Hence, we can assume that $\Hom_R(M/K,N/K)\neq0$ and $\Hom_R(M,N)\neq 0$. Given $f\in\Hom_R(M,N)$, since $K$ is fully invariant, there is a homomorphism $\overline{f}:M/K\to N/K$. On the other hand, since $M$ is quasi-projective, for any $g\in\Hom_R(M/K,N/K)$ there exists $f\in\Hom_R(M,N)$ such that $\pi f=g\pi$ where $\pi:M\to M/K$ is the canonical projection. Hence, there is a surjective homomorphism of rings without one $\Hom_R(M,N)\to \Hom_R(M/K,N/K)$. By hypothesis and Lemma \ref{factornil} we have that, $\Hom_R(M,N)$ and $\Hom_R(M/K,N/K)$ are nil-ideals of $\End_R(M)$ and $\End_R(M/K)$ respectively. It follows from Proposition \ref{maccsacc} and \cite[Lemma 1]{herstein1964nil} that $\Hom_R(M/K,N/K)$ contains a nilpotent ideal. In particular, there is $g\in\Hom_R(M/K,N/K)$ such that $gT$ is nilpotent, say $(gT)^n=0$, where $T=\End_R(M/K)$. It follows from \cite[18.4]{wisbauerfoundations} that $gT=\Hom_R(M/K,gTM/K)$. Therefore,
\begin{equation*}
\begin{split}
(gT(M/K))^{n+1} & ={gT(M/K)^n}_{M/K}gT(M/K) \\
& =\Hom_R(M/K,gT(M/K))gT(M/K)^n \\
& =\Hom_R(M/K,gT(M/K))^2gT(M/K)^{n-1} \\
& =\cdots \\
& =\Hom_R(M/K,gT(M/K))^ngT(M/K)=0.
\end{split}
\end{equation*}
\end{proof}
\begin{prop}\label{accnillocnil}
Let $M$ be $M^{(\Lambda)}$-projective for every index set $\Lambda$ and let $N$ be a fully invariant
nil-submodule of $M$. If $M$
satisfies acc on annihilators, then $N$ is locally nilpotent.
\end{prop}
\begin{proof}
Since $\mathfrak{L}(M)$ is a semiprime (fully invariant) submodule of $M$, $M/\mathfrak{L}(M)$ is projective in $\sigma[M/\mathfrak{L}(M)]$ and $M/\mathfrak{L}(M)$ is retractable because it is a semiprime module. By Lemma \ref{factornil}, $\frac{N+\mathfrak{L}(M)}{\mathfrak{L}(M)}$ is a nil-submodule of $M/\mathfrak{L}(M)$. Suppose $N\nsubseteq \mathfrak{L}(M)$. By Lemma \ref{nilpsubnil}, $\frac{N+\mathfrak{L}(M)}{\mathfrak{L}(M)}$ contains a nonzero nilpotent submodule $\frac{K}{\mathfrak{L}(M)}$. Hence, there exists $\ell>0$ such that $K^\ell\subseteq \mathfrak{L}(M)$ by Lemma \ref{epiproduct}. It follows from Corollary \ref{Lsp} that $K\subseteq \mathfrak{L}(M)$. Hence $\frac{K}{\mathfrak{L}(M)}=0$. Thus, $\frac{N+\mathfrak{L}(M)}{\mathfrak{L}(M)}=0$, that is, $N\subseteq \mathfrak{L}(M)$.
\end{proof}
\begin{defn}
Let $M$ be a module and let $N\leq M$. The \emph{annihilator of $N$ in $M$} is the submodule
\[\ann_M(N)=\bigcap\{\Ker f\mid f\in\Hom_R(M,N)\}.\]
The \emph{right annihilator of $N$ in $M$} is given by the submodule
\[\ann^r_M(N)=\sum\{K\leq M\mid N_MK=0\}.\]
\end{defn}
\begin{rem}
It is not difficult to see that $\ann_M(N)$ is a fully invariant submodule of $M$ and it is the largest submodule of $M$ such that $\ann_M(N)_MN=0$. When $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$, $\ann^r_M(N)$ is fully invariant and is the largest submodule of $M$ such that $N_M\ann^r_M(N)=0$ \cite[Remark 1.15]{mauacc}.
\end{rem}
\begin{lemma}\label{rannintersection}
Let $M$ be a module and let $\{N_i\}_I$ be a family of submodules of $M$. Then $\bigcap_{i\in I}\ann^r_M(N_i)=\ann^r_M(\sum_{i\in I}N_i)$.
\end{lemma}
\begin{proof}
We always have that $\ann^r_M(\sum_{i\in I}N_i)\subseteq \ann^r_M(N_i)$ for all $i\in I$. Hence $\ann^r_M(\sum_{i\in I}N_i)\subseteq\bigcap_{i\in I} \ann^r_M(N_i)$. On the other hand,
\[\left( \sum_{i\in I}N_i\right) _M\left( \bigcap_{i\in I}\ann^r_M(N_i)\right)=\sum_{i\in I}\left( {N_i}_M\bigcap_{i\in I}\ann^r_M(N_i)\right) =0.\]
Thus, $\bigcap_{i\in I}\ann^r_M(N_i)=\ann^r_M(\sum_{i\in I}N_i)$.
\end{proof}
\begin{lemma}\label{dccannr}
Let $M$ be a module. If $M$ satisfies acc on annihilators, then the set $\{\ann^r_M(N)\mid N\leq M\}$ satisfies dcc.
\end{lemma}
\begin{proof}
We just have to notice that $\ann^r_M(\ann_M(\ann^r_M(N)))=\ann^r_M(N)$.
\end{proof}
\begin{lemma}\label{dccl}
Suppose $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$. Then,
\begin{enumerate}
\item $\mathbf{l}_S\left( \bigcap_{f\in\mathbf{l}_S(X)}\Ker f\right)=\mathbf{l}_S(X)$ for any subset $X$ of $M$.
\item $\bigcap_{f\in Y}\Ker f=\bigcap\left\lbrace \Ker g\mid g\in\mathbf{l}_S\left( \bigcap_{f\in Y}\Ker f\right) \right\rbrace $ for any subset $Y$ of $\End_R(M)$.
\item $M$ satisfies acc on annihilators if and only if the set $\{\mathbf{l}_S(X)\mid X\subseteq M\}$ satisfies dcc.
\end{enumerate}
\end{lemma}
\begin{proof}
\textit{(1)} Since $X\subseteq\bigcap_{f\in\mathbf{l}_S(X)}\Ker f$, we have that $\mathbf{l}_S\left( \bigcap_{f\in\mathbf{l}_S(X)}\Ker f\right)\subseteq\mathbf{l}_S(X)$. The other inclusion is obvious.
\textit{(2)} It is clear that $\bigcap_{f\in Y}\Ker f\subseteq\bigcap\left\lbrace \Ker g\mid g\in\mathbf{l}_S\left( \bigcap_{f\in Y}\Ker f\right) \right\rbrace$. On the other hand, since $Y\subseteq\mathbf{l}_S\left( \bigcap_{f\in Y}\Ker f\right)$, we have the other inclusion.
\textit{(3)} It is clear from \textit{(1)} and \textit{(2)}.
\end{proof}
\begin{prop}\label{factorrightacc}
Suppose $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$ and that $M$ satisfies acc on annihilators. Let $\ann^r_M(N)$ be a right annihilator in $M$. Then $\overline{M}=M/\ann^r_M(N)$ satisfies acc on annihilators.
\end{prop}
\begin{proof}
Since $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$, there is a surjetive ring homomorphism $\rho:\End_R(M)\to \End_R(\overline{M})$. Let $\overline{X}$ be any subset of $\overline{M}$ and set $\overline{S}=\End_R(\overline{M})$. We claim that the inverse image under $\rho$ of $\mathbf{l}_{\overline{S}}(\overline{X})$ is $\mathbf{l}_S(N_M\left\langle X\right\rangle)$ where $X$ is the inverse image of $\overline{X}$ under the canonical projection $\pi:M\to \overline{M}$. For each $\bar{f}\in\mathbf{l}_{\overline{S}}(\overline{X})$ there is an $f\in\End_R(M)$ such that $f(X)\subseteq \ann^r_M(N)$. Note that $f(\left\langle X\right\rangle)=\left\langle f(X)\right\rangle \leq \ann^r_M(N)$. By Lemma \ref{fprod}, $f(N_M\left\langle X\right\rangle )=N_Mf(\left\langle X\right\rangle )=0$. Therefore, $f\in\mathbf{l}_S(N_M\left\langle X\right\rangle )$. This implies that $\rho^{-1}(\mathbf{l}_{\overline{S}}(\overline{X}))\subseteq\mathbf{l}_S(N_M\left\langle X\right\rangle )$. Now, let $g\in\mathbf{l}_S(N_M\left\langle X\right\rangle )$. Again by Lemma \ref{fprod}, $0=g(N_M\left\langle X\right\rangle )=N_Mg(\left\langle X\right\rangle )$. Hence $g(\left\langle X\right\rangle )\leq \ann^r_M(N)$. Therefore, $\rho(g)(\overline{X})=\rho(g)\pi(X)=\pi g(X)=0$. Thus, $\mathbf{l}_S(N_M\left\langle X\right\rangle )\subseteq\rho^{-1}(\mathbf{l}_{\overline{S}}(\overline{X}))$, proving the claim. It follows from Lemma \ref{dccl} that $\overline{M}$ satisfies acc on annihilators.
\end{proof}
Let $M$ be a module and $K\leq N\leq M$. Let $\mathbf{l}_N(K)$ denote the intersection $\ann_M(K)\cap N$. On the other hand, let $\mathbf{r}_N(K)=\ann^r_M(K)\cap N$. If $N$ is a fully invariant submodule of $M$ and $K\leq N$, then $\mathbf{l}_N(K)$ and $\mathbf{r}_N(K)$ are fully invariant in $M$.
\begin{cor}\label{dccrn}
Let $M$ be a module. If $M$ satisfies acc on annihilators then the set $\{\mathbf{l}_N(K)\mid K\leq N\}$ satisfies acc and the set $\{\mathbf{r}_N(K)\mid K\leq N\}$ satisfies dcc.
\end{cor}
\begin{proof}
Let $K\leq N\leq M$. We claim that $\mathbf{l}_N(K)=\ann_M(\ann^r_M(\mathbf{l}_N(K)))\cap N$. Since $\mathbf{l}_N(K)_M\ann^r_M(\mathbf{l}_N(K))=0$, we have that $\mathbf{l}_N(K)\subseteq \ann_M(\ann^r_M(\mathbf{l}_N(K)))\cap N$. On the other hand, $\mathbf{l}_N(K)\subseteq \ann_M(K)$. Hence $\ann^r_M(\mathbf{l}_N(K))\supseteq \ann^r_M(\ann_M(K))$. Therefore, $\ann_M(\ann^r_M(\mathbf{l}_N(K)))\subseteq \ann_M(\ann^r_M(\ann_M(K)))=\ann_M(K)$. Thus, $\ann_M(\ann^r_M(\mathbf{l}_N(K)))\cap N\subseteq \ann_M(K)\cap N=\mathbf{l}_N(K)$ proving the claim. Now, if
\[\mathbf{l}_N(K_1)\subseteq\mathbf{l}_N(K_2)\subseteq\mathbf{l}_N(K_3)\subseteq\cdots\]
is an ascending chain, applying $\ann^r(-)$ to the chain, we get a descending chain
\[\ann^r_M(\mathbf{l}_N(K_1))\supseteq \ann^r_M(\mathbf{l}_N(K_2))\supseteq \ann^r_M(\mathbf{l}_N(K_3))\supseteq\cdots\]
By Lemma \ref{dccannr}, there exists $\ell>0$ such that $\ann^r_M(\mathbf{l}_N(K_\ell))= \ann^r_M(\mathbf{l}_N(K_{\ell+i}))$ for all $i\geq 0$. It follows that
\[\mathbf{l}_N(K_\ell)=\ann_M(\ann^r_M(\mathbf{l}_N(K_\ell)))\cap N = \ann_M(\ann^r_M(\mathbf{l}_N(K_{\ell+i})))\cap N=\mathbf{l}_N(K_{\ell+i})\]
for all $i\geq 0$. Analogously, the set $\{\mathbf{r}_N(K)\mid K\leq N\}$ satisfies dcc.
\end{proof}
\begin{lemma}\label{rannncero}
Suppose $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$ and let $N< M$ be a fully invariant nil-submodule. If $M$
satisfies acc on annihilators, then $\mathbf{r}_N(N)\neq 0$.
\end{lemma}
\begin{proof}
Set $\Gamma=\{\mathbf{r}_N(K)\mid K\leq N\text{ and } K \text{ finitely generated}\}$. It follows from Corollary \ref{dccrn} that $\Gamma$ has minimal elements. Let $\mathbf{r}_N(K)$ be a minimal element in $\Gamma$ and let $n\in N$. Then $K+Rn$ is a finitely generated submodule of $N$ and so $\mathbf{r}_N(K)=\mathbf{r}_N(K+Rn)\subseteq\mathbf{r}_N(Rn)$. Hence, $\mathbf{r}_N(K)\subseteq\bigcap_{n\in N}\mathbf{r}_N(Rn)=\bigcap_{n\in N}\ann^r_M(Rn)\cap N=\ann^r_M(\sum_{n\in N}Rn)\cap N=\ann^r_M(N)\cap N=\mathbf{r}_N(N)$ by Lemma \ref{rannintersection}.
Now, by Lemma \ref{fgnilp} and Proposition \ref{accnillocnil}, $K$ is nilpotent, that is, there exists $\ell>1$ such that $K^\ell=0$ but $K^{\ell-1}\neq 0$. Therefore $0\neq K^{\ell-1}\subseteq\mathbf{r}_N(K)$.
\end{proof}
\begin{prop}\label{subm}
Suppose $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$ and let $N<M$ be a fully invariant nil-submodule such that $\mathbf{l}_N(N^j)=0$ for all $j>0$. If $M$
satisfies acc on annihilators, then there exist submodules $A_1,...,A_k,...\subseteq N$ such that:
\begin{enumerate}
\item ${A_1}_M\cdots_M{A_k}\neq 0$ for all $k>0$; and
\item ${A_1}_M\cdots_M{A_k}_MA_n=0$ if $n\leq k$.
\end{enumerate}
\end{prop}
\begin{proof}
We construct the submodules inductively. Let $A_1=\mathbf{r}_N(N)\neq 0$ by Lemma \ref{rannncero}. Note that ${A_1}_M{A_1}=0$. To get $A_2$, consider $\ann^r_M(A_1)$. Then $\frac{N+\ann^r_M(A_1)}{\ann^r_M(A_1)}$ is a nil-submodule of $M/\ann^r_M(A_1)$ by Lemma \ref{factornil}. Moreover, $M/\ann^r_M(A_1)$
satisfies acc on annihilators by
Proposition \ref{factorrightacc}. Write $\overline{N}=\frac{N+\ann^r_M(A_1)}{\ann^r_M(A_1)}$ and $\overline{M}=M/\ann^r_M(A_1)$. By Lemma \ref{rannncero}, $\mathbf{r}_{\overline{N}}(\overline{N})\neq 0$. Hence there exists $0\neq T/\ann^r_M(A_1)\leq \overline{M}$ such that $\overline{N}_{\overline{M}}\left( T/\ann^r_M(A_1)\right) =0$. This implies that $N_MT\subseteq \ann^r_M(A_1)$ and so ${A_1}_M{N}_MT=0$ but ${A_1}_MT\neq 0$. Let $A_2=\mathbf{r}_N({A_1}_MN)$. Therefore ${A_1}_M{A_2}\neq 0$. We have that ${A_1}_M{A_2}\subseteq N$, then ${A_1}_M{A_2}_M{A_1}=0$. On the other hand, ${A_1}_M{A_2}\subseteq {A_1}_MN$. Hence, ${A_1}_M{A_2}_M{A_2}\subseteq {A_1}_MN_M{A_2}=0$ because definition of $A_2$. Inductively, each $A_{i+1}=\mathbf{r}_N({A_1}_M\cdots_M{A_i}_MN)$.
\end{proof}
\section{Nilpotency of the prime radical of a Goldie module}\label{Nilpotency}
In this section, the main theorem is proved and as corollary we get that the prime radical of a Goldie module $M$ is nilpotent. In \cite[Theorem 1]{lanski1969nil} Lanski proved that nil subrings of a left Goldie ring are nilpotent. We will make use of this result applied to the endomorphism ring of a Goldie module. For, we start with the following lemmas.
\begin{lemma}\label{accmoduloann}
Let $M$ be a quasi-projective module with endomorphism ring $S=\End_R(M)$ and $N=\bigcap_{f\in I}\Ker f$ for some ideal $I$ of $S$. If $M$ satisfies acc on annihilators then so does $M/N$.
\end{lemma}
\begin{proof}
Note that $N$ is a fully invariant submodule of $M$ because $I$ is an ideal. It is enough to show that inverse image, under the canonical projection $\pi:M\to M/N$, of an annihilator in $M/N$ is an annihilator in $M$. Let $\overline{Y}\subseteq\End_R(M/N)$. Consider $\bigcap_{\bar{f}\in\overline{Y}}\Ker \bar{f}$ in $M/N$. Hence $\bigcap_{\bar{f}\in\overline{Y}}\Ker \bar{f}=A/N$ for some submodule $A$ of $M$. Since $M$ is quasi-projective, given $\bar{f}\in\overline{Y}$ there exists $f\in\End_R(M)$ such that $\pi f=\bar{f}\pi$. Let $\widehat{Y}=\{f\in\End_R(M)\mid \pi f=\bar{f}\pi\text{ for some }\bar{f}\in\overline{Y}\}$. It follows that $f(A)\subseteq N$ for all $f\in \widehat{Y}$ and so, $gf(A)=0$ for all $f\in \widehat{Y}$ and all $g\in I$. Let $Y=\{gf\mid g\in I\;\text{and}\;f\in \widehat{Y}\}$. Therefore, $A\subseteq\bigcap_{gf\in Y}\Ker gf$. Now, let $x\in\bigcap_{gf\in Y}\Ker gf$ and let $\bar{f}\in\overline{Y}$. Then $\bar{f}\pi(x)=\pi(f(x))$. Since $g(f(x))=0$ for all $g\in I$, $f(x)\in N$. Hence $\pi(f(x))=0$. It follows that $x\in A$. Thus $A=\bigcap_{gf\in Y}\Ker gf$.
\end{proof}
Recall that a module $M$ is said to be \emph{retractable} if $\Hom_R(M,N)\neq 0$ for every nonzero submodule $N$ of $M$.
\begin{lemma}\label{fmret}
Let $M$ be $M^{(\Lambda)}$-projective for every index set $\Lambda$ and let $N\leq M$. If $M$ is retractable, then
\begin{enumerate}
\item $M/\ann_M^r(N)$ is retractable.
\item $M/\ann_M(N)$ is retractable.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) Let $K/\ann_M^r(N)$ be a submodule of $M/\ann_M^r(N)$ and suppose that
\[\Hom_R(M/\ann_M^r(N),K/\ann_M^r(N))=0.\]
Since $\ann_M^r(N)$ is fully invariant in $M$, any homomorphism $f:M\to K$ induces a homomorphism $\overline{f}:M/\ann_M^r(N)\to K/\ann_M^r(N)$ such that $\pi f=\overline{f}\pi$ where $\pi:M\to M/\ann_M^r(N)$ is the canonical projection. This implies that for all $f\in\Hom_R(M,K)$, $f(M)\subseteq \ann_M^r(N)$. Therefore $0=N_M(M_MK)=(N_MM)_MK=N_MK$. Hence $K\subseteq\ann_M^r(N)$ and so $K/\ann_M^r(N)=0$. Thus $M/\ann_M^r(N)$ is retractable.
(2) Let $K/\ann_M(N)$ be a submodule of $M/\ann_M(N)$ and suppose that
\[\Hom_R(M/\ann_M(N),K/\ann_M(N))=0.\]
As in the previous proof, we get $0=tr^M(K)_MN=(M_MK)_MN=M_M(K_MN)=tr^M(K_MN)$. This implies that $K_MN=0$ because $M$ is retractable. Thus, $K\subseteq \ann_M(N)$. Hence $M/\ann_M(N)$ is retractable.
\end{proof}
\begin{lemma}\label{mgolsgol}
Suppose $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$ and retractable. If $M$ is Goldie then $S=\End_R(M)$ is a right Goldie ring.
\end{lemma}
\begin{proof}
By Proposition \ref{maccsacc}, $S$ satisfies acc on right annihilators. Since $M$ has finite uniform dimension, so does $S_S$ by \cite[Theorem 2.6]{HaghanyStudy}. Thus, $S$ is a right Goldie ring.
\end{proof}
\begin{thm}\label{main}
Suppose $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$ and retractable. If $M$ is a Goldie module and $N<M$ is a fully invariant nil-submodule, then $N$ is nilpotent.
\end{thm}
\begin{proof}
There is an ascending chain of submodules of $M$,
\[\ann_M (N)\subseteq \ann_M (N^2)\subseteq \ann_M (N^3)\subseteq\cdots\]
By hypothesis, there exists a positive integer $k$ such that $\ann_M (N^k)=\ann_M (N^{k+i})$ for all $i>0$. Denote $\overline {N}=(N+\ann_M(N^k))/\ann_M (N^k)$ and $\overline {M}=M/\ann_M(N^k)$. We claim that $\mathbf{l}_{\overline {N}}(\overline {N}^j)=0$ for all $j>0$.
If $\mathbf{l}_{\overline {N}}(\overline {N}^j)=A/\ann_M(N^k)$ with $A\leq M$, then $A_MN^j\leq \ann_M(N^k)$. Hence, $A_MN^{k+j}=(A_MN^j)_MN^{k}=0$. Therefore $A\subseteq \ann_M(N^{k+j})=\ann_M(N^k)$. Thus $\mathbf{l}_{\overline {N}}(\overline {N}^j)=0$, proving the claim.
If $N$ is not nilpotent, then $\overline {N}$ is a nil-submodule of $\overline {M}$ by Lemma \ref{factornil}. Moreover, $\overline {M}$ is retractable and satisfies acc on annihilators by Lemma \ref{fmret} and Lemma \ref{accmoduloann}. Hence, there exist nonzero submodules $\overline{A_1},...,\overline{A_i},...\subseteq \overline{N}$ by Proposition \ref{subm} such that
\begin{itemize}
\item[(i)] ${\overline{A_1}}_{\overline{M}}\cdots_{\overline{M}}{\overline{A_i}}\neq 0$ for all $i>0$; and
\item[(ii)] ${\overline{A_1}}_{\overline{M}}\cdots_{\overline{M}}{\overline{A_i}}_{\overline{M}}\overline{A_j}=0$ if $j\leq i$.
\end{itemize}
Consider the inverse images of the submodules $\overline{A_i}$, $\ann_M(N^{k})\subseteq
A_1,..,A_i,...\subseteq N+\ann_M(N^k)$. Intersecting with $N$, we have
\[\mathbf{l}_N(N)\subseteq \ann_M(N^k)\cap N \subseteq (A_1\cap N),...,(A_i\cap N),...\subseteq N.\]
Suppose that $\mathbf{l}_N(N)\neq 0$. Then $A_i\cap N\neq 0$. Let $B_i$ denote the submodule $A_i\cap N$. Hence, the submodules $B_1,...,B_i,...$ satisfy:
\begin{enumerate}
\item ${B_1}_M\cdots_M{B_i}\neq 0$ for all $i>0$; and
\item ${B_1}_M\cdots_M{B_i}_M{B_j}\subseteq \mathbf{l}_N(N^k)$ if $j\leq i$.
\end{enumerate}
Condition (1) implies that there exists a sequence $\{f_1,f_2,...,f_i,...\}\subseteq\Hom_R(M,N)$, with $f_i:M\to B_i$ for all $i>0$, such that $f_i\cdots f_2f_1\neq 0$ for all $i>0$ and the condition (2) says that $g(f_jf_i\cdots f_2f_1(M))=0$ for all $g:M\to N^k$ if $j\leq i$. Note that $f_{i+1}f_i(M)\subseteq f_{i+1}(N)\subseteq N^2$. Consider $\{f_{j+1},...,f_{j+k}\}$ with $j\leq i$. Then $f_{j+k}\cdots f_{j+1}(M)\subseteq N^k$, that is, $f_{j+k}\cdots f_{j+1}:M\to N^k$. It follows that $f_{j+k}\cdots f_{j+1}f_jf_i\cdots f_2f_1=0$ for $j\leq i$. Thus, we just got the elements described in \cite[Lemma 8]{lanski1969nil}, that is, elements $\{f_1,f_2,...,f_i,...\}\subseteq S$ such that
\begin{enumerate}
\item $f_i\cdots f_2f_1\neq 0$ for all $i>0$; and
\item $f_{j+k}\cdots f_{j+1}f_jf_i\cdots f_2f_1=0$ for $j\leq i$.
\end{enumerate}
Since $M$ is a Goldie module, if follows from Lemma \ref{mgolsgol} that $S$ is a right Goldie ring. Following the proof of \cite[Theorem 1]{lanski1969nil}, we arrive to a contradiction and so $N$ is nilpotent.
In the case $\mathbf{l}_N(N)=0$, then we start the proof with $M=\overline{M}$ and $A_i=\overline{A_i}$ for all $i>0$.
\end{proof}
\begin{cor}\label{primenilgoldie}
Suppose $M$ is $M^{(\Lambda)}$-projective for every index set $\Lambda$ and retractable. If $M$ is a Goldie module, then the prime radical of $M$ is nilpotent.
\end{cor}
|
\section{Introduction}
\pagenumbering{arabic}
The need for \textit{real-time} communication of status update packets involves maintaining information freshness in many mission-critical applications. Age of Information (AoI) is a metric measuring the freshness of information at the receiver of a status update system. Since its introduction in \cite{ kaul2012real} for queuing models motivated from vehicular status update systems, the AoI metric has been found useful in numerous applications that require timely availability of information at the receiving end of a communication or remote-controlled system, such as tele-surgery or skill transfer in tactile Internet.
While AoI provides a measure of the freshness of information, it does not account for soft or hard deadlines that applications impose on the age. Motivated by the need to consider deadlines, practitioners have recently considered a threshold for the age and investigated when the threshold is exceeded.
In \cite{zhang2021aoi}, the authors propose a freshness-aware refreshing scheme in which the cached content items will be refreshed to the up-to-date version upon user request if the AoI exceeds a certain threshold. In \cite{li2020aoi}, the authors consider scheduling problems at the network edge when each source node has an AoI requirement which is called Maximum AoI Threshold (MAT). A threshold-based ALOHA scheme, in which each terminal attempts transmission with constant probability in each slot when the age exceeds the threshold is studied in \cite{yavascan2021analysis}. In \cite{jiang2020analyzing}, the authors propose a simple threshold policy that achieves the optimum asymptotically for AoI in a wireless multi-access network with the thresholds explicitly derived. We considered a variant of threshold-based age in our previous paper~\cite{zou2020age} wherein an update packet has a value associated with it and the value decays with time and becomes zero beyond a threshold.
In this paper, we expand upon the initial work on threshold-based age and present three metrics of relevance in various status update systems. Given a threshold for the instantaneous age, the metrics are defined as follows. The {\em overage probability} is the probability that the age of the most recent packet currently held by the receiver is larger than the threshold. The {\em stale update probability} is the probability that a delivered update is stale, i.e., its age has exceeded the threshold, when it is delivered to the receiver. This metric is relevant when the receiving system takes action immediately upon the receipt of the packet. Finally, the {\em average overage} is defined as the time average of the overage (i.e., age beyond the threshold), and is a measure of the average ``staleness'' of the update packet held by the receiver. We investigate these three metrics in three typical status update queuing systems, namely, M/G/1/1, M/G/1/$2^*$, and M/M/1, and obtain closed-form expressions for these metrics. We present numerical results for these metrics under different parameter settings and service time distributions. We also show a comparison between the average overage and the average AoI.
The overage probability has been studied in a few scenarios previously.
In \cite{pan2021coding}, the authors use bounded AoI to denote the probability that the AoI is under a given threshold, which is the opposite of our overage probability metric. The authors of \cite{Inoue2019} obtain the distribution of AoI under different queuing schemes, which can in turn be used to derive the overage probability. In \cite{Devassy2018,Zheng2018,hsu2020age,seo2019outage,franco2019analysis,kim2020sensing,hu2021status,hou2021,zhang2021,Champati2019}, the probability that the age at the receiver is larger than the given threshold is denoted as violation probability or outage probability. While much research has been done for the threshold model in AoI optimization problems, little attention has been paid to the definition and derivation of threshold-based metrics. In this paper, we call this probability the overage probability, and analyze it for the three queuing systems; we note that the overage probability has not been analyzed for the M/G/1/1 and M/G/1/2$^{*}$ systems before. Further, to the best of our knowledge, the other two metrics we propose here have not been considered before. Thus, our contributions in this paper include: (a) the proposal of two new metrics, (b) the derivation of expressions for all metrics for three commonly studied queuing systems, and (c) numerical results exploring various trade-offs.
\section{System Model}
We consider a point-to-point communication system with a transmitter sending status updates from a single source to a receiver as shown in Fig. \ref{fig:1}. The update packets arrive at the transmitter following a Poisson process with an arrival rate $\lambda$. A packet may be discarded in the queuing phase based on buffer size and the packet dropping policy; those that are not discarded enter the server and are received by the receiver after a random service time. Here, we only index those packets that are received by the receiver and use $S_i$ to denote the service time for the $i$th packet. $S_i$ is assumed to be iid with pdf $f_S(s)$. In this paper, we consider M/GI/1/1, M/GI/1/$2^*$, and M/M/1 queuing schemes. In M/GI/1/1, there is no buffer and packets arriving in busy state are discarded. In M/GI/1/$2^*$, there is a single data buffer and an arriving packet replaces the packet in the buffer (if any).
In M/M/1, there is infinite buffer space for the packets with first come first serve discipline. The first two systems with no and limited buffer have been extensively studied recently \cite{kaul2012status,costa2016age,CKam2016MILCOM, Yates2019, champati2020minimum} because it is known that limited buffering helps decrease AoI. We also study the M/M/1 system here because it was originally studied in \cite{kaul2012real} and our analysis enables a comparison of these new metrics with the AoI.
\begin{figure}[!t]
\centering{
\includegraphics[totalheight=0.05\textheight]{./figs/System.eps}}
\caption{\sl Status update packets arriving to a single server transmission queue.}
\label{fig:1}
\vspace{-0.25in}
\end{figure}
Inspired by \cite{li2020aoi} and \cite{pan2021coding}, we introduce a deadline/threshold $H$ for the age of the update packets. Unlike the hard threshold assumption that instantaneous age is not allowed to exceed the threshold in \cite{li2020aoi}, our threshold is assumed as a soft threshold as in \cite{pan2021coding}. Based on a given threshold, we introduce three metrics -- {\em overage probability}, {\em average overage}, and {\em stale update probability} -- and formally define them below.
Our metrics are closely related to the peak age. As discussed in \cite{costa2016age} for peak age, the AoI reaches a local maximum value every time before a packet is delivered. Based on the queue states and system time of the packets, peak age is different for different received packets. Therefore, if a fixed threshold for instantaneous AoI is given, some received packets' peak age will exceed the threshold while some peak ages will stay under the threshold. At the same time, the AoI is supposed to decrease at the instant when the packet is delivered to the receiver. At this instant, the instantaneous AoI is equal to the system time of this packet.
\subsection{Overage Probability}
The {\em overage probability} is the probability that the age of the packet currently held by the receiver is larger than the threshold. Let $\varepsilon_i$ denote the time period during which the instantaneous age is larger than the threshold $H$ during the interval between the delivery of packet $i-1$ and packet $i$ (also called the inter-departure period between packet $i-1$ and packet $i$). Let $Y_i$ denote the inter-departure period between packet $i-1$ and packet $i$ and $T_i$ denote the system time (i.e., waiting time plus service time) of packet $i$. Then, the peak age caused by packet $i$ is $\Delta_i^{peak}=T_{i-1}+Y_{i}$, and we have $\varepsilon_i$ as:
\begin{align}\label{epsi}
\varepsilon_i=\left\{\begin{matrix}
&0 &\Delta_i^{peak}<H \\
&\Delta_i^{peak}-H &\Delta_i^{peak}>H,\ T_{i-1}<H \\
&Y_i& T_{i-1}>H.
\end{matrix}\right.
\end{align}
Then the overage probability is:
\begin{align}\label{oageP}
P_{o}=\lim_{T \to \infty} \frac{1}{T}\sum_{i=1}^{n(T)}\varepsilon_i,
\end{align}
where $T$ is the observation period and $n(T)$ is the number of packets delivered to the receiver in observation period $T$.
\subsection{Average Overage}
The {\em average overage} is defined as the time average of the update age beyond the threshold, and is a measure of the average ``staleness'' of the update packet held by the receiver. The instantaneous AoI is the difference between the current time and the generation time of the packet held by the receiver:
\begin{align*}
\Delta(t)=t-u(t),
\end{align*}
where $u(t)$ is the generation time of the latest packet at the receiver at time $t$. The instantaneous overage is then given by:
\begin{align}\label{oage}
\Delta^o(t)=\max(\Delta(t)-H,0),
\end{align}
and the average overage is:
\begin{align}\label{aoage}
\mathbb{E}[\Delta^o]=\lim_{T \to \infty} \frac{1}{T}\int_{0}^{T} \Delta^o(t)dt.
\end{align}
\subsection{Stale Update Probability}
Finally we introduce the {\em stale update probability}. This is the probability that an update is stale, i.e., its age has exceeded the deadline, when it is delivered to the receiver. It is clear that packets with system time larger than the threshold are stale since those packets cannot make the age decrease below the threshold. Furthermore, packets that are discarded in the queuing system are also considered to be stale, they are never delivered and can be thought of as having infinite system time. We define $p_d$ as the probability that a packet is delivered to the receiver, i.e., not dropped. Then the stale update probability, $P_s$ is given by:
\begin{align}\label{Ps}
P_s=1-p_d+{\rm Pr^*}\{T_i>H\}
\end{align}
where ${\rm Pr^*}\{T_i>H\}$ is the probability that a packet is a delivered packet with system time large than threshold.
The three metrics are illustrated in Fig. \ref{fig:Overage} for an M/GI/1/1 system. At time $t_1$, packet 1 arrives to the system and starts its service. During the service time of packet 1, a packet arrives to the system and is dropped since there is no buffer to store this packet. This time is denoted by a cross in the figure. The service of packet 1 is finished at $t_1'$; packet 2 arrives to the system at $t_2$ and finishes its service at $t_2'$. We can see that packet 1 is a stale update packet whereas packet 2 is not since the system time of packet 2 is smaller than the threshold.
\begin{figure}[!t]
\centering{
\includegraphics[totalheight=0.21\textheight]{./figs/Overage.eps}}
\caption{\sl Illustration of overage probability, stale update probability, and average overage. }
\label{fig:Overage}
\end{figure}
The average overage is the time average area of the age beyond the threshold, and we use $Q_i$ to denote those areas during the inter-departure period between packet $i-1$ and packet $i$. These areas are shown shaded in Fig. \ref{fig:Overage}. From the figure, we can see that $\varepsilon_2$ and $Q_2$ are obtained using the expression for the second case (i.e., $\Delta_i^{peak}>H,\ T_{i-1}<H$) and $\varepsilon_3$ and $Q_3$ are obtained using the expression for the third case (i.e., $T_{i-1}>H$).
$Q_i$ and $\varepsilon_i$ are related as follows:
\begin{align}\label{Qi}
Q_i=\left\{\begin{matrix}
&0 &\Delta_i^{peak}<H \\
&\frac{1}{2}\varepsilon_i^2 &\Delta_i^{peak}>H,\ T_{i-1}<H \\
&\frac{1}{2}\varepsilon_i^2+\varepsilon_i(T_{i-1}-H)& T_{i-1}>H,
\end{matrix}\right.
\end{align}
and we can rewrite equation (\ref{aoage}) as:
\begin{align}\label{aoage2}
\mathbb{E}[\Delta^o]=\lim_{T \to \infty} \frac{1}{T}\sum_{i=1}^{n(T)}Q_i.
\end{align}
If we denote the number of generated packets by the source in observation period $T$ by $N(T)$, we will have $\lambda=\lim_{T \to \infty}\frac{N(T)}{T}$.
Therefore, we have:
\begin{align}\label{Po}
P_o=\lambda\lim_{T \to \infty}\frac{\sum_{i=1}^{n(T)}\varepsilon_i}{N(T)}=\lambda\mathbb{E}[\varepsilon],
\end{align}\vspace{-0.1in}
\begin{align}\label{Do}
\mathbb{E}[\Delta^o]=\lambda\lim_{T \to \infty}\frac{\sum_{i=1}^{n(T)}Q_i}{N(T)}=\lambda\mathbb{E}[Q],
\end{align}
where $\mathbb{E}[\varepsilon]$ and $\mathbb{E}[Q]$ are expectations taken over {\em all generated packets} by assuming a value of zero for $\varepsilon$ and $Q$ for dropped packets.
In the next three sections, we derive expressions for the three metrics for the three queuing systems.
\section{M/GI/1/1 Queue}\label{partA}
In the $M/GI/1/1$ queuing system, there is a single server and no buffer. Packets that arrive when the server is idle are taken to service immediately and those arriving in busy period are dropped. In view of the renewal structure, we have the following stationary probabilities for each state:
\begin{equation}\label{MG11:pb}
p_{I}=\frac{1}{\lambda T_{cycle}},\ p_{B}=\frac{\mathbb{E}[S]}{T_{cycle}},
\end{equation}
where $T_{cycle}=\frac{1}{\lambda}+\mathbb{E}[S]$ is the expected length of one renewal cycle; $I$ and $B$ indicate the idle and busy states. Clearly, we have $p_d=p_{I}$ which indicates that only the packets arrive in idle state can be delivered to the receiver. Since there is no waiting period for packets (as there is no buffer), we have $T_i=S_i$, $Y_i=X_i+S_i$ where $X_i$ denotes the idle period before packet $i$ arrives to the system and is an exponentially distributed random variable with coefficient $\lambda$. We now obtain expressions for the metrics. Closed-form expressions may be obtained for specific service time distributions. We show closed-form expressions for the M/M/1/1 system in Appendix.\ref{Amm11}.
\subsection{Overage probability}
First, we evaluate $\mathbb{E}[\varepsilon]$ which is equal to $\mathbb{E}[\varepsilon|Id]p_{I}$, since the packets arriving in busy state will be dropped in this system. For $\mathbb{E}[\varepsilon|Id]$, based on equation (\ref{epsi}), we have:
\begin{align}\label{EeMG11}
\nonumber \mathbb{E}[\varepsilon|Id]=\int_{0}^{H}\int_{0}^{\infty}\int_{H-s_i-s_{i-1}}^{\infty}(x+s_{i}+s_{i-1}-H)\\
\nonumber f_{X}(x)f_{S}(s_i)f_{S}(s_{i-1})dxds_ids_{i-1}\\
+\int_{H}^{\infty}\int_{0}^{\infty}\int_{0}^{\infty}(x+s_{i})
f_{X}(x)f_{S}(s_i)f_{S}(s_{i-1})dxds_ids_{i-1}.
\end{align}
Finally, we have overage probability as $P_o=\lambda\mathbb{E}[\varepsilon|Id]p_{I}$ from equation (\ref{Po}). The more general case of a GI/GI/1/1 system was considered in \cite{Champati2019}.
\subsection{Average Overage}
Similarly, we evaluate $\mathbb{E}[Q|Id]$ based on (\ref{Qi}):
\begin{align}\label{EQMG11}
\nonumber \mathbb{E}[Q|Id]=\int_{0}^{H}\int_{0}^{\infty}\int_{H-s_i-s_{i-1}}^{\infty}\frac{1}{2}(x+s_{i}+s_{i-1}-H)^2\\
\nonumber f_{X}(x)f_{S}(s_i)f_{S}(s_{i-1})dxds_ids_{i-1}\\
\nonumber +\int_{H}^{\infty}\int_{0}^{\infty}\int_{0}^{\infty}[\frac{1}{2}(x+s_{i})^2+s_{i-1}(x+s_{i})]\\
f_{X}(x)f_{S}(s_i)f_{S}(s_{i-1})dxds_ids_{i-1}.
\end{align}
Then we have $\mathbb{E}[\Delta^o]=\lambda\mathbb{E}[Q|Id]p_{I}$ from equation (\ref{Do}).
\subsection{Stale Update Probability}
Since $p_d=p_{I}$, based on equation (\ref{Ps}), we evaluate ${\rm Pr^*}\{T_i>H\}$ as:
\begin{align}
\nonumber {\rm Pr^*}\{T_i>H\}&=p_{I}{\rm Pr^*}\{T_i>H|Id\}\\&=p_{I}\int_{H}^{\infty}f_{S}(s_i)ds_i=p_{I}(1-F_S(H)),
\end{align}
where $F_S(s)$ is the cumulative distribution function (cdf) of random variable $S_i$.
Therefore, we have $P_s=1-p_{I}F_S(H)$ from equation (\ref{Ps}).
\section{M/GI/1/$2^*$ Queue}
In this system, there is a single packet buffer. The server is in either idle or busy state. Packets that arrive in the idle period are served immediately and those that arrive in the busy period are stored in the buffer and take the place of the old packet in the buffer. In view of the renewal structure, we have the following stationary probabilities for each state of the server:
\begin{align}\label{prob:MG12}
p_{I}=\frac{1}{\lambda T_{cycle}},\ p_{B}=\frac{\mathbb{E}[S]}{T_{cycle}MGF_{S}(\lambda)},
\end{align}
where we use $MGF_{S}(\lambda)$ to denote the moment generating function of the service distribution evaluated at $-\lambda$:
\begin{align}
MGF_{S}(\lambda)= \mathbb{E}[e^{-\lambda S}],
\end{align}
where $T_{cycle}=\frac{1}{\lambda}+\frac{\mathbb{E}[S]}{MGF_{S}(\lambda)}$ is the expected length of one renewal cycle.
Since only one packet that arrives during the busy period is served and the others are discarded in the buffer, we define the states $B_{1}$ and $B_2$ as the busy states of the server with zero and one packet waiting in the queue, respectively. The renewal cycle is as follows: after idle period, an arrival happens, and the system enters $B_{1}$ state. Now a time duration of service $S$ starts and if during the service period another arrival occurs, the system turns to $B_{2}$ state. This back-and-forth between $B_1$ and $B_2$ states continues until no packet arrives in one service time. Based on the analysis in our previous work \cite{zou2020age}, we have:
\begin{align*}
p_{B_{2}}=p_{B}\left(1+\frac{MGF_{S}(\lambda)-1}{\lambda \mathbb{E}[S]}\right),
\end{align*}
and the probability of $B_{1}$ state is $p_{B_{1}}=p_{B}-p_{B_{2}}$.
Clearly we have $p_d=p_{I}+p_{B_{1}}$ for M/GI/1/$2^*$ system. We are now ready to derive the three metrics for this queuing system. Closed-form expressions the M/M/1/$2^*$ system are shown in Appendix.\ref{Amm12}.
\subsection{Overage probability}
Note that a packet can be delivered to the receiver even if it arrives to the system during busy period. Therefore, we evaluate $\mathbb{E}[\varepsilon|Id]$ and $\mathbb{E}[\varepsilon|B]$ separately to get the expression for $\mathbb{E}[\varepsilon]$.
\subsubsection{$\mathbb{E}[\varepsilon|Id]$}
First, we assume that packet $i-1$ finds the system in idle state and starts service. Therefore, we have $T_{i-1}=S_{i-1}$ and we assume the inter-arrival time between $i-1$ and next packet is $X$ (the next packet may be dropped and may not be indexed). If $X<S_{i-1}$, we have $Y_i=S_i$. Otherwise if $X>S_{i-1}$, we have $Y_i=X-S_{i-1}+S_{i}$. Now, we have:
\begin{align*}\vspace{-0.1in}
\nonumber \mathbb{E}[\varepsilon|Id]=\int_{0}^{H}\int_{s_{i-1}}^{\infty}\int_{H-x}^{\infty}&(x+s_{i}-H)\\
\nonumber &f_{S}(s_i)f_{X}(x)f_{S}(s_{i-1})ds_idxds_{i-1}
\end{align*}\vspace{-0.2in}
\begin{align*}
\nonumber +\int_{H}^{\infty}\int_{s_{i-1}}^{\infty}\int_{0}^{\infty}&(x+s_{i}-s_{i-1})\\
\nonumber &f_{S}(s_i)f_{X}(x)f_{S}(s_{i-1})ds_idxds_{i-1}
\end{align*}\vspace{-0.2in}
\begin{align*}
\nonumber+\int_{0}^{H}\int_{0}^{s_{i-1}}\int_{H-s_{i-1}}^{\infty}&(s_{i-1}+s_{i}-H)\\
\nonumber &f_{S}(s_i)f_{X}(x)f_{S}(s_{i-1})ds_idxds_{i-1}
\end{align*}\vspace{-0.2in}
\begin{align}\label{EeMG12}
+\int_{H}^{\infty}\int_{0}^{s_{i-1}}\int_{0}^{\infty}s_{i}
&f_{S}(s_i)f_{X}(x)f_{S}(s_{i-1})ds_idxds_{i-1}.
\end{align}\vspace{-0.1in}
\subsubsection{$\mathbb{E}[\varepsilon|B]$}
Next, we assume that packet $i-1$ finds the system in busy state. We use $W$ to denote the waiting time for this packet in the buffer and we have the pdf for $W$ as: $f_W(w)=\frac{{\rm Pr}\{S>w\}}{\mathbb{E}[S]}$. Note that only when $X>W$, this packet will be indexed as the $i-1$th packet and $T_{i-1}=W+S_{i-1}$. If $W<X<W+S_{i-1}$, we have $Y_i=S_i$. Otherwise if $X>W+S_{i-1}$, we have $Y_i=X-W-S_{i-1}+S_{i}$. Now, we have:
\begin{align*}
\nonumber \mathbb{E}[\varepsilon|B]=&\\
\nonumber \int_{0}^{H}&\int_{0}^{H-s_{i-1}}\int_{H-s_{i-1}-w}^{\infty}\int_{w}^{w+s_{i-1}}(w+s_{i-1}+s_{i}-H)\\
\nonumber &\ \ \ f_{X}(x)f_{S}(s_i)f_{W}(w)f_{S}(s_{i-1})dxds_idwds_{i-1}
\end{align*}\vspace{-0.2in}
\begin{align*}
\nonumber +\int_{0}^{\infty}&\int_{H-s_{i-1}}^{\infty}\int_{0}^{\infty}\int_{w}^{w+s_{i-1}}s_{i}\\
\nonumber &\ \ \ f_{X}(x)f_{S}(s_i)f_{W}(w)f_{S}(s_{i-1})dxds_idwds_{i-1}
\end{align*}\vspace{-0.2in}
\begin{align*}
\nonumber +\int_{0}^{H}&\int_{0}^{H-s_{i-1}}\int_{w+s_{i-1}}^{\infty}\int_{H-x}^{\infty}(x+s_{i}-H)\\
\nonumber &\ \ \ f_{S}(s_i)f_{X}(x)f_{W}(w)f_{S}(s_{i-1})ds_idxdwds_{i-1}
\end{align*}\vspace{-0.2in}
\begin{align}\label{EeMG12B}
\nonumber +\int_{0}^{\infty}&\int_{H-s_{i-1}}^{\infty}\int_{w+s_{i-1}}^{\infty}\int_{0}^{\infty}(x-w-s_{i-1}+s_{i})\\
&\ \ \ f_{S}(s_i)f_{X}(x)f_{W}(w)f_{S}(s_{i-1})ds_idxdwds_{i-1}.
\end{align}
Finally, we have overage probability as $P_o=\lambda(\mathbb{E}[\varepsilon|Id]p_{I}+\mathbb{E}[\varepsilon|B]p_{B})$ from equation (\ref{Po}).
\subsection{Average Overage}
Similarly, we evaluate $\mathbb{E}[Q|Id]$ and $\mathbb{E}[Q|B]$ based on (\ref{Qi}):
\begin{align*}
\nonumber \mathbb{E}[Q|Id]=\int_{0}^{H}\int_{s_{i-1}}^{\infty}\int_{H-x}^{\infty}&\frac{1}{2}(x+s_{i}-H)^2\\
\nonumber &f_{S}(s_i)f_{X}(x)f_{S}(s_{i-1})ds_idxds_{i-1}
\end{align*}\vspace{-0.2in}
\begin{align*}
\nonumber +\int_{H}^{\infty}\int_{s_{i-1}}^{\infty}\int_{0}^{\infty}\biggl(\frac{1}{2}(x+&s_{i}-s_{i-1})^2\\
\nonumber+(x+s_{i}-s_{i-1})(s_{i}-H)\biggr) &f_{S}(s_i)f_{X}(x)f_{S}(s_{i-1})ds_idxds_{i-1}
\end{align*}\vspace{-0.2in}
\begin{align*}
\nonumber+\int_{0}^{H}\int_{0}^{s_{i-1}}\int_{H-s_{i-1}}^{\infty}&\frac{1}{2}(s_{i-1}+s_{i}-H)^2\\
\nonumber &f_{S}(s_i)f_{X}(x)f_{S}(s_{i-1})ds_idxds_{i-1}
\end{align*}\vspace{-0.2in}
\begin{align}\label{EQMG12}
\nonumber +\int_{H}^{\infty}\int_{0}^{s_{i-1}}\int_{0}^{\infty}&\biggl(\frac{1}{2}s_{i}^2+s_{i}(s_{i-1}-H)\biggr)\\
&f_{S}(s_i)f_{X}(x)f_{S}(s_{i-1})ds_idxds_{i-1}.
\end{align}\vspace{-0.2in}
\begin{align*}
\nonumber \mathbb{E}[Q|B]=&\\
\nonumber \int_{0}^{H}\int_{0}^{H-s_{i-1}}&\int_{H-s_{i-1}-w}^{\infty}\int_{w}^{w+s_{i-1}}\frac{1}{2}(w+s_{i-1}+s_{i}-H)^2\\
\nonumber &\ \ \ f_{X}(x)f_{S}(s_i)f_{W}(w)f_{S}(s_{i-1})dxds_idwds_{i-1}
\end{align*}\vspace{-0.2in}
\begin{align*}
\nonumber +\int_{0}^{\infty}\int_{H-s_{i-1}}^{\infty}&\int_{0}^{\infty}\int_{w}^{w+s_{i-1}}\biggl(\frac{1}{2}s_{i}^2+s_{i}(w+s_{i-1}-H)\biggr)\\
\nonumber &\ \ \ f_{X}(x)f_{S}(s_i)f_{W}(w)f_{S}(s_{i-1})dxds_idwds_{i-1}
\end{align*}\vspace{-0.2in}
\begin{align*}
\nonumber +\int_{0}^{H}\int_{0}^{H-s_{i-1}}&\int_{w+s_{i-1}}^{\infty}\int_{H-x}^{\infty}\frac{1}{2}(x+s_{i}-H)^2\\
\nonumber &\ \ \ f_{S}(s_i)f_{X}(x)f_{W}(w)f_{S}(s_{i-1})ds_idxdwds_{i-1}
\end{align*}\vspace{-0.2in}
\begin{align}\label{EQMG12B}
\nonumber +\int_{0}^{\infty}\int_{H-s_{i-1}}^{\infty}&\int_{w+s_{i-1}}^{\infty}\int_{0}^{\infty}\biggl(\frac{1}{2}(x-w-s_{i-1}+s_{i})^2\\
\nonumber &+(x-w-s_{i-1}+s_{i})(w+s_{i-1}-H)\biggr)\\ &f_{S}(s_i)f_{X}(x)f_{W}(w)f_{S}(s_{i-1})ds_idxdwds_{i-1}.
\end{align}
Then we have $\mathbb{E}[\Delta^o]=\lambda(\mathbb{E}[Q|Id]p_{I}+\mathbb{E}[Q|B]p_{B})$ from equation (\ref{Do}).
\subsection{Stale Update Probability}
Now, we evaluate ${\rm Pr^*}\{T_i>H\}$ as ${\rm Pr^*}\{T_i>H|Id\}p_{I}+{\rm Pr^*}\{T_i>H|B\}p_{B}$. Since the packet that arrives in idle state will be a delivered packet and its system time $T_{i}=S_{i}$, we have:
\begin{align}\label{EPMG12}
{\rm Pr^*}\{T_i>H|Id\}=\int_{H}^{\infty}f_{S}(s_i)ds_i=1-F_S(H).
\end{align}
For the packets that arrive in busy state, only the packets with $X>W$ can be delivered packets, and their corresponding system time is $T_{i}=W+S_{i}$. So we have:
\begin{align}\label{EPMG12B}
\nonumber {\rm Pr^*}\{T_i>H|B\}=\int_{0}^{\infty}\int_{H-s_{i}}^{\infty}\int_{w}^{\infty}f_{X}(x)f_{W}(w)&f_{S}(s_i)\\&dxdwds_i.
\end{align}
Therefore, we have $P_s=1-(p_{I}+p_{B_{1}})+{\rm Pr^*}\{T_i>H\}$ from equation (\ref{Ps}).
\section{M/M/1 Queue}
The analysis for overage probability has been done in \cite{hu2021status} as violation probability. Therefore, we present only the analysis for average overage and stale update probability in the M/M/1 system. Note that in this system, all the packets are delivered packets since the buffer size is unlimited. The packets are served following the first come first serve discipline. The service time for each packet follows an iid exponential distribution with mean $\frac{1}{\mu}$.
\subsection{Average Overage}
Based on equation (6) in \cite{hu2021status}, we have the joint probability density function of $T_{i-1}$ and $Y_{i}$ as:
\begin{align*}
\nonumber f_{T,Y}(y,t)&=\\&(\mu^2-\lambda\mu)e^{\lambda t-\mu y-\mu t}-\mu^2e^{-\mu y-\mu t}+\lambda \mu e^{-\lambda y-\mu t}.
\end{align*}
Therefore we have:
\begin{align}\label{EQMM1}
\nonumber \mathbb{E}[Q]=\int_{0}^{H}\int_{H-t}^{\infty}\frac{1}{2}(y+t-H)^2f_{T,Y}(y,t)dydt\\+\int_{H}^{\infty}\int_{0}^{\infty}\biggl(\frac{1}{2}y^2+y(t-H)\biggr)f_{T,Y}(y,t)dydt,
\end{align}
and we have $\mathbb{E}[\Delta^o]=\lambda\mathbb{E}[Q]$ from equation (\ref{Do}). Closed-form expressions are shown in Appendix.\ref{Amm1}.
\subsection{Stale Update Probability}
Based on equation (4) in \cite{hu2021status}, we have the pdf of system time $T_{i}$ as:
\begin{align*}
f_{T}(t)=\mu(1-\frac{\lambda}{\mu})e^{-\mu(1-\frac{\lambda}{\mu})t}.
\end{align*}
Since all the packets are delivered packets in this system, we have $p_d=1$ and ${\rm Pr^*}\{T_i>H\}$ as:
\begin{align}\label{PsMM1}
{\rm Pr^*}\{T_i>H\}=\int_{H}^{\infty}f_{T}(t)dt=e^{-\mu(1-\frac{\lambda}{\mu})H}.
\end{align}
Therefore, we have $P_s={\rm Pr^*}\{T_i>H\}=e^{-\mu(1-\frac{\lambda}{\mu})H}$ from equation (\ref{Ps}).
\section{Numerical Results}
In this section, we present numerical results for the three metrics. We also performed packet-based queue simulations offline for $10^6$ packets in order to verify the analytical results. We consider exponential distributed service times with $f_{S}(s)=\mu e^{-\mu s}$ and gamma distributed service times with $f_{S}(s)=\frac{\beta^\alpha}{\Gamma(\alpha)} \beta^{\alpha-1} e^{-\beta s}$ where $\alpha$ is the shape parameter.
\begin{figure}[!t]
\centering{
\includegraphics[totalheight=0.265\textheight]{./figs/Num1_sim.eps}}\vspace{-0.15in}
\caption{\sl Overage Probability and Stale Update Probability with respect to $H$ for $\lambda=1$ and exponential service time with $\mu=2$. Circles are simulation results.}
\label{Num_1}
\end{figure}
Fig. \ref{Num_1} shows the overage probability and stale update probability for the three queuing systems with respect to $H$ for $\lambda=1$ and exponential service time with $\mu=2$. We observe that when the threshold $H$ is small, $P_s$ is smaller than $P_o$ for all the three systems, but as $H$ increases, $P_s$ for M/M/1/1 and M/M/1/$2^*$ plateau, while the other probabilities keep decreasing. This is because the threshold has no effect on the queue behaviour, and with a large enough threshold, stale updates are mainly due to dropped packets in the system. Note that since the M/M/1 system serves all the packets, $P_s $ for M/M/1 keeps decreasing as the threshold increases.
\begin{figure}[!t]
\centering
\subfigure[]{
\includegraphics[totalheight=0.265\textheight]{./figs/Num2_sim.eps}}
\subfigure[]{
\includegraphics[totalheight=0.265\textheight]{./figs/Num2_sim_2.eps}}\vspace{-0.15in}
\caption{\sl Average AoI and Average Overage with respect to $\rho$ for exponential service time with fixed $\mu=2$. (a) All three queuing systems with $H=1$ (b) M/M/1 queue with respect to $\rho$ and $H$. Circles are simulation results.}
\label{Num_2}
\end{figure}
In Fig. \ref{Num_2}, we show average AoI and average overage with respect to $\rho$ where $\rho=\frac{\lambda}{\mu}$ for exponential service time with fixed $\mu=2$. In Fig. \ref{Num_2}(a), we compare the three queuing schemes with fixed $H=1$. It can be seen that the average overage has a similar behaviour to the average AoI with respect to $\rho$. For the M/M/1 queue, we can find optimum points for both average AoI and average overage while the average AoI and average overage keep decreasing as $\rho$ increases for M/M/1/1 and M/M/1/$2^*$. In Fig. \ref{Num_2}(b), we show average AoI and average overage for the M/M/1 queue with respect to $\rho$ and $H$. Note that the average overage is not actually equal to the average AoI minus $H$, because update packets with system times smaller than the threshold have zero overage. As $H$ becomes large, the average overage tends to zero, while the average AoI is unaffected by $H$, as can be seen in the figure. Further, the values of $\rho$ that minimize average AoI and average overage are slightly different.
\begin{figure}[!t]
\centering{
\includegraphics[totalheight=0.265\textheight]{./figs/Num3_1.eps}}\vspace{-0.15in}
\caption{\sl Overage Probability and Stale Update Probability for M/GI/1/1 with respect to $\lambda$ for $H=1$ and gamma distribution service time with $\mathbb{E}[S]=\frac{1}{2}$.}
\label{Num_3}
\end{figure}
In Fig. \ref{Num_3}, we show Overage Probability and Stale Update Probability for M/GI/1/1 queue with respect to $\lambda$ for $H=1$ and gamma distributed service time with $\mathbb{E}[S]=\frac{1}{2}$. We can observe that with increasing $\lambda$, $P_o$ decreases while $P_s$ increases. This indicates that a trade-off exists for this system between overage probability and stale update probability. This is owing to the fact that as $\lambda$ increases, more packets are dropped in the queue, while the peak age decreases since the idle period decreases. We can also observe that as $\alpha$ increases, $P_o$ increases while $P_s$ decreases.
\begin{figure}[!t]
\centering{
\includegraphics[totalheight=0.265\textheight]{./figs/Num4_1.eps}}\vspace{-0.15in}
\caption{\sl Overage Probability and Stale Update Probability for M/GI/1/$2^*$ with respect to $\mathbb{E}[S]$ for gamma distribution service time and $H=1$, $\lambda=1$.}
\label{Num_4}
\end{figure}
Finally, in Fig. \ref{Num_4}, we show Overage Probability and Stale Update Probability for M/GI/1/$2^*$ queue with respect to $\mathbb{E}[S]$ for gamma distribution service time and $H=1$, $\lambda=1$. We can observe that as $\mathbb{E}[S]$ increases, both $P_o$ and $P_s$ increase. With smaller $\mathbb{E}[S]$, both $P_{o}$ and $P_{s}$ have good performance with larger $\alpha$ (i.e., smaller variance for service time); however, as $\mathbb{E}[S]$ increases, $P_{o}$ and $P_{s}$ are lower with smaller $\alpha$ (i.e., larger variance for service time).
This is because when the expected value of service time is small, a large variance implies that a large service time has a reasonably high probability of occurring, whereas when the expected value of service time is large, large variance means that small service times have a reasonably high probability of occurring.
\section{Conclusion}
For a status update system with a soft deadline, this paper introduced three threshold-based age metrics -- {\em overage probability}, {\em average overage}, and {\em stale update probability} -- and evaluated them for three typical queuing systems. Expressions for the metrics are derived for three queueing systems, and several numerical results are obtained
Our numerical results show the behavior of these metrics under different parameter settings and different service time distributions. We compare the average overage with general average AoI to show their differences under different values of $\rho$. The results indicate a trade-off between these two for an M/M/1 queue. In the future, we intend to investigate these metrics for other queuing systems, and consider scheduling disciplines that minimize them.
\subsection{Closed form expressions for M/M/1/1}\label{Amm11}
In M/M/1/1 system, we have $f_S(s)=\mu e^{-\mu s}$ and $p_{I}=\frac{\mu}{\lambda+\mu}$.
\subsubsection{Overage Probability} \label{mm11op}
Based on (\ref{EeMG11}), we have:
\begin{align*}
\nonumber \mathbb{E}[\varepsilon|Id]=(\frac{1}{\lambda} + \frac{1}{\mu})e^{-\mu H} + \frac{He^{-\mu H}}{\lambda - \mu} + \frac{\mu^2(e^{-\lambda H} - e^{-\mu H})}{\lambda(\lambda - \mu)^2}
\end{align*}
and finally we have:
$P_o=\frac{\lambda\mu}{\lambda+\mu}\mathbb{E}[\varepsilon|Id]$.
\subsubsection{Average Overage}
Based on (\ref{EQMG11}), we have:
\begin{align*}
\nonumber &\mathbb{E}[Q|Id]=e^{-\mu H}(\frac{1}{\lambda^2}- H(\frac{1}{\lambda} + \frac{1}{\mu}) + \frac{1}{\mu^2} + \frac{1}{\lambda\mu}) \\
&+ (He^{-\mu H} +\frac{e^{-\mu H}}{\mu})(\frac{1}{\lambda} + \frac{1}{\mu})
\\&+ \frac{\mu^2}{(\lambda - \mu)(\lambda - \mu)^3}\biggl((e^{-\lambda H}+ e^{-\mu H}) (\lambda^2 - 2\lambda\mu+ \mu^2 )H^2\end{align*}
\begin{align*}
&+(e^{-\lambda H}- e^{-\mu H})( 2\lambda - 2\mu)H\biggr)\\& -\frac{\mu^2H^2(e^{-\lambda H} - e^{-\mu H})}{(\lambda - \mu)^2} + \frac{\mu^2(e^{-\lambda H} - e^{-\mu H})}{\lambda^2(\lambda - \mu)^2} \\&- 2\mu^2 H\frac{e^{-\lambda H}-e^{-\mu H}(\mu H - \lambda H + 1)}{(\lambda - \mu)^3} +\frac{\lambda H e^{-\mu H}}{\mu(\lambda - \mu)}
\end{align*}
and finally we have:
$\mathbb{E}[\Delta^o]=\frac{\lambda\mu}{\lambda+\mu}\mathbb{E}[Q|Id]$
\subsubsection{Stale Update Probability}
Since we have:
\begin{align*}
\nonumber F_S(H)=1-e^{-\mu H}
\end{align*}
we can get $P_s=1-p_{I}F_S(H)=1-\frac{\mu}{\lambda+\mu}+\frac{\mu}{\lambda+\mu}e^{-\mu H}$.
\subsection{Closed form expressions for M/M/1/2$^{*}$}\label{Amm12}
In M/M/1/2$^{*}$ system, we have $p_{I}=\frac{\mu^2}{\lambda^2+\lambda\mu+\mu^2}, p_{B}=\frac{\lambda^2+\lambda\mu}{\lambda^2+\lambda\mu+\mu^2}$ and $p_{B_1}=\frac{\lambda\mu}{\lambda^2+\lambda\mu+\mu^2}$
\subsubsection{Overage Probability}
Based on (\ref{EeMG12}) and (\ref{EeMG12B}), we have:
\begin{align*}
\mathbb{E}[\varepsilon|Id]&=He^{-H\mu }+\frac{e^{-H\mu }}{\mu }+\frac{e^{-H\lambda }-e^{-H\mu }}{\lambda }\\
&-\frac{e^{-H\mu }\left(e^{-H\lambda }-1\right)}{\lambda -\mu }+\frac{\mu e^{-H(\lambda+\mu)}}{\lambda ^2+\mu \lambda }\\
&+\frac{e^{-H\lambda }\left(1-e^{-H\mu }\right)}{\mu }+\frac{\lambda e^{-H\lambda }\left(e^{-H\mu }-1\right)}{\mu \left(\lambda -\mu \right)}
\end{align*}
\begin{align*}
\mathbb{E}[\varepsilon|B]&=\frac{e^{-H\mu} (1-e^{-H\lambda })}{\lambda }+\frac{ e^{-H(\lambda+\mu)}}{\lambda +\mu}\\
&-\frac{e^{-H\lambda }\left(e^{-H\mu }-1\right)}{\lambda }-\frac{e^{-H\lambda }\left(e^{-H\mu }-1\right)}{\mu }-He^{-H(\lambda+\mu)}\\
&+\frac{H\mu e^{-H\mu }}{\lambda }+\frac{\mu ^2e^{-H(\lambda+\mu)}}{\lambda {\left(\lambda +\mu \right)}^2}+\frac{2\mu e^{-H\mu }\left(e^{-H\lambda }-1\right)}{\lambda ^2}\\
&+\frac{H e^{-H(\lambda+\mu)}(\lambda-\mu)}{\lambda -\mu }+\frac{H\mu ^2e^{-H(\lambda+\mu)}}{\lambda ^2+\mu \lambda }\\
&+\frac{\lambda e^{-H\lambda }\left(e^{-H\mu }-1\right)}{\mu \left(\lambda -\mu \right)}-\frac{\mu e^{-H\mu }\left(e^{-H\lambda }-1\right)}{\lambda \left(\lambda -\mu \right)}
\end{align*}
Finally, we have overage probability as $P_o=\lambda(\mathbb{E}[\varepsilon|Id]p_{I}+\mathbb{E}[\varepsilon|B]p_{B})$.
\subsubsection{Average Overage}
Based on (\ref{EQMG12}) and (\ref{EQMG12B}), we have:
\begin{align*}
\mathbb{E}[Q|Id]&=\frac{e^{-H\mu }}{\mu ^2}+H\frac{e^{-H\lambda }-e^{-H(\lambda+\mu)}}{\mu }+\frac{e^{-H\mu }\left(H\mu +1\right)}{\mu ^2}\\
&-\frac{e^{-H\lambda }\left(e^{-H\mu }-1\right)}{\mu ^2}+\frac{e^{-H(\lambda+\mu)}}{\lambda \left(\lambda +\mu \right)}+\frac{ e^{-H\lambda }-e^{-H(\lambda+\mu)}}{\lambda ^2}\\
&+\frac{\mu e^{-H(\lambda+\mu)}}{\lambda ^2\left(\lambda +\mu \right)}+\frac{ e^{-H\mu }-e^{-H(\lambda+\mu)}}{\mu \left(\lambda -\mu \right)}\\
&-\frac{\lambda \left(e^{-H\lambda }-e^{-H(\lambda+\mu)}\right)}{\mu^2 \left(\lambda -\mu \right)}+\frac{He^{-H\lambda }\left(e^{-H\mu }-1\right)}{\mu }\\
&+\frac{\mu e^{-H(\lambda+\mu)}\left(H\left(\lambda +\mu \right)+1\right)}{\lambda {\left(\lambda +\mu \right)}^2}-\frac{H\mu e^{-H(\lambda+\mu)}}{\lambda \left(\lambda +\mu \right)}
\end{align*}
\begin{align*}
&\mathbb{E}[Q|B]=\frac{2e^{-H\left(\lambda +\mu \right)}}{\mu \left(\lambda +\mu \right)}-\frac{2e^{-H\left(\lambda +\mu \right)}-2e^{-H\mu }}{\lambda\mu }-\frac{e^{-H\left(\lambda +\mu \right)}}{{\left(\lambda +\mu \right)}^2}\\
&+\frac{e^{-H\left(\lambda +\mu \right)}\left(H\left(\lambda +\mu \right)+1\right)}{{\left(\lambda +\mu \right)}^2}-\frac{2He^{-H\left(\lambda +\mu \right)}}{\lambda +\mu }-\frac{\mu e^{-H\left(\lambda +\mu \right)}}{{\left(\lambda +\mu \right)}^3}\\
&-\frac{\mu e^{-H\left(\lambda +\mu \right)}\left(H\left(\lambda +\mu \right)+1\right)}{{\left(\lambda +\mu \right)}^3}\\
&+\frac{e^{-H\left(\lambda +\mu \right)}\left(H\lambda ^3+3H\lambda ^2\mu +\lambda ^2+3H\lambda \mu ^2+3\lambda \mu +H\mu ^3+\mu ^2\right)}{\lambda ^2{\left(\lambda +\mu \right)}^2}\\
&+\frac{2e^{-H\mu }\left(e^{-H\lambda }-1\right)}{\lambda ^2}+\frac{He^{-H\mu }}{\lambda }+\frac{He^{-H(\lambda+\mu)}}{\lambda }\\
&-\frac{e^{-H\left(\lambda +\mu \right)}\left(\lambda -\mu -\lambda e^{H\lambda }+\mu e^{H\mu }+H\lambda ^2-H\mu ^2\right)}{\lambda ^2\left(\lambda -\mu \right)}
\end{align*}
and finally we have:
$\mathbb{E}[\Delta^o]=\lambda(\mathbb{E}[Q|Id]p_{I}+\mathbb{E}[Q|B]p_{B})$.
\subsubsection{Stale Update Probability}
Based on (\ref{EPMG12B}), we have:
\begin{align*}
{\rm Pr^*}\{T_i>H|B\}=\frac{\mu e^{-H(\lambda+\mu)}}{\lambda +\mu }-\frac{\mu e^{-H\mu }\left(e^{-H\lambda }-1\right)}{\lambda }
\end{align*}
we can get $P_s=1-(p_{I}+p_{B_{1}})+{\rm Pr^*}\{T_i>H\}$ where ${\rm Pr^*}\{T_i>H\}={\rm Pr^*}\{T_i>H|Id\}p_{I}+{\rm Pr^*}\{T_i>H|B\}p_{B}$.
\subsection{Closed form expressions for M/M/1}\label{Amm1}
In M/M/1 system, the $\mathbb{E}[\varepsilon]$ is derived in \cite{hu2021status} as equation (29) and we have $P_o=\lambda\mathbb{E}[\varepsilon]$. Next we show the closed-form expression for average overage in M/M/1.
\subsubsection{Average Overage}
Based on (\ref{EQMM1}), we have:
\begin{align*}
\nonumber \mathbb{E}[Q]&= \frac{{e}^{-\mu H}}{\lambda ^2}-\frac{{e}^{-\mu H}}{\mu ^2}-\frac{{e}^{\lambda H-\mu H}}{\mu (\lambda -\mu)}\\&-\frac{{e}^{-\mu H }(\mu H +1)}{\mu ^2}-\frac{{e}^{-\mu H}({e}^{\lambda H}-1)}{\mu ^2}+\frac{H{e}^{\lambda H -\mu H }}{\lambda -\mu }\\&-\frac{{e}^{\lambda H -\mu H}(H\left(\lambda -\mu \right)-1)}{{(\lambda -\mu )}^2}-\frac{H{e}^{-\mu H }}{\lambda }\\&-\frac{\mu ({e}^{-\lambda H}-{e}^{-\mu H})}{\lambda ^2(\lambda -\mu)}+\frac{{e}^{-\mu H}({e}^{H\lambda }-1)}{\lambda \mu }+\frac{\lambda {e}^{\lambda H-\mu H}}{\mu ^2(\lambda -\mu)}\\&+\frac{{e}^{-\mu H }(\mu H+1)}{\lambda \mu }+\frac{\lambda {e}^{\lambda H -\mu H}(H(\lambda -\mu)-1)}{\mu {(\lambda -\mu)}^2}\\&-\frac{\lambda H {e}^{\lambda H -\mu H }}{\mu (\lambda -\mu)}
\end{align*}
and finally we have:
$\mathbb{E}[\Delta^o]=\lambda\mathbb{E}[Q]$
|
\section{Introduction}
Oncolytic viruses constitute a class of targeted anticancer agents that have unique mechanisms of action compared with other therapies. The premise is to genetically engineer viral particles to selectively replicate in and lyse tumour cells. Over the past decade, hundreds of patients with cancer have been treated in clinical trials with oncolytic viruses \cite{liu2007clinical}. Unfortunately, due to the heterogeneous nature of cancer, success is elusive, and there is a growing need to quantify the dependency of treatment outcome on cancer characteristics.
A number of mathematical models have been constructed to understand the dynamics of proliferation and diffusion of such viruses in cancerous and healthy tissues. Zurakowski and Wodarz \cite{zurakowski2007model} developed a mathematical model for the \textit{in vitro} behaviour of the oncolytic virus ONYX-015 infecting cancer cells. The virus, administered in conjunction with a drug that up-regulates a tumour cell's ability to intake viral particles, shows two distinct behaviours: a locally stable steady state and a nonlinear periodic cycle between viral particles and tumour cells, in agreement with dynamics that have been experimentally observed. Their model has also allowed them to suggest strategies that alter the amplitude of oscillations and drive tumour size to low levels. Bajzer \textit{et al.} \cite{bajzer2008modeling} introduced a mathematical model that also exhibits stable oscillations between virus and tumour. Crivelli \textit{et al.} \cite{crivelli2012mathematical} instead derived a cycle-specific age-structured model for virotherapy where cell-cycle-specific activity of viruses has been investigated. Through analysis and simulation of the model, the authors have described how varying minimum cycling time and aspects of viral dynamics may lead to complete eradication of the tumour. Other authors have also focused their modelling on the delay occurring between the initial virus infection of tumour cells and the second wave of infections when viruses burst \cite{cassidy2018mathematical,jenner2018heter}. This phenomenon has also been accounted for by a delay differential equation model for the lapse in the second generation viral infection \cite{kim2018hopf}.
From the experimental point view, particularly relevant to the present work are the findings by Kim \textit{et al.} \cite{KimPH2011}. These authors have developed an oncolytic virus, modified with immunogenic polymer polyethylene-glycol (PEG) and monoclonal antibody Herceptin, which has exhibited potent anti-tumour behaviour in murine models. While the treatment appears unable to eliminate the tumour completely, it significantly reduces the growth rate of cancer cells with respect to the case of an untreated tumour. Aiming to provide an explanation as to why this treatment seems to fully eradicate a tumour, we have recently proposed a mathematical model calibrated to these experimental results \cite{jenner2018mathematical}. Analysis of common treatment protocols has unveiled some drawbacks of existing strategies and suggested optimal scheduling to maximise therapeutic benefits.
In this work, we conduct a further investigation of this mathematical model with a particular emphasis on the biologically relevant parameters that control virus-tumour dynamics. After a discussion of the basis for our approach, we introduce a non-dimensional version of the system, which allows us to conduct local stability analysis and analytically determine which parameters can lead to incomplete tumour eradication, as often observed in experimental settings. Then, we present a bifurcation analysis of the model, leading to some nontrivial and, in some case, counterintuitive findings about the viral characteristics that drive a complete tumour eradication. By examining perturbations of the viral dosage and their effect on different dynamical regions, we show how we can achieve a better application of treatment protocols. In particular, the dynamical states displayed by the model when therapies are administered strongly affect the final outcomes. Finally, a discussion of the advantages and limitations of our approach concludes the paper.
\section{Model development}
The dynamics of a tumour treatment administered via a PEG-modified adenovirus conjugated with Herceptin can be captured using a minimal mathematical frameworks as explained in Ref.~\cite{jenner2018mathematical}. Assuming that the immune response is negligible and does not need to be incorporated in the equations, the following model can be proposed:
\begin{align}
\frac{dU}{d\tau} & = r\ln \left(\frac{K}{U}\right)U -\frac{\beta U\hat{V}}{U+I},\label{AEqs1}\\
\frac{dI}{d\tau} & = \frac{\beta U\hat{V}}{U+I}-d_II,\\
\frac{d\hat{V}}{d\tau} & = -d_V\hat{V}+\alpha d_I I, \label{AEqs4}
\end{align}
\noindent where $\tau$ is time, $\hat{V}$ represents the density of virus particles at the tumour site, $U$ is the density of susceptible but virus-uninfected tumour cells, $I$ is the density of virus-infected tumour cells and the term $U+I$ corresponds to the total tumour cell population.
Tumour growth is controlled by nutrients and spatial limitations and is described by a Gompertz function, i.e. $g(U) = r\ln(K/U)U$. Here, $K$ represents the carrying capacity of the tumour and $r$ is its growth rate. This type of expression is well-known to reproduce the experimentally observed evolution of a number of proliferating tumours quite accurately~\cite{laird1964dynamics}. In our framework, the likelihood of a virus infecting a tumour cell is assumed to depend on the number of tumour cells available to infect. To model this, a frequency-dependent function, rather than a simple mass-action term, is introduced: virus particles at the tumour site infect susceptible tumour cells according to the expression $\beta U\hat{V}/(U+I)$, where $\beta$ is the infectivity rate. This also differentiates this model from existing, well-known virus dynamics models, such as systems used for influenza or HIV modelling~\cite{DeLeenheer20031313, Wang200644}.
\begin{figure}[h!]
\centering
\includegraphics[width=80mm]{Schem11}
\caption{Flow diagram for the interaction between a population of uninfected tumour cells, $U$; virus-infected tumour cells, $I$; and virus particles, $V$. The diagram lists parameters relating to the original model Eqs.~(\ref{AEqs1})-(\ref{AEqs4}), in grey boxes and parameters relating to the non-dimensional form of the model, Eqs.~(\ref{E4})-(\ref{E6}), in blue boxes.}
\label{Schem11}
\end{figure}
After initial injection, any virus subsequently produced via replication within tumour cells will not have the PEG or Herceptin modifications. To account for this, only single average infectivity and decay rates $\beta$ and $d_V$ are used for the combined populations of original and replicated virus, noting that the population is dominated by naked (replicated) virus over the majority of the time course of the experiments. Fig.~\ref{Schem11} depicts the flow diagram of the three populations described in the equations, and we refer to the original study \cite{jenner2018mathematical} for a discussion on biologically relevant ranges of values of the parameters.
To proceed with our mathematical analysis, an appropriate change of variables detailed in Appendix A is used to scale the above equations into dimensionless form. The final result is given as follows:
\begin{align}
\frac{dU}{dt} &= m\ln\left(\frac{K}{U}\right)U-\frac{UV}{U+I},\label{E4}\\
\frac{dI}{dt} &= \frac{UV}{U+I}-\xi I,\\
\frac{dV}{dt} &= -\gamma V+\xi I\label{E6}
\end{align}
\noindent where $m = \displaystyle\frac{r}{\beta}, \xi = \displaystyle\frac{d_I}{\beta}, \gamma =\displaystyle \frac{d_V}{\beta}$ and $\hat{\beta} = \displaystyle\beta \alpha$ are dimensionless parameters, and $t$ represents a dimensionless ``time''. This model still follows the schematic given in Fig.~\ref{Schem11} and is the object of the present study. The three parameters $m, \xi$ and $\gamma$ that regulate the behaviour of the system represent tumour growth rate, viral death rate and viral potency (or infectivity), respectively. As a result of the non-dimensionalisation process, where parameters are all scaled by the infectivity rate (see \ref{sec:appendix}), the rate of conversion of uninfected cells $U$ to infected cells $I$ due to the viral load $V$, i.e. the term $\pm\displaystyle \frac{UV}{U+I}$, is not affected by any parameter.
\section{Local stability analysis}
A local stability analysis of Eqs.~(\ref{E4})-(\ref{E6}) shows a number of interesting features. Of particular relevance is the existence of a stable equilibrium corresponding to eradication, which is characterised by a singular Jacobian matrix. This solution can coexist with other equilibria, for example a stable spiral or a stable node, which instead corresponds to incomplete eradication of the tumour. As we will show shortly, this occurrence can give rise to bistability for some biologically relevant parameter ranges.
\subsection{Equilibrium solutions}
Setting the right-hand-side of Eqs.~(\ref{E4})-(\ref{E6}) to zero, three equilibria are found: (a) a solution at a value for the uninfected cells equalling the carrying capacity, indicating a treatment with no effect; (b) a non-zero solution representing incomplete eradication, characterised by a quiescent tumour despite the viral load being constant and non zero; and (c) an equilibrium at the origin corresponding to complete eradication of the tumour. The populations corresponding to such cases are
\begin{align*}
(a)& \quad U= K, \ \ I=0, \ \ \ V=0; \\
(b)& \quad U= K\exp\left(\displaystyle\frac{\xi}{m\gamma}(\gamma-1) \right)=U^*, \ \ I= \frac{K}{\gamma}(1-\gamma)\exp\left(\displaystyle\frac{\xi}{m\gamma}(\gamma-1) \right) =I^*, \\
&\quad V = \frac{K\xi}{\gamma^2}(1-\gamma)\exp\left(\displaystyle\frac{\xi}{m\gamma}(\gamma-1) \right)= V^*;\\
(c)& \quad U = 0, \ \ \ I=0, \ \ \ V=0.\\~\label{nonzeroequil}
\end{align*}
\noindent The Jacobian of the system is given by
\begin{align}
J = \left(\begin{array}{ccc}
m\ln\left(\displaystyle\frac{K}{U}\right)-m-\displaystyle\frac{VI}{(U+I)^2} & \displaystyle\frac{UV}{(U+I)^2} & -\displaystyle\frac{U}{(U+I)}\\[8pt]
\displaystyle\frac{VI}{(U+I)^2} & -\xi-\displaystyle\frac{UV}{(U+I)^2} &\displaystyle\frac{U}{U+I}\\[8pt]
0 & \xi& -\gamma
\end{array}\right),
\end{align}
and we discuss the character of the eigenvalues for the above equilibria below.
\subsection{Stability of ineffective treatment equilibrium: $U=K$, $I=0$, $V=0$}
The first equilibrium (a) corresponds to a failed treatment where uninfected tumour cells $U$ grow to the system's carrying capacity $K$ and no viral particle survives. Evaluating the Jacobian at this point gives
\begin{align*}
J = \left(\begin{array}{ccc}
-m & 0 & -1\\
0 & -\xi &1\\
0 & \xi& -\gamma
\end{array}\right),
\end{align*}
which gives rise to the following characteristic equation:
\begin{equation}\label{ch-eq-partial}
\rho(\lambda; m, \gamma, \xi) = -(\lambda+m)\left(\lambda^2+(\xi+\gamma)\lambda+\xi(\gamma-1)\right).
\end{equation}
The overall stability of this equilibrium depends on the roots $\lambda_2$ and $\lambda_3$ of the quadratic factor, because the root $\lambda_1 = -m$ of the linear factor is negative, since the growth rate $m>0$. After calculating $\lambda_2$ and $\lambda_3$, we find that the equilibrium is either a stable node or stable focus when $\xi+\gamma>0$ and $\xi(\gamma-1)>0$. Since the parameter values in this model are considered to be always positive, the first condition holds. The second condition implies that, if $\gamma<1$, the equilibrium is unstable, and vice versa for $\gamma>1$. As we will show shortly, a one-parameter continuation in $\gamma$ shows that at $\gamma = 1$ a branch point is present and the treatment is always ineffective for a decay rate $\gamma>1$. Intuitively, if the virus dies too quickly, no infection can occur.
\begin{figure}[h!]
\centering
\includegraphics[width=0.44\textwidth]{new1second}
\includegraphics[width=0.44\textwidth]{new2second}\\[5pt]
\includegraphics[width=0.44\textwidth]{new3second}
\includegraphics[width=0.44\textwidth]{paramspaceasfuncofUstar2}
\caption{Regions representing the stability of the nonzero equilibrium, (a)-(c), and the influence of system parameters on tumour cell numbers at the equilibrium value $U^*$. In (a), the section of parameter space where the non-zero equilibrium is stable is shown. Note that (b) represents the volume in $(\xi,m,\gamma)$ giving rise to a stable node solution for the equilibrium $(U^*, I^*, V^*)$, whereas (c) is the section for a stable spiral. Combining the regions in (b) and (c) gives the volume in (a). Plot (d) is the stable parameter space for different values of $U^*$, within the following intervals: orange for $0.2<U^*<0.25$, yellow for $0.35<U^*<0.4$, green for $0.5<U^* < 0.55$ and blue for $0.65<U^*<0.7$. Note that these ``slices'' are almost symmetrical.}\label{nsregion}
\end{figure}
\subsection{Stability of partial eradication solution: $U=U^*$, $I=I^*$, $V=V^*$}
The model admits a second, non-zero equilibrium where a small tumour mass coexists with virus particles. The characteristic equation for this solution, after substituting $U^*,I^*,V^*$ in the Jacobian, is given by
\begin{equation}\label{characeqn}
\rho(\lambda) = -\lambda^3-\lambda^2(\gamma+m+\xi)+\lambda\left(\gamma m(\xi-1)+\frac{\xi^2}{\gamma}-\xi(2m+\xi)\right)+\gamma m \xi (\gamma-1).
\end{equation}
\noindent For this cubic, the Routh-Hurwitz criterion is used to deduce the parameter values that produce three roots with negative real part. This criterion states that, given a general cubic of the form $\rho(\lambda)=a_0\lambda^3+a_1\lambda^2+a_2\lambda+a_3$, two conditions need to be met simultaneously for all roots to have negative real parts, i.e.
\begin{equation*} (a)~\frac{a_1a_2-a_0a_3}{a_1}<0 \qquad \text{and} \qquad (b)~a_3<0
\end{equation*}
with, in our case, $a_0 = -1$, $a_1=-(\gamma+m+\xi)$, $a_2= \left(\gamma m(\xi-1)+\frac{\xi^2}{\gamma}-\xi(2m+\xi)\right)$ and $a_3=\gamma m \xi (\gamma-1)$.
Condition (b) is easily satisfied for $0<\gamma<1$, given that all parameters are assumed to be positive. Condition (a) requires that $a_1a_2>a_0a_3$, since $a_1<0$. The region in the $\xi,m,\gamma$ parameter space that satisfies this condition can be numerically computed and is depicted in Fig.~\ref{nsregion}(a). Using the discriminant of Eq.~(\ref{characeqn}) and imposing the appropriate conditions, subsections of that region corresponding to a stable node or stable spiral are illustrated in Fig~\ref{nsregion}(b, c). Note that all regions are smooth and connected.
It is also interesting to consider which parameter regimes result in a low tumour burden (or threshold) $U_T$. To visualise how the value of the equilibrium $U^*$ changes as a function of parameter values, we can compute the regions of parameters space satisfying the following equality for a given threshold $U_T$:
\begin{equation}
\xi = \frac{m}{\gamma-1}\ln\left(\frac{U_T}{K}\right).\label{Ustarcontoursurfaces}
\end{equation}
\noindent Plots for four different $U_T$, varying within intervals, are shown in Fig.~\ref{nsregion}(d). The regions are roughly symmetric, with parameter $\gamma$ being the major contributor to changes in $U^*$ values. For example, when $\gamma \lessapprox 0.5$, there is a set of $\xi$ and $m$ values resulting in $0.20\lessapprox U^*\lessapprox 0.25$. Since $m$ represents the growth rate of tumours and $U^*$ is almost insensitive to its variations, our analysis indicates that a value of $\xi$ can always be chosen to decrease the volume of the tumour, as long as the decay rate $\gamma$ is low (i.e. the virus does not decay too quickly).
\subsubsection{Stability of full eradication solution: $U=0$, $I=0$, $V=0$}
The last equilibrium of the model represents the case of complete eradication, where all variables are zero. As anticipated, the Jacobian is singular due to the presence of logarithmic and rational terms in $U$ and $(U+I)$ respectively. An analytical treatment is not possible and, in particular, the presence of logarithmic terms $m\ln(K/U)$ or its source in Eq.~(\ref{E4}), i.e. $m\ln(K/U)U$, is not treatable with straightforward expansions for $U\to 0$. A different approach based on numerical integration and computation of eigenvalues under specific assumptions on $U$, $I$ and $V$ is instead used and will be discussed in detail in the next section.
As far as the equilibrium's stability is concerned, it turns out that the eradication solution can be stable or unstable, depending on the value of model parameters. As a general rule, we observe that parameter sets where $\xi$ is high, corresponding to a potent viral load, tend to yield a stable equilibrium as long as the decay rate $\gamma$ is not excessive. This suggests that the engineered virus has to be potent and sufficiently resilient: one characteristic alone is not sufficient. If, for example, the virus has potency $\xi$ but dies too fast, then the equilibrium turns into an unstable point and no eradication is possible. A clear picture of how eradication depends on viral characteristics will emerge with the aid of bifurcation plots, which are discussed in the next section.
\section{Characteristic dynamical regimes}
\label{section:4}
The model supports a number of dynamical regimes that are interesting both from the biological and mathematical point of view. In Fig.~\ref{Fdyn}, four distinctive behaviours associated with the equilibria previously described are presented. Case (1) is an example of an equilibrium solution where the virus co-exists with uninfected and infected tumour cells, i.e. the case $U=U^*$, $V=V^*$ and $I=I^*$. The time series is for an attracting node, but similar long-term dynamics exists for the case of an attracting spiral, with the only difference being an initial, oscillatory transient that then damps down to a plateau. Note how the uninfected cells $U$ are the first to reach the equilibrium $U^* = K\exp(\frac{\xi}{m\gamma}(\gamma-1))$, which corresponds, for the chosen parameters, to $U^* \approx 40.65$. Case (2) corresponds to stable oscillations, characterised also by an almost quiescent phase where the system variables are close to zero and periods of growth and decay of cells and virus. Generally, we observe that this ``refractory'' state tends to have a longer duration than the active phase. Also in this case, the uninfected cells $U$ are the first to grow, with a subsequent increase in the infected cells $I$ and then in the virus load $V$. As we will see shortly with a bifurcation analysis, the duration of the ``rest'' and ``active'' phases of oscillations depends on the system parameters and changes continuously from case (2) to the limiting case (3). This is an extreme scenario where the system oscillates between two long plateaus of quasi-complete eradication (i.e. $U=I=V\approx 0$) and quasi-ineffective treatment (i.e. $U\approx K=100, ~I=V\approx 0$). The inset shows the almost square-wave appearance of the system's trajectories on a long time scale, whereas the switch from the two states is illustrated in the main figure, showing how the growth in $I$ and $V$ causes the uninfected cell numbers to decrease. It is important to note that the system cannot stabilise on either equilibra, because for the parameters chosen and as it will be evident shortly from bifurcation results, both equilibria are unstable.
\begin{figure}[h!]
\centering
\includegraphics[width=0.48\textwidth]{panel1final}
\includegraphics[width=0.48\textwidth]{panel2final}\\[6pt]
\includegraphics[width=0.48\textwidth]{panel3final}
\includegraphics[width=0.48\textwidth]{panel4final}
\caption{Numerical simulations of Eqs.~(\ref{E4})-(\ref{E6}) demonstrating different types of dynamics, for initial conditions $U(0)=50, I(0) = 10, V(0)= 10$ and fixed parameters $m = 0.1, \gamma = 0.1$, whereas values for $\xi$ are increasing from case (1) to (4). Type (1) corresponds to a stable co-existence of virus and tumour due to incomplete eradication, occurring at $\xi = 0.01$, (2) depicts a stable oscillatory solution for $\xi = 0.06$, (3) shows stable long period oscillations of almost ``square wave'' shape for $\xi = 0.097$ and (4) is a case of complete eradication for $\xi=0.12$. Note that the carrying capacity is chosen as $K=100$.}\label{Fdyn}
\end{figure}
Finally, a complete eradication solution is depicted in case (4). Although for the chosen initial conditions and parameters the model shows a monotonic decline to zero for $U$, other examples have been found where $U$ first shows a maximum, followed by an exponential decrease. Also in this final case, as for the other three scenarios just discussed, we observe that $U$ is the fastest to reach its equilibrium value, with $I$ and $V$ following.
To appreciate where these regimes occur and how the parameters influence their existence, two bifurcation plots with respect to system variables $\xi$ and $\gamma$ versus $U$ are presented in Fig.~\ref{Fbif}. In both plots, stable branches are indicated with continuous lines, whereas unstable ones are dashed. The two black branches at $U=0$ and $U=K=100$ indicate the full eradication and failed treatment solutions, respectively. The red line indicates the partial eradication case, where a non-zero value for the tumour volume and the viral load is present. Numbers point to areas where the typical dynamics just discussed can be found.
For the case of a codimension one plot with respect to $\xi$ (Fig.~\ref{Fbif}(a)), two branch points are present: one at $U=100$ and $\xi=0$, where the partial eradication solution coalesces with the failed treatment case, and a second at $U=100$ and $\xi \approx 0.098$ where the oscillatory, stable branch (green line) terminates. This branch originates from a supercritical Hopf bifurcation (HB), which causes the initial partial eradication branch to lose its stability. Note how, at this value of $\xi$, a change in the stability of the eradication solution $U=0$ (black line) also happens, with a saddle-node bifurcation (SN) occurring and a stable, fully eradicating regime appearing for $\xi>\xi_{SN} \approx 0.098$. This eradication solution branch regains its stability at $\xi =0$ through another saddle-node bifurcation (SN). Note also that the partial and full eradication branches (i.e. red and black lines, respectively) do not intersect. Finally, let us remind to the reader that solutions for parameter values that are negative do not bear any biological value.
\begin{figure}[h!]
\centering
\includegraphics[width=0.49\textwidth]{bif_A}
\includegraphics[width=0.49\textwidth]{bif_B}
\caption{Examples of typical bifurcation plots in one parameter for the model, with the case of $\xi$ in (a) and $\gamma$ in (b), both versus $U$. Numbers correspond to the dynamical regimes illustrated in Fig.~\ref{Fdyn} and, for the case of periodic orbits originating from a Hopf bifurcation, only the maximum value of $U$ is shown. For (a), the other model parameter are $m=0.1$, $\gamma = 0.1$. Note that the switch to regime $4$ (complete eradication) occurs when the branch of periodic orbits (in green) ceases to exist, for a value $\xi \approx 0.098$. Similar results for a continuation in $\gamma$ are shown in (b), with the switch to case $(4)$ dynamics also occurring in correspondance of a branch point for the periodic orbit, at $\gamma \approx 0.0103$. An inset with a magnification on the area that shows the richest dynamical variability is also shown. The value of the other, fixed parameters are given in this case by $m=0.1$, $\xi = 0.01$. In both cases, solutions for negative $\xi$ and $\gamma$ have been included for reasons of consistency, but do not correspond to any biologically meaningful state.}\label{Fbif}
\end{figure}
It is worth noting that, along the red branch of coexisting solutions, $U$ can span a large range of values, with $U$ increasing as the viral potency $\xi$ decreases. For example, close to the HB, which occurs at $\xi_{HB}\approx 0.042$, a partial eradication solution for $\xi=0.04$ gives a tumour burden $U\approx 2$. Note also the extension of the plateau of the periodic branch (green) close to the $U=100$ unstable equilibrium, before the branch point. This indicates that a ``square wave'' type of oscillations can be present for a moderately extended parameter interval in $\xi$.
Although not shown in the diagram, the switch between the node and the spiral equilibrium typical of the partial eradication solution takes place along the red branch. For the chosen parameters in Fig.~\ref{Fbif}(a), this happens at $\hat{\xi} \approx 0.01675$, with spirals existing for a value $\xi$ such that $\hat{\xi}<\xi <\xi_{HB}$. Generally speaking and as shown in Fig.~\ref{nsregion}(b)-(c), the value at which the equilibrium type changes depends also on the other parameters $m,\gamma$ of the model.
The system's behaviour also shows a strong, nonlinear dependence on viral death rate $\gamma$, as illustrated in Fig.~\ref{Fbif}(b). With respect to the case of $\xi$, the sensitivity of the model to $\gamma$ is somewhat reversed: intuitively, a surge in potency should act on the model in a similar way as a reduction in death rate and vice versa. For example, the branch of oscillatory solutions (green) out of the supercritical Hopf bifurcation (HB) shows an increasing maximum in $U$ as $\gamma$ decreases, opposite to what happens for $\xi$ (see the inset, in particular).
The stable, impartial eradication solution branch (red) shows higher tumour volumes with increasing $\gamma$, and coalesces with the unstable $U=100$ branch (in black) at $\gamma = 1$. For $\gamma>1$, the ineffective treatment solution is stable, as previously found from the analysis of the characteristic equation corresponding to this solution, i.e. Eq.~(\ref{ch-eq-partial}). A virus with a decay rate $\gamma > 1$ has no effect on the tumour. It is important to note that a mechanism identical to what we observe in the bifurcation plot for $\xi$ allows the existence of case (4) solutions, i.e. complete eradication. At a value of $\gamma \approx 0.0103$, the inset shows the termination of the oscillatory solutions (in green) and the occurence of a saddle-node point in the full eradication branch, making complete destruction of the tumour possible. From the biological perspective, this indicates that a right balance between the potency of the virus and its mortality must be achieved for an eradication to occur, depending on the growth rate $m$ of the tumour. In particular, as $\gamma$ is increased from zero, the model goes from full eradication to oscillations with an amplitude that decreases with $\gamma$, and subsequently to incomplete eradication up until $\gamma = 1$.
As previously mentioned, the full eradication solution gives rise to a singular Jacobian, making a purely numerical approach to continuation impossible. For solutions where $U\neq 0$, results have been obtained by using AUTO~\cite{AUTO07} and XPPAUT~\cite{Bard2002} softwares. For the case of solutions occurring for $U=0$, a combination of numerical methods and symmetry arguments have been employed. We assume that $U < I < V$, as exemplified by case (4) shown in Fig.~\ref{Fdyn}. If $\epsilon > 0$ and small, and we impose that $U\to \epsilon^n$, $V\to \epsilon^m$ and $I \to \epsilon^l$ with $n>m>l$, then the eigenvalues of the Jacobian $J$ can be numerically computed with an increasing approximation for growing $n,m$ and $l$.
For example, in determining the stability of the full eradication branch in Fig.~\ref{Fbif}(a), we consider $U = 10^{-7}$, $V = 10^{-5}$ and $I = 10^{-4}$, substitute these values in the Jacobian and numerically evaluate the eigenvalues. For $\xi>\xi_{SN}\approx 0.0975$, all three eigenvalues turn out to be negative and real, whereas for $\xi<\xi_{SN}$ two are positive and one is negative. For example, choosing $\xi = 0.095$ gives eigenvalues $\lambda_1 \approx -0.15$, $\lambda_2 \approx -0.06$ and $\lambda_3 \approx 8\cdot 10^{-5}$. For the case $\xi = 0.099$, the first two eigenvalues are almost unchanged, but the last one changes sign and is $\lambda_3 \approx -2 \cdot 10^{-3}$. Similar results hold for the SN on the eradication branch for continuation in $\gamma$ (see Fig.~\ref{Fbif}(b)), and the method is consistent for all the parameters $m,\gamma$ and $\xi$ we have tested (not all shown here). These results have also been checked by integrating the equations of motion with XPPAUT, and confirming that the solution is indeed attracting when stable or repelling when unstable, as reported in the bifurcation diagrams.
One important feature of the model is that it does not support stable oscillations for all biologically meaningful combinations of parameters. For some choices, a different structure of bifurcation plots emerges, with significant consequences from the biological perspective. In this sense, a typical example for a continuation in $\xi$ is illustrated in Fig.~\ref{Fbista}(a). An unstable periodic branch (green) originates from a subcritical Hopf bifurcation (HB) and maintains its unstable character until it collapses with the $U=K=100$ (black) branch. For this diagram, viral potency $\gamma$ is the same as in Fig.~\ref{Fbif}(b), but a value of $m=0.5$ (moderately high growth rate) is chosen, whereas both previous diagrams have been obtained with a $m=0.1$ (moderate growth rate). A more aggressive tumour, assuming that the potency of the virus is the same, does not engage in oscillatory behaviour with the virus, but only partial or full eradication are possible (i.e. black and red lines).
It is interesting to stress that in this case, as shown in Fig.~\ref{Fbista}(b), the saddle-node (SN) on the full eradication $U=0$ branch (in black) occurs for a value $\xi_{SN}$ that is less than the value $\xi_{HB}$ at which the subcritical Hopf (HB) originates. This occurrence is due to the fact that the periodic branch shows increasing values of max $U$ for decreasing values of $\xi$ when it is unstable. This is the opposite of what happens for the stable periodic branch described in Fig.~\ref{Fbif}(a), where $\xi_{HB} < \xi_{SN}$ and the stability of the eradicated solution does not switch in this way.
\begin{figure}[h!]
\centering
\includegraphics[width=0.49\textwidth]{unst_1}
\includegraphics[width=0.49\textwidth]{unst_2}\\[5pt]
\includegraphics[width=0.49\textwidth]{bi1_sp}
\includegraphics[width=0.49\textwidth]{bi1_er}
\caption{Bifurcation plots and bistable solutions for fixed parameter values $m=0.5$, $\gamma=0.1$. The rectangle in (b) shows the area where two solutions of different nature coexist, delimited by $\xi_{SN} \approx 0.1359$ and $\xi_{HB} \approx 0.1388$. A spiralling solution to an incomplete eradication is shown in (c) and occurs for initial conditions $U(0)=60$, $I(0) = 10$, $V(0)=40$, for a parameter $\xi_{SN} < \xi = 0.136 < \xi_{HB}$ . A fully eradicated solution is shown in (d) and instead occurs for $U(0)=40$, $I(0) = 10$, $V(0)=5$, for the same value $\xi = 0.136$. Nullclines, i.e. the loci of points corresponding to $U'=0$ and $V'=0$, are in red and green.}\label{Fbista}
\end{figure}
The change in the order in which the SN and HB emerge as $\xi$ is increased is responsible for the generation of a region of bistability, where two separate and distinct equilibria exist for an interval of potency values. For values of $\xi$ in this region, different initial conditions can lead to different outcomes, as shown in Fig.~\ref{Fbista}(c)-(d): the initial dosage of viral load and the numbers of infected and uninfected tumour cells can strongly influence the final fate of the system and, as it will be clear shortly, lead to somewhat unexpected results. In the first case (Fig.~\ref{Fbista}(c)), a spiralling solution achieves an incomplete eradication, which belongs to the red branch in Fig.~\ref{Fbista}(b). Conversely, the second case shows a complete eradication to a vanishing tumour, after traversing two maxima in $U$ and $V$ respectively, corresponding to the black branch in Fig.~\ref{Fbista}(b). A small variation in the initial conditions can hence result in the therapy being effective or instead giving rise to a partial eradication.
The existence of this area of bistability is associated with the presence of a subcritical Hopf bifurcation whose loci of points in $\xi$ and $m$, and for different values of $\gamma$, are plotted in Fig.~\ref{2par}(a). Generalised Hopf points (GH) separate subcritical Hopf points (dashed lines) from supercritical Hopf bifurcations (continuous lines). Note that, if the the growth $m$ is sufficiently small, no Hopf bifurcation can be present and the system does not support oscillations, either stable or unstable. For example and as a result of the interruption of the Hopf branches shown in the inset of Fig.~\ref{2par}(a), any one-parameter bifurcation plot in $\xi$ for a fixed $\gamma =0.1$ and values of $m \lessapprox 0.008$, does not contain a stable or unstable oscillatory branch, since no Hopf point exists for such values. Biologically this indicates that there is a lower bound on the tumour growth rate for oscillations (stable or unstable) to exist, implying that a very slow growth in general leads to a complete eradication for sufficiently potent virus, as long as its death rate is not excessively pronounced.
A numerical analysis of the model for a range of $\xi,\gamma$ and $m$ values shows that limit cycle amplitudes for $U$ do not follow a clear pattern, as captured by Fig.~\ref{2par}(b). Oscillations of different amplitudes can be achieved by the system and, as observed previously, depending on the growth rate of the tumour they can be enhanced by increases in $\xi$ and decreases in $\gamma$.
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{2p_collections}
\includegraphics[width=.45\textwidth]{limitcycamp2}
\caption{Different, two-parameters continuations in (a) for $m$ and $\xi$ for branches of Hopf bifurcations at different values of $\gamma$. Branches of supercritical Hopf bifurcations are shown in continuous lines, whereas those for subcritical bifurcations are in dashed lines. Generalised Hopf points are indicated by GH. Note that the branches cease to exist for low values of $(m, \xi)$, indicating the system cannot support either stable or unstable oscillations when parameters are sufficiently small (see the inset). Plot of the corresponding amplitude of stable limit cycles for points in the $\xi,m,\gamma$ parameter space are in (b). The colour of the point corresponds to the maximal value of the amplitude of the limit cycle in $U$.}\label{2par}
\end{figure}
\section{The effect of dosage applications and their optimisation}
As shown, there are large sections of parameter space that give birth to regimes with dormant tumours or tumour-virus oscillations, which can give rise to different outcomes when coupled with clinical therapies. For the purpose of the present study, one of the simplest possible therapeutic practices is considered: the administration of constant dosages of viral loads via external injections and at given time intervals. If the treatment is over the course of $n$ injections with $\kappa$ number of days between injections, a virus injection protocol $u_V(t)$ can be summarised by the following, generic schedule:
\begin{equation}
u_V(t) = \left\{\begin{array}{cc} \displaystyle\frac{D_0}{n} \qquad & \qquad t = (i-1)\kappa, \hspace{0.5cm}
\text{where }i = 1, \ldots, n, \\[6pt] 0\qquad & \text{otherwise.} \end{array} \right.\label{Eq5}
\end{equation}
\noindent
Given this simple scheme, let us consider how dosage perturbations affect regions of the bifurcation diagrams and if they can result in either tumour eradication or a stable tumour size below a given threshold. The two typical scenarios we consider are oscillations and bistability.
\subsection{Effects of injections on a stable, oscillatory trajectory}
After exploring different areas of the parameter space that give rise to oscillations, the main finding reported in this study is that simple therapies given by Eq.~(\ref{Eq5}) do not alter the long term behaviour of the model independently of the amplitude or period of oscillations. If an oscillatory, stable state exists between virus and tumours, an increment in the viral load through injections does not achieve complete eradication to zero. From the dynamical point of view, an increase in viral load via external perturbation cannot force the system out of the basin of attraction of a stable limit cycle. Nonetheless, transient phenomena do exist and are worth discussing.
Let us consider two injections, i.e. $n=2$, for a system already in a stable oscillatory state. The number of days $\kappa$ between injections alters the size of the tumour and virus populations as the system returns to its stable state. In Fig.~\ref{D0sims}, a single period of two different stable limit cycles of the model is shown, with arrows representing the instants at which injections that increase the viral load have been administered. The corresponding maximum and minimum tumour size, along with the maximum virus count reached, is also presented.
\begin{figure}[h!]
\includegraphics[scale=.45]{injectionmap
\includegraphics[scale=.45]{injectionmap2}\\[5pt]
\includegraphics[scale=.45]{injectionmap3
\includegraphics[scale=.45]{injectionmap4
\caption{Perturbations in the days between two treatments $\kappa$. Two different limit cycle regimes have been plotted for $\gamma = 0.1, m = 0.2$ (a) $\xi = 0.06915$ and (c) $\xi = 0.06993$. The maximum and minimum uninfected cell number is plotted in (b) and (d) for the corresponding value of $\kappa$ represented by an upward arrow in (a) and (c). Note the different scales, since the oscillations considered in (a) and (c) have different amplitudes.}
\label{D0sims}
\end{figure}
Injections that occur at different phases of the cycles have different outcomes. In particular and for large or small oscillations, as can be seen in Fig.~\ref{D0sims}(b)-(d) for the red and magenta curves around $t\approx 62$, dosing the virus close to the minimum in tumour population provides a typical outcome: the tumour initially responds to the injection by undergoing the lowest resulting minimum, but this is followed by a rebounds that causes $U$ to reach the highest value (max $U$ in the plot) of all other tested injections. Note that, in some cases and for sufficiently high dosages, the minima achieved by $U$ can be pushed to values so low to become experimentally undetectable. Injections at other instants within one oscillation period yield rebounds proportional to the original amplitude of the limit cycle, with best results occurring for the lowest amplitudes.
Perturbing the number of days between injections $\kappa$, the total injection amount $D_0$ or the number of injections $n$ does not affect the long term dynamics (not shown here), which remains oscillatory in the long term.
\subsection{Effects of injections on a trajectory in the bistable region}
For a solution in a bistable region, the final outcome of any injection is highly dependent on the initial tumour size and viral load. In particular, due to the complex structure of the basin of attraction of the two competing solutions, i.e. full eradication and an incomplete quiescent state, doses that are higher than a specific threshold, which is in turn highly dependent on the system parameters, can lead to a partial eradication rather than a complete one.
As typical scenarios, consider the administration of single injections of increasing dosage as depicted in Fig.~\ref{bistabsim}. Depending on the initial uninfected tumour $U(0)$, injections can lead to different outcomes or even have no effect on the final state. Considering the case of a high tumour size (Fig.~\ref{bistabsim}(a), $U(0) = 100$), different dosages always result in final eradication. Some dosages can lead to transient oscillations in the $U-V$ plane, but eventually eradication is achieved for all plotted trajectories. If, instead, the initial tumour size is smaller (Fig.~\ref{bistabsim}(b), $U(0) = 50$), a full eradication is obtained only if the dose is either sufficiently low or sufficiently high, whereas there is a considerable interval of possible doses that push the system to a stable spiral corresponding to a dormant state, where eradication is not complete. Note that the first two low dosage injections, i.e. injections 1 and 2 in Fig.~\ref{bistabsim}(b), also lead to a final eradication state after few oscillations on the $U-V$ plane.
\begin{figure}[h!]
\centering
\includegraphics[width=0.48\textwidth]{bi-dose1}
\includegraphics[width=0.48\textwidth]{bi-dose2}
\caption{Typical cases of dependence on injected viral dosage $D_0$ for a system in a bistable scenario. Examples of two injections with increasing dosage (i.e. injections 1 and 2) are also sketched. The effect of these injections is to push the starting point to larger values of $V(0)$, depending on the dose that is administered. For the same initial tumour size, different dosages result in either tumour eradication or tumour stabilisation. Initial fixed conditions in (a) are given by $U(0)=100$, $I(0)=10$ and by $U(0)=50$, $I(0)=10$ for (b). In both cases, $V(0)$ varies from a minimum of $20$ to a maximum of $120$ in constant steps and the parameters are $m = 0.5$, $\gamma =0.1$ and $\xi = 0.138$.} \label{bistabsim}
\end{figure}
This result is interesting, as it suggests that, for given initial tumour size and characteristics of the virus, there can be a unique interval of dosage sizes that does not result in treatment success. Boosting the amount of virus does not always guarantee a successful outcome.
\section{Discussion}
The model proposed in this work shows a number of interesting features, both from the mathematical and the biological points of view. Firstly, a range of possible dynamical outcomes, based on the value of the model parameters and, in some cases, of the initial conditions, have been found. A number of nontrivial bifurcation scenarios have also emerged, with the presence of an important system equilibrium (i.e. full tumour eradication) that is characterised by a singular Jacobian. This occurrence has required the use of a hybrid combination of numerical continuation, symmetry considerations and integration of the model to map out the dynamics as a function of relevant model parameters.
The model provides a few insights into the interactions between an oncolytic virus and a tumour growing with a realistic proliferation law. One of the main limitations of the present approach is the endless influx of viral load that occurs in the model: once the viral cycle is set into motion, and unless viral death rate is excessive (i.e. $\gamma > 1$), there is no natural stopping mechanism for viral infections to continue endlessly. This simplification is, for example, responsible for the appearance of dormant, partially eradicated tumours, which, after an initial transient, perpetually coexist with a constant viral load. These dynamics are common for models with unlimited reservoirs of populations~\cite{Wilkie2013201} .
Another important constraint is represented by the limited number of parameters used and their inherent inability to fully account for tumour-virus dynamics in detail. For example, we have introduced two general terms $\xi$ and $\gamma$ that aim to capture the potency and death rate of the virus and depend on the virus infectivity rate, $\beta$. These parameters are meant to encapsulate a large variety of different viral characteristics and can be associated to features as diverse as burst size, reproduction rate, spreading ability, and diffusivity. A similar observation must be made for the growth rate $m$: this value condenses a large number of often independent and highly variable features of tumour growth, which are highly sensitive to nutrients, vascularisation, extra cellular matrix characteristics, and so on.
Notwithstanding these limits, the model shows that, for a given rate of growth, a tumour responds in different ways to viral particles that have different, generic characteristics. As shown in Fig.~\ref{Fbif}, an increase in viral potency $\xi$ or a decrease in viral death rate $\gamma$ drives the system through similar stages of typical dynamics, from partial eradication to tumour-virus oscillations. At sufficiently large values of $\xi$ and $\gamma$, for instance $\gamma > 1$, the scenarios are instead opposite, with full eradication and inefficient treatment, respectively.
A metastable regime that appears somewhat counterintuitive is represented by the so-called ``square-wave'' oscillations, which are observed in a small interval of biologically relevant parameters (see Fig.~\ref{Fdyn}(c)). Given the limitations of the model proposed here and the size of the parameter space where this dynamics takes place, it may be unlikely that such an extreme tumour expansions can be directly observed in a clinical setting. Nonetheless, the switch between a quasi-eradicated to a quasi-ineffective treatment regime points to the importance of achieving a complete wipe out of the tumour if a sudden resurgence is to be avoided.
The existence of an extended area of the parameter space where oscillations among system variables arise is also worth noticing. These regimes, which also tend to respond nonlinearly to external injections (see Fig.~\ref{bistabsim}), have been known for quite some time in clinical settings {\bf XXX add REF if existing - Adri, please}. One major finding for this model is that virotherapy can prevent oscillations from occurring if the potency is sufficiently strong or, alternatively, the virus tends to survive for sufficiently long times in the infected population. Furthermore, and this is particularly interesting, oscillations tend to have larger amplitudes and periods for increasing $\xi$ (or decreasing $\gamma$), before they disappear completely for sufficiently high (or low) values. This is worth reflecting on, especially from the clinical perspective. Designing a potent virus that is still not sufficiently resilient may turn out to be a riskier strategy, since it could trigger larger fluctuations in the tumour population. These oscillations also occur at relatively distant time intervals from each other and long periods of tumour inactivity may be misinterpreted as successful eradication. Looking at Fig.~\ref{Fbif}(a) and assuming that a low value of uninfected tumour cells $U$ represents a good outcome, a less potent virus, say with a $\xi \approx 0.04$, results in a quiescent tumour of a smaller size than the amplitude of oscillations caused by a highly potent virus with, for instance $\xi \approx 0.08$ (i.e. twice as potent). This is also true from the point of view of resilience, see in particular the inset of Fig.~\ref{Fbif}(b): a virus that remains active for longer, say $\gamma \approx 0.015$, produces oscillations with very high values of $U$, whereas a virus decaying twice as fast, say with $\gamma \approx 0.03$, produces a stable, silent tumour of a smaller size. All this shows that therapeutic strategies must be chosen carefully and thoughtfully, and that optimal design of an oncolytic virus must be targeted on the tumour characteristics, in particularly its proliferation rate. It could be quite interesting to test these theoretical findings in vitro.
Note also that, even when external interventions with extra viral dosages are taken into account, the answers provided by our analysis do not appear trivial. Firstly, the existence of bistability and dependence from initial conditions has important effects. As seen in Fig.~\ref{bistabsim}, different initial viral loads and dosages can result in different outcomes, often in unpredictable ways. It is not true that a larger initial viral load always results in eradication: there is a large interval of values of dosages for which eradication is not possible and, quite interestingly, the system privileges either a sufficiently high or sufficiently low viral load for eradication. Starting at a smaller viral load is successful because it first allows the tumour to initially grow to a larger size, which thus elicits a stronger viral response. This response can wipe out the tumour completely, with no risk of ending in a dormant phase. Although this feature has been observed previously, for example, in simpler systems in tumour-immune dynamics and predator-prey models~\cite{Davis1962, frascoli2013}, it is the first time, as far as the authors are aware, that it is noted in virotherapy modelling. Clearly, the fact that our model hypothesises that the virus can penetrate the tumour and diffuse within its cells with no hindrances, has to be taken into consideration and is one of the drivers of this effect. Notwithstanding this, the result points to the existence of a preferred threshold in the size, for some values of the system parameters, where a limited quantity of viral load is preferable over a larger amount. Although strategies resulting even in a controlled and partial growth of a tumour have to be evaluated and considered with extreme care, the fact that a low viral load can still produce positive outcomes should be investigated further in laboratory and clinical settings. We remind the reader that the present model does not allow for a thorough description of the dynamics of virus penetration and diffusion, which certainly play a fundamental role in the success of virotherapy.
Secondly, therapies that couple with external injections of viral loads could have very different outcomes depending on the state of the system. Not only, as we have just highlighted, they can perturb a trajectory that was meant to be of full eradication into a dormant state, but, as shown for oscillations in Fig.~\ref{D0sims}, they can have a transient, often negative effect on the whole system. If administered when the system resides on a stable oscillating state, these injections, depending on when in the cycle are provided, tend to increase the amplitude of few cycles of oscillations before the system goes back to its original fluctuations, with no ability of driving the model out of this phase. There is generally no positive relevant effect in reducing the magnitude of periodic behaviour in the long term. Strategies that instead optimise the quality of the oncolytic virus seem to be preferable, as (see Fig.~\ref{Fbif}) oscillations can be reduced or damped to zero either by increasing the potency or the life span of virus at the right amount.
In this sense and as also Fig.~\ref{2par}(a) explains, the finding that oscillations that exist for different values of the parameters are suppressed when the growth rate $m$ is sufficiently small is very relevant and informative for therapeutic choices. Rather than complex injection schedules or larger amounts of externally provided virus, this model seems to promote pharmacological interventions that aim at blocking or reducing the growth of the tumour. It will also be interesting, for further studies, to establish whether combination therapies or interventions specifically targeted at boosting the immune response (not modelled here) could also improve outcomes, and how changes in the diffusion and penetration efficiency of infection waves may change the trends observed in the present model.
\section*{Acknowledgements}
ALJ, FF and PSK gratefully acknowledge support for this work through the Australian Research Council Discovery Project DP180101512, ``Dynamical systems theory and mathematical modelling of viral infections''.
|
\section{Introduction}
\label{intro}
Graph Neural Networks (GNNs) are effective methods for analyzing graph data, and various downstream graph learning tasks such as node classification, similarity search, graph classification, and link prediction have benefited from its recent developments \cite{article15,article21,article22}. However, most of the existing GNNs frameworks are supervised learning methods that have many drawbacks \cite{article15,article21,article22,article34}. GNNs obtained by supervised learning tend to learn task-specific knowledge, and the learned feature representations are difficult to transfer to other tasks \cite{article4}. Besides, supervised learning requires labeled data as input, and the over-reliance on labeled information will lead to poor robustness. Moreover, obtaining labeled information for large amounts of training data is labor-intensive, especially in the presence of large-scale networks.
\par In recent years, there has been a tremendous development in graph contrastive learning. Contrastive learning adopts data augmentation to obtain semantically identical features and then maximize feature consistency across augmented views to learn the representation. Researchers initially applied the contrast learning framework in computer vision to overcome the drawbacks, , such as the need for data annotation and the poor transferability of the learned features \cite{article1,article4,article24,article25,article27,article31}.
\par Graph data augmentation is not simple to define in contrastive learning methods, in contrast to diverse data transformation techniques for images and text. Graph data augmentation is more complex because of the non-Euclidean nature of graphs \cite{article32}, and the existing research mainly focuses on improvements to graph data augmentation. Inspired by contrastive learning in images such as MOCO \cite{article24}, SimCLR \cite{article4}, and DIM \cite{article27}, \itshape etc\upshape., many methods in graph contrastive learning have been proposed.
Many studies involve data augmentation by changing the edges and nodes of the graph. GCC \cite{article38} performing random walks in the network to construct subgraphs about nodes for contrastive learning. CGNN \cite{article11} learns consistent representations of nodes through different sampled neighbors. GraphCL \cite{article10} designs four types of graph data augmentation: Node dropping, Edge perturbation, Attribute masking, and Subgraph, which are combined with various prior knowledge to select data expansion methods in practical application. GCA \cite{article2} designs augmentation schemes based on node centrality measures to highlight important connective structures on the topology level.
\par Some studies perform data augmentation through constructing local-global pairs and negative-sampled counterparts. DGI \cite{article32} and InfoGraph \cite{article37} migrates DIM \cite{article27} to graphs and propose unsupervised learning objectives based on Mutual Information (MI). GMI \cite{article13} focuses on arriving at graphical mutual information maximization in a node-level by directly maximizing MI between inputs and outputs of the encoder. \cite{article36,article18} further extends MI maximization to heterogeneous graphs. The unsupervised learning model of graph structure is trained by maximizing the MI between the graph’s local features and global features. GRACE \cite{article12} simplifies DGI \cite{article32} by obtaining graph node representations by maximizing node embeddings’ consistency between two graph views generated through structure and feature perturbations. GRACE even surpasses its supervised counterparts on transductive tasks.
\par Different from previous graph contrastive learning models, we improve the framework structures of graph contrastive learning rather than graph data augmentation methods. The idea behind our strategy is to align semantically identical graph view data in different latent spaces to obtain a more general and differentiated representation. Then train a superior graph contrastive learning model, where the importance weights of the embeddings in different latent spaces are learned adaptively by the attention mechanism. Inspired by hierarchy semantic alignment strategy strategy on Convolutional Neural Network (CNN) \cite{article17} and OhmNet \cite{article20}, a hierarchy‐aware unsupervised node feature learning approach for multi‐layer networks, we propose a contrasting framework for unsupervised graph representation learning, called Adaptive Multi-layer Contrastive Graph Neural Network (AMC-GNN), which introduces adaptive multi-layer contrastive loss into graph contrastive learning models.
Specifically, semantically identical graph data is first obtained by graph data augmentation, and then the consistency between the node embeddings of the two graph views is maximized by minimizing the adaptive multi-layer contrastive loss. By optimizing the embedding consistency of the middle layer and the final embedding space, the representation consistency of the embedding generated by the GNN encoder is improved. In addition, we also introduce an auxiliary training model to improve the performance of the model further.
\par Our contributions are as follows:
\par 1. We propose the adaptive multi-layer graph contrastive learning framework that can be generalized to existing GNN models. Extensive experiments have been conducted to demonstrate that AMC-GNN can provide comparable or better performance than supervised models in graph data tasks.
\par 2. We experimentally proved that AMC-GNN has stronger robustness than other unsupervised models under slight perturbation. And we also verified that AMC-GNN has better performance at different feature removal rates.
\par 3. We have explored the introduction of adaptive multi-layer contrastive loss and auxiliary models to improve performance through ablation experiments, and verified that adaptive multi-layer contrastive loss is the key to improved performance.
\section{Related works}
\label{sec:1}
\subsection{Contrastive Learning}
\label{sec:2}
Contrastive learning is a self-supervised learning method whose main idea is to train the feature encoder by making the positive samples as close as possible and the negative samples as far away as possible in the feature space. The effect of only using a single data augmentation method on learning representation is general \cite{article4}, so well-constructed embeddings is essential for learning good representations. In the field of computer vision, many studies have been conducted on data augmentation for contrastive learning \cite{article4,article24,article25,article27}. Generally, the same image is rotated, cropped, divided into subgraphs, and other transformations that do not change the image’s semantics to construct its positive sample pairs \cite{article1,article4,article31}. DIM \cite{article27}, AMDIM \cite{article28} uses the principle of maximizing MI to maximize the MI between local features and global features of the same image. SimCLR \cite{article4} proposed Projection Head, and SimCLR v2 \cite{article31} verified that a deeper Projection Head could improve the quality of feature representation, and similar structures were introduced in subsequent studies \cite{article24,article25}. For data such as text and audio, samples within a certain window are usually considered as positive pairs \cite{article29,article30}, considering their temporal order. In graph contrastive learning, graph data is not as informative as the geometric and structured information that images have, and most research has focused on exploring the augmentation of structured graph data \cite{article10}.
\subsection{Graph Representation Learning}
\label{sec:3}
Traditional unsupervised graph representation learning node2vec \cite{article14}, Deepwalk \cite{article19}, VGAE \cite{article33} focuses on local contrast, forcing neighboring nodes to have similar embeddings. These approaches over-emphasize the structural information encoded in graph proximity and have the disadvantage of being difficult to handle large-scale datasets \cite{article14,article19}. Recent work on graphs employs graph convolutional network (GCN) encoders that better than conventional methods.
In the graph field, most of the research has studied supervised models \cite{article15,article21,article22,article34}. The rise of contrastive learning has motivated great interest in studying unsupervised learning in GNNs \cite{article9,article32,article10,article11,article12,article38}. Many methods study data augmentation by changing the edges and nodes of the graph \cite{article38,article10,article11,article2}.
A series of contrastive learning methods seeking to maximize the Mutual Information (MI) between the input and its representations have been proposed. Inspired by DIM \cite{article27} and AMDIM \cite{article28}, DGI \cite{article32} propose contrasting learning between local and global representations on graphs to capture structural information better. HDGI \cite{article36}, DMGI \cite{article18} further combine the MI with the meta-paths in the heterogeneous network, learns the weights of different meta-paths, and fuses them to obtain the final graph node representation. GRACE \cite{article12} simplifies DGI \cite{article32} by obtaining graph node representations by maximizing node embeddings’ consistency between two graph views generated through structure and feature perturbations. GRACE highlight the importance of appropriately choosing negative samples, which is often neglected in previous InfoMax-based methods.
\par Previous work focused on graph data augmentation methods. However, the information contained in the latent spaces of the middle layers is ignored. Therefore, we propose a graph contrastive learning framework that enables embeddings to be closer in multiple feature spaces.
\section{Methodology}
\label{sec:4}
\subsection{Preliminaries}
\label{sec:5}
In unsupervised graph representation learning, $\mathcal{G}=(\mathcal{V},\mathcal{E})$ denotes the undirected graph, $\mathcal{V}=\left\{ {{v}_{i}}\left| 1\le i\le N \right. \right\}$ represents the node set , and $\mathcal{E}=\left\{ {{e}_{ij}}\left| 1\le i,j\le N \right. \right\}$ represents the edge set. We define the feature matrix and adjacency matrix as $\mathbf{X}\in {{R}^{N\times F}}$ and $\mathbf{A}\in {{R}^{N\times N}}$, respectively, where ${\*{{A}}_{i,j}}=1$, if $({{v}_{i}},{{v}_{j}})\in \mathcal{E}$, while ${\*{{A}}_{i,j}}=0$ means not. Our purpose is to learn a graph encoder that $f:{{R}^{N\times F}}\times {{R}^{N\times N}}\to {{R}^{N\times {F}'}}$ such that $f(\*{X},\*{A})=\*{H}=\left\{ {\*{h}_{1}},{\mathbf{h}_{2}},...,{\*{h}_{N}} \right\}$, where ${\*{h}_{i}}$ represents the embedding of node ${{v}_{i}}$, which can be used for downstream tasks such as node classification and community detection.
\subsection{Adaptive Multi-level Contrastive Graph Neural Networks}
\label{sec:6}
Graph contrastive learning framework generally composed of three components: Data Augmentation, Encoder, and Loss. The proposed AMC-GNN introduced the auxiliary training model and the adaptive multi-layer contrastive loss in Encoder and Loss compared to the previous models. Next, we will illustrate each component of AMC-GNN and the learning process in detail.
\subsubsection{Data Augmentation for Graphs}
\label{sec:7}
The generation of different positive and negative sample pairs is an essential part of contrastive learning, because efficient positive and negative sample pairs can provide the most informative representation for downstream classification tasks.
\paragraph{Random Sampling Neighbours(RN)}~{}
\newline
For any node ${{v}_{i}}$, use $\mathcal{N}(i)$ to denote its original neighborhood specified by GNN architectures. When aggregating neighborhood nodes, randomly select neighboring nodes in $\mathcal{N}(i)$ for aggregation. To obtain different neighborhoods, we randomly dropped edges with proportion $\rho $ based on the original neighborhood to obtain new neighborhoods, denoted by $D\left( \mathcal{N}(i),\rho \right)$. In graph contrastive learning, the best performance is achieved by comparing first-order neighbor coding and graph diffusion \cite{article5}. Therefore, we obtained two subneighborhoods $\mathcal{N}{{(i)}_{1}}$ and $\mathcal{N}{{(i)}_{2}}$ by two independent random sampling:
\begin{equation}
\label{equ:1}
\mathcal{N}{{(i)}_{1}},\mathcal{N}{{(i)}_{2}}\sim D\left( \mathcal{N}(i),\rho \right).
\end{equation}
To improve the computational efficiency, we used a simplified approach while performing random discarding of the entire graph connection. The original graph is randomly dropped twice independently to obtain two different graphs ${{\mathcal{G}}_{1}}$ and ${{\mathcal{G}}_{2}}$. The neighborhoods of any node ${{v}_{0}}$ in the new graphs ${{\mathcal{G}}_{1}}$ and ${{\mathcal{G}}_{2}}$, are used as $\mathcal{N}{{(i)}_{1}}$ and $\mathcal{N}{{(i)}_{2}}$.
\paragraph{Attribute Masking(AM)}~{}
\newline
We randomly mask some of the dimensions in the node features with zeros, and the proportion of the masked dimensions is $p$. Formally, we first construct a random mask vector $\*{m}\in {{\left\{ 0,1 \right\}}^{F}}$, with each dimensional component of $\*m$ independently drawn from a Bernoulli distribution with probability $1-p$, and ${{m}_{i}}$ is the $i$th component of the vector, where ${{m}_{i}}\sim \mathcal{B}\left( 1-p \right)$, $1\le i\le N$. Use $\*{\tilde{X}}$ to denote the feature matrix $\*X$ processed by the AM,
\begin{equation}
\label{equ:2}
\*X= [{\*{{x}}_{1}},{\*{{x}}_{2}},...,{\*{{x}}_{N}}] , \*{\tilde{X}}= [{{\*{\tilde{x}}}_{1}},{{\*{\tilde{x}}}_{2}},...,{{\*{\tilde{x}}}_{N}}],
\end{equation}
\noindent where ${{\*{\tilde{x}}}_{i}}={\*{x}_{i}}\circ \*m$, $\left( \circ \right)$ denotes the inner product.
\par Our framework jointly utilizes both RN and AM methods to obtain semantically identical graph data for subsequent graph contrastive learning. The generation of two semantically identical graph data ${{\mathcal{G}}_{1}}$ and ${{\mathcal{G}}_{2}}$ is affected by the hyperparameters ${{p}_{R,1}}$, ${{p}_{A,1}}$, ${{q}_{R,1}}$, ${{q}_{A,1}}$ and ${{p}_{R,2}}$, ${{p}_{A,2}}$, ${{q}_{R,2}}$, ${{q}_{A,2}}$. ${{p}_{*}}$, ${{q}_{*}}$ represent the hyperparameters for target and auxiliary encoder data augmentation, respectively, and the subscripts R and A denote RN and AM.
\subsubsection{GNN-based encoder}
\label{sec:8}
In our model, two GNNs in series are used as encoders, where the first model is the target encoder and the second model is the auxiliary training encoder. The idea behind this is to increase the number of layers of the encoder to generate more embeddings of different layers as input to the adaptive multi-layer contractive loss, adding more constraints in the process of target encoder optimization. ${{l}_{1}}$ and ${{l}_{2}}$ indicate the number of layers of the target and auxiliary models, respectively, and the general number of layers of the encoders are two. The first model’s output embeddings are used as the final outputs and the feature vectors for the second model’s input. Let ${{f}_{1}}(\cdot )$ and ${{f}_{2}}(\cdot )$ be target model and auxiliary training model, then feature representation in different layers could get through Eq. \ref{equ:3}.
\begin{equation}
\label{equ:3}
{\*{{h}}^{n}}=f_{1}^{n}(\*{\tilde{X}},\*{\tilde{A}}), {\*{{h}}^{m}}=f_{2}^{m-{{l}_{1}}}({\*{{h}}^{{{l}_{1}}}},{\*{\tilde{A}}}'),
\end{equation}
\noindent where $f_{*}^{k}(\cdot )$ means passing through the $k$-layer GNN, $1\le n\le {{l}_{1}}, {{l}_{1}}<m\le {{l}_{1}}+{{l}_{2}}$, $\*{\tilde{A}}$ and $\*{{\tilde{A}}}'$, respectively, formed from $\mathbf{A}$ after two independent RN.
\begin{figure}
\centering
\includegraphics[width=0.50\textwidth]{encoder.png}
\caption{The GNN-based encoder.}
\label{fig:1}
\end{figure}
\par For the $i$th node, where $1\le i\le N$, the node after one graph transformation is ${{u}_{i}}$ and after another graph transformation is ${{v}_{i}}$. Nodes ${{u}_{i}}$ and ${{v}_{i}}$ form embedded vectors $\*{h}_{{{u}_{i}}}^{k}$ and $\*{h}_{{{v}_{i}}}^{k}$ respectively after passing through $k$-layers of GNN. Then the neural network Projection Head is used to map the embedding to the contrast space to obtain the vectors $\*{z}_{{{u}_{i}}}^{k}$ and $\*{z}_{{{v}_{i}}}^{k}$, where $1\le k\le {{l}_{1}}+{{l}_{2}}$. The Projection Head is composed of a 2-layer multi-layer perceptron (MLP) to enhance the expression power of the critic \cite{article17,article26} to avoid the loss function that computes similarity from dropping some important features during training \cite{article4}. The process can be represented as:$\*z={\*{W}_{2}}\left( \sigma \left( {\*{W}_{1}}*\*{h} \right) \right)$, where $\sigma \left( \cdot \right)$ is the activation function and $\*W$ is linear layer.
\begin{figure}
\centering
\includegraphics[width=0.65\textwidth]{model.png}
\caption{The illustrative schematic diagram of our proposed framework AMC-GNN. The model consists of three components: 1. GNN-based encoder: two GCN encoders are used to generate the feature representation vectors; 2. Projection Heads: four different projection heads are used to project the resulting embedding vectors into the loss space; 3. Contrastive loss: calculates the sum of positive and negative sample contrastive losses for the middle layer as well as the final embedding space.}
\label{fig:2}
\end{figure}
\subsubsection{Adaptive Multi-layer Contrastive Loss}
\label{sec:9}
Different from previous graph contrastive learning algorithms, we extend the contrastive loss to learn a more distinguishing feature representation in different layers. Specifically, the previous contrastive GNN model uses only the embedding layer vectors generated in the last layer after training the encoder, while the optimization of the middle hidden layers is only performed by backpropagating the gradient to the earlier layers. In this case, we extend the contrastive loss proposed in \cite{article12} by adding each layer’s embeddings to the contrastive loss. Combined with the attention mechanism, information from different layers can be adequately fused.
\par We learn node embeddings by maximizing node-level agreement between embeddings in different layers. By minimizing the node embedding’s loss, the coded embedding of each node after two different transformations is made consistent while being far away from the other nodes coded embedding in the feature space. $\*{u}_{i}^{k}$ and $\*{v}_{i}^{k}$ are the embeddings at the output of $k$th layer of the GNN encoder and their loss functions are shown in Eq. \ref{equ:5}.
\begin{small}
\begin{equation}
\label{equ:5}
\mathcal{L}\left( \*{z}_{{{u}_{i}}}^{k},\*{z}_{{{v}_{i}}}^{k} \right)=\log \frac{\exp \left( s(\*{z}_{{{u}_{i}}}^{k},\*{z}_{{{v}_{i}}}^{k})/\tau \right)}{\exp \left( s(\*{z}_{{{u}_{i}}}^{k},\*{z}_{{{v}_{i}}}^{k})/\tau \right)+\sum\limits_{j=1}^{N}{{{\mathbb{I}}_{[i\ne j]}}\exp \left( s(\*{z}_{{{u}_{i}}}^{k},\*{z}_{{{v}_{j}}}^{k})/\tau \right)}+\sum\limits_{j=1}^{N}{{{\mathbb{I}}_{[i\ne j]}}\exp \left( s(\*{z}_{{{u}_{i}}}^{k},\*{z}_{{{u}_{j}}}^{k})/\tau \right)}},
\end{equation}
\end{small}
\noindent where $s(\*{x},\*{y})={\*{(x)}}^{T}\*{y}$ calculates the similarity of two vectors, and $\tau $ is an adjustable factor. ${{\mathbb{I}}_{[i\ne j]}}$ represents the indicator function, which is 1 when $i$ and $j$ are not equal and 0 otherwise. The numerator of the loss function is the positive pair, which encourages similar vectors to be close together, and the denominator is the positive pair and negative pairs, which pushes all other vectors apart from the positive pair. The loss function of $k$th-layer is calculated by Eq. \ref{equ:8-1}.
\begin{equation}
\label{equ:8-1}
\mathcal{L}^{k}=\frac{1}{2N}\sum\limits_{i=1}^{N}{\left[ \mathcal{L}\left( \*{z}_{{{u}_{i}}}^{k},\*{z}_{{{v}_{i}}}^{k} \right)+\mathcal{L}\left( \*{z}_{{{v}_{i}}}^{k},\*{z}_{{{u}_{i}}}^{k} \right) \right]}.
\end{equation}
Adding the losses of different layers directly does not necessarily adapt to different datasets. We learn the weights of the embedding losses of different layers adaptively by attention mechanism. We firstly transform the embedding through a nonlinear transformation, and then use one shared attention vector ${{\mathbf{q}}^{k}}\in {{{R}}^{N\times 1}}$ to get the attention value $\omega _{{{u}_{i}}}^{k}$ as follows:
\begin{equation}
\label{equ:6}
\omega _{{{u}_{i}}}^{k}={{({{\mathbf{q}}^{k}})}^{T}}\cdot \tanh ({{\mathbf{W}}^{k}}\cdot {{(\mathbf{z}_{{{u}_{i}}}^{k})}^{T}}+\mathbf{b}),
\end{equation}
\noindent Here $\mathbf{W}^{k}\in {{R}^{N\times F}}$ is the weight matrix and $\mathbf{b}\in {{R}^{N\times 1}}$ is the bias vector. We then normalize the attention values $\omega _{{{u}_{i}}}^{k}$ with softmax function to get the final weight:
\begin{equation}
\label{equ:7}
\alpha _{{{u}_{i}}}^{k}=\text{soft}\max (\omega _{{{u}_{i}}}^{k})=\frac{\exp (\omega _{{{u}_{i}}}^{k})}{\sum\limits_{k=1}^{M}{\exp (\omega _{{{u}_{i}}}^{k})}},
\end{equation}
\noindent Larger $\alpha _{{{u}_{i}}}^{k}$ implies the corresponding embedding is more important. Finally, the loss of all positive sample pairs is calculated to obtain the overall loss as in Eq. \ref{equ:8-2}.
\begin{equation}
\label{equ:8-2}
\mathcal{{L}}_{total}={\sum\limits_{k=1}^{M}{\alpha _{{{u}_{i}}}^{k}\cdot \mathcal{{L}}^{k}}}.
\end{equation}
By minimizing $\mathcal{L}_{total}$, the effect of maximizing the lower bound on the MI between positive sample pairs can be achieved \cite{article1}. The previous graph contrastive learning model uses only the embeddings generated in the last layer, while the optimization of the middle hidden layers is performed only by back-propagating the gradients to the earlier layers. The optimization of the middle layers is difficult to converge to the optimization objective due to the absence of labels. We use the adaptive multi-layer contrastive loss to learn the feature representations in different layers, ensuring that the middle layers are also well optimized, thereby enhancing the model’s performance. Though our method capitalizes on multiple layers, they are all part of the same model, therefore, incur no additional computational overhead in reasoning.
\begin{algorithm}[htbp]\small
\begin{spacing}{1.3}
\caption{Main steps of the AMC-GNN algorithm.}
\label{alg:Framwork}
\begin{algorithmic}[1]
\For{$epoch\leftarrow 1,2,...,N$};
\label{code:fram}
\State Generate two graph view ${{\mathcal{G}}_{1}}$ and ${{\mathcal{G}}_{2}}$ by performing data augmentation on $\mathcal{G}$;
\State Obtain node embeddings ${\*{U}^{k}}\text{=}f\left( {{\mathcal{G}}_{1}} \right)$;
\State Obtain node embeddings ${\*{V}^{k}}\text{=}f({{\mathcal{G}}_{2}})$ ;
\State ${\*{U}^{k}}$ and ${\*{V}^{k}}$ are mapped into the contrast space to get $\*{Z}_{u}^{k}$ and $\*{Z}_{v}^{k}$;
\State Update parameters by applying gradient descent to maximize Eq. \ref{equ:6}.
\EndFor
\end{algorithmic}
\end{spacing}
\end{algorithm}
\section{Experiments and Analysis}
\label{sec:10}
\subsection{Dataset Description}
\label{sec:11}
We conducted experiments on eight widely used datasets to compared AMC-GNN with previous graph contrastive learning methods. Cora, Citeseer \cite{article7}, Pubmed \cite{article6} are widely used citation networks where each node represents an article and the edges indicate the citation relationships between articles. DBLP is a co-authorship multi-dimensional graph-based on publication records recorded in computer science literature websites \cite{article35}. We divide the training set, validation set, and test set as its original literature for the above datasets. Coauthor-CS and Coauthor-Physics \cite{article8} are two co-authorship graphs, where nodes are authors connected by their co-authorship. Amazon-Computers and Amazon-Photo \cite{article8} are two networks of co-buy relationships. For the above datasets, we randomly select 10\% of the nodes as the training set, 10\% as the validation set, and the remaining nodes as the test set. The details of each dataset are shown in Table \ref{tab:1}.
\begin{table}[h]\small
\centering
\caption{Statistics of datasets.}
\label{tab:1}
\begin{tabular}{ccccc}
\hline\noalign{\smallskip}
Dataset &Nodes &Edges &Features &Classes\\
\noalign{\smallskip}\hline\noalign{\smallskip}
Cora &2,708 &5,429 &1,433 &7\\
Citeseer &3,327 &4,732 &3,703 &6\\
Pubmed &19,717 &44,338 &500 &3\\
DBLP &17,716 &105,734 &1,639 &4\\
Coauthor-CS &18,333 &81,894 &6,805 &15\\
Coauthor-Physics &34,493 &247,962 &8,415 &5\\
Amazon-Computers &13,752 &245,861 &767 &10\\
Amazon-Photo &7,650 &119,081 &745 &8\\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
\subsection{Experiment Setups}
\label{sec:12}
\paragraph{Transductive Learning}~{}
\newline
In transductive learning tasks, the training set is $D=\left\{ \*{X},{\*{y}_{tra}} \right\}$ and the test sample ${\*{X}_{test}}$ also appears in the training set. We compared two typical transductive learning GNN models, GCN \cite{article15} and SCG \cite{article34}. In the transductive learning task, our encoder can be represented as:
\begin{equation}
\label{equ:9}
{\*{H}^{(l+1)}}=\sigma ({{\tilde{\*{D}}}^{-\frac{1}{2}}}\tilde{\*{A}}{{\tilde{\*{D}}}^{-\frac{1}{2}}}{\*{H}^{l}}{\*{W}^{l}}),
\end{equation}
\noindent where $\tilde{\*{A}}$ is the symmetric normalized adjacency matrix, $\sigma (\cdot )$ denotes the activation function, $\tilde{\*{D}}$ is the diagonal degree matrix ${{\tilde{D}}_{ii}}=\sum\nolimits_{j}{{{{\tilde{A}}}_{ij}}}$, and ${\*{W}^{l}}$ is the trainable weight matrix and is ${\*{H}^{l}}$ the hidden node representation in the $l$th layer. In SCG, The activation function is omitted.
\paragraph{Inductive Learning}~{}
\newline
In inductive learning tasks, the training set is $D=\left\{ {\*{X}_{tra}}, {\*{y}_{tra}} \right\}$ and the test set ${\*{X}_{test}}$ does not appear in the training set. Compared with transductive learning, inductive learning is more flexible. Inductive learning can easily get the representation of a new node by learning a method of node representation, instead of a fixed representation of a node.
Typical inductive learning GNN models are GraphSage \cite{article22} and GAT \cite{article21}, where the model consists of two phases: sampling and aggregation. In the sampling phase, a certain number of neighboring vertices are sampled for each vertex using the connectivity information. In the aggregation phase, the information of neighboring nodes is continuously merged by a multi-layer aggregation function. The merged information is used to predict the node labels. The propagation of the $k$th layer is represented as Eq. \ref{equ:10} and Eq. \ref{equ:11}.
\begin{equation}
\label{equ:10}
\*{h}_{\mathcal{N}(v)}^{k}\leftarrow AGGREGAT{{E}_{k}}\left( \left\{ \*{h}_{u}^{k-1},\forall u\in \mathcal{N}(v) \right\} \right),
\end{equation}
\begin{equation}
\label{equ:11}
\*{h}_{v}^{k}\leftarrow \sigma \left( {{\*{W}}^{k}}\cdot CONCAT(\*{h}_{v}^{k-1},\*{h}_{\mathcal{N}(v)}^{k}) \right),
\end{equation}
\noindent where $\*{h}_{\mathcal{N}(v)}^{k}$ is the embedding of the vertex $v$ at the $k$th layer with ${{\*{h}}^{0}}=\*{x}$, $\mathcal{N}(v)$ is a set of vertices adjacent to $v$.
\paragraph{Baseline}~{}
\newline
Following \cite{article32}, we conduct the node classification task to make the comparison with different methods. In the baseline model, we use traditional methods including: node2vec \cite{article14}, DeepWalk \cite{article19}; We also use Graph Autoencoders (GAE, VGAE) \cite{article33} and graph contrastive learning methods that currently reach state of the art: DGI \cite{article32}, GCA \cite{article2}, GraphCL \cite{article10} and GRACE \cite{article12}. We also compare our experimental results with some classical supervised learning models, including GCN \cite{article15}, SCG \cite{article34}, GraphSage \cite{article22}, GAT \cite{article21}.
\paragraph{Parameter Settings}~{}
\newline
We implement our method with Pytorch and PyTorch Geometric \cite{article16}. We train the model for a fixed number of epochs, 1500, 1000 epochs for Pubmed, DBLP respectively, and 200 epochs for the rest datasets. Adam optimizer are used on all datasets, weight decay of 1e-4. The probability control parameters ${{p}_{R,1}}$, ${{p}_{A,1}}$, ${{q}_{R,1}}$, ${{q}_{A,1}}$ for the first view ${{\mathcal{G}}_{1}}$ and ${{p}_{R,2}}$, ${{p}_{A,2}}$, ${{q}_{R,2}}$, ${{q}_{A,2}}$ for the second view ${{\mathcal{G}}_{2}}$, are all selected between 0 and 1. We conduct experiments on a computer server with four NVIDIA Tesla V100S GPUs (24GB memory each). All dataset-specific hyperparameters are summarized in Table \ref{tab:2}.
\begin{table}[h]
\centering
\caption{Hypeparameter specifications.}
\label{tab:2}
\resizebox{\textwidth}{!}{
\begin{tabular}{ccccccccccccc}
\hline\noalign{\smallskip}
Dataset & ${{p}_{R,1}}$ & ${{p}_{R,2}}$ & ${{q}_{R,1}}$ & ${{q}_{R,2}}$ & ${{p}_{A,1}}$ & ${{p}_{A,2}}$ & ${{q}_{R,1}}$ & ${{q}_{R,2}}$ &$\tau $ &\tabincell{c}{Learning\\rate} & \tabincell{c}{Hidden
\\dimension} & \tabincell{c}{Activation\\function}\\
\noalign{\smallskip}\hline\noalign{\smallskip}
Cora &0.3&0.4&0.2&0.7&0.3&0.3&0.1&0&0.9&5e-4&128 &ReLu \\
Citeseer &0.3&0.2&0.5&0.5&0.1&0.4&0.1&0.1&0.9&1e-3&256 &PReLu \\
Pubmed &0.3&0.2&0.6&1.0&0.3&0&0.1&0.1&0.5&1e-3&256 &Relu \\
DBLP &0.3&0.2&0.3&0.3&0.2&0.3&0.2&0.2&0.45&1e-3&256 &Relu \\
Coauthor-CS &0.3&0.3&0.6&0.3&0.3&0.2&0.1&0&0.8&5e-4&128 &Relu\\
Coauthor-Physics&0.3&0.3&0.2&0.7&0.3&0.2&0.1&0&0.4&1e-3&256&Relu \\
Amazon-Computers&0.3&0.5&0.8&0.2&0.2&0.2&0.1&0&0.7&1e-3&256&Relu \\
Amazon-Photo&0.3&0.2&0.5&0.8&0.2&0.2&0.1&1&0.9&1e-3&256&Relu\\
\noalign{\smallskip}\hline
\end{tabular}}
\end{table}
\paragraph{Results}~{}
\newline
The experimental results are summarized in Table \ref{tab:3}, Training Data denotes the data required for method training. $\*{X}$, $\*{A}$, $\*{Y}$ represent the feature matrix, adjacency matrix and sample labels of the graph, respectively. We use the average after five runs with different random seeds as the results.
\begin{table}[h]
\centering
\caption{Table of experimental results. For transductive tasks and inductive tasks, we use accuracy in percentage and micro-averaged F1-score as results respectively. For clarity, the best performance of AMC-GNN is highlighted in boldface, and the best representations of other unsupervised and supervised graph models are underlined.}
\label{tab:3}
\resizebox{\textwidth}{!}{
\begin{tabular}{cccccccccc}
\hline\noalign{\smallskip}
Method & \tabincell{c}{Training\\Data} & Cora & Citeseer & Pubmed & DBLP &\tabincell{c}{Coauthor\\-CS} &\tabincell{c}{Coauthor\\-Physics} & \tabincell{c}{Amazon\\-Computers} & \tabincell{c}{Amazon\\-Photo}\\
\noalign{\smallskip}\hline\noalign{\smallskip}
Raw features &$X$ &64.8 &64.6 &84.8 &71.6 &90.4 &93.6 &73.8 &78.5\\
node2vec &$A$ &74.8 &52.3 &80.3 &78.8 &85.1 &91.2 &84.4 &89.7\\
DeepWalk &$A$ &75.7 &50.5 &80.5 &75.9 &84.6 &91.8 &85.7 &89.4\\
DeepWalk + features &$X,A$ &73.1 &47.6 &83.7 &78.1 &87.7 &94.9 &86.3 &90.1\\
GAE &$X,A$ &76.9 &60.6 &82.9 &81.2 &91.6 &94.9 &85.3 &91.6\\
VGAE &$X,A$ &78.9 &61.2 &83.0 &81.7 &92.1 &94.5 &86.4 &92.2\\
DGI &$X,A$ &82.6$\pm$0.4 &68.8$\pm$0.7 &86.0$\pm$0.1 &83.2$\pm$0.1 &\underline{91.4$\pm$0.1} &94.5$\pm$0.5 &84.0$\pm$0.3 &91.6$\pm$0.2\\
GCA &$X,A$ &83.3$\pm$0.3 &69.3$\pm$0.4 &85.7$\pm$0.1 &83.6$\pm$0.2 &91.1$\pm$0.1 &93.7$\pm$0.3 &84.3$\pm$0.2 &90.6$\pm$0.2\\
GraphCL &$X,A$ &\underline{83.5$\pm$0.3} &71.2$\pm$0.5 &84.6$\pm$0.1 &\underline{84.5$\pm$0.1} &91.1$\pm$0.1 &93.2$\pm$0.3 &85.3$\pm$0.2 &90.8$\pm$0.2\\
GRACE &$X,A$ &83.3$\pm$0.4 &\underline{72.1$\pm$0.5} &\underline{86.7}$\pm$0.1 &84.2$\pm$0.1 &89.8$\pm$0.3 &\underline{95.6$\pm$0.2} &\underline{87.3$\pm$0.2} &\underline{92.1$\pm$0.2}\\
\noalign{\smallskip}\hline\noalign{\smallskip}
AMC-GCN &$X,A$ &\textbf{84.8$\pm$0.4} &\textbf{72.8$\pm$0.5} &\textbf{87.1$\pm$0.1} &\textbf{84.9$\pm$0.2} &\textbf{92.4$\pm$0.2} &95.7$\pm$0.1 &88.9$\pm$0.2 &92.8$\pm$0.1\\
AMC-GAT &$X,A$ &84.6$\pm$0.3 &72.3$\pm$0.3 &86.9$\pm$0.2 &84.3$\pm$0.1 &91.5$\pm$0.1 &\textbf{95.9$\pm$0.1} &\textbf{89.5$\pm$0.2} &\textbf{93.1$\pm$0.2}\\
\noalign{\smallskip}\hline\noalign{\smallskip}
GCN &$X,A,Y$ &82.8 &72.0 &\underline{84.9} &82.7 &\underline{93.0} &\underline{95.7} &86.5 &92.4\\
SGC &$X,A,Y$ &80.6 &69.1 &84.8 &81.7 &92.1 &95.1 &83.9 &90.9\\
GAT &$X,A,Y$ &\underline{83.7} &\underline{72.5} &79.3 &\underline{83.7} &92.3 &95.5 &\underline{86.9} &\underline{92.6}\\
\noalign{\smallskip}\hline
\end{tabular}}
\end{table}
AMC-GCN and AMC-GAT represent AMC-GNN models with GCN and GAT as encoders, respectively. AMC-GCN, GCN, SGC are transductive learning methods. AMC-GAT, GAT are inductive learning methods. The proposed model AMC-GNN outperforms the state-of-the-art unsupervised models DGI , GraphCL, GCA, and GRACE on all the datasets used in the experiments, and the performance on some datasets even exceeds that of supervised learning methods. On the Cora dataset, AMC-GCN outperformed GraphCL by 1.3\% and outperformed DGI by 2.2\% under 200 epochs of training. On the Amazon-Computers dataset, AMC-GAT outperformed GRACE by 2.2\% and outperformed GCA by 5.2\% under 200 epochs of training. We visualized the features after model encoding, using different colored points to represent different classes of samples. We performed the analysis on the Cora dataset because it has the least number of nodes for clear presentation. T-SNE plots of the embeddings is given in Fig. \ref{fig:3}.
\begin{figure}[h]
\centering
\subfloat[Raw]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.5in]{cora_RAW_30.png}
\end{minipage}%
}%
\subfloat[DGI]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.5in]{cora_DGI_30.png}
\end{minipage}%
}%
\subfloat [GRACE]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.5in]{cora_GCN_30.png}
\end{minipage}%
}%
\subfloat [AMC-GCN]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.5in]{cora_MCG_30.png}
\end{minipage}%
}%
\caption{The t-SNE visualization of learned node representations on Cora. (a) Raw means the raw node features are used. (b) DGI indicates the features are learned by DGI. (c) GRACE indicates the features are learned by GRACE. (d) AMC-GCN indicates the features are learned by AMC-GCN. The Silhouette score for (a) (b) (c) (d) respectively is 0.005, 0.207, 0.156 and 0.243.}
\label{fig:3}
\centering
\end{figure}
\subsection{Ablation Studies}
\label{sec:13}
The experimental results in Table \ref{tab:3} demonstrate the effectiveness of adaptive multi-layer contrastive loss and the auxiliary training model in improving the model's performance. In this section, we will explore the implications of the adaptive multi-layer contrastive loss and auxiliary training model.
\par w/o auxiliary model means using the model without auxiliary model. w/o multi-layer represents the model without adaptive multi-layer contrastive loss function. The performance of the above models was tested, and the models used the same parameters as in Sect. \ref{sec:12}. We run five times and average the results for all methods, and the experimental results are shown in Table \ref{tab:4}.
To reflect the change of feature space, we selected the well known K-means \cite{article40} to cluster the vectors generated by the model, and calculate clustering evaluation indexes such as: Calinski-Harabaz Index(CHI) \cite{article39}, Davies-Bouldin Index(DBI) \cite{article23}, Silhouette Coefficient(SC) \cite{article3}. For CHI and SC, larger values indicate better clustering of features; for DBI, smaller values indicate better clustering of features. A high-performance coding model produces features that are closer together within classes, while features between classes are further apart and more clearly bounded.
\begin{table}[h]\scriptsize
\centering
\caption{The performance of AMC-GCN, w/o auxiliary model and w/o multi-layer in transductive node classification on four citation datasets. The best clustering results are highlighted in boldface.}
\label{tab:4}
\begin{tabular}{c|c|cccc}
\hline\noalign{\smallskip}
Dataset &Model &CHI &DBI &SC &Accuracy \\
\noalign{\smallskip}\hline
\multirow{3}*{Cora} &full method &\textbf{204.094} &\textbf{2.117} &\textbf{0.088}
&\textbf{84.8$\pm$0.4}
\\
&w/o auxiliary model &174.778 &2.287 &0.085 &84.1$\pm$0.4\\
&w/o multi-layer &128.494 &2.835 &0.033 &83.6$\pm$0.3\\ \hline
\multirow{3}*{Citeseer} &full method &\textbf{163.685} &\textbf{2.753} &\textbf{0.018} &\textbf{72.8$\pm$0.5}\\
&w/o auxiliary model &174.778 &2.287 &0.085 &72.5$\pm$0.5\\
&w/o multi-layer &128.494 &2.835 &0.033 &72.1$\pm$0.5\\ \hline
\multirow{3}*{Pubmed} &full method &\textbf{3152.101} &\textbf{2.064} &\textbf{0.070} &\textbf{87.1$\pm$0.1}\\
&w/o auxiliary model &1544.902 &3.008 &0.026 &86.8$\pm$0.1\\
&w/o multi-layer &1361.766 &2.664 &0.034 &86.5$\pm$0.2\\ \hline
\multirow{3}*{DBLP} &Full Method &\textbf{959.503} &\textbf{3.161} &\textbf{0.011}
&\textbf{84.9$\pm$0.2}\\
&w/o auxiliary model &586.525 &4.262 &-0.012 &84.5$\pm$0.2\\
&w/o multi-layer &425.860 &5.020 &-0.021 &84.2$\pm$0.3\\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
The experimental results on four data sets prove that introducing the auxiliary model and the adaptive multi-layer contrastive loss can make the embeddings have a better clustering effect. Compared to the previous graph contrastive learning model, in the feature space, the features of the same category are closer to each other, while the features of different categories are farther away. By learning better embeddings, the performance of the model improved. The clustering effect proves that the introduction of adaptive multi-layer contrastive loss is a crucial reason for improved performance.
\subsection{The analysis of Data Augmentation}
\label{sec:14}
Data augmentation is essential in graph contrastive learning, and in our experiments, we use a combination of different data augmentation approaches to generate different graph views. In this section, we will compare different data augmentation methods and analyze the role of data augmentation.
AMC-GCN(RN) and AMC-GCN(AM) indicate the model with Random Sampling Neighbours and Attribute Masking only respectively. The model parameters are set as Sect. \ref{sec:12}, and the performance of the model for node classification on four datasets: Cora, Citeseer, Pubmed and DBLP, is shown in Table \ref{tab:5}. Obviously, the performance of utilizing both RN and AM data augmentation methods is better than using only a single method, and the results demonstrate that data augmentation requires changes in both graph topology and node features to have better results.
\begin{table}[h]\small
\centering
\caption{The performance of model variants along with the original AMC-GCN model, AMC-GCN(RN) and AMC-GCN(AM).}
\label{tab:5}
\begin{tabular}{ccccc}
\hline\noalign{\smallskip}
Method &Cora &Citeseer &Pubmed &DBLP \\
\noalign{\smallskip}\hline\noalign{\smallskip}
AMC-GCN &\textbf{84.8$\pm$0.4} &\textbf{72.8$\pm$0.5} &\textbf{87.1$\pm$0.1} &\textbf{84.9$\pm$0.2}\\
AMC-GCN(RN) &84.5$\pm$0.2 &68.8$\pm$0.4 &85.4$\pm$0.2 &83.8$\pm$0.4\\
AMC-GCN(AM) &84.6$\pm$0.2 &70.7$\pm$0.3 &86.1$\pm$0.2 &84.4$\pm$0.2\\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
\subsection{Analysis of Attention Mechanism}
\label{sec:16}
We analyzed the attention distribution and the attention learning trend separately to explore whether the attention values learned by the attention mechanism are meaningful.
\paragraph{Analysis of attention distributions}~{}
\newline
AMC-GNN learns the importance weights of embedding in different layers through the attention mechanism. The training data set and parameter settings are the same as in Sect. \ref{sec:12}. We conduct the attention distribution analysis on four datasets: Cora, Citeseer, Pubmed, and DBLP by using AMC-GCN, where the results are shown in Fig. \ref{fig:4}. As we can see, for Cora, Citeseer, the attention values of explicit embeddings in the second layer are larger than other layers’. This suggests that the data in embeddings of the second layer ought to be more critical than other layers in feature space. For Pubmed, embeddings of first two layers contain more information. For DBLP, embeddings of the last layer are more useful for optimization.
\begin{figure}[h]
\centering
\subfloat[Cora]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.5in]{cora_box.pdf}
\end{minipage}%
}%
\subfloat[Citeseer]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.5in]{citeseer_box.pdf}
\end{minipage}%
}
\subfloat [Pubmed]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.5in]{pubmed_box.pdf}
\end{minipage}%
}%
\subfloat [DBLP]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.5in]{DBLP_box.pdf}
\end{minipage}%
}%
\caption{Analysis of attention distribution.}
\label{fig:4}
\centering
\end{figure}
\begin{figure}[h]
\centering
\subfloat[Cora]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.5in]{cora_attention_trend.pdf}
\end{minipage}%
}%
\subfloat[Citeseer]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.5in]{citeseer_attention_trend.pdf}
\end{minipage}%
}
\subfloat[Pubmed]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.5in]{pubmed_attention_trend.pdf}
\end{minipage}%
}%
\subfloat[DBLP]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.5in]{dblp_attention_trend.pdf}
\end{minipage}%
}%
\caption{The attention changing trends w.r.t epochs.}
\label{fig:5}
\centering
\end{figure}
\paragraph{Analysis of attention trends}~{}
\newline
We dissect the changing patterns of attention value during the preparation cycle. Here we accept Cora and Citeseer as models in Fig. \ref{fig:5}, where x-axis denotes the epoch and y-axis denotes the average attention value. Toward the start, the average attention values of the embeddings of all layers are practically similar, with the training epoch increasing, the attention values become unique. The attention value for embeddings of the second layer gradually increases, while the attention value for other embeddings keeps decreasing. This marvel is reliable with the conclusions in Fig. \ref{fig:4}, and we can see that AMC-GCN can learn the significance of embeddings in different layers gradually.
\subsection{Robustness to Sparse Features}
\label{sec:15}
When a small amount of features are removed, the impact on the model is minor. However, when the node features are too sparse, the nodes may not have enough features to maintain their original semantics and can cause a decrease in the accuracy of the model.
\par In this section, we use AM to randomly pollute the training data and explore the robustness of the model to sparse features. Specifically, we conduct experiments on four cited network datasets, ranging the masking ratio of node features from 0.4 to 0.9. During training, all other parameters of the models are set as Sect. \ref{sec:12}.
\begin{figure}[h]
\centering
\subfloat[Cora]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.5in]{cora_mask.pdf}
\end{minipage}%
}%
\subfloat[Citeseer]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.5in]{citeseer_mask.pdf}
\end{minipage}%
}
\subfloat [Amazon-Photo]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.5in]{photo_mask.pdf}
\end{minipage}%
}%
\subfloat [Amazon-Computer]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.5in]{computer_mask.pdf}
\end{minipage}%
}
\caption{The performance of DGI, GRACE, AMC-GCN and AMC-GAT in transductive node classification on four datasets with different features masking ratio.}
\label{fig:6}
\centering
\end{figure}
The results on the four datasets are shown in Fig. \ref{fig:6}. With different pollution rates, AMC-GCN and AMC-GAT outperform GRACE and DGI in most cases, proving that our proposed AMC-GNN model has more robustness against the dropout of features. We attribute the robustness of AMC-GNN to the superiority of multi-layer comparison because the outputs of different layers are considered simultaneously, avoiding the cumulative propagation of errors caused by dropped features in the network. As the proportion of dropped features increases, the performance of the model decreases. This is because feature dropout has changed the semantic labels of the nodes, and the excessive dropout of node features prevents GNN from extracting meaningful information from the nodes.
\subsection{Representation Stability Visualization}
\label{sec:17}
If the model is stable, then similar feature representations will be learned for nodes obtained by different data augmentation methods. We arbitrarily choose a node for 10 different dropping features with $p=0.4$ and get 10 different feature vectors ${\*{h}_{i}}$ after the trained target model, where $1\le i\le 10$.
Calculate the similarity matrix of the two sets of vectors $\*{S}\in {{\mathbb{R}}^{10\times 10}}$, where ${\*{S}_{ij}}=\frac{{\*{h}_{i}}\cdot {\*{h}_{j}}}{\left| {\*{h}_{i}} \right|\left| {\*{h}_{j}} \right|}$. The similarity matrix is visualized and the result are shown in Fig. \ref{fig:7}.
\begin{figure}[h]
\centering
\subfloat[DGI]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.4in]{5-DGI.pdf}
\end{minipage}%
}%
\subfloat[GRACE]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.4in]{5-GRACE.pdf}
\end{minipage}%
}%
\subfloat [AMC-GCN]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.4in]{5-MCG_GCN.pdf}
\end{minipage}%
}%
\subfloat [AMC-GAT]{
\begin{minipage}[h]{0.24\linewidth}
\centering
\includegraphics[width=1.4in]{5-MCG_GAT.pdf}
\end{minipage}%
}%
\caption{The representation stability visualization on Cora. (a)(b)(c)(d) are performance of DGI, GRACE, AMC-GCN and AMC-GAT respectively. Deeper color indicates higher the similarity of the node representation.}
\label{fig:7}
\centering
\end{figure}
We randomly selected a test node to calculate the $\*{S}$ matrices for the four methods DGI, GRACE, AMC-GCN and AMC-GAT and visualize them in Fig. \ref{fig:7}. The similarity matrices of different models are visualized using darker colors to indicate higher similarity. Under the condition of slight disturbance, higher similarity of node features indicates better model stability and better extraction of essential features of nodes. The average similarity of DGI, GRACE, AMC-GCN, AMC-GAT are 0.503, 0.838, 0.849 and 0.859.
\section{Conclusion}
\label{conclusion}
In this paper, we propose a new contrastive GNN called AMC-GNN. AMC-GNN is a novel generic framework that could provide a new perspective on the structure of GNNs. AMC-GNN learns more essential features of different classes of data by introducing auxiliary training models and adding adaptive Multi-layer contrastive losses. AMC-GNN uses the attention mechanism to learn the importance weights of the embeddings in different layers adaptively. We conducted comprehensive experiments on various widely used datasets. The experimental results show that our proposed method can learn more robust and essential features of graphs, outperforming existing state-of-the-art unsupervised graph contrastive learning methods.
\bibliographystyle{spmpsci}
|
\section{Introduction}\label{S:introduction}
The grand challenges of ensuring resilience and security in cyber-physical systems (CPS) have motivated the study and characterization of possible adversarial attacks against these complex systems. Reactive approaches based on the detection and identification algorithms are a significant aspect of comprehensive defense strategies against malicious attacks \cite{cardenas2008secure}.
Due to their distributed nature, cyber-physical systems such as the power grid or networks of autonomous aerial/ground vehicles can often be modeled as multi-agent systems \cite{pasqualetti2013attack,ren2007distributed}, where the communication network is susceptible to attacks \cite{cardenas2008secure}.
In particular, this paper considers the problem of detecting stealthy attacks, namely covert attack and zero-dynamics attack, using a scalable detection framework for a class of networked multi-agent systems seeking average consensus upon system's initial conditions, as a canonical cooperative task.
\emph{Literature review:}
In general, the detection of stealthy attacks is not a trivial problem for networked multi-agent systems. Challenges arise due to the large scale of networked systems and the limited communication capability of its subsystems (or agents), which restrict an effective information aggregation and transmission required to implement centralized approaches \cite{pasqualetti2015divide}. Moreover, the prevalent observer-based attack detectors are ineffective in detecting stealthy attacks, particularly zero-dynamics attack (ZDA) and covert attack that are the worst-case attack scenarios in terms of detectability, due to the fact that they are not observable in the system outputs \cite{pasqualetti2013attack,teixeira2012revealing}.
The conventional detection frameworks for stealthy attacks rely on modifying the system structure or adding redundancy in the system measurements to expose such attacks. For instance, a signal modulation acting on the system actuation to alter the system's input behavior was proposed in \cite{hoehn2016detection} for both covert attack and ZDA detection. Change in the system structure was proposed first in \cite{teixeira2012revealing} upon which the study in \cite{schellenberger2017detection} extended the system dynamics with a randomly switched auxiliary system to achieve non-repeating dynamics, preventing the realization of covert attacks. Most recently, for a class of networks with distinct Laplacian eigenvalues, the authors in \cite{mao2020novel} characterized an intermittent ZDA that remains undetectable regardless of the system's switched structure and obtained the conditions for their detectability. As for the covert attack, the authors in \cite{barboni2020detection} proposed a distributed architecture composed of two cascaded observers for each subsystem to detect the attacks. As another strategy, multi-rate sampling in sampled-data systems was studied in \cite{jafarnejadsani2018multirate,back2017enhancement} to change the direction of sampling zeros and thus to prevent ZDA. Also, distributed function calculation was proposed in \cite{sundaram2010distributed} that requires intensive communication in the network and full knowledge of network model for each node.
In terms of scalability, considerable effort has been dedicated to extending the existing decentralized and distributed estimation/fault detection methods to the attack detection strategies implementable using locally available information for large-scale systems. For instance, one can refer to secure distributed observers for sensor networks in \cite{mitra2016secure}, distributed attack detection schemes for power networks \cite{pasqualetti2013attack,barboni2020detection,teixeira2010networked,gallo2020distributed}, decentralized detection scheme for stochastic interconnected systems in \cite{anguluri2018attack}, and divide-and-conquer approach in \cite{pasqualetti2015divide}. However, few studies have addressed distributed/decentralized detection strategies for \emph{stealthy} attacks, namely covert attack and zero-dynamics attack \cite{barboni2020detection,gallo2020distributed}. Moreover, they do not address the communication topology switching and the privacy of the agents' information.
\emph{Statement of contributions:}
The contributions of this paper are threefold.
First, as a security objective, we consider the privacy of agents' initial condition and the agreement's final value (consensus) and propose enforced unobservability constraints on the network topology to preserve the network privacy at the global level.
Second, for scalability, we propose a glocal (global-local) attack detection structure for which the networked multi-agent system is partitioned into clusters (subsystems) with their respective globally and locally monitored agents that satisfy specific conditions related to the network privacy and the detectability of stealthy attacks (i.e., zero-dynamics attack and covert attack).
Finally, we derive the theoretical conditions for topology switching (Theorem \ref{th:Attack_detectability_switching}) under which local detectors trigger switches in the system's communication topology such that stealthy attacks become detectable for the global (centralized) observer. We further discuss different types of topology switching and their outcome for the detection of stealthy attacks.
The rest of the paper is organized as the following. Section \ref{S:problem_formulation} presents the preliminary definitions and the problem formulation. The privacy preserving problem and the attack detection framework are studied in Section \ref{S:privacy_obs}.
Section \ref{S:Results} demonstrates the simulation results. Finally, Section \ref{S:conclusion} concludes the paper.
\section{Problem Formulation}\label{S:problem_formulation}
\subsection{Preliminaries}
\noindent
\textbf{{Notation}.} We use $ \mathbb{R} $, $ \mathbb{R}_{> 0} $, $ \mathbb{R}_{\geq 0} $, $\ensuremath{\mathbb{C}}$, and $\ensuremath{\mathbb{N}}$ to denote the set of reals, positive reals, non-negative reals, complex, and natural numbers, respectively. Also, $\ensuremath{\mathbb{N}}_{0} = \ensuremath{\mathbb{N}} + \{0\}$.
We use $ x \coloneqq \mathrm{col}(x_1,x_2,\ldots,x_n) $ to denote (block-partitioned) vectors.
$\boldsymbol{1}_n$, $ \boldsymbol{0}_n$, $ I_n $ and $ 0_{n} $ stand for $n$-vector of all ones, the $n$-vector of all zeros, the identity $ n $-by-$ n $ matrix, and $ n $-by-$ n $ zero matrix, respectively\footnote{We may omit the subscripts when clear from the
context.}. $x^{(\rm m)}(t)$ stands for the
$\rm m$-th order time derivative of $x(t)$. In addition, $ |\cdot| $ denotes the cardinality of sets, and for any index set $\mathcal{F}$ with $|{\mathcal{F}}|=m$, $I_{\scriptscriptstyle {\mathcal{F}}} \in \ensuremath{\mathbb{R}}^{n \times m}$ is the concatenation of the $i$-th columns of $I_{n}$ where $i \in {\mathcal{F}}$.
For a matrix $M \in \ensuremath{\mathbb{R}}^{m \times n}$,
the range (column space) is defined as $\mathrm{Im}{(M)} = \braces{Mx \mid x \in \ensuremath{\mathbb{R}}^n} \subseteq \ensuremath{\mathbb{R}}^m$ and the nullspace is defined as $\mathrm{ker}{(M)} = \braces{x \mid Mx = \boldsymbol{0}} \subseteq \ensuremath{\mathbb{R}}^n$.
The support of vector $x \in \mathbb{R}^n$ is the set of nonzero components defined as $\ensuremath{\mathrm{supp}}(x)=\setdef{i\in\braces{1,\dots,n}}{x_i\neq0}$.
We also define the set of nonzero columns of the $ n $-by-$ n $ matrix $M$ by $\ensuremath{\mathrm{colsupp}}(M)=\setdef{i\in\braces{1,\dots,n}}{[M]_{:,i}\neq\boldsymbol{0}_{n}}$.
\noindent
\textbf{{Graph theory}.} Let $ \mathcal{G}=(\mathcal{V},\mathcal{E},\mathcal{A})$ denote a weighted undirected graph with the set of nodes $ \mathcal{V}= \{1,2,\dots,N\} $, set of edges $ \mathcal{E} \subseteq \mathcal{V} \times \mathcal{V}$, and adjacency matrix $\mathcal{A} \coloneqq [a_{ij}]\in \mathbb{R}^{\scriptscriptstyle N}_{\geq0}$.
For any pair of nodes $ i,j,\, i \neq j$, a path from $j$ to $i$ implies the edge $ (i,j) \in \mathcal{E}$ corresponding to $a_{ij}>0$, otherwise $a_{ij}=0$. The Laplacian matrix $\mathcal{L} \coloneqq [l_{ij}] \in
\ensuremath{\mathbb{R}}^{n \times n}$ is defined as $ l_{ii} = \sum_{j \neq i}^{}a_{ij}$ and $ l_{ij} = -a_{ij}$ if $i \neq j$.
By convention, $ \mathcal{N}_i=\{j\in \mathcal{V} \mid \mathcal{E}_{ij}\in \mathcal{E} \} $ denotes the set of neighbors of node $ i $. A cluster is defined as any subset $ \boldsymbol{\mathcal{P}}:=\{\mathcal{P}_1,\dots,\mathcal{P}_{\scriptscriptstyle |\boldsymbol{\mathcal{P}}|}\}\subseteq \mathcal{V} $ of the nodes of graph $ \mathcal{G} $ such that $ \cup_{i=1}^{|\boldsymbol{\mathcal{P}}|}\mathcal{P}_i=\mathcal{V} $ and $ \mathcal{P}_i\cap \mathcal{P}_j=\emptyset $ if $i\neq j$.
We make the convention that $ \mathcal{G}_{\sigma(t)} $ with a right-continuous switching signal $\sigma(t): \mathbb{R}_{\geq0}\rightarrow{\mathcal{Q}:=\{1,2,\dots,q\}}, \, q\coloneqq|\mathcal{Q}| $ denotes a finite set of graphs, indexed by finite set $\mathcal{Q}$, that each holds all properties of graph $\mathcal{G}$.
\begin{definition}\btitle{Graph component \cite{newman2018networks}}\label{def:graph_component}
A component in an undirected graph is an induced subgraph with a (maximal) subset of nodes such that each is reachable by some path from each of the others.
\end{definition}
\noindent
\textbf{{Systems theory.}} A linear system $ \dot{x}(t)=Ax(t)+Bu(t)$, $ y(t)=Cx(t)+Du(t) $, where $x(t) \in \ensuremath{\mathbb{R}}^n, u(t) \in \ensuremath{\mathbb{R}}^m, y(t)\in \ensuremath{\mathbb{R}}^p$, is represented by the tuple $\Sigma(A,B,C,D)$.
\begin{definition}[Zeroing direction and zero-dynamics attack {\cite[Ch. 3]{zhou1996robust}},\cite{mao2020novel}]\label{def:inv_zeros}
Scalar $ \lambda_0\in \mathbb{C} $ is a zero of the tuple $\Sigma(A,B,C,D)$ if, and only if, there exists zeroing direction $\mathrm{col}(\mb{x}_0,\mb{u}_0)\neq \mathrm{col}(\boldsymbol{0},\boldsymbol{0})$ associated with $ \lambda_0 $ such that
\begin{equation}\label{eq:inv_zeros}
\begin{bmatrix}
\lambda_0 I_n - A & -B \\ C & D
\end{bmatrix}
%
\begin{bmatrix}
\mb{x}_0 \\ \mb{u}_0
\end{bmatrix}
%
=
%
\begin{bmatrix}
\boldsymbol{0} \\ \boldsymbol{0}
\end{bmatrix}.
\end{equation}
Then, the signal $ u(t)=\mb{u}_0e^{\lambda_0t}$ is a zero-dynamics attack that generates non-zero state trajectories $x(t)=\mb{x}_0e^{\lambda_0t}$ while the output $ y =Cx+Du$ satisfies $y(t) =\boldsymbol{0}$.
\end{definition}
\subsection{Problem Statement}
\noindent
\textbf{{System model.}} Consider a graph $\mathcal{G}$ of order $ N $, we associate each node $i$ of the graph with an agent $ \Sigma_i $ that evolves according to the following dynamics\footnote{For brevity, we may omit the time
argument $t$ from expressions whenever possible in the rest of the paper.}:
\begin{equation}\label{eq:ol_sys}
\Sigma_{i}:
\left\{
\begin{array}{l}
\dot{x}_i(t) = v_i(t)\\
\dot{v}_i(t) = u_i(t)
\end{array},
\right. \qquad i\in \mathcal{V}, \\
\end{equation}
in which $ x_i(t) $ and $ v_i(t) $ denote the position and velocity, and $u_i(t) $ (to be determined) stands for the control channel through which each agent communicates with a set of neighbors $\mathcal{N}_i $ to perform a prespecified cooperative task.
\noindent
\textbf{{Control protocol}.} The objective is to reach an average consensus upon the initial conditions of the system, as follows:
\begin{equation}\label{eq:cond_consensus}
\lim \limits_{t\rightarrow{\infty}} \left|x_i(t)-x_j(t) \right| =0
\; \text{and} \;
\lim \limits_{t\rightarrow{\infty}} \left|v_i(t) \right| =0, \quad \forall \; i,j \in \mathcal{V},
\end{equation}
which can be achieved by exchanging local information through the following switching control protocol \cite{mao2020novel}:
\begin{align}\label{eq:ctrl_proto}
u_i &= - \gamma v_i -\alpha \sum_{j\in \mathcal{N}_i} a^{\sigma(t)}_{ij}(x_i-x_j) + u_{a_i}, \quad i \in \mathcal{V},
\end{align}
where $ a^{\sigma(t)}_{ij} $ is the entry of the symmetric adjacency matrix associated with the graph $ \mathcal{G}_{\sigma(t)} $
representing the switching communication network of agents $ \Sigma_i $'s. Also, $\alpha$ and $\gamma$ are the control gains. Finally, $u_{a_i} $ is the injected malicious signal in control channel of the $i$-th agent. We assume the unknown subset $\overline{\mathcal{F}} \subset \mathcal{V}$ represents the set of compromised agents, and we have $u_{a_i} =0 $ for an uncompromised agent $i$, i.e., if $ i \in \mathcal{V} \setminus \overline{\mathcal{F}} $.
\noindent
\textbf{{Closed-loop system.}} Given \eqref{eq:ol_sys} and \eqref{eq:cond_consensus}, let $ \mb{x} \coloneqq \mathrm{col}(x,v)$, where $x \coloneqq \mathrm{col}(x_1,\dots,x_{\scriptscriptstyle N})$, $v \coloneqq \mathrm{col}(v_1,\dots,v_{\scriptscriptstyle N})$, and $\mb{u}_a=\mathrm{col}(u_{a_i}),\, i \in \overline{\mathcal{F}}$. Then, the closed-loop system is given by
\begin{equation}\label{eq:cl_sys}
\Sigma:
\left\{
\begin{array}{l}
\underbrace{
\begin{bmatrix}
\dot{x} \\ \dot{v}
\end{bmatrix}
}_{\dot{\mb{x}}}
=
\underbrace{
\begin{bmatrix}
0_{} & I_{} \\
-\alpha \mathcal{L}_{\sigma(t)} & -\gamma I_{}
\end{bmatrix}
}_{\mb{A}_{\sigma(t)}}
\underbrace{
\begin{bmatrix}
x \\ v
\end{bmatrix}
}_{\mb{x}}
+ \underbrace{
\begin{bmatrix}
0_{\scriptscriptstyle } \\ I_{\scriptscriptstyle \overline{\mathcal{F}}}
\end{bmatrix}
}_{\mb{B}}\mb{u}_a, \\
\mb{x}(t_0)=\mb{x}_0,\\
\mb{y} ={\mb{C}}{\mb{x}}-\mb{u}_s, \ \ \mb{C}=\mathrm{diag}\braces{C_{\rm x}, C_{\rm v}},
\end{array}
\right.
\end{equation}
with the system measurements $\mb{y}=\mathrm{col}(\mb{y}_1,\cdots,\mb{y}_{\scriptscriptstyle |\mathcal{M}|})$ corresponding to the output matrix $\mb{C}$ such that:
\begin{equation}\label{eq:measurments}
\ensuremath{\mathrm{colsupp}}({C_{\rm k}}) \in \mathcal{M}_{\rm k} \subset \mathcal{V}, \ \ {\rm k} \in \braces{{\rm x},{\rm v}}, \ \ \mathcal{M} = \braces{\mathcal{M}_{\rm x}, \mathcal{M}_{\rm v}},
%
\end{equation}
where the to-be-selected set $\mathcal{M}$ represents the set of the monitored agents' index.
Also, $\mb{u}_s=\mathrm{col}(u_{s_1},\dots,u_{s_{\scriptscriptstyle|\mathcal{M}|}})$ is a vector of injected malicious signals in the compromised measurement sensor channels. Finally, the Laplacian $ \mathcal{L}_{\sigma(t)} $ in \eqref{eq:cl_sys} encodes the information exchange among agents.
\noindent
\textbf{{Adversary model.}}
Let $ \overline{\mathcal{F}} \subset \mathcal{V} $ denote the set of agents with a compromised (under attack) control channel, and
$ \underline{\mathcal{F}} \subset \mathcal{M} $ represent the set of agents with compromised sensor channels. The dynamics of the adversarial attack is given by\footnote{The matrix $\mb{B}$ in \eqref{eq:attack_model} is the same as in \eqref{eq:cl_sys}. it is designed by the attacker.}
\begin{equation}\label{eq:attack_model}
\Sigma_{\scriptscriptstyle \mathcal{A}}:
\left\{
\begin{array}{l}
\dot{\tilde{\mb{x}}}=\tilde{\mb{A}}_{{\sigma}(t)}\tilde{\mb{x}}
+{\mb{B}}\mb{u}_a(t),
\quad \tilde{\mb{x}}(t_a)=\tilde{\mb{x}}_0, \\
\mb{u}_s =\tilde{\mb{C}}_{\scriptscriptstyle}\tilde{\mb{x}},\\
\ensuremath{\mathrm{supp}}(\mb{u}_a)=\overline{\mathcal{F}}, \quad \ensuremath{\mathrm{supp}}(\mb{u}_s)=\underline{\mathcal{F}},
\end{array}
\right.
\end{equation}
where the vector attack $\mb{u}_a$ is generally a function of disclosed information, i.e., $\mb{u}_a \coloneqq f(\tilde{\mb{x}},u_i,\mb{y}_{},t)$ by which the attacker steers the system towards undesired states, and $t_a \ge t_0$ is the attack starting time.
For example, the attack signal is in the form of $\mb{u}_a(t)=\mb{u}_0e^{\lambda_0(t-t_a)}$ in the case of ZDA, where $\lambda_0$ and $\mb{u}_0$ are introduced in Definition \ref{def:inv_zeros}.
\noindent
\textbf{{Communication topology switching}.} The multi-agent system in \eqref{eq:cl_sys} operates in the \textit{normal mode} with the initial communication topology specified by $\sigma(t)=1\in \mathcal{Q},$ $ t\in [t_0, t_1)$ until switching to a \textit{safe mode} following the detection of an attack at the time $t_1 > t_a$. In the safe mode for $t \ge t_1$, the communication topology switching is specified by the switching signal $\sigma(t)=\braces{2,\dots,q}\in \mathcal{Q}$, $q\coloneqq|\mathcal{Q}|$ whose switching policy will be determined later (See Section \ref{S:detection}).
\begin{assumption}\btitle{Disclosed information}\label{assum:attacker} In the normal mode, where $\sigma(t)=1\in \mathcal{Q},$ $ t\in [t_0, t_1)$, the attacker
\begin{enumerate}
\item has perfect knowledge of the system model, i.e., $\Sigma_{\scriptscriptstyle \mathcal{A}}(\tilde{\mb{A}}_{ \sigma(t)},{\mb{B}},\tilde{\mb{C}}_{},\sigma=1)
=
\Sigma({\mb{A}}_{ \sigma(t)},\mb{B},\mb{C}_{},\sigma=1)$,
\item does not know the system's initial condition, i.e., $ \tilde{\mb{x}}(t_a) \neq {\mb{x}}(t_0) $, and $ \tilde{\mb{x}}(t_a) = \tilde{\mb{x}}_0 = \boldsymbol{0} $ in a covert attack.
\item has no knowledge of the system switching times $\braces{{t}_k}_{k=1}^{\mb{m}-1}$, $\mb{m}\in \ensuremath{\mathbb{N}}$ associated with the safe mode when $\sigma(t)=\braces{2,\dots,q}\in \mathcal{Q},$ $ t\in [t_1, +\infty)$,
\item starts the attack at $t_a\geq t_0=0$.
\end{enumerate}
\end{assumption}
\begin{assumption}\btitle{Defender's policy}\label{assum:defender}
The defender
\begin{enumerate}
\item selects the monitored agents and designs the attack detection framework,
\item designs the communication topology for the safe mode and its corresponding switching policy.
\end{enumerate}
\end{assumption}
For the detectability of adversarial attacks in switched systems, we will need the following technical result:
\begin{lemma}\btitle{Observability of linear switched systems \cite{tanwani2012observability}}\label{lemma:observability}
Given a system $\dot{\mb{x}}=\mb{A}_{\scriptscriptstyle \sigma(t)}\mb{x}$, with measurements $\mb{y}_{}=\mb{C}\mb{x}$, ($\mb{x} \in \ensuremath{\mathbb{R}}^n$ and $\mb{y} \in \ensuremath{\mathbb{R}}^p$), over the interval $t\in[t_{0},{t}_{\mb{m}})$ that includes switching instances $\braces{{t}_k}_{k=1}^{\mb{m}-1}$ for modes $\sigma(t)=k\in\mathcal{Q}$ with the dwell time ${\tau}_{k}={t}_{k}-{t}_{k-1}$, the output of system is given by $\mb{y}(t)=\mb{C}e^{\mb{A}_k(t-t_{k-1})}\prod_{l=k-1}^{1}e^{\mb{A}_l(\tau_{l})}\mb{x}(t_0), t\in[t_{k-1}, \, t_{k})$.
Then, (i) the system is observable and the initial condition $\mb{x}(t_0)$ is reconstructable from $\mb{y}(t)$ if, and only if, \eqref{eq:obsv_O} is full rank (i.e., $\boldsymbol{\mathcal{N}}_1^{\mb{m}} \coloneqq \mathrm{ker}(\boldsymbol{\mathcal{O}})=\{0\}$).
(ii) If \eqref{eq:obsv_O} is rank deficient, the unobservable subspace of the system for $t\in[t_{0},t_{\mb{m}})$, which is the largest $\mb{A}_{\scriptscriptstyle \sigma(t)}$-invariant subspace contained in $\mathrm{ker}(\mb{C})$, can be recursively computed using \eqref{eq:obsv_nullspace_1}-\eqref{eq:obsv_nullspace_2}.
\begin{align}
\label{eq:obsv_O}
\boldsymbol{\mathcal{O}}&=\mathrm{col}(\mathcal{O}_1,\mathcal{O}_2 e^{\mb{A}_1\tau_1},\cdots,\mathcal{O}_{\mb{m}}\prod_{i=\mb{m}}^{1}e^{\mb{A}_{i}\tau_{i}}),\\
\label{eq:obsv_nullspace_1}
\boldsymbol{\mathcal{N}}_{\mb{m}}^{\mb{m}} &= \mathrm{ker}(\mathcal{O}_{\mb{m}}),\\
\label{eq:obsv_nullspace_2}
\boldsymbol{\mathcal{N}}_{k}^{\mb{m}} &= \mathrm{ker}(\mathcal{O}_{k})
\cap \bracket{\bigcap_{i=k+1}^{\mb{m}} \mathrm{ker}\paren{\mathcal{O}_{i} \prod_{j=i-1}^{k}e^{\mb{A}_{j}\tau_{j}}}}, \\ \nonumber
\text{where}\\
\label{eq:obsv_nullspace_matrix}
\mathcal{O}_{k} &= \mathrm{col}\paren{\mb{C},\mb{C}\mb{A}_{k},\dots,\mb{C}\mb{A}_{k}^{\scriptscriptstyle 2N-1}}, \,\, 1\leq k \leq \mb{m}\!-\!1,\\
\label{eq:obsv_nullspace_Amatrix}
\mb{A}_k &= \mb{A}_{\sigma(t)}, \quad t\in [t_{k-1},t_{k}).
\end{align}
\end{lemma}
\begin{proposition}\btitle{Stealthy attacks}\label{propo:stealthy_attacks}
Consider system \eqref{eq:cl_sys}, under the attack model \eqref{eq:attack_model} and Assumption \ref{assum:attacker}, an attack is stealthy\footnote{The stealthy attacks defined by the condition \eqref{eq:stealthy} are also known as undetectable attacks in the literature \cite{pasqualetti2013attack}.} if the system output in \eqref{eq:cl_sys} satisfies
\begin{equation}\label{eq:stealthy}
\mb{y}_{}(\mb{x}_0,\mb{u}_a,\mb{u}_s,t)=\mb{y}_{}(\bar{\mb{x}}_0,\mb{0},\mb{0},t),\quad \forall \, t\in [t_0,t_1),
\end{equation}
where $\mb{x}_0$ and $\bar{\mb{x}}_0$ are the actual and possible initial states, respectively. Then, \eqref{eq:stealthy} can be realized in two senses
\begin{enumerate}
\item {Covert Attack}:
Under Assumption \ref{assum:attacker}, if the attacker sets the initial condition $\tilde{\mb x}(t_a)=\boldsymbol{0}$ or alternatively $\tilde{\mb x}(t_0)\in \boldsymbol{\mathcal{N}}_1^{1}=\mathrm{ker}(\mathcal{O}_1)$ in \eqref{eq:attack_model}, then the attack $\mb{u}_a$ on \eqref{eq:cl_sys} is covert, that is there exists a vector $\mb{u}_s$, injected in \eqref{eq:cl_sys}, canceling out the effect of $\mb{u}_a$ on the system output $\mb{y}(t)$.
%
\item {Zero-dynamics Attack (ZDA)}: the attacker can excite the zero dynamics of the system by an unbounded signal and remains stealthy with no need to alter the system measurements (i.e., $\mb{u}_s(t)=\boldsymbol{0}$ in \eqref{eq:cl_sys}) if $\tilde{\mb{x}}_0 \in \mathrm{ker}(\mb{C}_{})$ and $\mb{u}_a(t)=\mb{u}_0e^{ \lambda_0(t-t_a)},\, t_a=t_0$, where $\lambda_0$, $\tilde{\mb{x}}_0$ and $\mb{u}_0$ are obtained using Definition \ref{def:inv_zeros}.
\end{enumerate}
\end{proposition}
\begin{IEEEproof}
Clearly before an attack starts, \eqref{eq:stealthy} is met over $t\in [t_0,t_a)$. Consider $\mb{x}(t_a) $ as the system states when the attack starts,\\
({\textrm{i}}): in the case of covert attack, the output of the system \eqref{eq:cl_sys} with the initial normal mode $\sigma(t)=1$ over $t\in[t_a,t_{1})$ is given by
{\small
\begin{align}\label{eq:ym_solution}
\mb{y}_{}(t) &=
\mb{C}_{}
{e^{\mb{A}_1(t-t_{a})}}
\mb{x}(t_a)
+
\mb{C}_{}
{\int^{t}_{t_{a}}
e^{\mb{A}_{1}(t-\boldsymbol{\tau})}
\mb{B}\mb{u}_{a}(\boldsymbol{\tau})d{\boldsymbol{\tau}}}
- \mb{u}_s(t),
\end{align}
}
and the last term which is the output of the attacker's model \eqref{eq:attack_model} is given by
{\small
\begin{align}\label{eq:us_solution}
\mb{u}_{s}(t) &=
\tilde{\mb{C}}_{}
{e^{\tilde{\mb{A}}_1(t-t_{a})}}
\tilde{\mb{x}}(t_a)
+
\tilde{\mb{C}}
{\int^{t}_{t_{a}}
e^{\tilde{\mb{A}}_1(t-\boldsymbol{\tau})}
\mb{B}\mb{u}_{a}(\boldsymbol{\tau})d{\boldsymbol{\tau}}}.
\end{align}
}
Substituting \eqref{eq:us_solution} into \eqref{eq:ym_solution} and considering Assumption \ref{assum:attacker} yields
\begin{eqnarray}\label{eq:ym_solution_final}
\mb{y}_{}(t) =
\mb{C}_{} {e^{\mb{A}_1(t-t_{a})}}(\mb{x}(t_a)-\tilde{\mb{x}}(t_a)), \quad t\in[t_a,t_{1}).
\end{eqnarray}
The measurement \eqref{eq:ym_solution_final} matches the attack-free response if the attacker simply sets $\tilde{\mb x}(t_a)=\boldsymbol{0}$. Also, in the case $\tilde{\mb x}(t_a)\neq \boldsymbol{0},\, t_a=t_0=0$, it is immediate from lemma \ref{lemma:observability} that if $\tilde{\mb x}(t_0)\in \boldsymbol{\mathcal{N}}_1^{1}\neq\braces{0} \implies \mb{C}_{} {e^{\mb{A}_1(t-t_{a})}}\tilde{\mb{x}}(t_0)=\boldsymbol{0},\, t_a=t_0=0$ in \eqref{eq:ym_solution_final}, and thus $\mb{y}(t) = $ $ \mb{C}_{} {e^{\mb{A}_1(t-t_{a})}}\mb{x}(t_a),$ $ t \in [t_a,t_{1})$. In both of the cases, condition \eqref{eq:stealthy}, guaranteeing the covertness of the attack, is met. We, however, focus on the first case under Assumption \ref{assum:attacker}-{\rm (ii)}, therefore the system state $\mb{x}(t)$, without any jump, continuously holds the following
\begin{align}\label{eq:covert_trajec}
\mb{x}(t)&=\bar{\mb{x}}(t)+\tilde{\mb{x}}(t),
\end{align}
%
where
\begin{align}\label{eq:covert_trajec_pre}
%
\tilde{\mb{x}}(t)&=\boldsymbol{0}
\implies
\mb{x}(t) = \bar{\mb{x}}(t), \ \ &\forall\, t \in [t_0, t_a),
\\ \label{eq:covert_trajec_post}
\tilde{\mb{x}}(t)&=
{\int^{t}_{t_{a}}
e^{{\mb{A}}_{1}(t-\boldsymbol{\tau})}
\mb{B}\mb{u}_{a}(\boldsymbol{\tau})d{\boldsymbol{\tau}}},
\ \ &\forall\, t \in [t_a, t_1),
\end{align}
with $\bar{\mb{x}}(t),\, \forall\, t \in [t_0, t_1)$ denoting the state of the system in \eqref{eq:cl_sys} in the absence of covert attack (i.e. $\dot{\bar{\mb{x}}}=\mb{A}_{1}\bar{\mb{x}},\, \bar{\mb{x}}_0={\mb{x}}_0$).
\\
(\textrm{ii}): In the case of ZDA, let $t_a=t_0=0$ for simplicity, and $\bar{\mb{x}}_0 = \mb{x}_0 -\tilde{\mb{x}}_0$. Under Assumption \ref{assum:attacker} and using Definition \ref{def:inv_zeros}, the attacker can solve the following:
\begin{equation}\label{eq:zda_cond}
\begin{bmatrix}
\lambda_{\scriptscriptstyle 0} I_{} -\mb{A}_{1} & -\mb{B} \\ \mb{C}_{} & 0_{}
\end{bmatrix}
%
\begin{bmatrix}
\tilde{\mb{x}}_0 \\ \mb{u}_0
\end{bmatrix}
%
=
%
\begin{bmatrix}
\boldsymbol{0} \\ \boldsymbol{0}
\end{bmatrix},
\end{equation}
to design the ZDA signal $\mb{u}_a(t)=\mb{u}_0e^{ \lambda_0 t}$ causing unbounded system states
\begin{align}\label{eq:zda_trajec}
\mb{x}(t)=\bar{\mb{x}}(t)+\tilde{\mb{x}}_0e^{ \lambda_0 t},
\end{align}
while \eqref{eq:stealthy} is met, where $\bar x(t)$ is the state of the system in (5) assuming the initial condition $\bar x_0$ and no attack signal.
The second equation in \eqref{eq:zda_cond}, $\mb{C}\tilde{\mb{x}}_0=\boldsymbol{0}$, implies $\tilde{\mb{x}}_0\in \mathrm{ker}(\mb{C}_{})$.
It is an immediate result from Definition \ref{def:inv_zeros} that the attack signal $\mb{u}_a(t)=\mb{u}_0e^{\lambda_{\scriptscriptstyle 0}t}$ results in $\mb{u}_s(t)=\mb{C}\tilde{\mb{x}}(t)=\boldsymbol{0}$ in \eqref{eq:attack_model} while the system states $\tilde{\mb{x}}(t)=\tilde{\mb{x}}_0e^{\lambda_0t} \in \mathrm{ker}(\mb{C}), \, \forall \,t\in[t_0,t_{1})$ is unboundedly increasing.
Consider \eqref{eq:zda_trajec} and the superposition principle in linear systems,
then injecting the designed ZDA signal $\mb{u}_a(t)$ in \eqref{eq:cl_sys} yields the solution $\mb{y}=\mb{C}\mb{x}(t)=\mb{C}\bar{\mb{x}}(t)+\mb{C}\tilde{\mb{x}}_0e^{ \lambda_0 t}$, which by considering \eqref{eq:zda_cond} is equivalent to \eqref{eq:stealthy}, guaranteeing the stealthiness of ZDA for \eqref{eq:cl_sys}.
\end{IEEEproof}
Given the system and attack models above, we now state the two problems which this paper aims to address in the following:
\begin{problem}\btitle{Privacy-preserving average consensus}\label{prob:privacy}
Given the switching consensus system \eqref{eq:cl_sys},
we seek to preserve the following privacy requirements:
\begin{enumerate}
\item \label{privacy_requr_IC} neither system's initial states $\mb{x}(t_0)$ nor final agreement values ($x^*=\frac{1}{N}\sum_{i=1}^Nx_i(t_0)$, $v^*=0$) should be revealed or be reconstructable.
\item \label{privacy_requr_topol_reconst} the system's communication topology $\mathcal{G}_{\sigma(t)}$ should not be reconstructable.
\end{enumerate}
\end{problem}
\begin{problem}\btitle{Scalable attack detection}\label{prob:attack_detection}
Given the system in \eqref{eq:cl_sys} under the attack model \eqref{eq:attack_model}, we seek to develop a stealthy attack detection framework such that:
\begin{enumerate}
\item it features a decentralized and scalable structure.
\item it satisfies the privacy-preserving requirements defined in Problem \ref{prob:privacy}.
\end{enumerate}
\end{problem}
\section{Privacy Preservation and Attack Detection}\label{S:privacy_obs}
In this section, we describe the attack detection framework and characterize the conditions required to address Problems \ref{prob:privacy} and \ref{prob:attack_detection}.
\subsection{Attack Detection Scheme}\label{subS:attak_detection_scheme}
\begin{figure}[t]
\centering
\includegraphics[width=.9\linewidth]{figures/detection_scheme.pdf}
\caption{\small Attack detection architecture.}
\label{fig:detection_scheme}
\end{figure}
The proposed framework, depicted in Figure \ref{fig:detection_scheme}, is a two-level attack detection framework. It is privacy-preserving and relies on topology switching generating model discrepancy between the attacker model \eqref{eq:attack_model} and the actual system \eqref{eq:cl_sys}.
The system is decomposed into a set of subsystems based on the characteristics of its communication topology such as sparsity. Then, a set of monitored agents will be characterized such that each subsystem (the dynamics of agents within a cluster) is fully observable with respect to its locally available measurements while the main system \eqref{eq:cl_sys} is partially observable with respect to its globally available measurements \eqref{eq:measurments}.
We show how unobservability and system clustering can be used respectively to address Problem \ref{prob:privacy} and \ref{prob:attack_detection}.
Building upon global and (private) local measurements, the attack detection framework consists of a centralized observer, implemented in the control center, and local observer(s) in each cluster ($ \mathcal{P}_{\rm i} ,\, {\rm i}\in \{1,2\}$ in Figure \ref{fig:detection_scheme}). As increasing data transmission between agents and the centralized observer in the control center arises scalability and privacy concerns (cf. Problem \ref{prob:attack_detection}),
local observers play a vital role in our attack detection framework. They are hidden from the attacker because they are distributed among clusters of the multi-agent system, and their output is not sent to the control center but kept locally for attack detection.
If a local observer detects a stealthy attack, it triggers a network topology switch whereby the stealthy attack becomes detectable in the global measurements available for the centralized observer. The local decision making for network topology switches and indirect communication with the control center allow for agile reconfigurability in autonomous multi-agent systems (e.g. a network of autonomous aerial/ground vehicles) as well as eliminates the need for additional data exchange required, at global level, for monitoring and stealthy attack detection.
\subsection{Privacy Preservation}\label{Ss:privacy}
Problem \ref{prob:privacy} on privacy preservation can be addressed by imposing unobservability constraint on system \eqref{eq:cl_sys}. Indeed, one can select the set of monitored agents $\mathcal{M}$ in \eqref{eq:measurments} such that $(\mb{A}_{\sigma(t)},\mb{C})$ is not an observable pair on $t\in[t_0,+\infty)$, making the globally available measurement $\mb y$ in \eqref{eq:measurments} insufficient to reconstruct either the entire system states' information or the system's switching structure (cf. privacy requirements in Problem \ref{prob:privacy}).
The following lemma provides sufficient conditions to determine whether the global system measurement \eqref{eq:measurments} is consistent with the privacy requirements.
\begin{lemma}\btitle{Invariant unobservable subspace of system \eqref{eq:cl_sys}}\label{lemma:Inv_unobs_subspace}
The subspace $ \mathrm{span} \braces{\begin{smallmatrix} \boldsymbol{1}_{\! N} \\ \boldsymbol{0}_{\!N} \end{smallmatrix}}$ is an $\mb{A}_{\sigma(t)}$-invariant unobservable subspace of the switching system in \eqref{eq:cl_sys} provided that
it lies in $\mathrm{ker}(\mb{C}) $ and $\mathcal{G}_{\sigma(t)}$ features only connected undirected (or strongly connected and balanced directed) graphs.
\end{lemma}
\begin{IEEEproof} See Appendix \ref{app_Inv_unobs_subspace}.
\end{IEEEproof}
\begin{remark}\btitle{Generality of Lemma \ref{lemma:Inv_unobs_subspace}} The result suggests that monitoring only the agents' velocity causes the agents' positions not to be reconstructable independently for system \eqref{eq:cl_sys}. This is a generic solution to Problem \ref{prob:privacy} that holds for all undirected graphs. It is also worth noting that the monitored agents corresponding to set $\mathcal{M}$ in \eqref{eq:measurments} can be also selected differently from the results in Lemma \ref{lemma:Inv_unobs_subspace} for any particular graph.
\end{remark}
We next introduce the system partitioning method followed by observer design to address Problem \ref{prob:attack_detection}.
\subsection{System Partitioning}
Consider the communication graph $\mathcal{G}_{\sigma(t)}=(\mathcal{V},\mathcal{E},\mathcal{A})$ of the system \eqref{eq:cl_sys}, let the set of agents $\mathcal{V}$ be partitioned into disjoint clusters $\boldsymbol{\mathcal{P}} \coloneqq \{\mathcal{P}_1,\dots,\mathcal{P}_{\scriptscriptstyle |\boldsymbol{\mathcal{P}}|}\} $ such that $\cup_{\rm{i}=1}^{\scriptscriptstyle |\boldsymbol{\mathcal{P}}|}\mathcal{P}_{\rm{i}}=\mathcal{V}$ with $\mathcal{P}_{\rm{i}} \in \ensuremath{\mathbb{R}}^{N_{\rm{i}}}$ and inter-cluster couplings
$\mathcal{E}_{\rm{cut}} \coloneqq \setdef{\mathcal{E}_{{ij}}}{ {i}\in \mathcal{P}_{\rm i},\, {j}\in \mathcal{P}_{\rm j}, \, \mathcal{P}_{\rm i} \cap \mathcal{P}_{\rm j} = \emptyset}$.
Accordingly, after relabeling the system states, the system \eqref{eq:cl_sys} is partitioned into $|\boldsymbol{\mathcal{P}}|$ subsystems described as
\begin{equation}\label{eq:cluster_dyn2}
\Sigma_{\scriptscriptstyle \mathcal{P}_{{\rm i}}}:
\left\{
\begin{aligned}
%
\dot{\mb{x}}_{\rm{i}} &= \mb{A}_{ \sigma(t)}^{{\rm{i}}}\mb{x}_{\rm{i}}
+
\sum\nolimits_{\rm{j}\in \mathcal{N}_{\scriptscriptstyle \mathcal{P}_{\!i}}} \mb{A}_{{\sigma(t)}}^{\rm{ij}} \mb{x}_{\rm{j}} + \mb{B}^{{\rm i}} \mb{u}_{a_{\scriptscriptstyle \rm{i}}}, \\
\mb{y}_{{\rm i}_i} &=
\mb{C}_{{\rm i}_i}\mb{x}_{{\rm i}}, \quad i \in \mathcal{M}_{{\rm i}} \subset \mathcal{P}_{{\rm i}}, \\
\mb{x}_{{\rm i}}(0)&=
\mb{x}_{0_{{\rm i}}}, \qquad \, {\rm i} \in
\braces{1,\cdots,{|\boldsymbol{\mathcal{P}}|}},
\end{aligned}\right.
\end{equation}
with
\begin{align}
\mb{A}_{\sigma(t)}^{\rm{i}} &=
\begin{bmatrix}
0_{} & I_{} \\
-\alpha \mathcal{L}^{\rm}_{\sigma(t)} & -\gamma I_{}
\end{bmatrix}, \:
%
\mb{A}_{\sigma(t)}^{\rm{ij}} =
\begin{bmatrix}
0_{} & 0_{} \\
-\alpha \mathcal{L}^{\rm{ij}}_{\sigma(t)} & 0_{}
\end{bmatrix},\\
%
\mathcal{L}_{\sigma(t)} &= \begin{bmatrix}
\mathcal{L}^{\scriptscriptstyle 1}_{\sigma(t)} & \cdots & \mathcal{L}^{\s1,|\boldsymbol{\mathcal{P}}|}_{\sigma(t)} \\
\vdots & \ddots & \vdots\\
\mathcal{L}^{\scriptscriptstyle|\boldsymbol{\mathcal{P}}|,1}_{\sigma(t)} & \cdots & \mathcal{L}^{\scriptscriptstyle|\boldsymbol{\mathcal{P}}|}_{\sigma(t)}
\end{bmatrix}, \,
\mb{B}^{\rm i}= \begin{bmatrix}
0_{\scriptscriptstyle } \\ I_{\scriptscriptstyle \overline{\mathcal{F}}_{\rm i}}
\end{bmatrix},
\end{align}
where $\mb{x}_{ \rm i} \coloneqq \begin{bmatrix}(x)_{\rm i}^{\top} & (v)_{\rm i}^{\top} \end{bmatrix}^{\top} \! \in \ensuremath{\mathbb{R}}^{2 N_{{\rm i}}}$ with $(x)_{\rm i}$ and $(v)_i$ representing the vectors of position and velocity states belonging to cluster $\mathcal{P}_{\rm i} \subset \mathcal{V}$.
Also, $\mb{u}_{a_{\scriptscriptstyle \rm i}}$ associated with the set $\overline\mathcal{F}_{\rm i}$ is the vector-valued attack on actuator channels in the cluster as defined in \eqref{eq:cl_sys}.
The output signal $\mb{y}_{{\rm i}_i}(t)$, associated with the output matrix $\mb{C}_{{\rm i}_i}$, denotes the \emph{local} measurements that are available at node $i$ in cluster $\mathcal{P}_{\rm i}$.
Finally, $ \mathcal{N}_{\scriptscriptstyle \mathcal{P}_{{\rm i}}}:=\{{\rm j} \in \braces{1,\cdots,{|\boldsymbol{\mathcal{P}}|}} \mid \exists \, \mathcal{E}_{i,j} \in \mathcal{E}_{\rm{cut}}, i \in \mathcal{P}_{\rm i}, j \in \mathcal{P}_{\rm j} \} $ denotes the index set of the neighboring clusters of cluster $ \mathcal{P}_{\rm i} $.
We note that the decomposition of \eqref{eq:cl_sys} into \eqref{eq:cluster_dyn2} leads to a concatenated set $ \bar\mathcal{M} \coloneqq \braces{\mathcal{M},\mathcal{M}_1,\dots,\mathcal{M}_{|\boldsymbol{\mathcal{P}}|}} $, where the set $\mathcal{M}$ is associated with \emph{global} measurements \eqref{eq:measurments} available for the control center and sets $\mathcal{M}_{\rm i}\text{'s}, \; {\rm i} \in \boldsymbol{\mathcal{P}}$ are associated with the \emph{local} measurements $\mb{y}_{{\rm i}_i}$ available at a node $i$ in respective clusters $\mathcal{P}_1,\dots, \mathcal{P}_{\scriptscriptstyle |\boldsymbol{\mathcal{P}}|}$ in \eqref{eq:cluster_dyn2}.
We make the following assumptions:
%
\begin{assumption}\btitle{Local information}\label{assum:local_info}
\begin{enumerate}
\item \label{assum:local_info1} \emph{local knowledge}: in each cluster, the agent $i \in \mathcal{P}_{\rm i}$ serves as the {local control center} that has the local system model of the cluster (matrices $\mb{A}^{\rm i}_{\sigma(t)}$, $\mb{A}^{\rm ij}_{\sigma(t)}$ and $\mb{C}_{{\rm i}_i}$) and the {local measurement} $\mb{y}_{{\rm i}_i}(t)$.
\item \label{assum:local_info2} \emph{local measurements}: the measured output $\mb{y}_{{\rm i}_i}(t)$ in \eqref{eq:cluster_dyn2} is locally available at the node $i$ and, unlike global measurements, it is not sent to the control center to keep the output secure and inaccessible to the attacker.
\item \label{assum:local_info3} \emph{cross-cluster communication}: every local control center, i.e., the node $i$ in cluster $\mathcal{P}_{\rm i}$, considers coupling terms $\sum_{\rm j\in \mathcal{N}_{\scriptscriptstyle {\mathcal{P}_{{\rm i}}}}} \mb{A}_{{\sigma(t)}}^{\rm {ij}} \mb{x}_{\rm {j}}$ as unknown inputs to $\Sigma_{\scriptscriptstyle \mathcal{P}_{{\rm i}}}$. Moreover, inter-cluster couplings do not change, i.e., $\mb{A}_{{\sigma(t)}}^{\rm {ij}}=\mb{A}_{1}^{\rm {ij}}$, $ \forall\, t \in [t_0, +\infty)$. Thus there is no need for exchange of $\mb{x}_{\rm j}$'s information between local control centers.
\end{enumerate}
\end{assumption}
The assumption \ref{assum:local_info}-\ref{assum:local_info1} is common in the literature (cf. \cite{gallo2020distributed}) as the model-based detection of cyber attacks on exchanged data over a network requires augmented knowledge of the neighboring agents' model to estimate their states and further compare them with the received data. Minimizing the local information exchange affects the scalability and depends on the sparsity of the communication network as well as on applications.
\subsection{Observer Design and Attack Detectability Analysis}
As described in Section \ref{subS:attak_detection_scheme}, the attack detection framework is composed of a centralized observer for monitoring the system \eqref{eq:cl_sys} from the control center, and a set of local observers in clusters, that serve as local attack detectors and trigger for communication topology switching. In what follows, we describe the observer design procedure based on the conditions derived in the previous section.
\noindent
\textbf{{Decentralized observer.}} Consider the dynamics of the system partitions described in \eqref{eq:cluster_dyn2} and Assumption \ref{assum:local_info}, we use the unknown input observer (UIO) scheme in \cite{chen1996design} to estimate the cluster state $\hat{\mb{x}}_{\rm i}$ independent of the states $\mb{x}_{\rm j} $'s of the neighboring clusters (i.e. ${\rm j} \in \mathcal{N}_{\scriptscriptstyle \mathcal{P}_{{\rm i}}}$). This is achieved by considering the interconnection of local models as unknown inputs and rewriting them such that
\begin{equation}\label{eq:E_definition}
\sum_{{\rm j}\in \mathcal{N}_{\scriptscriptstyle \mathcal{P}_{{\rm i}}}} \mb{A}_{{\sigma(t)}}^{\rm{ij}} \mb{x}_{\rm{j}}
\coloneqq
\mb{E}_{}^{\rm i}\mb{x}^{d}_{\rm i}, \ \ {\sigma(t)}=1, \, \forall\, t\in [t_0, +\infty),
\end{equation}
where $\mb{E}_{}^{\rm i}$ is a full column rank\footnote{ The columns of $\mb{E}_{}^{ \rm i}$ for cluster $\mathcal{P}_{\rm i}$ are corresponding to the edge-cuts connecting $\mathcal{P}_{\rm i}$ to its neighboring clusters.} matrix and $\mb{x}^{d}_{\rm i}$ is a vector of the states of neighboring clusters that are received by cluster $\mathcal{P}_{\rm i}$. Now, introducing the UIO state $ \mb{z}_{\rm i} =\hat{\mb{x}}_{\rm i} - \mb{h}_{}^{\rm i }\mb{y}_{{\rm i}_i} $, the dynamics of the local UIO is given by
\begin{equation}\label{eq:obs_decent}
\Sigma^{\scriptscriptstyle \mathcal{Z}_{\rm i}}_{\scriptscriptstyle \mathcal{O}}:
\left\{
\begin{array}{l}
%
\dot{{\mb{z}}}_{\rm i}
=
\mb{F}_{\sigma(t)}^{\rm {i}}{\mb{z}}_{\rm i}
+
\paren{\mb{K}_{\sigma(t)}+\bar{\mb{K}}_{\sigma(t)}}\mb{y}_{{\rm i}_i},
\\
\hat{\mb{x}}_{\rm i} =\mb{z}_{\rm i}+\mb{h}_{}^{\rm i}\mb{y}_{ {\rm i}_i},\\
\hat{\mb{x}}_{\rm i}(0)=\mb{0}, \quad \mathcal{P}_{\rm i} \subset \mathcal{V}, \quad {\rm i} \in \braces{1,\cdots,|\boldsymbol{\mathcal{P}}|},
\end{array}
\right.
\end{equation}
where $\mb{F}_{\sigma(t)}^{\rm {i}}, \mb{K}_{\sigma(t)}, \bar{\mb{K}}_{\sigma(t)}, $ and $\mb{h}_{}^{\rm i}$ are matrices
satisfying conditions
\begin{align}\label{eq:uio_conds_1}
\mb{T}_{}^{\rm i}&=\paren{I-\mb{h}_{}^{\rm i}\mb{C}_{{\rm i}_i}}, \ \ \paren{\mb{h}_{}^{\rm i}\mb{C}_{{\rm i}_i}-I}\mb{E}_{}^{\rm i}=0,
\\ \label{eq:uio_conds_2}
%
\mb{F}_{\sigma(t)}^{\rm i}&=\paren{\bar{\mb{A}}_{\sigma(t)}^{\rm {i}} -\bar{\mb{K}}_{\sigma(t)} \mb{C}_{{\rm i}_i}}, \ \
%
{\mb{K}}_{ \sigma(t)}=\mb{F}_{\sigma(t)}^{\rm i} \mb{h}_{}^{\rm i}, \\ \label{eq:uio_conds_3}
%
\bar{\mb{A}}_{\sigma(t)}^{\rm i} &= \mb{A}_{\sigma(t)}^{\rm {i}} - \mb{h}_{}^{\rm i}\mb{C}_{{\rm i}_i} \mb{A}_{\sigma(t)}^{\rm{i}}.
\end{align}
Furthermore, $\mb{F}_{\sigma(t)}^{\rm i}$ is Hurwitz stable over $t\in [t_0,t_{\mb{m}})$ for all \emph{normal} and \emph{safe modes}.
Consider \eqref{eq:cluster_dyn2}, \eqref{eq:obs_decent} and let $ \mb{e}_{\rm i} := \mb{x}_{\rm i} - \hat{\mb{x}}_{\rm i}$, one can use the conditions in \eqref{eq:uio_conds_1}-\eqref{eq:uio_conds_3} to obtain the error dynamics of UIO as follows
\begin{equation}\label{eq:obs_decent_error}
\Sigma^{\mb{e}_{\rm i}}_{ \mathcal{O}}:
\left\{
\begin{array}{l}
\:\dot{{\mb{e}}}_{\rm i} \:\,= {\mb{F}_{\sigma(t)}^{\rm {i}}} \mb{e}_{\rm i}
+ \mb{T}_{}^{\rm i} \mb{B}^{\rm i}\mb{u}_{a_{\scriptscriptstyle \rm i}} ,
\quad {\mb{e}}_{\rm i}(0)=\mb{x}_{\rm i}(0),
\\
{\mb{r}}_{{\rm i}_i} =\mb{C}_{{\rm i}_i}\mb{e}_{\rm i} , \quad \mathcal{P}_{\rm i} \subset \mathcal{V}, \ \ {\rm i} \in \braces{1,\cdots,|\boldsymbol{\mathcal{P}}|}.
\end{array}
\right.
\end{equation}
In the absence of adversarial attacks, $\mb{u}_{a_{\scriptscriptstyle \rm i}}=\boldsymbol{0}$, it is straightforward to show that $ \lim_{t\rightarrow \infty} \mb{e}_{\rm i}(t) = \boldsymbol{0}$ as $\mb{F}_{\sigma(t)}^{\rm i}$ is Hurwitz stable in all modes. LMI-based approaches can be used to design \eqref{eq:uio_conds_2} such that \eqref{eq:obs_decent_error} remains stable under arbitrary switching \cite{chesi2011nonconservative}.
Recall Assumption \ref{assum:local_info}-\ref{assum:local_info2}, unlike the case of global measurements (cf. Proposition \ref{propo:stealthy_attacks}-({\rm{i}})), the local measurements $\mb{y}_{{\rm i}_i}$'s are hidden and thus cannot be altered by the attacker to cancel out the effect of the attack $\mb{u}_{a_{\rm i}}$ on the output of \eqref{eq:cluster_dyn2}. This difference also manifests itself in the residual of local observer \eqref{eq:obs_decent_error}. Therefore, in order to determine the stealthiness of attack $\mb{u}_{a_{\rm i}}$ with respect to the local residual signal $\mb{r}_{{\rm i}_i}$, it is necessary and sufficient to investigate whether the stealthiness conditions presented in Proposition \ref{propo:stealthy_attacks} are satisfied for the system in \eqref{eq:obs_decent_error}.
In the following proposition, we formally characterize the conditions for the detection of stealthy attacks using the local observer in \eqref{eq:obs_decent}.
\begin{proposition}\btitle{Attack detectability of local observers}\label{prop:attack_detectability_local}
For a strongly connected cluster $\mathcal{P}_{\rm i}$ with $\boldsymbol{\mathcal{E}}$ inter-clustering edges and $|{\overline{\mathcal{F}_{\rm i}}}|$ compromised agents, there exists a local observer given by \eqref{eq:obs_decent} to locally detect the stealthy attacks if
\begin{enumerate}
\item there is a $\mb{k}$-connected node $i \in \mathcal{P}_{\rm i}$ as the local monitored agent such that $\mb{k} \geq \boldsymbol{\mathcal{E}} + |{\overline{\mathcal{F}_{\rm i}}}|$,
%
\item $ \mathrm{rank}\paren{\mb{C}_{{\rm i}_i}\mb{E}_{}^{\rm i}}=\mathrm{rank}\paren{\mb{E}_{}^{\rm i}}$,
\item the matrix pencil $\mb{P}$ in \eqref{eq:pencil_uio} is full (column) rank,
\begin{align}\label{eq:pencil_uio}
\mb{P} =
\begin{bmatrix}
\lambda_0 I_{} -{\mb{A}}^{\rm i}_{\sigma(t)} & -{\mb{B}}^{\rm i} & {\mb{E}}^{\rm i}_{}
\\
\mb{C}_{{\rm i}_i} & 0 & 0
\end{bmatrix}.
%
\end{align}
\end{enumerate}
where the tuple $\paren{{\mb{A}}^{\rm i}_{\sigma(t)},{\mb{B}}^{\rm i},\mb{C}_{{\rm i}_i}}$ and matrix ${\mb{E}}^{\rm i}_{}$ are defined in \eqref{eq:cluster_dyn2} and \eqref{eq:E_definition}, respectively.
\end{proposition}
\begin{IEEEproof}
See Appendix \ref{app_attack_detectability_local}.
\end{IEEEproof}
\begin{remark}\btitle{Evaluation of the condition in \eqref{eq:pencil_uio}}\label{rmk:local_obs_measurements}
Conditions \emph{(i)}-\emph{(iii)} in Proposition \ref{prop:attack_detectability_local} are equivalent to necessary and sufficient conditions for the existence of UIO in \eqref{eq:obs_decent} \cite{chen1996design}.
It is worth noting that as matrix $\mb{B}^{\rm i}$ in \eqref{eq:pencil_uio} is unknown to the defender, it can be replaced with $I_{N_i}$, i.e., assuming all the nodes of the cluster are under attack, in analysis and selecting locally monitored agents associated with $\mb{C}_{{\rm i}_i}$. This, however, may require further communication between agents within a cluster. Alternatively, as in a set cover problem setting, a set of local monitoring agents that each of them satisfies the conditions \emph{(i)}-\emph{(iii)} for part of a cluster can be used to cover all of nodes of the cluster \cite{teixeira2014distributed}. Minimizing the number of local measurements versus the number of local observers is a trade-off problem which will be the subject of future work.
\end{remark}
\noindent
\textbf{{Centralized observer.}} Consider the dynamical system \eqref{eq:cl_sys}, a Luenberger-type centralized observer, derived based on the normal mode $\sigma(t)=1$, is given by
\begin{equation}\label{eq:obs_cent}
\Sigma_{\scriptscriptstyle \mathcal{O}}^{\scriptscriptstyle \mathcal{M}} :
\left\{
\begin{array}{ll}
\dot{\hat{\mb x}}
=
\mb{A}_{\sigma(t)} \hat{\mb x} + \mb{H}_{\sigma(t)}(\mb{y}_{\scriptscriptstyle} - \hat{\mb y}_{} ),
& \sigma(t)=1, \\
\hat{\mb{y}}_{\scriptscriptstyle }=\mb{C}_{\scriptscriptstyle }\hat{\mb x},
& \hat{\mb{x}}(0)=\boldsymbol{0}, \\
\mb{r}_{\scriptscriptstyle 0} = (\mb{y}_{\scriptscriptstyle } -\hat{\mb{y}}_{\scriptscriptstyle }), & \text{residual,}
\end{array}
\right.
\end{equation}
where $ \mb{H}_{\sigma(t)}$ is the observer gain and $\mb{r}_{\scriptscriptstyle 0}(t)$ denotes the residual signal available in the control center for monitoring purposes.
In order to design the observer gain $ \mb{H}_{\sigma(t)}$, the partial observability of pair $(\mb{A}_{ \sigma(t)},\mb{C})$ imposed in Section \ref{Ss:privacy} and the activated mode $\sigma(t)$ should be taken into account. An immediate solution is to define an LMI optimization problem finding a constant $\mb{H}_{\sigma(t)} \coloneqq \mb{H}$ by which $(\mb{A}_{\sigma(t)}-\mb{H}\mb{C})$ is (Hurwitz) stable in all modes \cite{chilali1996h,chen2004observer}.
From Assumption \ref{assum:attacker} and condition \eqref{eq:stealthy}, it is straightforward to show that
the attack $\mb{u}_a$ remains stealthy for the observer \eqref{eq:obs_cent} in the normal mode over the time span $t\in[t_0,t_1)$ where $\mb{A}_{\sigma(t)}=\mb{A}_{1}$.
Recall \eqref{eq:covert_trajec} and \eqref{eq:zda_trajec}, and let
\begin{align}\label{eq:est_error_normal}
\bar{\mb{e}}
&\coloneqq
\bar{\mb{x}}-\hat{\mb{x}}
\\ \label{eq:est_error}
\mb{e}
&\coloneqq
\mb{x}-\hat{\mb{x}}=\bar{\mb{x}}+\tilde{\mb{x}}-\hat{\mb{x}}=\bar{\mb{e}}+\tilde{\mb{x}}
\end{align}
be the estimation error of the states of an attack-free system ($\dot{\bar{\mb{x}}}=\mb{A}_{\sigma(t)}\bar{\mb{x}},\, \mb{y}=\mb{C}\bar{\mb{x}}$) and the under attack system in \eqref{eq:cl_sys}, respectively.
Then using \eqref{eq:cl_sys} and \eqref{eq:obs_cent}, the error dynamics of the centralized observer is given by
\begin{equation}\label{eq:obs_cent_error}
\Sigma_{\scriptscriptstyle \mathcal{O}}^{\mb{e}} :
\left\{
\begin{array}{l}
\dot{{\mb e}}
=
(\mb{A}_{1}-\mb{H}\mb{C}) {\mb e}
+ (\mb{A}_{\sigma(t)}-\mb{A}_{1}){\mb{x}}
+\mb{H}\mb{u}_s+ \mb{B}\mb{u}_a, \\
{\mb{e}}(0) = \mb{x}_0,\\
\mb{r}_{\scriptscriptstyle 0} = (\mb{y}_{\scriptscriptstyle } -\hat{\mb{y}}_{\scriptscriptstyle })=\mb{C}\mb{e}-\mb{u}_s=\mb{C}\bar{\mb{e}}, \quad \text{ residual},
\end{array}
\right.
\end{equation}
where for measurement $\mb{y}$ in \eqref{eq:obs_cent} we used the expression $\mb{y}=\mb{C}\mb{x}-\mb{u}_s$ as defined in \eqref{eq:cl_sys}. Consider \eqref{eq:attack_model} and \eqref{eq:stealthy}, $\mb{y}$ in \eqref{eq:obs_cent} also satisfies $\mb{y}=\mb{C}\mb{x}-\mb{u}_s=\mb{C}\mb{x}-\mb{C}\tilde{\mb{x}}=\mb{C}\bar{\mb{x}}$. Then using $\mb{y}=\mb{C}\bar{\mb{x}}$, \eqref{eq:cl_sys}, \eqref{eq:attack_model}, \eqref{eq:obs_cent}, \eqref{eq:est_error_normal}, the following dynamics is obtained
\begin{equation}\label{eq:obs_cent_error_normal}
\Sigma_{\scriptscriptstyle \mathcal{O}}^{\bar{\mb{e}}} :
\left\{
\begin{array}{l}
\dot{\bar{\mb e}}
=
(\mb{A}_{1}-\mb{H}\mb{C}) \bar{\mb e}
+ (\mb{A}_{\sigma(t)}-\mb{A}_{1})\bar{\mb{x}}, \\
\bar{\mb{e}}(0) = \bar{\mb{x}}_0,\\
\bar{\mb{r}}_{\scriptscriptstyle 0} =\mb{C}\bar{\mb{e}}, \quad \text{ residual}.
\end{array}
\right.
\end{equation}
Note that, during normal mode $\sigma(t)=1$ over the time span $\forall \, t \in [t_0, t_1)$, the residual $\mb{r}_{\scriptscriptstyle 0}$ in \eqref{eq:obs_cent_error} is the same as that of \eqref{eq:obs_cent_error_normal} that is the dynamics of the estimation error of system states in the absence of attacks. This implies that, in the case of a covert attack with $\mb{u}_s \neq 0 $, as long as signal $\mb{u}_s(t)$ cancels out the effect of $\mb{u}_a(t)$ on the output $\mb{y}(t)$, the residual $ \mb{r}_0(t) = \mb{C}\bar{\mb{e}}(t)$ converges to zero as $t_1 \rightarrow +\infty$, yielding the stealthiness of the covert attack, in the normal mode, for the centralized observer \eqref{eq:obs_cent}.
\\
In the case of a ZDA, $\mb{u}_s = 0$ in \eqref{eq:obs_cent_error} although \eqref{eq:stealthy} still holds that leads to the stealthiness of a ZDA for the observer \eqref{eq:obs_cent}. To show this, one need to verify the attack $\mb{u}_a$ remains in the zeroing direction of \eqref{eq:obs_cent_error}. Using Definition \ref{def:inv_zeros} for \eqref{eq:obs_cent_error} in the normal mode, we obtain
\begin{equation}\label{eq:obs_cent_INV_zero}
\begin{bmatrix}
\lambda_0 I_{} - (\mb{A}_{1}-\mb{H}\mb{C}) & -\mb{B} \\ \mb{C} & 0_{}
\end{bmatrix}
%
\begin{bmatrix}
\tilde{\mb{e}}(0) \\ \mb{u}_0
\end{bmatrix}
%
=
%
\begin{bmatrix}
\boldsymbol{0} \\ \boldsymbol{0}
\end{bmatrix},
\end{equation}
where $\tilde{\mb{e}}(0) \coloneqq \mb{e}(0)-\bar{\mb{e}}(0)=\mb{x}_0-\bar{\mb{x}}_0= \tilde{\mb{x}}_0$. Recall $\tilde{\mb{x}}_0 \in \mathrm{ker}(\mb{C})$ in \eqref{eq:zda_cond}, then the second equation of \eqref{eq:obs_cent_INV_zero} yields $\mb{C}\tilde{\mb{e}}(0)=\mb{C}\tilde{\mb{x}}_0=\boldsymbol{0}$. Applying $\mb{C}\tilde{\mb{e}}(0)=\boldsymbol{0}$ into the first equation of \eqref{eq:obs_cent_INV_zero} simplifies the matrix pencil in \eqref{eq:obs_cent_INV_zero} into that of \eqref{eq:zda_cond} over $t\in[t_0,t_1)$ where $\mb{A}_{\sigma(t)}=\mb{A}_{1}$. This ensures the stealthiness of ZDA in the normal mode for the observer \eqref{eq:obs_cent}.
The following Theorem provides conditions to address Problem \ref{prob:attack_detection}-(ii) by characterization of switching modes that lead to attack detection with respect to global measurements.
\begin{theorem}\btitle{Attack detectability under switching communication}\label{th:Attack_detectability_switching}
Consider system \eqref{eq:cl_sys} under stealthy attack modeled in \eqref{eq:attack_model}, and let intra-cluster topology switching satisfy
\begin{enumerate}
\item $ \mathrm{Im} (\Delta \mathcal{L}_{\mb{q}}) \cap \mathrm{ker}{([\mb{C}^{\top}_{\rm x}\; \mb{C}^{\top}_{\rm v}]^{\top})} =\emptyset,$
%
\label{cond:detection_condition_suff_0}
\item $\mathcal{L}_{\mb{q}}$ features distinct eigenvalues,
\label{cond:detection_condition_suff_1}
%
\item $[\mathcal{U}_{\mb{q}}]_{i,\ell}-[\mathcal{U}_{\mb{q}}]_{j,\ell} \neq 0,\ \ \forall \, \ell \in \mathcal{V}\setminus\{1\}, \, \forall \, i,j \in \mathcal{D}_{\rm c},\, \forall \, {\rm c} \in \{1,\cdots, \boldsymbol{{\rm c}}\} $,\label{cond:detection_condition_suff_2}
\end{enumerate}
where $\Delta \mathcal{L}_{\mb{q}} \! \coloneqq \! \mathcal{L}_{\sigma(t)}-\mathcal{L}_{1} $, with $ \sigma(t) \!=\! \mb{q} \! \in \! \mathcal{Q},$ $ t \in [t_1, +\infty)$, $\mb{C}^{\top}_{\rm x}$ and $ \mb{C}^{\top}_{\rm v}$ are given in \eqref{eq:cl_sys}-\eqref{eq:measurments} and $\mathcal{D}_{\rm c} \!\subset \! \mathcal{V},$ denotes the set of nodes in ${\rm c}$-th connected component of $\Delta \mathcal{L}_{\mb{q}}$ corresponding to agents involved in connected switching links, and finally $\mathcal{U}_{\mb{q}}$ is a unitary matrix ( $\!\mathcal{U}_{\mb{q}}\mathcal{U}_{\mb{q}}^{\top}=I$) diagonalizing Laplacian $\mathcal{L}_{\mb{q}}$.
\\
Then, ZDA and covert attacks undetectable for the centralized observer \eqref{eq:obs_cent} are impossible only if the topology switching satisfies conditions {\ref{cond:detection_condition_suff_0}-\ref{cond:detection_condition_suff_2}}. If additionally the system is not at its exact consensus equilibrium when the attack is launched, conditions {\ref{cond:detection_condition_suff_0}-\ref{cond:detection_condition_suff_2}} are sufficient for the detection of ZDA.
\end{theorem}
\begin{IEEEproof} See Appendix \ref{app_Attack_detectability_switching}. \end{IEEEproof}
\begin{remark}\btitle{Safe topology switching}\label{rmk:Safe_topology_switching}
For a given pair $(\mb{A}_{\sigma(t)},\mb{C})$ in \eqref{eq:cl_sys}, one can compute a set of switching modes by evaluating the conditions {\ref{cond:detection_condition_suff_0}-\ref{cond:detection_condition_suff_2}} of Theorem \ref{th:Attack_detectability_switching}. This could be performed through iterative algorithms changing graph connections.
Furthermore, if $ \boldsymbol{\mathcal{Z}} $ be an unknown subspace associated with system states affected by stealthy attack $\mb{u}_a(t)$ i.e. $\tilde{\mb{x}}(t) \in \boldsymbol{\mathcal{Z}}$. Then, in view of $\tilde{\mb{x}}_0 = \mb{x}_0- \bar{\mb{x}}_0$ (see Proposition \ref{propo:stealthy_attacks}), the discrepancy term $(\mb{A}_{\sigma(t)}-\mb{A}_{1}){\mb{x}}$ in the dynamical system \eqref{eq:obs_cent_error} will be bounded and vanishing if
\begin{equation}\label{eq:unsafe_set}
\boldsymbol{\mathcal{X}}_{\mb{q}} \cap \boldsymbol{\mathcal{Z}} =\emptyset.
\end{equation}
Therefore, if condition \eqref{eq:unsafe_set} holds, $(\mb{A}_{\sigma(t)}-\mb{A}_{1}){\mb{x}}$ does not effect the stability of the system, as a consequence of input-to-state stability property of consensus systems \cite{meng2018consensus}.
It is also noteworthy that although identifying $\boldsymbol{\mathcal{Z}}$ beforehand is practically impossible as $\mb{B}$ and $\tilde{\mb{x}}_0$ in \eqref{eq:attack_model} are unknown to the defender, local observers detecting stealthy attacks in a cluster can locally identify and trigger a safe switching mode that satisfies \eqref{eq:unsafe_set}.
\end{remark}
\subsection{Attack Detection Procedure}\label{S:detection}
The results in the previous section provide conditions for the detectability of stealthy attacks locally, at the cluster level, and globally, at a ground control station equipped with a centralized observer. As described earlier, the attack detection framework relies on switching communication links generating a discrepancy between the attacker model \eqref{eq:attack_model} and the actual system \eqref{eq:cl_sys}. To this end, at local level (clusters), unknown-input observers in \eqref{eq:obs_decent}, satisfying conditions of Proposition \ref{prop:attack_detectability_local}, locally detect stealthy attacks. Followed by the detection, a local observer triggers a topology switching, $\mathcal{G}_{\sigma(t)}$, that satisfies conditions \ref{cond:detection_condition_suff_0}-\ref{cond:detection_condition_suff_2} of Theorem \ref{th:Attack_detectability_switching}, yielding stealthy attack detection in the control center. This procedure is depicted in Algorithm \ref{alg:detection_procedure}.
\begin{algorithm}[H]
\caption{Topology switching for attack detection}\label{alg:detection_procedure}
\begin{algorithmic}[1]
\Procedure{Attack detection}{$\,\mathcal{G}_{\sigma(t)}$, Obs. in \eqref{eq:obs_cent}, \eqref{eq:obs_decent}}
\State \textbf{do} run global observer \eqref{eq:obs_cent} and local observers \eqref{eq:obs_decent}.
\If{$\mb{r}_{{\rm i}_i}(t) > \texttt{threshold} $}
\State \textbf{do} Identify a safe mode $\sigma(t)=\mb{q}\in \mathcal{Q}$ for $\mathcal{L}_{\sigma(t)}$ that satisfies conditions \ref{cond:detection_condition_suff_0}-\ref{cond:detection_condition_suff_2} in Theorem. \ref{th:Attack_detectability_switching}
\State \textbf{do} Trigger an identified safe mode $\sigma(t)=\mb{q}\in \mathcal{Q}$
\If{$ \mb{r}_0(t) > \texttt{threshold} $}
\State Stealthy attack is detected.
\EndIf
\EndIf
\EndProcedure
\end{algorithmic}
\end{algorithm}
As presented in Algorithm \ref{alg:detection_procedure}, the observers (attack detectors) require an appropriate threshold for their residuals to avoid false attack detection. These thresholds can be designed by considering an upper bound on the estimation error of observers in the attack-free case. An analytical analysis, however, will be the subject of future work.
\section{Simulation Results}\label{S:Results}
\begin{figure*}[t]
\centering \subfloat[Case 1: bounded residual]{\small
\hspace*{-10pt} \includegraphics[width=.25\linewidth]{figures/case1.pdf}
}
\subfloat[Case 2: unbounded residual]{\small
\hspace*{-10pt} \includegraphics[width=0.25\linewidth]{figures/case2.pdf}
}
\subfloat[Case 3: no detection]{\small
\hspace*{-10pt} \includegraphics[width=0.25\linewidth]{figures/test4.pdf}
}
\subfloat[Communication topology]{\small
\includegraphics[width=0.25\linewidth]{figures/C.pdf}}
\hspace*{-10pt}
%
\caption{\small Simulation results of privacy-preserving stealthy attack detection for a $19$-node multi-agent control system. The state trajectory $\mb{x}(t)$ consists of the agents' position (blue) and velocity (green) as well as red trajectories showing affected agents by the stealthy attack (ZDA). (a)-(c) the results of attack detection for three cases with their respective control topology switching depicted in (d). In all cases of (d), green nodes show the globally monitored agents by the centralized observer, blue nodes indicate the local control centers equipped with local observers, and red-bordered shows compromised agents and red-coloured nodes indicate compromised agents affected by the stealthy attack (ZDA). Finally, dashed lines (edges) show the switching communication links. In the local residual's figures, with slight abuse of notations (cf. \eqref{eq:obs_decent_error}), the scalar residual $\mb{r}_{{\rm i}_i}$ shows only the velocity estimation error of node $i$.}
\label{fig:simulation_results}
\vspace*{-1ex}
\end{figure*}
We use a numerical example to validate the performance of the attack detection framework. We consider a network of $N=19$ agents and investigate, in three cases, the effect conditions proposed in Proposition \ref{prop:attack_detectability_local} and Theorem \ref{th:Attack_detectability_switching} on stealthy attack detection. It is assumed that the network has been partitioned into three clusters $\mathcal{P}_1=\braces{1,\cdots,7}$, $\mathcal{P}_2=\braces{8,\cdots,12}$, $\mathcal{P}_3=\braces{13,\cdots,19}$. Each cluster is equipped with the local observer \eqref{eq:obs_decent} (specified by blue nodes in Figure \ref{fig:simulation_results}) whose local measurements are consistent with Assumption \ref{assum:local_info} and Proposition \ref{prop:attack_detectability_local}. More specifically, In cases 1 and 2, cluster $\mathcal{P}_1$ has two local observers that each has access to its neighboring agents' measurements. In cluster $\mathcal{P}_2$, however, we considered one local observer having more communication with other agents within the cluster for its realization (cf. Remark \ref{rmk:local_obs_measurements}). Similar analysis is applied to case 3. Moreover, there is a centralized observer with global measurements as $\mathcal{M}_{x}=\emptyset$, $\mathcal{M}_{v}=\braces{7,12,14}$ consistent with Lemma \ref{lemma:Inv_unobs_subspace}. In the simulations, the system's initial conditions are considered to be known for observers although this is not a requirement for the presented theoretical results. Also, the constant thresholds were selected by evaluating the observers' performance in different case studies.
In cases 1 and 2 (shown respectively in Figures \ref{fig:simulation_results}-(a) and \ref{fig:simulation_results}-(b) with their communication topology in Figure \ref{fig:simulation_results}-(d)) a ZDA occurs in cluster $\mathcal{P}_1$ and particularly affects agents $3$ and $4$. As depicted, ZDA is stealthy in the global residuals $\mb{r}_{0_i}\text{'s}, \, i\in\braces{1,12,14}$ before topology switching. It is, however, detectable in local residual $\mb{r}_{1_5}(t)$. The local control center, node $5$, can trigger either of case 1's or case 2's switching topologies shown in Figures \ref{fig:simulation_results}-(d). While the conditions \ref{cond:detection_condition_suff_0}-\ref{cond:detection_condition_suff_2} of Theorem \ref{th:Attack_detectability_switching} are met in both cases, only case 2 meets \eqref{eq:unsafe_set} of remark \ref{rmk:Safe_topology_switching}. Consequently, the global residual $\mb{r}_{0_1}(t)$ for case 1 is bounded and vanishing after topology switching while that of case 2 is unbounded.
In cases 3 (shown in Figure \ref{fig:simulation_results}(c) with its communication topology in \ref{fig:simulation_results}-(d)) a ZDA occurs in cluster $\mathcal{P}_2$ and particularly affects agents $11$. Note that, unlike in cases 1 and 2, none of the Theorem \ref{th:Attack_detectability_switching}'s conditions are met in case 3, yielding the global residuals $\mb{r}_{0_i}(t), \, i\in\braces{1,12,14}$ remain unaffected
by the switching topology. Consequently, stealthy attack is not detectable.
Moreover, comparing cases 1's bounded global residual with case 2's unbounded global residual, it is noteworthy that meeting condition \eqref{eq:unsafe_set} yields a trade-off between a faster attack detection at a price of further exposing system states to ZDA and a slower detection by keeping uncompromised system states bounded.
\section{Conclusions}\label{S:conclusion}
In this paper, a novel attack detection framework is developed to detect stealthy attacks against a class of multi-agent control systems seeking average consensus. The scalability of the approach is addressed using decentralized local observers. Also, the privacy preservation of the multi-agent system's state information is achieved by imposing unobservability conditions for the central (global) observer. Theoretical conditions were derived for the detectability of the stealthy attacks. The numerical example validates the theoretical results and illustrates the effectiveness of the proposed approach.
Also, a discussion was provided on different types of switching topologies and their outcome for stealthy attack detection.
Deriving sufficient and verifiable conditions on safe topology switching
as well as optimizing the number of local observers and their respective measurements will be subjects of future work.
\appendices
\section{}\label{appdx:Additional_Lemmas}
The followings are used in the Proof of Theorem \ref{th:Attack_detectability_switching}.
\begin{definition}\label{prop:app_def_diagonal}
The Laplcaian matrix of the graph composed of switching links between two communication graphs is block diagonalizable,
where each block, also called a component, encodes either a single (added/removed) switching link or a group of them are are connected.
\end{definition}
The above definition can be formally presented as follows: consider a network topology switching between two graphs with Laplacian matrices $\mathcal{L}_{\sigma(t)=\mb{q}'}$ and $\mathcal{L}_{\sigma(t)=\mb{q}},\, \mb{q}', \mb{q} \in \mathcal{Q},\, \mb{q}' \neq \mb{q}$ and let $\Delta \mathcal{L}_{\mb{q}} = \mathcal{L}_{\mb{q}} - \mathcal{L}_{\mb{q}'}$ denote the difference of their Laplacian matrices. Then, under Definition \ref{def:graph_component},
$\Delta \mathcal{L}_{\mb{q}}$ is associated with the induced graph $\Delta \mathcal{G}_{\mb{q}}=(\mathcal{V}_{\mb{q}},\Delta \mathcal{E}_{\mb{q}},\Delta \mathcal{A}_\mb{q}) $, that specifies connected graph component(s) corresponding to added/removed communication link(s) in the communication network such that
\begin{align}\label{eq:app_components_sets_1}
\mathcal{V}_{\mb{q}} &= (\cup_{{\rm c}=1}^{\boldsymbol{\rm c}} \mathcal{D}_{\rm c}) \cup \mathcal{D}_{\rm s}, \;\;\text{\rm s.t}\;\; \mathcal{V}_{\mb{q}} = \mathcal{V},
\\
(i,j) &\in \Delta \mathcal{E}_{\mb{q}} \;\;\text{if}\;\; [\Delta \mathcal{A}_\mb{q}]_{i,j} = a^{\mb{q}}_{ij}-a^{1}_{ij} \neq 0 \iff
\nonumber \\ \label{eq:app_components_sets_2}
& \hspace{100pt} [\Delta \mathcal{L}_{\mb{q}}]_{i,j} \neq 0,
\end{align}
where $\mathcal{D}_{\rm c}$ denotes the set of nodes (agents involved in switching links) in ${\rm c}$-th connected component with $|\mathcal{D}_{\rm c}| \geq 2$ and $\mathcal{D}_{\rm i'} \cap \mathcal{D}_{\rm j'} = \emptyset$ for any $i',j' \in \{1, \cdots, \boldsymbol{\rm c}\},\, i'\neq j'$. Also, $\mathcal{D}_{\rm s}$ denotes the set of singletons i.e. single nodes that are not involved in any switching link.
Then, there exists a permutation matrix ${\rm P}, \, {\rm P}{\rm P}^{\top} =I$ to relabel the nodes and represent the Laplacian matrix $\Delta \mathcal{L}_{\mb{q}}$ in block diagonal form, (cf. \cite[Ch. 6.12]{newman2018networks}), as follows
\begin{align}\label{eq:graph_component_lap}
{\rm P}\Delta \mathcal{L}_{\mb{q}}{\rm P}^{\top}
&=
\breve{\mathcal{L}}_\mb{q}
\nonumber \\
&=
\mathrm{diag}{ \{\Delta \mathcal{L}_{\mb{q}}(\mathcal{D}_{\rm 1}), \cdots, \Delta \mathcal{L}_{\mb{q}}(\mathcal{D}_{\boldsymbol{\rm c}}), \Delta \mathcal{L}_{\mb{q}}(\mathcal{D}_{\rm s}) \}},
\end{align}
where $\Delta \mathcal{L}_{\mb{q}}(\mathcal{D}_{\rm c})$ denotes the Laplacian matrix of the ${\rm c}$-th connected component and $\Delta \mathcal{L}_{\mb{q}}(\mathcal{D}_{\rm s})=0$.
\begin{lemma}\label{lemma:app_component_measur}
Consider system in \eqref{eq:cl_sys} with topology switching from normal mode $\sigma(t)=1$ to a safe mode $\sigma(t)=\mb{q} \in \mathcal{Q}$ and the measurements set $\mathcal{M}$ in \eqref{eq:measurments}, and let $\Delta \mathcal{L}_{\mb{q}} = \mathcal{L}_{\mb{q}} - \mathcal{L}_{1}$ denote the difference of the Laplacian matrices in safe and normal mode. Then
under condition
\begin{align}\label{eq:cond_suff_0}
\mathrm{Im} (\Delta \mathcal{L}_{\mb{q}}) \cap \mathrm{ker}{([\mb{C}^{\top}_{\rm x}\; \mb{C}^{\top}_{\rm v}]^{\top})} =\emptyset,
\end{align}
every connected graph component has at least one globally monitored node (agent), that is
\begin{align}\label{eq:app_component_measurements}
\mathcal{D}_{\rm c} \cap \mathcal{M} \neq \emptyset, \ \ \forall \, {\rm c} \in \{1, \cdots, \boldsymbol{{\rm c}}\} .
\end{align}
where $\mb{C}_{\rm x}$ and $\mb{C}_{\rm v}$ are diagonal elements of $\mb{C}$ in \eqref{eq:measurments} and $\mathcal{D}_{\rm c}$ denotes the set of nodes in ${\rm c}$-th connected component of $\Delta \mathcal{L}_{\mb{q}}$ as given in \eqref{eq:app_components_sets_1}.
\end{lemma}
\begin{IEEEproof}
We first show \eqref{eq:cond_suff_0} is invariant under permutation of $\Delta \mathcal{L}_{\mb{q}}$ which is introduced in \eqref{eq:graph_component_lap} and accordingly permutation of $[\mb{C}^{\top}_{\rm x}\; \mb{C}^{\top}_{\rm v}]^{\top}$. To this end, from the definition of nullspace we have
\begin{align}\label{eq:Lemma_kernel}
\mathrm{ker} \paren{\begin{bmatrix}
\mb{C}_{\rm x}\\
\mb{C}_{\rm v}
\end{bmatrix} \Delta\mathcal{L}_{\mb{q}}}
= \setdef{{x} \in \ensuremath{\mathbb{R}}^{N}}{ \begin{bmatrix}
\mb{C}_{\rm x}\\
\mb{C}_{\rm v}
\end{bmatrix} \Delta\mathcal{L}_{\mb{q}}x = 0},
\end{align}
from which we obtain either
\begin{align}\label{eq:former}
\Delta \mathcal{L}_{\mb{q}}{x} \notin \mathrm{Im}{(\Delta \mathcal{L}_{\mb{q}})} &\iff
\Delta \mathcal{L}_{\mb{q}}{x}=\boldsymbol{0},
\end{align}
%
or
\begin{align}\label{eq:latter}
\boldsymbol{0} \neq \boldsymbol{\mathcal{Y}} = \Delta \mathcal{L}_{\mb{q}}{x} \in \mathrm{Im}{(\Delta \mathcal{L}_{\mb{q}})} &\implies
\begin{bmatrix}
\mb{C}_{\rm x}\\
\mb{C}_{\rm v}
\end{bmatrix} \boldsymbol{\mathcal{Y}}=\boldsymbol{0},
\end{align}
where the latter, \eqref{eq:latter}, is in contradiction with condition \eqref{eq:cond_suff_0}.
Now under the permutation defined in \eqref{eq:graph_component_lap}, $\begin{bmatrix}
\mb{C}_{\rm x}\\
\mb{C}_{\rm v}
\end{bmatrix} \Delta\mathcal{L}_{\mb{q}}x = 0$ in \eqref{eq:Lemma_kernel} can be rewritten in block-partitioned diagonal form as
\begin{align}\label{eq:H_q_elements_stack_relabeled}
\begin{bmatrix}
\mb{C}_{\rm x}\\
\mb{C}_{\rm v}
\end{bmatrix}
{\rm P}^{\top}
\breve{\mathcal{L}}_\mb{q}
{\rm P} x
=
\begin{bmatrix}
\mb{C}_{\rm x}\\
\mb{C}_{\rm v}
\end{bmatrix}
{\rm P}^{\top}
\breve{\mathcal{L}}_\mb{q}
\chi
=
\begin{bmatrix}
\breve{\mb{C}}_{\rm x}\\
\breve{\mb{C}}_{\rm v}
\end{bmatrix}
\breve{\mathcal{L}}_\mb{q}
\chi
=
\boldsymbol{0},
\end{align}
in which $\chi={\rm P}x$ denotes the relabeled $x$ such that
\begin{align}\label{eq:app_states_relabeled}
\chi&=\mathrm{col}{(\chi_1, \dots , \chi_{\boldsymbol{\rm c}})} ={\rm P}x, \;\, \text{with}
\nonumber \\
\chi_{\rm c} &= \mathrm{col}{(x_i)},\ \ \forall \, i \in \mathcal{D}_{\rm c},\ \ \forall \, {\rm c} \in \{1, \cdots, \boldsymbol{{\rm c}}\}.
\end{align}
Also, $ \breve{\mb{C}}_{\rm k} = \mb{C}_{\rm k} {\rm P}^{\top} = \begin{bmatrix}\mb{C}^{\rm 1}_{\rm k} & \cdots & \mb{C}^{\boldsymbol{\rm c}}_{\rm k}\end{bmatrix}$, ${\rm k} \in \{\rm x, v\}$ is a block-partitioned binary matrix that specifies monitored agents of each component \footnote{Note that ${\rm P}^{\top}$ permutes the columns of binary matrix $\mb{C}_{\rm k}$ whose row-vector elements are $\mathfrak{e}^{\top}_i,\, \forall \, i \in \mathcal{M}_{\rm k},\, {\rm k} \in \{\rm x, v\}$.}.
To show the results in \eqref{eq:former} and \eqref{eq:latter} hold also for the transformed form in \eqref{eq:H_q_elements_stack_relabeled}, one need to verify the invariance of \eqref{eq:cond_suff_0} under the permutation by ${\rm P}$, that is
\begin{align}\label{eq:cond_suff_0_invariant}
\mathrm{Im} (\Delta \mathcal{L}_{\mb{q}}) \cap \mathrm{ker}{([\mb{C}^{\top}_{\rm x}\; \mb{C}^{\top}_{\rm v}]^{\top})} &=\emptyset
\iff
\nonumber \\
\mathrm{Im} (\breve{\mathcal{L}}_{\mb{q}}) \cap \mathrm{ker}{([\breve{\mb{C}}^{\top}_{\rm x}\; \breve{\mb{C}}^{\top}_{\rm v}]^{\top})} &=\emptyset.
\end{align}
To show this, from the range and nullspace definition, for subspaces in \eqref{eq:cond_suff_0_invariant} we have
\begin{align}\label{eq:app_Img_Lap}
\mathrm{Im}(\Delta \mathcal{L}_{\mb{q}})
&= \setdef{\boldsymbol{\mathcal{Y}} \in \ensuremath{\mathbb{R}}^{N}}{\boldsymbol{\mathcal{Y}} = \Delta \mathcal{L}_{\mb{q}} x(t)},
\\
\mathrm{ker} \paren{\begin{bmatrix}
\mb{C}_{\rm x}\\
\mb{C}_{\rm v}
\end{bmatrix}}
&=
\mathrm{ker}(
\mb{C}_{\rm x})
\cap
\mathrm{ker}(
\mb{C}_{\rm v})
\nonumber \\
&= \setdef{\boldsymbol{\mathcal{X}} \in \ensuremath{\mathbb{R}}^{N}}{ \mb{C}_{\rm x} \boldsymbol{\mathcal{X}} = 0,\, \mb{C}_{\rm v} \boldsymbol{\mathcal{X}} = 0 },
\end{align}
and
\begin{align}
\mathrm{Im}(\breve{\mathcal{L}}_{\mb{q}})
&=
\setdef{ \breve{\boldsymbol{\mathcal{Y}}} \in \ensuremath{\mathbb{R}}^{N}}{\breve{\boldsymbol{\mathcal{Y}}}
= \breve{\mathcal{L}}_{\mb{q}} \chi(t) = \breve{\mathcal{L}}_{\mb{q}}{\rm P}x(t)}
\nonumber \\
&=
\setdef{ \breve{\boldsymbol{\mathcal{Y}}} \in \ensuremath{\mathbb{R}}^{N}}{{\rm P}^{\top}\breve{\boldsymbol{\mathcal{Y}}}
= {\rm P}^{\top}\breve{\mathcal{L}}_{\mb{q}} {\rm P}x(t) = \boldsymbol{\mathcal{Y}} }
\nonumber \\
&= {\rm P}\, \mathrm{Im}{(\Delta \mathcal{L}_{\mb{q}})},
\end{align}
%
where we used \eqref{eq:graph_component_lap} and $\chi(t)={\rm P}x(t)$ as in \eqref{eq:app_states_relabeled} and \eqref{eq:app_Img_Lap}. Similarly,
%
\begin{align}
\mathrm{ker} \paren{\begin{bmatrix}
\breve{\mb{C}}_{\rm x}\\
\breve{\mb{C}}_{\rm v}
\end{bmatrix}}
&=
\mathrm{ker}(
\breve{\mb{C}}_{\rm x})
\cap
\mathrm{ker}(
\breve{\mb{C}}_{\rm v})
\nonumber \\
&= \setdef{\breve{\boldsymbol{\mathcal{X}}} \in \ensuremath{\mathbb{R}}^{N}}{ \breve{\mb{C}}_{\rm x} \breve{\boldsymbol{\mathcal{X}}} = 0,\, \breve{\mb{C}}_{\rm v} \breve{\boldsymbol{\mathcal{X}}} = 0 }
\nonumber \\
&= \setdef{\breve{\boldsymbol{\mathcal{X}}} \in \ensuremath{\mathbb{R}}^{N}}{ {\mb{C}}_{\rm x}{\rm P}^{\top} \breve{\boldsymbol{\mathcal{X}}} = 0,\, {\mb{C}}_{\rm v}{\rm P}^{\top} \breve{\boldsymbol{{\mathcal{X}}}} = 0 }
\nonumber \\
&= \setdef{\breve{\boldsymbol{\mathcal{X}}} \in \ensuremath{\mathbb{R}}^{N}}{ {\mb{C}}_{\rm x}{\boldsymbol{\mathcal{X}}} = 0,\, {\mb{C}}_{\rm v}{\boldsymbol{{\mathcal{X}}}} = 0,\, {\rm P} \boldsymbol{\mathcal{X}}=\breve{\boldsymbol{\mathcal{X}}}}
\nonumber \\
&= {\rm P}\, \mathrm{ker} \paren{\begin{bmatrix}
\mb{C}_{\rm x}\\
\mb{C}_{\rm v}
\end{bmatrix}}.
\end{align}
Then
\begin{align}
\mathrm{Im}(\breve{\mathcal{L}}_{\mb{q}})
\cap
\mathrm{ker} \paren{\begin{bmatrix}
\breve{\mb{C}}_{\rm x}\\
\breve{\mb{C}}_{\rm v}
\end{bmatrix}}
&=
{\rm P}\, \mathrm{Im}{(\Delta \mathcal{L}_{\mb{q}})}
\cap
{\rm P}\, \mathrm{ker} \paren{\begin{bmatrix}
\mb{C}_{\rm x}\\
\mb{C}_{\rm v}
\end{bmatrix}}
\nonumber \\
&=
{\rm P} \paren{ \mathrm{Im}{(\Delta \mathcal{L}_{\mb{q}})}
\cap
\mathrm{ker} \paren{\begin{bmatrix}
\mb{C}_{\rm x}\\
\mb{C}_{\rm v}
\end{bmatrix}}}
\nonumber \\
&=
{\rm P} \paren{ \emptyset} = \emptyset.
\end{align}
where we used fact 2.9.29 in \cite{bernstein2009matrix} and condition \ref{cond:detection_condition_suff_0}.
Now one can proof \eqref{eq:app_component_measurements} by contradiction. Assume \eqref{eq:app_component_measurements} does not hold, that is $\exists \, {\rm c'} \in \{1, \cdots, \boldsymbol{{\rm c}}\},$ s.t. $\mathcal{D}_{\rm c'} \cap \mathcal{M} = \emptyset$, under which
we have the ${\rm c'}$-th block in \eqref{eq:H_q_elements_stack_relabeled} such that
\begin{align}\label{eq:app_contra}
\begin{bmatrix}
\breve{\mb{C}}^{\rm c'}_{\rm x}\\
\breve{\mb{C}}^{\rm c'}_{\rm v}
\end{bmatrix}
\Delta \mathcal{L}_{\mb{q}}(\mathcal{D}_{\rm c'})
\chi_{\rm c'}(t) = \boldsymbol{0}, \ \ \breve{\mb{C}}^{\rm c'}_{\rm x} = \breve{\mb{C}}^{\rm c'}_{\rm x} = 0,
\end{align}
which holds for all $\chi_{\rm c'}(t)$ with $\Delta \mathcal{L}_{\mb{q}}(\mathcal{D}_{\rm c'})
\chi_{\rm c'}(t) \in \mathrm{Im}{(\Delta \mathcal{L}_{\mb{q}}(\mathcal{D}_{\rm c'})) } \subseteq \mathrm{Im}{(\breve{\mathcal{L}}_{\mb{q}}) }$ as in \eqref{eq:app_contra} $\mathrm{Im}{(\Delta \mathcal{L}_{\mb{q}}(\mathcal{D}_{\rm c'})) } \in \mathrm{ker}{\paren{\begin{bmatrix}
\breve{\mb{C}}^{\rm c'}_{\rm x}\\
\breve{\mb{C}}^{\rm c'}_{\rm v}
\end{bmatrix}}}
\implies
\mathrm{Im}{(\breve{\mathcal{L}}_{\mb{q}} )} \cap \mathrm{ker}{([\breve{\mb{C}}^{\top}_{\rm x}\; \breve{\mb{C}}^{\top}_{\rm v}]^{\top})} \neq \emptyset$
that contradicts \eqref{eq:cond_suff_0_invariant}.
\end{IEEEproof}
\section{Proof of Lemma \ref{lemma:Inv_unobs_subspace}}\label{app_Inv_unobs_subspace}
Note that the Laplacian matrix $\mathcal{L}_{\sigma(t)}$ of every connected undirected (or strongly connected and balanced directed) graph has only one zero eigenvalue, $\lambda = 0$, with the corresponding eigenvector $\boldsymbol{1}_N$ such that $\mathcal{L}_{\sigma(t)}\boldsymbol{1}_{N} = \boldsymbol{0}$ \cite{olfati2004consensus}. Then, given the structure of $\mb{A}_{\sigma(t)}$ in \eqref{eq:cl_sys},
$(\lambda = 0, w_r =\begin{bsmallmatrix} {\scriptscriptstyle (1/\sqrt{N})}\boldsymbol{1}_{\scriptscriptstyle \! N} \\ \boldsymbol{0}_{\!N} \end{bsmallmatrix})$ is an eigenpair of system matrix $\mb{A}_{\sigma(t)}$ associated with that of Laplacian $\mathcal{L}_{\sigma(t)}$ with $\sigma(t_{k-1})=\mb{q} \in \mathcal{Q}, \, \, t\in [t_{k-1}, t_k)$. Also,
it can be verified that the eigenpair $(\lambda = 0, w_r)$ lies in the unobservable subspace of system \eqref{eq:cl_sys} as it is a nontrivial solution to the PBH test for observability:
\begin{align}\label{eq:PBH_test}
\begin{bmatrix}
\lambda I_{} -\mb{A}_{\mb{q}} \\ \mb{C}_{}
\end{bmatrix}
%
w_r
%
&=
%
\boldsymbol{0} , \quad \, \lambda = 0 \in \ensuremath{\mathbb{C}},\\
\mb{C}_{} &= \mathrm{diag}\braces{0, C_{\rm v}}.
\end{align}
Therefore, one can conclude that the right eigenvector $w_r$ contained in $\mathrm{ker}(\mb{C}) $ belongs to $\mathrm{ker}(\mathcal{O}_k)$ that is defined in \eqref{eq:obsv_nullspace_matrix} \cite[Th. 15.8]{hespanha2018linear}.
Furthermore, as $(\lambda = 0, w_r)$ is the eigenpair associated with the equilibrium subspace \eqref{eq:cond_consensus} of every $\mb{A}_{\mb{q}}$ with Laplacian $\mathcal{L}_{\mb{q}}$, it is straightforward from Lemma \ref{lemma:observability} that $ \mathrm{span} \braces{w_r} = \mathrm{span} \braces{ \begin{bsmallmatrix} \boldsymbol{1}_{\! N} \\ \boldsymbol{0}_{\!N} \end{bsmallmatrix}} \subseteq \boldsymbol{\mathcal{N}}_1^{\infty} = \mathrm{ker}(\boldsymbol{\mathcal{O}}) $ over $t \in [t_0, +\infty)$.
\section{Proof of Proposition~\ref{prop:attack_detectability_local}}\label{app_attack_detectability_local}
Let $\sigma(t)=\mb{q} \in \mathcal{Q}, \, \, t\in [t_{k-1}, t_k) $ and consider the error dynamics of local observers in \eqref{eq:obs_decent_error}. According to Definition \ref{def:inv_zeros}, a ZDA for \eqref{eq:obs_decent_error} should satisfy
\begin{equation}\label{eq:obs_decent_INV_zero_1}
\begin{bmatrix}
\lambda_0 I_{} - {\mb{F}}^{{\rm i}}_{\mb{q}}
& -\mb{T}_{}^{\rm i} \mb{B}^{\rm i} \\ \mb{C}_{{\rm i}_i} & 0
\end{bmatrix}
%
\begin{bmatrix}
\tilde{\mb{e}}_{\rm i}(0) \\ \mb{u}_{0_{\rm i}}
\end{bmatrix}
%
=
%
\begin{bmatrix}
\boldsymbol{0} \\ \boldsymbol{0}
\end{bmatrix},
\end{equation}
where $\tilde{\mb{e}}_{\rm i}(0) := \mb{e}_{\rm i}(0)-\bar{\mb{e}}_{\rm i}(0)=\tilde{\mb{x}}_{0_{\rm i}}$.
Also, by considering \eqref{eq:uio_conds_2} and the fact that $ \mb{C}_{{\rm i}_i} \tilde{\mb{e}}_{\rm i}(0)=\mb{C}_{{\rm i}_i} \tilde{\mb{x}}_{\rm i}(0)=\boldsymbol{0}$ in the second equation of \eqref{eq:obs_decent_INV_zero_1}, matrix pencil \eqref{eq:obs_decent_INV_zero_1} can be rewritten as
\begin{equation}\label{eq:obs_decent_INV_zero_2}
\small
\underbrace{
\begin{bmatrix}
\lambda_0 I_{} - {\bar{\mb{A}}^{\rm i}_{\mb{q}} }
& - \mb{T}_{}^{\rm i}\mb{B}^{\rm i} \\ \mb{C}_{{\rm i}_i} & 0
\end{bmatrix} }_{\bar{\mb{P}}}
%
\begin{bmatrix}
\tilde{\mb{x}}_{\rm i}(0) \\ \mb{u}_{0_{\rm i}}
\end{bmatrix}
%
=
%
\begin{bmatrix}
\boldsymbol{0} \\ \boldsymbol{0}
\end{bmatrix}.
\end{equation}
It is immediate from Definition \ref{def:inv_zeros} that stealthy attack $\mb{u}_{a_{\rm i}}$ in \eqref{eq:obs_decent_error}, in the both ZDA and covert attack, looses its stealthiness with respect to the local residual $\mb{r}_{{\rm i}_i}$ if, and only if, there is no non-trivial zeroing direction associated with matrix pencil in \eqref{eq:obs_decent_INV_zero_1} or equivalently $\bar{\mb{P}}$ in \eqref{eq:obs_decent_INV_zero_2}, which in turn implies $\bar{\mb{P}}$ has full rank.
Moreover, from Definition \ref{def:inv_zeros} and condition \eqref{eq:E_definition}, it is straightforward that matrix pencil $\mb{P}$, defined in \eqref{eq:pencil_uio}, is associated with the zeroing direction of the local system \eqref{eq:cluster_dyn2}.
We now show how conditions ({i})-({iii}) establish the equivalence of rank sufficiency for $\mb{P}$ in \eqref{eq:pencil_uio} and $\bar{\mb{P}}$ in \eqref{eq:obs_decent_INV_zero_2}.
Given $\mb{P}$ in \eqref{eq:pencil_uio}, one can write
\begin{align}\label{eq:pencil_uio_pre}
%
& \begin{bmatrix}
I_{} - {\mb{h}}^{\rm i}_{} \mb{C}_{{\rm i}_i} & \lambda_0 {\mb{h}}^{\rm i}_{}
\\
0 & I_{} \\
{\mb{h}}^{\rm i}_{} \mb{C}_{{\rm i}_i} & - \lambda_0 {\mb{h}}^{\rm i}_{}
\end{bmatrix}
%
\mb{P} = \nonumber \\
%
&
\begin{bmatrix}
\lambda_0 I_{} -\bar{\mb{A}}^{\rm i}_{\mb{q}} & -(I_{} - {\mb{h}}^{\rm i}_{} \mb{C}_{{\rm i}_i}){\mb{B}}^{\rm i} & 0
\\
\mb{C}_{{\rm i}_i} & 0 & 0 \\
- {\mb{h}}^{\rm i}_{} \mb{C}_{{\rm i}_i} {\mb{A}}^{\rm i}_{\mb{q}} & {\mb{h}}^{\rm i}_{} \mb{C}_{{\rm i}_i}{\mb{B}}^{\rm i} & {\mb{E}}^{\rm i}_{}
\end{bmatrix},
\end{align}
where ${\mb{h}}^{\rm i}_{} := {\mb{E}}^{\rm i}_{} (\mb{C}_{{\rm i}_i}{\mb{E}}^{\rm i}_{})^{\dagger}$ is a solution to \eqref{eq:uio_conds_1} that exists under condition ({ii}) \cite[Lemma 1]{chen1996design}. Then, postmultiplying \eqref{eq:pencil_uio_pre} by
\begin{align}\label{eq:pencil_uio_post}
\begin{bmatrix}
I_{} & 0 & 0
\\
0 & I_{} & 0 \\
(\mb{C}_{{\rm i}_i}{\mb{E}}^{\rm i}_{})^{\dagger}\mb{C}_{{\rm i}_i}\mb{A}^{\rm i}_{\mb{q}} & (\mb{C}_{{\rm i}_i}{\mb{E}}^{\rm i}_{})^{\dagger}\mb{C}_{{\rm i}_i}\mb{B}^{\rm i} & I_{}
\end{bmatrix},
\end{align}
and considering \eqref{eq:uio_conds_2} yields
\begin{align}
\begin{bmatrix}
\lambda_0 I_{} -\bar{\mb{A}}^{\rm i}_{\mb{q}} & -{\mb{T}}^{\rm i}_{} {\mb{B}}^{\rm i} & 0
\\
\mb{C}_{{\rm i}_i} & 0 & 0 \\
0 & 0 & {\mb{E}}^{\rm i}_{}
\end{bmatrix}.
\end{align}
Since node $i \in \mathcal{P}_{\rm i}$ is $\mb{k}$-connected, we have $|\mathcal{N}_{i}|=\mb{k}$ and $\mb{k}\leq \mathrm{rank} \paren{ \mb{C}_{{\rm i}_i} } \leq 2\mb{k}$ (cf. \eqref{eq:measurments}). Then, from condition ({i}), one can verify that $\mathrm{rank} \paren{ \mb{C}_{{\rm i}_i} } \geq \mathrm{rank} \paren{ \mb{B}^{\rm i} } + \mathrm{rank} \paren{ \mb{E}^{\rm i}_{} }$ guarantees \eqref{eq:pencil_uio} is a tall or square matrix pencil having only a finite number\footnote{This condition is not valid for degenerate systems which are out of scope of this work.} of output-zeroing directions \cite[Ch. 2]{lee2017l1}. Also, the pre- and post-multiplied matrices in \eqref{eq:pencil_uio_pre} and \eqref{eq:pencil_uio_post} are full column rank. Therefore, we have %
\begin{align}\label{eq:pencil_equivalance}
\mathrm{rank} \paren{ \mb{P}}
=
\mathrm{rank}
\underbrace{
\begin{bmatrix}
\lambda_0 I_{} -\bar{\mb{A}}^{\rm i}_{\mb{q}} & -{\mb{T}}^{\rm i}_{} {\mb{B}}^{\rm i}_{}
\\
\mb{C}_{{\rm i}_i} & 0
\end{bmatrix}}_{\bar{\mb{P}}}
%
+
\mathrm{rank} \paren{ {\mb{E}}^{\rm i}_{}}.
\end{align}
Recall ${\mb{E}}^{\rm i}_{}$ is full column rank, and hence ${\mb{P}}$ in \eqref{eq:pencil_uio} is full rank if, and only if, $\bar{\mb{P}}$ in \eqref{eq:pencil_equivalance} is full rank. This guarantees that a locally undetectable stealthy attack is impossible.
\section{Proof of Theorem~\ref{th:Attack_detectability_switching}}\label{app_Attack_detectability_switching}
Consider \eqref{eq:obs_cent_error} over $ t\in [t_0,+\infty)$, and let the safe mode $ \sigma(t) \!=\! \mb{q} \! \in \! \mathcal{Q},$ $ t \in [t_1, +\infty)$ the continuous system residual $\mb{r}_0(t)$ and its successive derivatives can be rewritten as
\begin{align}\label{eq:Y}
\mb{R} &=
\mb{\mathcal{O}}_1 \mb{e}(t) - \boldsymbol{\mathcal{H}}_{}( \mb{H}\mb{C}) \mb{E}
+
\boldsymbol{\mathcal{H}}(\mb{B}) \mb{U}_a
+
\boldsymbol{\mathcal{H}}_{}( \mb{H}) \mb{U}_s
%
\nonumber \\
&
\qquad \qquad \qquad \qquad \quad \ \ \:{-}\:
\mb{U}_s
+
\boldsymbol{\mathcal{H}}_{}( \Delta \mb{A}_{\mb{q}}) \mb{X},
\end{align}
where
\begin{align}\label{eq:output_sequence}
\mb{R} &=
\begin{bmatrix}
\mb{r}^{\top}_0(t)& \dot{\mb{r}}^{\top}_0(t) & \cdots & {(\mb{r}_0^{\top}(t))^{({\rm d})}}
\end{bmatrix}^{\top},\\
\mb{U}_{\jmath} &=
\begin{bmatrix}
\mb{u}^{\top}_{{\jmath}}(t) & \dot{\mb{u}}^{\top}_{{\jmath}}(t) & \cdots & ({\mb{u}^{\top}_{{\jmath}}(t))^{({\rm d})}}
\end{bmatrix}^{\top},\\ \label{eq:output_sequence_E}
\mb{E} &=
\begin{bmatrix}
\mb{e}^{\top}(t) & \dot{\mb{e}}^{\top}(t) & \cdots & {(\mb{e}^{\top}(t))^{({\rm d})}}
\end{bmatrix}^{\top},\\ \label{eq:output_sequence_X}
\mb{X} &=
\begin{bmatrix}
\mb{x}^{\top}(t) & \dot{\mb{x}}^{\top}(t) & \cdots & {(\mb{x}^{\top}(t))^{({\rm d})}}
\end{bmatrix}^{\top},\\ \label{eq:H_q}
\boldsymbol{\mathcal{H}}_{}({b})&=
\begin{bmatrix}
0 & 0 & 0 & \cdots & 0 \\
\mb{C} {b} & 0 & 0 & \cdots& 0 \\
\mb{C}{\mb{A}}_{1} {b} & \mb{C} {b} & 0 & \cdots& 0 \\
\vdots & \vdots & \ddots & \ddots & \vdots\\
\mb{C}{\mb{A}}_{1}^{{\rm d}} {b} & \mb{C}{\mb{A}}_{1}^{{\rm d}-1} {b} & \cdots & \mb{C} {b} & 0
\end{bmatrix},
\end{align}
with $ {\jmath}\in \braces{a,s}$, $ {b}\in\braces{\mb{B},\mb{H}\mb{C},\mb{H},\Delta \mb{A}_{\mb{q}}}$, $\Delta \mb{A}_{\mb{q}} = (\mb{A}_{\mb{q}}-\mb{A}_{1}) $ and ${\rm d} \in \ensuremath{\mathbb{N}} \setminus \{1,2\}$.
From \eqref{eq:stealthy} in Proposition \ref{propo:stealthy_attacks} and \eqref{eq:est_error_normal}-\eqref{eq:est_error}, it can be easily verified that \eqref{eq:Y} is simplified to $\mb{R}=\mb{\mathcal{O}}_1 \bar{\mb{e}}(t) - \boldsymbol{\mathcal{H}}_{}( \mb{H}\mb{C}) \bar{\mb{E}} + \boldsymbol{\mathcal{H}}_{}( \Delta \mb{A}_{\mb{q}}) \mb{X}$ where $ \bar{\mb{E}}$ has the same form as \eqref{eq:output_sequence_E} while whose elements are $\bar{\mb{e}}$ and its derivatives. Therefore, in a stealthy attack case $\lim_{t_1 \rightarrow \infty} \mb{R} = \boldsymbol{0}$ during normal mode over $ t\in [t_0,t_1)$. The objective is to characterize the effect of switching communication, modeled as discrepancy $\Delta \mb{A}_{\mb{q}}$ in \eqref{eq:obs_cent_error} and \eqref{eq:Y}, on the stealthiness of attacks in the residual $\mb{r}_0(t)$ of centralized observer \eqref{eq:obs_cent} during safe mode over $ t\in [t_1, +\infty) $ (cf. Problem \ref{prob:attack_detection}).
Given the input-output matrix \eqref{eq:H_q} for the switching perturbations $\Delta \mb{A}_{\mb{q}}$ in \eqref{eq:Y}, note that $\boldsymbol{\mathcal{H}}_{}( \Delta \mb{A}_{\mb{q}}) \mb{X} = \boldsymbol{0} $ over $ t\in [t_1,+\infty)$ in \eqref{eq:Y} is
the necessary condition under which the stealthy attacks, modeled in \eqref{eq:attack_model}, remain undetectable in the residual $\mb{r}_0(t)$ of \eqref{eq:obs_cent_error}, regardless of the perturbation $\Delta \mb{A}_{\mb{q}} \mb{x}$ caused by topology switching. Therefore, $\boldsymbol{\mathcal{H}}_{}( \Delta \mb{A}_{\mb{q}}) \mb{X} \neq \boldsymbol{0} $ in \eqref{eq:Y} implying the system switching $\Delta \mb{A}_{\mb{q}}$ affects $\mb{R}(t),\, t\in[t_1,+\infty)$ in \eqref{eq:Y} guarantees attack detectability in $\mb{r}_0(t)$.
Consider Markov parameters $\mb{C}{\mb{A}}_{1}^{\rm{d}}\Delta \mb{A}_{\mb{q}},$ $ \rm{d} \in \ensuremath{\mathbb{N}}_{0}$ in \eqref{eq:H_q}, the term $\boldsymbol{\mathcal{H}}_{}( \Delta \mb{A}_{\mb{q}}) \mb{X} $ in \eqref{eq:Y} can be rewritten as
\begin{align}\label{eq:H_q_elements}
\sum\limits_{ l = 0}^{\rm d}
\mb{C}{\mb{A}}_{1}^{{\rm{d}}}\Delta \mb{A}_{\mb{q}}\mb{x}^{({\rm d}-l)}(t) = \boldsymbol{0}, \ \ \forall \, {\rm d} \in \ensuremath{\mathbb{N}}_{0}.
\end{align}
We show that under condition \ref{cond:detection_condition_suff_0}, the first two terms in \eqref{eq:H_q_elements} are non-zero (and so is $\boldsymbol{\mathcal{H}}_{}( \Delta \mb{A}_{\mb{q}}) \mb{X} \neq \boldsymbol{0} $) unless $\Delta \mb{A}_{\mb{q}}\mb{x}(t)=\boldsymbol{0},\ \ \forall \,t \in [t_1, +\infty)$.
By setting $\rm d = 0, 1$, and expanding \eqref{eq:H_q_elements} we obtain
\begin{align} \label{eq:H_q_elements_d0}
{\rm d}=0
\overset{\text{\eqref{eq:H_q_elements}}}{\Rightarrow} \ \
\mb{C}_{\rm v}\Delta\mathcal{L}_{\mb{q}}x(t)&=\boldsymbol{0},
\ \ \forall \,t \in [t_1, +\infty),
\\ \label{eq:H_q_elements_d1}
{\rm d}=1
\overset{\text{\eqref{eq:H_q_elements}}}{\Rightarrow} \ \
\mb{C}_{\rm v}\Delta\mathcal{L}_{\mb{q}}v(t)&=\boldsymbol{0}, \ \ \text{and,} \nonumber \\
\mb{C}_{\rm x}\Delta\mathcal{L}_{\mb{q}}x(t)&=\boldsymbol{0},
\ \ \forall \,t \in [t_1, +\infty),
\end{align}
where $\mb{C}_{\rm x}$ and $\mb{C}_{\rm v}$ are diagonal elements of $\mb{C}$ as given in \eqref{eq:cl_sys}-\eqref{eq:measurments}, $\Delta \mathcal{L}_{\mb{q}} = \mathcal{L}_{\mb{q}} - \mathcal{L}_{1}$ is the non-zero submatrix of $\Delta \mb{A}_{\mb{q}} = (\mb{A}_{\sigma(t)}-\mb{A}_{1})= \begin{bsmallmatrix}
0_{} & 0 \\
- \alpha \Delta \mathcal{L}_{\mb{q}} & 0
\end{bsmallmatrix}$, and ${\mb{x}}(t) = \mathrm{col}({x}(t),{v}(t))$ as in \eqref{eq:cl_sys}.
Then, using \eqref{eq:H_q_elements_d0} and \eqref{eq:H_q_elements_d1}, we have
\begin{align}\label{eq:H_q_elements_stack}
\begin{bmatrix}
\mb{C}_{\rm x}\\
\mb{C}_{\rm v}
\end{bmatrix}
\Delta\mathcal{L}_{\mb{q}}x(t)=\boldsymbol{0}, \ \ \forall \,t \in [t_1, +\infty).
\end{align}
Under condition \ref{cond:detection_condition_suff_0}, one can verify that \eqref{eq:H_q_elements_stack} implies
\begin{align}\label{eq:H_q_elements_kernel}
\Delta \mathcal{L}_{\mb{q}}{x}(t) \notin \mathrm{Im}{(\Delta \mathcal{L}_{\mb{q}})} \iff
\Delta \mathcal{L}_{\mb{q}}{x}(t)=\boldsymbol{0}, \ \ \forall \, t \in [t_1, +\infty).
\end{align}
otherwise, for any $x(t)$ such that $ \boldsymbol{0} \neq \Delta \mathcal{L}_{\mb{q}}{x}(t)= \boldsymbol{\mathcal{Y}} \in \mathrm{Im}{(\Delta \mathcal{L}_{\mb{q}})} $, we obtain $[\mb{C}^{\top}_{\rm x}\; \mb{C}^{\top}_{\rm v}]^{\top} \boldsymbol{\mathcal{Y}} = 0,\, \boldsymbol{\mathcal{Y}} \in \mathrm{ker}{([\mb{C}^{\top}_{\rm x}\; \mb{C}^{\top}_{\rm v}]^{\top})} $ for \eqref{eq:H_q_elements_stack}, which is in contradiction with condition \ref{cond:detection_condition_suff_0}.
Now considering the consensus protocol \eqref{eq:ctrl_proto}, it can be verified that $\Delta \mathcal{L}_{\mb{q}}$ (or equivalently $\Delta \mb{A}_{\mb{q}}$ in \eqref{eq:H_q_elements}), encodes connected graph component(s) corresponding to added/removed communication link(s) in the communication network (cf. Definition \ref{def:graph_component} and \ref{prop:app_def_diagonal}). Then an elementary transformation, by means of the permutation matrix ${\rm P}$ as defined in \eqref{eq:graph_component_lap}, transforms \eqref{eq:H_q_elements_kernel} into block-diagonal form as
\begin{align}\label{eq:H_q_elements_kernel_relabeled}
\Delta \mathcal{L}_{\mb{q}}{x}(t)
=
\boldsymbol{0}
\iff
\breve{\mathcal{L}}_{\mb{q}}\chi(t)
=
\boldsymbol{0},
\ \ \forall \, t \in [t_1, +\infty),
\end{align}
where the block-diagonal $\breve{\mathcal{L}}_{\mb{q}}$ is given in \eqref{eq:graph_component_lap} and $\chi(t)={\rm P}x(t)$ denotes the relabeled system states such that
\begin{align}\label{eq:states_relabeled}
\chi(t)&=\mathrm{col}{(\chi_1(t), \dots , \chi_{\boldsymbol{\rm c}}(t))} ={\rm P}x(t), \;\, \text{with}
\nonumber \\
\chi_{\rm c}(t) &= \mathrm{col}{(x_i(t))},\ \ \forall \, i \in \mathcal{D}_{\rm c},\ \ \forall \, {\rm c} \in \{1, \cdots, \boldsymbol{{\rm c}}\},
\end{align}
with $\mathcal{D}_{\rm c}$ being the set of nodes (agents involved in switching links) in ${\rm c}$-th connected component\footnote{Although the analysis here is at the global level, it is worth mentioning that $\Delta \mathcal{L}_{\mb{q}}$ at cluster levels i.e. $\mathcal{P}_{\rm i},\, {\rm i} \in \braces{1,\cdots,|\boldsymbol{\mathcal{P}}|}$ may have more than one connected component.} as in \eqref{eq:app_components_sets_1}.
Also, note that permutation matrix ${\rm P}$ is a binary nonsingular matrix by definition, and that the Laplacian matrix is zero row sum matrix and, if connected, its nullspace is spanned by $\boldsymbol{1}$, a vector of all ones \cite{olfati2004consensus}. Therefore, from \eqref{eq:H_q_elements_kernel_relabeled} and for nodes participated in (connected) switching links, i.e. $ \forall \, i,j \in \mathcal{D}_{\rm c},\, i\neq j$, one can conclude that
\begin{align}\label{eq:H_q_elements_kernel_1}
{x}_i^{}(t) - {x}_j^{}(t) &= 0 \Leftrightarrow
\nonumber \\
{x}_i^{}(t) &= {x}_j^{}(t), &\forall& \, i, j \in \mathcal{D}_{\rm c}, \ \ \forall\, {\rm c} \in \{1, \cdots, \boldsymbol{\rm c}\}, \nonumber \\
& &\forall& \, t \in [t_1, +\infty),
\end{align}
which by considering the continuity of the system states can be extended for its higher-order time derivatives and be rewritten as
\begin{align}\label{eq:H_q_elements_kernel_2_2}
(\mathfrak{e}^{\top}_{i}-\mathfrak{e}^{\top}_{j}) {x}^{(\rm m)}(t) = 0, \ \ &\forall \, i, j \in \mathcal{D}_{\rm c}, \ \
\forall \,{\rm c} \in \{1, \cdots, \boldsymbol{\rm c}\}, \nonumber \\
&\forall \, {\rm m} \in \ensuremath{\mathbb{N}}_0, \ \ \forall \,t \in [t_1, +\infty),
\end{align}
with $\mathfrak{e}_{i},\, \mathfrak{e}_{j}$ being $i$-th and $j$-th standard-basis vectors in $\ensuremath{\mathbb{R}}^{N}$.
Also, from \eqref{eq:H_q_elements}, \eqref{eq:H_q_elements_kernel_relabeled}, \eqref{eq:H_q_elements_kernel_2_2} and by considering the structure $\mb{A}_{\sigma(t)}$ and system state $\mb{x}(t) = \mathrm{col}(x(t),v(t))$ in \eqref{eq:cl_sys}, we obtain
\begin{align}\label{eq:H_q_elements_2}
\Delta \mb{A}_{\mb{q}}\mb{x}^{(\rm m)}(t) &= \boldsymbol{0}
\Leftrightarrow \nonumber \\
\Delta \mathcal{L}_{\mb{q}}{x}^{(\rm m)}(t) &= \boldsymbol{0}, \ \ \forall \, {\rm m}\in \ensuremath{\mathbb{N}}_0,\, \forall \, t \in [t_1, +\infty).
\end{align}
Therefore, under condition \ref{cond:detection_condition_suff_0}, one can conclude that unless \eqref{eq:H_q_elements_kernel}/\eqref{eq:H_q_elements_kernel_2_2} holds that is the system states (positions $x_i(t)$, $x_j(t)$ and their successive derivatives) of all agents within each graph component, i.e. agents involved in connected intra-cluster switching links, are respectively identical $\forall \, t \in [t_1,
+\infty)$, the left side of \eqref{eq:H_q_elements_d0} and \eqref{eq:H_q_elements_d1} is non-zero and so is \eqref{eq:H_q_elements}, implying $\Delta \mb{A}_{\mb{q}}$ affects $\mb{R}(t)$ whereby the attacks are detectable.
We now show under conditions \ref{cond:detection_condition_suff_1}-\ref{cond:detection_condition_suff_2} the domain of existence of \eqref{eq:H_q_elements_kernel} is shrank into the only case that the entire system states, except for those affected by stealthy attacks, are at an equilibrium.
Zero-dynamics attack (ZDA) case:
it can be shown that under condition \ref{cond:detection_condition_suff_0}, \eqref{eq:H_q_elements_kernel} holds (and so does \eqref{eq:H_q_elements_kernel_2_2}) only in the worst-case scenario, in the sense of attack detection, that none of the agents involved in intra-cluster switching links are affected by the ZDA in a safe mode.
To this end, consider \eqref{eq:H_q_elements_kernel_relabeled} under which ZDA remains stealthy in residual $\mb{r}_0(t)$ in the safe modes and recall
\begin{align}\label{eq:ZDA_traj}
\mb{x}(t)=\bar{\mb{x}}(t)+ \tilde{\mb{x}}(t), \ \ \tilde{\mb{x}}_0e^{ \lambda_0 t},\ \ \forall\, t \in [t_0, +\infty),
\end{align}
in a stealthy ZDA case with $\tilde{\mb{x}}_0e^{ \lambda_0 t_1} \in \mathrm{ker}(\mb{C})$ being the initial condition of ZDA (cf. \eqref{eq:inv_zeros}, and \eqref{eq:zda_cond} in Proposition \ref{propo:stealthy_attacks}) at $t=t_1$ for a safe mode.
Similar to \eqref{eq:obs_cent_INV_zero}, by evaluating ZDA condition \eqref{eq:inv_zeros} for the tuple $(\mb{A}_{\mb{q}}, \mb{B}, \mb{C})$ with $\mb{A}_{\mb{q}} = (\mb{A}_{1}+\Delta \mb{A}_{\mb{q}})$ and considering \eqref{eq:H_q_elements_2} we obtain
\begin{equation}\label{eq:obs_cent_INV_zero_2}
\begin{bmatrix}
\lambda_0 I_{} - (\mb{A}_{1}-\mb{H}\mb{C})
&
(\mb{A}_{\mb{q}}-\mb{A}_{1})
&
-\mb{B}
\\
\mb{C} & 0_{} & 0_{}
\end{bmatrix}
%
\begin{bmatrix}
\tilde{\mb{e}}(t_1) \\ \tilde{\mb{x}}(t_1) \\ \mb{u}_a(t_1)
\end{bmatrix}
%
=
%
\begin{bmatrix}
\boldsymbol{0} \\ \boldsymbol{0} \\ \boldsymbol{0}
\end{bmatrix},
\end{equation}
where as in \eqref{eq:obs_cent_INV_zero}, $\tilde{\mb{e}}(t_1) = \tilde{\mb{x}}(t_1) $ with $\tilde{\mb{x}}(t_1)=\tilde{\mb{x}}_0e^{ \lambda_0 t_1}$ and $ \mb{u}_a(t_1) = {\mb{u}}_0e^{ \lambda_0 t_1}$. Then \eqref{eq:obs_cent_INV_zero_2} is simplified to
\begin{equation}\label{eq:obs_cent_INV_zero_3}
\begin{bmatrix}
\lambda_0 I_{} - (\mb{A}_{\mb{q}}-\mb{H}\mb{C})
& -\mb{B} \\ \mb{C} & 0_{}
\end{bmatrix}
%
\begin{bmatrix}
\tilde{\mb{x}}(0) \\ \mb{u}_0
\end{bmatrix}
%
=
%
\begin{bmatrix}
\boldsymbol{0} \\ \boldsymbol{0}
\end{bmatrix},
\end{equation}
where further simplification, similar to that in \eqref{eq:obs_cent_INV_zero}, and expanding it out yields
\begin{align}\label{eq:}
\begin{bmatrix}
\lambda_0 I_{N} & -I_{N} & 0\\
\alpha (\mathcal{L}_{1} + \Delta \mathcal{L}_{\mb{q}}) & (\lambda_0+\gamma) I_{N} & -I_{\bar{\mathcal{F}}} \\
\mb{C}_{\rm x}& 0 & 0_{}\\
0 & \mb{C}_{\rm v} & 0
\end{bmatrix}
%
\begin{bmatrix}
\tilde{x}(t_0) \\ \tilde{v}(t_0) \\ \mb{u}_0
\end{bmatrix}
%
=
%
\begin{bmatrix}
\boldsymbol{0} \\ \boldsymbol{0} \\ \boldsymbol{0}
\end{bmatrix},
\end{align}
from which and also from \eqref{eq:zda_cond} we have
\begin{align}\label{eq:ZDA_ex_1}
\lambda_0 \tilde{x}_i(t_0) &= \tilde{v}_i(t_0),\ \ \forall\, i \in \mathcal{V},
\\ \alpha \mathcal{L}_{1} \tilde{x}(t_0) + (\lambda_0+\gamma) \tilde{v}(t_0)- I_{\bar{\mathcal{F}}} \mb{u}_0 &\overset{\eqref{eq:zda_cond}}{=} \boldsymbol{0},
\\ \label{eq:ZDA_ex_2}
\Delta \mathcal{L}_{\mb{q}} \tilde{x}(t_0) &= \boldsymbol{0},
\\ \label{eq:ZDA_ex_3}
\mb{C}_{\rm x}\tilde{x}(t_0)=\boldsymbol{0},
\ \
\mb{C}_{\rm v}\tilde{v}(t_0)&=\boldsymbol{0}.
\end{align}
Then one can conclude from \eqref{eq:measurments}, \eqref{eq:ZDA_traj}, \eqref{eq:ZDA_ex_1}, and \eqref{eq:ZDA_ex_3} that
\begin{align}\label{eq:ZDA_ex_4}
\tilde{x}_i(t_0)=\tilde{v}_i(t_0) = 0 \implies \tilde{x}_i(t)=\tilde{v}_i(t) = 0 , \ \ \forall \, i \in \mathcal{M} \subset \mathcal{V},
\end{align}
and by applying the same permutation as defined in \eqref{eq:graph_component_lap} and used in \eqref{eq:H_q_elements_kernel_relabeled} to equation \eqref{eq:ZDA_ex_2} as well as by considering \eqref{eq:ZDA_traj} and \eqref{eq:ZDA_ex_1} that
\begin{align}\label{eq:ZDA_ex_5}
\tilde{x}_i(t_0)&=\tilde{x}_j(t_0)
\implies
\tilde{x}_i(t)=\tilde{x}_j(t),
\ \ \forall \, i,j \in \mathcal{D}_{\rm c} \subset \mathcal{V},
%
\\ \label{eq:ZDA_ex_6}
\tilde{v}_i(t_0)&=\tilde{v}_j(t_0)
\implies
\tilde{v}_i(t)=\tilde{v}_j(t),
\ \ \hspace{.6ex}\forall \, i,j \in \mathcal{D}_{\rm c} \subset \mathcal{V}.
\end{align}
Also, as shown in Lemma \ref{lemma:app_component_measur}, under condition \ref{cond:detection_condition_suff_0} we have
\begin{align}\label{eq:graph_component_measurements}
\mathcal{D}_{\rm c} \cap \mathcal{M} \neq \emptyset, \ \ \forall \, {\rm c} \in \{1, \cdots, \boldsymbol{{\rm c}}\},
\end{align}
with set $\mathcal{D}_{\rm c}$ given in \eqref{eq:states_relabeled}.
\noindent
Now under \eqref{eq:graph_component_measurements}, it is concluded from \eqref{eq:ZDA_ex_4}, \eqref{eq:ZDA_ex_5}-\eqref{eq:ZDA_ex_6} that
\begin{align}\label{eq:zda_simp}
\tilde{x}_i(t)=\tilde{x}_j(t)=0,
\ \
\tilde{v}_i(t)=\tilde{v}_j(t)=0,
\ \ \forall \, i,j \in \mathcal{D}_{\rm c},
%
\end{align}
which by considering \eqref{eq:ZDA_traj} implies that \eqref{eq:H_q_elements_kernel_2_2} is simplified to
\begin{align}\label{eq:H_q_elements_kernel__}
(\mathfrak{e}^{\top}_{i}-\mathfrak{e}^{\top}_{j}) {x}^{(\rm m)}(t)
&= \nonumber \\
(\mathfrak{e}^{\top}_{i}-\mathfrak{e}^{\top}_{j}) \bar{x}^{(\rm m)}(t)
&=
0,
&\forall& \, i, j \in \mathcal{D}_{\rm c}, \ \ \forall \,{\rm c} \in \{1, \cdots, \boldsymbol{\rm c}\},
\nonumber \\
& &\forall& \, {\rm m} \in \ensuremath{\mathbb{N}}_0, \ \ \forall \,t \in [t_1, +\infty),
\end{align}
where $\bar{x}_i$ and $\bar{x}_j$ are the elements of state vector $\bar{\mb{x}}$ in \eqref{eq:ZDA_traj} denoting the states of an attack-free system that satisfies \eqref{eq:stealthy} (i.e. $\dot{\bar{\mb{x}}} = \mb{A}_{\mb{q}} \bar{\mb{x}}$ obtained using $\dot{\mb{x}}$-dynamics in \eqref{eq:cl_sys} with $\mb{B}\mb{u}_a=\boldsymbol{0} $ and unknown initial condition $\bar{\mb{x}}_0$ as defined in Proposition \ref{propo:stealthy_attacks}).
Then using the attack-free dynamics $\dot{\bar{\mb{x}}} = \mb{A}_{\mb{q}} \bar{\mb{x}}$, the term $(\mathfrak{e}^{\top}_{i}-\mathfrak{e}^{\top}_{j}) \bar{x}^{(\rm m)}(t)
=
0$ in \eqref{eq:H_q_elements_kernel__} can be rewritten as
\begin{align}\label{eq:H_q_elements_new1}
(\mathfrak{e}^{\top}_{i}-\mathfrak{e}^{\top}_{j}) \mathcal{L}^{\rm m}_{\mb{q}} \bar{{x}}(t) = 0,
\ \ \forall \, i, j \in \mathcal{D}_{\rm c},\, \forall \, {\rm m} \in \ensuremath{\mathbb{N}}_{0}, \, \forall \, t \in [t_1, \infty),
\\ \label{eq:H_q_elements_new2}
%
(\mathfrak{e}^{\top}_{i}-\mathfrak{e}^{\top}_{j}) \mathcal{L}^{\rm m}_{\mb{q}} \bar{{v}}(t) = 0,
\ \ \, \forall \, i, j \in \mathcal{D}_{\rm c},\, \forall \, {\rm m} \in \ensuremath{\mathbb{N}}_{0}, \, \forall \, t \in [t_1, \infty).
\end{align}
Moreover, note that \eqref{eq:H_q_elements_new1} and \eqref{eq:H_q_elements_new2} have the same form as equations (109a) and (109b) in \cite{mao2020novel}. Then under further conditions \ref{cond:detection_condition_suff_1} and \ref{cond:detection_condition_suff_2}, it can be verified using the same procedure as in \cite[Th. 2]{mao2020novel} that \eqref{eq:H_q_elements_new1} and \eqref{eq:H_q_elements_new2} yield
\begin{align}\label{eq:equib_1}
\bar{x}_i^{}(t) = \bar{x}_j^{}(t), \ \
\forall \, i, j \in \mathcal{V}, \ \ \forall \, t \in [t_1, +\infty),
\\ \label{eq:equib_2}
\bar{v}_i^{}(t) = \bar{v}_j^{}(t), \ \
\forall \, i, j \in \mathcal{V}, \ \ \forall \, t \in [t_1, +\infty),
\end{align}
which means the the entire states of the attack-free system have achieved consensus. Considering the equilibrium subspace \eqref{eq:cond_consensus} as a result of the consensus protocol \eqref{eq:ctrl_proto}, one can conclude that \eqref{eq:equib_1}-\eqref{eq:equib_2} and \eqref{eq:cond_consensus} coincide. Therefore, from \eqref{eq:H_q_elements_kernel__} and \eqref{eq:equib_1}-\eqref{eq:equib_2}, obtained under conditions \ref{cond:detection_condition_suff_0}-\ref{cond:detection_condition_suff_2}, one can conclude that stealthy ZDA is undetectable in $\mb{r}_0(t)$ of \eqref{eq:obs_cent} only in the worst-case scenario that intra-cluster switching links are between agents whose trajectories are not affected by ZDA as well as all of the system \eqref{eq:cl_sys}'s attack-free trajectories, characterized in \eqref{eq:H_q_elements_kernel__}, are at the consensus equilibrium \eqref{eq:cond_consensus}.
Covert attack case:
consider \eqref{eq:H_q_elements_kernel_relabeled} under which a covert attack remains stealthy in a safe mode and note that
\begin{align}\label{eq:covert_sol}
\mb{x}(t)&=\bar{\mb{x}}(t)+\tilde{\mb{x}}(t),\ \
\forall\, t \in [t_1, +\infty), \;\; \text{with}
\nonumber \\
\tilde{\mb{x}}(t) &=
{e^{\mb{A}_1(t-t_{1})}}
\tilde{\mb{x}}(t_1)
+
{\int^{t}_{t_{1}}
e^{\mb{A}_{1}(t-\boldsymbol{\tau})}
\mb{B}\mb{u}_{a}(\boldsymbol{\tau})d{\boldsymbol{\tau}}}
\end{align}
according to the attack model \eqref{eq:attack_model} and Proposition \ref{propo:stealthy_attacks}.
Given \eqref{eq:covert_sol}, \eqref{eq:H_q_elements_kernel_2_2} can be rewritten as
\begin{align}\label{eq:H_q_elements_kernel_2_3}
[(\mathfrak{e}^{\top}_{i}-\mathfrak{e}^{\top}_{j}) \ \ 0]\, \bar{\mb{x}}^{(\rm m)}(t)
=
[(\mathfrak{e}^{\top}_{i}-\mathfrak{e}^{\top}_{j}) \ \ 0]\, \tilde{\mb{x}}^{(\rm m)}(t), \ \ &\forall \, i, j \in \mathcal{D}_{\rm c},
\nonumber \\
\forall \, {\rm c} \in \{1, \cdots, \boldsymbol{\rm c}\}, \ \
\forall \, {\rm m} \in \ensuremath{\mathbb{N}}_0, \ \ \forall \,t \in [t_1, +\infty),
\end{align}
Notice that the attack-free system states, $\bar{\mb{x}}(t)$ in \eqref{eq:covert_sol}, converge to \eqref{eq:cond_consensus} as $t\rightarrow +\infty$, then the left side of \eqref{eq:H_q_elements_kernel_2_3} converges to zero and one can conclude from \eqref{eq:covert_sol} and \eqref{eq:H_q_elements_kernel_2_3} that continuous states $\tilde{\mb{x}}(t) = \mathrm{col}(\tilde{x}(t),\tilde{v}(t))$ exist in either of the following cases
\begin{align}\label{eq:covert_case1}
\hspace{-1ex}\text{case 1}:
\tilde{{x}}_{i}(t)=\tilde{{x}}_{j}(t) \neq 0,
\ \ &\forall \, i, j \in \mathcal{D}_{\rm c}, \, \forall\, t \in [t_1, +\infty)
\end{align}
\begin{align}\label{eq:eq:covert_case2}
\hspace{-1ex}\text{case 2}:
\tilde{{x}}_{i}(t)=\tilde{{x}}_{j}(t) = 0,
\ \ &\forall \, i, j \in \mathcal{D}_{\rm c}, \, \forall\, t \in [t_1, +\infty)
\end{align}
Note that here case 1 in \eqref{eq:covert_case1} implies the attack input $\mb{u}_{a}$ in \eqref{eq:covert_sol} has driven and kept the states of agents involved in switching into an unknown equilibrium over time span $\forall \, t \in [t_1, +\infty)$. Also, case 2's interpretation and analysis coincide with that of ZDA in \eqref{eq:zda_simp}. Then following the same analysis as the ZDA's, one can conclude that, under conditions \ref{cond:detection_condition_suff_0}-\ref{cond:detection_condition_suff_2}, covert attack is undetectable in $\mb{r}_0(t)$ of \eqref{eq:obs_cent} only in the worst-case scenarios that 1) intra-cluster switching links are between agents whose trajectories are identical over time under the effect of covert attack; and 2) intra-cluster switching links are between agents whose trajectories are not affected by covert attack as well as all of the system \eqref{eq:cl_sys}'s attack-free trajectories are at the consensus equilibrium \eqref{eq:cond_consensus}.
\bibliographystyle{IEEEtran}
|
\section*{Introduction}
The question of how we perceive and experience the world we live in is a fascinating, long-standing open question in neuroscience, philosophy and psychology. This conscious experience vanishes during dreamless sleep or under general anesthesia \cite{koch2016neural,mashour2006integrating}. Every day millions of surgeries all over the world would not be possible without anesthetics. However, despite a century of research, the mechanisms by which anesthetics cause a reversible loss of consciousness remain obscure \cite{brown2011general}. During anesthesia, old memories are preserved, but no new memory is formed. Furthermore, anesthetics act relatively selectively on consciousness, as many non-conscious brain activities, including sensory-evoked potentials, persist during anesthesia. Understanding anesthesia, apart from its benefits in designing and developing novel anesthetics, could help explain the mystery of consciousness.\par
The connections between theories of anesthesia, related to neural correlates of consciousness, and the Meyer Overton correlation—solubility of anesthetics in a non-polar, ‘hydrophobic’ medium—has not been completely understood, and direct anesthetic effects on synaptic receptors are variable and inconsistent. Further, single-cell organisms perform cognitive activities predominantly by cytoskeletal microtubules (MTs) and are inhibited by anesthetic gases even without synapses or networks \cite{ja2015anesthetics}. Bernard showed that anesthetics act directly on cytoplasm, depending on cytoskeletal proteins' dynamics comprising actin filaments and MTs \cite{perouansky2012quest}. Franks and Lieb found that anesthetics act directly within proteins in non-polar hydrophobic regions \cite{franks1984general}. This led some anesthetic researchers to search for new target protein/molecule for anesthetics. Moreover, Eckenhoff et al. found that anesthetics bind to 23 membrane proteins and 34 cytoplasmic proteins including actin, and tubulin \cite{xi2004inhalational,pan2007halothane}. Studies suggest anesthetics exert their effects via acting on protein reaction networks involved in neuronal growth, proliferation, division, and communication, which depend on MTs \cite{pan2008inhaled}. Although the affinity of anesthetics binding to tubulin is a thousand times weaker than to membrane protein, the abundance of tubulin is thousand to ten thousand times more than membrane protein sites.\par
It has also been proposed that anesthetics act on quantum electronic activity in neuronal hydrophobic regions rather than binding to specific receptors. Consistently, Turin et al. showed that specific electron spin resonance (ESR) signals, consistent with free electrons, can be observed during anesthesia \cite{turin2014electron}. The same authors proposed that the anesthetic action may involve some form of electron transfer. Moreover, Li et al. showed experimentally that isotopes of xenon with non-zero nuclear spin had reduced anesthetic potency in mice compared with isotopes with no nuclear spin \cite{li2018nuclear}. These findings are consistent with the idea that different nuclear spins of anesthetics can modulate the electron transferring process differently. Motivated by this line of thought, a recent study shows that radical pairs may explain the mechanism behind xenon-induced anesthesia \cite{smith2021radical}.\par
Quantum physics has been proposed to be part of the solution for the mystery of consciousness. In particular the holistic character of quantum entanglement might provide an answer to the binding problem \cite{Simon2019}. In the 1990s, Penrose and Hameroff proposed a theory of consciousness based on quantum computations in MTs \cite{hameroff2014consciousness,stuart1998quantum,matsuno2001internalist,hagan2002quantum}. Computational modeling suggested that electron resonance transfer among aromatic amino acid tryptophan (Trp) rings in tubulin (subunits of MTs) in a quantum electronic process could play roles in consciousness \cite{hameroff2002conduction}. More recently, Craddock et al. showed that anesthetic molecules might bind in the same regions and hence result in anesthesia \cite{craddock2012computational}. However, quantum electronic coherence beyond ultrafast timescales demands more supporting evidence and has been recently challenged experimentally \cite{cao2020quantum}. In contrast, quantum spin coherence could be preserved for much longer timescales \cite{hu2004spin}. For example, Fisher has proposed that phosphorus nuclear spins could be entangled in networks of Posner molecules, \ch{Ca9(PO4)_6}, which could form the basis of a quantum mechanism for neural processing in the brain \cite{fisher2015quantum}. However, this particular spin-based model model also requires more supporting evidence and recently has faced experimental challenges \cite{Chen2020}. \par
It is know that magnetic fields (MFs) can influence different brain functions \cite{romero2019neural,wang2019spontaneous,lenz2016repetitive,rauvs2014extremely,pacini1999effect,grehl2015cellular,manikonda2007influence}. Recently, it has been shown that shielding the geomagnetic field—exposure to hypomagnetic field (HMF)—influences adult hippocampal neurogenesis and hippocampus-dependent cognition in mice, where ROS are implicated \cite{zhang2021long}. There exists a considerable amount of evidence showing that MFs affect MTs \cite{vassilev1982parallel,glade2005brief,bras1998susceptibility,zhang201727,qian2009large,luo2016moderate}. Wang et al. show that exposure to HMF caused tubulin assembly disorder \cite{wang2008tubulin}. Moreover, Wu et al. observe that low-frequency sub-millitesla MF modulates the density of MTs in cells \cite{wu2018weak}. All these observations establish the magnetosensitivity of MTs for wide ranges of MF strengths. \par
Magnetosensitive reactions often involve radical molecules—transient molecules with an odd number of electrons \cite{rodgers2009chemical}. A radical pair is a short-lived reaction intermediate comprising two radicals formed in non-equilibrium states whose unpaired electron spins may be in a superposition of singlet (S) and triplet (T) states \cite{Steiner1989}, depending on the parent molecule's spin configuration \cite{Timmel1998}. The radical pair mechanism (RPM) is the most promising explanation for weak magnetic field effects on chemical reactivity \cite{hore2016radical}. Schulten was the first to propose the RPM to explain the magnetoreception of migratory birds \cite{schulten1978biomagnetic}, and to date, the RPM is the most well-established model for this phenomenon \cite{xu2021magnetic}. Recently, it has also been proposed that RPM may explain xenon induced general anesthesia \cite{smith2021radical}, lithium effects on hyperactivity \cite{zadeh2021entangled} and the magnetic field effects on the circadian clock \cite{zadeh2021radical}. \par
MTs are made of $\alpha-\beta$ tubulin heterodimers. The distribution and organization of MTs in cells are governed by a large number of MT-associated proteins (MAPs) \cite{brouhard2018microtubule}. MTs play crucial roles in cell shape, cell transport, cell motility, cell division \cite{nogales2001structural,wu2017microtubule,akhmanova2008tracking,akhmanova2015control,redwine2012structural,monroy2020combinatorial,monroy2018competition}, neuronal polarity \cite{burute2021matrix}, information processing of living systems \cite{sanchez2021microtubules}, synaptic activity \cite{dixit2008differential}, regulating the precise timing of nerve spikes \cite{singh2021cytoskeletal}, and Alzheimer’s disease (AD) \cite{congdon2018tau}.\par
Evidence suggests that oxidative stress is vital for regulating actin and MT dynamics \cite{wilson2015regulation}. MTs contain specific amino acid residues, including Trp, tyrosine (Tyr), and phenylalanine (Phe), susceptible to oxidation. This, in turn, affects the ability of MT to polymerize and causes the severing of actin microfilaments in neuronal and non-neuronal cells. Contrarily, ROS inhibition causes aberrations in actin polymerization, decreases neurite outgrowth, and affects neurons' normal development and polarization. \par
Various studies have proposed mathematical models for the dynamics and stability of MTs \cite{white2015microtubule,bowne2013microtubule}. Craddock et al. show that the dynamics of MT can be framed in a simple kinetic model \cite{craddock2012zinc}. In the context of the RPM, Player et al. show that quantum effects can be introduced to the chemical oscillator model by considering the quantum effects on the corresponding reaction rates in the chemical equations \cite{player2021amplification}. Taking the same approach, a new study shows that quantum effects can directly modulate the period of the circadian clock in \textit{Drosophila}, where the spin dynamics of the PRs are the key elements \cite{zadeh2021radical}.\par
\begin{figure}[ht]
\includegraphics[width=0.55\linewidth]{TrpH-O2.png}
\caption{Schematic presentation of \ch{TrpH^{.+}} [the aromatic molecule] and \ch{O2^{.-}} [the red molecule] radical pair, considered in the RPM model in the present work, similar to Ref. \cite{smith2021radical}. The radical pair undergoes interconversion between singlet and triplet states. Image rendered using Avogadro (https://avogadro.cc).}
\label{fig:schem}
\end{figure}
In the present work, we propose that there are naturally occurring RPs in the form of \ch{TrpH^{.+}} and \ch{O2^{.-}}, which is an important ROS, that play important roles in MT organization, as shown in Fig. \ref{fig:schem}. Our model predicts that the applied magnetic fields alter the spin dynamics of such RPs, similar to Ref. \cite{smith2021radical}, and hence modulate the assembly of MTs in cytoskeleton. It
further predicts that the effect of zinc on the MT density exhibits isotopic dependence.\par
In the following, we review the experimental results for the effects of applied magnetic field \cite{wu2018weak,wang2008tubulin} on the density of MTs. We then describe the quantum spin dynamics of our radical pair model by a simple kinetic model for the MT dynamics. Next, we introduce the quantum effect to the density of the MT, and we show that our model can reproduce the observed magnetic field effects; it further makes new predictions for experiments using low-frequency MFs. Lastly, the model predicts that the zinc effect on the MT density is isotope dependent.
\section*{Results}
\subsection*{Prior experiment}
Wang et al. report that shielding Earth's geomagnetic field (GMF) (0.025-0.065 mT) caused disorders in tubulin self-assembly \cite{wang2008tubulin}. They show that the absorbance at 350 nm, which is for monitoring tubulin self-assembly, was altered by exposure to 30 min HMF. Average gray volume per cell was reported based on the amount of fluorescence in each cell. About 95$\%$ of tubulin were assembled in the GMF, while much less of the tubulin assembled ($\sim$ 64$\%$) in the HMF. In that work, the magnitude of the residual GMF was 10–100 nT.
\subsection*{Radical pair model calculations}
We develop an RP model to reproduce the HMF effects on the MT density observed by Wang et al. \cite{wang2008tubulin}. Among the amino acids in MTs, as mentioned above, Trp is redox active \cite{saito1981formation}, as shown by its involvement in RP formation in the context of cryptochrome, and could feasibly participate in the creation of RPs. We propose that the magnetic field interacts with the spins of RPs on Trp and superoxide. Note, that superoxide is thought to form RPs with flavins in other contexts \cite{usselman2016quantum}. The correlated spins of RP are assumed to be in the [\ch{TrpH^{.+}} ... \ch{O2^{.-}}] form, following Ref. \cite{smith2021radical}, where the unpaired electron on each molecule couples to the nuclear spins in the corresponding molecule. Oxygen has zero nuclear spin and thus zero coupling with its nucleus.\par
We consider a simplified system in which the unpaired electron on \ch{TrpH^{.+}} is coupled to the Trp's $\beta$-proton with the largest isotropic HF coupling constant (HFCC) of 1.6046 mT \cite{lee2014alternative} among all the nuclei in Trp. We consider only Zeeman and HF interactions \cite{Efimova2008,hore2016radical}. For the RPs, we assume the $g$-values of a free electron (which is an excellent approximation in the low-field regime that we are considering here). The Hamiltonian for the RP system reads:
\begin{ceqn}
\begin{equation}
\hat{H}=\omega \hat{S}_{A_{z}}+a_A \mathbf{\hat{S}}_A.\mathbf{\hat{I}}_1+\omega \hat{S}_{B_{z}},
\end{equation}
\end{ceqn}
where $\mathbf{\hat{S}}_A$ and $\mathbf{\hat{S}}_B$ are the spin operators of radical electron on \ch{TrpH^{.+}} and \ch{O2^{.}}, respectively, $\mathbf{\hat{I}}_A$ is the nuclear spin operator of \ch{TrpH^{.+}}'s $\beta$-proton, $a_{A}$ is HFCC, and $\omega$ is the Larmor precession frequency of the electrons due to the Zeeman effect. In the model presented here, for zinc effects, $a_{B}$ corresponds to the nuclear spin of zinc. We assumed the initial state of RPs to be singlet states (see the Discussion section).\par
Using the Liouville-von Neumann equation for the spin state of the radical pair, we calculate the singlet yield resulting from the radical pair mechanism throughout the reaction. The ultimate triplet yield, $\Phi_T$, for periods much greater than the RP lifetime \cite{Hore2019} has the following form:
\begin{ceqn}
\begin{equation}
\Phi_T=\frac{3}{4}+\frac{k}{4(k+r)}-\frac{1}{M}\sum_{m=1}^{4M}\sum_{n=1}^{4M}|\bra{m}\hat{P}^S \ket{n}|^2 \frac{ k(k+r)}{(k+r)^2+(\omega_m-\omega_n)^2},
\label{eq:sy}
\end{equation}
\end{ceqn}
where $M$ is the nuclear spin multiplicity, $\hat{P}^S$ is the singlet projection operator, $\ket{m}$ and $\ket{n}$ are eigenstates of $\hat{H}$ with corresponding eigenenergies of $\omega_m$ and $\omega_n$, respectively, $k$ is the RP reaction rate, and $r$ is the RP spin-coherence rate (relaxation rate). In this model, we assumed the reaction rates for singlet and triplet have the same values. Of note, the singlet and triplet product of the RP system in [\ch{TrpH^{.+}} ... \ch{O2^{.-}}] are \ch{H2O2} and \ch{O2^{-}} \cite{usselman2016quantum}, respectively, which are the major ROS in redox regulation of biological activities and signaling \cite{sies2020reactive}.\par
Here we look at the dependence of the triplet yield on changes in the strength of the external static magnetic field for the [\ch{TrpH^{.+}} ... \ch{O2^{.-}}] radical complex, as shown in Fig. \ref{fig:TY-SB}, for $k=10^{6}$ s$^{-1}$, and $r=10^{5}$ s$^{-1}$ with $a_{A}=1.6046$ mT. One can notice a fairly strong HMF effect (the triplet yield goes from around 60\% to around 40\%). Our choices for the rate constants are discussed below and in the Discussion section. \par
\subsection*{Chemical kinetics model for MT dynamics}
We use a simple mathematical model for the dynamics of MT, following the work of Craddock et al. \cite{craddock2012zinc}. This model is based on the interconversion of the free tubulins, Tu, and the MT in cytoskeleton:
\begin{ceqn}
\begin{equation}\label{eq:chem}
\ch{Tu <>[ $k_{\mathrm{p}}$ ][ $k_{\mathrm{d}}$ ] MT},
\end{equation}
\end{ceqn}
where $k_p$ and $k_d$ are the polymerization rate and the depolymerization rate, respectively. The chemical equation reads as follows:
\begin{ceqn}
\begin{equation}
\frac{[dMT(t)]}{dt}=k_p [Tu(t)]-k_d [MT(t)],
\label{eq:mt}
\end{equation}
\end{ceqn}
which yields $[MT(t)]= \frac{k_p [P]}{k_p+k_d} (1-e^{-(k_p+k_d)t})$.
Craddock et al. suggest that changes in the concentration of zinc can be incorporated by modulating the polymerization rate \cite{craddock2012zinc}. In the present work, it is assumed that the total tubulin protein concentration [P] is 400 $\mu$M, such that [MT]+[Tu]=[P], $k_p$ = 90 s$^{-1}$ and $k_d$ = 150 s$^{-1}$. These values are chosen such that we could be able to reproduce the experimental findings, which are comparable to those in Ref. \cite{craddock2012zinc}.
\subsection*{Quantum effects on microtubule density}
The effects of triplet yield change can be introduced to the chemical equation of the MTs by modifying the kinetic rates, following the work of Player et al. \cite{player2021amplification,zadeh2021radical}. In the chemical kinetics equations for MT oscillations, Eq. \ref{eq:mt}, the corresponding rate is $k_{p}$, which is the MT polymerization rate \cite{craddock2012zinc}. The key assumption in our model is that this rate is influenced by a RP reaction. The effect of the triplet yield change on $k_{p}$ reads:
\begin{ceqn}
\begin{equation}\label{eq:yeild-chem-n}
k'_{p} \propto k_{p} \frac{\Phi_T'}{\Phi_T},
\end{equation}
\end{ceqn}
where $k'_{p}$, $\Phi_T$, and $\Phi'_T$ are the modified rate constant $k_{p}$, the triplet yield under natural quantum effects (only GMF and no isotope effects), and the triplet yield resulting from quantum effects due to the external MF effects and hyperfine interactions from isotope effects, respectively.
\subsubsection*{Magnetic field effects on microtubule density}
Here, we look at the explicit effects of an applied magnetic field on the density of microtubules. Using Eqs. \ref{eq:mt}, we explore the parameter space of relaxation rate $r$ and recombination rate $k$ in order to investigate the effects of shielding geomagnetic field on MT's density. Wang et al. report that the ratio MT density of geomagnetic field over hypomagnetic field is about 1.48 \cite{wang2008tubulin}. Fig. \ref{fig:contT} show that the ratio of the MT density of GMF over HMF can reach above 1.3, which has the right trend compared to the experimental findings. However the uncertainty of the experiment was not reported. Our model predicts a magnetic dependence of the MT density. Fig. \ref{fig:mt-sb} show the dependence of the MT density ratio of in GMF compared to applied static magnetic field, for $a_{A}=1.6046$ mT based the RP complex of [\ch{TrpH^{.+}} ... \ch{O2^{.-}}], $k=10^{6}$ s$^{-1}$, and $r=10^{5}$ s$^{-1}$. Fig. \ref{fig:mt-sb} indicates that exposure to static magnetic fields stronger than the geomagnetic field could decrease the the microtubule density ratio. The maximum MT density occurs around 0.05 mT, which is in the range of the GMF.
\subsubsection*{Zinc isotope effect on microtubule density}
The RPM typically leads to isotope effects \cite{smith2021radical,zadeh2021entangled}, and thus an isotope effect would be a good test of our proposal. It is known that alterations in zinc ion concentration in neurons influence the stability of polymerized MTs \cite{craddock2012zinc}. Zn is a positively charged ion. Thus it is natural to assume that \ch{Zn^{2+}} couples with a molecule with a negative charge. Among all stable isotopes of zinc, only \ce{^{67}Zn} has a nuclear spin of $I_{B}=-\frac{5}{2}$, with a natural abundance of 4$\%$. Here we explore the isotope effect of zinc on the density of MTs. In this model nuclear spin of zinc modulates the spin dynamics of the RPs via hyperfine interactions. The Hamiltonian for the RP system here reads:
\begin{ceqn}
\begin{equation}
\hat{H}=\omega \hat{S}_{A_{z}}+a_A \mathbf{\hat{S}}_A.\mathbf{\hat{I}}_1+\omega \hat{S}_{B_{z}}+a_B \mathbf{\hat{S}}_B.\mathbf{\hat{I}}_2,
\label{eq:zn-ham}
\end{equation}
\end{ceqn}
where $\mathbf{\hat{I}}_2$ is the nuclear spin operator of \ch{Zn^{+2}}, and $a_B=-11.39$ mT is the HFCC of \ch{Zn^{+2}}.\par
We look at the effect of \ce{^{67}Zn} on the MT density. We assume that \ce{^{67}Zn} interacts with the superoxide radical. Our model predicts that administration of \ce{^{67}Zn} increases the density of MTs compared to Zn with zero nuclear spin, as shown in Fig.\ref{fig:cont-zn}. We explored the parameter space of relaxation rate $r$ and recombination rate $k$ in order to investigate the effects of the \ce{^{67}Zn} treatment on MT's density, for $a_{A}=1.6046$ mT and $a_{B}=-11.39$ mT, as shown in Fig. \ref{fig:cont-zn}. Fig. \ref{fig:mt-sb-zn} shows the dependence of the MT density ratio of the administration of Zn over \ce{^{67}Zn} on the strength of applied magnetic field based on the RP complex of [\ch{TrpH^{.+}} ... \ch{O2^{.-}}]. The geomagnetic field is 0.05 mT. The magnetic field modulates rates $k_d$ and $k_p$ for $r=10^{5}$ s$^{-1}$ and $k=10^{6}$ s$^{-1}$. Our model predicts that administering \ce{^{67}Zn} increases the MT density compared to Zn without nuclear spin.
\section*{Discussion}
In the present work, our main goal was to explore whether an RP model can help explain the magnetic field effects on the MT density. We showed that the quantum effects influence the rates in the chemical kinetics of the MT dynamics, and then this results in a change in the density of MT in cytoskeleton. Our model reproduces the experimental findings of Ref. \cite{wang2008tubulin} fairly well, as shown in Fig. \ref{fig:contT}. Our model further predicts that exposure to static magnetic fields stronger than the geomagnetic field could decrease the the microtubule density ratio, with the maximum microtubule density occurs at a magnetic field strength in the range of the geomagnetic field, which might have evolutionary roots.\par
The present work predicts that the zinc effects on the MT density exhibits an isotope dependent manner. Isotope effects are generally a good indication for radical pairs. It thus would be interesting to perform an experiment to probe such isotope effects.\par
The \ch{O2^{.-}} radical is thought to have a fast spin relaxation rate \cite{Player2019,Hogben2009}. However, it has also been proposed that this fast spin relaxation can be decreased by the biological environment. Additionally, Kattnig and Hore show that scavenger species around the superoxide radical can also reduce its fast spin relaxation \cite{Kattnig2017,Kattnig2017b}. \par
The ground state of the oxygen molecule is singlet, and due to this, it is often assumed that in the RP complexes superoxide are formed in triplet states, as opposed to the case considered in the present work, where it is assumed to be formed as a singlet. Studies have shown the formation of superoxide from singlet oxygen \cite{saito1981formation}. Thus it seems reasonable that the initial state for RP formation could also be its excited singlet state, which is also which is its excited state (and is also a biologically relevant ROS) \cite{kerver1997situ,miyamoto2014singlet,kanofsky1989singlet}. Further, the transition of the initial RP state from triplet to singlet could also take place due to spin-orbit coupling \cite{goushi2012organic,fay2019radical}. \par
Let us also note that this model could be adapted for other RP complexes in the microtubule dynamics, e.g. involving flavin and tyr. Such radical pairs can be formed via flavoenzymes \cite{liu2016mical3,yamauchi1983disassembly}.\par
A well-known indication of the pathogenesis of tauopathy—loss-of-function effects on the microtubules and the gain-of-function effects of the toxic tau species—is oxidative stress, which contributes to tau phosphorylation and the formation of neurofibrillary tangles \cite{ballatore2007tau}. However, the mechanisms behind the connection between reactive oxygen species (ROS) generated by oxidative stress and tau hyperphosphorylation are elusive \cite{kurian2017oxidative}. Further, redox signaling and oxidative stress regulate cytoskeletal dynamics \cite{wilson2015regulation}. Proper balances between chemical reduction and oxidation (known as redox balance) are crucial for normal cellular physiology. Imbalances in the production of oxidative species lead to DNA damage, lipid peroxidation, and aberrant post-translational modification of proteins, which could induce injury, cell death, and disease \cite{sies2020reactive}. These findings further emphasize the role of radicals and likely the need for a RPM in the context of the brain. Our proposed RPM model for magnetic field effects and Zn isotope effects on the microtubule organization includes superoxide radical.\par
Microtubules not only play crucial roles in cell shape, cell transport, cell motility, and cell division, but also are important targets for curing devastating diseases, such as Alzheimer's disease \cite{gong2008hyperphosphorylation,alonso1996alzheimer,iqbal2010tau}, Parkinson's diseases \cite{feng2006microtubule,kett2012lrrk2}, and cancer \cite{parker2014microtubules,mukhtar2014targeting,jordan2004microtubules}. The dynamics of MAPs signaling proteins play critical roles in the MT network, hence in the process of synaptic plasticity and brain function. Studies suggest that memory is encoded in MT of neuronal dendrites and cell bodies. Anesthetic causes loss of consciousness and memory formation via acting on MTs \cite{craddock2012cytoskeletal,janke2010tubulin,janke2011post}. Disintegration and separation of the MT-associated protein tau have been observed in memory neurodegenerative diseases and disorders, e.g., in AD \cite{goldstein2012chronic}; MT-stabilizers is currently the target for during such diseases \cite{brunden2010epothilone,congdon2018tau,li2017tau}. However, the underlying mechanism for such diseases is mostly unknown.Thus this project also paves a potential path to study other functionalities of the body and the brain connected to MTs in the light of the RPM. \par
In conclusion, our results suggest that quantum effects may underlie the magnetic field effects on the microtubule dynamics. This is likely a similar mechanism to those behind magnetoreception in animals \cite{xu2021magnetic}, xenon-induced general anesthesia \cite{smith2021radical}, lithium treatment for mania \cite{zadeh2021entangled}, the magnetic field effects on the circadian clock \cite{zadeh2021radical}. Our work is thus another piece of evidence that quantum entanglement \cite{fisher2015quantum,Simon2019,Adams2020,gauger2011sustained,bandyopadhyay2012quantum,cai2010quantum,kominis2012magnetic,pauls2013quantum,tiersch2014approaches,zhang2014sensitivity} may play essential roles in the brain's functions, anesthesia, and consciousness \cite{Hameroff2014b,hameroff2014consciousness}. Particularly, the photo-emission of singlet oxygen could serve as quantum messengers to establish long-distance connections \cite{Kumar2016} that might be essential for consciousness. Our work also provides a potential connection between microtubule-based and spin-based quantum approaches to consciousness.
\section*{Methods}
\subsection*{DFT Analysis}
The ORCA package \cite{Neese2011} was used for our \ch{Zn^{2+}}-\ch{O2^{.-}} DFT calculations, and the molecular structure was optimized using PBE0/def2-TZVP. The orbitals obtained from the optimization calculations were used to calculate orbital energies as well as the hyperfine coupling constant $a_B$. Using RI-B2GP-PLYP/def2-QZVPP\cite{Goerigk2011}, we obtained $a_2=-11.39$ mT. In these calculations, relativistic effects were treated by a scalar relativistic Hamiltonian using the zeroth-order regular approximation (ZORA) \cite{vanLenthe1996}. Solvent effects were considered by using the conductor-like polarizable continuum model (CPCM) \cite{Marenich2009}, with a dielectric constant of 2. The resulting Mulliken charge and spin population of the [\ch{Zn^{2+}}-\ch{O2^{.-}}] complex indicates that the unpaired electron resides primarily on the \ch{O2} molecule but is extended slightly onto the zinc atom, see Table \ref{tab:mulliken}. The highest occupied molecular orbital (HOMO) of [\ch{Zn^{2+}}-\ch{O2^{.-}}] is shown in Fig. \ref{fig:homo}.\par
\begin{table}[tbhp]
\centering
\caption{\label{tab:mulliken}Mulliken charge and spin population of [\ch{Zn^{2+}}... \ch{O2^{.-}}].}
\begin{tabular}{ccc}
Atom & Charge Population &Spin Population\\
\hline
O & -0.203235 & 0.554431 \\
O & -0.203243 & 0.554413 \\
Zn & 1.406478 & -0.108844 \\
\hline
Sum & 1 & 1 \\
\end{tabular}
\end{table}
\section*{Data Availability}
The generated datasets and computational analysis are available from the corresponding author on reasonable request.
|
\section{Introduction}
Quantum computing provides a new paradigm for manipulating information according to the laws of quantum mechanics; it is expected to provide an advantage over classical computation for some scientific problems.\cite{Mike_Ike} As envisioned by Richard Feynman, one of those problems is simulating quantum mechanical systems.\cite{Feynman1982} Focusing on molecular quantum systems, quantum chemistry is poised to be among the fields of study that could benefit from the developments of quantum computation. An example for how this can be achieved is given by the phase estimation algorithm \cite{Daniel1997, Daniel1999} for computing energy eigenvalues and preparing energy eigenstates. This algorithm has been simulated and shown to work\cite{Aspuru-Guzik2005}; it will provide opportunities for solving problems that cannot be solved on classical computers. Unfortunately, phase estimation is quite challenging to implement, because it requires controlled time evolution of the system. Hence, it has extremely deep circuits if the time evolution of the molecule is treated exactly in a conventional basis for the Hamiltonian.
Current quantum hardware is limited due to noise and decoherence. This near-term hardware is called noisy intermediate-scale quantum (NISQ)\cite{Preskill2018}. They can only work with low-depth circuits on a modest number of qubits. This precludes direct application of many quantum algorithms, such as the phase-estimation algorithm. To be specific, we have to work with two constraints in the near term: (i) the number of qubits will remain relatively small and (ii) the allowed circuit depth will remain low due to gate fidelity and decoherence errors. To take advantage of existing and near-term quantum computers, the variational quantum eigensolver (VQE) has been proposed as a low-depth alternative to quantum phase estimation. \cite{Peruzzo2014} It is a hybrid quantum-classical method, and it shows great promise.\cite{Cao2019} VQE needs to be carried out both on quantum and classical computers: on the quantum computer, quantum states depending on a set of variational parameters are prepared, and the expectation value of the Hamiltonian is then measured. Next, that set of parameters is optimized on classical computers and the loop is repeated until converged. But, this approach suffers from the need for higher-depth circuits as the ansatz wavefunction becomes more complex and from the appearance of barren plateaus in the optimization space (which is exacerbated by the noise and decoherence of NISQ machines). Our approach attempts to resolve both of these issues by using a quadratic expansion of the energy in terms of the variational amplitudes, which allows for significantly lower depth in the required circuits and an optimization that is performed entirely on the classical computer. But, it does so at the expense of requiring significantly more measurements.
Originally proposed as a wave-function ansatz for quantum chemistry about four decades ago,\cite{Kutzelnigg1977, Koch1981, Kutzelnigg1983} unitary coupled cluster theory (UCC) has gained renewed attention because, in its factorized form, it can serve as an efficient state preparation method for the VQE. Usually, the UCC ansatz starts from the Hartree-Fock (HF) reference state in the occupation number representation; then the different UCC factors are applied to the HF states in sequence; this excites the HF state by including terms from unoccupied orbitals, but it also includes de-excitation terms when the prepared state includes terms that can be de-excited by the next UCC factor. General speaking, the circuit depth for UCC state preparation is proportional to the number of excitations applied, especially if they are the same rank excitation. For example, a molecule with $N$ orbitals, has $N^4$ doubles excitations, which are usually the most important excitations to include in the ansatz. As a result, including all (or just a fraction of) all possible singles and doubles excitations already requires a circuit depth that is prohibitively high for large molecules, especially so on NISQ hardware.
In this contribution, we show how to improve this situation, by requiring only a small subset of UCC factors to be applied to the reference state on the quantum computer. The insight behind this comes from the fact that most amplitudes $\theta$ have small absolute values. This suggests that they can be expanded in a Taylor series about $\theta=0$, truncated at quadratic order and then optimized. Doing this from the reference HF state is only viable when the molecule is weakly correlated, which holds predominantly near the equilibrium configuration bond distances and angles. But, how many UCC factors need to be treated exactly when carrying out this approach as the correlations are increased (due to stretching)? This is the question we address in this work. We find the number remains relatively small, implying that such an approach can enable more complex molecules to be treated on currently available NISQ machines. Similar ideas have been used in other contexts as well. A quadratic expansion about a density-matrix-renormalization group calculation was performed successfully for a carbon dimer\cite{carbon} and the approach was also investigated for a quantum-inspired algorithm using an ansatz that can be constructed solely from Clifford circuits, which can be easily simulated on a classical computer~\cite{clifford}. Our approach is similar to both of these earlier works, but has a number of differences as well.
\section{Theory and Method}
\subsection{Unitary Coupled Cluster Theory (UCC) in Factorized Form and Operator Identity for UCC Factors}
In unitary coupled cluster (UCC) theory, the trial wave-function is expressed in an exponential form, given by
\begin{align}
|\Psi_{\textrm{UCC}}\rangle = \exp (\hat{\sigma}) |\Psi_0\rangle ,
\end{align}
where $|\Psi_0\rangle $ is the reference state and the operator $\hat{\sigma}$ is an anti-Hermitian combination of particle-hole excitation and de-excitation:
\begin{align}
\hat{\sigma} &= \hat{T} - \hat{T}^{\dagger} ; \\
\hat{T} &= \sum_i^{occ}\sum_a^{vir} \theta_i^a \hat{a}_a^{\dagger} \hat{a}_i + \sum_{ij}^{occ}\sum_{ab}^{vir} \theta_{ij}^{ab} \hat{a}_{a}^{\dagger}\hat{a}_{b}^{\dagger} \hat{a}_{j}\hat{a}_{i} + \cdots ~.
\end{align}
Here, the rotation angles $\theta$ are the variational parameters. We use letters from the start of the alphabet $a$, $b$, $c,\ldots$ to denote the virtual orbitals, with respect to the reference state, and letters from the middle of the alphabet $i$, $j$, $k,\ldots$ to denote the occupied orbitals in the reference state. To simplify the notation, we express a general $n$-fold excitation operator as $\hat{a}^{ab\dots}_{ij\dots} = \hat{a}^{\dagger}_{a}\hat{a}^{\dagger}_{b} \dots \hat{a}_j \hat{a}_i$ (with the corresponding de-excitation operator being its Hermitian conjugate). We work with a factorized form for the UCC, which is given by
\begin{align}
|\Psi_{\textrm{UCC}}\rangle = U_{\textrm{UCC}} |\Psi_0\rangle = \prod_k e^{\theta_k\sigma_k}
=\prod_{ij\cdots}^{occ}\prod_{ab\cdots}^{vir} \exp [\theta_{ij\cdots}^{ab\cdots} (\hat{a}_{ij\cdots}^{ab\cdots}-\hat{a}_{ab\cdots}^{ij\cdots})] |\Psi_0\rangle .
\label{eq:ucc-fac}
\end{align}
The factorized form is generally different from an ansatz that puts all operators in one exponential. But, because we are doing a variational calculation and there is flexibility given by the needed chemical accuracy, this factorized form is usually sufficient to achieve chemical accuracy, if enough factors are included. Indeed, if factors are repeated, it can be used to approximate the original UCC ansatz via the Trotter product formula.
For a general UCC factor, we derived a general operator identity based on a hidden SU(2) algebra.\cite{evangelista,luogen,jia} It is
\begin{align}\label{operator_identity}
\begin{split}
& U^{a_1\dots a_n}_{i_1\dots i_n} = \exp[\theta(\hat{a}^{a_1\dots a_n}_{i_1\dots i_n}-\hat{a}^{i_1\dots i_n}_{a_1\dots a_n})] = 1 + \sin\theta (\hat{a}^{a_1\dots a_n}_{i_1\dots i_n}-\hat{a}^{i_1\dots i_n}_{a_1\dots a_n}) \\
&+(\cos\theta-1)[\hat{n}_{a_1}\dots\hat{n}_{a_n}(1-\hat{n}_{i_1})\dots(1-\hat{n}_{i_n}) +(1-\hat{n}_{a_1})\dots(1-\hat{n}_{a_n})\hat{n}_{i_1}\dots\hat{n}_{i_n}].
\end{split}
\end{align}
Here the number operators are of the form $\hat{n}=\hat{a}^\dagger\hat{a}$.
If rotation angle $\theta = 0$, the UCC factor becomes the identity: $U(0)=\mathbb{I}$.
We can take the derivative of Eq.~(\ref{operator_identity}):
\begin{align}\label{operator_identity_2}
\begin{split}
&\frac{d \hat{U}^{a_1\dots a_n}_{i_1\dots i_n} (\theta)}{d \theta} = \cos\theta (\hat{a}^{a_1\dots a_n}_{i_1\dots i_n}-\hat{a}^{i_1\dots i_n}_{a_1\dots a_n}) \\
&-\sin\theta [\hat{n}_{a_1}\dots\hat{n}_{a_n}(1-\hat{n}_{i_1})\dots(1-\hat{n}_{i_n}) +(1-\hat{n}_{a_1})\dots(1-\hat{n}_{a_n})\hat{n}_{i_1}\dots\hat{n}_{i_n}].
\end{split}
\end{align}
When we evaluate the derivative at $\theta =0$, it becomes the operator term in the exponent, namely
\begin{align}\label{Ud0}
\frac{d \hat{U}^{a_1\dots a_n}_{i_1\dots i_n} (\theta)}{d \theta} |_{\theta=0} = \hat{a}^{a_1\dots a_n}_{i_1\dots i_n}-\hat{a}^{i_1\dots i_n}_{a_1\dots a_n} = \hat{\sigma}^{a_1\ldots a_n}_{i_1\ldots i_n}.
\end{align}
The derivative of the wavefunction with respect to one rotation angle immediately follows as
\begin{align}\label{ucc_first_derivative}
\frac{ d U_{\textrm{UCC}}}{d \theta_k} = \hat{U}_n\cdots\hat{U}_{k+1}\frac{d \hat{U}_k}{d \theta_k}\hat{U}_{k-1}\cdots\hat{U}_2\hat{U}_1 .
\end{align}
Evaluated at $\theta_k=0$, a simple result is obtained:
\begin{align}
\frac{ d U_{\textrm{UCC}}}{d \theta_k} \Big{|}_{\vec{\theta} =0}= \frac{d \hat{U}_k}{d \theta_k} \Big{|}_{\theta_k=0} = \hat{\sigma}_k.
\end{align}
\subsection{Quadratic Angle Expansion}
A Taylor expansion of the expectation value of the energy around the point where all angles vanish ($\vec{\theta}=0$) becomes
\begin{align}\label{tot_energy}
\langle \hat{H} (\vec{\theta})\rangle = \langle \hat{H} (0)\rangle + \sum_{k} b_k \theta_k + \frac{1}{2}\sum_{k,m} A_{km}\theta_k\theta_m + \mathcal{O}(\theta^3)
\end{align}
where,
\begin{align}\label{b_vec}
b_k &=\frac{d\langle \hat{H}(\vec{\theta})\rangle}{d\theta_k}\Big{|}_{\vec{\theta} =0} = \langle \Psi_0| \hat{U}^{\dagger}(0)\hat{H}\frac{d \hat{U}(0)}{d\theta_k}|\Psi_0\rangle+\langle \Psi_0|\frac{d \hat{U}^{\dagger}(0)}{d\theta_k}\hat{H} \hat{U}(0)|\Psi_0\rangle\nonumber\\
&= \langle \Psi_0| \hat{H} \hat{\sigma}_k|\Psi_0\rangle+\langle \Psi_0| \hat{\sigma}_k^\dagger \hat{H} \hat|\Psi_0\rangle=2\text{Re}\langle \Psi_0|\hat{H}\hat{\sigma}_k|\Psi_0\rangle,
\end{align}
and
\begin{align}\label{Akm}
A_{km} = \frac{d^2\langle \hat{H}(\vec{\theta})\rangle}{d\theta_k d\theta_m}\Big{|}_{\vec{\theta} =0} &= \langle \Psi_0| \hat{U}^{\dagger}(0)\hat{H} \frac{d^2 \hat{U}(0)}{d\theta_k d\theta_m} |\Psi_0\rangle +2 \langle \Psi_0| \frac{\hat{U}^{\dagger}(0)}{ d\theta _k}\hat{H}\frac{d \hat{U}(0)}{d\theta_ m}|\Psi_0\rangle\nonumber\\
&+ \langle \Psi_0| \frac{d^2 \hat{U}^{\dagger}(0)}{d\theta_k d\theta_m}\hat{H} \hat{U}(0)|\Psi_0\rangle \nonumber\\
&= 2 \langle \Psi_0|\hat{\sigma}^{\dagger}_k \hat{H} \hat{\sigma}_m|\Psi_0\rangle + 2 \text{Re}\langle \Psi_0| \hat{H} (\hat{\sigma}_k \hat{\sigma}_m)|\Psi_0\rangle.
\end{align}
Note that the real part is not necessary in most calculations, since the wavefunction is usually expanded in terms of real coefficients; this can change in the presence of a magnetic field or with spin-orbit coupling.
The ordering of the operators, however, is important. Note that in
Eq.~(\ref{Akm}), we have used the following notation:
\begin{align}
(\hat{\sigma}_k \hat{\sigma}_m)=
\begin{cases}
\hat{\sigma}_k \hat{\sigma}_m,& \text{if } m\geq k\\
\hat{\sigma}_m \hat{\sigma}_k & \text{otherwise}
\end{cases}.
\end{align}
The ordering of the UCC factors matters in the second derivative matrix because of second term in Eq.~(\ref{Akm}). The de-excitation operator may also apply when the first operator can be de-excited. Since $\hat{\sigma}_k|\textrm{HF}\rangle$ and $\hat{\sigma}_k\hat{\sigma}_m|\textrm{HF}\rangle$ are both single determinants, the above expressions for $A$ and $b$ are just many-body Hamiltonian matrix elements in the Hartree-Fock basis and with respect to many-body product states (determinants). This means one can perform an initial minimization, about the point where all angles are zero, from the Hamiltonian, expressed in the Hartree-Fock basis and extended to include matrix elements for all states that are required in the $b_k$ vector and the $A_{km}$ matrix.
To minimize the energy with respect to the angles, we take the derivative of Eq.~(\ref{tot_energy}) about the point where each $\theta_i=0$ and set the derivative to zero in order to find the minimum energy.
We have
\begin{align}\label{dev}
\frac{d\langle \hat{H} (\vec{\theta})\rangle}{d\theta_i}\Big |_{\vec{\theta}=0} = b_i + \sum_j A_{ij} \theta_j = 0
\end{align}
In matrix form, the angles that minimize the energy are the solution of the following system of linear equations:
\begin{align}\label{linear_system}
\mathbf{A} \cdot \vec{\theta} = -\vec{b}.
\end{align}
Angles that minimize the energy can be obtained either by inversion of the matrix $\mathbf{A}$, or by solving linear system Eq.~(\ref{linear_system}) by row-reduction (or by using sparse-matrix iterative techniques).
The correlation energy, which is the difference between the total energy and the Hartree-Fock energy becomes
\begin{align}
E_{corr} = \langle \hat{H}(\vec{\theta}) \rangle - \langle \hat{H}(0) \rangle = \vec{b}\cdot\vec{\theta}_{min} + \frac{1}{2}\vec{\theta}^{T}_{min} \cdot \mathbf{A} \cdot \vec{\theta}_{min}.
\end{align}
Note how this calculation is quite straightforward and simple to carry out, it just requires generating Hartree-Fock matrix elements and solving the linear matrix equation.
\subsection{Exact UCC Reference}\label{tree_reference}
For this quadratic expansion to be accurate, it is necessary for the angles to be small. This is not the case for strongly correlated systems. To extend this scheme to correlated systems, we can treat factors with large angles more carefully. After separating factors with large angles $\{\hat{\sigma_l}\}$ and small angles $\{\hat{\sigma_s}\}$ into two groups, UCC factors with large angles are used to construct an exact UCC reference wave-function and angles are optimized:
\begin{align}\label{ucc_ref}
|\Psi^{\textrm{UCC}}_0\rangle = \prod_l e^{\theta_l\hat{\sigma_l}} |\textrm{HF}\rangle.
\end{align}
This reference state naturally contains more than one determinant, and its contribution to the correlation energy can be calculated as
\begin{align}
E_{0}^{corr} = \langle \Psi^{\textrm{UCC}}_0 | \hat{H}|\Psi^{\textrm{UCC}}_0\rangle -E_{\textrm{HF}}.
\end{align}
In the second step, angles for all UCC factors are expanded to second order, as described in last section. The difference is that, for factors used in the UCC reference, their angles are expanded with respect to their optimized values, instead of around zero. The presence of these nonzero angles in the reference state affects the computation of the $b$ vector and the $A$ matrix for all of their elements. The total energy after the quadratic expansion becomes, in this case,
\begin{align}
\langle \hat{H} (\vec{\theta})\rangle = \langle \Psi^{\textrm{UCC}}_0| \hat{H} | \Psi^{\textrm{UCC}}_0 \rangle + \sum_{k} b_k \theta_k + \frac{1}{2}\sum_{k,m} A_{km}\theta_k\theta_m + \mathcal{O}(\theta^3)
\end{align}
Again, $b_k$ and $A_{km}$ can be obtained by taking the corresponding derivatives. But, the two groups of UCC factors, corresponding to large and small angles, have different expressions for the derivatives. For the vector $b_k$, if $k$ is within the group of small angles, elements can be calculated simply by replacing the reference state $|\Psi_0\rangle$ in Eq.~(\ref{b_vec}) by $|\Psi_0^{\textrm{UCC}}\rangle$ as defined in Eq.~(\ref{ucc_ref}); otherwise, $b_k$ can be calculated as a derivative of UCC wave-function, by inserting Eq.~(\ref{operator_identity_2}) into Eq.~(\ref{ucc_first_derivative}). We have an exact expression for the derivative of each UCC factor. Furthermore, we know exactly where to place the derivative operator within the calculation. The matrix $A_{km}$ can be obtained in a similar way---if two factors $k,m$ are both small angle factors, then Eq.~(\ref{Akm}) can be employed with $|\Psi_0^{\textrm{UCC}}\rangle$ replacing $|\Psi_0\rangle$ . If one or two of $k,m$ are large angle factors, derivatives with respect to $|\Psi_0^{\textrm{UCC}}\rangle$ need to be taken, just like the case for $b_k$, already discussed. It is a simple exercise to work out the exact formulas (we do not write them out here).
The energy minimization is again carried out to find $\theta_{min}$ for both large and small angles. Then the large angles used to construct the UCC reference wave-function are updated by $\theta'_l = \theta_l + \theta_l^{min}$. These $\theta'_l$ values are then used to construct the new UCC reference wavefunction, which is used for the new quadratic angle expansion. This procedure is iterated until the calculation has converged to a fixed point for both the large and the small angles. The total correlation energy now has three contributions:
\begin{align}
E_{corr} = E_{0}^{corr} + \vec{b}\cdot\vec{\theta}_{min} + \frac{1}{2}\vec{\theta}^{T}_{min} \cdot \mathbf{A} \cdot \vec{\theta}_{min}.
\end{align}
The computational procedure described above can be carried out as a hybrid quantum-classical algorithm. The preparation of the UCC reference state needs to be performed on a quantum computer. Then the quantities $b_k$ and $A_{km}$ are measured with respect to the prepared UCC reference state. As we will show later, only a small number of factors are required in the UCC reference state, which greatly reduces the circuit depth on a quantum computer. As a trade off, the $b_k$ and $A_{km}$ require many more measurements. But of course, even a standard VQE procedure will require calculating at least something like the $b_k$ for a gradient-based minimizer, or to choose the best operator to pick from an operator pool. If this tradeoff is worthwhile requires a more detailed analysis. But, note that there are likely some measurements that can be replaced by cubic expansion about $\vec{\theta}=0$, because they might not depend strongly on the large angles. This analysis can be performed on a classical computer prior to the quantum computation and can greatly reduce the required measurements.
One potential issue in this procedure is the appearance of an instability in the minimization step for strongly correlated cases. The origin of this instability is that the inverse of $\mathbf{A}$ may be ill-defined (due to zero or near zero eigenvalues). To overcome this, we use a pseudo-inverse of $\mathbf{A}$, which is constructed as follows. First, $\mathbf{A}$ is diagonalized via a similarity transformation with the matrix $Q$: $\mathbf{A} = \mathbf{Q} \mathbf{\Lambda} \mathbf{Q}^{-1}$. Then, a cutoff $\epsilon$ is applied to the eigenvalues included in $\Lambda$. If an eigenvalue is smaller than $\epsilon$, the corresponding diagonal element of $\Lambda^\prime$ is set to zero. The reduced set of eigenvalues is placed in the diagonal matrix $\Lambda'$, and the pesudoinverse of $\mathbf{A}$ is calculated as $\mathbf{A}^{-1}\approx \mathbf{Q} \mathbf{\Lambda'}^{-1} \mathbf{Q}^{-1} $, where the terms in $\bf{\Lambda}^\prime$ that were set equal to zero are \textit{not} inverted in computing the pseudoinverse---they remain set equal to zero. Thisapproach is the same as a singular-value decomposition of a matrix, relative to a cutoff $\epsilon$. Then, the angles that minimize the energy are found from $\theta_{min} = -\mathbf{A}^{-1}
\cdot \vec{b} $.
\section{Results and Discussion}
\subsection{Near equilibrium}
We implemented this quadratic UCC (qUCC) method using integrals generated by {\it PySCF}. We performed calculations on a set of small molecules with the {\it ccpvdz} basis set. To benchmark the calculations, we use almost exact energies from the semistochastic heat-bath configuration interaction calculation as a reference.\cite{yao_2020} The results are summarized in Fig.~\ref{tt}.
While an exact UCC calculation is always variational, the qUCCSD approximation need not have its energy bounded by the full configuration interaction (FCI) from below, because of the potential error from the truncated Taylor expansion. This means it need not be strictly variational. For several molecules in Fig.~\ref{tt}, the errors are indeed negative. However, if we compare the absolute error, qUCCSD is more accurate than CCSD, which has similar computational scaling. Comparing to CCSD(T), qUCCSD is not as accurate, but is not expected to be either. What is surprising is that qUCCSD is consistently not that far away from CCSD(T).
\begin{figure}
\centering
\includegraphics[width=0.85\linewidth,clip]{eq.eps}
\caption{ Error, (E-E$_{FCI}$), from CCSD, CCSD(T) and qUCCSD calculations on a subset of Gaussian-2 molecules. The basis set is {\it ccpvdz}. O$_2$ and S$_2$ are spin triplets, and NO is a spin doublet. All others are spin singlets. } \label{tt}
\end{figure}
\subsection{Chemical Bond Stretching}
Thje quadratic expansion of the unitary coupled cluster theory using HF ground state as reference, as described earlier, shows many similar characteristics with the older XCC approaches \cite{Bartlett1989}, and other linearized coupled-cluster methods \cite{Bartlett2007} developed by Rodney Bartlett and his colleagues. But even at second order, the derivatives from a factorized form of the UCC depend on the ordering, while for a conventional UCC, they do not. The common weakness of the XCC methods is that they diverge when near-degeneracy is present in the HF states, so they cannot treat level crossings or near level crossings in the potential energy surface. This often occurs for bond stretching and breaking situations. One way to overcome this difficulty for the quadratic expansion of the factorized form of the unitary coupled cluster theory is to use multi-determinant reference states, as detailed in Section.~\ref{tree_reference}.
To demonstrate this method, we performed calculations for the bond stretching of a H$_2$O molecule with the 6-31G basis set. In total, this system has 74 single and 2240 double excitations. We first did a geometry optimization, and then optimized the H-O bond length. It becomes 0.96\AA ~and the optimized H-O-H angle is 104.5$^{\circ}$. To study the situations of bond stretching, we gradually increase the bond from its equilibrium length to 2.74\AA, while keeping the H-O-H angle fixed. Standard quantum chemistry methods inclduing full configuration interaction (FCI) and coupled-cluster singles and doubles (CCSD) were carried to test against the qUCCSD with exact UCC reference states. Here all the qUCCSD calculations were performed with 28 factors of large angles and $\epsilon = 0.1$ for pseudo-inverse procedure. Those 28 large angles factors are determined from initial MP2 amplitudes. Results can be found in Fig.~\ref{ucc_reference}.
As expected, CCSD correlation energies are close to the exact FCI results when the molecular geometries are close to equilibrium. But, CCSD has difficulties when away from equilibrium; CCSD violates the variational principle and gives lower energies then FCI when O-H bond length is larger then 2.2 \AA. We also find that qUCCSD with UCC reference usually gives better results than CCSD. Close to equilibrium, qUCCSD gives similar results as CCSD, and when far equilibrium, qUCCSD with UCC reference states has better behaviour then CCSD, since it never violates variational principle.
\begin{figure}
\centering
\includegraphics[width=0.85\linewidth,clip]{corr_error.eps}
\caption{ Panel (a) and panel (b): total correlation energies for the water molecule as function of the H-O bond length. Here, we plot results for FCI (black circles), CCSD (red squares) and qUCCSD (blue diamonds) with a UCC reference state (that contains 28 exact UCC factors). Note that the ranges of the plots differ in the two panels. Panel (c) error of the CCSD (red squares) and qUCCSD (blue diamonds) results relative to the FCI as a function of the H-O bond length. Note how the CCSD calculation becomes nonvariational at strong coupling. } \label{ucc_reference}
\end{figure}
\section{Conclusion}
One of the main issues slowing down the ability for quantum computers to show an advantage over classical computers is that current hardware can only run codes that are quite short in circuit depth. In this situation, the only advantage will occur due to the larger memory that a quantum computer has in storing quantum states. The algorithm discussed here, denoted qUCC, is a methodology that will allow quantum computers to aid in determining the electronic structure of complex molecules much sooner, because it trades off circuit depth for additional measurements. If those additional measurements can be carried out, then we might see a quantum advantage sooner than later.
To illustrate how this methodology works, we showed that the number of exact UCC factors needed in an electronic structure calculation on a quantum computer can be greatly reduced from that of a standard approach. Our test case (bond stretching of water molecule) shows better results can be achieved with only a small fraction of factors (28 versus 2314) when constructing the wavefunction. Our result compares the qUCC approach to a CC approach, but results would be similar for a comparison to a standard UCC approach as well. Our algorithm is one way to utilize low-depth quantum circuits to treat molecules with large basis sets, which has been a major obstacle for applying quantum computing to quantum chemistry.
\begin{acknowledgement}
JC and HPC are supported by the Department of Energy, Basic Energy Sciences as part of the Center for Molecular Magnetic Quantum Materials, an Energy Frontier Research Center under Award No. DE-SC0019330. JKF is supported from the National Science Foundation under grant number CHE-1836497. JKF is also funded by the McDevitt bequest at Georgetown University. This research used resources of the National Energy Research Scientific Computing Center (NERSC), a U.S. Department of Energy Office of Science User Facility operated under Contract No. DE-AC02-05CH11231, and University of Florida Research Computing systems.
We also acknowledge useful discussions with Rodney Bartlett, Garnet Chan, Joseph Lee, John Staunton, Cyrus Umrigar, Luogen Xu and Dominika Zgid.
\end{acknowledgement}
|
\section{Introduction}
In Deep Learning, the recurrent neural network (RNN) is well-known as one of the most popular models to model sequential data and is widely used in practice for tasks in natural language processing (NLP). One of the characters of RNN is that it performs the same operation for all the input of the sequence.
Consider a input sequence $x=(X_1,X_2,...,X_L)$. A RNN with the form
\begin{equation}\label{rnne}
h_{l}(x)=\phi(\bm{W}h_{l-1}+\bm{A}X_l),
\end{equation}
is trying to learn functions $f_l(X_1, X_2,...X_l)$ as
\begin{equation}
\begin{aligned}
&h_{1}(x)=f_1(X_1)\\
&h_{2}(x)=f_2(X_1, X_2)\\
&\ \ \ \ \ \ \ \ \ \ \ \ \ \ \vdots\\
&h_{L}(x)=f_L(X_1,X_2,...X_L)
\end{aligned}
\end{equation}
Due to the complex nonlinearity, the loss is generally non-convex, and it is very difficult to give a theoretical guarantee. Recently, there are some works \cite{allenzhu2019convergence,cao,NEURIPS201962dad6e2,DBLP,sa,rate} trying to give a theoretical explanation that why gradient descent can allow an overparametrized network to attain arbitrarily low training error and ample generalization ability. These papers show that, under some assumptions, we have:
\begin{itemize}
\item[$\bullet$] \emph{Multi-layer feed-forward networks \cite{allenzhu2019convergence,DBLP} and recurrent neural networks \cite{rate} with large hidden size can attain zero training error, regardless of whether the data is properly labeled or randomly labeled.}
\item[$\bullet$] \emph{For multi-layer feed-forward networks, functions with the form $F^*(x)=\sum_{r=1}^{C} \phi_{r}(\beta^T_{r}X), X\in\mathbb{R}^d,\beta_r\in\mathbb{R}^d, ||\beta_r||=1$
are learnable i.e. fitting the training data with a provably small generalization error, if $\phi$ is analytic and the ``complexity'' is low enough \cite{NEURIPS201962dad6e2,sa,cao}.}
\item[$\bullet$] \emph{The ``complexity'' of function $\phi$ can be measured by a matrix derived from the NTK (Neural Tangent Kernel) of the network \cite{sa,cao}.}
\item[$\bullet$] \emph{For recurrent neural networks \cite{allenzhu2019sgd}, if the input sequence is normalized, i.e., $x=(X_1,X_2,...,X_L)$, $||X_1||=1$, $||X_l||=\epsilon$ with $\epsilon$ very small, functions with the form $F^*(x)=\sum_{l=1}^L \sum_{r=1}^{C_l} \phi_{l,r}(\beta^T_{l,r}X_l)$ are learnable, where $m$ is the size of matrix $\bm{W}$, and $\mathscr{C}=\sum_{i=0}^\infty a_iR^i $ is a series representing the complexity of learnable functions.}
\end{itemize}
These works show the provable learning ability of deep learning.
\begin{comment}
But there are still some issues that were not addressed in these papers. For RNNs, the method in \cite{allenzhu2019sgd} requires a normalized condition for $\bm{A}$ and $X_l$ in (\ref{rnne}) that $||\bm{A}X_l||\leq \epsilon_x$ and shows that, for a function $F^*(x)$ with complexity $\mathscr{C}$, it is learnable with error $O(\epsilon^{1/3}_x\mathscr{C})$. Thus $ ||X_l||$ (or equally, $||\bm{A}||$) should be very small and the scale is dependent on the complexity of functions. The dependence of $||\bm{A}X_l||$ on $\mathscr{C}$ makes the results unrealistic in practice.
On the other hand, the result in \cite{allenzhu2019sgd} shows that RNNs can learn functions which are the summation of functions like $\psi(\beta^T_{l}X_l)$. One may ask, since $h_{L}(x)$ is a function of $\{X_1,X_2,...X_L\}$, is it possible to go beyond and learn more complex functions?
\end{comment}
But there are still some important issues that were not addressed.
\begin{itemize}
\item[$\bullet$] Firstly, for RNNs, the method in \cite{allenzhu2019sgd} requires a normalized condition for $\bm{A}$ and $X_l$ in (\ref{rnne}) that $||\bm{A}X_l||\leq \epsilon_x$ for all $l\leq L$ and shows that for a function $F^*(x)$ with the complexity $\mathscr{C}$, it is learnable with error $O(\epsilon^{1/3}_x\mathscr{C})$. Thus $ ||X_l||$ (or equally, $||\bm{A}||$) should be very small and the scale is dependent on the complexity of functions. The dependence of $||\bm{A}X_l||$ on $\mathscr{C}$ makes the results unrealistic in practice since generally the norm of input will not be so small.
\item[$\bullet$]Secondly, the result in \cite{allenzhu2019sgd} shows that RNNs can learn functions which are the summation of functions like $\psi(\beta^T_{l}X_l)$. But this is only a linear combination of the functions of the input at different positions and does not consider the nonlinear interaction of the inputs. One may ask, since $h_{L}(x)$ is a function of $\{X_1,X_2,...X_L\}$, is it possible to go beyond and learn more complex functions?
\end{itemize}
In order to study these problems, we consider the binary classification problem: for every input $x_i$, the label ($+1$ or $-1$) of $x_i$ can be expressed by the sign of a target function $F^*(x_i)$. We consider Elman recurrent neural networks with ReLU activation
\begin{equation}\label{rnn}
\begin{aligned}
&h_{l}(x)=\phi(\bm{W}h_{l-1}+\bm{A}X_l)\\
&f(\bm{W},x)=\bm{B}^Th_L(x)\in\mathbb{R}.\\
&x=(X_1,X_2,...,X_L),X_l\in\mathbb{R}^d,\bm{W}\in\mathbb{R}^{m\times m},\\
&\bm{A}\in\mathbb{R}^{m\times d},\bm{B}\in\mathbb{R}^m,\phi(x)=\max (x,0)
\end{aligned}
\end{equation} to learn two types of target functions:
\begin{itemize}
\item[$\bullet$]{ Additive Concept Class:}
\begin{equation}\label{cc1}
\begin{aligned}
&F^*(x)=\sum_{l=1}^{L} \sum_{r=1} \psi_{l,r}(\beta^T_{l,r}X_l/||X_l||),\\
&\psi_{l,r}(x)=\sum_{i=0}^\infty c_ix^{i},\\
\end{aligned}
\end{equation}
\item[$\bullet$]{N-variables Concept Class:}
\begin{equation}\label{cc2}
\begin{aligned}
&F^*(x)=\sum_{r}\psi_{r}(\langle \beta_{r}, [X_{l_1},...,X_{l_{N}}]\rangle),\\
&\psi_{r}(x)=\sum_{i=0}^\infty c_ix^{i}.
\end{aligned}
\end{equation}
\end{itemize}
For these two types of function, we study the following questions:
\begin{itemize}
\item[$\bullet$] Can RNN learn additive concept class functions (\ref{cc1}) without the normalized condition with {\bf reasonable complexity on the sequence size $L$}?
\item[$\bullet$] Can RNN learn functions in N-variables Concept Class (\ref{cc2}) which can not be written as the summation of $f(X_l)$ with {\bf reasonable complexity on $N$ and $L$}?
\end{itemize}
{\bf Our Result. }
We answer the two questions and give a provable generalization error bound. Our results are stated as follows:
\begin{theo}(Informal) For a function $F^*(X_1,X_2,...,X_L)$ with the form as in (\ref{cc1}) or (\ref{cc2}), there is a power series named the complexity $\mathscr{C}(F^*)$ dependent on the Taylor expansion coefficient in (\ref{cc1}) and (\ref{cc2}). For (\ref{cc1}), $\mathscr{C}(F^*)$ is almost-polynomial in $L$. For (\ref{cc2}), when $N$ or $l_0=\max(l_1,..,l_N)-\min(l_1,..,l_N)$ is small, $\mathscr{C}(F^*)$ is almost-polynomial in $L$. Under this definition of complexity $\mathscr{C}(F^*)$, $F^*$ is learnable using RNN with $m$ hidden nodes and ReLU activation in (\ref{rnn}) in $\mathcal{O}(\mathscr{C}(F^*)^2)$ steps with $\mathcal{O}(\mathscr{C}(F^*)^2)$ samples if $m\geq poly(L,\mathscr{C}(F^*))$.
\end{theo}
{\bf Contribution.} We summarize the contributions as follows:
\begin{itemize}
\item[$\bullet$] \emph{In this paper, we prove that RNN without normalized condition can efficiently learn some notable concept classes with {\bf both time and sample complexity scaling almost polynomially in the input length $L$}.}
\item[$\bullet$] \emph{Our results go beyond the ``additive'' concept class. We prove a novel result that RNN can learn more complex function of the input such as N-variables concept class functions. And ``long range correlation functions'' with small $N$ (e.g. $N=2$, $f(\beta^T[X_l,X_{l+l_0}])$ ) are learnable with {\bf complexity scaling almost polynomially in the input length $L$ and correlation distance $l_0$}.}
\item[$\bullet$] \emph{Technically, we study the ``backward correlation'' of RNN network. In RNN case, using a crucial observation on the degeneracy of deep network, we show that the {\bf ``backward correlation'' $\frac{1}{m}\langle \text{Back}_l(x_i),\text{Back}_{l}(x_j) \rangle$ will decay polynomially rather than exponentially in input length $L$}. This shows the complexity of learning RNN with ReLU activation function is polynomial in the size of input sequence $L$.}
\end{itemize}
{\bf Notions.}
For two matrices $\bm{A}, \bm{B}\in \mathbb{R}^{m\times n}$, we define $\langle \bm{A}, \bm{B} \rangle= \text{Tr}(A^TB)$. We define the asymptotic
notations $\mathcal{O}(\cdot), \Omega(\cdot),poly(\cdot)$ as follows. $a_n, b_n$ are two sequences. $a_n=\mathcal{O}(b_n)$ if $\lim \sup_{n\to \infty}|a_n/b_n|< \infty$, $a_n=\Omega(b_n)$ if
$\lim \inf_{n\to \infty}|a_n/b_n|>0$, $a_n=poly(b_n)$ if there is $k\in \mathbb{N}$ that $a_n=O((b_n)^k)$. $\widetilde{\mathcal{O}}(\cdot), \widetilde{\Omega}(\cdot), \widetilde{poly}(\cdot)$ are notions which hide the logarithmic factors in $\mathcal{O}(\cdot), \Omega(\cdot), poly(\cdot)$. $||\cdot ||$ and $||\cdot||_2$ denote the 2-norm of matrices. $||\cdot ||_1$ denote the 1-norm. $||\cdot ||_F$ is the Frobenius-norm. $||\cdot||_0$ is the number of non-zero entries.
For elements $A_{i.j}, B_{i,j}$ of symmetric matrix $\bm{A}, \bm{B}$. We abuse the notion $A_{i.j}\succeq B_{i.j}$ to denote $\bm{A}\succeq \bm{B}$, i.e. $\bm{A}-\bm{B}$ is a positive semidefinite matrix.
\section{Preliminaries}
\subsection{Function Complexity}\label{fc}
For a analytic function $\psi(z)$, we can write it as $\psi(z)= c_0+ \sum_{i=1}^\infty c_{i}z^{i}$. We define the following notion to measure the complexity to learn such functions.
\begin{equation}
\mathscr{C}(\psi,R)= 1+\sum_{i=1}^\infty i\cdot |c_{i}|R^{i}.
\end{equation}
\begin{equation}
\mathscr{C}_N(\psi, R)= 1+\sum_{i=1}^\infty L^{1.5N} C_1^N\cdot \sqrt{C_{N,i}}\cdot (i/N)^N \cdot |c_{i}|R^{i}
\end{equation}
where $C_1>100$ is an large absolute constant and $C_{N,i}$ is the largest combination number $\frac{i!}{n_1!n_2!...n_N!}$ for $n_1,n_2...n_N>0, n_1+n_2+...n_N=i$,
\begin{exa} \cite{sa}
Consider $\psi(z)= arctan(z/2)$. Then
\begin{equation}
\psi(z)=\sum_{i=1}\frac{(-1)^{i-1}2^{1-2i}}{2i-1}z^{2i-1}
\end{equation}
In this case, $$\mathscr{C}(\psi,1)=1+\sum_{i=1}^\infty i\cdot |c_{i}|\leq 1+\sum_{i=1}^\infty 2^{1-2i}\leq \mathcal{O}(1).$$
\end{exa}
\begin{exa}
In the case $N=2$, $C_{2,i}=i$, $(i/2)^2\leq i^2$. $\psi(z)= exp(z)$
$$\mathscr{C}_2(\psi,1)\leq1+\sum_{i=1}^\infty L^3C_1^2 \pi i^{2.5}/i!\leq \mathcal{O}(1)$$
\end{exa}
\subsection{Concept Class}\label{fcc}
For the input sequence $\{X_l\}$, we assume $C_{min}\leq ||X_l||\leq C_{max}$, for all $1\leq l\leq L$ and $C_{max}/C_{min}\sim C_0$. Under this condition, we consider two types of target functions with the following form:\\
{\bf Additive Concept Class.}
\begin{equation}\label{cop}
F^*(x)=\sum_{l=1}^{L} \sum_{r=1}^{C_l} \psi_{l,r}(\beta^T_{l,r}X_l/||X_l||).
\end{equation}
Here for all $l,r$, $\psi_{l,r}$ is analytic and $||\beta_{l,r}||_2\leq 1$.
We define
\begin{equation}\label{com1}
\mathscr{C}(F^*)=L^{3.5}\sum_{l=1}^L \sum_{r=1}^{C_l} \mathscr{C}(\psi_{l,r},C_0\sqrt{ L}),
\end{equation}
to be the complexity of the target function.
\begin{rem}
If we consider function $\psi(\beta^TX_l)$ and $||X_l||=1$ for all $l$, the above complexity will become $\mathscr{C}(\psi,\mathcal{O}(\sqrt{L}))$. This is similar with that in \cite{allenzhu2019sgd} but this complexity requirement is much weaker than that in \cite{allenzhu2019sgd}. For example, the complexity of $arctan(z/2)$ in \cite{allenzhu2019sgd} is not finite, as shown in \cite{sa}.
\end{rem}
{\bf N-variables Concept Class.}
\begin{equation}\label{cop2}
F^*(x)=\sum_{r}\psi_{r}(\langle \beta_{r}, [X_{l_1},...,X_{l_{N}}]\rangle/\sqrt{N}\max ||X_{l_n}||).
\end{equation}
For all $r$, $\psi_{l,a,r}(x,y)$ is an analytic function $\psi_{r}(x)=c_0+ \sum_{i=1}^\infty c_ix^{i}$. $\beta_{r}\in \mathbb{R}^{dN}$, $||\beta_{r}||_2\leq 1$. Let $l_0=\max(l_1,..,l_N)-\min(l_1,..,l_N)$.
We define
\begin{equation}\label{com2}
\mathscr{C}(F^*)=\min(L^{2}\mathscr{C}_N(\psi_{r},C_0\sqrt{ L}), L^{3.5}\mathscr{C}(\psi_{r},2^{l_0}C_0\sqrt{ L})).
\end{equation}
\begin{rem}
The complexity $\sum_{r} \mathscr{C}_N(\psi_{r},C_0\sqrt{ L})$ and $\sum_{r}\mathscr{C}(\psi_{r},2^{l_0}C_0\sqrt{ L})$ are exponential in $N$ and $l_0$ respectively. And $\mathscr{C}(F^*)$ is less or equal than both. Thus if either $l_0$ or $N$ is small, $\mathscr{C}(F^*)$ will be polynomial in $L$. Especially when $N$ is small(e.g. N=2), even if $l_0=L-1$, functions with the form $f(\beta^T[X_l,X_{l+l_0}])$ are still learnable with a low complexity.
\end{rem}
\begin{comment}
\subsection{Data Set}\label{ds}
We consider the two classification problem $x=(X_1,X_2,...,X_L)$, $||X_l||\leq \mathcal{O}(1)$ and $y=\pm 1$. If this condition is not satisfied, it is equal scale down matrix $\bm{A}$.
A distribution ${\cal D}=\{x\in \mathbb{R}^d,y\in \mathbb{R}\}$ is called $(\delta,\lambda_0)$-non-degenerate if there are $n$ i.i.d. samples $\{x_i,y_i\}_{i=1}^n$ sampling randomly from ${\cal D}$, with probability at least $1-\delta$, for the matrix $\bm{M}\in \mathbb{R}^{n\times n}$ with
$$M_{ij}=\mathbb{E}_{w\sim N(0,I_d)} \mathbbm{1}\{w^Tx_i>0\}\mathbbm{1}\{w^Tx_j>0\},$$
we have $\lambda_{min}(M)\geq \lambda_0$. Using the methods in \cite{JMLR:v7:braun06a}, let $\pi(x)$ be the distribution function of $x$ in data set ${\cal D}$. $\{x_i\}^n$ are independently drawn from $\pi$. Then $\lambda_0$ is roughly the $n$-th eigenvalue of the Hilbert-Schmidt integral operator:
\begin{equation}
T_kf(x)=\int k(x,x')f(x')d\pi(x').
\end{equation}
Using this results, \cite{bietti2021deep} has provided an estimation on $\lambda_0$ if the distribution function of ${\cal D}$ is the uniform distribution on the d-dimension sphere by considering the spherical harmonics function expansion and compare the coefficients.
\end{comment}
\subsection{Results on Positive Definite Matrices and Functions}
We say a function $\phi(\cdot,\cdot):\mathbb{R}^d\times \mathbb{R}^d\to \mathbb{R}$ is positive definite if for all $n \in \mathbb{N}$, any $\{x_1,...,x_n\}\subseteq \mathbb{R}^d, \{c_1,...,c_n\}\subseteq \mathbb{R}$,
\begin{equation}
\sum_{i,j}c_ic_j\phi(x_i,x_j)\geq 0.
\end{equation}
The following basic properties in chapter 3 of \cite{Christian} are very useful in our proof.
\begin{pro}
If $\phi(\cdot,\cdot)$ is positive definite function, let matrix $\bm{M}\in \mathbb{R}^{n\times n}$, $\{x_1,...,x_n\}\subseteq \mathbb{R}^d$, and $M_{i,j}= \phi(x_i,x_j)$. Then $\bm{M}$ is a semi-positive definite matrix.
\end{pro}
\begin{pro}\label{c1}
If $\phi_1(\cdot,\cdot)$ and $\phi_1(\cdot,\cdot)$ are positive definite, $\phi(x_i,x_j)=\phi_1(x_i,x_j)\cdot \phi_2(x_i,x_j)$ is also a positive definite function.
\end{pro}
\begin{pro}\label{c2}
Let $\phi(\cdot, \cdot)$ be a positive definite function, and $\psi(x)=\sum_{i=0}^\infty c_ix^i$, $c_i\geq 0$. Then $\psi(\phi(\cdot,\cdot))$ is also a positive definite function.
\end{pro}
For a positive definite matrix $\bm{M}\in \mathbb{R}^{n\times n}$, there is a result in \cite{sa},
\begin{pro}\label{claa}(Section E of \cite{sa}.)
Let $\bm {X}=(x_1,...x_n)\in \mathbb{R}^{d\times n}$ and $\bm{K}_{p}\in \mathbb{R}^{n\times n}$ is a matrix with $(K_{p})_{i,j}=(x_i^Tx_j)^{p}$. Suppose there is $\alpha>0$, such that $\bm{M}\succeq \alpha^2 \bm{K}_{p}$. Let $y=((\beta^Tx_1)^{p},...,(\beta^Tx_n)^{p})\in \mathbb{R}^{n}$. We have $\sqrt{y^T(\bm{M})^{-1}y}\leq ||\beta||_2^{p}/\alpha$.
\end{pro}
\section{Main Results}
Assume there is an unknown data set ${\cal D}=\{x,y\}$. The inputs have the form $x=(X_1,X_2,...X_L)\in (\mathbb{R}^d)^L$. $||X_l||\leq \mathcal{O}(1)$ for all $1\leq l\leq L$. For every input $x_i$, there is a label $y_i=\pm1$.
The neural network with input $x$ is
\begin{equation}\label{ini}
\begin{aligned}
&h_0(x)=\phi(\bm{M}_0),\\
&h_{l}(x)=\phi(\bm{W}h_{l-1}+\bm{A}X_l),\\
&f(\bm{W},x)=\bm{B}^Th_L(x).
\end{aligned}
\end{equation}
Here $ \bm{W}\in\mathbb{R}^{m\times m}, \bm{A}\in \mathbb{R}^{m\times d}, \bm{B}, \bm{M}_0 \in \mathbb{R}^{ m}$. The entries of $\bm{M}_0$, $\bm{W}$ and $\bm{A}$ are respectively i.i.d. generated from $N(0,\frac{2}{m})$, $N(0,\frac{2}{m})$ and $N(0,\frac{2}{L^3\cdot m})$. The entries of $\bm{B}$ are i.i.d. generated from $N(0,\frac{1}{m})$.
The goal of learning RNN is to minimize the population loss:
\begin{equation}\label{ris}
L_{ {\cal D}}(\bm{W})=\mathbb{E}_{(x,y)\sim {\cal D}} \ell(y\cdot f(\bm{W},x)),
\end{equation}
by optimizing the empirical loss
\begin{equation}
L_{ S}(\bm{W})=\frac{1}{n}\sum_{i=1}^n\ell(y_i\cdot f(\bm{W},x_i)),
\end{equation}
using SGD.
Here $\ell(x)=\log(1+exp(-x))$ is the cross-entropy loss. Consider the SGD algorithm on this RNN.
\begin{algorithm}\label{a1}
\caption{Training RNN with SGD}
{\bf Input:} Data set ${\cal D}$, learning rate $\eta$.\\
The entries of $\bm{W}^0,\bm{A}$ are i.i.d. generated from $N(0,\frac{2}{m})$. The entries of $\bm{B}$ are i.i.d. generated from $N(0,\frac{1}{m})$.\\
\For {$t=1,2,3...n$}{
Randomly sample $(x_t, y_t)$ from the data set ${\cal D}$.\\
$\bm{W}^t=\bm{W}^{t-1}-\eta\nabla_{\bm{W}^{t-1}}\ell(y_t\cdot f(\bm{W}^{t-1},x_t))$.
}
\end{algorithm}
Let the complexity $\mathscr{C}^*$ of $F^*(\cdot)$ be defined in (\ref{com1}) and (\ref{com2}). The 0-1 error for ${\cal D}$ is $ L^{0-1}_{{\cal D}}(\bm{W})=\mathbb{E}_{(x,y)\sim {\cal D}}\mathbbm{1}\{y\cdot f(\bm{W},x)<0\}$. We have:
\begin{theo}\label{tt}
Assume there is $\delta\in (0,e^{-1}]$. Supposing for ${\cal D}=\{x_i,y_i\}$, there is a function $F^*$ belonging to the concept class (\ref{cop}) or (\ref{cop2}) such that $y_i\cdot F^*(x_i)\geq 1$ for all $i$. Let $\bm{W}^k$ be the output of Algorithm \ref{a1}. There is a parameter $m^*(n,\delta,L,\mathscr{C}^*)=poly(n,\delta^{-1},L,\mathscr{C}^*)$ such that, with probability at least $1-\delta$, if $m>m^*(n,\delta,L)$, there exits parameter $\eta= \mathcal{O}(1/ m)$ that satisfies
\begin{equation}
\frac{1}{n}\sum_{k=1}^nL^{0-1}_{{\cal D}}(\bm{W}^k)\leq \widetilde{\mathcal{O}}[\frac{(\mathscr{C}^*)^2}{n}]+\mathcal{O}(\frac{\log(1/\delta)}{n}).
\end{equation}
\end{theo}
\begin{rem}
This theorem induces that, to achieve {\bf population $0-1$ error}(rather than empirical loss) being less than $\epsilon$, it is enough to train the network using Algorithm \ref{a1} with $\widetilde{\Omega}((L\cdot\mathscr{C}^*)^2/\epsilon)$ steps. As defined in section \ref{fc} and \ref{fcc}, when $N$ is small, for the two types of concept class, $(\mathscr{C}^*)^2$ is almost-polynomial in input length $L$. Thus they can be learned effectively.
\end{rem}
\begin{rem}
This theorem can also be generalized to ``sequence labeling'' loss such as $\frac{1}{n}\sum_{i=1}^n\sum_{l=1}^L\ell(y_i\cdot f_l(\bm{W},x_i))$ with $f_l(\bm{W},x)=\bm{B}^Th_l(x)$. This is because the matrix $$H^l_{i,j}=\frac{1}{m}\langle \nabla f_l(\bm{W},x_i), \nabla f_l(\bm{W},x_j)\rangle$$ with different $l$ are almost ``orthogonal'' by a similar argument to (\ref{ot}) in Theorem \ref{tt6}. Then RNN can learn a function $f_l=sign(F^*_l(x))$ with $F^*_l(x)$ belonging to functions in section \ref{fcc}. See Remark \ref{g1} in the supplementary materials.
\end{rem}
\begin{comment}
\begin{rem}
Note that in this theorem, different from that in \cite{cao,sa}, we make no assumptions on the data distribution ${\cal D}$ and the results do not dependent on the minimum eigenvalue of kernel matrix. Therefore this is a PAC learning type of result as in \cite{allenzhu2019sgd}. In \cite{cao,sa}, matrix perturbation is used and $m$ should be larger than $\lambda_0^{-1}$ with $\lambda_0=\lambda_{min}(\bm{H})$. Using their methods, we can show there is a parameter $m^*(n,\delta,L,\lambda_0)=poly(n,\delta^{-1},L,\lambda_0^{-1})$ such that if $m>m^*(n,\delta,L,\lambda_0)$,
$$\frac{1}{n}\sum_{k=1}^nL^{0-1}_{{\cal D}}(\bm{W}^k)\leq \widetilde{\mathcal{O}}[\frac{(\mathscr{C}^*)^2}{n}]+\mathcal{O}(\frac{\log(1/\delta)}{n}).$$
However, when $d\to \infty$, $\lambda_0$ can be very large for some input distribution. But in fact the dependence of $\lambda_0$ in \cite{cao,sa} is due to the matrix perturbation. If we consider approximating $F^*(x_i)+\epsilon$ with $\epsilon$ is nonzero but small rather than $F^*(x_i)$, the dependence of $\lambda_0$ will not be necessary. Instead, we require $m>m^*(\mathscr{C}^*)\geq poly(\mathscr{C}^*)$.
\end{rem}
\end{comment}
\section{Sketch Proof of the Main Theorem}
The first step to prove the main theorem \ref{tt} is the following generalization of Corollary 3.10 in \cite{cao}.
\begin{theo}\label{mt}
Under the condition of Theorem \ref{tt}, let $n$ samples in the training set be $\{x_i,y_i\}_{i=1}^n$. $\widetilde{y}=[F^*(x_1),F^*(x_2),...F^*(x_n)]^T$. Let $\bm{H}$ be a matrix with $H_{i,j}= \frac{1}{m}\langle \nabla_{\widetilde{W}} f(\widetilde{\bm{W}},x_i),\nabla_{\widetilde{W}} f(\widetilde{\bm{W}},x_j) \rangle$. The entries of $\widetilde{\bm{W}}$ are i.i.d. generated from $N(0,\frac{2}{m})$.
If there is a matrix $\bm{H}^\infty \in \mathbb{R}^{n\times n}$ satisfying
\begin{equation}\label{condition1}
\bm{H} +\bm{\epsilon}^T\bm{\epsilon} \succeq \bm{H}^\infty \text{ with } ||\bm{\epsilon}||_F\leq 0.01/\mathcal{O}(\mathscr{C}^*),
\end{equation}
and $\sqrt{\widetilde{y}^T (\bm{H}^\infty)^{-1}\widetilde{y}}\leq \mathcal{O}(\mathscr{C}^*)$,
there exits $m^*(n,\delta^{-1},L,\mathscr{C}^*)=poly(n,\delta^{-1},L,\mathscr{C}^*)$ such that, with probability at least $1-\delta$, if $m>m^*$,
\begin{equation}
\frac{1}{n} \sum_{k=1}^n L^{0-1}_{{\cal D}}(\bm{W}^k)\leq \widetilde{\mathcal{O}}[\frac{\widetilde{y}^T (\bm{H}^\infty)^{-1}\widetilde{y}}{n}]+\mathcal{O}(\frac{\log(1/\delta)}{n}).
\end{equation}
\end{theo}
\begin{rem}
In order to show Theorem \ref{tt} using this theorem, we need to carefully pick out the exponential parts of $L$. Using the methods in \cite{rate} and \cite{cao}, we can show that $m^*(L,n,\sqrt{\widetilde{y}^T (\bm{H}^\infty)^{-1}\widetilde{y}})\geq poly(n,L,\sqrt{\widetilde{y}^T (\bm{H}^\infty)^{-1}\widetilde{y}})$ is enough. $\sqrt{\widetilde{y}^T (\bm{H}^\infty)^{-1}\widetilde{y}}$ is dealt with by calculating the forward and backward correlation in section \ref{f1} and \ref{f2}.
\end{rem}
The proof of theorem \ref{mt} is in fact a combination of the results in \cite{cao} and \cite{rate}. The really matter thing is how large can $\sqrt{\widetilde{y}^T (\bm{H}^\infty)^{-1}\widetilde{y}} \ $ be. We can show that:
\begin{theo}\label{mc}
Under the condition of Theorem \ref{mt}, with probability at least $1-\delta$, there exits matrix $\bm{H}^\infty$ satisfying (\ref{condition1}) and
\begin{equation}\label{ec}
\sqrt{\widetilde{y}^T (\bm{H}^\infty)^{-1}\widetilde{y}}\leq \mathcal{O}(\mathscr{C}^*).
\end{equation}
\end{theo}
Theorem \ref{tt} is a direct corollary of the above two theorems.
\subsection{Calculation on Kernel Matrix}
The proof of (\ref{ec}) relies on a direct calculation to construct a kernel matrix $\bm{H}^\infty$. We consider two input $x_i$ and $x_j$. Let $X_{i,l}$ and $X_{j,l}$ be the $l-th$ input of $x_i$ and $x_j$. Let $D_l\in \mathbb{R}^{m\times m}$ and $D_l'\in \mathbb{R}^{m\times m}$ be diagonal matrices that,
\begin{equation}
\begin{aligned}
&(D_l)_{k,k}=\mathbbm{1}\{\bm{W}h_{l-1}(x_i)+\bm{A}X_{i,l}>0\}\\
&(D'_l)_{k,k}=\mathbbm{1}\{\bm{W}h_{l-1}(x_j)+\bm{A}X_{j,l}>0\}
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
\text{Back}_l=BD_LW\cdots D_{l+1}W, \text{Back}'_l=BD'_LW\cdots D'_{l+1}W
\end{aligned}
\end{equation}
Then
\begin{equation}
\frac{1}{m}\langle \nabla_{\widetilde{W}} f(\widetilde{\bm{W}},x_i),\nabla_{\widetilde{W}} f(\widetilde{\bm{W}},x_j) \rangle=\frac{1}{m}\sum_{l,l'}\langle \text{Back}_l(x_i)\cdot D_l,\text{Back}_{l'}(x_j)\cdot D_{l'}' \rangle\cdot \langle h_l(x_i), h_{l'}(x_j)\rangle
\end{equation}
Generally $H_{i,j}= \frac{1}{m}\langle \nabla_{\widetilde{W}} f(\widetilde{\bm{W}},x_i),\nabla_{\widetilde{W}} f(\widetilde{\bm{W}},x_j) \rangle$ is hard to deal with. However, in the $m\to \infty$ limit, we can use some techniques to do the calculation.
\subsubsection{Forward Correlation}\label{f1}
\begin{theo}\label{tt5}
For fixed $i,j$, under the condition in Theorem \ref{mt}, with probability at least $1-exp(-\Omega(\log^2m))$,
\begin{equation}
|\langle h_l(x_i),h_l(x_j) \rangle -K^l_{i,j}|\leq \mathcal{O}(l^{16}\cdot\log^2 m/\sqrt{m})
\end{equation}
And let $Q_l=\sqrt{(1+\frac{1}{L^3}\sum_{k=1}^l||X_{i,k}||^2)\cdot(1+\frac{1}{L^3}\sum_{k=1}^l||X_{j,k}||^2) }$,
\begin{equation}\label{ek}
\begin{aligned}
&K^1_{i,j}= Q_1\cdot \sum_{r=0}^\infty \mu^2_r [(1+\frac{1}{L^3}X_{i,1}^TX_{j,1})/Q_1]^r\\
&K^{l}_{i,j}=Q_l\cdot \sum_{r=0}^\infty \mu^2_r (\{\frac{1}{L^3}X_{i,l}^TX_{j,l}+ K^{l-1}_{i,j}\}/Q_l)^r
\end{aligned}
\end{equation}
In the above equations, $\mu_r=\frac{1}{\sqrt{2\pi}}\int_0^\infty \sqrt{2}xh_r(x)e^{-\frac{x^2}{2}}dx$, $h_r(x)=\frac{1}{\sqrt{r!}}(-1)^re^{\frac{x^2}{2}}\frac{d^r}{dx^r}e^{-\frac{x^2}{2}}.$
\end{theo}
\subsubsection{Backward Correlation}\label{f2}
\begin{theo}\label{tt6} For $l\neq l'$, with probability at least $1-exp(-\Omega(\log^2m))$,
\begin{equation}\label{ot}
|\frac{1}{m}\langle \text{Back}_l(x_i)\cdot D_l,\text{Back}_{l'}(x_j)\cdot D_{l'}' \rangle| \leq \mathcal{O}(\frac{L^4\log^4 m}{m^{1/4}}).
\end{equation}
For $l=l'$, there is $F^l_{i,j}$ that, with probability at least $1-exp(-\Omega(\log^2m))$,
\begin{equation}
|\frac{1}{m}\langle \text{Back}_l(x_i)\cdot D_l,\text{Back}_{l}(x_j)\cdot D_{l'}' \rangle- F^l_{i,j}|\leq \mathcal{O}(\frac{L^4\log^4 m}{m^{1/4}}).
\end{equation}
where
\begin{equation}
\Sigma(x)=\frac{1}{2}+\frac{arcsin(x)}{\pi},
\end{equation}
\begin{equation}\label{ef}
F^l_{i,j}\succeq \frac{1}{K}\Sigma( \{\frac{1}{L^3}\langle X_{i,l}, X_{j,l} \rangle +K^{l-1}_{i,j}\}/Q_l).
\end{equation}
and $0<K\leq \mathcal{O}(1/L^4).$
\end{theo}
\begin{rem}
We should note that this theorem is one of the key differences between this work and the methods in \cite{allenzhu2019sgd}. In fact, we must show that there is a constant $K>0$ such that $\frac{1}{m}\langle \text{Back}_l(x_i),\text{Back}_{l}(x_j) \rangle -K $ is still positive definite. However, is $K$ large enough thus $1/K\geq poly(L)$ rather than $1/K \leq exp(-\Omega(L))$ ? This is not a trivial question. One can only get $K \geq \frac{1}{2^L}$ using naive estimation. In \cite{allenzhu2019sgd}, $||\bm{A}X_l||\leq \epsilon_x$ is required to make sure $\text{Back}'_l=\text{Back}_l(x_i)-\text{Back}_l(x_j)$ samll. However after $k$ steps of training, we can show the approximation error is roughly $\mathcal{O}(||\text{Back}'||\cdot ||\bm{W}^k-\bm{W}^0||)$ and $||\bm{W}^k-\bm{W}^0||_F \sim \sqrt{\widetilde{y}^T(\bm{H}^\infty)^{-1}\widetilde{y}} \sim \mathscr{C}(F^*)$. Thus the dependence of $\epsilon_x $ on $\mathscr{C}(F^*)$ is hard to be dealt with using this method. In this paper, we do not need the normalized condition. Our methods rely on a crucial observation that the function $\lim_{l\to \infty} h_l(x_i)^Th_l(x_j)/(||h_l(x_i)||\cdot||h_l(x_j||)$ will degenerate to a constant function.
\end{rem}
\subsubsection{Sketch Proof of Theorem \ref{mc}}
In order to estimate the complexity, we use the results in the last subsection and Proposition \ref{claa},\ref{c1} and \ref{c2}.
Proposition \ref{claa} shows that, in order to estimate $\sqrt{\widetilde{y}^T (\bm{H}^{\infty})^{-1}\widetilde{y}}$, we need to show
\begin{equation}
\bm{H}^\infty \succeq \xi_p\cdot (\bm{X}_l^T\bm{X}_l)^{\circ p}
\end{equation}
with $\xi_p>0$ for all $p\in \mathbb{N}, 1\leq l\leq L$.
Here $\bm{X}_l\in \mathbb{R}^{n\times d}=[X_{1,l}, X_{2,l}... X_{n,l}]$ and
\begin{equation}
[(\bm{X}_l^T\bm{X}_l)^{\circ p}]_{i,j}=\{X_{i,l}^TX_{j,l}\}^p.
\end{equation}
We will show that, there is a matrix $H^{\infty}$. With probability at least $1-\delta$, $H_{ij}=H^{\infty}_{ij}\pm \mathcal{O}(\frac{L^4\log^4 m}{m^{1/4}})$ for all $i,j\in [n]$, and,
\begin{equation}\label{eh}
H^{\infty}_{i,j}\succeq\frac{1}{\mathcal{O}(L^4)}\cdot Q_l\Sigma(\{\frac{1}{L^3}\langle X_{i,l}, X_{j,l} \rangle + K^{l-1}_{i,j}\}/Q_l).
\end{equation}
for all $l$.
\begin{comment}
To prove Equation (\ref{eh}), we need to calculate
\begin{equation}
\begin{aligned}
&\frac{1}{m}\langle \nabla_{\widetilde{W}} f(\widetilde{W},x_i),\nabla_{\widetilde{W}} f(\widetilde{W},x_j) \rangle\\
&=\frac{1}{m}\sum_{l,l'}\langle Back_l(x_i)\cdot D_l,Back_{l'}(x_j)\cdot D_{l'}' \rangle\cdot \langle h_l(x_i), h_{l'}(x_j)\rangle.
\end{aligned}
\end{equation}
Theorem \ref{tt6} induces that if $l\neq l'$, $$\frac{1}{m}\langle Back_l(x_i)\cdot D_l, Back_{l'}(x_j)\cdot D_{l'}' \rangle \approx 0$$
and
$$\frac{1}{m}\langle Back_l(x_i)\cdot D_l,Back_{l}(x_j)\cdot D_{l'}' \rangle \approx F^l_{i,j}$$
Meanwhile
$$F^l_{i,j}\succeq \frac{1}{K}\Sigma( \{\frac{1}{L^3}\langle X_{i,l}, X_{j,l} \rangle +K^{l-1}_{i,j}\}/Q_l)$$
$$\langle h_l(x_i),h_l(x_j) \rangle \succeq \Omega(1) .$$
Proposition \ref{c1} says if $\phi_1(\cdot,\cdot)$ and $\phi_1(\cdot,\cdot)$ are positive definite, $\phi(x_i,x_j)=\phi_1(x_i,x_j)\cdot \phi_2(x_i,x_j)$ is also a positive definite function.
Combing all these facts, directly we have $$ H_{i, j} \succeq F^l_{i,j}$$
and
$$ F^l_{i,j}\succeq \frac{1}{K}\Sigma( \{\frac{1}{L^3}\langle X_{i,l}, X_{j,l} \rangle +K^{l-1}_{i,j}\}/Q_l).$$
Thus (\ref{eh}) follows.
\end{comment}
Based on (\ref{eh}), we can show the following results:
For all $1\leq l\leq L$ and all $k$
\begin{equation}
H^{\infty}_{i,j}\succeq \frac{1}{\mathcal{O}(L^4)} \Sigma(\{K^{l}_{i,j}+\frac{1}{L^3}X_{i,l}^TX_{j,l}\}/Q_{l})\succeq \Omega( \frac{1}{L^7}) \cdot (\frac{1}{\mathcal{O}(L)})^{k}\cdot \frac{1}{k^2} (X_{i,l}^TX_{j,l})^{k}/(||X_{i,l}||\cdot X_{j,l}||)^{k}.
\end{equation}
This deduces the complexity for the Additive Concept Class in section \ref{fc},
\begin{equation}
\sqrt{\widetilde{y}^T (\bm{H}^{\infty})^{-1}\widetilde{y}}\leq \mathcal{O}(\mathscr{C}^*).
\end{equation}
As for N-Variables Concept Class,
\begin{equation}
\begin{aligned}
H^\infty_{i,j}\succeq& \frac{1}{C_1^N L^4 \cdot L^{2N} \cdot C_{N,p}\cdot(p/N)^N}\\
&\cdot (X_{i,r_1}^TX_{j,r_1} +X_{i,r_2}^TX_{j,r_2}... + X_{i,r_N}^TX_{j,r_N})^p/(N\cdot \max_n(||X_{i,r_n}||)\cdot \max_n(||X_{j,r_n}||)) ^{p}
\end{aligned}
\end{equation}
with some large constant $C_1>0$. Meanwhile, for any $l\leq L, a<l$, let $Z_{i, l,a}= [X_{i,l}, X_{i,l-1},... X_{i,l-a}] $. We have:
\begin{equation}
H^\infty_{i,j}\succeq \Omega( \frac{1}{L^7}) \cdot (\frac{1}{\mathcal{O}(L)})^{k}\cdot \frac{1}{k^2} (Z_{i,l,a}^TZ_{j,l,a})^{k}/(||Z_{i,l,a}||\cdot Z_{j,l,a}||\cdot 2^a)^{k}
\end{equation}
Then from definition of complexity in section \ref{fcc} and Proposition \ref{claa}, we can prove
\begin{equation}
\sqrt{\widetilde{y}^T (\bm{H}^{\infty})^{-1}\widetilde{y}}\leq \mathcal{O}(\mathscr{C}^*).
\end{equation}
Therefore (\ref{ec}) follows.
\section{Dissicusion}
In this paper, we use a new method to avoid the normalized conditions. The main idea is to provide an esitmation for $\sqrt{\widetilde{y}^T (\bm{H}^\infty)^{-1}\widetilde{y}}$ in the RNN case directly. However, the value of $\sqrt{\widetilde{y}^T (\bm{H}^\infty)^{-1}\widetilde{y}}$ is only explicitly calculated for the two-layer case in \cite{sa}. In the RNN cases, the neural tangent kernel matrix involves the depth and the weight sharing in the network and difficult to deal with.
In \cite{allenzhu2019sgd}, their method is to reduce the RNN case to $$f_L\approx\sum_lBack^{(0)} \cdot \mathbbm{1}_{\langle W, h_{l-1}\rangle+AX_l\geq 0} W^* \cdot h_{l-1},$$
which is similar to a summation of $L$ two-layer networks. And this reduction requires the following operations in \cite{allenzhu2019sgd}:
\begin{enumerate}[1)]
\item Introduce new randomness to keep the independence of rows in the random initialization matrices W and A at different depths. Then estimate the perturbation.
\item Show the "off-target" Backward Correlation is zero.
\item Estimate the "on target" Backward Correlation by introducing a normalized input sequence $x^{(0)}$.
\item Explicitly construct the approximation.
\end{enumerate}
These steps strongly rely on the normalized condition $||X_l||\ll 1 $ and this is apparently unrealistic. Instead, we calculate the kernel matrix and we introduce many new estimation to avoid this condition.
We should note that this expression $$f_L\approx\sum_lBack^{(0)} \cdot \mathbbm{1}_{\langle W, h_{l-1}\rangle+AX_l\geq 0} W^* \cdot h_{l-1}$$ is additive in itself. Thus the nonlinear interaction between different positions considered in this paper, especially N-variable target functions, {\bf cannot be deduced} using the from this method. In the previous proof, \cite{allenzhu2019sgd} is to use these steps to reduce the RNN function to a summation of two-layer networks and ignore the correlation between inputs from different locations and this heavily relies on the normalized condition. In our method, we need to consider the information in $\text{Back}$ to show the non-linear correlation between the inputs at different positions and prove N-variable target functions are learnable, while \cite{allenzhu2019sgd}. requires the normalized condition to make sure $\text{Back}\approx \text{Back}^{(0)}$ to be roughly a constant. This is one of the most different parts between this work and \cite{allenzhu2019sgd}.
In our case, since we do no use the normalized condition, we must show the polynomial decay of the constant part in $Back$. As mentioned in Remark 4.2, in our case, it is generally non-trivial to show $\sqrt{\widetilde{y}^T (\bm{H}^\infty)^{-1}\widetilde{y}}\leq O(\mathscr{C}^*)$ with $\mathscr{C}^*$ polynomial in $L$. Our methods rely on a detailed estimation on the degeneracy of long RNN based on Theorem \ref{tt5}.
\section{Related Work}
{\bf Overparameterized neural network.}
In \cite{tian2017symmetry-breaking} and \cite{du2018when}, it is shown that, for a single-hidden-node ReLU network, under mild assumptions, the loss function is one point convex in a very large area. However, in \cite{safran2018spurious}, the authors pointed out that such good properties are rare for networks with multi-hidden nodes, and indicated that an over-parameterization assumption is necessary. Similarly, \cite{2016Gradient} showed that over-parameterization can help in the training process of a linear dynamic system i.e., linear RNN. A different way to show over-parameterization is important as in \cite{freeman2016topology}, this work proved that in the two-layer case if the number of the hidden nodes is large enough, the sub-level sets of the loss will be nearly connected. Their method can also be applied to deep networks with a skip connection in \cite{ijcai2020-387} to study the properties of loss surfaces.
Recent breakthroughs were made in understanding the neural tangent kernel(NTK) \cite{NEURIPS20185a4be1fa, alemohammad2021the} of the neural network near the area of the random initialization. In \cite{NEURIPS201854fe976b}, \cite{DBLP}, \cite{allenzhu2019convergence} and \cite{rate}, it is shown that deep networks with a large hidden size can attain zero training error, under some assumptions of input non-degeneracy. This explains the empirical results \cite{zhang2017understanding} that DNN can fit training data with even random labels.
There are also some provable convergence results with over-parameterization going beyond NTK. The loss surface of the two-layer over-parameterized network with quadratic activation function was studied in \cite{du2018on} and \cite{MahdiTheoretical}. They showed that all the bad local minima are eliminated by over-parameterization. For ReLU activation function, in \cite{NEURIPS2019_5857d68c}, it is shown that there exits some functions can not be learned by any kernel functions but learnable with less error by a network with a skip connection. \cite{ li2020learning} provided a convergence result for learning a specific two-layer neural network which can not be learned by any kernel method, including Neural Tangent Kernel.
{\bf Generalization Ability of Deep Learning}
Classical VC theory cannot explain the generalization ability of deep learning because the VC-dimension of neural networks is at least linear in the number of parameters \cite{BartlettHLM19}. Recently, \cite{NEURIPS201962dad6e2} showed that overparameterized neural networks can learn some notable concept classes of target functions with rich types. Moreover, their work goes beyond the NTK linearization and provides new results on the non-convex
interactions of the three-layer network. Meanwhile, \cite{sa} provided a fine-grained analysis on the generalization error and showed the connections to the matrix of the neural tangent kernel. The results were generalized to the multi-layer case in \cite{cao}. Similar results were also studied in \cite{ji2020polylogarithmic} and \cite{chen2020overparameterization}.
\begin{comment}
Following the method in \cite{NEURIPS201962dad6e2}, \cite{allenzhu2019sgd} generalized the generalization results to RNN network to learn functions like
$F^*(x)=\sum_{l}\sum_{r} \psi_{l,r}(\beta^T_{l,r}X_l).$ For RNNs, when the system is not stable, previous works \cite{zhang2018stabilizing,chen2019generalization} require sample complexity exponentially grows in the input size $L$. In \cite{allenzhu2019sgd}, the complexity of sampling and learning are polynomial in the input size $L$. Yet this work requires a normalized condition $||X_l||\leq \epsilon_x$. To learn a function $F^*(x)$ with complexity $C'$, it is learnable with error $O(\epsilon)$. Thus $||X_l||\leq \mathcal{O}(\frac{\epsilon^{1/3}}{L^4C'^3})$ should be very small to learn very complex functions.
\end{comment}
Ref. \cite{NEURIPS201962dad6e2} also considered the generalization error bounds beyond the first-order NTK. It has been shown in \cite{NEURIPS201962dad6e2} that a three-layer ReLU network can provable learn some notable composite functions and dropout can help to reduce the Rademacher Complexity of the network thus reduce the generalization error bounds. The proof is based on the second-order NTK expansion and saddle points escaping arguments. Higher-order NTK are also studied in \cite{Bai2020Beyond} with provable generalization error bounds. Moreover, it is shown in \cite{NEURIPS2020_fb647ca6} that comparing with the general NTK, deep networks with neural representation can achieve improved sample complexities, while for the first-order NTK, depth may not provide benefits for the learning ability \cite{bietti2021deep}.
\section{Conclusion and Future Work}
In this paper, we studied the problem of what type of function can be learned by RNN. In this work, we showed that RNNs can provably learn the two types of functions, the additive concept class and the N-variables concept class in \emph{almost-polynomial in input length many iterations and samples} starting from random initialization. For the additive concept class, we proved the result without the normalized condition and showed the almost-polynomial complexity in input length $L$. For the N-variable concept class, we showed that RNN with ReLU activation function can provably learn functions like $\psi(\langle \beta,[X_{l_1},...,X_{l_N}]\rangle )$. The complexity of learning such functions grows exponentially with either $N$ or $l_0=\max(l_1,...l_N)- \min(l_1,...l_N)$, but when one of them is small, the complexity is almost-polynomial in the input length $L$.
One of the limitations is that this work relies on the NTK linearization of RNN. One probably direction is to consider the non-convex interactions in RNN and learn more complex functions using the method in \cite{NEURIPS201962dad6e2}. Meanwhile, this work studied RNN with ReLU activation function. This did not consider the ``gate'' structure in RNN. We believe that a study on GRU, LSTM, and MGU may lead to learning more complex functions with long-term memory.
\section*{Acknowledgement}
We would like to thank Professor Wenyu Zhang for his valuable discussion, and Shuai Wang for the great help in writing. We also thank the anonymous reviewers and area chair for their helpful comments.
This research was funded by the Fundamental Research Funds for the Central Universities (Grant number 2020YJS012).
\bibliographystyle{apalike}
|
\section{}
Nanoscale optofluidic cavities incorporating plasmonic and photonic crystal resonators have recently emerged as a powerful platform for chemical and biological sensing \cite{r49}, nanomanipulation \cite{r11} and optical nano-assembly \cite{r28,r50}. Owning to the nanoscale feature of the cavities, the directed transport of nanoparticles and biomolecules to the region of highest electromagnetic field enhancement is critical to device performance. Though plasmonic nanocavities can support localized electromagnetic hotspots, they suffer from intrinsic material loss that gives rise to low quality factor (Q factor) resonances with broad spectral linewidths. This makes realizing multi-resonant plasmonic cavities for wavelength switchable trapping and long-range particle transport extremely challenging. Dielectric photonic crystal (PhC) cavities, on the other hand confine light by means of a defect in an otherwise periodic arrangement of high index dielectric photonic structures \cite{r13,r19} leading to low-loss, narrow linewidth resonances that can be leveraged for wavelength switchable trapping applications. To date, the deterministic transport of particles for interaction with the enhanced field near resonant PhC cavities has been met with challenges. Prior reported attempts to achieve transport of particles to PhC cavities rely on pressure-driven flow \cite{r15,r21}. Unfortunately, this has limited particle capture rate because only the particles in the fluid boundary layer near the cavity can interact with the electromagnetic hotspots. Additionally, such pressure driven flow does not provide the mechanism for actively transporting particles from one nano-cavity to the next.\\
\indent In this letter, we investigate the physics of light-induced near-field trapping, attractive negative thermophoresis and long-range electrohydrodynamic transport of nanoparticles in PhC cavities for directional delivery of particles and trapping at the cavity region by switching the input wavelength.\par
\begin{figure}[htp!]
\includegraphics[scale=0.325]{fig1.jpg}
\caption{\label{fig1} Schematic of the PhC-based multiplexed long-range electrohydrodynamic transport and trapping system. The inset shows the forces experienced by a particle trapped at the bowtie. Here “opt”, “et” and “th” denote “optical force”, “thermophoretic force” and “electrothermal force”, respectively.}
\end{figure}
As shown in Figure 1, a PhC nanobeam with a bowtie defect at the center is placed beside a bus waveguide to enable evanescent coupling from the side. For multiplexed long-range nanoparticle transport and trapping across cavities, we design the system to contain ($n = 3$) engineered bowtie PhC nanobeams (BPCNs) cascaded along the bus waveguide, each having a specific resonance wavelength. Light is coupled into the device through the bus waveguide and then extremely localized in the bowtie defect. The figure illustrates that a fundamental TE mode with the wavelength of $\lambda_1$ is propagating in the bus waveguide and coupled to the middle BPCN ($\Lambda_1$), inducing the cavity resonance, and switching the cavity to the "ON" state. With a bus waveguide width of 480 nm and coupling gap of 150 nm, the loaded Q of the side-coupled BPCN is $\mathrm{1.5 \times 10^4}$ and the coupling efficiency \cite{r31} is 62$\mathrm{\%}$ in a water environment. The peak electric field amplitude is 138 times higher in the bowtie compared to that of the light input into the bus waveguide, resulting in an electromagnetic field intensity enhancement of $\sim19000$. For details about the design of the BPCNs, please see Supporting Information S1. The electric field profile of the side-coupled BPCN is shown in Figure \ref{fig2}(a), where the extreme light localization at the center of the bowtie can be easily observed. To obtain different resonance wavelengths for the other two BPCNs, we slightly adjust the period of the PhC holes by -4 nm and +4 nm, respectively. It is imperative to note that in practical experiments, it is not necessary to carefully engineer the dimensions of the BPCN to obtain different resonant wavelengths since fabrication uncertainties can naturally introduce resonance shift due to the high Q characteristic of the bowtie PhC. Moreover, resistive heaters can be incorporated to carefully tune the PhC cavities to the desired resonance wavelengths. Figure \ref{fig2}(b) shows the transmission spectra presenting the multi-resonant property of our system. \par
\begin{figure}[h!]
\includegraphics[scale=0.4]{fig2.jpg}
\caption{\label{fig2} (a) Electric field enhancement distribution of the BPCN cavity on resonance. The field enhancement is calculated by normalization to the amplitude of the electric field of the input fundamental TE mode in the bus waveguide. Inset: Zoom-in view of the bowtie region. (b) Transmission spectra for the three BPCNs, demonstrating the multi-resonant property of the system. The resonant wavelengths are $\lambda_1 = 1635.33 \mathrm{nm}, \lambda_2 = 1623.65 \mathrm{nm}, \lambda_3 = 1647.28 \mathrm{nm}$.}
\end{figure}
Next, we characterize the optical trapping performance of the optimized BPCN system. An enhanced optical gradient trapping force requires a spatially confined electromagnetic field, which is provided by the bowtie. The time averaged optical force exerted on a nanoscale object is calculated by integrating the Maxwell’s stress tensor (MST) \cite{r32} over an arbitrary surface enclosing the nanoscale object, which is given by
\begin{eqnarray}\label{eq1}
\langle\mathbf{F}\rangle=\oint_{S}\langle\overrightarrow{\mathbf{T}}\rangle \cdot d \mathbf{S},
\end{eqnarray}
where $\langle\overrightarrow{\mathbf{T}}\rangle$ is the time averaged Maxwell’s stress tensor given by:
\begin{eqnarray}\label{eq2}
\langle\overrightarrow{\mathbf{T}}\rangle=\frac{1}{2}\operatorname{Re}\left[\varepsilon \mathbf{E} \mathbf{E}^{*}+\mu \mathbf{HH}^{*}-\frac{1}{2}\left(\varepsilon|\mathbf{E}|^{2}+\mu|\mathbf{H}|^{2}\right)I\right]
\end{eqnarray}
Here $\mathbf{EE}^{*}$ and $\mathbf{HH}^{*}$ are the outer products of the fields; \emph{I} is the identity matrix; and $\varepsilon$ and $\mu$ are the permittivity and the permeability of the medium surrounding the object, respectively. The effective transverse trapping potential resulting from the optical force is given by \cite{r33}
\begin{eqnarray}\label{eq3}
U\left(\mathbf{r}_{0}\right)=\int_{\infty}^{r_0} \mathbf{F}(\mathbf{r}) d \mathbf{r}.
\end{eqnarray}
Figure \ref{fig3}(a) illustrates the force spectra of a 10 nm diameter ($D$ = 10 nm) PbSe quantum dot (refractive index \emph{n} = 4.73 + 0.24i \cite{r35}) positioned 21 nm above the center of the bowtie, showing a strong pulling force along the z direction (0.44 pN/2.5mW). This trapping force is at least an order of magnitude higher than those achieved using Mie-resonant dielectric nanoantenna \cite{r34} and dielectric nanoantenna supporting anapole states \cite{r46}. Figure \ref{fig3}(b) to \ref{fig3}(d) show the trapping potentials as well as the corresponding trapping forces when moving the quantum dot along the \emph{x}, \emph{y} and \emph{z} directions, respectively. The depth of the trapping potential well is around 2 $k_\mathrm{B}T$/2.5mW for the $x$ and $y$ directions and 10 $k_\mathrm{B}T$/2.5mW for the \emph{z} direction. Here $k_\mathrm{B}$ is the Boltzmann constant. This provides sufficient potential depth to stably confine the nanoparticle near the bowtie region.\par
\begin{figure}[h!]
\includegraphics[scale=0.45]{fig3.jpg}
\caption{\label{fig3} Optical trapping characterization for a 10nm PbSe quantum dot placed 21 nm above the bowtie surface. (a) Trapping force spectra for the quantum dot. (b) – (d) Trapping potential as well as trapping forces when moving the quantum dot along the \emph{x}, \emph{y} and \emph{z} directions, respectively. The vertical dashed lines in (b) and (c) denote the center of the bowtie. }
\end{figure}
Next, we explore the impact of the particle size and the vertical distance between the particle and the bowtie on the optical forces. We consider three different particle diameters (\emph{D} = 5, 10, and 20 nm) for the quantum dot. Figure \ref{fig4}(a) shows the depth of the optical trapping potential well along the $y$ direction ($U_{y}^\mathrm{opt}$) with respect to $F_{y}^\mathrm{opt}$ and Figure \ref{fig4}(b) shows the maximum absolute value of the pulling force $F_{z}^\mathrm{opt}$ as a function of the distance (\emph{z}) measured from the quantum dot bottom to the bowtie surface (shown in the inset), respectively.
\begin{figure}[h!]
\includegraphics[scale=0.6]{fig4.jpg}
\caption{\label{fig4} Depth of the trapping potential well along the y direction ($U_{y}^\mathrm{opt}$) due to $F_{y}^\mathrm{opt}$ (a) and the maximum absolute value of the pulling force $F_{z}^\mathrm{opt}$ (b) at different distances (z) from the bowtie surface (shown in the inset of (b)). z is varied as 3, 9, 15, and 21 nm. Both y axes are in log scale. (c) Transverse thermophoretic trapping potential for 10 nm particle at \emph{z} = 21 nm. The red dash line denotes the trace in (d). (d) Transverse trapping potential for the optical trapping and thermophoretic trapping potentials along the y axis.}
\end{figure}
Given the stable trapping requirement of 10 $k_\mathrm{B}T$ for the trapping potential well, Figure \ref{fig4}(a) indicates that the minimum power required for stable trapping of the three particles considered along the y direction is around 17 mW, 3 mW and 0.7 mW, respectively, when $z$ is 3 nm, which is less than a half of the power required by a plasmonic nanoaperture \cite{r33} taking into account the refractive index differences. Figure \ref{fig4}(b) shows that the maximum pulling force along the \emph{z} direction is 0.97 pN/2.5mW, 4.31 pN/2.5mW and 14.30 pN/2.5mW for the three particles considered at the same distance (\emph{z} = 3 nm). This pulling force drops exponentially as the particle moves away from the bowtie. Additional discussions on optical forces can be found in Supporting Information S3.\\
\indent Optical power dissipated in the water layer near the bowtie establishes a thermal gradient. To calculate the temperature field distribution, the computed electric field distribution is used to determine the heat source density, which gives the heat dissipated per unit volume and is expressed as \cite{r43,r56}
\begin{eqnarray}\label{eq4}
q(\mathbf{r})=\frac{1}{2} \operatorname{Re}\left[\mathbf{J}_{d}^{*}(\mathbf{r}) \cdot \mathbf{E}(\mathbf{r})\right]=\frac{\omega}{2} \operatorname{Im}(\varepsilon)|\mathbf{E}(\mathbf{r})|^{2},
\end{eqnarray}
where $\mathbf{J}_{d}=i\omega\mathbf{D}$ with $\mathbf{D}=\varepsilon\mathbf{E}$ is the displacement current density and $\varepsilon$ is the relative permittivity of the specific material. Since water is the only lossy material in this system, the power density dissipated into water serves as the source term in the heat diffusion equation for computation of the temperature around the PhC as well as in the surrounding fluid and substrate. The temperature field in the system is determined by solving the steady-state heat equation given by
\begin{eqnarray}\label{eq5}
\nabla \cdot\left[-\kappa \nabla T(\mathbf{r})+\rho c_{p} T(\mathbf{r}) \mathbf{u}(\mathbf{r})\right]=q(\mathbf{r}).
\end{eqnarray}
The first term on the left is the heat conduction term, while the second term is the convection term, which depends on the velocity of the fluid. $T(\mathbf{r})$ and $\mathbf{u}(\mathbf{r})$ are spatial temperature and fluid velocity field, respectively, and the material properties $\kappa$, $\rho$ and $c_p$ are thermal conductivity, density and specific heat capacity, respectively.\\
\indent Thermophoresis phenomenon \cite{r54,r55} is the motion of particles or molecules in the presence of thermal gradients and provides an attractive trapping force under negative thermophoresis \cite{r38,r39,r41}. Details of the calculation of the negative thermophoretic force are provided in the Supplementary Information S4. To compare with the optical trapping, Figure \ref{fig4}(c) illustrates the simulated transverse thermophoretic trapping potential ($\mathrm{U}_{xy}^\mathrm{th}$) for 10 nm particle at \emph{z} = 21 nm. The thermophoretic trapping potential has the same order of magnitude as the optical trapping potential shown in Figure 3, whereas the thermophoretic trapping potential well is much broader than the optical one. For emphasis, the transverse optical trapping and thermophoretic trapping potential along the $y$ axis are shown in Figure \ref{fig4}(d). The asymmetry of the thermophoretic trapping potential curve results from the bus waveguide which affects the heat diffusion. The superposition of the optical trapping and the thermophoretic trapping achieves a long range and 1.7 times deeper trapping potential well.\\
\begin{figure}[h!]
\includegraphics[angle=270,scale=0.475]{fig5.jpg}
\caption{\label{fig5} (a) Temperature field distribution of the $xy$ plane 300 nm above the bowtie. The radial velocity vector plot of electrothermal flow induced around the resonant BPCN is superimposed on the temperature profile. Arrow length represents the magnitude of the flow velocity. (b) Temperature field distribution of the $xz$ plane and (c) the $yz$ plane. (d) Illustration of three cascaded BPCNs placed beside a bus waveguide. The right three panels show corresponding radial velocity profiles of the induced electrothermal flow along $x$ and $y$ direction (300 nm above the surface) around the three BPCNs in different states. The electrothermal flow shows a long-range characteristic ($\sim 50 \mathrm{\mu m}$). (a) to (c) corresponds to the $\Lambda_1$ panel.}\par
\end{figure}
Next, we demonstrate the long-range and rapid transport of individual nanoparticles to the vicinity of the bowtie for optical trapping by generating wavelength-dependent electrohydrodynamic microfluidic vortices based on the electrothermal flow effect in a microfluidic channel. The thermal gradient induced in the water layer near the bowtie results in a gradient in the permittivity and electrical conductivity of the water medium near the bowtie. An applied a.c. electric field acts on these gradients to create a volumetric body force in the fluid due to the electrothermal effect \cite{r22,r52}.\\
\indent By leveraging the configuration of the cascaded BPCNs, the body force of the electrothermal flow in our system is not only space dependent, but also \emph{wavelength dependent}. That means the spatial distribution of the local temperature gradient can additionally be controlled by the wavelength of the input light, as shown in Figure \ref{fig5}(d). Following a perturbative expansion \cite{r22}, the wavelength dependent time-averaged electrothermal body force per unit volume at a.c. frequency $\omega$ can be expressed as:
\begin{eqnarray}\label{eq6}
\left\langle F_{\text{ET}}\right\rangle \hat{z}=\frac{1}{2} \varepsilon E_{z}^{2}\left[\frac{\sigma^{2} \varepsilon(\alpha-\gamma)}{\sigma^{2}+\omega^{2} \varepsilon^{2}}-\frac{1}{2} \alpha\right] \frac{\partial T(z, \lambda)}{\partial z} \hat{z},
\end{eqnarray}
\begin{eqnarray}\label{eq7}
\left\langle F_{\text{ET}}\right\rangle \hat{r}=-\frac{1}{4} \varepsilon \alpha E_{z}^{2} \frac{\partial T(r, \lambda)}{\partial r} \hat{r},
\end{eqnarray}
where $\lambda=\lambda_1,\lambda_2,\lambda_3…\lambda_n$ and $n$ is the number of the cascaded BPCNs; $\varepsilon$, $\tau=\varepsilon/\sigma$, $\sigma$ and $\omega$ are the fluid permittivity, charge relaxation time, electrical conductivity, and applied a.c. frequency, respectively; $\alpha$ and $\gamma$ are expressed as $\alpha=(1 / \varepsilon)(\partial \varepsilon / \partial T)$, $\gamma=(1 / \sigma)(\partial \sigma / \partial T)$ and are given as $-0.004 \mathrm{~K}^{-1}$ and $0.02 \mathrm{~K}^{-1}$, respectively \cite{r53}. Eq. \ref{eq6} and Eq. \ref{eq7} describe the axial and radial components of the electrothermal body force.\\
\indent The velocity field distribution of the fluidic flow when a given BPCN is excited is determined from the solution of the incompressible Navier–Stokes equations given by
\begin{eqnarray}\label{eq8}
\rho_{0}(\mathbf{u}(\mathbf{r}) \cdot \nabla) \mathbf{u}(\mathbf{r})+\nabla p(\mathbf{r})-\eta \nabla^{2} \mathbf{u}(\mathbf{r})=\mathbf{F},
\end{eqnarray}
\begin{eqnarray}\label{eq9}
\nabla \cdot \mathbf{u}=0.
\end{eqnarray}
The forcing term $\mathbf{F}$ in Eq. 8 describes the body force per unit volume acting on the fluid element, which is given by Eq. \ref{eq6} and Eq. \ref{eq7}. We note that contribution from buoyancy-driven convection is negligible in this system (see Supporting Information S7 for details).\\
\indent Figure \ref{fig5}(a) shows the temperature field distribution in the \emph{xy} plane 300 nm above the PhC surface. A temperature rise of 10.6 K is observed and the superimposed radial velocity vector plot of the flow shows that the induced electrothermal flow is directed radially inwards towards the thermal hotspot generated by water around the bowtie and serves as a powerful means to deliver suspended particles to the bowtie region. The maximum flow velocity is about 29.8 $\mathrm{\mu m/s}$, which is directed to the bowtie and hence much more efficient than traditional particle delivery methods such as pressure-driven flow and slow Brownian motion. Furthermore, this flow velocity is at least 20 times greater than the 1 $\mathrm{\mu m/s}$ thermoplasmonic convection flow velocity achievable with an array of optimized plasmonic bowtie nanoantenna \cite{r51}. Figure \ref{fig5}(b) and (c) show the temperature profile in the \emph{xz} and the \emph{yz} plane. The maximum temperature rise is 43.6 K under only 2.5 mW input power due to the high field enhancement. The superimposed velocity vectors verify the induced electrothermal vortex flow shown in Figure \ref{fig1}.\\
\indent Figure \ref{fig5}(d) demonstrates the concept of multiplexed nanoparticle transport and nano-optical trapping with the cascaded BPCNs shown in Figure \ref{fig1}. When the input wavelength is tuned to $\lambda_1$, only the middle BPCN ($\Lambda_1$) with a resonance wavelength of $\lambda_1$ is excited while the other two are off resonance and not excited. Therefore, the electrothermal flow is only induced around $\Lambda_1$ (see the radial velocity plot in the middle panel). By integrating over the radial velocity curve when $\Lambda_2$ is on, we estimate that it takes about 8 seconds for the flow to transport a nanoparticle 25 $\mathrm{\mu m}$ from the vicinity of $\Lambda_1$ to $\Lambda_2$. The high Q characteristic of the bowtie PhC permits to integrate multiple BPCNs along the low-loss bus waveguide to provide the means to achieve long range particle hand-off from tens of microns to millimeter scale distances by simply switching the wavelength of the input light. We note that the electrothermal flow along the $y$ direction shows a larger magnitude of velocity in comparison to the electrothermal flow along the $x$ direction. This is attributed to the asymmetric spatial distribution of the in-plane thermal hotspots. It is evident that there is a higher temperature gradient along the y direction and hence a stronger electrothermal flow velocity in comparison to that along the $x$ direction. The slightly different radial velocity values presented in the three panels are attributed to the different field enhancements of the three BPCNs.\\
\indent We have proposed and systematically studied a cascaded bowtie photonic crystal nanobeam system that can achieve multiplexed long-range electrohydrodynamic transport and optical trapping of nanoscale particles. Compared with traditional 1D photonic crystal nanobeams, our bowtie photonic crystal cavity can more strongly confine and enhance the electromagnetic field while maintaining a high quality factor. The extremely localized field provides a strong field gradient that is ideal for trapping sub-20 nm particles. Furthermore, the localized water absorption near the cavities serves as heat sources to generate negative thermophoresis that can assist in the optical trapping process. Finally, we harness the localized water absorption to induce on-demand electrothermal flow that can efficiently transport nanoparticles to the vicinity of the localized field of the bowtie photonic crystal cavity region for enhanced optical trapping. Our proposed multiplexed platform could enable millimeter scale transport and hand-off of particles across cavities in miniaturized optofluidic chips by simply switching the wavelength. We envision that our system will be a promising platform in many fields of biology and quantum information, such as in single molecule characterization and assembly of single photon sources.\\
\begin{acknowledgments}
J.C.N and S.M.W conceived and guided the project. J.A.A designed the photonic crystal cavity. J.A.A and S.Y performed the wave-optics simulations in Lumerical. S.Y. implemented the photonic crystal design in COMSOL and performed the wave-optics and Multiphysics simulations in COMSOL. C.H contributed to the Multiphysics simulations and K.P.A contributed to the Lumerical simulations. J.A.A wrote the bowtie PhC design section and S.Y. wrote the rest of the manuscript. J.C.N and S.M.W contributed to editing the manuscript.\\
The authors acknowledge financial support from the National Science Foundation (NSF ECCS-1933109) and Vanderbilt University.
\end{acknowledgments}
|
\section{Introduction}
\label{sec:intro}
Semantic segmentation has played crucial roles in many applications like autonomous vehicle and augmented reality.
Recent advances in this field are mainly attributed to the development of deep neural networks, whose success depends heavily on the availability of a large-scale annotated dataset for training.
However, creating large training datasets is prohibitively expensive since it demands manual annotation of pixel-level class labels.
To mitigate this problem, synthetic image datasets have been introduced~\cite{gta5,synthia}.
They provide a large amount of labeled images for training at minimal cost of construction.
Also, they can simulate scenes that are rarely observed in the real world yet must be considered in training (\emph{e.g.}, accidents in autonomous driving scenarios).
\input{figures/fig1_teaser}
When learning semantic segmentation using synthetic images, it is essential to close the gap between the synthetic and real domains caused by their appearance differences so as to avoid performance degradation of learned models on real-world images.
Most of existing solutions to this issue belong to the category of domain adaptation, which aims at adapting models trained on synthetic images (\emph{i.e.}, source domain) to real-world images (\emph{i.e.}, target domain).
In general, domain adaptation methods assume a single, particular target domain and train models using images from both of labeled source and unlabeled target domains~\cite{hoffman2016fcns,tsai2018learning,tsai2019domain,vu2019advent,zou2018unsupervised,li2019bidirectional,zou2019confidence,pan2020unsupervised,NEURIPS2020_243be281}.
Unfortunately, this setting limits applicability of learned models since, when deployed, models can face multiple and diverse target domains (\emph{e.g.}, geolocations and weather conditions in the case of autonomous vehicle) that are latent at the training stage.
As a more realistic solution to the problem, we study \emph{domain generalization} for semantic segmentation.
The goal of this task is to learn models that generalize well to various target domains without having access to their images in training.
A pioneer work~\cite{DRPC} achieves the generalization by forcing segmentation models to be invariant to random style variations of input image.
However, this method is costly since it applies an image-to-image translator~\cite{zhu2017unpaired} to every synthetic image multiple times for the style randomization.
Further, random styles are given by a small number of images sampled from ImageNet~\cite{Imagenet}, thus often irrelevant to target applications and hard to cover a wide range of real-world image styles.
Meanwhile, follow-up research~\cite{chen2020automated} encourages the representations learned using synthetic images to be similar with those of an ImageNet pretrained network.
This method is unfortunately also limited by the knowledge of ImageNet.
In this paper, we propose a WEb-image assisted Domain GEneralization scheme, dubbed \emph{WEDGE}, which overcomes the limitations of previous work by using real and application-relevant images crawled from web repositories (\emph{e.g.}, Flickr).
The crawling process demands no or minimal human intervention as it only asks search keywords that are determined directly by target application (\emph{e.g.}, ``driving + road'' for autonomous driving) or classes appearing in the source domain images.
Moreover, unlike those of ImageNet, the retrieved images can be used for self-supervised learning as well as for stylization since they are expected to be relevant to target application.
As illustrated in Fig.~\ref{fig:overall},
WEDGE utilizes images crawled from the Web in two different ways.
First, it replaces neural styles of synthetic training images with those of web-crawled images on-the-fly during training.
This helps enhance the generalization by giving illusions of diverse real images while exploiting groundtruth labels of synthetic images.
For this purpose, we introduce a \emph{style injection} module that conducts the style manipulation in a feature level at low cost.
As it is substantially more efficient than the image-to-image translator used in~\cite{DRPC}, it allows to perform the stylization on-the-fly using a large number of web images as style references in training.
Second, the web-crawled images are used as additional training data with pseudo segmentation labels.
To this end, the entire training procedure is divided into two stages.
In the first stage, a segmentation model is trained with the style injection module, and the web-crawled images are used only for stylization.
The learned model is then applied to the web-crawled images to estimate their pseudo labels.
The second stage is identical to the former, except that it also utilizes the web-crawled images as training data by taking their pseudo labels as supervision.
To demonstrate the efficacy of WEDGE, we adopt each of the GTA5~\cite{gta5} and SYNTHIA~\cite{synthia} datasets as the source domain for training, and evaluate trained models on three different real image datasets~\cite{cityscapes,BDD100k,mapillary}.
Experimental results demonstrate that WEDGE enables segmentation models to generalize well to multiple unseen real domains and clearly outperforms existing methods.
In summary, the contribution of this paper is three-fold:
\vspace{-1mm}
\begin{itemize}[leftmargin=4mm]
\itemsep=1mm
\item To the best of our knowledge, WEDGE is the first that attempts to utilize web-crawled images for domain generalizable semantic segmentation. %
These images facilitate self training based on the realistic data which may better approximate unseen testing domains.
\item We introduce style injection to domain generalizable semantic segmentation.
Through web-crawled images, it helps achieve the generalization by giving diverse illusions of reality to the network being trained using labeled synthetic images. Also, the superiority of our particular style injection method over other potential candidates is demonstrated empirically.
\item WEDGE outperforms existing domain generalization techniques~\cite{DRPC,chen2020automated,choi2021robustnet} in every experiment.
\end{itemize}
\section{Related Work}
\label{sec:relatedwork}
\noindent \textbf{Domain adaptive semantic segmentation.}
With the advent of synthetic datasets~\cite{gta5,synthia}, unsupervised domain adaptation has been widely studied for semantic segmentation.
This task is allowed to use both of labeled synthetic and unlabeled real data for training.
Most of existing solutions to the task are divided into two categories: (1) \emph{Distribution alignment}~\cite{hoffman2016fcns,zhang2017curriculum,tsai2018learning,tsai2019domain,vu2019advent}, and (2) \emph{Self-training}~\cite{zou2018unsupervised,li2019bidirectional,zou2019confidence,pan2020unsupervised}.
The former aims at aligning distributions of synthetic and real data in a common feature space.
To this end, Hoffman~\emph{et al.}~\cite{hoffman2016fcns} propose feature-level domain-adversarial learning.
Based on this, follow-up studies have been proposed to align the distributions in the image level~\cite{hoffman2018cycada} and in the prediction space~\cite{tsai2018learning,tsai2019domain,vu2019advent}.
On the other hand, self-training methods exploit unlabeled data for supervised learning through their pseudo labels.
In particular, they focus on improving the quality of pseudo labels by learning semantic segmentation and image translation bidirectionally~\cite{li2019bidirectional}, alleviating class imbalance issues in pseudo labels~\cite{zou2018unsupervised,zou2019confidence}, and learning texture invariant representation using randomly stylized source data~\cite{kim2020learning}.
However, applications of these methods are limited since they assume a single target domain.
In contrast, we seek generalization to multiple latent target domains without having access to data from the domains.
\vspace{1mm} \noindent \textbf{Domain generalizable semantic segmentation.}
The goal of domain generalization is to learn models that well generalize to unseen domains~\cite{muandet2013domain,gan2016learning}.
Early approaches address this task mostly for classification~\cite{li2018domain,li2017deeper,li2018learning,pan2018two,nam2019reducing,zhou2021domain}, but recent research demonstrates its potential for semantic segmentation~\cite{pan2018two,DRPC,chen2020automated}.
For example, Pan~\emph{et al.}~\cite{pan2018two} tackle this problem by a feature normalization operation designed for learning domain invariant features, while Chen~\emph{et al.}~\cite{chen2020automated} encourage the representation learned on a source domain to be similar with that of an ImageNet pretrained model.
Meanwhile, Yue~\emph{et al.}~\cite{DRPC} propose to learn features invariant to random style variations of input, and establish
an evaluation protocol for domain generalizable semantic segmentation.
The main difference of ours from the previous work is that ours explores and exploits real images on the Web which enable models to experience a variety of real domains during training with no human intervention.
\vspace{1mm} \noindent \textbf{Neural style transfer.}
A pioneer work by Gatys~\emph{et al.}~\cite{gatys2016image} shows that an image style can be captured by the Gram matrix of a feature map, and Johnson~\emph{et al.}~\cite{johnson2016perceptual} further enhance this idea to transfer a neural style to arbitrary content images.
Huang~\emph{et al.}~\cite{AdaIN} demonstrate that the channel-wise mean and standard deviation of a feature map represent image style effectively.
Also, Nam and Kim~\cite{nam2018batch} and Kim~\emph{et al.}~\cite{Kim2020UGATIT} propose to use different normalization operations complementary to each other for style transfer.
Recently, content-aware style transfer methods~\cite{park2019arbitrary,liu2021adaattn,huo2021manifold}
are emerged to catch more details of local style patterns and to preserve content better.
Park~\emph{et al.}~\cite{park2019arbitrary} and Liu~\emph{et al.}~\cite{liu2021adaattn} calculate similarities between content and style features as attention weights, and perform semantic-aware style transfer upon the similarity.
Huo~\emph{et al.}~\cite{huo2021manifold} suppose that features passing through a network form a manifold per each semantic region, and suggest a new style transfer technique based on manifold alignment.
Style injection in WEDGE is motivated in particular by the techniques presented in~\cite{AdaIN,huo2021manifold}.
However, it is distinct from the previous work in that it aims to perform feature stylization, rather than image stylization.
\vspace{1mm} \noindent \textbf{Style-based domain generalization.}
Recent studies on domain generalization~\cite{nam2019reducing,zhou2021domain} presume that a domain can be represented by neural styles of its images and achieve the generalization via style augmentation.
Zhou~\emph{et al.}~\cite{zhou2021domain} augment feature statistics by mixing among their input features as style interpolation.
Using AdaIN~\cite{AdaIN}, Nam~\emph{et al.}~\cite{nam2019reducing} propose the style-agnostic network by adversarial learning to make the feature extractor less style-biased.
Unlike the studies~\cite{nam2019reducing,zhou2021domain} which explore a limited range of styles in the source domains, we first propose a external web-crawled images as style reference for domain generalization.
Also, we suggest locally style injection between semantically similar features, which is suitable for semantic segmentation task that need to maintain fine content (\emph{e.g.}, shape or boundary) of objects on an input image.
\vspace{1mm} \noindent \textbf{Learning using data on the web.}
Modern recognition models tend to be data-hungry, yet the amount of training data is usually limited.
Data on the Web have been exploited to alleviate this issue.
Early studies utilize web-crawled images and videos for learning concept recognition by using their search keywords as pseudo labels~\cite{chen2013neil,divvala2014learning,chen2015webly}, and for object localization via clustering images~\cite{chen2013neil,divvala2014learning} or by motion segmentation~\cite{prest2012learning}.
Motivated by recent advances in pseudo labeling,
a large-scale web data have been used for supervised learning with their pseudo labels, which is known as webly supervised learning.
For image classification,
Niu~\emph{et al.}~\cite{niu2018webly} present a reliable way to utilize search keywords as pseudo class labels.
For semantic segmentation, Hong~\emph{et al.}~\cite{Hong2017_webly} and Lee~\emph{et al.}~\cite{lee2019frame} compute pseudo labels by segmenting web videos using attentions drawn by an image classifier.
Motivated by these methods, we present the first that makes use of web-crawled images for domain generalization.
\section{Proposed Method}
As illustrated in Fig.~\ref{fig:overall}, WEDGE is divided into three steps as follows.
\begin{itemize}[leftmargin=5mm]
\item[1.] \textbf{Crawling images from web repositories automatically.}
\item[2.] \textbf{The first stage of training with style injection (SI)}: Learning a segmentation model on the synthetic dataset while injecting styles of the web-crawled images to its intermediate features for training.
\item[3.] \textbf{The second stage of training using pseudo labels (PI)}: Further training the model using the web-crawled images and their pseudo segmentation labels as well as the synthetic dataset.
\end{itemize}
Details of each step are presented in the remainder of this section.
\subsection{Crawling Images from the Web}
\label{sec:crawl_images}
We collect 4,904 images by crawling on Flickr, through the search keyword ``driving + road'' to find images relevant to the target application scenario, \emph{i.e.}, autonomous driving.
Examples of the collected images are presented in Fig.~\ref{fig:pseudo_labels_qual}.
Using these images for domain generalization has several advantages.
First of all, they offer a large variety of real image styles as illustrated in~Fig.~\ref{fig:pseudo_labels_qual},
which \emph{potentially cover testing domains}.
This is vital for achieving generalization to unseen domains.
Second, they are not random but mostly relevant to target applications due to the use of search keywords and thus can be used for supervised learning given their pseudo labels.
Last, they are accessible with minimal human intervention since the crawling process above is fully automated given a query.
The web-crawled images are often different from synthetic domain images in terms of semantic layout, and could partly contain irrelevant contents due to the ambiguity of search keywords and errors of the search engine.
WEDGE is robust against these issues for the following reasons.
In the first stage of training, the style injection module exploits only styles of the web images while disregarding their contents.
In the second stage, irrelevant parts of an image tend to be ignored in pseudo segmentation labels
due to their unreliable class predictions (\emph{i.e.}, low confidence).
\input{figures/fig3_sty_inj}
\subsection{Stage 1: Learning with Style Injection (SI)}
\label{sec:style_inject}
We propose a content-aware style injection method which transfers styles between semantically similar features of synthetic and web-crawled images.
This method can inject diverse styles while better preserving the content of an image than conventional methods relying on global feature statistics (\emph{e.g.}, AdaIN~\cite{AdaIN} or Gram matrix approximation~\cite{gatys2016image}).
At each iteration of training the segmentation network, a synthetic image $I^s$ is coupled with a web-crawled image $I^w$ that is randomly sampled.
Let $F^{d,l}\in \mathbb{R}^{H_{d} \times W_{d} \times C}$ be the feature map of $I^d$ from the $l^\textrm{th}$ convolution block of the network where $d\in \{s, w\}$.
First, we compute a cross correlation between the $F^{s,l}$ and $F^{w,l}$ as an affinity matrix $\mathbf{\Sigma}^{l}\in \mathbb{R}^{N_{s}\times N_{w}}$, where
$N_{d}=H_{d}W_{d}$ ($d\in \{s, w\}$):
\begin{align}
\mathbf{\Sigma}^{l}_{i,j} = \frac{F^{s,l}_{i}{F^{w,l}_{j}}^{\top}}{\|F^{s,l}_{i}\|\|F^{w,l}_{j}\|}.
\label{eq:similarity}
\end{align}
Then we find the projection matrix $M$ that minimizes the distance between features of the projected feature map $F^{s,l}M^{\top}$ and web-crawled image feature map $F^{w,l}$ weighted by the affinity matrix $\mathbf{\Sigma}^{l}$;
the role of the projection matrix $M$ is to project a synthetic feature onto a subspace of the semantically similar web-crawled features.
The objective function is formulated by
\begin{align}
\min_{M}J(M) = \frac{1}{N_{\mathbf{\Sigma}}}\sum^{N_{s}}_{i=1}\sum^{N_{w}}_{j=1} \mathbf{\Sigma}_{ij}^{l} \|F^{s,l}M^{\top}-F_{j}^{w,l}\|,
\label{eq:obj_fun}
\end{align}
where $N_{\mathbf{\Sigma}}$ is sum of all elements of $\mathbf{\Sigma}^{l}$.
The closed form solution of the optimization problem is given by Huo~\emph{et al.}~\cite{huo2021manifold} as
\begin{align}
P = UV^{\top},
\end{align}
where $U$ and $V$ are derived from singular value decomposition of $F^{s,l}\mathbf{\Sigma}^{l}{F^{w,l}}^{\top}$, \emph{i.e.}, $F^{s,l}\mathbf{\Sigma}^{l}{F^{w,l}}^{\top} = U\Lambda{V}^{\top}$.
With the projection matrix $M$, we transform a synthetic feature map $F^{s,l}$ to the projected feature map $F^{s,l}M^{\top}$ which is fed into the ${l+1}^\textrm{th}$ convolutional block.
\Fig{styinj} illustrates this process.
The overall pipeline of our style injection method follows that of the manifold alignment based style transfer (MAST)~\cite{huo2021manifold}.
However, unlike MAST that computes a discrete affinity matrix using $k$ nearest neighbor assignment, our method uses cosine similarity to compute the continuous affinity matrix in~\Eq{similarity}.
It considers similarity of all features to produce a content-aware projection matrix and does not require hyperparameter $k$ nor $\mathcal{O}(n^2)$ time complexity to assign the nearest neighbors;
our style injection process is designed as non-parametric, which enables effective and low-cost feature adjustment.
The style injection is applied to multiple convolution blocks of the network, in particular lower blocks since features of deeper layers are known to be less sensitive to style variations.
More details for implementation can be found in Sec.~\ref{sec:setting}.
Injecting styles of web-crawled images to synthetic training images enlarges the training dataset by a multiple of the number of web-crawled images, which is tremendous regarding the size of the training dataset, as well as making them look diverse and realistic in feature spaces.
In addition, there are several advantages of the content-aware style injection for domain generalization of semantic segmentation over the conventional approaches~\cite{gatys2016image,AdaIN}.
First, it enables style injection between semantically similar regions of web-crawled and synthetic images, which is more natural and effective for semantic segmentation.
Second, since different styles are injected to different semantic regions on an image, it helps keep boundaries between the semantically different regions in style-injected features.
We empirically verify the superiority of our style injection method over other potential style injection candidates in~\Sec{ablation}.
Finally, the network is trained by the pixel-wise cross-entropy loss with the segmentation label of the synthetic image $I^s$.
Let $P^s$ and $Y^s$ denote the segmentation prediction and the groundtruth label of $I^s$, respectively.
Then the loss is formulated as
\begin{equation}
\mathcal{L}_{\textrm{seg}}(P^s, Y^s) = - \frac{1}{N}\sum_{i=1}^H\sum_{j=1}^W\sum_{k=1}^C Y_{ijk}^s \log P_{ijk}^s, \label{eq:loss_pixel_crs_ent}
\end{equation}
where $N=H\times W$.
Although this loss is applied only to the synthetic domain, its gradients with respect to parameters will act as if the network takes real domain images as input thanks to the style injection.
Note that, in this stage, $I^w$ is used only as a style reference.
\subsection{Stage 2:~Learning Using Pseudo Labels (PL)}
\label{sec:pseudo_label}
Once the first stage is completed, the learned model can be used to generate pseudo labels of the web images.
The pseudo labels allow us to exploit the web images for supervised learning of the segmentation network, which further enhances the generalization capability of the model by learning it directly on a variety of real-world images.
Let $P^w \in \mathbb{R}^{H \times W \times C}$ be the segmentation prediction of the network given $I^w$ as input.
The pseudo segmentation label of $I^w$, denoted by $\widetilde{Y}^w \in \{0,1\}^{H \times W \times C}$, is obtained by choosing pixels with highly reliable predictions and labeling them with the classes of maximum scores:
\iftrue
\begin{equation}
\widetilde{Y}^w_{ijc} =
\begin{cases}
1, & \textrm{ if } c = \underset{k}{\operatorname{argmax}} \ P^w_{ijk}
\ \textrm{ and } \ h(P^w_{ij}) < \tau \\
0, & \textrm{ otherwise}
\end{cases},
\label{eq:pseudo}
\end{equation}
where $P^w_{ij} \in \mathbb{R}^C$ denotes the class probability distribution of the pixel $(i,j)$, $h(\cdot)$ indicates the entropy, and $\tau$ is a hyperparameter.
Note that we regard the prediction $P^w_{ij}$ unreliable when its entropy is high, \emph{i.e.}, $h(P^w_{ij}) \geq \tau$; in this case, the pixel is assigned no label and ignored during training.
Examples of the pseudo labels are presented in Fig.~\ref{fig:pseudo_labels_qual}.
\fi
The second stage of training utilizes both of the synthetic and the web-crawled images for supervised learning of the network.
It is basically the same with the first stage including the style injection, except that the segmentation loss is now applied to $P^w$ as well as $P^s$.
The total loss for the second stage is thus given by a linear combination of two segmentation losses:
\begin{equation}
\mathcal{L}(P^s, Y^s, P^w, \widetilde{Y}^w) = \mathcal{L}_\textrm{seg}(P^s, Y^s) + \mathcal{L}_\textrm{seg}(P^w, \widetilde{Y}^w),
\label{eq:pseudo_loss}
\end{equation}
where $\mathcal{L}_\textrm{seg}$ is the pixel-wise cross-entropy loss as given in Eq.~\eqref{eq:loss_pixel_crs_ent}.
\section{Experiments}
\label{sec:experiment}
In this section, we first present experimental settings in detail, then demonstrate the effectiveness of WEDGE through extensive results.
Effectiveness of style injection, pseudo labeling, and other design choices of WEDGE are investigated by ablation studies.
Due to page limit, the effectiveness of feature-level style injection, the sensitivity of our method to hyperparameters, the impact of style injection points, comparisons with using domain-specific web images, and more qualitative results are given in supplementary material.
\subsection{Experimental Setting}
\label{sec:setting}
\vspace{1mm} \noindent \textbf{Source datasets.}
As a synthetic source domain for training, we use either the GTA5~\cite{gta5} or the SYNTHIA~\cite{synthia} datasets.
GTA5 consists of 24,966 images and shares the same set of 19 semantic classes with the real test datasets.
Note that we remove 36 images of smallest file sizes from the dataset since they are non-informative, \emph{e.g.}, blacked-out images.
Meanwhile, SYNTHIA contains 9,400 images, whose annotations cover only 16 classes of the real test datasets.
Thus, we take only these 16 classes into account when evaluating models trained on SYNTHIA.
\vspace{1mm} \noindent \textbf{Test datasets.}
As unseen target domains for evaluation, we choose
the validation splits of the Cityscapes~\cite{cityscapes}, BDD100k Segmentation (BDDS)~\cite{BDD100k} and Mapillary~\cite{mapillary} datasets.
The Cityscapes and BDDS datasets have 500 and 1,000 validation images, respectively, and they are labeled for the same 19 classes.
2,000 validation images of the Mapillary dataset are annotated for 66 classes.
By following the protocol of~\cite{he2020segmentations}, we merge these classes to obtain the same 19 classes of the Cityscapes dataset.
\vspace{1mm} \noindent\textbf{Web-crawled images.}
From Flickr, we search for images whose widths are larger than or equal to 760 pixels, and with no copyright reserved (\emph{i.e.}, CC0) for their public use in future work, using the search keyword ``driving + road".
As a result, 4,904 web images in total are collected.
Note that, given these conditions, the crawling process was done automatically, and the retrieved images are used as-is without modification.
\vspace{1mm} \noindent \textbf{Networks and their training details.}
Following the current state of the art~\cite{DRPC}, we adopt DeepLab-v2~\cite{deeplab_v2} with various backbone networks, VGG16~\cite{vggnet}, ResNet50~\cite{resnet}, and ResNet101~\cite{resnet}, as our segmentation networks.
They are first pretrained on ImageNet~\cite{Imagenet},
then trained with the source dataset and our web-crawled images using SGD with momentum of 0.9 and weight decay of $\expnum{5}{4}$;
the initial learning rate is $\expnum{2}{4}$ for the first stage (SI) and $\expnum{1}{4}$ for the second stage (PL).
We set $\tau$ in~\Eq{pseudo} to $\expnum{5}{2}$ for all experiments.
\input{tables/comp_dg_gta}
\input{tables/comp_dg_syn}
\input{tables/dgweb}
\vspace{1mm} \noindent \textbf{Where to inject styles.}
Styles of web images are injected into the feature maps output by the $1^\textrm{th}$ and $2^\textrm{nd}$ residual blocks for ResNet101 and ResNet50.
In the case of VGG16, we inject styles into the outputs of its $2^\textrm{nd}$ and $3^\textrm{rd}$ convolutional blocks as it is shallower than the ResNet architectures.
The effects of injection points on performance are investigated in the supplementary material.
\subsection{Comparisons with the State of the Art}
\label{sec:comp_dg}
WEDGE~is compared with existing domain generalization techniques, IBN-Net~\cite{pan2018two}, AGS~\cite{chen2020automated}, DRPC~\cite{DRPC} and RobustNet~\cite{choi2021robustnet},
using two source domains (GTA5, SYNTHIA), three test domains (Cityscapes, BDDS, Mapillary), and three different backbone networks (VGG16, ResNet50, ResNet101).
As summarized in Table~\ref{tab:comp_dg_gta} and~\ref{tab:comp_dg_syn}, WEDGE clearly outperforms all the previous arts in all the 18 experiments;
its outstanding performance suggests the advantages of using web-crawled images for domain generalization.
\subsection{In-depth Analysis on WEDGE}
\label{sec:ablation}
\vspace{2mm} \noindent \textbf{Detailed performance analysis.}
To investigate the contribution of each training stage in WEDGE, we measure its performance at each stage for all experiments we have conducted so far.
The results are summarized in Table~\ref{tab:dgweb}.
As shown in the table, the first stage using style injection most contributes to the performance in most experiments, which demonstrates the effectiveness of using web-crawled images and our style injection module for domain generalization.
This achievement is remarkable, especially when considering that web images could be erroneous or irrelevant to the test domains.
Thanks to our style injection modules, WEDGE can exploit only diverse and realistic styles of web images while disregarding their contents that may be irrelevant.
The second stage also leads to non-trivial performance improvement, particularly in the generalization from SYNTHIA to BDDS, which imply the semantics or layouts of pseudo labels on SYNTHIA is more similar to those of BDDS than the other datasets.
\input{figures/fig5_qual}
\vspace{1mm} \noindent \textbf{Qualitative results.}
Fig.~\ref{fig:qual} presents qualitative results of WEDGE and its baseline.
The examples in the figure show that the first stage of WEDGE recovers most of the ill-classified pixels, and even finds out objects that are missing in the baseline results.
Also, its second stage further improves the segmentation quality by correcting dotted errors and capturing fine details of object shapes.
\vspace{1mm} \noindent \textbf{Comparison of style injection methods.}
We compare our style injection method with other potential candidates based on existing style transfer techniques~\cite{AdaIN,huo2021manifold} to demonstrate its advantages.
Note that these techniques are also used for injecting styles of web images on-the-fly within the same framework.
As summarized in~\Tbl{sty_inj_analy}, while using AdaIN~\cite{AdaIN} and MAST~\cite{huo2021manifold} also improves performance, our method achieves the best in both SI and PL stages except for the GTA to BDDS case in the SI stage.
Moreover, our method is more efficient than MAST since it does not need $k$ nearest neighbor search, whose time complexity is $\mathcal{O}(n^2)$, that is required for MAST.
\input{tables/sty_inj_analy}
\input{tables/abl_query}
\vspace{1mm} \noindent \textbf{Effect of the number of web-crawled images.}
We investigate the effect of the number of web-crawled images by evaluating performance of a segmentation model trained by WEDGE with different numbers of web-crawled images.
\input{figures/fig6_abl_web}
In Fig.~\ref{fig:abl_web}, these models are compared in terms of segmentation quality on the three target datasets.
As shown in the figure, the generalization capability of the model can be substantially improved by using only 1,000 web images, while using the whole web dataset further improves performance.
To be specific, when using 1,000 web images, the average mIoU over the 3 test datasets is 42.4\%, lacking only 2.4\% compared to the average performance of our final model.
The results also indicate that WEDGE consistently enhances the generalization performance when increasing the number of web-crawled images.
\vspace{1mm} \noindent \textbf{Effect of using task-relevant web images.}
The contribution of our crawling strategy is demonstrated by comparing WEDGE with its variants relying on other types of style references instead of the task-relevant web images.
Specifically, we utilize images sampled from the ImageNet dataset and web images crawled by the search keyword ``indoor'', both of which are irrelevant to the target task.
Also, the number of style references is set to 5,000 for fair comparisons to WEDGE.
Note that since these images are totally irrelevant to the target task, they are not suitable for pseudo labeling thus are used only for style injection.
As shown in \Tbl{abl_query}, our method using task-relevant web images (\emph{i.e.}, ``driving+road'') clearly outperforms the others.
Using the real yet irrelevant images improves performance, suggesting the robustness of our method,
but the results are still inferior to those of our method, meaning that our crawling strategy is useful and using relevant images matters.
\input{tables/abl_inj}
\vspace{1mm} \noindent \textbf{Style injection vs. domain-adversarial learning.}
To further understand the advantages of the style injection, we compare WEDGE with AdaptOutput~\cite{tsai2018learning} in the web-image assisted domain generalization setting.
Specifically, AdaptOutput is trained using the same synthetic dataset and adapted to the web images by domain-adversarial learning; this variant of AdaptOutput is denoted by AdaptOutput$_{web}$.
Since AdaptOutput$_{web}$ does not exploit pseudo labels of web images, it is compared to the first stage of WEDGE using only style injection for a fair comparison.
As shown in Table~\ref{tab:abl_inj}, AdaptOutput$_{web}$ improves performance marginally or even degrades performance, while the first stage of WEDGE consistently and substantially enhances the segmentation quality.
These results suggest that the domain adversarial learning is not an effective method for utilizing web images in the context of domain generalization.
Aligning the synthetic and web-crawled images cannot efficiently extend the scope of feature spaces to cover unseen testing domains. In contrast, increasing the diversity of training images (feature styles) increases the generalization capability.
\section{Conclusion}
\label{sec:conclusion}
We propose WEDGE, the first web-image assisted domain generalization scheme for learning semantic segmentation.
It explores and exploits web images that depict large diversity of real world scenes, which potentially cover latent test domains thus help improve generalization capability of trained models.
We propose effective ways to utilize the web-crawled images, namely style injection and pseudo labeling, which lead to consistent and substantial performance improvement over its baseline on test domains.
WEDGE clearly outperformed existing domain generalization techniques in all experiments, and is even as competitive as domain adaptation methods using test domain data that are not accessible in domain generalization.
Extensive ablation studies also demonstrated that WEDGE is able to utilize noisy and irrelevant web-crawled images reliably and is not sensitive to their number in training.
\pagebreak
\section{Advantages of Feature-level Style Injection}
\label{sec:ben_styinj}
Since WEDGE injects style representations in feature levels, one may wonder its advantages over image-level style transfer.
This section demonstrates the effectiveness of WEDGE, especially its style injection (SI) module, compared to image-level style transfer.
To this end, we adopt AdaIN~\cite{AdaIN}, exploiting feature statistics as style representation like WEDGE.
We generate 100,000 stylized GTA5~\cite{gta5} images by AdaIN using web-crawled images as style references; a few examples are shown in~\Fig{adain}.
We then train a segmentation model on the stylized GTA5 dataset.
As summarized in~\Tbl{ada_styinj}, we compare the model of the 1$^\textrm{st}$ stage of WEDGE (SI only) with the model trained on the stylized GTA5 images generated by AdaIN.
The results show WEDGE using SI outperforms AdaIN on all experimental settings except G$\rightarrow$C and G$\rightarrow$B with VGG16~\cite{resnet}.
Moreover, our feature-level approach has another benefit over the image-level counterpart in terms of efficiency.
AdaIN requires an additional network for style transfer.
On the other hand, SI in WEDGE is non-parametric and adjusting feature statistics of source images by those of web-crawled images, thus demands a much lower computational cost than AdaIN.
\begin{table}[t!]
\caption{Quantitative results in mIoU and parameters of domain generalization from (G)TA5~\cite{gta5} to (C)ityscapes~\cite{cityscapes} , (B)DDS~\cite{BDD100k} and (M)appillary~\cite{mapillary}.
}
\centering
\vspace{-1mm}
\scalebox{0.9}{
\begin{tabular}{cccccc}
\toprule
\multicolumn{1}{c|}{Methods} & \multicolumn{1}{c|}{~Backbone~~} & \multicolumn{1}{c|}{~Params~~} & ~G~$\rightarrow$~C~ & G~$\rightarrow$~B & ~G~$\rightarrow$~M~ \\ \midrule
\multicolumn{1}{c|}{\multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}Deeplab-v2~\cite{deeplab_v2}\\ +AdaIN\end{tabular}}} & \multicolumn{1}{c|}{VGG16} & \multicolumn{1}{c|}{53.1M} & 35.33 & 34.49 & 40.17\\
\multicolumn{1}{c|}{} & \multicolumn{1}{c|}{ResNet50} & \multicolumn{1}{c|}{48.6M} & 33.31 & {34.02} & 38.55\\
\multicolumn{1}{c|}{} & \multicolumn{1}{c|}{ResNet101} & \multicolumn{1}{c|}{67.6M} & 39.41 & 36.20 & 41.50 \\ \midrule
\multicolumn{1}{c|}{\multirow{3}{*}{~~WEDGE (SI)~~}} & \multicolumn{1}{c|}{VGG16} & \multicolumn{1}{c|}{29.6M} & {35.33} & {34.48} & {40.54} \\
\multicolumn{1}{c|}{} & \multicolumn{1}{c|}{ResNet50} & \multicolumn{1}{c|}{25.1M} & {36.25} & 36.30 & {42.75} \\
\multicolumn{1}{c|}{} & \multicolumn{1}{c|}{ResNet101} & \multicolumn{1}{c|}{44.0M} & {43.55} & {40.35} & {47.30} \\
\bottomrule
\end{tabular}
}
\label{tab:ada_styinj}
\vspace{-3mm}
\end{table}
\section{Details of Style Injection}
\label{sec:styinj}
Style representations of web-crawled images are injected into the feature maps output by $1^\textrm{th}$ and $2^\textrm{nd}$ residual blocks for both ResNet101 and ResNet50 ~\cite{resnet}, and those of $2^\textrm{nd}$ and $3^\textrm{rd}$ blocks for VGG16~\cite{vggnet}.
To verify the effectiveness of our method, this section presents ablation studies with various combinations of injection points.
We present experimental results with ResNet101 combined with four different combinations in ~\Tbl{styinj_101}.
The results show that semantic segmentation performance is degraded when $4^\textrm{th}$ residual block is included.
We suspect this is because deeper features are known to contain semantic information rather than styles, which makes them inappropriate for style injection.
As a result, using the output feature maps from the \{$1^\textrm{st}$, $2^\textrm{nd}$\} residual blocks turn out to be the most effective combination for ResNet101.
Therefore, we choose our injection points based on these observations when applying style injection to other backbone networks.
\begin{table}[t!]
\caption{Performance of the models with ResNet101 backbone on the setting from (G)TA5 to (C)ityscapes, (B)DDS and (M)appillary.
}
\centering
\scalebox{0.92}{
\begin{tabular}{cccc|cccc}
\toprule
\multicolumn{4}{c|}{Style injection points} & \multirow{2}{*}{~G~$\rightarrow$~C~} & \multirow{2}{*}{~G~$\rightarrow$~B~} & \multirow{2}{*}{~G~$\rightarrow$~M~} & \multirow{2}{*}{~Average~} \\
~~~~1~~~~ & ~~~~2~~~~ & ~~~~3~~~~ & ~~~~4~~~~ & & & & \\ \midrule
\cmark & \cmark & & & 43.55 & 40.35 & 47.30 & 43.73 \\
\cmark & \cmark & \cmark & & 44.03 & 39.30 & 47.30 & 43.54 \\
\cmark & \cmark & \cmark & \cmark & 41.01 & 38.49 & 46.46 & 41.99 \\
& \cmark & \cmark & & 42.00 & 39.03 & 44.93 & 41.99 \\
& & \cmark & \cmark & 37.92 & 35.02 & 38.47 & 37.20 \\
& \cmark & \cmark & \cmark & 38.64 & 35.19 & 41.39 & 38.44 \\
\bottomrule
\end{tabular}
}
\label{tab:styinj_101}
\end{table}
\section{Comparison with using domain-specific web images}
\label{sec:morekeywords}
Since our task at hand is domain generalization that assumes arbitrary target domains, we employ the keyword that does not indicate any specific domains.
Nevertheless, we experiment with the keywords ``driving + \{rain, show, fog\}''.
As summarized in~\Tbl{table_more_keywords}, these specific keywords are not as useful as the general one ``driving + road'' in our framework.
\section{More Qualitative Results}
\label{sec:morequal}
We present qualitative examples of semantic segmentation results by WEDGE for both of Stage 1 (SI) and Stage 2 (SI+PL) in~\Fig{supp_qual1} and~\Fig{supp_qual2}.
In these figures, the semantic segmentation results are color-coded by following the standard Cityscapes color map~\cite{cityscapes}; the colors associated to the classes are exhibited in~\Tbl{colorcode}.
\section{Comparison with Adaptation Methods}
\label{sec:morequan}
Since domain adaptation and generalization share similar objectives, we also compare WEDGE to the state-of-the-art domain adaptation techniques in the two adaptation settings, GTA5 to Cityscapes and SYNTHIA to Cityscapes.
We provide the class-wise IoU of WEDGE with the previous domain adaptation and generalization in \Tbl{comp_gta} and \Tbl{comp_syn}, for more detailed comparison.
To unify the performance of domain adaptation and generalization methods,
we report segmentation accuracies obtained by with VGG16 and ResNet101 backbones on the GTA5~\cite{gta5} to Cityscapes~\cite{cityscapes} and SYNTHIA~\cite{synthia} to Cityscapes settings.
Because domain adaptation exploits test domain data, which are in contrast not accessible in domain generalization, direct comparisons between them are unfair.
Nevertheless, WEDGE is as competitive as or even outperforms recently proposed domain adaptation techniques~\cite{tsai2018learning,luo2019taking,vu2019advent,tsai2019domain}.
Further, it achieves 88\% and 90\% of the best domain adaptation performance with VGG16 and ResNet101 backbones, respectively, in the GTA5 to Cityscapes settings;
\textit{these results are impressive when considering that WEDGE does not have access to test domain images at all for training.}
\begin{table}[!t]
\caption{The color code of classes on the test datasets.}
\vspace{-2mm}
\centering
\scalebox{0.97}{
\begin{tabular}{cccccccccc}
\cellcolor{aero}road & \cellcolor{bicy}sidewalk & \cellcolor{bird}\brt{building} & \cellcolor{boat}\brt{wall} & \cellcolor{bottle}fence & \cellcolor{bus}pole & \cellcolor{car}traffic light & \cellcolor{cat}traffic sign & \cellcolor{chair}vegetation & \cellcolor{cow}terrain \\
\cellcolor{table}sky & \cellcolor{dog}person & \cellcolor{horse}rider & \cellcolor{mbike}\brt{car} & \cellcolor{person}\brt{truck} & \cellcolor{plant}\brt{bus} & \cellcolor{sheep}\brt{train} & \cellcolor{sofa}\brt{motorbike} & \cellcolor{train}\brt{bicycle} & \\
\end{tabular}
}
\vspace{-1mm}
\label{tab:colorcode}
\end{table}
\begin{figure*}[t!]
\begin{center}
\includegraphics[width=1 \linewidth]{supp/fig_tau_pseudo3.pdf}
\end{center}
\vspace{-4mm}
\caption{Examples of the pseudo labels with the different thresholding hyper-parameter $\tau$. (a) Input images. Pseudo labels with (b) $\tau=0.1$. (c) $\tau=0.05$ (Ours). (d) $\tau=0.01$. (e) $\tau=0.005$.}
\label{fig:pseudo_tau}
\vspace{-4mm}
\end{figure*}
\input{supp/fig_qual}
\clearpage
\input{supp/table_da_comp}
\section{Sensitivity to Hyper-parameter $\tau$}
\label{sec:tau}
This section demonstrates the impact of the thresholding parameter $\tau$ on the quality of pseudo labels in terms of semantic segmentation performance.
Specifically, pseudo segmentation labels are generated by using $\tau$, which is a hyper-parameter that filters out unreliable predictions.
To this end, we design multiple variants of our model that are trained from different pseudo segmentation labels generated from various $\tau$.
The pseudo segmentation label of $I^w$, denoted by $\widetilde{Y}^w \in \{0,1\}^{H \times W \times C}$, is obtained by choosing pixels with highly reliable predictions and labeling them with the classes of maximum scores:
\begin{equation}
\widetilde{Y}^w_{ijc} =
\begin{cases}
1, & \textrm{ if } c = \underset{k}{\operatorname{argmax}} \ P^w_{ijk}
\ \textrm{ and } \ h(P^w_{ij}) < \tau \\
0, & \textrm{ otherwise}
\end{cases},
\label{eq:pseudo}
\end{equation}
\begin{wrapfigure}{r}{0.44\textwidth}
\vspace{-11mm}
\begin{center}
\includegraphics[width=\linewidth]{supp/fig_tau_v5.pdf}
\end{center}
\vspace{-6mm}
\caption{
Performance of our ResNet101 backbone model trained with different thresholded pseudo labels.
$\textrm{G}_{avg}$ is the average performance of the three test domains.
The performances across the set of hyper-parameters $\tau$ except 0.1 are marginally different.
}
\label{fig:abl_tau}
\vspace{-5mm}
\end{wrapfigure}
where $h(\cdot)$ indicates the entropy, $P^w_{ij} \in \mathbb{R}^C$ denotes the class probability distribution of the pixel $(i,j)$, and $\tau$ is a hyper-parameter.
We sample $\tau$ from \{0.1, 0.05, 0.01, 0.005\}, where $\tau=0.05$ means our model in the main paper.
As summarized in~\Fig{abl_tau}, the results are marginally different across the variation of the hyper-parameters except 0.1, but the setting we adopt in the paper is slightly better than the others.
Examples of the pseudo labels of web-crawled images are presented in~\Fig{pseudo_tau}, which demonstrates both pros and cons of different threshold values.
With a moderate thresholding (\emph{e.g.}, 0.1), the pseudo labels cover more real texture or parts of an object but have more noisy semantic labels.
With a strict thresholding, on the other hand, the pseudo labels have more accurate semantic information but cover smaller regions of web-crawled images.
The thresholding hyper-parameter we choose is in the middle, and leads to the best performance.
\section{Examples of Web-crawled Images}
\vspace{-1mm}
\label{sec:webimage}
This section exhibits a part of our web-crawled dataset.
Qualitative examples of the web-crawled images are presented in Fig.~\ref{fig:supp_webimage},
which demonstrates the diversity of the images in terms of time, geolocation, weather condition, and so on.
Such diversity enables WEDGE to achieve the generalization to latent real domains.
Note that these images often depict entities and semantic layouts that are irrelevant to those of source (and target) domains.
However, they are used as-is with no manual filtering process since the style injection and pseudo labeling of WEDGE offer reliable and effective ways to utilize them.
|
\section{Derivation of the nonlinear Schr\"odinger equation (2).}
Here we outline the derivation of the NLS equation (2) from the main text. Some details of the derivation procedure similar to those described in details in earlier publications~\cite{edgesol06,edgesol07}, are not reproduced here.
We start with the main nonlinear parabolic equation~(1)
\begin{align}
\label{main}
i\frac{\partial \psi}{\partial z}=-\frac{1}{2}\nabla^2 \psi-\mathcal{R}({\bm r},z)\psi-|\psi|^2\psi
\end{align}
We use the same notations as in the main text, i.e., $\nabla=(\partial_x,\partial_y)$, ${\bm r}=(x,y)$,
\begin{align}
\mathcal{R}({\bm r})=R_0({\bm r})+R_d({\bm r},z)
\end{align}
with
\begin{align}
R_0({\bm r})=R_0({\bm r}+\ell\hat{\bm i})=R_0({\bm r}+\ell\hat{\bm j})
\end{align}
term that describes perfect periodic 2D square lattice and
\begin{align}
R_d({\bm r},z)=\begin{cases}
R_0({\bm r}-\alpha z\hat{\bm j})-R_0({\bm r}), & |x|<L/2
\\
0, & |x|>L/2
\end{cases}
\end{align}
term that describes the \textit{defect} (tilted layer).
Thus:
\begin{eqnarray}
\mathcal{R}({\bm r},z)=\mathcal{R}({\bm r}+\ell\hat{\bm j},z)= \mathcal{R}({\bm r},z+Z)
\end{eqnarray}
Consider the linear problem
\begin{align}
\label{main_lin}
i\frac{\partial \tpsi}{\partial z}=-\frac{1}{2}\nabla^2 \tpsi-\mathcal{R}({\bm r},z)\tpsi
\end{align}
and the ansatz
\begin{align}
\tpsi({\bm r},z)=e^{ibz}\phi({\bm r},z), \qquad b\in\left[-\frac{\pi}{Z},\frac{\pi}{Z} \right)
\end{align}
where the function $\phi({\bm r},z)$ is periodic along the $z-$axis
\begin{align}
\phi({\bm r},z)=\phi({\bm r},z+Z)
\end{align}
and solves the equation
\begin{align}
\label{main_phi}
i\frac{\partial \phi}{\partial z}=-\frac{1}{2}\nabla^2 \phi-\mathcal{R}({\bm r},z)\phi+b\phi
\end{align}
Next we use the fact that due to periodicity of the array along the $y-$axis, the function $\phi(x,y,z)$ can be searched in the form of a Bloch wave:
\begin{align}
\label{Bloch_funct}
\phi({\bm r},z)=e^{iky}u_{\nu k}({\bm r},z),
\end{align}
where
\begin{align}
u_{\nu k}({\bm r},z)=u_{\nu k}({\bm r}+\ell\hat{\bm j},z)=u_{\nu k}({\bm r},z+Z),
\end{align}
with Bloch momentum $k$ in the first Brillouin zone
\begin{align}
k \in\left[-\frac{\pi}{\ell},\frac{\pi}{\ell} \right)
\end{align}
and with index $\nu$ enumerating modes at a given value of $k$ (in the absence of the level crossing).
Finally, we turn to the $x-$dependence of the solution. For a mode to be localized along $x$ axis in the defect (tilted) layer, we have to impose zero boundary conditions
\begin{align}
\lim_{x\to\pm \infty} u_{\nu k}({\bm r},z) =0
\end{align}
Since at $|x|>L/2$ the lattice is unperturbed, in these domains one can look for a solution in the form
\begin{align}
u_{\nu k}({\bm r},z)= e^{-|\lambda| x}w_{\nu k}({\bm r},z)
\end{align}
where
\begin{equation}
\begin{split}
w_{\nu k}({\bm r},z)=&w_{\nu k}({\bm r}+\ell\hat{\bm i},z)= \\
\qquad &w_{\nu k}({\bm r}+\ell\hat{\bm j},z)=w_{\nu k}({\bm r},z+Z),
\end{split}
\end{equation}
$\lambda$ is a constant and we used the fact that the system is symmetric with respect to the inversion $x\to-x$ considering the decay at $x\to-\infty$ and at $x\to \infty$ characterized by the same exponent.
We are interested in the simplest soliton solutions, which are constructed (bifurcate from) on linear defect modes localized in the $x$-direction. Taking into account that in realistic applications one deals with finite structures we solve the eigenvalue problem (\ref{main_phi}) on sufficiently large, but finite $x$-window of width $X$, using plane-wave expansion method. Even though this method assumes periodic boundary conditions:
\begin{align}
u_{\nu k}(x-X/2,y,z)=u_{\nu k}(x+X/2,y,z)
\end{align}
they are nevertheless fully consistent with exponential localization of the defect modes, since the latter have amplitude of the order of $10^{-9}$ at the borders of the integration domain. For the parameters used in the present work and for selected $k$ values we have found the number of localized modes equal to the number of the tilted layers in the defect. These defect modes coexist with bulk modes that extend far beyond the defect, into the bulk of stationary array. The solution of the linear eigenvalue problem (\ref{main_phi}) eventually yields the Floquet exponent $b$ as a function of $\nu$ and $k_0$ (see (\ref{Bloch_funct})), that is depicted in the panels (a)-(d) of Fig.~2 of the main text.
A quasi-one-dimensional envelope soliton can be constructed by imposing broad envelope on a mode localized in the defect region and having a Bloch wave number $k_0$. The procedure of the multiple-scale expansion combined with the averaging over $Z$-period is described in all details in Refs.~\cite{edgesol06,edgesol07}. Therefore, here we outline only the main steps.
Supposing that the soliton is constructed on the mode with index $\nu$, we perform the multiscale expansion by introducing a formal small parameter $0<\mu\ll 1$, two sets of formally independent scaled variables $(y_0,y_1,y_2,...):=(y, \mu y, \mu^2 y,...)$ and $(z_0,z_1,z_2...):=(z, \mu z, \mu^2 z,...)$, and look for a solution of Eq.~(\ref{main}) in the form of the expansion
\begin{equation}
\begin{split}
\label{ansatz}
&\psi= e^{ib_{\nu k_0} z_0}\phi,\\ &\phi=\mu A(\eta, z_2)\phi_{\nu k_0} ({\bm r},z_0) + \mu^2 \phi^{(1)} + \mu^3 \phi^{(2)} +\cdots
\end{split}
\end{equation}
where $\eta=y_1+b_{\nu k_0}'z_1$: $A(\eta, z)=A(y_1+b_{\nu k_0}' z_1, z_2)$ [the convention that in the arguments of the amplitude only the slowest variables are indicated is used]. The Bloch state $\phi_{\nu k_0}$ depends only on the "fast" variables $(x,y_0,z_0)$; the variable $x$ is not scaled. All slow variables are considered as independent.
Defining the linear operator
\begin{equation}
\begin{split}
\label{H0}
&H_0:= i\frac{\partial}{\partial z_0}+\frac{1}{2}\left(\frac{\partial^2 }{\partial x^2}+\frac{\partial^2 }{\partial y_0^2}\right) + \mathcal{R}({\bm r}_0,z)+b_{\nu k_0}, \\ &{\bm r}_0=(x,y_0)
\end{split}
\end{equation}
and using that in the slow variables
\begin{align}
\frac{\partial}{\partial y}=\frac{\partial}{\partial y_0}+\mu\frac{\partial}{\partial y_1}+\mu^2\frac{\partial}{\partial y_2}+\cdots
\\
\frac{\partial}{\partial z}=\frac{\partial}{\partial z_0}+\mu\frac{\partial}{\partial z_1}+\mu^2\frac{\partial}{\partial z_2}+\cdots
\end{align}
we substitute (\ref{ansatz}) into (\ref{main}) and collect the terms with different orders of $\mu$. This gives (only the three leading orders are shown):
\begin{equation}
\begin{split}
\label{line1}
0&= A H_0\phi_{\nu k_0}
\\
&+\mu \left(H_0 \phi^{(1)}+i\frac{\partial A}{\partial z_1}
\phi_{\nu k_0}+\frac{\partial \phi_{\nu k_0}}{\partial y_0}\frac{\partial A}{\partial y_1}\right)
\\
&+\mu^2 \left(H_0 \phi^{(2)}+i\frac{\partial A}{\partial z_2}
\phi_{\nu k_0}+\frac{\partial \phi_{\nu k_0}}{\partial y_0}\frac{\partial A}{\partial y_2}+\frac{\partial^2 \phi^{(1)}}{\partial y_0\partial y_1}+ \right.
\\
&\qquad\qquad \left.\frac{1}{2}\frac{\partial^2 A}{\partial y_1^2}\phi_{\nu k_0}+|A|^2A|\phi_{\nu k_0}|^2\phi_{\nu k_0}\right)
\end{split}
\end{equation}
The expressions in each of the lines in this equation must be zero independently. The ansatz (\ref{ansatz}) ensures that the leading order in the first line of (\ref{line1}) is zero identically. In order to resolve the orders of $\mu$ in the second line of (\ref{line1}) and $\mu^2$ in the third line of (\ref{line1}) we expand $\phi^{(1,2)}$ over the complete set of eigenfunctions ($j=1,2$)
\begin{align}
\label{expan_first}
\phi^{(j)} =\sum_{\nu'} C_{\nu' k_0}^{(j)}(y_1,z_0)\phi_{\nu'k_0}(y_0,z_0)
\end{align}
where, as explained in~\cite{edgesol06,edgesol07}, $C_{\nu' k_0}^{(j)}(y_1,z_0)=C_{\nu'k_0}^{(j)}(y_1,z_0+Z)$ and it is enough to account only the eignefunctions of the different bands with the same $k_0$.
To proceed we define time averaging over the period $Z$:
\begin{align}
\langle f\rangle_T=\frac{1}{Z}\int_{0}^{Z}f({\bm r},z)dz
\end{align}
and the inner product
\begin{align}
\label{inner}
(f,g)=\int_{-\infty}^{\infty} dx\int_0^\ell dy f^*({\bm r},z) g({\bm r},z).
\end{align}
Equating the second line of (\ref{line1}) to zero we obtain~\cite{edgesol06}
\begin{equation}
\begin{split}
\label{mu2order_1}
i\frac{\partial A}{\partial z_1}\phi_{\nu k_0}+&\frac{\partial A}{\partial y_1}\frac{\partial \phi_{\nu k_0}}{\partial y_0}
=\\
&\sum_{\nu'}\left[\frac 1i \frac{ \partial C_{\nu'k_0}^{(1)} }{\partial z_0} + (b_{\nu k_0}-b_{\nu' k_0})C_{\nu' k_0}^{(1)}\right]\phi_{\nu'k_0}.
\end{split}
\end{equation}
Taking into account that $A$ depends on $y_1$ and $z_1$ through $\eta$, where the relation
\begin{align}
\label{omega1}
b_{\nu k_0}^\prime = \left\langle \left(\phi_{\nu k_0},i\frac{\partial \phi_{\nu k_0}}{\partial y_0} \right)\right\rangle_Z
\end{align}
is obtained by the generalized $k\cdot p$ method described in~\cite{edgesol06,edgesol07}, one can ensure that application of $\langle (\phi_{\nu k_0},\cdot)\rangle_Z$ to the secon line of (\ref{line1}) yields zero. On the other hand, by applying to the same line $\langle (\phi_{\nu' k_0},\cdot)\rangle_Z$ with $\nu'\neq\nu $, one can obtain that $ C_{\nu'k_0}^{(1)} \propto \partial A/\partial y_0$ (the proportionality coefficients can be found in ~\cite{edgesol06}.
The generalized $k\cdot p$ method also allows to express $b_{\nu k_0}^{\prime\prime}$ through the Floquet-Bloch eigenstates~\cite{edgesol06}:
\begin{equation}
\begin{split}
\label{dispersion}
\frac{1}{2}\frac{\partial^2 A}{\partial y_1^2}-&i\left\langle \sum_{\nu'\neq \nu} \left(\phi_{\nu k_0},i\frac{\partial \phi_{\nu' k_0}}{\partial y_0} \right)
\frac{\partial C_{\nu'k_0}^{(1)} }{\partial y_1} \right\rangle_Z
=\\
&-\frac{b_{\nu k_0}^{\prime\prime}}{2}\frac{\partial^2 A}{\partial \eta^2}
\end{split}
\end{equation}
Now, equating the third line of (\ref{line1}) to zero, applying to it $\langle (\phi_{\nu k_0},\cdot)\rangle_Z$, using the relations (\ref{omega1}) and (\ref{dispersion}), employing the Fredholm alternative (i.e., eliminating secular terms), and assuming that the envelope $A$ is independent of the slow variable $y_2$, we arrive at the effective NLS equation (2) from the main text, where $\mu$ was set to one~\cite{edgesol06,edgesol07}.
\section{Interaction with a defect}
\begin{figure*}[t]
\centering\includegraphics[width=0.9\linewidth]{si-defect.png}
\caption{Interaction of soliton from Fig.~4 of the main text with the defect in the form of missing waveguide (see Visualization 2). Profiles of the array and field modulus distributions are shown at the same distances $z$. Green arrows indicate instantaneous defect position. }
\label{figureSup}
\end{figure*}
To study the impact of defects in the underlying structure on soliton propagation we consider the interaction of a bright soliton with a defect in the form of missing waveguide in a tilted layer. As an input, we use the same soliton as in Fig. 4 from the main text of the Letter with $M=5$. Input soliton is located at the point $y = 0$, while the missing waveguide is initially located far from soliton, at the point $y = -99$ (see left panels in Fig.~\ref{figureSup}). The figure shows field modulus distributions at different distances, when soliton only slightly touches the defect (second set of panels), when it fully overlaps with it (third set of panels), and when defect is passed (fourth to sixths sets of panels). One can see that even though the interaction with the defect causes notable backscattering and radiation into the bulk of the structure, the soliton survives after collision with defect and only its peak amplitude slightly decreases.
|
\section{Introduction}
If $\widetilde X_2$ is a del Pezzo surface of degree one, then it is the double cover of a quadric cone $X_2$ branched over a nonsingular genus four curve $X$ and the vertex of the cone. The quadric cone $X_2$ containing $X$ identifies a vanishing even theta characteristic $\theta_0$ of $X$.
%
Let $\kappa$ denote the canonical class of $\widetilde X_2$. The intersection pairing on $\widetilde X_2$ gives $\Pic(\widetilde X_2)$ the structure of a lattice, which admits a decomposition of the form
\[
\Pic(\widetilde X_2) = \gen{\kappa} \oplus \Pic(\widetilde X_2)^\perp
\]
where $\Pic(\widetilde X_2)^\perp$ denotes the sublattice of divisor classes orthogonal to the canonical class under the intersection pairing. The sublattice $\Pic(\widetilde X_2)^\perp$ is a root lattice of type $E_8$, and such a lattice has $135$ sublattices $\Lambda_{D_8} \subset \Pic(\widetilde X_2)^\perp$ of type $D_8$. In this article, we consider those del Pezzo surfaces of degree one whose Picard lattice has a Galois invariant sublattice of type $D_8$. The Galois action on $\Pic(\widetilde X_2)$ permutes the $240$ exceptional curves of $\widetilde X_2$ and acts on $\Pic(\widetilde X_2)^\perp$ though the Weyl group of type $E_8$, denoted $W_{E_8}$. As it turns out, the stabilizer in the Weyl group $W_{E_8}$ of a $D_8$-sublattice is of index $135$, and in fact it is isomorphic to the Weyl group $W_{D_8}$. Further details about del Pezzo surfaces can be found in \cite[Chapter~8]{Dol2012}. We show how to transmute the solution to the classical inverse Galois problem for subgroups of $W_{D_8}$ into a solution to the inverse Galois problem for del Pezzo surfaces of degree one with a Galois invariant sublattice $\Lambda_{D_8} \subset \Pic(\widetilde X_2)$ of type $D_8$.
\begin{theorem} \label{thm: main theorem}
Let $\rho\: \Gal(\mathbb{Q}^\mathrm{sep}/\mathbb{Q}) \rightarrow W_{D_8}$ be a continuous homomorphism. If $G := \Im(\rho)$ is the Galois group of some irreducible polynomial over $\mathbb{Q}$, then there exists a nonsingular del Pezzo surface of degree one $\widetilde X_2$ such that each $\sigma \in \Gal(\mathbb{Q}^\mathrm{sep}/\mathbb{Q})$ permutes the $240$ exceptional curves of $\widetilde X_2$ as described by $\rho(\sigma) \subset S_{240}$.
\end{theorem}
In particular, the subfield of $\mathbb{Q}^\mathrm{sep}$ fixed by $\rho^{-1}(\id)$ is the splitting field of the $240$ exceptional curves. The group $W_{D_8}$ admits a transitive permutation representation on a set of size $16$ (see Section~\ref{sec: background}). If $\rho'\: W_{D_8} \rightarrow S_{16}$ is the associated morphism and if $G \subseteq W_{D_8}$ is a subgroup for which a solution to the inverse Galois problem is known, then there exists a (possibly reducible) polynomial $f$ of degree $16$ such that $G \cong \Gal(f)$ and each $\sigma \in G$ permutes the roots of $f$ as described by $\rho'(\sigma)$. A solution to the inverse Galois problem is known for several subgroups of $W_{D_8}$, including $W_{D_8}$ itself.
\begin{theorem} \label{thm: inverse Galois polynomials}
Let $G$ be a group which acts transitively on a set of size at most $16$. Then there is an irreducible polynomial $f$ over $\mathbb{Q}$ such that $\Gal(f) \cong G$.
\end{theorem}
\begin{proof}
A list of examples can be obtained from the online database of Kl\"uners and Malle \cites{KlunersMalleDatabaseArticle, KlunersMalleDatabase2021}.
\end{proof}
Our techniques are related to those in the article by Elsenhans and Jahnel~\cite{ElsenhansJahnel2019}, where they study plane quartics in terms of a set of eight points in $\mathbb{P}^3$ called a Cayley octad.
In \cite{Corn2007}, Corn showed that the Brauer group of a del Pezzo surface is one of a finite number of explicit possibilities. Specifically, for a del Pezzo surface $X$ over $\mathbb{Q}$ there is a canonical isomorphism $\Br X / \Br \mathbb{Q} \cong \H^1(\Gal(\mathbb{Q}^\mathrm{sep}/\mathbb{Q}), \Pic X^\mathrm{sep})$ coming from the Hochschild-Serre spectral sequence. Corn's result is obtained by considering the action of an abstract group on the Picard lattice and using group cohomology to determine the list of possibilities. Corn's result raises an interesting question, \emph{which Brauer groups from Corn's list actually occur for a del Pezzo surface defined over $\mathbb{Q}$?}
%
Theorem~\ref{thm: main theorem} allows us to resolve this question for the groups in Corn's list \cite[Theorem~4.1]{Corn2007} of exponent $1$, $2$, or $4$; we list these in Table~\ref{tbl: Corn list}.
\begin{table}
\label{tbl: Corn list}
\begin{center}
\begin{tabular}{rl}
Degree & Groups of exponent $1,2,4$ \\ \hline \\[-1em]
$5 \leq d \leq 9$ & $\{\id\}$ \\
$d=4$ & any of the groups above, $\mathbb{Z}/2\mathbb{Z}, (\mathbb{Z}/2\mathbb{Z})^2$ \\
$d=3$ & any of the groups above \\
$d=2$ & any of the groups above, $(\mathbb{Z}/2\mathbb{Z})^s \ \ (3 \leq s \leq 6)$, $\mathbb{Z}/4\mathbb{Z} \oplus (\mathbb{Z}/2\mathbb{Z})^t \ \ (0 \leq t \leq 2)$, $(\mathbb{Z}/4\mathbb{Z})^2$ \\
$d=1$ & any of the groups above, $(\mathbb{Z}/2\mathbb{Z})^s \ \ (7 \leq s \leq 8)$, $\mathbb{Z}/4\mathbb{Z} \oplus (\mathbb{Z}/2\mathbb{Z})^t \ \ (0 \leq t \leq 4)$, \\ & $(\mathbb{Z}/4\mathbb{Z})^2 \oplus (\mathbb{Z}/2\mathbb{Z})^t \ \ (1 \leq t \leq 2)$ \\
\end{tabular}
\bigskip
\caption{The list of possible $\Br X/\Br \mathbb{Q}$ for del Pezzo surfaces by degree.}
\end{center}
\end{table}
\begin{corollary}
Any group from row $d$ in Table~\ref{tbl: Corn list} is of the form $\Br X/\Br \mathbb{Q}$ for some del Pezzo surface $X/\mathbb{Q}$ of degree~$d$.
\end{corollary}
\begin{proof}
First we consider the case where $d=1$. If $G$ is a subgroup of $W_{D_8}$ for which a solution to the inverse Galois problem is known, then the action of $G \subseteq W_{D_8} \subset W_{E_8}$ on the exceptional curves of a del Pezzo surface of degree one can be realized as a Galois action over $\mathbb{Q}$. We can explicitly compute the group cohomology for the action of $G$ on the root lattice $\Lambda_{E_8}$ of type $E_8$ and check for the desired Brauer group. A script that does this for all of the subgroups of $W_{D_8}$ is available at the link in Subsection~\ref{sec: sub: software}. (Of course, Corn carries out a similar calculation to prove \cite[Theorem~4.1]{Corn2007}.) For each of the groups $M$ in Table~\ref{tbl: Corn list}, our computation shows there is a subgroup $G \subset W_{D_8}$ of order at most $8$ such that $M \cong \H^1(G, \Lambda_{E_8})$. In particular, a solution to the inverse Galois problem for such $G$ is known.
For del Pezzo surfaces of degree $2 \leq d \leq 4$, we may apply the result of \cite{ElsenhansJahnel2019}. Alternatively, we search for a $G$-action on $\Pic(\widetilde X_2)^\perp$ such that there is an isomorphism of $G$-lattices
\[
\Pic(\widetilde X_2)^\perp \cong \Lambda \oplus \gen{e_1} \oplus \cdots \oplus \gen{e_{d-1}}
\]
where $e_j^2 = -1$, the summands are orthogonal, and where $\H^1(G, \Lambda)$ is as desired. A del Pezzo surface of degree one with such a Picard lattice has a set of $d-1$ pairwise orthogonal exceptional curves defined over $\mathbb{Q}$; these can be blown-down to obtain a del Pezzo surface of degree $d$ with Brauer group $\H^1(G, \Lambda)$.
\end{proof}
\subsection{Software} \label{sec: sub: software}
The computations for this paper were performed using the \texttt{magma} computer algebra system \cite{Magma}. The scripts are available at
\begin{center}
\url{https://github.com/a-kulkarn/SixteenPointsScripts.git}
\end{center}
The total computation time to run all of the scripts is approximately $1$ hour on commercial hardware (processor: AMD Ryzen$^{\text{tm}}$ Threadripper$^{\text{tm}}$ 2970WX).
\subsection*{Acknowledgements}
I would like to thank Eran Assaf for comments on the early drafts of this article.
The author has been supported by the Simons Collaboration on Arithmetic Geometry, Number Theory, and Computation (Simons Foundation grant 550033).
\section{Background} \label{sec: background}
\begin{comment}
\begin{corollary}
Let $L/\mathbb{Q}$ be a number field whose Galois group is a subgroup of $W_{D_8}$ and let $K$ be the normal $S_8$-subextension. Then there exists a normal basis $\alpha_1, \ldots, \alpha_8$ of $K$ such that $\mathbb{Q}(\alpha_1, \ldots, \alpha_8)$ generate the $S_8$ subextension $K$ and $L := K(\sqrt{\alpha_1}, \ldots, \sqrt{\alpha_8})$.
\end{corollary}
\begin{proof}
Kummer theory.
\end{proof}
\end{comment}
Let $K/\mathbb{Q}$ be an \'etale algebra, and let $K = \prod_{j} K_j$ be a decomposition into simple factors. If $K$ is simple, then the \defi{norm} of an element $\alpha \in K$ is the usual field norm $N_{K/\mathbb{Q}}$. Otherwise, we define the norm of $\alpha$ by $N_{K/\mathbb{Q}}(\alpha) := \prod_j N_{K_j/\mathbb{Q}}(\alpha_j)$.
\subsection{\'Etale algebras with a $W_{D_8}$ action}
The (abstract) group $W_{D_8}$ is isomorphic to an index $2$ subgroup of the wreath product $S_2 \wr S_8$. Specifically, it is the extension of $S_8$ by the subgroup of elements of $\mu_2^8$ whose entries multiply to $1$. We have the exact sequence
\[
\xym{
0 \ar[r] & \mu_2^7 \ar[r] & W_{D_8} \ar[r] & S_8 \ar[r] & 1
}.
\]
Essentially by definition, the wreath product $S_2 \wr S_8$ admits an action on a set of $16$ elements $\Omega$ (the eight elements on which $S_8$ acts endowed with signs).
\begin{construction} \label{cons: etale algebra of degree 16}
Given $\rho\: \Gal(\mathbb{Q}^\mathrm{sep}/\mathbb{Q}) \rightarrow W_{D_8}$ a continuous homomorphism with image $G$, we describe how to construct an \'etale algebra $L$ of degree $16$ such that $\Hom(L, \mathbb{Q}^\mathrm{sep})$ is isomorphic to $\Omega$ as a $G$-set. We may choose a set of orbit representatives $\beta_1, \ldots, \beta_r$ such that $\Omega = G\cdot \beta_1 \cup \cdots \cup G \cdot \beta_r$. Functoriality between \'etale algebras and continuous Galois actions on a finite set allows us to identify an \'etale algebra $\widehat L$ from $\rho$ \cite[Chapter 8]{milneFT}. Functoriality also allows us to define an \'etale subalgebra $L$ of degree $16$ as the subalgebra element-wise fixed by $\Stab(G; \beta_1, \ldots, \beta_r)$.
\end{construction}
The $16$ elements of $\Omega$ are identified with the $16$ homomorphisms $L \rightarrow \mathbb{Q}^\mathrm{sep}$ via the $\Hom$ functor. The natural quotient of $G$-sets $\Omega \rightarrow (\Omega/{\pm})$ induces an inclusion of \'etale algebras $K \longhookrightarrow L$, where the degree of $K$ is equal to $8$. Up to isomorphism we may write $K = \prod_j K_j$ and
\begin{equation} \label{eq: L presentation}
L \cong \prod_{j} K_j[x]/(x^2 - \alpha_j) \tag{$\dagger$}
\end{equation}
for some finite field extensions $K_j/\mathbb{Q}$ and some $\alpha_j \in K_j$ with the property that $\prod_j N_{K_j/\mathbb{Q}}(\alpha_j) \in \mathbb{Q}^{\times 2}$. We call $K$ the \defi{distinguished subalgebra}.
\subsection{Zariski density of generators}
\newcommand{\mba^{\!K}}{\mathbb{A}^{\!K}}
\newcommand{\mba^{\!A}}{\mathbb{A}^{\!A}}
\newcommand{\mathrm{Pol}^8}{\mathrm{Pol}^8}
\newcommand{\mathrm{sq}}{\mathrm{sq}}
We denote by $\mathrm{Pol}^8$ the scheme whose set of $R$-valued points is the set of monic polynomials of degree $8$ with coefficients in $R$, where $R$ is a commutative ring containing $\mathbb{Q}$. Of course, $\mathbb{A}^8_\mathbb{Q} \cong \mathrm{Pol}^8$.
If $K$ is an \'etale $\mathbb{Q}$-algebra of dimension $8$, we may endow $K$ with the structure of an affine scheme, specifically the affine space $\mathbb{A}^8_\mathbb{Q}$. We denote this scheme by $\mba^{\!K}$. The functor of points is given by $\mba^{\!K}\!(R) = K \otimes_\mathbb{Q} R$.
The scheme $\mba^{\!K}$ is simply the Weil restriction $\mathrm{Res}_{K/\mathbb{Q}} \mathbb{A}^1_K$. If $R$ is a commutative ring containing $\mathbb{Q}$, then we denote by $\mba^{\!K}_R$ the scheme $\mba^{\!K} \times_{\Spec \mathbb{Q}} \Spec R$.
If $\beta \in \mba^{\!K}\!(R)$, we denote by $[\beta]$ the endomorphism of $R \otimes_\mathbb{Q} K$ defined by multiplication-by-$\beta$. There is a canonical morphism of schemes
\[
\mapdef{\chi^K}{\mba^{\!K}}{\mathrm{Pol}^8}{\beta}{\mathrm{coeffs}(\det(xI - [\beta]))}.
\]
As an example, if $K = \prod_{j=1}^8 \mathbb{Q}$ is the split \'etale $\mathbb{Q}$-algebra of dimension $8$, then $\chi^K$ is simply the map defined by the elementary symmetric functions.
\begin{lemma} \label{lem: minimal polynomials are dense}
The set $\chi^K(\mba^{\!K}\!(\mathbb{Q}))$ is Zariski dense in $\mathrm{Pol}^8$.
\end{lemma}
\begin{proof}
Let ${\widehat K}$ be a splitting field for $K$. Since $\mba^{\!K}$ is just affine space, $\mba^{\!K}\!(\mathbb{Q})$ is dense in $\mba^{\!K}_{\widehat K}$. On the other hand, if $A := {\mathbb{Q}}^8$ is the split \'etale ${\widehat K}$ algebra, diagonalization gives an isomorphism $\psi\: \mba^{\!K}_{\widehat K} \rightarrow \mba^{\!A}_{\widehat K}$. Diagonalization does not change the characteristic polynomial, so the diagram
\[
\xym{
\mba^{\!K}_{\widehat K} \ar[d]_-{\chi^K_{\widehat K}} \ar[r]^\psi & \mba^{\!A}_{\widehat K} \ar[dl]^-{\chi^A_{\widehat K}} \\
\mathrm{Pol}^8_{\widehat K}
}
\]
of morphisms over ${\widehat K}$ commutes. But $\chi^A$ is just the morphism defined by elementary symmetric functions, so it is surjective as a morphism of schemes, and thus so too is $\chi^K$.
\end{proof}
\begin{corollary} \label{cor: sqrt primitive basis}
Given an \'etale algebra $L$ arising from Construction~\ref{cons: etale algebra of degree 16}, with distinguished subalgebra $K$, the set
\[
\left\{(\alpha_1, \ldots, \alpha_r) \in \prod_j K_j : L \cong \prod_j {K_j[x]}/{\gen{x^2-\alpha_j}}, \ K_j \cong \mathbb{Q}(\alpha_j)\right\}
\]
has a dense image under $\chi^K$ in $\mathrm{Pol}^8$.
\end{corollary}
\begin{proof}
Choose a presentation for $L$ as in Equation~\eqref{eq: L presentation}. Since the set $\alpha \cdot K^{\times 2}$ is Zariski dense in $\mba^{\!K}$ for any unit $\alpha \in K^\times$, and contains a dense subset of primitive elements, the result follows.
\end{proof}
\subsection{Tensors}
We denote by $\Sym_2 \mathbb{Q}^{m+1}$ the space of $(m+1) \times (m+1)$ symmetric matrices over $\mathbb{Q}$ and by $\Sym^2 \mathbb{Q}^{n+1}$ the space of quadratic forms over $\mathbb{Q}$ in $n+1$ variables.
Let $\mathcal{A} \in \mathbb{Q}^{n+1} \otimes \Sym_2 \mathbb{Q}^{m+1}$ be a tensor, symmetric in the last two entries. We view $\mathcal{A}$ as an $(n+1) \times (m+1) \times (m+1)$ array of elements of $\mathbb{Q}$. We may think of such an array as being an ordered collection $A_0, \ldots, A_n$ of symmetric $(m+1) \times (m+1)$ matrices, the \emph{slices} of $\mathcal{A}$. We denote the contraction of $\mathcal{A}$ along a vector $v \in \mathbb{Q}^{n+1}$ by $\mathcal{A}(v, \cdot, \cdot)$. More generally, we will contract along an element of $R^n$, where $R$ is a $\mathbb{Q}$-algebra. Similarly, we will denote the contraction by an element $y \in R^{m+1}$ by $\mathcal{A}(\cdot, y, \cdot)$ or $\mathcal{A}(\cdot, \cdot, y)$, depending along which axis we contract.
Denote $\mathbb{P}^{n} := \Proj \mathbb{Q}[x_0,\ldots,x_n]$, $\mathbb{P}^m := \Proj \mathbb{Q}[y_0, \ldots, y_m]$, and denote the dual projective space of $\mathbb{P}^n$ by $\widehat \mathbb{P}^n$. We will also denote $\mathbf{x} := (x_0, \ldots, x_n)$ and $\mathbf{y} := (y_0, \ldots, y_m)$.
If $x \in \mathbb{Q}^{n+1}$, the contraction $\mathcal{A}(x, \mathbf{y}, \mathbf{y})$ is the quadratic form
\[
\mathbf{y}^T\left(x_0A_0 + \ldots + x_nA_n \right)\mathbf{y}.
\]
If $x \in \mathbb{P}^n(\mathbb{Q})$ is a point, then the contraction
$
\mathcal{A}(x, \cdot, \cdot)
$
is a symmetric $(m+1) \times (m+1)$ matrix with entries in $\mathbb{Q}$, well-defined up to scaling.
Write $(q_0, \ldots, q_n)$ for the quadrics defined by the $n+1$ slices of $\mathcal{A}$. The $q_i$ define the rational map
\[
\begin{tabu}{cccc}
\psi\: & \mathbb{P}^m &\dashrightarrow & \widehat \mathbb{P}^n \\
& y & \mapsto & \mathcal{A}(\cdot, y, y) = (q_0(y), \ldots, q_n(y))
\end{tabu}.
\]
\section{Proof of the main result}
\subsection{Construction of $16$ points in $\mathbb{P}^4$}
Given a hyperelliptic genus $3$ curve $Y$ of the form
\[
Y\: y^2 = f(z_0, z_1)
\]
such that $f(z_0, z_1)$ is homogeneous, square-free, and $f(0,1), f(1,0) \in \mathbb{Q}^{\times 2}$, we describe a method to construct $16$ points in $\mathbb{P}^4$ based on the construction in \cite[Section 6]{KulkarniVemulapalli}.
Define $K := \mathbb{Q}[z]/(f(z,1))$ and $L := \mathbb{Q}[z]/(f(z^2,1))$. An elementary calculation shows that the conditions on $f(z_0, z_1)$ allow us to write
\[
f(z_0, z_1) = b(z_0, z_1)^2 - z_0z_1^3c(z_0,z_1) = -
\det
\begin{bmatrix}
z_0 z_1^3 & b(z_0, z_1) \\
b(z_0, z_1) & c(z_0, z_1)
\end{bmatrix}
\]
for some homogeneous polynomials $b(z_0, z_1), c(z_0, z_1)$ of degree $4$. If $\alpha_1, \ldots, \alpha_8$ are the roots of $f(z, 1)$, then the set
\[
\mathfrak{C} := \left\{(\pm\sqrt{\alpha_j} : (\pm\sqrt{\alpha_j})^{-1} \cdot b(\alpha_j, 1) : \alpha_j^2 : \alpha_j : 1) \ : \ 1 \leq j \leq 8\right\}
\]
of points in $\mathbb{P}^4$ is split over $\widehat L$, the Galois closure of $L/\mathbb{Q}$. One can check that when the $\alpha_j$ are distinct, the Vandermonde matrix of degree $2$ forms evaluated at the points of $\mathfrak{C}$ has corank $5$, so $\mathfrak{C}$ is contained in the intersection of $4$ quadrics. Generically, $\mathfrak{C}$ is a complete intersection of $4$ quadrics in $\mathbb{P}^4$. If $L$ is presented as in Equation~\eqref{eq: L presentation} such that $\alpha_1, \ldots, \alpha_r$ are primitive elements of $K_1, \ldots, K_r$ (respectively), then we have that
$\{\pm \sqrt{\alpha_j} : 1 \leq j \leq 8\} = \Hom_\mathbb{Q}(L, \mathbb{Q}^\mathrm{sep})$. Thus, the permutation action of $\Gal(\mathbb{Q}^\mathrm{sep}/\mathbb{Q})$ on the $16$ points of $\mathfrak{C}$ is identical to the action of $\Gal(\mathbb{Q}^\mathrm{sep}/\mathbb{Q})$ on $\Omega$.
We can explicitly describe the linear space of quadrics containing $\mathfrak{C}$. Write $\mathbb{P}^4 := \Proj(\mathbb{Q}[y_0, y_1, y_2, y_3, y_4])$, which admits a natural projection to $\mathbb{P}^2 := \Proj(\mathbb{Q}[y_2, y_3, y_4])$. Under the Veronese embedding
\[
\mapdef{\nu_2}{\mathbb{P}^1}{\Proj(\mathbb{Q}[y_2, y_3, y_4])}{(z_0:z_1)}{(z_0^2 : z_0z_1 : z_1^2)}
\]
the quartic forms $z_0z_1^3, b(z_0, z_1), c(z_0, z_1)$ can be identified with the three quadratic forms $y_3y_4$, $b(y_2, y_3, y_4)$, $c(y_2, y_3, y_4)$; the identified quadric forms are unique modulo $y_3^2 - y_2y_4$. The determinantal representation for $f(z_0, z_1)$ becomes a quadratic determinantal representation, i.e., is a tensor in $\Sym^2 \mathbb{Q}^3 \otimes \Sym_2 \mathbb{Q}^2$.
\begin{construction} \label{cons: tensor Recillas}
Let $\mathbb{P}^2 := \Proj(k[y_2, y_3, y_4])$. Given a conic $C \subset \mathbb{P}^2$ and a tensor $\mathcal{B} \in \Sym^2 k^3 \otimes \Sym_2 k^2$ representing a $2 \times 2$ matrix of quadratic forms, we can construct a tensor $\mathcal{A} \in k^4 \otimes (\Sym_2 k^2 \oplus \Sym_2 k^3)$ as follows: Writing $\mathcal{B}(\mathbf{y}, \cdot) = \begin{bmatrix} a & b \\ b & c \end{bmatrix}$, we have that
\begin{align*}
a(\mathbf{y}) = \sum_{2 \leq i,j \leq 4} a_{ij} y_iy_j, \quad
b(\mathbf{y}) = \sum_{2 \leq i,j \leq 4} b_{ij} y_iy_j, \quad
c(\mathbf{y}) = \sum_{2 \leq i,j \leq 4} c_{ij} y_iy_j.
\end{align*}
We may write the defining equation for $C$ as $\sum_{2 \leq i,j \leq 4} d_{ij} y_iy_j$. (The terms with $i \neq j$ appear twice.) Define the slices of $\mathcal{A}$ to be
\begin{alignat*}{3}
&A_0 :=
\begin{bmatrix}
-1 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 0 & 0 \\
0 & 0 & a_{22} & a_{23} & a_{24} \\
0 & 0 & a_{32} & a_{33} & a_{34} \\
0 & 0 & a_{42} & a_{43} & a_{44} \\
\end{bmatrix},
\quad
&&A_1 :=
\begin{bmatrix}
0 & -\frac{1}{2} & 0 & 0 & 0 \\
-\frac{1}{2} & 0 & 0 & 0 & 0 \\
0 & 0 & b_{22} & b_{23} & b_{24} \\
0 & 0 & b_{32} & b_{33} & b_{34} \\
0 & 0 & b_{42} & b_{43} & b_{44} \\
\end{bmatrix},
\\
&A_2 :=
\begin{bmatrix}
0 & 0 & 0 & 0 & 0 \\
0 & -1 & 0 & 0 & 0 \\
0 & 0 & c_{22} & c_{23} & c_{24} \\
0 & 0 & c_{32} & c_{33} & c_{34} \\
0 & 0 & c_{42} & c_{43} & c_{44} \\
\end{bmatrix},
\quad
&&A_3 :=
\begin{bmatrix}
0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 \\
0 & 0 & d_{22} & d_{23} & d_{24} \\
0 & 0 & d_{32} & d_{33} & d_{34} \\
0 & 0 & d_{42} & d_{43} & d_{44} \\
\end{bmatrix}.
\end{alignat*}
The tensor $\mathcal{A}$ generically defines a genus $4$ curve via an intersection of the two symmetric determinantal varieties
$X_2 := Z(4x_0x_2 - x_1^2)$ and
$X_3 := Z\left(\det \mathcal{A}^{(2)}(\mathbf{x}, \cdot, \cdot) \right)$,
where
\[
\mathcal{A}^{(2)}(\mathbf{x}, \cdot, \cdot) :=
x_0
\begin{bmatrix}
a_{22} & a_{23} & a_{24} \\
a_{32} & a_{33} & a_{34} \\
a_{42} & a_{43} & a_{44} \\
\end{bmatrix}
+ x_1
\begin{bmatrix}
b_{22} & b_{23} & b_{24} \\
b_{32} & b_{33} & b_{34} \\
b_{42} & b_{43} & b_{44} \\
\end{bmatrix}
+ x_2
\begin{bmatrix}
c_{22} & c_{23} & c_{24} \\
c_{32} & c_{33} & c_{34} \\
c_{42} & c_{43} & c_{44} \\
\end{bmatrix}
+ x_3
\begin{bmatrix}
d_{22} & d_{23} & d_{24} \\
d_{32} & d_{33} & d_{34} \\
d_{42} & d_{43} & d_{44} \\
\end{bmatrix}.
\]
From the four symmetric matrices $A_0, \ldots, A_3$ we obtain an intersection of four quadrics in $\mathbb{P}^4$.
\end{construction}
\begin{comment}
\begin{example} \label{ex: main example 1}
Let
\[
f(z_0, z_1) := 4z_0^8 + z_0^7z_1 + 67z_0^6z_1^2 + 63z_0^5z_1^3 + 58z_0^4z_1^4 + 100z_0^3z_1^5 + 32z_0^2z_1^6 + z_1^8 \in \mathbb{F}_{101}[z_0, z_1].
\]
Then the tensor given by the construction above, represented as a matrix of linear forms, is
\[
\left(\begin{array}{rrrrr}
x_0 & 51x_1 & 0 & 0 & 0 \\
51x_1 & x_2 & 0 & 0 & 0 \\
0 & 0 & 100x_1 + 96x_2 & 0 & 51x_3 \\
0 & 0 & 0 & 85x_1 + 100x_2 + 100x_3 & 50x_0 + 100x_1 + 50x_2 \\
0 & 0 & 51x_3 & 50x_0 + 100x_1 + 50x_2 & 99x_1 + 94x_2
\end{array}\right)
\]
and the resulting genus $4$ curve in $\mathbb{P}^3$ is defined by the common vanishing locus of
\begin{align*}
q(x) &:= x_0x_2 + 25x_1^2, \\
g(x) &:= 76x_0^2x1 + 77x_0^2x_2 + x_0x_1^2 + 56x_0x_1x_2 + 53x_0x_2^2 + 70x_1^3 + 35x_1^2x_2 \\
& \quad + 99x_1^2x_3 + 9x_1x_2^2 + 84x_1x_2x_3 + 4x_1x_3^2 + 42x_2^3 + 66x_2^2x_3 + 76x_2x_3^2 + 76x_3^3.
\end{align*}
Computer algebra can be used to verify that the genus $4$ curve is nonsingular. (See subsection~\ref{sec: sub: software}.)
\end{example}
\end{comment}
\begin{example} \label{ex: main example 1}
Let
\[
f(z_0, z_1) := (z_0^2 - z_1^2)(z_0^2 - 4z_1^2)(z_0^2 - 9z_1^2)(z_0^2 - 16z_1^2) \in \mathbb{F}_{101}[z_0, z_1].
\]
Choosing
\[
b(z_0, z_1) := z_0^4 + 86z_0^2z_1^2 + 24z_1^4 \sim y_2^2 + 86y_3^2 + 24y_4^2, \qquad c(z_0, z_1) := -z_0z_1^3 \sim -y_3y_4,
\]
we have that $f(z_0, z_1) = b(z_0,z_1)^2 - z_0z_1^3 c(z_0,z_1)$.
The tensor given by Construction~\ref{cons: tensor Recillas}, represented as a matrix of linear forms, is
\[
\left(\begin{array}{rrrrr}
-x_0 & -51x_1 & 0 & 0 & 0 \\
-51x_1 & -x_2 & 0 & 0 & 0 \\
0 & 0 & x_1 & 0 & -51x_3 \\
0 & 0 & 0 & 86x_1 + x_3 & 51(x_0-x_2) \\
0 & 0 & -51x_3 & 51(x_0-x_2) & 24x_1
\end{array}\right)
\]
and the resulting genus $4$ curve in $\mathbb{P}^3$ is defined by the common vanishing locus of
\begin{align*}
q(x) &:= x_0x_2 - x_1^2, \\
g(x) &:= 25x_0^2x_1 + 51x_0x_1x_2 + 44x_1^3 + 24x_1^2x_3 + 25x_1x_2^2 + 29x_1x_3^2 + 25x_3^3.
\end{align*}
Computer algebra can be used to verify that the genus $4$ curve is nonsingular. (See subsection~\ref{sec: sub: software}.)
\end{example}
\begin{lemma} \label{lem: data for nonsingular genus 4 curve}
Let $L/\mathbb{Q}$ be an \'etale algebra arising from Construction~\ref{cons: etale algebra of degree 16}, let $K$ be the distinguished subalgebra, and let
\[
U := \left\{(\alpha_1, \ldots, \alpha_r) \in \prod_j K_j : L \cong \prod_j {K_j[x]}/{\gen{x^2-\alpha_j}}, \ K_j \cong \mathbb{Q}(\alpha_j)\right\}.
\]
Let $V \subseteq U$ be the subset such that for any $\alpha \in V$, Construction~\ref{cons: tensor Recillas} applied to the characteristic polynomial of $\alpha$ produces a nonsingular genus $4$ curve and a base locus $\mathfrak{C}$ of dimension $0$ and degree $16$ of the linear space of quadrics. Then $V$ is non-empty.
\end{lemma}
\begin{proof}
Example~\ref{ex: main example 1} shows that there is a non-empty open subscheme of $\mathrm{Pol}^8$ where the constructed curve is nonsingular. The result follows from Corollary~\ref{cor: sqrt primitive basis}.
\end{proof}
\subsection{Cycles on double covers}
In \cite{Reid1972quadrics}, Reid showed that the algebraic cycles on the complete intersection of 3 quadrics correspond to algebraic cycles within the Prym variety of the natural double cover of the degeneracy locus. An alternative reference for this construction is \cite{Tyurin1975}. In this section, we show how a similar construction allows us to identify pairs of points on the intersection of $4$ quadrics in $\mathbb{P}^4$ with the $112$ exceptional curves on a del Pezzo surface of degree one.
The following lemma contains some results we will freely use about quadrics.
\begin{lemma} \label{lem: properties of quadrics}
Let $k$ be a field of characteristic not equal to $2$, let $A \in \Sym_2 k^{m+1}$ be of rank $r>0$, and let $Q := Z(\mathbf{y}^TA\mathbf{y}) \subset \mathbb{P}_k^m$ be the associated quadric. Then:
\begin{enumerate}[(a)]
\item
Any singular quadric is a cone over a non-singular quadric.
\item
The singular locus of $Q$ is contained in every maximal isotropic subspace. In other words, every maximal isotropic subspace is a cone over an isotropic subspace of the nonsingular part. Furthermore, the singular locus of $Q$ is the linear subspace $\mathbb{P}(\ker A) \subset \mathbb{P}^m$.
\item
Over $\bar{k}$, the dimension of a maximal isotropic subspace is $\floor{\frac{r}{2}} + \dim(\ker A) - 1$.
\item
If $r$ is even, there are two distinct families of maximal isotropic subspaces of $Q$. If $r$ is odd, then there is a unique family of maximal isotropic subspaces.
\item
If $r$ is even, and $V,W$ are maximal, then $V,W$ are contained in the same family if and only if $\dim(V) - \dim (V \cap W) \equiv 0 \pmod 2$.
\end{enumerate}
\end{lemma}
\begin{proof}
See \cite{GriffithsHarris1994}.
\end{proof}
If $\mathcal{L}$ is a maximal isotropic subspace of a quadric $Q$, we will denote by $[\mathcal{L}]$ the family of maximal isotropic subspaces containing $\mathcal{L}$.
\begin{proposition} \label{prop: defining the generator cover}
Let $S_r^m$ be the space of quadrics of even rank $r$. Let $x \in S_r^m$ and $[\mathcal{L}]$ denote a family of maximal isotropic subspaces of $x$. Then the choice of generator $(x,[\mathcal{L}]) \mapsto x$ defines a nontrivial double cover branched over $S_{r-1}^m$.
\end{proposition}
\begin{proof}
See \cite[Section 5]{Tyurin1975}.
\end{proof}
\begin{definition}
The \emph{generator double cover} is the morphism $\gencover\: (x,[\mathcal{L}]) \mapsto x$ given by Proposition~\ref{prop: defining the generator cover}.
\end{definition}
The generator double cover allows us to identify the secants between $16$ points in $\mathbb{P}^4$ with exceptional curves of a del Pezzo surface of degree $1$. In our particular case of a block-diagonal tensor $\mathcal{A} \in k^4 \otimes (\Sym_2 k^2 \oplus \Sym_2 k^3)$, the locus of quadrics of rank at most $4$ is $X_5 := Z(\det \mathcal{A}(\mathbf{x}, \cdot, \cdot))$, which is the union of the quadric cone $X_2$ where the first block degenerates and the symmetroid cubic $X_3$ where the second block degenerates. The generator double cover $\gencover\: \widetilde X_5 \rightarrow X_5$ is branched over the locus of quadrics of rank at most $3$, which consists of the genus four curve $X = X_2 \cap X_3$ as well as the singularities of $X_2$ and $X_3$. Generically, the web of quadrics defined by $\mathcal{A}$ does not contain any quadrics of rank $2$.
We obtain by restriction a double cover $\gencover\: \widetilde X_2 \rightarrow X_2$ branched along a genus $4$ curve and the vertex of the cone; in other words, $\widetilde X_2$ is a del Pezzo surface of degree one.
Let $\mathfrak{C}$ be the intersection of all the quadrics in the web defined by $\mathcal{A}$, which we saw before is generically a complete intersection of four quadrics in $\mathbb{P}^4$. If $\ell$ is a secant of $\mathfrak{C}$ and $V_\ell \subset k^5$ is the affine cone over $\ell$, then we define a subscheme of $\widetilde X_2$ by
\[
\tau(\ell) := \left\{(x, [\mathcal{L}]) \in \widetilde X_2 : \mathcal{L} := V_\ell + \ker \mathcal{A}(x, \cdot, \cdot) \text{ is a maximal isotropic subspace of } \mathcal{A}(x, \mathbf{y}, \mathbf{y}) \right\}.
\]
Observe that $\tau(\ell)$ is well-defined; the kernel of a quadric $Q$ of rank $4$ generically does not meet $V_\ell$, so the space $V_\ell + \ker Q$ is a maximal isotropic subspace of $Q$. Notice that the automorphism $\eta\: \mathbb{P}^4 \rightarrow \mathbb{P}^4$ defined by $\eta\: (y_0:y_1:y_2:y_3:y_4) \mapsto (-y_0:-y_1:y_2:y_3:y_4)$ acts invariantly on all of the quadratic forms in the web defined by $\mathcal{A}$.
If $\mathcal{A}(x, \mathbf{y}, \mathbf{y})$ is a quadratic form of rank $4$ vanishing on $\ell$, then it must also vanish on $\eta(\ell)$.
Furthermore, the intersection of the maximal isotropic subspaces $V_\ell + \ker \mathcal{A}(x, \cdot, \cdot)$ and $\eta(V_\ell) + \ker \mathcal{A}(x, \cdot, \cdot)$ is generically $\ker \mathcal{A}(x, \cdot, \cdot)$, and thus the two maximal isotropic subspaces lie in opposite families.
In other words, the curves $\tau(\ell), \tau(\eta \ell)$ on $\widetilde X_2$ only intersect along the branch locus and are exchanged by $\Aut(\widetilde X_2/X_2)$.
The last result we need is a theorem from \cite{KulkarniVemulapalli}. To clarify the statement of the theorem, any nonhyperelliptic genus $4$ curve $X$ with a vanishing even theta characteristic $\theta_0$ has a unique vanishing even theta characteristic. This allows us to partition the $2$-torsion classes of the Jacobian variety into one of two types:
\begin{itemize}
\item (odd) $\epsilon \in \Jac(X)[2]$ is of the form $[\theta- \theta_0]$ for some odd theta characteristic of $X$.
\item (even) $\epsilon \in \Jac(X)[2]$ is not even.
\end{itemize}
There are $120$ odd $2$-torsion classes and $135$ nontrivial even $2$-torsion classes.
\begin{theorem}[{\cite[Theorem~1.1.3]{KulkarniVemulapalli}}]
\label{thm: genusfoursummary}
Let $k$ be a field of characteristic not $2$ or $3$. Then:
\begin{enumerate}[(a)]
\item
There is a canonical bijection between:
\[
\left \{ \begin{array}{c}
k \text{-isomorphism classes of tuples } (X, \epsilon, \theta_0),\\
\text{where } X \text{is a smooth genus $4$ curve with vanishing even theta} \\
\text{characteristic } \theta_0 \text{ with a rational divisor class defined over } k, \\
\text{and } \epsilon \text{ is a nontrivial even $2$-torsion class}
\end{array}
\right \} \longleftrightarrow \left \{ \begin{array}{c}
\text{nondegenerate orbit classes of } \\
k^4 \otimes \left(\Sym_2 k^2 \oplus \Sym_2 k^3 \right) \\
\text{under the action of } \\
\GL_4(k) \times \GL_2(k) \times \GL_3(k)
\end{array} \right \}
\]
\end{enumerate}
\noindent
Let $\mathcal{A} \in k^4 \otimes (\Sym_2 k^2 \oplus \Sym_2 k^3)$ be a nondegenerate tensor and let $\theta_0$ and $\epsilon$ be the associated line bundles on $X$. Then:
\begin{enumerate}[(b)]
\item
The images of the $120$ secants of $\mathfrak{C}$ under $\psi$ define $56 + 8$ tritangent planes of $X$. Viewing $X \cap H$ as a divisor of $X$, eight of these tritangents satisfy $X \cap H = 2D$ where $D \in |\theta_0|$. The other $56$ tritangents satisfy $X \cap H = 2D$, where $D$ is the effective representative of an odd theta characteristic of $X$.
\item
Let $e_2$ denote the Weil pairing on $\Jac(X)[2]$. Then the $56$ distinct odd theta characteristics constructed from the secants of $\mathfrak{C}$ are precisely the odd theta characteristics $\theta$ such that $e_2(\theta \otimes \theta_0^\vee, \epsilon) = 0$.
\end{enumerate}
\end{theorem}
\begin{comment}
\label{thm: genusfoursummary}
Let $\mathcal{A} \in k^4 \otimes (\Sym_2 k^2 \oplus \Sym_2 k^3)$ be a nondegenerate tensor and let $\theta_0$ and $\epsilon$ be the associated line bundles on $X$. Then:
\begin{enumerate}[(a)]
\item
The images of the $120$ secants of $\mathfrak{C}$ under $\psi$ define $56 + 8$ tritangent planes of $X$. Viewing $X \cap H$ as a divisor of $X$, eight of these tritangents satisfy $X \cap H = 2D$ where $D \in |\theta_0|$. The other $56$ tritangents satisfy $X \cap H = 2D$, where $D'$ is the effective representative of an odd theta characteristic of $X$.
\item
Let $e_2$ denote the Weil pairing on $\Jac(X)[2]$. Then the $56$ distinct odd theta characteristics constructed from the secants of $\mathfrak{C}$ are precisely the odd theta characteristics $\theta$ such that $e_2(\theta \otimes \theta_0^\vee, \epsilon) = 0$.
\end{enumerate}
\end{comment}
The inclusion $X \longhookrightarrow \widetilde X_2$ induces a restriction morphism of divisor classes. We denote by $\Pic(\widetilde X_2)^\perp$ the divisor classes orthogonal to the canonical class under the intersection pairing. There is an exact sequence
\[
\xym{
0 \ar[r] & 2\Pic(\widetilde X_2)^\perp \ar[r] & \Pic(\widetilde X_2)^\perp \ar[r] & \Pic(X)[2] \ar[r] & 0.
}
\]
Additionally, the anti-canonical class of $\widetilde X_2$ restricts to the unique vanishing even theta characteristic $\theta_0$ of $X$. If $e \in \Pic(\widetilde X_2)$ is an exceptional curve and $\kappa$ is the canonical class, then $e + \kappa \in \Pic(\widetilde X_2)^\perp$, the restriction of $e$ to $X$ defines an odd theta characteristic, and the sublattice generated by
\[
R_{D_8} := \{ e + \kappa \in \Pic(\widetilde X_2) : e^2 = -1, \ \ e_2((e + \kappa)|_{X}, \epsilon) = 0\}
\]
is a root lattice of type $D_8$; the set of roots of this lattice is precisely $R_{D_8}$.
\begin{proposition} \label{prop: Identification of $G$-sets}
Let $\ell$ be a secant of $\mathfrak{C}$. Then the curve $\tau(\ell)$ is one of the $112$ exceptional curves of $\widetilde X_2$ with the property that
\[
e_2((\tau(\ell) + \kappa)|_{X}, \epsilon) = 0.
\]
Furthermore, each of the $112$ exceptional curves of this type is of the form $\tau(\ell)$ for some secant of $\mathfrak{C}$.
\end{proposition}
\begin{proof}
Follows immediately from Theorem~\ref{thm: genusfoursummary} and the discussion above.
\end{proof}
The automorphism $\eta$ restricts to an automorphism of $\mathfrak{C}$. A pair of geometric points of $\mathfrak{C}$ is either:
\begin{itemize}
\item Type $8$: a pair of the form $\{p, \eta(p)\}$, or
\item Type $112$: a pair not of the form $\{p, \eta(p)\}$.
\end{itemize}
We see that any $G$-action preserves the type of a pair $\{p,q\}$. In particular,
Proposition~\ref{prop: Identification of $G$-sets} shows that the $G$-set of pairs of type $112$ is isomorphic (as a $G$-set) to the $112$ roots of the $D_8$ lattice identified in Proposition~\ref{prop: Identification of $G$-sets}.
\subsection{Proof of the main theorem}
Let $\rho\: \Gal(\mathbb{Q}^\mathrm{sep}/\mathbb{Q}) \rightarrow W_{D_8}$ be a continuous homomorphism with image $G$. By Construction~\ref{cons: etale algebra of degree 16}, we construct an \'etale algebra $L$ of degree $16$ on which Galois acts via $G$, as well as its distinguished subalgebra $K$ of degree $8$.
We may choose an element $(\alpha_1, \ldots, \alpha_r) \in V \subset K$ as in Lemma~\ref{lem: data for nonsingular genus 4 curve} and let $f$ be its characteristic polynomial over $\mathbb{Q}$. Construction~\ref{cons: tensor Recillas} allows us to construct a locus of $16$ points with a $G$-action from $f$, as well as a nonsingular genus $4$ curve $X$ contained in a quadric cone $X_2$. The domain of the generator double cover $\gencover\: \widetilde X_2 \rightarrow X_2$ is a del Pezzo surface of degree one, and the Galois action on the $112$ roots of the $D_8$ sublattice from Proposition~\ref{prop: Identification of $G$-sets} has the Galois action prescribed by $\rho$.
\renewcommand{\MR}[1]{}
\begin{bibdiv}
\begin{biblist}
\bib{Magma}{article}{
author={Bosma, Wieb},
author={Cannon, John},
author={Playoust, Catherine},
title={The Magma algebra system. I. The user language},
note={Computational algebra and number theory (London, 1993)},
journal={J. Symbolic Comput.},
volume={24},
date={1997},
number={3-4},
pages={235--265},
issn={0747-7171},
review={\MR{1484478}},
doi={10.1006/jsco.1996.0125},
}
\bib{Corn2007}{article}{
author={Corn, Patrick},
title={The Brauer-Manin obstruction on del Pezzo surfaces of degree 2},
journal={Proc. Lond. Math. Soc. (3)},
volume={95},
date={2007},
number={3},
pages={735--777},
issn={0024-6115},
review={\MR{2368282}},
doi={10.1112/plms/pdm015},
}
\bib{Dol2012}{book}{
author={Dolgachev, Igor V.},
title={Classical algebraic geometry},
note={A modern view},
publisher={Cambridge University Press, Cambridge},
date={2012},
pages={xii+639},
isbn={978-1-107-01765-8},
review={\MR{2964027}},
doi={10.1017/CBO9781139084437},
}
\bib{ElsenhansJahnel2019}{article}{
author={Elsenhans, Andreas-Stephan},
author={Jahnel, J\"{o}rg},
title={On plane quartics with a Galois invariant Cayley octad},
journal={Eur. J. Math.},
volume={5},
date={2019},
number={4},
pages={1156--1172},
issn={2199-675X},
review={\MR{4015450}},
doi={10.1007/s40879-018-0292-3},
}
\bib{GriffithsHarris1994}{book}{
author={Griffiths, Phillip},
author={Harris, Joseph},
title={Principles of algebraic geometry},
series={Wiley Classics Library},
note={Reprint of the 1978 original},
publisher={John Wiley \& Sons, Inc., New York},
date={1994},
pages={xiv+813},
isbn={0-471-05059-8},
review={\MR{1288523}},
doi={10.1002/9781118032527},
}
\bib{KlunersMalleDatabaseArticle}{article}{
author={Kl\"{u}ners, J\"{u}rgen},
author={Malle, Gunter},
title={A database for field extensions of the rationals},
journal={LMS J. Comput. Math.},
volume={4},
date={2001},
pages={182--196},
review={\MR{1901356}},
doi={10.1112/S1461157000000851},
}
\bib{KlunersMalleDatabase2021}{article}{
author={Kl\"{u}ners, J\"{u}rgen},
author={Malle, Gunter},
title={A Database for Number Fields},
date={2021},
note={online database},
eprint={http://galoisdb.math.upb.de/home}
}
\bib{KulkarniVemulapalli}{article}{
author={Kulkarni, Avinash},
author={Vemulapalli, Sameera},
title={On Intersections of symmetric determinantal varieties and theta characteristics of canonical curves},
date={2021}
eprint={arXiv:2109.08740}
}
\bib{milneFT}{book}{
author={Milne, James S.},
title={Fields and Galois Theory (v5.00)},
year={2021},
note={Available at www.jmilne.org/math/},
pages={142}
}
\bib{Reid1972quadrics}{thesis}{
author = {Reid, Miles},
title = {The complete intersection of two or more quadrics},
note = {Ph.D. dissertation},
publisher = {Cambridge University},
year = {1972}
}
\bib{Tyurin1975}{article}{
author = {Tyurin, A. N.},
title = {On intersections of quadrics},
year = {1975},
journal = {Russian Mathematical Surveys},
doi = {http://dx.doi.org/10.1070/RM1975v030n06ABEH001530},
volume = {30},
issue = {6}
}
\end{biblist}
\end{bibdiv}
\end{document}
|
\section{Introduction}
Technological advances for collecting and measuring information in the biomedical field and beyond have led to an explosion in high-dimensional data. Such data can be used to identify patterns or markers and predict an outcome of interest. However, in fields such as genomics, proteomics, and chemometrics, the high-dimensional data is often functional and possesses complicated correlation structures. These complexities pose challenges to statistical and machine learning methods that are used for analyzing the data.
As a motivating concrete example, we consider the context of predicting phenotypes and identifying biomarkers based on mass spectrometry (MS) data. MS technology measures the mixtures of proteins/peptides of tissues or fluids and produces an MS spectrum
\citep{cruz2008comparison}. The resulting experimental data consists of discretely observed functional spectra, with typically tens of thousands of observed locations and just a few hundred samples. Moreover, data at neighboring locations tends to be highly correlated, with the strength of such correlations varying across the mass-to-charge ($m/z$) range. In addition, MS data tends to be noisy due to chemical noise, misalignment, calibration and other issues \citep{cruz2008comparison}.
To overcome such technical challenges,
many approaches have been proposed which differ in how they deal with the complexity of the functional data. The most common approaches rely on a two-stage process, where the first stage reduces the dimension of the functional predictors (e.g. through functional principal component analysis), followed by the second stage which applies a classification technique on the reduced features \citep[e.g.][]{hall2001functional, ferre2006multilayer}. More sophisticated techniques employ generalized functional linear regression to estimate the functional coefficients through basis expansion. Different choices of basis functions have been considered, including splines \citep{James2002,cardot2003spline}, wavelets \citep{brown2001bayesian}, and step functions \citep{Grollemund2018}. In addition, regularization is typically employed to avoid overfitting, through prior distributions or penalty functions, including lasso \citep{zhao2012wavelet}, Bayesian variable selection priors \citep{zhu2010bayesian}, and random series priors \citep{li2018bayesian}.
An overview is provided in \citet{reiss2017methods}. However, most of these methods do not perform variable selection to identify intervals or regions of the functional inputs which are relevant for predicting the class labels. Variable selection is important in this setting as retaining a large number of non-discriminative variables exacerbates the risk of overfitting. Moreover, variable selection allows
identification of disease markers and improves model interpretability. Relevant literature has demonstrated the importance of including a variable selection component in classification and prediction models with functional data, for example, fused lasso and its Bayesian analogues \citep{tibshirani2005sparsity, casella2010penalized}, Bayesian fused shrinkage \citep{song2018bayesian}, and Bayesian sparse step functions \citep{Grollemund2018}. For multivariate functions of the input space, various extensions of penalizations or Bayesian shrinkage and spike-and-slab priors have been proposed to incorporate spatial smoothness in the functional coefficient and variable selection \citep[e.g][]{goldsmith2014,li2015spatial, kang2018scalar}.
In this work, we focus on the discriminant analysis (DA) framework, which provides an alternative approach to generalized functional regression for classification tasks. In DA, the conditional distribution of the inputs ${\bm x}$ given the class label $y$ is modelled, which is then flipped via Bayes theorem to obtain the classification rule for $y$ given ${\bm x}$. The standard approach assumes that the class conditional distribution is a multivariate normal. A number of extensions have been developed for high-dimensional functional data, including penalized linear DA \citep{hastie1995penalized} and functional linear DA \citep{james2001functional}. Notably, \cite{murphy2010variable} proposed a quadratic DA which incorporates variable selection through constraints on the covariance matrices, whereas \cite{ferraty2003} proposed a nonparametric DA based on kernel density estimation. In the Bayesian paradigm, \cite{stingo2011variable} developed a quadratic DA which includes latent binary indicators for variable selection with a Markov random field prior to incorporate known structures. \cite{stingo2012bayesian} extended this by applying a wavelet transformation to account for smoothness in the functional inputs. \cite{GUTIERREZ201456} developed a robust Bayesian nonparametric (BNP) version of quadratic DA through a two-stage approach that first selects variables based on a fitted Gaussian process (GP) model for the functional data and individual quadratic DA across variables, and at the second stage, employs BNP mixture models for flexible discriminant analysis based on the selected variables.
In this article, we propose a novel and scalable Bayesian DA which performs variable selection and classification jointly, by combining recent developments in deep GPs \citep{dunlop2018deep} to flexibly model the functional inputs and incorporating Ising priors to identify differentially-distributed locations within a unified model framework. While powerful, GP models suffer from a well-known computational burden due to the need to store and invert large covariance matrices. Several approaches have been proposed in literature to enable scalability of GP models \citep[e.g.][]{Kumar2009, Hensman2013, salimbeni2017doubly, Geoga2020}. In this work, we develop a scalable inference algorithm that ameloriates the computational burden by utilizing the link between GPs and stochastic partial differential equations (SPDEs) to construct sparse precision matrices \citep{Lindgren2011, grigorievskiy2017}
and combine various variational inference schemes.
Specifically, we focus on a two-level architecture of deep GPs with exponential covariance functions due to the attractive balance between flexibility and computational efficiency, and describe how to extend this framework to other settings.
We apply and demonstrate the utility of our model on proteomics-related mass spectrometry datasets, while also highlighting its relevance to other applications with functional data, including temporal gene expression \citep{leng2006classification}, nuclear magnetic resonance spectroscopy \citep{allen2013regularized}, chemometrics \citep{murphy2010variable}, agricultural production based temporal measurements \citep{Grollemund2018}, speech recognition \citep{hastie1995penalized}, and pregnancy loss based on hormonal indicators \citep{bigelow2009bayesian}, among others. Lastly, we emphasize that our proposed Bayesian model offers explainability as well as uncertainty quantification, which are desirable qualities for increasing trust and widespread acceptance of data-driven tools in biomedicine and other scientific fields.
The paper is structured as follows. We present our proposed model in Section \ref{sec:model} and the developed posterior inference scheme in Section \ref{sec:inference}. In Section \ref{improveEfficiency}, we provide details of the implementation tricks required for scalability and efficiency of the algorithm.
In Section \ref{sec:results}, we present numerical results to demonstrate the performance of our
methodology in simulated and real datasets. Section \ref{sec:conc} summarizes and concludes with future directions.
\section{Model}\label{sec:model}
\begin{figure}[!t]
\centering
\subfloat[Controls]{\includegraphics[width=0.5\textwidth]{spectra_nocovid.jpg}}
\subfloat[SARS-CoV-2]{\includegraphics[width=0.5\textwidth]{spectra_covid.jpg}}\\
\subfloat[Class-averaged spectra]{\includegraphics[width=0.65\textwidth]{avgspec_detpeaks_covid.pdf}\label{fig:covid_avg}}
\caption{Illustration of the SARS-CoV-2 data \citep{nachtigall2020detection}, with the square root intensities of the preprocessed spectra for (a) healthy controls and (b) SARS-CoV-2 positive patients (class-average given in red). The class averages are compared in (c) and the detected peaks, identified through the standard two-stage pipeline in \citet{nachtigall2020detection}, are marked with stars.}
\label{fig:covid_spectra}
\end{figure}
Consider a set of $n$
functional inputs
$\{ x_i (t) \}_{i=1}^n$ defined on the domain $\mathcal{D} \subset \bbR$ and their corresponding class labels $\{ y_i \}_{i=1}^n$, where $x_i(t) \in \bbR$, $y_i \in \mathcal{Y}$, and $\mathcal{Y}$ is a set of class labels.
As a motivating example, Figure \ref{fig:covid_spectra} depicts a set of processed MS spectra for healthy controls and SARS-CoV-2 positive patients \citep{nachtigall2020detection}, with the $x$-axis indicating the mass-to-charge ratio ($m/z$) and the $y$-axis indicating the intensity of the protein or peptide ions. The commonly used pipeline for analysing such data follows a two-stage approach. At the first stage, peak detection and extraction protocols are employed to yield a set of $m/z$ values and their intensity readings corresponding to the detected peaks (for SARS-CoV-2, the class-averaged spectra and detected peaks are compared in Figure \ref{fig:covid_avg}). At the second stage, information from these peak regions are used to predict the outcome of interest and identify differentially-expressed proteins.
However, the initial steps of peak detection and extraction are critical for classification and may exclude important information from regions along the functional trajectory which
are useful for classification \citep{liu2009comparison}.
A viable solution is to build a unified data analysis framework that utilizes the entire high-dimensional inputs and
incorporates variable selection within the model. In the context of DA, the classical vanilla model
is well-known to suffer from poor classification accuracy in high-dimensional datasets
\citep{Bickel2004, Fan2008}.
The two-stage approach, which first selects variables according to specified criteria and then classifies with DA on the selected variables \citep[some examples in][]{Fan2008, DuarteSilva2011, Cui2015}, provides some improvements. However, information is lost when variable selection and classification are performed in separate stages. For example, two variables with adjusted p-values of 0.003 and 0.03 may be selected, but their difference in significance levels is not accounted for at the classification stage. Some proposed methods that circumvent this loss of information by incorporating variable selection directly within a DA model have been proposed have led to good classification performance in several examples \citep{Witten2011, Romanes2020}.
In the same spirit, we propose a novel unified model that performs the variable selection and the classification steps simultaneously.
For brevity, we present our method in the context of a binary classification problem (although an extension to more than two classes is straightforward). Our proposed approach builds on a DA framework that employs GPs to flexibly model the entire functional trajectory:
\begin{equation} \label{eqn::datadist}
x_i \; \vert \; y_i=k, \mu_k, \sigma_k, \psi_i \overset{\text{ind}}{\sim}
\GP(\mu_k, K_{\psi_i} + \mathcal{E} _k ),
\end{equation}
where $y_i = k \in \lbrace 0,1 \rbrace$ refers to the class label; $\mu_k (t) \in \bbR$ is the group-specific mean function; $K_{\psi_i}$ is a covariance function (or kernel) with observation-specific parameters $\psi_i$; $\mathcal{E} _k$ is a white noise kernel with class- and location-dependent variance $\sigma_k^2 (t) > 0$; and $\GP(\mu, K + \mathcal{E} )$ denotes a Gaussian process with mean function $\mu$ and covariance function $K+\mathcal{E} $. Here, $K$ and $\mathcal{E} $ induce the marginal variance $\text{Var} \{ x_i(t) \} = K(t,t) + \mathcal{E} (t)$ and the pairwise covariance $\text{Cov} ( x_i(t), x_i(t^\prime ) ) = K(t,t^\prime)$. Each stochastic process may be decomposed as
$$
x_i (t) = \mu_k (t) + z_i (t) + \epsilon_{k,i}(t),
$$
where
$z_i \mid \psi_i \sim \GP ( 0, K_{\psi_i})$ is an observation-specific latent process and $\epsilon_{k,i} (t) \in \bbR$ is a white-noise process with
variance $\sigma_k^2 (t) > 0$. Note the latent process $z_i$ accounts for the covariances between the values of the stochastic process at multiple locations, and $\epsilon_{k,i} (t)$ allows location-varying noisy errors, which are often present in functional data, such as the MS trajectories in Figure \ref{fig:covid_spectra}. We allow for variable selection in our proposed model by defining a binary signal process $\gamma (t) \in \{0,1\}$ such that
\begin{equation*}
\mu_k (t) = \gamma (t) \widetilde{\mu}_k (t) + (1 - \gamma (t) ) \widetilde{\mu}_\emptyset (t) \;\; \text{and} \;\; \sigma_k^2 (t) = \gamma (t) \widetilde{\sigma}_k^2 (t) + (1 - \gamma (t) ) \widetilde{\sigma}_\emptyset^2 (t),
\end{equation*}
where $\widetilde{\mu}_k$ and $\widetilde{\sigma}_k^2$ are the group-specific mean and noise variance processes at discriminative locations and $\widetilde{\mu}_\emptyset$ and $\widetilde{\sigma}_\emptyset^2$ are the common mean and noise variance processes at non-discriminative locations. In the context of MS data, $\gamma$ allows for detection of relevant $m/z$ values within the classification model. Thus, the entire aligned spectra are the inputs of the classification model, combining the steps of peak detection, feature extraction, and classification into a unified modeling framework. An example of a pair of $\mu_1$ and $\mu_0$ are depicted in the Supplementary Material.
The above parameterization allows us to directly infer the components $\widetilde{\mu}_k$ and $\widetilde{\mu}_\emptyset$ for the mean functions and $\widetilde{\sigma}_k^2$ and $\widetilde{\sigma}_\emptyset^2$ for the noise variance functions, as well as the signal process $\gamma$, which determines the regions where the mean and variance differ between groups.
\paragraph{Two-level non-stationary Gaussian processes.}
In several examples such as in our motivating example of MS data, the observed functions are often unevenly rough which is indicative of a non-stationary covariance structure. This behavior is evident in Figure \ref{fig:covid_spectra}, where the spectra are flatter in some regions and change more rapidly in others. To account for this behavior, we assign a non-stationary covariance kernel \citep{Paciorek2003} for $K_{\psi_i}$.
Specifically, the kernel parameter, $\psi_i = (\tau, \nu_i)$, consists of the magnitude $\tau>0$ and a location-varying log length-scale process $\nu_i$, i.e., $K_{\psi_i} = K_{NS;\tau, \nu_i}$, and hence we may write
\begin{align*}
z_i | \tau, \nu_i \sim \GP(0, K_{NS;\tau, \nu_i}).
\end{align*}
At the second level, we place Gaussian process priors on the log length-scale processes with
$$\nu_{i} (t) = R(t) + \zeta_i, \quad \text{ and } \quad R \sim \GP(\mu_\nu, K_{S;\tau_2, \lambda}),$$
where $K_{S;\tau_2, \lambda}$ is a stationary covariance kernel with marginal scale $\tau_2$ and length scale $\lambda$. Here, each observation-specific log length-scale process has been decomposed into a common component $R(t)$ to account for the location-varying covariance structure common across all observed functions, and an observation-specific perturbation $\zeta_i \in \bbR$ to allow for between-spectra variation in smoothness across the entire domain. This behavior is illustrated in Figure \ref{fig:ns_illus}. This decomposition is flexible yet also reduces the computational cost of having to infer $n$ observation-specific location-varying log length-scales (details are provided in the Supplementary Material).
\begin{figure}[!t]
\centering
\subfloat[Common log length-scale $R(t)$]{\includegraphics[width=0.5\textwidth]{S.pdf}}
\subfloat[Non-stationary process $z_i(t)$]{\includegraphics[width=0.5\textwidth]{z.pdf}}
\caption{Illustration of the two-level non-stationary GP. Larger/smaller values of the common log-length scale process $R(t)$ in (a) result in flatter/more wiggly behavior for all observed latent $z_i(t)$. In addition, the log-length scale process for each observation is perturbed by $\zeta_i$, allowing for some $z_i(t)$ (e.g. with $\zeta_i = 4$) to be flatter than others across the entire domain. }
\label{fig:ns_illus}
\end{figure}
Due to the observed roughness in our motivating dataset (refer to Figure \ref{fig:covid_spectra}), we focus on the exponential covariance function, a member of the Mat\'ern family with smoothness parameter $\nu=1/2$. In one-dimension, this is also known as the Ornstein-Uhlenbeck process, and it is the continuous-time counterpart of the first-order autoregressive model AR(1). Motivated by the link between GPs and SPDEs \citep{Lindgren2011, monterrubio2018}, we employ an SDE representation of the nonstationary processes:
\begin{align}
dz_i &= -\frac{1}{\exp(\nu_i)} z_i dt + \sqrt{\frac{2 \tau}{\exp(\nu_i)}} d \omega_{1},
\label{eq:SDElev1}\\
dR &= -\frac{1}{\lambda} R dt + \sqrt{\frac{2 \tau_{2}}{\lambda}} d \omega_{2}, \label{eq:SDElev2}
\end{align}
where $ \nu_{i} (t) = R(t) + \zeta_i$ and $\omega_{1}$ and $\omega_{2}$ are Wiener processes.
While appropriate for rough MS data,
the SDE representation can be extended to other choices of covariance functions, such as Mat\'ern processes with smoother realisations \citep{grigorievskiy2017}.
The resulting two-level GP model
provides the flexibility needed to capture non-stationarities, but this could be further extended with a non-stationary covariance kernel also for the log length-scale process by using deeper architectures \citep{dunlop2018deep,zhao2020deep}.
\subsection{Finite-time discretization}\label{sec:finite}
In practice, the functional realizations ${\bm x}_i = (x_i(t_1), \ldots, x_i(t_T))$ are discretely-observed at locations $t_1 < \ldots < t_T$\footnote{For ease of notation, we assume the locations $t_1 < \ldots < t_T$ are common across all observations, but the discretized model could be extended accordingly.}. Following our proposed model for the process $x_i$ in \eqref{eqn::datadist}, the likelihood of the discretely-observed vector ${\bm x}_i$ is a multivariate normal density:
$$
{\bm x}_i \; \vert \; y_i, \bm \mu_{y_i}, \bm \nu_i, \tau, \bm \sigma_{y_i}^2 \sim \left\{ \begin{array}{ll}
\N(\bm \mu_1, Q_{NS;\tau, \bm \nu_i} ^{-1} + D_{\epsilon1} ), & \mbox{ if $y_i = 1$;} \\ [1ex]
\N(\bm \mu_0, Q_{NS;\tau, \bm \nu_i} ^{-1} + D_{\epsilon0} ), & \mbox{ if $y_i = 0$},
\end{array} \right.
$$
where $\bm \mu_k = ( \mu_k(t_1), \ldots, \mu_k (t_T ))^\top$; $\bm \sigma_{k}^2 = ( \sigma_k^2 (t_1), \ldots, \sigma_k^2 (t_T) )^\top$ ; $ \bm \nu_i = \zeta_i + {\bm R} $ with ${\bm R} = (R(t_1),\ldots, R(t_T))^\top$; $D_{\epsilon k} = \text{dg} ( \sigma_k^2 (t_1), \ldots, \sigma_k^2 (t_T) )$; dg forms a diagonal matrix from elements of a vector, and $Q_{NS;\tau, \bm \nu_i}$ is a precision matrix parameterized $\tau$ and $\bm \nu_i$. Again, the discretely-observed ${\bm x}_i$ can be decomposed as:
$${\bm x}_i = \bm \mu_k + {\bm z}_i + \bm \epsilon_{k,i}, \quad \text{ for } y_i = k,$$
where ${\bm z}_i = (z_i(t_1), \ldots, z_i(t_T) )^\top $ is the observation-specific latent vector with precision matrix $Q_{NS;\tau, \bm \nu_i}$; $\bm \epsilon_{k,i}$ is the white noise vector with covariance matrix $ D_{\epsilon k}$;
$$ \bm \mu_k = {\bm \gamma} \odot \widetilde{\bm \mu}_k + ({\bm 1}-{\bm \gamma}) \odot \widetilde{\bm \mu}_\emptyset, \quad \text{ and } \quad \bm \sigma_k = {\bm \gamma} \odot \widetilde{\bm \sigma}_k + ({\bm 1}-{\bm \gamma}) \odot \widetilde{\bm \sigma}_\emptyset,$$
where $\odot$ denotes element-wise product. At the second level, the common log length-scale vector ${\bm R}$ has a multivariate normal distribution,
$$ {\bm R} \, \vert \, \lambda, \tau_2 \sim \N( {\bm 0}, Q_{S; \lambda, \tau_2}^{-1} ). $$
To overcome the computational bottleneck of GPs, we construct sparse precision matrices $Q_{NS;\tau, \bm \nu_i}$ and $Q_{S; \lambda, \tau_2}$ by discretizing the SDE representations in \eqref{eq:SDElev1} and \eqref{eq:SDElev2}, respectively, on the grid $t_1 < \ldots < t_T$ using the Euler-Maruyama scheme:
\begin{align}
z_i(t_j) &= z_i (t_{j-1}) -\frac{1}{\exp(\nu_i(t_{j-1}))} z_i(t_{j-1}) \delta_j + \sqrt{\frac{2 \tau}{\exp(\nu_i(t_{j-1}))}} w_{1,j}, \label{eq:em1}\\
R(t_{j}) &= R(t_{j-1}) -\frac{1}{\lambda} R(t_{j-1}) \delta_j + \sqrt{\frac{2 \tau_{2}}{\lambda}} w_{2,j}, \label{eq:em2}
\end{align}
where $\nu_i(t_{j}) = \zeta_i + R(t_{j}) $; $w_{1,j} \sim \text{N}(0, \delta_j) \independent w_{2,j} \sim \text{N}(0, \delta_j) $; and $\delta_j = t_j - t_{j-1}$.
For notational simplicity, we assume the locations are equally-spaced, i.e., $t_j - t_{j-1} = \delta$. Through \eqref{eq:em1} and \eqref{eq:em2}, the precision matrices of ${\bm z}_i$
and ${\bm R}$ have a tridiagonal structure (specific form in the Supplementary Material), which alleviates the complexity and enhances the efficiency of our posterior inference algorithm. A detailed discussion is provided in Section \ref{improveEfficiency}.
\subsection{Choice of priors}
To reflect our prior belief that the underlying variable selection process $\gamma$ is smooth, we assign a linear chain Ising prior \citep{li2010} to account for smoothness. Following this choice of prior, the conditional distribution of $\gamma(t)$ given its corresponding set of neighbors with locations in $\mathcal{N}_t \subset \lbrace t_1, \ldots, t_T \rbrace$ is
\begin{align*}
\bbP ( \gamma (t) = 1 \, \vert \, \{ \gamma (t^\prime) \}_{t^\prime \in \mathcal{N}_t} ) = \expit \left \{ -\alpha + \sum_{t^\prime \in \mathcal{N}_t} \beta(t, t^\prime) \gamma (t^\prime) \right \},
\end{align*}
where $\expit(x) = \{1+ \exp(-x) \}^{-1}$, $\alpha \in \bbR$, and $\beta(t,t^\prime) > 0$. Here, a larger value of $\alpha$ corresponds to more sparsity in $\gamma$, whereas $\beta$ controls the correlation and smoothness between the values of $\gamma$ at neighboring locations.
In our context, we define $\mathcal{N}_t= \lbrace t-1, t+1 \rbrace$ and $\beta(t,t') =\beta$.
Motivated by the non-stationary behavior also evident in the class-averaged functions (see Figure \ref{fig:covid_spectra}), we also assign hierarchical GP priors for the mean functions
\begin{align*}
\widetilde{\mu}_k (t) \, \vert \, \widetilde{\tau}_k, \widetilde{\nu}_k \, &\sim \, \GP(0, K_{NS; \widetilde{\tau}_k, \widetilde{\nu}_k}), \\
\widetilde{\nu}_k (t) \, \vert \, \widetilde{\eta}, \widetilde{\lambda} \; &\sim \; \GP (\mu_{\widetilde{\nu}} , K_{S; \widetilde{\eta}, \widetilde{\lambda} } ),
\end{align*}
with hyperpriors $\widetilde{\tau}_k \sim \text{InvGa}( A_{\widetilde{\tau}}, B_{\widetilde{\tau}} )$ for $k = 0, 1, \emptyset$;
$\widetilde{\eta} \sim \text{InvGa}(A_{\widetilde{\eta}},B_{\widetilde{\eta}})$; and $\widetilde{\lambda} \sim \text{LogN}( \mu_{\widetilde{\lambda}}, \sigma_{\widetilde{\lambda}}^2)$. Here, the hyperparameters of the length scale $\widetilde{\lambda}$ are specified such that the resultant prior is weakly informative to mitigate potential non-identifiability issues \citep[][and further details in the Supplementary Material]{Betancourt2017robust}.
For the noise variance, we assign an independent prior at each location
$$
\widetilde{\sigma}_k^2 (t) \, \sim \, \text{InvGa} (A_\epsilon, B_\epsilon), \text{ for } k =0,1, \;\; \text{and} \;\; \widetilde{\sigma}_\emptyset^2 (t) \, \sim \, \text{InvGa} (A_\epsilon, B_\epsilon).
$$
For the magnitude $\tau$ of the $z_i$ and the magnitude $\tau_2$ of ${\bm R}$, we assume $\tau \sim \text{InvGa}(A_\tau, B_\tau)$ and $\tau_2 \sim \text{InvGa}(A_{\tau_2}, B_{\tau_2})$, respectively. The observation-specific perturbations are assigned the priors $\zeta_i \overset{\text{iid}}{\sim} \text{N}(0, \sigma_\zeta^2)$, where $\sigma_\zeta^2=1$. For the length scale $\lambda$, we assign the weakly informative hyperprior $\lambda \sim \text{LogN}( \mu_{\lambda}, \sigma_{\lambda}^2)$.
A directed acyclic graph (DAG) that summarizes the parameters in our proposed model is provided in Figure \ref{fig:DAG}.
\begin{figure}[!h]
\centering
{\includegraphics[width=0.8\textwidth]{DAG.pdf}}
\caption{DAG representation of the proposed model. The colour fills denote the inference approach adopted: blue for MAP; red for SVB; white for CAVI; and gray denotes observed data. More details of the various inference approaches is provided in Section \ref{sec:inference}. }
\label{fig:DAG}
\end{figure}
\section{Inference and prediction}\label{sec:inference}
While Markov chain Monte Carlo is a popular technique for computing posterior distributions in Bayesian modelling,
an alternative approximate Bayesian inference method known as \textit{variational Bayes} has gained popularity in the literature. Variational Bayes has
been shown to be a fast posterior computation method that yields reasonably accurate approximations in several problems.
Consider fitting a model parameterized by ${\bm \theta}$ to the observed data $\mathcal{D}$.
In variational Bayes, the actual posterior $p({\bm \theta} \, \vert \, \mathcal{D})$ is approximated by a density $q({\bm \theta})$ from a family of distributions $\mathcal{F}$ that maximizes the evidence lower bound (ELBO)
\begin{equation}
\label{eqn::ELBO}
\mathbb{E}_{q({\bm \theta}) } \left [ \log \left \{ \frac{p({\bm \theta}, \mathcal{D}) }{q({\bm \theta})} \right\} \right ].
\end{equation}
A common choice for $\mathcal{F}$ is the mean-field family on the partition $\{ {\bm \theta}_1, \ldots, {\bm \theta}_L \}$ of ${\bm \theta}$:
$$
q({\bm \theta}) = \prod_{l=1}^L q_l({\bm \theta}_l),
$$
where $L \le \text{dim} ({\bm \theta})$. Without any further parametric assumptions, it has been shown \citep{Ormerod2010} that the optimal choice for each product component $q_l$ is
\begin{equation}
\label{eqn::CAVIupdate}
q_l ({\bm \theta}_l) \propto \exp \left [ \bbE_{- {\bm \theta}_l} \log \{ p({\bm \theta}, \mathcal{D}) \} \right ],
\end{equation}
where the above expectation is taken with respect to $\prod_{l^\prime \neq l} q_{l^\prime} ({\bm \theta}_{l^\prime})$. This choice of product component is known as \textit{coordinate ascent variational inference} (CAVI). Note that in some cases, the RHS in \eqref{eqn::CAVIupdate} is intractable. In such cases, we adopt alternative methods such as stochastic variational Bayes (SVB) or maximum a posteriori (MAP) to circumvent the intractability.
\subsection{Posterior inference}\label{sec:vb}
Let ${\bm \theta}$ denote the vector of all model parameters (excluding the hyperparameters $\zeta_i$, $\lambda$, $\tau_2$, $\widetilde{\lambda}$, $\widetilde{\eta}$, $\alpha$, and $\beta$).
We specify the mean-field family for the approximate posterior:
\begin{align*}
q({\bm \theta}) &= q(\tau) q({\bm R}) \prod_{i=1}^n \{ q({\bm z}_i) \} \times \prod_{k \in \{ \emptyset, 0, 1 \} } \left \{ q(\widetilde{\bm \mu}_k ) q( \widetilde{\bm \nu}_k) q( \widetilde{\tau}_k) \prod_{j=1}^T q(\widetilde{\sigma}_{kj}^{2}) \right \} \times \prod_{j=1}^T \{ q(\gamma_j) \}.
\end{align*}
In the following, we provide details on the functional form of each product component. Note that the product components for the location-varying log length-scales, i.e. ${\bm R}$ and $\bm \nu_i$, are computed using SVB updates, while CAVI updates are employed for all other parameters in ${\bm \theta}$. The remaining hyperparameters $\zeta_i$, $\lambda$, $\tau_2$, $\widetilde{\lambda}$, $\widetilde{\eta}$, $\alpha$, and $\beta$ are optimized using MAP estimation, with the marginal likelihood approximated by the ELBO \eqref{eqn::ELBO}. In Figure \ref{fig:DAG}, parameters in white, red, and blue fill are updated with CAVI, SVB and MAP respectivly, whereas gray fill denotes observed quantities. For the rest of this section, we use the notation $\bbE$ to denote an expectation with respect to the variational posterior and a subscript $j$ to denote the value of the functional at location $t_j$, e.g. $\widetilde{\mu}_k (t_j) = \mu_{kj}$ and $\gamma_j = \gamma(t_j)$. Furthermore, as proof of concept, we present our approximate posteriors in the case whereby each stochastic process $x_i$ is observed at the same set of equally-spaced locations $t_1, \ldots, t_T$, with $t_{j+1} - t_j = \delta$.
\paragraph{Mean function parameters.} The product component for $\widetilde{\bm \mu}_k$ is:
\begin{equation*}
q(\widetilde{\bm \mu}_k) = \text{N} (m_{\widetilde{\bm \mu}_k}, Q_{\widetilde{\bm \mu}_k}^{-1} ),
\end{equation*}
where, for $k =0,1$ and $k = \emptyset$, respectively, we have
\begin{align*}
&m_{\widetilde{\bm \mu}_k} = Q_{\widetilde{\bm \mu}_k}^{-1} \bbE ( \widetilde{D}_{\epsilon k}^{-1} ) {\bm W} \left ( {\bm X}^{(k)} - m_{{\bm Z}}^{(k)} \right )^\top {\bm 1}_{n_k} \; \text{and} \; Q_{\widetilde{\bm \mu}_k} = n_k \mathbf{W} \bbE ( \widetilde{D}_{\epsilon k}^{-1} ) + \bbE Q_{NS; \widetilde{\tau}_{k}, \widetilde{\boldsymbol{\nu}}_k }, \\% \text{ for } k=0,1;\\
&m_{ \widetilde{\bm \mu}_\emptyset} = Q_{\widetilde{\bm \mu}_\emptyset}^{-1} \bbE ( \widetilde{D}_{\epsilon \emptyset}^{-1} ) ({\bm I} - {\bm W}) \left ( {\bm X} - m_{\bm Z} \right )^\top {\bm 1}_n\; \text{and} \; Q_{\widetilde{\bm \mu}_\emptyset} = n (\mathbf{I} - \mathbf{W} ) \bbE ( \widetilde{D}_{\epsilon \emptyset}^{-1} ) + \bbE Q_{NS; \widetilde{\tau}_{\emptyset}, \widetilde{\boldsymbol{\nu}}_\emptyset };
\end{align*}
with
$n_k$ denoting the number of training observations in group $k = 0, 1$ and ${\bm 1}_n$ denotes a column-vector of size $n$ with all entries equal to one. Here, ${\bm w} = \bbE ({\bm \gamma})$, ${\bm W} = \text{dg} ( {\bm w})$, and $m_{{\bm Z}} = \bbE ( {\bm Z})$ with ${\bm Z} = [{\bm z}_1, \ldots, {\bm z}_n]^\top$, ${\bm X} = [{\bm x}_1, \ldots, {\bm x}_n]^\top$, and ${\bm X}^{(k)}$ and $m_{{\bm Z}}^{(k)}$ denoting the rows of ${\bm X}$ and $m_{{\bm Z}}$, respectively, corresponding to training observations from class $k$.
The product component for the magnitude $\widetilde{\tau}_{k}$ is:
\begin{equation*}
q(\widetilde{\tau}_{k}) = \text{InvGa} (a_{\widetilde{\tau}_k}, b_{\widetilde{\tau}_k} ),
\end{equation*}
where $a_{\widetilde{\tau}_k} = A_{\widetilde{\tau}} + T/2$ and $ b_{\widetilde{\tau}_k} = B_{\widetilde{\tau}} + \text{\tr}[ \bbE( \widetilde{\bm \mu}_k \widetilde{\bm \mu}_k^\top) \bbE C_{NS; \widetilde{\boldsymbol{\nu}}_k }]/2$, and $\tr$ denotes the trace operator. Here, $C_{NS; \widetilde{\boldsymbol{\nu}}_k } = \widetilde{\tau}_{k}Q_{NS; \widetilde{\tau}_{k}, \widetilde{\boldsymbol{\nu}}_k }$ represents the non-stationary precision matrix with unit marginal scale
(computational details in the Supplementary Material).
For the location-varying log length-scale $\widetilde{\bm \nu}_k$, non-conjugacy between the priors for $\widetilde{\bm \nu}_k$ and $\widetilde{\bm \mu}_k$ leads to an intractable CAVI update. To circumvent this intractability, we adopt SVB and specify a Gaussian form for $q(\widetilde{\bm \nu}_k)$, i.e.,
$$
q(\widetilde{\bm \nu}_k) = \text{N}(m_{\widetilde{\bm \nu}_k}, (\Omega_{\widetilde{\bm \nu}_k}\Omega_{\widetilde{\bm \nu}_k}^{\top})^{-1} ),
$$
where $m_{\widetilde{\bm \nu}_k}$ and $\Omega_{\widetilde{\bm \nu}_k}$ are chosen to maximize the evidence lower bound:
\begin{align*}
&\text{ELBO} (m_{\widetilde{\bm \nu}_k},\Omega_{\widetilde{\bm \nu}_k}) =\bbE \left [ \log p(\widetilde{\bm \mu}_k \, \vert \, \widetilde{\bm \nu}_k) + \log p(\widetilde{\bm \nu}_k) - \log q(\widetilde{\bm \nu}_k) \right ] \\
&=\tfrac{1}{2} {\bm 1}^\top \bbE( \widetilde{\bm \nu}_k) - \tfrac{1}{2} m_{1/\tau} \text{\tr}[ \bbE( \widetilde{\bm \mu}_k \widetilde{\bm \mu}_k^\top) \bbE C_{NS; \widetilde{\boldsymbol{\nu}}_k }]\\
&- \tfrac{1}{2} \bbE \left\{ (\widetilde{\bm \nu}_k - \mu_{\widetilde{\nu}} )^\top
Q_{S;\widetilde{\eta},\widetilde{\lambda}}
(\widetilde{\bm \nu}_k - \mu_{\widetilde{\nu}} ) \right \} - \bbE \log q(\widetilde{\bm \nu}_k; m_{\widetilde{\bm \nu}_k}, \Omega_{\widetilde{\bm \nu}_k}),
\end{align*}
and $m_{1/\widetilde{\tau}_k} = \bbE(1/ \widetilde{\tau}_k )$.
To reduce the computational complexity,
we specify a sparse Cholesky decomposition for the variational precision matrix. In particular, $\Omega_{\widetilde{\bm \nu}_k}$ is a 1-banded lower triangular matrix, leading to a tridiagonal precision matrix.
Note that, in addition to the computational savings, this form is further motivated as the approximate posterior precision matrix maintains a similar structure to the prior precision matrix.
Details about the steps for this SVB update are provided in the Supplementary Material.
The magnitude $\widetilde{\eta}$ and length-scale $\widetilde{\lambda}$ of the log length-scale processes are updated as MAP estimates, i.e., the maximizer of the objective
\begin{equation*}
\text{OBJ} (\widetilde{\eta},\widetilde{\lambda}) = \sum_{k \in \{0,1, \emptyset\}} \bbE \log \phi( \widetilde{\bm \nu}_k ; \mu_{\widetilde{\bm \nu}}, Q_{S; \widetilde{\eta}, \widetilde{\lambda} }^{-1} ) + \log p( \widetilde{\eta} ) + \log p(\widetilde{\lambda}).
\end{equation*}
\paragraph{Noise variance parameters.} The product component for $\widetilde{\sigma}_k^2$ is:
\begin{equation*}
q(\widetilde{\sigma}_{kj}^2) = \text{InvGa} (r_{kj}, s_{kj}).
\end{equation*}
Here, for $k = 0,1$ and $k = \emptyset$, respectively, we have
\begin{align*}
r_{kj} = A_{\epsilon} + n_kw_j/2, \quad &\text{ and } \quad
s_{kj} = B_\epsilon + \tfrac{w_j}{2} \sum_{i:y_i=k} \bbE (x_{ij} - \widetilde{\mu}_{kj} - z_{ij})^2; \\%\\
r_{\emptyset j} = A_{\epsilon} + n(1-w_j)/2, \quad &\text{ and } \quad
s_{\emptyset j} = B_{\epsilon} + \tfrac{1-w_j}{2} \sum_{i=1}^n \bbE (x_{ij} - \widetilde{\mu}_{\emptyset j} - z_{ij})^2,
\end{align*}
where,
$\bbE (x_{ij} - \widetilde{\mu}_{kj} - z_{ij})^2 = (x_{ij} - m_{\widetilde{\mu}_{kj}} - m_{z_{ij}} )^2 + (Q_{\widetilde{\bm \mu}_k}^{-1})_{jj} + (Q_{{\bm z}_i}^{-1})_{jj}$, $\bbE ( \widetilde{\mu}_{kj} ) = m_{\widetilde{\mu}_{kj}}$, $\bbV( \widetilde{\bm \mu}_{k} ) = Q_{\widetilde{\bm \mu}_k}^{-1} $, $\bbV ( {\bm z}_i ) = Q_{{\bm z}_i}^{-1}$, and the subscript $jj$ denotes the $(j,j)$-th entry of a matrix.
\paragraph{Latent process.} The product component for ${\bm z}_i$:
\begin{equation*}
q({\bm z}_i) = \text{N} (m_{{\bm z}_i}, Q_{{\bm z}_i}^{-1}),
\end{equation*}
where
\begin{align}
&Q_{{\bm z}_i} = {\bm W} \bbE ( \widetilde{D}_{\epsilon y_i}^{-1}) + ( {\bm I} - {\bm W} ) \bbE (\widetilde{D}_{\epsilon \emptyset}^{-1}) + \bbE (Q_{NS;\tau, \boldsymbol{\nu}_i}) , \label{eq:zvar}\\
&m_{{\bm z}_i} = Q^{-1}_{{\bm z}_i} \left \{ {\bm W} \bbE ( \widetilde{D}_{\epsilon y_i}^{-1}) ({\bm x}_i - m_{ \widetilde{\bm \mu}_{y_i}}) + ({\bm I} - {\bm W}) \bbE (\widetilde{D}_{\epsilon \emptyset}^{-1}) ({\bm x}_i - m_{ \widetilde{\bm \mu}_{\emptyset}}) \right \}. \label{eq:zmean}
\end{align}
\paragraph{Covariance parameters.} For the magnitude $\tau$ of the latent process,
$ q(\tau) = \text{InvGa} (r_\tau, s_\tau)
$
where, $r_\tau = A_{\tau} + nT/2$ and
\begin{align*}
s_\tau = b_\tau + \tfrac{1}{2} \left[ \sum_{i=1}^n \text{\tr}\left(\bbE[{\bm z}_{i}{\bm z}_{i}^\top] \bbE[ C_{NS; \bm \nu_i}]\right)
\right],
\end{align*}
with $C_{NS; \bm \nu_i} = \tau Q_{NS; \tau, \bm \nu_i}$ denoting the precision matrix with unit precision. For the common log length-scale vector ${\bm R}$,
we adopt a SVB update by specifying its approximate density:
\begin{equation*}
q({\bm R}) = \N( m_{{\bm R}}, (\Omega_{{\bm R}} \Omega_{{\bm R}}^{ \top})^{-1} ),
\end{equation*}
where the variational parameters $m_{{\bm R}}$ and $\Omega_{{\bm R}}$ are chosen to maximize:
\begin{align*}
\text{ELBO}(m_{\bm R}, \Omega_{{\bm R}}) = \bbE \Big [ \sum_{i=1}^n \log \phi ({\bm z}_{i}; {\bm 0}, Q_{NS;\tau, \boldsymbol{\nu}_i}^{-1} ) + \log \phi({\bm R}; {\bm 0}, Q_{S;\tau_2, \lambda}^{-1} )-\log q({\bm R}; m_{\bm R}, \Omega_{{\bm R}}) \Big ].
\end{align*}
The details of the SVB update are similar to the update for $\widetilde{\bm \nu}_k$ and thus are omitted. The observation-specific perturbations $\zeta_i$ are updated as the maximizers of the objective:
\begin{align*}
\text{OBJ}(\zeta_i) = \bbE \Big [ \log \phi ({\bm z}_{i}; {\bm 0}, Q_{NS;\tau, \boldsymbol{\nu}_i}^{-1} ) \Big ] + \log p( \zeta_i ).
\end{align*}
The magnitude $\tau_2$ and length-scale $\lambda$ of the common location-varying log length-scale ${\bm R}$ are updated as MAP estimates, i.e., maximizers of the objective
\begin{equation*}
\text{OBJ} (\tau_2,\lambda) = \bbE \log \phi( {\bm R} ; 0, Q_{S; \tau_2, \lambda}^{-1} ) + \log p( \tau_2 ) + \log p(\lambda).
\end{equation*}
\paragraph{Feature selection parameter.} For the binary indicator $\gamma_j$ at each location,
\begin{align*}
w_j &= q(\gamma_j = 1) = \expit \left [ - \frac{u_{j}}{2} - \tfrac{1}{2} {\bm 1}^\top \mathbf{g}_j - \alpha + \beta (w_{j+1} + w_{j-1}) \right ],
\end{align*}
where $u_j = \tfrac{n_1}{2} \bbE \log (\widetilde{\sigma}_{1j}^{2}) + \tfrac{n_0}{2} \bbE \log \widetilde{\sigma}_{0j}^{2}) - \tfrac{n}{2} \bbE \log (\widetilde{\sigma}_{\emptyset j}^{2})$, $\mathbf{g}_j = (g_{1j}, \ldots, g_{n j})^\top$, and
\begin{align*}
g_{ij} &= \bbE(\widetilde{\sigma}_{y_i j}^{2}) \left \{ (x_{ij} - m_{ \widetilde{\mu}_{y_i} j} - m_{z_{i} j})^2 + (Q_{ \widetilde{\bm \mu}_{y_i}}^{-1})_{jj} + (Q_{{\bm z}_i}^{-1})_{jj} \right \} \\
&
- \bbE(\widetilde{\sigma}_{\emptyset j}^{2}) \left \{ (x_{ij} - m_{ \widetilde{\mu}_{\emptyset} j} - m_{z_{i}j})^2 + (Q_{\widetilde{\bm \mu}_\emptyset}^{-1})_{jj} + (Q_{{\bm z}_i}^{-1})_{jj} \right \}.
\end{align*}
The hyperparameters $\alpha$ and $\beta$ controlling sparsity and correlation of the Ising prior are updated as the maximizer of the objective:
\begin{align*}
&\text{OBJ}(\alpha,\beta)= \bbE \left [ \log p({\bm \gamma} \, \vert \, \alpha, \beta) + \log p(\alpha) + \log p(\beta) \right ],
\end{align*}
where in the case of the linear chain Ising prior, a closed form expression is available for the partition function (see \cite{salinas2001} and details in the Supplementary Material).
\subsection{Classification}
Upon convergence of the variational parameters in the posterior inference phase, we proceed to derive a classification rule for a new process $x_{n+1} (t)$ that follows the distribution as described in equation \eqref{eqn::datadist}.
This requires the predictive distribution $p(y_{n+1}, {\bm z}_{n+1} \, \vert \, \mathcal{D}, {\bm x}_{n+1})$, where $\mathcal{D}$ denotes all observed data. To simplify computations, we make the following mean field approximation for the predictive distribution of $y_{n+1}$ and ${\bm z}_{n+1}$:
$$
p({\bm z}_{n+1}, y_{n+1} \, \vert \, \mathcal{D}, {\bm x}_{n+1}) \approx q({\bm z}_{n+1}) q(y_{n+1}).
$$
Following equation (\ref{eqn::CAVIupdate}), we adopt a CAVI update for ${\bm z}_{n+1}$, i.e.,
\begin{align*}
q({\bm z}_{n+1}) = \N(m_{{\bm z}_{n+1}}, \Sigma_{{\bm z}_{n+1}} ),
\end{align*}
where $m_{{\bm z}_{n+1}}$ and $\Sigma_{{\bm z}_{n+1}}$ are similar in form to the variational updates for ${\bm z}_{i}$ in \eqref{eq:zvar}-\eqref{eq:zmean} but also take into account the unknown class label $y_{n+1}$:
\begin{align*}
&Q_{{\bm z}_{n+1}} = {\bm W} \left \{ \xi_1 \bbE \widetilde{D}_{\epsilon 1}^{ -1} + (1-\xi_1) \bbE \widetilde{D}_{\epsilon 0}^{ -1} \right \} + ({\bm I} - {\bm W}) \bbE \widetilde{D}_{\epsilon \emptyset}^{ -1} + \bbE (Q_{NS;\tau, \boldsymbol{\nu}_{n+1}}), \\ &m_{{\bm z}_{n+1}} = Q_{{\bm z}_{n+1}}^{-1} {\bm W} \left \{ \xi_1 \bbE \widetilde{D}_{\epsilon 1}^{-1} ({\bm x}_{n+1} - m_{\widetilde{\bm \mu}_1} ) + (1-\xi_1) \bbE \widetilde{D}_{\epsilon 0}^{-1} ({\bm x}_{n+i} - m_{\widetilde{\bm \mu}_0}) \right \} \\
&\hspace{1.5cm} + Q_{{\bm z}_{n+1}}^{-1} ({\bm I} - {\bm W}) \bbE \widetilde{D}_{\epsilon \emptyset}^{ -1} ({\bm x}_{n+i} - m_{\widetilde{\bm \mu}_\emptyset}),
\end{align*}
where $\xi_1 = q(y_{n+1}=1)$. Note that $Q_{{\bm z}_{n+1}}$ depends on the estimate of the observation-specific log length-scale $v_{n+1} (t) = R(t) + \zeta_{n+1}$. Since the approximate posterior for $R$ has been computed in the posterior inference phase, we only need an estimate for the perturbation $\zeta_{n+1}$. This may be computed via MAP estimation as the maximizer of:
\begin{align*}
\text{OBJ}(\zeta_{n+1}) = \bbE \Big [ \log \phi ({\bm z}_{n+1}; {\bm 0}, Q_{NS;\tau, \boldsymbol{\nu}_{n+1}}^{-1} ) \Big ] + \log p( \zeta_{n+1} ).
\end{align*}
For the group label $y_{n+1}$, we adopt a CAVI update:
\begin{align*}
&\xi_1 = q(y_{n+1} = 1) \\
&= \expit \Bigg [ - \tfrac{1}{2} {\bm w}^\top \bbE \left \{ \log (\widetilde{\bm \sigma}_{1}^{2} ) - \log (\widetilde{\bm \sigma}_{0}^{2}) \right \} - \tfrac{1}{2} \text{QDA} ({\bm x}_{n+1}) - \tfrac{1}{2} \tr \left \{ {\bm W} \bbE(\widetilde{D}_{\epsilon 1}^{-1}) ( Q_{\widetilde{\bm \mu}_1}^{-1} + Q_{{\bm z}_{n+1}}^{-1} ) \right \} \\
&+ \tfrac{1}{2} \tr \left \{ {\bm W} \bbE(\widetilde{D}_{\epsilon 0}^{-1}) ( Q_{\widetilde{\bm \mu}_0}^{-1} + Q_{{\bm z}_{n+1}}^{-1} ) \right \} + \log \left ( \frac{n_1}{n_0} \right ) \Bigg ],
\end{align*}
where $\log (\widetilde{\bm \sigma}_{k}^{2} )$ denotes the element-wise log of the vector $\widetilde{\bm \sigma}_{k}^{2}$ and
\begin{align*}
\text{QDA} ({\bm x}_{n+1}) &= ({\bm x}_{n+1} - m_{\widetilde{\bm \mu}_1} - m_{{\bm z}_{n+1}})^\top {\bm W} \bbE( \widetilde{D}_{\epsilon 1}^{ -1} ) ({\bm x}_{n+1} - m_{\widetilde{\bm \mu}_1} - m_{{\bm z}_{n+1}}) \\
&\hspace{0.5cm} - ({\bm x}_{n+1} - m_{\widetilde{\bm \mu}_0} - m_{{\bm z}_{n+1}})^\top {\bm W} \bbE( \widetilde{D}_{\epsilon 0}^{ -1} ) ({\bm x}_{n+1} - m_{\widetilde{\bm \mu}_0} - m_{{\bm z}_{n+1}}).
\end{align*}
The posterior inference and classification phases of the algorithm are terminated when the change in values of the variational parameters and MAP estimates are sufficiently small. Pseudocodes for the posterior inference and classification algorithms are provided in Algorithms S1 and S2 in the Supplementary Material.
\section{Improving Scalability and Efficiency}
\label{improveEfficiency}
A na{\"i}ve implementation of the inference algorithm described in the previous section would be impractical as it involves computationally costly operations with large matrices. For example,
the variational parameters for the latent processes ${\bm z}_i$
require solving systems of $T$ linear equations for the evaluation of $m_{{\bm z}_i}$ and computing the diagonals of $Q_{{\bm z}_i}^{-1}$, with a total cost of $O(nT^3)$ for both operations. Moreover, if a full variational precision matrix is specified for the location-varying length-scales, the SVB updates may be slow to converge as the variational posterior would have $O(T^2)$ variational parameters . In the motivating SARS-CoV-2 example with $T=25,001$, these steps would clearly be infeasible.
In this section, we briefly describe the computational shortcuts that we have adopted to reduce the computational complexity of our entire inference algorithm from $O(nT^3)$ to $O(nT)$, with further details in the Supplementary Material.
In particular, a careful inspection of all steps in the variational algorithm reveals that we can avoid the costly $O(nT^3)$ operations and only require two types of operations which admit computationally efficient implementations: (1) solve a system of linear equations of the form $\mathbf{Q} \mathbf{a} = \mathbf{b}$, where $\mathbf{Q}$ is a $T \times T$ tridiagonal matrix and $\mathbf{a}, \mathbf{b} \in \mathbb{R}^T$; (2) computing the main and first off-diagonal entries of the inverse tridiagonal precision matrix. This can be attributed to both the tridiagonal structure of the precision matrices obtained through the specified kernel and SDE discretization, detailed in Section \ref{sec:finite}, and the sparse Cholesky decomposition of the variational precision matrix for the location-varying log length-scales, specified in Section \ref{sec:vb}. Additionally, the banded form of the Cholesky decomposition reduces the number of variational parameters in the SVB updates to $O(T)$.
To compute efficiently the main diagonal and first off-diagonal entries, we adopt the sparse inverse subset algorithm \citep{Takahashi1973, durrande2019}. The algorithm begins by computing the inverse of the 1-banded Cholesky decomposition of the precision matrix, which also turns out to be a lower triangular 1-banded matrix, and then utilizes a recursive algorithm for computing the required entries. For solving the required system of linear equations, we utilize Thomas' algorithm \citep{Higham2002} that exploits efficiently the tridiagonal structure of $\mathbf{Q}$ and results in a low computational complexity $O(T)$.
Lastly, we note that for applications with smoother functional realizations, Mat\'ern kernels with a larger smoothness parameter can be used. This also leads to sparse banded precision matrices but with a higher bandwidth $b$ and increased computational complexity of $O(b^2T)$.
\section{Numerical results}\label{sec:results}
In this section, we study the performance of our proposed Gaussian process DA (\texttt{GPDA}) in four simulation settings and two publicly available proteomics datasets. We compare with seven other methods - variational nonparametric DA \citep[\texttt{VNPDA,}][]{Yu2020CSDA}, penalized linear DA with fused lasso penalty \citep[\texttt{penLDA-FL},][]{Witten2011}, random forest, sparse linear DA \citep[\texttt{SparseLDA},][]{Clemmensen2011}, variational linear DA \citep[\texttt{VLDA},][]{Yu2020SC}, and both the $L_2$-regularized and $L_1$-regularized support vector machine (\texttt{SVM}) with linear kernels \citep{Cortes1995, Fan2008Liblinear}. For the proteomics datasets, we also compare with the traditional two-stage algorithm which involves peak detection and followed by classification using linear DA and quadratic DA. These classifiers are selected as competing methods as their implementations are publicly available, have reasonably low computation time for high-dimensional datasets. Moreover, they perform variable selection except the $L_2$-regularized support vector machine.
\subsection{Simulated datasets}
In all simulations, we set $T=5000$. Each simulation setting is repeated 50 times. At each repetition, we draw a training dataset of size $n_{train}=100$ and a testing dataset of size $n_{test}=500$ from the distribution:
$$
{\bm x}_i \; \vert \; y_i, {\bm \theta}^\star \sim \left\{ \begin{array}{ll}
\N_T ( \bm \mu_1^\star , \boldsymbol{\Sigma}_{i1}^\star ), & \mbox{ if $y_i = 1$; and} \\ [1ex]
\N_T ( \bm \mu_0^\star , \boldsymbol{\Sigma}_{i0}^\star ), & \mbox{ if $y_i = 0$},
\end{array} \right.
$$
where the superscript $\star$ denotes the simulation setting.
The class labels are generated from: $y_i \sim \text{Bernoulli} (0.5)$.
Full details on the simulation settings for the mean function and noise variances are provided in the Supplementary Material.
\paragraph{Description.} In {\textbf{Simulation 1}}, we study the performance of the methods when a large proportion (40\%) of the locations have weak predictive power, whereas the rest of the locations do not have any predictive power. The \texttt{GPDA} model is correctly specified, i.e., the covariance function of the $i$-th observation is
$
\boldsymbol{\Sigma}_{ik}^\star = D_{\epsilon, k}^\star + Q_{NS; \tau^\star, \bm \nu_i^\star}^{-1}, \;\; \text{and} \;\; {\bm R}^\star \sim \text{N}({\bm 0}, Q_{S; \tau_2^\star, \lambda^\star}^{-1}),
$
and we fix $\tau^\star = 4.5$, $\tau_2^\star = 2$, $\lambda^\star = 500$, and $\zeta_i^\star = 0.5\exp\{i^{0.05}\} - 1.5$.
For {\textbf{Simulation 2}}, we consider a similar scenario to Simulation 1 but with a much smaller proportion (5\%) of the locations having strong predictive power, whereas the rest of the locations do not have any predictive power. The \texttt{GPDA} model is again correctly specified. We fix $\tau^\star = 1.5$, $\tau_2^\star = 2$, $\lambda^\star = 500$, and $\zeta_i^\star = 0.5\exp\{i^{0.05}\} - 1.5$.
For {\textbf{Simulation 3}}, we assess the performance of the methods when the locations are mutually independent, the noise variances are equal between groups, and a small proportion (10\%) of the locations are weak signals, i.e. $\texttt{VLDA}$ is correctly specified. This is a boundary case whereby the true log length scale $\bm \nu_i \rightarrow - \infty$. Lastly, {\textbf{Simulation 4}} allows us to assess the performance of the methods when the $\text{GPDA}$ model is misspecified. In particular, the true covariance matrix has a uniform structure with all diagonal entries equal $1$ and the off-diagonal entries equal $0.95$. A small proportion (10\%) of the locations have strong predictive power.
\begin{figure}[!t]
\centering
{\includegraphics[width=0.8\textwidth]{NonStatSimulationClassification.pdf}}
\caption{Classification errors rates (\%) for Simulations 1 to 4.}
\label{fig:SimResultClassErr}
\end{figure}
\begin{figure}[!t]
\centering
{\includegraphics[width=0.8\textwidth]{NonStatSimulationMCC.pdf}}
\caption{$MCC \times 100\%$ for Simulations 1 to 4.}
\label{fig:SimResultMCC}
\end{figure}
\paragraph{Results.}
We assess the classification performance of the methods using the classification error rate, true positive rate and true negative rate, and the variable selection performance using the Matthews correlation coefficient. Results are provided in Figures \ref{fig:SimResultClassErr} and \ref{fig:SimResultMCC}, with further details in the Supplementary Material. \texttt{GPDA} achieved good classification and variable selection in comparison to the alternative methods. In Simulation 1, \texttt{GPDA}, \texttt{SVM-L1}, and \texttt{SparseLDA} attained comparably high classification accuracies. Moreover, \texttt{GPDA} attained the second highest mean MCC, whereas \texttt{SVM-L1} and \texttt{SparseLDA} did not perform well for variable selection. \texttt{penLDA-FL} outperformed \texttt{GPDA} in feature selection in this simulation setting as it is well-known to perform well as a feature selector when the signal strengths are weak. In Simulation 2, \texttt{GPDA} and \texttt{SVM-L1} attained comparably low classification errors, while \texttt{GPDA} yielded the highest mean MCC. In Simulation 3, \texttt{GPDA} achieved the second lowest mean classification error rate. This demonstrates its ability to perform well even in the case when a simpler model fits the data well. Moreover, it outperforms \texttt{VLDA} in feature selection as the Ising prior works well when the true discriminative process ${\bm \gamma}^\star$ is smooth. In Simulation 4, \texttt{GPDA}, \texttt{SVM-L1}, \texttt{SVM-L2}, and \texttt{SparseLDA} achieved comparably high classification accuracies, while \texttt{GPDA} achieved the highest mean MCC. This demonstrates our proposed method's robustness when the true precision matrix is not sparse.
\subsection{Proteomics datasets}
We consider two datasets that aim to predict and identify markers of 1) SARS-CoV-2 in nasal swabs using matrix-assisted laser desorption/ionization (MALDI) MS \citep{nachtigall2020detection} 2) breast cancer in plasma using surface-enhanced laser desorption and ionization (SELDI) protein MS \citep{shi2006declining}. We assess the performance via five-fold cross validation (CV) classification accuracies. Moreover, for methods that perform variable selection, we compute the variable selection rate at each location. To investigate the utility of a high-dimensional approach versus the traditional approach in bioinformatics, we include a comparison of the competing methods with \texttt{LDA-traditional} and \texttt{QDA-traditional} as benchmark methods. These methods are similar to the traditional approach in that they first employ a peak detection method to identify peak locations, and followed by fitting a low-dimensional classification model using the identified peak locations.
\paragraph{Data description.} 1) COVID-19 has shaken up the world; in only a year and half from its first appearance, approximately 190 million people have been infected, over four million have died, and over half of the world population has experienced some form of lockdown\footnote{John Hopkins Coronavirus Resource Center \url{https://coronavirus.jhu.edu/}}. To improve testing capacity in countries that lack resources to handle large-scale PCR testing, the SARS-CoV-2 dataset was collected using equipment and expertise commonly found in clinical laboratories in developing countries. The dataset contains samples from 362 individuals, of which 211 were SARS-CoV-2 positive and 151 were negative by PCR testing.
The processed spectra contain $T= 25,001$ variables and are depicted in Figure \ref{fig:covid_spectra}. 2) Breast cancer is a common and deadly disease, and improvements in early detection and screening are needed for improved treatment and survival. Towards this goal, this dataset was collected to investigate and identify markers from plasma that discriminate between controls and breast cancer patients. The processed spectra contain $T=10,451$ variables. Due to heterogeneity in breast cancers, in the following, we focus on discriminating between healthy controls and HER2 (with $n=119$).
More details on the raw data and required processing steps for both examples are in the Supplementary Material.
\begin{figure}[!t]
\centering
{\includegraphics[width=1.0\textwidth]{ClassAccCombine.pdf}}
\caption{Classification accuracy (\%) for SARS-CoV-2 (left) and breast cancer (right).}
\label{fig:CombineClassAcc}
\end{figure}
\begin{figure}[!t]
\centering
{\includegraphics[width=1.0\textwidth]{VariableSelectionRateCombine.pdf}}
\caption{Variable selection rate for SARS-CoV-2 (left) and breast cancer (right). }
\label{fig:CombineVarbSelect}
\end{figure}
\paragraph{Results.} Classification accuracies of all methods for breast cancer and SARS-CoV-2 are presented in Figure \ref{fig:CombineClassAcc}. Other supporting plots can be found in the Supplementary Material. In both datasets, \texttt{GPDA} attained amongst the highest classification accuracy, with high true positive and true negative rates (refer to the Supplementary Material). \texttt{GPDA}'s good classification performance can be attributed to its ability to account for the highly non-stationary correlation structure that is evident from the location-varying roughness in the observed spectra in Figures \ref{fig:covid_spectra} and S6 (Supplementary Material). Moreover, the posterior expected log length-scales (Figures S8 and S10 in the Supplementary Material) exhibit an overall increasing trend that is congruent with the spectra being flatter at higher values of $m/z$. We also observe that methods which unify variable selection and classification in a single framework generally performed better than the traditional peak detection methods, suggesting that the two-stage peak-detection approach leads to a substantial loss of information for both datasets. Figure \ref{fig:CombineVarbSelect} summarizes the variable selection frequencies. Note that although both \texttt{GPDA} and \texttt{SVM-L1} performed well in terms of classification accuracy in the SARS-CoV-2 dataset, \texttt{GPDA} has identified many more locations as discriminative locations than \texttt{SVM-L1}. This disparity may be attributed to a subtle difference in the variable selection component of both methods. Specifically, the variable selection component in \texttt{GPDA} identifies all discriminative locations, whereas the $L_1$ penalization in \texttt{SVM-L1} identifies the optimal set of locations that minimizes classification error. We see a different trend for the breast cancer data where the classification accuracy of \texttt{SVM-L1} is clearly lower. For both datasets, \texttt{SVM-L2} attains mean accuracy comparable to \texttt{GPDA}. However, unlike \texttt{SVM-L2}, \texttt{GPDA} identifies discriminative locations and hence may be used to identify differences in protein compositions. This information may be useful for developing new diagnostic tools and an effective anti-viral treatment. The other competing methods considered showed less comparable performance.
\section{Conclusion}
\label{sec:conc}
In this paper, we developed a comprehensive and unified framework for classification and variable selection with high-dimensional functional data.
To account for the non-stationary and rough nature of the realized functions, we introduced a two-level non-stationary GP model with carefully chosen kernel structures, and in combination with the DA framework. Moreover, this is coupled with an Ising prior to allow for smoothness in the variable selection component. The model poses serious computational challenges due to its complexity and high dependence between parameters, hierarchical layers, and latent variables. To deal with these challenges, we proposed an inference scheme that exploits a number of advances in GPs and variational methods, as well as computational tricks, to improve the scalability and
efficiency of our entire inference algorithm from $O(nT^3)$ to $O(nT)$. The performance of our approach in comparison to competing methods is demonstrated in several simulated and real MS datasets. Results indicate that our method performs consistently as the best or second best in all scenarios. In addition, for the proteomics data, we illustrated how combining the steps of peak detection, feature extraction, and classification into a unified modeling framework, that accounts for the complicated dependence structure in both the inputs and variable selection, outperforms the traditional two-stage approaches commonly used in practice. We focused in this work on one-dimensional functional data, however future work will explore extensions for multivariate
structured functional inputs, e.g. images. Other choices of kernels and deeper architectures can also be implemented to account for smoother and/or more complex structures arising in other applications.
\section*{Supplementary material}
All supplementary content and codes for this article may be downloaded from the repository: \url{https://github.com/weichangyu10/GPDAPublic}.
\bibliographystyle{apalike}
|
\section{Introduction}
\label{sec:Introduction}
\input{1_Introduction}
\section{Related Work}
\label{sec:Relatedwork}
\input{2_Relatedwork}
\section{Proposed Method}
\label{sec:Methodology}
\input{3_0_Methodology}
\section{Experiments}
\label{sec:Experiments}
\input{4_0_Experiments}
\section{Conclusion}
\label{sec:Conclusion}
\input{5_Conclusion}
\section*{Acknowledgment}
This work was supported in part by the National Natural Science Foundation of China under Grant 61602197, and in part by the Cognitive Computing and Intelligent Information Processing (CCIIP) Laboratory, Huazhong University of Science and Technology, Wuhan, China.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
\section{Introduction}
Recent advances in deep learning have resulted in great progress in both the computer vision and natural language processing communities. These achievements make it possible to connect vision and language, and facilitate multi-modal learning tasks such as image-text matching, visual question answering, visual grounding and image captioning. Image captioning aims to automatically describe an image’s content using a natural language sentence. The task is challenging since it requires one to recognize key objects in an image, and to understand their relationships with each other. Inspired by machine translation, encoder-decoder architecture became the most successful approach for image caption task. Encoder encodes the visual(object region or grid visual feature) or semantic(attribute words) information into feature vectors. Given the input features, decoder iteratively generates the output caption words. Based on this architecture, various improvements have been applied into encoder, decoder and training method to further improve the performance.
Previously, there are three kinds of method to improve the model.
The first one is based on attention mechanism. To establish the fine-grained connections of caption words and their related image regions, attention can be seamlessly inserted into the framework. Then, a lot of works focus on improving attention measurement to enhance the interaction between visual content and natural sentence to boost image captioning. The second one treat image captioning as a multi-modal problem, by gradually fusing visual and semantic information or exploiting semantic and visual information simultaneously, they can identify the equivalent visual signals especially when predicting highly abstract words. The last one is aim to address the exposure bias of generated captions by using the cross-entropy loss, reinforcement learning (RL)-based algorithm~\cite{E_1, E_2,E_3} are designed to directly optimize the non-differentiable evaluation metrics (e.g., BLEU, METEOR, ROUGE, CIDEr and SPICE~\cite{F_1, F_2, F_3, F_4, F_5}).
Despite those reinforcement learning based algorithms which are compatible to different models, there existing two main limitations in both attention focused mechanism and multi-modal focused mechanism:
1) Even with visual sentinel mechanism, typical attention focused mechanisms are still incapable of inter-modal interactions, thus they are arduous to identify the equivalent visual signals especially when predicting highly abstract words.
2) Multi-modal focused models typically fuse visual branch(object region or grid visual feature) and semantic branch(attribute words) with concatenation or summation, these method are usually shallow and may fail to fully understand the complex relationships among information from two different modalities. Besides, even attention mechanism are used in each branch, since these primitive semantic features have no geometry information as guidance, noise from different modal will deteriorate the attention measurement.
To address the limitation in attention focused models, we extend the Transformer model for machine translation to a Visual-Semantic Transformer model for image captioning. Different from previous Transformer based captioning models, there are four parallel branches in the encoder (VV:Pure Visual, VS:Semantic fused to Visual, SV:Visual fused to Semantic, SS:Pure Semantic). Different branch stands for different ratios of visual information to semantic information. By dynamically adjust each branch's contribution at different time step, model can not only maintain sophisticated attention measurement in each branch, but can also generate words with abstract semantic meaning.
To address the limitation in multi-modal focused models, rather than using MIL model to extract attribute words, we use Dense-Caption model to extract dense captions which describe sub-images and each dense caption has a bounding box. By properly taking both content and geometry information into a fusion module, model can not only encode this complex multi-modal relationship to a deeper degree, but can also generate inter geometry features between two different modals. Then, inter and intra geometry information will contribute to a better attention measurement in each self-attention encoder branch. Thus we call the model as Geometry-Entangled Visual-Semantic Transformer(GEVST).
To summarize, the main contributions of this study are three-fold:
\begin{itemize}
\item Inter- and intra-geometry and content relation in both fusion and self-attention is first proposed in the GEVST model. Our model is not only capable of fusing two modality information based on both content and geometry information, by feeding both inter and intra geometry features into self-attention module, model can also get a better attention measurement which improves image captioning performance.
\item To dynamically give a more fine-grained ratio and higher order interaction between visual and semantic, we perform a cross-attention with 4 encoding branches, these branches' outputs with multi-VS ratio are then summed together after modulation.
\item Extensive experiments on the benchmark \textsf{MSCOCO}~\cite{G_1, G_2} image captioning dataset are conducted to quantitatively and qualitatively prove the effectiveness of the proposed models. The experimental results show that the GEVST significantly outperforms previous state-of-the-art approaches with a single model.
\end{itemize}
The rest of the paper is organized as follows:
In section II, we review the related work of image captioning approaches, especially those attention focused and multi-modal focused mechanisms.
In section III, we introduce our visual-semantic fusion module which take both content and geometry information into consideration.
In section IV, we introduce our Geomerty-Entangled self-attention module, and will also elaborate the architecture of four parallel branches in our GEVST model.
In section V, we introduce our extensive experimental results for evaluation and use the benchmark MSCOCO image captioning dataset to evaluate our proposed approaches.
Finally, we conclude this work in section VI.
\section{Related Work}
In this section, we briefly review the most relevant research on image captioning, especially those attention focused and multi-modal focused mechanisms.
\subsection{Image Captioning}
Previously, the research on image captioning can be mainly divided into three categories: template-based approaches~\cite{A_1, A_2}, retrieval-based approaches~\cite{A_3, A_4, A_5}, and generation-based approaches~\cite{A_6, A_7, A_8, A_10, A_11}.
The template-based approaches address the task using a two-stage strategy, it first detects the objects, their attributes, and object relationships in an image, and then fills the extracted information in a pre-designed sentence template. Both Conditional Random Field (CRF) and Hidden Markov Model (HMM)~\cite{A_12, A_13} are used to predict labels based on the detected objects, attributes, and prepositions, and then generate caption sentences with a template. Obviously, those captions generated by the template based approaches are highly depend on the quality of the templates and usually follow the syntactical structures. However, the diversity of the generated captions is severely restricted, as a result, the generated captions are often rigid, and lack naturalness.
To ease the diversity problem, is to directly retrieve the target sentences for the query image from a large-scale caption database with respect to their cross-modal similarities in a multi-modal embedding space. The core idea~\cite{A_3} is to propose an approach to match the image-caption pairs based on the alignment of visual segments and caption segments. For prediction, the cross-modal matching over the whole caption database is performed to generate the caption for one image. Further improvements~\cite{A_4, A_5} are focusing on different metrics or loss functions to learn the cross-modal matching model. However, the result captions are fixed, and may fail to describe the object combinations in a new image. To increase caption's diversity, caption database has to be as large as possible. Whereas, for large database, there is a big problem in retrieval efficiency. Thus, the diversity problem has not been completely resolved.
Different from template-based and retrieval-based models, in recent years, With the successes of \emph{sequence generation} in machine translation~\cite{A_14, A_15}, generation-based model has emerged as a promising solution to image captioning. In this kind of architecture, model tries to build an end-to-end system with encoder-decoder framework, thus, it can generate novel captions with more flexible syntactical structures. Vinyals et al. propose an encoder(CNN Based)-decoder(LSTM Based) architecture as its backbones. After that, this kind of encoder-decoder architecture became the mainstream for generation-based approaches. Similar architectures are also proposed by Donahue et al.~\cite{A_16} and Karpathy et al.~\cite{A_11}. Due to the flexibility and excellent performance, most of subsequent improvements are based on the generation-based models.
\subsection{Attention Focused Model}
Within the encoder-decoder framework, one of the most important improvements for generation-based models is the attention mechanism. Xu et al.~\cite{A_6} proposed soft/hard attention mechanism to automatically focus
on the salient regions when generating corresponding words. The attention model is a pluggable module that can be seamlessly inserted into previous approaches to remarkably improve the caption quality.
The attention model is further improved in~\cite{B_1, B_2, A_8, B_3, B_4, B_6, B_7, B_8, B_10}.
Lu et al. presented an adaptive attention encoder-decoder model for automatically deciding when to rely on visual or language signals.
Chen et al. proposed a spatial- and channel-wise attention model to attend to visual features.
Anderson et al. introduced a bottom-up module, that uses a pre-trained object detector to extract region-based image features, and a top-down module that utilizes soft attention to dynamically attend to these object
Gu et al. adopted multi-stage learning for coarse-to-fine image captioning by stacking three layers of LSTMs, where each layer of LSTM was different from other layers of LSTMs.
Zhou et al. designed a novel salience-enhanced re-captioning framework via two-phase learning is developed for the optimization of image caption generation
Huang et al. proposed attention on attention module enhances visual attention by further measuring the relevance between the attention result and the query.
Wang et al. proposed a hierarchical attention network to combine text, grids, and regions with a relation module to exploit the inherent relationship among diverse features.
Herdade et al. utilized bounding boxes of regions to model location relationships between regions in a relative manner.
Cornia et al. presented a mesh-like structure transformer to exploit both low-level and high-level contributions.
Through all of these works, it is confirmed that improving attention measurement is an effective way to enhance the interaction between visual content and natural sentence and thus boost image captioning.
However, typical attention mechanisms are arduous to identify the equivalent visual signals especially when predicting highly abstract words. This phenomenon is known as the semantic gap between vision and language.
\subsection{Multi-Modal Focused Model}
This problem in those attention focused model can be overcome by providing semantic attributes that are homologous to language~\cite{C_1, A_7, C_2, C_3, C_5, C_6, C_7}.
Hao et al. use multiple instance learning to train to detect words(nouns, verbs, and adjectives) that commonly occur in captions.
Based on the these detected words, You et al. leveraged the high-level semantic information directly with semantic attention.
As it was justified that the combination of the two complementary attention paradigms can alleviate the harmful impacts of the semantic gap. Therefor, Li et al. proposed a two-layered LSTM that the visual and semantic
attentions are separately conducted at each layer.
Yu et al. designed a multi-modal transformer model which simultaneously captures intra- and inter-modal interactions in a unified attention block.
Huang et al. also introduced two modules to couple attribute detection with image captioning as well as prompt attributes by predicting appropriate attributes at each time step.
Another way to bridge the modality gap is to employ graph convolutional neural networks(GCN).
Yao et al. built a GCN model to explore the spatial and semantic relationships. Then, by late fusion, they combined two LSTM language models that are independently trained on different modalities.
Similarly, Xu et al. used a multi-modal graph convolution network to modulate scene graphs into visual representations.
Whereas, the interactions between different modals are usually shallow and may fail to fully understand the complex relationships among these two modal information. Besides, even attention mechanism are used in each branch, since their semantic features have no geometry information as guidance.
\subsection{Caption Related Task}
Dense-caption task is also a task aimed at caption generation. Dense-caption performs intensive captions\cite{D_1, D_2, D_3} on the image. Beside giving caption which describes a certain sub-image, it can also give the geometry information of this caption. And since the output is a complete sentence rather than attribute words, it can provide more advanced semantic and linguistic information. Like image-caption, Visual Question Answering(VQA) also aim at handling the modal-crossing problem. In VQA, although attention mechanism helps model to focus on the visual content relevant to the question, it's still insufficient to model complex reasoning features. Rather than simple attention mechanism, multi-modal fusion plays the key roles in better performance. In VQA, there are many visual semantic fusion methods\cite{D_4, D_5, D_6, D_7} are proved to be more powerful than just concatenation or element-wise addition.
\subsection{The Transformer Model}
\label{sec:The_Transformer_Model}
The Transformer model was first proposed for machine translation, and has been successfully applied to many natural language processing tasks.
All intra- and inter-modality interactions between word and image-level features are modeled via scaled dot-product attention, without using recurrence.
Attention operates on three sets of vectors, namely a set of queries ${Q}$, keys ${K}$ and values ${V}$,
and takes a weighted sum of value vectors according to a similarity distribution between query and key vectors.
In the case of scaled dot-product attention, the operator can be formally defined as below,
\begin{equation}
\label{eq:scaled dot-product attention}
{F} = Attention({Q}, {K}, {V}) = softmax(\frac{{Q}{K}^T}{\sqrt{d}}){V},
\end{equation}
where ${F}$ is a matrix of $n_q$ attended query vectors, ${Q}$ is a matrix of $n_q$ query vectors, ${K}$ and ${V}$ both contain $n_k$ keys and values, all with the same dimension, and $d$ is a scaling factor.
Instead of performing a single attention function for the queries,
multi-head attention is to allow the model to attend to diverse information from different representation sub-spaces.
The multi-head attention contains $h$ parallel ‘heads’ with each head corresponding to an independent scaled dot-product attention function. The attended features ${F}$ of the multi-head attention functions is given as follows:
\begin{equation}
\begin{aligned}
\label{eq:multi-head-attention}
& {F} = MultiHead({Q}, {K}, {V}) = Concat({H_1},\cdots,{H_h})W^O \\
& {H_i} = Attention({Q}{W^Q_i}, {K}{W^K_i}, {V}{W^V_i})
\end{aligned}
\end{equation}
where ${W^Q_i}, {W^K_i}, {W^V_i} \in \mathbb{R}^{{d}\times\frac{d}{h}}$, are the independent head projection matrices.
${W^O_i} \in \mathbb{R}^{{d}\times{d}}$ denotes the linear transformation.
Note that the bias terms in linear layers are omitted for the sake of concise expression,
and the subsequent descriptions follow the same principle.
The Transformer is a deep end-to-end architecture that stacks attention blocks to form an encoder-decoder strategy.
Both the encoder and the decoder consist of $N$ attention block,
and each attention block contains the multi-head attention modules.
The multi-head attention module learns the attended features that consider the pairwise interactions between two input features,
In the encoder, each attention block is self-attentional such that the queries, keys and values refer to the same input features.
In the decoder contains a self-attention layer and a guided attention layer.
It first models the self-attention of given input features and then takes the output features of the last encoder attention block to guide the attention learning.
\subsection{Data Preparation}
\label{sec:Data_Preparation}
\input{3_1_Data_Preparation}
\subsection{Visual-Semantic Fusion}
\label{sec:Visual_Semantic_Fusion}
\input{3_2_Visual_Semantic_Fusion}
\subsection{Geometry-Entangled Self-Attention}
\label{sec:Geometry_Entangled_Attention}
\input{3_3_Geometry_Entagled_Attention}
\subsection{Multi-Branches Encoder-Decoder}
\label{sec:Multi_Branches_Encoder_Decoder}
\input{3_4_Multi_Branches_Encoder_Decoder}
\subsection{Datasets}
\label{sec:Data_Preparation}
\input{4_1_Datasets}
\subsection{Implementation Details}
\label{sec:Data_Preparation}
\input{4_2_Implementation_Details}
\subsection{Comparison with the State-of-the-Art}
\label{sec:Comparison_with_the_State-of-the-Art}
\input{4_3_Comparison_with_the_State-of-the-Art}
\subsection{Qualitative Analysis}
\label{sec:Qualitative_Analysis}
\input{4_5_Qualitative_Analysis}
\subsection{Ablation Studies}
\label{sec:Ablation_Studies}
\input{4_4_Ablation_Studies}
|
\section{Introduction \label{sec:introduction}}
Spontaneous time-reversal symmetry breaking has long been attracted much attention,
as it leads to intriguing physical phenomena, such as the anomalous Hall effect and the magneto-optical Kerr effect.
Modern understanding of these phenomena has been achieved based on the Berry phase mechanism~\cite{Nagaosa_review,xiao2010berry}.
Although such phenomena were originally studied in the ferromagnetic state,
it has recently been recognized that similar phenomena can occur in a certain class of antiferromagnetic (AFM) states without the uniform magnetization~\cite{smejkal2021anomalous}.
For example, the collinear AFM ordering with the mirror symmetry breaking as the uniform magnetization, results in the anomalous Hall effect~\cite{PhysRevB.55.8060, li2019quantum, PhysRevB.102.075112, PhysRevB.103.L180407}.
Thus, the AFM materials can also exhibit the same physical properties as ordinary ferromagnetic ones, which is advantageous for functional materials without leakage of a magnetic field.
The AFM state also exhibits multiferroic phenomena when both spatial inversion $(\mathcal{P})$ and time-reversal $(\mathcal{T})$ symmetries are broken simultaneously while their product $(\mathcal{PT})$ symmetry is preserved.
The typical example is the linear magnetoelectric effect in the AFM insulators, e.g., Cr$_{2}$O$_{3}$~\cite{popov1999magnetic}, Ga$_{2-x}$Fe$_x$O$_3$~\cite{popov1998magnetoelectric, doi:10.1143/JPSJ.74.1419}, LiCoPO$_4$~\cite{van2007observation, zimmermann2014ferroic}, and Ba$_2$CoGe$_2$O$_7$~\cite{PhysRevB.84.094421}, and in the AFM metals, e.g., UNi$_{4}$B~\cite{hayami2014toroidal, hayami2015toroidal, saito2018evidence} and Ce$_3$TiBi$_5$~\cite{doi:10.7566/JPSCP.30.011189, doi:10.7566/JPSJ.89.033703}.
Moreover, the nonreciprocal optical and transport properties have been studied~\cite{tokura2018nonreciprocal, ma2021topology, oyamadaa2018anomalous, PhysRevResearch.2.043081, PhysRevX.11.011001}.
Among them, multiferroic phenomena {\it within} the linear response theory have been understood by regarding the fact that the AFM states accompany the uniform orderings of the electronic odd-parity magnetic-type multipoles~\cite{Yanase_JPSJ.83.014703,hayami2014toroidal,hayami2014spontaneous,hayami2016emergent,
watanabe2017magnetic,
yanagi2018manipulating, yanagi2018theory, PhysRevB.98.020407,
hayami2018classification, watanabe2018group,
PhysRevB.98.060402,PhysRevB.98.020407,thole2018magnetoelectric, yatsushiro2019atomic,
thole2020concepts, jpsj_yatsushiro2020odd,PhysRevB.104.045117, PhysRevB.104.054412},
such as the magnetic toroidal (MT) dipole
~\cite{doi:10.1080/00150199408213381, PhysRevB.76.214404, spaldin2008toroidal, Kopaev_2009,
hayami2014toroidal, hayami2014spontaneous,hayami2016emergent,
PhysRevB.97.134423, PhysRevB.98.060402,PhysRevB.98.020407,yatsushiro2019atomic, jpsj_yatsushiro2020odd}.
Meanwhile, the microscopic understanding of the {\it nonlinear} transports in AFMs has not been fully achieved except for several works~\cite{PhysRevResearch.2.043081,PhysRevLett.127.277201, PhysRevLett.127.277202} and symmetry analyses~\cite{PhysRevB.104.054412, zhang2020higher}.
For example, it remains unclear which model parameters are essentially important to induce nonlinear transports and how the odd-parity magnetic-type multipoles are related to them.
To be clear this point and obtain an intuitive understanding of the nonlinear transport, it is useful to extract the essential model parameters, without which the nonlinear transport coefficients vanish, from various hopping processes, spin-orbit coupling, and order parameters in the microscopic model Hamiltonian.
Such an understanding provides a guideline to
explore new functional AFM materials with a giant nonlinear transport, and its efficient bottom-up design in combination with the {\it ab initio} calculations.
In this paper, we elucidate the microscopic essential model parameters for the second-order nonlinear conductivity in the $\mathcal{PT}$-symmetric collinear AFMs by focusing on the role of the MT moment.
By analyzing a minimal model on a two-dimensionally-stacked zigzag chain based on the nonlinear Kubo formula,
we show that the effective coupling between the MT moment and one of the antisymmetric spin-orbit interactions (ASOIs) plays an essential role in inducing the longitudinal and transverse components of the nonlinear conductivity.
Moreover, we find that the nonlinear conductivities are highly enhanced near the transition temperature in the case that the AFM molecular field is comparable to the ASOI in a multi-band system.
We also discuss the relevance between the transverse nonlinear conductivity and the linear magnetoelectric coefficient by comparing the ASOI and temperature dependences.
The organization of this paper is as follows.
In Sec.~\ref{sec:model}, we introduce a minimal model on a two-dimensionally stacked zigzag chain.
After showing the relation of an MT moment to the nonlinear conductivity and the linear magnetoelectric coefficient in Sec.~\ref{sec:formalism}, the numerical results are presented in Sec.~\ref{sec:numerical}.
In Sec.~\ref{sec:discussion}, we discuss the essential model parameters and the semi-quantitative evaluation of the nonlinear conductivity.
We summarize this paper in Sec.~\ref{sec:summary}.
In Appendix~\ref{sec:MP_express},
we present the functional forms of the odd-parity magnetic and MT multipoles.
In Appendix~\ref{sec:extract_parameter}, we show the analytic expressions for the essential model parameters in the asymmetric band modulation, nonlinear conductivities, and linear magnetoelectric coefficient.
Finally, we present the numerical result of the nonlinear transverse conductivity in the presence of the additional interlayer hopping in Appendix~\ref{sec:interlayer_hopping}.
\begin{figure}[t!]
\centering
\includegraphics[width=88mm]{Fig1.eps}
\caption{(a), (b) Schematic pictures of (a) a two-sublattice zigzag chain and (b) its stacking along the $z$ direction.
(c) The temperature ($T$) dependence of the MT moment $T_x^{\rm MF}$ at $\alpha_1=0.4$ and $\alpha_2=0.1$.
The AFM structure with the MT moment along the $x$ direction $T_x$ is shown in the inset.
(d) The energy bands measured from the chemical potential $\mu$ at $k_z=0$ for three temperatures.
\label{fig:fig1}}
\end{figure}
\section{Model \label{sec:model}}
We consider a minimal two-dimensional system where the zigzag chain along the $x$ direction [Fig.~\ref{fig:fig1}(a)] is stacked along the $z$ direction [Fig.~\ref{fig:fig1}(b)].
The tight-binding Hamiltonian is given by
\begin{align}
\label{eq:hamiltonian}
\mathcal{H}=& \mathcal{H}_{\rm hop}^{\rm AB} + \mathcal{H}_{\rm hop} + \mathcal{H}_{\rm ASOI} + \mathcal{H}_{\rm int},\\
\label{eq:hop_different}
\mathcal{H}_{\rm hop}^{\rm AB} =& \sum_{\bm k} \sum_{\sigma}
\left\{\varepsilon^{\rm AB}({\bm k}) c_{{\bm k}{\rm A}\sigma}^\dagger c_{{\bm k}{\rm B}\sigma}
+ {\rm H.c.}
\right\},\\
\label{eq:hop_same}
\mathcal{H}_{\rm hop} =& \sum_{\bm k} \sum_{\sigma}
\varepsilon({\bm k})(c_{{\bm k}{\rm A}\sigma}^\dagger c_{{\bm k}{\rm A}\sigma}+c_{{\bm k}{\rm B}\sigma}^\dagger c_{{\bm k}{\rm B}\sigma}), \\
\label{eq:ASOI}
\mathcal{H}_{\rm ASOI} =& \sum_{\bm k} \sum_{\sigma\sigma'}
{\bm g}({\bm k}) \cdot {\bm \sigma}^{\sigma\sigma'}(c_{{\bm k}{\rm A}\sigma}^\dagger c_{{\bm k}{\rm A}\sigma'}-c_{{\bm k}{\rm B}\sigma}^\dagger c_{{\bm k}{\rm B}\sigma'}), \\
\label{eq:int}
\mathcal{H}_{\rm int}=&
J_{\rm AF}
\sum_{\braket{ij}}
\hat{M}_{i{\rm A}}^{z} \hat{M}_{j{\rm B}}^{z},
\end{align}
where $c_{{\bm k} l \sigma}^\dagger$ ($c_{{\bm k} l \sigma}$) is the creation (annihilation) operator of
electrons at wave vector ${\bm k}$, sublattice $l= {\rm A}, {\rm B}$, and spin $\sigma = \uparrow, \downarrow$.
The hopping Hamiltonian $\mathcal{H}_{\rm hop}^{\rm AB}$ in Eq.~(\ref{eq:hop_different}) includes the nearest-neighbor hopping between A and B sublattices as $\varepsilon^{\rm AB}({\bm k})=-2t_1\cos (k_xa/2)$, while $\mathcal{H}_{\rm hop}$ includes the hoppings within the same sublattices along the $x$ and $z$ directions as $\varepsilon({\bm k})= -2t_2\cos{(k_xa)} -2t_3\cos{(k_zc)}$.
$\mathcal{H}_{\rm ASOI}$ in Eq.~(\ref{eq:ASOI}) represents the ASOI that arises from the relativistic spin-orbit coupling as ${\bm g}({\bm k}) = [-\alpha_2 \sin{(k_zc)}, 0, \alpha_1 \sin{(k_xa)}]$.
The ASOI in Eq.~\eqref{eq:ASOI} has the sublattice-dependent staggered form satisfying the global inversion symmetry~\cite{doi:10.1143/JPSJ.81.034702, Yanase_JPSJ.83.014703}.
$\mathcal{H}_{\rm int}$ in Eq.~\eqref{eq:int} represents the Ising-type AFM exchange interaction of the nearest-neighbor A-B bond with $J_{\rm AF} > 0$ where $\hat{M}_{i{\rm A}({\rm B})}^z = \sum_{\sigma\sigma'} c_{i{\rm A}({\rm B})\sigma}^\dagger \sigma^z_{\sigma\sigma'} c_{i{\rm A}({\rm B})\sigma'}$ is the $z$ component of the magnetic dipole operator and $c_{il\sigma}^\dagger$ and $c_{il\sigma}$ are the Fourier transforms of $c_{\bm{k}l\sigma}^\dagger$ and $c_{\bm{k}l\sigma}$, respectively.
We adopt the Hartree-type mean-field approximation as
\begin{align}
&J_{\rm AF}\sum_{\braket{ij}} \hat{M}_{i{\rm A}}^z \hat{M}_{j{\rm B}}^z \notag\\
&\to
\tilde{J}_{\rm AF}\sum_i \left(
\braket{\hat{M}_{\rm A}^z} \hat{M}_{i{\rm B}}^z + \braket{\hat{M}_{\rm B}^z} \hat{M}_{i{\rm A}}^z - \braket{\hat{M}_{\rm A}^z}\braket{\hat{M}_{\rm B}^z}\right),
\end{align}
where $\braket{\cdots}$ represents the statistical average and $\tilde{J}_{\rm AF}=2J_{\rm AF}$ is the renormalized coupling constant taking into account the two nearest-neighbor atomic sites.
We set the model parameters as $(t_1, t_2, t_3, J_{\rm AF})=(0.1, 1, 0.5, 2.5)$, electron filling as $1/5$, and the lattice constant as $a=c=1$ in the following discussion; $t_2$ is set as the energy unit.
The model in Eq.~(\ref{eq:hamiltonian}) exhibits the MT moment when the global inversion symmetry is broken under the staggered AFM ordering,
as shown in the inset of Fig.~\ref{fig:fig1}(c)~\cite{Yanase_JPSJ.83.014703,hayami2014toroidal}.
In the present system, the staggered AFM moment along the $z$ direction is equivalent to the uniform MT moment along the $x$ direction; $T_x^{\rm MF} \equiv (\braket{\hat{M}_{\rm A}^z} - \braket{\hat{M}_{\rm B}^z})/2$~\cite{hayami2015spontaneous}; see also Appendix~\ref{sec:MP_express}.
The $T$ dependence of $T_x^{\rm MF}$ at $\alpha_1=0.4$ and $\alpha_2=0.1$ is shown in Fig.~\ref{fig:fig1}(c), where
$T_x^{\rm MF}$ is self-consistently determined for the two-sublattice unit cell by taking over $200^2$ grid points in the Brillouin zone.
$T_x^{\rm MF}$ becomes nonzero below the transition temperature $T_{\rm N}$ and saturates below $T \simeq 0.2 T_{\rm N}$.
Almost the same behavior is obtained for $\alpha_1,\alpha_2 \lesssim 0.5$.
Reflecting $T_x^{\rm MF} \neq 0$, the electronic band structure is asymmetrically modulated along the $k_x$ direction, as shown in Fig.~\ref{fig:fig1}(d)~\cite{Yanase_JPSJ.83.014703,hayami2015spontaneous}.
This asymmetric band modulation is understood from the effective coupling between
$T_x^{\rm MF}$ and the ASOI $\alpha_1$ in the doubly degenerate bands with the $\mathcal{PT}$ symmetry, i.e.,
\begin{align}
\label{eq:E}
\varepsilon_{\pm}({\bm k})&=\varepsilon ({\bm k}) \pm X({\bm k}),\\
X({\bm k})&=\sqrt{(\alpha_1 s_x- \tilde{T}_x^{\rm MF})^2 + \alpha^2_2 s^2_z + 4t^2_1c^2_{x/2}},
\end{align}
where $s_x=\sin k_x$, $s_z=\sin k_z$,
$c_{x/2}=\cos k_x/2$, and $\tilde{T}_x^{\rm MF} =\tilde{J}_{\rm AF}T_x^{\rm MF}$.
The factor $(\alpha_1 s_x- \tilde{T}_x^{\rm MF})^2$ includes the coupling between $\tilde{T}_x^{\rm MF}$ and $\alpha_1$ with the odd function of $k_x$.
This asymmetric band modulation due to the coupling between $\alpha_1$ and $\tilde{T}_x^{\rm MF}$ becomes a source of the nonlinear transport as will be discussed in the following sections; see also Appendix~\ref{sec:extract_parameter}.
\section{Second-order nonlinear conductivity and linear response coefficient \label{sec:formalism}}
\subsection{Second-order nonlinear conductivity \label{sec:formalism_nonlinear}}
The second-order nonlinear conductivity tensor $\sigma_{\mu\nu\lambda}$ defined as $J_{\mu}=\sigma_{\mu\nu\lambda}E_{\nu} E_{\lambda}$ ($\mu,\nu,\lambda=x,y,z$) is calculated on the basis of the second-order Kubo formula~\cite{PhysRevResearch.2.043081}.
In the clean limit, the intraband contribution is dominant, which is given by
\begin{align}
\label{eq:ncon}
\sigma_{\mu\nu\lambda} =
\frac{e^3\tau^2}{\hbar^3}\frac{1}{V} \sum_{\bm k} \sum_{n} \frac{\partial^2 \varepsilon_n({\bm k})}{\partial k_\mu\partial k_\nu }
\frac{\partial \varepsilon_n({\bm k})}{\partial k_\lambda} \frac{\partial f [\varepsilon_n ({\bm k})]}{\partial \varepsilon_n({\bm k})},
\end{align}
where $e(>0)$, $\tau$, $\hbar$, and $V$ are the elementary charge, relaxation time, the reduced Planck constant, and the system volume, respectively~\footnote{There is no contribution from the Berry curvature dipole~\cite{sodemann2015} because of the $\mathcal{PT}$ symmetry, while the interband contribution in the $\mathcal{T}$-breaking system~\cite{gao2014} is neglected by considering the clean limit.}.
$f[\varepsilon_n({\bm k})]$ is the Fermi distribution function for the $n$th-band eigen energy $\varepsilon_n({\bm k})$.
The intraband contribution in Eq.~\eqref{eq:ncon} represents the Drude-type one with the dissipation $\tau^{-2}$, whose expression eventually coincides with that obtained by the Boltzmann formalism~\cite{ideue2017bulk, PhysRevB.99.045121, PhysRevResearch.2.043081, gao2019semiclassical}.
Hereafter, we use the scaled $\sigma_{\mu\nu\lambda}$ as $\bar{\sigma}_{\mu\nu\lambda} =\sigma_{\mu\nu\lambda} /(e^3\tau^2\hbar^{-3})$.
From Eq.~(\ref{eq:ncon}), one finds the relation $\sigma_{\mu\nu\nu}=\sigma_{\nu\mu\nu}$ by integration by parts.
This indicates that the Drude-type nonlinear conductivity $\sigma_{\mu\nu\lambda}$ is the totally symmetric rank-3 tensor with 10 independent components: $\sigma_{xxx}$, $\sigma_{yyy}$, $\sigma_{zzz}$, $\sigma_{xyy}$, $\sigma_{yzz}$, $\sigma_{zxx}$, $\sigma_{xxy}$, $\sigma_{yyz}$, $\sigma_{zzx}$, and $\sigma_{xyz}$.
As $\sigma_{\mu\nu\lambda}$ is a third-rank polar time-reversal-odd tensor, i.e., $ \sigma_{\mu\nu\lambda} \to -\sigma_{\mu\nu\lambda}$ under $\mathcal{P}$ or $\mathcal{T}$ operation but $\sigma_{\mu\nu\lambda} \to \sigma_{\mu\nu\lambda}$ under $\mathcal{PT}$ operation, it becomes nonzero when both the spatial inversion and time-reversal symmetries are absent.
From the multipole viewpoint, above symmetry requirement means that the nonzero tensor components are related to the active odd-parity MT multipoles~\cite{dubovik1974multipole, dubovik1986axial, dubovik1990toroid, spaldin2008toroidal, hayami2018microscopic}: three rank-1 MT dipoles $(T_x, T_y, T_z)$ and seven rank-3 MT octupoles $(T_{xyz}, T_x^\alpha, T_y^\alpha, T_z^\alpha, T_x^\beta, T_y^\beta, T_z^\beta)$, whose correspondence is given by~\cite{PhysRevB.104.054412}
\begin{align}
\label{eq:rank12}
\sigma &=
\begin{pmatrix}
\sigma_{xxx} & \sigma_{yxx} & \sigma_{zxx} \\
\sigma_{xyy} & \sigma_{yyy} & \sigma_{zyy} \\
\sigma_{xzz} & \sigma_{yzz} & \sigma_{zzz} \\
\sigma_{xyz} & \sigma_{yyz} & \sigma_{zyz} \\
\sigma_{xzx} & \sigma_{yzx} & \sigma_{zzx} \\
\sigma_{xxy} & \sigma_{yxy} & \sigma_{zxy} \\
\end{pmatrix}^{\rm T}
\notag\\
&\leftrightarrow
\begin{pmatrix}
3{T}_x + 2T_{x}^\alpha
&T_y - T_{y}^\alpha -T_{y}^\beta
& T_z - T_{z}^\alpha + T_{z}^\beta \\
%
T_x - T_{x}^\alpha + T_{x}^\beta
&3{T}_y + 2T_{y}^\alpha
&T_z - T_{z}^\alpha - T_{z}^\beta \\
%
T_x - T_{x}^\alpha -T_{x}^\beta
& T_y -T_{y}^\alpha + T_{y}^\beta
&3{T}_z + 2T_{z}^\alpha \\
T_{xyz}
&T_z - T_{z}^\alpha - T_{z}^\beta
& {T}_y - T_{y}^\alpha + T_{y}^\beta \\
%
{T}_z - T_{z}^\alpha + T_{z}^\beta
& T_{xyz}
& {T}_x -T_{x}^\alpha - T_{x}^\beta \\
%
{T}_y - T_{y}^\alpha - T_{y}^\beta
& {T}_x - T_{x}^\alpha + T_{x}^\beta
& T_{xyz}\\
\end{pmatrix}^{\rm T},
\end{align}
where the functional forms of dipoles and octupoles are summarized in Appendix~\ref{sec:MP_express}.
The correspondence in Eq.~\eqref{eq:rank12} is obtained by decomposing $\sigma_{\mu\nu\lambda}$ into the tensor components with the same rotational symmetry to the dipoles and octupoles (See also Appendix~\ref{sec:MP_express}).
When the MT dipole and/or MT octupole in Eq.~\eqref{eq:rank12} are activated in an AFM metal, the corresponding tensor component of $\sigma_{\mu\nu\lambda}$ becomes nonzero.
From Eq.~(\ref{eq:rank12}), one finds that MT dipole $T_\mu$ is relevant to the longitudinal component $\sigma_{\mu\mu\mu}$ and the transverse components $\sigma_{\mu\nu\nu}$ and $\sigma_{\nu\mu\nu}$ ($\nu\neq\mu$).
It means that both nonreciprocal conductivity and nonlinear transverse conductivity are expected to be realized in the presence of the MT dipole, i.e., ferrotoroidal metals~\cite{litvin2008ferroic,schmid2008some,PhysRevB.104.054412}.
In the present system under the magnetic point-group $m'mm$ with the nonzero MT moment $T_x^{\rm MF}$,
five components $\sigma_{xxx}$, $\sigma_{xyy}$, $\sigma_{yxy}$, $\sigma_{xzz}$, and $\sigma_{zzx}$ can be nonzero, since $T_x$, $T_x^\alpha$, and $T_x^\beta$ in Eq.~\eqref{eq:rank12} belong to the totally symmetric irreducible representation~\cite{PhysRevB.104.054412}.
Among them, $\sigma_{xyy}$ and $\sigma_{yxy}$ vanish owing to $k_y=0$ in the present two-dimensional system.
In addition to the nonzero contribution from the linear conductivity $\sigma_{xx}$,
$\sigma_{xxx}$ results in the nonreciprocal current, while $\sigma_{xzz}$ without linear $\sigma_{xz}$ leads to the pure second-order transverse current, respectively.
\subsection{Linear response coefficient \label{sec:formalism_linear}}
In the presence of the MT moment $T_x^{\rm MF}$, the linear magnetoelectric tensor $\alpha_{\mu\nu}$ in $M_\mu=\alpha_{\mu\nu}E_{\nu}$ ($\mu,\nu=x,y,z$) is also finite.
We calculate the linear magnetoelectric tensor by the linear response theory as
\begin{align}
\label{eq:a_munu}
\alpha_{\mu\nu} =& \frac{eg\mu_{\rm B} \hbar}{2Vi} \sum_{\bm k} \sum_{n \neq m} \frac{ f[\varepsilon_n ({\bm k})]-f[\varepsilon_m({\bm k})]}{[\varepsilon_n({\bm k})-\varepsilon_m({\bm k})]^2 + (\hbar \delta)^2} \sigma_{\mu{\bm k}}^{nm} \varv_{\nu{\bm k}}^{mn},
\end{align}
where $g$ and $\mu_{\rm B}$ are the g factor ($g=2$) and Bohr magneton, respectively.
$\sigma_{\mu{\bm k}}^{nm}=\braket{n{\bm k}|\sigma_{\mu}|m{\bm k}}$ and $\varv_{\nu{\bm k}}^{mn}=\braket{m{\bm k}|\varv_{\nu{\bm k}}|n{\bm k}}$ are the matrix elements of spin $\sigma_{\mu}$ and velocity $\varv_{\nu{\bm k}}=\partial \mathcal{H}/(\hbar \partial k_{\nu})$ for the eigenstate $\ket{n{\bm k}}$.
We use the scaled $\bar{\alpha}_{\mu\nu} = \alpha_{\mu\nu}/(e\mu_{\rm B}\hbar)$ in the following discussion.
As $\alpha_{\mu\nu}$ in a $\mathcal{PT}$ symmetric system is relevant to the rank-0--2 odd-parity multipoles: magnetic monopole $M_0$, MT dipoles $(T_x, T_y, T_z)$, and magnetic quadrupoles $(M_u, M_v, M_{yz}, M_{zx}, M_{xy})$ (see also Appendix~\ref{sec:MP_express}), the relation is represented as follows~\cite{hayami2018classification, watanabe2018group}:
\begin{align}
\alpha &=
\begin{pmatrix}
\alpha_{xx} & \alpha_{xy} & \alpha_{xz}\\
\alpha_{yx} & \alpha_{yy} & \alpha_{yz}\\
\alpha_{zx} & \alpha_{zy} & \alpha_{zz}\\
\end{pmatrix} \\
&\leftrightarrow
\begin{pmatrix}
M_0 -M_u+M_v & M_{xy}+T_z & M_{zx}-T_y\\
M_{xy}-T_z & M_0-M_u-M_v & M_{yz}+T_x \\
M_{zx}+T_y & M_{yz}-T_x & M_0+2M_u \\
\end{pmatrix}.
\label{eq:formalism_ME_mul}
\end{align}
Since $T_x$ and $M_{yz}$ become active for $T_x^{\rm AF}\neq 0$ in the present system,
$\alpha_{yz}$ and $\alpha_{zy}$ are expected to be nonzero.
As $\alpha_{zy}$ is zero due to the two dimensionality, we only consider $\alpha_{yz}$.
For the following discussion, we also present the linear Hall conductivity
\begin{align}
\label{eq:s_xz}
\sigma_{xz} =& \frac{e^2\hbar}{Vi} \sum_{\bm k} \sum_{n \neq m} \frac{ f[\varepsilon_n ({\bm k})]-f[\varepsilon_m({\bm k})]}{[\varepsilon_n({\bm k})-\varepsilon_m({\bm k})]^2 + (\hbar \delta)^2} \varv_{x{\bm k}}^{nm} \varv_{z{\bm k}}^{mn}.
\end{align}
We use the scaled value $\bar{\sigma}_{xz}=\sigma_{xz}/(e^2\hbar H_y)$ in the following, where $H_y$ is the Zeeman field along the $y$ direction.
\section{Numerical Result \label{sec:numerical}}
\subsection{Longitudinal second-order conductivity $\sigma_{xxx}$ \label{sec:numerical_xxx}}
\begin{figure}[t!]
\centering
\includegraphics[width=85mm]{Fig2.eps}
\caption{
(a) The longitudinal second-order conductivity $\bar{\sigma}_{xxx}$ for $\alpha_1=0.1$--$0.5$ as a function of $T$ at $\alpha_2=0.1$.
The inset shows $\bar{\sigma}_{xxx}/\alpha_1$.
(b) The upper- and lower-band contributions to $\bar{\sigma}_{xxx}$ at $\alpha_1=0.4$.
\label{fig:fig2}}
\end{figure}
We first show the numerical result of the longitudinal nonlinear conductivity $\bar{\sigma}_{xxx}$.
Figure~\ref{fig:fig2}(a) shows $\bar{\sigma}_{xxx}$ as a function of $T$ for various $\alpha_1=0.1$--$0.5$ at $\alpha_2=0.1$.
The $T$ dependence for different $\alpha_1$ is qualitatively similar; $\bar{\sigma}_{xxx}$ is largely enhanced just below $T=T_{{\rm N}}$, and shows maximum with decrease of $T$.
While further decreasing $T$, $\bar{\sigma}_{xxx}$ shows the sign change, and then reaches a negative value at the lowest $T$.
The nonzero $\sigma_{xxx}$ is closely related to the formation of the asymmetric band structure under $T_x^{\rm MF} \neq 0$, since $\sigma_{xxx}$ has the same symmetry as $T_x^{\rm MF}$~\cite{PhysRevB.104.054412}.
As the asymmetric band modulation is caused by the coupling between $\tilde{T}_x^{\rm MF}$ and $\alpha_1$,
they are indispensable for nonzero $\sigma_{xxx}$.
Indeed, $\bar{\sigma}_{xxx}$ vanishes for $\alpha_1=0$ or $\tilde{T}_x^{\rm MF}=0$.
Moreover, $\bar{\sigma}_{xxx}$ is well scaled by $\bar{\sigma}_{xxx}/\alpha_1$ at low temperatures $T \lesssim 0.7T_{\rm N}$ for small $\alpha_1$.
See Sec.~\ref{sec:Essential model parameters} for the essential model parameters in details.
Meanwhile, $\bar{\sigma}_{xxx}$ is not scaled by $\alpha_1$ for $0.7\lesssim T/T_{\rm N}\leq 1$ in the region where $\bar{\sigma}_{xxx}$ is drastically enhanced.
This is attributed to the rapid increase of $\tilde{T}_x^{\rm MF}$ and resultant drastic change of the electronic band structure near the Fermi level.
As $\bar{\sigma}_{xxx}$ in Eq.~(\ref{eq:ncon}) includes the factors $\partial^2 \varepsilon_n({\bm k})/\partial k_x^2 $ and $\partial \varepsilon_n({\bm k})/\partial k_x$, the small $X(\bm{k})$ appearing in the denominator of $\partial^2 \varepsilon_n({\bm k})/\partial k_x^2 $ and $\partial \varepsilon_n({\bm k})/\partial k_x$
gives a dominant contribution.
When considering the small order parameter compared to the ASOI, i.e., $\tilde{T}_x^{\rm MF} \lesssim \alpha_1$, $X(\bm{k})$ can become small when the Fermi wavenumber $k^{\rm F}_x$ satisfies $\tilde{T}_x^{\rm MF} \simeq \alpha_1 \sin k^{\rm F}_x$, which results in a large enhancement of $\bar{\sigma}_{xxx}$.
Such an enhancement is remarkable when the upper and lower bands are closely located in the paramagnetic state with small $X(\bm{k})$ as shown in Fig.~\ref{fig:fig1}(d), which can be realized for small $t_1=0.1$ and $\alpha_2=0.1$.
In short, there are two conditions for large $\bar{\sigma}_{xxx}$:
One is the large essential parameters, such as $\alpha_1$, $T_x^{\rm MF}$, and $J_{\rm AF}$,
and the other is to satisfy $\tilde{T}_x^{\rm MF} \simeq \alpha_1 \sin k^{\rm F}_x$ in a multi-band system.
These conditions can be experimentally controlled by electron/hole doping and temperature.
The sign change of $\bar{\sigma}_{xxx}$ in $T$ dependence is owing to the multiband effect.
As shown in Fig.~\ref{fig:fig1}(d), the band bottom is shifted in the opposite direction for the upper and lower bands, which means that the opposite sign of the coupling $\alpha_1 \tilde{T}_x^{\rm MF}$ results in the opposite contribution to $\bar{\sigma}_{xxx}$.
This is demonstrated by decomposing $\bar{\sigma}_{xxx}$ into the upper- and lower-band contributions, as shown in Fig.~\ref{fig:fig2}(b).
The results indicate that the dominant contribution of $\bar{\sigma}_{xxx}$ arises from the upper band for $0.9\lesssim T/T_{\rm N}\leq 1$, while that arises from the lower band for $T/T_{\rm N}\lesssim 0.9$.
The suppression of the upper-band contribution for low $T$ is because it becomes away from the Fermi level by the development of $T_x^{\rm MF}$.
\subsection{Transverse second-order conductivity $\sigma_{xzz}$ \label{sec:numerical_xzz}}
\begin{figure}[htb!]
\centering
\includegraphics[width=90mm]{Fig3.eps}
\caption{
The transverse second-order nonlinear conductivity $\bar{\sigma}_{xzz}$ for several $\alpha_1$ and $\alpha_2$ with $\alpha_1=\alpha_2$.
The inset represents $\bar{\sigma}_{zxx}/(\alpha_1\alpha_2^2)$.
\label{fig:fig3}}
\end{figure}
Next, let us discuss the transverse nonlinear conductivity $\bar{\sigma}_{xzz}$.
Figure~\ref{fig:fig3} shows the $T$ dependence of $\bar{\sigma}_{xzz}$ for $0.02 \leq \alpha_1, \alpha_2 \leq 0.1$ with $\alpha_1=\alpha_2$.
The behavior of $\bar{\sigma}_{xzz}$ against $T$ is similar to $\bar{\sigma}_{xxx}$ except for the sign change; $\bar{\sigma}_{xzz}$ becomes nonzero below $T=T_{\rm N}$ and shows the maximum near $T_{\rm N}$.
While decreasing $T$, $\bar{\sigma}_{xzz}$ is suppressed and shows an almost constant value.
Similar to $\sigma_{xxx}$, the origin of nonzero
$\sigma_{xzz}$ is the asymmetric band modulation under $T_x^{\rm MF} \neq 0$ via the effective coupling $\tilde{T}_x^{\rm MF}\alpha_1$.
Besides, we find another contribution from $\alpha_2$ for nonzero $\sigma_{xzz}$ in contrast to $\sigma_{xxx}$, where
$\bar{\sigma}_{xzz}$ is well scaled by $\alpha_1 \alpha_2^2$ as shown in the inset of Fig.~\ref{fig:fig3}, as discussed in Sec.~\ref{sec:Essential model parameters}.
The additional parameter dependence for $\alpha_2^2$ is owing to an additional symmetry between $k_z$ and $k_z+\pi$ for $\alpha_2=0$, which gives the opposite-sign contribution to $\sigma_{xzz}$ so that totally $\sigma_{xzz}=0$.
\subsection{Comparison to magnetoelectric coefficient $\alpha_{yz}$ \label{sec:numerical_yz}}
\begin{figure}[htb!]
\centering
\includegraphics[width=90mm]{Fig4.eps}
\caption{
(a) The magnetoelectric coefficient $\bar{\alpha}_{yz}$ and
(b) the quantity $\bar{\sigma}_{xz}\bar{\alpha}_{yz}$ with the same parameters as Fig.~\ref{fig:fig3}.
$\bar{\sigma}_{xz}$ is calculated by supposing the magnetic field $H_y=0.01$.
The insets of (a) and (b) represent $\bar{\alpha}_{yz}/\alpha_2$ and $\bar{\sigma}_{xz} \bar{\alpha}_{yz}/(\alpha_1\alpha_2^2)$, respectively.
\label{fig:fig4}}
\end{figure}
We also present another MT-moment-driven phenomena, the magnetoelectric response, and compare its parameter and $T$ dependence to the nonlinear conductivities obtained in the previous section.
Figure~\ref{fig:fig4}(a) shows the $T$ dependence of $\bar{\alpha}_{yz}$ for $0.02 \leq \alpha_1, \alpha_2 \leq 0.1$ with $\alpha_1=\alpha_2$, whose behavior is similar to the transverse nonlinear conductivity $\sigma_{xzz}$ in Fig.~\ref{fig:fig3} except for the sign.
$\bar{\alpha}_{yz}$ is nonzero even if $\alpha_1=0$ that is different from the nonlinear conductivities, whereas $\alpha_2$ and $\tilde{T}_x^{\rm MF}$
are essential to obtain the finite $\bar{\alpha}_{yz}$, as detailed in Sec.~\ref{sec:Essential model parameters}.
As shown in the inset of Fig.~\ref{fig:fig3}(a), $\bar{\alpha}_{yz}$ is well scaled as $\bar{\alpha}_{yz}/\alpha_2$ for small $\alpha_2$.
Moreover, it is noteworthy to comment on the relation between the transverse nonlinear conductivity and a combination of the linear magnetoelectric and Hall coefficients, since the nonlinear transverse transport in the $\mathcal{PT}$-symmetric AFMs can be understood as the Hall transport driven by the induced magnetization through the linear magnetoelectric response at the phenomenological level~\cite{hayami2014toroidal, oyamadaa2018anomalous}.
We show the $T$ dependence of $\bar{\sigma}_{xz}\bar{\alpha}_{yz}$ in Fig.~\ref{fig:fig4}(b) for the same parameters in Fig.~\ref{fig:fig3}.
The small magnetic field $H_y=0.01$ is introduced to mimic the induced magnetization in $\alpha_{yz}$.
Compared to the results in Fig.~\ref{fig:fig3} and \ref{fig:fig4}(b),
one finds the resemblance between the $T$ dependences of $\bar{\sigma}_{xzz}$ and $\bar{\sigma}_{xz}\bar{\alpha}_{yz}$, both of which are scaled by $\alpha_1 \alpha_2^2$.
A good qualitative correspondence in these responses indicates that the interpretation of dividing subsequent two linear processes for nonlinear conductivity is reasonable in the present model.
The overall quantitative difference $\bar{\sigma}_{xz}\bar{\alpha}_{yz}/\bar{\sigma}_{xzz}\sim 10^{-2}$ may be ascribed to the magnitude of the used internal magnetic field ($H_{y}=0.01$) that should be replaced by the true internal field.
However, it is hard to estimate it quantitatively.
\section{Discussion \label{sec:discussion}}
\subsection{Essential model parameters \label{sec:Essential model parameters}}
We discuss the parameter dependences of the asymmetric band modulation, nonlinear conductivity, and the linear magnetoelectric and Hall coefficients at the level of the microscopic model Hamiltonian.
For this purpose, we try to extract the essential parameters for each response from various hoppings, spin-orbit coupling, and internal/external field in the model Hamiltonian based on the method in Refs.~\onlinecite{hayami2020bottom,doi:10.7566/JPSJ.91.014701}.
In the following, we discuss the important model parameters in each case one by one, and the results are summarized in Table~\ref{tab_params}.
The derivation is shown in Appendix~\ref{sec:extract_parameter}.
First, the essential parameters for the asymmetric band modulation~\cite{hayami2020bottom} are given by $\tilde{T}_x^{\rm MF}\alpha_1$, as shown in Appendix~\ref{sec:abm}.
The result is consistent with the eigenvalues in Eq.~(\ref{eq:E}).
Next, the essential model parameters for $\sigma_{xxx}$~\cite{doi:10.7566/JPSJ.91.014701} (see also Appendix~\ref{sec:ncon}) are given by
\begin{align}
\sigma_{xxx}=&\alpha_1 \tilde{T}_x^{\rm MF} \left[t_1^2F(t_1,t_2,t_3,\alpha_1,\alpha_2, \tilde{T}_x^{\rm MF})\right. \notag\\
&\left. \qquad \quad+t_2F'(t_1,t_2,t_3,\alpha_1,\alpha_2, \tilde{T}_x^{\rm MF})\right],
\end{align}
where $F$ and $F'$ represent the arbitrary functions.
Note that only the even power of $\alpha_1$ and $\tilde{T}_x^{\rm MF}$ appears in $F$ and $F'$.
Thus, one finds that the coupling of $\alpha_1$ and $\tilde{T}_x^{\rm MF}$ is always necessary to induce $\sigma_{xxx}$, which is consistent with the numerical result presented in Sec.~\ref{sec:numerical_xxx}.
Moreover, $\sigma_{xxx}$ is closely related to the asymmetric band modulation because both of them are characterized by the same essential model parameters.
Similarly, the essential model parameters of $\sigma_{xzz}$ are given by
\begin{align}
\sigma_{xzz}= \alpha_1\tilde{T}_x^{\rm MF} \left[\alpha_2^2 t_2 F(t_1,t_2,t_3,\alpha_1,\alpha_2,\tilde{T}_x^{\rm MF})\right],
\label{eq:para_xzz}
\end{align}
where the even power of $\alpha_1$, $\alpha_2$, and $\tilde{T}_x^{\rm MF}$ appears in $F$.
Equation~\eqref{eq:para_xzz} shows that the coupling of $\alpha_1$ and $\tilde{T}_x^{\rm MF}$ is essential to induce $\sigma_{xzz}$ as similar to $\sigma_{xxx}$, which is consistent with the numerical result in Sec.~\ref{sec:numerical_xzz}.
Moreover, Eq.~\eqref{eq:para_xzz} indicates that $t_2$ and even power of $\alpha_2$ are also necessary for $\sigma_{xzz}$ in the present model in Eq.~\eqref{eq:hamiltonian}.
In a similar way, the essential model parameters to induce $\alpha_{yz}$ and $\sigma_{xz}$ are given by
\begin{align}
\alpha_{yz} &= \alpha_{2} \tilde{T}_x^{\rm MF} \left[t_3 F(t_1,t_2,t_3,\alpha_1,\alpha_2,\tilde{T}_x^{\rm MF})\right], \\
\sigma_{xz} &= \alpha_{1} \alpha_{2} H_{y} \left[t_3 F(t_1,t_2,t_3,\alpha_1,\alpha_2, H_y,\tilde{T}_x^{\rm MF})\right].
\end{align}
This indicates that nonzero $\alpha_{yz}\sigma_{xz}$ needs nonzero $\alpha_1\alpha_2^2 \tilde{T}_x^{\rm MF}$, which shows a good agreement with the condition for $\sigma_{xzz}$.
The common essential model-parameter dependence in small parameter region was already confirmed in Secs.~\ref{sec:numerical_xzz} and \ref{sec:numerical_yz}.
It is noteworthy that the above approach to extract the essential model parameters can be straightforwardly applied even when introducing the other model parameters.
For example, let us consider the additional interlayer A-B hopping $t_4$ in the model Hamiltonian.
In this situation, one finds that there is no longer simple correlation between $\sigma_{xzz}$ and $\sigma_{xz}\alpha_{yz}$; the essential model parameters for the former are $\alpha_1 \tilde{T}_x^{\rm MF}$ rather than $\alpha_1\alpha_2^2 \tilde{T}_x^{\rm MF}$, while those for the latter still remains the same as $\alpha_1\alpha_2^2 \tilde{T}_x^{\rm MF}$ as discussed in Appendix~\ref{sec:extract_parameter}.
In other words,
the factor $\alpha^2_2 t_2$ in the square bracket in Eq.~(\ref{eq:para_xzz}) is not truly the essential factor.
Indeed, the numerical results in the presence of $t_4$ give a different temperature dependence from each other, as shown in Appendix~\ref{sec:interlayer_hopping}.
Thus, the correspondence between $\sigma_{xzz}$ and $\sigma_{xz}\alpha_{yz}$ occurs depending on the hopping in the effective model, which is clarified by performing a procedure in Appendix~\ref{sec:extract_parameter}.
\begin{table}[h]
\caption{
Model parameters necessary for the asymmetric band modulation and response tensors indicated by the checkmark (\checkmark).
In the last two columns, model parameters are decomposed into the essential and semi-essential parts.
}
\label{tab_params}
\begin{center}
\begin{tabular}{lcccccccc}
\hline \hline
& $t_2$ & $t_3$ & $\alpha_1$ & $\alpha_2$ & $\tilde{T}_x^{\rm MF}$ & $H_{y}$ & essential &
semi-essential \\ \hline
asymmetric & & & \checkmark & & \checkmark & & $\alpha_1 \tilde{T}_x^{\rm MF}$ \\
band modulation & \\\hline
$\sigma_{xxx}\,\, (t_4 = 0)$ & & & \checkmark & & \checkmark & & $\alpha_1 \tilde{T}_x^{\rm MF}$ & $t_1^2, t_2$ \\
$\sigma_{xxx}\,\, (t_4 \neq 0)$ & & & \checkmark & & \checkmark & & $\alpha_1 \tilde{T}_x^{\rm MF}$ & $t_1^2, t_2, t_4$ \\
$\sigma_{xzz}\,\, (t_4 = 0)$ & \checkmark & & \checkmark & \checkmark & \checkmark & & $\alpha_1 \tilde{T}_x^{\rm MF}$ & $\alpha_2^2t_2$ \\
$\sigma_{xzz}\,\, (t_4 \neq 0)$ & & & \checkmark & & \checkmark & & $\alpha_1 \tilde{T}_x^{\rm MF}$ & $\alpha_2^2t_2, t_4$ \\ \hline
$\alpha_{yz}\,\, (t_4 = 0)$ & & \checkmark & & \checkmark & \checkmark & & $\alpha_2 \tilde{T}_x^{\rm MF}$ & $t_3$ \\
$\alpha_{yz}\,\, (t_4 \neq 0)$ & & & & \checkmark & \checkmark & & $\alpha_2 \tilde{T}_x^{\rm MF}$ & $t_3, t_4$ \\ \hline
$\sigma_{xz}\,\,(t_4 = 0)$ & & \checkmark & \checkmark & \checkmark & & \checkmark & $\alpha_1 \alpha_2 H_y$ & $t_3$ \\
$\sigma_{xz}\,\, (t_4 \neq 0)$ & & & \checkmark & \checkmark & & \checkmark & $\alpha_1 \alpha_2 H_y$ & $t_3, t_4$ \\
\hline \hline
\end{tabular}
\end{center}
\end{table}
\subsection{Quantitative evaluation}
Finally, we discuss the order estimate of the nonlinear conductivity
for $\alpha_1=0.5$ and $\alpha_2=0.1$ by the ratio $\sigma_{xxx}/(\sigma_{xx})^2$ with being independent of
the relaxation time in the clean limit.
By putting the typical values as $a\sim 0.5$~[nm] and $|t_2|=0.2$~eV,
we obtain $\sigma_{xxx}/(\sigma_{xx})^2 \sim 10^{-3} \hbar a^{2} e^{-1} |t_2|^{-1}
\sim 10^{-18}$~[m$^3$ A$^{-1}$] for $T\to 0$ and
$10^{-17}$~[m$^3$ A$^{-1}$] near $T_{\rm N}$, which is comparable to the value in the 2D
nonmagnetic Rashba system under the magnetic field~\cite{ideue2017bulk}.
Further enhancement can be achieved by tuning the model parameters and electron filling.
\section{Summary \label{sec:summary}}
In summary, we investigated the microscopic essence for the second-order nonlinear conductivity in the $\mathcal{PT}$-symmetric collinear AFM with the MT moment on a two-dimensionally stacked zigzag chain by focusing on the role of the MT moment.
Based on the nonlinear Kubo formula in the clean limit, we found that the effective coupling between the ASOI and the MT moment is essential for the nonlinear conductivity.
By analyzing both the longitudinal and transverse components of the nonlinear conductivity while changing the ASOI and the temperature, we showed that their large enhancement can be achieved near the transition temperature, provided that the AFM molecular field is comparable to the ASOI in a multi-band system.
We also discussed the similarity and difference between the transverse nonlinear transport and the combined response consisting of the linear magnetoelectric and Hall coefficients.
The present result elucidates the essential model parameters for MT-related physical phenomena, such as the nonlinear conductivity and the linear magnetoelectric effect, in $\mathcal{PT}$-symmetric collinear AFMs.
The similar analysis can be applied to examine the role of the MT moment for any collinear AFMs with the MT moment in the zigzag structure, e.g.,
CeRu$_2$Al$_{10}$~\cite{Tursina:wm6046, doi:10.1143/JPSJ.80.073701}, Ce$_3$TiBi$_5$~\cite{doi:10.7566/JPSJ.89.033703,doi:10.7566/JPSCP.30.011189}, and $\alpha$-YbAl$_{1-x}$Mn$_x$B$_4$~\cite{PhysRevResearch.3.023140},
and other ferrotoroidal metals/semiconductors with locally noncentrosymmetric crystal structures, such as Mn$_2$Au~\cite{barthem2013revealing, PhysRevLett.127.277202}, $R$B$_4$ ($R=$Dy, Er)~\cite{WILL1981349, WILL197931},
CuMnAs~\cite{wadley2015antiferromagnetic,PhysRevLett.127.277201}, PrMnSbO~\cite{PhysRevB.82.100412}, NdMnAsO~\cite{PhysRevB.83.144429}, and $X_y$Fe$_{2-x}$Se$_2$ ($X=$K, Tl, Rb)~\cite{Bao_2011, Pomjakushin_2011, PhysRevLett.109.077003}, once the model Hamiltonian is given.
The measurements of the linear magnetoelectric effect and the nonlinear conductivity for these materials are also useful to investigate their microscopic mechanisms.
Moreover, the analysis is straightforwardly extended to the AFMs with the other odd-parity magnetic-type multipole moments, such as the MT octupole, since they are characterized by the same spatial inversion and time-reversal symmetries.
Our study will stimulate a further investigation of the multiferroic and conductive phenomena in the $\mathcal{PT}$-symmetric AFM metals.
\begin{acknowledgements}
We thank Y. Motome and Y. Yanagi for the fruitful discussion.
This research was supported by JSPJ KAKENHI Grant Numbers
JP19K03752, JP19H01834, JP21H01037, and by JST PRESTO (JPMJPR20L8).
M.Y. and R.O. are supported by a JSPS research fellowship and JSPS KAKENHI (Grant No. JP20J12026
and JP20J21838).
\end{acknowledgements}
|
\section{Introduction}
With the recent development of deep learning and the accumulation of electronic health records (EHR), also known as time-series data, there has been an increasing effort in clustering EHR data in order to discover meaningful patterns throughout longitudinal health information. Moreover, chronic diseases, such as Parkinson's disease (PD) and Alzheimer's disease (AD), can have various outcomes even with a limited number of patients. Such diseases are heterogeneous in nature and often evolve at unique patterns that trigger distinct responses to therapeutic interventions based upon different conditions \cite{kehagia2010neuropsychological}. Thus, it has become crucial to develop a disease progression modeling (DPM) system to capture certain progression patterns, provide early detection to critical situations, and yield clinically helpful information to improve the quality of care.
Traditionally, DPM or disease clustering/staging is developed by domain experts with extensive clinical experience, in which disease stages are defined separately and based solely on the values of one or a few biomarkers \cite{ferrer1997chronic,auer1997gds}. Nevertheless, developing a DPM system requires long-term observation and human labor, and the result is often based on known biomarkers and acknowledged covariants, which makes it difficult to develop a DPM system for disease with limited knowledge on biomarkers that have not been well-studied. In recent years, the rapid growth of data-driven machine learning methods has motivated a great effort in developing DPM models. There are two main approaches when it comes to DPM: 1) The problem is formed as a risk prediction task with label information based on patient representation that is extracted from the last layer of the model. \cite{lee2020temporal,gao2020stagenet,ma2018health,sun2019probabilistic,zhang2019data}. 2) The problem is formed as a traditional unsupervised, patient clustering/subtyping problem where the model is trained to separate the patient into multiple groups \cite{wang2014unsupervised,fortuin2018som,mou2019t}. Leveraging disease outcomes during the training process can prevent the model from forming heterogeneous clusters. However, for certain diseases, diagnosis labels are often unavailable at each patient visit due to limited knowledge of the disease. Moreover, deep learning models that are designed for supervised tasks may not perform well when training in an unsupervised fashion. Therefore, there is a need for developing a DPM framework that can handle both situations with respect to the availability of training labels. However, most developed deep learning models for disease progression modeling suffers from the following limitations:
\begin{itemize}
\item \textbf{Irregularity and heterogeneity}: Many diseases are heterogeneous in nature and EHR data often has high internal complexity. Due to the complexity of effectively encoding various health conditions into patient representation, accurate DPM still remains a challenging problem.
\item \textbf{Long-term Dependency}: RNNs are long known to suffer from modeling long-term dependency since it tends to forget earlier information when the input sequence is long. Disease progression modeling, especially for chronic disease, requires long-term observation of the patient in order to provide a comprehensive view for decision making.
\item \textbf{Target Awareness}: Most rnn-based methods derive patient representations directly from the hidden states of the model. Such an approach neglect the contribution of target-relevant information. In fact, real-world clinical decisions made by doctors are often based upon past diagnoses as well.
\end{itemize}
To address these challenges, we propose Temporal Clustering with External Memory Network (TC-EMNet) for disease progression modeling via both supervised and unsupervised settings. TC-EMNet leverages a variation autoencoder framework and a memory network to deal with data irregularity and long-term dependency problems of RNNs respectively. At each time step, TC-EMNet takes EHR medical records as input and encodes the input feature using a recurrent neural network to get hidden representations. Then TC-EMNet samples from the hidden state to form a latent representation. Meanwhile, the hidden state is stored in a global-level memory network, which in turn outputs a memory representation based on current memory cells. The memory representation is then concatenated with the current latent representation to form the patient representation at the current time step. When the training label is available, the model also employs a patient-level memory network to process label information up to most recent visit and outputs target-aware memory representation. We combine memory representations from global-level and patient-level memory networks using a calibration process. TC-EMNet is trained with reconstruction objective under unsupervised setting and prediction objective under supervised setting.
In this paper, our contributions are four fold:
\begin{itemize}
\item We propose a novel deep learning framework, namely TC-EMNet for disease progression modeling under both supervised and unsupervised settings.
\item TC-EMNet uses a combined recurrent neural network and variational auto-encoder (VAE) architecture to capture the irregularity in data and heterogeneity nature of the disease.
\item Under superviesd setting, TC-EMNet employs dual memory network architecture to leverage both hidden representations from the input data and clinical diagnosis to produce accurate patient representations.
\item Experiments on two world datasets show that TC-EMNet yields competitive clustering performance over state-of-the-art methods and is able to find clinically interpretable disease clusters/stages.
\end{itemize}
The remainder of the paper is organized as follows. Section II briefly reviews existing works related to DPM, temporal clustering, and VAE. Section III describes the technical details of the proposed model (TC-EMNet). Section IV and V present experimental results and discussions. Finally, Section VI concludes the paper.
\begin{figure*}[t]
\centering
\includegraphics[width=0.85\textwidth]{pics/framework4.png}
\caption{
Overview of the proposed framework. At each timestamp, hidden representation from the encoder network is updated with the memory state to produce disease clusters/stages based on current and previous observations.
}
\label{framework}
\end{figure*}
\section{Related Work}
\subsection{Disease Progression Modeling}
Disease progression modeling (DPM) plays a very important role in the healthcare domain, especially for chronic diseases such as Parkinson's Disease (PD) and Alzheimer's Disease (AD). A well-performed disease progression modeling system can not only provide early detection or diagnosis but also discover clinically meaningful patterns for certain groups of trajectories.
Most probabilistic models for DPM are based on the hidden markov model (HMM).
For example, \cite{alaa2019attentive} derived a deep probabilistic model based on sequence-to-sequence architecture to model progression dynamic on UK Cystic Fibrosis Registry. \cite{wang2014unsupervised} introduced a continuous-time Markov process to learn a discrete representation of each progression state. Moreover, deep learning methods have also been developed for disease progression modeling. \cite{liu2018joint} proposed a CNN-based model to jointly learn features from MR images combined with demographic information to predict Alzheimer's Disease progression patterns. \cite{teng2020stocast} designed a prediction framework using generative models to forecast the distribution of patients' outcomes. DPM can be regarded as a classification problem, where diagnosis labels are leveraged in favor of model training. On the other hand, DPM can also be seen from an unsupervised perspective where the goal is to discover potential disease states or patient subtypes throughout patients' medical history \cite{dennis2019disease}. However, DPM still remains a challenge due to the high complexity of data introduced by irregular progression patterns for certain chronic diseases.
\subsection{Temporal clustering}
Temporal clustering, widely known as time-series clustering, is a data-driven method to cluster patients into subgroups based on time-series observation. Temporal clustering can be considered a challenging task often because of the high dimensionality of the dataset and multiple time steps for each data sample. Recent advances have been focused on leveraging the latent representation learned by recurrent neural network (RNN) for temporal clustering, which was motivated by the success of RNN modeling time-series data. Moreover, due to the emerging availability of electronic health records (EHR) that introduced large-scale and normalized context for individual patients, the deep learning approach become capable of learning more comprehensive patterns and achieving better performance on several critical tasks. \cite{baytas2017patient} introduces a time-aware mechanism to long short term memory cells to capture progression patterns with irregular time-interval. \cite{lee2020temporal} proposed an actor-critic algorithm for predictive clustering where, instead of defining a similarity measure for clustering, a cluster embedding is trained to represent each disease stage. \cite{yin2020identifying} proposed an auto-encoder to reconstruct relevant features for sepsis with attention and showed that the proposed model can identify interpretable patient subtypes. Nevertheless, there is only limited literature that focuses on DPM using temporal clustering techniques.
\subsection{Variational Autoencoder}
Variational autoencoder (VAE) is a type of generative model that can handle complicate distributions. VAEs are effective against modeling complex data structures and are widely adopted to solve many real-world problems range from image generation to anomaly detection \cite{burgess2018understanding, higgins2016beta, kuo2017variational}. It has also several successful applications with healthcare data \cite{shickel2017deep}. \cite{jun2019stochastic} proposed to use VAE to impute missing values for electronic health data with uncertainty-aware attention. Experiments on real-world datasets show that VAE can capture the complexity of EHR distribution. \cite{teng2020stocast} leveraged the VAE framework to forecast disease states for Parkinson's Disease (PD) and Alzheimer's Disease (AD). Nonetheless, the latent representation learned from VAE can be drawn from unrealistic distribution if trained without any constraints.
\section{Methodology}
\subsection{Problem Definition}
Let $x \subseteq \boldsymbol{X} $ and $y \subseteq \boldsymbol{Y} $ be the random variables for input feature space and label space accordingly. Here we focus on a clustering problem, where we are given a population of time-series data
$D = \{(x_{t}^{n}, y_{t}^{n})_{t=1}^{T}\}_{n=1}^{N}$ consisted of paired sequences of observations $(x, y)$ for $N$ patients. $t \subseteq {1, ..., T}$ denotes the time stamps for each patients at which the observations are made.
We aim to identify $K$ clusters for time-series data, each corresponding to a disease stage. Each cluster consists of homogeneous data samples, represented by the centroids based on certain similarity measures.
\subsection{Method}
This section presents our proposed framework. Here we discuss disease progression modeling under both supervised and unsupervised settings, where our proposed question requires estimating the underlying distribution of all possible disease stages. Such a DPM framework can help the doctors identify meaningful characteristics in both times when a disease has certain diagnosis labels but possible underlying disease stages and when a disease has no well-defined diagnosis labels.
The framework consists of three components: the encoder, the memory network, and the clustering network. For each patient, a recurrent neural network is deployed to encode the patient's information. The memory network controls the overall long-term information at each timestamp. Specifically, when a hidden representation $h_{i}$ is generated based on current and previous observation $X_{<i}$ at timestamp $i$, the hidden state is read by the memory network and updates the memory storage. Next, a latent variable $z_{i}$ is drawn from the prior distribution $p_{\theta}(z_{i}|X_{<i})$ conditioned on the hidden state that is generated from the memory network. Then, we either yield prediction outcomes or reconstruct the current observation $X_{i}$ accordingly. We take the hidden presentation from the last layer of the model for clustering.
\subsubsection{Encoder Network}
The encoder network takes the current observation and the hidden state from the previous timestamp and yields the hidden representation that can interact with the external network. Specifically, a LSTM cell is adopted to generate and update the hidden state:
\begin{equation}
h_{t} = LSTM(X_{t}, h_{t-1}),
\label{encoding}
\end{equation}
where $X_{t}$ is the current observation at timestamp $t$ and $h_{t-1}$ is the hidden state from previous step. At each timestamp, the encoder network maps a sequence of time-series input $x_{1:t}$ to a hidden representation $z_{t} \subseteq Z$, where $Z$ is the subspace of latent representation. The hidden representation will be interacting with the external memory network to form an accurate representation.
\subsubsection{Memory Network}
Long-term information plays an important role in disease progression modeling, since, in the context of chronic disease, the health conditions from the past will affect the current disease stages of the patient. In addition, historical information should be stored in an efficient way such that it can provide useful guidance towards the patient's current health state at different timestamps. To this end, we propose an external memory network to capture long-term information throughout the progression modeling process. Our proposed memory network is closely related to \cite{sukhbaatar2015end}, which has several successful applications in the field of natural language processing. Similarly, we define memory slots to represent historical information that can be extracted and retrieved at any given timestamp. At each timestamp, the hidden state from the encoder network is recorded and written into the memory cells. By pushing through a series of observations, the memory network will process continuous representations for each individual visit so that a more comprehensive review of the patient can be utilized during the clustering/staging process.
\paragraph{Memory Reading}
We denote a clinical sequences record $r_{t}, t=1,...,T$, where t stands for index or timestamps of the given record. In memory network, after receiving a hidden representation $z_{t}$ from the encoder network, the network will produce an external representation $e_{t}$ based on reading weight $w_{t}^{l}, l=1,...,T$ of the memory slots. Specifically, $e_{t}$ can be expressed as:
\begin{equation}
\begin{aligned}
& e_{t} = \sum_{l} w_{t}^{l}m_{t}, \\
& w_{t}^l = \textbf{softmax}(\alpha_{t}, m_{t}, h_{t}) \\
& = \frac{\exp{(\alpha_{t}^{l})}C(h_t, \mathbf{M}_{t})}{\sum_{j} \exp{(\alpha_{t}^{j})}C(h_t, \mathbf{M}_{t})}
\end{aligned}
\label{memory_read}
\end{equation}
where $l=1,...L$ denotes the number of memory slots, $m_{t}\subseteq\mathbb{R}^{1 \times D}$ is the memory representation with hidden size $D$. $\alpha$ is the strength vector that can be learned through the reading operation and $C(\dots)$ is the cosine similarity measure. Memory reading operation is built upon the idea that not all records in the sequence contribute equally to the current health state of the patient. Hence, the weights are computed using the softmax function based on the cosine similarity of the current hidden states and all the previous memories.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{pics/memroy_network.png}
\caption{
Overview of the proposed memory network. Hidden states are first written into the memory cells and read by the clustering network to produce a comprehensive representation.
}
\label{memory}
\end{figure}
\paragraph{Memory Writing}
Memory writing stores latent representation into memory slots. We use a fixed number of slots to denote the overall memory size. The dimension of the continuous space for each memory slot is $d$ and we use $D$ to denote the dimension of hidden representation $z_{t}$. The hidden state is non-linearly projected into the memory space using a $d\times D$ matrix \textbf{A}, $m_{t} = Az_{t}$, where $m_{t}$ is the new input memory representation. Memory writing aims to filter out non-related information and stores only personalized information based on the current hidden state. Mathematically, memory writing can be expressed as:
\begin{equation}
\begin{aligned}
\mathbf{M}_{t} = r_{t}\mathbf{M}_{t-1} + v_{t}h_{t},
\end{aligned}
\label{memory_write}
\end{equation}
where $r_{t}$ and $v_{t}$ are gated vectors that control the information flow between the previous and current memory vector.
\subsubsection{Clustering Network}
After obtaining the representation of the observation $h_{t}$ through the encoding network, i.e the prior network, and updating the memory cell $m_{t}$ at current timestamp $t$, we follow the traditional framework of variational autoencoder (VAE) \cite{lopez2017conditional} to compute the mean and standard deviation vectors through the posterior network. We assume that the output is a Gaussian distribution. The computation process can be expressed as:
\begin{equation}
\begin{aligned}
& \mathbf{\mu}_{z_{t}} = f_{posterior}([h_{t}, x_{i}]) \\
& \mathbf{\sigma}_{z_{t}} = f_{posterior([h_{t}, x_{i}])}
\end{aligned}
\label{vae}
\end{equation}
where $h_{t}$ is the hidden state and $x_{i}$ is the observation at timestamp $t$. $f_{posterior}$ is posterior functions described by feed-forward neural networks. We then draw samples from the posterior Gaussian distribution using the reparameterization trick:
\begin{equation}
z_{t} = \mu_{z_{t}} + \sigma_{z_{t}}\odot\epsilon,
\end{equation}
where $\epsilon \sim \mathcal{N}(0, 1)$, and $z_{t}$ is the latent representation. $\odot$ indicates element-wise multiplication. The reparameterization trick allows the gradient to backpropagate through the sampling process. Lastly, depends on the availability of diagnosis labels, the clustering network will be trained on two different objectives. When diagnosis label is used, the clustering network is directly trained to predict the label information:
\begin{equation}
\hat{Y_{t}} = f_{pred}(f_{con}([z_{t}, m_{t}])),
\end{equation}
where $f_{pred}$ is a feed-forward network that outputs probabilities of each label. When diagnosis label is not available, we trained the framework to reconstruct the observation $x_{i}$ from the latent variable $z_{t}$ conditioned on the memory state $m_{t}$, denote as:
\begin{equation}
\hat{X_{t}} = f_{recon}(f_{con}([z_{t}, m_{t}])),
\label{recon}
\end{equation}
where $\hat{X_{t}}$ is the reconstructed input, $f{recon}$ is a feed forward network and $f_{con}$ is the concatenation. During cluster phase, we use euclidean distance-based k-means algorithm on the latent variable $z_{t}$.
\begin{algorithm}[t]
\caption{TC-EMNet}
\label{algo}
\begin{algorithmic}[1]
\item Initialize encoder and decoder network parameters $\theta, \sigma$;
\item Initialize memory embedding and memory slots;
\For {(every time stamp $t$)}
\State Compute patient hidden encoding through Encoder
\State network via Eq. (\ref{encoding});
\State Read from global-level memory network to extract
\State recent memory representations via Eq. (\ref{memory_read});
\If{diagnosis is available}
\State Read from patient-level memory network
\State to extract recent memory representations
\State via Eq. (\ref{memory_read});
\State Compute memory representation via Eq. (\ref{cali});
\EndIf
\State Compute loss via Eq. (\ref{vae}) - (\ref{recon});
\State Write to corresponding memory slots via Eq. (\ref{memory_write});
\EndFor
\item Update parameters by optimizing Eq. (\ref{objective1}), (\ref{objective2})\ accordingly;
\end{algorithmic}
\end{algorithm}
\subsubsection{Dual Memory Network Architecture}
Under a clinical setting, doctors often provide diagnosis labels based on patients' current and past medical events. Such information can be target health conditions or a diagnosis. Under supervised setting when the label is available during training, we further utilize a patient-level memory network to capture diagnosis information during each visit. Compared to global-level memory network, patient-level memory network at current memory slot $M_{t}$ can only access diagnosis up to previous timestamp, namely, $\{Y_{i} | i = 1,2,...,t-1\}$. patient-level memory network only reads and writes diagnosis information which later is combined with a global-memory network for clustering. We propose a calibration process to integrate representations from two memory networks, as follows:
\begin{equation}
\begin{aligned}
& h_{t}^{global} = M_{global}(X; \alpha_{t}^{global}, m_{t}^{global}, h_{t}^{global}), \\
& h_{t}^{patient} = M_{patient}(Y_{<t}; \alpha_{t}^{patient}, m_{t}^{patient}, h_{t}^{patient}), \\
& h_{t}^{final} = h_{t}^{global} \odot \sigma(f_{embed}(h_{t}^{patient} )),
\end{aligned}
\label{cali}
\end{equation}
where $M_{global}$ and $M_{patient}$ is the global-level and patient-level memory network respectively. This memory calibration process can be regarded as a point-wise attention mechanism.
\subsection{Objective Function and Optimization}
Here, we present our training objectives and optimization process. As mentioned in previous sections, the entire network can be trained from end to end using maximum likelihood estimation (MLE). To solve the intractable marginalization for the latent variable $z_{i}$, we use the variational lower bound parameterized by $q_{\phi}$ to approximate the true distribution, which we assume to be Gaussian. After the memory work reading and writing, We use the latent variable $z_{i}$ at timestamp $i$ to identify the disease stages. Here we restrict the latent variable to be a multivariate Gaussian distribution, which enforces the same for the posterior. We learn the generative parameter $\theta$ using maximum likelihood estimation (MLE):
\begin{equation}
\begin{aligned}
\theta = \operatorname*{argmin}_\theta \sum_{i} \log \int [ p_{\theta}(x_{i} | z_{i}, X_{<i}) \\
\times p_{\theta}(z_{i} | X_{i}) ] \,dz_{i}
\end{aligned}
\end{equation}
However, the marginalization of $z_{i}$ is intractable for complicated functions (for instance neural networks). Thus, we need to derive a variational lower bound (i.e. variational Bayesian method) to approximate the logarithm of the marginal probability of the observation, which is as follows:
\begin{equation}
\begin{aligned}
& \log p_{\theta} (x_{i} | z_{i}, X_{<i}) \\
& = \log (\mathbf{E}_{q\phi} [\frac{p_{\theta}(x_{i}, z_{i} | X_{<i})}{q_{\phi}(z_{i} | x_{i}, X_{<i})])} \\
& \geq \mathbf{E}_{q\phi} [\log p_{\theta}(x_{i}, z_{i} | X_{<i})] \\
& - \mathbf{E}_{q\phi} [\log q_{\phi}(z_{i} | x_{i}, X_{<i})],
\end{aligned}
\end{equation}
where the inequality can be obtained using Jensen's inequality and the variational lower bound involves the probability $q_{\phi}$ that are parameterized by $\phi$, which eventually approximate the intractable true posterior distribution $ p_{\theta}(z_{i} | X_{i})$. Since health-related data is often associated with high-dimensional and general more complicated distribution, we introduce the latent variable $z_{t}$ to capture the internal stochasticity from the data. We can train the entire clustering network end-to-end using stochastic optimization techniques. After obtaining the variational lower bound, the optimization follows the KL divergence that is the difference of log-likelihood and the variational lower bound:
\begin{equation}
\begin{aligned}
& \mathcal{L}_{variation}(\theta, \phi) \\
&= -KL[q_{\phi}(z_{t}|h_{t}, X_{<i}) || p_{\theta}(z_{t}|h_{t}, X_{<i})] \\
& = \textbf{E}_{\log q_{\phi}(z_{t}|z_{t}, X_{<i})}
- \textbf{E}_{\log p_{\theta}(z_{i}, x_{i}|X_{<i})}
\end{aligned}
\label{objective3}
\end{equation}
where $\phi$ and $\theta$ represents the model parameter and proxy posterior accordingly. The equation holds if the distribution of $q_{\phi}$ is equal to the true distribution. When diagnosis label is used during training, we use the cross-entropy loss to directly predict the outcome from the combined latent representation denoted as:
\begin{equation}
\begin{aligned}
& \mathcal{L}_{objective}(\theta,\phi,x_{t}) = \\
& \alpha\mathcal{L}_{variation}(\theta,\phi) + \mathcal{L}_{CE}(Y, \hat{Y}; \theta,\phi,x_{t}),
\end{aligned}
\label{objective1}
\end{equation}
When the model is trained in a unsupervised manner, the overall objective function combined with the reconstruction loss becomes:
\begin{equation}
\begin{aligned}
& \mathcal{L}_{objective}(\theta,\phi,x_{t}) = \\
& \alpha\mathcal{L}_{variation}(\theta,\phi) + \mathcal{L}_{recon}(X, \hat{X};\theta,\phi,x_{t}),
\end{aligned}
\label{objective2}
\end{equation}
where we use the mean square error (MSE) for reconstruction loss and $\alpha$ is a hyperparameter to prevent VAE from KL vanishing problem. We adopt a linear annealing schedule for $\alpha$ based on training steps denoted as:
\begin{equation}
\begin{aligned}
\alpha = \min(1, \frac{\textit{training step}}{x}),
\end{aligned}
\label{objective3}
\end{equation}
where $x$ is a threshold value. Last but not least, we use the k-means algorithm \cite{wagstaff2001constrained} on the patient representation to perform clustering.
\begin{table*}[!ht]
\centering
\caption{Results of proposed methods and other methods on ADNI datasets. $\downarrow$ indicates that the smaller the better (0=best, and 1=worst). $\uparrow$ indicates that the greater the better (0=worst, and 1=best).}
\label{ADNI_results}
\begin{tabular}{lcccccc}
\hline
\multicolumn{1}{l}{} & \multicolumn{3}{c}{w/o label} & \multicolumn{3}{c}{with label} \\
\hline
Model & Purity $\uparrow$ & NMI $\uparrow$ & RI $\uparrow$ & Purity $\uparrow$ & NMI $\uparrow$ & RI $\uparrow$\\
\hline
RNN & 0.6799$\pm$0.00 & 0.1415$\pm$0.01 & 0.1406$\pm$0.02 & 0.8532$\pm$0.00 & 0.4020$\pm$0.01 & 0.3805$\pm$0.01\\
Bi-LSTM & 0.6810$\pm$0.02 & 0.1540$\pm$0.02 & 0.1559$\pm$0.02 & 0.8674$\pm$0.00 & 0.4092$\pm$0.01 & 0.4042$\pm$0.02\\
RETAIN & 0.6903$\pm$0.02 & 0.1787$\pm$0.01 & 0.1671$\pm$0.01 & 0.7144$\pm$0.02 & 0.2572$\pm$0.01 & 0.1838$\pm$0.03\\
Dipole & 0.6839$\pm$0.00 & 0.1707$\pm$0.01 & 0.1452$\pm$0.00 & 0.8904$\pm$0.01 & 0.4674$\pm$0.01 & 0.4776$\pm$0.02\\
StageNet & 0.6943$\pm$0.01 & \bfseries{0.2002}$\pm$0.01 & 0.1791$\pm$0.01 & 0.8513$\pm$0.01 & 0.4045$\pm$0.03 & 0.3744$\pm$0.01\\
AC-TPC & - & - & - & 0.8214$\pm$0.03 & 0.3362$\pm$0.07 & 0.3827$\pm$0.09\\
\hline
VAE & 0.6651$\pm$0.02 & 0.1023$\pm$0.02 & 0.1117$\pm$0.02 & 0.6495$\pm$0.04 & 0.1718$\pm$0.05 & 0.1042$\pm$0.04\\
Memory Network & 0.6887$\pm$0.02 & 0.1392$\pm$0.01 & 0.1584$\pm$0.02 & 0.8262$\pm$0.01 & 0.3603$\pm$0.01 & 0.3538$\pm$0.02\\
$\text{TC-EMNet}^{-u}$ & \bfseries{0.7040}$\pm$0.01 & 0.1967$\pm$0.02 & \bfseries{0.1891}$\pm$0.02 & 0.8904$\pm$0.00 & 0.4679$\pm$0.01 & 0.4889$\pm$0.01\\
$\text{TC-EMNet}^{-s}$ & - & - & - & \bfseries{0.9126}$\pm$0.01 & \bfseries{0.4789}$\pm$0.01 & \bfseries{0.4923}$\pm$0.02\\
\hline
\end{tabular}
\end{table*}
\begin{table*}[!h]
\centering
\caption{Results of proposed methods and other methods on PPMI datasets. $\downarrow$ indicates that the smaller the better (0=best, and 1=worst). $\uparrow$ indicates that the greater the better (0=worst, and 1=best).}
\label{PPMI_results}
\begin{tabular}{lcccccc}
\hline
\multicolumn{1}{l}{} & \multicolumn{3}{c}{w/o label} & \multicolumn{3}{c}{with label} \\
\hline
Model & Purity $\uparrow$ & NMI $\uparrow$ & RI $\uparrow$ & Purity $\uparrow$ & NMI $\uparrow$ & RI $\uparrow$\\
\hline
RNN & 0.7221$\pm$0.00 & 0.3089$\pm$0.01 & 0.3120$\pm$0.01 & 0.7640$\pm$0.02 & 0.4222$\pm$0.04 & 0.3663$\pm$0.03\\
Bi-LSTM & 0.7264$\pm$0.00 & 0.3170$\pm$0.00 & 0.2976$\pm$0.01 & 0.7674$\pm$0.03 & 0.4456$\pm$0.05 & 0.3575$\pm$0.05\\
RETAIN & 0.5241$\pm$0.02 & 0.1188$\pm$0.01 & 0.0619$\pm$0.01 & 0.7510$\pm$0.01 & 0.4072$\pm$0.03 & 0.3361$\pm$0.01\\
Dipole & 0.7233$\pm$0.00 & 0.3200$\pm$0.00 & 0.3153$\pm$0.00 & 0.8033$\pm$0.01 & 0.4947$\pm$0.01 & 0.4476$\pm$0.02\\
StageNet & 0.7252$\pm$0.01 & 0.3305$\pm$0.00 & 0.3234$\pm$0.01 & 0.7839$\pm$0.01 & 0.4700$\pm$0.03 & 0.3840$\pm$0.01\\
AC-TPC & - & - & - & 0.8151$\pm$0.01 & 0.4984$\pm$0.03 & \bfseries{0.5129}$\pm$0.01\\
\hline
VAE & 0.7161$\pm$0.00 & 0.3576$\pm$0.01 & 0.3153$\pm$0.00 & 0.7942$\pm$0.01 & 0.4452$\pm$0.00 & 0.3782$\pm$0.01\\
Memory Network & 0.6996$\pm$0.01 & 0.2809$\pm$0.01 & 0.2581 $\pm$0.02 & 0.7689$\pm$0.01 & 0.4482$\pm$0.01 & 0.4597$\pm$0.01\\
$\text{TC-EMNet}^{-u}$ & \bfseries{0.7452}$\pm$0.00 & \bfseries{0.3773}$\pm$0.00 & \bfseries{0.3742}$\pm$0.01 & 0.8256$\pm$0.00 & \bfseries{0.5053}$\pm$0.00 & 0.4823$\pm$0.01\\
$\text{TC-EMNet}^{-s}$ & - & - & - & \bfseries{0.8339}$\pm$0.00 & 0.5035$\pm$0.00 & 0.4993$\pm$0.01\\
\hline
\end{tabular}
\end{table*}
\section{Experiments}
We evaluated our proposed model on two real-world datasets, Alzheimer's Disease Neuroimaging Initiative (ADNI) and Parkinson's Progression Markers Initiative (PPMI) dataset. All dataset can be accessed on IDA website\footnote[1]{\url{https://ida.loni.usc.edu/}}. The code can be found on GitHub\footnote[2]{\url{https://github.com/Ericzhang1/TC-EMNet.git}}.
\subsection{Datasets}
\subsubsection{ADNI Dataset}
Alzheimer's disease (AD) is a chronic neurodegenerative disease that is often related to behavior and cognitive impairment. ADNI is a longitudinal study that aims to explore early detection and tracking of AD based on imaging, biomarkers, and genetic data collected throughout the process \cite{jack2008alzheimer}. The dataset consists of a total of 11651 visits over 1346 patients with 6 months intervals. For each patient, 21 variables are collected and processed, including 16 time-varying features (brain function, cognitive tests) and 5 static features (background, demographics). 3 diagnose labels are assigned by doctors at each visit for the patient, including control normal (CN), Mild Cognitive Impairment (MCI), and AD, which indicates the severity of how AD symptoms have progressed on each patient.
\subsubsection{PPMI Dataset}
Parkinson's Progression Markers Initiative (PPMI) is a longitudinal study aiming to evaluate patients' progression on Parkinson's disease (PD) based on biomarkers \cite{marek2011parkinson}. The dataset consists of a total of 13685 visits over 2145 patients with irregular time intervals. For each patient, 79 features based on motor and non-motor symptoms are collected, including cognitive assessment, lab tests, demographic information, and biospecimens. Since the dataset does not provide a diagnosis label per visit for each patient, we use Hoehn and Yahr (HY) scores as labels for our evaluation. HY scores, ranges from 0 to 5, indicate the severity of patients' symptoms of Parkinson's disease. We use the mean and last occurrence carried forward method to impute missing values.
\subsection{Baselines}
We compare our proposed model to several state-of-the-art methods, ranged from vanilla RNNs to multi-layer attention models. Since here we consider disease progression modeling under both supervised and unsupervised settings, we adjusted the architecture of the baseline models to fit the objective accordingly. For baselines that cannot be modified interchangeably, we did not collect the result under the corresponding setting. For all experiments, we use k-means clustering on the hidden representations from the last layer to report the clustering performance.
\begin{itemize}
\item \textbf{RNN} \cite{mikolov2010recurrent}: A single RNN cell with an additional layer of feed-forward neural network. The model is trained with cross-entropy loss and reconstruction objective accordingly.
\item \textbf{Bi-LSTM} \cite{huang2015bidirectional}: Similar to RNN model, a Bi-directional LSTM is used with a reconstruction objective, the model takes both directions of the sequence data into account and is showed to capture richer information compare to single direction.
\item \textbf{RETAIN} \cite{choi2016retain}: An interpretable deep learning model that is based on recurrent neural network and reverse time attention mechanism. The RETAIN model learns the importance of hospital records through attention weights. We modify the last layer of RETAIN and train the model based on the prediction and reconstruction objective.
\item \textbf{Dipole} \cite{ma2017dipole}: A interpretable bidirectional recurrent neural network that employs attention mechanism to leverage both past and future visits. We use concatenation-based attention mechanism for testing and, similar to RETAIN, we adjust the last layer of the model accordingly.
\item \textbf{StageNet} \cite{gao2020stagenet}: A recent risk prediction model that learned to extract disease progression patterns during training and leveraged modified LSTM cell with an attention mechanism. The progression pattern at each timestamp is re-calibrated accordingly using a convolution network.
\item \textbf{AC-TPC} \cite{lee2020temporal}: A recent deep predictive clustering network that consists of an encoder network, selector, and a predictor. The model is first initialized using a prediction objective and then optimized to train a cluster embedding using the actor-critic algorithm. This method cannot be trained without label information.
\item \textbf{VAE} \cite{kusner2017grammar}: A vanilla variational autoencoder model using a LSTM cell as encoder and trained with prediction and variation objective respectively. Note that this baseline method can be served as an ablation example against our proposed method.
\item \textbf{Memory Network}: A vanilla global-level memory network with reading and writing mechanism described previously. The network reads and writes the EHR sequence directly and the k-means algorithm is applied directly to the hidden memory representation.
\item \textbf{$\text{TC-EMNet}^{-u}$}: Unsupervised version of TC-EMNet. When the training label is not available, only a global-level memory network is used to produce memory representation. We also train the model for the prediction task and set it as an ablation example against supervised version of TC-EMNet.
\item \textbf{$\text{TC-EMNet}^{-s}$}: Supervised version of TC-EMNet. When the training label is available, a patient-level memory network is used to combine with a global-level memory network to produce target-aware memory representations.
\end{itemize}
\begin{table}[h]
\centering
\caption{Hyperparameter Searching Space}
\label{hyper}
\begin{tabular}{cc}
\hline
Hyperparameter & Range\\
\hline
hidden size & $[32,64,128,256]$\\
latent variable size & $[32,64,128,256]$\\
x & $[500, 700, 900]$ \\
learning rate & $[1e-2,1e-3,1e-4,1e-5]$\\
batch size & $[64, 128]$\\
\hline
\end{tabular}
\end{table}
\begin{figure*}
\centering
\subcaptionbox{Bilstm}{\includegraphics[width=0.24\textwidth]{pics/cluster/bilstm_0.png}}
\hfill
\subcaptionbox{StageNet}{\includegraphics[width=0.24\textwidth]{pics/cluster/stage_0.png}}%
\hfill
\subcaptionbox{Dipole}{\includegraphics[width=0.24\textwidth]{pics/cluster/dipole_0.png}}%
\hfill
\subcaptionbox{Ours}{\includegraphics[width=0.24\textwidth]{pics/cluster/ours_0.png}}%
\hfill
\subcaptionbox{Bilstm}{\includegraphics[width=0.24\textwidth]{pics/cluster/bilstm.png}}
\hfill
\subcaptionbox{StageNet}{\includegraphics[width=0.24\textwidth]{pics/cluster/stage.png}}%
\hfill
\subcaptionbox{Dipole}{\includegraphics[width=0.24\textwidth]{pics/cluster/dipole.png}}%
\hfill
\subcaptionbox{Ours}{\includegraphics[width=0.24\textwidth]{pics/cluster/ours.png}}%
\hfill
\caption{Visualization of the clusters for ADNI (first row) and PPMI (second row) using PCA: Bilstm (1st column), StageNet (2nd column),
Dipole (3rd column), Ours (4th) column). }
\label{result_compare}
\end{figure*}
\begin{table*}[t]
\centering
\caption{Most significant features in each cluster measured by first order gradient for ADNI and PPMI dataset.}
\begin{tabular}{ccccccccc}
\multicolumn{9}{c}{ADNI Dataset}\\
\hline
& \multicolumn{8}{c}{Features}\\
\hline
Cluster \RomanNumeralCaps {1}& \multicolumn{2}{c}{RAVLT\_learning} & \multicolumn{2}{c}{Ventricles} & \multicolumn{2}{c}{WholeBrain} & \multicolumn{2}{c}{ICV} \\
& \multicolumn{2}{c}{RAVLT\_perc\_forgetting} & \multicolumn{2}{c}{RAVLT\_forgetting} & \multicolumn{2}{c}{ADAS13} & \multicolumn{2}{c}{RAVLT\_immediate} \\
\hline
Cluster \RomanNumeralCaps{2} & \multicolumn{2}{c}{ICV} & \multicolumn{2}{c}{RAVLT\_perc\_forgetting} & \multicolumn{2}{c}{ADAS13} & \multicolumn{2}{c}{Ventricles} \\
& \multicolumn{2}{c}{serial} & \multicolumn{2}{c}{RAVLT\_immediate} & \multicolumn{2}{c}{CDRSB} \\
\hline
Cluster \RomanNumeralCaps{3} & \multicolumn{2}{c}{RAVLT\_perc\_forgetting} & \multicolumn{2}{c}{serial
} & \multicolumn{2}{c}{ICV} & \multicolumn{2}{c}{RAVLT\_learning}\\
& \multicolumn{2}{c}{Entorhinal} & \multicolumn{2}{c}{Hippocampus} & \multicolumn{2}{c}{Ventricles} & \multicolumn{2}{c}{WholeBrain}\\
\hline
\multicolumn{9}{c}{}\\
\multicolumn{9}{c}{PPMI Dataset}\\
\hline
& \multicolumn{8}{c}{Features}\\
\hline
Cluster \RomanNumeralCaps {1}& \multicolumn{2}{c}{Global Spontaneity of Movement} & \multicolumn{2}{c}{Speech} & \multicolumn{2}{c}{Anxious Mood} & \multicolumn{2}{c}{Arising from Chair} \\
& \multicolumn{2}{c}{Right leg} & \multicolumn{2}{c}{Getting Out of Bed} & \multicolumn{2}{c}{Pronation-Supination (left)} & \multicolumn{2}{c}{} \\
\hline
Cluster \RomanNumeralCaps{2} & \multicolumn{2}{c}{Posture} & \multicolumn{2}{c}{Rest tremor amplitude} & \multicolumn{2}{c}{Dopamine} & \multicolumn{2}{c}{Rigidity} \\
& \multicolumn{2}{c}{Saliva + Drooling} & \multicolumn{2}{c}{Anxious Mood} & \multicolumn{2}{c}{Global Spontaneity of Movement} \\
\hline
Cluster \RomanNumeralCaps{3} & \multicolumn{2}{c}{Postural Stability} & \multicolumn{2}{c}{Cognitive Impairment
} & \multicolumn{2}{c}{Rest Tremor Amplitude} & \multicolumn{2}{c}{Pronation-Supination (left)}\\
& \multicolumn{2}{c}{Dopamine} & \multicolumn{2}{c}{Standing} & \multicolumn{2}{c}{Rigidity} & \multicolumn{2}{c}{}\\
\hline
Cluster \RomanNumeralCaps {4}& \multicolumn{2}{c}{Pronation-Supination (left)} & \multicolumn{2}{c}{Standing} & \multicolumn{2}{c}{Postural Stability} & \multicolumn{2}{c}{Chewing} \\
& \multicolumn{2}{c}{Cognitive Impairment} & \multicolumn{2}{c}{Dopamine} & \multicolumn{2}{c}{Right Hand} & \multicolumn{2}{c}{} \\
\hline
Cluster \RomanNumeralCaps{5} & \multicolumn{2}{c}{Dopamine} & \multicolumn{2}{c}{Cognitive Impairment} & \multicolumn{2}{c}{ Hallucinations} & \multicolumn{2}{c}{Chewing} \\
& \multicolumn{2}{c}{Dressing} & \multicolumn{2}{c}{Pronation-Supination (left)} & \multicolumn{2}{c}{Arising from Chair} \\
\hline
Cluster \RomanNumeralCaps{6} & \multicolumn{2}{c}{Rigidity} & \multicolumn{2}{c}{Serial
} & \multicolumn{2}{c}{Rigidity} & \multicolumn{2}{c}{Standing}\\
& \multicolumn{2}{c}{Apathy} & \multicolumn{2}{c}{Constipation Problems} & \multicolumn{2}{c}{Cognitive Impairment} & \multicolumn{2}{c}{Dopamine}\\
\hline
\end{tabular}
\label{features}
\end{table*}
\begin{table}[h]
\centering
\caption{Complexity comparison between models}
\begin{tabular}{ cc }
\hline
Model & \# of trainable parameters \\
\hline
Dipole & 279k \\
StageNet & 283k \\
AC-TPC & 143k \\
$\text{TC-EMNet}^{-u}$ & 163k \\
$\text{TC-EMNet}^{-s}$ & 174k \\
\hline
\end{tabular}
\label{complexity}
\end{table}
\subsection{Model Training and Implementation Details}
As mentioned previously, our proposed network is continuous and differentiable. We can train the network using stochastic optimization techniques. All neural networks in the proposed network are feed-forward networks. We implemented our solution using Pytorch \cite{paszke2019pytorch} and trained the model on a single Nvidia Volta V100 GPU with 16GB memory. We adopt gradient accumulation when dealing with out-of-memory problems. We select hyperparameters through random search as shown in table \ref{hyper}.
For our model, we set both hidden size and latent variable size to be 128. We adopt Adam optimizer with a learning rate of $1e-3$. The model is trained with batch size $32$ for $70$ epochs. $x$ is set to $700$. We split the dataset into training, validation, and testing set with a ratio of $3/1/1$ and report the performance of $5$ fold cross-validation for both datasets.
A detailed description of the optimization process of our proposed framework can be found in Algorithm \ref{algo}. The average running time of our proposed framework on both datasets is about 2 hours for cross-validation.
For the implementation of other baseline methods, we implement RNN and Bi-lstm methods with Pytorch. We adopt implementations from Pyhealth \cite{zhao2021pyhealth} for RETAIN, Dipole, and StageNet. And we adopt implementation from \cite{lee2020temporal} for AC-TPC. All baseline methods share the same hyperparameter searching space.
\subsection{Evaluation Metrics}
To evaluate the clustering performance of our model, we use purity score (purity), normalized mutual information (NMI) \cite{vinh2010information}, and adjusted rand index (ARI) \cite{hubert1985comparing}. Purity score is ranged between $0$ to $1$, indicating the extent to which a cluster is consist of single class. NMI ($0$ to $1$) represents the mutual information between each clusters with $1$ being perfect clustering. ARI derives from the Rand index and measures the percentage of the correct cluster assignment. Mathematically, the metrics can be expressed as follows:
\begin{equation}
\begin{aligned}
& \textbf{Purity} = \frac{1}{N}\sum^{j}\max|c_{j} \cap l_{j}|, \\
& \textbf{NMI} = \frac{2\cdot\textbf{I}(c_{j}, l_{j})}{[H(C)+H(L)]},\\
& \textbf{ARI} = \frac{RI - E(RI)}{\max({RI}) - E(RI)},
\end{aligned}
\end{equation}
where $N$ is the total number of samples, $c_{j}$ and $l_{j}$ denotes the cluster assignment and true label respectively, $I(\cdot)$ is the mutual information function and $H(\cdot)$ is the entropy, $E(\cdot)$ and \textit{RI} are the expectation value and Rand index accordingly.
\section{Results}
\subsection{Clustering Performance}
A quantitative comparison of the clustering performance on ADNI and PPMI dataset is shown in table \ref{ADNI_results} and table \ref{PPMI_results} respectively. We set the cluster assignments to the number of class/diagnosis for each dataset, i.e. $3$ for ADNI (diagnosis label) and $6$ (NHY score) for PPMI. We want the model to identify the individual disease stages both when there is only limited knowledge known to a certain disease, i.e class/diagnosis is not available and when diagnosis label is available, and thus provide insightful and interpretable information to help discover corresponding treatment to individual treatment. We compare our proposed method with the aforementioned baselines in terms of clustering performance. It is clear that our method has demonstrated competitive performance against all baseline methods across all evaluation metrics for both datasets. We note that it is generally difficult to identify clusters without the presence of label information as indicated by low NMI and RI scores. However, TC-EMNet outperforms baseline by a large margin in terms of NMI and RI scores when clustering with label. Training under supervised setting yields significantly better clustering performance compared to training under unsupervised setting. This is due to fact that the correlation between diagnosis and input features is encoded into each hidden representation. Although AC-TPC has better performance in terms of RI on the PPMI dataset. The method relies on pre-training the model with over $1000$ epochs, which could result in the model memorizing the input data. Both Dipole and StageNet have comparable performance. However, it is worth mentioning that both models have leveraged attention over multi-layer RNNs, which introduces additional complexity to the model. A detailed comparison between the trainable parameters is shown in table~\ref{complexity}. Furthermore, we find that when training with label information, RI score can be negatively impacted compared to training without labels. Such phenomena are observed for multiple baseline methods. One explanation could be directly leveraging label information overwhelms the training process since labels possess strong prediction power compared to input features, making the model more biased towards dominated class when dealing with imbalanced datasets; thus, RI may drop as there are more false positives and false negatives. It also can be observed that leveraging external memory effectively captures long-term information and the TC-EMNet is capable of learning complexity from the input data. The patient-level memory network constructively binds with the global-level memory network to produce more comprehensive memory representations.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{pics/feat.png}
\caption{
Significant feature values of cluster centroids on ADNI dataset. The distribution of clusters is very different, which means distinct subtypes.
}
\label{feat1}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{pics/feat2.png}
\caption{
Significant feature values of cluster centroids on PPMI dataset. The distribution of clusters is very different, which means distinct subtypes.
}
\label{feat2}
\end{figure}
\subsection{Disease Stage}
In order to interpret the disease stages and progression patterns found by TC-EMNet. We first selected three baseline models that have comparable performance against TC-EMNet and visualized the hidden representations in $2D$ space using PCA \cite{martinez2001pca}. The results are shown in Fig \ref{result_compare}. We observed that in general most methods can produce distinct clusters for the ADNI dataset. However, for PPMI dataset, most baseline methods failed at producing effective clusters, whereas TC-EMNet produces distinct clustering results. This shows that TC-EMNet is able to constructively model long-term information between each visit in order to find effective representations. Next, we compute feature importance for every cluster based on the weights from the last layer of the network. The results are shown in table \ref{features}. It can be observed that for both datasets, each cluster is determined by a diverse range of features, which means it is easier to identify each patients' progression patterns through observation. We also compute the centroid values for each cluster and plot the distribution in Fig \ref{feat1}, \ref{feat2} for ADNI and PPMI datasets respectively. For ADNI dataset, our proposed model has determined significant features such as RAVLT\_learning, RAVLT\_perc\_forgetting, ICV, ventricles. Rey's Auditory Verbal Learning Test (RAVLT) scores are helpful in testing episodic memories and are very important indicators in identifying a patient's progression in Alzheimer's disease \cite{moradi2017rey}. In particular, the learning test (RAVLT\_learning) and percent forgetting test (RAVLT\_perc\_forgetting) are highly correlated and thus become crucial biomarkers for early detection in AD. It can be observed in Fig \ref{feat1} that three clusters produced by our model have wide distribution for RAVLT testing values, which suggests three different patient subtypes. As for PPMI dataset, our model has found that the dopamine dysregulation syndrome (Dopamine) is a significant feature in identifying clusters. Studies have discovered that under clinical settings early characterization of Dopamine can aid the treatment for motor and non-motor complications for Parkinson's disease \cite{evans2004dopamine}. There are also studies that showed that cognitive impairment (Cognitive impairment) is a strong indicator for Parkinson's disease. Difference in cognitive impairment scores can reflect advanced progression in PD \cite{verbaan2007cognitive}.
\begin{comment}
\subsection{Model Limitations and Future Works}
This work also has some limitations. First of all, the model relies on the completeness of the EHR records. Although dataset used in this study does not require specific data processing techniques. This criteria is hard to met in reality due to the missingness and noise in most time-series data. Given the consideration of missing values, data imputation may introduce undesired bias that can negative affect patient representation. Future work is needed to develop methods that are robust against these problems. Secondly, our proposed model requires per visit diagnosis to produce accurate representations. Such condition is unlikely to be met under real-world clinical setting. Thus, enable model to derive target-aware information directly from input data and leverage other forms of external knowledge could be one of future directions as well.
\end{comment}
\section{Conclusion}
In this paper, we propose TC-EMNet for disease progression modeling on time-series data. TC-EMNet leverages VAE to model data irregularity and an external memory network to capture long-term dependency. We developed TC-EMNet to perform patient clustering/subtyping under both supervised and unsupervised settings. Under supervised setting, TC-EMNet leverages a dual memory network architecture to extract target-aware information from diagnosis to compute patient representations. Throughout the experiment on two real-world datasets, we showed that our model outperforms state-of-the-art methods and is able to identify interpretable disease stages that are clinically meaningful. TC-EMNet yields competitive clustering performance with limited complexity. In the real-world clinical setting, we hope that our model could help physicians identify patients' progression patterns and discover potential disease stages to gain more understanding about chronic and other heterogeneous diseases.
\section*{Acknowledgment}
This paper was funded in part by the National Science Foundation under award number CBET-2037398.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Orlicz cohomology has been studied recently (see \cite{C,GK19,K17,K21,KP2,KP}) as a natural generalization of $L^p$-cohomology. These cohomology theories provide quasi-isometry invariants (and therefore have applications to classification problems, see for example \cite{C,Pa08,S}), and have connections with Sobolev and Poincaré type inequalities (\cite{GK19,GT16}) and harmonic functions (\cite{KP2,P}).
There exist several definitions of $L^p$-cohomology depending on the context (simplicial complexes, Riemannian manifolds, topological groups, general metric spaces), which are related by equivalence theorems (see \cite{BR,G,GKS88,Pa95,S}). The Orlicz case is not so explored yet in this aspect. We can mention, as results obtained in this direction, the equivalence between
Orlicz cohomology of a star-bounded simplicial complex and the cohomology of the corresponding Orlicz-Sullivan complex (\cite{K21}), and the equivalence between the simplicial and de Rham Orlicz cohomology for Riemannian manifolds in the case of degree 1 (\cite{C}). In a previous work (\cite{S}) a generalization of this second result in all degrees was given in the case of Lie groups. Here we present an improvement of that.
In the following theorem $L^\phi H^k(M)$ denotes the $k$-space of $L^\phi$-cohomology of the Riemannian manifold $M$ and $\ell^\phi H^k(X)$ the $k$-space of $\ell^\phi$-cohomology of the simplicial complex $X$, where $\phi$ is a Young function. The reduced versions of de Rham and simplicial Orlicz cohomology are denoted by $L^\phi \overline{H}^k(M)$ and $\ell^\phi \overline{H}^k(X)$. See $\S$\ref{OC} for the precise definitions.
\begin{theorem}\label{main}
Let $M$ be a Riemannian manifold of dimension $n$ that admits a continuous triangulation $X$ such that:
\begin{enumerate}
\item[(a)] there exists a uniform bound for the degree of every vertex; and
\item[(b)] every simplex of $X$ is biLipschitz homeomorphic to the standard one of the same dimension, where the Lipschitz constant does not depend on the simplex.
\end{enumerate}
Then for any Young function $\phi$ and $k=0,\ldots, n$, the topological vector spaces $L^\phi H^k(M)$ and $\ell^\phi H^k(X)$ are isomorphic, and so are the Banach spaces $L^\phi \overline{H}^k(M)$ and $\ell^\phi \overline{H}^k(X)$.
\end{theorem}
The existence of a triangulation as in Theorem \ref{main} can be ensured by imposing certain geometric hypothesis on the manifold $M$, such as bounded curvature and positive injectivity radius (see \cite{A}).
In \cite{C} it is stablished that, if there exists a quasi-isometry between two simplicial complexes with certains properties, then their (reduced) $\ell^\phi$- cohomology are isomorphic. Using this result we can directly deduce the following:
\begin{corollary}
Let $M_1$ and $M_2$ be two uniformly contractible Riemannian manifolds admiting triangulations as in Theorem \ref{main}. If they are quasi-isometric, then for every Young function $\phi$ and $k=0,\ldots,n$ the topological vector spaces $L^\phi H^k(M_1)$ and $L^\phi H^k(M_2)$ are isomorphic, and so are the Banach spaces $L^\phi \overline{H}^k(M_1)$ and $L^\phi \overline{H}^k(M_2)$.
\end{corollary}
A metric space $X$ is \textit{uniformly contractible} if there exists an increasing function $\varphi:[0,+\infty)\to [0,+\infty)$ such that for every $x\in X$ and $r>0$ the ball $B(x,r)=\{y\in X:|x-y|< r\}$ is contractible in $B(x,\varphi(r))$.
\section{Preliminaries}
\subsection{Cochain complexes and bicomplexes}
A \textit{cochain complex} (or simply \textit{complex} from here on) will be a sequence of vector topological spaces $C^*=\left\{C^k\right\}_{k\in{\mathbb{N}}}$ together with a sequence of continuous linear maps $d=d_k:C^k\to C^{k+1}$ such that $d\circ d=0$. We denote it by $(C^*,d_*)$ or $(C^*,d)$ and say that $d$ is the \textit{derivative} of the complex.
The \textit{cohomology} of the complex $(C^*,d)$ is the family of vector topological spaces
$$H^k=\frac{\operatorname{Ker} d_k}{\operatorname{Im} d_{k-1}},$$
where $d_{-1}=0$.
If $(C^*,d)$ is a complex of Banach spaces, we also consider its reduced cohomology by the family of Banach spaces (or more generally, Frechét spaces)
$$\overline{H}^k=\frac{\operatorname{Ker} d_k}{\overline{\operatorname{Im} d_{k-1}}}.$$
A \textit{cochain map} between two cochain complexes $(C^*,d_*)$ and $(D^*,\delta_*)$ is a family of continuous linear maps $f_k:C_k\to D_k$ that commutes with the derivatives, i.e. $\delta_k\circ f_k=f_{k+1}\circ d_k$ for every $k$. A cochain map naturally defines a continuous linear map between the corresponding (relative) cohomology spaces.
Two cochain maps $f_*,g_*:(C^*,d_*)\to (D^*,\delta_*)$ are \textit{homotopic} if there exists a family of continuous linear maps $p_k:C^k\to D^{k-1}$ such that
\begin{equation}
\left\{\begin{array}{cc}
p_{k+1}\circ d_k+\delta_{k-1}\circ p_k= f_k - g_k &\text{ if }k\geq 1 \\
p_1\circ d_0 = f_0-g_0 &.
\end{array}\right.
\end{equation}
This implies that $f_*$ and $g_*$ define the same map in cohomology. We say that two complexes $(C^*,d_*)$ and $(D^*,\delta_*)$ are \textit{homotopically equivalent} if there exist cochain maps $f_*:C^*\to D^*$, $g_*:D^*\to C^*$ such that $f_*\circ g_*$ and $g_*\circ f_*$ are homotopic to the identity. If $D^*\subset C^*$, $d_*=\delta_*$ and $g_*$ is the inclusion we said that $(C^*,d_*)$ retracts to $(D^*,\delta_*)$.
If two complexes are homotopically equivalent, then their cohomologies are isomorphic (in the sense of vector topological spaces). If they are in addition complexes of Banach (or Fréchet) spaces, then their reduced cohomologies are also isomorphic.
\medskip
The method we will use to prove Theorem \ref{main} is a bicomplex argument as the used, for example, in \cite[Theorem 8.9]{BT}, \cite{G} and \cite{Pa95}.
By a \textit{bicomplex} we mean a family of topological vector spaces $\{C^{k,m}\}_{k,m\in{\mathbb{N}}}$ together with linear continuous maps $d':C^{k,m}\to C^{k+1,m}$ and $d'':C^{k,m}\to C^{k,m+1}$ sucht that $d'\circ d'=0$, $d''\circ d''=0$ and $d'\circ d''+d''\circ d'=0$. It will be denoted by $(C^{*,*},d',d'')$. The \textit{rows} and \textit{colums} of the bicomplex are the complexes of the form $(C^{*,m},d')$ and $(C^{k,*},d'')$ respectively.
Observe that, if $(C^{*,*},d',d'')$ is a bicomplex, then one can define the \textit{kernels horizontal} and \textit{vertical} complexes $(E^*,d')$ and $(F^*,d'')$, where $E^k=\operatorname{Ker} d''|_{C^{k,0}}$ and $F^m=\operatorname{Ker} d'|_{C^{0,m}}$. Observe that the condition $d'\circ d''+d''\circ d'=0$ implies that they are cochain complexes.
And important fact about bicomplexes that we will use in the proof of Theorem \ref{main} is the following result, whose proof can be found in \cite{Pa95,G,S}.
\begin{theorem}\label{bicomplejo}
Let $(C^{*,*},d',d'')$ a bicomplex. Suppose that every row $(C^{*,m},d')$ retracts to $(F^0\to 0\to 0 \to \cdots)$ and every colum $(C^{k,*},d'')$ retracts to $(E^0\to 0\to 0\to \cdots)$. Then the kernels horizontal and vertical complexes of the bicomplex are homotopically equivalent.
\end{theorem}
\subsection{Young functions and Orlicz spaces}
By a \textit{Young function} we mean a pair and convex function $\phi:{\mathbb{R}}\to [0,+\infty)$ that satisfy $\phi(t)=0$ if, and only if, $t=0$.
If $(Z,\mu)$ is a measure space and $f:Z\to{\mathbb{R}}$ is a measurable function, we define
$$\rho_\phi(f)=\int_Z \phi\left(f(x)\right)d\mu(x).$$
The Orlicz space of $(Z,\mu)$ associated to $\phi$ is the Banach space $L^\phi(Z)=L^\phi(Z,\mu)$ of classes of functions $f:Z\to{\mathbb{R}}$ for which there exists a constant $\alpha>0$ such that
$\rho_\phi(f/\alpha)<+\infty$, equipped with the Luxemburg norm
$$\|f\|_{L^\phi}=\inf\left\{\alpha>0 : \rho_\phi\left(\frac{f}{\alpha}\right)d\mu\leq 1\right\}.$$
If $Z$ is countable and $\mu$ is the counting measure, we write $L^\phi(Z)=\ell^\phi(Z)$.
Observe that, if we consider the Young function $\phi_p(t):=|t|^p$ with $p\geq 1$, then $L^\phi(Z)=L^p(Z)$. Other examples of Young functions are
$$\phi_{p,\kappa}(t)=\frac{|t|^p}{\log(e+|t|^{-1})},\ (p\geq 1, \kappa\geq 0),$$
which are used in \cite{C} for studying the large scale geometry of Heintze groups. These function have some nice properties: they are $N$-functions with the $\Delta_2$-condition if $(p,\kappa)\neq (1,0)$ (see definitions in \cite[Chapter II]{RR}). An example of Young function that does not satisfy the previous conditions can be $\phi(t)=e^{|t|}-1$.
\begin{remark}\label{ObsEqivalencia}
If $\lambda\geq 1$ is any constant, then $\rho_{\lambda\phi}(f)=\lambda\rho_{\phi}(f)\geq \rho_\phi(f)$, which implies $\|f\|_{L^\phi}\leq \|f\|_{L^{\lambda\phi}}$. Furthermore, the convexity of $\phi$ implies $\rho_{\lambda\phi}(f)\leq \rho_\phi(\lambda f)$, hence $\|f\|_{L^{\lambda\phi}}\leq \lambda\|f\|_{L^\phi}$. We conclude that for every $\lambda>0$ the spaces $L^\phi(Z)$ and $L^{\lambda\phi}(Z)$ are isomorphic, and
$$K^{-1}\|\ \|_{L^\phi}\leq \|\ \|_{L^{\lambda\phi}}\leq K \|\ \|_{L^\phi},$$
where $K=\max\{\lambda,\lambda^{-1}\}$.
\end{remark}
For more details about Orlicz spaces we refer to \cite{RR}.
\subsection{Orlicz cohomology}\label{OC}
We say that a simplicial complex $X$ equipped with a length distance has \textit{bounded geometry} if it has finite dimension and there exist a constant $C>0$ and a function $N:[0,+\infty)\to{\mathbb{N}}$ such that
\begin{enumerate}
\item[(c)] the diameter of every simplex is bounded by $C$; and
\item[(d)] for every $r\geq 0$, the number of simplices contained in a ball of radius $r$ is bounded by $N(r)$.
\end{enumerate}
Observe that the triangulation in Theorem \ref{main} (satisfying conditions (a) and (b)) has bounded geometry.
Denote by $X_k$ the set of $k$-simplices in $X$ and consider the cochain complex $(\ell^\phi(X_*),\delta)$,
with the usual coboundary operator $\delta$ defined by $\delta\theta(\sigma)=\theta(\partial\sigma)$, where if $\Delta=(v_0,\ldots,v_k)\in X_k$,
$$\partial\Delta=\sum_{i=0}^k \partial_i\Delta=\sum_{i=0}^k(v_0,\ldots,\hat{v_i},\ldots,v_k).$$
It is not difficult to see, using bounded geometry, that $\delta:\ell^\phi(X_k)\to \ell^\phi(X_{k+1})$ is welld-defined and continuous for every $k$.
The cohomology of this complex is called the \textit{simplicial Orlicz cohomology} of $X$ associated to the Young function $\phi$ (or, more simply, the $\ell^\phi$\textit{-cohomology} of $X$) and denoted by $\ell^\phi H^*(X)$. Since $(\ell^\phi(X_*),\delta)$ is a complex of Banach spaces, we also consider the \textit{reduced simplicial Orliz cohomology} of $X$ associated to $\phi$ (or \textit{reduced $\ell^\phi$-cohomology} of $X$) as its reduced cohomology. We denote it by $\ell^\phi \overline{H}^*(X)$.
\medskip
Now take a Riemannian manifold $M$ and denote by $\Omega^k(M)$ the space of all (smooth) differential $k$-forms on $M$. Consider
$$L^\phi\Omega^k(M)=\{\omega\in\Omega^k(M) : \|\omega\|_{L^\phi},\|d\omega\|_{L^\phi}<+\infty\},$$
equipped with the norm $|\omega|_{L^\phi}=\|\omega\|_{L^\phi}+\|d\omega\|_{L^\phi}$. Here $d$ denotes the usual exterior derivative and $\|\omega\|_{L^\phi}$ is the Luxemburg norm of the function
$$x\mapsto|\omega|_x=\sup\left\{|\omega_x(u_1,\ldots,u_k)| : u_i\in T_xM \text{ for }i=1,\ldots,k, \text{ with } \|u_i\|_x=1\right\}$$
in the measure space $(M,dV)$, where $\|\ \|_x$ is the Riemannian norm on the tangent space $T_xM$ and $dV$ is the Riemannian volume on $M$.
We can consider the cohomology of the complex $(L^\phi\Omega^k(M),d)$, which we call the \textit{smooth Orlicz-de Rham cohomology} of $M$ associated to $\phi$ (or \textit{smooth $L^\phi$-cohomology} of $M$) and denote it by $L^\phi H^{*}_{s}(M)$.
Since $(L^\phi \Omega^*(M),d)$ is not a complex of Banach we take $(L^\phi C^*(M),d)$ the complex of their completions where $d$ is the continuous extension of the exterior derivative. From this we obtain the \textit{Orlicz-de Rham cohomology} of $M$ associated with $\phi$ (or $L^\phi$\textit{-cohomology} of $M$), denoted by $L^\phi H^*(M)$
and the \textit{reduced Orlicz-de Rham cohomology} of $M$ (or \textit{reduced} $L^\phi$\textit{-cohomology} of $M$), denoted by $L^\phi\overline{H}^*(M)$.
In \cite{KP} it is prove that, under certain hypothesis on $\phi$, the Orlicz-de Rham cohomology is isomorphic to the smooth version. We give (at the end of this work) a proof of that for any Young function under some hypotesis on the manifold $M$.
\begin{remark}\label{ObsDef}
A \textit{measurable $k$-form} on $M$ is a function $x\mapsto\omega_x$, where $\omega_x$ is an alternating $k$-linear form on $T_xM$, such that the coefficients of $\omega$ for every parmetrization of $M$ are all measurable. We denote by $L^\phi(M,\Lambda^k)$ the space of $L^\phi$-integrable measurable $k$-forms up to almost everywhere zero forms. It is a Banach space equipped with the Luxemburg norm $\|\ \|_{L^\phi}$.
Since $L^\phi\Omega^k(M)\subset L^\phi(M,\Lambda^k)$ and the inclusion is continuous, one can prove using Hölder's inequality ($\|fg\|_{L^1}\leq 2\|f\|_{L^\phi}\|g\|_{L^{\phi^*}}$, where $\phi^*$ is complementary Young function of $\phi$, see \cite[Section 3.3]{RR}) that $L^\phi C^k(M)$ can be seen as a space of measurable $k$-forms in $L^\phi(M,\Lambda^k)$ that have weak derivatives in $L^\phi(M,\Lambda^{k+1})$.
We say that a measurable $(k+1)$-form $\varpi$ is the weak derivative of $\omega\in L^\phi(M,\Lambda^k)$ if for every differential $(n-k-1)$-form with compact support $\beta$ one has
$$\int_M \varpi\wedge\beta=(-1)^{k+1}\int_M \omega\wedge d\beta.$$
\end{remark}
Throughout this work we will say $k$-\textit{form} to mean a measurable $k$-form and \textit{differential $k$-form} to mean a smooth one.
\section{Poincare's lemma for Orlicz cohomology}
The following result is previously proved in \cite{GK19}. We present a proof based on Lemma 8 of \cite{Pa95} in order to show a construction that will be used in the proof of Theorem \ref{main}.
\begin{lemma}\label{Poincare}
Let $B=B(0,1)$ be the open unit ball in the Euclidean space ${\mathbb{R}}^n$.
Then both complexes $(L^\phi \Omega^*(B),d)$ and $(L^\phi C^*(B),d)$ retract to the complex $({\mathbb{R}}\to 0\to 0\to \ldots)$.
\end{lemma}
\begin{proof}
For a fixed $x\in B$ we consider $\varphi_x:[0,1]\times B\to B$, $\varphi_x(t,y)=ty+(1-t)x$ and $\eta_t:B\to [0,1]\times B$, $\eta_t(y)=(t,y)$. We denote $\frac{\partial}{\partial t}=(1,0)\in [0,1]\times B$ and define the $\frac{\partial}{\partial t}$-contraction of a $k$-form $\omega$ as the $(k-1)$-form
$$\iota_{\frac{\partial}{\partial t}}\omega_y (u_1,\ldots,u_{k-1})=\omega_y\left(\mathsmaller{\frac{\partial}{\partial t}},u_1,\ldots,u_{k-1}\right).$$
Now we define, for $y\in B$, $\omega\in L^\phi\Omega^k(B)$ and $u_1,\ldots,u_{k-1}\in T_yB={\mathbb{R}}^n$,
\begin{align*}
\chi_x(\omega)_y(u_1,\ldots,u_{k-1})&=\left(\int_0^1 \eta_t^*\left(\iota_{\frac{\partial}{\partial t}}\varphi_x^*\omega\right)dt\right)_y(u_1,\ldots,u_{k-1})\\ &=\int_0^1 \eta_t^*\left(\iota_{\frac{\partial}{\partial t}}\varphi_x^*\omega\right)_y(u_1,\ldots,u_{k-1})dt,
\end{align*}
where $\eta_t^*$ and $\varphi^*$ are the classical pull-back transformations of the respective functions.
Observe that the coefficients of
$\eta_t^*\left(\iota_{\frac{\partial}{\partial t}}\varphi_x^*\omega\right)_y$ are smooth on $t,x$ and $y$. Thus, using the Leibniz integral rule, we can observe that $\chi_x(\omega)$ is a differential $(k-1)$-form.
For every $(k-1)$-simplex $\sigma$ we have
\begin{align*}
\int_\sigma \chi_x(\omega)
= \int_\sigma\int_0^1 \eta_s^*(\iota_{\frac{\partial}{\partial t}}\varphi_x^*\omega)ds
= \int_{[0,1]\times\sigma}\varphi_x^*\omega
= \int_{\varphi_x([0,1]\times\sigma)}\omega
= \int_{C_\sigma}\omega,
\end{align*}
where $C_\sigma$ is the cone centered at $x$, defined as follows: If $\sigma=(x_0,\ldots,x_{k-1})$, then $C_\sigma=(x,x_0,\ldots,x_{k-1})$.
Suppose that $\sigma$ is a $k$-simplex in $B$ with $\partial \sigma=\tau_0+\cdots+\tau_k$ and $\omega\in \Omega^k(B)$. Then, using Stoke's theorem, we have
\begin{align*}
\int_\sigma \chi_x(d\omega) = \int_{C_\sigma} d\omega = \int_{\partial C_\sigma}\omega = \int_\sigma \omega-\sum_{i=0}^k \int_{C_{\tau_i}}\omega =\int_\sigma \omega- \int_{\partial\sigma}\chi_x(\omega)=\int_\sigma \omega- \int_{\sigma}d\chi_x(\omega).
\end{align*}
Since this is true for every $k$-simplex, we conclude that
\begin{equation}\label{dos}
\chi_x d+d\chi_x=Id
\end{equation}
(see for example \cite[Chapter IV]{W}). Observe that if $\omega$ is closed, then $\chi_x(\omega)$ is a primitive of $\omega$, so it is enough to prove the classic Poincaré's lemma. However, in our case we need a $L^\phi$-primitive, so we take a convenient average.
Define
$$h(\omega)=\frac{1}{\operatorname{Vol}\left(\frac{1}{2}B\right)}\int_{\frac{1}{2}B}\chi_x(\omega)dx,$$
where $\frac{1}{2}B=B\left(0,\frac{1}{2}\right)$.
For the same argument as before $h(\omega)$ is smooth and
\begin{equation*}
dh(\omega)=\frac{1}{\operatorname{Vol}\left(\frac{1}{2}B\right)}\int_{\frac{1}{2}B}d\chi_x(\omega)dx.
\end{equation*}
Here is important the fact that we are integrating on a ball whose closure is strictly contained in $B$. Therefore, using (\ref{dos}) we have
\begin{equation}\label{homot}
dh(\omega)+h(d\omega)=\omega
\end{equation}
for all $\omega\in\Omega^k(B)$ with $k\geq 1$.
We have to prove that $h$ is well-defined and continuous from $L^\phi\Omega^k(B)$ to $L^\phi\Omega^{k-1}(B)$. To this end we first bound $|\chi_x(\omega)|_y$ for $y\in B$ and $\omega\in\Omega^k(B)$. Since $\iota_{\frac{\partial}{\partial t}}\varphi^*\omega$ is a form on $[0,1]\times B$ that is zero in the direction of $\frac{\partial}{\partial t}$, we have
$|\eta_t^*(\iota_{\frac{\partial}{\partial t}}\varphi^*\omega)|_y=|\iota_{\frac{\partial}{\partial t}}\varphi^*\omega|_{(t,y)}$ for every $t\in (0,1)$ and $y\in B$. After a direct calculation we get the estimate $$\left|\iota_{\frac{\partial}{\partial t}}\varphi^*\omega\right|_{(t,y)}\leq t^{k-1}|y-x||\omega|_{\varphi(t,y)}.$$
Hence, using the assumption that $t\in (0,1)$, we can write
\begin{equation}\label{uno}
|\chi(\omega)|_y\leq \int_0^1|y-x||\omega|_{\varphi(t,y)}dt.
\end{equation}
Consider the function $u:{\mathbb{R}}^n\to{\mathbb{R}}$ defined by $u(z)=|\omega|_z$ if $z\in B$ and $u(z)=0$ in the other case. Using (\ref{uno}) and the change of variables $z=ty+(1-t)x$ we have
\begin{align*}
\operatorname{Vol}\left(\mathsmaller{\frac{1}{2}}B\right)|h(\omega)|_y
&\leq \int_{B\left(ty,\frac{1-t}{2}\right)}\int_0^1|z-y|u(z)(1-t)^{-n-1}dtdz \\
&= \int_{B(y,2)}|z-y|u(z)\left(\int_0^1 \mathds{1}_{B\left(ty,\frac{1-t}{2}\right)}(z)(1-t)^{-n-1} dt\right)dz.
\end{align*}
Observe that $\mathds{1}_{B\left(ty,\frac{1-t}{2}\right)}(z)=1$ implies that $|z-y|\leq 2(1-t)$. Therefore
$$\int_0^1\mathds{1}_{B\left(ty,\frac{1-t}{2}\right)}(z)(1-t)^{-n-1}dt\leq \int_0^{1-\frac{1}{2}|z-y|}(1-t)^{-n-1}dt= \int_{\frac{1}{2}|z-y|}^1 r^{-n-1}dr\preceq \frac{1}{|z-y|^n}.$$
This implies
$$Vol(\mathsmaller{\frac{1}{2}}B)|h(\omega)|_y\preceq \int_{B(y,2)}|z-y|^{1-n}u(z)dz,$$
where $f\preceq g$ means $f\leq \text{const.} g$. (We also write $f\asymp g$ if $f\preceq g$ and $g\preceq f$.)
Using this estimate we have
\begin{align*}
\|h(\omega)\|_{L^\phi}
&\preceq \inf\left\{\gamma>0 : \int_B\phi\left(\int_{B(y,2)}|z-y|^{1-n}\frac{u(z)}{\alpha}dz\right)dy\leq 1\right\}.
\end{align*}
Since $\int_{B(y,2)}|z-y|^{1-n}dz<+\infty$, we can use Jensen's inequality and write
\begin{align*}
\|h(\omega)\|_{L^\phi}
&\preceq \inf\left\{\alpha>0 : \operatorname{Vol}(B(0,3))\int_B\int_{B(0,3)}\phi\left(\frac{u(z)}{\operatorname{Vol}(B(0,3))\alpha}\right)\frac{1}{|z-y|^{n-1}} dz dy\leq 1\right\}\\
&=\inf\left\{\alpha>0 : \operatorname{Vol}(B(0,3))\int_{B(0,3)}\phi\left(\frac{|\omega|_z}{\operatorname{Vol}(B(0,3))\alpha}\right)\left(\int_B\frac{dy}{|z-y|^{n-1}}\right)dz\leq 1\right\}.
\end{align*}
We have that there exists a constant $K>0$ such that $\int_B\frac{dy}{|z-y|^{n-1}}\leq K$ for all $z\in B(0,3)$, thus
$$\|h(\omega)\|_{L^\phi}\preceq \operatorname{Vol}(B(0,3))\|\omega\|_{L^{\tilde{K}\phi}}\preceq \|\omega\|_{L^\phi},$$
where $\tilde{K}=K\operatorname{Vol}(B(0,3))$.
By the identity $dh(\omega)=\omega-h(d\omega)$ we also have
$$\|dh(\omega)\|_{\phi}\leq \|\omega\|_{\phi}+\|h(d\omega)\|_{\phi}\preceq \|\omega\|_{\phi}+\|d\omega\|_{\phi}.$$
We conclude that $h$ is well-defined and bounded for the norm $|\ \ |_{L^\phi}$ in all degrees $k\geq 1$.
If $\omega=df$ for certain function $f$ we observe that
$$\eta_t^*(\iota_{\frac{\partial}{\partial t}}\varphi^*_x df) (y)= df_{\varphi_x(t,y)}(y-x)=(f\circ\gamma)'(t),$$
where $\gamma$ is the curve $\gamma(t)=\varphi_x(t,y)$. Then $\chi_x(df)(y)=f(y)-f(x)$, from which we get
$$h(df)=f-\frac{1}{Vol\left(\mathsmaller{\frac{1}{2}}B\right)}\int_{\frac{1}{2}B} f.$$
We define $h:L^\phi\Omega^0(B)\to L^p\Omega^{-1}(B)={\mathbb{R}}$ by
$h(f)=\frac{1}{Vol\left(\mathsmaller{\frac{1}{2}}B\right)}\int_{\frac{1}{2}B} f$,
which is clearly continuous because $\frac{1}{2}B$ has finite Lebesgue measure. Therefore
\begin{equation}\label{homot2}
h(f)+h(df)=f.
\end{equation}
This shows that $(L^\phi\Omega^*,d)$ retracts to $({\mathbb{R}}\to 0\to 0\to \cdots)$.
Note that, since $h$ is bounded, then it can be extended continuously to $L^\phi C^k(B)$ for every $k\geq 0$. Equalities \eqref{homot} and \eqref{homot2} are also true for every $\omega\in L^\phi C^k(B)$, which finishes the proof.
\end{proof}
\section{Proof of Theorem \ref{main}}
If $v$ is a vertex of $X$ we denote by $U_v$ the \textit{open star} of $v$, that is, the interior of the union of all simplices containing $v$. Observe that if $\Delta=(v_0,\ldots,v_m)\in X_m$, then
$$U_\Delta=U_{v_0}\cap\cdots\cap U_{v_m}$$
contains the interior of the simplex $\Delta$ as subset of $M$. By condition $(b)$ there exists a uniform constant $L\geq 1$ such that if $\Delta$ is a simplex of $X$, then $U_\Delta$ is $L$-biLipschitz homeomorphic to the open ball $B\subset{\mathbb{R}}^n$. Moreover, condition $(a)$ allows us to take a constant $N\in {\mathbb{N}}$ such that every point cannot belong to more than $N$ simplices of $X$.
The triangulation $X$ is the nerve of the covering ${\mathcal{U}}=\left\{U_v : v\in X_0\right\}$. In fact, Theorem \ref{main} can be stated in terms of the existence of a convenient open covering.
For $k,m\geq 0$ we consider $C^{k,m}$ as the space of alternating functions
\begin{equation}\label{forma}
\omega=\prod_{\Delta\in X_m} \omega_\Delta,\ \omega_\Delta\in L^\phi C^k(U_\Delta).
\end{equation}
Where by \textit{alternating} we mean that for $i\neq j$,
$$\omega_{(v_0,\ldots,v_i,\ldots,v_j,\ldots,v_m)}=-\omega_{(v_0,\ldots,v_j,\ldots,v_i,\ldots,v_m)}.$$
If $\omega\in C^{k,m}$, we consider
$$\varrho_\phi(\omega)=\sum_{\Delta\in X_m}\int_{U_\Delta}\phi\left(|\omega_\Delta|\right)dV,$$
where $dV$ is the volume on $M$. From this we take the Luxemburg norm
$$\|\omega\|_{C_\phi}=\inf\left\{\alpha>0 : \varrho_\phi\left(\frac{\omega}{\alpha}\right)\leq 1\right\}.$$
We define, for $\omega$ as in \eqref{forma},
$$d'\omega=(-1)^m\prod_{\Delta\in X_m}d\omega_\Delta,$$
and take
$$C_\phi^{k,m}=\left\{\omega\in C^{k,m} : \|\omega\|_{C_\phi},\|d'\omega\|_{C_\phi}<+\infty\right\},$$
which is naturally equipped with the norm $|\omega|_{C_\phi}=\|\omega\|_{C_\phi}+\|d'\omega\|_{C_\phi}$. Observe that $d': C_\phi^{k,m}\to C_\phi^{k+1,m}$ is continuous and satisfy $d'\circ d'=0$.
For $\omega\in C_\phi^{k,m}$ and $\Delta\in X_{m+1}$ we put
$$(d''\omega)_{\Delta}=\sum_{i=0}^{m+1} (-1)^i \omega_{\partial_i\Delta}.$$
\begin{lemma}
$(C_\phi^{*,*},d',d'')$ is a bicomplex.
\end{lemma}
\begin{proof}
One can directly verify that $d''\circ d''=0$ and $d'\circ d''+d''\circ d'=0$ so we need to prove that $d''$ is well-defined and continuous from $C_\phi^{k,m}$ to $C_\phi^{k,m+1}$.
Let $\omega\in C_\phi^{k,m}$. It is clear that $d''\omega$ is alternating. Moreover, for every $\alpha>0$ we have,
\begin{align*}
\varrho_\phi\left(\frac{d''\omega}{\alpha}\right) &=\sum_{\Delta\in X_{m+1}}\int_{U_\Delta}\phi\left(\frac{1}{\alpha}\left|\sum_{i=0}^{m+1}(-1)^i\omega_{\partial_i\Delta}\right|\right) dV\\
& \leq \frac{1}{m+2} \sum_{\Delta\in X_{m+1}}\sum_{i=0}^{m+1} \int_{U_\Delta} \phi\left(\frac{m+2}{\alpha}|\omega_{\partial_i\Delta}|\right)dV\\
& \leq \frac{N}{m+2} \sum_{\Delta'\in X_m}\int_{U_{\Delta'}} \phi\left(\frac{m+2}{\alpha}|\omega_{\Delta'}|\right)dV\\
&=\varrho_{\frac{N}{m+2}\phi}\left(\frac{m+2}{\alpha}\omega\right).
\end{align*}
Where in the second line we use Jensen's inequality and in the third we use the fact that every $m$-simplex of $X$ can be in the boundary of at most $N$ $(m+1)$-simplices. This estimate and an argument as in Remark \ref{ObsEqivalencia} allow to conclude that $\|d''\omega\|_{C_\phi}\preceq \|\omega\|_{C_\phi}$. Using the identity $d'\circ d''+d''\circ d'=0$ we also have $\|d'd''\omega\|_{C_\phi}=\|d''d'\omega\|_{C_\phi}\preceq \|d''\omega\|_{C_\phi}$, thus $d''\omega\in C_\phi^{k,m+1}$ and $d''$ is continuous for the norm $|\ \ |_{C_\phi}$.
\end{proof}
We write $E^k_\phi=\operatorname{Ker} d''|_{C_\phi^{k,0}}$ and $F^m_\phi=\operatorname{Ker} d'|_{C_\phi^{0,m}}$. These spaces conform the horizontal and vertical complexes of kernels of the bicomplex $(C_\phi^{*,*},d',d'')$.
\begin{lemma}\label{retraccion1}
For every $m$ the complex $(C_\phi^{*,m},d')$ retracts to $(F_\phi^m\to 0\to 0\to \cdots)$.
\end{lemma}
\begin{proof}
We have to define a family of continuous linear maps $H^k:C_\phi^{k,m}\to C_\phi^{k-1,m}$ such that
\begin{equation}\label{H}
\left\{\begin{array}{cc}
H^{k+1}\circ d'+d'\circ H^k= Id &\text{ if }k\geq 1 \\
H^1\circ d' + inc\circ H^0 =Id, &
\end{array}\right.
\end{equation}
where $C_\phi^{-1,m}=F_\phi^m$.
Remember that for every $\Delta\in X_m$ there exists a $L$-biLipschitz homeomorphism $f_\Delta:B\to U_\Delta$. By Rademacher's theorem it is differentiable almost everywhere, so we can consider the pull-back by $f$ of a $k$-form $\omega$ defined almost everywhere and denoted, as usual, by $f^*_\Delta\omega$.
Since $f_\Delta$ is $L$-biLipschitz one can easily prove that
$$|f_\Delta^*\omega|_x\leq L^k |\omega|_{f_\Delta(x)}\text{ and }|J_x f_\Delta|\leq L^n,$$
where $J_xf_\Delta$ denotes the Jacobian at $x$ of $f_\Delta$. This implies that $f^*_\Delta$ is well-defined and continuous from $L^\phi C^k(U)$ to $L^\phi C^k(B)$. The same can be done for the inverse $f^{-1}_\Delta$.
By Lemma \ref{Poincare} there exists a family of maps $h=h_k:L^\phi C^k(B)\to L^\phi C^{k-1}(B)$ such that
\begin{equation}\label{hh}
\left\{\begin{array}{cc}
h_{k+1}\circ d+d\circ h_k= Id &\text{ if }k\geq 1 \\
h_1\circ d + inc\circ h_0 =Id, &
\end{array}\right.
\end{equation}
(where $L^\phi C^{-1}(B)={\mathbb{R}}$), and
\begin{equation}\label{Estimacionh}
\int_B \phi\left(\frac{|h(\omega)|_y}{\alpha}\right)dy \leq C \int_{B}\phi\left(\frac{|\omega|_y}{Vol(B(0,3))\alpha}\right)dz
\end{equation}
with $C$ a positive constant.
Now we define $H^k:C_\phi^{k,m}\to C_\phi^{k-1,m}$ by
$$(H^k \omega)_\Delta =(-1)^m (f_\Delta^{-1})^*\circ h \circ f_\Delta^* \omega_\Delta.$$
Observe that \eqref{H} follows directly from \eqref{hh} and the fact that the pull-back commutes with $d$. Moreover
\begin{align*}
\varrho_\phi\left(\frac{H^k\omega}{\alpha}\right) & = \sum_{\Delta\in X_m}\int_{U_\Delta}\phi\left(\frac{1}{\alpha}|(f_\Delta^{-1})^*\circ h\circ f_\Delta^* \omega_\Delta|_x\right)dV(x) \\
&\leq \sum_{\Delta\in X_m}\int_{U_\Delta} \phi\left(\frac{L^k}{\alpha}|h\circ f_I^* \omega_\Delta|_{f_\Delta^{-1}(x)}\right)dV(x)\\
&\leq L^n\sum_{\Delta\in X_m}\int_B \phi\left(\frac{L^k}{\alpha}|h\circ f_\Delta^* \omega_\Delta|_y\right)dy\\
&\leq L^n C \sum_{\Delta\in X_m}\int_{B} \phi\left(\frac{L^k|f_\Delta^*\omega_I|_y}{Vol(B(0,3))\alpha}\right)dy,
\end{align*}
Using that $|f_\Delta^*\omega_I|_y\leq L^k |\omega|_{f_\Delta(y)}$ we obtain
\begin{align*}
\varrho_\phi\left(\frac{H^k\omega}{\alpha}\right) & \leq L^{2n} C \sum_{\Delta\in X_m}\int_{U_\Delta} \phi\left(\frac{L^{2k}|\omega|_{x}}{Vol(B(0,3))\alpha}\right)dV(x)\leq \varrho_{L^{2n}C\phi}\left(\frac{L^{2k}\omega}{Vol(B(0,3))\alpha}\right).
\end{align*}
From here we deduce that $\|H^k\omega\|_{C_\phi}\preceq \|\omega\|_{C_\phi}$. Applying this and \eqref{H}, we obtain $\|d'H^k\omega\|_{C_\phi}\preceq \|\omega\|_{C_\phi}+\|d'\omega\|_{C_\phi}$, which shows that $H^k$ is continuous for the norm $|\ \ |_{C_\phi}$.
\end{proof}
\begin{lemma}\label{retraccion2}
For every $k$ the complex $(C_\phi^{k,*},d'')$ retracts to $(E_\phi^k\to 0\to 0\to \cdots)$.
\end{lemma}
\begin{proof}
We have to construct a family of linear and continuous maps $P^m:C_\phi^{k,m}\to C_\phi^{k,m-1}$ such that
\begin{equation}\label{IdentidadP}
\left\{\begin{array}{cc}
P^{m+1}\circ d''+d''\circ P^m= Id &\text{ if }m\geq 1 \\
P^1\circ d'' + inc\circ P^0 =Id. &
\end{array}\right.
\end{equation}
Here we denote $C_\phi^{k,-1}=E_\phi^k$.
We consider a smooth partition of unity $\{\eta_v\}_{v\in X_0}$ for the covering ${\mathcal{U}}$. It can be taken such that $|d\eta_v|$ is uniformly bounded independently of $v\in X_0$ because of the properties of the covering. Then we define
$$(P^m \omega)_{\Delta} = (-1)^m\sum_{v\in X_0} \eta_v \omega_{\Delta v}\ \text{ if }m\geq 1;\text{ and}$$
$$(P^0\omega)_{v_0}=\sum_{v\in X_0} \eta_v \omega_v|_{U_{v_0}}.$$
Where $\Delta v=(v_0,\ldots,v_m,v)$ if $\Delta=(v_0,\ldots,v_m)$, and $\omega_{\Delta v}=0$ if $\Delta v\notin X_{m+1}$. Remember that in the previous expressions the sums do not have more than $N$ terms. It is clear that if $\omega\in C_\phi^{k,0}$, then $d''P^0\omega=0$.
If $\omega\in C_\phi^{k,m}$ for $m\geq 1$, we have
\begin{align*}
(P^{m+1}\circ d'' \omega)_\Delta &= (-1)^{m+1}\sum_{v\in X_0} \eta_v (d''\omega)_{\Delta v}=(-1)^{m+1}\sum_{v\in X_0} \eta_v \left( \sum_{i=0}^{m+1} (-1)^i \omega_{\partial_i(\Delta v)}\right)\\
&= (-1)^{m+1}\sum_{v\in X_0} \sum_{i=0}^m (-1)^i \eta_v \omega_{(\partial_i\Delta)v} + \sum_{v\notin \Delta}\eta_v \omega_\Delta.
\end{align*}
On the other hand
\begin{align*}
(d''\circ P^m \omega)_\Delta &= \sum_{i=0}^{m} (-1)^i (P^m \omega)_{\partial_i\Delta}= (-1)^m \sum_{i=0}^m \sum_{v\in X_0} (-1)^i \eta_v \omega_{(\partial_i\Delta)v}\\
&= (-1)^m \sum_{i=0}^m \sum_{v\in X_0} (-1)^i \eta_v \omega_{(\partial_i\Delta)} + \sum_{v\in \Delta} \eta_v \omega_{\Delta}
\end{align*}
This shows that for every $\Delta\in X_m$,
$$(P^m\circ d'' \omega)_\Delta+(d''\circ P^m \omega)_\Delta=\omega_\Delta.$$
Now suppose that $m=0$. If $x\in U_{v_0}$, we have
\begin{align*}
(P^1\circ d''\omega)_{v_0} &= -\sum_{v\in X_0} \eta_v (d''\omega)_{(v_0,v)}=-\sum_{v\in X_0} \eta_v \left(\omega_v-\omega_{v_0(x)}\right)\\
& = \sum_{v\in X_0}\eta_v\omega_{v_0}-\sum_{v\in X_0}\eta_v\omega_v =\omega_{v_0}-(inc\circ P^0 \omega)_{v_0}.
\end{align*}
To finish the proof let us show that $P^m$ is continuous (and that its image is where it must be). Take $\omega\in C_\phi^{k,m}$ for any $m\geq 0$, then, using Jensen's inequality and the fact that $\eta_v$ is supported in $U_v$ and is bounded by $1$, we obtain
$$\varrho_\phi\left(\frac{P^m \omega}{\alpha}\right)\leq \varrho_{(m+1)\phi}\left(\frac{N\omega}{\alpha}\right),$$
which allows to conclude that $\|P^m\omega\|_{C_\phi}\preceq \|\omega\|_{C_\phi}$. Moreover, if $m\geq 1$, we have
$$|(d'\circ P^m\omega)_\Delta| \leq \sum_{v\in X_0} |d\eta_v| |\omega_{\Delta v}| + |P^m(d'\omega)_\Delta|.$$
We can apply the previous estimation in the second term and the fact that $|d\eta_v|$ is uniformly bounded independently from $v$ to get a $\|d'\circ P^m\omega\|_{C_\phi}\preceq \|\omega\|_{C_\phi}+\|d'\omega\|_{C_\phi}$. An analogous estimate can be proved for $m=0$. As a conclusion we have that $P^m\omega\in C_\phi^{k,m}$ and $P^m$ is continuous for the norm $|\ \ |_{C_\phi}$.
\end{proof}
\begin{lemma}\label{horizontal}
There exists a family of isomorphisms $\mathcal{E}:E_\phi^*\to L^\phi C^*(M)$ such that $\mathcal{E}\circ d'=d\circ\mathcal{E}$.
\end{lemma}
\begin{proof}
If $\omega\in E_\phi^k$, then for every $v_0,v_1\in X_0$ we have $\omega_{v_0}|_{U_{v_1}}=\omega_{v_1}|_{U_{v_0}}$ (a.e.). This allows to define a measurable $k$-form $\tilde{\omega}=\mathcal{E} \omega$ such that $\omega_v=\tilde{\omega}|_{U_v}$ for every $v\in X_0$. On the other hand, the inverse of $\Psi$ is clearly defined by
$$(\mathcal{E}^{-1}\tilde{\omega})_v=\tilde{\omega}|_{U_v}.$$
To see that both maps $\mathcal{E}$ and $\mathcal{E}^{-1}$ are well-defined and continuous embeddings observe that
\begin{align*}
\varrho_\phi\left(\frac{\omega}{\alpha}\right) &= \sum_{v\in X_0}\int_{U_v} \phi \left(\frac{|\omega_v|}{\alpha}\right)dV = \sum_{v\in X_0}\int_{U_v} \phi \left(\frac{|\tilde{\omega}|}{\alpha}\right)dV \leq (n+1) \int_M \phi \left(\frac{|\tilde{\omega}|}{\alpha}\right)dV\\
&\leq (n+1)\sum_{v\in X_0}\int_{U_v} \phi \left(\frac{|\tilde{\omega}|}{\alpha}\right)dV = (n+1)\varrho_\phi\left(\frac{\omega}{\alpha}\right)
\end{align*}
This shows that $\|\omega\|_{C_\phi}\asymp \|\tilde{\omega}\|_{L^\phi}$. It is clear that both $\mathcal{E}$ and $\mathcal{E}^{-1}$ commute with the derivatives, which, together with the previous estimate imply $\|d'\omega\|_{C_\phi}\asymp \|d\tilde{\omega}\|_{L^\phi}$ and finish the proof.
\end{proof}
\begin{lemma}\label{vertical}
There exists a family of isomorphisms $\mathcal{F}:F_\phi^*\to \ell^\phi(X_*)$ such that $\mathcal{F}\circ d''=\delta \circ\mathcal{F}$.
\end{lemma}
\begin{proof}
If $\omega\in F_\phi^m$, then for every $\Delta\in X_m$ the $k$-form $\omega_\Delta$ is an essentially constant function. We define $\mathcal{F} \omega: X_m\to {\mathbb{R}}$ such that $\mathcal{F} \omega (\Delta)$ is the essential value of $\omega_\Delta$. The inverse of $\mathcal{F}$ is given by $(\mathcal{F}^{-1} \theta)_\Delta\equiv \theta(\Delta)$. It is easy to verify that $\mathcal{F}\circ d''=\delta \circ\mathcal{F}$.
To see that $\mathcal{F}$ and $\mathcal{F}^{-1}$ are well-defined and continuous embeddings suppose that $V\geq 1$ is such that $V^{-1}\leq \operatorname{Vol}(U_\Delta)\leq V$ for every simplex $\Delta$. Then, using Jensen's inequality, we have for every $\alpha$,
\begin{align*}
\sum_{\Delta\in X_m} \phi\left(\frac{|\mathcal{F} \omega(\Delta)|}{\alpha}\right) &= \sum_{\Delta\in X_m} \phi\left(\int_{U_\Delta}\frac{|\omega_\Delta|}{\operatorname{Vol}(U_\Delta)\alpha}\right)dV\\
&\leq V \sum_{\Delta\in X_m}\int_{U_\Delta} \phi\left(\frac{|\omega_\Delta|}{\alpha}\right) dV\\
&\leq V^2 \sum_{\Delta\in X_m} \phi\left(\frac{|\mathcal{F} \omega(\Delta)|}{\alpha}\right).
\end{align*}
This shows that $\|\mathcal{F} \omega\|_{\ell^\phi}\asymp \|\omega\|_{C_\phi}=|\omega|_{C_\phi}$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{main}]
Combining Theorem \ref{bicomplejo} with Lemmas \ref{retraccion1} and \ref{retraccion2} we have that the complexes $(E_\phi^*,d')$ and $(F_\phi^*,d'')$ are homotopically equivalent. Then, aplying Lemmas \ref{horizontal} and \ref{vertical}, we conclude that $(L^\phi C^*(M),d)$ and $(\ell^\phi (X_*),d)$ are homotopically equivalent, which implies the existence of a family of isomorphisms between the corresponding cohomology spaces.
\end{proof}
Observe that the previous argument can be done considering $\Omega_\phi^{k,m}$ instead of $C_\phi^{k,m}$, which is the space of alternating functions
\begin{equation}
\omega=\prod_{\Delta\in X_m} \omega_\Delta,\ \omega_\Delta\in L^\phi \Omega^k(U_\Delta)
\end{equation}
with $\|\omega\|<+\infty$. Using the above arguments we can obtain the following result:
\begin{theorem}\label{mainDif}
Take $M$ and $X$ as in Theorem \ref{main} and suppose in addition that
\begin{enumerate}
\item[(e)] $X$ is a smooth triangulation, and
\item [(f)] the biLipschitz homeomorphisms given by condition $(b)$ are indeed diffeomorphisms.
\end{enumerate}
Then for any Young function $\phi$ and $k=0,\ldots,n$, there exists an isomorphism between $L^\phi H_s^k(M)$ and $\ell^\phi H^k(x)$. In particular $L^\phi H_s^k(M)$ is isomorphic to $L^\phi H^k(M)$.
\end{theorem}
If $M$ is a compact Riemannian manifold, then it satisfies the conditions of Theorem \ref{mainDif}. In this case it is easy to see that $L^\phi\Omega^*(M)=\Omega^k(M)$, thus the $L^\phi$-cohomology of $M$ is coincides with its classic de Rham cohomolgy (in the sense of vector spaces).
\section*{Acknowledgments}
I am deeply grateful to Yaroslav Kopylov for his helpful ideas and discussions.
This work was supported by the \textit{Mathematical Center in Akademgorodok} under the agreement No. 075-15-2019-1675 with the \textit{Ministry of Science and Higher Education} of the Russian Federation.
|
\section{Introduction}
Based on the previous work \cite{xu2021sharp}, we continue to discuss the operator
\begin{equation}\label{OSO}
Tf(x)=\int_{\R} e^{i\lambda S(x,y)}K(x,y)\psi(x,y)f(y)dy,
\end{equation}
where $K(x,y)$ is a $C^2$ function away from the diagonal satisfying
\begin{equation}\label{SinCond}
\abs{K(x,y)}\leq E\abs{x-y}^{-\mu}, \ \ \abs{\partial_y^i K(x,y)}\leq E\abs{x-y}^{-\mu-i},
\end{equation}
where the $0<\mu<1$, $E$ is an constant and $i=1, 2$. This kind of operators were introduced in \cite{liu1999model} in which the author also established the sharp $L^2$ decay estimates for \eqref{OSO} with arbitrary homogeneous polynomial phases. Look at the operators \eqref{OSO}, they are closely related to the usual degenerate oscillatory integral operators studied by \cite{phong1992oscillatory}, \cite{phong1994models}, \cite{phong1997newton} and also can be seen as variants of the operators considered in \cite{phong1986hilbert}, \cite{ricci1987harmonic}. Readers may find more backgrounds about these operators as well as the relationship in \cite{pan1998complete}, \cite{pan19972}, \cite{shi2019damping}, \cite{xiao2017endpoint}, \cite{xu2021sharp} and the references therein. Now, returning to our main topic. For \eqref{OSO}, an intersting question is to seek for the range of the exponent $p$ such that the sharp decay rate $-\frac{1-\mu}{n}$ obtained on $L^2$ spaces can be preserved on $L^p$ spaces. Obviously, the range of $p$ is determined by the phase function $S(x,y)$. If $S(x,y)$ is a homongeneous polynomial, it can be written as
\begin{equation}\label{phase}
S(x,y)=\sum_{k=1}^{n-1}a_kx^{n-k}y^k.
\end{equation}
Observe the term above, the reason we ignore the pure $x-$term or $y-$term is that these pure terms have no effect on norm estimate.
In \cite{xu2021sharp}, by imposing an additional condition on $K(x,y)$,
\begin{equation}\label{AddCondi}
\abs{\partial_x^i K(x,y)}\leq E\abs{x-y}^{-\mu-i}, \tag{AC}
\end{equation}
we proved that
\begin{theorem}\label{PreMain-Result}
If $S(x,y)$ is a homogeneous polynomial of the form \eqref{phase} with $a_1a_{n-1}\neq 0$ and the Hessian is not of the form $c(y-x)^{n-2}$, while
$K(x,y)$ safisfies \eqref{SinCond} and \eqref{AddCondi}.
Then the sharp decay estimate
\begin{equation}\label{Main-Inq}
\|Tf\|_{L^p}\leq C \lambda^{-\frac{1-\mu}{n}}\|f\|_{L^p}
\end{equation}
holds for $\frac{n-2\mu}{n-1-\mu}\leq p\leq \frac{n-2\mu}{1-\mu}$, and $C$ is independent of $\lambda$.
\end{theorem}
Observe that the phases in this result are just a class of homogeneous polynomials whose Hessian vanish only on lines away from axes. On the other extreme side, if the Hessian of the phase vanishes only on axes, this correponds to the monomial cases which have been extensively discussed in \cite{pan1998complete}, \cite{pan19972}. It should be pointed out that the operators considered in \cite{pan1998complete}, \cite{pan19972} are global integral operators which integrate over the entire real axis. To transfrom the boundedness results therein to the decay estimates for \eqref{OSO}, it suffices to apply the routine scaling argument which we omit here.
Before we state our main results, some notations should be introduced. For a homogeneous polynomial of the form \eqref{phase}, we denote
\begin{equation*}
k_m=\min\{k: a_k\neq 0\},\quad k_M=\max\{k: a_k\neq 0\}.
\end{equation*}
In view of the argument above, if we aim to establish sharp $L^p$ decay estimates for operators with arbitrary homogeneous polynomial phases, it remains to discuss the case $k_m\neq k_M$. Then we obtain the following theorem.
\begin{theorem}\label{Main-Result}
If $S(x,y)$ is a homogeneous polynomial of the form \eqref{phase} and $k_m\neq k_M$, while $K(x,y)$ safisfies \eqref{SinCond} and \eqref{AddCondi}.
Then the sharp decay estimate
\begin{equation}
\|Tf\|_{L^p}\leq C \lambda^{-\frac{1-\mu}{n}}\|f\|_{L^p}
\end{equation}
holds for $\frac{n-2\mu}{n-2\mu-k_m(1-\mu)}\leq p\leq \frac{n-2\mu}{(1-\mu)(n-k_M)}$, and $C$ is independent of $\lambda$.
\end{theorem}
When dealing with degenerate oscillatory integral operators, we should analyze the Hessian $S^{''}_{xy}(x,y)$ of the phase function since we shall make use of the operator van der Corput lemma. By direct factor decomposition, we can rewrite the Hessian as
\begin{equation}\label{Hessian}
S^{''}_{xy}(x,y)=cx^\gamma y^\beta\prod_{l=1}^s(y-\alpha_lx)^{m_l}\prod_{l=1}^rQ_l(x,y),
\end{equation}
where $c$ and $\alpha_j$ are nonzero and each $Q_l(x,y)$ is a positive definite quadratic form, obviously,
\begin{align}
&\gamma+\beta+\sum_{l=1}^sm_l+2r=n-2,\label{Degree}\\
&k_m=\beta+1,\\
&k_M=n-(\gamma+1).
\end{align}
Using these notations, we can reformulate Theorem \ref{Main-Result} as the following theorem which we will prove.
\begin{theorem}\label{Main-Result-1}
If the Hessian of the phase function $S(x,y)$ is of the form \eqref{Hessian} and $s\neq 0$ or $r\neq 0$, while $K(x,y)$ safisfies \eqref{SinCond} and \eqref{AddCondi}.
Then the sharp decay estimate
\begin{equation}\label{Main-Inq}
\|Tf\|_{L^p}\leq C \lambda^{-\frac{1-\mu}{n}}\|f\|_{L^p}
\end{equation}
holds for $\frac{n-2\mu}{n-2\mu-(\beta+1)(1-\mu)}\leq p\leq \frac{n-2\mu}{(1-\mu)(\gamma+1)}$, and $C$ is independent of $\lambda$.
\end{theorem}
Now we roughly describe the strategy of our proof. \\
\emph{Step} 1: Seperating the operator $T$ into two parts by inserting a cut-off function
\begin{equation*}
T=T_1+T_2.
\end{equation*}
The support of $T_1$ is in a neighborhood of the singular line, $x-y=0$, of $K(x,y)$. Simple Schur test yields \eqref{Main-Inq} for $T_1$. In this step, the decay comes from the integrability of $K(x,y)$, so we do not concern about what the phase function is.\\
\emph{Step }2: We move to treat $T_2$ and analyze the singular varieties of the Hessian. In view of the form of \eqref{Hessian}, it may vanish on $x-$axis, $y-$axis or a line crossing the origin. Specifically, we consider the following three cases respectively.\\
\emph{Case }1: $\gamma\neq 0$ or $\beta\neq 0$.\\
This means the singular varieties of the Hessian consist of at least one axis. Since $x$ and $y$ are in the same status in this operator, so it suffices to consider for instance $\gamma\neq 0$, and the case $\beta\neq 0$ can be dealt with by duality and interchanging the roles of $x$ and $y$. Furthermore, we divide $T_2$ into three parts according to the singular varieties,
\begin{equation*}
T_2=T_X+T_\Delta+T_Y.
\end{equation*}
\eqref{Main-Inq} for $T_X$ and $T_Y$ can be derived by inserting them into two families of analytic operators and employing lifting trick and complex interpolation respectively. For $T_\Delta$, we shall apply local Riesz-Thorin interpolation to get the conclusion.\\
\emph{Case }2: The Hessian is of the form $c(y-x)^{n-2}$.\\
In this case, local Riesz-Thorin interpolation can not give the endpoint estimates. So we insert this operator into a family of damped oscillatory integral operators and establish $L^2\to L^2$ as well as $H^1\to L^1$ boundedness results for different complex exponents. At last, complex interpolation implies the final result.\\
\emph{Case }3: Otherwise.\\
This case has been treated in \cite{xu2021sharp}, specifically Theorem \ref{PreMain-Result}.\\
The novelty of this paper is the observation that, apart from some special cases, orthogonality is powerful enough to give the sharp $L^p$ decay estimates for one-dimensional degenerate oscillatory integral operators. The orthogonality displays in two aspects: on one side, if the varieties of the Hessian contain lines apart from axes, around one such line, the geometry gurantees the orthogonality even on $L^p$ spaces; on the other side, when we deal with the region close to axes, we use the othogonality of damped oscillatory integral operators. In fact, following this strategy, we can recover the main results of \cite{shi2017sharp}. Moreover, together with the works of \cite{pan1998complete} and \cite{pan19972}, we step forward to understand the $L^p$ mapping properties of \eqref{OSO} with more degenerate phases.
\\
This article is organized as follows: Section 2 is devoted to \emph{Step }1. In Section 3, we deal with \emph{Case} 1. \emph{Case} 2 will be treated in Section 4. Since the basic strategy in the present paper is similar with \cite{xu2021sharp}, so we omit some repeating and routine arguments and just state those different techniques as well as some necessary contents.\\
{\bf Notation:} In this paper, the constant $C$ independent of $\lambda$ and the test function $f$ is not necessarily the same one in each occurrence. Readers interested in the precise parameters, the constant $C$ depend on, may refer to \cite{xu2021sharp}. Throughout this paper, we use $\norm{\cdot}$ to denote $\norm{\cdot}_{L^2\to L^2}$ and $\norm{\cdot}_{L^p}$ to denote $\norm{\cdot}_{L^p\to L^p}$.
\section{Preliminaries}
Two elements arise in \eqref{OSO}, one is the singular integral, the other one is the oscillatory integral. We need to clarify different contributions of these two kind of integrals to the final decay. Specifically, inserting a cut-off function into \eqref{OSO} we get
\begin{align*}
Tf(x)&=\int_{\R} e^{i\lambda S(x,y)}K(x,y)\psi(x,y)f(y)dy\\
&=\int_{\R} e^{i\lambda S(x,y)}K(x,y)\phi\left(\lambda^{\frac{1}{n}}(x-y)\right)\psi(x,y)f(y)dy+\\
&\quad\int_{\R} e^{i\lambda S(x,y)}K(x,y)\left[1-\phi\left(\lambda^{\frac{1}{n}}(x-y)\right)\right]\psi(x,y)f(y)dy\\
&:=T_1f(x)+T_2f(x),
\end{align*}
where $\phi\in C_0^\infty(\R)$ and
\begin{equation*}
\phi(x)\equiv
\begin{cases}
0, &\quad \abs{x}\geq 1,\\
1, &\quad \abs{x}\leq \frac{1}{2}.
\end{cases}
\end{equation*}
The kernel of $T_1$
\begin{equation*}
K_1(x,y)=e^{i\lambda S(x,y)}K(x,y)\phi\left(\lambda^{\frac{1}{n}}(x-y)\right)\psi(x,y)
\end{equation*}
is absolutely integrable and
\begin{equation*}
\sup_x\int \abs{K_1(x,y)}dy\leq C\lambda^{-\frac{1-\mu}{n}}, \quad \sup_y\int \abs{K_1(x,y)}dx\leq C\lambda^{-\frac{1-\mu}{n}}.
\end{equation*}
Therefore the following Schur test leads to \eqref{Main-Inq} for $T_1$.
\begin{lemma}\label{Schur-Test}
If the operator
\begin{equation*}
Vf(x)=\int K(x,y)f(y)dy,
\end{equation*}
has a kernel $K(x,y)$ satisfying
\begin{equation*}
\sup_x\int \abs{K(x,y)}dy\leq A_1, \quad \sup_y\int \abs{K(x,y)}dx\leq A_2,
\end{equation*}
then
\begin{equation*}
\norm{V}_{L^p\rightarrow L^p}\leq \left(\frac{A_1}{p}+\frac{A_2}{p'}\right),
\end{equation*}
where $1\leq p\leq+\infty$.
\end{lemma}
The proof of this lemma is easy, interested readers may find more details in \cite{xu2021sharp}, so we omit it here. A more general version of this lemma has been given by \cite{prama2005}. \\
Choose a cut-off function $\Psi\in C_0^{\infty}$ such that $\,{\rm supp}\,\Psi\subset[\frac{1}{2},2]$ and $\sum_{l\in \mathbb{Z}}\Psi(2^lx)\equiv 1$. Dyadically decompose $T_2$ as
\begin{align*}
T_2f(x)&=\sum_{\sigma_1,\sigma_2=\pm}\sum_{j,k}\int_{\R} e^{i\lambda S(x,y)}K(x,y)\left[1-\phi\left(\lambda^{\frac{1}{n}}(x-y)\right)\right]\Psi_j(\sigma_1x)\Psi_k(\sigma_2y)\psi(x,y)f(y)dy\\
&:=\sum_{j,k}T_{j,k}^{\sigma_1,\sigma_2}f(x)
\end{align*}
where $\Psi_j(x)=\Psi(2^jx), \Psi_k(x)=\Psi(2^kx)$. For convenience, we focus only on the case $\sigma_1=+, \sigma_2=+$, the remaining cases can be dealt with similarly. We shall still use $T_2$ and $T_{j,k}$ to denote $\sum_{j,k}T_{j,k}^{+,+}$ and $T_{j,k}^{+,+}$ respectively.
Since we only consider the first quadrant, we may suppose that $0<\alpha_1<\alpha_2<\cdots<\alpha_s$. Now we finish the preparation work and are clear about all the varieties of the Hessian. The next step is to properly decompose the operator according to these varieties.
\section{Case 1: $\gamma\neq 0$ or $\beta\neq 0$.}
As we have stated in Section 1, we may assume $\gamma\neq 0$. Suppose that $\mathcal{K}$ is a positive constant depending on $\alpha_1,\cdots,\alpha_s$. Let $j\gg k$($j\ll k$) represent $j>k+\mathcal{K}$($j<k-\mathcal{K}$) such that the size of $y$-variable ($x$-variable) is dominant in the Hessian $S_{xy}^{''}$, while $j\sim k$ naturally means $|j-k|\leq \mathcal{K}$.
We further divide $T_2$ into three groups as follows.
\begin{align*}
T_2f(x)&=\sum_{j\gg k}T_{j,k}f(x)+\sum_{j\sim k}T_{j,k}f(x)+\sum_{j\ll k}T_{j,k}f(x)\\
&:=T_Yf(x)+T_\Delta f(x)+T_Xf(x).
\end{align*}
Our goal is to establish \eqref{Main-Inq} for $T_X, T_\Delta$ and $T_Y$ individually.\\
We insert $T_X$ and $T_Y$ into the following two families of analytic damped oscillatory integral operators
\begin{align}
\notag T_Y^zf(x)&=\sum_{j\gg k}\int_{\R}e^{i\lambda S(x,y)}K(x,y)\abs{D(x,y)}^{z}\left[1-\phi\left(\lambda^{\frac{1}{n}}(x-y)\right)\right]\cdot\\
\notag&\quad\quad\quad \Psi_j(x)\Psi_k(y)\psi(x,y)f(y)dy,\\
&:=\sum_{j\gg k}D^Y_{j,k}f(x),\label{T_X}\\
\notag T_X^zf(x)&=\sum_{j\ll k}\int_{\R}e^{i\lambda S(x,y)}K(x,y)\abs{D(x,y)}^{z}\left[1-\phi\left(\lambda^{\frac{1}{n}}(x-y)\right)\right]\cdot\\
\notag &\quad\quad\quad\Psi_j(x)\Psi_k(y)\psi(x,y)f(y)dy\\
&:=\sum_{j\ll k}D^X_{j,k}f(x).\label{T_Y}
\end{align}
The damped oscillatory operators in one dimension have been studied intensively, if the damping factor is the Hessian of the phase function, readers can find more results in \cite{phong1998damped}, if the damping factor is unrelated to the phase function, readers may refer to \cite{pramanik2002convergence} and \cite{prama2005}. In \cite{yang2004sharp} the author used damping estimates to establish $L^p$ decay estimates for oscillatory integral operators.
Our goal of this section is to establish the following $L^2$ decay estimates for damped oscillatory integral operators.
\begin{theorem}\label{dampL2}
If the Hessian of the phase function $S(x,y)$ is of the form \eqref{Hessian} and $\gamma\neq 0$, and if we set
\begin{equation}\label{DamFac1}
D(x,y)=cx^\gamma\prod_{l=1}^s(y-\alpha_lx)^{m_l}\prod_{l=1}^rQ_l(x,y),
\end{equation}
then for $\,{\rm Re}\,(z)=\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(n-2-\beta)(\beta+1)}\geq 0$, we have
\begin{align}
&\norm{T_Y^zf}_{L^2}\leq C\lambda^{-\frac{n-2\mu}{2n}\cdot \frac{1}{\beta+1}}\norm{f}_{L^2},\label{dampL2-1}\\
&\norm{T_X^zf}_{L^2}\leq C\lambda^{-\frac{n-2\mu}{2n}\cdot \frac{1}{\beta+1}}\norm{f}_{L^2}.\label{dampL2-2}
\end{align}
If we set
\begin{equation}\label{DamFac2}
D(x,y)=cx^\gamma,
\end{equation}
then for $\,{\rm Re}\,(z)=\frac{1}{2\gamma}\cdot\frac{2(1-\mu)(\gamma+1)-n+2\mu}{n-2\mu-(1-\mu)(\gamma+1)}\geq 0$, we have
\begin{align}
&\norm{T_Y^zf}_{L^2}\leq C\lambda^{-\frac{n-2\mu}{2n}\cdot\frac{1-\mu}{n-2\mu-(1-\mu)(\gamma+1)}}\norm{f}_{L^2},\label{dampL2-3}\\
&\norm{T_X^zf}_{L^2}\leq C\lambda^{-\frac{n-2\mu}{2n}\cdot\frac{1-\mu}{n-2\mu-(1-\mu)(\gamma+1)}}\norm{f}_{L^2}.\label{dampL2-4}
\end{align}
\end{theorem}
Note that this theorem extends Theorem 2 in \cite{xu2021sharp} to more general phases. To establish these estimates, we start with a local version, then we make use of them to get the global estimates. In fact, the procedure to get the local version of these estimates is routine and cumbersome because it needs some notations and definitons. The following part is basically same with the conterpart in \cite{xu2021sharp}, however, it is necessary for the rigorous statement, so we keep them here. Readers may skip these and directly see Lemma \ref{Orth_2}. \\
Let us introduce the local damped operators
\begin{equation}\label{DampO}
D(\mathcal{B})f(x)=\int_{\R} e^{i\lambda S(x,y)}\abs{D(x,y)}^{z}K(x,y)\left[1-\phi\left(\lambda^{\frac{1}{n}}(x-y)\right)\right]\psi(x,y)f(y)dy,
\end{equation}
where $\psi\in C_0^{\infty}$ and $\,{\rm supp}\,\psi\subset \mathcal{B}$. Now we consider two operators $D(\mathcal{B}_1)$ and $D(\mathcal{B}_2)$ with supports in $\mathcal{B}_1$ and $\mathcal{B}_2$ respectively. Here both $\mathcal{B}_1$ and $\mathcal{B}_2$ are rectangular boxes with sides parallel to the axes; in addition, we suppose that $\mathcal{B}_2$ is the minor box and will be contained in a horizontal translate of the major box $\mathcal{B}_1$. Now, we repeat the statements and assumptions in \cite{phong1998damped}.
\begin{align*}
&\mathcal{B}_1=\{(x,y): a_1<x<b_1, c_1<y<d_1\}, \rho_1=d_1-c_1;\\
&\widetilde{\mathcal{B}_1}=\left\{(x,y): a_1-\frac{1}{10}(b_1-a_1)<x<b_1+\frac{1}{10}(b_1-a_1), c_1<y<d_1\right\};\\
&\mathcal{B}_1^*=\{(x,y): a_1-(b_1-a_1)<x<b_1+(b_1-a_1), c_1<y<d_1\};\\
&\mathcal{B}_2=\{(x,y): a_2<x<b_2, c_2<y<d_2\}, \rho_2=d_2-c_2.
\end{align*}
\begin{enumerate}
\item[(A1)] We define the span \emph{span}($\mathcal{B}_1,\mathcal{B}_2$), as the union of all line segments parallel to the $x$-axis, which joints a point $(x,y)\in\mathcal{B}_1$ with a point $(z,y)\in\mathcal{B}_2$. While we also assume that $S_{xy}^{''}$ does not change sign in the span \emph{span}($\mathcal{B}_1,\mathcal{B}_2$) and satisfies
\begin{align}
&\nu\leq \min_{\widetilde{\mathcal{B}_1}}\abs{S_{xy}^{''}}\leq A\nu,\\
&\max_{\emph{span}(\mathcal{B}_1,\mathcal{B}_2)}\abs{S_{xy}^{''}}\leq A\nu.
\end{align}
\item[(A2)] $\mathcal{B}_2\subset \mathcal{B}_1^*$, this implies $\rho_2\leq \rho_1$.
\end{enumerate}
For the cut-off functions $\psi_j(x,y)$, we also assume that
\begin{enumerate}
\item[(A3)] $\sum_k\rho_j^{k}\abs{\partial_y^{k}\psi_j}\leq B$.
\end{enumerate}
Based on these concepts, repeating the proof of Lemma 2 in \cite{xu2021sharp}, we can obtain that
\begin{lemma}\label{Orth_2}
Under the assumptions (A1)-(A3),
\begin{align}
&\norm{D(\mathcal{B}_1)D(\mathcal{B}_2)^{*}}\leq C \lambda^{\frac{2\mu}{n}}(\lambda \nu)^{-1}
\sup_{\mathcal{B}_2}\abs{D(x,y)}^{\,{\rm Re}\,(z)}\cdot\sup_{\widetilde{\mathcal{B}_1}}\abs{D(x,y)}^{\,{\rm Re}\,(z)},\label{Orth_2-1}\\
&\norm{D(\mathcal{B}_2)D(\mathcal{B}_1)^{*}}\leq C \lambda^{\frac{2\mu}{n}}(\lambda \nu)^{-1}
\sup_{\mathcal{B}_2}\abs{D(x,y)}^{\,{\rm Re}\,(z)}\cdot\sup_{\widetilde{\mathcal{B}_1}}\abs{D(x,y)}^{\,{\rm Re}\,(z)}.\label{Orth_2-2}
\end{align}
\end{lemma}
\begin{remark}\label{DualLemma}
In the operator \eqref{DampO}, interchanging the roles of $x$ and $y$ and assuming the same assumptions (A1)-(A3), then \eqref{Orth_2-1} and \eqref{Orth_2-2} also hold for operators $D(\mathcal{B}_1)^{*}D(\mathcal{B}_2)$ and $D(\mathcal{B}_2)^{*}D(\mathcal{B}_1)$ respectively. Especially, if $\mathcal{B}_1=\mathcal{B}_2$ which we redenote by $\mathcal{B}$, we have
\begin{equation}\label{DampL2}
\norm{D(\mathcal{B})}\leq C\lambda^{\frac{\mu}{n}}(\lambda \nu)^{-\frac{1}{2}}
\sup_{\mathcal{B}}\abs{D(x,y)}^{\,{\rm Re}\,(z)}.
\end{equation}
\end{remark}
Now we are ready to prove Theorem \ref{dampL2}.
\begin{proof}
Recall that
\begin{equation*}
T_X^zf(x)=\sum_{j\ll k}D^X_{j,k}f(x),\quad
T_Y^zf(x)=\sum_{j\gg k}D^Y_{j,k}f(x).
\end{equation*}
We first prove \eqref{dampL2-2} assuming $\beta\neq 0$. If $j\ll k, j'\ll k'$ and also $j<j'$, then from Lemma \ref{Orth_2}, we know that
\begin{align}
&\norm{D_{j,k}^X\left(D^X_{j',k'}\right)^{*}}\notag\\
&\leq C\lambda^{\frac{2\mu}{n}}\left[\lambda 2^{-j(n-2-\beta)}2^{-\beta k}\right]^{-1}\left[2^{-j(n-2-\beta)}\right]^{\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(n-\beta-2)(\beta+1)}}\left[2^{-j'(n-2-\beta)}\right]^{\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(n-2-\beta)(\beta+1)}}\notag\\
&=C\lambda^{\frac{2\mu-n}{n}}2^{j\frac{2(n-2-\beta)(\beta+1)-[n-2\mu-2(1-\mu)(\beta+1)]}{2(\beta+1)}}2^{-j'\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(\beta+1)}}2^{k\beta}.\label{OscEst-1}
\end{align}
In the inequality above there is no $k'$ because $D_{j,k}^X\left(D^X_{j',k'}\right)^{*}=0$ if $\abs{k-k'}>2$, so we identify $k$ with $k'$ here. In what follows, we also use this to avoid cumbersome argument.\\
On the other hand, we have the trivial size estimate
\begin{align}
&\norm{D_{j,k}^X\left(D^X_{j',k'}\right)^{*}}\notag\\
&\leq \norm{D_{j,k}^X}\cdot\norm{\left(D^X_{j',k'}\right)^{*}}\notag\\
&\leq 2^{-j\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(\beta+1)}}2^{j\mu}2^{-\frac{j}{2}}2^{-\frac{k}{2}}2^{-j'\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(\beta+1)}}2^{j'\mu}2^{-\frac{j'}{2}}2^{-\frac{k}{2}}\notag\\
&=2^{-j\frac{n-2\mu-(\beta+1)}{2(\beta+1)}}2^{-j'\frac{n-2\mu-(\beta+1)}{2(\beta+1)}}2^{-k}\label{SizEst-1}
\end{align}
By convex combination, for any $\theta(0\leq \theta\leq 1)$, we know that
\begin{align*}
&\norm{D_{j,k}^X\left(D^X_{j',k'}\right)^{*}}\\
&\leq {\eqref{OscEst-1}}^{\theta}\cdot{\eqref{SizEst-1}}^{1-\theta}.
\end{align*}
By setting $\theta=\frac{1}{\beta+1}$, we have
\begin{align*}
&\norm{D_{j,k}^X\left(D^X_{j',k'}\right)^{*}}\\
&\leq {\eqref{OscEst-1}}^{\frac{1}{\beta+1}}\cdot{\eqref{SizEst-1}}^{\frac{\beta}{\beta+1}}\\
&\leq C\lambda^{-\frac{n-2\mu}{2n(\beta+1)}}2^{j\frac{2(n-2-\beta)+\beta-(n-2)}{2(\beta+1)}}2^{-j'\frac{n-2-\beta}{2(\beta+1)}}.
\end{align*}
Given \eqref{Degree}, it follows
\begin{equation*}
2(n-2-\beta)+\beta-(n-2)=n-2-\beta.
\end{equation*}
This fact yields
\begin{equation*}
\norm{D_{j,k}^X\left(D^X_{j',k'}\right)^{*}}\leq C\lambda^{-\frac{n-2\mu}{2n(\beta+1)}}2^{(j-j')\frac{n-2-\beta}{2(\beta+1)}}.
\end{equation*}
Repeat the above argument for $\left(D^X_{j,k}\right)^{*}D_{j',k'}^X$ and assume $k<k'$, there is
\begin{equation*}
\norm{\left(D^X_{j,k}\right)^{*}D_{j',k'}^X}\leq C\lambda^{-\frac{n-2\mu}{2n(\beta+1)}}2^{(k-k')\frac{\beta}{2(\beta+1)}}.
\end{equation*}
Invoking Cotlar-Stein Lemma we can get \eqref{Orth_2-2}. If $\beta=0$, the inequality above can not guarantee the almost orthogonality, thus we turn to seek for the orthogonality between operators whose supports are in larger regions. Rewrite
\begin{equation*}
T^z_Xf(x)=\sum_{j\ll k}D_{j,k}^Xf(x):=\sum_j D^{X}_jf(x).
\end{equation*}
Suppose $j<j'$, Lemma \ref{Orth_2} implies
\begin{align*}
&\norm{D_{j}^X\left(D^X_{j'}\right)^{*}}\\
&\leq C\lambda^{\frac{2\mu}{n}}\left[\lambda 2^{-j(n-2)}\right]^{-1}\left[2^{-j(n-2)}\right]^{\frac{1}{2}}\left[2^{-j'(n-2)}\right]^{\frac{1}{2}}\notag\\
&=C\lambda^{\frac{2\mu-n}{n}}2^{(j-j')\cdot\frac{n-2}{2}},
\end{align*}
and the almost orthogonality lemma shows that \eqref{Orth_2-2} is ture. Now we proceed to treat $T_Y$, in fact, in this case, for $k\ll j, k'\ll j'$, from Lemma \ref{Orth_2}, we know that
\begin{align}
&\norm{D_{j,k}^Y\left(D^Y_{j',k'}\right)^{*}}\notag\\
&\leq C\lambda^{\frac{2\mu}{n}}\left[\lambda 2^{-j\gamma}2^{-(n-2-\gamma) k}\right]^{-1}\left[2^{-j\gamma-k(n-2-\gamma-\beta)}\right]^{\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(n-\beta-2)(\beta+1)}}\cdot\notag\\
&\quad \left[2^{-j'\gamma-k(n-2-\gamma-\beta)}\right]^{\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(n-2-\beta)(\beta+1)}}\notag\\
&=C\lambda^{\frac{2\mu-n}{n}}2^{j\gamma\frac{2(n-2-\beta)(\beta+1)-[n-2\mu-2(1-\mu)(\beta+1)]}{2(n-2-\beta)(\beta+1)}}2^{-j'\gamma\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(n-2-\beta)(\beta+1)}}\cdot\notag\\
&\quad 2^{k\left[(n-2-\gamma)-(n-2-\gamma-\beta)\frac{n-2\mu-2(1-\mu)(\beta+1)}{(n-2-\beta)(\beta+1)}\right]}.\label{OscEst-2}
\end{align}
In view of these fancy exponents, we list them as follows:
\begin{align*}
&2^{j}:\quad\quad \gamma\frac{2(n-2-\beta)(\beta+1)-[n-2\mu-2(1-\mu)(\beta+1)]}{2(n-2-\beta)(\beta+1)};\\
&2^{j'}:\quad\quad -\gamma\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(n-2-\beta)(\beta+1)};\\
&2^{k}:\quad \quad (n-2-\gamma)-(n-2-\beta-\gamma)\frac{n-2\mu-2(1-\mu)(\beta+1)}{(n-2-\beta)(\beta+1)}
\end{align*}
Similarly, we have the trivial size estimate
\begin{align}
&\norm{D_{j,k}^Y\left(D^Y_{j',k'}\right)^{*}}\notag\\
&\leq \left[2^{-j\gamma-k(n-2-\gamma-\beta)}\right]^{\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(n-\beta-2)(\beta+1)}}\cdot \left[2^{-j'\gamma-k(n-2-\gamma-\beta)}\right]^{\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(n-2-\beta)(\beta+1)}}2^{k(2\mu-1)}2^{-\frac{j+j'}{2}}\notag\\
&=2^{-j\left[\gamma\cdot\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(n-\beta-2)(\beta+1)}-\frac{1}{2}\right]}2^{-j'\left[\gamma\cdot\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(n-\beta-2)(\beta+1)}-\frac{1}{2}\right]}2^{-k\left[(n-2-\gamma-\beta)\cdot\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(n-\beta-2)(\beta+1)}+1-2\mu\right]},\label{SizEst-2}
\end{align}
and list these exponents
\begin{align*}
&2^{j}:\quad\quad -\frac{1}{2}-\gamma\cdot\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(n-\beta-2)(\beta+1)};\\
&2^{j'}:\quad\quad -\frac{1}{2}-\gamma\cdot\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(n-\beta-2)(\beta+1)};\\
&2^{k}:\quad \quad 2\mu-1-(n-2-\gamma-\beta)\cdot\frac{n-2\mu-2(1-\mu)(\beta+1)}{(n-\beta-2)(\beta+1)}.
\end{align*}
Again, we use convex combination to obtain
\begin{align*}
&\norm{D_{j,k}^Y\left(D^Y_{j',k'}\right)^{*}}\\
&\leq {\eqref{OscEst-2}}^{\frac{1}{\beta+1}}\cdot{\eqref{SizEst-2}}^{\frac{\beta}{\beta+1}}.
\end{align*}
These dyadic terms and corresponding exponents are as follows.
\begin{align}
&2^{j}:\quad\quad \frac{\gamma}{\beta+1}-\frac{\beta}{2(\beta+1)}-\gamma\cdot\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(n-\beta-2)(\beta+1)};\tag{a}\\
&2^{j'}:\quad\quad -\frac{\beta}{2(\beta+1)}-\gamma\cdot\frac{n-2\mu-2(1-\mu)(\beta+1)}{2(n-\beta-2)(\beta+1)};\tag{b}\\
&2^{k}:\quad \quad \frac{n-2-\gamma}{\beta+1}+\frac{\beta(2\mu-1)}{\beta+1}-(n-2-\gamma-\beta)\cdot\frac{n-2\mu-2(1-\mu)(\beta+1)}{(n-\beta-2)(\beta+1)}.\tag{c}
\end{align}
After tedious calculation, we can see that
\begin{equation*}
\text{(a)}+\text{(c)}=\text{-(b)}.
\end{equation*}
On account of $k\ll j<j'$, we can conclude
\begin{align*}
\norm{D_{j,k}^X\left(D^X_{j',k'}\right)^{*}}&\leq C\lambda^{-\frac{n-2\mu}{2n(\beta+1)}}2^{j\text{(a)}}2^{j'\text{(b)}}2^{k\text{(c)}}\\
&\leq C\lambda^{-\frac{n-2\mu}{2n(\beta+1)}}2^{-\text{(b)}(j-j')}.
\end{align*}
A similar estimate also holds for $\left(D^X_{j,k}\right)^{*}D_{j',k'}^X$, then Cotlar-Stein Lemma implies \eqref{dampL2-1}.\\
The proofs of \eqref{dampL2-3} and \eqref{dampL2-4} are similar, the only difference is that when we use convex combination we set $\theta=\frac{1-\mu}{n-2\mu-(1-\mu)(\gamma+1)}$ instead of $\frac{1}{\beta+1}$. Thus we complete our proof.
\end{proof}
To get the $L^p$ estimate, the following endpoint estimates are necessary. The proof can be easily verified and details can be found in \cite{xu2021sharp}.
\begin{theorem}\label{EndPo1}
For the damped operators $T^z_X$ and $T_Y^z$, if $\gamma\neq 0$, whenever
\begin{equation*}
D(x,y)=cx^\gamma\prod_{l=1}^s(y-\alpha_lx)^{m_l}\prod_{l=1}^rQ_l(x,y),\quad\quad \,{\rm Re}\,(z)=-\frac{1-\mu}{n-2-\beta},
\end{equation*}
or
\begin{equation*}
D(x,y)=cx^\gamma,\quad\quad \,{\rm Re}\,(z)=-\frac{1-\mu}{\gamma},
\end{equation*}
we always have
\begin{align}
&\norm{T_X^zf}_{L^{1,\infty}}\leq C\norm{f}_{L^1},\label{EndPo-1}\\
&\norm{T_Y^zf}_{L^1}\leq C\norm{f}_{L^1}.\label{EndPo-2}
\end{align}
\end{theorem}
For the sake of interpolation, we also need the following lemma with change of power weights. This lifting trick can be found in \cite{pansamsze1997}, see also \cite{shi2018uniform}, \cite{shi2017sharp} for details of proof.
\begin{lemma}\label{InterpolationLem}
Let $dx$ be the Lebesgue measure on $\R$. Assume $V$ is a linear operator defined on all simple functions with respect to $dx$. If there exist two constant $A_1, A_2>0$ such that
\begin{enumerate}
\item $\norm{Vf}_{L^\infty(dx)}\leq A_1\norm{f}_{L^1(dx)}$ for all simple functions $f$,
\item $\norm{\abs{x}^aVf}_{L^{p_0}(dx)}\leq A_2\norm{f}_{L^{p_0}(dx)}$ for some $1<p_0, a\in \R$ satisfying $ap_0\neq -1$,
\end{enumerate}
then for any $\theta\in (0,1)$, there exists a constant $C=C(a,p_0,\theta)$ such that
\begin{equation}
\norm{\abs{x}^bVf}_{L^{p}(dx)}\leq CA_1^{\theta}A_2^{1-\theta}\norm{f}_{L^{p}(dx)}
\end{equation}
for all simple function $f$, where $b$ and $p$ satisfy $b=-\theta+(1-\theta)a$ and $\frac{1}{p}=\theta+\frac{1-\theta}{p_0}$.
\end{lemma}
We have finished the preparation works for $T_X$ and $T_Y$, it remains to deal with $T_\Delta$. The crucial observation for $T_\Delta$ is that along the lines, on which the Hessian vanish, orthogonaltiy ensure that we can reduce the infinite sum into a finite sum. Thus we claim that
\begin{align}
\|T_\Delta f\|_{L^p}&\leq C\lambda^{-\frac{1-\mu}{n}}\|f\|_{L^p},\quad \quad \frac{n-2\mu}{n-2\mu-(\beta+1)(1-\mu)}\leq p\leq \frac{n-2\mu}{(1-\mu)(\gamma+1)}.\label{T_DelEnd}
\end{align}
This result in fact has been essentially given by \cite{xu2021sharp}, in which $\gamma=0, \beta=0$. If we repeat that proof line by line, we can see that the vulues of $\gamma$ and $\beta$ have no effect on our result, so we omit the tedious manipulation here.\\
Now we give the proof of Theorem \ref{Main-Result} in the case $\gamma\neq 0$ or $\beta\neq 0$.
\begin{proof}
In Theorem \ref{Main-Result}, observe the range of $p$, we first assume $ \frac{n-2\mu}{(1-\mu)(\gamma+1)}\leq 2$, then
\begin{equation*}
\gamma+1\geq \frac{n-2\mu}{2(1-\mu)}.
\end{equation*}
Given \eqref{Degree}, this also implies
\begin{equation*}
\beta+1\leq \frac{n-2\mu}{2(1-\mu)}-\frac{(n-2)\mu+2(1-\mu)}{1-\mu}<\frac{n-2\mu}{2(1-\mu)}.
\end{equation*}
For the right endpoint $p=\frac{n-2\mu}{(1-\mu)(\gamma+1)}$,we choose the damping factor as \eqref{DamFac2}. Combining \eqref{dampL2-3} with \eqref{EndPo-2} and using Stein's complex interpolation yields \eqref{Main-Result} for $T_Y$, while applying Lemma \ref{InterpolationLem} to \eqref{dampL2-4} together with \eqref{EndPo-1} implies \eqref{Main-Result} for $T_X$. The other endpoint $p=\frac{n-2\mu}{n-2\mu-(\beta+1)(1-\mu)}$ can be similary derived by choosing the damping factor as \eqref{DamFac1} and using Stein's complex interpolation or the lifting trick Lemma \ref{InterpolationLem}.
Thus we complete the argument when $ \frac{n-2\mu}{(1-\mu)(\gamma+1)}\leq 2$. \\
Now we proceed with $\frac{n-2\mu}{n-2\mu-(\beta+1)(1-\mu)}\leq 2, \frac{n-2\mu}{(1-\mu)(\gamma+1)}>2$ and reduce them as
\begin{equation*}
\gamma+1< \frac{n-2\mu}{2(1-\mu)},\quad \beta+1\leq \frac{n-2\mu}{2(1-\mu)}.
\end{equation*}
The left endpoint $p=\frac{n-2\mu}{n-2\mu-(\beta+1)(1-\mu)}$ can be obtained by the same argument above wheras the right endpoint $p=\frac{n-2\mu}{(1-\mu)(\gamma+1)}$ shall be given by duality argument. Specifically, if we desire \eqref{Main-Inq} for $p>2$, it suffices to establish
\begin{equation*}
\norm{T^*g}_{L^{p'}}\leq C\norm{g}_{L^{p'}},
\end{equation*}
where the adjoint operator $T^*$ is similar with $T$ and defined by
\begin{equation}
T^{*}g(y)=\int_{\R} e^{-i\lambda S(x,y)}\overline{K(x,y)} \overline{\psi(x,y)}g(x)dx.
\end{equation}
Compare this operator with \eqref{OSO}, they are essentially same by interchanging the roles of $x$ and $y$. So the disired result natually hold if we replace $x$ with $y$, $\gamma$ with $\beta$ in all the above arguments. As for the last case $\frac{n-2\mu}{n-2\mu-(\beta+1)(1-\mu)}> 2, \frac{n-2\mu}{(1-\mu)(\gamma+1)}>2$, \eqref{Main-Inq} follows by duality argument. Therefore we complete our proof.
\end{proof}
\section{Case 2: The Hessian is of the form $c(y-x)^{n-2}$.}
For the remaining case, i.e. $S_{xy}^{''}(x,y)=c(y-x)^{n-2}$, unlike the above arguments, we shall not seperate the operators and turn to establish corresponding estimates on the space which is equipped with more delicate localized properties, for intance, Hardy space $H^1$. The strategy to prove \eqref{Main-Result} in this case is applying Stein's complex interpolation to a class of complex operators for which we establish $L^2\to L^2$ and $H^1\to L^1$ boundedness results with respect to different complex exponents.\\
The class of complex operators we consider here are of the form
\begin{equation*}
D_\lambda^z f(x)=\int_{\R}e^{iS(x,y)}K(x,y)\psi(x,y)\abs{x-y}^{z}\left[1-\phi(\lambda^{\frac{1}{n}}(x-y))\right]f(y)dy.
\end{equation*}
Observe that the support of this operator is outside a neighborhood, having width about $\lambda^{-\frac{1}{n}}$, of the line $y=x$. Thus this operator is essentially a nondegenerate oscillatory integral operator. So the $L^2\to L^2$ decay estimates are comparably easy to prove. In what follows we always assume the Hessian is of the form $c(y-x)^{n-2}$.
\begin{theorem}\label{SpecL2}
If $\,{\rm Re}\,(z)=\frac{n-2}{2}$, then
\begin{equation}\label{SpecDampL2-1}
\norm{D_\lambda^z f}_{L^2(\mathbb{R})}\leq C\lambda^{\frac{\mu}{n}-\frac{1}{2}}\norm{f}_{L^2(\mathbb{R})},
\end{equation}
and if $\mu-1\leq \,{\rm Re}\,(z)<\frac{n-2}{2}$, then
\begin{equation}\label{SpecDampL2-2}
\norm{D_\lambda^z f}_{L^2(\mathbb{R})}\leq C\lambda^{\frac{\mu-1-\,{\rm Re}\,(z)}{n}}\norm{f}_{L^2(\mathbb{R})}.
\end{equation}
\end{theorem}
\begin{proof}
We first prove \eqref{SpecDampL2-2} because it is comparably easy. We decompose the operator $D^z_\lambda$ as
\begin{equation*}
D^z_\lambda f(x)=\sum_k\int_{\R}e^{i\lambda(x-y)^n}K(x,y)\psi(x,y)\abs{x-y}^{z}\Psi_m(y-x)\left[1-\phi(\lambda^{\frac{1}{n}}(x-y))\right]f(y)dy,
\end{equation*}
where the function $\Psi_m$ is same with what we have used in Section 2. On account of the support of $\left[1-\phi(\lambda^{\frac{1}{n}}(x-y))\right]$, the sum over $m$ is in fact a finite sum and $m\lesssim \log(\lambda^{\frac{1}{n}})$. We now invoke \eqref{DampL2} and obtain
\begin{align*}
\norm{D_\lambda^z}&\lesssim \sum_m \lambda^{\frac{\mu}{n}}\left(\lambda 2^{-m(n-2)}\right)^{-\frac{1}{2}}(2^{-m})^{\,{\rm Re}\,(z)}\\
&\lesssim \lambda^{\frac{\mu-1-\,{\rm Re}\,(z)}{n}}.
\end{align*}
Therefore we arrive at \eqref{SpecDampL2-2}. Now we turn to give \eqref{SpecDampL2-1}. Recall the notations of \eqref{T_X} and \eqref{T_Y}, we rewrite $D_\lambda^z$ as
\begin{align*}
D_\lambda^zf(x)&=T_Y^zf(x)+T_X^zf(x)+\sum_{j\sim k}\int_{\R}e^{i\lambda S(x,y)}K(x,y)\abs{y-x}^{z}\left[1-\phi\left(\lambda^{\frac{1}{n}}(x-y)\right)\right]\cdot\\
&\quad\Psi_j(x)\Psi_k(y)\psi(x,y)f(y)dy\\
&:=T_Y^zf(x)+T_X^zf(x)+T_\Delta^zf(x).
\end{align*}
By setting $\beta=0$, \eqref{dampL2-1} and \eqref{dampL2-2} have implied the correponding estimates for $T_Y^z$ and $T_X^z$ above. So it suffices to verify \eqref{SpecDampL2-1} for $T_\Delta^z$. Since
\begin{align*}
T_\Delta^zf(x)&=\sum_{j\sim k}\int_{\R}e^{i\lambda S(x,y)}K(x,y)\abs{y-x}^{z}\left[1-\phi\left(\lambda^{\frac{1}{n}}(x-y)\right)\right]
\Psi_j(x)\Psi_k(y)\psi(x,y)f(y)dy\\
&:=\sum_{j\sim k}D_{j,k}^{\Delta}f(x),
\end{align*}
then by orthogonality, it suffices to focus on one such $D_{j,k}^{\Delta}$. So we further decompose $D_{j,k}^{\Delta}$ as
\begin{align*}
D_{j,k}^{\Delta}f(x)&=\sum_{m}\int_{\R}e^{i\lambda S(x,y)}K(x,y)\abs{y-x}^{z}\left[1-\phi\left((x-y)\lambda^{\frac{1}{n}}\right)\right]\Psi_m(x-y)
\Psi_j(x)\Psi_k(y)\cdot\\
&\quad\psi(x,y)f(y)dy\\
&:=\sum_m D_{j,k,m}^{\Delta}f(x).
\end{align*}
We apply Lemma \ref{Orth_2} to $D_{j,k,m}^{\Delta}$ and get
\begin{equation*}
\norm{D_{j,k,m}^{\Delta}\left(D_{j,k,m'}^{\Delta}\right)^*}\lesssim \lambda^{\frac{2\mu}{n}-1}2^{-|m-m'|/2}.
\end{equation*}
Similar estimates also hold for $\left(D_{j,k,m}^{\Delta}\right)^*D_{j,k,m'}^{\Delta}$ by interchanging the roles of $x$ and $y$. Therefore $\norm{D_{j,k}^{\Delta}}\lesssim \lambda^{\frac{\mu}{n}-\frac{1}{2}}$, this completes the proof.
\end{proof}
To establish the $L^p$ estimate, we also need to prove that the damped oscillatory integral operator with critical negative exponent maps $H^1(\mathbb{R})$ into $L^1(\mathbb{R})$.
\begin{theorem}
If $\,{\rm Re}\,(z)=\mu-1$, then
\begin{equation}
\norm{D_\lambda^z f}_{L^1(\mathbb{R})}\leq C\norm{f}_{H^1(\mathbb{R})}.
\end{equation}
\end{theorem}
\begin{proof}
Suppose that a $H^1$-atom $b$ satisfies
\begin{align*}
&\,{\rm supp}\,{b}\subset I:=\left[C_I-\frac{\abs{I}}{2},C_I-\frac{\abs{I}}{2}\right],\\
&\norm{b}_{L^\infty}\leq \frac{1}{\abs{I}},\\
&\int_{I}b(y)dy=0.
\end{align*}
Therefore
\begin{align*}
\norm{D_\lambda b}_{L^1(\mathbb{R})}=&\int_{\R}\abs{D_\lambda b(x)}dx\\
=&\int_{\abs{x-C_I}\leq 2\abs{I}}\abs{D_\lambda b(x)}dx+\int_{\abs{x-C_I}> 2\abs{I}}\abs{D_\lambda b(x)}dx\\
:=&I_1+I_2.
\end{align*}
By H\"{o}lder's inequality, we know that
\begin{equation*}
I_1\leq \abs{I}^{\frac{1}{2}}\norm{D_\lambda b}_{L^2}\leq C\abs{I}^\frac{1}{2}\norm{b}_{L^2}\leq C.
\end{equation*}
The second inequality results from Theroem \ref{SpecL2}. We now focus on $I_2$. For simplicity, we set
\begin{equation*}
\Phi(x,y)=K(x,y)\psi(x,y)\abs{x-y}^{\mu-1}.
\end{equation*}
Thus
\begin{align*}
I_2\leq &\int_{\abs{x-C_I}> 2\abs{I}}\abs{\int_{\R}e^{i\lambda(x-y)^n}\left[\Phi(x,y)-\Phi(x,C_I)\right]\left[1-\phi(\lambda^{\frac{1}{n}}(x-y))\right]b(y)dy}dx+\\
&\int_{\abs{x-C_I}> 2\abs{I}}\abs{\int_{\R}e^{i\lambda(x-y)^n}\Phi(x,C_I)\left[1-\phi(\lambda^{\frac{1}{n}}(x-y))\right]b(y)dy}dx\\
:=&I_3+I_4.
\end{align*}
To bound $I_3$, we need to analyze the difference between $\Phi(x,y)$ and $\Phi(x,C_I)$. In fact, from mean value theorem, we know that
\begin{align*}
\abs{\Phi(x,y)-\Phi(x,C_I)}\leq &C\left(\abs{\partial_yK(x,
\xi)}\abs{x-\xi}^{\mu-1}\abs{\psi(x,\xi)}+\abs{K(x,
\xi)}\abs{x-\xi}^{\mu-2}\abs{\psi(x,\xi)}+\right.\\
&\left.\abs{K(x,
\xi)}\abs{x-\xi}^{\mu-1}\abs{\partial_y\psi(x,\xi)}\right)\abs{y-C_I}\\
\lesssim &\abs{I}\left(\abs{\partial_yK(x,
\xi)}\abs{x-\xi}^{\mu-1}\abs{\psi(x,\xi)}+\abs{K(x,
\xi)}\abs{x-\xi}^{\mu-2}\abs{\psi(x,\xi)}+\right.\\
&\left.\abs{K(x,
\xi)}\abs{x-\xi}^{\mu-1}\abs{\partial_y\psi(x,\xi)}\right)\\
\leq &\abs{I}\left(\abs{x-\xi}^{-2}\norm{\psi}_{L^\infty}+\abs{x-\xi}^{-2}\norm{\psi}_{L^\infty}+\abs{x-\xi}^{-1}\norm{\partial_y\psi(x,\cdot)}_{L^1}\right).
\end{align*}
Taking absolute value for every term in the integrand of $I_3$, by means of the upper bound above, we can conclude that
\begin{align*}
I_3\lesssim &\int_{\abs{x-C_I}\geq 2\abs{I}}\abs{I}\left(2\abs{x-\xi}^{-2}\norm{\psi}_{L^\infty}+\abs{x-\xi}^{-1}\norm{\partial_y\psi(x,\cdot)}_{L^1}\right)dx\\
\lesssim &\abs{I}\left(2\abs{I}^{-1}\norm{\psi}_{L^\infty}+\abs{I}^{-1}\norm{\partial_y\psi}_{L^1}\right)\\
\leq &C.
\end{align*}
For $I_4$, according to the length of the interval $I$, we divide our argument into two different cases.\\
\emph{Case 1:}$\abs{I}\geq 1$.
On account of the support of $1-\phi(\lambda^{\frac{1}{n}}(x-y))$, and also
\begin{equation*}
\abs{\Phi(x,C_I)}\lesssim \abs{x-C_I}^{-1},
\end{equation*}
H\"{o}lder's inequality implies
\begin{align*}
I_4\lesssim &\left(\int_{\abs{x-C_I}>\max\{2\abs{I},\lambda^{-\frac{1}{n}}\}}\abs{x-C_I}^{-2}dx\right)\lambda^{-\frac{1}{n}}\norm{b}_{L^2}\\
\lesssim &\min\left\{\lambda^{\frac{1}{2n}}, \abs{I}^{\frac{1}{2}}\right\}\lambda^{-\frac{1}{n}}\abs{I}^{-\frac{1}{2}}\\
\leq &C.
\end{align*}
Otherwise, if $\abs{I}<1$, then
\begin{align*}
I_4\leq &\int_{\abs{x-C_I}> 2\abs{I}}\abs{x-C_I}^{-1}\abs{\int_{\R}e^{i\lambda(x-y)^n}\left[1-\phi(\lambda^{\frac{1}{n}}(x-y))\right]b(y)dy}dx\\
\leq &\int_{\abs{x-C_I}> 2\abs{I}}\abs{x-C_I}^{-1}\abs{\int_{\R}e^{i\lambda(x-y)^n}\left[\phi(\lambda^{\frac{1}{n}}(x-C_I))-\phi(\lambda^{\frac{1}{n}}(x-y))\right]b(y)dy}dx+\\
&\int_{\abs{x-C_I}> 2\abs{I}}\abs{x-C_I}^{-1}\abs{\int_{\R}e^{i\lambda(x-y)^n}\left[1-\phi(\lambda^{\frac{1}{n}}(x-C_I))\right]b(y)dy}dx\\
:=&I_5+I_6.
\end{align*}
Similar to what we have done for $I_3$, it is necessary to analyze the difference between $\phi(\lambda^{\frac{1}{n}}(x-C_I))$ and $\phi(\lambda^{\frac{1}{n}}(x-y))$. By means of mean value theorem, it is easy to check that
\begin{equation*}
\abs{\phi(\lambda^{\frac{1}{n}}(x-C_I))-\phi(\lambda^{\frac{1}{n}}(x-y))}\leq \lambda^{\frac{1}{n}}\abs{\partial_y\phi(\lambda^{\frac{1}{n}}(x-\xi))}\abs{y-C_I}.
\end{equation*}
Since the support of $\partial_y\phi(\lambda^{\frac{1}{n}}(x-y))$ is restricted in the region $\abs{x-\xi}\approx \lambda^{-\frac{1}{n}}$, so
\begin{align*}
I_5\leq &\int_{\abs{x-C_I}> 2\abs{I}} \abs{x-C_I}^{-1}\lambda^{\frac{1}{n}}\abs{\partial_y\phi(\lambda^{\frac{1}{n}}(x-\xi))}\abs{y-C_I}\abs{y-C_I}\norm{b}_{L^1}dx\\
\leq &\abs{I}^{-1}\norm{\partial_y\psi}_{L^\infty}\abs{I}\lambda^{\frac{1}{n}}\lambda^{-\frac{1}{n}}\\
\leq & C.
\end{align*}
It remains to deal with $I_6$, actually
\begin{align*}
I_6\leq & \int_{\abs{x-C_I}>\max\{2\abs{I},\lambda^{-\frac{1}{n}}\}}\abs{x-C_I}^{-1}\abs{\int e^{i\lambda(x-y)^n}b(y)dy}dx\\
=&\int_{r>\abs{x-C_I}>\max\{2\abs{I},\lambda^{-\frac{1}{n}}\}}\abs{x-C_I}^{-1}\abs{\int e^{i\lambda(x-y)^n}b(y)dy}dx+\\
&\int_{r\leq\abs{x-C_I}}\abs{x-C_I}^{-1}\abs{\int e^{i\lambda(x-y)^n}b(y)dy}dx\\
:=&I_7+I_8.
\end{align*}
Here, $r$ is a parameter which will be determined later. Since
\begin{equation*}
\abs{e^{i\lambda(x-y)^n}-e^{i\lambda(x-C_I)^n}}\leq \abs{\lambda(x-\xi)^{n-1}}\abs{y-C_I},
\end{equation*}
then by the vanishing property of the atom $b$, we have
\begin{align*}
I_7=&\int_{r>\abs{x-C_I}>\max\{2\abs{I},\lambda^{-\frac{1}{n}}\}}\abs{x-C_I}^{-1}\abs{\int \left[e^{i\lambda(x-y)^n}-e^{i\lambda(x-C_I)^n}\right]b(y)dy}dx\\
\leq & \int_{r>\abs{x-C_I}>\max\{2\abs{I},\lambda^{-\frac{1}{n}}\}}\abs{x-C_I}^{-1}\abs{\lambda(x-\xi)^{n-1}}\abs{y-C_I}\norm{b}_{L^1}dx\\
\leq &\int_{r>\abs{x-C_I}>\max\{2\abs{I},\lambda^{-\frac{1}{n}}\}}\lambda\abs{x-C_I}^{n-2}\abs{I}\norm{b}_{L^1}dx\\
\lesssim &\lambda\abs{I}r^{n-1}.
\end{align*}
If we set
\begin{equation*}
\lambda\abs{I}r^{n-1}=1,
\end{equation*}
then $I_7\leq 1.$ Now we continue to treat $I_8$, by change of variables,
\begin{align*}
I_8\approx &\sum_{l\geq 1}\int_{2^lr\leq \abs{u}\leq 2^{l+1}r}\frac{1}{2^lr}\abs{\int e^{i\lambda(u-\abs{I}v)^n}\abs{I}b\left(\abs{I}v+C_I\right)dv}du\\
=&\sum_{l\geq 1}\int_{1\leq \abs{\tilde{u}}\leq 2}\abs{\int e^{i\lambda(2^lr\tilde{u}-\abs{I}v)^n}\abs{I}b\left(\abs{I}v+C_I\right)dv}d\tilde{u}\\
\leq &\sum_{l\geq 1}\left(\lambda\abs{I}\left(2^lr\right)^{n-1}\right)^{-\epsilon}\norm{\abs{I}b\left(\abs{I}v+C_I\right)}_{L^2(dv)}\\
\leq &\left(\sum_{l\geq 1}2^{-\epsilon l}\right)\abs{I}\abs{I}^{-1}\left(\lambda\abs{I}r^{n-1}\right)^{-\epsilon}\\
\lesssim &1.
\end{align*}
Here we employ the uniform operator van der Corput lemma in \cite{carbery1999multidimensional}, thus our proof is complete.
\end{proof}
{\bf Acknowledgement:} The author would like to acknowledge financial support from Jiangsu Natural Science Foundation, Grant No. BK20200308.
|
\section*{Appendix}
\subsection{Proofs}
\label{appendix:proofs}
\textbf{Proposition~\ref{prop:safety2}:} Under Assumption 1, Algorithm~\ref{alg:general2} is $\epsilon + 1/(1+|{\mathcal{A}}|)$-safe (with respect to ${\hat{Y}}, {\hat{Z}}$).
\begin{proof
Given a sequence of data points $(Y_1, Z_1), \cdots, (Y_T, Z_T)$, denote the subsequence of ``unsafe'' data as $(Y_{c_1}, Z_{c_1}), \cdots, (Y_{c_M}, Z_{c_M})$ where $Z_{c_t}$ is the $t$-th unsafe example (i.e. $f(Z_{c_t}) < f_0$), so $M = |{\mathcal{A}}|$. Suppose that $\hat{Z}$ is also unsafe, i.e. $f(\hat{Z}) < f_0$. Let $\lbag \cdot \rbag$ denote an unordered bag (i.e. it is a set that can have repeated elements). We can bound the safety by
\begin{align*}
\Pr[w(\hat{Y}) = 0] &= \Pr[q > 1 - \epsilon] \\
&= {\mathbb{E}}\Bigl[ \Pr\bigl[q > 1 - \epsilon \mid \lbag (Y_{c_1}, Z_{c_1}), \cdots, (Y_{c_M}, Z_{c_M}), (\hat{Y}, \hat{Z}) \rbag \bigr] \Bigr] & \text{Tower} \\
&= {\mathbb{E}}\Bigl[ \Pr\bigl[ |\lbrace t \mid g(Y_{c_t}) < g(\hat{Y}) \rbrace| + U > (1 - \epsilon)(M+1) \\
&\qquad - 1 \mid \lbag (Y_{c_1}, Z_{c_1}), \cdots, (Y_{c_M}, Z_{c_M}), (\hat{Y}, \hat{Z}) \rbag \bigr]\Bigr] & \text{Definition}
\end{align*}
By the assumption of exchangeability we are equally likely to observe any permutation of $\lbag (Y_{c_1}, Z_{c_1}), \cdots, (Y_{c_M}, Z_{c_M}), (\hat{Y}, \hat{Z}) \rbag$. Intuitively, it is equally likely for $g(\hat{Y})$ to be the largest, 2nd largest, etc, among $g(Y_{c_1}), \cdots, g(Y_{c_M}), g(\hat{Y})$. Formally, the random variable $|\lbrace t \mid g(Y_{c_t}) < g(\hat{Y}) \rbrace| + U$ takes on all values $\lbrace 0, 1, \cdots, M \rbrace$ with equal probability. Therefore,
\begin{align*}
&\Pr\Bigl[ |\lbrace t \mid g(Y_{c_t}) < g(\hat{Y}) \rbrace| + U > (1 - \epsilon)(M+1) \\
&\qquad\qquad - 1 \mid \lbag (Y_{c_1}, Z_{c_1}), \cdots, (Y_{c_M}, Z_{c_M}), (\hat{Y}, \hat{Z}) \rbag \Bigr] \\
&\qquad = 1 - \Pr\Bigl[ |\lbrace t \mid g(Y_{c_t}) < g(\hat{Y}) \rbrace| + U \leq (1 - \epsilon)(M+1) \\
&\qquad\qquad - 1 \mid \lbag (Y_{c_1}, Z_{c_1}), \cdots, (Y_{c_M}, Z_{c_M}), (\hat{Y}, \hat{Z}) \rbag \Bigr] \\
&\qquad \leq 1 - \frac{\lceil(1 - \epsilon)(M+1)-1 \rceil}{M+1} \\
&\qquad = \frac{\lfloor M+1 - (1-\epsilon)(M+1) + 1 \rfloor}{M+1} \\
&\qquad \leq \frac{1 + \epsilon M + \epsilon}{M+1} \\
&\qquad = \epsilon + \frac{1}{M+1}
\end{align*}
We can combine this with the original result to get
\begin{align*}
\Pr[w(\hat{Y}) = 0]
&= {\mathbb{E}}\Bigl[ \Pr\bigl[ |\lbrace t \mid g(Y_{c_t}) < g(\hat{Y}) \rbrace| + U > (1 - \epsilon)(M+1) \\
&\qquad - 1 \mid \lbag (Y_{c_1}, Z_{c_1}), \cdots, (Y_{c_M}, Z_{c_M}), (\hat{Y}, \hat{Z}) \rbag \bigr] \Bigr] \\
&\leq {\mathbb{E}}\left[\epsilon + \frac{1}{M+1} \right] \\
&= \epsilon + \frac{1}{M+1}
\end{align*}
\end{proof}
\subsection{Lower bound on the false positive rate}
\label{appendix:lower_bound}
Consider a function $w$ that maps a dataset ${\mathcal{D}} = (g(X_1), Y_1), \cdots, (g(X_T), Y_T)$ of unsafe examples, and a new data point $g({\hat{X}})$, to $\lbrace 0, 1\rbrace$. We argue that any $w$ that gives a distribution-free false negative rate guarantee should depend only on the ordering between $g(X_1), \cdots, g(X_T), g({\hat{X}})$, and not on their specific values. In other words, $w$ should take the form defined by
\begin{align}
w({\mathcal{D}}, {\hat{Z}}) = \left\lbrace \begin{array}{ll} \phi \left( \# \lbrace t, g({\hat{X}}) < g(X_t) \rbrace \right) & \text{ with probability } \gamma \\
1 & \text{ with probability } 1-\gamma \end{array} \right. \label{eq:test_function_form}
\end{align}
for some deterministic function $\phi$ and real number $\gamma$. We know that when the data is exchangeable, $\# \lbrace t, g({\hat{X}}) < g(X_t) \rbrace$ is uniformly distributed on $\lbrace 0, 1, \cdots, T \rbrace$.
\textbf{Case 1} Suppose $\phi$ takes the value $0$ for at least one possible input; then the false negative rate is given by
\begin{align}
\text{FNR} \geq \gamma/(1+T)
\end{align}
and the false positive rate is given by
\begin{align}
\text{FPR} \geq 1-\gamma
\end{align}
so combined we have
\begin{align}
\text{FPR} \geq 1-\gamma \geq 1 - (1+T) \text{FNR} \geq 1 - (1+T) \epsilon
\end{align}
\textbf{Case 2} Suppose $\phi$ takes the value $0$ for none of the inputs; then the false negative rate is given by
\begin{align}
\text{FNR} = 0, \text{FPR} = 1
\end{align}
so we would still (trivially) have $\text{FPR} \geq 1-(1+T)\epsilon$.
So far we have shown that if $w$ were to take the specific form of Eq. (\ref{eq:test_function_form}), then the false positive rate must be lower bounded by $1-(1+T)\epsilon$. In other words, when $\epsilon = o(1/T)$, the false positive rate tends to $1$ when $T$ is large.
\subsection{Additional Experimental Details: Driver Alert System}
\textbf{Safety score:}
We define the safety score by the Mahalanobis distance between the ego-vehicle and the agent, where the first eigenvector is aligned with the ego-vehicle's velocity vector, and the second eigenvector is orthogonal to the ego-vehicle; the magnitude of the first eigenvector is the magnitude of the velocity, and the magnitude of the second eigenvector is approximately half of a car width (we use 1m). Intuitively, this means that agents that are along the ego-vehicle's velocity vector appear closer than agents in the perpendicular direction. This metric is similar to time to collision (TTC), but it is continuous whereas TTC is not --- TTC is infinite unless two vehicles are exactly on a collision course.
\textbf{Dataset details: }
The nuScenes dataset includes 952 scenes collected across Boston and Singapore, divided into a 697/105/150 train/val/test split (the same split used for the original Trajectron++). Each scene is 20 seconds long. The Kaggle Lyft Motion Prediction dataset is a subset of the full Lyft Level 5 dataset (chosen over the full dataset for computational reasons). It includes approximately 16k scenes, divided into an 70\%/15\%/15\% train/val/test split. Each scene is 25 seconds long. Both datasets include labeled ego-vehicle trajectories as well as labeled detections and trajectories for other agents in the scene. Note that for both of these datasets, because the training split was used to train the Trajectron++ model, we used the validation split as the input training data for Algorithm~\ref{alg:general2}.
\textbf{Additional experimental results:}
We demonstrate empirically on the nuScenes dataset that the sum of $\epsilon$ and the false positive rate must be high when there are few (e.g. $< 1/T$) samples, which is consistent with what our theory from Section~\ref{sec:fpr} would predict. Figure~\ref{fig:fpr_vs_samples} plots the epsilon bound as well as the false negative and false positive rates vs. the number of unsafe samples in the validation dataset; we see that when $\epsilon$ decreases as $1/T$, the false positive rate is relatively flat and low.
\begin{figure}[h]
\centering
\includegraphics[width=0.5\linewidth]{figures/vary_samples.pdf}
\caption{Epsilon bound, false negative rate, and false positive rate on the nuScenes dataset while varying the number of unsafe samples. Consistent with our theory from Section~\ref{sec:fpr}, the sum of $\epsilon$ and the false positive rate is high when there are few samples. }
\label{fig:fpr_vs_samples}
\end{figure}
We also demonstrate empirically on the Kaggle Lyft dataset that the variance on the false negative rate over different train/test splits is low. Table~\ref{tab:variance} displays the variance on the false negative rate calculated over the 100 trials at each $\epsilon$ value. All of the variances are well below 0.003, suggesting that the test sequence false negative rates are clustered around $\epsilon$ (rather than having some sequences that fail on zero examples and others with catastrophic failures). As further evidence, in Figure~\ref{fig:box-plot}, we provide a representative box plot of the false negative rates over the 100 trials with $\epsilon = 0.04$. The variances are indeed clustered around $0.04$.
\begin{table}[h]
\centering
\setlength{\tabcolsep}{0.5em}
\begin{tabular}{c | c c c c c}
$\epsilon$ & $0.02$ & $0.04$ & $0.06$ & $0.08$ & $0.10$ \\
\hline
Variance & 0.00096 & 0.0019 & 0.0014 & 0.0023 & 0.0024 \\
\end{tabular}
\vspace{2mm}
\caption{\label{tab:variance} Variance on the test sequence false negative rates at different $\epsilon$.}
\end{table}
\vspace{-2mm}
\begin{figure}[H]
\centering
\includegraphics[width=0.5\linewidth]{figures/box_plot.pdf}
\caption{Box plot of the 100 false negative rates calculated over randomized train/test splits with $\epsilon = 0.04$.}
\label{fig:box-plot}
\end{figure}
\subsection{Additional Experimental Details: Robotic Grasping Experiments}
\textbf{Model and dataset details: }
The Grasp Quality Convolutional Neural Network (GQ-CNN) from~\cite{mahler2017dex} is a model that classifies whether a candidate robotic grasp will be successful. The inputs to a GQ-CNN are a point cloud representation of an object, $\mathbf{y}$, and a candidate grasp, $\mathbf{u}$. A GQ-CNN outputs the predicted probability, $Q_{\theta}(\mathbf{y}, \mathbf{u})$, that the candidate grasp will be able to successfully pick and transport the object.
We use this predicted probability as the safety score, $g = Q_{\theta}(\mathbf{y}, \mathbf{u})$. We consider a candidate grasp ``unsafe'' if it will not be able to successfully pick the object
(i.e.\ the true label is $Z = 0$). Note that this is exactly the ROC curve threshold tuning setup.
We use the DexNet dataset of synthetic objects grasped with a parallel jaw gripper~\cite{mahler2017dex}, which includes approximately 500k pick attempts not used in training the GQ-CNN model. These are divided into a 50\%/50\% train/test split. Each example is labeled a success if the robot successfully picks and places the object, and a failure otherwise.
\section{Introduction}
Monitoring a system for faults, or detecting if unsafe situations will occur is a key problem for high-stakes robotics applications, and indeed the field of fault detection has long been the state of practice for building reliable systems~\cite{visinsky1994robotic,visinsky1994expert,visinsky1995dynamic,vemuri1998neural,khalastchi2018fault,muradore2011pls,crestani2015enhancing,Ding2013fault,Patton1997Observer,Harichi2015Model,Harirchi2017Guaranteed}. With the advent of learning-enabled components in robotic systems, robots are performing increasingly complex safety-critical tasks, so reliability has become increasingly important. At the same time, it is less clear how to ensure reliability for these learned systems.
In this work, we present a sample efficient and principled method for detecting unsafe situations based on the statistical inference technique of conformal prediction~\cite{Vovk2005Algorithmic}. Our method provides \textit{provable} false negative rates for warning systems (i.e.\ among the situations in which an alert should be issued, fewer than $\epsilon$ occur without an alert), while achieving low false positive rates (few unnecessary alerts are issued).
As a running example, we use our method to design an alert system to warn a human operator of impending danger in a driving application (illustrated in Figure~\ref{fig:system}).
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{figures/system_4.pdf}
\caption{We design a warning system that achieves a \textit{provable} false negative rate sample efficiently. Among the situations that are dangerous (i.e.\ lead to an unsafe future situation in the absence of corrective action), fewer than $\epsilon$ occur without an alert.}
\label{fig:system}
\end{figure}
\textbf{Related Work} Traditional fault detection techniques include hardware redundancy, signal processing, and plausibility tests~\cite{Ding2013fault,visinsky1994robotic,visinsky1994expert,visinsky1995dynamic,vemuri1998neural}. However,
hardware redundancy requires extra components, signal processing works well only for processes in steady state, and plausibility tests do not catch faults that lead to a physically plausible system.
Additionally, these methods typically lack performance guarantees. Model-based fault detection techniques~\cite{Ding2013fault,Patton1997Observer,Harichi2015Model,Harirchi2017Guaranteed} involve using a model of the system to determine whether a fault has occurred; they assume that users have a very accurate model of the system dynamics, which is difficult to obtain in practice.
Another common approach for detecting unsafe states employs supervised learning to train a classifier model for labeling states as unsafe, and then the classifier hyperparameters are adjusted until empirically the false negative rate is low. In practice this is typically accomplished by plotting a receiver operating characteristic (ROC) curve and tuning the classification threshold to achieve low false negative rate. However, this approach requires training a new classification model, and provides no performance guarantees.
To guarantee the false negative rate of a learned warning system, the standard statistical learning framework could be used under standard i.i.d. assumptions~\cite{Luxburg2011Statistical}. A practitioner could collect additional data and use a validation dataset to provably certify the false negative rate. However, the key problem is \textit{data efficiency}, because collecting data for unsafe situations can be very expensive~\cite{Foody2009Sample,Luxburg2011Statistical,Calfiore2006Scenario}.
\textbf{Main Question} Can we tune a warning system and guarantee low false negative rate with only a handful of data points?
For example, with only 30 data samples of dangerous situations, can we tune a warning system to have \textit{provable} 5\% false negative rate? This problem is easy if we allow trivial systems that always issue a warning, but such systems are not practically useful.
If we restrict our attention to non-trivial systems,
this problem is seemingly impossible because even if a \textit{fixed} warning system successfully identifies all 30 dangerous situations, due to statistical fluctuations, we cannot prove that its false negative rate is less than 5\% (with high confidence). If proving that a fixed predictor achieves safety is difficult, tuning a predictor to provably achieve safety seems only more challenging.
\textbf{Our Contribution} We answer our main question affirmatively. We adapt a statistical inference framework known as conformal prediction to a robotics setting in order to tune systems to achieve provable safety guarantees (e.g. 5\% false negative rate) with extremely limited data (e.g. 30 samples). We only require a single assumption: the training samples are exchangeable with each test sample, i.e. for each test sample, if we permute the concatenated sequence of the training samples and the test sample, there is no reason to believe that any permutation is more or less likely to occur. This is a weaker assumption than the i.i.d. assumption typically used in the statistical learning framework.
The assumption is reasonable in practice, even in situations in which the statistical inference assumptions fail (e.g.\ situations with temporal correlations between different test samples). In a driving scenario, for instance, the training dataset could include scene snippets sampled from different scenes, and these will be i.i.d. At test time, there may be one scene with several snippets. These snippets are obviously not independent; however, they are individually exchangeable with the training dataset.
While this answer seems too good to be true, the key insight here is that we provide a type of guarantee that is different from standard statistical learning guarantees.
Consider a sequence of test samples $Z_1, \cdots, Z_N$, and event indicators $F_1, \cdots, F_N$ for whether our warning system fails on each test sample.
\begin{itemize}
\item \textbf{Statistical learning guarantee.} In the statistical learning framework, we assume that the test samples $Z_1, \cdots, Z_N$ are i.i.d., so the failure events $F_1, \cdots, F_N$ are also i.i.d. --- we guarantee the failure probability for a sequence of i.i.d. failure events.
\item \textbf{Conformal prediction guarantee.} In the conformal prediction framework, the test samples are not necessarily i.i.d., so the failure events $F_1, \cdots, \allowbreak F_N$ can be correlated --- we guarantee the \textit{marginal} failure probability for each failure event. In other words, we know that each test sample has a low probability of failure (i.e. $F_n = 1$ with low probability), but the failures could be correlated. For example, conditioning on $F_n = 1$ might increase or decrease the probability that $F_{n+1} = 1$ (while in the i.i.d. case, $F_n$ and $F_{n+1}$ are independent events).
\end{itemize}
The usefulness of the conformal guarantee depends on the intended application. Consider the driver alert system example: for individual drivers, collisions are rare and most drivers will not encounter more than one. Hence, there is little reason to worry about whether the warning failures are correlated between collisions. In other words, the conformal guarantee can convey confidence to individual users who rarely encounter multiple failures.
On the other hand, the conformal guarantee may convey less confidence to a company with a large fleet of vehicles. For example, if $F_n = 1$ increases the probability that $F_{n+1} = 1$, then it is possible to have multiple simultaneous failures. However, this is not a limitation of our method, but rather an unavoidable consequence of the weaker (not i.i.d.) assumptions: if the test data is correlated (which we have no control over), then failure events of a warning system are inherently correlated. The weaker assumption is usually necessary because most robotics applications are deployed in time series or sequential decision making setups, so data from nearby time steps are correlated and not i.i.d.
Since standard statistical learning guarantees are not applicable due to violation of the i.i.d. assumption, having some (conformal) guarantee is better than none.
Furthermore, we will show empirically in Section~\ref{sec:experiments-driving} that failures are not highly correlated on two real-world driving datasets. Therefore, despite the lack of formal guarantees, there is strong empirical evidence suggesting that simultaneous failures do not occur in practice.
Thus, our contribution is four-fold: 1) We introduce a new notion of safety guarantee that is satisfactory for many use cases and has extremely good sample efficiency. 2) We show how to leverage the statistical inference tool of conformal prediction for robotics applications. 3) We instantiate a framework for applying conformal prediction to robotic safety.
4) We validate our framework experimentally on both a driver alert safety system and a robotic grasping system, showing that the conformal guarantees hold in practice, without issuing too many false positive alerts (e.g. less than 1\% for many setups).
\textbf{Organization} The rest of this paper is organized as follows. In Section~\ref{sec:background}, we review conformal prediction. In Section~\ref{sec:framework}, we describe our problem setup, introduce our framework, and demonstrate that specific choices for elements of our framework lead to instantiations such as tuning an ROC curve threshold to limit false negatives (though we enrich this classic method with new guarantees). We then explain the differences between the conformal prediction guarantees and the statistical learning guarantees, and discuss when our guarantees should be applied. Finally, in Sections~\ref{sec:experiments-driving} and~\ref{sec:experiments-grasp}, we evaluate our framework on a driver alert safety system and on a robotic grasping system.
\section{Overview of Conformal Prediction}
\label{sec:background}
This section provides an overview of conformal prediction, the general framework that we adapt for robotics safety. It may be skipped without breaking the flow of the paper.
Consider a prediction problem where the input feature is denoted by $X$ and the label is denoted by $Z$. Conformal prediction~\cite{Shafer2008Tutorial} is a class of methods that can produce prediction sets (i.e. a set of labels), such that the true label belongs to the predicted set with high probability.
In its standard form,
conformal prediction requires two components: a sequence of validation data $(X_1, Z_1), \cdots, (X_T, Z_T)$ and a non-conformity score $\psi$, which is any function from the input feature $X$ and the label $Z$ to a real number. Intuitively, the non-conformity score should measure the ``unusualness'' of the label $Z$ when the input feature is $X$. An example non-conformity score is $\psi(X, Z) = |h(X) - Z|$ where $h$ is some fixed prediction function --- intuitively, $Z$ is ``unusual'' if the prediction function has large error.
The conformal prediction algorithm computes the non-conformity score for all samples in the validation set. Given a new test example with input feature ${\hat{X}}$, the conformal prediction algorithm then ``tries'' all possible labels $z$, and measures the non-conformity score $\psi({\hat{X}}, z)$. A label is rejected if the computed non-conformity score is greater than $1 - \epsilon$ of the non-conformity scores in the validation set. Any label that is not rejected is included in the prediction set. Intuitively, the true label is unlikely to have a non-conformity score higher than $1 - \epsilon$ of validation samples; hence the true label is unlikely to get rejected.
If the training data and the new test data point $({\hat{X}}, {\hat{Z}})$ are exchangeable, i.e. the probability of observing any permutation of $(X_1, Z_1), \cdots, (X_T, Z_T), ({\hat{X}}, {\hat{Z}})$ is equally likely, then conformal prediction has very strong validity guarantees: the true label will be within the
prediction set with $1 - \epsilon \pm 1/(T+1)$ probability. We note that this guarantee holds regardless of the nonconformity function $\psi$.
There are many extensions of conformal prediction, and the most relevant extension to our safety application is
Mondrian conformal prediction~\cite{Vovk2003Mondrian,Vovk2005Algorithmic}, which partitions the input data into several categories such that each data point belongs to exactly one category, and guarantees validity separately for each category. Our work is based on Mondrian conformal prediction; because we wish to limit the false negative rate in warning systems, we need class-conditional validity for samples in the ``unsafe'' class.
Works that apply conformal prediction to robotics settings include~\cite{Chen2020ReactiveMP,Cai2020RealtimeOD,Nouretdinov2011MRI,Gammerman2008ClinicalMS}. ~\cite{Chen2020ReactiveMP} uses conformal prediction to predict a set of possible future motion trajectories from out of a set of 17 basis trajectories, ~\cite{Cai2020RealtimeOD} uses some ideas from conformal prediction for detecting out of distribution samples in cyber-physical systems, and~\cite{Nouretdinov2011MRI,Gammerman2008ClinicalMS} use conformal prediction for medical diagnosis. However, these works consider very different targeted problems, while we consider the problem of warning systems and provide a general framework for using conformal prediction on a variety of robotics applications.
\vspace{-1mm}
\section{Conformal Prediction Framework for Robotics Applications}
\label{sec:framework}
\subsection{Problem Setup}
\label{sec:setup}
We consider a model-based planning application where we have some existing simulator or model, and given the current observations (denoted by random variable $X$), the simulator or model predicts the future states of the system (denoted by $Y$) in the absence of a warning. For instance, many applications have off-the-shelf simulators: an autonomous driving software might simulate the future trajectories of all traffic participants (up to some time horizon), or an aircraft control software might forward simulate the dynamics of the aircraft. We will use the random variable $Z$ to denote the true unknown future states of the system in the absence of a warning, e.g. the true future trajectories of traffic participants, or the true future dynamics of an aircraft.
In our setup, depending on the model or simulator available, $Y$ could have the same type as $Z$ (e.g. both $Y$ and $Z$ are random variables that represent the future trajectories of traffic participants), or $Y$ could have a different type from $Z$ (e.g.\ $Y$ might represent some but not all aspects about the future, such as the direction of movement for traffic participants, or the distance from collision).
\textbf{Assessing Safety} We assume that if we know the \textit{true} future state of the system $Z$, we can assess whether it is safe or not. Specifically, there exists some \textbf{safety score} denoted by $f(Z)$; we specify some threshold (denoted by $f_0$), and wish to be alerted if the safety score drops below this threshold (i.e.\ if $f(Z) < f_0$). Most applications have natural safety scores. For instance, an autonomous driving safety score $f$ could be the distance to or time from collision; an aircraft control safety score could be the (negative) absolute difference between the orientation of the aircraft and its ideal orientation.
In addition, we assume that the user provides a \textbf{surrogate safety score} $g: Y \mapsto \mathbb{R}$ that maps from the simulator prediction to a ``safety score,'' where a higher score indicates ``safe'' and a lower score indicates ``unsafe''. Ideally the surrogate safety score $g(Y)$ should be highly correlated with the true safety score $f(Z)$, but technically $g$ can be any function. None of our technical results depend on any assumptions about $g$; however, the choice of $g$ affects the empirical performance in terms of false positive rate (i.e. how often our warning system issues unnecessary alerts). When $Y$ and $Z$ have the same type, we can simply choose $g := f$; when $Y$ and $Z$ have different types we need to choose $g$ on a case-by-case basis.
\textbf{Warning Function} We wish to design a warning function (denoted as a function $w(Y)$) that given the simulation or model output $Y$, decides to issue a warning ($w(Y) = 1$) or not ($w(Y) = 0$). Note that $Y$ depends on previous states, so the warning function implicitly depends on previous observations through $Y$.
Formally we define ``safety'' as the following requirement:
\begin{definition}
\label{def:safe}
For some $0 < \epsilon < 1$, we say that the warning system $w$ is $\epsilon$-safe (with respect to $Y,Z$, $f$, and $f_0$) if
\begin{align*}
\Pr[w(Y) = 1 \mid f(Z) < f_0] \geq 1 - \epsilon.
\end{align*}
\end{definition}
In words, whenever the true future safety score $f(Z)$ is below $f_0$, the warning system should issue a warning ($w(Y) = 1$) with at least $1-\epsilon$ probability. Another way to think of this is that the false negative rate is at most $\epsilon$. The main difficulty here is that the warning function $w$ can depend on only the simulated future $Y$ rather than the true future $Z$ (which is not yet observed when the warning is issued), and the simulation might not come with any performance guarantees.
A trivial warning system that always issues a warning (i.e. $w_{\mathrm{trivial}}(Y) \equiv 1$) is always $\epsilon$-safe for any $\epsilon > 0$. However, such a warning system is not useful. A useful warning system should issue as few warnings as possible when safe. Therefore, we should also consider its false positive rate
\begin{align*}
\mathrm{FPR}(w) = \Pr[w(Y) = 1 \mid f(Z) \geq f_0].
\end{align*}
The false positive rate is of lower priority for safety because issuing an unnecessary warning might only be an inconvenience, while failing to issue a warning when the situation is unsafe can lead to catastrophic outcomes. In summary, our goal is to design a warning function $w(\cdot)$ such that:
\textbf{Goal:} Provably achieve $\epsilon$-safety for small $\epsilon$ (e.g. $0.02$), while achieving low false positive rate (FPR).
\textbf{Examples} A few examples that illustrate this problem setup are as follows:
(1) In a driver alert system, users may want an assurance that among the instances in which the driver is in a dangerous situation, the system will issue a warning the vast majority of the time. The safety score in this case could be the time to collision (TTC), or the nearest distance from another car.
(2) In a multi-arm robot collaboration system, users may want an assurance that among the instances in which the robot arms may collide, the system will issue a warning the majority of the time. The safety score could be the nearest distance to another robot arm.
(3) In a warehouse robotic box-stacking system, users may want an assurance that among the instances in which the boxes will topple, the system will issue a warning the majority of the time. The safety score could be the probability of a stable stack.
Example 3 can be thought of as ROC curve threshold tuning. If the model used is a binary classifier that predicts whether there is a stable stack,
we can use the predicted probability
as $g$. Note that in this special case, our method also tunes the threshold, but adds guarantees on the false negative rate and practical guidelines for sample complexity.
\vspace{-1mm}
\subsection{Analysis of the Trade-off Between the FNR and FPR}
\label{sec:fpr}
In this section, we analyze the fundamental trade-off between the false negative rate (FNR) and the false positive rate (FPR) of a warning system. For example, a trivial system that always issues a warning will have a 0\% FNR but 100\% FPR. Conversely, a system that never issues a warning will have a 0\% FPR but 100\% FNR. This suggests a trade-off between the achievable FNR and FPR.
\paragraph{Infinite validation data regime} Even with infinite validation data, we may not be able to achieve both perfect FNR and perfect FPR because of inherent limitations of the safety score. For instance, at one extreme, if the safe and unsafe examples have identical safety score distributions, then there is no way to distinguish them according to Definition~\ref{def:safe}. At the other extreme, if the safe and unsafe examples have disjoint safety score distributions, then we can distinguish them perfectly (i.e. achieve 0\% FPR and 0\% FNR). A typical real world scenario will likely fall somewhere in between the two extremes, as illustrated in Figure~\ref{fig:optimal_fpr}. The key quantity is the amount of overlap between the safety score distribution for safe vs. unsafe examples, which will dictate the optimal achievable trade-off between the FPR and the FNR.
\begin{figure}
\centering
\includegraphics[width=0.5\linewidth]{figures/optimal_fpr.pdf}
\caption{Even in the limit of infinite validation data, the best false positive rate achievable (for a given $\epsilon$-safety level) is determined by the distribution of the safe samples and the unsafe samples under the surrogate safety score function $g$.}
\label{fig:optimal_fpr}
\vspace{-5mm}
\end{figure}
\paragraph{Finite validation data regime} The lack of sufficient validation data is another source of error that degrades the best achievable FNR/FPR trade-off. Intuitively,
because we need to \textit{provably guarantee} the FNR for $\epsilon$-safety, in the absence of sufficient validation data, we must be conservative and issue more warnings than necessary. For example, with zero validation data we have no choice but to issue a warning for nearly every example, leading to a high FPR.
In fact, in Appendix~\ref{appendix:lower_bound} we show that if we have fewer than $T$ data samples, we cannot guarantee better than $O(1/T)$ FNR without incurring an FPR of close to $1$.
Our conformal algorithm can guarantee an $O(1/T)$ FNR while the FPR is not much higher than in the infinite data regime, demonstrating the (asymptotic) optimality of the conformal algorithm presented below.
\subsection{Algorithm to Achieve Guaranteed Safety Assurances}
In this section, we will describe an algorithm that achieves $\epsilon^*$-safety with a low (nontrivial) false positive rate. The setup is as described in Section \ref{sec:setup}, with current observations $X$, simulator $Y$, and true unknown future states $Z$.
If the simulation $Y$ is perfect and has the same type as the ground truth future state $Z$, i.e. $Z = Y$ almost surely, then we can simply set $g = f$ and choose $w(Y) = \mathbb{I}(g(Y) < f_0)$, and this $w$ will automatically satisfy our definition of safety. However, in most applications, it is difficult to provide any guarantees on the accuracy of the simulation. For example, in autonomous driving situations, traffic participants can behave in unexpected and hard to predict ways.
When we are uncertain about the simulation accuracy, we will require an additional training dataset. With a dataset of (simulated future state, true future state) pairs $(Y_1, Z_1)$, $(Y_2, Z_2), \cdots, (Y_T, Z_T)$, where $T$ is the number of samples, we can guarantee $\epsilon$-safety. Let $({\hat{Y}}, {\hat{Z}})$ denote a new test sample. We require only a single assumption on the dataset:
\begin{assumption} \label{assump:exchangeable}
The sequence $(Y_1, Z_1)$, $\cdots$, $(Y_T, Z_T)$, $({\hat{Y}}, {\hat{Z}})$ is exchangeable, i.e. the probability of observing any permutation of the sequence is equally likely.
\end{assumption}
Exchangeability is a strong assumption. However, it is weaker than typical i.i.d. assumptions that underlie most machine learning methods with performance guarantees: if a sequence of data is i.i.d., then it is also exchangeable.
In addition, if the distribution shifts, it is not prohibitively costly to collect a new training dataset from the shifted distribution. This is because we require only a very small dataset (e.g. in most of our experiments, the training dataset contains only about 50 examples of unsafe situations)
Based only on Assumption 1, we design an algorithm to guarantee safety on test data.
The algorithm can be thought of as an instantiation of the conformal prediction framework.
\begin{algorithm}
\begin{algorithmic}[1]
\State \textbf{Input} A set of training data $(Y_1, Z_1), \cdots, (Y_T, Z_T)$, surrogate safety score $g$, true safety score $f$, and threshold $f_0$; a new simulation ${\hat{Y}}$.
\State Compute ${\mathcal{A}} = \lbrace g(Y_t) \mid f(Z_t) < f_0, t= 1, \cdots, T \rbrace$
\State Sample $U$ uniformly in $\lbrace 0, 1,\cdots, |\lbrace a \in {\mathcal{A}} \mid a = g({\hat{Y}}) \rbrace | \rbrace$
\State Compute $q = \frac{\ | \lbrace a \in {\mathcal{A}} \mid a < g({\hat{Y}}) \rbrace | + U + 1}{|{\mathcal{A}}|+1}$
\State If $q \leq 1 - \epsilon$ then \textbf{output} 1, otherwise \textbf{output} 0 \label{alg2:line1}
\end{algorithmic}
\caption{Approximate $\epsilon$-safety}
\label{alg:general2}
\end{algorithm}
Intuitively, the procedure is as follows. We first compute a predicted safety score (based on the simulator outputs) for each unsafe sample in the training dataset (Line 2). We then sample a number from a uniform distribution (Line 3). We next compute the quantile value for the new test simulation, i.e.\ the proportion of validation samples with a lower safety score than ${\hat{Y}}$ (Line 4), with a small randomization factor from the previous step. If this quantile value is smaller than $1 - \epsilon$ (i.e. fewer than $1 - \epsilon$ of the unsafe samples from the training set have a lower safety score), we say that this may be an unsafe situation; otherwise, we say that it is safe (Line 5). The following proposition (proved in Appendix~\ref{appendix:proofs}) shows that Algorithm~\ref{alg:general2} can guarantee safety.
\begin{prop}
\vspace{-1mm}
\label{prop:safety2}
Under Assumption 1, Algorithm~\ref{alg:general2} is $\epsilon + 1/(1+|{\mathcal{A}}|)$-safe (with respect to ${\hat{Y}}, {\hat{Z}}$).
\end{prop}
To use Proposition~\ref{prop:safety2} to provide safety guarantees, we choose $\epsilon$ based on the number of samples available $|{\mathcal{A}}|$. Specifically, if the desired safety level is $\epsilon^*$, then we can choose any $\epsilon < \epsilon^*$ in Algorithm~\ref{alg:general2} such that
\begin{align}
\epsilon + 1/(1+|{\mathcal{A}}|) \leq \epsilon^*\label{eq:sample_complexity1}
\end{align}
In other words, if our choice of $\epsilon$ satisfies Eq. (\ref{eq:sample_complexity1}), then Algorithm~\ref{alg:general2} will be $\epsilon^*$-safe.
Intuitively, choosing a large $\epsilon$ decreases the false positive rate (FPR). This is because according to Algorithm~\ref{alg:general2} Line \ref{alg2:line1}, choosing a larger $\epsilon$ decreases the number of times that a warning is output. Therefore, based on the number of samples in the training dataset $|{\mathcal{A}}|$, we choose the largest $\epsilon$ that satisfies Eq. (\ref{eq:sample_complexity1}) (i.e. we choose $\epsilon = \epsilon^* - 1/(1+|{\mathcal{A}}|)$). We will call $1/(1+|{\mathcal{A}}|)$ the discretization error.
Proposition~\ref{prop:safety2} also reveals the sample complexity of the conformal prediction algorithm. If the number of unsafe examples is too small ($|{\mathcal{A}}| \leq 1/\epsilon^* - 1$), then we must choose $\epsilon < 0$ to ensure $\epsilon^*$-safety according to Proposition~\ref{prop:safety2}. Algorithm~\ref{alg:general2} with $\epsilon < 0$ will trivially always output $1$ (i.e. always issue a warning). On the other hand, if the number of unsafe examples exceeds the threshold ($|{\mathcal{A}}| > 1/\epsilon^* - 1$), then there will be an $\epsilon > 0$ that ensures $\epsilon^*$-safety according to Proposition~\ref{prop:safety2}. Consequently, Algorithm~\ref{alg:general2} will not be trivial. In practice, we find that to get good results and a low false positive rate, it is sufficient to have sample count $|{\mathcal{A}}|$ that exceeds the threshold by a small margin, such as $|{\mathcal{A}}| = 1.5/\epsilon^*-1$. For example, to achieve a 5\% false positive rate, it is sufficient to have only about 30 unsafe examples.
\subsection{Comparing Conformal Prediction with PAC Learning}
We further compare the statistical learning and conformal prediction guarantees. We first clarify the notation and formally define the different assumptions. Consider a sequence of training data $(Y_1, Z_1), \cdots, (Y_T, Z_T)$ and a sequence of test data $({\hat{Y}}_1, {\hat{Z}}_1), \cdots, ({\hat{Y}}_N, {\hat{Z}}_N)$.
Let $c_1, \cdots, c_M$ denote the unsafe subsequence of test data, i.e.
$
({\hat{Y}}_{c_1}, {\hat{Z}}_{c_1}), \cdots, ({\hat{Y}}_{c_{M}}, {\hat{Z}}_{c_M})
$
is the subsequence of $({\hat{Y}}_1, {\hat{Z}}_1), \cdots, ({\hat{Y}}_N, {\hat{Z}}_N)$ such that, for all $m$, $f({\hat{Z}}_{c_m}) < f_0$.
Two possible assumptions that we could make on the training and test data sequence are shown in Assumptions~\ref{assump:exchangeable2} and~\ref{assump:iid}.
In particular, marginal exchangeability (Assumption~\ref{assump:exchangeable2}) is the same as Assumption~\ref{assump:exchangeable} from the previous section. The only difference here is that we explicitly state that we only require exchangeability with \textit{each} test data point.
\begin{assumption}[marginal exchangeability]\label{assump:exchangeable2}
For each $n$, the sequence $(Y_1, Z_1)$, $\cdots$, $(Y_T, Z_T)$, $({\hat{Y}}_n, {\hat{Z}}_n)$ is exchangeable.
\end{assumption}
\begin{assumption}[i.i.d.]\label{assump:iid}
The training / test data sequence $(Y_1, Z_1)$, $\cdots$, $(Y_T, Z_T)$, $({\hat{Y}}_1, {\hat{Z}}_1), \cdots, ({\hat{Y}}_N, {\hat{Z}}_N)$ is drawn from an i.i.d. distribution.
\end{assumption}
Given a warning function, we use the random variables ${\hat{F}}_{c_1}, \cdots, {\hat{F}}_{c_M}$ to denote failure of the warning function, i.e. ${\hat{F}}_{c_m} = {\mathbb{I}}(w({\hat{Y}}_{c_m}) = 0)$.
Note that ${\hat{F}}_{c_m}$ depends on $w$, but we drop this dependence from our notation.
A learning algorithm is a function that takes as input the training data $(Y_1, Z_1), \cdots, (Y_T, Z_T)$ and outputs a warning function $w: X \to \lbrace 0, 1\rbrace$. There are two main paradigms for designing learning algorithms with guarantees.
\paragraph{PAC Learning} Under Assumption~\ref{assump:iid}, a learning algorithm is $(\epsilon,\delta)$-safe if with $1-\delta$ probability (with respect to randomness of the training data) the learned warning function $w$ satisfies for some $\epsilon' < \epsilon$
\begin{align*}
{\hat{F}}_{c_1}, \cdots, {\hat{F}}_{c_M} \sim \mathrm{Bernoulli}(\epsilon') \numberthis\label{eq:pac_guarantee}
\end{align*}
\paragraph{Conformal Learning} For completeness we restate the conformal learning guarantee. A learning algorithm is $\epsilon$-safe if the learned function $w$ satisfies for some $\epsilon' < \epsilon$
\begin{align*}
{\hat{F}}_{c_m} \sim \mathrm{Bernoulli}(\epsilon'), \text{ for all } m = 1, \cdots, M \numberthis\label{eq:conformal_guarantee}
\end{align*}
\textbf{Comparing Assumptions} Conformal learning requires weaker assumptions.
Assumption~\ref{assump:exchangeable2} is much weaker than Assumption~\ref{assump:iid}, and hence is applicable to a much larger class of problems. For example, consider an autonomous driving application where the training data are snippets from randomly sampled driving scenes (no two training data points come from the same driving scene), and the test data $({\hat{Y}}_1, {\hat{Z}}_1), \cdots, ({\hat{Y}}_N, {\hat{Z}}_N)$
is a sequence of driving snippets from a random driving scene. The test data points are not independent because they are from the same scene, and hence Assumption~\ref{assump:iid} is violated. However, Assumption~\ref{assump:exchangeable2} holds because the training data and any \textit{single} test sample are snippets from randomly sampled driving scenes.
\textbf{Comparing Sample Complexity} Conformal learning requires $\Theta(1/\epsilon)$ training examples of unsafe situations (Proposition~\ref{prop:safety2}), while standard analysis in PAC learning requires $\Theta(1/\epsilon^2)$ examples. For example, consider the following $(\epsilon,\delta)$-safe algorithm: based on the simulation $Y$ and the surrogate safety function $g$, we consider the family of warning functions $w_\theta(Y) = {\mathbb{I}}(g(Y) < \theta)$. Our goal is to estimate the false negative rate of $w_\theta$ (denoted by $\epsilon^*(\theta)$ for each $\theta$ and select the smallest $\theta$ such that $\epsilon^*(\theta) \leq \epsilon$.
To estimate $\epsilon^*$, we compute the (empirical) false negative rate (denoted by $\hat{\epsilon}(\theta)$ on the training data, i.e.
\begin{align}
\hat{\epsilon}(\theta) = 1/M \sum_m {\mathbb{I}}( g(Y_{c_m}) \geq \theta)
\end{align}
and use a standard concentration inequality (such as Hoeffding) to bound the difference between $\epsilon^*(\theta)$ and $\hat{\epsilon}(\theta)$. Specifically, with probability $1-\delta$
\begin{align}
\epsilon^*(\theta) \in \hat{\epsilon}(\theta) \pm \sqrt{\frac{ \log(1/\delta)}{2M}} \label{eq:hoeffding}
\end{align}
Note that Eq. (\ref{eq:hoeffding}) is already the tightest bound possible up to constants~\cite{Foody2009Sample}. To verify that $w_\theta$ has $\leq \epsilon$ false negative rate, we have to check that $\hat{\epsilon}(\theta) + \sqrt{\frac{ \log(1/\delta)}{2M}} \leq \epsilon$, which requires
\begin{align*}
\sqrt{\frac{ \log(1/\delta)}{2M}} \leq \epsilon \iff M \geq \frac{\log(1/\delta)}{2\epsilon^2}
\end{align*}
i.e. we should have at least $\Theta(1/\epsilon^2)$ samples.
In words, even a fixed $w_\theta$ requires $\Theta(1/\epsilon^2)$ samples to verify its false negative rate according to Eq. (\ref{eq:hoeffding}). Thus, finding $w_\theta$ to provably achieve low false negative rate should require at least as many, if not more, training examples.
\textbf{Comparing Usefulness of Guarantees}
PAC learning and conformal learning both have advantages. PAC learning has the advantage that its i.i.d. error rate guarantee in Eq. (\ref{eq:pac_guarantee}) is stronger than the marginal error rate guarantee in Eq. (\ref{eq:conformal_guarantee}). For example, if the downstream user is very sensitive to high variance (i.e. it is unacceptable for all test examples to fail simultaneously even if the probability is vanishingly small) then the i.i.d. error rate guarantee in Eq. (\ref{eq:pac_guarantee}) might be necessary. Nevertheless, the risk can be reduced by alternative methods such as financial tools (insurance).
On the other hand, the conformal learning guarantee in Eq. (\ref{eq:conformal_guarantee}) has the advantage that it always holds, while the PAC learning guarantee in Eq. (\ref{eq:pac_guarantee}) only holds with $1-\delta$ probability.
To summarize, conformal learning requires much weaker assumptions and fewer samples, and its guarantees always hold (rather than with $1-\delta$ probability). PAC learning offers stronger guarantees when its assumptions and sample complexity requirements are met.
\section{Experiments: Driver Alert System}
\label{sec:experiments-driving}
We empirically validate the guarantees of our framework on a driver alert safety system using real driving data. The system should warn the driver if the driver may get into an unsafe situation, without issuing too many false alarms. We show that the false negative rate (the percentage of unsafe situations that the system fails to identify) is indeed bounded according to Proposition~\ref{prop:safety2}, while the FPR remains low.
\subsection{Experimental Setup}
\textbf{Methods} We evaluate our framework on the setup described in Section~\ref{sec:setup}.
We use Trajectron++~\cite{Salzmann2020Trajectron} as our future dynamics model (i.e. in the notation of Section~\ref{sec:setup}, \ $Y$ is the output of Trajectron++ and $g = f$). We choose the safety score $f$ as a weighted distance metric, where agents in the direction of the ego-vehicle velocity vector are considered ``closer'' than agents in the orthogonal direction. Refer to the Appendix for a more detailed explanation.
\textbf{Datasets} We use the nuScenes~\cite{nuscenes2019} and the Kaggle Lyft Motion Prediction~\cite{kaggleLyft} autonomous driving datasets.
Each dataset contains multiple scenes, and each scene contains multiple trajectories. The trajectories in a scene are correlated with each other, but the different scenes are sufficiently distinct from each other to be considered exchangeable. To generate a dataset of exchangeable trajectories, we
sample a single trajectory uniformly at random from each scene.
\textbf{Data Splitting} To compute average performance, we use random train and test splits.
For both datasets, we first pool together all available data points, randomly shuffle them, and separate them back into training and test splits (with the same size as the original splits).
We ran 100 trials for each experiment, and averaged over the results.
\subsection{Results and Discussion}
In Fig.~\ref{fig:vary-threshold}, \ref{fig:vary-epsilon}, and \ref{fig:vary-label-freq} we vary several parameters (safety threshold $f_0$, safety guarantee $\epsilon$, and proportion of unsafe situations) for nuScenes. We show qualitatively similar results for the Lyft dataset in Figure~\ref{fig:lyft}. Our main observations:
1. The false negative rate (i.e. safety) is always within the theoretical bound in Proposition~\ref{prop:safety2}.
We achieve these false negative rates with very little data.
nuScenes has 50-70 unsafe examples in the training dataset, and Lyft has about 50. Yet, even with these few examples, we can ensure a false negative rate to within 1 or 2\% of the desired $\epsilon$.
2. The false positive rate (FPR) is generally very good --- well below 1\% on the nuScenes dataset.
We use an off-the-shelf trajectory predictor trained on a small academic dataset; a more accurate trajectory predictor trained on industry-sized datasets might be expected to provide a more discriminative safety score (as in Figure 2), and thus a further improved FPR.
Note that as shown in Figure~\ref{fig:fpr_vs_samples} in the Appendix, there is a tradeoff between $\epsilon$ and the FPR when there are few (e.g. $< 1/T$) samples, which is consistent with what our theory from Section~\ref{sec:fpr} would predict.
\begin{figure*}[tb]
\centering
\begin{subfigure}[t]{0.475\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/vary_threshold.pdf}
\caption{False negative and false positive rates on the nuScenes dataset with varying $f_0$ and $\epsilon = 0.05$. The theoretical upper bound on epsilon is shown in {\color{OliveGreen} green}. The false negative rate ({\color{red} red}) is below the upper bound and the false positive rate ({\color{blue} blue}) is very low.}
\label{fig:vary-threshold}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.475\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/vary_epsilon.pdf}
\caption{False negative and false positive rates on the nuScenes dataset with varying $\epsilon$ and $f_0 = 25$. The theoretical upper bound on epsilon is shown in {\color{OliveGreen} green}. The false positive rate ({\color{blue} blue}) improves with higher $\epsilon$. }
\label{fig:vary-epsilon}
\end{subfigure}
\vskip\baselineskip
\begin{subfigure}[t]{0.475\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/vary_frequency.pdf}
\caption{False negative and false positive rates on the nuScenes dataset with varying proportions of unsafe samples in the training set. The theoretical upper bound on epsilon is shown in {\color{OliveGreen} green}. Here, $\epsilon = 0.05$ and $f_0 = 25$.}
\label{fig:vary-label-freq}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.475\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/lyft_epsilon.pdf}
\caption{False negative and false positive rates on the Kaggle Lyft dataset with a varying $\epsilon$ value. Here, $f_0 = 3.5$, and there are approx. 50 unsafe examples in the training dataset.}
\label{fig:lyft}
\end{subfigure}
\caption{False negative rate, false positive rate, and theoretical upper bound on epsilon for the nuScenes and Lyft datasets while varying several parameters.}
\vspace{-2mm}
\end{figure*}
3. One previously unmentioned benefit of our approach is that our method is robust to label frequency shift --- the frequency of unsafe situations can differ between the training data and test data. Observe that the output of Algorithm~\ref{alg:general2} depends only on the unsafe examples; consequently, the safety guarantee in Proposition~\ref{prop:safety2} still holds if we increase or decrease the number of safe examples. For example, the training data collection process could intentionally focus on unsafe situations, so that unsafe examples are over-represented in the training data. We empirically simulate this in Figure~\ref{fig:vary-label-freq} where we increase the proportion of unsafe examples in the training set (by deleting safe examples). The performance of our algorithm does not change qualitatively.
In the comparison of PAC learning and conformal learning, we argued that the main advantage of PAC learning is that the failures are i.i.d., so the total number of failures should have low variance (due to the Central Limit Theorem). However, we show empirically that users need not be overly concerned about highly correlated failures, as long as the test samples are not inherently highly correlated.
We find that the variance on the false negative rate from different train/test splits is very low. With $\epsilon = 0.06$, for instance, it was only $0.0014$, and this variance is representative among the various $\epsilon$ experiments. We show more evidence of this in the Appendix.
\section{Experiments: Robotic Grasping}
\label{sec:experiments-grasp}
Finally, we validate the guarantees of our framework on a robotic grasping system that should warn the user when the robot will fail to pick and transport an object. Picking is a core problem in warehouse robotics~\cite{correll2016analysis,eppner2016lessons,hernandez2016team,yu2016summary,zeng2017multi,mahler2017dex,mahler2017suction,mahler2019learning}, and failures hurt throughput (potentially even stopping the assembly line). Failures can also lead to dropped or damaged goods.
We again evaluate our framework on the setup described in Section~\ref{sec:setup}, using the Grasp Quality Convolutional Neural Network (GQ-CNN) from~\cite{mahler2017dex,mahler2017suction,mahler2019learning} as our predictor model on the DexNet 4.0 dataset of synthetic objects grasped with a parallel-jaw gripper~\cite{mahler2019learning}. This model classifies whether a candidate robotic grasp will be successful, and we use the probability of a successful pick as the safety score.
We consider a candidate grasp ``unsafe'' if it will not be able to successfully pick the object
(i.e.\ $Z = 0$). Note that this is exactly an ROC curve tuning setup.
We averaged over 100 trials of Algorithm~\ref{alg:general2} with randomized train/test splits.
With $\epsilon = 0.05$, we achieved a false negative rate of 0.05, and a false positive rate of 0.11. With $\epsilon = 0.1$, we achieved a false negative rate of 0.10 and a false positive rate of 0.04. The conformal guarantees of our framework hold.
\section{Conclusion}
In this work, we introduce a broadly applicable framework that uses conformal prediction to tune warning systems for robotics applications. This framework allows us to achieve provable safety assurances with very little data. We demonstrate empirically that the guarantees on the false negative rate hold for a driver alert system and for a robotic grasping system. In future work, we would like to explore conformal prediction in non-exchangeable scenarios~\cite{tibshirani2019conformal}, conditional safety~\cite{feldman2021improving}, and deployment in industry-scale applications. Another important future direction is to study the impact of the predictor on the data that it is trying to predict~\cite{perdomo2020performative}.
{\small
\bibliographystyle{splncs04}
|
\section{Introduction}
Fine-grained species classification is essential in monitoring biodiversity. Diversity of life is the central tenet to biology and preserving biodiversity is key to a more sustainable life. Monitoring biodiversity requires identifying living organisms at the lowest taxonomic level possible. The traditional approach to identification uses published morphological dichotomous keys to identify the collected sample. This identification involves a tedious process of manually assessing the presence or absence of a long list of morphological traits arranged at hierarchical levels. The analysis is often performed in a laboratory setting by a well-trained human taxonomist and is difficult to do at scale. Fortunately, advances in technology have addressed this challenge to some extent through the use of DNA barcodes. DNA barcoding is a technique that uses a short section of DNA from a specific gene, such as \textit{cytochrome C oxidase I (COI)}, found in mitochondrial DNA, and offers specific information about speciation in living organisms and can achieve nearly perfect classification accuracy at the species level~\cite{lunt1996,hebert2004}.
As it is costly to obtain the label information for fine-grained classification of species, Zero-Shot Learning (ZSL) that handles missing label information is a suitable task. In ZSL, side information is used to associate seen and unseen classes. Heretofore, popular choices for side-information were manually annotated attributes~\cite{lampert_cvpr,farhadi}, word embeddings~\cite{cmt,devise,w2v} derived from free-form text or the WordNet hierarchy \cite{wordnet,sje}. It is often assumed that an exhaustive list of visual attributes characterizing all object classes (both \textit{seen} and \textit{unseen}) can be determined based only on seen classes.
However, taking insects as our object classes, if no seen class species have antennae, the attribute list may not contain \textit{antenna}, which may in fact be necessary to distinguish unseen species that are very similar to seen classes but have antennae as a characteristic trait.
In the United States alone, more than 40\% of all insect species (>70,000) remain undescribed \cite{stork2018many}, which is a clear sign of the limitations of existing identification techniques that rely on visual attributes.
Similarly, free-form text is unlikely to contain sufficiently descriptive information about fine-grained objects to generate discriminative vector embeddings. For example, \textit{tiger beetle} is a class in the ImageNet dataset. The Wikipedia page on the tiger beetle contains a detailed description of this large group of beetles in the Cicindelinae subfamily.
However,
the \textit{tiger beetle} group itself contains thousands of known species and the Wikipedia pages for these species either do not exist or are limited to short text that does not necessarily contain any information about species' morphological characteristics. WordNet hierarchy may not be useful either as most of the species names do not exist in WordNet.
Given that DNA information can be readily available for training~\cite{BOLD2007,ratnasingham2013dna}, species-level DNA information can be used as highly specific side information to replace high-level semantic information in ZSL. For seen classes, species-level DNA information can be obtained by finding the consensus nucleotide sequence among samples of a given species or by averaging corresponding sequence embeddings of samples. For unseen classes, species-level DNA information can be obtained from actual samples, if available, in the same way as seen classes, or can be simulated in a non-trivial way to represent potentially existing species.
Our approach uses DNA as side information for the first time for zero-shot classification of species. In fine-grained, large-scale species classification, no other side information can explain class dichotomy better than DNA, as new species are explicitly defined based on variations in DNA.
The hierarchical Bayesian model leverages the implicit inter-species association of DNA and phenotypic traits and ultimately allows us to establish a Bayesian hierarchy based on DNA similarity between unseen and seen classes. We compare DNA against word representations for assessing class similarity and show that the Bayesian model that uses DNA to identify similar classes achieves favorable results compared to the version that uses word representations on a well-known ZSL benchmark species dataset involving slightly less than 200 bird species. In the particular case of an insect dataset with over 1000 species, when visual attributes or word representations may not offer feasible alternatives, we show that our hierarchical model that relies on DNA to establish class hierarchy significantly outperforms all other embedding-based methods and feature generating networks.
Our contributions are on three fronts. First, we introduce DNA as side information for fine-grained ZSL tasks, implement a Convolutional Neural Net (CNN) model to learn DNA barcode embeddings, and show that the embeddings are robust and highly specific for closed-set classification of species, even when training and test sets of species are mutually exclusive. We use the benchmark CUB dataset as a case study to show that DNA embeddings are competitive to word embeddings as side information. Second, we propose a fine-grained insect dataset involving $21,212$ matching image/DNA pairs from $578$ genera and $1,213$ species as a new benchmark dataset and discuss the limitations of current ZSL paradigms for fine-grained ZSL tasks when there is no strong association between side information and image features. Third, we perform extensive studies to show that a simple hierarchical Bayesian model that uses DNA as side information outperforms state of the art ZSL techniques on the newly introduced insect dataset by a wide margin.
\section{Related Work}
\paragraph{Zero-Shot Learning.} Early ZSL literature is dominated by methods that embed image features into a semantic space and perform various forms of nearest neighbor search to do inference \cite{devise,cmt,ale}. As the dimensionality of semantic space is usually much smaller than the feature space this leads to the hubness problem. In an effort to alleviate the hubness problem, \cite{zhang_dem,shitego_rigdereg_hubness} change the direction of the embedding from semantic space to image feature space. This was followed by a line of work that investigates bidirectional embedding between semantic and image spaces through a latent space ~\cite{sse,Rel_net,sje,conse,sync}.
In \cite{long2017,guo2017}, a new strategy of synthesizing features for unseen classes and converting the challenging ZSL problem into traditional supervised learning is introduced ~\cite{Lisgan,lsrgan,gan_Chen18,Cycle_gan,gan_zsl,gan_Zhu19,vae_Mishra18,cada_vae, vae_Arora18L}. Although feature generating networks (FGNs) currently achieve state-of-the-art results in ZSL, they suffer from the same problem as earlier lines of work in ZSL: hypersensitivity towards side information not strongly correlated with visual attributes. The vulnerability of both embedding and FGN-based methods toward sources of side information different than visual attributes, such as word vectors or WordNet \ hierarchy, is investigated in \cite{sje,cada_vae,lsrgan}. Another limitation of FGNs is that features generated for unseen classes are significantly less dispersed than actual features due to the generator failing to span more than a small subset of modes available in the data. Recent deep generative models mitigate this problem by proposing different loss functions that can better explore inter-sample and inter-class relationships \cite{wgan,bao2017cvae,cacheux2019modeling,jiang2019transferable,vyas2020leveraging}. However, these methods fail to scale well with an increasing number of classes with an especially high inter-class similarity \cite{liu2021cross}.
\paragraph{Side Information in ZSL.} Side information serves as the backbone of ZSL as it bridges the knowledge gap between seen and unseen classes. Earlier lines of work \cite{dap_iap,ale} use visual attributes to characterize object classes. Although visual attributes achieve compelling results, obtaining them involves a laborious process that requires manual annotation by human experts not scalable to data sets with a large number of fine-grained object classes. When dealing with fine-grained species classification, apart from scalability, a more pressing obstacle is how to define subtle attributes potentially characteristic of species that have never been observed.
As an alternative to manual annotation, several studies \cite{elhoseiny2013,devise,sje,latem,qiao2016,ba2015} proposed to learn side information that requires less effort and minimal expert labor such as textual descriptions, distributed text representations, like Word2Vec~\cite{w2v} and GloVe~\cite{glove}, learned from large unsupervised text corpora, taxonomical order built from a pre-defined ontology like WordNet \cite{wordnet}, or even human gaze reaction to images~\cite{humangaze_zsl}. The accessibility, however, comes at the cost of performance degradation~\cite{sje,cada_vae}. A majority of ZSL methods implicitly assume strong correlation between side information and image features, which is true for handcrafted attributes but less likely to be true for text representations or taxonomic orders. Consequently, all these methods experience significant decline in performance when side information is not based on visual attributes.
\begin{figure}
\centering
\includegraphics[scale=0.72]{figures/NIPS_image_samples_final.png}
\caption{Image samples from the INSECT dataset. Rows represents a small subset of species from three orders: Hymenoptera, Coleoptera and Diptera, respectively. The first word in names indicate genus, the two words together define the species name.}
\label{fig:image_samples}
\end{figure}
\section{Barcode of Life Data and DNA Embeddings} \label{BOLD}
In this study, we present a newly compiled fine-grained INSECT dataset containing $21,212$ matching image/DNA pairs from $1,213$ species for the evaluation of generalized ZSL techniques (see Fig. \ref{fig:image_samples} for sample images). Unlike existing benchmark ZSL datasets, this new dataset uses DNA as side information and can be best characterized with the high degree of similarity among classes. Among the existing benchmark datasets, SUN contains the largest number of classes ($717$) but classes in SUN represent a wide range of scene categories related to transportation, indoor and outdoors, nature, underwater etc., and as such can be considered a relatively coarse-grained dataset compared to the INSECT dataset we are introducing in this study.
All insect images and associated DNA barcodes in our dataset come from the \underline{B}arcode \underline{o}f \underline{L}ife \underline{D}ata System (BOLD) \cite{BOLD2007, ratnasingham2013dna}. BOLD is an open-access database in which users can upload DNA sequences and other identifying information for any living organism on Earth. The database provides approximately $658$ base pairs of the mitochondrial DNA barcode extracted from the \textit{cytochrome c oxidase I} (COI) gene along with additional information such as country of origin, life-stage, order, family, subfamily, and genus/species names.
\paragraph{Data Collection.} We collected image/DNA pairs of insects that originate from three orders: Diptera (true flies), Coleoptera (beetles) and Hymenoptera (sawflies, wasps, bees, and ants). While the dataset is in general clean, manual effort was devoted to further curate the dataset. Only cases with images and matching DNA barcodes of adult insects are included. Images from each species were visually inspected and poor quality images were deleted. Only species with larger than ten instances were included. The final dataset consisted of $21,212$ images and $1,213$ insects species of which $254$ belong to Diptera ($133$ genera), $564$ to Coleoptera ($315$ genera) and $395$ to Hymenoptera ($130$ genera). We extracted image features, namely image embeddings, using a pre-trained (on ImageNet 1000 classes) ResNet101 model~\cite{resnet}. Images are resized to $256\times256$ and center-cropped before fed to the ResNet model. No other pre-processing is applied to the images.
\begin{wraptable}{r}{0.4\textwidth}
\vspace{-10pt}
\begin{tabular}{c| c c c}
& $Y^{all}$ & $Y^s$ & $Y^{u}$ \\ \hline
\#Images & 21,212 & 3,525 & 2,425 \\
\#Classes & 1,213 & 1,080 & 121\\ \hline
\end{tabular}
\caption{ZSL split details. $Y^{s}$, and $Y^{u}$ denote the seen and unseen test sets, whereas $Y^{all}$ represents entire data. There are $15,262$ $(21,212-3,525-2,425)$ samples left for the training set.}
\label{tab:dataset_summary}
\end{wraptable}
\paragraph{Data Split.} We randomly chose 10\% of all species as unseen classes for the test set, which left us with $1,092$ seen and $121$ unseen classes. In the same fashion, we randomly chose 10\% of the $1,092$ training classes as unseen classes for the validation set. Samples from seen classes were split by a $80/20$ ratio in a stratified fashion to create seen portion of the train and test datasets. In the dataset there were a few hundred cases where multiple image views (dorsal, ventral, and lateral) of the same insect were present. To avoid splitting these cases between train and test, we made sure all instances of the same insect are included in the training set. As a result, $12$ of the $1,092$ seen classes in the training set were not represented in the test set. The total number of images and classes available in Train/Test splits are summarized in Table~\ref{tab:dataset_summary}.
\begin{figure}
\centering
\includegraphics[scale=0.4]{figures/NIPS_DNA_diagram.pdf}
\caption{Attribute extraction from mitochondrial DNA.
}
\label{fig:dna_attributes}
\end{figure}
\paragraph{DNA Embeddings.} We trained a Convolutional Neural Network (CNN) to learn a vector representation of DNA barcodes in the Euclidean space. First, the consensus sequence of all DNA barcodes in the training set with 658bp is obtained. Then, all sequences are aligned with respect to this consensus sequence using a progressive alignment technique implemented in MATLAB R2020A (Natick, MA, USA). A total of five tokens are used, one for each of the four bases, \textit{Adenine}, \textit{Guanine}, \textit{Cytosine}, \textit{Thymine}, and one for \textit{others}. All ambiguous and missing symbols are included in the \textit{others} token. In pre-processing, barcodes are one hot encoded into a 658x5 2D array, where 658 is the length of the barcode sequence (median of the nucleotide length of the DNA data).
To train the CNN model, a balanced subset of the training data is subsampled, where each class size is capped at 50 samples. The CNN is trained with $14,723$ barcodes from $1,092$ classes. No barcodes from the $121$ unseen classes are employed during model training. The training set is further split into two as train ($80\%$) and validation ($20\%$) by random sampling. We used 3 blocks of convolutional layers each followed by batch normalization and 2D max-pooling. The output of the third convolutional layer is flattened and batch normalized before feeding the data into a fully-connected layer with $500$ units. The CNN architecture is completed by a softmax layer. We used the output of the fully-connected layer as the embeddings for DNA. Class level attributes are computed by the mean embedding of each class. The DNA-based attribute extraction is illustrated in Figure~\ref{fig:dna_attributes}. The details of the model architecture is depicted in Figure 3 in Supplementary material. We used ADAM optimizer for training the model for five epochs with a batch size of $32$ (with a step-decay initial learning rate = $0.0005$ and drop factor$=0.5$, $\beta_1=0.9$, $\beta_2=0.999$). The model is developed in Python with Tensorflow-Keras API.
\paragraph{Predictive accuracy of DNA embeddings.} Although the insect barcodes we used are extracted from a single gene (COI) of the mitochondrial DNA with a relatively short sequence length of 658 base pairs, they are proven to have exceptional predictive accuracy; the CNN model achieves a 99.1\% accuracy on the held-out validation set.
Note that, we only used the data from training seen classes to train the CNN model. In order to validate the generalizability of embeddings to unseen data, we trained a simple K-Nearest Neighbor classifier ($K=1$) on the randomly sampled $80\%$ of the DNA-embeddings of unseen classes and tested on the remaining $20\%$. The classifier had a perfect accuracy for all $121$ but one classes with an overall accuracy of 99.8\%.
To demonstrate that the approach can be easily extended to larger members of the animal kingdom, we compiled approximately $26,000$ DNA barcodes from $1,047$ bird species to train another CNN model (ceteris paribus) to learn the DNA embeddings for CUB dataset (see the Supp. materials for details). The CNN model achieved a compelling 95.60\% on the held-out validation set.
\section{Bayesian Zero-shot Learning}
Object classes in nature already tend to emerge at varying levels of abstraction, but the class hierarchy is more evident when classes represent species and species are considered the lowest taxonomic rank of living organisms. We build our approach on a two layer hierarchical Bayesian model that was previously introduced and evaluated on benchmark ZSL datasets with promising results \cite{bzsl}. The model assumes that there are latent classes that define the class hierarchy in the image space and uses side information to build the Bayesian hierarchy around these latent classes. Two types of Bayesian priors are utilized in the model: global and local. As the name suggests, global priors are shared across all classes, whereas local priors represent latent classes, and are only shared among similar classes. Class similarity is evaluated based on side information in the Euclidean space. Unlike standard Bayesian models where the posterior predictive distribution (PPD) forms a compromise between prior and likelihood, this approach utilizes posterior predictive distributions to blend local and global priors with data likelihood for each class. Inference for a test image is performed by evaluating posterior predictive distributions and assigning the sample to the class that maximizes the posterior predictive likelihood.
\paragraph{Generative Model.} The two-layer generative model is given below.
\begin{gather}
\boldsymbol{x_{jik}} \sim N(\boldsymbol{\mu_{ji}},\Sigma_{j}), \quad
\boldsymbol{\mu_{ji}} \sim N(\boldsymbol{\boldsymbol{\mu_{j}}},\Sigma_{j}\kappa_{1}^{-1}), \quad
\boldsymbol{\mu_{j}} \sim N(\boldsymbol{\mu_{0}},\Sigma_{j}\kappa_{0}^{-1}), \quad
\Sigma_{j} \sim W^{-1}(\Sigma_{0},m)
\label{eq:i2gmm}
\end{gather}
where $j, i, k$ represent indices for local priors, classes, and image instances, respectively. We assume that image feature vectors $\boldsymbol{x_{jik}}$ come from a Gaussian distribution with mean $\boldsymbol{\mu_{ji}}$ and covariance matrix $\Sigma_{j}$, and are generated independently conditioned not only on the global prior but also on their corresponding local priors.
Each local prior is characterized by the parameters $\boldsymbol{\mu_{j}}$ and $\Sigma_{j}$.
$\boldsymbol{\mu_{0}}$ is the mean of the Gaussian prior defined over the mean vectors of local priors, $\kappa_{0}$ is a scaling constant that adjusts the dispersion of the means of local priors around $\boldsymbol{\mu_{0}}$. A smaller value for $\kappa_{0}$ suggests that means of the local priors are expected to be farther apart from each other whereas a larger value suggests they are expected to be closer. On the other hand, $\Sigma_{0}$ and $m$ dictate the expected shape of the class distributions, as under the inverse Wishart distribution assumption the expected covariance is $E(\Sigma|\Sigma_{0},m)=\frac{\Sigma_{0}}{m-D-1}$, where $D$ is the dimensionality of the image feature space. The minimum feasible value of $m$ is equal to $D+2$, and the larger the $m$ is the less individual covariance matrices will deviate from the expected shape.
The hyperparameter $\kappa_{1}$ is a scaling constant that adjusts the dispersion of the class means around the centers of their corresponding local priors. A larger $\kappa_{1}$ leads to smaller variations in class means relative to the mean of their corresponding local prior, suggesting a fine-grained relationship among classes sharing the same local prior. Conversely, a smaller $\kappa_{1}$ dictates coarse-grained relationships among classes sharing the same local prior. To preserve conjugacy of the model, the proposed model constrains classes sharing the same local prior to share the same covariance matrix $\Sigma_{j}$. Test examples are classified by evaluating posterior predictive distributions (PPD) of seen and unseen classes. As illustrated in Fig. \ref{fig:surrogate_classes} the PPD in general incorporates three sources of information: the data likelihood that arises from the current class, the local prior that results from other classes sharing the same local prior as the current class, and global prior defined in terms of hyperparameters. PPDs for seen classes include the global prior and data likelihood and are derived in the form a Student-t distribution whereas for unseen classes the data likelihood does not exist as no image samples are available for these classes. We leave the details of derivations to the supplementary material and here explain the formation of surrogate classes in terms of only local and global priors.
\iffalse
\begin{algorithm}[t]
\textbf{Input:} Training data, $\phi(seen)$, $\phi(unseen)$\\
\textbf{Output:} PPD parameters for each seen class ($\boldsymbol{\bar{\mu}_{jc}}, \bar{v}_{jc}, \bar{\Sigma}_{jc} $) and unseen class ($\boldsymbol{\bar{\mu}_{j}}, \bar{v}_{j}, \bar{\Sigma}_{j} $)
\begin{algorithmic}[1]
\State Set hyper-parameters: $\kappa_0, \kappa_1, m, s, K$
\State Compute $\boldsymbol{\mu_0}$ (mean of class means) and $\Sigma_0$ (mean of class covariances scaled by s)
\For{each seen class $\omega_{jc}$}\Comment{Images available}
\State Calculate current class params: $\boldsymbol{\bar{x}_{jc}}, n_{jc}, S_{jc}$
\State Calculate $S_{\boldsymbol{\mu}}$ (Eq x1 from Supp. mat.)
\State Calculate PPD by combining \textit{global prior}
and \textit{data driven likelihood}: $\boldsymbol{\bar{\mu}_{jc}}, \bar{v}_{jc}, \bar{\Sigma}_{jc}$ (Eq \ref{eq:i2gmm_seen})
\EndFor
\For{each unseen class $\omega_{j}$}\Comment{No image available}
\State Find K most similar seen classes:
\State $\mathcal{L}^2(\phi(\omega_{j}),\phi(seen))$
\For{ each selected seen class $\omega_{ji}$}
\State Calculate class params: $\boldsymbol{\bar{x}_{ji}}, n_{ji}, S_{ji}$
\EndFor
\State Calculate intermediate terms: $\tilde{\kappa}_j, S_{\boldsymbol{\mu}}$ (Eq x1, x2 from Supp. Mat.)
\State Calculate PPD parameters using \textit{local}
and \textit{global priors}: $\boldsymbol{\bar{\mu}_{j}}, \bar{v}_{j}, \bar{\Sigma}_{j}$ (Eq \ref{eq:i2gmm_unseen})
\EndFor
\end{algorithmic}
\caption{Modeling seen and unseen classes in BZSL \zeynep{if space becomes an issue, this algorithm can be removed}}
\label{algo:main}
\end{algorithm}
The final PPD for seen classes is derived in the form of a Student-t distribution as follows,
\begin{align}
&P(\boldsymbol{x}|\{\boldsymbol{\bar{x}_{ji}}, S_{ji}\}_{t_i=j}, \boldsymbol{\bar{x}_{jc}}, S_{jc}, \boldsymbol{\mu_{0}}, \kappa_0, \kappa_1) = T(\boldsymbol{x}|\boldsymbol{\bar{\mu}_{jc}},\bar{\Sigma}_{jc},\bar{v}_{jc}) \nonumber \\
&\boldsymbol{\bar{\mu}_{jc}} = \frac{n_{jc}\boldsymbol{\bar{x}_{jc}}+\frac{\kappa_0\kappa_1}{\kappa_0 + \kappa_1}\boldsymbol{\mu_0}}{n_{jc}+\frac{\kappa_0\kappa_1}{\kappa_0 + \kappa_1}}, \:
\bar{v}_{jc} = n_{jc}+m-D+1,\:
\bar{\Sigma}_{jc} = \frac{(\Sigma_{0}+S_{jc}+S_{\mu})(n_{jc}+\frac{\kappa_0\kappa_1}{\kappa_0+\kappa_1}+1)}{(n_{jc}+\frac{\kappa_0\kappa_1}{\kappa_0 + \kappa_1})\bar{v}_{jc}}
\label{eq:i2gmm_seen}
\end{align}
where, $S_{\mu}$ is defined as in Equation (x1) in the supplementary material. The index $c$ in Equation (\ref{eq:i2gmm_seen}) represents the current seen class, whose PPD is being derived. Derivation details are provided in the supplementary material.
\fi
\begin{wrapfigure}{r}{0.40\textwidth}
\centering
\vspace{-10pt}
\includegraphics[scale=0.5]{figures/NIPS_surrogate_classes.pdf}
\vspace{-10pt}
\caption{Class formations for PPD during inference.}
\label{fig:surrogate_classes}
\end{wrapfigure}
\paragraph{Surrogate classes.} According to the generative model in (\ref{eq:i2gmm}), groupings among classes are determined based on local priors. Thus, once estimated from seen classes, local priors can be used to define surrogate classes for unseen ones during inference. Associating each unseen class with a unique local prior forms the basis of our approach. The local prior for each unseen class is defined by finding the $K$ seen classes most similar to that unseen class. The similarity is evaluated by computing the $\mathcal{L}^2$ (\textit{Euclidean})
distance between class-level attribute or embedding vectors ($\phi$) obtained from the side information available. Once a local prior is defined for each unseen class the PPD for the corresponding surrogate class can be derived in terms of only global and local priors as in equation (\ref{eq:i2gmm_unseen}). Test examples are classified based on class-conditional likelihoods evaluated for both seen and surrogate classes.
\begin{align}
P&(\boldsymbol{x}|\{\boldsymbol{\bar{x}_{ji}}, S_{ji}\}_{t_i=j}, \boldsymbol{\mu_{0}}, \kappa_0, \kappa_1) = T(\boldsymbol{x}|\boldsymbol{\bar{\mu}_{j}},\bar{\Sigma}_{j},\bar{v}_{j}); \qquad
\boldsymbol{\bar{\mu}_{j}} = \frac{\sum_{i:t_{i}=j}\frac{n_{ji}\kappa_{1}}{(n_{ji}+\kappa_{1})}\boldsymbol{\bar{x}_{ji}}+\kappa_{0}\boldsymbol{\mu_{0}}}{\sum_{i:t_{i}=j}\frac{n_{ji}\kappa_{1}}{(n_{ji}+\kappa_{1})}+\kappa_{0}},\nonumber\\
\bar{v}_{j}&= \sum_{i:t_{i}=j}(n_{ji}-1)+m-D+1, \quad
\bar{\Sigma}_{j} = \frac{(\Sigma_{0}+\sum_{i:t_{i}=j} S_{ji})(\tilde{\kappa}_{j}+1)}{\tilde{\kappa}_{j} \bar{v}_{j}}\label{eq:i2gmm_unseen}
\end{align}
where, $\boldsymbol{\bar{x}_{ji}}, S_{ji}$ and $n_{ji}$ represent sample mean, scatter matrix and size of class $i$ associated with local prior $j$, respectively and $\tilde{\kappa}_j$ is defined as in Eq. (30) in the supplementary material\footnote{ The code and dataset are available at \url{https://github.com/sbadirli/Fine-Grained-ZSL-with-DNA}}.
\paragraph{Rationale for the hierarchical Bayesian approach and limitations.}
We believe that the hierarchical Bayesian model is ideally suited for fine-grained zero-shot classification of species when DNA is used as side information for the following reasons. The performance of the model in identifying unseen classes depends on how robust the local priors can be estimated. This in turn depends on whether or not the set of seen classes contain any classes similar to unseen ones. As the number of seen classes increases, seen classes become more representative of their local priors, more robust estimates of local priors can be obtained, and thus, unseen classes sharing the same local priors as seen classes can be more accurately identified.
On the other hand, if the class-level side information is not specific enough to uniquely characterize a large number of classes, then the model cannot evaluate class similarity accurately and local priors are estimated based on potentially incorrect association between seen and unseen classes. In this case having a large number of seen classes available may not necessarily help. Instead, highly specific DNA as side information comes into play for accurately evaluating class similarity. If a unique local prior can be eventually described for each unseen class, then unseen classes can be classified during test time without the model having to learn the mapping between side information and image features beforehand. Uniqueness of the local prior can only be ensured when the number of seen classes is large compared to the number of unseen classes. Thus, the ratio of the number of seen and unseen classes becomes the ultimate determinant of performance for the hierarchical Bayesian model. The higher this ratio is the higher the accuracy of the model will be. An experiment demonstrating this effect is performed in Section \ref{sec:seenclasseffect}.
If the same set of $K$ classes is found to be the most similar for two different unseen classes, then these two unseen classes will inherit the same local prior and thus they will not be statistically identifiable during test time. The likelihood of such a tie happening for fine-grained data sets quickly decreases as the number of classes increases. In practice we deal with this problem by replacing the least similar of the $K$ most similar seen classes by the next most similar seen class for one of the unseen classes.
\section{Experiments}
In this section we report results of experiments with two species datasets that use DNA as side information. Details of training and hyperparameter tuning are provided in the supplementary material along with the source code of our methods.
\subsection{Experiments with the INSECT dataset}
\label{sec:insectexp}
We compare our model (BZSL) against state-of-the-art (SotA) ZSL methods proved to be most competitive on benchmark ZSL datasets that use visual attributes or word vector representations as side information. Selected SotA models represent various ZSL categories: (1) Embedding methods with traditional~\cite{ale,eszsl} and end-to-end neural network~\cite{crnet} approaches, (2) FGNs using VAE~\cite{cada_vae} and GAN~\cite{lsrgan}, and (3) end-to-end few shot learning approach extended to ZSL~\cite{Rel_net}.
\begin{wraptable}{r}{0.5\textwidth}
\centering
\vspace{-10pt}
\begin{tabular}{l|c c c}
\textbf{Method} & \textbf{US} & \textbf{S} & \textbf{H} \\ \hline
CRNet~\cite{crnet} & $13.33$ & $19.70$ & $15.90$ \\
ALE~\cite{ale} & $2.86$ & $27.18$ & $5.17$ \\
RelationNet~\cite{Rel_net} & $3.25$ & $24.37$ & $5.73$ \\
CADA-VAE~\cite{cada_vae} & $14.55$ & $20.81$ & $17.10$ \\
ESZSL~\cite{eszsl} & $3.41$ & $18.61$ & $5.77$ \\
LsrGan~\cite{lsrgan} & $12.58$ & $30.41$ & $17.75$ \\ \hline
\textbf{BZSL} & $\mathbf{20.83}$ & $\mathbf{38.30}$ & $\mathbf{26.99}$\\
\end{tabular}
\caption{Generalized ZSL results on Insect data using DNA barcodes as attributes.}
\label{tab:gzsl_results}
\end{wraptable}
Table~\ref{tab:gzsl_results} displays seen and unseen accuracies and their harmonic mean on the INSECT data using DNA as the side information. Results suggest that the large number of seen classes along with the highly specific nature of DNA information in characterizing classes particularly favors the Bayesian method to more accurately estimate local priors and characterize surrogate classes. The harmonic mean achieved by the proposed method is 52\% higher than the harmonic mean achieved by the second best performing technique. Similar levels of improvements are maintained on both seen and unseen class accuracies. The next top performers are FGNs. CADA-VAE uses a VAE whereas LsrGan utilizes GAN to synthesize unseen class features, then both train a \textit{LogSoftmax} classifier for inference. Lower unseen class accuracies suggest that FGNs struggle to synthesize meaningful features in the image space. On the other hand, CRNet that uses end-to-end neural network to learn the embedding between semantic and image spaces renders slightly worse performance than FGNs. It seems, non-linear embedding also works better than a linear (ESZSL) and bilinear (ALE) ones for this specific dataset. RelationNet is amongst the ones with the lowest performance, as the method is explicitly designed for Few-shot learning and expects the side information to be strongly correlated with image features. The weak association between side information and image features affects the performance of both FGNs and embedding methods, but the traditional embedding methods suffer the most.
\begin{table}[t]
\centering
\begin{tabular}{l c c c|c c c|c c c}
& \multicolumn{3}{c}{Attributes} & \multicolumn{3}{c}{Word Vectors} & \multicolumn{3}{c}{DNA}\\
\textbf{Method} & \textbf{US} & \textbf{S} & \textbf{H} & \textbf{US} & \textbf{S} & \textbf{H} & \textbf{US} & \textbf{S} & \textbf{H}\\ \hline
CRNet~\cite{crnet} & $44.28$ & $59.84$ & $50.89$ & $22.75$ & $45.92$ & $30.43$ & $9.27$ & $56.56$ & $15.93$ \\
ALE~\cite{ale} & $25.15$ & $60.80$ & $35.59$ & $3.95$ & $48.57$ & $7.31$ & $3.50$ & $50.18$ & $6.54$ \\
RelationNet~\cite{Rel_net} & $11.66$ & $44.81$ & $18.50$ & $8.67$ & $36.16$ & $13.99$ & $5.33$ & $40.83$ & $9.42$ \\
CADA-VAE~\cite{cada_vae} & $47.15$ & $53.11$ & $49.95$ & $26.45$ & $41.98$ & $\mathbf{32.45}$ & $19.42$ & $37.05$ & $25.48$ \\
ESZSL~\cite{eszsl} & $15.58$ & $50.66$ & $23.84$ & $2.26$ & $23.86$ & $4.12$ & $5.99$ & $5.38$ & $5.67$ \\
LsrGan~\cite{lsrgan} & $47.65$ & $56.97$ & $\mathbf{51.89}$ & $24.63$ & $37.96$ & $29.88$ & $15.99$ & $33.57$ & $21.66$ \\ \hline
\textbf{BZSL} & $31.49$ & $50.61$ & $38.82$ & $22.43$ & $45.00$ & $29.94$ & $27.46$ & $48.14$ & $\mathbf{34.97}$ \\ \hline \\
\end{tabular}
\caption{Generalized ZSL results on CUB data using original visual attributes, word vectors, and DNA attributes. }
\label{tab:CUB_DNA}
\end{table}
\subsection{Experiments with the benchmark CUB dataset}
To demonstrate the utility of DNA-based attributes in a broader spectrum of species classification, we procured DNA barcodes, again from the BOLD system, for bird species in the CUB dataset. For this experiment, we derived 400 dimensional embeddings in order to have the same size with word vectors and eliminate the attribute size effect. There were 6 classes, 4 seen and 2 unseen, that did not have DNA barcodes extracted from COI gene in the BOLD system. These classes were excluded from the dataset but the proposed split from \cite{gbu_tpami} is preserved otherwise.
The results shown in Table~\ref{tab:CUB_DNA} validate our hypothesis that when side information is not strongly correlated with visual characteristics of object classes (like in word vectors or DNA) both embedding methods and FGNs display significant performance degradation. With the exception of the proposed Bayesian model, word vector representation yields better accuracy than DNA-based attributes for all models. This phenomenon can be explained by our observation that text fragments related to common animals/birds in the Wikipedia/Internet often include some morphological traits of the underlying species. Hence, word vector representation is expected to have higher degree of correlation to visual attributes than DNA information. Our model produces the best results, $34.97\%$ vs $32.45\%$ when the side information is not derived from visual characteristics of classes. This outcome validates the robustness of the Bayesian model to diverse sources of side information and emphasizes the need for more robust FGN or embedding based models in more realistic scenarios where hand-crafted visual attributes are not feasible.
\subsection{The effect of the number of seen classes on performance}
\label{sec:seenclasseffect}
Local priors are central to the performance of the hierarchical Bayesian model. Here, we perform experiments to show that as the number of seen classes increases while the number of unseen classes fixed, each unseen class can be associated with a larger pool of candidate seen classes and more informative local priors can potentially be obtained, which in turn leads to more accurate identification of unseen classes. To demonstrate this effect we run two experiments. In the first experiment we use the same set of unseen classes as in Section \ref{sec:insectexp} but gradually increase the number of seen classes used for training. In the second experiment we double the size of the unseen classes and gradually include the remaining classes into training as seen classes. The first experiment is also performed for CADA-VAE. LsrGan is skipped for this experiment due to long training time. To account for random subsampling of seen classes each experiment is repeated five times and error bars are included in each plot. There is a clear trend in these results that further highlights the intuition behind the hierarchical Bayesian model and explains why this model is well-suited for fine-grained ZSL. When $10\%$ of the classes are used as unseen, unseen class accuracy improves with increasing number of seen classes until it flatlines beyond the $60\%$ mark while seen class accuracy always maintained around the same level (see Fig. \ref{fig:f1}). When $20\%$ of the classes are used as unseen no flatlining effect in unseen class accuracy is observed even at $100\%$ mark, which suggest that there is still room for improvement in unseen class accuracy if more seen classes become available (see Fig. \ref{fig:f2}). For CADA-VAE unseen class accuracy initially improves and then flatlines beyond $80\%$ mark but this improvement comes at the expense of significant degradation in seen class accuracy, which suggest that as the number of seen classes increase generated features further confound the classifier as would be expected of an FGN for a fine-grained dataset.
\begin{figure}[!tbp]
\begin{subfigure}[b]{0.31\textwidth}
\includegraphics[width=\textwidth]{figures/seen_cls_exp_v1.pdf}
\caption{BZSL results in original setup ($Y^{s}_{tr}= 1,092$ and $Y^{u}=121$)}
\label{fig:f1}
\end{subfigure}
\quad
\begin{subfigure}[b]{0.31\textwidth}
\includegraphics[width=\textwidth]{figures/seen_cls_exp_v2.pdf}
\caption{BZSL results with $Y^{s}_{tr}= 983$ and $Y^{u}=230$}
\label{fig:f2}
\end{subfigure}
\quad
\begin{subfigure}[b]{0.31\textwidth}
\includegraphics[width=\textwidth]{figures/cadavae_seen_cls_exp_v1.pdf}
\caption{CADA-VAE in original setup ($Y^{s}_{tr}= 1,092$ and $Y^{u}=121$)}
\label{fig:f3}
\end{subfigure}
\caption{The effect of the number of seen classes on the performance of BZSL and CADA-VAE. Each experiment is repeated five times to account for random subsampling of seen classes.}
\end{figure}
\subsection{Trade-off between seen and unseen class accuracies}
The Bayesian model can leverage different hyperparameter settings to modify the operating point of the classifier to favor seen class accuracy over unseen one or vice versa. In this experiment, we investigate the effect of $\kappa_0$ and $\kappa_1$ on seen and unseen class accuracies. Recall that $\kappa_0$ adjusts the dispersion of surrogate-class centers with respect to the center of the overall data and $\kappa_1$ adjusts the dispersion of actual class centers with respect to their corresponding surrogate-class centers. The smaller these parameters are the higher the dispersion will be.
\begin{wrapfigure}{r}{0.50\textwidth}
\vspace{-10pt}
\centering
\includegraphics[scale=0.5]{figures/k0_vs_k1_m5d_s10_K3.pdf}
\vspace{-15pt}
\caption{Effects of $\kappa_0$ and $\kappa_1$ on INSECT data.}
\label{fig:kappas}
\vspace{-5pt}
\end{wrapfigure}
The impression from Figure~\ref{fig:kappas} reflects that unseen class accuracy is highest when $\kappa_1$ is close to $1$, more precisely $\kappa_1 \in [0.1, 1]$, and drops significantly lower in both directions, i.e., for $\kappa_1<<1$ and $\kappa_1>>1$. As expected the opposite of this pattern is observed for seen class accuracy. Although both seen and unseen class accuracies are highly responsive to the selection of $\kappa_1$, the changes are less receptive with respect to $\kappa_0$. Moving $\kappa_1$ towards zero encodes a local prior that imposes unrealistically large dispersion for centers of actual-classes sharing the same local prior, which violates the main assumption of our model that classes sharing the same local prior are supposed to be semantically similar classes. On the other hand moving $\kappa_1$ towards infinity encodes a local prior that imposes limited to no deviation among centers of actual classes which is another extreme that is not true for real-world datasets, i.e. classes are supposed to be statistically identifiable.
In both extremes unrealistic prior assumptions that cannot be reconciled with the characteristics of real-world data sets impede knowledge transfer between seen and unseen classes and lead to poor classification performance on unseen classes. On the other hand, the same extreme assumptions do not affect seen class accuracies at the same scale, because seen classes circumvent local priors and are modeled with the data likelihood. Since INSECT data is very fine-grained, harmonic mean peaks when $\kappa_1 \geq 1$. We also conduct an ablation study to investigate the effect of different components of the model on the performance, the results of which are reported in the supplementary material.
\section{Conclusions}
For the first time in the ZSL literature we use DNA as a side information and demonstrate its utility in evaluating class similarity for the purpose of identifying unseen classes in a fine-grained ZSL setting. On the CUB dataset, despite being trained with less than 30,000 very short sequences, we find DNA embeddings to be highly competitive with word vector representations trained on massive text corpora. We emphasize the importance of DNA as side information in zero-shot classification of highly fine-grained species datasets involving thousands of species, and on the INSECT dataset, show that a simple Bayesian model that readily exploits inherent class hierarchy with the help of DNA can significantly outperform highly complex models. We show that SotA ZSL methods that take the presence of an explicit association between visual attributes and image features for granted, suffer significant performance degradation when non-visual attributes such as word vectors and WordNet are used as side information. The same effect is observed with DNA embeddings as well. Although visual attributes tend to be the best alternative as side information for a coarse-grained species classification task, they quickly lose their appeal with an increasing number of classes. Considering the tens of thousands of \textit{described} species and even larger number of \textit{undescribed} species, DNA seems to be the only feasible alternative to side information for large-scale, fine-grained zero-shot classification of species.
These favorable results by a simpler model suggest that as the number of classes increases along with inter-class similarity, the complexity of the mapping between side information and image attributes emerges as a major bottleneck at the forefront of zero-shot classification. A promising future research avenue appears to be implementing hierarchically organized FGNs where each subcomponent only operates with a small subset of seen classes all sharing the same local prior.
This work does not present any foreseeable negative societal consequences beyond those already associated with generic machine learning classification algorithms.
\bibliographystyle{ieee.bst}
|
\section*{Overview}\addcontentsline{toc}{section}{\protect\numberline{}\hspace{-0.9cm}Overview}}
\newcommand{\textbf{\Huge{List of Algorithms}}}{\textbf{\Huge{List of Algorithms}}}
\newlistof{Algorithm}{exp}{\textbf{\Huge{List of Algorithms}}}
\newcounter{instructioncounter}
\renewcommand{\theinstructioncounter}{\textbf{(\theAlgorithm\alph{instructioncounter})}}
\newenvironment{Algorithm}[1][]
{\refstepcounter{Algorithm}
\begin{framed}\addcontentsline{exp}{Algorithm}{\protect\numberline{\theAlgorithm} #1}\par\begin{center}\textbf{Algorithm \theAlgorithm}\end{center}\begin{list}
{\bf{(\arabic{Algorithm}\alph{instructioncounter}})}{\usecounter{instructioncounter}}}{\end{list}\end{framed}}
\addbibresource{biblio.bib}
\begin{document}
\title{Swapping in lattice-based cell migration models}
\author{Shahzeb Raja Noureen\inst{1}\thanks{Corresponding author.
E-mail: <EMAIL>} \and \qquad Jennifer P. Owen\inst{1} \and \qquad Richard L. Mort\inst{2}
\and \qquad Christian A. Yates\inst{1} }
\institute{Centre for Mathematical Biology, Department of Mathematical Sciences, University of Bath, Claverton Down, Bath, BA2 7AY, UK.
\and
Division of Biomedical and Life Sciences, Faculty of Health and Medicine, Furness Building, Lancaster University, Bailrigg, Lancaster, LA1 4YG, UK.
}
\date{Received: date / Revised version: date}
\maketitle
\begin{abstract}
Cell migration is frequently modelled using on-lattice agent-based models (ABMs) that employ the excluded volume interaction. However, cells are also capable of exhibiting more complex cell-cell interactions, such as adhesion, repulsion, pulling, pushing and swapping. Although the first four of these have already been incorporated into mathematical models for cell migration, swapping is an interaction that has not been well studied in this context. In this paper, we develop an ABM to describe cell movement where an active agent can `swap' its position with another agent in its neighbourhood with a given swapping probability. We consider single-species and two-species systems. In both cases, we derive the corresponding macroscopic model and compare it with the average behaviour of the ABM. We see good agreement between the ABM and the macroscopic density. We also derive an expression for the cell-level diffusion coefficient in terms of the swapping probability and cell density. We conclude by showing applications of swapping by using the ABM to represent cell movement with proliferation and cell-cell adhesion.
\end{abstract}
\pagebreak
\section{Introduction}
Cell migration is an essential biological process required for the correct development of tissues and organs during embryonic development and their proper maintenance, by wound healing and tissue homeostasis, throughout life \parencite{mayor2016frc,giniunaite2020mcc,thomas2008tmm,wang2011asn}. In-depth understanding of cell migration is also important for identifying causes of neurocristopathies such as Hirschsprung's disease and neurofibromatosis, as well as developing new therapeutic targets to prevent metastasis in cancers \parencite{mayor2016frc,deroulers2009mtc}.
Traditionally, many biological problems have been modelled using deterministic methods. However, in cell migration, randomness can play a salient role in determining a cell's trajectory and fate and hence deterministic theory may not be appropriate. Extensive research has gone into modelling cell movement as a stochastic process. In one widely used approach, cells are modelled as agents whose positions evolve probabilistically in space and time according to a predefined set of rules. These models are commonly known as agent-based models (ABMs) or individual-based models (IBMs). The agent-based modelling paradigm can be sub-divided into off-lattice and on-lattice models, both of which have wide applicability to different problems within mathematical biology. \citet{gavagnin2018sdm} recently reviewed the most commonly used ABMs for cell movement.
In this paper, we only concern ourselves with on-lattice models of cell movement. In a lattice-based approach, the domain is divided into a series of compartments in which the cells reside. Cells take up space, preventing other cells from occupying the same space at the same time. For biological plausibility, it is often desirable that mathematical models of cell migration account for the single occupancy of sites. This realism is incorporated in an ABM via the volume exclusion principle, which states that a cell attempting to move into a neighbouring site successfully moves only if the neighbouring site is not already occupied at the time of moving.
Models with volume exclusion at their core have been used to describe the collective migration of cells for a wide range of biological applications. \citet{mort2016rdm} used an on-lattice ABM with the exclusion principle to model the invasion of the developing epidermis by melanoblasts (the embryonic precursors of melanocytes) and investigate the basis of piebaldism in mice. Lattice-based exclusion models have also been applied to wound healing \parencite{khain2007rcc}, migration of breast cancer cells \parencite{simpson2010mbc}, developmental processes on growing domains \parencite{baker2010fmt}, cells' responses to chemotaxis \parencite{charteris2014mca} and cells exhibiting pushing \parencite{yates2015ipe} and pulling \parencite{chappelle2019pmc} interactions in densely crowded environments.
\begin{figure}[t!]
\centering
\includegraphics[width=\textwidth]{figures/images/cell_swapping_1.png}
\caption{Two \textit{Fucci2a} labelled NIH 3T3 mouse embryonic fibroblasts swapping places with each other in culture \parencite{mort2014fbc}. The colour of the cells represents their cell cycle stage (Red = G1, Green = S/G2/M) in this case making it easy to observe the swap. In (a), we show the initial placement of cells: cell 1 (shown in red) is on the bottom-left of cell 2 (shown in green). In (b) a swap starts to take place and in (c) the swap is complete and now cell 1 is on the top-right of cell 2.}
\label{fig:cell_swapping_1}
\end{figure}
Biologically, although cells are excluded from the space occupied by other cells, their movement is not completely inhibited by them as typically assumed in volume exclusion models. For example melanoblasts are able to move freely between keratinocytes in the developing epidermis \parencite{mort2016rdm}. Experimental data suggests that cells are often able to move past each other (passing laterally, above or below) exchanging places with one another. In Figure \ref{fig:cell_swapping_1} we show experimental images of two \textit{Fucci2a} labelled NIH-3T3 fibroblasts exhibiting the swapping behaviour. Swapping has also been observed in blood cells such as leukocytes, erythrocytes and thrombocytes \parencite{Lan2014nsp} and in pattern formation for maintaining sharp boundaries between different groups of cells as part of a cell sorting mechanism \parencite{dahmann2011bfm}. These examples highlight the importance of incorporating swapping into models of cell migration. To the best of our knowledge, this has not yet been explored thoroughly from a mathematical perspective.
In this paper, we develop a mathematical model to describe and analyse cell-cell swapping. By modifying the movement rules of the traditional volume exclusion process, we show that swapping between agents has an effect on the migration of agents at different spatial resolutions. For movement of agents at an individual level, we derive an expression for the diffusion coefficient. To investigate how swapping manifests itself in the corresponding population-level model (PLM) we derive a set of partial differential equations (PDEs) describing the macroscopic dynamics of the agents. We compare numerical solutions of the PDEs with the averaged results from the ABM and comment on the agreement or discrepancy between them.
In the remainder of this paper we develop a model that allows for swapping to take place between pairs of neighbouring agents (Section \ref{sec:swapping_model}). We then derive the mean-field PDEs and compare
the average behaviour of the ABM to that of the PDEs and analyse the
individual-level dynamics of the agents, determining a dependency between the individual-level
diffusion coefficient, the parameter associated with swapping and the background domain density.
We then give examples that illustrate the applications of swapping (Section \ref{sec:examples}) and conclude the paper with a discussion and summary (Section \ref{sec:discussion}).
\section{Cell migration model with swapping}\label{sec:swapping_model}
We begin by describing an ABM for cell movement with swapping in Section \ref{sec:abm}. We will employ this model throughout the rest of the paper unless stated otherwise. We consider two scenarios. In the first, we investigate the effect of swapping in the single-species case, i.e. all the agents are identical to each other, and in the second, we consider the effects of swapping for situations in which there are two distinct types of agents populating the domain. In both cases, we derive the corresponding population-level models and compare these with the average behaviour of the ABM in Section \ref{sec:plm}. To analyse individual-level movement, we derive an analytical expression for the individual-level diffusion coefficient and show that it matches the diffusion coefficient estimated via simulation of the ABM in Section \ref{sec:individual_level_analysis}.
\subsection{On-lattice agent-based model}\label{sec:abm}
We model cell migration on a two-dimensional lattice. We discretise the domain into compartments (also known as `sites') such that there are $L_x$ compartments in the horizontal direction and $L_y$ compartments in the vertical direction. We assume that the compartments are square with side length $\mathrm{\Delta}$. Supposing that each compartment can contain no more than one agent, $\mathrm{\Delta}$ can be considered a rough proxy for a cell's diameter. A site $(i,j)$ for $i=1,...,L_x$ and $j=1,...,L_y$ can be either occupied by an agent or unoccupied. Occupancy is defined as a binary indicator, taking a value of unity if there is an agent at the site $(i,j)$ or 0 if the site is empty.
\begin{figure}[t!]
\begin{center}
\subfigure[]{
\includegraphics[width=0.22\textwidth]{figures/schematics/swap_lattice_1.pdf}
\label{figure:lattice_sd}
}
\subfigure[]{
\includegraphics[width=0.22\textwidth]{figures/schematics/swap_lattice_2.pdf}
\label{figure:lattice_swap1}
}
\subfigure[]{
\includegraphics[width=0.22\textwidth]{figures/schematics/swap_lattice_3.pdf}
\label{figure:lattice_swap2}
}
\end{center}
\caption{A schematic illustrating the swapping mechanism. Red sites are occupied with agents and black sites are unoccupied. The initial configuration of the lattice before is shown in \subref{figure:lattice_sd}. The agent chosen to move is at site $(i,j)$ and labelled 1. The target site is at position $(i+1,j)$ and the agent occupying the target site is labelled 2. Agent 1 attempts to move into the target site in \subref{figure:lattice_swap1}. The final configuration once the swapping move is complete is shown in \subref{figure:lattice_swap2}.}
\label{figure:swap_single_species}
\end{figure}
We initialise $N$ agents on the lattice (their initial positions dependent on the specific scenarios we investigate later in this section) and let their positions evolve in continuous time using the Gillespie algorithm \parencite{gillespie1977ess}. We let $r_m$ be the rate of movement of an agent chosen uniformly at random such that $r_m \delta t$ is the probability that the agent attempts to move during a finite time interval $[t,t+\delta t]$ of duration $\delta t$. The agent attempts to move into one of its four neighbouring sites with equal probability. In an exclusion process, if the chosen neighbouring site is empty, the focal agent successfully moves and its position is updated. However, if the site is already occupied by another agent, the move is aborted \parencite{gavagnin2018sdm,simpson2015dpg,yates2015ipe,chappelle2019pmc,simpson2009mss}. In the two-species case with agents of types M and X, we can have a different rate of movement for each type: $r_{m}^M$ for a type-M agent and $r_{m}^X$ for a type-X agent. The dynamics of the agents remain the same as described above.
Swapping works by modifying the rules of the exclusion process by allowing an exchange in the positions of two neighbouring agents if the target site is already occupied. We introduce the swapping parameter $\rho$ denoting the probability of a successful swap between a pair of neighbouring agents conditional on one of the agents attempting to move into the other's position. If $\rho=0$ then there are no swaps and we arrive back at the exclusion process. If $\rho> 0$ then we can have different levels of swapping based on the value of $\rho$. To implement swapping, we sample a random number $u$ from the uniform distribution over the unit interval $(0,1)$. If $u<\rho$, the agent at the site $(i,j)$ swaps with the agent at the target site (Figure \ref{figure:lattice_sd}-\subref{figure:lattice_swap2}), otherwise the move is aborted and the state of the system remains unchanged.
\begin{figure}[h!]
\begin{center}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/sss_density_Ps=0_t=0-eps-converted-to.pdf}
\label{figure:sss_Ps=0_t=0}
}
\setcounter{subfigure}{3}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/sss_density_Ps=0.5_t=0-eps-converted-to.pdf}
\label{figure:sss_Ps=0.5_t=0}
}
\setcounter{subfigure}{1}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/sss_density_Ps=0_t=100-eps-converted-to.pdf}
\label{figure:sss_Ps=0_t=100}
}
\setcounter{subfigure}{4}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/sss_density_Ps=0.5_t=100-eps-converted-to.pdf}
\label{figure:sss_Ps=0.5_t=100}
}
\setcounter{subfigure}{2}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/sss_density_Ps=0_t=1000-eps-converted-to.pdf}
\label{figure:sss_Ps=0_t=1000}
}
\setcounter{subfigure}{5}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/sss_density_Ps=0.5_t=1000-eps-converted-to.pdf}
\label{figure:sss_Ps=0.5_t=1000}
}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/sss_density_Ps=1_t=0-eps-converted-to.pdf}
\label{figure:sss_Ps=1_t=0}
}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/sss_density_Ps=1_t=100-eps-converted-to.pdf}
\label{figure:sss_Ps=1_t=100}
}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/sss_density_Ps=1_t=1000-eps-converted-to.pdf}
\label{figure:sss_Ps=1_t=1000}
}
\end{center}
\caption{Snapshots of lattice occupancy for the single-species swapping model at $t=0, 100,1000$ with $r_m=1$ for swapping probabilities $\rho=0$ [\subref{figure:sss_Ps=0_t=0}-\subref{figure:sss_Ps=0_t=1000}], $\rho=0.5$ [\subref{figure:sss_Ps=0.5_t=0}-\subref{figure:sss_Ps=0.5_t=1000}] and $\rho=1$ [\subref{figure:sss_Ps=1_t=0}-\subref{figure:sss_Ps=1_t=1000}]. Agents are initialised on a domain with dimensions $L_x=200$ and $L_y=20$ such that all the sites in the range $81 \leqslant x \leqslant 120$ are occupied with agents (green) [\subref{figure:sss_Ps=0_t=0},\subref{figure:sss_Ps=0.5_t=0},\subref{figure:sss_Ps=1_t=0}]. Further snapshots of the IBM at $t=100$ and $t=1000$ show the dispersal of agents with time. The column-averaged density of agents over 100 runs of the ABM is also plotted (shown in black). We impose reflective boundary conditions on all four boundaries of the domain.}
\label{figure:sss_profiles}
\end{figure}
\begin{figure}[h!]
\begin{center}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/two_species_rho=0_t=0-eps-converted-to.pdf}
\label{figure:two_species_rho=0_t=0}
}
\setcounter{subfigure}{3}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/two_species_rho=0.5_t=0-eps-converted-to.pdf}
\label{figure:two_species_rho=0.5_t=0}
}
\setcounter{subfigure}{1}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/two_species_rho=0_t=100-eps-converted-to.pdf}
\label{figure:two_species_rho=0_t=100}
}
\setcounter{subfigure}{4}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/two_species_rho=0.5_t=100-eps-converted-to.pdf}
\label{figure:two_species_rho=0.5_t=100}
}
\setcounter{subfigure}{2}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/two_species_rho=0_t=1000-eps-converted-to.pdf}
\label{figure:two_species_rho=0_t=1000}
}
\setcounter{subfigure}{5}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/two_species_rho=0.5_t=1000-eps-converted-to.pdf}
\label{figure:two_species_rho=0.5_t=1000}
}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/two_species_rho=1_t=0-eps-converted-to.pdf}
\label{figure:two_species_rho=1_t=0}
}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/two_species_rho=1_t=100-eps-converted-to.pdf}
\label{figure:two_species_rho=1_t=100}
}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/two_species_rho=1_t=1000-eps-converted-to.pdf}
\label{figure:two_species_rho=1_t=1000}
}
\end{center}
\caption{Snapshots of the lattice occupancy of the multi-species swapping model at $t=0,100,1000$ using $r_{m}^M=r_{m}^X=1$ for swapping probabilities $\rho=0$ [\subref{figure:two_species_rho=0_t=0}-\subref{figure:two_species_rho=0_t=1000}], $\rho=0.5$ [\subref{figure:two_species_rho=0.5_t=0}-\subref{figure:two_species_rho=0.5_t=1000}] and $\rho=1$ [\subref{figure:two_species_rho=1_t=0}-\subref{figure:two_species_rho=1_t=1000}]. Agents are initialised on a domain with dimensions $L_x=200$ and $L_y=20$ such that all the sites in the range $81 \leqslant x \leqslant 120$ are occupied by type-M agents (red) and the remaining sites are randomly populated with agents of type X (green) at a density of 0.5 [\subref{figure:two_species_rho=0_t=0},\subref{figure:two_species_rho=0.5_t=0},\subref{figure:two_species_rho=1_t=0}]. Further snapshots of the ABM at $t=100$ and $t=1000$ show the dispersal of agents with time. The column-averaged density of the two species over 100 runs of the ABM is also plotted (shown in black). We impose reflective boundary conditions on all four boundaries of the domain.}
\label{figure:multispecies_density}
\end{figure}
In Figure \ref{figure:sss_profiles} we present snapshots of the lattice occupancy for the single-species case at times $t=0,100,1000$ and for swapping probabilities $\rho=0,0.5,1$ with $r_m=1$. We see that swapping seems to have no effect on the dispersion of the agents at a macroscopic scale as the density profiles are indistinguishable regardless of the swapping probability. If two agents are identical to each other and unlabelled then exchanging positions by swapping is equivalent to an aborted movement attempt in the volume exclusion process and produces no change in the state of the system. On the other hand, by introducing a second species, as shown in Figure \ref{figure:multispecies_density} (with rates $r_m^M=r_m^X=1$), we see that swapping appears to speed up the mixing of the two species causing both to disperse faster than they would in the classical volume exclusion process. Initially, the effect of swapping is only pronounced at the interface between the red agents and the green agents where a red agent swapping with a green agent gives rise to a noticeable change in the state of the system. Away from this interface a swap between two unlabelled agents of the same type produces no noticeable effect. As time progresses, the system becomes more and more well-mixed making it more likely for two agents of different types to be next to each other and hence swapping makes a greater difference to the diffusion of the agents.
In the next section, we derive the macroscopic PDEs describing the evolution of the mean lattice occupancy. By analysing the PDEs we generate further insight into the behaviours seen in Figures \ref{figure:sss_profiles} and \ref{figure:multispecies_density}.
\subsection{Continuum model for average occupancy}\label{sec:plm}
In Section \ref{sec:ss_pde}, we derive the PDE corresponding to the single-species model, extending it to the two-species case in Section \ref{sec:sm_pde}.
\subsubsection{Single-species continuum model}\label{sec:ss_pde}\hfill\\
Let $C_{ij}^r(t)$ be the occupancy of site $(i,j)$ on the $r$th repeat at time $t$, where $C_{ij}^r(t)=1$ if the site $(i,j)$ is occupied and $C_{ij}^r(t)=0$ otherwise. The average occupancy of site $(i,j)$ at time $t$ after $R$ runs is then given by,
\begin{equation}
C_{ij}(t)=\frac{1}{R}\sum_{r=1}^R C_{ij}^r(t).
\end{equation}
By considering the possible movement events of the agent at the site $(i,j)$ during the small time step $\delta t$ \parencite{yates2015ipe,chappelle2019pmc}, we can write down the master equation for the occupancy of the site at time $t+\delta t$,
\begin{align}
C_{ij}(t+\delta t)-C_{ij}(t)&=\frac{r_m}{4}\delta t[ (1-C_{ij})(C_{i-1,j} + C_{i+1,j} + C_{i,j-1} + C_{i,j+1}) \nonumber \\
& \quad -C_{ij}(4-C_{i-1,j} - C_{i+1,j} - C_{i,j-1} - C_{i,j+1})] \nonumber \\
& \quad + \frac{r_m}{4}\rho \delta t C_{ij}(C_{i-1,j} + C_{i+1,j} + C_{i,j-1} + C_{i,j+1}) \nonumber\\
& \quad + \frac{r_m}{4}\rho \delta t C_{ij}(C_{i-1,j} + C_{i+1,j} + C_{i,j-1} + C_{i,j+1}) \nonumber\\
& \quad - \frac{r_m}{4}\rho \delta t C_{ij}(C_{i-1,j} + C_{i+1,j} + C_{i,j-1} + C_{i,j+1}) \nonumber\\
& \quad - \frac{r_m}{4}\rho \delta t C_{ij}(C_{i-1,j} + C_{i+1,j} + C_{i,j-1} + C_{i,j+1}). \label{eqn:ss_occupancy_eqn}
\end{align}
The site $(i,j)$ can gain occupancy if it is unoccupied at time $t$ and the agent from a neighbouring site moves in (first line on the RHS of equation \eqref{eqn:ss_occupancy_eqn}). Similarly, the site $(i,j)$ can lose occupancy if the site is occupied at time $t$ and the residing agent jumps out to a neighbouring compartment, leaving the site $(i,j)$ empty (second line in equation \eqref{eqn:ss_occupancy_eqn}). Movement of agents due to a successful swapping event is captured by the lines 3-6 in equation \eqref{eqn:ss_occupancy_eqn}. However, since the agents are assumed to be identical and unlabelled lines 3-6 cancel each other out, eliminating the effect of swapping. Consequently, we are left with lines 1 and 2 only. Taylor expanding these remaining terms around the site $(i,j)$ up to second-order and taking the limit as $\mathrm{\Delta} \to 0$ gives the diffusion equation, as expected:
\begin{equation}\label{eqn:diffusion_eqn}
\D C t = D \nabla^2 C.
\end{equation}
Here,
\begin{equation*}
D = \lim_{\mathrm{\Delta} \to 0} \frac{r_m \mathrm{\Delta}^2}{4},
\end{equation*}
is the diffusion coefficient given that $\mathrm{\Delta^2}/\delta t$ is held constant in the diffusive limit. Equation \eqref{eqn:diffusion_eqn} describes the evolution of the lattice occupancy over time. It is well-known that for the simple exclusion process, the occupancy is described by the diffusion equation \parencite{simpson2009mss,gavagnin2018sdm}. It makes sense therefore that in Figure \ref{figure:sss_profiles} swapping made no difference to the overall occupancy of the lattice. In Figure \ref{figure:density_compar_single_species} we compare the average column density of the ABM which is given by,
\begin{equation*}
\overline{C}_i(t)=\frac{1}{L_y}\sum_{j=1}^{L_y} C_{ij}(t),
\end{equation*}
to the solution of the one-dimensional analogue of equation \eqref{eqn:diffusion_eqn} with reflective boundary conditions by averaging the occupancy over the $y$ direction. As expected, we see an excellent agreement between the two density profiles.
\begin{figure}[t!]
\begin{center}
\subfigure[]{
\includegraphics[width=0.31\textwidth]{figures/plots/sss_density_only_rho=0-eps-converted-to.pdf}
\label{figure:sss_density_only_rho=0}
}
\subfigure[]{
\includegraphics[width=0.31\textwidth]{figures/plots/sss_density_only_rho=0.5-eps-converted-to.pdf}
\label{figure:sss_density_only_rho=0.5}
}
\subfigure[]{
\includegraphics[width=0.31\textwidth]{figures/plots/sss_density_only_rho=1-eps-converted-to.pdf}
\label{figure:sss_density_only_rho=1}
}
\end{center}
\caption{A comparison between the numerical solution of the one-dimensional version of the PDE \eqref{eqn:diffusion_eqn} and the averaged behaviour of the ABM with $r_m=1$ for $\rho=0$ \subref{figure:sss_density_only_rho=0}, $\rho=0.5$ \subref{figure:sss_density_only_rho=0.5} and $\rho=1$ \subref{figure:sss_density_only_rho=1}. The averaged densities are shown in black and the corresponding PDE approximation is shown in green. We present solutions at $t=0$, $t=100$ and $t=1000$. The black arrows show the direction of increasing time.}
\label{figure:density_compar_single_species}
\end{figure}
In the next Section, we derive the PLM for the two-species swapping case and compare the resulting PDEs with the average dynamics of the ABM.
\subsubsection{Two-species continuum model}\label{sec:sm_pde}\hfill\\
Let $M_{ij}^r(t)$ be the occupancy of site $(i,j)$ at time $t$ on the $r$th repeat of the ABM such that $M_{ij}^r(t)=1$ if the site is occupied by a type-M agent and 0 otherwise. Let $X_{ij}^r(t)$ be the same for a type-X agent. Then the average density of type-M and type-X agents after $R$ repeats is given by,
\begin{equation}
M_{ij}(t)=\frac{1}{R}\sum_{r=1}^R M_{ij}^r(t), \quad \text{and} \quad X_{ij}(t)=\frac{1}{R}\sum_{r=1}^R X_{ij}^r(t).
\end{equation}
By considering all the possible ways in which the site $(i,j)$ can gain or lose occupancy of either type-M or type-X agents during the time step $\delta t$, we can write down the corresponding occupancy master equations at time $t+\delta t$:
\begin{align}
M_{ij}(t+\delta t)-M_{ij}(t)&=\frac{r_{m}^M}{4}\delta t[(1-M_{ij}-X_{ij})(M_{i-1,j}+M_{i+1,j}+M_{i,j-1}+M_{i,j+1}) \nonumber\\
& \quad -M_{ij}(4-M_{i-1,j}-M_{i+1,j}-M_{i,j-1}-M_{i,j+1}-X_{i-1,j}-X_{i+1,j} -X_{i,j-1}-X_{i,j+1})] \nonumber \\
& \quad + \frac{(r_{m}^M+r_m^X)}{4}\rho \delta t X_{ij}(M_{i-1,j}+M_{i+1,j}+M_{i,j-1}+M_{i,j+1})\nonumber \\
&\quad -\frac{(r_{m}^M+r_m^X)}{4}\rho \delta t M_{ij}(X_{i-1,j}+X_{i+1,j}+X_{i,j-1}+X_{i,j+1}), \label{eqn:master_eqn_M}
\end{align}
\begin{align}
X_{ij}(t+\delta t)-X_{ij}(t)&=\frac{r_{m}^X}{4}\delta t[(1-X_{ij}-M_{ij})(X_{i-1,j}+X_{i+1,j}+X_{i,j-1}+X_{i,j+1}) \nonumber\\
& \quad -X_{ij}(4-X_{i-1,j}-X_{i+1,j}-X_{i,j-1}-X_{i,j+1}-M_{i-1,j}-M_{i+1,j} -M_{i,j-1}-M_{i,j+1})] \nonumber\\
&\quad + \frac{(r_{m}^M+r_m^X)}{4}\rho \delta t M_{ij}(X_{i-1,j}+X_{i+1,j}+X_{i,j-1}+X_{i,j+1})\nonumber \\
&\quad - \frac{(r_{m}^M+r_m^X)}{4}\rho \delta t X_{ij}(M_{i-1,j} + M_{i+1,j}+M_{i,j-1}+M_{i,j+1}). \label{eqn:master_eqn_X}
\end{align}
Here, we describe the terms in equation \eqref{eqn:master_eqn_M}. The terms in equation \eqref{eqn:master_eqn_X} carry similar interpretations. A site $(i,j)$ can gain occupancy of type M in one of three ways. In the first, the site is unoccupied and a type-M agent moves in from a neighbouring site (line 1 in equation \eqref{eqn:master_eqn_M}). In the second, the site can be occupied by a type-X agent, which initiates and completes a swap with a type-M agent at a neighbouring site (line 3 in equation \eqref{eqn:master_eqn_M}). In the third, the site can be occupied by a type-X agent and this time the type-M agent at the neighbouring site initiates a swap to exchange positions with the agent in site $(i,j)$ (line 3 in equation \eqref{eqn:master_eqn_M}). In all three cases, a type-M agent moves into the site $(i,j)$. Similarly, there are three ways for a type-M agent to move out of the site $(i,j)$ leading to a loss in the corresponding occupancy of type M. The site can either be occupied by a type-M agent and the agent jumps out to a previously empty site, leaving site $(i,j)$ empty (line 2 in equation \eqref{eqn:master_eqn_M}), or the site $(i,j)$ can be occupied by a type-M agent which initiates a swap with a type-X agent in its neighbourhood or the type-X agent in the neighbourhood initiates a swap to exchange positions with the agent in site $(i,j)$ (line 4 in equation \eqref{eqn:master_eqn_M}). In all three cases, a type-M agent moves out of the site $(i,j)$.
To obtain the continuum model, we Taylor expand the appropriate terms on the RHS of equations \eqref{eqn:master_eqn_M} and \eqref{eqn:master_eqn_X} around the site $(i,j)$ keeping terms of up to second order. By letting $\mathrm{\Delta} \to 0$ and $\delta t \to 0$ such that $\mathrm{\Delta}^2/\delta t$ is held constant, we arrive at the coupled PDEs,
\begin{align}
\D M t &= \nabla \cdot [D_1(X) \nabla M + D_2(M)\nabla X], \label{eqn:pde_M}\\
\D X t &= \nabla \cdot [D_3(M)\nabla X + D_4(X)\nabla M] \label{eqn:pde_X},
\end{align}
where,
\begin{equation}
D_1=D_M(1-X)+\rho(D_M+D_X)X, \qquad D_2=(D_M -\rho(D_M+D_X))M,
\end{equation}
\begin{equation}
D_3=D_X(1-M)+\rho(D_M+D_X)M, \qquad D_4=(D_X - \rho(D_M+D_X))X.
\end{equation}
Here,
\begin{equation*}
D_M=\lim_{\mathrm{\Delta} \to 0} \frac{r_{m}^M \mathrm{\Delta}^2}{4}, \quad \text{and} \quad D_X=\lim_{\mathrm{\Delta} \to 0} \frac{r_{m}^X \mathrm{\Delta}^2}{4},
\end{equation*}
are the macroscopic diffusion coefficients corresponding to species M and X, respectively. Setting $\rho=0$ in equations \eqref{eqn:pde_M} and \eqref{eqn:pde_X} leads to,
\begin{align}
\D M t &= D_M \nabla \cdot [(1-X)\nabla M + M\nabla X], \label{eqn:two_species_M_simpson_mss}\\
\D X t &= D_X \nabla \cdot [(1-M)\nabla X + X\nabla M], \label{eqn:two_species_X_simpson_mss}
\end{align}
which are the macroscopic equations for the two-species volume exclusion process \parencite{simpson2009mss}.
\begin{figure}[h!]
\begin{center}
\subfigure[]{
\includegraphics[width=0.31\textwidth]{figures/plots/density_plot_IC=1_rho=0-eps-converted-to.pdf}
\label{figure:IC=1_P_s=0}
}
\subfigure[]{
\includegraphics[width=0.31\textwidth]{figures/plots/density_plot_IC=1_rho=0.5-eps-converted-to.pdf}
\label{figure:IC=1_P_s=0.5}
}
\subfigure[]{
\includegraphics[width=0.31\textwidth]{figures/plots/density_plot_IC=1_rho=1-eps-converted-to.pdf}
\label{figure:IC=1_P_s=1}
}
\subfigure[]{
\includegraphics[width=0.31\textwidth]{figures/plots/density_plot_IC=2_rho=0-eps-converted-to.pdf}
\label{figure:IC=2_P_s=0}
}
\subfigure[]{
\includegraphics[width=0.31\textwidth]{figures/plots/density_plot_IC=2_rho=0.5-eps-converted-to.pdf}
\label{figure:IC=2_P_s=0.5}
}
\subfigure[]{
\includegraphics[width=0.31\textwidth]{figures/plots/density_plot_IC=2_rho=1-eps-converted-to.pdf}
\label{figure:IC=2_P_s=1}
}
\subfigure[]{
\includegraphics[width=0.31\textwidth]{figures/plots/density_plot_IC=4_rho=0-eps-converted-to.pdf}
\label{figure:IC=4_P_s=0}
}
\subfigure[]{
\includegraphics[width=0.31\textwidth]{figures/plots/density_plot_IC=4_rho=0.5-eps-converted-to.pdf}
\label{figure:IC=4_P_s=0.5}
}
\subfigure[]{
\includegraphics[width=0.31\textwidth]{figures/plots/density_plot_IC=4_rho=1-eps-converted-to.pdf}
\label{figure:IC=4_P_s=1}
}
\end{center}
\caption{A comparison between the numerical solution of the PDEs \eqref{eqn:pde_M} and \eqref{eqn:pde_X} and the averaged behaviour of the ABM for the migration process with swapping with different swapping probabilities: $\rho=0$ for [\subref{figure:IC=1_P_s=0}, \subref{figure:IC=2_P_s=0}, \subref{figure:IC=4_P_s=0}]; $\rho=0.5$ for [\subref{figure:IC=1_P_s=0.5}, \subref{figure:IC=2_P_s=0.5}, \subref{figure:IC=4_P_s=0.5}] and $\rho=1$ for [\subref{figure:IC=1_P_s=1}, \subref{figure:IC=2_P_s=1}, \subref{figure:IC=4_P_s=1}]. The initial conditions for [\subref{figure:IC=1_P_s=0}-\subref{figure:IC=1_P_s=1}] are the same as Figure \ref{figure:multispecies_density}. For [\subref{figure:IC=2_P_s=0}-\subref{figure:IC=2_P_s=1}] we initialised the region $1 \leqslant x \leqslant 100$ with agents of type M at a density of 0.9 and the remaining sites with agents of type X also at a density of 0.9. In [\subref{figure:IC=4_P_s=0}-\subref{figure:IC=4_P_s=1}] the lattice was initialised such that all the sites in the region $1 \leqslant x \leqslant 100$ are occupied by agents of type M and all the sites in the region $101 \leqslant x \leqslant 200$ are occupied by agents of type X. We present solutions at $t=0$, $t=100$ and $t=1000$ in all cases. The averaged densities $\overline{M}_i (t)$ and $\overline{X}_i(t)$ are shown in black and the approximate PDE solution trajectories are shown in colour (red for agents of type M and green for agents of type X). The black arrows show the direction of increasing time.}
\label{figure:density_compar}
\end{figure}
In Figure \ref{figure:density_compar}, we compare the column-averaged density of the ABM with the numerical solution of the one-dimensional analogue of equations \eqref{eqn:pde_M} and \eqref{eqn:pde_X} by averaging the PDEs over the $y$ direction\footnote{Note that in Figure \ref{figure:IC=4_P_s=0} in the fully occupied domain the agreement is trivially perfect since volume exclusion prohibits departure from the initial condition.}. In the $\rho=0$ case (Figure \ref{figure:density_compar}, first column), the PDE solutions and the ABM do not agree well as evidenced by the disparity between the two profiles. This discrepancy can also be seen in the
work of \citet{simpson2009mss} where the authors devise an ABM for multi-species exclusion processes (which corresponds to the $\rho=0$ version of our model) and compare their ABM with the corresponding continuum model. We remark that one reason for the discrepancy is that in crowded environments where movement of agents is continuously inhibited by other agents in the neighbourhood, lattice occupancies cannot be considered independent of each other and spatial correlations can arise \parencite{simpson2013emi,markham2013isc,markham2013smi}. Independence of lattice sites is a key assumption that is typically made when deriving the continuum models such as the one we have derived above \parencite{simpson2007sic,simpson2009mss,simpson2010mbc,simpson2015dpg,gavagnin2018sdm,chappelle2019pmc,yates2015ipe}. For non-zero swapping probabilities, we see significantly improved agreement between the deterministic and stochastic profiles (Figure \ref{figure:density_compar}, second and third columns). Swapping helps to retain the independence assumption, hence breaking down the spatial correlations and, in turn, improving the agreement between the macroscopic and individual-level densities. We also see that in the no-swapping case, crowding of the green agents behind the red agents leads to profiles for which the maximum densities over time are higher than the initial maximum density. This behaviour can be observed most clearly in the profiles for type-X agent (shown in green in Figure \ref{figure:IC=1_P_s=0}). This is because for multi-species diffusion systems with cross-diffusion there is no maximum principle for the individual species \parencite{rahman2016dmc,jungel2015bem,simpson2009mss}. We note that the enhanced diffusion engendered by swapping eliminates this effect. However, this does not necessarily mean that a maximum principle now hold for the systems under consideration.
\subsection{Individual level analysis}\label{sec:individual_level_analysis}
In Section \ref{sec:plm}, we saw that in the single-species case with unlabelled agents, swapping has no macroscopic effect on the diffusion of agents. In this section, we analyse the movement of agents at an individual level to assess how swapping impacts the movement of individual agents.
To do this, we initialise a $150$ by $150$ lattice and randomly seed it with agents at different background densities $\mathbf{c}=[0,0.25,0.5,0.75,1]$, where $c=0$ corresponds to one agent with no agents in the background to interact with and $c=1$ corresponds to a fully populated lattice. We let the positions of the agents evolve according to the ABM described in Section \ref{sec:abm} using the movement rate $r_m=1$ for a range of swapping probabilities $\boldsymbol\rho=[0, 0.25, 0.5, 0.75, 1]$. The positions $(X_t,Y_t)$ of agents starting in the region defined by the central square $[51,100]\times [51,100]$, are recorded over a regular time grid to create a `track'. Each track traces the path travelled by an agent over time. We impose periodic boundary conditions on the domain and if a tracked agent hits the boundary, we stop recording its position.
We analyse individual movement of agents using the sum of squared displacement (SSD) \parencite{simpson2015dpg,simpson2007sic},
\begin{equation}
S_t^{(x)}=\sum_{\ell=1}^t (X_\ell - X_{\ell-1})^2,
\end{equation}
and,
\begin{equation}
S_t^{(y)}=\sum_{\ell=1}^t (Y_\ell - Y_{\ell-1})^2.
\end{equation}
By calculating the SSD and its mean over an ensemble of tracks $\overline S_t^{(x)}$ and $\overline S_t^{(y)}$, we wish to establish a relationship between the individual-level diffusion coefficient, $D$, the background density, $c$, and the swapping probability, $\rho$. To this end, we fit a linear model of the form $\hat S_t=at$ to the mean SSD in each orthogonal direction and calculate the diffusion coefficient as,
\begin{equation}
\hat D(c_i,\rho_j)=\frac{a_x+a_y}{4},
\end{equation}
where $a_x=2D_x$ and $a_y=2D_y$ are the gradients of the lines fitted to $\overline S_t^{(x)}$ and $\overline S_t^{(y)}$, respectively.
Next, we analytically derive a relationship between the diffusion coefficient, $D$, swapping probability, $\rho$, and the background density, $c$, in order to compare to our simulation results. Let $P_{ij}(t)=P_{ij}$ denote the probability that a focal agent is at position $(i,j)$ at time $t$. We can write down the probability of being at position $(i,j)$ at time $t+\delta t$ as,
\begin{multline}\label{eqn:prob_master_eqn}
P_{ij}(t+\delta t)=\frac{r_m}{4}P_{i-1,j}[(1-c)+2c\rho]\delta t+\frac{r_m}{4}P_{i+1,j}[(1-c)+2c\rho]\delta t+\frac{r_m}{4}P_{i,j-1}[(1-c)+2c\rho]\delta t\\+\frac{r_m}{4}P_{i,j+1}[(1-c)+2c\rho]\delta t+P_{ij}[(1-r_m\delta t)+r_mc(1-\rho)\delta t-r_m c\rho\delta t].
\end{multline}
To understand this equation, we explain the terms that comprise it. The first term is obtained by considering an agent at site $(i-1,j)$ that has attempted to move into site $(i,j)$ with probability $r_m/4 \hspace{4pt} \delta t$. The agent successfully moves, either if the site $(i,j)$ is empty or via a swap. If the site $(i,j)$ is empty (i.e. $c=0$ in equation \ref{eqn:prob_master_eqn}), the agent successfully moves in. However, if the site is already occupied by another agent (i.e. $c=1$), then the two agents swap with probability $\rho$. The factor of 2 in the term $2\rho c$ corresponds to the two ways in which the position of the agent at site $(i,j)$ can change due to a swap: either the agent at $(i,j)$ initiates a swap with the agent at site $(i-1,j)$ or the agent at site $(i-1,j)$ initiates a swap with the agent at position $(i,j)$. Terms 2, 3 and 4 carry similar interpretations for an agent moving into site $(i,j)$ from sites $(i+1,j),(i,j-1)$ or $(i,j+1)$. The last term in equation \eqref{eqn:prob_master_eqn} gives the transition probabilities for the agent already occupying position $(i,j)$ not moving in $[t,t+\delta t]$, with probability $(1-r_m \delta t)$, or the focal agent attempting to swap with the agent at an occupied neighbouring site but that swapping event being aborted with probability $1-\rho$, or finally a neighbouring agent attempting to jump into site $(i,j)$ and successfully swapping with the focal agent with probability $\rho$.
By rearranging, dividing both sides of equation \eqref{eqn:prob_master_eqn} by $\delta t$ and taking the limit as $\delta t \to 0$ leads to the system of ODEs given by,
\begin{equation}
\frac{\mathrm{d} P_{ij}}{\mathrm{d} t}=\frac{r_m}{4}[(1-c)+2c\rho](P_{i-1,j}+P_{i+1,j}+P_{i,j-1}+P_{i,j+1})-r_m P_{i,j}[(1-c)+2c\rho],
\end{equation}
which describe the time evolution of the probability of finding an agent at position $(i,j)$ at time $t$.
The equations governing the evolution of the $m$th moments of the probability distribution $P_{ij}$ in the orthogonal directions $i$ and $j$ are defined as,
\begin{equation}\label{eqn:mth_moment_i}
\frac{\mathrm{d} \langle i^m \rangle(t)}{\mathrm{d} t} = \frac{\mathrm{d} \langle i^m \rangle}{\delta t} = \frac{\mathrm{d} }{\mathrm{d} t}\left(\sum_{i=1}^{L_x} i^m P_{ij} \right),
\end{equation}
and,
\begin{equation}\label{eqn:mth_moment_j}
\frac{\mathrm{d} \langle j^m \rangle(t)}{\mathrm{d} t} = \frac{\mathrm{d} \langle j^m \rangle}{\mathrm{d} t} = \frac{\mathrm{d} }{\mathrm{d} t}\left(\sum_{j=1}^{L_y} j^m P_{ij} \right).
\end{equation}
It can be shown that for $m=1$, $\langle i \rangle = \langle j \rangle = 0$ as expected for the unbiased process \parencite{codling2008rwm}. Since this gives us no information about the dispersion of the agents, we use the second moment to quantify the statistical fluctuations \parencite{simpson2015dpg}. Using equations \eqref{eqn:mth_moment_i} and \eqref{eqn:mth_moment_j} with $m=2$, it can be shown that,
\begin{equation*}
\frac{\delta \langle i^2\rangle}{\delta t}=\frac{\delta \langle j^2\rangle}{\delta t}=\frac{r_m}{2}[(1-c)+2c\rho].
\end{equation*}
Under the initial condition that $\langle i^2 \rangle(0)=\langle j^2 \rangle(0)=0$, the equations can be solved to give,
\begin{equation}
\langle i^2 \rangle = \langle j^2 \rangle = \frac{r_m}{2}[(1-c)+2c\rho]t.
\end{equation}
The individual-level diffusion coefficient is then given by,
\begin{equation}\label{eqn:D_theory}
D = D_0[(1-c)+2c\rho],
\end{equation}
where $D_0 = r_m/4$ is the diffusion constant. Letting $\rho=0$ we retrieve the diffusion constant for the basic exclusion process. Notice that compared to the simple exclusion process there is an extra term $2c\rho$ which accounts for swapping with probability $\rho$ when $c > 0$. The factor of 2 here is due to the two ways in which a swap can change the occupancy of site $(i,j)$: either the focal agent initiates the swap with the cell at the neighbouring site, or the agent at a neighbouring site initiates a swap with the focal agent.
\begin{figure}[t!]
\begin{center}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/heatmap_simul_2-eps-converted-to.pdf}
\label{figure:heatmap_simul}
}
\subfigure[]{
\includegraphics[width=0.475\textwidth]{figures/plots/heatmap_theory_2-eps-converted-to.pdf}
\label{figure:heatmap_theory}
}
\subfigure[]{
\includegraphics[width=0.5\textwidth]{figures/plots/D_vs_c_vs_rho-eps-converted-to.pdf}
\label{figure:D_vs_c_vs_rho}
}
\end{center}
\caption{Heat map showing the relationship between the cell-level diffusion coefficient, $D$, the domain density, $c$, and the swapping probability, $\rho$. To obtain the simulated relationship \subref{figure:heatmap_simul}, we initialised a 150 by 150 periodic domain with cell densities $\mathbf{c}=[0,0.25,0.5.0.75,1]$ and for the values of swapping probability $\boldsymbol \rho = [0,0.25,0.5,0.75,1]$ we let the positions of the agents evolve according to the ABM in Section \ref{sec:abm} with $r_m=1$. We tracked all the agents inside the region defined by the square $[51,100]\times[51,100]$. We stopped tracking a cell after it hit a boundary. Theoretical heatmap in \subref{figure:heatmap_theory} was obtained using equation \eqref{eqn:D_theory}. In \subref{figure:D_vs_c_vs_rho} we show the linear relationship between the diffusion coefficient, $D$, the domain density, $c$ and the swapping probability, $\rho$. The coloured circles are values from simulated tracks whereas the dotted black lines are given by equation \eqref{eqn:D_theory}.}
\label{figure:heatmaps}
\end{figure}
In Figure \ref{figure:heatmaps}, we compare the simulated approximation to the diffusion coefficient, $\hat D$ (shown in \subref{figure:heatmap_simul}) with the derived expression for $D$ (shown in \subref{figure:heatmap_theory}) as a pair of heat maps for the range of values $c$ and $\rho$ defined earlier. We also show a line plot that highlights the linear relationship between $D$ and $c$ for different value of $\rho$ (shown in \subref{figure:D_vs_c_vs_rho}).
We can see that there is excellent agreement between the analytical expression and the simulated results. We note that in the case of zero background density ($c=0$) $D$ is always 0.25 regardless of the swapping probability: for a single agent with no agents to interact with swapping can have no effect. We also note that for swapping probability $\rho=0.5$, the value of the diffusion coefficient is always 0.5 irrespective of the density: half the time a focal agent attempts to move into an occupied site, that movement will be rejected. If this were the only impact of swapping, we would expect the diffusion coefficient of a focal agent to be reduced. However, just as often as a focal agent attempts to move into an occupied neighbour's position, an occupied neighbour tries to move into the focal agent's position -- achieving this successfully with probability $\rho=0.5$. Assuming a well-mixed scenario, this exactly compensates for the number of aborted moves the focal agent makes, meaning diffusion is as if the focal agent were on an unoccupied domain ($D=0.25$), irrespective of density.
For $c=1$ and $\rho=1$ we note that $D=0.5$ (i.e. twice as large as for an agent moving on an unoccupied domain) as every attempted move by the focal agent is executed successfully and the focal agent is also moved equally often by neighbouring cells swapping into its position.
\section{Illustrative examples}\label{sec:examples}
In this section, we show examples of the situations in which swapping has important applications. In Section \ref{sec:example_prolif}, we build the swapping mechanism into a cell migration model with proliferation and in Section \ref{sec:example_adhesion}, we show how the swapping mechanism in conjunction with cell-cell adhesion can facilitate spontaneous pattern formation in densely crowded environments.
\subsection{Swapping model with cell proliferation}\label{sec:example_prolif}
We look at the role of swapping in cell migration with proliferation. For this example, we concern ourselves with the two-species cell migration model. The movement kinetics of the agents are the same as the swapping model described in Section \ref{sec:abm} but in addition to migrating, agents can attempt to proliferate, placing a daughter at a randomly chosen neighbouring site if the site is empty, otherwise the division event is aborted. The proliferation rates per unit time for the two species M and X are denoted by $r_{p}^M$ and $r_{p}^X$, respectively.
We initialise the domain with $L_x=100$ sites in the horizontal direction and $L_y=20$ sites in the vertical direction. We fill all the sites in the range $41 \leqslant x \leqslant 60$ with agents of type M and all the remaining sites with type-X agents at a density of 0.5. The movement rates of agents are set to $r_{m}^M=r_{m}^X=1$ and the proliferation rates as $r_{p}^M=0.01$ and $r_{p}^X=0$, i.e. only the agents of type M divide and the number of type-X agents are held constant. We let the system evolve according to the specified ABM.
In Figure \ref{figure:prolif_snapshots} we provide snapshots of the evolving lattice occupancy for $\rho=0, 0.5, 1$ (columns 1, 2 and 3, respectively) at $t=0, 500, 1000$ (rows 1, 2 and 3, respectively). We see, at the same time points, that cells are unsurprisingly more well-mixed in the case of non-zero swapping (second and third columns in Figure \ref{figure:prolif_snapshots}) compared to the zero-swapping situation (first column in Figure \ref{figure:prolif_snapshots}). We also see faster colonisation of the domain overall in the non-zero swapping cases than without swapping. This is because swapping allows the proliferating red agents to disperse more quickly into less dense regions, which in turn increases the probability of a successful division events for these agents. Without swapping, it takes longer for proliferative red agents to find the space to proliferate into. This trend of decreasing colonisation time with increasing swapping probability is reinforced in Figure \ref{fig:time_to_reach_K} where we see that the time to reach the domain's carrying capacity is a decreasing function of the swapping probability, $\rho$.
\begin{figure}[t!]
\begin{center}
\subfigure[]{
\includegraphics[width=0.30\textwidth]{figures/plots/prolif_agents_imagesc_rho=0_T=0-eps-converted-to.pdf}
\label{figure:prolif_agents_Ps=0_T=0}
}
\setcounter{subfigure}{3}
\subfigure[]{
\includegraphics[width=0.30\textwidth]{figures/plots/prolif_agents_imagesc_rho=0.5_T=0-eps-converted-to.pdf}
\label{figure:prolif_agents_Ps=0.5_T=0}
}
\setcounter{subfigure}{6}
\subfigure[]{
\includegraphics[width=0.30\textwidth]{figures/plots/prolif_agents_imagesc_rho=1_T=0-eps-converted-to.pdf}
\label{figure:prolif_agents_Ps=1_T=0}
}
\setcounter{subfigure}{1}
\subfigure[]{
\includegraphics[width=0.30\textwidth]{figures/plots/prolif_agents_imagesc_rho=0_T=500-eps-converted-to.pdf}
\label{figure:prolif_agents_Ps=0_T=500}
}
\setcounter{subfigure}{4}
\subfigure[]{
\includegraphics[width=0.30\textwidth]{figures/plots/prolif_agents_imagesc_rho=0.5_T=500-eps-converted-to.pdf}
\label{figure:prolif_agents_Ps=0.5_T=500}
}
\setcounter{subfigure}{7}
\subfigure[]{
\includegraphics[width=0.30\textwidth]{figures/plots/prolif_agents_imagesc_rho=1_T=500-eps-converted-to.pdf}
\label{figure:prolif_agents_Ps=1_T=500}
}
\setcounter{subfigure}{2}
\subfigure[]{
\includegraphics[width=0.30\textwidth]{figures/plots/prolif_agents_imagesc_rho=0_T=1000-eps-converted-to.pdf}
\label{figure:prolif_agents_Ps=0_T=1000}
}
\setcounter{subfigure}{5}
\subfigure[]{
\includegraphics[width=0.30\textwidth]{figures/plots/prolif_agents_imagesc_rho=0.5_T=1000-eps-converted-to.pdf}
\label{figure:prolif_agents_Ps=0.5_T=1000}
}
\setcounter{subfigure}{8}
\subfigure[]{
\includegraphics[width=0.30\textwidth]{figures/plots/prolif_agents_imagesc_rho=1_T=1000-eps-converted-to.pdf}
\label{figure:prolif_agents_Ps=1_T=1000}
}
\end{center}
\caption{Snapshots of the lattice occupancy with swapping probability $\rho=0$ in [\subref{figure:prolif_agents_Ps=0_T=0}-\subref{figure:prolif_agents_Ps=0_T=1000}], $\rho=0.5$ in [\subref{figure:prolif_agents_Ps=0.5_T=0}-\subref{figure:prolif_agents_Ps=0.5_T=1000}] and $\rho=1$ in [\subref{figure:prolif_agents_Ps=1_T=0}-\subref{figure:prolif_agents_Ps=1_T=1000}] at $t=0,500,1000$ for the cell migration process with swapping and proliferation. We initialise the domain as a $20$ by $100$ lattice where all the sites in the horizontal range $41$ to $60$ are occupied by the agents of type M (red) and the remaining sites are inhabited by agents of type X at a density of $0.5$. Both species diffuse at equal rates $r_{m}^M=r_{m}^X=1$. Rates of proliferation are given by $r_{p}^M=0.01$ and $r_{p}^X=0$.}
\label{figure:prolif_snapshots}
\end{figure}
\begin{figure}[t!]
\centering
\includegraphics[width=0.6\textwidth]{figures/plots/time_to_reach_K-eps-converted-to.pdf}
\caption{Time to reach the carrying capacity. The red circles show the mean time for the number of agents to reach the carrying capacity of the domain for a range of swapping probabilities (averaged over 100 repeats). The blue error bars represent the mean plus or minus one standard deviation.}
\label{fig:time_to_reach_K}
\end{figure}
\begin{figure}[t!]
\begin{center}
\subfigure[]{
\includegraphics[width=0.30\textwidth]{figures/plots/prolif_agents_density_rho=0_T=0-eps-converted-to.pdf}
\label{figure:prolif_agents_density_rho=0_T=0}
}
\setcounter{subfigure}{3}
\subfigure[]{
\includegraphics[width=0.30\textwidth]{figures/plots/prolif_agents_density_rho=0.5_T=0-eps-converted-to.pdf}
\label{figure:prolif_agents_density_rho=0.5_T=0}
}
\setcounter{subfigure}{6}
\subfigure[]{
\includegraphics[width=0.30\textwidth]{figures/plots/prolif_agents_density_rho=1_T=0-eps-converted-to.pdf}
\label{figure:prolif_agents_density_rho=1_T=0}
}
\setcounter{subfigure}{1}
\subfigure[]{
\includegraphics[width=0.30\textwidth]{figures/plots/prolif_agents_density_rho=0_T=500-eps-converted-to.pdf}
\label{figure:prolif_agents_density_rho=0_T=500}
}
\setcounter{subfigure}{4}
\subfigure[]{
\includegraphics[width=0.31\textwidth]{figures/plots/prolif_agents_density_rho=0.5_T=500-eps-converted-to.pdf}
\label{figure:prolif_agents_density_rho=0.5_T=500}
}
\setcounter{subfigure}{7}
\subfigure[]{
\includegraphics[width=0.30\textwidth]{figures/plots/prolif_agents_density_rho=1_T=500-eps-converted-to.pdf}
\label{figure:prolif_agents_density_rho=1_T=500}
}
\setcounter{subfigure}{2}
\subfigure[]{
\includegraphics[width=0.30\textwidth]{figures/plots/prolif_agents_density_rho=0_T=1000-eps-converted-to.pdf}
\label{figure:prolif_agents_density_rho=0_T=1000}
}
\setcounter{subfigure}{5}
\subfigure[]{
\includegraphics[width=0.30\textwidth]{figures/plots/prolif_agents_density_rho=0.5_T=1000-eps-converted-to.pdf}
\label{figure:prolif_agents_density_rho=0.5_T=1000}
}
\setcounter{subfigure}{8}
\subfigure[]{
\includegraphics[width=0.30\textwidth]{figures/plots/prolif_agents_density_rho=1_T=1000-eps-converted-to.pdf}
\label{figure:prolif_agents_density_rho=1_T=1000}
}
\end{center}
\caption{Density profiles for cell migration process with swapping and proliferation. All parameters and initial conditions are the same as Figure \ref{figure:prolif_snapshots}. Here, we present the column densities at $t=0,500,1000$ for $\rho=0$ in [\subref{figure:prolif_agents_density_rho=0_T=0}-\subref{figure:prolif_agents_density_rho=0_T=1000}], $\rho=0.5$ in [\subref{figure:prolif_agents_density_rho=0.5_T=0}-\subref{figure:prolif_agents_density_rho=0.5_T=1000}] and $\rho=1$ in [\subref{figure:prolif_agents_density_rho=1_T=0}-\subref{figure:prolif_agents_density_rho=1_T=1000}] averaged over 100 repeats of the ABM described above. We also plot the corresponding mean-field PDE solutions with $r_p^M=0.01$ in red for species $M$ and in green for species $X$.}
\label{figure:prolif_density_plots}
\end{figure}
The mean-field PDEs describing the approximate population-level dynamics of the agent are given by,
\begin{align}
\D M t &= \nabla \cdot [D_1(X) \nabla M + D_2(M)\nabla X]+ r_p^M M\left(1-(M+X)\right), \label{eqn:prolif_PDE_M}\\
\D X t &= \nabla \cdot [D_3(M)\nabla X + D_4(X)\nabla M] \label{eqn:prolif_PDE_X},
\end{align}
where $D_1, D_2, D_3$ and $D_4$ are as defined previously. Notice that proliferation of type-M agents gives rise to an additive source term in equation \eqref{eqn:prolif_PDE_M}. The derivation of this source is standard and can be found in \citet{plank2012mcc,simpson2010cip}, for example.
In Figure \ref{figure:prolif_density_plots} we compare the average column density of the ABM with the numerical solution of the one-dimensional analogue of the mean-field PDEs obtained by averaging over the $y$ direction with $r_p^M=0.01$. We chose this value for the proliferation rate as we wanted to keep the ratio $r_m^M/r_p^M \ll 1$ \parencite{fadai2019aae,simpson2010cip}. There are two reasons for this: firstly, a modelling choice to prevent agents clustering into proliferation-induced patches and secondly, it is biological realism that given the parameters of the model we could expect real biological cells will attempt proliferation events less frequently than movement events. We see good agreement between the two profiles for non-zero swapping probability. However, when the swapping probability is set to 0, we see discrepancies arising that are amplified as time increases. Recall that the disparity between the PDE and ABM profiles can be also observed in Figure \ref{figure:multispecies_density} due to spatial correlations that are not accounted for by the mean-field PDEs. The addition of proliferation into the model increases the spatial correlations between site occupancies, leading to greater disparity. One way to partially rectify this problem is by modelling the higher order moments in the PDE description \parencite{markham2013isc,markham2013smi,simpson2013emi} however for the purposes of this work, we simply note that allowing swapping breaks up the correlations more effectively than in its absence leading to better agreement between the ABM and the population-level densities.
\subsection{Swapping model with cell-cell adhesion}\label{sec:example_adhesion}
Another interesting application of swapping is the formation of patterns in densely crowded environments. In this section, we use a cell-cell adhesion model with swapping to investigate how biologically plausible patterns can form starting from a randomly seeded domain. Our model is based on the similar cell-cell adhesion model studied previously by \parencite{simpson2010mbc,charteris2014mca,khain2007rcc} who consider adhesion between identical agents. Here, we extend the model to incorporate two types of agents with swapping to facilitate the movement events.
For the purpose of this paper, we assume adhesion between two species, M and X, on a fully populated domain. For a simple exclusion-based ABM the agents on the fully populated domain would not successfully move at all. This is since the exclusion principle forbids cells from occupying the already occupied lattice sites. In our model, the movement of the agents and the formation of patterns will be facilitated by the swapping mechanism. In an on-lattice adhesion model, agents can adhere to other agents in their neighbourhood, making them less likely to successfully complete the movement event. As well as the number of agents in the neighbourhood, the strength of adhesion determines how likely an agent is to successfully move. In a simple model with species M and X, $0\leqslant p\leqslant 1$ characterises the strength of adhesion between two type-M agents and $0\leqslant q\leqslant 1$ the strength of adhesion between two type-X agents.
We assume cell movement in a densely crowded domain where the underlying lattice is fully populated with type-M and X agents, their positions chosen uniformly at random. When an agent is chosen to move into an occupied neighbouring site, we check the feasibility of swapping by sampling a random number $u$ from the standard uniform distribution and comparing it with the swapping probability, $\rho$. A swapping move breaks existing interactions between the two swapping agents and their neighbours and makes new connections following a successful swap. Since the movement of the focal agent and the target agent depends on their respective neighbours, the success of a swapping move depends on whether the respective neighbouring sites are occupied by type-M or type-X agents. Let the agent chosen to move be the agent at site $(i,j)$ and let $Z_{ij}=\{(i-1,j),(i+1,j),(i,j-1),(i,j+1) \}$ denote the set containing the positions of sites in the neighbourhood of the focal agent on the two-dimensional lattice. If the focal agent is a type-M agent then the probability of its successful movement into an empty site is given by,
$$p_\text{move}^\text{agent}=(1-p)^{\mathrm{\Sigma}_{z\in Z} M_z}.$$ However, if the focal agent is a type-X agent then, $$p_\text{move}^\text{agent}=(1-q)^{\mathrm{\Sigma}_{z\in Z} X_z}.$$ Here, $M_z$ are binary taking a value of unity if the site with position $z\in Z_{ij}$ is occupied by a type-M agent or 0 otherwise. Therefore, $\sum_{z\in Z} M_z$ is the sum of occupancies of the sites in $Z_{ij}$ that are occupied by type-M agents. Likewise, $\sum_{z\in Z} X_z$ is the sum of occupancies of the sites in $Z_{ij}$ that are occupied by type-X agents. Similarly, the probability of movement of a type-M agent occupying the target site is given by, $$p_\text{move}^\text{targ}=(1-p)^{\mathrm{\Sigma}_{y\in Y_z} M_y}.$$ However, if the agent occupying the target site is a type-X agent then, $$p_\text{move}^\text{targ}=(1-q)^{\mathrm{\Sigma}_{y \in Y_z} X_y}.$$ Here, the set $Y_z$ contains the positions of sites in the neighbourhood of the target site, $z$. Therefore, $\sum_{y\in Y_z} M_y$ is the sum of occupancies of all the sites in $Y_z$ that are occupied by type-M agents and $\sum_{y\in Y_z} X_y$ denotes the sum of occupancies of all the sites in $Y_z$ that are occupied by type-X agents.
The overall probability of a successful swap given a movement event has been attempted is therefore a product of the swapping probability, the probability of the focal agent breaking links with its neighbours in order to move out and the probability of the target breaking links with its neighbours in order to move in, i.e.,
\begin{equation*}
p_\text{swap} = \rho p_\text{move}^\text{agent} p_\text{move}^\text{targ}.
\end{equation*}
Here we are considering that the link between the two swapping agents has to be broken in order for the swap to take place. As an example, consider the case where the focal agent at site $(i,j)$ attempts to jump into the neighbouring site $(i+1,j)$ (Figure \ref{figure:adh_lattice_all}). The focal agent is a type-M agent, shown in red and labelled as 1. The agent at the target site $(i+1,j)$ is a type-X agent, shown in green and labelled as 2. Since the focal agent is a red agent, it is adhesive to other red agents in its von-Neumann neighbourhood. There is only one site (with position $(i,j-1)$) neighbouring agent 1 that is occupied by a red agent. Therefore, $\sum_{z\in Z_{ij}} M_z = 1$. Since the agent at the target site is a green agent, it is adhesive to the other green agents in its von Neumann neighbourhood. The number of green agents neighbouring the target site agent is also 1 situated at the site with position $(i+2,j)$. Hence, $\sum_{y\in Y_{i+2,j}} X_y = 1$. For an arbitrary $\rho$, $p$ and $q$, the probability of swap in this situation can be written as $p_\text{swap}=\rho(1-p)(1-q)$.
\begin{figure}[t!]
\begin{center}
\subfigure[]{
\includegraphics[width=0.3\textwidth]{figures/schematics/adh_lattice_1.pdf}
\label{figure:adh_lattice_1}
}
\subfigure[]{
\includegraphics[width=0.3\textwidth]{figures/schematics/adh_lattice_2.pdf}
\label{figure:adh_lattice_2}
}
\subfigure[]{
\includegraphics[width=0.3\textwidth]{figures/schematics/adh_lattice_3.pdf}
\label{figure:adh_lattice_3}
}
\end{center}
\caption{A schematic illustrating swapping in the cell-cell adhesion model. A red site represents a type-M agent and a green site represent a type-X agent. Agent 1 at site $(i,j)$ attempts to swap with agent 2 at site $(i+1,j)$. We have coloured in black the neighbouring sites that are unimportant in this context (i.e. occupancy of these sites does not affect the probability of a successful swap happening between agent 1 and agent 2). The initial configuration of the lattice is shown in \subref{figure:adh_lattice_1}. Agent 1 attempts to swap with agent 2 \subref{figure:adh_lattice_2} and the state of the lattice after the successful swap is shown in \subref{figure:adh_lattice_3}.}
\label{figure:adh_lattice_all}
\end{figure}
\begin{figure}[h!]
\begin{center}
\subfigure[]{
\includegraphics[width=0.31\textwidth]{figures/plots/initial_condition_pattern-eps-converted-to.pdf}
\label{figure:pattern_t=0}
}
\subfigure[]{
\includegraphics[width=0.31\textwidth]{figures/plots/pattern_BCs=periodic_rm=1_rx=1_p=0.9_q=0.7_r=0_Ps=1_T=15000-eps-converted-to.pdf}
\label{figure:pattern_b}
}
\subfigure[]{
\includegraphics[width=0.31\textwidth]{figures/plots/pattern_BCs=periodic_rm=1_rx=1_p=0.98_q=0.98_r=0_Ps=1_T=15000-eps-converted-to.pdf}
\label{figure:pattern_c}
}
\end{center}
\caption{Pattern formation in a crowded environment. The domain is initialised as a square lattice with $L=20$ sites in the horizontal and vertical directions, respectively. Initially, the domain is fully populated by type-M and type-X agents at a density of 0.5 each, where the positions of the agents are assigned uniformly at random (as shown in \subref{figure:pattern_t=0}). We present the state of the lattice at $t=15,000$ for two different sets of adhesion strengths: $p=0.90$ and $q=0.70$ \subref{figure:pattern_b} and $p=q=0.98$ \subref{figure:pattern_c}. In both cases, we let $\rho=1$ and the movement rates of the two species $r_{m}^M=r_{m}^X=1$.}
\label{figure:patterns}
\end{figure}
In Figure \ref{figure:patterns},we present some results that demonstrate the importance and impact of swapping in a model of adhesion-mediated pattern formation. For this, we consider a square lattice with $L=20$ sites in both the horizontal and vertical direction. As before, one compartment can accommodate no more than a single agent at a time. We impose periodic boundary conditions. We seed the lattice with type-M and type-X agents at a density of 0.5 each, where the initial positions of the agents on the lattice are assigned uniformly at random (Figure \ref{figure:pattern_t=0}). We let positions of the agents evolve according to the kinetics described above using the movement rate $r_{m}^M = r_{m}^X=1$ and $\rho=1$.
Snapshots of the evolving lattice occupancy at $t=15,000$ for two different sets of $p$ and $q$ values are shown in Figure \ref{figure:pattern_b} and \subref{figure:pattern_c}. We can see self-organisation of agents into clusters of like type agents. The characteristic size of the aggregates is sensitive to the magnitude of the adhesion strengths. The clusters are bigger where self-adhesion within one species is stronger than within the other ($p=0.9$ and $q=0.7$) whereas we see more labyrinthine patterns when both the adhesion strengths are very strong and equal ($p=q=0.98$). It is evident from these figures that in densely crowded domains, swapping plays a vital role in allowing agents to organise themselves to form patterns. Without swapping, no agent movement and hence no pattern formation would be possible.
\section{Summary, conclusion and further work}\label{sec:discussion}
In many contexts, cell movement is modelled as a volume exclusion process. In this paper, motivated by real-life examples, we devised an ABM that allows for swapping to take place between pairs of neighbouring cells. Our model maintains the important carrying capacity component of volume exclusion models, but allows the flexibility of movement observed even amongst some densely packed cell configurations. We considered single-species and two-species cases. In both situations, we derived the corresponding PLM and compared it with the column-averaged densities of the ABM. In the single-species case, we found that swapping has no macroscopic effect on the agents' macroscopic dispersal compared to the exclusion process: two identical agents exchanging positions is identical, from a macroscopic point of view, to the impact of an aborted movement event on identical agents. To understand how swapping affects agent movement at an individual level, we analysed simulated agent tracks to determine the dependency of the individual level diffusion coefficient on the background domain density and the swapping probability. We found that swapping enhanced the movement of agents in all cases compared to the volumed exclusion model. Using the probability master equation, we were able to analytically derive an expression for the diffusion coefficient that confirms the relationship obtained via the simulated tracks.
In the two-species case, we found that swapping enhances the movement at the macroscopic scale by allowing agents to mix more compared to the two-species pure volume exclusion model. We derived the corresponding population-level model and found excellent agreement between the ABM and the macroscopic model in all cases with a sufficiently large swapping probability. We note that compared to the two-species volume exclusion model, the agreement improved due to the increased mixing of agents caused by swapping at the interface between the two species. This retains the assumption of independence of lattice occupancy on which the derivation of the PDEs relies. Models that relax the necessary independence assumption have been extensively studied by \citet{markham2013isc,markham2013smi,simpson2011cmm,simpson2013emi} who identified the importance of spatial correlations arising from the lack of independence between of lattice sites in cell migration.
In Section \ref{sec:examples}, we demonstrated the importance of swapping via a series of examples. In the first application, we considered a cell migration model with proliferation. We found that swapping accelerates the proliferation process by allowing the agents to disperse more and breaking spatial correlations. We found that the time to reach the domain's carrying capacity varies inversely with the swapping probability. Deriving the PLM and comparing it to the average density of the ABM, we showed that there is a good agreement between the two profiles for sufficiently large swapping probability. For the $\rho=0$ case, we note discrepancies that are caused by the build up of spatial correlations between sites.
For the second example, by incorporating swapping into a cell-cell adhesion model, we showed that agents can spontaneously rearrange themselves into clusters to form patterns even on densely populated domains. Here, swapping is vital for the formation of these patterns. Without swapping, the agents would not be able to move at all and the patterns would never be achieved. Even on a more sparsely populated domain, swapping would greatly increase the speed at which the patterns can form.
In conclusion, motivating our study by real-life examples, we have developed a cell migration model that incorporates swapping as a viable movement process. As well as adding biological realism, our model has the added benefit of better agreement between the corresponding continuum description and the ABM compared to the classical volume exclusion process. We also saw that the swapping ABM with cell-cell adhesion, when applied to cells in densely crowded environments, leads to pattern formation. We once again stress that the patterns would be unattainable under the traditional volume exclusion model. The results in this paper hint that swapping is an important and overlooked mechanism in the context of modelling real biological scenarios and merits further explorations in conjunction with experimental data.
\section*{Acknowledgements}
Shahzeb Raja Noureen is supported by a scholarship from the EPSRC Centre for Doctoral Training
in Statistical Applied Mathematics at Bath (SAMBa), under the project EP/S022945/1. This research made use of the Balena High Performance Computing (HPC) Service at the University of Bath. The images in Figure \ref{fig:cell_swapping_1} were taken by Dr. Matthew Ford (Centre for Research in Reproduction and Development, McGill, Canada). Richard Mort was supported by North West Cancer Research (NWCR Grant CR1132) and the NC3Rs (NC3Rs grant NC/T002328/1). We would like to thank the members of Christian Yates’ mathematical biology reading group for constructive and helpful comments on a preprint of this paper. Finally, we would like to thank Fraser Waters for his help in managing some of the code related to this project.
\printbibliography
\end{document}
\section{Appendices}
\subsection{Derivation of macroscopic model for cell migration with proliferation}
We derive the macroscopic description of the cell migration model described in section \ref{sec:example_prolif} given by equation \eqref{eqn:prolif_PDE_M} and \eqref{eqn:prolif_PDE_X}. Since we are considering proliferation of onl
|
\section{Introduction}
Stochastic gradient descent (SGD) is the backbone of optimization for neural networks, going back at least as far as \citet{lecun_gradient-based_1998}, and SGD is the de-facto tool for optimizing the parameters of modern neural networks \citep{krizhevsky_imagenet_2012,he_deep_2015,brown_language_2020}. A central reason for the success of stochastic gradient descent is its efficiency in the face of large datasets -- a noisy estimate of the loss function gradient is generally sufficient to improve the parameters of a neural network and can be computed much faster than a full gradient over the entire training set.
At the same time, folk wisdom dictates that small-batch SGD is not only faster but also has a unique bias towards good loss function minima that cannot be replicated with full batch gradient descent.
Some even believe that stochastic sampling is the fundamental force behind the success of neural networks. These popular beliefs are linked to various properties of SGD, such as its gradient noise, fast escape from saddle points, and its uncanny ability to avoid sub-optimal local minima \citep{hendrik_machine_2017, lecun_yann_2018}. It is common to under-saturate compute capabilities and retain small batch sizes, even if enough compute is available to reap these apparent benefits.
These properties are also attributed in varying degrees to all mini-batched first-order optimizers, such as Adam \citep{kingma_adam:_2015} and others \citep{schmidt_descending_2020}.
But why does stochastic mini-batching really aid generalization? In this work, we set out to isolate mechanisms which underlie the benefits of SGD and use these mechanisms to replicate the empirical benefits of SGD without stochasticity.
In this way, we provide a counterexample to the hypothesis that stochastic mini-batching, which leads to noisy estimates of the gradient of the loss function, is fundamental for the strong generalization success of over-parameterized neural networks.
We show that a standard ResNet-18 can be trained with batch size 50K (the entire training dataset) and still achieve $95.68\% (\pm 0.09)$ validation accuracy on CIFAR-10, which is comparable to the same network trained with a strong SGD baseline, provided data augmentation is used for both methods (see \cref{fig:loss_visualization}). We then extend these findings to train without (random) data augmentations, for an {\em entirely} non-stochastic full-batch training routine with exact computation of the full loss gradient, while still achieving over 95\% accuracy.
Because existing training routines are heavily optimized for small-batch SGD, the success of our experiments requires us to eschew standard training parameters in favor of more training steps, aggressive gradient clipping, and explicit regularization terms.
The existence of this example raises questions about the role of stochastic mini-batching, and by extension gradient noise, in generalization. In particular, it shows that the practical effects of such gradient noise can be captured by explicit, non-stochastic, regularization.
This shows that deep learning succeeds even in the absence of mini-batched training.
A number of authors have studied relatively large batch training, often finding trade-offs between batch size and model performance \citep{yamazaki_yet_2019,mikami_massively_2019, you_limit_2020}. However, the goal of these studies has been first and foremost to accelerate training speed \citep{goyal_accurate_2018,jia_highly_2018}, with maintaining accuracy as a secondary goal. In this study, we seek to achieve high performance on full-batch training at all costs. Our focus is not on fast runtimes or ultra-efficient parallelism, but rather on the implications of our experiments for deep learning theory. In fact, the extremely high cost of each full-batch update makes GD far less efficient than a conventional SGD training loop.
We begin our discussion in \cref{sec:perspectives} by reviewing the literature on SGD and describing various studies that have sought to explain various successes of deep learning through the lens of stochastic sampling. Then, in \cref{sec:exp_data_aug} and \cref{sec:noaug}, we explain the hyper-parameters needed to achieve strong results in the full-batch setting and present benchmark results using a range of settings, both with and without data augmentation.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figures/figure_1.pdf}
\caption{One-dimensional loss landscapes (random direction) of models trained with gradient descent. Default full-batch gradient descent (left) produces sharp models that neither train nor generalize well, yet it can be modified to converge to flatter minima with longer training, gradient clipping and appropriate regularization (right).}
\label{fig:loss_visualization}
\end{figure}
\section{Perspectives on Generalization via SGD}\label{sec:perspectives}
The widespread success of SGD in practical neural network implementations has inspired theorists to investigate the gradient noise created by stochastic sampling as a potential source of observed generalization phenomena in neural networks.
This section will cover some of the recent literature concerning hypothesized effects of stochastic mini-batch gradient descent (SGD). We explicitly focus on generalization effects of SGD in this work. Other possible sources of generalization for neural networks have been proposed that do not lean on stochastic sampling, for example generalization results that only require overparametrization \citep{neyshabur_towards_2018,advani_high-dimensional_2020}, large width \citep{golubeva_are_2021}, and well-behaved initialization schemes \citep{wu_towards_2017, mehta_extreme_2020}.
We will not discuss these here. Furthermore, because we wish to isolate the effect of stochastic sampling in our experiments, we fix an architecture and network hyperparameters in our studies, acknowledging that they were likely chosen because of their synergy with SGD.
\paragraph{Notation.} We denote the optimization objective for training a neural network by $\mathcal{L}(x, \theta)$, where $\theta$ represents network parameters, and $x$ is a single data sample. Over a dataset $X$ of $N$ data points, $\lbrace x_i \rbrace_{i=1}^N$, the neural network training problem is the minimization of
\begin{equation}
L(\theta) = \frac{1}{N} \sum_{x \in X} \mathcal{L}(x, \theta).
\end{equation}
This objective can be optimized via first-order optimization, of which the simplest form is descent in the direction of the negative gradient with respect to parameters $\theta$ on a batch $B$ of data points and with step size $\tau_k$:
\begin{equation}\label{eq:sgd}
\theta^{k+1} = \theta^k - \tau_k \frac{1}{|B|} \sum_{x \in B} \nabla \mathcal{L}(x, \theta^k).
\end{equation}
Now, full-batch gradient descent corresponds to descent on the full dataset $B = X$, stochastic gradient descent corresponds to sampling a single random data point $B = \lbrace x\rbrace \sim X$ (with or without replacement), and mini-batch stochastic gradient descent corresponds to sampling $S$ data points $B= \lbrace x_j \rbrace_{j=1}^S, x_j \sim X$ at once. When sampling without replacement, the set is commonly reset after all elements are depleted.
The update equation \cref{eq:sgd} is often analyzed as an update of the full-batch gradient that is contaminated by gradient noise arising from the stochastic mini-batch sampling, in this setting gradient noise is defined via:
\begin{equation}\label{eq:sgd_noise}
\theta^{k+1} = \theta^k - \tau_k \underbrace{\nabla L(\theta^k)}_{\text{full loss gradient}} + \tau_k \left(\underbrace{\frac{1}{|X|} \sum_{x \in X} \nabla \mathcal{L}(x, \theta^k) - \frac{1}{|B|} \sum_{x \in B} \nabla \mathcal{L}(x, \theta^k)}_{\text{gradient noise $g_k$}}\right).
\end{equation}
Although stochastic gradient descent has been used intermittently in applications of pattern recognition as far back as the 90's, its advantages were debated as late as \citet{wilson_general_2003}, who in support of SGD discuss its efficiency benefits (which would become much more prominent in the following years due to increasing dataset sizes), in addition to earlier ideas that stochastic training can escape from local minima, and its relationship to Brownian motion and ``quasi-annealing'', both of which are also discussed in practical guides such as \citet{lecun_efficient_1998}.
\paragraph{SGD and critical points.}
While early results from an optimization perspective were concerned with showing the effectiveness and convergence properties of SGD \citep{bottou_large-scale_2010}, later ideas focused on the generalization of stochastic training via navigating the optimization landscape, finding global minima, and avoiding bad local minima and saddlepoints. \citet{ge_escaping_2015} show that stochastic descent is advantageous compared to full-batch gradient descent (GD) in its ability to escape saddle points. Although the same conditions actually also allow vanilla gradient descent to avoid saddle-points \citep{lee_gradient_2016}, full-batch descent is slowed down significantly by the existence of saddle points compared to stochastically perturbed variants \citep{du_gradient_2017-1}. Random perturbations also appear necessary to facilitate escape from saddle points in \citet{jin_nonconvex_2019}. It is also noted by some authors that higher-order optimization, which can alleviate these issues, does perform better in the large-batch regimes \citep{martens_optimizing_2020,yadav_making_2020,anil_scalable_2021}.
Related works further study a \textit{critical mini-batch size} \citep{ma_power_2018,jain_parallelizing_2018} after which SGD behaves similarly to full-batch gradient descent (GD) and converges slowly. The idea of a critical batch size is echoed for noisy quadratic models in \citet{zhang_which_2019}, and an empirical measure of critical batch size is proposed in \citet{mccandlish_empirical_2018}.
There are also hypotheses \citep{haochen_shape_2020} that GD necessarily overfits at sub-optimal minima as it trains in the linearized neural tangent kernel regime of \citet{jacot_neural_2018-2,arora_fine-grained_2019-1}. Additional regularization, for example via clever weight decay scheduling as in \citet{xie_understanding_2021} improves training in large(r)-batch settings, possibly by alleviating such overfitting effects.
Overall, the convergence behavior of SGD has to be understood in tandem with the setting of overparametrized neural networks, as theoretical analysis based on convex optimization theory \citep{dauber_can_2020,bassily_stability_2020,amir_never_2021,amir_sgd_2021} is able to show the existence and construction of convex loss functions on which SGD converges to optimal generalization error orders of magnitude fast than full-batch GD.
It is hence unclear though whether the analysis of sub-optimal critical points can explain the benefits of SGD, given that modern neural networks can generally be trained to reach global minima even with deterministic algorithms (for wide enough networks see \citep{du_gradient_2019}). The phenomenon is itself puzzling as sub-optimal local minima do exist and can be found by specialized optimization techniques \citep{yun_critical_2018,goldblum_truth_2020}, but they are not found by first-order descent methods with standard initialization. It has been postulated that ``good'' minima that generalize well share geometric properties that make it likely for SGD to find them \citep{huang_understanding_2020}.
\paragraph{Flatness and Noise Shapes.}
One such geometric property of a global minimizer is its \textit{flatness} \citep{hochreiter_flat_1997}. Empirically, \citet{keskar_large-batch_2016-1} discuss the advantages of small-batch stochastic gradient descent and propose that finding flat basins is a benefit of small-batch SGD: Large-batch training converges to models with both lower generalization and sharper minimizers. Although flatness is difficult to measure \citep{dinh_sharp_2017}, flatness based measures appear to be the most promising tool for predicting generalization in \citet{jiang_fantastic_2019}.
The analysis of such stochastic effects is often facilitated by considering the stochastic differential equation that arises for small enough step sizes $\tau$ from \cref{eq:sgd} under the assumption that the gradient noise is effectively a Gaussian random variable:
\begin{equation}\label{eq:firstorder_sde}
\diff\theta_t = -\nabla L(\theta_t)\diff t + \sqrt{\tau \Sigma_t} \diff W_t,
\end{equation}
where $\Sigma_t$ represents the covariance of gradient noise at time $t$, and $W_t$ is a Brownian motion modeling it.
The magnitude of $\Sigma_t$ is inversely proportional to mini-batch size \citep{jastrzebski_width_2018}, and it is also connected to the flatness of minima reached by SGD in \citet{dai_towards_2018} and \citet{jastrzebski_width_2018} if $\Sigma_t$ is isotropic. Analysis therein as well as in \citet{le_bayesian_2018} provides evidence that the step size should increase linearly with the batch size to keep the magnitude of noise fixed. However, the anisotropy of $\Sigma_t$ is strong enough to generate behavior that qualitatively differs from Brownian motion around critical points \citep{chaudhari_stochastic_2018,simsekli_tail-index_2019} and isotropic diffusion is insufficient to explain generalization benefits in \citet{saxe_information_2019}.
The shape of $\Sigma_t$ is thus further discussed in
\citet{zhu_anisotropic_2019} where anisotropic noise induced by SGD is found to be beneficial to reach flat minima in contrast to isotropic noise, \citet{zhou_towards_2020} where it is contrasted with noise induced by Adam \citep{kingma_adam:_2015}, and \citet{haochen_shape_2020} who discuss that such parameter-dependent noise, also induced by label noise, biases SGD towards well-generalizing minima. Empirical studies in \citet{wen_empirical_2020,wu_noisy_2020} and \citet{li_validity_2021} show that large-batch training can be improved by adding the right kind of anisotropic noise.
Notably, in all of these works, the noise introduced by SGD is in the end both \textit{unbiased} and (mostly) Gaussian, and its disappearance in full-batch gradient descent should remove its beneficial effects. However, \cref{eq:firstorder_sde} only approximates SGD to first-order, while for non-vanishing step sizes $\tau$, \citet{li_stochastic_2017} find that a second-order approximation,
\begin{equation}\label{eq:2ndorder_sde}
\diff\theta_t = -\nabla \left( L(\theta_t) + \frac{\tau}{4}||\nabla L(\theta)||^2 \right) \diff t + \sqrt{\tau \Sigma_t} \diff W_t,
\end{equation}
does include an implicit bias proportional to the step size. Later studies such as \citet{li_towards_2020,li_reconciling_2020} discuss the importance of large initial learning rates, which are also not well modeled by first-order SDE analysis but have a noticeable impact on generalization.
Analysis of flatness through other means, such as dynamical system theory \citep{wu_how_2018,hu_diffusion_2018}, also derives stability conditions for SGD and GD, where among all possible global minima, SGD both converges to flatter minima than GD and also can escape from sharp minima. \citet{xing_walk_2018} analyze SGD and GD empirically in response to the aforementioned theoretical findings about noise shape, finding that both algorithms (without momentum) significantly differ in their exploration of the loss landscape and that the structure of the noise induced by SGD is closely related to this behavior. \citet{yin_gradient_2018} introduce gradient diversity as a measure of the effectiveness of SGD:
\begin{equation}\label{eq:diversity}
\Delta_D(\theta) = \frac{\sum_{x \in X} \left|\left|\nabla \mathcal{L}(x_i, \theta)\right|\right|^2}{N^2 \left|\left| \nabla L(\theta) \right|\right|^2},
\end{equation}
which works well up to a critical batch size proportional to $\Delta_D(\theta)$. Crucially gradient diversity is a ratio of per-example gradient norms to the full gradient norm. This relationship is also investigated as gradient coherence in \citet{chatterjee_coherent_2020} as it depends on the amount of alignment of these gradient vectors.
Additional analysis of SGD as a diffusion process is facilitated in \citet{xie_diffusion_2020}.
\paragraph{An explicit, non-stochastic bias?}
Several of these theoretical investigations into the nature of generalization via SGD rely on earlier intuitions that this generalization effect would not be capturable by explicit regularization: \Citet{arora_implicit_2019} write that ``standard regularizers may not be rich enough to fully encompass the implicit regularization brought forth by gradient-based optimization'' and further rule out norm-based regularizers rigorously. Similar statements have already been shown for the generalization effects of overparametrization in \citet{arora_optimization_2018} who show that no regularizer exists that could replicate the effects of overparametrization in deep linear networks.
Yet, \citet{barrett_implicit_2020-1,smith_origin_2020,dandi_implicit_2021} find that the implicit regularization induced by GD and SGD can be analyzed via backward-error analysis and a scalar regularizer can be derived. The implicit generalization of mini-batched gradient descent with batches $B \in \mathcal{B}$ can be (up to third-order terms and sampling without replacement) described explicitly by the modified loss function
\begin{equation}\label{eq:sgd_bias}
L(\theta) + \frac{\tau}{4 |\mathcal{B}|} \sum_{B \in \mathcal{B}} \left|\left|\frac{1}{|B|} \sum_{x \in B} \nabla \mathcal{L}(x, \theta)\right|\right|^2,
\end{equation}
which simplifies for gradient descent to
\begin{equation}\label{eq:gd_bias}
L(\theta) + \frac{\tau}{4} \left|\left| \nabla L(\theta) \right|\right|^2,
\end{equation}
as found in \citet{barrett_implicit_2020-1}. Assumptions on approximation up to third-order can be exchanged for assumptions on Lipschitz smoothness of the Hessian, see \citet{dandi_implicit_2021}.
Training with this regularizer can induce the generalization benefits of larger learning rates, even if optimized with small learning rates, and induce benefits in generalization behavior for small batch sizes when training moderately larger batch sizes.
However, \citet{smith_origin_2020} ``expect this phenomenon to break down for very large batch sizes''.
Related are discussions in \citet{roberts_sgd_2018} and \citet{poggio_loss_2020}, who show a setting in which SGD can be shown to converge to a critical point where $\nabla \mathcal{L}(x_i, \theta)=0$ holds separately for each data point $x$, a condition which implies that the regularizer of \cref{eq:sgd_bias} is zero.
\paragraph{Large-batch training in practice.}\label{sec:ref_large_batch}
In response to \citet{keskar_large-batch_2016-1}, \citet{hoffer_train_2017} show that the adverse effects of (moderately) large batch training can be mitigated by improved hyperparameters -- tuning learning rates, optimization steps, and batch normalization behavior. A resulting line of work suggests hyperparameter improvements that successively allow larger batch sizes, \citep{you_large_2017} with reduced trade-offs in generalization. Yet, parity in generalization between small and large batch training has proven elusive in many applications, even after extensive hyperparameter studies in \citet{de_automated_2017,golmant_computational_2018,masters_revisiting_2018} and \citet{smith_generalization_2020}. \citet{golmant_computational_2018} go on to discuss that this is not only a problem of generalization in their experiments but also one of optimization during training, as they find that the number of iterations it takes to even reach low training loss increases significantly after the critical batch size is surpassed. Conversely, \citet{shallue_measuring_2019} find that training in a large-batch regime is often still possible, but this is dependent on finding an appropriate learning rate that is not predicted by simple scaling rules, and it also depends on choosing appropriate hyperparameters and momentum that may differ from their small-batch counterparts. A reduction of possible learning rates that converge reliably is also discussed in \citet{masters_revisiting_2018}, but a significant gap in generalization is observed in \citet{smith_generalization_2020} even after grid-searching for an optimal learning rate.
Empirical studies continue to optimize hyperparameters for large-batch training with reasonable sacrifices in generalization performance, including learning rate scaling and warmup \citep{goyal_accurate_2018,you_large-batch_2019}, adaptive optimizers \citep{you_large_2017,you_large_2019}, omitting weight regularization on scales and biases \citep{jia_highly_2018}, adaptive momentum \citep{mikami_massively_2019}, second-order optimization \citep{osawa_large-scale_2019}, and label smoothing \citet{yamazaki_yet_2019}. Yet, \citet{you_limit_2020} find that full-batch gradient descent cannot be tuned to reach the performance of SGD, even when optimizing for long periods, indicating a fundamental ``limit of batch size''.
The difficulty of achieving good generalization with large batches has been linked to instability of training. As discussed in \citet{cohen_gradient_2020,gilmer_loss_2021,giladi_at_2019}, training with GD progressively increases the sharpness of the objective function until training destabilizes in a sudden loss spike. Surprisingly however, the algorithm does not diverge, but quickly recovers and continues to decrease non-monotonically, while sharpness remains close to a stability threshold. This phenomenon of non-monotone, but effective training close to a stability threshold is also found in \citet{lewkowycz_large_2020}, where flat minima can be found for both GD and SGD if the learning rate is large enough, namely close to the stability threshold. Large but stable learning rates within a ``catapult" phase can succeed and generalize, and the training loss decreases non-monotonically.
\subsection{A more subtle hypothesis}
From the above literature, we find two main advantages of SGD over GD. First, its optimization behavior appears qualitatively different, both in terms of stability and in terms of convergence speed beyond the critical batch size. Secondly, there is evidence that the implicit bias induced by large step size SGD on mini batches can be replaced with explicit regularization as derived in \cref{eq:2ndorder_sde} and \cref{eq:sgd_bias} - a bias that approximately penalizes the per-example gradient norm of every example
In light of these apparent advantages, we hypothesize that \textit{ we can modify and tune optimization hyperparameters for GD and also add an explicit regularizer
in order to recover SGD's generalization performance without injecting any noise into training.}
This would imply that gradient noise from mini-batching is not necessary for generalization, but an intermediate factor; while modeling the bias of gradient noise and its optimization properties is sufficient for generalization, mini-batching by itself is not necessary and these benefits can also be procured by other means.
This hypothesis stands in contrast to possibilities that
gradient noise injection is either necessary to reach state-of-the-art performance (as in \citet{wu_noisy_2020,li_validity_2021}) or that no regularizing function exists with the property that its gradient replicates the practical effect of gradient noise \citep{arora_optimization_2018}.
A ``cultural'' roadblock in this endeavor is further that existing models and hyperparameter strategies have been extensively optimized for SGD, with a significant number of hours spent improving performance on CIFAR-10 for models trained with small batch SGD, which begets the question whether these mechanisms are by now self-reinforcing?
\section{Full-batch GD with randomized data augmentation}\label{sec:exp_data_aug}
We now investigate our hypothesis empirically, attempting to set up training so that strong generalization occurs even without gradient noise from mini-batching. We will thus compare \textit{full-batch} settings in which the gradient of the full loss is computed every iteration and \textit{mini-batch} settings in which a noisy estimate of the loss is computed.
Our central goal is to \textit{reach good full-batch performance without resorting to gradient noise}, via mini-batching or explicit injection. Yet, we will occasionally make remarks regarding \textit{full-batch} in practical scenarios outside these limitations.
For this, we focus on a well-understood case in the literature and train a ResNet model on CIFAR-10 for image classification.
We consider a standard ResNet-18 \citep{he_deep_2015,he_bag_2019} with randomly initialized linear layer parameters \citep{he_delving_2015} and batch normalization parameters initialized with mean zero and unit variance, except for the last in each residual branch which is initialized to zero \citep{goyal_accurate_2018}. This model and its initialization were tuned to reach optimal performance when trained with SGD. The default random CIFAR-10 data ordering is kept as is.
We proceed in several stages from baseline experiments using standard settings to specialized schemes for full-batch training, comparing stochastic gradient descent performance with full-batch gradient descent. Over the course of this and the next section we first examine full-batch training with standard data augmentations, and later remove randomized data augmentations from training as well to evaluate a completely noise-less pipeline.
\subsection{Baseline SGD}
We start by describing our baseline setup, which is well-tuned for SGD. For the entire \Cref{sec:exp_data_aug}, every image is randomly augmented by horizontal flips and random crops after padding by 4 pixels.
\textbf{Baseline SGD:}
For the SGD baseline, we train with SGD and a batch size of 128, Nesterov momentum of $0.9$ and weight decay of $0.0005$. Mini-batches are drawn randomly without replacement in every epoch. The learning rate is warmed up from 0.0 to 0.1 over the first 5 epochs and then reduced via cosine annealing to 0 over the course of training \citep{loshchilov_sgdr_2017}. The model is trained for 300 epochs. In total, $390 \times 300 = 117,000$ update steps occur in this setting.
With these hyperparameters, mini-batch SGD (sampling without replacement) reaches a validation accuracy of $95.70\% (\pm 0.05)$, which we consider a very competitive modern baseline for this architecture. Mini-batch SGD provides this strong baseline largely independent from the exact flavor of mini-batching as can be seen in \Cref{tab:baseline_gradient_noise}, reaching the same accuracy when sampling with replacement. In both cases the gradient noise induced by random mini-batching leads to strong generalization. If batches are sampled without replacement and in the same order every epoch, i.e. without shuffling in every epoch, then mini-batching still provides its generalization benefit. The apparent discrepancy between both versions of shuffling is not actually a SGD effect, but shuffling benefits the batch normalization layers also present in the ResNet-18. This can be seen by replacing batch norm with group normalization \citep{wu_group_2018}, which has no dependence on batching. Without shuffling we find $94.44\% (\pm 0.21)$ for SGD and with shuffling $94.55\% (\pm 0.16)$ for group normalized ResNets; a difference of less than $1\sigma$.
Overall any of these variations of mini-batched stochastic gradient descent lead to strong generalization after training. As a validation of previous work, we also note that the gap between SGD and GD is not easily closed by injecting simple forms of gradient noise, such as additive or multiplicative noise, as can also be seen in \Cref{tab:baseline_gradient_noise}.
\begin{table}
\centering
\begin{tabular}{l|c|c|c|c|c}
Source of Gradient Noise & Batch size & Val. Accuracy \%\\
\hline
Sampling without replacement & 128 &$95.70 (\pm 0.11)$ \\
Sampling with replacement & 128 & $95.70 (\pm 0.05)$ \\
Sampling without replacement (fixed across epochs) & 128 & $95.25 (\pm 0.07)$\\
\hline
Additive $n=0.01$ & 50'000 & $61.41 (\pm 0.09)$\\
Multiplicative $m=0.01$ & 50'000 & $79.25 (\pm 0.14)$\\
- & 50'000 & $75.42 (\pm 0.13)$ \\
\end{tabular}
\caption{Summary of validation accuracies on the CIFAR-10 validation dataset for \textit{baseline} types of gradient noise in experiments with data augmentations considered in \Cref{sec:exp_data_aug}.
}
\label{tab:baseline_gradient_noise}
\end{table}
With the same settings, we now switch to full-batch gradient descent. We replace the mini-batch updates by full batches and accumulate the gradients over all mini-batches. To rule out confounding effects of batch normalization, batch normalization is still computed over blocks of size 128 \citep{hoffer_train_2017}, although the assignment of data points to these blocks is kept fixed throughout training so that no stochasticity is introduced by batch normalization.
In line with literature on large-batch training, applying full-batch gradient descent with these settings reaches a validation accuracy of only $75.42\% (\pm 00.13)$, yielding a $\sim 20\%$ gap in accuracy between SGD and GD.
In the following experiments, we will close the gap between full-batch and mini-batch training. We do this by eschewing common training hyper-parameters used for small batches, and re-designing the training pipeline to maintain stability without mini-batching.
\subsection{Stabilizing Training}\label{sec:gentle}
Training with huge batches leads to unstable behavior. As the model is trained close to its \textit{edge of stability} \citep{cohen_gradient_2020}, we soon encounter spike instabilities, where the cross entropy objective $\mathcal{L}(\theta)$ suddenly increases in value, before quickly returning to its previous value and improving further. While this behavior can be mitigated with small-enough learning rates and aggressive learning rate decay (see supp. material), small learning rates also mean that the training will firstly make less progress, but secondly also induce a smaller implicit gradient regularization, i.e. \cref{eq:gd_bias}. Accordingly, we seek to reduce the negative effects of instability while keeping learning rates from vanishing. In our experiments, we found that very gentle \text{warmup} learning rate schedules combined with aggressive \textit{gradient clipping} enables us to maintain stability with a manageable learning rate.
\begin{table}
\centering
\begin{tabular}{l|c|c|c|c|c|}
Experiment & Mini-batching & Epochs & Steps & Modifications & Val. Acc.\%\\
\hline
Baseline SGD & \cmark & 300& 117,000 & - & $95.70 (\pm 0.11)$ \\
SGD regularized & \cmark & 300 & 117'000 & reg & $95.81 (\pm 0.18)$ \\
\hline
Baseline FB & \xmark & 300& 300 & - & $75.42 (\pm 0.13)$ \\
FB train longer &\xmark & 3000 & 3000 & - & $87.36 (\pm 1.23)$\\
FB clipped & \xmark & 3000 &3000 & clip & $93.85 (\pm 0.10)$\\
FB regularized & \xmark & 3000 &3000 & clip+reg & $95.54 (\pm 0.09)$ \\%$ 94.70 (\pm 0.17)$ \\
FB strong reg. & \xmark & 3000 &3000 & clip+reg+bs32 & $95.68 (\pm 0.09)$ \\%$ 94.97 (\pm 0.05)$ \\
\hline
FB in practice & \xmark & 3000 & 3000 & clip+reg+bs32+shuffle & $95.91 (\pm 0.14)$ \\%{95.60} (\pm 0.21)$ \\
\end{tabular}
\caption{Validation accuracies on the CIFAR-10 validation set for each experiment with data augmentations considered in \Cref{sec:exp_data_aug}. All validation accuracies are averaged over 5 runs.}
\label{tab:baseline_comparison}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.44\textwidth]{figures/fbaug_2.pdf}
\qquad
\includegraphics[width=0.44\textwidth]{figures/fbaug_clip_detailed.pdf}
\caption{Cross-Entropy Loss on the training and validation set and full loss (including weight decay) during training for full-batch gradient descent. Left: training as described in \Cref{sec:trainlonger} without clipping, right: with gradient clipping. Clipped steps are marked in black. Validation computed every 100 steps.
\label{fig:training_stability}
\end{figure}
\paragraph{Gentle learning rate schedules.}\label{sec:trainlonger}
Because full-batch training is notoriously unstable, the learning rate is now warmed up from 0.0 to 0.4 over 400 steps (each step is now an epochs) to maintain stability, and then decayed by cosine annealing (with a single decay without restarts) to 0.1 over the course of 3000 steps/epochs.
The initial learning rate of $0.4$ is not particularly larger than in the small-batch regime, and it is extremely small by the standards of a linear scaling rule \citep{goyal_accurate_2018}, which would suggest a learning rate of $39$, or even a square-scaling rule \citep{hoffer_train_2017}, which would predict a learning rate of $1.975$ when training longer. As the size of the full dataset is certainly larger than any critical batch size, we would not expect to succeed in fewer steps than SGD. Yet, the number of steps, $3000$, is simultaneously huge, when measuring efficiency in passes through the dataset, and tiny, when measuring parameter update steps. Compared to the baseline of SGD, this approach requires a ten-fold increase in dataset passes, but it provides a 39-fold decrease in parameter update steps. Another point of consideration is the effective learning rate of \citet{li_exponential_2019}. Due to the effects of weight decay over 3000 steps and limited annealing, the effective learning rate is not actually decreasing during training.
Training with these changes leads to full-batch gradient descent performance of $87.36\% (\pm 1.23)$, which is a $12\%$ increase over the baseline, but still ways off from the performance of SGD. We summarize validation scores in \Cref{tab:baseline_comparison} as we move across experiments.
\paragraph{Gradient Clipping.} We clip the gradient over the entire dataset to have an $\ell^2$ norm of at most $0.25$ before updating parameters.
Training with all the previous hyperparameters and additional clipping obtains a validation accuracy of $93.85 (\pm 0.10)$. This is a significant increase of $6\%$ over the previous result due to a surprisingly simple modification, even as other improvements suggested in the literature (label smoothing \citep{yamazaki_yet_2019}, partial weight decay \citep{jia_highly_2018}, adaptive optimization \citep{you_large_2017}, sharpness-aware minimization \citep{foret_sharpness-aware_2021} fail to produce significant gains, see appendix).
Gradient clipping is used in some applications to stabilize training \citep{pascanu_difficulty_2013}. However in contrast to its usual application in mini-batch SGD,
where a few batches with high gradient contributions might be clipped in every epoch, here the entire dataset gradient is clipped. As such, the method is not a tool against heavy-tailed noise \citep{gorbunov_stochastic_2020}, but it is effectively a limit on the maximum distance moved in parameter space during a single update. Because clipping simply changes the size of the gradient update but not its direction, clipping is equivalent to choosing a small learning rate when the gradient is large. Theoretical analysis of gradient clipping for GD in \citet{zhang_why_2019-1} and \citet{zhang_improved_2020} supports these findings, where it is shown that clipped descent algorithms can converge faster than unclipped algorithms for a class of functions with a relaxed smoothness condition. Clipping also does not actually repress the spike behavior entirely. To do so would require a combination of even stronger clipping and reduced step sizes, but the latter would reduce both training progress and regularization via \cref{eq:gd_bias}.
We note that with the addition of gradient clipping we now reach similar training loss than the original SGD baseline (see more details in the appendix), so that the modifications so far are sufficient to let fullbatch gradient descent mimic the optimization properties of mini-batched SGD. Yet, test error is still at $\sim6\%$, compared to the baseline of $\sim 4\%$.
\subsection{Bridging the gap with Explicit Regularization}\label{sec:explicit}
\begin{figure}
\centering
\includegraphics[width=0.44\textwidth]{figures/final_fbaug_gradreg_lr04_1.pdf}
\qquad
\includegraphics[width=0.44\textwidth]{figures/final_fbaug_highreg_lr04_0.pdf}
\caption{Cross-Entropy Loss on the training and validation set and full loss (including weight decay) during training for full-batch gradient descent. Clipped steps are marked in black. Validation computed every 100 steps. Training with gradient regularization: \textit{FB regularized}) on the left and \textit{FB strong reg.} on the right, both with a learning rate of 0.4.}
\label{fig:training_stability2}
\end{figure}
Finally, there is still the bias of mini-batch gradient descent towards solutions with low gradient norm per batch described in \cref{eq:2ndorder_sde} and \cref{eq:sgd_bias} to consider. This bias, although a 2nd-order effect, is noticeable in our experiments. We can replicate this bias as an explicit regularizer via \cref{eq:sgd_bias}. However, computing exact gradients of this regularizer directly is computationally expensive due to the computation of repeated Hessian-vector products in each accumulated batch, especially within frameworks without forward automatic differentiation which would allow for the method of \citet{pearlmutter_fast_1994} for fast approximation of the Hessian. As such, we approximate the gradient of the regularizer through a finite-differences approximation and compute
\begin{equation
\label{eq:approx}
\nabla \frac{1}{2}\left|\left| \nabla \mathcal{L}(x, \theta)\right|\right|^2 \approx \frac{\nabla \mathcal{L} \left(x, \theta + \varepsilon \nabla\mathcal{L}(x, \theta)\right) - \nabla\mathcal{L}(x, \theta) }{\varepsilon}.
\end{equation}
This approximation only requires one additional forward-backward pass, given that $\nabla\mathcal{L}(x, \theta)$ is already required for the main loss function. Its accuracy is similar to a full computation of the Hessian-vector products (see supplementary material).
In all experiments, we set $\varepsilon=\frac{0.01}{||\nabla \mathcal{L}(x, \theta)||}$, similar to \citep{liu_darts_2018}. To compute \cref{eq:sgd_bias}, the same derivation is applied for averaged gradients $\frac{1}{|B|}\sum_{x \in B} \nabla \mathcal{L}(x, \theta)$.
\paragraph{Gradient Penalty.} We regularize the loss via the gradient penalty in \cref{eq:sgd_bias} with coefficient $\frac{\alpha \tau_k}{4}$. We set $\alpha=1.0$ for these experiments.
We use this regularizer entirely without sampling, computing it over the fixed mini-batch blocks $B \in \mathcal{B}$, already computed for batch normalization, which are never shuffled. We control the strength of the regularization via a parameter $\alpha$. Note that this regularizer can be computed in parallel across all batches in the dataset. Theoretical results from \citet{smith_origin_2020} do not guarantee that the regularizer can work in this setting, especially given the relatively large step sizes we employ. However, the regularizer leads to the direct effect that not only $||\nabla L(\theta)||$ is small after optimization, but also $||\sum_{x \in B} \nabla \mathcal{L}(x, \theta)||$, i.e. the loss on each mini-batch. Intuitively, the model is optimized so that it is still optimal when evaluated only on subsets of the training set (such as these mini-batches).
Applying this regularizer on top of clipping and longer training leads to a validation accuracy of $ 94.70\% (\pm 0.17)$ for a regularizer accumulated over the default batch size of $|B|=128$. This can be further increased to $ 94.97 (\pm 0.05)$ if the batch size is reduced to $32$ (reducing the SGD batch size in the same way does not lead to additional improvement, see supp. material). Reducing $|B|$ is a beneficial effect as the regularizer \cref{eq:sgd_bias} is moved closer to a direct penalty on the per-example gradient norms of \citet{poggio_loss_2020}, yet computational effort increases proportionally. We visualize the new training behavior in \cref{fig:training_stability2}.
\paragraph{Double the learning rate.} We again increase the initial learning rate, now to 0.8 at iteration 400, which then decays to $0.2$ over the course of 3000 steps/epochs.
This second modification of the learning rate is interestingly only an advantage after the regularizer is included. Training with this learning rate and clipping, but without the regularizer (i.e. as in \Cref{sec:gentle}), reduces that accuracy slightly to $93.75\% (\pm 0.13)$. However, the larger learning rate does improve the performance when the regularizer is included, reaching $ 95.54 (\pm 0.09)$ if $|B|=128$ and $95.68 (\pm 0.09)$ if $|B|=32$, which is finally fully on par with SGD. The regularized full-batch gradient descent is actually stable for a large range of learning rates than the baseline version, reaching strong performance for learning rates in at least the interval $[0.4, 1.6]$, see more details in the appendix.
Overall, we find that after all modifications, both full-batch (with random data augmentations) and SGD behave similarly, achieving significantly more than $95\%$ validation accuracy. \Cref{fig:loss_visualization_detailed} visualizes the loss landscape around the found solution throughout these changes. Noticeably both clipping and gradient regularization correlate with a flatter landscape. We also verify that this behavior cannot be explained simply by a strong regularization: We also train the SGD baseline with the discussed regularization term, see \textit{SGD regularized}. Arguably this means the regularization is applied explicitly on top of the implicit effect that is already present. However, this fails to have a significant effect, verifying that instead of elevating both, the regularization indeed bridges the gap between SGD and GD.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figures/figure_3.pdf}
\caption{One-dimensional loss landscapes visualizations (random direction) of models trained with gradient descent, going from SGD (left) to GD with successive modifications (right). Whereas the models trained with unmodified gradient descent (middle) are noticeably sharper than the model trained with stochastic gradient descent (left), the final model trained with modified gradient descent (right) replicates the qualitative properties of the SGD model.}
\label{fig:loss_visualization_detailed}
\end{figure}
\begin{remark}[The Practical View]
Throughout these experiments with full-batch GD, we have decided not to shuffle the data in every epoch to rule out confounding effects of batch normalization. If we turn on shuffle again, we reach $95.91\% (\pm 0.14)$
validation accuracy (with separate runs ranging between $96.11\%$ and $95.71\%$), which even slightly exceeds SGD.
This is the practical view, given that shuffling is nearly for free in terms of performance, but of course potentially introduces a meaningful source of gradient noise - which is why it is not our main focus.
\end{remark}
Furthermore, to verify that this behavior is not specific to the ResNet-18 model considered so far, we also evaluate related vision models with exactly the same hyperparameters. Results are found in \Cref{tab:more_vision_models}, where we find that our methods generalize to the ResNet-50, ResNet-152 and a DenseNet-121 without any hyperparameter modification. For VGG-16, we do make a minimal adjustment and increase the clipping to $1.0$, as the gradients there are scaled differently, to reach parity with SGD.
\begin{table}
\small
\centering
\begin{tabular}{l|c|c|c|c|c|c}
Experiment & ResNet-18 & ResNet-50 & Resnet-152 & DenseNet-121 & VGG-16\\
\hline
Baseline SGD & $95.70 (\pm 0.11)$ & $95.65 (\pm 0.18)$ & $95.80 (\pm 0.15)$ & $95.80 (\pm 0.25)$ & $94.42 (\pm 0.89)$\\
\hline
Baseline FB & $75.42 (\pm 0.13)$ & $55.03 (\pm 3.89)$ & $77.26 (\pm 0.28)$ & $77.26 (\pm 0.28)$ & $46.95 (\pm 19.51)$ \\
FB train longer & $87.36 (\pm 1.23)$ & $85.12 (\pm 0.00)$ & $91.67 (\pm 0.35)$ & $89.56 (\pm 0.00)$ & $89.36 (\pm 1.19)$\\
FB clipped & $93.85 (\pm 0.10)$ & $92.55 (\pm 0.36)$ & $92.48 (\pm 0.46)$ & $92.59 (\pm 0.16)$ & $92.55 (\pm 0.21)$\\
FB regularized & $95.54 (\pm 0.09)$ & $95.84 (\pm 0.09)$ & $95.98 (\pm 0.12)$ & $95.92 (\pm 0.09)$ & $93.86 (\pm 0.18)$\\
FB strong reg. & $95.68 (\pm 0.09)$ & $96.06 (\pm 0.04)$ & $96.21 (\pm 0.12)$ & $96.08 (\pm 0.13)$ & $93.91 (\pm 0.17)$\\
\hline
FB in practice & $95.91 (\pm 0.14)$ & $96.50 (\pm 0.13)$ & $96.37 (\pm 0.56)$ & $96.43 (\pm 0.10)$ & $94.44 (\pm 0.07)$\\
\end{tabular}
\caption{Validation accuracies on the CIFAR-10 validation set for each of the experiments with data augmentations considered in \Cref{sec:exp_data_aug} for multiple modern CNNs. No hyperparameters were altered when evaluating the other architectures, aside from rescaling the clipping value for VGG to $1.0$.}
\label{tab:more_vision_models}
\end{table}
\section{Full-batch GD in the totally non-stochastic setting}\label{sec:noaug}
A final question remains -- if the full-batch experiments shown so far work to capture the effect of mini-batch SGD, what about the stochastic effect of random data augmentations on gradient noise? It is conceivable that the generalization effect is impacted by the noise variance of data augmentations. As such, we repeat the experiments of the last section in several variations.
\paragraph{No Data Augmentation.}\label{sec:exp_fixed}
If we do not apply any data augmentations and repeat previous experiments, then GD with clipping and regularization at $89.17\%$, substantially beats SGD with default hyperparameters at $84.32\% (\pm 1.12)$ and nearly matches SGD with newly tuned hyperparameters at $90.07 (\pm 0.48)$, see \Cref{tab:fixed_data_comparison}.
Interestingly, not only does the modified GD match SGD, the modified GD is even more stable, as it works well with the same hyperparameters as described in the previous section, and we must tune SGD even though it benefits from the same regularization implicitly.
\paragraph{Enlarged CIFAR-10}
To analyze both GD and SGD in a setting were they enjoy the benefits of augmentation, but without stochasticity, we replace the random data augmentations with a fixed increased CIFAR-10 dataset. This dataset is generated by sampling $N$ random data augmentations for each data point before training. These samples are kept fixed during training and never resampled, resulting in an $N$-times larger CIFAR-10 dataset. This dataset contains the same kind of variations that would appear through data augmentation, but is entirely devoid of stochastic effects on training. For a fixed initialization, training behaves deterministically and reaches the same solution up to floating-point and algorithm implementation precision, but incorporates the increase in data available through the augmented dataset.
If we consider this experiment for a $10\times$ enlarged CIFAR-10 dataset, then we do recover a value of $95.11 \%$. Note that we present this experiments only because of its implications for deep learning theory; computing the gradient over the enlarged CIFAR-10 is $N$-times as expensive, and there are additional training expenses incurred through increased step numbers and regularization. For this reason we do not endorse training this way as a practical mechanism. Note that SGD still have an advantage over $10\times$ CIFAR -- SGD sees 300 augmented CIFAR-10 datasets, once each, over its 300 epochs of training. If we take the same enlarged CIFAR-10 dataset and train SGD by selecting one of the 10 augmented versions in each epoch, then SGD reaches $95.20\% (\pm 0.09)$.
Overall, we find that we can reach more than $95\%$ validation accuracy entirely without stochasticity, after disabling gradient noise induced via mini-batching, shuffling as well as via data augmentations. The gains of $\sim 6\%$ compared to the setting without data augmentations are realized only through the increased dataset size. This shows that noise introduced through data augmentations does not appear to influence generalization in our setting and is by itself also not necessary for generalization. We further verify that this is not an effect of limited numerical precision or non-determinism in the appendix.
\begin{table}
\centering
\begin{tabular}{l|c|c|c|c|c|}
Experiment & Fixed Dataset & Mini-batching & Steps& Modifications & Val. Acc. \\
\hline
Baseline SGD & CIFAR-10& \cmark & $117,000$ & - & $ 84.32 (\pm 1.12)$ \\
Baseline SGD* & CIFAR-10& \cmark & $117,000$ & - & $90.07 (\pm 0.48)$ \\
FB strong reg. & CIFAR-10& \xmark & $3000$ & clip+reg+bs32 & $89.17 (\pm 0.24)$\\
\hline
Baseline SGD & $10\times$ CIFAR-10& \cmark & $117,000$ & - & $95.20 (\pm 0.09)$\\
FB train longer & $10\times$ CIFAR-10 & \xmark & $3000$ & - & $88.44 (-)$\\
FB strong reg. & $10\times$ CIFAR-10 & \xmark & $3000$ & clip+reg+bs32 & $95.11(-)$\\
\end{tabular}
\caption{Validation accuracies on the CIFAR-10 validation set for \textit{fixed} versions of the dataset with no random data augmentations in \Cref{sec:exp_fixed}. Hyperparameters fixed from the previous section except for SGD marked*, where the learning rate is doubled to 0.2 for a stronger baseline.}
\label{tab:fixed_data_comparison}
\end{table}
\section{Discussion \& Conclusions}
SGD, which was originally introduced to speed up computation, has become a mainstay of neural network training. The hacks and tricks at our disposal for improving generalization in neural models are the result of millions of hours of experimentation in the small batch regime. For this reason, it should come as no surprise that conventional training routines work best with small batches. The heavy reliance of practitioners on small batch training has made stochastic noise a prominent target for theorists, and SGD is and continues to be the practical algorithm of choice, but the assumption that stochastic mini-batching by itself is the unique key to reaching the impressive generalization performance of popular models may not be well founded.
In this paper, we show that full-batch training matches the performance of stochastic small-batch training for a popular image classification benchmark. We observe that (i) with randomized augmentations, full-batch training can match the performance of even a highly optimized SGD baseline, reaching $95.67\%$ for a ResNet-18 on CIFAR-10, (ii) without any form of data augmentation, fully non-stochastic training beats SGD with standard hyper-parameters, matching it when optimizing SGD hyperparameters, and (iii) after a 10$\times$ fixed dataset expansion, full-batch training with no stochasticity exceeds $95\%$, matching SGD on the same dataset. Nonetheless, our training routine is highly inefficient compared to SGD (taking far longer run time), and stochastic optimization remains a great practical choice for practitioners in most settings.
The results in this paper focus on commonly used vision models. While the scope may seem narrow, the existence of these counter-examples is enough to show that stochastic mini-batching, and by extension gradient noise, is not required for generalization. It also strongly suggests that any theory that relies exclusively on stochastic properties to explain generalization is unlikely to capture the true phenomena responsible for the success of deep learning.
Stochastic sampling has become a focus of the theory community in efforts to explain generalization. However, experimental evidence in this paper and others suggests that strong generalization is achievable with large or even full batches in several practical scenarios. If stochastic regularization does indeed have benefits in these settings that cannot be captured through non-stochastic, regularized training, then those benefits are just the cherry on top of a large and complex cake.
\section*{Reproducibility Statement}
We detail all hyperparameters in the main body and provide all additional details in \cref{sec:experimental_setup}. Our open source implementation can be found at \url{https://github.com/JonasGeiping/fullbatchtraining} and contains the exact implementation with which these results were computed and we further include all necessary scaffolding we used to run distributed experiments on arbitrarily many GPU nodes, as well as model checkpointing to run experiments on only a single machine with optional GPU. Overall, we thus believe that our experimental evaluation is accessible and reproducible for a wide range of interested parties.
\section*{Funding Statement}
This research was made possible by the OMNI cluster of the University of Siegen which contributed a notable part of its GPU resources to the project and we thank the Zentrum für Informations- und Medientechnik of the University of Siegen for their support. We further thank the University of Maryland Institute for Advanced Computer Studies for additional resources and support through the Center for Machine Learning cluster.
This research was overall supported by the universities of Siegen and Maryland and by the ONR MURI program, AFOSR MURI Program, and the National Science Foundation Division of Mathematical Sciences.
{\small
|
\section{Supplemental Material: Thermodynamics of a minimal algorithmic cooling refrigerator}
The Supplemental Material contains details about (I) the analytical solution of the dynamics of the three-qubit heat bath algorithmic cooling refrigerator for arbitrary cycle number, (II) the theoretical evaluation of heat, work, target qubit polarization, as well as cooling power and coefficient of performance for arbitrary initial polarizations, {(III) the general properties of the imperfect compression map, (IV) the experimental implementation of the cooling algorithm, and (V) the experimental evaluation of the thermodynamic performance of the refrigerator.}
\section{Analytical solution of the qubit dynamics in Liouville space}
This section presents the exact solution of the dynamics of the three-qubit system in Liouville space \cite{gya20}. We begin with a brief reminder of the vectorization technique in order to set the notation \cite{gil09}.
\subsection{Vectorization method}
The Liouville representation provides a compact and efficient method to describe quantum processes by mapping operators in Hilbert space onto vectors in an enlarged space.
Let the space of operators on a Hilbert space be \(\mathsf{B}(\mathbf{H})\). Let us further consider its representation in terms of the Hilbert space and its dual, \(\mathsf{B}(\mathbf{H}) \cong \mathbf{H} \otimes \mathbf{H}^*\), or in other words, in terms of bras and kets. Vectorization is then the isomorphism mapping
\begin{equation}
\opn{vec} \colon \mathbf{H} \otimes \mathbf{H}^*
\to
\mathbf{H} \otimes \mathbf{H} ,
\end{equation}
where we define the Liouville space as \(\mathbf{L} = \mathbf{H} \otimes \mathbf{H}\).
This map naturally extends to linear maps on the space of operators itself. In the case of linear transformations, such as the operator-sum representation of quantum channels, we have \(\mathcal{M}[\bullet] = \sum_\mu M_\mu \bullet M_\mu^\dag \in \mathsf{B}(\mathbf{H}) \otimes \mathsf{B}(\mathbf{H}^*)\), with \(\mathsf{B}(\mathbf{H}^*) \cong \mathsf{B}(\mathbf{H})^*\). Adopting the character of Hilbert space of the spaces of operators themselves, vectorization extends to them yielding
\begin{equation}
\text{(extended)}\quad
\opn{vec} \colon \mathsf{B}(\mathbf{H}) \otimes \mathsf{B}(\mathbf{H}^*)
\to
\mathsf{B}(\mathbf{H}) \otimes \mathsf{B}(\mathbf{H}) .
\end{equation}
Since we are working with qubit spaces, we have \(\mathbf{H} = \mathbb{C}^2\) and \(\mathsf{B}(\mathbf{H}) = \mathbb{M}^{2 \times 2}_\mathbb{C}\), where \(\mathbb{M}^{2 \times 2}_\mathbb{C}\) is the space of 2-by-2 matrices with complex entries. Starting from these building blocks, vectorization is the same as rearranging columns and rows of matrices, with different ways of doing this related by reshuffling their components.
With the conventions taken in the main text, the action of vectorization on density matrices is to stack their columns in a single-column vector \cite{hor12}. We denote the vectorized density matrices as \(\vec{\rho} = \opn{vec}(\rho)\). Quantum channels, in operator-sum representation with Kraus operators \(E_\mu\), are accordingly mapped to the matrices
\begin{equation} \label{eq:supermatrix}
\Phi_\mathcal{M} = \sum_\mu M_\mu \otimes (M_\mu^\dag)^\intercal ,
\end{equation}
that act on \(\vec\rho\) from the left as regular matrix multiplication. These rules extend to the states and channels on the tensor product of target and two reset qubits Hilbert spaces, the composite Liouville state is thus \(\mathbf{L} = \mathbf{L}_1 \otimes \mathbf{L}_2 \otimes \mathbf{L}_3\).
\subsection{Solutions for target and reset qubits}
In order to evaluate the states of the target qubit and of the two reset qubits after an arbitrary number of refrigeration cycles \(n\), we need to compute the matrix in Liouville space of the combined quantum channel consisting of damping \(\mathcal{D}\), compression \(\mathcal{C}\) and refresh \( \mathcal{R}\) maps. As a first step, we treat the damping channel \(\mathcal{D}[\bullet]= \sum_{j} \Gamma_j \bullet \Gamma_j^\dag \), given by Eqs.~(2)-(3) of the main text, and use formula \eqref{eq:supermatrix} to build the corresponding Liouville superoperator \(\Phi_\mathcal{D} = \sum_j \Gamma_\mu \otimes (\Gamma_\mu^\dag)^\intercal\), a 4-by-4 matrix acting on the target subspace.
As a second step, we consider the compression channel, given by Eqs.~(4)-(6) of the main text. {Since we focus on the target qubit for the time being, we introduce the reduced compression channel \(\mathcal{C}_\text{red}[\rho_1] = \tr_{23}\big\{ \sum_{k = 1, 2} K_k \big( \rho_1 \otimes \rho_2(0) \otimes \rho_3(0) \big) K_k^\dag \big\}\), that acts on the target qubit space alone.}
{We determine the Kraus operators \(C_j\) of the operator-sum representation $\mathcal{C}_\text{red}[\bullet]= \sum_{j}
C_j \bullet C_j^\dag$} by extending the treatment that can be found in Ref.~\cite{nie00}, Section 8.2.3, to nonunitary operations. It involves introducing a purification of the mixed state of reset qubits, that is, the map
\begin{equation}
\rho_2(0) \otimes \rho_3(0) \mapsto \ket{\rho_2, \rho_3} \in \mathbf{H}_2 \otimes \mathbf{H}_3 \otimes \mathbf{H}_\text{R} ,\quad \text{with }
\ket{\rho_2, \rho_3} = \sum_{ij} \ket{ij} \ket{ij}_\text{R} \sqrt{ \bra{i}\rho_2(0)\ket{i} \bra{j}\rho_3(0)\ket{j} } ,
\end{equation}
such that \(\tr_\text{R} \ketbra{\rho_2, \rho_3}{\rho_2, \rho_3} = \rho_2(0) \otimes \rho_3(0)\), where \(\mathbf{H}_\text{R}\) is an artificial, reference Hilbert space introduced as part of the purification. By explicitly evaluating the trace and the purification, we find,
\begin{equation}
C_{(k), ij}^{i'j'} = \bra{i'j'} \bra{ij}_\text{R} \big( K_k \otimes I_\text{R} \big) \sum_{rs} \sqrt{ p_2^r p_3^s } \ket{rs}\ket{rs}_\text{R} ,\quad \text{with } p_2^r = \bra{r}\rho_2(0)\ket{r},\ p_3^s = \bra{s}\rho_3(0)\ket{s} ,
\end{equation}
where \(I_\text{R}\) is the identity on the reference Hilbert space. The bras here stem from the trace while the sum over double-primed indices and their kets stem from the purified state. This leads to the operators%
\begin{equation}
C_{(k), ij}^{i'j'} = \bra{i'j'} K_k \ket{ij} \sqrt{p_2^i p_3^j} .
\end{equation}
To simplify the notation, we group the indices \(i, j, i', j'\) and \(k\), each binary, into the new index \(\mu\) ranging over 32 values, and thus yielding 32 operators. Most of them are identically 0 through this procedure, however, with six others remaining for \(k = 1\) and other four for \(k = 2\). The corresponding superoperator \(\Phi_\mathcal{\tilde C}\) is again given by Eq.~\eqref{eq:supermatrix}.
\begin{figure}[t]
\centering
\includegraphics{S1a}
\includegraphics{S1b}
\caption{Heat \(Q(n)\), Eq.~\eqref{12}, and work \(W(n)\), Eq.~\eqref{16}, as a function of the number \(n\) of cycles, for various values of damping coefficient \(\gamma\) and of the mixing angle \(\theta\). Parameters are \(\epsilon_1(0) = 0\), \(\epsilon_2(0)=\epsilon_3(0)=0.6\).}
\label{fig:qw}
\end{figure}
The state of the target qubit, which is needed to determine heat and cooling power of the heat bath algorithmic cooling refrigerator, is obtained after concatenation of the two maps \(\mathcal{D}\) and \(\mathcal{C}_\text{red}\), which in vectorized form reads \(\Phi_{\mathcal{C}_\text{red}} \Phi_\mathcal{D}\). The target qubit states after a number \(n\) of cycles are then obtained by calculating \((\Phi_{\mathcal{C}_\text{red}}\Phi_\mathcal{D})^n\). We find for arbitrary initial polarizations \(\epsilon_1(0)\), \(\epsilon_2(0)\) and \(\epsilon_3(0)\)
\begin{equation}
\label{9}
\tilde \rho_1(n)
= \opn{unvec}\big\{ \big( \Phi_{\mathcal{C}_\text{red}} \Phi_\mathcal{D} \big)^n \vec\rho_1(0) \big\}
= \frac{1}{2}
\begin{pmatrix}
1 - \epsilon{1}(n) & 0 \\ 0 & 1 + \epsilon{1}(n)
\end{pmatrix}.
\end{equation}
The polarization \(\epsilon{1}(n, \theta, \gamma)\) is here explicitly given by
\begin{equation} \label{eq:target-qubit}
\epsilon{1}(n) =
\frac{
2 (\epsilon{2}(0) + \epsilon{3}(0)) \sin^2\theta - \gamma F(\theta)
+ \big[ 2 \sin^2\theta \big( (1 + \epsilon{2}(0) \epsilon{3}(0)) \epsilon{1}(0) - \epsilon{2}(0) - \epsilon{3}(0) \big) + \gamma (1 + \epsilon{1}(0)) F(\theta) \big] e^{-nG(\theta,\gamma) }
}{ ( \gamma - 1 ) F(\theta) + 4 }
\end{equation}
where
\begin{equation}
\begin{split}
F(\theta) ={}& 3 + (1 + \epsilon{2}(0)\epsilon{3}(0)) \cos(2\theta) - \epsilon{2}(0)\epsilon{3}(0) \xrightarrow{\epsilon{2}(0) = \epsilon{3}(0)} f(\theta) , \\
G(\theta, \gamma) ={}& \ln\left( \frac{4}{ (1 - \gamma) F(\theta) } \right) \xrightarrow{\epsilon{2}(0) = \epsilon{3}(0)} g(\theta, \gamma) .
\end{split}
\end{equation}
These expressions are symmetric under exchange of the initial polarizations of the two reset qubits (\(\epsilon{2}(0) \leftrightarrow \epsilon{3}(0)\)). {The ideal asymptotic polarization \(2\epsilon/(1+\epsilon^2)\), obtained for \(\gamma=0\), \(\theta=\pi/2\) and \(\epsilon_1(0)=0\), \(\epsilon_2(0) = \epsilon_3(0) = \epsilon\), agrees with the one derived in Refs.~\cite{rod16,zai21}. For different reset spin polarizations, this ideal limit gets modified to \([\epsilon_2(0)+ \epsilon_3(0)]/[1+\epsilon_2(0) \epsilon_3(0)]\). }
The quantum map \(\Phi_{\mathcal{C}_\text{red}} \Phi_\mathcal{D}\) is found without difficulties because the compression step is evaluated with respect to a tensor product with fixed reset states, allowing for the method of Ref.~\cite{nie00} to be applied. However, this simplification does not occur when changing reset qubits are involved, as in the refresh operation.
\begin{figure}[t]
\centering
\includegraphics{S2a.pdf}
\includegraphics{S2b.pdf}
\includegraphics{S2c.pdf}
\includegraphics{S2d.pdf}
\caption{Heat \(Q(n)\), Eq.~\eqref{12}, work \(W(n)\), Eq.~\eqref{16}, and polarization of the target qubit \(\epsilon_1(n)\), Eq.~\eqref{9}, as well as the corresponding temperature \(T_c(n)\), as a function of the number \(n\) of cycles, for fixed values of the damping coefficient \(\gamma=0.1\) and of the mixing angle \(\theta=\pi/3\), for different initial polarizations of the reset qubit 3. Parameters are \(\epsilon_1(0) = 0\), \(\epsilon_2(0) =0.3\). Changing the initial polarization of the reset qubit may either increase or decrease the values of these thermodynamic quantities.
Dashed lines in the lower right plot correspond to the respective temperatures of the reset qubits with matching colors (the red line also corresponds to the polarization \(\epsilon_2(0)\)).
}
\label{fig:diffpol}
\end{figure}
As a next step, we deal with the refresh operator \(\mathcal{R}[ \bullet ]\) given by Eq.~(7) of the main text. To that end, we consider an extension \(\mathcal{D}_\text{ext}\) of the damping channel that acts trivially (through the identity) on the reset qubits \(2\) and \(3\). We accordingly define the channel \(\mathcal{E}=\mathcal{R}[ \mathcal{C}[ \mathcal{D}_\text{ext}[\bullet] ] ]\)
for the ensemble of three qubits, as
\begin{equation} \label{eq:full-state}
\rho(n) = \mathcal{E}[ \rho(n-1) ]
= \mathcal{R}\left[ \sum_k K_k \big( \mathcal{D}_\text{ext} [\rho(n-1)] \big) K_k^\dag \right],
\end{equation}
where \(\rho(n) = \rho_1(n) \otimes \rho_2(0) \otimes \rho_3(0)\). The role of the refresh operation is to keep the reset states equal to \(\rho_2(0) \otimes \rho_3(0)\) at the beginning and at the end of each cooling cycle. We denote the corresponding superoperator as \(\Phi_\mathcal{E}\).
The determination of the coefficient of performance and, in turn, of the work applied during the compression stage requires the knowledge of the reset qubit states before the refresh stage. The latter are given by
\begin{equation} \label{eq:bath-qubit}
\begin{split}
\tilde \rho_i(n) = \tr_{1, j \neq i}\big\{
\sum_k K_k \big( \mathcal{D}_\text{ext} [\rho(n-1)] \big) K_k^\dag\big\}
= \frac{1}{2}
\begin{pmatrix}
1 - \epsilon{i}(n) & 0 \\
0 & 1 + \epsilon{i}(n)
\end{pmatrix} ,
\end{split}
\end{equation}
where \(i\) and \(j\) are either 2 or 3, {and \(\rho(n-1) = \rho_1(n-1) \otimes \rho_2(0) \otimes \rho_3(0)\)}. The polarizations \(\epsilon{i}(n)\) explicitly read
\begin{multline}
\epsilon_i(n) =
\frac{
2 I_i(\gamma) \sin^2\theta + 4\gamma ( 1 - \epsilon{i}(0) \cos^2\theta )
}
{ 2 [ ( \gamma - 1 ) F(\theta) + 4 ] } \\
+ \frac{
\sin^2\theta ( 1 + \epsilon{2}(0) \epsilon{3}(0) ) }{F(\theta)} \frac{
\big[ 2 \sin^2\theta \big( (1 + \epsilon{2}(0) \epsilon{3}(0)) \epsilon{1}(0) - \epsilon{2}(0) - \epsilon{3}(0) \big) + \gamma (1 + \epsilon{1}(0) F(\theta) \big]
}
{ ( \gamma - 1 ) F(\theta) + 4 } e^{-nG(\theta,\gamma)} ,
\end{multline}
with
\begin{equation}
I_i(\gamma) = \big( (\gamma - 1) \epsilon{i}(0)^2 + \epsilon{i}(0) + \gamma \big) \epsilon{j \neq i}(0) - \epsilon{i}(0) + 1 .
\end{equation}
{The reset qubit states only change within a single stroke (before they are refreshed). Their dependence upon the \(n-1\) previous applications of the cooling cycle is implicit in the target qubit input state \(\rho_1(n-1)\). This is schematically represented as \(\epsilon{1}(n) \to \epsilon{1}(n+1)\) while \(\epsilon{2, 3}(0) \to \epsilon{2, 3}(n+1)\).}
\begin{figure}[t]
\centering
\includegraphics{S3a.pdf}
\includegraphics{S3b.pdf}
\caption{Cooling power \(J(n)\), Eq.~\eqref{15}, and coefficient of performance (COP) \(\zeta(n)\), Eq.~\eqref{17}, as a function of the number \(n\) of cycles, for fixed values of the damping coefficient \(\gamma=0.1\) and of the mixing angle \(\theta=\pi/3\), for different initial polarizations of the reset qubit 3.
Parameters are \(\epsilon_1(0) = 0\), \(\epsilon_2(0) = 0.3 \).
}
\label{fig:diffpol2}
\end{figure}
\section{Theoretical evaluation of the thermodynamic quantities}
We now turn to the evaluation of the thermodynamic quantities of the heat bath algorithmic cooling refrigerator and, in particular, of their fundamental upper bounds in the reversible limit (\(\gamma=0, \theta= \pi/2\)), for arbitrary initial polarizations.
The heat per cycle follows directly from Eq.~\eqref{9} from the definition given in the main text and reads
\begin{equation}
\label{12}
Q(n) = \big[
2 ( (1 + \epsilon{2}(0) \epsilon{3}(0)) \epsilon{1}(0) - \epsilon{2}(0) - \epsilon{3}(0) ) \sin^2\theta + \gamma (1 + \epsilon{1}(0)) F(\theta)
\big]
\frac{ e^{-nG(\theta,\gamma)} }{4}.
\end{equation}
The corresponding cooling power is found to be
\begin{equation}
\label{15}
\begin{split}
J(n) ={}& { Q(n + 1) - Q(n) }
={} \left[\frac{(1 - \gamma) F(\theta)}{4} - 1 \right] {Q(n)} \\
\xrightarrow[\theta = \theta_n]{\gamma \to 0} {}& J_\text{max}(n)=
[ 1+\epsilon{2}(0) \epsilon{3}(0) ][ \epsilon{2}(0) + \epsilon{3}(0) - \epsilon{1}(0) ( 1 + \epsilon{2}(0) \epsilon{3}(0) ) ] \frac{ e^{- nG(\theta_n, 0) } }{ 2 - 2\epsilon{2}(0) \epsilon{3}(0) } \\
&+ \gamma \big[2 \epsilon{2}(0)+2 \epsilon{3}(0)+\epsilon{2}(0)^{2} \epsilon{3}(0)^{2}
- 1 \\
&\quad- \big( 1+\epsilon{2}(0) \epsilon{3}(0) \big)\big( \epsilon{1}(0) [ n - 3 + (1+n) \epsilon{2}(0) \epsilon{3}(0) ] - n [ \epsilon{2}(0)+\epsilon{3}(0) ] \big)
\big]
\frac{ e^{- nG(\theta_n, 0) } }{ 2 - 2\epsilon{2}(0) \epsilon{3}(0) } .
\end{split}
\end{equation}
The cooling power \(J(n)\) is proportional to the heat \(Q(n)\) since the finite difference of an exponential is again an exponential. The fundamental upper bound \(J_\text{max}(n)\), Eq.~\eqref{15}, generalizes Eq.~(8) of the main text to arbitrary initial polarizations \(\epsilon_i(0)\) of the three qubits. {
Contrary to polarization, maximum power is not achieved for \(\theta \to \pi/2\). This happens because the steady state value \(J(\infty)\) is reached faster in this limit due to large values of the power in the first two rounds. This then leads to a suppressed heat removal from the target qubit and a reduced power output. To maximize the refrigeration power \(J(n)\), a solution is to suppress the exponential decay in \(n\) by decreasing the angle \(\theta\) (and therefore the decay coefficient \(G(\theta, \gamma)\)) with the number of strokes. The optimal value of \(\theta\) is
\begin{equation}
\theta_n =
\begin{cases}
\pi/2 ,\quad& n = 0 \\
\begin{cases}
\pi/2 ,\quad& (\epsilon{1}(0), \epsilon{2}(0), \epsilon{3}(0)) \in \mathbb{J} \\
\theta_\text{opt} ,\quad& \text{otherwise}
\end{cases} ,\quad& n = 1 \\
\theta_\text{opt} ,\quad& n \geq 2 ,
\end{cases}
\end{equation}
where the angle \(\theta_\text{opt}\) reads
\begin{equation}
\theta_\text{opt} = \frac{1}{2} \arccos\left( \frac{2 \epsilon{2}(0)\epsilon{3}(0) + n \epsilon{2}(0)\epsilon{3}(0) + n - 6}{(2 + n) (1 + \epsilon{2}(0)\epsilon{3}(0))} \right) .
\end{equation}
The condition \(\mathbb{J}\) is given by
\begin{equation}
\mathbb{J} = \left\{
(\epsilon{1}(0), \epsilon{2}(0), \epsilon{3}(0)) \in [0, 1]^3
\ \Big|\
0 \leq \epsilon{1}(0) < \frac{1}{\sqrt{3}} \cap
\left(
0 \leq \epsilon{2}(0) < \frac{1}{3 \epsilon{1}(0)} \cap
0 \leq \epsilon{3}(0) < \frac{1}{3 \epsilon{2}(0)}
\right)
\right\} ,
\end{equation}
where we always assumed that \( (\epsilon{2}(0), \epsilon{3}(0)) \geq \epsilon{1}(0) \geq 0 \).
}
On the other hand, using Eq.~\eqref{eq:bath-qubit} and the definition of the work done on the qubit system, we obtain
\begin{multline}
\label{16}
W(n) = 4 \sin^2\theta \frac{ (1 + \epsilon{2}(0)) (1 + \epsilon{3}(0)) }{(\gamma - 1) F(\theta) + 4}
+ \left(
1
+ 4 \gamma \sin^2\theta \frac{
(1 + \epsilon{2}(0) \epsilon{3}(0)) (\gamma - 1)
}{(\gamma - 1) F(\theta) + 4}
\right) \\
\times
\big(
2 [ (1 + \epsilon{2}(0) \epsilon{3}(0)) \epsilon{1}(0) - \epsilon{2}(0) - \epsilon{3}(0) ] \sin^2\theta + \gamma (1 + \epsilon{1}(0)) F(\theta)
\big) \frac{e^{-nG(\theta,\gamma)}}{4} .
\end{multline}
The coefficient of performance eventually follows as
\begin{equation}
\label{17}
\begin{split}
\zeta(n) ={}& \frac{
- [ (\gamma - 1) F(\theta) + 4 ] R(\theta, \gamma) e^{-n G(\theta,\gamma)}
}{
\big[
(\gamma - 1) F(\theta) + 4
- 4 (1 + \epsilon{2}(0) \epsilon{3}(0)) (1 - \gamma) \sin^2(\theta)
\big] R(\theta, \gamma) e^{-nG }
+ 16 \sin^2(\theta) (1 + \epsilon{2}(0)) (1 + \epsilon{3}(0))
} \\
{}&\xrightarrow[\theta = \pi/2]{\gamma \to 0}
\zeta_\text{max}(n)= 1
+ \frac{ 4 \gamma }{ 1+\epsilon{2}(0) \epsilon{3}(0) }
\left(
1 + \frac{
\left(1+\epsilon{2}(0)\right)\left(1+\epsilon{3}(0)\right)
}{
\epsilon{1}(0)-\epsilon{2}(0)-\epsilon{3}(0)+\epsilon{1}(0) \epsilon{2}(0) \epsilon{3}(0)
} e^{nG(\pi/2, 0) }
\right)
\end{split}
\end{equation}
where \( R(\theta, \gamma) = 2 \sin^2(\theta) [ (1 + \epsilon{2}\epsilon{3}) \epsilon{1}(0) - \epsilon{2} - \epsilon{3} ] + \gamma (1 + \epsilon{1}(0)) F(\theta) \). The fundamental upper bound \(\zeta_\text{max}(n)\), Eq.~\eqref{17}, generalizes Eq.~(8) of the main text to arbitrary initial polarizations \(\epsilon_i(0)\) of the three qubits.
Heat \(Q(n)\), Eq.~\eqref{12}, and work \(W(n)\), Eq.~\eqref{16} are shown in Fig.~\ref{fig:qw} as a function of the number \(n\) of cycles, for various values of decay rate \(\gamma\) and of the mixing angle \(\theta\). The influence of unequal initial polarizations of the reset qubits is illustrated in Figs.~\ref{fig:diffpol} and \ref{fig:diffpol2} for fixed values of the damping rate \(\gamma=0.1\) and of the mixing angle \(\theta=\pi/3\). Work \(W(n)\) and polarization \(\epsilon_1(n)\) of the target qubit, Eq.~\eqref{eq:target-qubit}, are increased when \(\epsilon_3(0) >\epsilon_2(0)\), whereas heat \(Q(n)\) is decreased. At the same time, cooling power \(J(n)\), Eq.~\eqref{15}, and COP \(\zeta(n)\), Eq.~\eqref{17}, are also both increased when \(\epsilon_3(0) >\epsilon_2(0)\).
{\section{Generality of the properties of the imperfect compression gate}}
{The imperfect compression map \(\mathcal{C}\) parametrized by the mixing angle \(\theta\) preserves the target steady state of perfect compression \(\mathcal{C}_{\theta = \pi/2}\) and, at the same time, slows down the convergence to the steady state for vanishing dissipation \(\gamma = 0\). We show in this Section that these features generically hold for a family of imperfect compression maps given by a convex combination of two unitaries, namely the ideal compression and the identity. The structure of these generalized nonideal compression maps will provide additional physical insight into their remarkable properties.}
{Let us consider the trace preserving completely positive map that chooses between the application of the perfect compression gate and the identity (i.e.~do nothing) with a probability distribution \((p_1=\sin^2\theta, p_2=\cos^2\theta)\):
\begin{equation}
\label{21}
\mathcal{C}_\text{gen}[\bullet] =
\sin^2\theta \,U \bullet U^\dag + (\bullet) \cos^2\theta,
\end{equation}
where \(U= \exp(-\mathrm{i} \pi V/2)\) with \(V=\ketbra{100}{011} + \ketbra{011}{100}\) is the unitary describing ideal compression swap. Such a convex combination of unitary operations (sometimes called random external fields map \cite{ben06}) is unital by construction and thus leaves the maximally mixed state invariant \cite{ben06}. The operator-sum representation of the map \(\mathcal{C}_\text{gen}\) on states diagonal in the energy eigenbasis (and only on those) is identical to that of the compression map \(\mathcal{C}\). The random map \(\mathcal{C}_\text{gen}\) may thus be regarded as a generalization of the compression map \(\mathcal{C}\).}
\begin{figure}[t]
\centering
\includegraphics{S4.pdf}
\caption{{Graph representing the causal tree of possible evolution paths of fine-grained realizations of the generalized nonideal compression map \eqref{21} given as a convex combination of the ideal compression (with probability \(p_1 = \sin^2\theta\)) and the identity map (with probability \(p_2=\cos^2\theta\)). Nonzero dissipation \(\gamma \neq 0\) leads to different state states of the target qubit
that depend on the angle \(\theta\). The symbol \(\circ\) indicates composition of maps.}}
\label{fig:graph}
\end{figure}
{Let us next show that the map \eqref{21} preserve the state state of the target quit. In analogy to Eq.~\eqref{eq:full-state}, we introduce the concatenated three-qubit map \(\mathcal{E}_{\theta}=\mathcal{R}[ \mathcal{C}_\text{gen}[ I [\bullet] ] ]\) that combines compression and refresh maps (in the absence of dissipation). We denote by \(\rho_\text{ss}=\mathcal{E}^\infty_{\pi/2}(\rho_0)\) the steady state of \(\mathcal{E}_{\pi/2}\). We then have
\begin{equation}
\begin{split}
\lim_{n \to \infty} \mathcal{E}_{\pi/2}^n(\rho_1(0))
= \lim_{n \to \infty} \mathcal{E}_{\pi/2}(\rho_1(n-1)) ={}& \lim_{n \to \infty}
\tr_\text{reset}\{
U (\rho_1(n-1) \otimes \rho_\text{reset}) U^\dag
\} \\
={}& \tr_\text{reset}\{
U (\rho_\text{ss} \otimes \rho_\text{reset}) U^\dag
\} \\
\equiv{}& \rho_\text{ss} .
\end{split}
\end{equation}
Thus, in this limit, the concatenation of dilation, unitary, and trace, acts as an identity operation on \(\rho_\text{ss}\). With this property, we can prove that \(\mathcal{E}_\theta\) has a steady state
which in fact does not depend on \(\theta\):
\begin{equation}
\begin{split}
\lim_{n \to \infty} \mathcal{E}_\theta(\rho_1(n-1)) ={}& \lim_{n \to \infty}
\sin^2\theta \tr_\text{reset}\{
U (\rho_1(n-1) \otimes \rho_\text{reset}) U^\dag
\}
+ \lim_{n \to \infty} \cos^2(\theta) \rho_1(n-1) \\
={}& \sin^2\theta \tr_\text{reset}\{
U (\rho_\text{ss} \otimes \rho_\text{reset}) U^\dag
\}
+ \cos^2(\theta) \rho_\text{ss} \\
={}& \rho_\text{ss} . \label{eq:ss}
\end{split}
\end{equation}
In summary, at each step the map combines two states sharing the same asymptotic value. As a result, the steady state \(\rho_\text{ss}\) is preserved.}
{Physically, the angle \(\theta\) interpolates between a map which implements the one-shot ideal compression at every single step (\(\theta = \pi/2\)), and an identity map that not does nothing (\(\theta = 0\)). In this one-shot regime, the features of the nonideal compression may be intuitively understood: at every cycle the compression brings the state closer to its stationary value, but in some cycles nothing happens. As a consequence, the steady state is unchanged and the convergence time increases. This property holds approximately when \(\gamma\) is non-zero, but very small. We also emphasize that this behavior does not depend on the number of reset qubits.}
{We further note that in combination with dissipation, the second line in Eq.~\eqref{eq:ss} is no longer valid. The dissipation map \(\mathcal{D}\) not only modifies the overall quantum operation, it also introduces an asymmetry between each realization of the imperfect compression. Consider the graph in Fig.~\ref{fig:graph}: In contrast to the undamped case, each branch in this tree, representing the possible fine-grained paths the system can take, leads to its own steady state. The average will constitute of a typical evolution in this branch, and will thus depend on \(\theta\) to the extent that this typical path depends on the weight of the probability distribution \(p\). For \({\gamma = 0}\), the majority of branches will asymptotically consist of compressions and only a single branch, the uppermost one, consists of only identity operations.}
{\section{Experimental details}}
{This section provides additional details about the experimental implementation of the three-qubit heat bath algorithmic cooling refrigerator using a system of a NV center in diamond \cite{{zai21}}.}
{
The experimental setup consists of a homebuilt confocal
microscope, a permanent magnet for the creation of the external magnetic field
and equipment for electron and nuclear spin manipulation as shown in Fig.~\ref{fig:polgate}. The setup operates at ambient conditions, i.e. room temperature and atmospheric pressure and is used exclusively to work with single NV centers. The diamond sample is embedded into a sapphire waver of 2 mm thickness and a diameter of 50 mm. The sapphire waver is mounted on a 3-axis piezoelectric scanner with a travel range of 100 \(\mu\)m x 100 \(\mu\)m x 25 \(\mu\)m and subnanometer resolution.}
\begin{figure}[t]
\centering
\includegraphics[width=0.54\textwidth]{S5.png}
\caption{{Sketch of the experimental setup. The setup consists of a homebuilt confocal microscope, a permanent magnet and microwave (MW) and radio frequency
(RF) sources. The 520nm laser is operated at a power close to NV center saturation (0.1mW to 0.5mW before the objective). An additional 637nm laser is used for electron spin repolarization (charge state control) and thus has a power of less than 10 \(\mu\)W. O.D is the standard optical detection setup where the fluorescence is filtered by a 650-nm long-pass filter and a 50-\(\mu\)m pinhole, and then detected by a single-photon-counting avalanche photodiode. At the bottom are shown the sample, substrate and the confocal image displaying the location of the NV center.}}
\label{fig:polgate}
\end{figure}
{\subsection{Reset polarization}}
{The SWAP gate used for the reset steps is adapted for the efficient generation of a variable degree of nuclear spin polarization. As compared to the implementation of the traditional SWAP gate using three CNOT gates, here only two CNOT gates are enough. The final electron spin state after application of the SWAP gate is indeed irrelevant, as it only acts as source of polarization and can be easily repolarized with a green or red laser pulse into \(\ket{m_s= 0}\). Therefore, the third controlled rotation is not required and the SWAP gate simplifies to two controlled spin rotations. Furthermore, to achieve variable polarization transfer to the nuclear spins, the second electron controlled nuclear rotation does not necessarily need to cover the full angle \(\theta=\pi\) but can be replaced by a rotation of variable angle, \(R_{y,\theta}\) as shown in Fig.~\ref{fig:polgate}.}
{For the choice of the magnetic field (\(540\) mT) used in the experiment, direct optical nuclear spin polarization due to GSLAC and ESLAC is not possible as it requires much lower fields (\(\sim 50-100\) mT). The choice for such large fields is to achieve high fidelity single shot readout of the nuclear spins, by improving the nuclear spin life-time that scales quadratically with the field strength \cite{SSR}. The \(^{14}\)N nuclear spin life time reaches close to a ms at such field strengths.}
\begin{figure}[t]
\centering
\includegraphics[width=0.59\textwidth]{S6.pdf}
\vspace{-15mm}
\caption{{Pulse sequence for variable degree polarization transfer from electron spin to the two nuclear spins used in the experiment. The
electron initially is in state \(\ket{m_s = 0}\), while the target 13C nuclear spin is in a fully mixed state. To remove any remaining polarization, before the polarization step, a \(50 \mu \)s long \(\frac{\pi}{2}\)-pulse is performed on the nuclear spin. The actual polarization transfer part of the sequence consists of a \(80 \mu\)s red laser pulse for electron reset, a nuclear spin
controlled electron \(\pi\)-pulse (\(6 \mu \text{s}/20 \mu \)s for \({}^{13}C_1/{}^{13}C_2\)) and an electron spin controlled nuclear spin rotation of variable duration (\(0 \mu\)s to \(100 \mu\)s). To increase the nuclear
spin polarization, the polarization transfer part can be repeated
\(k\)-times. Finally, the spin state is read out with single-shot readout (SSR). The experiment was performed for angles \(\theta\) between \(0\) and \(2\pi\).}}
\label{fig:circuit}
\end{figure}
{\subsection{Gate implementation}}
{The total gate duration of cooling operation \(U\) is \(\sim 284 \mu\)s. An optimal pulse-duration for the nuclear spin gates was chosen to be around \(50\mu\)s to omit heating of the sample due to the
large RF power and to omit crosstalk to other nuclear spin transitions. The electron spin controlled nuclear spin phase gates do not change the state of the electron spin, thereby avoiding any decoupling errors during the gate operation. Furthermore, the electron spin state remaining in state \(\ket{m_s= 0}\) during the long nuclear spin operations will preserve its coherence over the electron spin relaxation timescales of \(T_{1e}\sim 5.7\) ms. The electron spin \(2\pi\)-pulses take at total duration of \(84 \mu\)s \cite{zai21}. While the coherences decay on a timescale of \(T^\text{Hahn}_{2,e} \sim395 \mu\)s. The electron spin gates were optimized with help of the optimal control platform DYNAMO \cite{dynamo} to realize fast and robust Hahn gates despite electron decoherence on timescales of \(T^\text{Hahn}_{2,e}\) and a dense electron spin spectrum \cite{QEC}.}\\
\section{Experimental determination of the thermodynamic performance}
We evaluate the heat from the set of qubit polarizations \(\epsilon{1}(n)\) measured using single-shot readout \cite{zai21}. Using the definition of the heat \(Q(n)\) given in the main text, we concretely have
\begin{equation}
Q_\text{exp}(n) = - [ \epsilon{1}(n+1) - \epsilon{1}(n) ] .
\end{equation}
The cooling power \(J_\text{exp}(n)\) follows from the finite difference \(Q_\text{exp}(n+1) - Q_\text{exp}(n)\).
Since the states \(\tilde \rho_i(n)\) of the reset qubits after the compression stage are entirely determined by the target qubit polarization \(\epsilon{1}(n-1)\) via Eq.~\eqref{eq:bath-qubit}, the work \(W(n)\) may be directly evaluated from the target polarization data (without having to measure the polarizations of the target qubits) and a non-cumulative version of Eq.~\eqref{16}, where the dependence on \(n\) is implicit in using \(\epsilon{1}(n)\) instead of \(\epsilon{1}(0)\). That is,
\begin{equation}
W_\text{exp}(n) = \sin^2(\theta) \big[ \gamma (\epsilon{2}(0) \epsilon{3}(0) + 1) + (\gamma - 1) (\epsilon{2}(0) \epsilon{3}(0) + 1) \epsilon{1}(n) + \epsilon{2}(0) + \epsilon{3}(0) \big]
+ \epsilon{1}(n) - \epsilon{1}(n+1) .
\end{equation}
We use the experimentally obtained values of \(\epsilon{1}(n)\) to determine the work using the above relation.
\begin{figure}[t]
\centering
\includegraphics{S7.pdf}
\caption{
Carnot coefficient of performance (COP) \(\zeta_\text{C}(n)\) as a function of the number \(n\) of cycles, for experimentally observed target qubit polarizations, and reset qubit polarization \(\epsilon{2}(0) {\sim} 0.58\).
}
\label{fig:expcarnot}
\end{figure}
Using the single-shot readout errors of the three qubits polarizations,
of \({\sim}3\%\) for \(\epsilon{1}(n)\) and \(\epsilon{2}(0)\), and \({\sim}1\%\) for \(\epsilon{3}(0)\), we estimate the relative error of \(W_\text{exp}(n)\) as the standard deviation calculated through conventional linear propagation methods. Likewise, we estimate the relative error of \(Q_\text{exp}(n)\) to be bounded by \({\sim}3\%\) \cite{zai21,qft}. The errors bars for \(\zeta_\text{exp}(n)\) and \(J_\text{exp}(n)\) follow from the definition of these quantities.
Figure~\ref{fig:expcarnot} additionally shows the experimental Carnot COP \({\zeta_\text{C}}_\text{exp}(n)\) and the corresponding theoretical expectation \({\zeta_\text{C}}(n)\) for \(\gamma = 10^{-4}\) and \(\theta = \pi/3.4\). The latter are determined uniquely through the temperatures associated to the two polarizations \(\epsilon{1}(n)\) and \(\epsilon{2}(0)\).
\end{document}
|
\section{Introduction}
Intersections are the most complicated scenarios in urban traffic, where traffic jams and vehicle collisions frequently occur~\cite{azimi2014stip}. With the rapid development of vehicle-to-everything (V2X) technology, the central coordinator deployed at an intersection can guide connected and automated vehicles (CAVs) through it. This guarantees the high-efficiency and conflict-free cooperation of CAVs~\cite{chen2020mixed}. After receiving the scheduled arrival time from the coordinator, each CAV optimizes its speed trajectory in pursuit of high traffic efficiency and low fuel consumption~\cite{xu2017v2i, guler2014using}. Several methods have been proposed to solve the longitudinal control problems of CAVs, including model predictive control~\cite{asadi2011predictive,yang2017eco}, fuzzy logic~\cite{milanes2009controller,onieva2015multi}, and optimal control~\cite{li2015eco,jiang2017eco}.
Apart from vehicle control, the CAV scheduling problem is also frequently discussed in research on intersections. CAVs approach from different directions with different destinations; thus, their trajectories inevitably intersect in the middle of the intersection. Therefore, staggering the arrival times of CAVs is an essential functionality of intersection management. Previous studies have also found that scheduling CAVs is the key factor influencing the traffic efficiency at intersections~\cite{li2006cooperative, meng2017analysis}. To schedule the CAVs approaching intersections, the most straightforward method is a first-in-first-out (FIFO) strategy, wherein the CAVs that enter first are scheduled to leave the intersection first~\cite{malikopoulos2018decentralized, dresner2008multiagent}. Similar concepts are applied in reservation-based~\cite{dresner2004multiagent}, batch-based~\cite{tachet2016revisiting}, and platoon-based methods~\cite{jin2013platoon}.
Evidently, these ad hoc scheduling methods have low computational burden; however, they are less likely to obtain a high-efficiency scheduling plan. Therefore, other studies have also proposed optimization-based methods to solve the scheduling problem. Several researchers formulated this problem as a mixed integer program (MIP) problem~\cite{lu2018mixed, wang2020mixed, ge2021real}. Other methods, such as Monte Carlo tree search~\cite{xu2019cooperative}, dynamic programming~\cite{yan2009autonomous}, and minimum clique cover (MCC)~\cite{chen2021conflict}, were also proposed to schedule CAVs. It is worth mentioning that because scheduling CAVs is a discrete problem rather than a continuous one, a graph-based method is another promising method to solve this problem. Apart from the depth-first spanning tree algorithm proposed in~\cite{xu2018distributed}, a Petri net~\cite{lin2019graph} and a conflict duration graph~\cite{deng2020conflict} have also been used in modeling the scheduling problem. Although these studies widely investigated the CAV scheduling problem, the intersection scenarios remained limited to where lane changing was prohibited.
In the aforementioned studies on intersections, the CAVs were assumed to run in their target lanes,~\textit{i}.\textit{e}., only their longitudinal control was considered. In practice, however, CAVs approach from random lanes and have different target lanes; therefore, it is necessary to extend this research to scenarios that permit lane changing. Earlier studies on this topic focused on obtaining a smooth CAV speed trajectory~\cite{li2005cooperative, li2007cooperative}; however, traffic efficiency was not investigated completely. With regard to CAV scheduling, which we are concerned with, a few studies formulated this problem as an MIP~\cite{lu2019trajectory} or a linear programming problem~\cite{hu2019trajectory}. This was done by assuming that lane changing maneuvers are accomplished in a given time interval. \cite{xu2020bi} proposed a practical bi-level framework, where the high-efficiency arrival plans and collision-free path planning are solved on the upper and lower levels separately. Several other prospective studies focused on changing lane directions dynamically rather than allocating CAVs to constant directional lanes~\cite{he2018erasing, cai2021multi, mitrovic2019combined}. However, flexible lane directions are unsuitable for a mixed traffic environment, where human-driven vehicles (HDVs) and CAVs coexist.
In summary, most of the existing studies were conducted in scenarios where lane changing was prohibited. Therefore, only CAV scheduling or vehicle longitudinal control were studied. In this study, we focus on a scenario where lane changing is permitted. CAVs are required to change to their target lanes on approaching the intersection. Most of the aforementioned studies circumvent this problem by assuming that the CAVs are running in their target lanes~\cite{li2006cooperative, malikopoulos2018decentralized, dresner2004multiagent, tachet2016revisiting, xu2018distributed}. A few studies considered lane changing behavior in their scheduling algorithm~\cite{lu2019trajectory,hu2019trajectory,xu2020bi}; however, the optimality of the scheduling was not extensively discussed. Moreover, although CAV cooperation control at intersections has undeniable advantages in terms of collision avoidance, constant traffic signal phase and timing (SPAT) control also has advantages in terms of vehicle evacuation, especially for large traffic volumes. In this respect, the comparison between CAV cooperation control and constant traffic SPAT remains deficient.
Thus, the main contributions of this study are as follows. We propose a two-stage cooperation framework to decouple the longitudinal and lateral control of CAVs. In the first stage, a formation control method is used to guide the CAVs into their target lanes. An iterative framework is developed to solve the multi-vehicle target assignment and path planning problem, preventing the deadlock problem in the single vehicle lane changing algorithm. In the second stage, the CAV arrival time is optimized to increase the traffic efficiency. Specifically, a heuristic graph-based solution is proposed to solve the scheduling problem with low computational burden. As opposed to specially considering lane changing during algorithm design~\cite{lu2019trajectory,hu2019trajectory,xu2020bi}, the decoupling framework has an improved generalization ability in merging with other scheduling methods. In simulations, we compare the proposed algorithm with constant traffic SPAT, and our results verify the effectiveness of the proposed algorithm.
The remainder of this paper is organized as follows. Section~\ref{Sec:ProblemStatement} describes the scenario and introduces the two-stage cooperation framework. Section~\ref{Sec:StageOne} explains the multi-vehicle target assignment and path planning method. Section~\ref{Sec:StageTwo} presents the graph-based CAV scheduling method. The simulation results are discussed in Section~\ref{Sec:Simulation}, and Section~\ref{Sec:Conclusion} concludes the paper.
\section{Two-stage cooperation framework}
\label{Sec:ProblemStatement}
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{Traffic_Scenario}
\caption{Illustration of the traffic scenario. Vehicles are colored in red, black, and blue to represent turning left, going straight, and turning right, respectively. Red circles, orange squares, and green arrows represent different types of potential collision points.}
\label{fig:Intersection}
\end{figure}
In~\cite{chen2021conflict}, we introduced a graph-based minimum clique cover (MCC) method applied to an intersection with complex vehicle conflict relationships, as shown in Fig.~\ref{fig:Intersection}. In this previous study, changing lanes was prohibited to guarantee vehicle safety,~\textit{i}.\textit{e}., the CAVs are assumed to be in their target lane from the beginning. In this study, the same intersection is used as the traffic scenario; however, lane change is permitted. The CAVs have different destinations; therefore, it is apparent that a scenario where lane changing is permitted is more realistic. Moreover, two assumptions remain to be clarified.
\begin{assumption}
\label{Ass:Com}
CAVs transmit their velocities and positions to the central cloud coordinator through ideal wireless communication, ~\textit{e}.\textit{g}., V2I communication~\cite{gerla2014internet}, where communication delay and packet loss do not occur.
\end{assumption}
\begin{assumption}
\label{Ass:Auto}
The CAVs are capable of fully autonomous driving, implying that they are assumed to have perfect steering performance in their lane changing and turning behavior.
\end{assumption}
\begin{figure*}[!t]
\centering
\includegraphics[width=\linewidth]{Road_Segmentation}
\caption{Lane changing and car-following zones. When arriving at the lane changing zone, the CAVs perform lane change maneuvers to arrive at their target lanes. In the car-following zone, the CAVs modify their car-following distances to arrive at the stopping lane according to the scheduled plan.}
\label{fig:RoadSegmentation}
\end{figure*}
Considering the limit range of V2X communication, previous studies usually set a certain range of the \emph{control zone} to design cooperation algorithms~\cite{chen2020mixed, xu2017v2i, chen2021conflict, xu2018distributed}. In our study, we aim to decouple the longitudinal and lateral control problem in CAVs; therefore, the \emph{control zone} is divided into a \emph{lane changing zone} and a \emph{car-following zone}, as shown in Fig.~\ref{fig:RoadSegmentation}, whose lengths are denoted as $ L_{\mathrm{ctrl}} $, $ L_{\mathrm{LCZ}} $, and $ L_{\mathrm{CFZ}} $, respectively.
Most of the existing studies only considered the trajectory planning of CAVs,~\textit{i}.\textit{e}., their longitudinal control, implying that their target lane selection and lane changing behavior was neglected in intersection management. In this study, we propose a two-stage cooperation framework to decouple the longitudinal and lateral control of CAVs, as shown in Fig.~\ref{fig:RoadSegmentation}. The lane changing behavior is strongly related to the surrounding CAVs, and an \emph{edge coordinator} is deployed at the lane changing zone to coordinate the movement of the CAVs to their target lanes. When entering the lane changing zone, task assignment is first accomplished to allocate the preferred lanes to the CAVs. We then develop multi-vehicle path planning to guarantee a collision-free scenario during lane changing. Therefore, when exiting the lane changing zone, all CAVs are aligned to their target positions. As the CAVs enter the car-following zone, the \emph{central coordinator} collects the information from all of them at the intersection and generates an optimal arrival time plan to increase the traffic efficiency. Subsequently, the CAVs modify their car-following distances according to the scheduling result in the car-following zone. Owing to the arrival time of the CAVs being staggered, each CAV travels through the intersection without idling at the stopping line, and thus, vehicle idling is avoided. Notably, in Fig.~\ref{fig:RoadSegmentation}, the lane lines in the lane changing zone are white dashed lines while those in the car-following zone are white solid lines. This implies that lane change is only permitted in the lane changing zone, which corresponds to the current traffic rules.
\begin{remark}
Decoupling the longitudinal and lateral control of CAVs into two stages has the following advantages. First, many studies have been conducted on the cooperative lane changing behavior of CAVs on straight roads and also arrival time scheduling at intersections. Solving these aforementioned problems separately makes it feasible for us to inherit from these existing studies. Second, although we assume ideal wireless communication in Assumption~\ref{Ass:Com} as in most studies in this field, the effective communication range of the wireless device should not be neglected entirely. Previous experimental tests of the dedicated short-range communications (DSRC) device~\cite{xu2017dsrc} have proved that when the communication distance exceeds $ 500 \mathrm{m} $ and the vehicle velocity is $ 120 \mathrm{km/h}$, the average packet loss rate exceeds $ 20\% $ and the average round-trip time delay exceeds $ 25\mathrm{ms} $. In previous studies, the $ L_{\mathrm{LCZ}} $ and $ L_{\mathrm{CFZ}} $ were usually set to $ 400 - 1000 \mathrm{m}$. Thus, two types of coordinators are deployed to realize intersection management.
\end{remark}
\section{Multi-vehicle Target Assignment and Path Planning}
\label{Sec:StageOne}
In~\cite{xu2021coordinated}, Cai~\textit{et al}.~ proposed a relative coordinate system (RCS) to realize the formation control of CAVs. In~\cite{cai2021formation}, a bi-level conflict-based search (CBS) was further deployed to solve the multi-vehicle collision-free path planning problem. The original methods were applied to a straight road scenario to increase traffic efficiency. We propose that similar methods can be employed to solve the lane changing problem of CAVs.
\subsection{Relative Coordinate System}
\begin{figure*}[!t]
\centering
\includegraphics[width=0.8\linewidth]{Relative_Coordinate_System}
\caption{Relative coordinate system (RCS). When entering the lane changing zone, the approaching CAVs are aligned to their nearest coordinates in the RCS. In this zone, the Hungarian algorithm is used to generate the target position assignment and conflict-based search (CBS) is used to find collision-free paths for the CAVs. When exiting the lane changing zone, each CAV adjusts to its target destination.}
\label{fig:RelativeCoordinateSystem}
\end{figure*}
First, we introduce the RCS in the formation control method. As shown in Fig.~\ref{fig:RelativeCoordinateSystem}, when CAVs arrive at the starting point of the lane changing zone, they are not likely to be in their desired lanes. Hence, the RCS is used to perform the lane change task for the CAVs. Notably, there are two commonly observed formation geometric structures in formation control,~\textit{i}.\textit{e}., the parallel and interlaced structures~\cite{kato2002vehicle, marjovi2015distributed}. In~\cite{cai2021formation}, the interlaced structure was selected as the occupation geometric structure to increase the flexibility in multi-vehicle formation coordination. However, in this study, there exists a car-following zone where we adjust the car-following distance. Therefore, we focus on a feasible lane change solution in the lane changing zone. Thus, the parallel structure is selected, where each point in the RCS can be occupied.
When developing conflict-free path planning for a group of CAVs, they are first assigned to the closest points in the RCS. If the two-dimensional position of the CAV $ i $ is $ (x_{i}, \; y_{i}) $, its relative coordinates in the RCS, $ (x^{r}_{i}, \; y^{r}_{i}) $, are determined using
\begin{equation}
\begin{aligned}
\min \; &(x_{i} - x^{r}_{i})^2 + (y_{i} - y^{r}_{i})^2,\\
\mathrm{subject\; to: }\; & x^{r}_{i} \in \mathbb{N}, \; y^{r}_{i} \in \mathbb{N}.
\end{aligned}
\end{equation}
\subsection{Vehicle-target Assignment}
\label{Sec:Hungary}
After assigning the CAVs to the primary positions in the RCS, their target positions are also assigned based on their destinations. Notably, we use a parallel occupation geometric structure; thus, the CAVs fully occupy the RCS as shown in the car-following zone in Fig.~\ref{fig:RelativeCoordinateSystem}. If there are $ N $ CAVs at primary positions $ (x^{r}_{i}, \; y^{r}_{i}), \; i \leq N, \; i \in \mathbb{N}^+$, there also exist $ N $ target positions $ (x^{t}_{j}, \; y^{t}_{j}), \; j \leq N, \; j \in \mathbb{N}^+ $. Every CAV $ i $ should be assigned to a target position $ j $.
Before assigning the target positions to the CAVs, the cost of this assignment should be defined. We use the distance from each primary position to each target position as the cost for one CAV, where the Euclidean distance is used as the evaluation index. Hence, we can obtain the cost matrix $ {\mathcal{C}} $ as follows:
\begin{equation}
\begin{aligned}
{\mathcal{C}} & = \left[c_{ij}\right] \in \mathbb{R}^{N \times N},\; i,j \in \mathbb{N}^{+},\\
c_{ij} & = \sqrt{(x^{t}_{k} - x^{r}_{i}) ^{2} + (y^{t}_{k} - y^{r}_{i})},
\end{aligned}
\end{equation}
where the element in the $ i $-th row and $ j $-th column represents the cost to assign vehicle $ i $ to target $ j $.
Each CAV has a preferred lane; therefore, the preference matrix $ {\mathcal{L}} $ is obtained to define the CAV preferences.
\begin{equation}
\begin{array}{c}
{\mathcal{L}} = \left[l_{ij}\right] \in \mathbb{R}^{N \times N},\; i,j \in \mathbb{N}^{+}, \\
l_{ij} = \left\{
\begin{array}{rl}
1, &\text{if vehicle $ i $ can be assigned to target $ j $,} \\
M, &\text{otherwise,}
\end{array}\right.
\end{array}
\end{equation}
where the element in the $ i $-th row and $ j $-th column represents whether vehicle $ i $ can be assigned to target $ j $ considering their preferred lane. $ M $ is a positive number sufficiently large to prevent vehicle $ i $ from being assigned to target $ j $.
The assignment matrix $ {\mathcal{A}} $, whose element in the $ i $-th row and $ j $-th column represents whether vehicle $ i $ is assigned to target $ j $, is defined as follows:
\begin{equation}
\begin{array}{c}
{\mathcal{A}} = \left[a_{ij}\right] \in \mathbb{R}^{N \times N},\; i,j \in \mathbb{N}^{+}, \\
a_{ij} = \left\{
\begin{array}{rl}
1, &\text{if vehicle $ i $ is assigned to target $ j $,} \\
0, &\text{otherwise.}
\end{array}\right.
\end{array}
\end{equation}
Subsequently, the assignment problem can be modeled as a 0-1 integer programming problem:
\begin{equation}
\label{equ:AssignmentProblem}
\begin{aligned}
\min \; & \sum_{i=1}^{N}\sum_{j=1}^{N} \left(c_{ij} \times l_{ij} \times a_{ij} \right), \\
\mathrm{subject\; to: }\; & \sum_{i=1}^{N} a_{ij} = 1, \\
& \sum_{j=1}^{N} a_{ij} = 1, \\
& i,j \in \mathbb{N}^{+},
\end{aligned}
\end{equation}
where $ N $ is the number of CAVs, $ c_{ij} $ is the cost matrix, $ l_{ij} $ is the preference matrix, and $ a_{ij} $ is the assignment result.
The Hungarian algorithm~\cite{kuhn1955hungarian} is commonly used to solve assignment problems such as~\eqref{equ:AssignmentProblem}, and we employ this algorithm in our method to generate a feasible assignment with the lowest cost. In the following section, for each group of CAVs, we aim to obtain not only the best assignment but also several sub-optimal assignments. For simplicity, we use $ A_{k} $ to represent the $ k $-th optimal assignment $ {\mathcal{A}} $ and $ C_{k} $ to denote the corresponding cost $ {\mathcal{C}} $.
\subsection{Conflict Types in Lane Changing Behavior}
\begin{figure}[!t]
\centering
\subcaptionbox{Node Conflict \uppercase\expandafter{\romannumeral1} \label{fig:Node_Conflict}}
{\includegraphics[width=0.45\linewidth]{Node_Conflict}}
\subcaptionbox{Node Conflict \uppercase\expandafter{\romannumeral2} \label{fig:Node_Conflict2}}
{\includegraphics[width=0.45\linewidth]{Node_Conflict2}}
\subcaptionbox{Edge Conflict \label{fig:Edge_Conflict}}
{\includegraphics[width=0.45\linewidth]{Edge_Conflict}}
\subcaptionbox{Intermediate Conflict \label{fig:Intermediate_Conflict}}
{\includegraphics[width=0.45\linewidth]{Intermediate_Conflict}}
\caption{Conflict types of CAVs in lane change behavior.}
\label{fig:ConflictRelationships}
\end{figure}
After obtaining the assignment, we need to develop path planning for the CAVs, which should not have conflicts in their trajectories. First, the collision types in the lane changing behavior of CAVs should be clarified. Notably, we regulate the CAVs to move only along orthogonal directions in the RCS,~\textit{i}.\textit{e}., movement along diagonal directions in the RCS is prohibited. Movement along diagonal directions in the RCS produces large acceleration and deceleration in lane change behavior, which is hazardous and also increases the computational burden in multi-vehicle path planning. Hence, we constrain the movement of the CAVs to orthogonal directions in the RCS for simplicity.
The conflict types in the path planning of the CAVs are shown in Fig.~\ref{fig:ConflictRelationships}. The first type is the node conflict. As shown in Fig.~\ref{fig:Node_Conflict}, if two CAVs are scheduled into the same coordinates in the RCS, they will inevitably collide. A similar situation occurs in Fig.~\ref{fig:Node_Conflict2}, where one CAV is scheduled into the coordinates of another. The second type is the edge conflict, as shown in Fig.~\ref{fig:Edge_Conflict}, where the scheduled coordinates of two CAVs are interchanged. The third conflict is the intermediate conflict, as shown in Fig.~\ref{fig:Intermediate_Conflict}. In this scenario, one CAV is scheduled to the coordinates of another when the second CAV is on the verge of moving away. Unlike the previous two conflicts, the occurrence of an intermediate conflict depends on the size of the vehicle and the duration of the lane change behavior. For safety concerns, we include this type of conflict in our further analysis.
\subsection{Collision-free Path Planning}
Studies on multi-agent coordination have thoroughly investigated the multi-agent path planning problem. One of the most well-known methods is the CBS ~\cite{sharon2015conflict}, which constructs a constraint tree to obtain the optimal solution in the same manner as single-agent path planning. Other multi-agent path planning methods have also been proposed to solve the problem,~\textit{e}.\textit{g}., cooperative path planning~\cite{silver2005cooperative}, the swap method~\cite{luna2011push}, and $ M^{*} $~\cite{wagner2011m}. The primary task of the CAVs in the lane changing zone in this study is to successfully change their lanes to the generated target coordinates. Hence, in this study, the CBS is used to solve the multi-vehicle path planning problem.
$ A^{*} $~\cite{hart1968formal} is a well-known path planning algorithm for a single agent. The input of $ A^{*} $ is the starting point, ending point, and obstacle points. The output is a feasible path from the starting point to the ending point, which avoids the obstacles. Single-agent path planning is not the focus of this study; therefore, we omit the details of the $ A^{*} $ algorithm (hereinafter referred to as A-STAR). Interested readers may refer to~\cite{hart1968formal} for further details.
For each assignment $ A_{k} $ of the $ k $-th iteration, we define $ P_{k} $ as the corresponding collision-free path set. $ P_{k} $ is written as
\begin{equation}
\begin{aligned}
P_{k} &= [p_{it}]\in \mathbb{R}^{N \times T},\; i \in \mathbb{N}^{+},t \in \mathbb{N}, \\
p_{it} &= (x^{r}_{it}, y^{r}_{it}),
\end{aligned}
\end{equation}
where $ (x^{r}_{it}, y^{r}_{it}) $ are the coordinates of CAV $ i $ in the RCS at time $ t $. Notably, there are $ N $ CAVs in $ A_{i} $; therefore, there are $ N $ rows in $ P_{i} $ representing $ N $ paths. Considering the limited length of the lane changing zone, the maximum number of steps is defined as $ T $. For simplicity, we use $ p_{i} $ to represent the planning path for CAV $ i $ in $ T $ time steps and $ p_{t} $ to represent the coordinates of all $ N $ CAVs at time step $ t $.
The collision-free path set $ P_{k} $ should comply with the following rules. First, the starting points $ (x^{r}_{i0}, y^{r}_{i0}) $ and ending points $ (x^{r}_{iT}, y^{r}_{iT}) $ of CAV $ i $ should correspond to the assignment in $ A_{k} $. Second, at each time step $ t $, the coordinate set $ p_{t} $ should contain no conflict relationships, as shown in Fig.~\ref{fig:ConflictRelationships}. To solve the multi-vehicle path planning problem, the CBS algorithm is proposed in Algorithm~\ref{algo:CBS}.
\begin{algorithm}[t]
\caption{Conflict-based Search}
\label{algo:CBS}
\begin{algorithmic}[1]
\Require{The $ k $-th assignment $ A_{k} $ for a group of $ N $ CAVs}
\Ensure{Collision-free planning path set $ P_{k} $}
\State Initialize conflict point set $ F_{i} = \varnothing, \, i \in \mathbb{N}^{+} $
\While{True}
\For{each CAV $ i $ in assignment $ A_{k} $}
\State $ p_{i} $ = A-STAR($ (x^{r}_{i0}, y^{r}_{i0}) $, $ (x^{r}_{iT}, y^{r}_{iT}) $, $ F_{i} $)
\State update the path $ p_{i} $ in $ P_{k} $
\EndFor
\If{there exists conflict $ f $ for CAV $ i \in \mathbb{N}^{+}$ in $ P_{k} $}
\State $ F_{i} = F_{i} \cup f $
\Else
\State \Return $ P_{k} $
\EndIf
\EndWhile
\end{algorithmic}
\end{algorithm}
The essential part of the CBS algorithm is examining the conflicts $ f $ in the current planning path set $ P_{k} $, adding them into the corresponding conflict point set $ F_{i} $ for CAV $ i $, and executing the next round of path planning. The iteration ends when there exist no conflicts in $ P_{k} $,~\textit{i}.\textit{e}., the planning path set $ P_{k} $ is collision-free. It can be argued that this method may not find the collision-free path planning solution for a large number of CAVs. However, the feasibility of the algorithm also depends on the maximum number of steps. In the worst case, provided that the maximum time step $ T $ is sufficiently large, the CAVs can always find a collision-free path planning solution. Notably, because the assignment $ A_{k} $ does not consider the conflict while the planning path $ P_{k} $ is collision-free, the cost of assignment $ A_{k} $ $ C_{k} $ never exceeds that of the planning path $ P_{k} $, which is denoted as $ C'_{k} $. We interpret that this non-decreasing property of the CBS guarantees the optimality of the algorithm.
\subsection{Iterative Solution for Assignment and Path Planning}
\label{Sec:PathPlanning}
\begin{figure*}[!t]
\centering
\subcaptionbox{Algorithm interpretation\label{fig:CBS_algorithm}}
{\includegraphics[width=0.52\linewidth]{CBS_algorithm}}
\subcaptionbox{Algorithm flow chart \label{fig:CBS_flow}}
{\includegraphics[width=0.45\linewidth]{CBS_flow}}
\caption{Iterative framework for solving the target assignment and path planning problem. Yellow blocks represent target assignment sections. Green blocks stand for path planning sections. Purple ones are the optimality check process.}
\label{fig:ConflictBasedSearching}
\end{figure*}
In the previous sections, we have shown how we generate the best assignment $ A_{1} $ according to the initial positions and target lanes of the CAVs. We have also introduced how we obtain a collision-free planning path set $ P_{1} $ from the best optimal assignment $ A_{1} $. However, $ P_{1} $ is the local optimal solution generated from $ A_{1} $. We have not checked all the assignments; therefore, a global optimal solution has not been found. Thus, we propose an iterative framework to find this global optimal solution, as shown in Fig.~\ref{fig:ConflictBasedSearching}. In the figure, we use the colors yellow to illustrate the part of the target assignment solved using the Hungarian algorithm (Section~\ref{Sec:Hungary}) and green to denote the part of the multi-vehicle path planning solved using A*-based CBS (Section~\ref{Sec:PathPlanning}).
As shown in Fig.~\ref{fig:CBS_flow}, the CAVs are first assigned to their target coordinates. The Hungarian algorithm provides the optimal assignment $ A_{1} $ with the lowest cost $ C_{1} $,~\textit{i}.\textit{e}., each CAV is assigned to unique target coordinates based on their permitted lanes with the shortest total distance. We then develop the initial path planning for each CAV and obtain the initial collision-free path $ P_{1} $ with the cost $ C_{1}' $. However, we cannot claim that $ P_{1} $ is the global optimal solution because $ P_{k} $, generated using another $ A_{k} $, could have a lower cost. Thus, we generate the second-best assignment $ A_{2} $ using the Hungarian algorithm, which has the lowest cost except for $ A_{1} $. $ A_{2} $ is not the best assignment; therefore, the cost $ C_{1} \geq C_{2} \geq C_{k}, \, k>2 $. It is apparent that if $ C_{2} > C_{1}' $, $ P_{1} $ is the global optimal solution because the path $ C_{2}' $ generated from $ P_{2} $ has a larger cost than $ C_{2} $ and all the remaining assignments also have larger costs. In other circumstances, if $ C_{2} \leq C_{1}' $, we further generate the planning path $ P_{2} $ and the third-best optimal assignment $ A_{3} $, and the iteration continues. The determination of the optimality is colored in purple in Fig.~\ref{fig:ConflictBasedSearching}.
\begin{theorem}
The planning path set $ P_{\min} $ generated in Fig.~\ref{fig:CBS_flow} is the global optimal solution,~\textit{i}.\textit{e}., the cost $ C_{\min}' $ is the lowest cost of the collision-free path set.
\end{theorem}
\begin{IEEEproof}
Without loss of generality, we set $ C_{i}' = \min(C_{1}', C_{2}', C_{3}', \dots, C_{k}') $, implying that the generated planning path set $ P_{\min} = P_{i} $. If there exists a planning path set $ P_{j} $ with a lower cost $ C_{j}' \leq C_{i}' $, it must be generated from the assignment $ A_{j} $ and $ C_{j}'\geq C_{j} $. $ P_{j} $ is not explored in the previous path plannings; therefore, it must be generated from an assignment that has not been explored,~\textit{i}.\textit{e}., $ j > k $. Thus, we have $ C_{j} \geq C_{k} $. From the algorithm, we have $ C_{k} > C_{\min}' = C_{i}' $; therefore, we summarize the inequality equations as $ C_{j}'\geq C_{j} \geq C_{k} \geq C_{i}' $, which contradicts the assumption $ C_{j}' \leq C_{i}' $. Therefore, $ C_{i}' $ is the lowest cost and $ P_{\min} $ is the global optimal solution.
\end{IEEEproof}
\subsection{Vehicle Control}
After obtaining the planning path set $ P_{\min} $, generated from the iterative solution method, each CAV has a collision-free path to its target position. CAV control is not the focus in this study; therefore, we simplify the CAV control process to ease the calculation burden. In Assumption~\ref{Ass:Auto}, we have assumed that the CAV has perfect steering performance, which exempts us from a lateral controller design. Therefore, given adequate time, CAVs are able to execute the planned movement in one step. Hence, a second-order vehicle model is used, and the deviation from the current position of CAV $ i $ to its target position at time $ t $ is defined as follows:
\begin{equation}
\begin{array}{l}
\delta_{\mathrm{p}}^{(i,t)} = x_{i,t} - x^{r}_{it},\\
\delta_{\mathrm{v}}^{(i,t)} = v_{i,t}(t)-v_{\mathrm{p}},
\end{array}
\end{equation}
where $ (x_{i,t}, y_{i,t}) $, and $ v_{i,t}(t) $ represent the position and velocity, respectively, of CAV $ i $ at time $ t $, and $ v_{\mathrm{p}} $ is the designed platoon velocity. A linear feedback controller is designed as follows:
\begin{equation}
\label{equ:Feedback_1}
u_{i}=-k_{p} \delta_{\mathrm{p}}^{(i, j)}-k_{v} \delta_{\mathrm{v}}^{(i, j)}.
\end{equation}
The design of the RCS coordinates and the controller also considers the CAV control in the car-following zone, which is described in Section~\ref{Sec:VehicleControl}. We obtain the lane change timing for each CAV from the planning path set $ P_{\min} $; however, the lane changing behavior is accomplished by the lane changing model in the traffic simulator SUMO~\cite{erdmann2015sumo}. Other key control parameters are listed in Table~\ref{tab:Parameters}.
\section{CAV Scheduling at Unsignalized Intersections}
\label{Sec:StageTwo}
As shown in Fig.~\ref{fig:RoadSegmentation}, the CAVs run in their target lanes after driving through the lane changing zone. The trajectories of the CAVs intersect in the middle of the intersection; thus, in the car-following zone, the CAVs have to schedule collision-free arrival plans. In~\cite{xu2018distributed}, a method using a \emph{virtual platoon} was proposed, which projects the CAVs from different lanes onto a virtual lane. Thus, CAVs from different lanes can drive through the intersection as if they were in the same lane. Therefore, the central coordinator only needs to schedule the CAVs and transmit their target positions in the virtual platoon. Subsequently, CAVs can be controlled in terms of platooning behavior.
\subsection{Vehicle Model}
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{MCC_Scenario}
\caption{Traffic scenario in the car-following zone. Potential conflict points are denoted by red circles, orange squares, and green triangles. CAVs are colored in red, black, or blue based on their different destinations.}
\label{fig:MCC_Scenario}
\end{figure}
The incoming CAVs are indexed from $ 1 $ to $ N $ according to their arrival sequence in the car-following zone, as shown in Fig.~\ref{fig:MCC_Scenario}. For each vehicle $ i \left(i \leq N, i \in \mathbb{N}^+\right) $, the second-order dynamic model is given by
\begin{equation}
\label{equ:StateSpaceEquation}
\left\{
\begin{array}{l}
\dot{x_{i}}(t) = v_{i}(t), \\
\dot{v_{i}}(t) = u_{i}(t).
\end{array} \right.
\end{equation}
$ u_{i}(t) $ represents the input of vehicle $ i $ at time $ t $. There also exist velocity and acceleration constraints on the vehicle.
\begin{equation}
\label{equ:Constraints}
\begin{aligned}
0 &\le v_{i} \le v_{\max}, \\
u_{\min} &\le u_{i} \le u_{\max}.
\end{aligned}
\end{equation}
\subsection{Conflict Analysis}
The CAVs from different lanes form multiple conflict points. Despite the complicated conflict scenarios, they can be classified into the following conflict modes. Without loss of generality, several CAVs are selected in Fig.~\ref{fig:MCC_Scenario} to illustrate the conflict relationship. We define four conflict types in this study.
\begin{enumerate}
\item \emph{Crossing Conflict}: CAVs from different lanes have the potential to collide while crossing the conflict points, indicated by the 24 red circles. For example, CAV $ 2 $ and CAV $ 3 $ have a crossing conflict point.
\item \emph{Diverging Conflict}: The CAVs change lanes in the lane changing zone, and lane changing and passing are not permitted in the car-following zone. Thus, vehicles on the same lane cannot pass the intersection simultaneously, as indicated by the 14 orange squares. For example, CAV $ 5 $ and CAV $ 6 $ have a diverging conflict point.
\item \emph{Converging Conflict}: Vehicles from different lanes cannot drive into the same lane simultaneously, as shown by the six green arrows. For example, CAV $ 1 $ and CAV $ 4 $ have a converging conflict point.
\item \emph{Reachability Conflict}: CAVs cannot pass the intersection simultaneously because of the acceleration and velocity constraints, regardless of whether or not they have the abovementioned conflicts. CAV $ 1 $ and CAV $ 7 $ have a reachability conflict.
\end{enumerate}
The first three conflict types are route conflicts~\cite{roess2004traffic}, where CAVs have intersections in their trajectories along their paths. The fourth conflict type is caused by the velocity and acceleration constraints on the CAVs. For instance, in Fig.~\ref{fig:MCC_Scenario}, CAV $ 7 $ arrives at the car-following zone when CAV $ 1 $ nearly reaches the stop line with the designed virtual platoon velocity. In this case, CAV $ 7 $ cannot catch up with CAV $ 1 $ at the stop line, regardless of whether or not they have any conflict relationships. This is because of the constraints on the vehicle velocity and acceleration.
Expanding equation~\eqref{equ:StateSpaceEquation} and~\eqref{equ:Constraints}, we obtain the evaluation condition as follows:
\begin{equation}
\label{equ:Reachability}
\frac{L_\mathrm{prec}}{v_\mathrm{p}} < \frac{L_\mathrm{CFZ}}{v_{\max}} + \frac{v_{\max}}{2u_{\max}},
\end{equation}
where $ L_\mathrm{prec} $ and $ L_\mathrm{CFZ} $ are the distances from the stop line to the preceding CAV and to the beginning of the car-following zone, respectively, and $ v_\mathrm{p} $ is the designed virtual platoon velocity. Namely, the CAVs that are very close to the intersection should not be considered in the scheduling of new incoming CAVs. Notably, most of the existing studies assumed that the CAVs can reach the stop line under all circumstances,~\textit{i}.\textit{e}., the reachability conflict was ignored.
We define different conflict sets to describe the conflict relationship of the CAVs. For each CAV $ i \left(i \leq N, i \in \mathbb{N}^+\right) $, the crossing, diverging, converging, and reachability sets are defined as $ \mathcal{C}_{i} $, $ \mathcal{D}_{i} $, $ \mathcal{V}_{i} $, and $ \mathcal{R}_{i} $, respectively. Notably, because the conflict sets are determined when the CAV reaches the car-following zone, the CAV indexes in the conflict sets are smaller than those of the CAVs at the border of the car-following zone,~\textit{i}.\textit{e}., the elements in the conflict sets satisfy~\eqref{equ:ElementRelationship}.
\begin{equation}
\label{equ:ElementRelationship}
i < j, \, \mathrm{if}\ i \in \mathcal{C}_{j} \cup \mathcal{D}_{j} \cup \mathcal{V}_{j} \cup \mathcal{R}_{j}.
\end{equation}
\subsection{Description of Graph-based Conflicts}
\label{Sec:ConflictDescription}
Based on the conflict set analysis, we further define a conflict directed graph (CDG) $ \mathcal{G}_{N+1} $ to represent the conflict relationship between the CAVs.
\begin{definition}[Conflict Directed Graph]
\label{def:CDG}
The CDG is denoted as $ \mathcal{G}_{N} = \left( \mathcal{V}_{N},\mathcal{E}_{N} \right) $. If there are $ N $ CAVs in the car-following zone, we have the node set $ \mathcal{V}_{N} = \{1,2,\dots,N\} $. The unidirectional edge set is defined as $ \mathcal{E}_{N}^{u}=\{(i,j) \mid i \in \mathcal{D}_{j} \cup \mathcal{R}_{j}\} $, and the bidirectional edge set is defined as $ \mathcal{E}_{N}^{b}=\{(i,j) \mid i \in \mathcal{C}_{j} \cup \mathcal{V}_{j}\} $. The edge set is the union of these two sets as $ \mathcal{E}_{N} = \mathcal{E}_{N}^{u} \cup \mathcal{E}_{N}^{b} $.
\end{definition}
The CDG of the scenario in Fig.~\ref{fig:MCC_Scenario} is depicted in Fig.~\ref{fig:ConflictDirectedGraph}. The nodes in the CDG represent the CAVs in the car-following zone. The red unidirectional edges represent the diverging and reachability conflicts. The existence of a unidirectional edge $ (i,j) $ implies that CAV $ j $ is not permitted to pass CAV $ i $ or CAV $ j $ is unable to catch up to CAV $ i $ because it satisfies~\eqref{equ:Reachability}. Thus, CAV $ j $ cannot reach the intersection earlier than CAV $ i $. The black bidirectional edges denote the crossing and converging conflicts, implying that the arrival sequences of CAVs $ i $ and $ j $ can be interchanged.
\begin{figure}[!t]
\centering
\subcaptionbox{Conflict Directed Graph\label{fig:ConflictDirectedGraph}}
{\includegraphics[width=0.53\linewidth]{Conflict_Directed_Graph}}
\subcaptionbox{Coexisting Undirected Graph\label{fig:CoexistUndirectedGraph}}
{\includegraphics[width=0.41\linewidth]{Coexist_Undirected_Graph}}
\caption{Fig.~\ref{fig:ConflictDirectedGraph} is the conflict directed graph (CDG). The red unidirectional edges represent the diverging and reachability conflicts, whereas the black bidirectional edges represent the crossing and converging conflicts. Fig~\ref{fig:CoexistUndirectedGraph} is the coexisting undirected graph (CUG), which is the complement graph of the CDG and describes the coexistence relationship of the vehicles.}
\label{fig:ConflictDirectedScenario}
\end{figure}
It is straightforward that the CDG describes all the conflict relationships of the CAVs. From a different perspective, another method to describe the conflict relationships of the CAV is to describe their coexistence relationships.
\begin{definition}[Coexisting Undirected Graph]
\label{def:CUG}
The coexisting undirected graph (CUG) is defined as the complement graph of the CDG $ \mathcal{G}_{N} $. Thus, $ \overline{\mathcal{G}}_{N} = \left(\overline{\mathcal{V}}_{N},\overline{\mathcal{E}}_{N}\right) $, where $ \overline{\mathcal{V}}_{N} = \mathcal{V}_{N} $, $ \overline{\mathcal{E}}_{N} = \{(i,j) \mid i,j \in \overline{\mathcal{V}}_{N} , i \neq j, \mathrm{and} \ (i,j) \notin \mathcal{E}_{N}\} $.
\end{definition}
In this scenario, the CDG is depicted in Fig.~\ref{fig:ConflictDirectedGraph} and the CUG in Fig.~\ref{fig:CoexistUndirectedGraph}. The CDG edge $ \mathcal{E}_{N} $ implies that two CAVs have conflicts and the CUG is the complement graph of the CDG; therefore, the CUG edge $ \overline{\mathcal{V}}_{N} $ implies that the two CAVs are conflict-free,~\textit{i}.\textit{e}., they can pass through the intersection simultaneously.
\subsection{Minimum Clique Cover Scheduling}
Generally, all the scheduling methods at unsignalized intersections are designed to find a high-efficiency collision-free passing order of the CAVs,~\textit{i}.\textit{e}., to schedule the arrival sequence based on their conflict relationships described in Section~\ref{Sec:ConflictDescription}. In this study, we focus on scheduling the CAVs based on the CUG defined in Definition~\ref{def:CUG}. The CUG describes the coexistence of the CAVs, which may pass the intersection simultaneously. In graph theory, a clique is suitable for describing the coexistence relationship of CAVs. The definition of the clique is shown in Definition~\ref{def:clique}.
\begin{definition}[Clique\cite{luce1949method}]
\label{def:clique}
A clique $ C $ in an undirected graph $ G = (V, E) $ is a subset of the nodes, $ C \subseteq V, $ such that every two distinct nodes are adjacent. This is equivalent to the condition that the subgraph of $ G $ induced by $ C $ is a complete graph.
\end{definition}
Considering the cliques in the CUG $ \overline{\mathcal{G}}_{N} $, the CAVs in one clique are conflict-free,~\textit{i}.\textit{e}., they can pass through the intersection simultaneously. Therefore, the objective is to find the minimum number of vehicle groups in the CUG,~\textit{i}.\textit{e}., the minimum number of cliques covering all the nodes in the CUG $ \overline{\mathcal{G}}_{N} $. Therefore, we define the MCC problem as follows.
\begin{definition}[Minimum Clique Cover (MCC)~\cite{karp1972reducibility}]
\label{def:MCC}
A clique cover of a graph $ G=(V,E) $ is a partition of $ V $ into $ k $ disjoint subsets $ V_{1},V_{2},\dots, V_{k}$ such that for $ 1 \leq i \leq k $, the subgraph induced by $ V_{i} $ is a clique,~\textit{i}.\textit{e}., a complete graph. The MCC number of $ G $ is the minimum number of subsets in a clique cover of $ G $, denoted as $ \theta(G) $.
\end{definition}
The MCC number $ \theta(\overline{\mathcal{G}}_{N}) $ of the CUG represents the minimum number of cliques covering it. The cliques in the CUG represent the CAVs that can pass through the intersection simultaneously; therefore, these CAVs in the same clique can be scheduled to simultaneously drive through the intersection. For example, considering the CUG in Fig.~\ref{fig:CoexistUndirectedGraph}, the MCC number $ \theta(\overline{\mathcal{G}}_{N}) = 4 $, and the corresponding cliques are listed in Table~\ref{tab:MCC}. We conclude that for an arbitrary intersection scenario, the coexistence relationship of the incoming CAVs is depicted in the CUG $ \overline{\mathcal{G}}_{N} $. Thus, the MCC number $ \theta(\overline{\mathcal{G}}_{N}) $ represents the possible minimum passing order solution.
\begin{table}[!t]
\renewcommand{\arraystretch}{1.3}
\centering
\begin{tabular}{c|c|c|c|c}
\hline
\diagbox[width=8em]{\textbf{Solutions}}{\textbf{Subsets}} & $ {V}_{1} $ & $ {V}_{2} $ & $ {V}_{3} $ & $ {V}_{4} $ \\
\hline
\bfseries 1 & $ \{1,3,5\} $ & $ \{4,7\} $ & $ \{2\} $ & $ \{6\} $ \\
\bfseries 2 & $ \{1,3,6\} $ & $ \{4,7\} $ & $ \{2\} $ & $ \{5\} $\\
\bfseries 3 & $ \{1,2\} $ & $ \{3,5\} $ & $ \{4,7\} $ & $ \{6\} $\\
\bfseries 4 & $ \{1,2\} $ & $ \{3,6\} $ & $ \{4,7\} $ & $ \{5\} $\\
\bfseries 5 & $ \{1,5\} $ & $ \{3,6\} $ & $ \{4,7\} $ & $ \{2\} $\\
\bfseries 6 & $ \{1,6\} $ & $ \{3,5\} $ & $ \{4,7\} $ & $ \{2\} $\\
\hline
\end{tabular}
\caption{Possible MCC solutions of Fig.~\ref{fig:CoexistUndirectedGraph}. Note that $ \theta(\overline{\mathcal{G}}_{N}) = 4$ in this graph.}
\label{tab:MCC}
\end{table}
\begin{figure}[!t]
\centering
\includegraphics[width=0.7\linewidth]{Spanning_Tree}
\caption{Spanning tree generated by the MCC method. The CAVs in the same layer are collision-free and can pass the intersection simultaneously. Therefore, the overall depth of the spanning tree is $ 4 $, implying that all the CAVs can pass the intersection in the platoon passing time of $ 4 $ vehicles.}
\label{fig:Spanning_Tree}
\end{figure}
If we consider solution $ 1 $ from Table~\ref{tab:MCC} and place CAVs of the same clique into the same layer, a spanning tree $ \mathcal{G}_{N+1}' $ is generated as shown in Fig~\ref{fig:Spanning_Tree}. Node $ 0 $ is the virtual leading vehicle and the CAVs in the same layer are arranged to simultaneously pass the intersection. Notably, $ \theta(\overline{\mathcal{G}}_{N}) = 4 $ in the CUG in Fig.~\ref{fig:CoexistUndirectedGraph}, implying that the spanning trees generated by the MCC solutions have a minimum layer of $ 4 $. This also indicates that the theoretical evacuation times of these solutions are the same,~\textit{i}.\textit{e}., the theoretical values of $ t_\mathrm{evac} $ are the same. Thus, the evaluation index of $ t_\mathrm{evac} $ in~\eqref{equ:EvacuationTime} corresponds to $ \theta(\overline{\mathcal{G}}_{N}) $, which is the global optimal passing order considering the evacuation time. In addition, we further consider the average travel time delay (ATTD) among these solutions as a secondary index. The definition of $ t_\mathrm{ATTD} $ in~\eqref{equ:TravelTimeDelay} can be rewritten in graphical terms as
\begin{equation}
\label{equ:ATTD_MCC}
\begin{aligned}
\min \; & \sum_{i=1}^{k} d_{i}|V_{i}|, \, i \in \mathbb{N}^+,\\
\mathrm{subject\; to: }\; & \sum_{i=1}^{k} V_{i} = N,
\end{aligned}
\end{equation}
where $ V_{i} $ represents the nodes in CUG $ \overline{\mathcal{G}}_{N} $ and $ d_{i} $ is the spanning tree depth of the each CAV node $ V_{i} $.
In this scenario, it is evident that the subsets $ V_{i} $ should be arranged in the descending order to decrease the average $ d_{i} $ of $ N $ CAVs in~\eqref{equ:ATTD_MCC}. For example, in Table~\ref{tab:MCC}, we prefer to choose solution $ 1 $, and the corresponding optimized spanning tree is scheduled as $ \{1,3,5\} \rightarrow \{4,7\} \rightarrow \{2\} \rightarrow \{6\}$, as shown in Fig.~\ref{fig:Spanning_Tree}.
The CUG $ \overline{\mathcal{G}}_{N} $ only contains the coexistence information of the CAVs, whereas the passing order of the CAVs in the same lane should be strictly according to their relative positions. If the MCC method generates solutions that cannot be directly executed,~\textit{e}.\textit{g}., solution $ 2 $ in Table~\ref{tab:MCC}, the spanning tree is generated as $ \{1,3,6\} \rightarrow \{4,7\} \rightarrow \{2\} \rightarrow \{5\} $. However, this solution is not feasible because CAV $ 5 $ is ahead of CAV $ 6 $. We have proved that this can be solved by exchanging the unfeasible sequence~\cite{chen2021conflict}, as shown in Lemma~\ref{Pro:1}.
\begin{lemma}
\label{Pro:1}
If the MCC method leads to an unfeasible solution, where CAVs $ i $ and $ j $ have conflicting trajectories, exchanging the positions of $ i $ and $ j $ solves the conflict. The new solution is also an MCC solution.
\end{lemma}
The MCC problem is proved to be an NP-hard problem, which is difficult to solve during real deployment, especially when the CAV number $ N $ is large. Therefore, we apply a practical approach to solve the problem heuristically, as shown in lines~\ref{algo:MCCBegin} to~\ref{algo:MCCEnd} of Algorithm~\ref{algo:MCC}. The MCC problem of $ G $ is proved to be reduced to the graph coloring problem of $ \overline{G} $~\cite{garey1979guide}, and there are numerous heuristic methods to solve the graph coloring problem. As mentioned earlier, we intend to find the solutions with larger cliques. Thus, we first generate a node sequence $ K=(v_{1},v_{2},\dots,v_{N}) $ using a breadth-first search (BFS). Then, we greedily assign node $ v_{i} $ the smallest possible color,~\textit{i}.\textit{e}., the clique index according to the node sequence $ K $, forming the subset cliques $ V_{1},\dots,V_{k}$. Line~\ref{algo:MCC:Spanning} depicts the spanning process, which arranges the cliques into a spanning tree and has a constant calculation time.
\begin{algorithm}[tb]
\caption{Minimum Clique Cover Method}
\label{algo:MCC}
\begin{algorithmic}[1]
\Require{Coexisting Undirected Graph $\overline{\mathcal{G}}_{N} = \left(\overline{\mathcal{V}}_{N},\overline{\mathcal{E}}_{N}\right) $}
\Ensure{Spanning Tree $ \mathcal{G}_{N+1}' = \left( \mathcal{V}_{N+1},\mathcal{E}_{N+1}' \right) $}
\State Calculate the complement graph $\overline{\overline{\mathcal{G}}}_{N} = \mathcal{G}_{N}$ \label{algo:MCCBegin}
\State Find the breadth-first search sequence $ K=(v_{1},v_{2},\dots,v_{N}) $ of the nodes in $ \mathcal{G}_{N} $
\For{each node $ v_{i} $ of $ \mathcal{G}_{N} $ in the sequence $ K $}
\State assign node $ v_{i} $ the smallest possible clique index
\EndFor \label{algo:MCCEnd}
\State Rank $ V_{1}, V_{2}, \dots,V_{k}$ in the descending order and obtain the spanning $ \mathcal{G}_{N+1}' $ \label{algo:MCC:Spanning}
\State Exchange the conflicting CAVs of $ \mathcal{G}_{N+1}' $ in the same lane if necessary
\State \Return $ \mathcal{G}_{N+1}' $
\end{algorithmic}
\end{algorithm}
\subsection{Vehicle Control}
\label{Sec:VehicleControl}
A virtual platoon is constructed from the spanning tree in Fig.~\ref{fig:Spanning_Tree}; hence, the controller of CAV $ i $ is related to the communication topology design. Owing to the page limit, we only present the basic controller design here. Notions $ a_{ij} $, $ q_{ij} $, and $ l_{ij} $ are related to the designed predecessor--leader following (PLF) topology. Interested readers may refer to~\cite{chen2021conflict} for further details.
First, a union set $ \mathbb{I}_{i} $ is defined to describe the information exchange of CAV $ i $ as follows:
\begin{equation}
\mathbb{I}_{i}=\left\{j \mid a_{i j}=1\right\} \cup \left\{0 \mid q_{i i}=1\right\}.
\end{equation}
The distance and velocity errors are defined as
\begin{equation}
\begin{array}{l}
\delta_{\mathrm{p}}^{(i,j)} = p_{j}(t)-p_{i}(t)-d_\mathrm{f}\left(d_{j}-d_{i}\right) \\
\delta_{\mathrm{v}}^{(i,j)} = v_{i}(t)-v_{j}(t)
\end{array}, \, j \in \mathbb{I}_{i},
\end{equation}
where $ d_\mathrm{f} $ is the car-following distance, $ \delta_{\mathrm{p}}^{(i,j)} $ is the car-following distance error of CAV $ i $, and $ \delta_{\mathrm{v}}^{(i,j)} $ is the car-following velocity error considering all the CAVs in $ \mathbb{I}_{i} $.
A linear feedback controller is designed as follows:
\begin{equation}
\label{equ:Feedback_2}
\begin{aligned}
u_{i}
&=-\sum_{j \in \mathbb{I}_{i}} k_{p} \delta_{\mathrm{p}}^{(i, j)}-\sum_{j \in \mathbb{I}_{i}} k_{v} \delta_{\mathrm{v}}^{(i, j)},
\end{aligned}
\end{equation}
where $ k_{p} $ and $ k_{v} $ are the feedback gains of the distance and velocity errors of CAV $ i $, respectively. The same gains are set for all the CAVs because we consider a homogeneous scenario.
As mentioned earlier, we consider second-order vehicle dynamics as shown in~\eqref{equ:StateSpaceEquation}. We define the car-following errors as the new vehicle state.
\begin{equation}
\bar{\boldsymbol{x}}_{i}=\left[
\begin{array}{c}
\bar{x}_{i, 1} \\
\bar{x}_{i, 2} \\
\end{array}\right]=\left[
\begin{array}{c}
p_{0}-p_{i}-D_\mathrm{des}\left(d_{0}-d_{i}\right) \\
v_{i}-v_{\mathrm{p}} \\
\end{array}\right], i \in \mathbb{N}^{+}.
\end{equation}
The vehicle input remains the same,~\textit{i}.\textit{e}., $ \bar{u}_{i} = u_{i} $.
Therefore, the car-following vehicle dynamic model is
\begin{equation}
\dot{\bar{\boldsymbol{x}}}_{i}=\boldsymbol{A} \bar{\boldsymbol{x}}_{i}+\boldsymbol{B} \bar{u}_{i}, \, i \in \mathbb{N}^{+}.
\end{equation}
The linear feedback controller is simplified to
\begin{equation}
\bar{u}_{i}=-k_{p} \sum_{j}\left(l_{i j}+q_{i j}\right) \bar{x}_{j, 1}-k_{v} \sum_{j}\left(l_{i j}+q_{i j}\right) \bar{x}_{j, 2},\, j \in \mathbb{I}_{i}.
\end{equation}
Defining $ \boldsymbol{k} = \left[k_{p},k_{i}\right]^{T} $, we have
\begin{equation}
\bar{u}_{i}=-\sum_{j} \left(l_{ij}+q_{ij}\right) \boldsymbol{k}^{T}\bar{\boldsymbol{x}}_{i}, \, i \in \mathbb{N}^{+}.
\end{equation}
\begin{table}[!t]
\renewcommand{\arraystretch}{1.3}
\centering
\begin{tabular}{c|c|c|c}
\hline
{\textbf{Types}} & {\textbf{Parameter}} & {\textbf{Symbol}} & {\textbf{Value}} \\
\hline
\multirow{3}{*}{\textbf{Simulation}} & Simulation step & - & $ 0.1 \mathrm{s} $ \\
& Lane changing zone length & $ L_{\mathrm{LCZ}} $ & $ 500 \mathrm{m} $ \\
& Car-following zone length & $ L_{\mathrm{CFZ}} $ & $ 500 \mathrm{m} $ \\
\hline
\multirow{3}{*}{\textbf{RCS}} & Group size & - & $ 3 $ \\
& One-step time & - & $ 4 \mathrm{s}$ \\
& Lane changing time & - & $ 3 \mathrm{s}$ \\
\hline
\multirow{4}{*}{\textbf{Controller}} & Feedback gain of distance error & $ k_{p} $ & 0.1\\
& Feedback gain of velocity error & $ k_{i} $ & $ 0.3 $\\
& Steady platoon velocity & $ v_{\mathrm{p}} $ & $ 10 \mathrm{m/s} $\\
& Steady car-following distance & $ d_\mathrm{f} $ & $ 30 \mathrm{m} $\\
\hline
\multirow{4}{*}{\textbf{Constraints}} & Maximum acceleration & $ a_{\max} $ & $5 \mathrm{m/s}^{2} $ \\
& Minimum acceleration & $ a_{\min} $ & $ -6 \mathrm{m/s}^{2} $ \\
& Maximum velocity & $ v_{\max} $ & $ 15 \mathrm{m/s} $ \\
& Minimum velocity & $ v_{\min} $ & $ 0 \mathrm{m/s} $ \\
\hline
\end{tabular}
\caption{Parameters}
\label{tab:Parameters}
\end{table}
\section{Simulations}
\label{Sec:Simulation}
\subsection{Simulation Environment}
The traffic simulation was conducted in SUMO, which is widely used in traffic research\cite{lopez2018microscopic}. The simulation was run on an Intel Core i7-7700 @3.6 GHz processor. The intersection scenario and lane direction settings are the same as shown in Fig.~\ref{fig:Intersection}, and the overall length of the lane changing and car-following zones is $ 1000 \, m $. The CAV arrival is assumed to be a Poisson distributed flow, given by
\begin{equation}
\label{equ:Poisson}
P(X=k)=\frac{\lambda^{k}}{k!} e^{-\lambda}, k=0,1,\cdots,
\end{equation}
where $ X $ represents the arrival of the vehicle at the control zone. $ \lambda $ is the expected value as well as the variance of the Poisson distribution. Other key simulation parameters are listed in Table~\ref{tab:Parameters}.
\subsection{Evaluation Indexes}
As mentioned earlier, we aim to propose a cooperation method to improve both traffic safety and efficiency,~\textit{i}.\textit{e}., obtain the collision-free optimal CAV passing order. We assign $ t_{i}^{\mathrm{in}} $ as the time step when vehicle $ i $ enters the lane changing zone and $ t_{i}^{\mathrm{out}} $ as the time step when it arrives at the intersection. Several evaluation indexes have been proposed to measure the scheduling performance. In this study, we chose the evacuation time as the primary optimization target, as shown in Definition~\ref{def:EvacuationTime}. The evacuation time is related to the spanning tree depth, which is described in Section~\ref{Sec:StageTwo}.
\begin{definition}[Evacuation Time]
\label{def:EvacuationTime}
The evacuation time of $ N $ CAVs is defined as the time when the last CAV reaches the stop line; it is expressed as
\begin{equation}
\label{equ:EvacuationTime}
t_\mathrm{evc} = \max t_{i}^{\mathrm{out}}, i \le N, i \in \mathbb{N}^+.
\end{equation}
Considering $ N $ incoming CAVs, $ t_\mathrm{evc} $ represents the arrival time of the last CAV at the stop line. For $ N $ CAVs, smaller evacuation times indicate that these CAVs pass through the intersection in a shorter time. Thus, it demonstrates the overall traffic efficiency performance,~\textit{i}.\textit{e}., the overall benefits to the CAVs.
\end{definition}
In addition, the vehicle travels through the control zone in $ t_{i}^{\mathrm{out}} - t_{i}^{\mathrm{in}} $ time, whereas it travels through it under the free driving condition in $ {L_{\mathrm{ctrl}}}/{v_{\max}} $ time. Accordingly, the ATTD is selected to measure the average traffic efficiency of the vehicles, as described in Definition ~\ref{def:ATTD}.
\begin{definition}[Average Travel Time Delay]
\label{def:ATTD}
The ATTD is designed to evaluate the average traffic efficiency of $ N $ CAVs. It is expressed as
\begin{equation}
\label{equ:TravelTimeDelay}
t_\mathrm{ATTD} = \frac{1}{N} \sum_{i = 1}^{N} \left(t_{i}^{\mathrm{out}} - t_{i}^{\mathrm{in}}-\frac{L_{\mathrm{ctrl}}}{v_{\max}}\right),
\end{equation}
where $ L_{\mathrm{ctrl}} $ is the length of the control zone and $ t_\mathrm{ATTD} $ represents the average travel delay of the CAVs. The travel time of every CAV is considered in $ t_\mathrm{ATTD} $; therefore, it denotes the individual benefits for $ N $ CAVs, which is the secondary optimization target of traffic efficiency.
\end{definition}
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{ConstantSPAT}
\caption{Dual ring control in constant traffic SPAT. The green line indicates that the corresponding traffic light is set to green, and the yellow square denotes the clearance time.}
\label{fig:ConstantSPAT}
\end{figure}
\begin{figure*}[t]
\centering
\subcaptionbox{Evacuation time results comparison. \label{fig:VehicleNumber_Duration}}
{\includegraphics[width=\linewidth]{VehicleNumber_Duration}}
\subcaptionbox{ATTD results comparison. \label{fig:VehicleNumber_ATTD}}
{\includegraphics[width=\linewidth]{VehicleNumber_ATTD}}
\caption{Comparison of the traffic evacuation times and ATTDs of the algorithms for different numbers of input vehicles.}
\label{fig:VehicleNumber}
\end{figure*}
As mentioned earlier, the control zone is divided into the lane changing and car-following zones. Although we primarily focus on the feasible path planning solution in the lane changing zone as described in Section~\ref{Sec:StageOne}, here, we continue to measure the algorithm performance over the complete length of the control zone.
\subsection{Benchmark algorithm}
In~\cite{chen2021conflict}, we have proved the optimality of the MCC algorithm in scheduling CAVs. This study further considers lane changing behavior; therefore, we compare our algorithm with the constant traffic SPAT method. The lanes leading to the intersection are divided into the lane changing and car-following zones, as shown in Fig.~\ref{fig:RoadSegmentation}; therefore, two algorithms are applied to each zone.
\subsubsection{Lane changing zone}
The formation control method is used as described in Section~\ref{Sec:StageOne} to obtain a parallel structure formation; therefore, the proposed lane changing zone algorithm is named as the formation-control lane changing (FCLC). As mentioned earlier, SUMO provides a lane changing model~\cite{erdmann2015sumo} to simulate the driver lane changing behavior. We use the default lane changing model as the benchmark algorithm, where the lane change timing is determined by the lane changing model rather than the formation control methods. Namely, the default lane changing model is called Sumo lane changing (SumoLC).
\begin{figure*}[t]
\centering
\subcaptionbox{Evacuation time results comparison. \label{fig:Lambda_Duration}}
{\includegraphics[width=\linewidth]{Lambda_Duration}}
\subcaptionbox{ATTD results comparison. \label{fig:Lambda_ATTD}}
{\includegraphics[width=\linewidth]{Lambda_ATTD}}
\caption{Comparison of the traffic evacuation times and ATTDs of algorithms for different traffic volumes.}
\label{fig:Lambda}
\end{figure*}
\subsubsection{Car-following zone}
As described in Section~\ref{Sec:StageTwo}, an MCC formulation is used to solve the scheduling problem; therefore, the MCC is referred to as the proposed algorithm. Constant traffic light (ConstantTL) is set as a dual ring control~\cite{xu2017v2i}, as shown in Fig~\ref{fig:ConstantSPAT}. In $ t_{1} $, the traffic light is set to green for vehicles turning left from either the east or the west. The yellow square represents the clearance time when the lights turn yellow. $ t_{2} $ represents the vehicles going straight from either the east or the west, and so on. A time of $ 35 \, \mathrm{s} $ is set for each phase, implying that
\begin{equation}
t_{1} = t_{2} = t_{3} = t_{4} = 35 \, \mathrm{s},
\end{equation}
with $ 5 \, \mathrm{s} $ as the clearance time.
Theoretically, the combinations of two lane changing algorithms and two scheduling algorithms lead to four algorithms. However, SumoLC algorithm has its limitation that deadlock problem may occur if approaching lane is not long enough~\cite{erdmann2015sumo}. In our simulation, when vehicle number exceeds $ 40 $ or traffic volume reaches $ 1500 \, \mathrm{veh/h} $, lane changing failure occurs in one third of the results in SumoLC+MCC method. In comparison, deadlock does not happen in SumoLC+ConstantTL method because the control zone $ L_{\mathrm{ctrl}} $ is long enough. Moreover, FCLC guarantees the collision-free paths set is feasible. Similar results have been observed in~\cite{he2018erasing, cai2021formation}. If lane changing fails, CAVs cannot pass the intersection since there exist directional limitations on each lane. Therefore, we excluded this method. The comparison of the other three algorithms still provides convincing proof on the effectiveness of the proposed algorithm.
\subsection{Simulation Results for Different Numbers of Input Vehicles}
The first simulation was conducted for different numbers of vehicles. The scheduling problem is highly related to the distribution of incoming vehicles; therefore, ten iterations of the simulation were conducted for each number of vehicles and for each algorithm. Notably, identical vehicle distributions were applied for the algorithms,~\textit{i}.\textit{e}., ten sets of vehicle distributions were randomly generated for each number of vehicles and used as the same input for the simulated algorithms. Typically, the number of CAVs in one intersection is less than $ 100 $; thus, we set the number of vehicles to range from $ 20 $ to $ 100 $. The traffic volume was set as $ 2000 \, \mathrm{veh/lane} $.
The simulation result is shown in Fig.~\ref{fig:VehicleNumber}. The bar plot shows the average value of the ten sets of simulations. The standard deviations of each set of results are also provided. Generally, both the evacuation time and the ATTD increase with the number of vehicles because of vehicle queuing. It is apparent that the MCC in the lane changing zone significantly improves the traffic efficiency. The evacuation time is the primary optimization target in the MCC scheduling; therefore, the MCC algorithm reduces the evacuation time by $ 15.8\% $ to $ 67.9\% $ for $ 20 $ to $ 100 $ vehicles.
Regarding the ATTD perspective, it is evident that the FCLC negatively influences the ATTD. This is because when SumoLC is applied, CAVs are set to the maximum velocity $ v_{\max} = 15 \, \mathrm{m/s} $ if there are no other vehicles around. However, the FCLC aims to develop a steady parallel formation with a platoon velocity $ v_{\mathrm{p}} = 10 \, \mathrm{m/s} $. Therefore, the total evacuation time is unaffected when vehicle number is less than $ 80 $ because the intersection is not oversaturated; however, the ATTD is higher because of the delay in the lane changing zone. On the other side, the formation control in FCLC helps to generate a collision-free and feasible path planning set to MCC. Moreover, the steady parallel formation is convenient for MCC to adjust the car following distance to desired virtual platoon positions. MCC algorithm reduces $ 2.2 \% $ to $ 32.6 \% $ of the ATTD. Moreover, as the vehicle number increases, it has less improvement on ATTD since ATTD is not the priority concern in optimization.
\subsection{Simulation Results for Different Traffic Volumes}
The second simulation was conducted for different traffic volumes. Similar to the previous simulation, ten randomly generated repetitions of the simulation were conducted for each traffic volume. The number of vehicles was set to $ 50 $, and the traffic volume varied from $ 500 \, \mathrm{veh/h} $ to $ 2500 \, \mathrm{veh/h} $.
Generally, as the number of vehicles is constant, the evacuation time is decreased as the traffic volume increases. The MCC exhibits noticeable improvement in the traffic efficiency, which reduces the evacuation time by $ 18.8\% $ to $ 33.2\% $ and ATTD by $ 1.9\% $ to $ 14.5\% $. The comparison between the SumoLC and FCLC results agrees with the analysis of the simulation results for different numbers of vehicles. FCLC delays the travel time in lane changing zone because of velocity settings. FCLC+MCC provides a feasible solution with noticeable improvement in the traffic efficiency.
\section{Conclusions}
\label{Sec:Conclusion}
In this paper, a two-stage cooperation framework is proposed to improve vehicle safety and traffic efficiency at intersections where lane changing is permitted. In the first stage, we design an iterative method to solve the multi-vehicle target assignment and path planning problem. In contrast to existing single-vehicle lane changing algorithms, the deadlock problem is solved in our method and a feasible collision-free path planning is ensured. In the second stage, a graph-based method is proposed to schedule the CAV arrival time. A heuristic algorithm is established to solve the problem with a low computational burden. Traffic simulations verified the effectiveness of the proposed algorithm.
Future research directions include studying the relationship of the two decoupled stages further. In this study, the optimal arrival plan is calculated when the CAV arrives at the car-following zone. However, it can also be calculated when the CAV enters the lane changing zone. Thus, in the lane changing zone, a CAV can drive to its target position in virtual platoon. Another interesting topic involves considering HDVs in this scenario because lane changing would be permitted. There has been little research on intersections with mixed traffic where lane changing is permitted.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:Intro}
Biochemical processes mostly involve macromolecules such as proteins and nucleic acids in ionic or
salt solutions. The resultant electrostatic interactions are significant in understanding the
biological functions and structures of biomolecules, enzyme catalysis, molecular recognition, and
biomolecular encounter or association rates \cite{Fogolari2002,Neves-Petersen2003,Stein:2007,Stein:2010}.
Efficient modeling of these interactions remains a great challenge in computational biology because
of the complexity of biomolecular systems which are dominated by the effects of solvation on
biomolecular processes and by the long-range intermolecular interactions
\cite{DesHolm:98,LiStCaMaMe:13,ren2012}.
There are two main types of electrostatic models which can be used to model electrostatic
interactions in ionic solutions. The explicit approaches which treat both the solute and solvent in
atomic detail, are generally computationally demanding. This is because they require substantial
sampling and equilibration in order to converge properties of interest in an ensemble
\cite{ren2012,Jurrus2018}. On the other hand, continuum or implicit approaches treat the solvent
molecules as a continuum, by integrating out uninteresting degrees of freedom in order to circumvent
the need for sampling and equilibration \cite{Bashford:2000,ren2012,Jurrus2018}.
\begin{figure}[b]
\centering
\captionsetup{width=\linewidth}
\includegraphics[width=9.0cm]{Domain_PBE_new.png}
\caption{\label{fig:Biomolecular_system}
2D representation of a biomolecular system.}
\end{figure}
There exists a number of implicit solvation approaches for biomolecules \cite{Bashford:2000,Barone:1997,
ren2012}, but the most popular is based on the Poisson-Boltzmann equation (PBE), which was extensively
analyzed, for example, in \cite{Holst94}. The PBE is used for calculating the electrostatic potential and
energies of ionic solvated biomolecules. We present the PBE model in \Cref{sec:PBE_theory}.
It is impossible to obtain analytic solutions of the PBE for biomolecules with complex geometries and
highly singular charge distributions \cite{Holst94, Dong2008}. The numerical solution of the PBE was
pioneered by Warwicker and Watson in 1982 \cite{Warwicker1982}, where the electrostatic potential was
computed at the active site of an enzyme using the finite difference method (FDM). Besides the FDM
\cite{Baker2001, Wang2010}, other numerical techniques such as the finite element methods (FEM)
\cite{Baker2001, Holst2000} and the boundary element methods (BEM) \cite{Boschitsch2004, Zhou1993} have
hitherto successfully been used to solve the PBE, see \cite{Holst:2008} for a thorough review. However, the
numerical solution of the PBE is faced with a number of challenges. The most significant are the strong
charge singularities caused by the singular source terms (Dirac delta distribution), the nonlinearity
caused by the exponential nonlinear terms, the unbounded domain due to slow polynomial decay of the
potential with respect to distance, and imposing the correct jump or interface conditions
\cite{Xie:14, Mirzadeh:13}.
The presence of a highly singular right-hand side of (\ref{eqn:PBE}) which is described by a sum of Dirac
delta distributions introduces significant errors in the numerical solution of the PBE. To overcome this
problem, the PBE theory has recently received a major boost by the introduction of solution decomposition
(regularization) techniques which have been developed, for example, in
\cite{Xie:14,Mirzadeh:13,Chen:07,Chern:2003}, see the discussion in \S\ref{sec:soln_decomposition}.
The idea behind these regularization techniques is the avoidance of building numerical
approximations corresponding to the Dirac delta distributions by treating the biomolecular system (see
\Cref{fig:Biomolecular_system}), as an interface problem. This is coupled with the advantage that analytical
expansions in the molecular sub-region are possible, by the Newton kernel.
In this paper, for resolving the problem of strong singularities, we apply the method introduced recently
in \cite{BeKKKS:18}, for the computation of the free-space electrostatic potential of a linear PBE and
Poisson equation. For this purpose the range-separated canonical tensor format was applied, which was
introduced and analyzed in \cite{BKK_RS:18, BKK_RS:16}. We extend the results of \cite{BeKKKS:18} to the
case of nonlinear BPE and compare the method numerically for a number of biomolecules. Similar to
\cite{BeKKKS:18}, we apply the PBE only to the regular part of the solution corresponding to the
modified right-hand side via extraction of the long-range part in the discretized Dirac delta
distribution \cite{khor-DiracRS:2018}.
The RS tensor formats can be gainfully applied in computational problems which include functions with
multiple local singularities or cusps, Green kernels with intrinsic non-local behavior, and in various
approximation problems which are generated by radial basis functions. The grid-based canonical tensor
representation for the Newton kernel was developed in \cite{BeHaKh:08} and then gainfully applied in
tensor-based electronic structure calculations \cite{khor-ml-2009,Khor_bookQC_2018}. Tensor numerical
techniques for superfast computation of the collective electrostatic potentials of large finite lattice
clusters have been previously introduced in \cite{VeBoKh:Ewald:14}.
The splitting technique employed in this paper is based on the RS tensor decomposition of the discretized
Dirac delta distribution \cite{khor-DiracRS:2018}, which allows avoiding the nontrivial matrix
reconstruction as in (\ref{eqn:PBE_solndecomp}) and in \cite{Xie:14}. The only requirement in this approach
is a simple modification of the singular charge density of the PBE in the molecular region $\Omega_m$,
which does not change the FEM/FDM system matrix. The singular component in the total potential is recovered
explicitly by the short-range component in the RS tensor splitting of the Newton potential. The main
computational benefits of this approach are the localization of the modified singular charge density
within the molecular region and automatic maintaining of the continuity in the Cauchy data on the
interface. Furthermore, this computational scheme only includes solving a single system of FEM/FDM
equations for the regularized (or long-range) component of the decomposed potential.
The remainder of the paper is structured as follows. \Cref{sec:soln_decomposition} provides insights into
the existing solution decomposition techniques for the PBE model. \Cref{sec:RS_survey} describes the
basic rank-structured tensor formats and the short description of the range-separated tensor format
\cite{BKK_RS:18, BKK_RS:16} for representation of the electrostatic potential of multiparticle systems.
\Cref{sec:RS_2_PBE} explains how the application of the RS tensor format leads to the new regularization
scheme for solving the PBE. Finally, \Cref{sec:Numer_Tests} presents the numerical tests illustrating the
benefits of the proposed method and comparisons with the solutions obtained by the
standard FEM/FDM-based
PBE solvers.
\section{The Poisson-Boltzmann equation theory}
\label{sec:PBE_theory}
The PBE is a nonlinear elliptic partial differential equation (PDE) which computes a global solution for
the electrostatic potential within the biomolecule ($\Omega_m$) and in the surrounding ionic solution
($\Omega_s$), see \Cref{fig:Biomolecular_system} for the illustration of the two regions. For a monovalent
electrolyte (i.e., $1:1$ ion ratio), the dimensionless PBE is given by
\begin{equation}\label{eqn:PBE}
-\nabla\cdot(\epsilon(\bar{x})\nabla u(\bar{x})) + \bar{\kappa}^2(\bar{x})\sinh(u(\bar{x})) =
\sum_{i=1}^{N_m}q_i\delta(\bar{x}-\bar{x}_i),\quad \Omega \in
\mathbb{R}^3,
\end{equation}
subject to
\begin{equation}\label{eq:DH_solution}
u(\bar{x}) = \sum_{i=1}^{N_m}\frac{q_ie^{-\kappa(d-a_i)}}{4\pi\epsilon_s (1+\kappa a_i)d}
\quad \mbox{on} \, \, \partial{\Omega}, \quad d = \lVert \bar{x}-\bar{x}_i \rVert,
\quad \bar{x} = (x,y,z),
\end{equation}
where $u(\bar{x}) = {e_c\psi(\bar{x})}/{\kappa_B T}$ represents the dimensionless potential, $\psi(\bar{x})$
is the original electrostatic potential in centimeter-gram-second (cgs) units scaled to the thermal voltage
$(\kappa_B T)/e_c$, $q_i = \frac{4\pi e_c^2}{\kappa_B T}z_i$, $N_m$ is the total number of point partial charges in the
biomolecule, $\epsilon_s$ is the bulk solvent dielectric coefficient, and $a_i$ is the atomic radius.
Here, $\kappa_B T$, $\kappa_B$, $T$, $e_c$, and $z_i$ are the thermal energy, the Boltzmann constant, the absolute
temperature, the electron charge, and the non-dimensional partial charge of each atom, respectively. The
Debye-H\"uckel screening parameter, $\kappa^2 = {8\pi e_c^2 I}/{1000\epsilon_s \kappa_BT}$, describes ion
concentration and accessibility, and is a function of the ionic strength
$I = 1/2\sum_{j=1}^{N_{ions}}c_jz_j^2$, where $z_j$ and $c_j$ are charge and concentration of each ion.
The sum of Dirac delta distributions, located at atomic centers $\bar{x}_i$, represent the molecular charge
density. See \cite{Holst94, KwBFM2017} for more details concerning the PBE theory.
The dielectric coefficient $\epsilon(\bar{x})$ and kappa function $\bar{\kappa}^2(\bar{x})$ are piecewise
constant functions given by
\begin{eqnarray}\label{eqn:diel_kappa_def}
\epsilon(\bar{x}) =
\begin{cases}
\epsilon_m = 2 & \text{if } \bar{x} \in \Omega_m\\
\epsilon_s \,\,= 78.54 & \text{if } \bar{x} \in \Omega_s
\end{cases}, \quad \quad
\bar{\kappa}(\bar{x}) =
\begin{cases}
0 & \text{if } \bar{x} \in \Omega_m \\
\sqrt{\epsilon_s}\bar{\kappa} & \text{if } \bar{x} \in \Omega_s
\end{cases},
\end{eqnarray}
where $\Omega_m$ and $\Omega_s$ are the molecular and solvent regions, respectively, as shown in
\Cref{fig:Biomolecular_system}. Details of regarding the PBE theory and the significance of (\ref{eqn:PBE})
in biomolecular modeling can be found in \cite{SharpHonig90,Holst94,KwBFM2017}.
The PBE in (\ref{eqn:PBE}) can be linearized for small electrostatic potentials relative to the thermal
energy (i.e., $\psi(\bar{x}) \ll \kappa_BT$). Nevertheless, even when the linearization condition does not hold, the
solution obtained from the linearized PBE (LPBE) is close to that of the nonlinear PBE \cite{Fogolari99}.
The onset of substantial differences between the two models is attributed to the magnitude of the electric
field, hence, of the charge density at the interface between the solute and the solvent \cite{Fogolari99}.
The LPBE is given by
\begin{equation}\label{eqn:LPBE}
-\nabla\cdot(\epsilon(\bar{x})\nabla u(\bar{x})) + \bar{k}^2(\bar{x})u(\bar{x})
= \sum_{i=1}^{N_m}q_i\delta(\bar{x}-\bar{x}_i).
\end{equation}
The electrostatic potential can be used in a variety of applications, a few of which we highlight here.
First, the surface potential, (i.e., the electrostatic potential on the biomolecular surface), can be
used to obtain insights into possible docking sites for other molecules. Secondly, it can be used to compare
the interaction properties of related proteins by calculating similarity indices \cite{Wade_etal:2001}.
Finally, the electric field, which is the derivative of the potential around the solute, may be essential for
obtaining the rates of molecular recognition and encounter \cite{Fogolari2002,Jurrus2018}.
\section{Rank-structured tensor representation of electrostatic potentials}
\label{sec:RS_survey}
\subsection{Sketch of basic tensor formats}
\label{ssec:Tensor_formats}
Here, we recall the rank-structured tensor formats and briefly describe the range-separated tensor format
introduced in \cite{BKK_RS:18,BKK_RS:16} for tensor-based representation of multiparticle long-range
potentials. Rank-structured tensor techniques have recently gained popularity in scientific computing due
to their inherent property of reducing the grid-based solution of the multidimensional problems arising in
large-scale electronic and molecular structure calculations to essentially 1D computations
\cite{khor-ml-2009,KhKhFl_Hart:09}. In this concern, the so-called reduced higher order singular value
decomposition (RHOSVD) introduced in \cite{khor-ml-2009} is one of the salient ingredients in the
development of tensor methods in quantum chemistry, see details in \cite{Khor_bookQC_2018} and references
therein.
A tensor of order $d$ is defined as a real multidimensional array over a $d$-tuple index set
\begin{equation}
{\bf A} = [a_{i_1, \ldots, i_d}] \equiv [a(i_1, \ldots, i_d)] \in \mathbb{R}^{n_1
\times \cdots \times n_d},
\end{equation}
with multi-index notation $i= (i_1, \ldots, i_d)$, $i_{\ell} \in I_{\ell} := \{1, \dots,
n_{\ell}\}$. It is considered as an element of a linear vector space $\mathbb{R}^{n_1\times \cdots
\times n_d}$ equipped with the Euclidean scalar product $\langle \cdot,\cdot \rangle : \mathbb{V}_n
\times \mathbb{V}_n \rightarrow \mathbb{R}$, defined as
\begin{equation}
\langle {\bf A},{\bf B} \rangle := \sum\limits_{(i_1, \ldots, i_d) \in I} a_{i_1, \ldots, i_d}
b_{i_1, \ldots, i_d} \quad \mbox{for}\quad {\bf A}, \,{\bf B} \, \in \mathbb{V}_n.
\end{equation}
The storage size scales exponentially in the dimension $d$, i.e., $n^d$, resulting in the so-called ``curse of
dimensionality''.
To get rid of the exponential scaling in storage and the consequent drawbacks,
one can apply the rank-structured separable approximations of multidimensional tensors.
The simplest separable tensor is given by a rank-1 canonical tensor (i.e., tensor/outer product of
vectors in $d$ dimensions)
\begin{equation}\label{eqn:rank1_canon}
{\bf U} = {\bf u}^{(1)} \otimes \cdots \otimes {\bf u}^{(d)}
\in \mathbb{R}^{n_1 \times \cdots \times n_d},
\end{equation}
with entries computed as $u_{i_1, \ldots, i_d} = u_{i_1}^{(1)} \cdots u_{i_1}^{(d)}$, which requires only
$(n_1 + \ldots + n_d)\ll n^d$ numbers to store it. If $n_{\ell} = n$, then the storage cost is $dn \ll n^d$.
\begin{definition}
The $R$-term canonical tensor format is defined by a finite sum of rank-1 tensors
\begin{equation}\label{eqn:canonical}
{\bf U}_R = \sum_{k =1}^{R} \xi_k {\bf u}_k^{(1)} \otimes \cdots \otimes {\bf u}_k^{(d)},
\quad \xi_k \in \mathbb{R},
\end{equation}
where ${\bf u}_k^{(\ell)} \in \mathbb{R}^{n_{\ell}}$ are normalized vectors, and $R \in \mathbb{R}_+$ is
the canonical rank.
\end{definition}
The storage cost for this tensor format is bounded by $dRn$. The entries of the three-dimensional
canonical tensor (\ref{eqn:canonical}) are computed as the sums of elementwise products,
\begin{equation}
u_{i_1,i_2,i_3} = \sum_{k=1}^R \xi_k u_{i_1,k}^{(1)} \cdot u_{i_2,k}^{(2)} \cdot u_{i_3,k}^{(3)}.
\end{equation}
\begin{definition}
The rank-${\bf r}$ orthogonal Tucker format for a tensor ${\bf V}$ is
\begin{equation}\label{eqn:Tucker}
{\bf V} = \sum_{\nu_1 =1}^{r_1} \cdots \sum_{\nu_d =1}^{r_d} \beta_{\nu_1, \ldots, \nu_d}
{\bf v}_{\nu_1}^{(1)} \otimes \cdots \otimes {\bf v}_{\nu_d}^{(d)} \equiv
\boldsymbol{\beta} \times_1 V^{(1)} \times_2 V^{(2)}\ldots \times_d V^{(d)},
\end{equation}
where $\{ {\bf v}_{\nu_{\ell}}^{(\ell)}\}_{\nu_{\ell} = 1}^{r_{\ell}} \in \mathbb{R}^{n_{\ell}}$
is the set of orthonormal vectors for $\ell = 1, \ldots, d$.
$\times_{\ell}$ denotes the contraction along the mode $\ell$ with the orthogonal matrices
$V^{(\ell)} = [{\bf v}_1^{(\ell)}, \ldots, {\bf v}_{r_{\ell}}^{(\ell)}] \in \mathbb{R}^{n_{\ell}
\times r_{\ell}}$.
$\boldsymbol{\beta} = \beta_{\nu_1, \ldots, \nu_d} \in \mathbb{R}^{r_1 \times \cdots r_d}$ is
the Tucker core tensor. The storage cost is bounded by $drn + r^d$ with $r = |r| := \mbox{max}_{\ell}
r_{\ell}$.
\end{definition}
Rank-structured tensor approximations provide fast multilinear algebra with linear complexity scaling in
the dimension $d$ \cite{BKK_RS:18}. For instance, for the given canonical tensor representation
(\ref{eqn:canonical}), the Hadamard products, the Euclidean scalar product, and $d$-dimensional
convolution can be computed by univariate tensor operations in 1D complexity \cite{KhKh:06}.
\subsection{Outline on the RS tensor format for numerical modeling of multiparticle systems}
\label{ssec:Coulomb}
In what follows, first recall the canonical tensor representation of the non-local Newton kernel
$1/\|\bar{x}\|$, $\bar{x} \in \mathbb{R}^3$, by using sinc-quadratures and Laplace transform introduced
in \cite{BeHaKh:08}. The corresponding theoretical basis was developed in seminal papers
\cite{HaKhtens:04I,khor-rstruct-2006} on low-rank tensor product approximation of multidimensional
functions and operators. According to above papers, the Newton kernel is approximated in a computational
domain $\Omega=[-b,b]^3$, using the uniform $n\times n\times n$ 3D Cartesian grid. Then, using the Laplace
transform and sinc-quadrature approximation, this discretized potential is approximated by a canonical
rank $R$ tensor,
\begin{equation} \label{eqn:canon_repr}
\mathbf{P} \approx
\sum\limits_{k=1}^{R} {\bf p}^{(1)}_k \otimes {\bf p}^{(2)}_k \otimes {\bf p}^{(3)}_k
\in \mathbb{R}^{n^{\otimes 3}},
\end{equation}
with vectors ${\bf p}^{(\ell)}_k \in \mathbb{R}^n$, and the accuracy of this approximation decays
exponentially fast in the rank parameter $R$.
The canonical tensor representation of the Newton kernel was first applied in rank-structured grid-based
calculations of the multidimensional operators in electronic structure calculations,
\cite{khor-ml-2009,KhKhAn:12}, where it manifested high accuracy levels compared with the analytical based
computational methods.
In \cite{VeBoKh:Ewald:14}, the canonical tensor representation was applied in modeling
of the electrostatic potentials in finite rectangular three-dimensional lattices, where
it was proved that the rank of the collective long-range electrostatic potentials
of large 3D lattices remains as small as that of a canonical tensor for a single Newton kernel.
For lattices with defects and impurities it is increased by a small constant \cite{Khor_bookQC_2018}.
For modeling the electrostatic interaction potential in large molecular systems of general type, the
range-separated tensor format \cite{BKK_RS:18} is based on additive decomposition of the reference canonical
tensor $\textbf{P}_R$
\[
\mathbf{P}_R = \mathbf{P}_{R_s} + \mathbf{P}_{R_l},
\]
with
\begin{equation} \label{eqn:Split_Tens}
\mathbf{P}_{R_s} =
\sum\limits_{k\in {\cal K}_s} {\bf p}^{(1)}_k \otimes {\bf p}^{(2)}_k \otimes {\bf p}^{(3)}_k,
\quad \mathbf{P}_{R_l} =
\sum\limits_{k\in {\cal K}_l} {\bf p}^{(1)}_k \otimes {\bf p}^{(2)}_k \otimes {\bf p}^{(3)}_k.
\end{equation}
Here, ${\cal K}_l := \{k|k = 0,1, \ldots, R_l\}$ and ${\cal K}_s := \{k|k = R_l+1, \ldots, M\}$ are the
sets of indices for the long- and short-range canonical vectors determined depending of the claimed size
of effective support of the short-range part $\mathbf{P}_{R_s}$.
The total electrostatic potential is represented by a projected tensor
${\bf P}_0\in \mathbb{R}^{n \times n \times n}$ that can be constructed by a direct sum of
shift-and-windowing transforms of the reference tensor $\widetilde{\bf P}_R$, defined in the twice larger
domain $\widetilde{\Omega}_n$ (see \cite{VeBoKh:Ewald:14} for more details),
\begin{equation}\label{eqn:Total_Sum}
{\bf P}_0 = \sum_{\nu=1}^{N} {z_\nu}\, {\cal W}_\nu (\widetilde{\bf P}_R)=
\sum_{\nu=1}^{N} {z_\nu} \, {\cal W}_\nu (\widetilde{\mathbf{P}}_{R_s} + \widetilde{\mathbf{P}}_{R_l})
=: {\bf P}_s + {\bf P}_l.
\end{equation}
The shift-and-windowing transform ${\cal W}_\nu$ maps a reference tensor
$\widetilde{\bf P}_R\in \mathbb{R}^{2n \times 2n \times 2n}$ onto its sub-tensor
of smaller size $n \times n \times n$, obtained by first shifting the center of
the reference tensor $\widetilde{\bf P}_R$ to the grid-point $x_\nu$ and then restricting
(windowing) the result onto the computational grid $\Omega_n$.
It was proven in \cite{BKK_RS:18} that the Tucker and canonical rank parameters of the "long-range part"
in the tensor ${\bf P}_0$, defined by
\begin{equation}\label{eqn:Long-Range_Sum}
{\bf P}_l = \sum_{\nu=1}^{N} {z_\nu} \, {\cal W}_\nu (\widetilde{\mathbf{P}}_{R_l})=
\sum_{\nu=1}^{N} {z_\nu} \, {\cal W}_\nu
(\sum\limits_{k\in {\cal K}_l} \widetilde{\bf p}^{(1)}_k \otimes \widetilde{\bf p}^{(2)}_k
\otimes \widetilde{\bf p}^{(3)}_k)
\end{equation}
remain almost uniformly bounded in the number of particles,
\[
\mbox{rank}({\bf P}_l)\leq C \log^{3/2} N.
\]
The rank reduction algorithm is accomplished by the canonical-to-Tucker (C2T) transform through
the reduced higher order singular value decomposition (RHOSVD) \cite{khor-ml-2009} with a
subsequent Tucker-to-canonical (T2C) decomposition (see \cite{Khor_bookQC_2018} and references therein).
In turn, the tensor representation of the sum of short-range parts is considered as a sum of cumulative
tensors of small support characterized by the list of the 3D potentials coordinates and weights. The
total tensor is then represented in the range-separated tensor format \cite{BKK_RS:18}.
Here we recall the slightly simplified definition of the RS tensor format.
\begin{definition}\label{Def:RS-Can_format} (RS-canonical tensors \cite{BKK_RS:18}).
Given a reference tensor ${\bf A}_0$ such that $\mbox{rank}({\bf A}_0)\leq R_0$,
the separation parameter $\gamma \in \mathbb{N}$ and a set of points $x_\nu \in \mathbb{R}^{d}$,
$\nu=1,\ldots,N$,
the RS-canonical tensor format specifies the class of $d$-tensors
${\bf A} \in \mathbb{R}^{n_1\times \cdots \times n_d}$
which can be represented as a sum of a rank-${R}_L$ canonical tensor
\begin{equation}\label{eq:LR_tensor_sum}
{\bf A}_{R_L} = {\sum}_{k =1}^{R_L} \xi_k {\bf a}_k^{(1)} \otimes \cdots \otimes {\bf a}_k^{(d)}
\in \mathbb{R}^{n_1\times ... \times n_d}
\end{equation}
and a cumulated canonical tensor
\begin{equation}\label{eq:CCT_tensor}
\widehat{\bf A}_S={\sum}_{\nu =1}^{N} c_\nu {\bf A}_\nu ,
\end{equation}
generated by replication of the reference tensor ${\bf A}_0$ to the points $x_\nu$.
Then the RS canonical tensor is represented in the form
\begin{equation}\label{eqn:RS_Can}
{\bf A} = {\bf A}_{R_L} + \widehat{\bf A}_S=
{\sum}_{k =1}^{R_L} \xi_k {\bf a}_k^{(1)} \otimes \cdots \otimes {\bf a}_k^{(d)} +
{\sum}_{\nu =1}^{N} c_\nu {\bf A}_\nu,
\end{equation}
where $\mbox{diam}(\mbox{supp}{\bf A}_0)\leq 2 \gamma$ in the index size.
\end{definition}
The storage size for the RS-canonical tensor ${\bf A}$ in (\ref{eqn:RS_Can})
is estimated by (\cite{BKK_RS:18}, Lemma 3.9),
$$
\mbox{stor}({\bf A})\leq d R n + (d+1)N + d R_0 \gamma.
$$
Notice that the RS tensor decomposition of the collective electrostatic potential $\mathbf{P}_{0}$
can be obtained by setting ${\bf A}_0=\mathbf{P}_s$ and
${\bf A}_{R_L} = {\bf P}_l$.
\section{Solution decomposition techniques for the PBE}
\label{sec:soln_decomposition}
The presence of a highly singular right-hand side of (\ref{eqn:PBE}) implies that every singular charge
$z_i$ in (\ref{eqn:PBE}), the electrostatic potential $u(\bar{x})$ exhibits degenerate behavior at each atomic
position $\bar{x}_i$ in the molecular region $\Omega_m$.
To overcome this difficulty, the PBE theory has recently
Typical solution decomposition techniques for the PBE entail a coupling of two equations for the
electrostatic potential in the molecular ($\Omega_m$) and solvent ($\Omega_s$) regions, through the
boundary interface \cite{Chen:07, Chern:2003}. The equation inside $\Omega_m$ is simply the Poisson
equation, due to the absence of ions, i.e.,
\begin{equation}\label{eqn:PE}
-\nabla \cdot(\epsilon_m\nabla u) = \sum_{i=1}^{N_m}q_i \delta(\bar{x}-\bar{x}_i)
\quad \mbox{in} \,\, \Omega_m,
\end{equation}
On the other hand, there is absence of atoms in $\Omega_s$. Therefore, the density is purely given by the
Boltzmann distribution
\begin{equation}\label{eqn:Boltzmann_distn}
-\nabla \cdot(\epsilon_s\nabla u) + \bar{\kappa}^2\sinh(u) = 0 \quad \mbox{in} \,\, \Omega_s.
\end{equation}
The two equations (\ref{eqn:PE}) and (\ref{eqn:Boltzmann_distn}) are coupled together through the
interface boundary conditions
\begin{equation}\label{eqn:interface_condn}
\left[ u\right]_{\Gamma} = 0, \quad \mbox{and} \quad \left[ \epsilon \frac{\partial u}{\partial
n_{\Gamma}}\right]_{\Gamma} = 0,
\end{equation}
where $\Gamma := \partial\Omega_m = \partial\Omega_s \cap \Omega_m$ and $\left[ f\right]_{\Gamma} =
\lim\limits_{t \longrightarrow 0} f(\bar{x}+tn_{\Gamma}) - f(\bar{x}-tn_{\Gamma})$. Here, $n_{\Gamma}$
denotes the unit outward normal direction of the interface $\Gamma$.
Next, we highlight one of the solution decomposition techniques for the PBE in \cite{Chen:07} which
provides the motivation for the RS tensor format demonstrated in this paper. It is also implemented as an
option for the PBE solution in the well-known adaptive Poisson-Boltzmann software (APBS) package using the
FEM \cite{Bakersept2001}. To deal with the singular source term represented by the sum of Dirac delta
distributions in the PBE, the unknown solution $u(\bar{x})$ is decomposed as an unknown smooth function
$u^r(\bar{x})$ and a known singular function $G(\bar{x})$, i.e.,
\begin{equation}
u(\bar{x}) = G(\bar{x}) + u^r(\bar{x}),
\end{equation}
where
\begin{equation}\label{eqn:analytic_PE}
G(\bar{x})= \sum_{i=1}^{N_m}\frac{q_i}{\epsilon_m}\frac{1}{\|\bar{x} -\bar{x}_i\|},
\end{equation}
is a sum of the Newton kernels ($1/\|\bar{x}\|$), which solves the Poisson equation (\ref{eqn:PE}) in
$\mathbb{R}^3$. Substitute the decomposition into (\ref{eqn:PBE}), to obtain
\begin{equation}\label{eqn:PBE_solndecomp}
\begin{rcases}
\begin{aligned}
-\nabla \cdot(\epsilon\nabla u^r) + \bar{\kappa}^2(\bar{x})\sinh(u^r +G)
&= \nabla \cdot((\epsilon-\epsilon_m)\nabla G), & \, \mbox{in} \,\, \Omega \\
u^r &= g - G & \, \mbox{on} \,\, \partial{\Omega},
\end{aligned}
\end{rcases}
\end{equation}
where $g(\bar{x})$ is the boundary condition obtained from (\ref{eq:DH_solution}). The PBE in
(\ref{eqn:PBE_solndecomp}) is referred to as the regularized PBE (RPBE) in \cite{Chen:07}. Notice that
the singularities of the Dirac delta distribution are transferred to $G$, which is known analytically,
therefore, building the numerical approximation to $G$ is circumvented. Consequently, the cutoff
coefficients $\bar{\kappa}$ and $\epsilon-\epsilon_m$ are zero in $\Omega_m$, where the degenerate
behaviour is exhibited at each $\bar{x}_i$. This allows the RPBE to be a mathematically well-defined
equation for the regularized solution $u^r$. It is important to note that away from the ${\bar{x}_i}$, the
function $G$ is smooth \cite{Chen:07}.
The RPBE in (\ref{eqn:PBE_solndecomp}) can further be decomposed into the linear and nonlinear components,
$u^r(\bar{x}) = u^l(\bar{x}) + u^n(\bar{x})$, where $u^l(\bar{x})$ satisfies,
\begin{equation}\label{eqn:PBE_solndecompb}
\begin{rcases}
\begin{aligned}
-\nabla \cdot(\epsilon\nabla u^l) &= \nabla \cdot((\epsilon-\epsilon_m)\nabla G),
& \, \mbox{in} \,\, \Omega \\
u^l &= 0 & \, \mbox{on} \,\, \partial{\Omega},
\end{aligned}
\end{rcases}
\end{equation}
and $u^n(\bar{x})$ satisfies
\begin{equation}\label{eqn:PBE_solndecompc}
\begin{rcases}
\begin{aligned}
-\nabla \cdot(\epsilon\nabla u^n) + \bar{\kappa}^2(x)\sinh(u^n + u^l + G) &= 0,
& \, \mbox{in} \,\, \Omega \\
u^n &= g-G & \, \mbox{on} \,\, \partial{\Omega}.
\end{aligned}
\end{rcases}
\end{equation}
However, the following computational challenges are inherent in the aforementioned techniques. First, due
to regularization splitting of the solution by using the kappa and dielectric coefficients as cutoff
functions, discontinuities at the interface arise. Therefore, interface or jump conditions need to be
incorporated to eliminate the solution discontinuity (e.g., Cauchy data) at the interface of complicated
sub-domain shapes. Consequently, the long-range components of the free space potential are not completely
decoupled from the short-range parts at each atomic radius, in the ``so-called'' singular function $G$,
in the molecular domain $\Omega_m$. Secondly, the Dirichlet boundary conditions, for example, in
(\ref{eq:DH_solution}) have to be specified using some analytical solution of the LPBE. Thirdly, in
solution decomposition techniques, see, for instance, \cite{Xie:14}, multiple algebraic systems for the
linear and nonlinear boundary value problems have to be solved, thereby increasing the computational
costs. Thirdly, the system matrix is modified because of incorporating the interface conditions and also,
for instance, the smooth function ($G$), in the Boltzmann distribution term in (\ref{eqn:PBE_solndecomp}).
In this paper, we present a new approach for the regularization of the PBE by using the RS canonical
tensor format.
\section{The regularization scheme for the PBE via RS tensor format} \label{sec:RS_2_PBE}
In this section, we extend the approach introduced in \cite{BeKKKS:18} for linear PBE to the nonlinear case.
We present a new regularization scheme for the nonlinear PBE which is based on the range-separated
representation of the highly singular charge density, described by the Dirac delta distribution
in the target PBE (\ref{eqn:PBE}) \cite{khor-DiracRS:2018}.
Similar to \cite{BeKKKS:18} we modify the right-hand
side of the nonlinear PBE (\ref{eqn:PBE}) in such a way that
the short-range part in the solution $u$ can be pre-computed
independently by the direct tensor decomposition of the free space potential, and the initial elliptic
equation (or the nonlinear RPBE) applies only to the long-range component of the total potential. The latter is a
smooth function, hence the FDM/FEM approximation error can be reduced dramatically even on relatively
coarse grids in 3D.
\subsection{Regularization scheme for the nonlinear PBE (NPBE)} \label{ssec:nonlin_PBE}
To fix the idea, we first consider the weighted sum of interaction potentials in a large $N$-particle
system, generated by the Newton kernel, $1/{\|\bar{x}\|}$, at each charge location $\bar{x}_i$,
$\bar{x}\in \mathbb{R}^3$, i.e.,
\begin{equation}\label{eqn:analytic_PE2}
G(\bar{x})= \sum_{i=1}^{N_m}\frac{q_i}{\epsilon_m}\frac{1}{\|\bar{x} -\bar{x}_i\|},
\end{equation}
We recall that the sum of Newton kernels for a multiparticle system discretized by the $R$-term
sum of Gaussian type functions living on the $n^{\otimes 3}$ tensor grid $\Omega_n$ is represented
by a sum of long-range tensors in (\ref{eqn:Long-Range_Sum}) and a cumulated canonical tensor in (\ref{eq:CCT_tensor}),
respectively.
Since it is well known that (\ref{eqn:analytic_PE2}) solves the Poisson equation analytically, i.e.,
\begin{equation}\label{eqn:PE_substit}
-\nabla \cdot(\epsilon_m\nabla G(\bar{x})) = \sum_{i=1}^{N_m}q_i \delta(\bar{x}-\bar{x}_i)
\quad \mbox{in} \,\, \mathbb{R}^3,
\end{equation}
we can leverage this property in order to derive a smooth (regularized) representation, $f_r$, of the
Dirac delta distributions in the right-hand side of (\ref{eqn:PE_substit}). Consider the RS tensor
splitting of the multiparticle Newton potential into a sum of long-range tensors ${\bf P}_l$ in
(\ref{eqn:Long-Range_Sum}) and a cumulated canonical tensor ${\bf P}_s$ in (\ref{eq:CCT_tensor}), i.e.,
\begin{equation}\label{eqn:Newt_splitting}
G(\bar{x}) = {\bf P}_s(\bar{x}) + {\bf P}_l(\bar{x}).
\end{equation}
Substituting each of the components of (\ref{eqn:Newt_splitting}) into the discretized Poisson equation,
we derive the respective components of the molecular charge density
(or the collective Dirac delta distributions) as follows
\begin{equation}\label{eqn:Dirac_splitting}
f^s:= -A_{\Delta} {\bf P}_s, \quad
\mbox{and} \quad f^l:= -A_{\Delta} {\bf P}_l,
\end{equation}
where $A_{\Delta}$ is the 3D finite difference Laplacian matrix defined on the uniform rectangular grid as
\begin{equation}\label{eqn:Lapl_Kron3}
A_{\Delta} = \Delta_{1} \otimes I_2\otimes I_3 + I_1 \otimes \Delta_{2} \otimes I_3 +
I_1 \otimes I_2\otimes \Delta_{3},
\end{equation}
where $-\Delta_\ell = h_\ell^{-2} \mathrm{tridiag} \{ 1,-2,1 \} \in \mathbb{R}^{n_\ell \times n_\ell}$,
$\ell=1,2,3$, denotes the discrete univariate Laplacian and $I_\ell$, $\ell=1,2,3$, is the identity matrix
in each dimension.
See \cite{BeKKKS:18,khor-DiracRS:2018} for more details.
\Cref{fig:Long_short_RHS} depicts the behaviour of the modified representations of both the smooth and
singular components of the Dirac delta distributions using the formula in (\ref{eqn:Dirac_splitting}). The
charge density data is obtained from protein Fasciculin 1, an anti-acetylcholinesterase toxin from green
mamba snake venom \cite{DuMaBoFo:92}. Notice from the highlighted data cursors, that
the effective supports of both functions
are localized within the molecular region, with values dropping to zero outside this region. Furthermore,
\Cref{fig:Long-range_RHS} represents the function $f^l$, which we utilize as the modified right-hand side
to derive a regularized PBE model (RPBE) in the next step.
\begin{figure}[t]
\captionsetup{width=\linewidth}
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{Smooth_RHS_protein.png}
\caption{Long-range part of charge density.}
\label{fig:Long-range_RHS}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{Singular_RHS_protein.png}
\caption{Short-range part of charge density.}
\label{fig:Short-range_RHS}
\end{subfigure}
\caption{\label{fig:Long_short_RHS}
The long- and short-range parts of the charge density for protein Fasciculin 1 on $129^{\otimes 3}$
grid.}
\end{figure}
The nonlinear regularized PBE (NRPBE) can now be derived as follows. First, the unknown solution
(or target electrostatic potential) $u$ to the PBE (\ref{eqn:PBE}) can be decomposed as
\[
u = u^s + u^r,
\]
where $u^s$ is the known singular function (or short-range component) and $u^l$ is the unknown long-range
component to be determined. Therefore, the PBE (\ref{eqn:PBE}) can be rewritten as
\begin{equation}\label{eqn:PBE_splitting}
\begin{rcases}
\begin{aligned}
-\nabla\cdot(\epsilon\nabla (u^s + u^r)) + \bar{\kappa}^2\sinh(u^s + u^r) &= f^s + f^l
\quad \mbox{in} \,\, \mathbb{R}^3,\\
u^r &= g, \quad \mbox{on} \quad \partial\Omega,
\end{aligned}
\end{rcases}
\end{equation}
where the right-hand side of (\ref{eqn:PBE}) is replaced by $f^s + f^l $ due to (\ref{eqn:PE_substit}) and
(\ref{eqn:Dirac_splitting}) and $g$ is the Dirichlet boundary conditions defined in (\ref{eq:DH_solution}).
It was proved in \cite{khor-DiracRS:2018} and demonstrated in \cite{BeKKKS:18} that the function $f^s$ and the corresponding
short-range potential $u^s$ are localized within the molecular region $\Omega_m$ and vanishes on the
interface $\Gamma$. Moreover, from (\ref{eqn:diel_kappa_def}), the function $\bar{\kappa}$ is piecewise
constant and $\bar{\kappa} = 0$ in $\Omega_m$. Therefore, we can rewrite the Boltzmann distribution term
in (\ref{eqn:PBE_splitting}) as
\begin{equation}\label{eqn:Boltzmann_lr
\begin{aligned}
& \bar{\kappa}^2\sinh(u^s + u^r) = \bar{\kappa}^2\sinh(u^r), \quad & \mbox{ because } \, u^s = 0 \,
\mbox{ in } \Omega_s.
\end{aligned}
\end{equation}
Consequently, following the splitting of the Dirac delta distributions in (\ref{eqn:Dirac_splitting}), the
short-range component of the potential satisfies the Poisson equation, i.e.,
\begin{equation}\label{eqn:PE_sr}
-\nabla\cdot(\epsilon_m\nabla u^s) = f^s \quad \mbox{in} \,\, \mathbb{R}^3.
\end{equation}
It can be easily shown that \[u^s(\bar{x}) = {\bf P}_s \] is the cumulated canonical tensor in (\ref{eq:CCT_tensor})
which represents the precomputed short-range potential sum supported within the solute domain $\Omega_m$.
Subtracting (\ref{eqn:PE_sr}) from (\ref{eqn:PBE_splitting}) and using (\ref{eqn:Boltzmann_lr}),
we obtain the NRPBE as follows
\begin{equation}\label{eqn:RPBE_decomp}
-\nabla \cdot(\epsilon\nabla u^r(\bar{x})) + \bar{\kappa}^2(\bar{x})\sinh(u^r(\bar{x}))
= f^l, \quad \mbox{in} \,\,\, \Omega,
\end{equation}
subject to
\begin{equation}\label{eqn:RPBE_bc}
u^r(\bar{x}) = \bar{\kappa}^2(\bar{x})\mathbf{P}_l \quad \mbox{on} \,\, \partial{\Omega}.
\end{equation}
We recall that the regularization scheme for linear PBE introduced in \cite{BeKKKS:18}
reads as follows,
\begin{equation}\label{eqn:LRPBE_decomp}
-\nabla \cdot(\epsilon\nabla u^r(\bar{x})) + \bar{\kappa}^2(\bar{x})u^r(\bar{x})
= f^l(\bar{x}), \, \mbox{in} \,\,\, \Omega,
\end{equation}
subject to the Dirichlet boundary conditions
\begin{equation}\label{eqn:LRPBE_decomp_bc}
u^r(\bar{x}) = \bar{\kappa}^2(\bar{x})\mathbf{P}_l \quad \mbox{on} \,\, \partial{\Omega}.
\end{equation}
In this way, (\ref{eqn:RPBE_decomp}) -- (\ref{eqn:RPBE_bc}) generalizes the regularization scheme
(\ref{eqn:LRPBE_decomp}) -- (\ref{eqn:LRPBE_decomp_bc}) to the nonlinear case.
Notice that by construction, the short-range potential vanishes on the interface $\Gamma$, hence it
satisfies the discrete Poisson equation in (\ref{eqn:PE}) with the respective charge density $f^s$ and zero
boundary conditions on $\Gamma$. Therefore, we recall (see \cite{BeKKKS:18} for the detailed discussion)
that this equation can be subtracted from the full
linear discrete PE system, such that the long-range component of the solution, $\mathbf{P}_l$, will satisfy
the same linear system of equations (same interface conditions), but with a modified charge density
corresponding to the weighted sum of the long-range tensors $f^l$ only.
\section{Numerical approach to solving the NRPBE} \label{sec:RPBE_numerical_approach}
Consider the uniform 3D $n^{\otimes 3}$ rectangular grid in $\Omega = [-b,b]^3$ with the mesh parameters
$dx,dy,dz < 0.5$. One standard way of solving the NRPBE in (\ref{eqn:RPBE_decomp}) is that it is first
discretized in space to obtain a nonlinear system in matrix-vector form
\begin{equation}\label{eq:FOM}
A(u_{\mathcal{N}}^r) = b^r,
\qquad \mbox{ in }\, \mathbb{R}^3,
\end{equation}
where $A(u_{\mathcal{N}}^r) \in \mathbb{R}^{\mathcal{N}\times \mathcal{N}}$, $b^r
\in \mathbb{R}^{\mathcal{N}}$, and $u_{\mathcal{N}}^r$ is the discretized solution vector. Here,
$\mathcal{N}$ is of $\mathcal{O}(10^6)$.
Then system (\ref{eq:FOM}) can be solved using several existing techniques. For
example, the nonlinear relaxation methods has been implemented in the Delphi software
\cite{Rocchia_2001}, the nonlinear conjugate gradient (CG) method has been implemented in University
of Houston Brownian Dynamics (UHBD) software \cite{Brock_1992}, the nonlinear multigrid (MG) method
\cite{Oberoi_1993} and the inexact Newton method have been implemented in the adaptive
Poisson-Boltzmann solver (APBS) software \cite{Holst:95}.
In this study, we apply a different approach of solving (\ref{eqn:RPBE_decomp})
\cite{Mirzadeh:13, Shestakov:2002, Ji:2018}. In particular, an iterative approach is first applied to
the continuous NRPBE in (\ref{eqn:RPBE_decomp}), where at the $(n+1)$th iteration step, the NRPBE is
approximated by a linear equation via the Taylor series truncation. The expansion point of the Taylor
series is the continuous solution $(u^r)^n$ at the nth iteration step.
Consider $(u^r)^n$ as the approximate solution at the $n$th iterative step, then the nonlinear
term $\sinh((u^r)^{n+1})$ at the $(n+1)$th step is approximated by its truncated Taylor series
expansion as follows
\begin{equation}\label{eq:Taylor_expansion_sinh}
\sinh((u^r)^{n+1}) \approx \sinh((u^r)^n)
+ ((u^r)^{n+1} - (u^r)^n)\cosh((u^r)^n).
\end{equation}
Substituting the approximation (\ref{eq:Taylor_expansion_sinh}) into (\ref{eqn:RPBE_decomp}), we
obtain
\begin{multline}\label{eq:PBE_approx_sinh}
-\nabla\cdot(\epsilon(\bar{x})\nabla (u^r)^{n+1})
+ \bar{\kappa}^2(\bar{x})\cosh((u^r)^n)(u^r)^{n+1}
= -\bar{\kappa}^2(\bar{x})\sinh((u^r)^n) \\
+ \bar{\kappa}^2(\bar{x})\cosh((u^r)^n)(u^r)^n + b^r.
\end{multline}
The equation in (\ref{eq:PBE_approx_sinh}) is linear, and can then be numerically solved by first
applying spatial discretization. In this regard, we first define
\begin{equation}\label{eqn:hyperbolic_cosine_vec}
\cosh\odot u_{\mathcal{N}}^r =: w = \begin{bmatrix}
w_1 \\
w_2 \\
\vdots \\
w_{\mathcal{N}}
\end{bmatrix},
\end{equation}
where $\odot$ is the elementwise operation on a vector.
Then, we construct the corresponding diagonal matrix from (\ref{eqn:hyperbolic_cosine_vec}) of the
form \[B = \mbox{diag}(w_1, w_2, \ldots, w_{\mathcal{N}}).\]
Finally, we obtain the following iterative linear system
\begin{equation}\label{eqn:Nonaffine_form_iterative_FOM}
A_1(u_{\mathcal{N}}^r)^{n+1} + A_2B^{n}(u_{\mathcal{N}}^r)^{n+1}
= -A_2\sinh\odot(u_{\mathcal{N}}^r)^{n} + A_2B^{n}(u_{\mathcal{N}}^r)^{n} + b_1^r + b_2,
\end{equation}
where $A_1$ is the Laplacian matrix and $A_2$ is a diagonal matrix containing the $\bar{\kappa}^2$
function. Note that the diagonal matrix $B^{n}$ changes at each iteration step, therefore, it cannot
be precomputed. The vectors $b_1^r$ and $b_2$ are the regularized approximation of the Dirac delta
distributions and the Dirichlet boundary conditions, respectively.
Let
\begin{equation}\label{eqn:affine_A_iter}
A(\cdot) = A_1 + A_2B^n
\end{equation}
and
\begin{equation}\label{eqn:affine_F_iter}
F : \mbox{right-hand side of} \: (\ref{eqn:Nonaffine_form_iterative_FOM}),
\end{equation}
we obtain
\begin{equation}\label{eq:PBE_system_iterative}
A((u_{\mathcal{N}}^r)^n)(u_{\mathcal{N}}^r)^{n+1} = F((u_{\mathcal{N}}^r)^n),
\quad n = 0,1, \ldots.
\end{equation}
Then, at each iteration, system (\ref{eq:PBE_system_iterative}) is a linear system w.r.t.
$(u^r_{\mathcal{N}})^{n+1}$, which can be solved by any linear system solver of choice. In this study,
we employ the aggregation-based algebraic multigrid method (AGMG) \footnote{\textbf{AGMG} implements an
aggregation-based algebraic multigrid method, which solves algebraic systems of linear equations, and is
expected to be efficient for large systems arising from the discretization of scalar second order
elliptic PDEs \cite{Notay:2010}.} \cite{Notay:2010}. \Cref{alg:Iterative_NRPBE} summarizes the detailed
iterative approach of solving (\ref{eq:PBE_system_iterative}). This approach of first linearization,
then discretization is shown to be more efficient than the standard way of first discretization and
then linearization, via, for example, the Newton iteration. The advantage of the proposed approach is
that it avoids computing the Jacobian of a huge matrix. It is observed that it converges faster than
the standard Newton approach.
\begin{algorithm}[t]
\caption{Iterative solver for the NRPBE}\label{alg:Iterative_NRPBE}
\begin{algorithmic}[1]
\Require Initialize the potential $(u_{\mathcal{N}}^r)^0$, e.g., $(u_{\mathcal{N}}^r)^0 = 0$
and the tolerance $\delta^0 = 1$.
\Ensure The converged NRPBE solution $(u_{\mathcal{N}}^r)^n$ at $\delta^n \leq 10^{-8}$.
\While{$\delta^n \geq 10^{-8}$}
\State Solve the linear system (\ref{eq:PBE_system_iterative}) for
$(u_{\mathcal{N}}^r)^{n+1}$ using AGMG.
\State $\delta^{n+1} \gets \|(u_{\mathcal{N}}^r)^{n+1} - (u_{\mathcal{N}}^r)^n\|_2$.
\State $(u_{\mathcal{N}}^r)^n \gets (u_{\mathcal{N}}^r)^{n+1}$.
\EndWhile
\State \textbf{end while}
\end{algorithmic}
\end{algorithm}
The benefits of the RS tensor format as a solution decomposition technique over the existing techniques
in the literature are highlighted as follows. First, the efficient splitting of the short- and long-range
parts in the target tensor circumvents the need to modify jump conditions at the interface and the
use of $\epsilon$ and $\bar{\kappa}$ as cut-off functions, e.g., in (\ref{eqn:PBE_solndecomp}). Secondly,
the long-range part in the RS tensor decomposition of the Dirac delta distributions
\cite{khor-DiracRS:2018} vanishes at the interface and, therefore, the modified charge density in
(\ref{eqn:Dirac_splitting}) generated by this long-range component remains localized in the solute
region. Thirdly, the boundary conditions are obtained from $\mathbf{P}_l$, the long-range part of the
free space potential sum, thereby avoiding the computational costs involved in solving some
external analytical function at the boundary. Lastly, only a single system of algebraic equations is
solved for the regularized component of the collective potential which is then added to the directly
precomputed short-range contribution, $u^s(\bar{x})$. This is more efficient than, for instance, in
\cite{Xie:14}, where the regularized PBE model is subdivided into the linear interface and the
nonlinear interface problems which are solved independently, with respective boundary and interface
conditions.
\section{Efficient computation of electrostatic energies and forces for the PBE} \label{sec:Energy_force_PBE}
\subsection{Electrostatic energies for the PBE}\label{sec:Energy_PBE}
An important application of the PBE's electrostatic potential is the electrostatic solvation free energy,
which is useful in biophysics and biomedicine \cite{Vergara-Perez2016, Holst:2008}. It is defined as the
free energy required to transfer a biomolecule from a uniform dielectric continuum to an inhomogeneous
medium, which is generally divided into nonpolar and polar terms \cite{Dong2008}. The polar contribution
to the solvation free energy is given by
\begin{equation}\label{eqn:solvation_energy}
\Delta G_{solv}^{polar} = G_{elec}^{Solv} - G_{elec}^{Ref},
\end{equation}
where $G_{elec}^{Ref}$ (reference energy) is the total biomolecular electrostatic free energy in the
reference or vacuum state (solute homogeneous dielectric medium) and $G_{elec}^{Solv}$ (solvated energy)
is that in the solvated state (inhomogeneous dielectric medium, for instance, a protein in aqueous medium)
\cite{Vergara-Perez2016,Dong2008}. The electrostatic energy by definition, represents the work
required to assemble the biomolecule, and is given by
\begin{equation}\label{eqn:elec_energy}
G_{elec} = \frac{1}{2}\sum_{i=1}^{N_m}q_i u(\bar{x}_i),
\end{equation}
where $u(\bar{x}_i)$ is the mean electrostatic potential acting on the atom located at $\bar{x}_i$ with
charge $q_i$ \cite{Vergara-Perez2016}.
The APBS software package, for example, has an option which implements the multigrid or multilevel finite
difference calculations using PMG \footnote{\textbf{PMG} is a Parallel algebraic MultiGrid code for general
semilinear elliptic equations\cite{www2}.} \cite{Bakersept2001, Chen:07}, solve the PBE twice for the
components of the solvation free energy in (\ref{eqn:solvation_energy}). This is computationally demanding,
especially if the biomolecule under investigation is very large, such that the computational domain is
chosen large enough in order to accurately approximate the boundary conditions, leading to high degrees
of freedom $\mathcal{O}(10^6)$ in the discretized PBE \cite{Holst94,Holst:2008}. Notice that by the RS
tensor format, we solve only for the solvated energy ($G_{elec}^{Solv}$) component numerically using the
linearized variant of RPBE in (\ref{eqn:RPBE_decomp}) while the reference energy ($G_{elec}^{Ref}$) is
determined directly from the precomputed long-range component of the free space electrostatic potential
$\mathbf{P}_l$ in (\ref{eqn:Long-Range_Sum}).
However, since the emergence of \cite{Chen:07}, options to calculate the regularized linear and nonlinear
PBE as in (\ref{eqn:PBE_solndecomp}) using the finite element toolkit (FEtk), have also been incorporated
in the APBS, which yields the solvation energy without the need for reference calculations. This is a
positive step in the history of APBS, although, as we observed earlier, the regularization scheme in
(\ref{eqn:PBE_solndecomp}) suffers from the inability to efficiently separate the long- and short-range
components.
A striking feature of the electrostatic energy of interaction is that it is entirely driven by the
long-range electrostatic potential. This is because the short-range components do not communicate with
their neighbours due to their localization (effective local support) in the atomic volumes. To justify
this claim, we prove in Lemma \ref{lemma:Energy_long}, that the solvation free energy calculation
depends only on the long-range components of the free-space potential and that in the solvated state.
\begin{lemma}\label{lemma:Energy_long}
Let the total free-space (reference) electrostatic potential in (\ref{eqn:analytic_PE2}) be given by the
sum $u_{ref}^{tot}(\bar{x}) = u_{ref}^{s}(\bar{x}) + u_{ref}^{r}(\bar{x})$ and that of the solvated state
of the PBE decomposition in (\ref{eqn:RPBE_decomp}) by $u_{solv}^{tot}(\bar{x}) = u_{ref}^{s}(\bar{x})
+ u_{solv}^{r}(\bar{x})$, $\bar{x} \in \mathbb{R}^3$, using the RS tensor splitting scheme. Then the
solvation free energy in (\ref{eqn:solvation_energy}) is given by the regularized form:
\begin{equation}
\Delta G_{solv}^{r} = \frac{1}{2}\sum_{i=1}^{N_m}q_i u_{solv}^{r}(\bar{x}_i) -
\frac{1}{2}\sum_{i=1}^{N_m}q_i u_{ref}^{r}(\bar{x}_i).
\end{equation}
\end{lemma}
\begin{proof} We substitute the components of the total electrostatic potentials from the
\begin{equation}\label{eqn:solvation_energy_expan}
\begin{rcases}
\begin{aligned}
\Delta G_{solv}^{polar} &= G_{elec}^{Solv} - G_{elec}^{Ref},\\
&= \frac{1}{2}\sum_{i=1}^{N_m}q_i u_{solv}^{tot}(\bar{x}_i)
- \frac{1}{2}\sum_{i=1}^{N_m}q_i u_{ref}^{tot}(\bar{x}_i),\\
&= \frac{1}{2}\sum_{i=1}^{N_m}q_i (u_{ref}^{s}(\bar{x}_i) + u_{solv}^{r}(\bar{x}_i)) -
\frac{1}{2}\sum_{i=1}^{N_m}q_i (u_{ref}^{s}(\bar{x}_i) + u_{ref}^{r}(\bar{x}_i)),\\
&= \frac{1}{2}\sum_{i=1}^{N_m}q_i \left(u_{solv}^{r}(\bar{x}_i) - u_{ref}^{r}(\bar{x}_i)\right)\\
&= \Delta G_{solv}^{r}.
\end{aligned}
\end{rcases}
\end{equation}
\end{proof}
The result for the electrostatic potential energy in (\ref{eqn:elec_energy}), which is based on the
effective support of the short-range part of the total potential, was presented in (Lemma 4.2 in
\cite{BKK_RS:18}). Therefore, by applying the RS tensor format to the PBE, we can leverage this property
to greatly reduce the computational costs and errors, which are inherent in the current software packages.
Other electrostatic energies, such as the electrostatic free energy, the folding free energy, and the
binding free energy can be calculated in a similar way \cite{Dong2008}. We shall revisit the calculation
of these energies in detail in forthcoming papers.
\subsection{Electrostatic forces for the PBE}\label{sec:Forces_PBE}
Electrostatic forces on atoms are critical quantities for a range of biomolecular simulations, for
example, the molecular structure optimization and equilibrium molecular dynamics \cite{Wells:2015}.
Electrostatic or polar solvation forces, which are obtained by the derivative of the electrostatic
energy with respect to the interatomic distance \cite{Dong2008,Gilson1993}, can also be computed
efficiently using the results in Lemma \ref{lemma:Energy_long}. The traditional computation of these
forces is given by
\begin{equation}\label{eqn:elec_forces}
\mathbf{f}_i = -\frac{\partial G(r)}{\partial r_i}
= -\sum\limits_j \tilde{r}_{ij}\frac{\partial G(r)}{\partial r_{ij}},
\end{equation}
where $r_{ij}$ is the interatomic distance and $G$ is the electrostatic energy in (\ref{eqn:elec_energy}).
Since the electrostatic forces are long-ranged by nature \cite{Fogolari2002,BeKKKS:18, BKK_RS:16}, it
suffices to compute these forces from the resulting electrostatic solvation energies obtained in
(\ref{eqn:solvation_energy_expan}). We claim this in the following result.
\begin{lemma}\label{lemma:force_long}
Let the regularized electrostatic solvation energy be derived as in (\ref{eqn:solvation_energy_expan}) via the
long-range electrostatic potential component. Then, the corresponding regularized electrostatic
force of interaction $\mathbf{f}^r$ can be computed from the derivative of the regularized solvation
energy $\Delta G_{solv}^{r}$.
\end{lemma}
\begin{proof}
Substituting the regularized solvation energy $\Delta G_{solv}^{r}$ into (\ref{eqn:elec_forces}), we
obtain the required regularized electrostatic forces of interaction, i.e.,
\begin{equation}
\mathbf{f}_i^r = -\frac{\partial (\Delta G^r(\mathbf{r}))}{\partial \mathbf{r}_i}
= -\sum\limits_j \tilde{r}_{ij}\frac{\partial (\Delta G^r(\mathbf{r}))}{\partial r_{ij}},
\end{equation}
which completes the proof.
\end{proof}
\section{Numerical results}
\label{sec:Numer_Tests}
In this section, we consider $n^{\otimes 3}$ 3D uniform Cartesian grids in a box $[-b,b]^3$ with equal step size
$h =2b/(n-1)$ for computing the electrostatic potentials of the PBE on a modest PC with the following
specifications: Intel (R) Core (TM) $i7-4790$ CPU @ 3.60GHz with 8GB RAM. The FDM is used to discretize the
PBE in this work and the numerical computations are implemented in the MATLAB software, version R2017b.
\subsection{Numerical results for LPBE}
\label{ssec:Numer_Tests}
First, we validate our FDM solver for the classical LPBE by comparing its solution with that of
the APBS software package (version 1.5-linux64), which uses the multigrid (PMG) accelerated FDM
\cite{Bakersept2001}. Here, we consider the protein Fasciculin 1, with 1228 atoms. The
\Cref{fig:APBS_FDM_LPBE} shows the electrostatic potential of the PBE on a $n\times n$ grid surface
with $n=129$ at the cross-section of the volume box ($60\,\mbox{\AA}$) in the middle of the $z$-axis computed
by the FDM solver and the corresponding error between the two solutions. Here, we use the ionic strength
of $0.15M$ and the dielectric coefficients $\epsilon_m =2$ and $\epsilon_s = 78.54$, respectively. The
numerical results show that the FDM solver provides as accurate results as those of the APBS with a
discrete $L_2$ error of $\mathcal{O}(10^{-4})$ in the full solution.
\begin{figure}[t]
\captionsetup{width=\linewidth}
\centering
\begin{subfigure}[b]{0.42\textwidth}
\centering
\includegraphics[width=\textwidth]{FDM_PBE_soln_protein.png}
\caption{LPBE solution by the FDM solver.}
\label{fig:LPBE_FDM}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.42\textwidth}
\centering
\includegraphics[width=\textwidth]{APBS_FDM_err_protein.png}
\caption{APBS vs FDM solution error.}
\label{fig:Error_APBS_FDM}
\end{subfigure}
\caption{\label{fig:APBS_FDM_LPBE}
The electrostatic potential for the protein Fasciculin 1 computed by the FDM solver (left)
and the error between the APBS and FDM solutions (right) on $129^{\otimes 3}$ grid, at $0.15M$ ionic
strength.}
\end{figure}
The corresponding electrostatic potential energy for the aforementioned LPBE solvers on a sequence of
fine grids is given in the \Cref{table:Energy_compare}. The results for solvation free energy of protein
varieties are presented in \cite{KwBFM2017}. To validate the claim in \Cref{remk:LPBE_NPBE_diff}, we
provide in the \Cref{table:Energy_LPBE_NPBE}, the comparison between the total electrostatic
potential energies $\Delta G_{elec}$ in kJ/mol, between the LPBE and the nonlinear PBE (NPBE)
computations on a sequence of fine grids using the APBS software package.
\begin{table}[t]
\centering
\captionsetup{width=\linewidth}
\begin{tabular}{|c|c|c|c|c|}
\hline
$h$& $\mathcal{N}$ & $\Delta G_{elec}$, FDM & $\Delta G_{elec}$, APBS
& Relative error\\ \hline
{0.465} & $129^3$ & 91,232.9217 & 91,228.0388 & 5.3524e-5\\ \hline
{0.375} & $161^3$ & 130,611.0021 & 130,606.0444 & 3.7962e-5\\ \hline
{0.320} & $193^3$ & 170,159.4204 & 170,154.3821 & 2.9610e-5\\ \hline
\end{tabular}
\caption{Comparison of the total electrostatic potential energies $\Delta G_{elec}$ in kJ/mol, between FDM
and APBS on a sequence of fine grids.}
\label{table:Energy_compare}
\end{table}
\begin{remark}\label{remk:LPBE_NPBE_diff}
We reiterate that the solutions obtained from the LPBE and the nonlinear PBE are very close to each
other, even when the linearization condition does not hold \cite{Fogolari2002}. This is especially
manifested in protein molecules whose charge densities are small. However, in biomolecules with large
charge densities, for example, the DNA, significant differences might be observed at the solute-solvent
interface \cite{Fogolari99, Fogolari2002}. Moreover, the solution of the LPBE is usually used as the
initial guess for the nonlinear PBE.
\end{remark}
\begin{table}[t]
\centering
\captionsetup{width=\linewidth}
\begin{tabular}{|c|c|c|c|c|}
\hline
$h$& $\mathcal{N}$ & $\Delta G_{elec}$, LPBE & $\Delta G_{elec}$, NPBE
& Relative error\\ \hline
{0.465} & $129^3$ & 91,228.0575 & 91,227.8354 & 2.4345e-6\\ \hline
{0.375} & $161^3$ & 130,606.0630 & 130,605.8448 & 1.6707e-6\\ \hline
{0.320} & $193^3$ & 170,154.4401 & 170,154.1862 & 1.4922e-6\\ \hline
\end{tabular}
\caption{Comparison of the total electrostatic potential energies $\Delta G_{elec}$ in kJ/mol, between the
LPBE and the NPBE computations on a sequence of fine grids.}
\label{table:Energy_LPBE_NPBE}
\end{table}
\begin{remark}\label{remk:Energy_gridsize}
Notice from the \Cref{table:Energy_compare} and the \Cref{table:Energy_LPBE_NPBE} that the electrostatic
potential energies $\Delta G_{elec}$ increase with decreasing grid/mesh size, $h$. This is caused by the
short-range electrostatic potential behaviour in $1/\|\bar{x}\|$ as $\|\bar{x}\|$ $ \to 0$.
\end{remark}
\subsection{Accuracy of the nonlinear RPBE based on the RS tensor format}
\label{ssec:RPBE_RS_accuracy}
Here, we provide the results for the calculation of electrostatic potential for the nonlinear RPBE (NRPBE)
based on the RS tensor format and compare the results with those of the traditional NPBE for various
proteins. This is because the nonlinear RPBE implementation in the APBS package does not currently
function for reasons not specified. On the other hand, the LRPBE variant is not optimized for the in
APBS package and users are recommended to use the multigrid-based methods (which use PMG) for
improved performance. Therefore, we cannot compare the results for the regularized PBE models between
APBS and the RS tensor-based model.
First, we consider the protein Fasciculin 1 consisting of 1228 atoms of varying atomic radii as shown in
the \Cref{table:Fasciculin_atoms}. Notice that 322 of the total atoms have zero radius, which implies that
we must annihilate them from the RS tensor format calculations so that they are not assigned Newton kernels.
Therefore, we consider the smallest atom in the protein as that with 1 $\mbox{\AA}$ radius, (i.e., the
Hydrogen atom).
\begin{table}[t]
\centering
\captionsetup{width=\linewidth}
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
\multicolumn{8}{|c|}{Atomic radii in $\mbox{\AA}$} \\ \hline
Atomic radius & 0.00 & 1.00 & 1.40 & 1.50 & 1.70 & 1.85 & 2.00\\ \hline
Number of atoms & 322 & 333 & 195 & 82 & 104 & 10 & 182\\ \hline
\end{tabular}
\caption{Atomic radii and the corresponding number of atoms for the constituent atoms of the protein
Fasciculin 1.}
\label{table:Fasciculin_atoms}
\end{table}
We provide the comparisons between the electrostatic potential of NRPBE, based on the RS tensor format,
with that of the traditional NPBE. The \Cref{fig:APBS_NPBE_NRPBE_solns} shows the solutions from the
two models and the corresponding error on $129^{\otimes 3}$ uniform Cartesian grid and a
$60\mbox{\AA}$ domain length, at $0.15M$ ionic strength.
\begin{remark}\label{rmk:molecular_solute_solns}
Notice that the error is predominant within the molecular region, where the solution is singular.
However, in the solute region, which is dominated by the long-range regime, the error is small, in
the $\mathcal{O}(10^{-5})$.
\end{remark}
\begin{figure}[t]
\centering
\captionsetup{width=\linewidth}
\resizebox{16cm}{10.0cm}{%
\begin{tikzpicture}
\node (APBS_protein_pot_129) at (0,3.0) {\begin{tabular}{l}
{\includegraphics[width=0.48\textwidth]{APBS_NPBE_classic_protein_129.png}}
\end{tabular}
};
\node (NRPBE_protein_pot_129) at (0,-3.0) {\begin{tabular}{l}
{\includegraphics[width=0.48\textwidth]{FDM_NRPBE_soln_protein_129.png}}
\end{tabular}
};
\node (Error_NPBE_NRPBE_129) at (10,0) {\begin{tabular}{l}
{\includegraphics[width=0.48\textwidth]{APBS_FDM_NRPBE_err_protein_129.png}}
\end{tabular}
};
\draw[->,draw=blue,thick] (3.5,0) -- (Error_NPBE_NRPBE_129.west);
\end{tikzpicture}}
\caption{\label{fig:APBS_NPBE_NRPBE_solns
Absolute error between the solutions of the traditional NPBE and the NRPBE for the protein
Fasciculin 1. Please, insert the data cursor in solute area (for the error).}
\end{figure}
\Cref{fig:APBS_FDM_NPBE_view} provides the cross-sectional view of the electrostatic potential shown
in the \Cref{fig:APBS_NPBE_NRPBE_solns}, for demonstrating the accuracy of the numerical treatment of the
solution singularities inherent in the NRPBE model as compared with
the traditional NPBE model. Notice that the NRPBE is capable of capturing
exactly, the short-range component of the total potential sum because this part is precomputed analytically
thereby avoiding the numerical errors generated by the traditional NPBE solver.
\begin{remark}\label{rmk:singularities_accuracy}
\Cref{fig:NRPBE_FDM} contains densely populated singularities/cusps as a result of explicit
treatment of each atomic charge by the short-range part of the RS tensor
whereas the \Cref{fig:NPBE_APBS}, displays sparsely populated singularities, most of which are not sharp due
to the redundant smoothing/smearing effect of the atomic charges by the cubic spline interpolation.
\end{remark}
\begin{figure}[t]
\captionsetup{width=\linewidth}
\centering
\begin{subfigure}[b]{0.42\textwidth}
\centering
\includegraphics[width=\textwidth]{APBS_NPBE_classic_protein_129_view.png}
\caption{Cross-sectional view of NPBE solution in the \Cref{fig:APBS_NPBE_NRPBE_solns}.}
\label{fig:NPBE_APBS}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.42\textwidth}
\centering
\includegraphics[width=\textwidth]{FDM_NRPBE_soln_protein_129_view.png}
\caption{Cross-sectional view of NRPBE solution in the \Cref{fig:APBS_NPBE_NRPBE_solns}.}
\label{fig:NRPBE_FDM}
\end{subfigure}
\caption{\label{fig:APBS_FDM_NPBE_view}
The cross-sectional view of the electrostatic potentials in the \Cref{fig:APBS_NPBE_NRPBE_solns}.}
\end{figure}
Secondly, we provide results for a 180-residue cytokine solution NMR structure of a murine-human chimera
of leukemia inhibitory factor (LIF) \cite{HinMauZhaNic:98} consisting of 2809 atoms. The corresponding
variation in atomic radii and the corresponding atomic occurrences are shown in the \Cref{table:2809_atoms}.
\Cref{fig:APBS_NPBE_NRPBE_solns_2809} shows the comparison between the electrostatic potential of NRPBE,
with that of the classical NPBE and the corresponding error on a $129^{\otimes 3}$ grid and a
$65\mbox{\AA}$ domain length, at $0.15M$ ionic strength.
\begin{table}[t]
\centering
\captionsetup{width=\linewidth}
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
\multicolumn{8}{|c|}{Atomic radii in $\mbox{\AA}$} \\ \hline
Atomic radius & 0.2245 & 0.4500 & 0.9000 & 1.3200 & 1.3582 & 1.4680 &$\geq$ 1.7000\\ \hline
Number of atoms & 315 & 6 & 6 & 1032 & 54 & 6 & 1390\\ \hline
\end{tabular}
\caption{Atomic radii and the corresponding number of atoms for the constituent atoms of a 180-residue
cytokine solution NMR structure of a murine-human chimera of leukemia inhibitory factor (LIF).}
\label{table:2809_atoms}
\end{table}
\begin{figure}[H]
\centering
\captionsetup{width=\linewidth}
\resizebox{16cm}{10.0cm}{%
\begin{tikzpicture}
\node (APBS_protein_pot_129) at (0,3.0) {\begin{tabular}{l}
{\includegraphics[width=0.48\textwidth]{APBS_NPBE_classic_2809_129.png}}
\end{tabular}
};
\node (NRPBE_protein_pot_129) at (0,-3.0) {\begin{tabular}{l}
{\includegraphics[width=0.48\textwidth]{FDM_NRPBE_soln_2809_129.png}}
\end{tabular}
};
\node (Error_NPBE_NRPBE_129) at (10,0) {\begin{tabular}{l}
{\includegraphics[width=0.48\textwidth]{APBS_FDM_NRPBE_err_2809_129.png}}
\end{tabular}
};
\draw[->,draw=blue,thick] (3.5,0) -- (Error_NPBE_NRPBE_129.west);
\end{tikzpicture}}
\caption{\label{fig:APBS_NPBE_NRPBE_solns_2809
Absolute error between the solutions of the traditional NPBE and the NRPBE for the
murine-human chimera of leukemia inhibitory factor (LIF).
PLEASE insert data cursor for the solute area of the error.}
\end{figure}
\begin{remark}\label{rmk:accuracy_2809}
In a similar vein, we notice in the \Cref{fig:APBS_FDM_NPBE_view_2809} that the error is predominant
within the molecular region, where the solution is singular. The scaling of the total potential
demonstrates that the solution singularity in the NRPBE model is more accurately captured than in the
traditional NPBE model. It is also worth mentioning that the small atomic radii, ($< 0.9 \mbox{\AA}$),
in the \Cref{table:2809_atoms} are treated independently in terms of the RS tensor splitting of the
short- and long-range potentials.
\end{remark}
\begin{figure}[H]
\captionsetup{width=\linewidth}
\centering
\begin{subfigure}[b]{0.42\textwidth}
\centering
\includegraphics[width=\textwidth]{APBS_NPBE_classic_2809_129_view.png}
\caption{Cross-sectional view of NPBE solution in the \Cref{fig:APBS_NPBE_NRPBE_solns}.}
\label{fig:NPBE_APBS_2809}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.42\textwidth}
\centering
\includegraphics[width=\textwidth]{FDM_NRPBE_soln_2809_129_view.png}
\caption{Cross-sectional view of NRPBE solution in the \Cref{fig:APBS_NPBE_NRPBE_solns}.}
\label{fig:NRPBE_FDM_2809}
\end{subfigure}
\caption{\label{fig:APBS_FDM_NPBE_view_2809}
The cross-sectional view of the electrostatic potentials in the
\Cref{fig:APBS_NPBE_NRPBE_solns_2809}.}
\end{figure}
\subsection{Runtimes and Computational Speed-ups}
\label{ssec:RPBE_RS_runtimes}
We compare the runtimes of computing both the classical and regularized PBE models in the
\Cref{table:Runtimes_speedups} for the protein Fasciculin 1 in an $n^3=129^3$ domain of $60\,\mbox{\AA}$
length at an ionic strength of $0.15M$. Notice that the runtimes for the LPBE and the LRPBE are almost
equal because the linear systems are solved by the same solver (i.e., AGMG). On the other hand, the
runtime for solving the nonlinear system for the NRPBE is half that of the NPBE due to the absence of
the Dirac delta distributions and their corresponding solution singularities in
our scheme, and which increase the computational costs in NPBE.
\begin{table}[t]
\centering
\captionsetup{width=\linewidth}
\begin{tabular}{|c|c|c|c|}
\hline
\multicolumn{4}{|c|}{Runtime (seconds) and speed-up} \\
\hline
& LPBE & LRPBE & Speed-up\\
\hline
Solve linear system & 5.26 & 6.34 & $\approx$ 1\\ \hline
Total runtime & 15.25 & 16.47 & $\approx$ 1\\ \hline
& NPBE & NRPBE & Speed-up\\
\hline
Solve nonlinear system & 24.23 & 12.30 & 1.97\\ \hline
Total runtime & 34.40 & 28.30 & 1.21\\ \hline
\end{tabular}
\caption{Runtimes and speed-ups for LPBE, LRPBE, NPBE, and NRPBE.}
\label{table:Runtimes_speedups}
\end{table}
\section{Conclusions}\label{sec:Conclusions}
In this paper, we apply the RS tensor format for a solution decomposition
of the nonlinear PBE for computation of electrostatic potential of large solvated biomolecules.
The efficacy of the tensor-based regularization scheme established in \cite{BeKKKS:18} for the linear PBE,
is based on the unprecedented properties of the grid-based RS tensor splitting
of the Dirac delta distribution \cite{khor-DiracRS:2018}. Similar to the linear case, the key computational
benefits are attributed to the localization of the modified Dirac delta distributions within the molecular
region and the automatic maintaining of the continuity of the Cauchy data on the solute-solvent interface.
Moreover, our computational scheme entails solving only a single system of algebraic equations for the
regularized component of the collective electrostatic potential discretized by the FDM.
The total potential
is obtained by adding this solution to the directly precomputed low-rank tensor representation of the
short-range contribution.
The main properties of the presented scheme are demonstrated by the various numerical tests.
For instance, the \Cref{fig:APBS_FDM_NPBE_view} and the \Cref{fig:APBS_FDM_NPBE_view_2809}
vividly demonstrate that the traditional PBE model does not accurately capture the solution
singularities which originate from the short-range component of the total target electrostatic
potential in the numerical approximation. In the RPBE, the Dirac delta distribution is replaced by a
smooth long-range function from (\ref{eqn:Dirac_splitting}). It only requires one to solve for the
long-range electrostatic potential numerically and add this solution to the short-range component
which is computed a priori using the canonical tensor approximation to the Newton kernel. The
resultant total potential sum is of high accuracy which demonstrated by \Cref{fig:NRPBE_FDM} and the
\Cref{fig:NRPBE_FDM_2809}.
\section*{Acknowledgement}
The authors thank the following organizations for financial and material support on
this project: International Max Planck Research School (IMPRS) for Advanced Methods in Process
and Systems Engineering and Max Planck Society for the Advancement of Science (MPG).
\begin{footnotesize}
\bibliographystyle{unsrtnat}
|
\section{Introduction}
There is an increasing surge of interest in analyzing biomedical data to improve health. Biostatisticians and machine learning researchers are keen to access personal health information for a deeper understanding of diagnostics, disease development, and potential preventive or treatment options \cite{telenti2020treating}.
In the US, healthcare and clinical data are often collected by local institutions. For many situations, combining these datasets would increase statistical power in hypothesis testing and provide better means to investigate regional differences and subpopulation bias (e.g., due to differences in disease prevalence or social determinants). However, such an information harmonization process needs to respect the privacy of individuals, as healthcare data contain sensitive information about personal characteristics and health conditions. As a minimum requirement, HIPAA (Health Insurance Portability and Accountability Act)\cite{hippa} specifies PHIs (protected health information) and regulations to de-identify the sensitive information (i.e., safe harbor mechanism). But HIPAA compliance does not mean full protection of the data, as several studies demonstrated re-identifiability of HIPAA de-identified data \cite{bonomi2018linking, janmey2018re, sweeney2017re}. Ethical healthcare data sharing and analysis should also respect the “minimum necessary” principle to reduce the unnecessary risk of potential data leakage, which might increase the likelihood of information leakage.
\begin{figure}[H]
\centering
\includegraphics[scale=0.4]{fed.png}
\caption{\textbf{Schema of federated learning model in multiple geographically distributed healthcare institutions.} The local institutions periodically exchange intermediate statistics and update the convergence situation of the global model.}
\label{federated}
\end{figure}
The recent development of federated learning, which intends to build a shared global model without moving local data from their host institutions (Fig. \ref{federated}), shows good promise in addressing the challenge in data sharing mentioned above. Despite the exciting progress, there is still an important limitation as existing models cannot effectively handle mixed-effects (i.e., both fixed and random effects), which is very important to analyzing non-independent, multilevel/hierarchical, longitudinal, or correlated data. \textcolor{black}{Also, due to the sampling errors (i.e., smaller sample size in local sites), variances from these local statistics are larger than those of the global model. These issues, if not addressed appropriately, would lead to failure in global optimization.} The goal of this paper is to improve existing techniques and provide practical solutions with open-source implementation and to allow ordinary biomedical/healthcare researchers to build federated mixed effect learning models for their studies.
\section{Related Work}
Federated learning for healthcare data analysis is not a new topic, and there have been many previous studies. However, most of the existing methods assume the observations are independent and identically distributed\cite{glm, regression}.
In the presence of non-independence due to hierarchical structures (e.g., due to institutional or regional differences), existing federated models have strong limitations in ignoring the regional differences. The generalized linear mixed model (GLMM), which takes the heterogeneous factors into consideration, is more amenable to accommodate the heterogeneity
across healthcare systems. There have been very few studies in this area and one relevant work is a privacy-preserving Bayesian GLMM model\cite{zhu2020privacy}, which proposed an Expectation-Maximization (EM) algorithm to fit the model collaboratively on horizontally partitioned data. The convergence process is relatively slow (due to the Metropolis–Hastings sampling in the E-step) and it is also not very stable (likely to be trapped in local optima\cite{xu1996convergence} in high-dimensional data). In the experiment, a loose threshold (i.e., 0.08) was used as a convergence condition \cite{zhu2020privacy} while typical federated learning algorithms \cite{glore} in healthcare use much stringent convergence threshold (i.e., $10^{-6}$).
Another related work to fit GLMM in a federated manner is the distributed penalized quasi-likelihood (dPQL) algorithm\cite{luo2021dpql}. This algorithm reduces the computational complexity by considering the target function of penalized quasi-likelihood, which is motivated from Laplacian approximation. The model has communication efficiency over the EM approach and can converge in a few shots. However, the target function PQL can have first order asymptotic bias \cite{lin1996bias} due to the Laplacian approximation of the integrated likelihood.
There is an alternative strategy, Gauss-Hermite (GH), which supports high-order approximation. It is computationally more intensive and requires special techniques to handle the numerical instability of the logSumExp operation (due to the overflow issue when the dimensionality grows in the sum of the exponential terms). We will explain both models in this manuscript and compare their performance on simulated and real-world data.
\section{Methods}
In this section, we will discuss the statistic model along with challenges to be tackled. A high-level schema of the method is shown in algorithm \ref{alg1}.
\subsection{Notation}
Before we introduce the formation of GLMM, let us define some notations.
\begin{table}[H]
\centering
\begin{tabular}{rl|rl}
$i$ & Index of sites & $l_i$ & Log-likelihood function for site $i$ \\
$j$ & Index of patients in a specific site & $\bm\beta$ & Parameters of fixed effect \\
$k$ & Index of Hermite polynomial & $\mu_i$ & Parameters of random effect in site $i$ \\
$K$ & Order of Hermite polynomial & $\tau$ & Hyper-parameters \\
$m$ & Number of sites & $\bm\theta$ & Parameter space $(\bm\beta,\tau)$ \\
$n_i$ & Number of patients in site $i$ & $X_{ij}$ & \begin{tabular}[c]{@{}l@{}}A vector represents the data of $j$-th \\ patient in $i$-th site\end{tabular} \\
$\mathcal L_i$ & Likelihood function for site $i$ & $y_{ij}$ & The outcome of patient $j$ from site $i$ \\
$\lambda$ & The parameter of regularization term & $p$ & Number of variables
\end{tabular}
\end{table}
\subsection{Fitting GLMM with quasi-likelihood}
Let us provide the formation of the GLMM. Define $\mathbb P$ is the distribution of interest and depending on patient-level data $X_{ij},\ y_{ij}$. Define $\phi$ as the distribution of random effects. We can compose the joint distribution as following
\[
\prod_{j=1}^{n_i}\mathbb P(\bm\theta|X_{ij},y_{ij})\phi(\mu_{i};\tau)
\]
Now we have the log-likelihood function of the joint distribution:
\begin{equation}\label{problem}
\log\{\mathcal L(\bm\theta)\}=\sum_{i=1}^{m}\log\biglbrac{\int_{\mu_{i}}\bigmbrac{\prod_{j=1}^{n_i}\mathbb P(\bm\theta|X_{ij},y_{ij})}\phi(\mu_{i};\tau)\text{d}\mu_{i}}
\end{equation}
From the log-likelihood function Eq. \eqref{problem}, one can see that it does not support direct linear decomposition.
In order to support federated learning, we will leverage approximation strategies to make the objective linearly decomposable with simple summary statistics.
We will compare Laplace approximation and Gauss-Hermite approximation in the following sections.
\subsection{Laplace (LA) approximation}
\textcolor{black}{
With the help of Laplace approximation, the integration from Eq.\eqref{problem} can be approximated by an exponential family expression.
\begin{equation}\label{eq1}
\int_{\mu_{i}}f_{\theta}(\mu_{i})\text{d}\mu_{i}=\int_{\mu_{i}}e^{\log{f_{\bm\theta}(\mu_{i})}}\text{d}\mu_{i}\triangleq\int_{\mu_{i}}e^{g(\mu_{i},\theta)}\text{d}\mu_{i}
\end{equation}
After the deduction (\ref{laplace_approx}), the intractable problem is solved and the objective is to maximize the following formula with respect to $\theta$, where $g$ is an exponential family function defined above (Eq.\eqref{eq1})
}
\[
\sum_{i=1}^{n_i}\bigbrac{g(\hat\mu_{i},\bm\theta)-\dfrac{n_i}{2}\log\bigbrac{g_{\mu\mu}(\hat\mu_{i},\bm\theta)})}
\],
for which the terms are linearly decomposable from local sites. Site $i$ needs to calculate the following aggregated data:
\begin{itemize}
\item $p\times p$ matrix:
\begin{equation}\label{la_hessian}
\dfrac{\hat\omega_{\beta\beta}\hat\omega-\hat\omega_\beta\hat\omega_{\beta}}{\hat\omega^2}+\hat\mu_{\beta\beta}g_\mu+\hat\mu_\beta(\hat\mu_{\beta}g_{\mu\mu}+g_{\mu\beta})+\hat\mu_{\beta}g_{\mu\beta}+g_{\beta\beta}
\end{equation}
\item $p$ - dim vector:
\begin{equation}\label{la_score}
\dfrac{\hat\omega_\beta}{\hat\omega}+\hat\omega^2g_{\mu\beta}(\hat\mu_i)g_\mu+g_\beta
\end{equation}
\item scalar of random effect: $\hat\mu_i$ and first order derivative of $\tau$ by
\[
\dfrac{\hat\omega_\tau}{\hat\omega}+\hat\omega^2g_{\mu\tau}(\hat\mu_i)g_\mu+g_\tau
\]
\end{itemize}
where $\hat\omega=\sqrt{-\dfrac{1}{g_{\mu\mu}(\hat\mu_{i0})}}$
\subsection{Gauss-Hermite (GH) approximation}
Gauss-Hermite approximation \cite{liu1994note} implements Hermite interpolation concerning Eq. \eqref{eq1}.
\textcolor{black}{
And after the deduction in \ref{guass_hermite_approx}, notice that when the order of Hermite polynomial $K=1$, the objective function is identical to the method with Laplace approximation. Because GH is more generalizable, we will describe the distributed federated learning model on the GLMM problem with the formation of Gauss-Hermite approximation Eq. \eqref{GHA}. For each site $i$, the followings need to calculate and transmit:
}
\begin{itemize}
\item $p\times p$ matrix:
\begin{equation}\label{gh_hessian}
\dfrac{\hat\omega_{\beta\beta}\hat\omega-\hat\omega_\beta\hat\omega_{\beta}}{\hat\omega^2}+\dfrac{1}{\sum_{k=1}^Kf_k}\sum_{k=1}^K\dfrac{\partial}{\partial\bm\beta}(f_{k_\mu}\hat\mu_\beta+f_{k_\omega}\hat\omega_\beta+f_{k_\beta})-\dfrac{1}{(\sum_{k=1}^Kf_k)^2}\|\sum_{k=1}^l(f_{k_\mu}\hat\mu_\beta+f_{k_\omega}\hat\omega_\beta+f_{k_\beta})\|^2_2
\end{equation}
\item $p$ - dim vector:
\begin{equation}\label{gh_score}
\dfrac{\hat\omega_\beta}{\hat\omega}+\dfrac{1}{\sum_{k=1}^Kf_k}\sum_{k=1}^K(f_{k_\mu}\hat\mu_\beta+f_{k_\omega}\hat\omega_\beta+f_{k_\beta})
\end{equation}
\item scalar of random effect: $\hat\mu_i$ and first order derivative of $\tau$ by
\[
\dfrac{\hat\omega_\tau}{\hat\omega}+\dfrac{1}{\sum_{k=1}^Kf_k}\sum_{k=1}^K(f_{k_\mu}\hat\mu_\tau+f_{k_\omega}\hat\omega_\tau+f_{k_\tau})
\]
\end{itemize}
\subsection{\textcolor{black}{Training} Penalization GLMM with GH approximation}
The convergence of the approximation of the likelihood function may be compromised due to over-fitting. Also, for those spatially correlated data, the convergence of them may lead to a complex model. Hence, $L2$ regularization is added to the local log-likelihood function of Gauss-Hermite approximation form, and as shown below
\begin{equation}\label{local_max}
l_i=\log\mathcal L_i=\log\bigbrac{\sqrt{2\pi}\hat\omega\sum_{k=1}^Kh_k\exp\biglbrac{g(\hat\mu_{i}+\sqrt{2\pi}\hat\omega x_k;\bm\theta)+x_k^2}} - \lambda\|\bm\beta\|^2_2
\end{equation}
note that when $K=1$, it is represented as regularized Laplace approximation to the problem. To evaluate and find the optimum $\lambda$, we steadily increased the value of $\lambda$ in range $[0,10]$ by $1$. Set $\lambda_{\text{opt}}$ as the optimized regularization term with largest $\sum_i^ml_i$. And choose $\hat{\bm\beta}_{\text{opt}}$ as the optimized estimator for $\bm\beta$.
Due to the limited computation digits, computers are not able to calculate the correct results of the local log-likelihood function $l_i$ of the Gauss-Hermite approximation form as stated above. Such problem is also known as the Log-Sum-Exponential problem and can be solved by shifting the center of the exponential sum for easier computation,
\[
\log\sum_{k=1}^K\exp\biglbrac{g(\hat\mu_{i}+\sqrt{2\pi}\hat\omega x_k;\bm\theta)+x_k^2}=a+\log\sum_{k=1}^K\exp\biglbrac{g(\hat\mu_{i}+\sqrt{2\pi}\hat\omega x_k;\bm\theta)+x_k^2-a}
\]
where $a$ is an arbitrary number.
Thus, the global problem of maximizing $\sum_i^ml_i$ can be divided into several local maximization problems \eqref{local_max}. Each local site $i$ will update the regression intermediates, and they will be combined to update the iteration status.
Specifically, in each iteration of the federated GLMM algorithm, the following statistics are exchanged from each site to contribute aggregated data for the global model
\begin{table}[H]
\centering
\begin{tabular}{ll}
\hline
\multicolumn{1}{c}{LA} & \multicolumn{1}{c}{GH} \\ \hline
number of variables $p$ & number of variables $p$ \\
$p\times p$ matrix (Eq. \ref{la_hessian}) & $p\times p$ matrix (Eq. \ref{gh_hessian}) \\
$p$ - dim vector (Eq. \ref{la_score}) & $p$ - dim vector (Eq. \ref{gh_score}) \\
$p$ - dim vector $\bm\beta$ & $p$ - dim vector $\bm\beta$ \\
scalar $\lambda$ & scalar $\lambda$ \\
scalar $\hat\mu_i$ & scalar $\hat\mu_i$ \\
\textcolor{black}{scalar first order derivative $\tau$} & \textcolor{black}{scalar first order derivative $\tau$} \\
& \textcolor{black}{ scalar $K$} \\ \hline
\end{tabular}
\end{table}
Detailed derivatives with the logistic regression setting of the optimization are presented in the appendix \ref{optimizaiton}.
\begin{algorithm}[H]\label{alg1}
\SetAlgoLined
\KwData{Local data $X_i$ from site $i$}
\KwResult{Global model with coefficients $\hat{\bm\beta}_\text{global}$, test P-values, upper and lower bound}
Initialization: coefficients $\bm\beta$, random effects $\mu_i$ and $\tau$, regularization term $\lambda$\;
\For{$\lambda=0$ to $10$}{
\While{$\Delta\mu\neq 0$}{
Maximized $\mu_i$ with respect to $\bm\theta$
\While{$\Delta\bm\theta\neq 0$}{
1. Approximate the log-likelihood functions $l_\text{approx}$ with Laplace or Gauss-Hermite;
2. Calculate intermediate statistics. For Laplace, Eq. \ref{la_hessian}, \ref{la_score}; for Gauss-Hermite, Eq. \ref{gh_hessian}, \ref{gh_score};
3. Send the intermediate statistics to center server, and then the center server will aggregate them;
4. Update $\bm\theta$ and $l_\text{approx}({\bm\theta},\lambda)$ in center server and send back to each client $i$;
}
}
}
\textbf{Return:} The largest $l_{\text{approx}}(\hat{\bm\theta},\hat\lambda)$, the coefficients $\hat{\bm\beta}_\text{global}=\hat{\bm\beta}$, \textcolor{black}{hyperparameter $\hat\tau$}, and the regularization term $\hat\lambda
\caption{Distributed GLMM with approximation methods}
\end{algorithm}
\section{Experiments}
\textcolor{black}{Our algorithm is developed in Python with packages \textit{pandas}, \textit{numpy}, \textit{scipy}, and the benchmark algorithm is glmer function in R package `lme4'.}
\subsection{Benchmarking the methods using synthetic data}
To test the performance of our proposed methods, we first designed a stress test based on a group of synthetic data, which include 8 different settings (Tab.\ref{summary_data}), and each set contains 20 datasets. In each dataset, it consists of 4 categorical variables with value in $\{0,1\}$; 6 categorical variables with value in range $[-1,1.5]\in\mathbb R$; 1 outcome variable with value in $\{0,1\}$; Site ID, represents the id of which site the entry belongs to; Site sample size, represents the number of samples in this specific setting; Log-odds ratio for each sample; Number of true positive, true negative, \textcolor{black}{true} positive, false positive, false negative.
To evaluate which method can reach better performance, we proposed the following evaluation measurements: discrimination of the estimated coefficients $\hat{\bm\beta}$, the test power of each coefficient, and the precision and recall of the number of significant coefficients.
\begin{table}[H]\label{summary_data}
\caption{{The summary of data in each setting.}}
\centering
\begin{tabular}{cccc}
\hline
\multicolumn{1}{c}{Setting} & \multicolumn{1}{c}{\begin{tabular}[c]{@{}c@{}}Number of \\ sites\end{tabular}} & \multicolumn{1}{c}{\begin{tabular}[c]{@{}c@{}}Sample size \\ in each site\end{tabular}} & \multicolumn{1}{c}{variance} \\ \hline
1 & 2 & 500 & small \\
2 & 2 & 500 & large \\
3 & 10 & 500 & small \\
4 & 10 & 500 & large \\
5 & 2 & 30 & small \\
6 & 2 & 30 & large \\
7 & 10 & 30 & small \\
8 & 10 & 30 & large \\ \hline
\end{tabular}
\label{summary_data}
\end{table}
The valuation experiments were conducted among federated GLMM with Laplace approximation, federated GLMM with Gauss-Hermite approximation, and centralized GLMM (all of the data stored in single host) in the R package. And the stress test will be run in 160 different datasets in 8 different settings as mentioned in Tab.\ref{summary_data}. \textcolor{black}{All of the data in different settings were randomly separated into training sets and validation sets with a ratio of 7:3. And we trained the federated learning model on training data sets, then by slowly increasing the regularization term $\lambda$, we chose the optimum model with the best Akaike information criterion and Bayesian information criterion performance on the validation sets. All testing was performed on 2017 iMac with 16 GB memory, CPU (4.2 GHz Quad-Core Intel Core i7), macOS Big Sur version 11.6, Python 3.8, and R version 3.5.0.}
Although we tested the data sets with the state-of-art benchmark algorithm for centralized GLMM in R, the regression is not perfect for the ground truth coefficients we used to generate the data (Fig.\ref{coef}). So, it is also important to have the P-values of variables into consideration when interpreting the model.
Thus, We made comparisons among centralized GLMM, Laplace method, and Gauss-Hermite method concerning the p-values of coefficients. Tables in the appendix captured the performance of different methods. Fig.\ref{precision} shows the precision and recall results of centralized, Laplace, and Gauss-Hermite methods. Noted that we set our Gauss-Hermite approximation to 2-degree. See tables in Appendix (Tab.\ref{R}, \ref{LA}, \ref{GH}).
\begin{figure}[H]
\centering
\includegraphics[scale=0.6]{coef_diff.png}
\caption{\textbf{The difference from coefficients to the true parameters that are used to generate data.} (Left) The distributed GLMM with Laplace approximation; (Middle) The distributed GLMM with 2-degree Gauss-Hermite approximation. Reminds that $X_1$ is the intercept; (Right) The benchmark of centralized GLMM in R package. }
\label{coef}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.48]{precision.png}
\includegraphics[scale=0.48]{recall.png}
\caption{\textbf{The precision and recall among centralized, Laplace, and Gauss-Hermite method under significance level $\alpha=0.05$.} (Left) The precision of the test compared to the true value. (Right) The recall of the test compared to the true value. }
\label{precision}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.6]{power.png}
\caption{\textbf{The curve of test power among centralized, Laplace, and Gauss-Hermite methods.} (Left) The power of the test of the Laplace method. (Middle) The power of the test of the 2-degree Gauss-Hermite method. (Right) The power of the test of the Centralized method. Power was calculated as the two-sided t-test on p-values among different methods.}
\label{power}
\end{figure}
\begin{table}[h]
\centering
\caption{\color{black}\textbf{The convergence rates on approximation methods LA and GH.} (Both LA and GH held the same convergence threshold $10^{-3}$. The mean values and standard deviations (in parentheses) were given)}
\color{black}\begin{tabular}{c|cc|cc}\hline
& \multicolumn{2}{|c|}{LA} & \multicolumn{2}{c}{GH} \\ \hline
Setting & Steps & Runtime (s) & Steps & Runtime (s) \\ \hline
1 & 22.875 (21.623) & 47.953 (20.513) & 34.850 (9.213) & 104.460 (10.614) \\
2 & 21.500 (21.977) & 40.947 (36.466) & 35.000 (8.711) & 100.940 (19.940) \\
3 & 29.867 (31.719) & 108.931 (65.486) & 34.900 (6.138) & 1259.285 (231.956) \\
4 & 27.846 (24.034) & 84.343 (76.502) & 36.650 (6.310) & 1342.695 (250.603) \\
5 & 59.722 (42.057) & 10.631 (3.945) & 33.750 (10.146) & 12.568 (2.116) \\
6 & 67.188 (48.994) & 10.499 (4.054) & 31.400 (11.081) & 11.430 (3.064) \\
7 & 96.286 (53.635) & 96.501 (38.632) & 37.450 (3.818) & 369.165 (41.998) \\
8 & 116.083 (46.479) & 91.304 (62.410) & 37.150 (4.295) & 309.693 (36.621) \\ \hline
\end{tabular}
\label{convergence}
\end{table}
The simulation results showed the federated Gauss-Hermite approximation performed better than the method based on Laplace approximation on every variable. Also, the federated Gauss-Hermite method achieved higher test power (Fig.\ref{power}). \textcolor{black}{When considering the convergence rates between the two approximation methods, both showed less convergence efficiency in Setting 7 and 8 (Tab.\ref{convergence}). The result Indicates that more local sites and smaller sample sizes will make the federated GLMM more inefficient to converge. Also, GH approximation method will required more computation time compared with LA approximation.} In sum, one-degree increase of the approximation function in LA with our developed GH method, GH outperformed LA methods for federated GLMM implementation.
\subsection{Mixed-effects logistic regression on mortality for patients with COVID-19}
We analyzed the data of COVID-19 electronic health records collected by Optum$^\circledR$ from February 2020 to January 28, 2021, from a network of healthcare providers. The dataset has been de-identified and based on HIPAA statistical de-identification rules and managed by Optum$^\circledR$ customer data user agreement. In this database, there are 56,898 unique positive tested COVID-19 patients. After removing the patients with missing data, the final cohort contains 4,531 patients who died and the rest population (41,781) survived. The database contains a regional variable with five levels (Midwest, Northwest, South, West, Others/unknown) to provide privacy-preserving area information to indicate where the samples were collected.
We have conducted a GLMM model (considering region-distinct random effect) using this dataset with the following predictors: age, gender, race, ethnicity, Chronic obstructive pulmonary disease (COPD), Congestive heart failure (CHF), Chronic kidney disease (CKD), Multiple sclerosis (MS), Rheumatoid arthritis (RA), LU (other lung diseases), High blood pressure (HTN), ischemic heart disease (IHD), diabetes (DIAB), Asthma (ASTH), obesity (Obese). Our proposed method with GH approximation performed the best with both the smallest \textcolor{black}{Akaike information criterion} (AIC) and \textcolor{black}{Bayesian information criterion} (BIC) according to the table of the goodness of fit (Tab.\ref{goodness}). \textcolor{black}{And the performance of different methods can be shown in (Tab.\ref{measures}).}
\begin{table}[H]
\centering
\caption{\textbf{Statistics of goodness of fit among different methods}}
\centering
\begin{tabular}{rccc}
\hline
\multicolumn{1}{l}{} & Log-likelihood & AIC & BIC \\ \hline
R & 13562.9 & 27165.9 & 27340.8 \\
LA & -13695.0 & 27428.0 & 27594.1 \\
\textbf{GH} & \textbf{-11.8} & \textbf{61.6} & \textbf{227.7} \\ \hline
\end{tabular}
\label{goodness}
\end{table}
\begin{figure}[H]
\centering
\includegraphics[scale=0.55]{all_roc.png}
\caption{\textbf{The ROC curve with Area Under Curve (AUC) among centralized, Laplace, and Gauss-Hermite methods.} The orange ROC curve is the centralized method without regularization and the Laplace approximation(i.e., R implementation in the `lme4' package, which does not have an option for including regularization). AUC values are also included, a higher AUC value implicates better performance of the model. The green ROC curve is the 2-degree Gauss-Hermite method with regularization.}
\label{roc}
\end{figure}
\color{black}
\begin{table}[h]
\centering
\caption{\color{black}\textbf{Statistics of performances among different methods}(95\% CIs were generated by Wilson Score interval)}
\color{black}\begin{tabular}{cllllll}
\hline\hline
\multicolumn{1}{l}{} & & Precision & Recall & F1-score & AUC & threshold \\ \hline\hline
\multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}Centralized\\ method\\ with LA\end{tabular}} & Value & 0.1507 & 0.6204 & 0.2425 & 0.6789 & 0.0900 \\
& Lower bound (0.95) & 0.1474 & 0.6160 & 0.2386 & 0.6700 & \\
& Upper bound (0.95) & 0.1539 & 0.6248 & 0.2464 & 0.6878 & \\ \hline
\multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}GH\\ with\\ regularization\end{tabular}} & Value & 0.1705 & 0.6546 & 0.2705 & 0.7178 & 0.0108 \\
& Lower bound (0.95) & 0.1670 & 0.6503 & 0.2664 & 0.7091 & \\
& Upper bound (0.95) & 0.1739 & 0.6589 & 0.2745 & 0.7265 & \\ \hline\hline
\end{tabular}
\label{measures}
\end{table}
\color{black}
We also compared the ROC curves (Fig.\ref{roc}) between our proposed GH method and centralized method to check their performance. And the result showed that GH approximation (AUC=0.72) outperforms the centralized method without regularization (AUC=0.68). Indicating GH-based GLMM method has better classification performance than the GLMM based on LA approximation.
In our proposed model, it showed variables: Unknown race, Chronic kidney disease (CKD), Multiple sclerosis (MS), and other lung diseases (LU) are not significant to the mortality of COVID-19. The result of the regression is in the Appendix (Tab.\ref{r_result}, \ref{gh_result}, \ref{la_result}).
\section{Discussion}
\textcolor{black}{We developed solutions to address the limited digit problem (i.e., overflow issue of fixed-length object types due to extremely large numbers in local estimation) using an alternative loss-less estimation of log-sum-exponential term, and the singularity issue (involved in Newton optimization) with an adaptive regularization strategy to avoid inverting low-rank matrices without imposing too much unnecessary smoothness.}
We \textcolor{black}{further} compared two federated GLMM algorithms \textcolor{black}{with our developed federated solutions} (LA vs. GH) and demonstrated the performance of the federated GLMM based on the GH method surpassed the method based on LA in terms of the accuracy of estimation, power of tests, and AUC. Although the GH method is requiring slightly more computations than the LA method, it is still acceptable for more accurate results. For example, in the prediction of COVID-19 mortality rates, the accuracy of prediction will be more reliable, as we have shown in the previous section. During the optimization iterations, we noticed that some sites have already achieved convergence in very few steps. If those sites stop communicating with the central server, they can be released from extra computations. We would investigate more efficient algorithms based on such a strategy of `lazy regression' for minimizing communication for federated learning models.
Another limitation of the proposed federated GLMM model is not yet differentially private and iterative summary statistics exchange can lead to incremental information disclosure, which might increase the re-identification risk over time. There are several strategies to improve the model based on secure operations like homomorphic encryption and differential privacy, which we have previously studied in GLM models \cite{Kim2019-gs}. Finally, in practice, there can be extra heterogeneity that cannot be explained by random intercepts only, it is of interest to further develop our algorithms toward GLMM that allows multiple random effects including random coefficients in the regression models.
\section{Acknowledgement}
XJ is CPRIT Scholar in Cancer Research [RR180012]; and he was supported in part by Christopher Sarofim Family Professorship, UT Stars award, UTHealth startup, the National Institute of Health (NIH) under award number [R01AG066749, R01GM114612, U01TR002062]; and the National Science Foundation (NSF) RAPID [\#2027790]. NM is supported by NSERC Discovery Grants [RGPIN-2015-04147]. JT and YC's research is supported in part by NIH under award number [R01AI130460 and R01LM012607] and Patient-Centered Outcomes Research Institute (PCORI) Project Program Award (ME-2019C3-18315). All statements in this report, including its findings and conclusions, are solely those of the authors and do not necessarily represent the views of the Patient-Centered Outcomes Research Institute (PCORI), its Board of Governors or Methodology Committee.
\bibliographystyle{elsarticle-num}
|
\section{Introduction \label{Sec.Introduction}}
There is ongoing interest in investigating the correlations between the normal-state and superconducting properties of heavy fermion (HF) superconduconductors when subjected to a variation in a nonthermal control parameter such as
pressure $p$, concentration of charge carriers $n$, stoichiometry/doping $x$, disorder, or magnetic field $H$.\cite{Gegenwart08-QC-HF-metals,Coleman07-HF-Review,Jaccard99-HF-SC-Mag-HP-Review,Yuan03-CeCu2Si2-SC-Phases-HF,Lohneysen07-FL-Instability-QPT-Reivew}
Driven by ease and convenience, the evolution of these correlations is often investigated within a \textit{T-p} phase diagram; there,
an extrapolation of the phase boundaries down to zero temperature often leads to one or two quantum
critical/crossover points.\cite{Gegenwart08-QC-HF-metals,Coleman07-HF-Review,Jaccard99-HF-SC-Mag-HP-Review,Yuan03-CeCu2Si2-SC-Phases-HF,Lohneysen07-FL-Instability-QPT-Reivew,Mathur98-HF-SC-Mag-Mediation,Miyake17-HF-QCF-Scattering-Review} One is usually attributed to a magnetic instability at $p_{m}$; often, it is accompanied by a superconducting dome wherein the superconductivity is considered to be driven by spin fluctuations.\cite{Mathur98-HF-SC-Mag-Mediation} The other point, at $p_{v}$, is usually assigned to a valence instability\cite{Miyake17-HF-QCF-Scattering-Review} which, as well, is often accompanied by a superconducting dome with the superconductivity considered to be driven by valence fluctuations.\cite{Holmes04-CeCu2Si2-ValenceFluctuation, Miyake17-HF-QCF-Scattering-Review}
A surge of superconducting dome within the neighborhood of a magnetic quantum critical point is also manifested in other superconducting family, such as high-Tc cuprates and Fe-based pnictides and chalcogenides.\cite{Pines13-SpinFluctuation-SUC, Scalapino12-UnconSCs-CommonThread} Much of the understanding of the superconducting and normal-state properties were obtained from theoretical and experimental investigation of three regions of the dome: the two neighboring the emergence and disappearance of the superconductivity (at the left and right $T_c \rightarrow 0$) while the third is at the middle wherein $T_c$ is a maximum. Analysis of the character of the left and middle regions is usually complicated by the additional influence of competing antiferromagnetic, spin glass, charge and stripes, or pseudogap phases. In contrast, investigation of the region at the right $T_c \rightarrow 0$ limit is much simpler since only two states are involved, namely the superconducting and the FL phases.\cite{Maier20-Overdoped-End-Cuprates-PhaseDiagram,Li21-Overdoped-End-Cuprates-SUC-to-Normal,Dean13-SpinFluctuation-Underdoped-Overdoped,Lee20-Overdoped-Cuprate-Landau-BCS,Lee17-Cuperate-Overdoped-Disorder-Superfluid-Density}
This work, in line with Refs.~\onlinecite{Maier20-Overdoped-End-Cuprates-PhaseDiagram,Li21-Overdoped-End-Cuprates-SUC-to-Normal,Dean13-SpinFluctuation-Underdoped-Overdoped,Lee20-Overdoped-Cuprate-Landau-BCS,Lee17-Cuperate-Overdoped-Disorder-Superfluid-Density}, investigate the evolution of (and correlation among) the superconducting and normal states within the FL region of the HF \textit{T-p} phase diagram. As shown below, in contrast to these references, (i) we consider explicitly the normal-state FL character, (ii) we do not invoke any additional impurity scattering; rather we consider the spin-fluctuations to be operating within patches of the sample (the average dimension of a patch is longer than the mean free path and coherent length while their concentration is reflected in the excess residual resistivity), and most importantly (iii) we consider that the very same fluctuation-mediated electron-electron scattering channel is responsible for the evolution of both the superconductivity and the FL character.
It is worth recalling that various experiments reveal the presence of phase-boundary curve, e.g. $T^{*}_\text{{FL}}(p^{*})$ within the HF {\it T-p} phase diagram, across which both
the transport and thermodynamic properties switch over from non-Fermi-liquid (NFL) into a Fermi-liquid (FL) behavior.
Such a switch is best illustrated by the electronic contribution to the low-temperature resistivity, $ \rho (T,p)$,
of typical HF superconductors: Within the NFL phase,
\begin{equation}\label{Eq-total-resistivity-NFL}
\rho(T>T^{*}_{\text{{FL}}}[p^{*}], p\leq p^{\text{*}})-\rho_{x} = \rho_{\circ} +A_{\text{NFL}}T^{n}, \mathrm{\, \, \, \, \, \, [n<2],}
\end{equation}
while within the FL phase,
\begin{equation} \label{Eq-total-resistivity-FL}
\rho(T_{c}<T<T^{*}_{\text{FL}}[p^{*}],p>p^{\text{*}})-\rho_{x} =\rho_{\circ}+AT^{2},
\end{equation}
where $\rho_{x}=\rho_{\circ}^{n.f.}+\rho_{n.f.}(T)$ represents all non-fluctuation-related (e.g. defects, phonon, magnetic) contributions (see end of Subsec.~\ref{subsec.Sketch-Analysis}). Within the FL region (which starts at the switching of NFL into FL character and ends when $T_c \rightarrow 0$), we denote the excess contribution of the residual resistivity by $\rho_{\circ}$, that of the {\it quadratic-in-T} coefficient by $A$, and that of the superconductivity transition by $T_c$; all these contributions are attributed to a fluctuation-mediated electron-electron (e-e) scattering channel (see below).
Generally, these $\rho_{\circ}(p),\, A(p),$ and $T_{c}(p)$ parameters \textemdash $\,$ derived from experimental $\rho (T,p)$ curves \textemdash manifest a dramatic and non-monotonic variation around each of the critical $p_{m}$ and $p_{v}$ points (see, e.g., Figs.~\ref{Fig2-CeCu2Ge2},\ref{Fig3-CeCu2Si2},\ref{Fig4-CeCoIn5}).
In this work, we are interested in analyzing and rationalizing the baric evolution of $\rho_{\circ}(p)$, $A(p)$, and $T _{c}(p)$ within the FL region of various archetype HF superconductors. Our analysis identified two universal correlations among these $\rho_{\circ}$, $A$, and $T_c$ parameters. We argue that the fluctuation-mediated electron-electron scattering channel is responsible for the emergence of the superconductivity, the FL character and the correlations among them.
The text below is organized as follows. We first recall some helpful insight regarding the fluctuation-mediated e-e scattering process.\cite{Mathur98-HF-SC-Mag-Mediation,Miyake17-HF-QCF-Scattering-Review} In Subsec.~\ref{SubSec.Emprical-Derivation} we extract, identify, and generalize two main empirical correlations. In Subsec.~\ref{subSec.Theoretical-Derivation} we argue that these
correlations are driven by a fluctuation-mediated e-e interaction
channel.
Such fluctuations can be either spin fluctuation\cite{Mathur98-HF-SC-Mag-Mediation} or valence fluctuation\cite{Miyake17-HF-QCF-Scattering-Review} (generically referred to as fluctuation). Although we consider below the spin-fluctuation exchange mechanism \cite{Mathur98-HF-SC-Mag-Mediation}, generalization to valence fluctuation\cite{Miyake17-HF-QCF-Scattering-Review} is implicitly assumed. On adopting such a mechanism and applying standard Migdal-Eliashberg description of superconductivity and Boltzmann's transport theory, we derive analytic expressions that compare favorably with the empirically-obtained correlations.
Comparison to other fluctuation-bearing/defect-bearing superconductors will be briefly discussed in Sec.~\ref{Sec.Discussion-Conclusion}; there, we demonstrate the generality of our approach by discussing the Kadowaki-Woods and gap-to-$T_c$ ratios of these HF superconductors.
\begin{figure}[hbtp]
\centering
\includegraphics[scale=0.20,trim={0cm 0cm 0cm 0cm}, clip]{Fig1-Magnetic-Defectals-Model.eps}%
\caption{ A sketch of a fluctuation-mediated e-e interaction channel.
(a) An antiferromagnet structure representing an initial state in the phase diagram of a typical heavy fermion superconductor. The tiny arrows denote localized magnetic moments of atomic entities (represented by the small solid circles). (b) The nonmagnetic state after the quench of localized magnetic moments and, consequently, the magnetic structure.
Within the direct lattice, the two electrons with ${\bf k}_1$ and ${\bf k}_2$ scatter into final states ${\bf k}_1^\prime$ and
${\bf k}_2^\prime$.
(c) A $\bf {\delta g}= 0$ two-electrons scattering process in the reciprocal space: As that ${\bf q}={\bf k}_1^\prime-{\bf k}_1-{\bf g}$, the phase space for scattering is quite limited; a Fermi-liquid sate is possible only at very low temperatures.
(d) The $\bf {\delta g}\neq 0$ two-electrons scattering process in the reciprocal space: as that ${\bf q}={\bf k}_1^\prime-{\bf k}_1-{\bf g}-\delta{\bf g}$, the phase space available for scattering is considerably enlarged, leading to a robust superconducting and Fermi-liquid states. (e) Feynman diagram of the $\bf {\delta g}= 0$ process [as in panel~(c)] wherein only longitudinal modes are involved. (f) The Feynman diagram of the $\bf {\delta g}\neq 0$ process [as in panel~(d)] which is mediated by exchanging all kinematically unconstrained modes.
}\label{Fig1-Magnetic-Defectals-Model}
\end{figure}
\section{Analysis \label{Sec.Analysis}}
\subsection{Some Preliminaries on the fluctuation-mediated e-e interaction channel \label{subsec.Sketch-Analysis}}
It is recalled that a typical \textit{T-p} phase diagram of most quantum-critical HF superconductors exhibits a series of distinct electronic states \big(see, e.g., Figs.~\ref{Fig2-CeCu2Ge2}(a)-\ref{Fig3-CeCu2Si2}(a) and Refs.~[\onlinecite{Jaccard99-HF-SC-Mag-HP-Review,Holmes04-CeCu2Si2-ValenceFluctuation,Sidorov02-CeCoIn5-Q-Criticality}]\big).
Often, the initial state is an antiferromagnetic, represented in Fig.~\ref{Fig1-Magnetic-Defectals-Model}(a) as simple two-dimension Neel-type arrangment.
Beyond a critical value of the control parameter, $p_m$ in the present cases, the localized moments and the three-dimensional magnetic structure are quenched leading to a series of "nonmagnetic" states which can be of Kondo-type, non-conventional superconducting, non-FL, or FL character.
Nevertheless, some remnant magnetic fluctuations persist within various micro-sized patches which are distributed randomly within an otherwise metallic matrix.\cite{Note-Patch-Formation}
These magnetic fluctuations are considered to mediate an e-e scattering process \cite{Mathur98-HF-SC-Mag-Mediation}: as shown in Figs.~\ref{Fig1-Magnetic-Defectals-Model}(b-f), the two electrons with ${\bf k}_1$ and ${\bf k}_2$ scatter into final states ${\bf k}_1^\prime$ and
${\bf k}_2^\prime$. In particular, the electron, initially at a state ${\bf k}_1$, goes into a final state ${\bf k}_1^\prime$ after being scattered by a mode with wavevector ${\bf q}$; in doing so, it transfers an amount of energy and quasi-momentum satisfying ${\bf k}_1^\prime-{\bf k}_1-{\bf q}= {\bf g}+\bf {\delta g}$ where ${\bf g}$ is a vector in a reciprocal magnetic lattice while $ \bf{\delta g}$ is an uncertainty in ${\bf g}$ due to the inherent nature of the process leading to fluctuation (most obvious for the doping process).
We consider that, within the Fermi-liquid state of these HF superconductors, an increase in pressure leads to a reduction of the size and density of the fluctuation-bearing patches,\cite{Note-Patch-Formation} till eventually one reaches, at very high pressure, the fluctuation-free state wherein no modes are available for mediation. This, as will be detailed below in Figs.~\ref{Fig2-CeCu2Ge2}, \ref{Fig3-CeCu2Si2},~and~\ref{Fig4-CeCoIn5}, leads to a complete removal of fluctuation-related features, namely the superconductivity ($T_c \rightarrow 0$), the Fermi-liquid state ($A \rightarrow A_{\circ}$), the residual resistivity ($\rho \rightarrow \rho_{\circ}^{\circ}$), and their correlations.
Let us now discuss the fluctuation-based mediation within the FL region. First, we consider the case satisfying $\bf{\delta g}=0$ condition of Figs.~\ref{Fig1-Magnetic-Defectals-Model}(c)~and~\ref{Fig1-Magnetic-Defectals-Model}(e). Here, the quasi-momentum ${\bf q}={\bf k}_1^\prime-{\bf k}_1-{\bf g}$ is conserved exactly. Accordingly, only longitudinal modes with a well defined polarization, $\hat{{\bf e}}({\bf q}={\bf k}_1^\prime-{\bf k}_1)$, will be involved. This means that the phase space available for momentum relaxation will be quite limited leading to, if any, a very small $T_c$ and $A$.
Secondly, in contrast to above, the surge of fluctuation-based mediation satisfying ${\bf \delta g}\neq 0$ condition, Figs.~\ref{Fig1-Magnetic-Defectals-Model}(d)~and~\ref{Fig1-Magnetic-Defectals-Model}(f), provides a source for short wavelength (large ${\bf q}$) phase interference. This $\bf {\delta g}\neq 0$ condition implies that the quasi-momentum ${\bf q}={\bf k}_1^\prime-{\bf k}_1-{\bf g}-\delta{\bf g}$ is no longer conserved since $\bf {\delta g}$ is arbitrary.\cite{Bergmann71-Disorderd-SCs, Bergmann76-SC-AmorphousMetals-Review,Gurvitch86-Disorder-induced-transition-AT2}
Then, multiple modes \big[longitudinal and transverse, of all polarizations
$\hat{{\bf e}}({\bf q}\neq {\bf k}_1^\prime-{\bf k}_1-{\bf g})$\big] become kinematically available, Fig.~\ref{Fig1-Magnetic-Defectals-Model}(f). As shown in Fig.~\ref{Fig1-Magnetic-Defectals-Model}(d), the phase space available for momentum relaxation is considerably enlarged: this leads to an enhanced $T_c$ and $A$.
Considering the above partitioning of a fluctuation-bearing sample into two spacial regions, Matthiessen's rule suggests a sum of two types of contributions: (i) one type is related to normal (non-fluctuation-bearing) patches; these are the residual resistivity $\rho_{\circ}^\circ$, the coefficient $A_{\circ}$, the coupling constant $\lambda_{\circ}$, and the mean free path $\ell _{\circ}$ ($\ell_{\circ} \propto \frac{1}{\rho_{\circ}^{\circ}}$). (ii) The other type is related to fluctuation-bearing patches within which, we consider, a fluctuation-mediated e-e scattering channel is operating and as such leading to excess contributions denoted as $A$, $\lambda$, $\ell$, and $\rho_{\circ}$. Below we consider $\rho_{\circ}$ to be arising from those batches and as such it measures the strength of the fluctuation-related channel while $\rho_{\circ}^\circ$ as a measure of all normal processes.
The total contribution is denoted as $\rho_{\circ}^{tot}$, $A_{tot}$, $\lambda_{tot}$, or $\ell_{tot}$. Often, the non-fluctuation-bearing $\rho_{\circ}^{\circ}$ and $A_{\circ}$ contributions are obtained by extrapolation. In our case here, these values (shown as large green circle in Figs.~\ref{Fig2-CeCu2Ge2},~\ref{Fig3-CeCu2Si2},~and~\ref{Fig4-CeCoIn5} and tabulated in Table~\ref{Tab-Fit-Values-A-Ro-Theta-F}) were estimated from the very-high-pressure region whereat $T_c \rightarrow 0$, indicating, we assume, the quench of the fluctuations. Then each excess, fluctuation-related contribution $X = X_{tot} -X_{\circ}$, is obtained from \cite{note-X-approx-Xtot}
\begin{subequations}
\begin{align}
&\rho_{\circ} = \rho_{\circ}^{tot}-\rho_{\circ}^{\circ} , \\
& A= A_{tot}- A_{\circ} , \\
&\lambda = \lambda_{tot}-\lambda_{\circ} , \\
&\ell = 1/(\frac{1}{\ell_{tot}}-\frac{1}{\ell_{\circ}}) .
\end{align}
\label{Eq-X-fluctuation-is-Xtot-Xo}
\end{subequations}
For theoretical analysis, it is more convenient to measure the strength of the fluctuation-related scattering channel via the effective mean free path, the {\it scaling length $\ell \propto 1/\rho_{\circ}$}.
Then any variation in the control parameter (e.g. pressure, alloying, or defect incorporation\cite{Note-Heavy-Fermion-SUC-2021-NonMag-Impurities}) would be reflected in $\ell$ and manifested as a variation in the superconductivity, the FL character, and the correlation among $\rho_{\circ} (\ell)$, $T_c (\ell)$ and $A (\ell)$; within the FL region of Figs.~\ref{Fig2-CeCu2Ge2}-\ref{Fig4-CeCoIn5}, a pressure increase leads to a reduction of these parameters indicating an increase in fluctuation-related $\ell$. The functional dependence of $\ell$ on, e.g., pressure will not be discussed in this work.
\begin{figure*}[hbtp]
\centering
\includegraphics[scale=0.6,trim={0cm 0cm 0cm 0cm}, clip]{Fig2-CeCu2Ge2.eps}%
\caption{
\textit{T-p} phase diagram and baric evolution of the parameters of CeCu$_{2}$Ge$_2$ (data taken from Ref.~\onlinecite{Jaccard99-HF-SC-Mag-HP-Review}). (a) A semi-log \textit{T-p} phase diagram showing the evolution of $T_{N}(p)$, $T_{c}(p)$, and the exponent \textit{n(p)}. Vertical dashed arrows represent $p_{m} \approx 9.4\,$ GPa and $p_{v} \approx 16\,$ GPa.\cite{Jaccard99-HF-SC-Mag-HP-Review} The dashed lines are visual guides while the curved arrows signal the direction of pressure increase. The red symbols and hatched area mark the FL region. Evolution of (b) $T_{c}(p)$, (c) $A_{tot}(p)$ in a semi-log plot, and (d) $\rho_{\circ}^{tot}(p)$.
As discussed in text, for $p \rightarrow$20~GPa, the pressure-induced strong reduction of the fluctuation leads to $T_c \rightarrow T_c^{\circ} \approx 0$, $A \rightarrow A_{\circ}$ and $\rho_{\circ}^{tot}\rightarrow \rho_{\circ}^{\circ}$: the limits at 20~GPa (large green circles) are shown in Table~\ref{Tab-Fit-Values-A-Ro-Theta-F}.
(e) Correlation of $[A_{tot}(p)-A_{\circ}]$ with $[\rho_{\circ}^{tot}-\rho_{\circ}^{\circ}]$. The solid line is a best fit to Eq.~\ref{Eq-A-vs-rho}: $A_0$ and $A_2$ given in Table~\ref{Tab-Fit-Values-A-Ro-Theta-F}.
(f) $T_c$ \textit{versus} $[\rho_{\circ}-\rho_{\circ}^{\circ}]$: The solid line is calculated based on Eq.~(\ref{Eq-Tc-vs-(Ro-Roo)}) of Subsec.~\ref{SubsubSec.Tc-A-Correlation}. (g) $ln(T_{c}) \textit{ versus } (A_{tot}-A_{\circ})^{-\frac{1}{2}}$. The solid red line is the best fit to linearized Eq.~\ref{Eq-Tc-of-A}; $\theta$ and $\mathcal{F}$ parameters are given in Table~\ref{Tab-Fit-Values-A-Ro-Theta-F}.
}
\label{Fig2-CeCu2Ge2}
\end{figure*}
\begin{figure*}[hbtp]
\centering%
\includegraphics[scale=0.6,trim={0cm 0cm 0cm 0cm}, clip]{Fig3-CeCu2Si2.eps}%
\caption{ (a) The \textit{T-p} phase diagram and baric evolution of (b) $T_{c}(p)$, (c) $A_{tot}(p)$, and (d) $\rho_{\circ}^{tot}(p)$ of CeCu$_{2}$Si$_{2}$. All curves were taken from the continuous line (the visual guide) of Ref.~\onlinecite{Holmes04-CeCu2Si2-ValenceFluctuation}. For a discussion on the critical pressures $p_m$ and $p_v$ and the related superconductivity regions (SC~1 and SC~II) see Ref.~\onlinecite{Holmes04-CeCu2Si2-ValenceFluctuation}. The red symbols and hatched area mark the FL region while the dashed lines are visual guides. It is noted that, due to non-stoichiometry, defects, impurities or disorder, the reported parameters of CeCu$_{2}$Si$_{2}$ vary considerably from sample to sample; nonetheless, the overall evolution as well as the derived correlations are consistently similar.\cite{Rosch99-disorder-SpinFluctuation-QCP,Kambe97-CeCu2Si2-Purity,Sheikin00-CeCu2Si2-HF-R-P,Gegenwart98-CeCu2Si2-Brakup-HF} %
As $p \rightarrow 7$~GPa, $T_c \rightarrow 0$, $A \rightarrow A_{\circ}$ and $\rho_{\circ}\rightarrow \rho_{\circ}^{\circ}$ (see text): the extrapolation to 7~GPa (large green circles) are shown in Table~\ref{Tab-Fit-Values-A-Ro-Theta-F}.
(e) A plot of [$A(p)-A_{\circ}$] \textit{versus} [$\rho_{\circ}(p)-\rho_{\circ}^{\circ}$]. The best fit to Eq.~\ref{Eq-A-vs-rho} is shown as solid red curve; the fit parameters $A_0$ and $A_2$ are given in Table~\ref{Tab-Fit-Values-A-Ro-Theta-F}. (f) $T_c$ \textit{versus} [$\rho_{\circ}(p)-\rho_{\circ}^{\circ}$]: For the solid calculated line, see Eq.~(\ref{Eq-Tc-vs-(Ro-Roo)}) in Subsec.~\ref{SubsubSec.Tc-A-Correlation}. (g) $ln(T_{c})$ \textit{versus} $(A-A_{\circ})^{-\frac{1}{2}}$. The straight red line is a best fit to linearization of Eq.~(\ref{Eq-Tc-of-A}): $\theta$ and $\mathcal{F}$ are given in Table~\ref{Tab-Fit-Values-A-Ro-Theta-F}.
}%
\label{Fig3-CeCu2Si2}
\end{figure*}
\begin{table*}
\scriptsize
\begin{center}
\caption{Representative fit parameters of CeCu$_{2}$Ge$_{2}$, CeCu$_{2}$Si$_{2}$ and CeCoIn$_{5}$. Values of A$_o$ and A$_2$ were obtained from a fit of Eq.~\ref{Eq-A-vs-rho} to Figs.~\ref{Fig2-CeCu2Ge2}(e),\ref{Fig3-CeCu2Si2}(e),\ref{Fig4-CeCoIn5}(e)): As $A_1 \approx 0$, it was fixed as $A_1 \equiv 0$.
For all compounds, $A_2$, $\theta$ and $\mathcal{F}$ are close to each other, except $A_2$ of \ce{CeCoIn5} which is orders of magnitude higher than the others: this together with the high value of $T_c$ is most probably related to a relatively high strength of the fluctuation-related e-e scattering channel of \ce{CeCoIn5}.
It is worth noting that the strength and evolution of the fit parameters depend critically on the criteria for determining \textit{T$_{c}$} (10-90\%, 50\%, onset, zero point etc.) and $A$ (the sample geometry, which is often not precisely determined). Additionally, the widely different methods for sample synthesis and thermal treatment adopted by different groups give rise to a corresponding variation in sample-dependent properties.
Although these non-fluctuation-related material properties are accounted for by
$T_{\circ}$, $A_{\circ}$, and $\rho_{\circ}^{\circ}$, it is expected to lead to a wide scatter in the fit parameters. The scatter in the experimental curves leads to a large error value.}
\begin{tabular}[c]{ccccccc}
\hline\hline
$HFs$ & $\rho _{\circ}^{\circ}$ & A$_o$ & A$_2$ & $ \theta$ & $\mathcal{F}$ \\
- & $\mu \Omega \text{cm}$ & $\mu \Omega \text{cm/K}^{2}$ & $ \times 10^6 \Omega^{-1} \text{cm}^{-1}/\text{K}^{2}$& K & $\times 10^{-3} (\Omega \text{cm/K}^{2})^{1/2}$\\ %
\hline
CeCu$_2$Ge$_2$ & 10(1) & 2.9(5)$\times 10^{-3}$ & 2.0(2)$\times 10^{-4}$ & 4.4(2) & 0.36(1) \\ %
CeCu$_2$Si$_2$ & 18.5(5) & 0.010(5) & 7.0(5)$\times 10^{-4}$ & 2.6(1) & 0.49(1) \\ %
CeCoIn$_5$ & 0.130(5) & 0.082(4) & 6.0(2) & 3.2(1) & 0.18(1) \\ %
\hline
\label{Tab-Fit-Values-A-Ro-Theta-F}%
\end{tabular}
\end{center}
\end{table*}
\subsection{Empirical analysis: extraction of correlations among \texorpdfstring{$\rho_{\circ}$, $A$, and \textit{T$_{c}$}}{} \label{SubSec.Emprical-Derivation}}
Below, we review the baric evolution of $\rho_{\circ}(p)$, $A(p)$, and $T _{c}(p)$ of three representative HF superconductors for which detailed resistivity curves $\rho_{tot}(T,p)$ (essentials for determining these $\rho_{\circ}$, $A$, and $T_c$ parameters) were reported. Special attention will be directed towards identifying the correlations.
\subsubsection{\texorpdfstring{\textrm{CeCu$_{2}X_{2}$ ($X$=Si, Ge)}}{} \label{SubsubSec.CeCu2X2}}
The \textit{T-p} phase diagram and the baric evolution of $T_c$, $A_{tot}$, and $\rho_{\circ}^{tot}$ of \ce{CeCu2Ge2} are shown in Figs.~\ref{Fig2-CeCu2Ge2}(a)-\ref{Fig2-CeCu2Ge2}(d) (Refs.~\onlinecite{Jaccard99-HF-SC-Mag-HP-Review,Honda13-CeCu2Ge2-Mag-SC}) while those of \ce{CeCu2Si2} in Figs.~\ref{Fig3-CeCu2Si2}(a)-\ref{Fig3-CeCu2Si2}(d) (Refs.~\onlinecite{Yuan03-CeCu2Si2-SC-Phases-HF,Gegenwart98-CeCu2Si2-Brakup-HF,Bellarbi84-CeCu2Si2-Valence-Instability, Rueff11-CeCu2Si2-VQCP,Holmes04-CeCu2Si2-ValenceFluctuation}). As evident, these plots highlight the dramatic baric evolution in particular around the critical points.
The phase diagram of each of \ce{CeCu2X2} manifests a FL character at sufficiently higher pressure; this is evidenced as a {\it quadratic-in-T} resistivity contribution, identified in Figs.~\ref{Fig2-CeCu2Ge2}(a)~and~\ref{Fig3-CeCu2Si2}(a) by the hatched area and the {\it n=2} notation.
At 20~GPa for \ce{CeCu2Ge2} and 7~GPa for \ce{CeCu2Si2},
$T_c \rightarrow 0$; we identify this as a signal of a pressure-induced quench of the fluctuation-related FL contributions. Based on Eqs.~\ref{Eq-X-fluctuation-is-Xtot-Xo}, all parameters extrapolate to non-fluctuation-related contributions:
$A\rightarrow A_{\circ}$ and $\rho_{\circ} \rightarrow \rho_{\circ}^{\circ}$ (see Figs.~\ref{Fig2-CeCu2Ge2}~and~\ref{Fig3-CeCu2Si2} and Table~\ref{Tab-Fit-Values-A-Ro-Theta-F}).
Then for the purpose of empirical identification of any possible correlation {\it within the FL region}, we plot $[A_{\text{tot}}(p)- A_{\circ}] \, \, \it{versus} \,\, [\rho_{\circ}^{tot}(p)-\rho_{\circ}^{\circ}]$ in Figs.~\ref{Fig2-CeCu2Ge2}(e) and \ref{Fig3-CeCu2Si2}(e)], $T_c \, \, \it{versus} \,\, [\rho_{\circ}^{tot}(p)-\rho_{\circ}^{\circ}]$ in Figs.~\ref{Fig2-CeCu2Ge2}(f) and \ref{Fig3-CeCu2Si2}(f)], and $ln[T_{c}(p)] \, \, \it{versus} \,\, 1/\sqrt{A_{\text{tot}}(p)-A_{\circ}}$ in Figs.~\ref{Fig2-CeCu2Ge2}(g) and \ref{Fig3-CeCu2Si2}(g).
Figures~\ref{Fig2-CeCu2Ge2}(e) and \ref{Fig3-CeCu2Si2}(e) show that within {\it the FL state, well above the critical pressures region}, \cite{Gegenwart08-QC-HF-metals,Coleman07-HF-Review,Miranda05-disorder-driven-NFL-Review,Sheikin00-CeCu2Si2-HF-R-P,Gegenwart98-CeCu2Si2-Brakup-HF} one obtains
\begin{equation}
(A - A_{\circ}) \approx A_{2} (\rho_{\circ}-\rho_{\circ}^{\circ}) ^{2}, \label{Eq-A-propto-rho2}
\end{equation}
The $A_{\circ}$ and $A_{2}$ parameter of each \ce{CeCu2X2} are given in Table~\ref{Tab-Fit-Values-A-Ro-Theta-F}. It is noted that the absence of a linear-in-$\rho_{\circ}$ term rules out any Koshino-Taylor contribution.
On the other hand, Figs.~\ref{Fig2-CeCu2Ge2}(g)~and~\ref{Fig3-CeCu2Si2}(g) indicate that
\begin{equation}
\ln(T_{c}) \propto (A-A_{\circ})^{-\frac{1}{2}}.
\label{Eq-lnTc-versus-Inv-sqrtA}
\end{equation}
A closer look at Eqs.~(\ref{Eq-A-propto-rho2}~and~\ref{Eq-lnTc-versus-Inv-sqrtA}) suggests that a relation between $T_c$ and $(\rho_{\circ}-\rho_{\circ}^{\circ})$ can be derived; indeed the solid red curves in Figs.~\ref{Fig2-CeCu2Ge2}(f)~and~\ref{Fig3-CeCu2Si2}(f) were calculated based on Eq.~\ref{Eq-Tc-vs-(Ro-Roo)} which is a derived equation (see Subsec.~\ref{SubsubSec.Tc-A-Correlation}).
\begin{figure*}[hbtp]
\centering%
\includegraphics[scale=0.6,trim={0cm 0cm 0cm 0cm}, clip]{Fig4-CeCoIn5.eps}%
\caption{(a) \textit{T-p} phase diagram and baric evolution of (b) $T_{c}(p)$, (c) $A_{tot}(p)$, and (d) semilog plot of $\rho_{\circ}^{tot}(p)$ of CeCoIn$_5$.
All curves were taken from the continuous line (the visual guide) of Ref.~\onlinecite{Sidorov02-CeCoIn5-Q-Criticality}. For more detail on this phase diagram, in particular, the position of critical $p_m$, see Refs.~\onlinecite{Sidorov02-CeCoIn5-Q-Criticality,Paglione03-CeCoIn5-H-QCP, Bianchi03-CeCoIn5-H-QCP, Bauer05-CeCoIn5-xSn-x-HCP}.
The red symbols and hatched area identify the FL region while the dashed lines are visual guides.
For $p\rightarrow$4.4~GPa, $T_c\rightarrow 0$, $A\rightarrow A_{\circ}$ and $\rho_{\circ}\rightarrow \rho_{\circ}^{\circ}$ (see text and Table~\ref{Tab-Fit-Values-A-Ro-Theta-F}).
(e) A plot of $A_{tot}-A_{\circ}$ \textit{versus} $\rho_{\circ}^{tot}-\rho_{\circ}^{\circ}$. The solid curve is the best fit to Eq.~\ref{Eq-A-vs-rho} (see Table~\ref{Tab-Fit-Values-A-Ro-Theta-F}). (f) $T_c$ \textit{versus} ($\rho_{\circ}^{tot}-\rho_{\circ}^{\circ}$): The solid calculated line is based on Eq.~(\ref{Eq-Tc-vs-(Ro-Roo)}).
(g) $ln(T_{c})$ \textit{versus} $(A-A_{\circ})^{-\frac{1}{2}}$. The best fit to the linearization of Eq.~(\ref{Eq-Tc-of-A}) (see Table~\ref{Tab-Fit-Values-A-Ro-Theta-F} for the fit parameters).
}\label{Fig4-CeCoIn5}
\end{figure*}
\subsubsection{\texorpdfstring{CeCoIn$_{5}$ }{} \label{SubsubSec.CeCoIn5}}
The \textit{T-p} phase diagram of CeCoIn$_{5}$ [Refs.~\onlinecite{Sidorov02-CeCoIn5-Q-Criticality,Ronning06-CeCoIn5-Pressure-QCP,Paglione03-CeCoIn5-H-QCP,Bianchi03-CeCoIn5-H-QCP}] is shown in Fig.~\ref{Fig4-CeCoIn5}(a) while the baric evolution of $T_{c}(p)$, $A_{tot}(p)$ and $\rho_{\circ}^{tot}(p)$ are shown in Figs.~\ref{Fig4-CeCoIn5}(b),~\ref{Fig4-CeCoIn5}(c)~and~\ref{Fig4-CeCoIn5}(d), respectively.
A comparison of the features of CeCoIn$_{5}$ with those of CeCu$_{2}X_{2}$ ($X$=Si, Ge) suggests an overall similarity in that:
\begin{itemize}
\item Figure~\ref{Fig4-CeCoIn5}(e) indicates that $A(p)-A_{\circ}$ is large but still \textit{quadratic-in-$[\rho_{\circ}(p)-\rho_{\circ}^{\circ}$}],
\item Figure~\ref{Fig4-CeCoIn5}(f) reveals that the evolution of $T_c[\rho_{\circ}(p)-\rho_{\circ}^{\circ}]$ can be described by Eq.~\ref{Eq-Tc-vs-(Ro-Roo)} (see Subsec.~\ref{SubsubSec.Tc-A-Correlation}).
\item Figure~\ref{Fig4-CeCoIn5}(g) shows that $\ln[T_{c}(p)]$ is linear in $[A(p)-A_{\circ}]^{-\frac{1}{2}}$ .
\end{itemize}
\begin{figure*}[hbtp]
\centering
\includegraphics[scale=0.6,trim={0cm 0cm 0cm 0cm}, clip]{Fig5-Tc-A-AllHFS.eps}%
\caption{(a) The generalized plot of $\frac{T_{c}(p)}{\theta}$\textit{versus} $\lambda=\frac{(A-A_{\circ})^{\frac{1}{2}}}{\mathcal{F}}$ of the tabulated HF superconductors. \cite{Jaccard99-HF-SC-Mag-HP-Review, Holmes04-CeCu2Si2-ValenceFluctuation,Sheikin00-CeCu2Si2-HF-R-P,Sidorov02-CeCoIn5-Q-Criticality, Kotegawa06-CeNiG3-SC-AFM,Araki02-CeRh2Si2-SC-pressure,Nakashima04-CeNiGe3,Nakashima06-Ce2Ni3Ge5-SC-Pressure,Onuki08-CePt3Si-CeIr3Si,Hassinger08-URu2Si2-HiddenOrder-Nesting,deVisser84-UPt3-FL,
Ott83-UBe13-Unconv-SC}
(b) For comparison, we include the plot of the same relation in conventional superconductors (data taken from Ref.~[\onlinecite{Nunes12-FermiLiquid-SUC}]). (c) A collapse of data from all HF and conventional superconductors on the single $T_c/\theta =exp(-1/\lambda)$ curve: an unsurprising result considering panels (b and c); nevertheless, considering the diversity of the superconducting materials, this plot is highly nontrivial.
Each pair of sample-dependent $\theta$ and $\mathcal{F}$ was evaluated within the pressure range wherein Eq.~(\ref{Eq-Tc-of-A}) is valid [see Figs.~\ref{Fig2-CeCu2Ge2}(g), \ref{Fig3-CeCu2Si2}(g), and \ref{Fig4-CeCoIn5}(g)]. Representative values of $\theta$ and $\mathcal{F}$ are shown in Table~\ref{Tab-Fit-Values-A-Ro-Theta-F}. %
As evident, the distribution region of the pairs ($T_{c}/\theta$, $\lambda$) in both HF and strong-coupled superconductors are similar, with the upper limit being $\approx (0.6,2.1)$ (for $\lambda$ limit of strong superconductors, see Ref.~\onlinecite{Gurvitch86-Disorder-induced-transition-AT2} and also Fig.~4 of
Ref.~{\onlinecite{Nunes12-FermiLiquid-SUC}}). As mentioned for \ce{CeCu2X2} and \ce{CeCoIn5}, $\lambda$ is corrected using $\frac{(A-A_{\circ})^{\frac{1}{2}}}{\mathcal{F}}$. Similar plot, but with no $\frac{(A-A_{\circ})^{\frac{1}{2}}}{\mathcal{F}}$ correction, was reported for both pnictide and chalcognides superconductors.\cite{18-Castro-Tc-A-Correlation,18-KxFe2-ySe2} Finally, it is emphasized that the relatively high values of $\frac{T_{c}(p)}{\theta}$ (those close to 1) should not be substituted in Eq.~\ref{Eq-gap-to-Tc-ratio} since this equation is valid only within the $\frac{T_c}{\theta} \ll 1$ condition.\cite{Carbotte90-SCs-Boson-Exchanged-Reivew}
}%
\label{Fig5-Tc-A-AllHFS}
\end{figure*}
\subsection{Theoretical analysis: interpretation of the correlations among \texorpdfstring{$\rho_{\circ}$, $A$, and \textit{T$ _{c}$} }{} \label{subSec.Theoretical-Derivation}}
The aforementioned experimental evidences emphasize that in spite of the wide differences in material properties of these HF superconductors, one observes:
(i) A similarity in the overall evolution of the phase diagrams as well as in the baric evolution of their $T_{c}(p)$, $A(p)$, and $\rho_{\circ}(p)$.
(ii) A nontrivial manifestation of neighboring superconductivity and FL phases.
(iii) Two correlations (expressed in Eqs.~\ref{Eq-A-propto-rho2}~and~\ref{Eq-lnTc-versus-Inv-sqrtA}) and a derived one; all manifested in Figs.~\ref{Fig2-CeCu2Ge2},\ref{Fig3-CeCu2Si2},\ref{Fig4-CeCoIn5}.
Below we discuss the significance of these features in terms of a fluctuation-mediated e-e scattering channel which, due to a modification in the kinematic constraints, is endowed with a significantly enlarged phase space for scattering, much larger than the traditional Baber or Umklapp e-e scattering channels.\cite{21-FL-SC-Defectal-Reconciled}
Furthermore, the empirical analysis highlights the distinct and nontrivial contrast between the properties of the FL state of a HF superconductor and that of conventional weakly-correlated superconductor: It is remarkable that $A$ is more than five orders of magnitude higher than the contribution expected for a typical Fermi-liquid metal ($A_{\circ} \simeq 10^{-7}\mu\Omega$cm/K$^2$).\cite{Lawrence73-ee-Scattering-in-Res,Lawrence76-ee-Scattering-NobleMetals,MacDonald80-e-F-enhanced-e-e-Interaction,MacDonald81-Umkalpp-resistivity, Patton75-FermiLiquid-Superfluid,Pethick86-FL-Theory-UPt3}
Another contrast is that a conventional FL superconductor, in contrast to a HF one, does not manifest a correlation among $\rho_{\circ}$, $T_c$ and $A$: $\rho_{\circ}$ is determined by the electron-impurity scattering, $\rho_{\circ}\sim|V_{imp}|^2$; $T_c\sim e^{-1/\lambda}$ is associated with electron-phonon coupling, $\lambda \sim |V_{ep}|^2$; while $A$ depends on e-e interaction, $A\sim |V_{ee}|^2$.
It is worth emphasizing that, although two exceptional cases were reported to yield a similar BCS-like correlation among $T_c$ and $A$ (see Ref.~\onlinecite{note-Two-exceptions} and Subsec.~\ref{SubSec.Emprical-Derivation}), none of them manifest a correlation between $A$ and $\rho_{\circ}$ nor between $T_c$ and $\rho_{\circ}$; in fact, Anderson theorem excludes the latter relation. It is, then, quite puzzling that the aforementioned empirical analysis, see Figs.~\ref{Fig2-CeCu2Ge2}-\ref{Fig4-CeCoIn5}, delineate a distinct FL phase within which one identifies a surge of superconductivity and pressure-dependent correlations among $\rho_{\circ}$, $A$ and $T_c$. This puzzle can be resolved if \textit{the manifestation of superconductivity and FL state as well as the correlation of $T_c$ (hallmark of superconductivity) and $A$ (hallmark of normal FL state) are assumed to be driven by a retarded, spin-fluctuation \cite{Mathur98-HF-SC-Mag-Mediation} or valence-fluctuation \cite{Miyake17-HF-QCF-Scattering-Review} mediated e-e interaction.}
This is reminiscent of the defect-induced, phonon-mediated e-e scattering in defectal-bearing systems (see Ref.~\onlinecite{21-FL-SC-Defectal-Reconciled})
The basic idea is discussed in Subsect.\ref{subsec.Sketch-Analysis}: within a HF superconducting sample, the fluctuation-related quasiparticles can be created or annihilated and that their mediation of the e-e interaction, $V^{f}_{ee}$,\cite{Mathur98-HF-SC-Mag-Mediation, Miyake17-HF-QCF-Scattering-Review} is the driving mechanism behind the surge of each of these features.
We illustrate this argument by considering the spin-fluctuation case, as in Ref.~\onlinecite{Mathur98-HF-SC-Mag-Mediation}. There, $I^2\chi(\omega)$ [$I$ is an exchange coupling while $\chi(\omega)$ is a dynamic susceptibility] plays the same role as that of Eliashberg's electron-phonon spectral function $\alpha^{2}\mathcal{F}\left(\omega\right)$. $V^{sf}_{ee}$ can then be obtained from the integral over frequencies of such an Eliashberg's type spectral function.
Below, we give a short summary of the main expressions describing $T_c(\ell)$, $A(\ell)$ and their correlations with each other as well as with $\rho_{\circ}(\ell)$.
\subsubsection{Expression of \texorpdfstring{$T_c(\ell)$}{} \label{SubsubSec.Tc-Calculation}}
Adopting the traditional strategy for calculating $T_c(\ell)$, we solved the set of coupled linearized Eliashberg's equations within the BCS limit. In this limit, $T_c(\ell)$ is determined by the superconducting coupling constant
$\lambda(\ell) =N(\epsilon_F)V_{ee}^{sf}(\ell)=
2\int d\omega\frac{I^2\chi(\omega)}{\omega_{opt}^{2}+\omega^{2}}$, where $\omega_{opt}$ is an optimal frequency at which $\lambda $ is maximal, and can thus be written as: \cite{21-FL-SC-Defectal-Reconciled}
\begin{equation}
T_c(\ell)=\theta\, exp \Bigg ({\frac{-1-\lambda(\ell)}{\lambda(\ell) -\mu^*}}\Bigg ),
\label{Eq-Tc-vs-Lambda}
\end{equation}
where $\mu^*$ is the Coulomb pseudopotential and $\theta$ is an energy scale that depends on the cut-off bandwidth $\omega_c^{sf}$.
\subsubsection{Expression of \texorpdfstring{$A(\ell)$}{} \label{SubsubSec.A-Calculation} }
Applying a variational approach to the linearized version of Boltzmann's transport equations within the relaxation time approximation (wherein the inverse scattering time is calculated by the use of Fermi's golden rule) we obtain\cite{21-FL-SC-Defectal-Reconciled} $\rho (\ell)=A(\ell)T^2$ and
\begin{equation}
A(\ell)=F^2_{\ell}\left|\frac{\lambda(\ell) -\mu^*}{1+\lambda(\ell) }\right|^2.
\label{Eq-A-vs-Lambda}
\end{equation}
Here, $F_{\ell}$ represents the efficiency of momentum relaxation and the availability of phase space for scattering.\cite{21-FL-SC-Defectal-Reconciled}
Within the FL phase, $\ell$ is long, $1\ll k_F\ell_p <\infty$, and $\rho_{\circ}$ is small ($k_F$ = Fermi wave number); as such, Eq.~(\ref{Eq-A-vs-Lambda}) can be expanded around $\lambda_\circ=\lambda (\ell \rightarrow \infty)$ of the normal matrix as
\begin{equation}
A(\ell) \simeq a_\circ + a_1 (\delta\lambda) + a_2 (\delta\lambda)^2 + \mathcal{O}[(\delta\lambda)^3],
\label{Eq-A-vs-deltaLambda}
\end{equation}
wherein $a_\circ =A(\ell \rightarrow \infty) = \frac{|\lambda_\circ-\mu^*|^2}{1+\lambda_\circ}$ refer to the negligibly small kinematically-constrained non-fluctuating contributions; the second and third term \big[containing $\delta\lambda= \lambda(\ell)- \lambda_\circ$ and the coefficients $a_1=2F^2_{\ell}(|\lambda_\circ-\mu^*|(1+\mu^*))/(1+\lambda_\circ)^3$ and
$a_2=F^2_{\ell}(1+\mu^*-2|\lambda_\circ-\mu^*|)(1+\mu^*)/(1+\lambda_\circ)^4$\big] denote contributions from all kinematically unconstrained relaxation processes after incorporating the fluctuation-mediated channel.
The second-order polynomial expression of $A(\ell)$, Eq.~\ref{Eq-A-vs-deltaLambda}, is reminiscent of the empirical \textit{quadratic-in-$\rho_{\circ}$} of Eq.~\ref{Eq-A-propto-rho2}. Below we look for an analytical description of $A(\rho_{\circ})$. Let us start by recalling that, in a typical Fermi liquid, the frequency dependence of the imaginary part of the self energy is given by ${\cal I}m\Sigma(\omega)\sim\omega^2$.
Then, on considering the relevant energy scale to be set by $k_B T$, one obtains the characteristic FL {\it quadratic-in-T} resistivity. Intuitively, this is related to the fact that, for a given temperature $T$, $N(E_F) k_B T$ single particles within the Fermi surface are participating in the fluctuation-mediated two-particle channel (each single particle can scatter into one another via this fluctuation-mediated scattering process): this generates the well-known $AT^2$ resistivity contribution. Specifically, our analysis showed that this channel leads to a FL state with $A$ as in Eq.~\ref{Eq-A-vs-Lambda} and superconductivity with $T_c$ as in Eq.~\ref{Eq-Tc-vs-Lambda}. Moreover, at $T=0$ limit, the relevant energy scale is set by the {\it fuzziness} of the Fermi surface which is determined by $\delta g$. Just as the case for the $\omega^2$-leading-to-$T^2$-contribution, we expect $\rho_{\circ} \propto \delta\lambda \propto (\delta g)^2$.\cite{note-Lambda-propto-rho0} This allows us to establish a correlation among $\rho_{\circ}$ and each of $T_c$ and $A$ (each is a function of $\ell$). Guided by these considerations, as well as the empirical relation of Eq.~\ref{Eq-A-propto-rho2}, we consider
\begin{equation}
A(\ell)\simeq A_{\circ} + A_1 (\rho_{\circ}) + A_2 (\rho_{\circ})^2 ,
\label{Eq-A-vs-rho}
\end{equation}
where $A_i(i=0,1,2$) are functions of $F^2_{\ell}, ~\mu^*, \text{ and } \lambda_\circ$. Eq.~\ref{Eq-A-vs-rho} suggests three limiting contributions to $A(\ell)$: (i) The host contribution (dominant $A_{\circ} \neq 0$ when $A_1 \approx 0$ and $A_2 \approx 0$); (ii) a single-particle-like Koshino-Taylor-type $(A-A_{\circ}) \propto \rho_{\circ}$, contribution (dominant $A_1 \neq0$ contribution); and (iii) an $(A-A_{\circ}) \propto (\rho_{\circ})^2$ contribution (a dominant $A_2 $ and a negligible $A_1$). Results of Figs.~\ref{Fig2-CeCu2Ge2}(e), \ref{Fig3-CeCu2Si2}(e), \ref{Fig4-CeCoIn5}(e) belong to the third case.
\subsubsection{ Correlation between \texorpdfstring{$T_c(\ell)$ and $A(\ell)$}{} \label{SubsubSec.Tc-A-Correlation}}
Combining the expressions for $T_c(\ell)$ in Eq.~\ref{Eq-Tc-vs-Lambda} and $A(\ell)$ in Eq.~\ref{Eq-A-vs-Lambda}, we arrive at:
\begin{equation}
T_c(\ell) = \theta\, e^{-\mathcal{F}/\sqrt{A(\ell)}}.
\label{Eq-Tc-of-A}
\end{equation}
This universal and exact kinematic scaling relation (valid for long $\ell$, low $\rho_{\circ}$: the Fermi-liquid region) is the essence of the relations of Eq.~(\ref{Eq-lnTc-versus-Inv-sqrtA}) and Figs.~\ref{Fig2-CeCu2Ge2}(g),~\ref{Fig3-CeCu2Si2}(g),~and~\ref{Fig4-CeCoIn5}(g). Eq.~(\ref{Eq-Tc-of-A}) is most remarkably manifested in Fig.~\ref{Fig5-Tc-A-AllHFS}: an increase (decrease) of pressure leads to a downwards (upwards) flow of $T_c(\ell)$ towards weaker (stronger) couplings, without ever leaving the curve given by Eq.~ (\ref{Eq-Tc-of-A}).
Finally, a simplified but approximate relation between $T_c$ and $(\rho_{\circ}-\rho_{\circ}^{\circ})$ can be obtained in the specific case wherein the {\it quadratic-in-$\rho_{\circ}$} term in Eq.~\ref{Eq-A-vs-rho} is dominant. Substituting Eqs.~(\ref{Eq-A-propto-rho2}~and~\ref{Eq-A-vs-Lambda}) into Eq.~\ref{Eq-Tc-of-A}, we obtain
\begin{equation}
T_c(\ell) \approx \theta\, exp\big(\frac{-\mathcal{F}}{\sqrt{A_2}\big[\rho_{\circ}(\ell)-\rho_{\circ}^{\circ}\big]}\big.
\label{Eq-Tc-vs-(Ro-Roo)}
\end{equation}
On substituting $\theta$, $\mathcal{F}$, $A_2$, and $\rho_{\circ}^{\circ}$ of Table~\ref{Tab-Fit-Values-A-Ro-Theta-F} into this equation, one obtains the solid curves of Figs.~\ref{Fig2-CeCu2Ge2}(f), \ref{Fig3-CeCu2Si2}(f), \ref{Fig4-CeCoIn5}(f); this excellent description of experimental data with no fitting parameters is no surprise since, as mentioned above, Eq.~(\ref{Eq-Tc-vs-(Ro-Roo)}) is a derived one.
\section{Discussion and Conclusions \label{Sec.Discussion-Conclusion}}
The similarity of the phase diagrams shown in Figs.~\ref{Fig2-CeCu2Ge2},\ref{Fig3-CeCu2Si2},\ref{Fig4-CeCoIn5}, as well as those of other HF superconductors,\cite{Gegenwart08-QC-HF-metals,Coleman07-HF-Review} suggests a generalized \textit{T-p} phase diagram that highlights the similarity in the cascade of distinct electronic states and in the overall evolution of $T_{c}(\ell)$, $A(\ell)$, and $\rho_{\circ}(\ell)$. Of particular interest to this work is the nontrivial manifestation, in all phase diagrams, of a superconductivity, a FL character and their correlations (see Figs.~\ref{Fig2-CeCu2Ge2},~\ref{Fig3-CeCu2Si2},~\ref{Fig4-CeCoIn5},~and~\ref{Fig5-Tc-A-AllHFS}).
We argue above that the surge of all these features and correlations is driven by the spin-fluctuation/valence-fluctuation mediated e-e channel operating within the FL range of the studied HF superconductors.\cite{note-No-Contradicting-Difference-in-Phase-DIagram}
It happened that for these fluctuations to be well defined quasiparticles and for the Migdal-Eliashberg theoretical framework to be applicable, one needs to ensure that $\ell$ is long which for the spin-fluctuation case translates into $\chi(\omega)$ and $V^{sf}_{ee}$ being weak and, as a consequence, reduced $\rho_{\circ}$, $A$, and \textit{T$_c$}. In fact this argument, turned around, can be used to define the FL range which can be reached by an application of higher pressure (see Figs.~\ref{Fig2-CeCu2Ge2}- \ref{Fig4-CeCoIn5}), strong magnetic field,\cite{Paglione03-CeCoIn5-H-QCP, Bianchi03-CeCoIn5-H-QCP, Bauer05-CeCoIn5-xSn-x-HCP} or incorporation of nonmagnetic impurities: \cite{Smith84-UBe13-Impurity-Incorporation,Adrian88-CeCu2Si2-CeCu6-Lattice-Disorder,Adrian88-CeCu2Si2-CeCu6-UPt3-Lattice-Disorder,Adrian89-HF-SUCs-LatticeDisorder} All promote the weakening of the fluctuation-mediated scattering process.
Based on these arguments, the following inferences can be drawn:
First, Figs.~\ref{Fig2-CeCu2Ge2}- \ref{Fig4-CeCoIn5} indicate that on moving out of the FL region but towards the quantum critical points, one observes a continuous enhancement in $\rho_{\circ}$, $A$, and $T_c$. This continuity is suggestive of a similarity in the scattering interaction, the strength of which is enhanced on reducing the pressure.
However, Figs.~\ref{Fig2-CeCu2Ge2}-\ref{Fig4-CeCoIn5} also indicate that below the FL region our analytical expressions do not reproduce the observed baric evolution of $A$ or $T_c$.
This shortcoming is related to the breakdown of the long-$\ell$ condition. Nevertheless, this does not invalidate our analysis within the FL region.
It is worth repeating that the availability of a FL state is not a precondition for the applicability of our approach. Rather, the surge of the e-e scattering channel gives rise to the FL state, the superconductivity and their correlation.
Second, it was reported that an application of a magnetic field ($H>H_{2}$) within the superconducting dome of CeCoIn$_{5}$ leads to a quench of superconductivity in favor of a FL normal-state.\cite{Paglione03-CeCoIn5-H-QCP, Bianchi03-CeCoIn5-H-QCP, Bauer05-CeCoIn5-xSn-x-HCP} We attribute this field-induced FL character within the NFL region to a field-induced weakening of the strength of the spin-fluctuation and as such to an increase in $\ell$, in reminiscence of the aforementioned increase in applied pressure or in the incorporation of nonmagnetic impurities. Further analysis is underway.
Third, we argued above that within the long-$\ell$ region, the spin-fluctuation/valence-fluctuation modes are mediating the e-e interaction and this in turn leads to the scaling between $\frac{T_c}{\theta}$ and $(A_{tot}-A_{\circ})$ (see Fig.~\ref{Fig5-Tc-A-AllHFS}(a)). This is similar to the emergent phonon-mediated scaling reported for other non-HF superconductors: Fig.~\ref{Fig5-Tc-A-AllHFS}(b) demonstrates the scaling in conventional superconductors.\cite{21-FL-SC-Defectal-Reconciled,Nunes12-FermiLiquid-SUC}
Similar scaling was reported for the Fe-based pnictides\cite{18-Castro-Tc-A-Correlation} and chalcogenides:\cite{18-KxFe2-ySe2} Although the superconductivity in these series are considered to be driven by spin-fluctuation mediated pairing\cite{Scalapino12-UnconSCs-CommonThread} however we did not include these Fe-based materials in Fig.\ref{Fig5-Tc-A-AllHFS} since these reports did not include a correction for the non-fluctuation-related contribution: $A_{tot}$ was used instead of $(A_{tot}-A_{\circ})$.
We extend our present analysis by including a discussion\cite{21-FL-SC-Defectal-Reconciled} of two specific parameters of the HF superconductors: (i) the Kadowaki-Woods ratio and (ii) the gap-to-$T_c$ ratio. We show that the influence of the fluctuation processes on both ratios can be accounted for by including an additional factor which embodies the material properties of these HF superconductors.
(i) Jacko \textit{et al.}\cite{Jacko09-Kadowaki-Woods-Strongly-correlated-Metals} accounted for the wide difference among the Kadowaki-Woods ratio of a variety of strongly coupled systems by demonstrating that
\begin{equation}
\frac{A}{\gamma^2}= \Big( \frac{81}{4\pi\hbar k_B^2 e^2}\Big) \Big(\frac{1}{d^2 n N^2(\epsilon_F)\langle v_{0x}^2\rangle}\Big)= \Big(f_{con}\Big)\Big( \frac{1}{f_{mat}}\Big),
\label{Eq-KDW-ratio}
\end{equation}
wherein $\langle v_{0x}^2\rangle$ is an average of the carrier velocity squared
(a measure of the anisotropies), $n$ is the carrier density, $N(\epsilon_F)$ is the density of states at the Fermi level,
and $d \sim 1$ is a dimensionless number.
Apparently, due to the material-dependent second factor, $\frac{A}{\gamma^2}$ of Eq.~\ref{Eq-KDW-ratio} is nonuniversal.
This shortcoming would be most evident for the ratio of the fluctuation-bearing HF superconductors within their Fermi-liquid region (see Figs.~\ref{Fig2-CeCu2Ge2}-\ref{Fig4-CeCoIn5}).
Our evaluation of the Kadowaki-Woods ratio (Eq.~\ref{Eq-KDW-ratio}) of these HF superconductors gives
\begin{equation}
\frac{A}{\gamma^2}=\Big(\frac{81}{4\pi\hbar k_B^2 e^2}\Big)\Big(
\frac{F_{\ell}}{d^2 n N^2(\epsilon_F)\langle v_{0x}^2\rangle}\Big)=\Big(f_{con}\Big)\Big( \frac{F_{\ell}}{f_{mat}}\Big) ,
\label{Eq-KDW-ratio-Defectal}
\end{equation}
wherein the material-dependent $F_{\ell}$ factor is as described in Eq.~(\ref{Eq-A-vs-Lambda}). It is not close to 1 as assumed during the derivation of Eq.~(\ref{Eq-KDW-ratio}).\cite{Jacko09-Kadowaki-Woods-Strongly-correlated-Metals} Rather $F_{\ell} > 1$: a larger \textit{apparent} ratio because of the easing of the kinematic constraints in these fluctuation-bearing systems. As in Ref.~\onlinecite{Jacko09-Kadowaki-Woods-Strongly-correlated-Metals}, the universal character is restored only when expressed as
$\frac{A}{\gamma^2}\Big(\frac{f_{mat}}{F_{\ell}}\Big)=\frac{A.f_{mat}^{'}}{\gamma^2}=f_{con}$.
(ii) Carbotte\cite{Carbotte90-SCs-Boson-Exchanged-Reivew} derived an approximate expression for the gap-to-$T_c$ ratio of strong-coupled superconductor. Modified to the case of a fluctuation-bearing superconductor within $\frac{T_c}{\theta} \ll 1$ region, this expression reads as
\begin{subequations}
\begin{align}
\frac{2\Delta(\ell)}{k_B T_c(\ell)} \approx 3.53\left\{1+12.5 \left[\frac{T_c(\ell)}{\theta}\right]^2 \ln{\left[\frac{\theta}{2T_c(\ell)}\right]}\right\},\\
\approx 3.53\left\{1+12.5 \left[e^{-\mathcal{F}/\sqrt{A(\ell)}}\right]^2 \ln{\left[\frac{1}{2e^{-\mathcal{F}/\sqrt{A(\ell)}}}\right]}\right\}
\end{align}
\label{Eq-gap-to-Tc-ratio}
\end{subequations}
Thus, within $\frac{T_c}{\theta} \ll 1$ region, the gap-to-$T_c$ ratio can be fine-tuned: From $2\Delta(\ell \rightarrow \infty)/k_B T_c(\ell \rightarrow \infty)=3.53$, the universal BCS value, up to a higher nonuniversal value by varying the control-parameter that modifies $\ell$.
Finally, it is worth adding that our internally-consistent empirical and theoretical analyses are based on a clear identification of the difference between the fluctuation-related and normal (non-fluctuation-related) contributions within the fluctuation-related FL region (which starts when the resistivity manifests the $T^2$ contribution and ends when $T_c \rightarrow 0$).
In summary, we investigated the superconductivity, the FL transport, and their correlations within the FL region of the \textit{T-p} phase diagram of representative quantum-critical HF superconductors.
Empirically, on varying the control parameters, (i) the normal-state resistivities manifest the characteristic FL, $\rho_o + \textit{AT}^2$, character with $A\propto \rho_{\circ}^2$ and
(ii) the superconducting state manifests a correlation of $T_c$ with $A$ ($ln\frac{T_{c}}{\theta}\propto A^{-\frac{1}{2}}$).
We attribute the surge of these superconducting and FL transport features and their correlations to a fluctuation-mediated e-e scattering channel.
Theoretically, on adopting many-body techniques, we derive analytic expressions for $T_c( \rho_{\circ})$ and $A(\rho_{\circ})$ and their correlations that reproduce satisfactorily the aforementioned empirical correlations.
\section*{Methods}
We analyzed the extensively reported pressure-dependent resistivities of various HF superconductors.
Within the FL region, we looked for any correlation among the superconductivity (as measured by $T_c$), the Fermi-liquid character (as measured by $A$), and the excess in the residual resistivity ($\rho_{\circ}$ as a measure of the strength of the scattering channel) when the control parameter is varied.
Using graphical and analytical procedures, we managed to identify the empirical expression discussed in Subsec.~\ref{SubSec.Emprical-Derivation}.
More importantly, we managed to identify the spin-fluctuation/valence-fluctuation mediated e-e scattering channel as being the driving mechanism behind these correlations.
Guided by the inferences drawn from the empirical expressions and the identified channel, we formulated the theoretical framework outlined in Subsec.~\ref{SubSec.Emprical-Derivation} and Sec.~\ref{subSec.Theoretical-Derivation}: Basically, we started with the spin-fluctuation/valence-fluctuation exchange mechanism which is different from the traditional mechanisms in that it takes into consideration the modification in the kinematic constraints which, in turn, lead to a significant enlargement in the phase space available for scattering. Then, after applying the standard theories of Migdal-Eliashberg (superconductivity) and Boltzmann (transport), we managed to derive the analytic expressions of Subsec.~\ref{SubSec.Emprical-Derivation} which satisfactorily explain the empirical observations.
\section*{Acknowledgments}
We acknowledge partial financial support from Brazilian agency CNPq.
|
\section[#1:\ \emph{#2}]{#1:{\ \emph{#2}}}
}
\renewcommand{\AA}[0]{\mathbb A}\newcommand{\NN}[0]{\mathbb N}
\newcommand{\BB}[0]{\mathbb B} \newcommand{\OO}[0]{\mathbb O}
\newcommand{\CC}[0]{\mathbb C} \newcommand{\PP}[0]{\mathbb P}
\newcommand{\DD}[0]{\mathbb D} \newcommand{\QQ}[0]{\mathbb Q}
\newcommand{\EE}[0]{\mathbb E} \newcommand{\RR}[0]{\mathbb R}
\newcommand{\FF}[0]{\mathbb F} \renewcommand{\SS}[0]{\mathbb S}
\newcommand{\GG}[0]{\mathbb G} \newcommand{\TT}[0]{\mathbb T}
\newcommand{\HH}[0]{\mathbb H} \newcommand{\UU}[0]{\mathbb U}
\newcommand{\II}[0]{\mathbb I} \newcommand{\VV}[0]{\mathbb V}
\newcommand{\JJ}[0]{\mathbb J} \newcommand{\WW}[0]{\mathbb W}
\newcommand{\KK}[0]{\mathbb K} \newcommand{\XX}[0]{\mathbb X}
\newcommand{\LL}[0]{\mathbb L} \newcommand{\YY}[0]{\mathbb Y}
\newcommand{\MM}[0]{\mathbb M} \newcommand{\ZZ}[0]{\mathbb Z}
\newcommand{\cA}[0]{\mathcal A} \newcommand{\cN}[0]{\mathcal N}
\newcommand{\cB}[0]{\mathcal B} \newcommand{\cO}[0]{\mathcal O}
\newcommand{\cC}[0]{\mathcal C} \newcommand{\cP}[0]{\mathcal P}
\newcommand{\cD}[0]{\mathcal D} \newcommand{\cQ}[0]{\mathcal Q}
\newcommand{\cE}[0]{\mathcal E} \newcommand{\cR}[0]{\mathcal R}
\newcommand{\cF}[0]{\mathcal F} \newcommand{\cS}[0]{\mathcal S}
\newcommand{\cG}[0]{\mathcal G} \newcommand{\cT}[0]{\mathcal T}
\newcommand{\cH}[0]{\mathcal H} \newcommand{\cU}[0]{\mathcal U}
\newcommand{\cI}[0]{\mathcal I} \newcommand{\cV}[0]{\mathcal V}
\newcommand{\cJ}[0]{\mathcal J} \newcommand{\cW}[0]{\mathcal W}
\newcommand{\cK}[0]{\mathcal K} \newcommand{\cX}[0]{\mathcal X}
\newcommand{\cL}[0]{\mathcal L} \newcommand{\cY}[0]{\mathcal Y}
\newcommand{\cM}[0]{\mathcal M} \newcommand{\cZ}[0]{\mathcal Z}
\newcommand{\mathbb{R}}{\mathbb{R}}
\newcommand{\fA}[0]{\mathfrak A} \newcommand{\fN}[0]{\mathfrak N}
\newcommand{\fB}[0]{\mathfrak B} \newcommand{\fO}[0]{\mathfrak O}
\newcommand{\fC}[0]{\mathfrak C} \newcommand{\fP}[0]{\mathfrak P}
\newcommand{\fD}[0]{\mathfrak D} \newcommand{\fQ}[0]{\mathfrak Q}
\newcommand{\fE}[0]{\mathfrak E} \newcommand{\fR}[0]{\mathfrak R}
\newcommand{\fF}[0]{\mathfrak F} \newcommand{\fS}[0]{\mathfrak S}
\newcommand{\fG}[0]{\mathfrak G} \newcommand{\fT}[0]{\mathfrak T}
\newcommand{\fH}[0]{\mathfrak H} \newcommand{\fU}[0]{\mathfrak U}
\newcommand{\fI}[0]{\mathfrak I} \newcommand{\fV}[0]{\mathfrak V}
\newcommand{\fJ}[0]{\mathfrak J} \newcommand{\fW}[0]{\mathfrak W}
\newcommand{\fK}[0]{\mathfrak K} \newcommand{\fX}[0]{\mathfrak X}
\newcommand{\fL}[0]{\mathfrak L} \newcommand{\fY}[0]{\mathfrak Y}
\newcommand{\fM}[0]{\mathfrak M} \newcommand{\fZ}[0]{\mathfrak Z}
\renewcommand{\aa}[0]{\textbf{\textit{a}}}\newcommand{\nn}[0]{\textbf{\textit{n}}}
\newcommand{\bb}[0]{\textbf{\textit{b}}}\newcommand{\oo}[0]{\textbf{\textit{o}}}
\newcommand{\cc}[0]{\textbf{\textit{c}}}\newcommand{\pp}[0]{\textbf{\textit{p}}}
\newcommand{\dd}[0]{\textbf{\textit{d}}}\newcommand{\qq}[0]{\textbf{\textit{q}}}
\newcommand{\ee}[0]{\textbf{\textit{e}}}\newcommand{\rr}[0]{\textbf{\textit{r}}}
\newcommand{\ff}[0]{\textbf{\textit{f}}}\renewcommand{\ss}[0]{\textbf{\textit{s}}}
\renewcommand{\tt}[0]{\textbf{\textit{t}}}
\newcommand{\hh}[0]{\textbf{\textit{h}}}\newcommand{\uu}[0]{\textbf{\textit{u}}}
\newcommand{\ii}[0]{\textbf{\textit{i}}}\newcommand{\vv}[0]{\textbf{\textit{v}}}
\newcommand{\jj}[0]{\textbf{\textit{j}}}\newcommand{\ww}[0]{\textbf{\textit{w}}}
\newcommand{\kk}[0]{\textbf{\textit{k}}}\newcommand{\xx}[0]{\textbf{\textit{x}}}
\newcommand{\yy}[0]{\textbf{\textit{y}}}
\newcommand{\mm}[0]{\textbf{\textit{m}}}\newcommand{\zz}[0]{\textbf{\textit{z}}}
\renewcommand{\bar}[1]{\overline{#1}}
\newcommand{\one}[0]{\mathbf{1}}
\newcommand{\eps}[0]{\varepsilon}
\newcommand{\Hom}[0]{\text{Hom}}
\newcommand{\leg}[2]{\left(\frac{#1}{#2}\right)}
\newcommand{\tr}[0]{\text{Tr}}
\newcommand{\n}[0]{\text{N}}
\newcommand{\disc}[0]{\text{disc}}
\newcommand{\ev}[0]{\text{eval}}
\renewcommand{\mod}[1]{\ (\text{mod }#1)}
\newcommand{\fpx}[0]{\mathbb F_p^\times}
\newcommand{\lr}[1]{\left(#1\right)}
\newcommand{\flr}[1]{\left\lfloor #1\right\rfloor}
\newcommand{\ceil}[1]{\left\lceil#1\right\rceil}
\newcommand{\supp}[0]{\text{supp}}
\newcommand{\spec}[0]{\mathrm{Spec}}
\newcommand{\diam}[0]{\mathrm{diam}}
\newcommand{\Ker}[0]{\text{Ker}}
\newcommand{\qu}[1]{``#1''}
\renewcommand{\emptyset}{\varnothing}
\newcommand{\lL}[0]{\textbf{\textit{L}}}
\renewcommand{\thefootnote}{\fnsymbol{footnote}}
\author[B. Hanson]{Brandon Hanson} \address{Brandon Hanson \\University of Maine\\
Orono, ME}
\email{<EMAIL>}
\author[O. Roche-Newton]{Oliver Roche-Newton} \address{Oliver Roche-Newton, Johann Radon Institute for Computational and Applied Mathematics\\
Linz, Austria}
\email{<EMAIL>}
\author[S. Senger]{Steven Senger} \address{Steven Senger \\ Missouri State University\\
Springfield, MO}
\email{<EMAIL>}
\date{}
\begin{document}
\baselineskip=17pt
\title{Convexity, Superquadratic Growth, and Dot Products}
\date{}
\begin{abstract}
Let $P \subset \mathbb R^2$ be a point set with cardinality $N$. We give an improved bound for the number of dot products determined by $P$, proving that,
\[
|\{ p \cdot q :p,q \in P \}| \gg N^{2/3+c}.
\]
A crucial ingredient in the proof of this bound is a new superquadratic expander involving products and shifts. We prove that, for any finite set $X \subset \mathbb R$, there exist $z,z' \in X$ such that
\[
\left|\frac{(zX+1)^{(2)}(z'X+1)^{(2)}}{(zX+1)^{(2)}(z'X+1)}\right| \gtrsim |X|^{5/2}.
\]
This is derived from a more general result concerning growth of sets defined via convexity and sum sets, and which can be used to prove several other expanders with better than quadratic growth. The proof develops arguments from \cite{HRNR}, and uses predominantly elementary methods.
\end{abstract}
\maketitle
\section{Introduction}
\subsection{Dot products}
For a finite set $P$ of points in $\mathbb R^2$, let $\Lambda(P)$ denote the set of all dot products determined by $P$, that is
\[
\Lambda(P):= \{ p \cdot q : p,q \in P\},
\]
where the dot product of $p=(p_1,p_2)$ and $q=(q_1,q_2)$ is $p \cdot q= p_1q_1+p_2q_2$. In the spirit of the Erd\H{o}s distinct distance problem, one expects that the set $\Lambda(P)$ should be large for any $P$. The bound
\begin{equation} \label{STbasic}
|\Lambda(P)| \gg |P|^{2/3}
\end{equation}
follows from an application of the Szemer\'{e}di-Trotter Theorem.\footnote{Here and throughout this paper, the notation
$X\gg Y$, $Y \ll X,$ $X=\Omega(Y)$, and $Y=O(X)$ are all equivalent and mean that $X\geq cY$ for some absolute constant $c>0$. $X \approx Y$ and $X=\Theta (Y)$ denote that both $X \gg Y$ and $X \ll Y$ hold. $X \gg_a Y$ means that the implied constant is no longer absolute, but depends on $a$. We also use the notation $X \gtrsim Y $ and $Y \lesssim X$ to denote that $X \gg Y/(\log_2 Y)^c$ for some absolute constant $c>0$.} The Erd\H{o}s distance problem was resolved up to logarithmic factors in a remarkable paper of Guth and Katz \cite{GK}, which proved that the bound
\begin{equation} \label{GKbound}
|D(P)| \gg \frac{ |P|}{\log |P|}
\end{equation}
holds for any finite point set $P \subset \mathbb R^2$, where $D(P)$ is the set of Euclidean distances determined by pairs of points in $P$. It is widely believed that a bound similar to \eqref{GKbound} also holds for $\Lambda(P)$. However, in stark contrast to the distance problem, this question remains wide open, and the apparent connection between these two problems may be somewhat misleading.
For instance, the closely related unit distance conjecture claims that any fixed distance can occur at most $N^{1+\epsilon}$ times among a point set of size $N$. The bound
\begin{equation} \label{4/3}
|\{ (p,q) \in P \times P : \|p-q\|=1 \}| \ll N^{4/3}
\end{equation}
was established by Spencer, Szemer\'{e}di and Trotter \cite{SST}. This bound is widely believed to be sub-optimal, but there has been no significant improvement to \eqref{4/3} in 37 years. On the other hand, a construction in \cite{IRNR} shows that there exists a point set $P \subset \mathbb R^2$ such that
\[
|\{ (p,q) \in P \times P : p \cdot q=1 \}| \gg N^{4/3}.
\]
In this paper, we give the following improvement to \eqref{STbasic}.
\begin{Theorem} \label{thm:main}
For any finite set $P \subset \mathbb R^2$
\begin{equation} \label{dpgoal}
|\Lambda(P)| \gtrsim |P|^{\frac{2}{3}+\frac{1}{3057}}.
\end{equation}
\end{Theorem}
We remark that a better bound, with exponent $\frac{2}{3}+\frac{1}{39}$, for the case of skew symmetric bilinear forms was obtained in \cite{IRNR}. This covers, for instance, the set of wedge/cross products determined by a point set, i.e. $p \wedge q = p_1q_2-p_2q_1$.
\subsection{Superquadratic growth}
An essential ingredient in the proof of Theorem \ref{thm:main} is a new sum-product type result which shows that a certain expander involving products and shifts gives superquadratic growth. We will prove the following theorem.
\begin{Theorem} \label{cor:prodshift2}
Let $X$ be a finite set of positive reals. Then there exist
$z,z' \in X$ such that
\begin{equation} \label{5/2}
\left|\frac{(zX+1)^{(2)}(z'X+1)^{(2)}}{(zX+1)^{(2)}(z'X+1)}\right| \gtrsim |X|^{5/2}.
\end{equation}
\end{Theorem}
Here, $A^{(k)}$ denotes the $k$-fold product set $\{a_1 \cdots a_k : a_1,\dots, a_k \in A \}$. A key quantitative feature of Theorem \ref{cor:prodshift2} is that the exponent in the right hand side of \eqref{5/2} is strictly larger than $2$. We refer to the set in Theorem \ref{cor:prodshift2} as a \textit{superquadratic expander}. There are many sets defined by a combination of arithmetic operations that, one may suspect, grow in such a way. However, given such a set, it is often difficult to prove that it really is a superquadratic expander, particularly as methods from incidence geometry typically lose their potency in this range. The existence of other superquadratic expanders has been established in \cite{BRNZ}, \cite{R}, \cite{S}, \cite{RNW}.
Theorem \ref{cor:prodshift2} is derived from a more general result, the forthcoming Theorem \ref{MainExpander}. The statement of Theorem \ref{MainExpander} is a little technical, depending on some further definitions, and we do not state it in full yet. However, Theorem \ref{MainExpander} can also be used to derive several other new sum-product results. For instance, we will prove the following result about superquadratic growth under addition for cubes of shifts.
\begin{Corollary} \label{cor:cubes}
Let $A \subset \mathbb R$ be a finite set. Then there exists $a,a' \in A$ such that
\[
|2(a+A)^3 + 2(a'+A)^3 - 2 (a+A)^3 - (a'+A)^3| \gtrsim |A|^{5/2}.
\]
\end{Corollary}
Further applications of Theorem \ref{MainExpander} will be given in the forthcoming Section \ref{sec:other}.
The proof of Theorem \ref{MainExpander} uses similar elementary methods to an earlier paper of Rudnev and the first two authors \cite{HRNR}, which was itself built up from a simple and beautiful observation concerning sumsets of convex sets in a paper of Ruzsa, Shakan, Solymosi and Szemer\'{e}di \cite{RSSS}. Our first step towards proving Theorem \ref{MainExpander} and its consequences is to give a simple proof of the following quadratic expander.
\begin{Theorem} \label{thm:quad}
Let $A \subset \mathbb R$ be a finite set and let $f$ be a strictly convex or concave function. Then there exists $a,a' \in A$ such that
\[
|f(a+A) + f(a'+A) - f(a+A) | \gg |A|^{2}.
\]
\end{Theorem}
Theorem \ref{thm:quad} gives the optimal order of growth. This can be seen by considering the case
\[
A=\{1,2,\dots,N\} ,\,\,\,\,\,\,\, f(x)=x^2.
\]
A further example illustrating the optimality of Theorem \ref{thm:quad} is given in Section \ref{sec:expanders}.
\subsection{Sketch of the proof of Theorem \ref{thm:main}}
The proof of Theorem \ref{thm:main} is rather long, and so we sketch it here in an attempt to convey some of the main ideas behind the proof to the reader. In this sketch we are using Theorem \ref{cor:prodshift2} as a black box.
In fact, let us begin the sketch by assuming something a little stronger than Theorem \ref{cor:prodshift2}: suppose that there exists $z \in X$ such that
\[
\left|\frac{(zX+1)^{(4)}}{(zX+1)^{(3)}}\right| \gtrsim |X|^{5/2}.
\]
That is, we make the simplifying assumption that $z=z'$ in Theorem \ref{cor:prodshift2}. After several applications of Pl\"{u}nnecke-Ruzsa type inequalities, it would then follow that, for any $Y \subset \mathbb R$,
\begin{equation} \label{2/3+}
|Y(zX+1)| \gtrsim |Y|^{\frac{7}{8}}|X|^{\frac{1}{4}+\frac{1}{128}}.
\end{equation}
Quantitatively, the additional $1/128$ in the exponent is important here. Later in the sketch, we will apply this bound in a case whereby $|Y|=N^{2/3}$ and $|X|=N^{1/3}$, in which case \eqref{2/3+} yields the non-trivial bound
\begin{equation} \label{calc}
|Y(zX+1)| \gtrsim N^{\frac{2}{3} + \frac{1}{384}}.
\end{equation}
To prove Theorem \ref{thm:main}, we begin with an application of the Szemer\'{e}di-Trotter Theorem which tells us that we obtain an exponent better than $2/3$ unless we are in a very specific situation. Roughly speaking, this allows us to assume that $P$, our given set of $N$ points, is supported on exactly $N^{1/3}$ lines through the origin, each containing $N^{2/3}$ points from $P$. Furthermore, since the problem is rotation invariant, we may assume that one of these lines is the $x$-axis. This in fact implies that the set of $x$-coordinates involved in $P$ also has size around $N^{2/3}$, otherwise we have too many dot products between a fixed point from $P$ on the $x$-axis and the remaining elements of $P$. Pigeonholing, we deduce that there is some vertical line $\ell_0$, with equation $x=x_0$, which supports around $N^{1/3}$ points of $P$. Simplifying slightly, let's assume that all of the $N^{1/3}$ lines through the origin contain a point on $P \cap \ell_0$.
Let $S$ denote the set of slopes of these $N^{1/3}$ lines, each containing a point from $P \cap \ell_0$ and $N^{2/3}$ points from $P$ in total. Now apply \eqref{2/3+} for the set $S$. It follows that there is some slope $s \in S$ such that the bound
\[
|Y(sS+1)| \gtrsim |Y|^{\frac{7}{8}}|S|^{\frac{1}{4}+\frac{1}{128}}
\]
holds for any $Y \subset \mathbb R$. However, if we consider the set of dot products between points of $P \cap \{y=sx\}$ and $P \cap \ell_0$, this is equal to the set $x_0Y(sS+1)$, where $Y$ is the set of the $N^{2/3}$ $x$-coordinates of $P \cap \{y=sx\}$. Recalling the calculation in \eqref{calc}, we obtain an exponent better than $2/3$, as required.
If we remove the simplifying assumption that $z=z'$, things become a bit trickier when it comes to constructing the set that plays the role of $Y$. We need to choose a set of $x$-coordinates that works with respect to two lines $y=sx$ and $y=s'x$ simultaneously, and it is not immediately obvious that these lines share many $x$-coordinates.
However, an extra piece of information comes to our aid here. We may consider the dot products between $P \cap \{y=sx\} $ and $P \cap \{y=s'x\}$. This turns out to be essentially a product set of the two sets of $x$-coordinates. Each set has size $N^{2/3}$, and their product set has approximately the same size, otherwise we are done. We can use this information, and some fairly basic additive combinatorial trickery, to deduce the existence of a good set $Y$ which works simultaneously for both of the fixed lines.
\subsection{Structure of this paper}
We begin Section \ref{sec:expanders} by proving Theorem \ref{thm:quad}. The main goal of the section is to prove our general result on superquadratic growth, which is Theorem \ref{MainExpander}. We then apply this bound to prove our main result on growth of products with shifts, Theorem \ref{cor:prodshift2}. In Section \ref{sec:dp} we use Theorem \ref{cor:prodshift2}, or rather a slightly refined statement containing some extra information, to prove our main result, Theorem \ref{thm:main}. In Section \ref{sec:other}, we give several other applications of Theorem \ref{MainExpander}. This includes the proof of Corollary \ref{cor:cubes}.
\section{Superquadratic expansion} \label{sec:expanders}
\subsection{An elementary proof of a quadratic expander}
The basic argument underlying this section is contained in the forthcoming proof of Theorem \ref{thm:quad}. In fact we record the following more general result, which will prove useful.
\begin{Theorem} \label{thm:quadagain}
Let \[A=\{b_1<\ldots<b_N\}\] be a finite set of real numbers, and let $f$ be a strictly convex or strictly concave function. Suppose that $ h<h'$ are parameters and
\begin{equation} \label{spacing}
\min\{b_{i+1}-b_i:1\leq i\leq N-1\}\geq h'-h.
\end{equation}
Then
\[
|f(h+A)+f(h'+A)-f(h+A)| \gg |A|^2.
\]
In particular, there exists $a,a' \in A$ such that
\[
|f(a+A)+f(a'+A)-f(a+A)| \gg |A|^2.
\]
\end{Theorem}
\begin{proof}
Observe that the intervals
\[
(b_i+h, b_i+h') : 1 \leq i \leq N
\]
all have length $h'-h$ and do not overlap because of the spacing condition. Therefore, by convexity of $f$, the intervals
\[
(f(b_i+h), f(b_i+h')) : 1 \leq i \leq N
\]
are increasing (or decreasing, if $f$ is concave) in length as $i$ increases, and they also do not overlap. This enables us to squeeze the smaller intervals into the bigger ones. In particular, for all $1\leq i<j \leq N$ we have
\[
f(b_j+h)<f(b_j+h)+(f(b_i+h') - f(b_i+h)) \leq f(b_j+h').
\]
This shows the existence of at least $j-1$ elements of $f(h+A)+f(h'+A)-f(h+A)$ in the interval $(f(b_j+h), f(b_j+h'))$. Summing over all intervals, it therefore follows that
\[
|f(h+A)+f(h'+A)-f(h+A)| \geq \sum_{j=1}^N (j-1) \gg N^2.
\]
For the final statement of the theorem, let $a$ and $a'$ be two elements of $A$ such that $a'-a$ is the minimum positive element of $A-A$.
\end{proof}
As we mentioned in the introduction, Theorem \ref{thm:quad} is optimal in the case when $A=[N]$\footnote{Here we are adopting the standard notation $[N]=\{1,\ldots,N\}$.} and $f(x)=x^2$. Furthermore, if we take $A=\{\log b : b \in B \}$ and $f(x)=e^x$, Theorem \ref{thm:quad} gives the bound
\[
|bB+b'B-bB| \gg |B|^2.
\]
This bound is also optimal, again in the case $B=[N]$. In fact, a closer look at the proof shows that we obtain the bound
\begin{equation} \label{2var}
|bB + (b'-b)B| \gg |B|^2.
\end{equation}
A very similar bound to \eqref{2var} was obtained by Garaev using a different elementary argument. He made the observation that, if $B$ is a set of positive integers with maximal element $b_{max}$ and with $d_{min}=b'-b$ being the smallest positive element in $B-B$, then
\[
|b_{max}B+(b'-b)B| = |B|^2,
\]
as there are no repetitions in this sum set.
There is some flexibility in Theorem \ref{thm:quad}, hinted at in Theorem \ref{thm:quadagain}, as to the choice of the elements $a$ and $a'$. Indeed, we chose them so that $a'-a$ is a minimal positive element of $A-A$, but we could also have taken another close pair of elements, with minimal modification of the proof. In particular, the same proof works if we replace $a,a'$ with a \textit{second nearest consecutive pair} $b,b' \in A$. That is, we choose $b,b'$ to be consecutive elements of $A$ such that
\[
[c<c' \text{ are consecutive and } c'-c < b' -b ] \Rightarrow (c,c')=(a,a').
\]
We then have the following refined version of the statement of Theorem \ref{thm:quadagain}, which will be useful later.
\begin{Theorem} \label{thm:quadagainagain}
Let $A \subset \mathbb R$ be a finite set and let $f$ be a strictly convex or concave function. Let $a<a'$ be a nearest consecutive pair of elements in $A$ and let $b<b'$ be a second nearest consecutive pair. Then
\begin{equation} \label{nearest}
|f(a+A)+f(a'+A)-f(a+A)| \gg |A|^2.
\end{equation}
and \begin{equation} \label{2ndnearest}
|f(b+A)+f(b'+A)-f(b+A)| \gg |A|^2.
\end{equation}
\end{Theorem}
\begin{proof}
The proof of Theorem \ref{thm:quadagain} immediately implies \eqref{nearest}. To prove \eqref{2ndnearest}, define $A' \subset A$ to be the set consisting of every second element of $A$. That is, if $A=\{b_1<\dots<b_N \}$, $A'$ is the subset
\[
A'=\{b_2<a_4<b_{2M} \},
\]
where $M= \lfloor N/2 \rfloor$. Note that,
\[
\min \{ b_{2(i+1)} - b_{2i} : 1 \leq i \leq \lfloor N/2 \rfloor \} \geq b'-b
\]
and so the spacing condition \eqref{spacing} is satisfied. Therefore
\begin{align*}
|f(b+A)+f(b'+A)-f(b+A)|\geq |f(b+A')+f(b'+A')-f(b+A')| \gg |A|^2,
\end{align*}
as required.
\end{proof}
In this paper, we are particularly interested in the expander consisting of products of shifts of products, and so we record the following corollary.
\begin{Corollary} \label{cor:prodshiftquad}
Let $X$ be a set of positive reals. Suppose that $z_1,z_1', z_2,z_2' \in X$ with $z_1 < z_1'$ and $z_2<z_2'$, and with the additional properties
\begin{itemize}
\item $\log z_1 < \log z_1'$ is a nearest consecutive pair of elements in $\log X$, and
\item $\log z_2 < \log z_2'$ is a second nearest consecutive pair of elements in $\log X$.
\end{itemize}
Then
\[
\left | \frac{(z_1X+1)(z_1'X+1)}{(z_1X+1)} \right | \gg |X|^2
\]
and
\[
\left | \frac{(z_2X+1)(z_2'X+1)}{(z_2X+1)} \right | \gg |X|^2.
\]
\end{Corollary}
\begin{proof}
Apply Theorem \ref{thm:quadagainagain} with $A= \log X$ and $f(t)=\log(e^t+1)$.
\end{proof}
\subsection{More variables, more growth}
We would like to do better than Theorem \ref{thm:quadagain} using differences of differences, in the spirit of \cite{HRNR}. We begin to work towards this goal by proving the following lemma.
\begin{Lemma}\label{Expansion2}
Let $X=\{x_1<\ldots<x_N\}$ be a finite set of real numbers and suppose that $F$ is a function which is strictly increasing and strictly convex (or concave) on the interval $[x_1,x_N]$. Then
\[|\lr{(X\times F(X))+(X\times F(X))-(X\times F(X))}\cap[x_1,x_N]\times(F(x_1),F(x_N)]|\gg \lr{\frac{N}{\log N}}^3.\]
\end{Lemma}
\begin{proof}
This argument is taken from \cite{HRNR}. We have
\[(X\times F(X))+(X\times F(X))-(X\times F(X))=(X+X-X)\times(F(X)+F(X)-F(X)).\]
We will estimate the number of elements in the appropriate interval coming from of each factor on the right separately.
Let \[D=\{x_{i+1}-x_i:i\in[N-1]\},\] let \[I_d=\{i\in [N-1]:x_{i+1}-x_i=d\},\]
and let $r(d)=|I_d|$. Since
\[N-1=\sum_{0\leq k\leq \log N}\sum_{\substack{d\in D\\ 2^{k-1}\leq r(d)\leq 2^k}}r(d)\]
there is some integer $L$ and set $D'\subseteq D$, say of size $m$, such that
\begin{equation}\label{mLBound}
mL\gg \frac{N}{\log N},
\end{equation} and $L/2\leq r(d)\leq L$ for each $d\in D'$. If $d,d'\in D'$ are such that $d\leq d'$ then for each $i\in I_{d'}$ we have
\[x_i+d\in (x_i,x_{i+1}]\cap (X+X-X).\]
It follows that, if $D'=\{d_1<\ldots<d_m\}$, we have \begin{equation}\label{KBound}
|(X+X-X)\cap (x_1,x_N]|\geq \sum_{j=1}^m j|I_{d_j}|\gg \frac{mN}{\log N}.
\end{equation}
Now fix $d\in D'$ and consider all the intervals \[(F(x_i),F(x_{i+1})]=(F(x_i),F(x_{i}+d)],\ i\in I_d. \] There are at least $L/2$ such intervals, they are visibly disjoint, and each has a distinct length in view of the convexity of $F$. Thus if $F(x_j+d)-F(x_j)<F(x_i+d)-F(x_i)$ for some $i,j\in I_d$, we have
\[F(x_i)+F(x_j+d)-F(x_j)\in (F(x_i,F(x_{i+1})]\cap (F(X)+F(X)-F(X)).\]
and by considering such contributions across all $d\in D'$, we deduce
\[|(F(X)+F(X)-F(X))\cap (F(x_1),F(x_N)]|\gg \frac{(Lm)^2}{m}.\] Finally, combining the estimates \eqref{mLBound} and \eqref{KBound} concludes the proof.\end{proof}
The following is merely a two-dimensional extension of Theorem \ref{thm:quadagain}, obtained by applying the one-dimensional result in each coordinate independently, and stated in a way that is convenient to our remaining arguments. The lemma extends with no difficulty to arbitrary cartesian products in any dimension.
\begin{Lemma}\label{Expansion1}
Suppose that we have a real numbers $h_1,h_1',h_2,h_2'$ with $h_1<h_1'$ and $h_2 < h_2'$. Suppose that
\[
A=\{a_1<\ldots<a_N\}
\]
is a finite set of reals
which satisfies the spacing condition \[\min\{a_{i+1}-a_i:i\in[N-1]\}\geq \max \{ h_1'-h_1, h_2'-h_2 \}.
\]
Suppose $f_1$ and $f_2$ are strictly increasing and strictly convex functions. Then, for any $k_1,k_2$ with $N/2\leq k_1,k_2\leq N$, the box
\[
\cB_{k_1,k_2}=(f_1(a_{k_1}+h_1),f_1(a_{k_1}+h_1')] \times (f_2(a_{k_2}+h_2),f_2(a_{k_2}+h_2')]
\]
contains all points of the form
\[
(f_1(a_{k_1}+h_1)+f_1(a_{j_1}+h_{1}')-f_1(a_{j_1}+h_1), f_2(a_{k_2}+h_2)+f_2(a_{j_2}+h_{2}')-f_2(a_{j_2}+h_2)),
\] with $1\leq j_1,j_2\leq N/2$. Furthermore, the boxes $\cB_{k_1,k_2}$ are disjoint as $k_1$ and $k_2$ vary.
\end{Lemma}
\begin{proof}
From the spacing condition, we deduce that the boxes
\[
(a_{k_1}+h_1, a_{k_1}+h_1'] \times (a_{k_2}+h_2, a_{k_2}+h_2']
\]
are disjoint.
That the boxes $\cB_{k_1,k_2}$ are disjoint follows from this and the fact that each of $f_1$ and $f_2$ is strictly increasing.
To conclude the proof, note that for each $l\in \{1,2\}$ we have $j_l\leq k_l$. The fact that $f_l$ is strictly increasing and strictly convex yields
\[0<f_l(a_{j_l}+h_l')-f_l(a_{j_l}+h_{l})\leq f_l(a_{k_l}+h_l')-f_l(a_{k_l}+h_l).\]
\end{proof}
Lemma \ref{Expansion1} will produce for us a basic amount of expansion, and can be applied rather generally. From there, we will work locally, but in order to find further expansion we will require additional convexity. Such convexity will be found in a pair of discrete derivatives of the function $f$, provided $f$ is itself sufficiently convex. To that end, let $f_1,f_2:[u,\infty)\to\RR$ be continuous functions defined on some interval, which are each strictly increasing and strictly convex. For a non-negative integer $W$, we say the quadruple $(h_1,h_1',h_2,h_2')$ is $W$-good for $(f_1,f_2)$ if $h_1'>h_1>0$, $h_2'>h_2>0$, and if there is a partition
\[[u,\infty)=I_1\sqcup \cdots\sqcup I_W\]
such that each of the curves
\[
\{(f_1(t+h_1')-f_1(t+h_1),f_2(t+h_2')-f_2(t+h_2)) : t\in I_W\}
\]
is the graph of a strictly convex or strictly concave function.
This definition just allows us to pass to a large subset of a curve which is convex, should the full curve fail to be so. With it in hand, we are ready to state and prove our main result about superquadratic expanders in its most general form. Again, we remark that the spacing condition serves only to provide additional freedom in our future choice of parameters $h_1,\ h_2,\ h_1',$ and $h_2'$.
\begin{Theorem}\label{MainExpander}
Let $f_1,f_2:[a_1,\infty)\to\RR$ be continuous, strictly increasing and strictly convex functions, and suppose $(h_1,h_1',h_2,h_2')$ is $W$-good for $(f_1,f_2)$, for some $W$. Suppose $A=\{a_1<\ldots<a_N\}$ is a finite set of positive real numbers satisfying the spacing condition
\[\min\{a_{i+1}-a_i:i\in[N-1]\}\geq \max\{h_1'-h_1, h_2'-h_2\}.\] Then
\begin{multline*}|2f_1(A+h_1)-2f_1(A+h_1)+2f_1(A+h_1')-f_1(A+h_1')|\cdot\\\cdot|2f_2(A+h_2)-2f_2(A+h_2)+2f_2(A+h_2')-f_2(A+h_2')| \gg \frac{N^5}{(W\log N)^3}.
\end{multline*}
\end{Theorem}
\begin{proof}
Apply Lemma \ref{Expansion1} to deduce that the disjoint boxes $\cB_{k_1,k_2}$ contain all points of the form
\[(f_1(a_{k_1}+h_1)+f_1(a_{l_1}+h_1')-f_1(a_{l_1}+h_1),f_2(a_{k_2}+h_2)+f_2(a_{l_2}+h_2')-f(a_{l_2}+h_2))\]
whenever $1\leq l_1,l_2\leq N/2\leq k_1,k_2\leq N$.
Observe that for fixed values of $k_1$ and $k_2$, the set of such points can be re-written as the cartesian product
\[(f_1(a_{k_1}+h_1),f_2(a_{k_2}+h_2))+g_1(A')\times g_2(A')\]
where
\[A'=\{a_i:i\in[N/2]\}\] and \[g_j(t)=f_j(t+h_j')-f_j(t+h_j).\]
By $W$-goodness, there is an interval $I$ such that
\[
t\in I\mapsto (f_1(a_{k_1}+h_1),f_2(a_{k_2}+h_2)) + (g_1(t),g_2(t))
\]
is the graph of a strictly convex (or concave) function, say $F$, and such that $A''=I\cap A'$ has size at least $N/3W$. Define
\[
X=f_1(a_{k_1}+h_1)+g_1(A''),
\]
which is the projection of these points to the $x$-axis, so that
\[
F(X)=f_2(a_{k_2}+h_2)+g_2(A'').
\]
Apply Lemma \ref{Expansion2} to set $X$ and the function $F$, in order to deduce that
\[
|((X\times F(X))+(X\times F(X))-(X\times F(X)))\cap \cB_{k_1,k_2}|\gg \lr{\frac{N}{W\log N}}^3.
\]
We note that
\[
X+X-X\subseteq 2f_1(A+h_1)-2f_1(A+h_1)+2f_1(A+h_1')-f_1(A+h_1')\]
and
\[
F(X)+F(X)-F(X)\subseteq 2f_2(A+h_2)-2f_2(A+h_2)+2f_2(A+h_2')-f_2(A+h_2').
\]
The boxes $\cB_{k_1,k_2}$ are disjoint, so we conclude the proof by adding the contributions from each choice of $k_1$ and $k_2$.
\end{proof}
\subsection{Proof of Theorem \ref{cor:prodshift2}}
We will now use Theorem \ref{MainExpander} to prove Theorem \ref{cor:prodshift2}. We instead prove it in the form of the following statement, which contains some useful extra information.
\begin{Corollary} \label{cor:prodshift2again}
Let $X$ be a set of positive reals. Suppose that $z_1,z_1', z_2,z_2' \in X$ with $z_1 < z_1'$ and $z_2<z_2'$, and with the additional properties
\begin{itemize}
\item $\log z_1 < \log z_1'$ is a nearest consecutive pair of elements in $\log X$, and
\item $\log z_2 < \log z_2'$ is a second nearest consecutive pair of elements in $\log X$.
\end{itemize}
Then
\begin{equation} \label{longprod}
\left|\frac{(z_1X+1)^{(2)}(z_1'X+1)^{(2)}}{(z_1X+1)^{(2)}(z_1'X+1)}\right| \cdot \left|\frac{(z_2X+1)^{(2)}(z_2'X+1)^{(2)}}{(z_2X+1)^{(2)}(z_2'X+1)}\right|
\gtrsim |X|^5
\end{equation}
In particular, there exist $z,z' \in X$ such that
\begin{equation}\label{again}
\left|\frac{(zX+1)^{(2)}(z'X+1)^{(2)}}{(zX+1)^{(2)}(z'X+1)}\right| \gtrsim |X|^{5/2}.
\end{equation}
\end{Corollary}
\begin{proof}
Let $A=\log(X)$ and write
\[
A=\{a_1<a_2<\dots a_N\}.
\]
Let $A'$ be the set of all elements of $A$ with even indices, relabelled as $A'=\{b_1,b_2,\dots b_{\lfloor N/2 \rfloor} \}$ and observe that $A'$ satisfies the spacing condition
\[
\min \{b_{i+1}-b_i : i \in [\lfloor N /2 \rfloor -1] \} \geq \max \{\log z_1' - \log z_1, \log z_2' - \log z_2\}= \log (z_2'/z_2).
\]
Define $X'=\exp(A')$. Now take $f_1(t)=f_2(t)=f(t):=\log(e^t+1)$ and observe that
\begin{align*}
&|2f(A'+\log(z_1))-2f(A'+\log(z_1))+2f(A'+\log(z_1'))-f(A'+\log(z_1'))|
\\&=\left|\frac{(z_1X'+1)^{(2)}(z_1'X'+1)^{(2)}}{(z_1X'+1)^{(2)}(z_1'X'+1)}\right|.
\end{align*}
Similarly,
\begin{align*}
&|2f(A'+\log(z_2))-2f(A'+\log(z_2))+2f(A'+\log(z_2'))-f(A'+\log(z_2'))|
\\&=\left|\frac{(z_2X'+1)^{(2)}(z_2'X'+1)^{(2)}}{(z_2X'+1)^{(2)}(z_2'X'+1)}\right|.
\end{align*}
With an appropriate value of $W$, we have from Theorem \ref{MainExpander} that
\[
\left|\frac{(z_1X'+1)^{(2)}(z_1'X'+1)^{(2)}}{(z_1X'+1)^{(2)}(z_1'X'+1)}\right| \cdot \left|\frac{(z_2X'+1)^{(2)}(z_2'X'+1)^{(2)}}{(z_2X'+1)^{(2)}(z_2'X'+1)}\right| \gg \frac{|X'|^5}{(W\log |X'|)^3}.
\]
We turn to estimating $W$. The curve in question can be parameterized as a shift of
\[\lr{\log(z_1't+1)-\log(z_1t+1),\log(z_2't+1)-\log(z_2t+1)},\,\,\, t>0\]
and so
\[\frac{dy}{dx}=\frac{z_2'-z_2}{z_1'-z_1}\cdot \frac{z_1z_1't^2 +(z_1+z_1')t+1}{z_2z_2't^2 +(z_2+z_2')t+1}.\]
Therefore
\begin{multline*}
\frac{d^2y}{dx^2}=\frac{z_2'-z_2}{z_1'-z_1} \cdot\frac{(z_1z_1'(z_2+z_2')-z_2z_2'(z_1+z_1'))t^2+2(z_1z_1'-z_2z_2')t+(z_1+z_1'-z_2-z_2')}{(z_2z_2't^2 +(z_2+z_2')t+1)^2},
\end{multline*}
which is not identically zero for positive values of $z_1,\ z_1',\ z_2$ and $z_2'$ unless $z_1=z_2$ and $z_1'=z_2'$. As long as this is guaranteed, there are at most $4$ points at which this curve can change convexity, and so we can take $W=5$. This completes the proof.
\end{proof}
We record one more important corollary here, which is that the pair $z,z'$ satisfying \eqref{again} also satisfies the bound
\[
\left | \frac{(zX+1)(z'X+1)}{(zX+1)} \right | \gg |X|^2
\]
by Corollary \ref{cor:prodshiftquad}. We also make a note here of the obvious fact that $z \neq z'$.
\begin{Corollary} \label{cor:combo}
Let $X$ be a set of positive reals. Then there exists $z,z' \in X$, with $z \neq z'$, such that both
\begin{equation} \label{quadcor}
\left | \frac{(zX+1)(z'X+1)}{(zX+1)} \right | \gg |X|^2
\end{equation}
and
\begin{equation}\label{supercor}
\left|\frac{(zX+1)^{(2)}(z'X+1)^{(2)}}{(zX+1)^{(2)}(z'X+1)}\right| \gtrsim |X|^{5/2}.
\end{equation}
hold.
\end{Corollary}
\subsection{The Pl\"{u}nnecke-Ruzsa inequality and applications}
Before we are ready to start the proof of Theorem \ref{thm:main}, we need to manipulate Pl\"{u}nnecke's inequality in order to deduce from Corollary \ref{cor:combo} a sum-product type theorem which is tailored for the dot product problem. We will prove Corollary \ref{cor:SP}, the key feature of which is that we use only product sets (as opposed to including ratio sets) and still obtain better than quadratic growth. We believe it is of independent interest.
We will use the standard Pl\"{u}nnecke-Ruzsa inequality. See Petridis \cite{P} for a modern proof of the statement.
\begin{Theorem} \label{thm:Plun1}
Let $X$ be a finite set in an additive abelian group. Then for any positive integers $k$ and $l$
\[
|kX-lX| \leq \frac{ |X+X|^{k+l}}{|X|^{k+l-1}}
\]
\end{Theorem}
The Ruzsa Triangle Inequality will be used later in the paper. This is the following result.
\begin{Theorem} \label{thm:RTI}
Let $X,Y,Z$ be finite sets in an additive abelian group. Then
\[
|Y-Z| \leq \frac{|X-Y||X-Z| }{|X|}.
\]
\end{Theorem}
A quick observation is that Theorem \ref{thm:RTI} gives the bound
\begin{equation} \label{RTI2}
|Y-Z|=|(-Y)-(-Z)|\leq \frac{|X-(-Y)||X-(-Z)| }{|X|}=\frac{|X+Y||X+Z| }{|X|}.
\end{equation}
We will also use the following consequence of the Pl\"{u}nnecke-Ruzsa inequality, which was observed by Garaev \cite{G}.
\begin{Theorem} \label{thm:Plun2}
Let $X_1,\dots X_k$ and $Y$ be finite subsets in an additive abelian group. Then
\[
|X_1+\dots + X_k| \leq \frac{|X_1+Y|\cdots |X_k+Y|}{|Y|^{k-1}}.
\]
\end{Theorem}
We also need a non-trivial bound for the size of $A(A+1)$. The following result of Garaev and Shen \cite{GS} is sufficient for our purposes.
\begin{Theorem} \label{thm:GS} For any finite set $A \subset \mathbb R$ and any non-zero $\lambda \in \mathbb R$,
\[
|A(A+\lambda)| \gg |A|^{5/4}.
\]
\end{Theorem}
We remark that improvements to Theorem \ref{thm:GS} are known, and a recent result of Stevens and Warren \cite{SW} gives the better exponent $49/38$. We use Theorem \ref{thm:GS} in the proof of Theorem \ref{thm:main} in order to simplify the exposition slightly. If we instead apply the result of Stevens and Warren, we obtain a further small improvement, (namely $c=1/2739$ instead of $1/3057$).
We are now ready to derive the following consequence of Corollary \ref{cor:combo}.
\begin{Corollary} \label{cor:SP}
Let $X$ be a set of positive reals. Then there exists $z,z' \in X$ such that
\[
|(zX+1)^{(4)}(z'X+1)^{(4)}| \gtrsim |X|^{33/16} .
\]
\end{Corollary}
\begin{proof} Let $z, z' \in X$ be elements given by Corollary \ref{cor:combo} which satisfy the bounds
\begin{equation} \label{quadcoragain}
\left | \frac{(zX+1)(z'X+1)}{(zX+1)} \right | \gg |X|^2
\end{equation}
and
\begin{equation} \label{5/2again}
\left|\frac{(zX+1)^{(2)}(z'X+1)^{(2)}}{(zX+1)^{(2)}(z'X+1)}\right| \gtrsim |X|^{5/2}.
\end{equation}
First we will prove the inequality
\begin{equation} \label{goal1}
|(zX+1)^{(2)}(z'X+1)^{(2)}| \gg |X|^{13/8} .
\end{equation}
Indeed, let us label
\[
S=(zX+1)(z'X+1).
\]
By \eqref{RTI2} and \eqref{quadcoragain}, we have
\begin{equation} \label{inter}
|X|^2 \ll |S/S| \leq \frac{|SS|^2}{|S|}.
\end{equation}
Furthermore, since $z \neq z'$, it follows from Theorem \ref{thm:GS} that
\[
|S|=\left|\frac{z'}{z}(zX+1)(z'X+1)\right|=\left|\left(z'X+\frac{z'}{z}\right)\left(\left(z'X+\frac{z'}{z}\right)+\left(1-\frac{z'}{z}\right)\right) \right| \gg |X|^{5/4}.
\]
Plugging this into \eqref{inter} and rearranging yields \eqref{goal1}.
Now define
\[
T=(zX+1)^{(2)}(z'X+1)^{(2)}.
\]
By \eqref{RTI2} again and \eqref{5/2again},
\[
|X|^{5/2} \lesssim |T/T| \leq \frac{|TT|^2}{|T|}.
\]
We can lower bound $|T|$ using \eqref{goal1}. After doing this and rearranging, it follows that
\[
|TT| \gtrsim |X|^{\frac{33}{16}},
\]
as required.
\end{proof}
\section{Dot products} \label{sec:dp}
\subsection{Controlling the number of slopes with the Szemer\'{e}di-Trotter Theorem}
The next lemma uses the Szemer\'{e}di-Trotter Theorem to give a lower bound for the size of $\Lambda(P)$ in terms of the number of lines through the origin. The result is well-known, and can be used to quickly deduce the threshold bound $|\Lambda(P)| \gg |P|^{2/3}$. Moreover, should it be the case that $P$ is incident to at least $|P|^{1/3+c}$ lines through the origin, the exponent $2/3$ can be improved.
\begin{Lemma} \label{lem:dotbasic}
Let $P \subset \mathbb R^2$ be a finite point set. Let $\mathcal L$ denote the set of all lines through the origin incident to $P$. Then there exists $p \in P$ such that
\[
| \{ p \cdot q : q \in P \}| \gg |P|^{1/2}|\mathcal L|^{1/2}.
\]
\end{Lemma}
\begin{proof}
We may assume without loss of generality that each line in $\mathcal L$ has the form $y= \lambda x$, with $\lambda \neq 0$. Otherwise, we can rotate the whole point set so that it avoids the co-ordinate axes, since all of the quantities involved are rotation invariant. This is just done in order to simplify some notation.
Let $\ell_{\lambda}$ denote the line $y=\lambda x$. For each $\ell_{\lambda} \in \mathcal L$, we choose a point $p_{\lambda} \in P$ on the line $\ell_{\lambda}$ arbitrarily. Consider the set
\[
\{p_{\lambda} \cdot q : q \in P\}
\]
of all dot products determined by $p_{\lambda}$. Note that $p_{\lambda} \cdot q= p_{\lambda} \cdot q'$ if and only if $q$ and $q'$ lie on a line which is perpendicular to $\ell_{\lambda}$. Let $ S_{\lambda}$ denote the set of all lines with slope $-1/\lambda$ which are incident to $P$. It therefore follows that
\begin{equation} \label{equiv}
|S_{\lambda}|=|\{p_{\lambda} \cdot q : q \in P\}|.
\end{equation}
Define $ S$ to be the set of lines
\[
S= \bigcup_{ \ell_{\lambda} \in \mathcal L} S_{\lambda}.
\]
Note that this union is disjoint, since elements of $ S_{\lambda}$ and $ S_{\lambda'}$ have distinct slopes for $\lambda \neq \lambda'$. By the Szemer\'{e}di-Trotter Theorem
\begin{equation} \label{STapp}
|\mathcal L||P|=\sum_{\ell_{\lambda} \in \mathcal L}I(P, S_{\lambda})=I(P, S) \ll |P|^{2/3}|S|^{2/3}+|P|+|S|
\end{equation}
If $|P|$ dominates the right hand side of \eqref{STapp}, then we must have $|\mathcal L| \ll 1$. In this case, there is some line $\ell \in \mathcal L$ such that $|\ell \cap P| \gg |P|$. This set determines $\Omega(|P|)$ dot products, since the set of dot products determined by set of points on a line is in one-to-one correspondence with the product set of the set of magnitudes.
If the first term on the right hand side of \eqref{STapp} is dominant then a rearrangement gives
\begin{equation} \label{linesbound}
|S| \gg |P|^{1/2}|\mathcal L|^{3/2},
\end{equation}
while if $|S|$ dominates the right hand side of \eqref{STapp} then we notice that by definition $|P|\geq|\mathcal L|,$ so we have \[|S| \gg |P||\mathcal L| \geq |P|^{1/2}|\mathcal L|^{3/2}\] as well, and so \eqref{linesbound} holds in either case. Because $S$ is a disjoint union,
\[
\sum_{\ell_{\lambda} \in \mathcal L}| S_{\lambda}| \gg |P|^{1/2}|\mathcal L|^{3/2} .
\]
It then follows from \eqref{equiv} that there is some $p_{\lambda} \in P$ such that
\[
|\{p_{\lambda} \cdot q : q \in P\}| \gg |P|^{1/2}|\mathcal L|^{1/2}.
\]
\end{proof}
We will take this lemma as a launch point for our theorem. The lemma cannot itself be improved without additional information. Indeed, if $P=[N]\times [N]$ then $\cL$ has size of order $N^2$. Meanwhile, any single point $(m,n)$ makes at most $2N^2$ distinct dot products with the rest of $P$, so this result is optimal. The key will be to use a few different lines from $\cL$, whose slopes will introduce different arithmetic constraints and allow us to invoke an expansion result. This lemma puts us in a regime where this approach is viable.
\subsection{Proof of Theorem \ref{thm:main}}
\begin{proof}
We repeat some notation from the proof of the previous subsection: $\mathcal L$ denotes the set of lines through the origin covering $P$, and $\ell_{\lambda} \in \mathcal L$ is the line with equation $y=\lambda x$. Fix $c=\frac{1}{3057}$. The proof is presented in such a way that $c$ can be viewed as a parameter, and we will calculate the optimal choice for $c$ at the end of the proof.
The idea of the proof is to use the assumption of few dot products in order to deduce arithmetic combinatorial constraints associated to the set of points $P$. We do this by considering dot products arising from various subsets of $P$, each of which is chosen to impose a different constraint. The resulting combination of constraints will combine in such a way as to violate the expansion results proved in the previous section.
\emph{Step 1. There is a large subset, $P'$, of points lying on a small set, $\cL'$, of rich lines through the origin. The $x$-axis is in $\cL'$, and the other lines in $\cL'$ have positive slope.}
Lemma \ref{lem:dotbasic} implies the desired bound \eqref{dpgoal} if
\[
|\mathcal L| \geq |P|^{1/3+2c}.
\]
Therefore, we can henceforth assume that
\begin{equation} \label{assumption}
|\mathcal L| \leq |P|^{1/3+2c}.
\end{equation}
We perform a dyadic decomposition of the line set $\mathcal L$ according to the number of points from $P$ they contain. Note that
\[
\sum_{j=1}^{\lceil \log_2 |P| \rceil } \sum_{\ell \in \mathcal L : 2^{j-1} \leq |\ell \cap P| < 2^j}|\ell \cap P|=\sum_{\ell \in \mathcal L} |\ell \cap P|=|P|.
\]
Therefore, there is a set $\mathcal L' \subset \mathcal L$ and some integer $M$ such that for all $ \ell \in \mathcal L'$, $M \leq |\ell \cap P| \leq 2M$ and
\begin{equation} \label{prod}
|\mathcal L'| M \gtrsim |P|.
\end{equation}
It then follows from \eqref{assumption} that
\begin{equation} \label{Mbound}
M \gtrsim |P|^{2/3-2c}.
\end{equation}
For the remaining part of the claim, observe first that at least half of the slopes of the lines in $\mathcal L'$ have the same sign. If the most popular sign is negative, we rotate the whole point set by $90$ degrees so that at least half of the lines in $\mathcal L'$ have positive slope. We can then make a further rotation so that the line with smallest positive slope goes to the $x$-axis.
Abusing notation slightly, we redefine the remaining set of lines as $\mathcal L'$. Let $P'$ denote the set of points from $P$ lying on the lines of $\mathcal L'$. We have $|P'| \geq M| \mathcal L'| \gtrsim |P|$.
\emph{Step 2. There is a large subset, $\cL''$, of lines which intersect the vertical line $x=x_0$ in $P'$.}
Fix a point $p \in P' $ lying on the $x$-axis. The set of dot products
\[
\{ p \cdot q : q \in P'\}
\]
is in bijection with $L_0$, the set of vertical lines incident to $P'$. If we denote by $X$ the set of all $x$-coordinates appearing in $P'$, then $|X|=|L_0|$, and we may therefore assume that $|X| \leq |P|^{2/3+c}$.
For each $x_0 \in X$, define
\[
\mathcal L(x_0) = \{ \ell \in \mathcal L' : \ell \cap \{x=x_0\} \in P'\},
\]
and observe that
\[
\sum_{x_0 \in X} |\mathcal L(x_0)|=|P'| \gtrsim |P|.
\]
It therefore follows by the pigeonhole principle that there exists $x_0 \in X$ such that
\begin{equation} \label{L0}
|\mathcal L(x_0) | \gtrsim \frac{ |P|}{|X|} \geq |P|^{1/3-c}.
\end{equation}
Let $\mathcal L'':=\mathcal L(x_0)$ be the set satisfying \eqref{L0}.
\emph{Step 3. A first appeal to expansion -- the set, $S$, of slopes from $\cL''$ contains two elements, $s$ and $s'$, such that \begin{equation} \label{case1}
|(sS+1)^{(4)}(s'S+1)^{(4)}| \gtrsim |S|^{33/16}.
\end{equation}}
This is a direct application of Corollary \ref{cor:SP} (with the role of $X$ played by $S$).
\emph{Step 4. The sets, $X(s)$ and $X(s')$, of $x$-coordinates from $\ell_s\cap P'$ and $\ell_{s
'}\cap P'$ respectively, have multiplicative structure controlled by $\Lambda(P)$. In particular, there is a non-zero $a \in \mathbb R$, for which the intersection \[
Z=X(s) \cap a X(s')
\]
is large:
\begin{equation} \label{Zbound}
|Z| \gtrsim |P|^{2/3 -8c}.
\end{equation}
}
Recall that all lines in $\mathcal L''$ have approximately $M$ points from $P'$ on them, and so
\[
|X(s)|, |X(s')| \approx M.
\]
Now consider the subset of $\Lambda(P)$ consisting of all dot products between points of $P' \cap \{y=sx \}$ and $P' \cap \{y=s'x \}$. This is the set
\[\Lambda_1=
\{(x,sx) \cdot (x',s'x') : x \in X(s), x' \in X(s') \}=X(s)X(s')(1+ss').
\]
Note that $|\Lambda(P)| \geq |\Lambda_1| = |X(s)X(s')|$. Therefore, we may assume that
\begin{equation} \label{smallproduct}
|X(s)X(s')| \leq |P|^{2/3+c}.
\end{equation}
Similarly, by considering only dot products involving points from $P' \cap \{y=sx \}$, we may assume that
\[
|X(s)X(s)| \leq |P|^{2/3+c}.
\]
It therefore follows from the Ruzsa Triangle Inequality, and specifically \eqref{RTI2}, that
\[
|X(s)/X(s')| \leq \frac{|X(s)X(s')||X(s)X(s)|}{|X(s)|} \leq \frac{|P|^{4/3+2c}}{M} .
\]
On the other hand
\[
M^2 \approx |X(s)||X(s')|=\sum_{a \in X(s)/X(s')} |X(s) \cap a X(s')|.
\]
Therefore, by \eqref{Mbound}, there exists $a \in X(s)/X(s')$ such that
\[
|Z|=|X(s) \cap a X(s')| \geq \frac{M^3}{|P|^{4/3+2c}} \gtrsim |P|^{2/3 -8c}.
\]
\emph{Step 5. The multiplicative structure of $Z$ is controlled by $\Lambda(P)$:
\begin{equation} \label{aim}
|(
sS+1)Z|, |(s'S+1)Z| \leq | \Lambda(P)|.
\end{equation}
}
Consider the following subset of $\Lambda(P)$:
\[
\Lambda_2:=\{(x,sx)\cdot (x_0,s_1 x_0) : x \in Z, s_1 \in S \}.
\]
This is indeed a subset of $\Lambda(P)$, since $(x,sx)$ and $(x_0,s_1 x_0)$ are elements of $P$. Furthermore,
\[
|\Lambda_2|=|\{(xx_0(1+ss_1) : x \in Z, s_1 \in S \}|=|Z(1+sS)|.
\]
The first claimed estimate in \eqref{aim} follows immediately. Similarly, the set
\[
\Lambda_3=\{(x,s'x)\cdot (x_0,s_1 x_0) : x \in a^{-1}Z, s_1 \in S \}
\]
is a subset of $\Lambda(P)$ and has cardinality
\[
|\{xx_0(1+s's_1) : x \in a^{-1}Z, s_1 \in S\}|=|Z(1+s'S)|.
\] This shows the second claimed estimate.
\emph{Step 6. A second appeal to expansion -- the sets $(sS+1)Z$ and $(s'S+1)Z$ cannot both be too small.}
Observe first that
\begin{equation}\label{Sbound}
|S|=|\mathcal L''| \geq |P|^{1/3-c}
\end{equation}
by \eqref{L0}.
Now apply Theorem \ref{thm:Plun2} in the multiplicative setting with $Y=Z$ and $k=8$, along with the bounds \eqref{case1}, \eqref{Sbound} and \eqref{Zbound}, to obtain
\begin{align*} \label{case11}
|(sS+1)Z|^4|(s'S+1)Z|^4 &\geq |Z|^7|(sS+1)^{(4)}(s'S+1)^{(4)}|\\
&\gtrsim |S|^{33/16}|Z|^7 \\
&\gtrsim |P|^{16/3+1/48-929c/16}.
\end{align*}
It follows from this and \eqref{aim} that
\[
|\Lambda(P)| \gtrsim |P|^{2/3 + 1/384 - 929c/128}.
\]
We optimise by setting $1/384-929c/128=c$. That is, $c=1/3057$. This concludes the proof.
\end{proof}
\section{Further expander results}
\subsection{Consequences of Theorem \ref{MainExpander}} \label{sec:other}
Theorem \ref{MainExpander} is flexible and provides other types of superquadratic expansion. For instance we can use it to prove Corollary \ref{cor:cubes}. We recall the statement below.
\begin{Corollary} \label{cor:cubesagain}
Let $A \subset \mathbb R$ be a finite set. Then there exists $a,a' \in A$ such that
\[
|2(a+A)^3 + 2(a'+A)^3 - 2 (a+A)^3 - (a'+A)^3| \gtrsim |A|^{5/2}.
\]
\end{Corollary}
\begin{proof}
Let $(h_1,h_1')$ to be a nearest consecutive pair of elements in $A$ and let $(h_2,h_2')$ be a second nearest consecutive pair of elements in $A$. Write $A=\{a_1<a_2<\dots<a_N\}$, and let $B$ be the subset of $A$ consisting of all elements with even index. Relabelling, we have $B=\{b_1<b_2<\dots<b_M\}$, with $M = \lfloor N/2 \rfloor \gg N$.
Now apply Theorem \ref{MainExpander} with this set $B$ and with $f_1(x)=f_2(x)=x^3$. Note that our choices for $B$ and $(h_1,h_1',h_2,h_2')$ imply that the spacing condition of Theorem \ref{MainExpander} is satisfied. We claim that $W=2$ is permissible in the application of the theorem. Assuming the claim, it then follows that
\begin{multline*}|2(A+h_1)^3-2(A+h_1)^3+2(A+h_1')^3-(A+h_1')^3|\cdot\\\cdot|2(A+h_2)^3-2(A+h_2)^3+2(A+h_2')^3-(A+h_2')^3| \gg \frac{N^5}{(\log N)^3}.
\end{multline*}
Therefore, at least one of the two components of this product is greater $N^{5/2}(\log N)^{-3/2}$, and the proof is complete.
It remains to check that $W=2$ is permissible. The parametric curve we are considering is
\[
((t+h_1')^3- (t+h_1)^3, (t+h_2')^3- (t+h_2)^3).
\]
We calculate that
\[
\frac{d^2y}{dx^2}= \frac{h_2'-h_2}{(h_1'-h_1)^2} \cdot \frac{2(h_1'+h_1-h_2'-h_2)}{3(2t+h_1'+h_1)^3}.
\]
The sign of this expression changes at most once, and therefore the curve can be split into two parts, each of which is strictly convex or concave.
\end{proof}
Theorem \ref{MainExpander} yields several other superquadratic expanders of a similar form, simply by taking $f$ to be a convex function and checking the arguments and corresponding calculations, as we have just done in the proof of Corollary \ref{cor:cubesagain}. For instance, it can be used to prove that there exists $a,a' \in A$ such that
\begin{equation} \label{prodshiftalt}
\left|\frac{(a+A)^{(2)}(a'+A)^{(2)}}{(a+A)^{(2)}(a'+A)} \right | \gtrsim |A|^{5/2}.
\end{equation}
We omit the proof of \eqref{prodshiftalt} as it repeats many of the same calculations in the proof of Corollary \ref{cor:prodshift2again}.
On the other hand, there are some functions where the argument breaks down. For instance, it cannot be used to prove the bound
\[
|2(A+A)^2 + 2(A+A)^2 - 2 (A+A)^2 - (A+A)^2| \gtrsim |A|^{5/2}.
\]
Indeed, this is reassuring, since this bound is not true. If we take $A=\{1,\dots,N\}$ then this is a set of integers contained in the interval $[-12N^2,16N^2]$, and thus has size $O(N^2)$.
On the other hand, we can still use Theorem \ref{MainExpander} to deduce results about growth which concerning sums and differences of $(A+A)^2$. For instance, we can deduce the following result.
\begin{Corollary}
For any $A \subset \mathbb R$, there are elements $a,a'\in A$ such that
\begin{multline} \label{SquaresAndCubes}
|2(A+a)^2-2(A+a)^2+2(A+a')^2-(A+a')^2| \cdot
\\ \cdot |2(A+a)^3-2(A+a)^3+2(A+a')^3-(A+a')^3| \gtrsim |A|^5.
\end{multline}
In particular,
\[
\max \{ |4(A+A)^2-3(A+A)^2|,|4(A+A)^3-3(A+A)^3| \} \gtrsim |A|^{5/2}.
\]
\end{Corollary}
\begin{proof}
We apply Theorem \ref{MainExpander} with $f_1(t)=t^2$ and $f_2(t)=t^3$. This time we take $(h_1,h_1')=(h_2,h_2')$ equal to any pair of nearest neighbours in $A$, which will play the role of $(a,a')$. In this case, the curve in question is parameterized by
\[((t+h_1')^2-(t+h_1)^2,(t+h_1')^3-(t+h_1)^3)\]
which is a parabola, and hence $W=1$ is permissible.
\end{proof}
One can interpret the estimate \eqref{SquaresAndCubes} as saying that, if the set $4(A+A)^2-3(A+A)^2$ attains its minimal possible value $O(N^2)$, then we obtain a cubic order of growth for the set $|4(A+A)^3-3(A+A)^3|$, which is better than that given by Corollary \ref{cor:cubesagain}. The estimate \eqref{SquaresAndCubes} is seen to be best possible in view of the case $X=[N]$.
\subsection{A superquadratic bound for three products and shifts}
We conclude the paper by recording the following consequence of the work in Section \ref{sec:expanders}.
\begin{Corollary} \label{cor:prodshift3}
For any $A \subset \mathbb R$,
\[
|(AA+1)(AA+1)(AA+1)| \gtrsim |A|^{2+\frac{1}{32}}.
\]
\end{Corollary}
In order to prove this, we will need the fact that
\begin{equation} \label{quadprod}
|(AA+1)(AA+1)| \gtrsim |A|^2
\end{equation}
holds for any $A \subset \mathbb R$. This follows from a simple application of the Szemer\'{e}di-Trotter Theorem. The proof of \eqref{quadprod} is a very small modification of the proof of the main result in \cite{RN}, and also exists implicitly in earlier work of Jones \cite{J}. We give the full statement and proof for completeness.
\begin{Lemma} \label{lem:triples}
For any $A \subset \mathbb R$, there exists $a,a' \in A$ such that
\[
|(aA+1)(a'A+1)| \gg \frac{|A|^2}{\log |A|}.
\]
\end{Lemma}
\begin{proof}
Consider the sum
\begin{equation} \label{Esum}
\sum_{a,a' \in A} E^*(aA+1,a'A+1),
\end{equation}
where $E^*(X,Y)$ denotes the usual multiplicative energy between two sets, that is, the number of solutions to the equation
\[
x_1y_1=x_2y_2 ,\,\,\,\,\,x_i \in X, y_i \in Y.
\]
The quantity in \eqref{Esum} is equal to the number of solutions to the equation
\[
(b_1+a^{-1})(b_2+a'^{-1})=(b_3+a^{-1})(b_4+a'^{-1}), \,\,\,\, a,a',b_1,b_2,b_3,b_4 \in A.
\]
Each solution corresponds to a collinear triple $(-a'^{-1},-a^{-1}), (b_4,b_1), (b_2,b_3)$ of three elements of $(A \cup -A^{-1}) \times (A \cup -A^{-1}) $. Therefore, by the Szemer\'{e}di-Trotter Theorem,
\[
\sum_{a,a' \in A} E^*(aA+1,a'A+1) \ll |A|^4 \log |A|.
\]
By pigeonholing, there exists some $a,a' \in A$ such that $E^*(aA+1,a'A+1) \ll |A|^2\log|A|$. An application of the ubiquitous Cauchy-Schwarz bound
\[
E^*(X,Y) \geq \frac{|X|^2|Y|^2}{|XY|}
\]
completes the proof.
\end{proof}
\begin{proof}[Proof of Corollary \ref{cor:prodshift3}]
By \eqref{again} in the statement of Corollary \ref{cor:prodshift2again}, we have
\[
\left|\frac{(AA+1)(AA+1)(AA+1)(AA+1)}{(AA+1)(AA+1)(AA+1)(AA+1)}\right| \gtrsim |A|^{5/2}.
\]
By Theorem \ref{thm:Plun1} and then Theorem \ref{thm:Plun2}, it follows that
\begin{align*}
|A|^{5/2} & \lesssim \left|\frac{(AA+1)(AA+1)(AA+1)(AA+1)}{(AA+1)(AA+1)(AA+1)(AA+1)}\right|
\\ & \leq \frac{|(AA+1)(AA+1)(AA+1)(AA+1)|^4}{|(AA+1)(AA+1)|^3}
\\& \leq \frac{|(AA+1)(AA+1)(AA+1)|^{16}}{|(AA+1)(AA+1)|^{15}}.
\end{align*}
Rearranging and applying Lemma \ref{lem:triples} completes the proof.
\end{proof}
A very similar argument, instead using the bound \eqref{prodshiftalt} as a starting point, can be used to deduce the bound
\[
|(A+A)(A+A)(A+A)| \gtrsim |A|^{2+\frac{1}{32}}.
\]
This improves the bound on the bound $|(A+A)(A+A)(A+A)| \gtrsim |A|^{2+\frac{1}{392}}$, which was proven in \cite{RNS}. We omit the details to avoid repetition.
\subsection*{Acknowledgements} Oliver Roche-Newton was supported by the Austrian Science Fund FWF Projects P 30405-N32 and P 34180. Brandon Hanson was supported by NSF grant 2001622. We are grateful to Ali Mohammadi, Misha Rudnev and Ilya Shkredov for various helpful discussions. Special thanks go to Audie Warren for patiently listening to the details of the paper during its unstable construction phase.
|
\chapter*{basic pictures}
\section{\textbackslash tanglepairingI}
{ \psset{unit=0.2}
\begin{pspicture}[showgrid=false](-11,-7)(11,7)
\rput{-45}(0,0){
\pscircle[linecolor=lightgray](4,4){2.5}
\pscircle[linecolor=lightgray](-4,-4){2.5}
\psline[linecolor=white,linewidth=\stringwhite](1,-4)(-2,-4)
\psline[linewidth=\stringwidth,linecolor=gray](1,-4)(-2,-4)
\psline[linecolor=white,linewidth=\stringwhite](-4,1)(-4,-2)
\psline[linewidth=\stringwidth,linecolor=gray](-4,1)(-4,-2)
\psline[linecolor=white,linewidth=\stringwhite](-6,-4)(-9,-4)
\psline[linewidth=\stringwidth,linecolor=gray](-6,-4)(-9,-4)
\psline[linecolor=white,linewidth=\stringwhite](-4,-6)(-4,-9)
\psline[linewidth=\stringwidth,linecolor=gray](-4,-6)(-4,-9)
\psline[linecolor=white,linewidth=\stringwhite](-1,4)(2,4)
\psline[linewidth=\stringwidth,linecolor=gray](-1,4)(2,4)
\psline[linecolor=white,linewidth=\stringwhite](4,-1)(4,2)
\psline[linewidth=\stringwidth,linecolor=gray](4,-1)(4,2)
\psline[linecolor=white,linewidth=\stringwhite](6,4)(9,4)
\psline[linewidth=\stringwidth,linecolor=gray](6,4)(9,4)
\psline[linecolor=white,linewidth=\stringwhite](4,6)(4,9)
\psline[linewidth=\stringwidth,linecolor=gray](4,6)(4,9)
\pscircle[linestyle=dotted,linewidth=\stringwidth](4,4){5}
\pscircle[linestyle=dotted,linewidth=\stringwidth](-4,-4){5}
\psecurve[linewidth=\stringwidthThick]{C-C}(8,-3)(-1,4)(-10,-3)(-9,-4)(-8,-3)
\psecurve[linewidth=\stringwidth]{C-C}(-3,8)(4,-1)(-3,-10)(-4,-9)(-3,-8)
\psecurve[linewidth=\stringwhite,linecolor=white](-8,3)(1,-4)(10,3)(9,4)(8,3)
\psecurve[linewidth=\stringwhite,linecolor=white](3,-8)(-4,1)(3,10)(4,9)(3,8)
\psecurve[linewidth=\stringwidth]{C-C}(-8,3)(1,-4)(10,3)(9,4)(8,3)
\psecurve[linewidth=\stringwidth]{C-C}(3,-8)(-4,1)(3,10)(4,9)(3,8)
\rput{45}(-4,-4){$T_1$}
\rput{45}(4,4){$T_2$}
}%
\rput{45}(0,0){\rput(4,1){\mystar{0}}}
\rput{45}(0,0){\rput(-4,9){\mystar{0}}}
\end{pspicture}
\section{\textbackslash tanglepairingII}
\begin{pspicture}[showgrid=false](-11,-7)(11,7)
\rput{-45}(0,0){
\pscircle[linecolor=lightgray](4,4){2.5}
\pscircle[linecolor=lightgray](-4,-4){2.5}
\psline[linecolor=white,linewidth=\stringwhite](1,-4)(-2,-4)
\psline[linewidth=\stringwidth,linecolor=gray](1,-4)(-2,-4)
\psline[linecolor=white,linewidth=\stringwhite](-4,1)(-4,-2)
\psline[linewidth=\stringwidth,linecolor=gray](-4,1)(-4,-2)
\psline[linecolor=white,linewidth=\stringwhite](-6,-4)(-9,-4)
\psline[linewidth=\stringwidth,linecolor=gray](-6,-4)(-9,-4)
\psline[linecolor=white,linewidth=\stringwhite](-4,-6)(-4,-9)
\psline[linewidth=\stringwidth,linecolor=gray](-4,-6)(-4,-9)
\psline[linecolor=white,linewidth=\stringwhite](-1,4)(2,4)
\psline[linewidth=\stringwidth,linecolor=gray](-1,4)(2,4)
\psline[linecolor=white,linewidth=\stringwhite](4,-1)(4,2)
\psline[linewidth=\stringwidth,linecolor=gray](4,-1)(4,2)
\psline[linecolor=white,linewidth=\stringwhite](6,4)(9,4)
\psline[linewidth=\stringwidth,linecolor=gray](6,4)(9,4)
\psline[linecolor=white,linewidth=\stringwhite](4,6)(4,9)
\psline[linewidth=\stringwidth,linecolor=gray](4,6)(4,9)
\pscircle[linestyle=dotted,linewidth=\stringwidth](4,4){5}
\pscircle[linestyle=dotted,linewidth=\stringwidth](-4,-4){5}
\psecurve[linewidth=\stringwidth]{C-C}(-2,-1)(1,-4)(4,-1)(1,2)
\psecurve[linewidth=\stringwidth]{C-C}(2,1)(-1,4)(-4,1)(-1,-2)
\psecurve[linewidth=\stringwidthThick]{C-C}(3,8)(4,9)(3,10)(-4.3,4.3)(-10,-3)(-9,-4)(-8,-3)
\psecurve[linewidth=\stringwidth]{C-C}(-3,-8)(-4,-9)(-3,-10)(4.3,-4.3)(10,3)(9,4)(8,3)
\rput{45}(-4,-4){$T_1$}
\rput{45}(4,4){\reflectbox{$T_2$}}
}%
\rput{45}(0,0){\rput(9,-4){\mystar{90}}}
\rput{45}(0,0){\rput(-4,9){\mystar{0}}}
\end{pspicture}
}
{
\psset{unit=0.25\baselineskip}
\section{\textbackslash InlineTrivialTangle}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\pscircle[linestyle=dotted,dotsep=1pt](0,0){2}
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\end{pspicture}
}
{ \psset{unit=0.2}
\section{\textbackslash Tpii}
\begin{pspicture}[showgrid=false](-11,-7)(11,7)
\rput{-45}(0,0){
\rput{180}(0,0){\rput{90}(8,0){
\psecurve[linewidth=\stringwidth,linecolor=red](14,2)(9,4)(3,3)(4,9)(2,14)
\psecurve[linewidth=\stringwhite,linecolor=white](6,-6)(4,-1)(5,5)(-1,4)(-6,6)
\psecurve[linewidth=\stringwidth,linecolor=red](6,-6)(4,-1)(5,5)(-1,4)(-6,6)
\psecurve[linewidth=\stringwhite,linecolor=white](9,4)(3,3)(4,9)(2,14)
\psecurve[linewidth=\stringwidth,linecolor=red](9,4)(3,3)(4,9)(2,14)
\psline[linecolor=red]{->}(1,4.42)(0,4.1)
\psline[linecolor=red]{<-}(3.9,8)(3.58,7)
}}
\pscircle[linestyle=dotted,linewidth=\stringwidth](4,4){5}
\pscircle[linestyle=dotted,linewidth=\stringwidth](-4,-4){5}
\pscircle[linecolor=lightblue](4,4){2.5}
\psline[linecolor=white,linewidth=\stringwhite](-1,4)(2,4)
\psline[linewidth=\stringwidth,linecolor=blue](-1,4)(2,4)
\psline[linecolor=white,linewidth=\stringwhite](4,-1)(4,2)
\psline[linewidth=\stringwidth,linecolor=blue](4,-1)(4,2)
\psline[linecolor=white,linewidth=\stringwhite](6,4)(9,4)
\psline[linewidth=\stringwidth,linecolor=blue](6,4)(9,4)
\psline[linecolor=white,linewidth=\stringwhite](4,6)(4,9)
\psline[linewidth=\stringwidth,linecolor=blue](4,6)(4,9)
\psline[linecolor=blue]{<-}(-0.5,4)(1.5,4
\psline[linecolor=blue]{<-}(4,-0.5)(4,1.5
\psline[linecolor=blue]{<-}(6.5,4)(8.5,4
\psline[linecolor=blue]{<-}(4,6.5)(4,8.5
\psecurve[linewidth=\stringwidthThick]{c-c}(8,-3)(-1,4)(-10,-3)(-9,-4)(-8,-3)
\psecurve[linewidth=\stringwidth]{c-c}(-3,8)(4,-1)(-3,-10)(-4,-9)(-3,-8)
\psecurve[linewidth=\stringwhite,linecolor=white](-8,3)(1,-4)(10,3)(9,4)(8,3)
\psecurve[linewidth=\stringwhite,linecolor=white](3,-8)(-4,1)(3,10)(4,9)(3,8)
\psecurve[linewidth=\stringwidth]{c-c}(-8,3)(1,-4)(10,3)(9,4)(8,3)
\psecurve[linewidth=\stringwidth]{c-c}(3,-8)(-4,1)(3,10)(4,9)(3,8)
\rput{45}(4,4){\textcolor{blue}{$T$}}
}%
\rput{45}(0,0){\rput(4,1){\mystar{0}}}
\rput{45}(0,0){\rput(-4,9){\mystar{0}}}
\end{pspicture}
\section{\textbackslash Tmii}
\begin{pspicture}[showgrid=false](-11,-7)(11,7)
\rput{-45}(0,0){
\rput{180}(0,0){\rput{90}(8,0){
\psecurve[linewidth=\stringwidth,linecolor=red](2,14)(4,9)(3,3)(9,4)(14,2)
\psecurve[linewidth=\stringwhite,linecolor=white](-6,6)(-1,4)(5,5)(4,-1)(6,-6)
\psecurve[linewidth=\stringwidth,linecolor=red](-6,6)(-1,4)(5,5)(4,-1)(6,-6)
\psecurve[linewidth=\stringwhite,linecolor=white](4,9)(3,3)(9,4)(14,2)
\psecurve[linewidth=\stringwidth,linecolor=red](4,9)(3,3)(9,4)(14,2)
\psline[linecolor=red]{->}(1,4.42)(0,4.1)
\psline[linecolor=red]{<-}(8,3.9)(7,3.58)
}}
\pscircle[linestyle=dotted,linewidth=\stringwidth](4,4){5}
\pscircle[linestyle=dotted,linewidth=\stringwidth](-4,-4){5}
\pscircle[linecolor=lightblue](4,4){2.5}
\psline[linecolor=white,linewidth=\stringwhite](-1,4)(2,4)
\psline[linewidth=\stringwidth,linecolor=blue](-1,4)(2,4)
\psline[linecolor=white,linewidth=\stringwhite](4,-1)(4,2)
\psline[linewidth=\stringwidth,linecolor=blue](4,-1)(4,2)
\psline[linecolor=white,linewidth=\stringwhite](6,4)(9,4)
\psline[linewidth=\stringwidth,linecolor=blue](6,4)(9,4)
\psline[linecolor=white,linewidth=\stringwhite](4,6)(4,9)
\psline[linewidth=\stringwidth,linecolor=blue](4,6)(4,9)
\psline[linecolor=blue]{<-}(-0.5,4)(1.5,4
\psline[linecolor=blue]{->}(4,-0.5)(4,1.5
\psline[linecolor=blue]{->}(6.5,4)(8.5,4
\psline[linecolor=blue]{<-}(4,6.5)(4,8.5
\psecurve[linewidth=\stringwidthThick]{c-c}(8,-3)(-1,4)(-10,-3)(-9,-4)(-8,-3)
\psecurve[linewidth=\stringwidth]{c-c}(-3,8)(4,-1)(-3,-10)(-4,-9)(-3,-8)
\psecurve[linewidth=\stringwhite,linecolor=white](-8,3)(1,-4)(10,3)(9,4)(8,3)
\psecurve[linewidth=\stringwhite,linecolor=white](3,-8)(-4,1)(3,10)(4,9)(3,8)
\psecurve[linewidth=\stringwidth]{c-c}(-8,3)(1,-4)(10,3)(9,4)(8,3)
\psecurve[linewidth=\stringwidth]{c-c}(3,-8)(-4,1)(3,10)(4,9)(3,8)
\rput{45}(4,4){\textcolor{blue}{$T$}}
}%
\rput{45}(0,0){\rput(4,1){\mystar{0}}}
\rput{45}(0,0){\rput(-4,9){\mystar{0}}}
\end{pspicture}
\section{\textbackslash Tpn}
\begin{pspicture}[showgrid=false](-11,-7)(11,7)
\rput{-45}(0,0){
\rput{180}(0,0){\psscalebox{-1 1}{\rput{90}(0,0){
\psecurve[linewidth=\stringwidth,linecolor=red](14,2)(9,4)(5,2)(5,4)(3,4
\psecurve[linewidth=\stringwidth,linecolor=red](5,4)(3,4)(3,6)(-1,4)(-6,6)
\psecurve[linewidth=\stringwhite,linecolor=white](2,14)(4,9)(2,5)(4,5)(4,3)(6,3)(4,-1)(6,-6)
\psecurve[linewidth=\stringwidth,linecolor=red](2,14)(4,9)(2,5)(4,5)(4,3
\psecurve[linewidth=\stringwidth,linecolor=red](4,5)(4,3)(6,3)(4,-1)(6,-6)
\psecurve[linewidth=\stringwhite,linecolor=white](9,4)(5,2)(5,4)(3,4)
\psecurve[linewidth=\stringwidth,linecolor=red](9,4)(5,2)(5,4)(3,4)
\psecurve[linewidth=\stringwhite,linecolor=white](5,4)(3,4)(3,6)(-1,4)
\psecurve[linewidth=\stringwidth,linecolor=red](5,4)(3,4)(3,6)(-1,4)
\psline[linewidth=\stringwidth,linecolor=red,linestyle=dotted](3.4,4.6)(4.6,3.4)
\psline[linecolor=red]{->}(-0,4.35)(1,5.2)
\psline[linecolor=red]{<-}(3.6,8)(2.8,7)
}}
\rput{-135}(6,6){\textcolor{red}{$n\Bigg\{$}}
}
\pscircle[linestyle=dotted,linewidth=\stringwidth](4,4){5}
\pscircle[linestyle=dotted,linewidth=\stringwidth](-4,-4){5}
\pscircle[linecolor=lightblue](4,4){2.5}
\psline[linecolor=white,linewidth=\stringwhite](-1,4)(2,4)
\psline[linewidth=\stringwidth,linecolor=blue](-1,4)(2,4)
\psline[linecolor=white,linewidth=\stringwhite](4,-1)(4,2)
\psline[linewidth=\stringwidth,linecolor=blue](4,-1)(4,2)
\psline[linecolor=white,linewidth=\stringwhite](6,4)(9,4)
\psline[linewidth=\stringwidth,linecolor=blue](6,4)(9,4)
\psline[linecolor=white,linewidth=\stringwhite](4,6)(4,9)
\psline[linewidth=\stringwidth,linecolor=blue](4,6)(4,9)
\psline[linecolor=blue]{->}(-0.5,4)(1.5,4)
\psline[linecolor=blue]{->}(4,6.5)(4,8.5)
\psecurve[linewidth=\stringwidthThick]{c-c}(8,-3)(-1,4)(-10,-3)(-9,-4)(-8,-3)
\psecurve[linewidth=\stringwidth]{c-c}(-3,8)(4,-1)(-3,-10)(-4,-9)(-3,-8)
\psecurve[linewidth=\stringwhite,linecolor=white](-8,3)(1,-4)(10,3)(9,4)(8,3)
\psecurve[linewidth=\stringwhite,linecolor=white](3,-8)(-4,1)(3,10)(4,9)(3,8)
\psecurve[linewidth=\stringwidth]{c-c}(-8,3)(1,-4)(10,3)(9,4)(8,3)
\psecurve[linewidth=\stringwidth]{c-c}(3,-8)(-4,1)(3,10)(4,9)(3,8)
\rput{45}(4,4){\textcolor{blue}{$T$}}
}%
\rput{45}(0,0){\rput(4,1){\mystar{0}}}
\rput{45}(0,0){\rput(-4,9){\mystar{0}}}
\end{pspicture}
\section{\textbackslash Tmn}
\begin{pspicture}[showgrid=false](-11,-7)(11,7)
\rput{-45}(0,0){
\rput{180}(0,0){\psscalebox{-1 1}{\rput{90}(0,0){
\psecurve[linewidth=\stringwidth,linecolor=red](2,14)(4,9)(2,5)(4,5)(4,3)
\psecurve[linewidth=\stringwidth,linecolor=red](4,5)(4,3)(6,3)(4,-1)(6,-6)
\psecurve[linewidth=\stringwhite,linecolor=white](14,2)(9,4)(5,2)(5,4)(3,4)(3,6)(-1,4)(-6,6)
\psecurve[linewidth=\stringwidth,linecolor=red](14,2)(9,4)(5,2)(5,4)(3,4)
\psecurve[linewidth=\stringwidth,linecolor=red](5,4)(3,4)(3,6)(-1,4)(-6,6)
\psecurve[linewidth=\stringwhite,linecolor=white](4,9)(2,5)(4,5)(4,3)
\psecurve[linewidth=\stringwidth,linecolor=red](4,9)(2,5)(4,5)(4,3)
\psecurve[linewidth=\stringwhite,linecolor=white](4,5)(4,3)(6,3)(4,-1)
\psecurve[linewidth=\stringwidth,linecolor=red](4,5)(4,3)(6,3)(4,-1)
\psline[linewidth=\stringwidth,linecolor=red,linestyle=dotted](3.4,4.6)(4.6,3.4)
\psline[linecolor=red]{->}(-0,4.35)(1,5.2)
\psline[linecolor=red]{<-}(3.6,8)(2.8,7)
}}
\rput{-135}(6,6){\textcolor{red}{$n\Bigg\{$}}
}
\pscircle[linestyle=dotted,linewidth=\stringwidth](4,4){5}
\pscircle[linestyle=dotted,linewidth=\stringwidth](-4,-4){5}
\pscircle[linecolor=lightblue](4,4){2.5}
\psline[linecolor=white,linewidth=\stringwhite](-1,4)(2,4)
\psline[linewidth=\stringwidth,linecolor=blue](-1,4)(2,4)
\psline[linecolor=white,linewidth=\stringwhite](4,-1)(4,2)
\psline[linewidth=\stringwidth,linecolor=blue](4,-1)(4,2)
\psline[linecolor=white,linewidth=\stringwhite](6,4)(9,4)
\psline[linewidth=\stringwidth,linecolor=blue](6,4)(9,4)
\psline[linecolor=white,linewidth=\stringwhite](4,6)(4,9)
\psline[linewidth=\stringwidth,linecolor=blue](4,6)(4,9)
\psline[linecolor=blue]{->}(-0.5,4)(1.5,4)
\psline[linecolor=blue]{->}(4,6.5)(4,8.5)
\psecurve[linewidth=\stringwidthThick]{c-c}(8,-3)(-1,4)(-10,-3)(-9,-4)(-8,-3)
\psecurve[linewidth=\stringwidth]{c-c}(-3,8)(4,-1)(-3,-10)(-4,-9)(-3,-8)
\psecurve[linewidth=\stringwhite,linecolor=white](-8,3)(1,-4)(10,3)(9,4)(8,3)
\psecurve[linewidth=\stringwhite,linecolor=white](3,-8)(-4,1)(3,10)(4,9)(3,8)
\psecurve[linewidth=\stringwidth]{c-c}(-8,3)(1,-4)(10,3)(9,4)(8,3)
\psecurve[linewidth=\stringwidth]{c-c}(3,-8)(-4,1)(3,10)(4,9)(3,8)
\rput{45}(4,4){\textcolor{blue}{$T$}}
}%
\rput{45}(0,0){\rput(4,1){\mystar{0}}}
\rput{45}(0,0){\rput(-4,9){\mystar{0}}}
\end{pspicture}
\section{\textbackslash Tpiin}
\begin{pspicture}[showgrid=false](-12,-7)(12,7)
\rput{-45}(0,0){
\pscircle[linestyle=dotted,linewidth=\stringwidth](4,4){5}
\pscircle[linestyle=dotted,linewidth=\stringwidth](-4,-4){5}
\rput{180}(0,0){\psscalebox{-1 1}{\rput{90}(0,0){
\psecurve[linewidth=\stringwidth,linecolor=red](14,2)(9,4)(5,2)(5,4)(3,4
\psecurve[linewidth=\stringwidth,linecolor=red](5,4)(3,4)(3,6)(-1,4)(-6,6)
\psecurve[linewidth=\stringwhite,linecolor=white](2,14)(4,9)(2,5)(4,5)(4,3)(6,3)(4,-1)(6,-6)
\psecurve[linewidth=\stringwidth,linecolor=red](2,14)(4,9)(2,5)(4,5)(4,3
\psecurve[linewidth=\stringwidth,linecolor=red](4,5)(4,3)(6,3)(4,-1)(6,-6)
\psecurve[linewidth=\stringwhite,linecolor=white](9,4)(5,2)(5,4)(3,4)
\psecurve[linewidth=\stringwidth,linecolor=red](9,4)(5,2)(5,4)(3,4)
\psecurve[linewidth=\stringwhite,linecolor=white](5,4)(3,4)(3,6)(-1,4)
\psecurve[linewidth=\stringwidth,linecolor=red](5,4)(3,4)(3,6)(-1,4)
\psline[linewidth=\stringwidth,linecolor=red,linestyle=dotted](3.4,4.6)(4.6,3.4)
\psline[linecolor=red]{->}(-0,4.35)(1,5.2)
\psline[linecolor=red]{<-}(3.6,8)(2.8,7)
\psline[linecolor=red]{->}(8,3.6)(7,2.8)
\psline[linecolor=red]{<-}(4.35,0)(5.2,1)
}}
\rput{-135}(7.5,7.5){\pscircle*[linecolor=white](0,0){1}}
\rput{-135}(6.5,6.5){\textcolor{red}{$2n\Bigg\{$}}
}
\pscircle[linecolor=lightblue](4,4){2.5}
\psline[linecolor=white,linewidth=\stringwhite](-1,4)(2,4)
\psline[linewidth=\stringwidth,linecolor=blue](-1,4)(2,4)
\psline[linecolor=white,linewidth=\stringwhite](4,-1)(4,2)
\psline[linewidth=\stringwidth,linecolor=blue](4,-1)(4,2)
\psline[linecolor=white,linewidth=\stringwhite](6,4)(9,4)
\psline[linewidth=\stringwidth,linecolor=blue](6,4)(9,4)
\psline[linecolor=white,linewidth=\stringwhite](4,6)(4,9)
\psline[linewidth=\stringwidth,linecolor=blue](4,6)(4,9)
\psline[linecolor=blue]{->}(-0.5,4)(1.5,4)
\psline[linecolor=blue]{<-}(4,-0.5)(4,1.5)
\psline[linecolor=blue]{<-}(6.5,4)(8.5,4)
\psline[linecolor=blue]{->}(4,6.5)(4,8.5)
\psecurve[linewidth=\stringwidthThick]{c-c}(8,-3)(-1,4)(-10,-3)(-9,-4)(-8,-3)
\psecurve[linewidth=\stringwidth]{c-c}(-3,8)(4,-1)(-3,-10)(-4,-9)(-3,-8)
\psecurve[linewidth=\stringwhite,linecolor=white](-8,3)(1,-4)(10,3)(9,4)(8,3)
\psecurve[linewidth=\stringwhite,linecolor=white](3,-8)(-4,1)(3,10)(4,9)(3,8)
\psecurve[linewidth=\stringwidth]{c-c}(-8,3)(1,-4)(10,3)(9,4)(8,3)
\psecurve[linewidth=\stringwidth]{c-c}(3,-8)(-4,1)(3,10)(4,9)(3,8)
\rput{45}(4,4){\textcolor{blue}{$T$}}
}%
\rput{45}(0,0){\rput(4,1){\mystar{0}}}
\rput{45}(0,0){\rput(-4,9){\mystar{0}}}
\end{pspicture}
\section{\textbackslash Tmiin}
\begin{pspicture}[showgrid=false](-12,-7)(12,7)
\rput{-45}(0,0){
\pscircle[linestyle=dotted,linewidth=\stringwidth](4,4){5}
\pscircle[linestyle=dotted,linewidth=\stringwidth](-4,-4){5}
\rput{180}(0,0){\psscalebox{-1 1}{\rput{90}(0,0){
\psecurve[linewidth=\stringwidth,linecolor=red](2,14)(4,9)(2,5)(4,5)(4,3)
\psecurve[linewidth=\stringwidth,linecolor=red](4,5)(4,3)(6,3)(4,-1)(6,-6)
\psecurve[linewidth=\stringwhite,linecolor=white](14,2)(9,4)(5,2)(5,4)(3,4)(3,6)(-1,4)(-6,6)
\psecurve[linewidth=\stringwidth,linecolor=red](14,2)(9,4)(5,2)(5,4)(3,4)
\psecurve[linewidth=\stringwidth,linecolor=red](5,4)(3,4)(3,6)(-1,4)(-6,6)
\psecurve[linewidth=\stringwhite,linecolor=white](4,9)(2,5)(4,5)(4,3)
\psecurve[linewidth=\stringwidth,linecolor=red](4,9)(2,5)(4,5)(4,3)
\psecurve[linewidth=\stringwhite,linecolor=white](4,5)(4,3)(6,3)(4,-1)
\psecurve[linewidth=\stringwidth,linecolor=red](4,5)(4,3)(6,3)(4,-1)
\psline[linewidth=\stringwidth,linecolor=red,linestyle=dotted](3.4,4.6)(4.6,3.4)
\psline[linecolor=red]{->}(-0,4.35)(1,5.2)
\psline[linecolor=red]{<-}(3.6,8)(2.8,7)
\psline[linecolor=red]{->}(8,3.6)(7,2.8)
\psline[linecolor=red]{<-}(4.35,0)(5.2,1)
}}
\rput{-135}(7.5,7.5){\pscircle*[linecolor=white](0,0){1}}
\rput{-135}(6.6,6.6){\textcolor{red}{$2|n|\Bigg\{$}}
}
\pscircle[linecolor=lightblue](4,4){2.5}
\psline[linecolor=white,linewidth=\stringwhite](-1,4)(2,4)
\psline[linewidth=\stringwidth,linecolor=blue](-1,4)(2,4)
\psline[linecolor=white,linewidth=\stringwhite](4,-1)(4,2)
\psline[linewidth=\stringwidth,linecolor=blue](4,-1)(4,2)
\psline[linecolor=white,linewidth=\stringwhite](6,4)(9,4)
\psline[linewidth=\stringwidth,linecolor=blue](6,4)(9,4)
\psline[linecolor=white,linewidth=\stringwhite](4,6)(4,9)
\psline[linewidth=\stringwidth,linecolor=blue](4,6)(4,9)
\psline[linecolor=blue]{->}(-0.5,4)(1.5,4)
\psline[linecolor=blue]{<-}(4,-0.5)(4,1.5)
\psline[linecolor=blue]{<-}(6.5,4)(8.5,4)
\psline[linecolor=blue]{->}(4,6.5)(4,8.5)
\psecurve[linewidth=\stringwidthThick]{c-c}(8,-3)(-1,4)(-10,-3)(-9,-4)(-8,-3)
\psecurve[linewidth=\stringwidth]{c-c}(-3,8)(4,-1)(-3,-10)(-4,-9)(-3,-8)
\psecurve[linewidth=\stringwhite,linecolor=white](-8,3)(1,-4)(10,3)(9,4)(8,3)
\psecurve[linewidth=\stringwhite,linecolor=white](3,-8)(-4,1)(3,10)(4,9)(3,8)
\psecurve[linewidth=\stringwidth]{c-c}(-8,3)(1,-4)(10,3)(9,4)(8,3)
\psecurve[linewidth=\stringwidth]{c-c}(3,-8)(-4,1)(3,10)(4,9)(3,8)
\rput{45}(4,4){\textcolor{blue}{$T$}}
}%
\rput{45}(0,0){\rput(4,1){\mystar{0}}}
\rput{45}(0,0){\rput(-4,9){\mystar{0}}}
\end{pspicture}
}
{
\psset{unit=0.29}
\section{\textbackslash splittangle}
\newcommand{5.656854249492381}{5.656854249492381}
\begin{pspicture}[showgrid=false](-6.5,-1)(6.5,12)
\rput{90}(0,0){%
\rput{-45}(0,0){%
{
\psarc[linewidth=\stringwidth](-1,-1){5}{0}{90}
\psarc[linewidth=\stringwidth](9,9){5}{180}{270}
}
\pscircle[linewidth=\stringwidth,linestyle=dotted](4,4){5}
}
\rput(5.656854249492381,0){%
\rput(-2.25,0){%
\pscircle[fillcolor=white,fillstyle=solid](0,0){10pt}%
\rput{-90}(0,0){$T_2$}%
}
\rput(2.25,0){%
\pscircle[fillcolor=white,fillstyle=solid](0,0){10pt}%
\rput{-90}(0,0){$T_1$}%
}%
}%
\rput{-45}(0,0){%
\rput{45}(4,4){%
\rput(5;45){\mystar{15}}
}%
}%
}
\end{pspicture}
}
{
\psset{unit=0.25}
\section{\textbackslash ECSCIspecials}
\begin{pspicture}[showgrid=false](-2,-9)(19,9)
\psline*[linecolor=lightgray](-2,9)(10,9)(10,-9)(-2,-9)(-2,9)
\psline[style=para](10,-9)(10,9)
\psline[style=curveblue](5,-7)(5,7)
\psline[style=curveblue](15,-7)(15,7)
\psline[style=curvered, linearc=3]%
(0,0)(2,2)(16,2)
\psline[style=curvered, linearc=3]%
(0,0)(2,-2)(16,-2)
\pscircle[style=sp](0,0){3pt}%
\rput(10,2){%
\pscircle[linecolor=red,fillcolor=red,fillstyle=solid](0,0){2pt}
\uput{0.5}[45]{0}(0,0){\footnotesize\(\QGrad{-2}\)}
}
\rput(10,-2){%
\pscircle[linecolor=red,fillcolor=red,fillstyle=solid](0,0){2pt}
\uput{0.5}[-45]{0}(0,0){\footnotesize\(\QGrad{-4}\)}
}
\rput(5,-2){%
\pscircle[linecolor=black,fillcolor=black,fillstyle=solid](0,0){2pt}
\rput{-90}(0,0){\psline[arrows=->,arrowsize=1.5pt 1.5, linearc=0.2](-0.3,2)(-0.3,0.3)(-2,0.3)}
}
\rput(5,2){%
\pscircle[linecolor=black,fillcolor=black,fillstyle=solid](0,0){2pt}
\rput{-90}(0,0){\psline[arrows=->,arrowsize=1.5pt 1.5, linearc=0.2](-0.3,2)(-0.3,0.3)(-2,0.3)}
}
\rput(15,-2){%
\pscircle[linecolor=black,fillcolor=black,fillstyle=solid](0,0){2pt}
\rput{90}(0,0){\psline[arrows=->,arrowsize=1.5pt 1.5, linearc=0.2](-0.3,2)(-0.3,0.3)(-2,0.3)}
}
\rput(15,2){%
\pscircle[linecolor=black,fillcolor=black,fillstyle=solid](0,0){2pt}
\rput{90}(0,0){\psline[arrows=->,arrowsize=1.5pt 1.5, linearc=0.2](-0.3,2)(-0.3,0.3)(-2,0.3)}
}
\rput[l](16.5,2){\textcolor{red}{\(\a_+\)}}
\rput[l](16.5,-2){\textcolor{red}{\(\a_-\)}}
\rput[b](5,7.5){\textcolor{blue}{\(\gamma_i\)}}
\rput[t](15,-7.5){\textcolor{blue}{\(\gamma_i\)}}
\end{pspicture}
\section{\textbackslash ECSCIIspecials}
\begin{pspicture}[showgrid=false](-12,-9)(12,9)
\psline*[linecolor=lightgray](-5,5)(-5,-9)(5,-9)(5,5)
\psline*[linecolor=lightgray](-5,5)(-10,5)(-10,9)(-5,9)
\psline*[linecolor=lightgray](5,5)(10,5)(10,9)(5,9)
\psline[style=para](5,-9)(5,5)(10,5)
\psline[style=para](-5,-9)(-5,5)(-10,5)
\psline[style=sp]%
(-5,9)(-5,5)(5,5)(5,9)
\psdot[dotstyle=asterisk,dotsize=9pt,dotangle=15](0,8)
\psline[linecolor=black]{->}(5,3)(-5,3)
\rput[b](0,3.25){\textcolor{black}{\(D\)}}
\psline[style=curveblue](8,-7)(8,8)
\psline[style=curveblue](-8,-7)(-8,8)
\psline[style=curvered](10,0)(-10,0)
\psline[style=curvered](10,-4)(-10,-4)
\pscircle[linecolor=blue,fillcolor=white,fillstyle=solid](8,5){2pt}
\pscircle[linecolor=blue,fillcolor=white,fillstyle=solid](-8,5){2pt}
\rput(5,0){%
\pscircle[linecolor=red,fillcolor=red,fillstyle=solid](0,0){2pt}
\uput{0.5}[135]{0}(0,0){\footnotesize\(\QGrad{4j-3-2n}\)}
}
\rput(5,-4){%
\pscircle[linecolor=red,fillcolor=red,fillstyle=solid](0,0){2pt}
\uput{0.5}[135]{0}(0,0){\footnotesize\(\QGrad{2n+1-4j}\)}
}
\rput(-5,0){%
\pscircle[linecolor=red,fillcolor=red,fillstyle=solid](0,0){2pt}
\uput{0.5}[-45]{0}(0,0){\footnotesize\(\QGrad{4j-1-2n}\)}
}
\rput(-5,-4){%
\pscircle[linecolor=red,fillcolor=red,fillstyle=solid](0,0){2pt}
\uput{0.5}[-45]{0}(0,0){\footnotesize\(\QGrad{2n+3-4j}\)}
}
\rput(8,0){%
\pscircle[linecolor=black,fillcolor=black,fillstyle=solid](0,0){2pt}
\rput{90}(0,0){\psline[arrows=->,arrowsize=1.5pt 1.5, linearc=0.2](-0.3,2)(-0.3,0.3)(-2,0.3)}
}
\rput(8,-4){%
\pscircle[linecolor=black,fillcolor=black,fillstyle=solid](0,0){2pt}
\rput{90}(0,0){\psline[arrows=->,arrowsize=1.5pt 1.5, linearc=0.2](-0.3,2)(-0.3,0.3)(-2,0.3)}
}
\rput(-8,0){%
\pscircle[linecolor=black,fillcolor=black,fillstyle=solid](0,0){2pt}
\rput{-90}(0,0){\psline[arrows=->,arrowsize=1.5pt 1.5, linearc=0.2](-0.3,2)(-0.3,0.3)(-2,0.3)}
}
\rput(-8,-4){%
\pscircle[linecolor=black,fillcolor=black,fillstyle=solid](0,0){2pt}
\rput{-90}(0,0){\psline[arrows=->,arrowsize=1.5pt 1.5, linearc=0.2](-0.3,2)(-0.3,0.3)(-2,0.3)}
}
\rput[l](10.5,0){\textcolor{red}{\(\a_+\)}}
\rput[l](10.5,-4){\textcolor{red}{\(\a_-\)}}
\rput[t](8,-7.5){\textcolor{blue}{\(\gamma_i\)}}
\rput[t](-8,-7.5){\textcolor{blue}{\(\gamma_i\)}}
\end{pspicture}
\section{\textbackslash AGCCCzeroSpecials}
\begin{pspicture}[showgrid=false](-9,-8)(9,8)
\psline*[linecolor=lightgray](-9,0)(-9,8)(9,8)(9,0)(-9,0)
\psline[style=para](9,0)(-9,0)
\psline[style=curveblue](-5,4)(5,4)
\psline[style=curveblue](-5,-4)(5,-4)
\psline[style=curvered](0,6)(0,-6)%
\pscircle[style=sp](0,6){3pt}%
\rput(0,0){%
\pscircle[linecolor=red,fillcolor=white,fillstyle=solid](0,0){2pt}
\uput{0.5}[-135]{0}(0,0){\footnotesize\(\QGrad{-4n}\)}
}
\rput(0,4){%
\pscircle[linecolor=black,fillcolor=black,fillstyle=solid](0,0){2pt}
\rput{180}(0,0){\psline[arrows=->,arrowsize=1.5pt 1.5, linearc=0.2](-0.3,2)(-0.3,0.3)(-2,0.3)}
}
\rput(0,-4){%
\pscircle[linecolor=black,fillcolor=black,fillstyle=solid](0,0){2pt}
\rput{0}(0,0){\psline[arrows=->,arrowsize=1.5pt 1.5, linearc=0.2](-0.3,2)(-0.3,0.3)(-2,0.3)}
}
\rput[t](0,-6.5){\textcolor{red}{\(\a_{\nicefrac{1}{2n}}\)}}
\rput[l](5.5,4){\textcolor{blue}{\(\Sigma\)}}
\rput[r](-5.5,-4){\textcolor{blue}{\(\Sigma\)}}
\end{pspicture}
\section{\textbackslash AGCCCzeroRationals}
\begin{pspicture}[showgrid=false](-14,-7)(14,9)
\psline*[linecolor=lightgray](-8,9)(-8,4)(8,4)(8,9)
\psline*[linecolor=lightgray](-14,4)(-8,4)(-8,-7)(-14,-7)
\psline*[linecolor=lightgray](14,4)(8,4)(8,-7)(14,-7)
\psline[style=para](-8,-7)(-8,4)(8,4)(8,-7)
\psline[style=sp](8,9)(8,4)(14,4)
\psline[style=sp](-8,9)(-8,4)(-14,4)
\psdot[dotstyle=asterisk,dotsize=9pt,dotangle=15](11,7)
\psdot[dotstyle=asterisk,dotsize=9pt,dotangle=15](-11,7)
\psarcn[linecolor=black]{->}(8,4){2.3}{-90}{-180}
\psarcn[linecolor=black]{->}(-8,4){2.3}{0}{-90}
\uput{0.6}[-135]{0}(8,4){\textcolor{black}{\(S\)}}
\uput{0.6}[-45]{0}(-8,4){\textcolor{black}{\(S\)}}
\psline[style=curveblue](-10,-2)(10,-2)
\psline[style=curvered](0,7)(0,-4)
\psline[style=curvered,linearc=1](9,0)(0,0)(0,-2)
\psline[style=curvered,linearc=1](-9,0)(0,0)(0,-2)
\pscircle[style=sp](0,-4){3pt}%
\pscircle[style=sp](0,7){3pt}%
\pscircle[style=sp](11,-4){3pt}%
\pscircle[style=sp](-11,-4){3pt}%
\pscircle[linecolor=blue,fillcolor=blue,fillstyle=solid](8,-2){2pt}
\pscircle[linecolor=blue,fillcolor=blue,fillstyle=solid](-8,-2){2pt}
\rput(8,0){%
\pscircle[linecolor=red,fillcolor=red,fillstyle=solid](0,0){2pt}
\uput{0.5}[135]{0}(0,0){\footnotesize\(\QGrad{-1}\)}
}
\rput(-8,0){%
\pscircle[linecolor=red,fillcolor=red,fillstyle=solid](0,0){2pt}
\uput{0.5}[45]{0}(0,0){\footnotesize\(\QGrad{+1}\)}
}
\rput(0,4){%
\pscircle[linecolor=red,fillcolor=white,fillstyle=solid](0,0){2pt}
\uput{0.5}[-45]{0}(0,0){\footnotesize\(\QGrad{0}\)}
}
\rput(0,-2){%
\pscircle[linecolor=black,fillcolor=black,fillstyle=solid](0,0){2pt}
\rput{0}(0,0){\psline[arrows=->,arrowsize=1.5pt 1.5, linearc=0.2](-0.3,1)(-0.3,0.3)(-2.5,0.3)}
}
\rput[b](-1.75,0.5){\textcolor{red}{\(\a_{\nicefrac{1}{2n}}\)}}
\rput[l](9.5,0){\textcolor{red}{\(n>0\)}}
\rput[b](3,5){\textcolor{red}{\(n=0\)}}
\rput[r](-9.5,0){\textcolor{red}{\(n<0\)}}
\rput[l](10.5,-2){\textcolor{blue}{\(P\)}}
\end{pspicture}
}
\section{\textbackslash pretzeltangleKh}
{
\psset{unit=0.3}
\begin{pspicture}[showgrid=false](-7.3333333,-7.3333333)(7.3333333,7.3333333
\psellipticarc[linecolor=lightgray!50!white](0,0)(6.5,1.1){0}{180}
\rput{-45}(0,0){\rput(-4,-4){%
\psset{linewidth=\stringwidth}
\psecurve[linecolor=black]
(-5,3)%
(-1,4)(3,3)(4,-1)%
(3,-5)
\psecurve[linewidth=\stringwhite,linecolor=white](5,7)%
(2,6)%
(2,2)%
(6,2)%
(7,5)
\psecurve[linecolor=black](5,7)%
(2,6)%
(2,2)%
(6,2)%
(7,5)
\psecurve[linewidth=\stringwhite,linecolor=white]
(-1,4)(3,3)(4,-1)%
(3,-5)
\psecurve[linecolor=black]
(-1,4)(3,3)(4,-1)%
(3,-5)
\psecurve[linecolor=black]%
(2,2)%
(6,2)%
(7,5)(3,5)(4,9)(3,13)
\psecurve[linewidth=\stringwhite,linecolor=white]%
(13,3)(9,4)(5,3)(5,7)%
(2,6)%
(2,2)%
\psecurve[linecolor=black]%
(13,3)(9,4)(5,3)(5,7)%
(2,6)%
(2,2)%
\psecurve[linewidth=\stringwhite,linecolor=white]%
(6,2)%
(7,5)(3,5)(4,9)
\psecurve[linecolor=black]%
(6,2)%
(7,5)(3,5)(4,9)
}}
\psellipticarc[linewidth=3pt,linecolor=white](0,-0.13)(6.5,1.1){180}{0}
\psellipticarc[linecolor=gray](0,0)(6.5,1.1){180}{0}
\pscircle(0,0){6.5}
\psecurve[linewidth=1.5\stringwidth,linecolor=white](5;135)(4;65)(4.5;30)(5.6;35)(5.7;90)
\psecurve[style=para](-4,2.5)(-3,3)(4;65)(4.5;30)(5.6;35)(5.7;90)(-3.8,3.8)(4;135)
\rput{-90}(0,0){%
\psecurve[linewidth=1.5\stringwidth,linecolor=white](5;-135)(4;-65)(4.5;-30)(5.6;-35)(5.7;-90)
\psecurve[style=para](-4,-2.5)(-3,-3)(4;-65)(4.5;-30)(5.6;-35)(5.7;-90)(-3.8,-3.8)(4;-135)}
\rput{-45}(0,0){
\pscircle[style=sp](5;0){2pt}%
\pscircle[style=sp](5;90){2pt}%
\pscircle[style=sp](5;270){2pt}%
}
\pscustom[style=sp,linearc=0.35,linejoin=2,linecolor=white,linewidth=3pt]{
\psline[liftpen=1](-3,3)(-2.5,4.2)(-3.8,3.8)
\psline[liftpen=1](-3.8,3.8)(-4.2,2.5)(-3,3)
\closepath
}
\pscustom[style=sp,linearc=0.35,linejoin=2]{
\psline[liftpen=1](-3,3)(-2.5,4.2)(-3.8,3.8)
\psline[liftpen=1](-3.8,3.8)(-4.2,2.5)(-3,3)
\closepath
}
\rput{45}(0,0){
\rput(0,4.75){\scalebox{1.1}[0.55]{\huge$\ast$}}
}%
\end{pspicture}
}
{\psset{unit=1.1}
\section{\textbackslash pretzeltangleDownstairsKhr}
\begin{pspicture}(-2,-2)(2,2)
\KhparaStart
\rput(0,-0.5){\psset{xunit=1.3, yunit=1.5}
\psecurve[linecolor=blue]%
(1,1.5)(-0.7,1.4)(-1.3,1)(-1,0.6)
(1,1.2)(1.2,1)
(0.4,0.5)(-1,0.4)(-1.4,1)(-1,1.4)
(0,1.45)
(1,1.4)(1.4,1)(1,0.4)(-0.4,0.7)
(-1.18,1)(-1,1.2)
(1,0.6)(1.3,1)(0.7,1.4)(-1,1.5)(-1.5,1)(-1.3,0.4)
(0,0.1)
(1.3,0.4)(1.5,1)(1,1.5)(-0.7,1.4)(-1.3,1)
}
\psecurve[linecolor=blue,linestyle=dashed]%
(1.4,1)(1.1,1.25)(-0.4,0)(-1.2,-1.1)(1.3,-0.95)(0,-1.4)(-1.4,-1.2)(-1.3,-0.55)(0,0.1)(1.4,1)(1.1,1.25)(-0.4,0)
\KhparaEnd
\end{pspicture}
}
{
\psset{unit=0.25}
\section{\textbackslash pretzeltangleUpstairsKhr}
\begin{pspicture}[showgrid=false](-5,-1)(29,9)
\multido{\i=-16+8}{6}{
\multido{\ii=-12+8}{3}{
\rput(\i,\ii){%
\Khpara
}
}
}
\rput{180}(4,4){%
\psecurve[linestyle=dashed,style=curveblue]%
(-16,-8)(-7.5,-4.5)(-0.5,0.5)(8.5,3.5)(16,8)
}
\rput(2,4){%
\psline[style=curveblue,linearc=2]%
(-6,-1.5)(-1,-1.5)(2,1.5)(15,1.5)(18,-1.5)(26,-1.5)%
}%
\end{pspicture}
}
{\psset{unit=1.1}
\section{\textbackslash pretzeltangleDownstairsBNr}
\begin{pspicture}(-2,-2)(2,2)
\KhparaStart
\psecurve[style=curveblue]%
(3,1.4)(1,1)
(-1.3,1.3)(-1.3,0.7)
(1.3,0.7)(1.3,1.3)
(-1.5,1.5)(-1.5,0.5)
(1.5,0.6)(1.4,1.4)
(-0.2,0.2)
(-1.4,-1)(-1,-1.4)
(1,-1)(3,-1.4)
\KhparaEnd
\end{pspicture}
}
{
\psset{unit=0.25}
\section{\textbackslash pretzeltangleUpstairsBNr}
\begin{pspicture}[showgrid=false](-5,-1)(21,9)
\multido{\i=-8+8}{6}{
\multido{\ii=-12+8}{3}{
\rput(\i,\ii){%
\Khpara
}
}
}
\rput(2,4){%
\psline[style=curveblue,linearc=2](-6,-4)(3.5,1.5)(17,1.5)(18,0)%
\pscircle[style=sp](18,0){2pt}%
\pscircle[style=sp](-6,-4){2pt}%
}%
\end{pspicture}
}
{
\psset{unit=1.1,arrowsize=1.5pt 2}
\section{\textbackslash PairingTrefoilArcINTRO}
\begin{pspicture}(-2,-2)(2,2)
\KhparaStartLabels
\psecurve[style=curveblue]%
(-2,-2)(-1,-1)%
(1.7,-0.6)(1.75,-1.5)%
(0,-1.9)%
(-1.3,-1.4)(-1.3,-0.8)%
(1,1)(1.2,2)
\rput(-0.305,-0.67){\pscircle[linecolor=blue,fillcolor=blue,fillstyle=solid](0,0){2pt}}
\rput(-0.84,-1.74){\pscircle[linecolor=blue,fillcolor=blue,fillstyle=solid](0,0){2pt}}
\rput(-0.55,-0.08){\pscircle[linecolor=blue,fillcolor=blue,fillstyle=solid](0,0){2pt}}
\rput(0.37,0.42){\pscircle[linecolor=blue,fillcolor=white,fillstyle=solid](0,0){2pt}}
{\psset{arrowsize=3pt 2}
\psline[linecolor=blue]{->}(-0.280,-0.66)(-0.305,-0.67)
\psline[linecolor=blue]{->}(-0.86,-1.732)(-0.84,-1.74)
\psline[linecolor=blue]{->}(-0.534,-0.071)(-0.55,-0.08)
}
\setlength{\fboxsep}{2pt}
\rput(0.2;105){\colorbox{white}{\textcolor{blue}{$S$}}}
\rput(-1.4,-1){\colorbox{white}{\textcolor{blue}{$D$}}}
\rput(1,-0.35){\colorbox{white}{\textcolor{blue}{$S^2$}}}
\KhparaEnd
\end{pspicture}
\section{\textbackslash PairingTrefoilLoopINTRO}
\begin{pspicture}(-2,-2)(2,2)
\KhparaStartLabels
\psbezier[linecolor=red](0.5,1)(0.5,1.7)(1.5,1.7)(1.5,1)
\psbezier[linecolor=red](0.4,-1)(0.4,-1.8)(1.6,-1.8)(1.6,-1)
\psbezier[linecolor=red](0.4,-1)(0.4,0)(1.5,0)(1.5,1)
\psbezier[linecolor=red](0.5,1)(0.5,0)(1.6,0)(1.6,-1)
{\psset{arrowsize=3pt 2}
\psline[linecolor=red]{->}(0.79,0.258)(0.78,0.27)
\psline[linecolor=red]{->}(0.77,0.282)(0.78,0.27)
}
\rput(1,-1.55){\colorbox{white}{\textcolor{red}{$S^2$}}}
\rput(1,1.5){\colorbox{white}{\textcolor{red}{$D$}}}
\rput(1.22,0.27){\pscircle[linecolor=red,fillcolor=white,fillstyle=solid](0,0){2pt}}
\rput(0.78,0.27){\pscircle[linecolor=red,fillcolor=white,fillstyle=solid](0,0){2pt}}
\KhparaEnd
\end{pspicture}
\section{\textbackslash PairingBothINTRO}
\begin{pspicture}(-2,-2)(2,2)
\KhparaStartLabels
\rput(0.59,-0.39){%
\pscircle[linecolor=lightgray,fillcolor=lightgray,fillstyle=solid](0,0){10pt}%
}%
\psbezier[linecolor=red](0.5,1)(0.5,1.7)(1.5,1.7)(1.5,1)
\psbezier[linecolor=red](0.4,-1)(0.4,-1.8)(1.6,-1.8)(1.6,-1)
\psbezier[linecolor=red](0.4,-1)(0.4,0)(1.5,0)(1.5,1)
\psbezier[linecolor=red](0.5,1)(0.5,0)(1.6,0)(1.6,-1)
\psecurve[style=curveblue]%
(-2,-2)(-1,-1)%
(1.7,-0.6)(1.75,-1.5)%
(0,-1.9)%
(-1.3,-1.4)(-1.3,-0.8)%
(1,1)(1.2,2)
\rput(1.22,0.27){\pscircle[linecolor=red,fillcolor=white,fillstyle=solid](0,0){2pt}}
\rput(-0.305,-0.67){\pscircle[linecolor=blue,fillcolor=blue,fillstyle=solid](0,0){2pt}}
\pscircle[linecolor=black,fillcolor=black,fillstyle=solid](1.46,-0.46){2pt}
\pscircle[linecolor=black,fillcolor=black,fillstyle=solid](0.59,0.56){2pt}
\rput(0.59,-0.39){%
\pscircle[linecolor=black,fillcolor=black,fillstyle=solid](0,0){2pt}%
\rput{0}(-0.03,0.03){\psline[arrows=->,arrowsize=1.5pt 1.5, linearc=0.15](0.15,0.27)(-0.05,0.05)(-0.3,0)}%
}%
\KhparaEnd
\end{pspicture}
}
\section{\textbackslash GeographyCovering}
{
\psset{unit=0.35}
\begin{pspicture}(-18,-4)(22,4)
\multido{\i=-16+8}{5}{
\rput(\i,0){%
\psdot[dotstyle=asterisk,dotsize=9pt,dotangle=15](0,2)
\pscircle[style=sp](4,2){2pt}%
\pscircle[style=sp](0,-2){2pt}%
\pscircle[style=sp](4,-2){2pt}%
}
}
\rput(2,2){%
\psline[style=curveblue,linearc=5]%
(-16,0)(-14,-0.75)(-2,-0.75)(2,0.75)(14,0.75)(16,0)%
}%
\rput(-6,3){\textrm{$\overbrace{\hspace{4.6cm}}^{\text{2n punctures}}$}}
\rput(10,1){\textrm{$\underbrace{\hspace{4.6cm}}_{\text{2n punctures}}$}}
\rput(2,-2){%
\psline[style=curveblue,linearc=5]%
(-12,0)(-10,-0.75)(-2,-0.75)(2,0.75)(10,0.75)(12,0)%
}%
\rput(-4,-1){\textrm{$\overbrace{\hspace{3.2cm}}^{\text{n punctures}}$}}
\rput(8,-3){\textrm{$\underbrace{\hspace{3.2cm}}_{\text{n punctures}}$}}
\rput(18,3){\textcolor{blue}{\footnotesize$\tilde{\s}_{2n}(0)$}}
\rput(14,-1){\textcolor{blue}{\footnotesize$\tilde{\r}_{n}(0)$}}
\end{pspicture}
}
\section{\textbackslash Blanksphere}
{\psset{unit=0.45}
\begin{pspicture}(-5,-5)(5,5)
\psdot[dotstyle=asterisk,dotsize=9pt,dotangle=15](-2,2)
\pscircle[style=sp](2,2){2pt}%
\pscircle[style=sp](-2,-2){2pt}%
\pscircle[style=sp](2,-2){2pt}%
\psellipticarc[linecolor=lightgray, linestyle=dotted,dotsep=2pt](0,0)(5,1){0}{180}
\pscircle(0,0){5}
\psellipticarc[linecolor=lightgray](0,0)(5,1){180}{360}
\end{pspicture}
}
{\psset{unit=1.1}
\section{\textbackslash slopeZeroNbh}
\begin{pspicture}(-1.7,-1.7)(1.7,1.7)
\psline*[linearc=0.7,linecolor=lightgreen](0,-0.3)(1.7,-0.3)(1.7,-1.7)(-1.7,-1.7)(-1.7,-0.3)(0,-0.3)
\KhparaStartWOshade
\KhparaEnd
\end{pspicture}
\section{ }
\begin{pspicture}(-1.7,-1.7)(1.7,1.7)
\psline*[linearc=0.7,linecolor=lightgreen](0,-0.3)(1.7,-0.3)(1.7,-1.7)(-1.7,-1.7)(-1.7,-0.3)(0,-0.3)
\psline*[linearc=0.7,linecolor=lightgreen](0,0.3)(1.7,0.3)(1.7,1.7)(-1.7,1.7)(-1.7,0.3)(0,0.3)
\KhparaStartWOshade
\KhparaEnd
\end{pspicture}
\section{\textbackslash KissingHearts}
\begin{pspicture}(-1.7,-1.7)(1.7,1.7)
\KhparaStartWOshade
\rput{90}(0,0){\psscalebox{-1 1}{
\psset{style=curveblue}
\psbezier(0.5,1)(0.5,1.55)(1.2,1.35)(1.2,1)
\psbezier(0.8,1)(0.8,1.35)(1.5,1.55)(1.5,1)
\psbezier(0.8,1)(0.8,0.7)(1.2,0.7)(1.2,1)
\psbezier(0.4,-1)(0.4,0)(1.5,0)(1.5,1)
\psbezier(0.5,1)(0.5,0)(1.6,0)(1.6,-1)
\psbezier(0.4,-1)(0.4,-1.75)(1.3,-1.5)(1.3,-1)
\psbezier(0.7,-1)(0.7,-1.5)(1.6,-1.75)(1.6,-1)
\psbezier(0.7,-1)(0.7,-0.6)(1.3,-0.6)(1.3,-1)
}}
\KhparaEnd
\end{pspicture}
}
{ \psset{unit=0.2,linewidth=\stringwidth}
\section{\textbackslash tanglepairing}
\begin{pspicture}[showgrid=false](-11,-7)(11,7)
\rput{-45}(0,0){
\pscircle[linecolor=lightgray](4,4){3}
\pscircle[linecolor=lightgray](-4,-4){3}
\psline[linecolor=white,linewidth=\stringwhite](1,-4)(-2,-4)
\psline[linewidth=\stringwidth,linecolor=gray](1,-4)(-2,-4)
\psline[linecolor=white,linewidth=\stringwhite](-4,1)(-4,-2)
\psline[linewidth=\stringwidth,linecolor=gray](-4,1)(-4,-2)
\psline[linecolor=white,linewidth=\stringwhite](-6,-4)(-9,-4)
\psline[linewidth=\stringwidth,linecolor=gray](-6,-4)(-9,-4)
\psline[linecolor=white,linewidth=\stringwhite](-4,-6)(-4,-9)
\psline[linewidth=\stringwidth,linecolor=gray](-4,-6)(-4,-9)
\psline[linecolor=white,linewidth=\stringwhite](-1,4)(2,4)
\psline[linewidth=\stringwidth,linecolor=gray](-1,4)(2,4)
\psline[linecolor=white,linewidth=\stringwhite](4,-1)(4,2)
\psline[linewidth=\stringwidth,linecolor=gray](4,-1)(4,2)
\psline[linecolor=white,linewidth=\stringwhite](6,4)(9,4)
\psline[linewidth=\stringwidth,linecolor=gray](6,4)(9,4)
\psline[linecolor=white,linewidth=\stringwhite](4,6)(4,9)
\psline[linewidth=\stringwidth,linecolor=gray](4,6)(4,9)
\pscircle[linestyle=dotted,linewidth=\stringwidth](4,4){5}
\pscircle[linestyle=dotted,linewidth=\stringwidth](-4,-4){5}
\psecurve[linewidth=\stringwidth]{C-C}(8,-3)(-1,4)(-10,-3)(-9,-4)(-8,-3)
\psecurve[linewidth=\stringwidth]{C-C}(-3,8)(4,-1)(-3,-10)(-4,-9)(-3,-8)
\psecurve[linewidth=\stringwhite,linecolor=white](-8,3)(1,-4)(10,3)(9,4)(8,3)
\psecurve[linewidth=\stringwhite,linecolor=white](3,-8)(-4,1)(3,10)(4,9)(3,8)
\psecurve[linewidth=\stringwidth]{C-C}(-8,3)(1,-4)(10,3)(9,4)(8,3)
\psecurve[linewidth=\stringwidth]{C-C}(3,-8)(-4,1)(3,10)(4,9)(3,8)
\rput{45}(-4,-4){$Q_{-\nicefrac{p}{q}}$}
\rput{45}(4,4){$T$}
}
\end{pspicture}
\section{\textbackslash ratixo}
\begin{pspicture}[showgrid=false](0,-7)(11,7)
\rput{-45}(0,0){
{\psset{linewidth=\stringwidth}
\psarc(-1,-1){5}{0}{90}
\psarc(9,9){5}{180}{270}
}
\pscircle[linestyle=dotted](4,4){5}
}
\end{pspicture}
\section{\textbackslash ratoxi}
\begin{pspicture}[showgrid=false](0,-7)(11,7)
\rput{-45}(0,0){
{\psset{linewidth=\stringwidth}
\psarc(-1,9){5}{-90}{0}
\psarc(9,-1){5}{90}{180}
}
\pscircle[linestyle=dotted](4,4){5}
}
\end{pspicture}
\section{\textbackslash ratixi}
\begin{pspicture}[showgrid=false](0,-7)(11,7)
\rput{-45}(0,0){
\psline(-1,4)(9,4)
\psline[linewidth=\stringwhite,linecolor=white](4,-1)(4,9)
\psline(4,-1)(4,9)
\pscircle[linestyle=dotted](4,4){5}
}
\end{pspicture}
\section{\textbackslash ratixii}
\begin{pspicture}[showgrid=false](0,-7)(11,7)
\rput{-45}(0,0){
\psecurve(14,2)(9,4)(3,3)(4,9)(2,14)
\psecurve[linewidth=\stringwhite,linecolor=white](6,-6)(4,-1)(5,5)(-1,4)(-6,6)
\psecurve(6,-6)(4,-1)(5,5)(-1,4)(-6,6)
\psecurve[linewidth=\stringwhite,linecolor=white](9,4)(3,3)(4,9)(2,14)
\psecurve(9,4)(3,3)(4,9)(2,14)
\pscircle[linestyle=dotted](4,4){5}
}
\end{pspicture}
\section{\textbackslash ratixiii}
\begin{pspicture}[showgrid=false](0,-7)(11,7)
\rput{-45}(0,0){
\psecurve(14,2)(9,4)(4,2)(4,6)(-1,4)(-6,6)
\psecurve[linewidth=\stringwhite,linecolor=white](6,-6)(4,-1)(6,4)(2,4)(4,9)(2,14)
\psecurve(6,-6)(4,-1)(6,4)(2,4)(4,9)(2,14)
\psecurve[linewidth=\stringwhite,linecolor=white](9,4)(4,2)(4,6)(-1,4)
\psecurve(9,4)(4,2)(4,6)(-1,4)
\pscircle[linestyle=dotted](4,4){5}
}
\end{pspicture}
\section{\textbackslash crossingcircle}
\begin{pspicture}[showgrid=false](0,-7)(11,7)
\rput{-45}(0,0){%
\rput{45}(4,4){%
\psellipticarc[linecolor=gray](0,0)(4,2){0}{180}
}
\psline[linewidth=\stringwhite,linecolor=white](-1,4)(9,4)
\psline{->}(-1,4)(9,4)
\psline[linewidth=\stringwhite,linecolor=white](4,-1)(4,9)
\psline{->}(4,-1)(4,9)
\rput{45}(4,4){%
\psellipticarc[linewidth=\stringwhite,linecolor=white](0,0)(4.35,2.35){180}{360}
\psellipticarc[linecolor=gray](0,0)(4,2){170}{370}
\rput(0,-3){$c$}
}
\pscircle[linestyle=dotted](4,4){5}
}
\end{pspicture}
}
{
\psset{unit=0.25}
\section{\textbackslash AGCCCnoupperbound}
\begin{pspicture}[showgrid=false](-2,-9)(30,9)
\psline*[linecolor=lightgray](4,-6)(30,-6)(30,6)(4,6)
\psline[style=para](4,-6)(30,-6)
\psline[style=para](4,6)(30,6)
\psline*[linecolor=lightgray](4,-6)(-2,-6)(-2,-9)(4,-9)
\psline[style=para](4,-6)(4,-9)
\psline*[linecolor=lightgray](4,6)(-2,6)(-2,9)(4,9)
\psline[style=para](4,6)(4,9)
\psline[style=sp](-2,-6)(4,-6)(4,6)(-2,6)
\psdot[dotstyle=asterisk,dotsize=9pt,dotangle=15](0,0)
\psline[style=curveblue](17,-8)(27,8)
\psline[style=curvered](26,0)(-1,-8)
\pscircle[style=sp](26,0){3pt}%
\rput(18.333,-6){%
\pscircle[linecolor=blue,fillcolor=white,fillstyle=solid](0,0){2pt}
\uput{0.75}[-70]{0}(0,0){\footnotesize\(\QGrad{x}\)}
}
\rput(5.6666,-6){%
\pscircle[linecolor=red,fillcolor=white,fillstyle=solid](0,0){2pt}
\uput{0.5}[-90]{0}(0,0){\footnotesize\(\QGrad{-4n}\)}
}
\rput(21.1,-1.5){%
\pscircle[linecolor=black,fillcolor=black,fillstyle=solid](0,0){2pt}
\rput{59}(0,0){\psline[arrows=->,arrowsize=1.5pt 1.5, linearc=0.2](-4,3)(-1.2,0.5)(-3.5,0.5)}
\uput{0.5}[-50]{0}(0,0){\footnotesize\(\QGrad{x+4n}\)}
}
\rput[b](12,-3.5){\textcolor{red}{\(\a_{\nicefrac{1}{2n}}\)}}
\rput[l](22,2.5){\textcolor{blue}{\(\gamma\)}}
\end{pspicture}
}
\end{document}
\chapter*{KhCurves-pics-Inline: inline diagrams}
{
\psset{unit=0.3}
\section{\textbackslash DotC}
\begin{pspicture}(-0.3,-0.3)(0.3,0.3)
\pscircle[fillcolor=white,fillstyle=solid](0,0){0.25}
\end{pspicture}
\section{\textbackslash DotB}
\begin{pspicture}(-0.3,-0.3)(0.3,0.3)
\pscircle[fillcolor=black,fillstyle=solid](0,0){0.25}
\end{pspicture}
\section{\textbackslash DotCblue}
\begin{pspicture}(-0.3,-0.3)(0.3,0.3)
\pscircle[linecolor=blue,fillcolor=white,fillstyle=solid](0,0){0.25}
\end{pspicture}
\section{\textbackslash DotBblue}
\begin{pspicture}(-0.3,-0.3)(0.3,0.3)
\pscircle[linecolor=blue,fillcolor=blue,fillstyle=solid](0,0){0.25}
\end{pspicture}
\section{\textbackslash DotCred}
\begin{pspicture}(-0.3,-0.3)(0.3,0.3)
\pscircle[linecolor=red,fillcolor=white,fillstyle=solid](0,0){0.25}
\end{pspicture}
\section{\textbackslash DotBred}
\begin{pspicture}(-0.3,-0.3)(0.3,0.3)
\pscircle[linecolor=red,fillcolor=red,fillstyle=solid](0,0){0.25}
\end{pspicture}
}
{
\psset{unit=0.25\baselineskip
\section{\textbackslash Circle}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\pscircle(0,0){1.9}
\end{pspicture}
\section{\textbackslash resolution}
\begin{pspicture}(-2,-2)(2,2)
\psset{unit=1.75}
\psline[linecolor=red](1,-1)(-1,1)
\psline[linecolor=blue](-1,-1)(1,1)
\psdot(0,0)
\rput{-45}(0,0){
\psline[linearc=0.25,arrowsize= 1pt 2]{<-}(0.3,1.3)(0.3,0.3)(1.3,0.3)
}
\rput{135}(0,0){
\psline[linearc=0.25,arrowsize= 1pt 2]{<-}(0.3,1.3)(0.3,0.3)(1.3,0.3)
}
\end{pspicture}
\section{\textbackslash Li}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psLi
\end{pspicture}
\section{\textbackslash Lo}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psLo
\end{pspicture}
\section{\textbackslash LiRed}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\psset{linecolor=red}
\pscircle[linestyle=dotted,dotsep=1pt](0,0){2}
\rput{}(0,2){\textcolor{red}{$\ast$}}
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\end{pspicture}
\section{\textbackslash LoRed}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\psset{linecolor=red}
\pscircle[linestyle=dotted,dotsep=1pt](0,0){2}
\rput(0,2){\textcolor{red}{$\ast$}}
\pscustom{
\psline(0,2)(0,1)
\psarc(1,1){1}{-180}{-90}
\psline(1,0)(2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarc(-1,-1){1}{0}{90}
\psline(-1,0)(-2,0)
}%
}
\end{pspicture}
\section{\textbackslash LiBlue}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\psset{linecolor=blue}
\pscircle[linestyle=dotted,dotsep=1pt](0,0){2}
\rput{}(0,2){\textcolor{blue}{$\ast$}}
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\end{pspicture}
\section{\textbackslash LoBlue}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\psset{linecolor=blue}
\pscircle[linestyle=dotted,dotsep=1pt](0,0){2}
\rput(0,2){\textcolor{blue}{$\ast$}}
\pscustom{
\psline(0,2)(0,1)
\psarc(1,1){1}{-180}{-90}
\psline(1,0)(2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarc(-1,-1){1}{0}{90}
\psline(-1,0)(-2,0)
}%
}
\end{pspicture}
\section{\textbackslash Ni}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\pscircle[linestyle=dotted,dotsep=1pt](0,0){2}
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\end{pspicture}
\section{\textbackslash No}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){-45}{
\pscircle[linestyle=dotted,dotsep=1pt](0,0){2}
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\end{pspicture}
\section{\textbackslash Lil}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\rput(0,2){$\ast$}
\rput{45}(1,-1){
\psline[linecolor=black](1;90)(1.5;90)
}
\rput{45}(-1,1){
\psline[linecolor=black](1;-90)(1.5;-90)
}
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\end{pspicture}
\section{\textbackslash Lol}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\rput(0,2){$\ast$}
\rput{135}(1,1){
\psline[linecolor=black](1;90)(1.5;90)
}
\rput{-45}(-1,-1){
\psline[linecolor=black](1;90)(1.5;90)
}
\pscustom{
\psline(0,2)(0,1)
\psarc(1,1){1}{-180}{-90}
\psline(1,0)(2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarc(-1,-1){1}{0}{90}
\psline(-1,0)(-2,0)
}%
}
\end{pspicture}
\section{\textbackslash Nil}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\rput{45}(1,-1){
\psline[linecolor=black](1;90)(1.5;90)
}
\rput{45}(-1,1){
\psline[linecolor=black](1;-90)(1.5;-90)
}
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\end{pspicture}
\section{\textbackslash Nol}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){-45}{
\rput{45}(1,-1){
\psline[linecolor=black](1;90)(1.5;90)
}
\rput{45}(-1,1){
\psline[linecolor=black](1;-90)(1.5;-90)
}
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\end{pspicture}
\section{\textbackslash CrossingL}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\psline(2,0)(-2,0)
\psline(0,2)(0,0.5)
\psline(0,-2)(0,-0.5)
}
\end{pspicture}
\section{\textbackslash CrossingR}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){-45}{
\psline(2,0)(-2,0)
\psline(0,2)(0,0.5)
\psline(0,-2)(0,-0.5)
}
\end{pspicture}
\section{\textbackslash CrossingLDot}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\pscircle[linestyle=dotted,dotsep=1pt](0,0){2}
\psline(2,0)(-2,0)
\psline(0,2)(0,0.5)
\psline(0,-2)(0,-0.5)
}
\psrotate(0,0){45}{\rput(0,2){$\ast$}}
\end{pspicture}
\section{\textbackslash CrossingRDot}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){-45}{
\pscircle[linestyle=dotted,dotsep=1pt](0,0){2}
\psline(2,0)(-2,0)
\psline(0,2)(0,0.5)
\psline(0,-2)(0,-0.5)
}
\psrotate(0,0){45}{\rput(0,2){$\ast$}}
\end{pspicture}
\section{\textbackslash CrossingPos}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\psline{<-}(2.4,0)(-2,0)
\psline{<-}(0,2.4)(0,0.5)
\psline(0,-2)(0,-0.5)
}
\end{pspicture}
\section{\textbackslash CrossingNeg}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){-45}{
\psline{->}(2,0)(-2.4,0)
\psline{<-}(0,2.4)(0,0.5)
\psline(0,-2)(0,-0.5)
}
\end{pspicture}
\section{\textbackslash CrossingPosMarkedi}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\psline{<-}(2.3,0)(-2,0)
\psline{<-}(0,2.2)(0,0.5)
\psline(0,-2)(0,-0.5)
}
\psrotate(0,0){45}{\rput(0,2){$\ast$}}
\end{pspicture}
\section{\textbackslash CrossingNegMarkedi}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){-45}{
\psline{->}(2,0)(-2.2,0)
\psline{<-}(0,2.3)(0,0.5)
\psline(0,-2)(0,-0.5)
}
\psrotate(0,0){45}{\rput(0,2){$\ast$}}
\end{pspicture}
\section{\textbackslash CrossingPosMarkedii}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\psline{<-}(2.4,0)(-2,0)
\psline{<-}(0,2.4)(0,0.5)
\psline(0,-2)(0,-0.5)
}
\psrotate(0,0){135}{\rput(0,2){$\ast$}}
\end{pspicture}
\section{\textbackslash CrossingNegMarkedii}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){-45}{
\psline{->}(2,0)(-2.4,0)
\psline{<-}(0,2.4)(0,0.5)
\psline(0,-2)(0,-0.5)
}
\psrotate(0,0){135}{\rput(0,2){$\ast$}}
\end{pspicture}
\section{\textbackslash CrossingPosMarkediii}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\psline{<-}(2.4,0)(-2,0)
\psline{<-}(0,2.4)(0,0.5)
\psline(0,-2)(0,-0.5)
}
\psrotate(0,0){-135}{\rput(0,2){$\ast$}}
\end{pspicture}
\section{\textbackslash CrossingNegMarkediii}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){-45}{
\psline{->}(2,0)(-2.4,0)
\psline{<-}(0,2.4)(0,0.5)
\psline(0,-2)(0,-0.5)
}
\psrotate(0,0){-135}{\rput(0,2){$\ast$}}
\end{pspicture}
\section{\textbackslash CrossingPosMarkediv}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\psline{<-}(2.2,0)(-2,0)
\psline{<-}(0,2.3)(0,0.5)
\psline(0,-2)(0,-0.5)
}
\psrotate(0,0){-45}{\rput(0,2){$\ast$}}
\end{pspicture}
\section{\textbackslash CrossingNegMarkediv}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){-45}{
\psline{->}(2,0)(-2.3,0)
\psline{<-}(0,2.2)(0,0.5)
\psline(0,-2)(0,-0.5)
}
\psrotate(0,0){-45}{\rput(0,2){$\ast$}}
\end{pspicture}
\section{\textbackslash LiO}
\begin{pspicture}[showgrid=false](-3.5,-2)(3.5,2)
\psrotate(0,0){45}{
\pscustom{
\psline(-1,3)(-1,2)
\psarcn(-2,2){1}{0}{-90}
\psline(-2,1)(-3,1)
}%
\pscircle(0,0){1}
\pscustom{
\psline(1,-3)(1,-2)
\psarcn(2,-2){1}{180}{90}
\psline(2,-1)(3,-1)
}%
}
\psrotate(0,0){45}{\rput(-1,3){$\ast$}}
\end{pspicture}
\section{\textbackslash LiOl}
\begin{pspicture}[showgrid=false](-3.5,-2)(3.5,2)
\psrotate(0,0){45}{
\rput{45}(0,0){
\psline[linecolor=black](1;90)(1.5;90)
}
\rput{45}(-2,2){
\psline[linecolor=black](1;-90)(1.5;-90)
}
\pscustom{
\psline(-1,3)(-1,2)
\psarcn(-2,2){1}{0}{-90}
\psline(-2,1)(-3,1)
}%
\pscircle(0,0){1}
\pscustom{
\psline(1,-3)(1,-2)
\psarcn(2,-2){1}{180}{90}
\psline(2,-1)(3,-1)
}%
}
\psrotate(0,0){45}{\rput(-1,3){$\ast$}}
\end{pspicture}
\section{\textbackslash LilO}
\begin{pspicture}[showgrid=false](-3.5,-2)(3.5,2)
\psrotate(0,0){45}{
\rput(0,2){
\psline[linecolor=black](1;-135)(1.5;-135)
}
\rput(-2,0){
\psline[linecolor=black](1;45)(1.5;45)
}
\pscustom{
\psline(-1,3)(-1,2)
\psarc(0,2){1}{180}{270}
\psarcn(0,0){1}{90}{-180}
\psarc(-2,0){1}{0}{90}
\psline(-2,1)(-3,1)
}%
\pscustom{
\psline(1,-3)(1,-2)
\psarcn(2,-2){1}{180}{90}
\psline(2,-1)(3,-1)
}%
}
\psrotate(0,0){45}{\rput(-1,3){$\ast$}}
\end{pspicture}
\section{\textbackslash LiOr}
\begin{pspicture}[showgrid=false](-3.5,-2)(3.5,2)
\psrotate(0,0){45}{
\rput{45}(2,-2){
\psline[linecolor=black](1;90)(1.5;90)
}
\rput{45}(0,0){
\psline[linecolor=black](1;-90)(1.5;-90)
}
\pscustom{
\psline(-1,3)(-1,2)
\psarcn(-2,2){1}{0}{-90}
\psline(-2,1)(-3,1)
}%
\pscircle(0,0){1}
\pscustom{
\psline(1,-3)(1,-2)
\psarcn(2,-2){1}{180}{90}
\psline(2,-1)(3,-1)
}%
}
\psrotate(0,0){45}{\rput(-1,3){$\ast$}}
\end{pspicture}
\section{\textbackslash LirO}
\begin{pspicture}[showgrid=false](-3.5,-2)(3.5,2)
\psrotate(0,0){45}{
\rput(0,-2){
\psline[linecolor=black](1;45)(1.5;45)
}
\rput(2,0){
\psline[linecolor=black](1;-135)(1.5;-135)
}
\pscustom{
\psline(1,-3)(1,-2)
\psarc(0,-2){1}{0}{90}
\psarcn(0,0){1}{-90}{0}
\psarc(2,0){1}{180}{270}
\psline(2,-1)(3,-1)
}%
\pscustom{
\psline(-1,3)(-1,2)
\psarcn(-2,2){1}{0}{-90}
\psline(-2,1)(-3,1)
}%
}
\psrotate(0,0){45}{\rput(-1,3){$\ast$}}
\end{pspicture}
\section{\textbackslash NiO}
\begin{pspicture}[showgrid=false](-3.5,-2)(3.5,2)
\psrotate(0,0){45}{
\pscustom{
\psline(-1,3)(-1,2)
\psarcn(-2,2){1}{0}{-90}
\psline(-2,1)(-3,1)
}%
\pscircle(0,0){1}
\pscustom{
\psline(1,-3)(1,-2)
\psarcn(2,-2){1}{180}{90}
\psline(2,-1)(3,-1)
}%
}
\end{pspicture}
\section{\textbackslash NiOl}
\begin{pspicture}[showgrid=false](-3.5,-2)(3.5,2)
\psrotate(0,0){45}{
\rput{45}(0,0){
\psline[linecolor=black](1;90)(1.5;90)
}
\rput{45}(-2,2){
\psline[linecolor=black](1;-90)(1.5;-90)
}
\pscustom{
\psline(-1,3)(-1,2)
\psarcn(-2,2){1}{0}{-90}
\psline(-2,1)(-3,1)
}%
\pscircle(0,0){1}
\pscustom{
\psline(1,-3)(1,-2)
\psarcn(2,-2){1}{180}{90}
\psline(2,-1)(3,-1)
}%
}
\end{pspicture}
\section{\textbackslash NilO}
\begin{pspicture}[showgrid=false](-3.5,-2)(3.5,2)
\psrotate(0,0){45}{
\rput(0,2){
\psline[linecolor=black](1;-135)(1.5;-135)
}
\rput(-2,0){
\psline[linecolor=black](1;45)(1.5;45)
}
\pscustom{
\psline(-1,3)(-1,2)
\psarc(0,2){1}{180}{270}
\psarcn(0,0){1}{90}{-180}
\psarc(-2,0){1}{0}{90}
\psline(-2,1)(-3,1)
}%
\pscustom{
\psline(1,-3)(1,-2)
\psarcn(2,-2){1}{180}{90}
\psline(2,-1)(3,-1)
}%
}
\end{pspicture}
\section{\textbackslash NiOr}
\begin{pspicture}[showgrid=false](-3.5,-2)(3.5,2)
\psrotate(0,0){45}{
\rput{45}(2,-2){
\psline[linecolor=black](1;90)(1.5;90)
}
\rput{45}(0,0){
\psline[linecolor=black](1;-90)(1.5;-90)
}
\pscustom{
\psline(-1,3)(-1,2)
\psarcn(-2,2){1}{0}{-90}
\psline(-2,1)(-3,1)
}%
\pscircle(0,0){1}
\pscustom{
\psline(1,-3)(1,-2)
\psarcn(2,-2){1}{180}{90}
\psline(2,-1)(3,-1)
}%
}
\end{pspicture}
\section{\textbackslash NirO}
\begin{pspicture}[showgrid=false](-3.5,-2)(3.5,2)
\psrotate(0,0){45}{
\rput(0,-2){
\psline[linecolor=black](1;45)(1.5;45)
}
\rput(2,0){
\psline[linecolor=black](1;-135)(1.5;-135)
}
\pscustom{
\psline(1,-3)(1,-2)
\psarc(0,-2){1}{0}{90}
\psarcn(0,0){1}{-90}{0}
\psarc(2,0){1}{180}{270}
\psline(2,-1)(3,-1)
}%
\pscustom{
\psline(-1,3)(-1,2)
\psarcn(-2,2){1}{0}{-90}
\psline(-2,1)(-3,1)
}%
}
\end{pspicture}
\section{\textbackslash LiDotL}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psdot(-0.5,0)
\psrotate(0,0){45}{
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\psrotate(0,0){45}{\rput(0,2){$\ast$}}
\end{pspicture}
\section{\textbackslash LiDotR}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psdot(0.5,0)
\psrotate(0,0){45}{
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\psrotate(0,0){45}{\rput(0,2){$\ast$}}
\end{pspicture}
\section{\textbackslash LoDotT}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psdot(0,0.5)
\psrotate(0,0){-45}{
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\psrotate(0,0){45}{\rput(0,2){$\ast$}}
\end{pspicture}
\section{\textbackslash LoDotB}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psdot(0,-0.5)
\psrotate(0,0){-45}{
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\psrotate(0,0){45}{\rput(0,2){$\ast$}}
\end{pspicture}
\section{\textbackslash NiDotL}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psdot(-0.5,0)
\psrotate(0,0){45}{
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\end{pspicture}
\section{\textbackslash NiDotR}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psdot(0.5,0)
\psrotate(0,0){45}{
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\end{pspicture}
\section{\textbackslash NoDotT}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psdot(0,0.5)
\psrotate(0,0){-45}{
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\end{pspicture}
\section{\textbackslash NoDotB}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psdot(0,-0.5)
\psrotate(0,0){-45}{
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\end{pspicture}
\section{\textbackslash TwistTwo}
\begin{pspicture}[showgrid=false](-3.41421356237,-2)(3.41421356237,2
\psrotate(0,0){-45}{\rput(-1,-1){%
\psline(0,-2)(0,-0.5)
\psline(2,2.5)(2,4)
\pscustom{
\psline(0,0.5)(0,1)
\psarcn(1,1){1}{180}{90}
\psline(2,2)(4,2)
}%
\pscustom{
\psline(-2,0)(1,0)
\psarc(1,1){1}{-90}{0}
\psline(2,1)(2,1.5)
}%
}}
\end{pspicture}
\section{\textbackslash TwistNegTwo}
\begin{pspicture}[showgrid=false](-3.41421356237,-2)(3.41421356237,2
\psrotate(0,0){45}{\rput(-1,1){%
\psline(0,2)(0,0.5)
\psline(2,-2.5)(2,-4)
\pscustom{
\psline(0,-0.5)(0,-1)
\psarc(1,-1){1}{180}{270}
\psline(2,-2)(4,-2)
}%
\pscustom{
\psline(-2,0)(1,0)
\psarcn(1,-1){1}{90}{0}
\psline(2,-1)(2,-1.5)
}%
}}
\end{pspicture}
\section{\textbackslash TwistThree}
\begin{pspicture}[showgrid=false](-4.82842712475,-2)(4.82842712475,2
\psrotate(0,0){-45}{\rput(0,0){%
\psline(-2,-4)(-2,-2.5)
\psline(2,2.5)(2,4)
\pscustom{
\psline(0,0.5)(0,1)
\psarcn(1,1){1}{180}{90}
\psline(2,2)(4,2)
}%
\pscustom{
\psline(-2,-1)(-2,-1.5)
\psarcn(-1,-1){1}{180}{90}
\psline(-1,0)(1,0)
\psarc(1,1){1}{-90}{0}
\psline(2,1)(2,1.5)
}%
\pscustom{
\psline(-4,-2)(-1,-2)
\psarc(-1,-1){1}{-90}{0}
\psline(0,-1)(0,-0.5)
}%
}}
\end{pspicture}
\section{\textbackslash TwistNegThree}
\begin{pspicture}[showgrid=false](-4.82842712475,-2)(4.82842712475,2
\psrotate(0,0){45}{\rput(0,0){%
\psline(-2,4)(-2,2.5)
\psline(2,-2.5)(2,-4)
\pscustom{
\psline(0,-0.5)(0,-1)
\psarc(1,-1){1}{180}{270}
\psline(2,-2)(4,-2)
}%
\pscustom{
\psline(-2,1)(-2,1.5)
\psarc(-1,1){1}{180}{270}
\psline(-1,0)(1,0)
\psarcn(1,-1){1}{90}{0}
\psline(2,-1)(2,-1.5)
}%
\pscustom{
\psline(-4,2)(-1,2)
\psarcn(-1,1){1}{90}{0}
\psline(0,1)(0,0.5)
}%
}}
\end{pspicture}
\section{\textbackslash TwistTwoDot}
\begin{pspicture}[showgrid=false](-3.41421356237,-2)(3.41421356237,2
\psrotate(0,0){-45}{\rput(-1,-1){%
\psline(0,-2)(0,-0.5)
\psline(2,2.5)(2,4)
\pscustom{
\psline(0,0.5)(0,1)
\psarcn(1,1){1}{180}{90}
\psline(2,2)(4,2)
}%
\pscustom{
\psline(-2,0)(1,0)
\psarc(1,1){1}{-90}{0}
\psline(2,1)(2,1.5)
}%
}}
\psrotate(0,0){45}{\rput(-1,1){\rput(0,2){$\ast$}}}
\end{pspicture}
\section{\textbackslash TwistNegTwoDot}
\begin{pspicture}[showgrid=false](-3.41421356237,-2)(3.41421356237,2
\psrotate(0,0){45}{\rput(-1,1){%
\psline(0,2)(0,0.5)
\psline(2,-2.5)(2,-4)
\pscustom{
\psline(0,-0.5)(0,-1)
\psarc(1,-1){1}{180}{270}
\psline(2,-2)(4,-2)
}%
\pscustom{
\psline(-2,0)(1,0)
\psarcn(1,-1){1}{90}{0}
\psline(2,-1)(2,-1.5)
}%
\rput(0,2){$\ast$}
}}
\end{pspicture}
\section{\textbackslash TwistThreeDot}
\begin{pspicture}[showgrid=false](-4.82842712475,-2)(4.82842712475,2
\psrotate(0,0){-45}{\rput(0,0){%
\psline(-2,-4)(-2,-2.5)
\psline(2,2.5)(2,4)
\pscustom{
\psline(0,0.5)(0,1)
\psarcn(1,1){1}{180}{90}
\psline(2,2)(4,2)
}%
\pscustom{
\psline(-2,-1)(-2,-1.5)
\psarcn(-1,-1){1}{180}{90}
\psline(-1,0)(1,0)
\psarc(1,1){1}{-90}{0}
\psline(2,1)(2,1.5)
}%
\pscustom{
\psline(-4,-2)(-1,-2)
\psarc(-1,-1){1}{-90}{0}
\psline(0,-1)(0,-0.5)
}%
}}
\psrotate(0,0){45}{\rput(0,0){\rput(-2,4){$\ast$}}}
\end{pspicture}
\section{\textbackslash TwistNegThreeDot}
\begin{pspicture}[showgrid=false](-4.82842712475,-2)(4.82842712475,2
\psrotate(0,0){45}{\rput(0,0){%
\psline(-2,4)(-2,2.5)
\psline(2,-2.5)(2,-4)
\pscustom{
\psline(0,-0.5)(0,-1)
\psarc(1,-1){1}{180}{270}
\psline(2,-2)(4,-2)
}%
\pscustom{
\psline(-2,1)(-2,1.5)
\psarc(-1,1){1}{180}{270}
\psline(-1,0)(1,0)
\psarcn(1,-1){1}{90}{0}
\psline(2,-1)(2,-1.5)
}%
\pscustom{
\psline(-4,2)(-1,2)
\psarcn(-1,1){1}{90}{0}
\psline(0,1)(0,0.5)
}%
\rput(-2,4){$\ast$}
}}
\end{pspicture}
\section{\textbackslash CrossingLBasepoint}
\begin{pspicture}[showgrid=false](-4,-2)(4,2)
\psline(4,0)(-4,0)
\psline(0,2)(0,0.5)
\psline(0,-2)(0,-0.5)
\psdot[dotstyle=asterisk,dotsize=6pt,dotangle=0](2,0)
\end{pspicture}
\section{\textbackslash CrossingRBasepoint}
\begin{pspicture}[showgrid=false](-4,-2)(4,2)
\psline(0,2)(0,-2)
\psline(4,0)(0.5,0)
\psline(-4,0)(-0.5,0)
\psdot[dotstyle=asterisk,dotsize=6pt,dotangle=0](-2,0)
\end{pspicture}
\section{\textbackslash LiT}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\rput(0,2){$\ast$}
\rput{45}(1,-1){
\psline[linecolor=black, linewidth=3.5pt](1;90)(1.5;90)
}
\rput{45}(-1,1){
\psline[linecolor=black, linewidth=3.5pt](1;-90)(1.5;-90)
}
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\end{pspicture}
\section{\textbackslash LoT}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\rput(0,2){$\ast$}
\rput{135}(1,1){
\psline[linecolor=black, linewidth=3.5pt](1;90)(1.5;90)
}
\rput{-45}(-1,-1){
\psline[linecolor=black, linewidth=3.5pt](1;90)(1.5;90)
}
\pscustom{
\psline(0,2)(0,1)
\psarc(1,1){1}{-180}{-90}
\psline(1,0)(2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarc(-1,-1){1}{0}{90}
\psline(-1,0)(-2,0)
}%
}
\end{pspicture}
\section{\textbackslash NiT}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\rput{45}(1,-1){
\psline[linecolor=black, linewidth=3.5pt](1;90)(1.5;90)
}
\rput{45}(-1,1){
\psline[linecolor=black, linewidth=3.5pt](1;-90)(1.5;-90)
}
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\end{pspicture}
\section{\textbackslash NoT}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){-45}{
\rput{45}(1,-1){
\psline[linecolor=black, linewidth=3.5pt](1;90)(1.5;90)
}
\rput{45}(-1,1){
\psline[linecolor=black, linewidth=3.5pt](1;-90)(1.5;-90)
}
\pscustom{
\psline(0,2)(0,1)
\psarcn(-1,1){1}{0}{-90}
\psline(-1,0)(-2,0)
}%
\pscustom{
\psline(0,-2)(0,-1)
\psarcn(1,-1){1}{180}{90}
\psline(1,0)(2,0)
}%
}
\end{pspicture}
}
{
\psset{unit=0.3}
\section{\textbackslash singleB}
\begin{pspicture}(-0.3,-0.3)(0.3,0.3)
\psrotate(0,0){-45}{
\psline*(0,0.2)(-0.2,0.2)(-0.2,-0.2)(0.2,-0.2)(0.2,0.2)(0,0.2)
}
\end{pspicture}
\section{\textbackslash singleW}
\begin{pspicture}(-0.3,-0.3)(0.3,0.3)
\psrotate(0,0){-45}{
\psline[fillcolor=white,fillstyle=solid](0,0.2)(-0.2,0.2)(-0.2,-0.2)(0.2,-0.2)(0.2,0.2)(0,0.2)
}
\end{pspicture}
}
\section{\textbackslash ThreeTwistTangle}
{\psset{unit=0.05
\begin{pspicture}[showgrid=false](0,-5.3)(12,5.3)
\rput{-45}(0.29289321881,0)
\psecurve[linecolor=black](14,2)(9,4)(4,2)(4,6)(-1,4)(-6,6)
\psecurve[linewidth=0.5\stringwhite,linecolor=white](6,-6)(4,-1)(6,4)(2,4)(4,9)(2,14)
\psecurve[linecolor=black](6,-6)(4,-1)(6,4)(2,4)(4,9)(2,14)
\psecurve[linewidth=0.5\stringwhite,linecolor=white](9,4)(4,2)(4,6)(-1,4)
\psecurve[linecolor=black](9,4)(4,2)(4,6)(-1,4)
\pscircle[linestyle=dotted,dotsep=1pt](4,4){5.2}
\pst@object{bpt}{0}(-1,4)
}
\end{pspicture}
}
\section{\textbackslash ThreeTwistTangleBlue}
{\psset{unit=0.05,linecolor=blue}
\begin{pspicture}[showgrid=false](0,-6)(12,6)
\rput{-45}(0.29289321881,0)
\psecurve(14,2)(9,4)(4,2)(4,6)(-1,4)(-6,6)
\psecurve[linewidth=0.5\stringwhite,linecolor=white](6,-6)(4,-1)(6,4)(2,4)(4,9)(2,14)
\psecurve(6,-6)(4,-1)(6,4)(2,4)(4,9)(2,14)
\psecurve[linewidth=0.5\stringwhite,linecolor=white](9,4)(4,2)(4,6)(-1,4)
\psecurve(9,4)(4,2)(4,6)(-1,4)
\pscircle[linestyle=dotted,dotsep=1pt](4,4){5.2}
\pst@object{bpt}{0}(-1,4)
}
\end{pspicture}
}
{
\psset{unit=0.3}
\section{\textbackslash DotCarc}
\begin{pspicture}(-0.4,-0.6)(0.4,0.6)
\psline(0,-0.6)(0,0.6)
\pscircle[fillcolor=white,fillstyle=solid](0,0){0.25}
\end{pspicture}
\section{\textbackslash DotBarc}
\begin{pspicture}(-0.6,-0.6)(0.6,0.6)
\psline(-0.6,0)(0.6,0)
\pscircle[fillcolor=black,fillstyle=solid](0,0){0.25}
\end{pspicture}
}
{\psset{unit=0.25\baselineskip
\section{\textbackslash TrivialTwoTangle}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\pscircle[linestyle=dotted,dotsep=1pt](0,0){2}
\psline(-2,0)(2,0)
\end{pspicture}
\section{\textbackslash CircleDot}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\pscircle(0,0){1.9}
\psrotate(0,0){45}{\rput(0,1.9){$\ast$}}
\end{pspicture}
\section{\textbackslash Nio}
\begin{pspicture}[showgrid=false](-2,-2)(2,2)
\psrotate(0,0){45}{
\pscircle[linestyle=dotted,dotsep=1pt](0,0){2}
\psline(0,2)(0,-2)
\psline(-2,0)(2,0)
}
\end{pspicture}
}
\section{\textbackslash }
\section{\textbackslash }
\section{\textbackslash }
\end{document}
\section{\texorpdfstring{Splitness detection for \(\BNr\) and \(\Khr\)}{Split tangle detection for BNr and Khr}}\label{sec:split_tangle_detection}
\begin{figure}[b]
\centering
\begin{subfigure}{0.3\textwidth}
\centering
\(\splittangle\)
\caption{}\label{fig:split:tangle}
\end{subfigure}
\begin{subfigure}{0.3\textwidth}
\centering
\(\slopeZeroNbh\)
\caption{}\label{fig:split:slope}
\end{subfigure}
\begin{subfigure}{0.3\textwidth}
\centering
\(\KissingHearts\)
\caption{}\label{fig:split:kissinghearts}
\end{subfigure}
\caption{(a) A split tangle and (b) the region in \(S^2_{4,\ast}\) supporting the multicurve invariants of any such tangle, according to Theorem~\ref{thm:split_tangle_detection}. Figure~(c) illustrates the generalized figure-eight curves \(\mathbf{e}_n(0)\) for \(n=2\); note that \(\mathbf{e}_1(0)=\mathbf{r}_1(0)\). }
\label{fig:split}
\end{figure}
Recall that a Conway tangle $T\subset B^3$ is \emph{split} if there exists an essential curve in $\partial B^3 \smallsetminus \partial T$ that bounds a disk in $B^3 \smallsetminus T$. If the slope of this curve is 0, we call the tangle \emph{horizontally split}. Equivalently, a tangle is horizontally split if it can be written as a disjoint union of two-ended tangles \(T_1\) and \(T_2\) as in Figure~\ref{fig:split:tangle}.
In this section, we show that \(\Khr\) and \(\BNr\) detect this property:
\begin{theorem
\label{thm:split_tangle_detection}
For any Conway tangle \(T\) the following conditions are equivalent:
\begin{enumerate}
\myitem{(1)} \label{enu:split_detection:T} \(T\) is horizontally split;
\myitem{(2a)} \label{enu:split_detection:shapesBNr}
Up to some bigrading shift, each component of \(\BNr(T)\) is equal to the horizontal arc
\(
\a
\mathrel{\widehat{=}}
[\DotB]
\)
or a generalized figure-eight curve
\(
\mathbf{e}_k(0)
\mathrel{\widehat{=}}
\Big[
\begin{tikzcd}[nodes={inner sep=2pt},column sep=14pt]
\DotB
\arrow{r}{H^k}
&
\DotB
\end{tikzcd}
\Big]
\)
for some \(k>0\);
\myitem{(2b)} \label{enu:split_detection:shapesKhr} Up to some bigrading shift, each component of \(\Khr(T)\) is equal to \(\mathbf{r}_1(0)\);
\myitem{(3a)} \label{enu:split_detection:BNr} Up to homotopy, \(\BNr(T)\) is entirely contained in the shaded region in Figure~\ref{fig:split:slope};
\myitem{(3b)} \label{enu:split_detection:Khr} Up to homotopy, \(\Khr(T)\) is entirely contained in the shaded region in Figure~\ref{fig:split:slope};
\myitem{(4a)} \label{enu:split_detection:nodotsBNr}
The complex \(\DD(T)\mathrel{\widehat{=}}\BNr(T)\) contains no generator $\DotC$;
\myitem{(4b)} \label{enu:split_detection:nodotsKhr}
The complex \(\DD_1(T)\mathrel{\widehat{=}}\Khr(T)\) contains no generator $\DotC$.
\end{enumerate}
\end{theorem}
\begin{remark}
There is an analogous detection result for the Heegaard Floer tangle invariant \(\HFT\) \cite[Theorem~4.1]{LMZ}.
\end{remark}
By the naturality of \(\BNr\) and \(\Khr\) under twisting \cite[Theorem~1.13]{KWZ}, it follows that a tangle is split if and only if \(\Khr(T)\) consists of rational components of the same slope, or equivalently, if and only if \(\BNr(T)\) consists only of generalized figure-eight curves and arcs of the same slope.
\begin{proof}
We start with the implication \(\ref{enu:split_detection:T}\Rightarrow\ref{enu:split_detection:shapesBNr}\).
Bar-Natan associates with the two-ended tangles \(T_1\) and \(T_2\) the invariants \(\KhTl{T_1}\) and \(\KhTl{T_2}\), which are chain complexes over the cobordism category whose objects are crossingless two-ended tangles. Thanks to delooping \cite[Observation~4.18]{KWZ}, we can write these as complexes over the subcategory generated by the trivial tangle~\(\TrivialTwoTangle\). The morphisms in this subcategory can be represented by linear combinations of cobordisms without closed components, ie identity cobordisms with some number of handles attached.
By Bar-Natan's gluing formalism, \(\KhTl{T}\) is then a tensor product of the complexes \(\KhTl{T_1}\) and \(\KhTl{T_2}\). In particular, the objects of \(\KhTl{T}\) are equal to \(\No\) and the differential consists of linear combinations of identity cobordisms with some number of handles attached to one of the components. Since we are working with coefficients in \(\mathbb{F}\), the $(4Tu)$-relation \cite[Definition~4.3]{KWZ} allows us to move all handles on these cobordisms to the component containing the basepoint $\ast$ of \(\Lo\). Attaching a handle to the component with a basepoint corresponds to multiplying by $H$ \cite[Definition~4.10]{KWZ}, so \(\KhTl{T}\) is a chain complex over the graded algebra \(\mathbb{F}[H]\). Therefore, up to homotopy, $\DD(T)$ is a direct sum of complexes of the form
\[
\left[
\begin{tikzcd}[nodes={inner sep=2pt},column sep=14pt]
\DotB
\end{tikzcd}\right]
\qquad\text{or}\qquad
\Big[
\begin{tikzcd}[nodes={inner sep=2pt},column sep=14pt]
\DotB
\arrow{r}{H^k}
&
\DotB
\end{tikzcd}
\Big]
\qquad\text{for some integer \(k>0\),}
\]
as required.
The implication \(\ref{enu:split_detection:shapesBNr}\Rightarrow\ref{enu:split_detection:shapesKhr}\) follows from the definition of \(\Khr(T)\) as the curve corresponding to the mapping cone \(\DD_1(T)\) of the identity map on \(\DD(T)\) multiplied by \(H\).
The equivalences $\ref{enu:split_detection:BNr}\Leftrightarrow \ref{enu:split_detection:nodotsBNr}$ and $\ref{enu:split_detection:Khr}\Leftrightarrow \ref{enu:split_detection:nodotsKhr}$ and the implications $\ref{enu:split_detection:shapesBNr} \Rightarrow \ref{enu:split_detection:BNr}$ and $\ref{enu:split_detection:shapesKhr} \Rightarrow \ref{enu:split_detection:Khr}$
are obvious.
The equivalence $\ref{enu:split_detection:nodotsBNr}\Leftrightarrow \ref{enu:split_detection:nodotsKhr}$ follows from the observation that any complex \(\DD(T)\) corresponding to a curve \(\BNr(T)\) contains a generator \(\DotC\) if and only if the same is true for its mapping cone \(\DD_1(T)\).
The implication \(\ref{enu:split_detection:nodotsBNr}\Rightarrow\ref{enu:split_detection:T}\) remains. This direction relies on a detection result for annular Khovanov homology; this was established by Xie using annular instanton homology.
We know that the complex $\DD(T)$ representing $\BNr(T)$ only contains generators $\DotB$.
This is equivalent to saying that the tangle invariant $\KhTl{T}$, as a homotopy equivalence class of chain complexes over $\Cob_{/l}(\Lo\oplus \Li)$, has a representative containing only generators $\Lo$.
Let \(T_a(\infty)\) be the annular link shown in Figure~\ref{fig:annular}.
Its annular Khovanov homology $\AKh(T_a(\infty);\mathbb{F})$ can be computed from $\KhTl{T}$ via gluing arguments similar to~\cite[Section~5]{BarNatanKhT}. It is concentrated in annular grading zero, because in that computation, every circle has winding number zero around the annulus.
By the universal coefficient theorem, $\AKh(T_a(\infty);\C)$ is also concentrated in annular grading zero. We can now apply Xie's detection result~\cite[Corollary~1.6]{Xie} to deduce that the link $T_a(\infty)$ is contained in a three-ball embedded in the solid torus $S^1\times D^2$.
We conclude with Lemma~\ref{lem:sphere_in_solid_torus} below.
\end{proof}
\begin{figure}[t]
\centering
\labellist
\pinlabel $T$ at 15 45
\pinlabel $E$ at 85 45
\endlabellist
\includegraphics[width=4.5cm]{figures/annular}
\caption{The annular link \(T_a(\infty)\) in $S^1\times D^2$. The shaded disk on the right shows the essential disk \(E\) used in the proof of Lemma~\ref{lem:sphere_in_solid_torus}.}
\label{fig:annular}
\end{figure}
\begin{lemma}\label{lem:sphere_in_solid_torus}
If $T_a(\infty)$ is contained in a three-ball inside a solid torus, then $T$ must be horizontally split.
\end{lemma}
\begin{proof}
Without loss of generality, we may assume that \(T\) has no unlinked closed components.
Let \(S\) be the boundary of the three-ball containing \(T_a(\infty)\) and let \(E\) be the essential disk in $S^1\times D^2$ shown in Figure~\ref{fig:annular}.
Without loss of generality, we may assume that \(S\) intersects \(E\) transversely, so that \(E\cap S\) is a union of circles.
Clearly, \(E\cap S\neq\varnothing\) since \(S\) separates \(E\cap T_a(\infty) \neq\varnothing\) from \(\partial E\).
We now consider those circles as subsets of \(S\). One of them, let us call it \(C\), is innermost, so it bounds a disk \(D_S\) in \(S\) disjoint from any other circles.
The circle \(C\) also bounds a disk \(D_E\) in \(E\).
Suppose \(D_E\) is disjoint from \(T_a(\infty)\). Then \(D_E\cup D_S\) is disjoint from \(T_a(\infty)\) and bounds a three-ball \(B\) in the solid torus. Since \(T\) has no unlinked closed component, \(B\) is disjoint from \(T_a(\infty)\). Therefore, there exists an isotopy of \(S\) that is the identity outside a small neighbourhood of \(B\) and inside this neighbourhood removes the circle \(C\) from \(E\cap S\neq\varnothing\) as well as any other component of \(E\cap S\neq\varnothing\) that lies in \(D_E\).
After repeating this procedure a finite number of times, we may assume that \(D_E\) intersects \(T_a(\infty)\) non-trivially. Since the sphere \(D_S\cup D_E\) intersects \(T_a(\infty)\) in an even number of points and \(D_S\) is disjoint from \(T_a(\infty)\), \(D_E\) must contains both intersection points of \(E\) with \(T_a(\infty)\). So \(E\smallsetminus D_E\) is an annulus and \(D_S\cup (E\smallsetminus D_E)\) is an essential disk which, after isotopy, certifies that \(T\) is a split tangle.
\end{proof}
\section{Introduction}\label{sec:intro}
Here are two classical open conjectures in low dimensional topology:
\begin{csconj}
Given a non-trivial knot \(K\subset S^3\) and \(r,r'\in\operatorname{\mathbb{Q}P}^1\), suppose there exists an orientation-preserving diffeomorphism \(S^3_r(K)\cong S^3_{r'}(K)\). Then \(r=r'\).
\end{csconj}
\begin{ccconj}
Any crossing change that preserves the isotopy class of a knot must occur at a nugatory crossing, meaning the crossing circle (see Figure~\ref{fig:crossingcircle}) bounds an embedded disk in the complement of the knot.
\end{ccconj}
The Cosmetic Surgery Conjecture originates with Gordon \cite[Conjecture 6.1]{Gordon} and the Cosmetic Crossing Conjecture is due to Lin.
Both problems appear in Kirby's Problem List \cite[Problem 1.81 A, Bleiler; Problem 1.58, Lin]{Kirby}.
This paper explores what Khovanov homology can say about these conjectures from the perspective of the multicurve technology developed by Kotelskiy, Watson, and Zibrowius \cite{KWZ}.
\begin{figure}[t]
\centering
\begin{subfigure}{0.17\textwidth}
\centering
\(\crossingcircle\)
\caption{}
\label{fig:crossingcircle}
\end{subfigure}
\begin{subfigure}{0.17\textwidth}
\centering
\(\ratoxi\)
\caption{\(Q_{0}\)}
\label{fig:rat:oxi}
\end{subfigure}
\begin{subfigure}{0.17\textwidth}
\centering
\(\ratixi\)
\caption{\(Q_{1}\)}
\label{fig:rat:ixi}
\end{subfigure}
\begin{subfigure}{0.17\textwidth}
\centering
\(\ratixiii\)
\caption{\(Q_{\nicefrac{1}{3}}\)}
\label{fig:rat:ixiii}
\end{subfigure}
\begin{subfigure}{0.28\textwidth}
\centering
\(\tanglepairing\)
\caption{\(T(\nicefrac{p}{q})\)}
\label{fig:rat:pairing}
\end{subfigure}
\caption{A crossing circle \(c\) (a), some examples of rational tangles~(b--d), and the \(\nicefrac{p}{q}\)-rational filling \(T(\nicefrac{p}{q})\) of a Conway tangle \(T\)~(e)}
\label{fig:rat}
\end{figure}
\subsection{Cosmetic surgeries}
Following Sakuma \cite{Sakuma}, a strongly invertible knot is a pair \((K,h)\), where \(K\) is a knot in \(S^3\) and \(h\) is an orientation-preserving involution of \(S^3\) mapping \(K\) to itself and reversing a choice of orientation of \(K\).
Strongly invertible knots $(K,h)$ and $(K',h')$ are equivalent if there exists an orientation-preserving homeomorphism $f$ on $S^3$ for which $f(K)=K'$ (so that $K$ and $K'$ are equivalent knots) and $h = f^{-1}\circ h'\circ f$.
Any strong inversion \(h\) on a knot \(K\subset S^3\) restricts to the hyperelliptic involution of the torus on the boundary of the knot exterior, and hence can be extended to an involution \(h_r\) on \(S^3_r(K)\) for any slope \(r\in\operatorname{\mathbb{Q}P}^1\).
(The latter statement can be found in \cite{Holzmann}. See also \cite{Auckly} for a nice exposition.)
This extension is unique up to homotopy.
In this article we prove the following equivariant version of the Cosmetic Surgery Conjecture:
\begin{theorem}\label{thm:intro:ecsc}
Given a non-trivial strongly invertible knot \((K,h)\) and \(r,r'\in\operatorname{\mathbb{Q}P}^1\), suppose that there exists an orientation-preserving diffeomorphism \(f\co S^3_r(K)\rightarrow S^3_{r'}(K)\) such that $h_{r'}\circ f= f \circ h_r$. Then \(r=r'\).
\end{theorem}
A strongly invertible knot $(K,h)$ gives rise to a Conway (ie four-ended) tangle $T \subset B^3$ constructed as follows. According to the Smith conjecture, the fixed point set $\operatorname{Fix}(h)$ in $S^3$ is an unknot intersecting $K$ in two points, and thus restricts to a pair of arcs in the knot exterior $X_K=S^3\smallsetminus\nu(K)$. Therefore, taking the quotient produces a Conway tangle $T=\operatorname{Im}(\operatorname{Fix}(h)\cap X_K)$ inside the three ball $B^3=X_K / h$. For an illustration, see \cite[Figure~4]{Watson2012}.
Cosmetic surgeries along a strongly invertible knot $(K,h)$ are closely related to cosmetic tangle fillings of the quotient tangle $T$, due to the Montesinos trick \cite{Montesinos}:
The fixed point set of the involution $h_{\nicefrac{p}{q}}$ on $S^3_{\nicefrac{p}{q}}(K)$ restricted to the surgery solid torus is a pair of arcs, which descends in the quotient of the solid torus to a trivial Conway tangle. Montesinos gives an explicit correspondence between the rational surgery slope and a rational parameterization of the trivial tangle in the quotient. In particular, $S^3_{\nicefrac{p}{q}}(K)$ is the two-fold branched cover of the rational tangle filling $T(\nicefrac{p}{q}) = Q_{\nicefrac{-p}{q}}\cup T$ illustrated in Figure~\ref{fig:rat}, where $Q_{\nicefrac{-p}{q}}$ denotes the rational tangle of slope $\nicefrac{-p}{q}$ and the tangle $T$ is the quotient tangle of $(K,h)$.
Conversely, a cosmetic crossing change on a knot induces a cosmetic surgery on the two-fold branched cover. Hence, the two cosmetic conjectures are related, and the analogue of Theorem~\ref{thm:intro:ecsc} is:
\begin{theorem}\label{thm:intro:cosmeticrational}
Let \(T\) be a Conway tangle in the three-ball with an unknot closure. Suppose \(T(r)\) and \(T(r')\) are isotopic links in the three-sphere for some \(r, r' \in \operatorname{\mathbb{Q}P}^1\). Then \(r = r'\) or \(T\) is rational.
\end{theorem}
Indeed Theorem~\ref{thm:intro:cosmeticrational} is equivalent to Theorem~\ref{thm:intro:ecsc} by taking two-fold branched covers. The condition of having an unknot closure is equivalent to considering surgeries on knots in $S^3$ and non-triviality of the knot in $S^3$ is equivalent to the tangle not being rational.
Gordon and Luecke's solution to the knot complement problem implies that no non-trivial knots admit cosmetic surgeries when one of the slopes is $\infty$ \cite{GordonLuecke}. Boyer and Lines showed that if $\Delta''_K(t)\neq0$, rational surgeries along $K$ are always distinct, where $\Delta_K(t)$ is the Alexander polynomial \cite[Proposition 5.1]{BoyerLines}, and a similar result may be formulated in terms of the Jones polynomial \cite{IchiharaWu}. Note that the obstructions of \cite{BoyerLines, IchiharaWu} do not disqualify strongly invertible knots from admitting cosmetic surgeries because all Alexander polynomials of knots can be realized by strongly invertible knots \cite{Sakai}.
Strong restrictions on cosmetic surgeries can be formulated in terms of Heegaard Floer homology, as shown in work of Wang \cite{WangJiajun}, Ozsv\'ath and Szab\'o \cite{OzSzRational}, and Ni and Wu \cite{NiWu}.
Using immersed curves, Hanselman's work \cite{HanselmanCSC} extends these results; in particular, he shows that cosmetic surgery slopes have to be either \(\pm2\) or \(\nicefrac{\pm1}{n}\).
In an entirely different direction, hyperbolic geometry techniques have been used to bound the lengths of cosmetic filling slopes \cite{FuterPurcellSchleimer}.
Using the strategies above and others, the conjecture has been established for knots of genus one, cables, connected sums and three-braids \cite{WangJiajun, Tao1, Tao2, Varvarezos}.
Absent from this summary are developments utilizing Khovanov homology.
Our main tool for approaching this problem is the multicurve technology for Khovanov homology developed by Kotelskiy, Watson, and Zibrowius \cite{KWZ,KWZthinness}, which we use to prove Theorem~\ref{thm:intro:cosmeticrational}.
This theory assigns an immersed multicurve in the four-punctured sphere to a Conway tangle and a suitable Lagrangian Floer homology of these multicurves computes the Khovanov homology.
\subsection{Cosmetic Crossings}
As illustrated by Theorem~\ref{thm:intro:cosmeticrational}, the immersed curve invariants for Khovanov homology provide a powerful tool for studying the behavior of Khovanov homology under tangle fillings. Note that the Cosmetic Crossing Conjecture can be viewed in terms of comparing the $+1$ and $-1$ tangle fillings of a Conway tangle; hence, the Khovanov tangle invariants provide a natural tool for studying the Cosmetic Crossing Conjecture. In fact, Theorem~\ref{thm:intro:cosmeticrational} immediately implies the Cosmetic Crossing Conjecture for the unknot, originally due to Scharlemann-Thompson \cite{ST}. This conjecture is still open in general, but has been established for knots which are two-bridge \cite{Torisu} or fibered \cite{Kalfagianni, Rogers}, as well as for large classes of knots which are genus one \cite{BFKP, Ito} or alternating \cite{LidmanMoore}. In the second half of this paper, we illustrate the utility of this theory by giving elementary proofs of two other known results about the Cosmetic Crossing Conjecture.
First, we recall that there is a generalization of the Cosmetic Crossing Conjecture. Let $c$ be a crossing circle for a knot $K$ as in Figure~\ref{fig:crossingcircle}.
Then, performing $\nicefrac{-1}{n}$-Dehn surgery on $c$ produces a new knot $K_n$ which corresponds to adding $n$ full right-handed twists at the crossing. The Generalized Cosmetic Crossing Conjecture predicts that if $c$ is a non-nugatory crossing, then $K_n$ is not isotopic to $K_m$ for $n \neq m$.
We first give a new proof of a recent result of Wang on the Generalized Cosmetic Crossing Conjecture, which also used Khovanov homology:
\begin{theorem}[Wang \cite{Wang}] \label{thm:josh}
Let $K$ be a knot obtained by a non-trivial band surgery on a split link $L$. If $K_n$ is obtained by inserting $n \in \mathbb{Z}$ twists into the band, then $K_n$ is not isotopic to $K_m$ for any $n \neq m$.
\end{theorem}
We also prove that the generalized crossing conjecture holds ``asymptotically'':
\begin{theorem}\label{thm:agccc}
Let $K$ be an unoriented knot and $c$ a crossing circle for a non-nugatory crossing. Let $\{K_n\}_{n \in \mathbb{Z}}$ be the associated sequence of knots obtained by inserting twists at $c$. Then there exists an integer $N$ such that $\{K_n\}_{|n|\geq N}$ are pairwise different.
\end{theorem}
\begin{remark}
While we could not find Theorem~\ref{thm:agccc} written explicitly in the literature, it is certainly known to experts using standard techniques from three-manifold topology. Rather than give this alternative proof here in full, we illustrate this by sketching a proof for a suitably generic case. Suppose that $K \cup c$ is a hyperbolic link. Then $K_n$ is obtained by performing $\nicefrac{-1}{n}$-surgery on $c$. Hence, $K_n$ is hyperbolic for all but finitely many $n$ and further the hyperbolic volume of $K_n$ converges to the hyperbolic volume of $K \cup c$, which is strictly greater than that of any $K_n$. The asymptotics of this convergence is described by the work of Neumann-Zagier \cite{NeumannZagier} and precludes having more than finitely many knots in the sequence of fixed volume.
It is interesting that Khovanov homology and the hyperbolic volume establish the same result in this setting.
We note that the behaviour of the Jones polynomial under twisting in relation with hyperbolic geometry has been considered, see \cite{CK} for example.
\end{remark}
Along the way, we establish the following technical result about the Khovanov multicurve invariants, which may be of independent interest:
\begin{theorem}\label{thm:intro:split}
The invariants \(\Khr(T)\) and \(\BNr(T)\) detect if the Conway tangle \(T\) is split.
\end{theorem}
Note that Theorem~\ref{thm:intro:split} has been established for the analogous knot Floer homology multicurve invariant by Lidman, Moore, and Zibrowius \cite{LMZ}.
\subsection*{Outline}
In Section~\ref{sec:background}, we give the requisite background on the immersed curve invariants for tangles. In Section~\ref{sec:split_tangle_detection}, we prove Theorem~\ref{thm:intro:split}. In Section~\ref{sec:ecsc}, we prove Theorem~\ref{thm:intro:cosmeticrational} (and hence Theorem~\ref{thm:intro:ecsc}). In Section~\ref{sec:agccc}, we prove Theorem~\ref{thm:agccc} and Theorem~\ref{thm:josh}.
\subsection*{Acknowledgements}
We thank Joshua Wang for useful and motivating discussions around the Generalized Cosmetic Crossing Conjecture, and in particular for posing the main question that we answer in Section~\ref{sec:agccc}.
\section{The Generalized Cosmetic Crossing Conjecture holds asymptotically}\label{sec:agccc}
Throughout this section we fix a Conway tangle $T$ with connectivity $\No$ and without any closed components. Furthermore, we consider the family of knots \(\{K_n\}_{n\in\mathbb{Z}}\) shown in Figure~\ref{fig:Tpmiin} and defined by $K_n=T(\nicefrac{1}{2n})$ for \(n\in\mathbb{Z}\).
Equivalently, each knot \(K_n\) is the result of a band surgery on a fixed two-component link $T(0)$ and the knots \(K_n\) and \(K_{n+1}\) are obtained from each other by adding a full twist to the band.
This point of view explains the restrictions placed on the tangle~\(T\).
\begin{figure}[b]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\(\Tpiin\)
\caption{}\label{fig:Tpiin}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\centering
\(\Tmiin\)
\caption{}\label{fig:Tmiin}
\end{subfigure}
\caption{The knots $K_n$ for \(n\geq0\) (a) and \(n\leq0\) (b)}
\label{fig:Tpmiin}
\end{figure}
If $T$ is horizontally split, then all $K_n$ are equal to each other. We now restate the two conjectures on cosmetic crossings in terms of Conway tangles:
\begin{conjecture}[Cosmetic Crossing Conjecture]
Suppose \(T\) is not horizontally split.
Then \(K_0\) and \(K_1\) are different as unoriented knots.
\end{conjecture}
\begin{conjecture}[Generalized Cosmetic Crossing Conjecture]
Suppose \(T\) is not horizontally split.
Then the unoriented knots \(\{K_n\}_{n\in \mathbb{Z}}\) are pairwise different.
\end{conjecture}
Recently, Wang showed that these conjectures hold assuming $T(0)$ is a split link \cite{Wang}. While he states these conjectures for oriented knots, we are not aware of any counterexamples to the conjectures as stated above for unoriented knots. Note, however, that a crossing change may result in the mirror of the original knot, as illustrated by the \((3,-3,\pm1)\)-pretzel knots, see the remarks to \cite[Problem~1.58]{Kirby}.
Below we restate and prove Theorem~\ref{thm:agccc}, showing that the Generalized Cosmetic Crossing Conjecture holds ``asymptotically'', that is, for $n$ large enough:
\begin{theorem}\label{thm:main}
Suppose \(T\) is not horizontally split.
Then there exists an integer \(N\) such that the knots \(\{K_n\}_{|n|\geq N}\) are pairwise different as unoriented knots.
\end{theorem}
\begin{lemma}\label{lem:only_zero_slopes_implies_GCCC}
Suppose \(\Khr(T)\) only contains curves of slope 0 of which at least one is special. Then \(K_n\not\cong K_m\) for any \(n\neq m\).
\end{lemma}
\begin{figure}[b]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\(\AGCCCzeroRationals\)
\caption{The canonical generators of
\(\operatorname{HF}(\textcolor{red}{\a_{\nicefrac{1}{2n}}},\textcolor{blue}{P})\) shown in the planar cover of $S^2_{4,\ast}$}\label{fig:agccc:zero:rationals}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\centering
\(\AGCCCzeroSpecials\)
\caption{The canonical generators of
\(\operatorname{HF}(\textcolor{red}{\a_{\nicefrac{1}{2n}}},\textcolor{blue}{\Sigma})\) shown in the planar cover of $S^2_{4,\ast}$}\label{fig:agccc:zero:specials}
\end{subfigure}
\bigskip\\
\begin{subfigure}[b]{0.38\textwidth}
\centering
\centering
\labellist
\pinlabel $S$ at 27 79
\pinlabel $D$ at 51 78
\pinlabel $S$ at 49 61
\pinlabel $D$ at 33 58
\pinlabel $\textcolor{red}{\a_{\nicefrac{1}{2n}}}$ at 120 30
\pinlabel \tiny $\textcolor{violet}{-4n}$ at 86 61
\pinlabel \tiny $\textcolor{violet}{-1}$ at 19 33
\endlabellist
\includegraphics[scale=1.2]{figures/downstairs}
\caption{Schematic intersection picture for the curves $\textcolor{red}{\a_{\nicefrac{1}{2n}}}$, $\textcolor{blue}{P}$, and $\textcolor{blue}{\Sigma}$ in $S^2_{4,\ast}$
}\label{fig:agccc:zero:downstairs}
\end{subfigure}
\begin{subfigure}[b]{0.55\textwidth}
\centering
\(\AGCCCnoupperbound\)
\caption{A generator with unbounded quantum grading of
\(\operatorname{HF}(\textcolor{red}{\a_{\nicefrac{1}{2n}}},\textcolor{blue}{\gamma})\) shown in the planar cover of $S^2_{4,\ast}$}\label{fig:agccc:zero:noupperbound}
\end{subfigure}
\caption{Illustration of intersection points between various curves in the proofs of Lemmas~\ref{lem:only_zero_slopes_implies_GCCC} and~\ref{lem:nonzero_slope_implies_limit_iinfinitty}. Numbers near generators (intersection points) indicate their quantum gradings.
In (c), the dots $\bullet$ indicate the intersection points of the curve $\textcolor{red}{\a_{\nicefrac{1}{2n}}}$ with $\textcolor{blue}{P}$ (contained in the bottom region) and $\textcolor{blue}{\Sigma}$ (contained in the top region).
}
\label{fig:agccc:zero}
\end{figure}
\begin{proof}
Let us write \(\Khr(T)=\textcolor{blue}{\Sigma}\cup \textcolor{blue}{P}\), where \(\textcolor{blue}{\Sigma}\) consists of special components and \(\textcolor{blue}{P}\) of rational components. By assumption \(\textcolor{blue}{\Sigma}\neq\varnothing\). Also \(\textcolor{blue}{P}\neq\varnothing\), since the pairing of
the arc \(\BNr(\Lo)\) with \(\Khr(T)\) computes the Khovanov homology of the link $T(0)$ and as such is non-zero.
Since the connectivity of the tangle \(T\) is \(\No\), the induced orientation of \(K_n\) on the tangle \(T\) and its rational filling are as shown in Figure~\ref{fig:Tpiin}, up to overall orientation reversal.
Define the arc of slope $\nicefrac{1}{2n}$:
\[
\textcolor{red}{\a_{\nicefrac{1}{2n}}}
\mathrel{\widehat{=}}
\begin{cases*}
\left[
\begin{tikzcd}[nodes={inner sep=2pt},column sep=14pt,ampersand replacement=\&, red]
\GGzqh{\DotCred}{}{-4n}{}
\arrow{r}{S}
\&
\GGzqh{\DotBred}{}{1-4n}{}
\arrow{r}{D}
\&
\GGzqz{\DotBred}{}{3-4n}{}
\arrow[dotted]{r}{S^2}
\&
\quad
\arrow[dotted]{r}{}
\&
\GGzqh{\DotBred}{}{-3}{}
\arrow{r}{D}
\&
\GGzqz{\DotBred}{}{-1}{}
\end{tikzcd}\right]
& for \(n>0\)
\\
\left[
\begin{tikzcd}[nodes={inner sep=2pt},column sep=14pt,ampersand replacement=\&, red]
\GGzqz{\DotCred}{}{0}{}
\end{tikzcd}\right]
& for \(n=0\)
\\
\left[
\begin{tikzcd}[nodes={inner sep=2pt},column sep=14pt,ampersand replacement=\&, red]
\GGzqz{\DotBred}{}{1}{}
\arrow{r}{D}
\&
\GGzqz{\DotBred}{}{3}{}
\arrow[dotted]{r}{}
\&
\quad
\arrow[dotted]{r}{S^2}
\&
\GGzqz{\DotBred}{}{-4n-3}{}
\arrow{r}{D}
\&
\GGzqz{\DotBred}{}{-4n-1}{}
\arrow{r}{S}
\&
\GGzqz{\DotCred}{}{-4n}{}
\end{tikzcd}
\right]
& for \(n<0\)
\end{cases*}
\]
By the same calculation as in the proof of Theorem~\ref{thm:intro:cosmeticrational}, Case 2 (using \cite[Example~6.2, Proposition~4.8]{KWZ}), these are the arc invariants of the mirrors of the rational fillings of \(T\).
Then, by the pairing theorem,
\[
\Khr(K_n)
\cong
\operatorname{HF}(\textcolor{red}{\a_{\nicefrac{1}{2n}}},\Khr(T))
=
\operatorname{HF}(\textcolor{red}{\a_{\nicefrac{1}{2n}}},\textcolor{blue}{\Sigma})
\oplus
\operatorname{HF}(\textcolor{red}{\a_{\nicefrac{1}{2n}}},\textcolor{blue}{P})
\]
We now study how both summands behave when varying \(n\); Figure~\ref{fig:agccc:zero:downstairs} depicts the schematic picture of intersections in $S^2_{4,\ast}$. After pulling the multicurves \(\textcolor{blue}{P}\) and \(\textcolor{red}{\a_{\nicefrac{1}{2n}}}\) sufficiently tight, the intersection points generating \(\operatorname{HF}(\textcolor{red}{\a_{\nicefrac{1}{2n}}},\textcolor{blue}{P})\) all sit close to the end of \(\textcolor{red}{\a_{\nicefrac{1}{2n}}}\) corresponding to the generator \(\GGzqz{\DotBred}{}{-1}{0}\) for \(n>0\), \(\GGzqz{\DotCred}{}{0}{0}\) for \(n=0\), and \(\GGzqz{\DotBred}{}{+1}{0}\) for \(n<0\).
The corresponding curve segments are shown in Figure~\ref{fig:agccc:zero:rationals}. We see that the quantum grading of \(\operatorname{HF}(\textcolor{red}{\a_{\nicefrac{1}{2n}}},\textcolor{blue}{P})\) is independent of \(n\); see Section~\ref{subsec:background:bigrading} for how gradings are computed.
We now investigate the pairing of $\textcolor{red}{\a_{\nicefrac{1}{2n}}}$ with special curves. After pulling all multicurves sufficiently tight, the intersection points generating \(\operatorname{HF}(\textcolor{red}{\a_{\nicefrac{1}{2n}}},\textcolor{blue}{\Sigma})\) all sit close to the end of \(\textcolor{red}{\a_{\nicefrac{1}{2n}}}\) corresponding to the generator \(\GGzqz{\DotCred}{}{-4n}{0}\), see Figure~\ref{fig:agccc:zero:specials}. Thus, the shift in quantum grading is as follows:
\[
\operatorname{HF}(\textcolor{red}{\a_{\nicefrac{1}{2n}}},\textcolor{blue}{\Sigma})
\cong
q^{4n}
\operatorname{HF}(\textcolor{red}{\a_{\infty}},\textcolor{blue}{\Sigma})
\cong
q^{4n-4m}
\operatorname{HF}(\textcolor{red}{\a_{\nicefrac{1}{2m}}},\textcolor{blue}{\Sigma})
\]
Therefore \(K_n\not\cong K_m\) if \(n\neq m\).
\end{proof}
\begin{lemma}\label{lem:nonzero_slope_implies_limit_iinfinitty}
Suppose \(\Khr(T)\) contains a curve of a non-zero slope. Then there exists $N$ such that the unoriented knots \(\{K_n\}_{|n|\geq N}\) are all different.
\end{lemma}
\begin{proof}
Let us write \(\Khr(T)=\textcolor{blue}{\gamma_1}\cup \cdots\cup\textcolor{blue}{\gamma_m}\) for some integer \(m>0\).
For each \(i=1,\dots,m\), let \((p_i,q_i)\) be a pair of mutually prime integers such that the slope of \(\textcolor{blue}{\gamma_i}\) is \(\nicefrac{p_i}{q_i}\). By assumption, there is some \(i\in\{1,\dots,m\}\) such that \(p_i\neq0\).
Let
\[
M=\max\left\{\left.\tfrac{|q_i|}{2|p_i|}\right| i\in\{1,\dots,m\}\co p_i\neq0\right\}
\]
Then for \(|n|> M\), the slope \(\nicefrac{1}{2n}\) of the curve \(\textcolor{red}{\a_{\nicefrac{1}{2n}}}\) is distinct from \(\nicefrac{p_i}{q_i}\). Therefore, by Lemma~\ref{lem:pairing_linear_curves:dimension_formula},
\[
\dim\operatorname{HF}(\textcolor{red}{\a_{\nicefrac{1}{2n}}},\textcolor{blue}{\gamma_i})
=
\ell_i\cdot
|q_i-p_i\cdot 2n|
\]
where \(\ell_i\) is the length of \(\textcolor{blue}{\gamma_i}\). As we have seen in the proof of Lemma~\ref{lem:only_zero_slopes_implies_GCCC}, if \(p_i=0\), the dimension of \(\operatorname{HF}(\textcolor{red}{\a_{\nicefrac{1}{2n}}},\textcolor{blue}{\gamma_i})\) is independent of \(n\). If \(p_i\neq0\), the sign of the expression
\(\nicefrac{q_i}{p_i}-2n\) is the same for all \(n>M\).
Therefore, \(\dim\Khr(K_n)\) is a strictly increasing function in \(n\) for $n>M$. The same argument shows that it is strictly decreasing for $n<-M$.
Thus the knots \(\{K_n\}_{n>M}\) are pairwise different, and so are the knots \(\{K_n\}_{n<-M}\).
It remains to distinguish the two families. For this we first prove that the quantum grading of $\{\Khr(K_n)\}_{n\gg0}$ is unbounded above and bounded below.
The existence of a lower bound follows from the following two observations: First, the quantum gradings of the generators of $\textcolor{red}{\a_{\nicefrac{1}{2n}}}$ are bounded above by $-1$. Second, every intersection point generating the Lagrangian Floer homology between $\textcolor{red}{\a_{\nicefrac{1}{2n}}}$ and a rational or special curve can be represented by a homogeneous morphism containing a component labelled by an algebra element of quantum grading greater than or equal to $-2$ (namely one of the algebra elements \(\id,S,S^2, D\in\BNAlgH\)); this follows from an elementary argument about straight lines in the covering space $S^2_{4,\ast}$. Thus, if the minimal grading of a generator of $\DD_1(T)\mathrel{\widehat{=}}\textcolor{blue}{\Khr(T)}$ is $\mu$, the formula from Section~\ref{subsec:background:bigrading} for computing the quantum grading of generators of $\operatorname{HF}(\textcolor{red}{\a_{\nicefrac{1}{2n}}},\textcolor{blue}{\Khr(T)})$ gives us $\mu-(-1)+(-2)$ as a lower bound.
Next, we show that the quantum grading of $\{\Khr(K_n)\}_{n\gg0}$ has no upper bound.
By assumption, there exists a component $\textcolor{blue}{\gamma}$ of $\textcolor{blue}{\Khr(T)}$ of non-zero slope. For $n\gg0$, we may assume that the slope of $\textcolor{blue}{\gamma}$ is bigger than the slope $\nicefrac{1}{2n}$ of the arc $\textcolor{red}{\a_{\nicefrac{1}{2n}}}$. Then, there exists an intersection point close to the generator $\GGzqz{\DotCred}{}{-4n}{}$, which looks like Figure~\ref{fig:agccc:zero:noupperbound}. Clearly, the quantum grading of this generator is unbounded.
Analogous arguments imply that the quantum grading of $\{\Khr(K_n)\}_{n\ll0}$ is unbounded below and bounded above. This proves that there exists $N\gg M$ such that knots in \(\{K_n\}_{n>N} \cup \{K_n\}_{n<-N}\) are pairwise different.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:main}]
We will show that the reduced Khovanov homology of the knots \(\{K_n\}_{|n|\geq N}\) are pairwise different by studying how the invariant \(\Khr(T)\) pairs with the arc \(\textcolor{red}{\a_{\nicefrac{1}{2n}}}\coloneqq\BNr(Q_{\nicefrac{1}{2n}})\).
First, suppose \(\Khr(T)\) contains only rational components of slope $0$: In this case Theorem~\ref{thm:split_tangle_detection} implies that $T$ is horizontally split, contradicting the assumption in Theorem~\ref{thm:main}. Next, suppose \(\Khr(T)\) only contains curves of slope 0 of which at least one is special: This case is covered by Lemma~\ref{lem:only_zero_slopes_implies_GCCC}.
The last case of \(\Khr(T)\) containing curves of non-zero slopes is covered by Lemma~\ref{lem:nonzero_slope_implies_limit_iinfinitty}.
\end{proof}
\subsection{Non-trivial band detection}
Joshua Wang asked if it is possible to recover~\cite[Theorems~1.1 and~1.8]{Wang} using our techniques:
\begin{theorem}[Split closure property]\label{thm:split_link_closure_detection}
Suppose \(T(0)\) is a split link $K \cup K'$. Then \(\Khr(T)\) only contains components of slope \(0\).
\end{theorem}
\begin{proof}
Placing the reduction point to the top-left end of $T$, we consider the reduced Khovanov homology of the split link \(\Khr(T(0))=\Khr(K\cup K')\). Since only one of the two components is reduced (say $K$), we may consider the basepoint action on \(\Khr(K\cup K')\) with respect to a basepoint on $K'$, which we place near the bottom left end of $T$. Keeping in mind that we work with $\mathbb{F} $ coefficients, we have
\[\Khr(K\cup K')=\Khr(K) \otimes_{\mathbb{F}} \Kh(K') \]
and the basepoint action of $\mathbb{F}[x]/(x^2)$ on $\Khr(K\cup K')$ is induced by the basepoint action of $ \mathbb{F}[x]/(x^2)$ on $ \Kh(K')$. Over $\mathbb{F} $ the latter basepoint action on unreduced Khovanov homology is well-known to be free~\cite[Corollary~3.2.C]{Shumakovitch}, and so the basepoint action of \(\mathbb{F}[x]/(x^2)\) on \(\Khr(K\cup K')\) is also free.
We now come back to the tangle $T$ and leverage the description of Khovanov homology in terms of the Floer homology of curves:
$$\Khr(K\cup K')= \Khr(T(0))=\operatorname{HF}(\textcolor{red}{\a_0}, \textcolor{blue}{\Khr(T)})$$
Let us suppose there is a curve $\textcolor{blue}{\gamma}$ in \(\textcolor{blue}{\Khr(T)}\) of slope $\nicefrac p q \neq 0$. By adding twists to the lower two punctures we may assume that $\nicefrac p q$ is positive and as close to $0$ as we want. After pulling the curves tight in the planar cover, each intersection between $\textcolor{red}{\a_0}$ and $\textcolor{blue}{\gamma}$ locally looks as in Figure~\ref{fig:interpicture}. (Here we implicitly use the idea from Remark~\ref{rem_idea}.)
\begin{figure}[t]
\centering
\labellist
\pinlabel $D$ at 25 58
\pinlabel $S$ at 57 45
\pinlabel $S$ at 105 45
\pinlabel $S$ at 208 105
\pinlabel $D$ at 238 96
\pinlabel $S$ at 167 105
\pinlabel $D$ at 138 57
\pinlabel \tiny$\textcolor{red}{0}$ at 154 79
\pinlabel \tiny$\textcolor{blue}{1}$ at 9 47
\pinlabel \tiny$\textcolor{blue}{2}$ at 40 44
\pinlabel \tiny$\textcolor{blue}{3}$ at 122 71
\pinlabel \tiny$\textcolor{blue}{4}$ at 154 68
\pinlabel \tiny$\textcolor{blue}{5}$ at 222 92
\pinlabel \tiny$\textcolor{blue}{6}$ at 254 89
\endlabellist
\includegraphics[scale=1.2]{figures/interpicture}
\caption{The lifts of the arc $\textcolor{red}{\a_0}$ and a rational or special curve $\gamma$ of sufficiently shallow positive slope, based at a common point of intersection }
\label{fig:interpicture}
\end{figure}
Thus the complexes over the algebra $\mathcal B$ associated to the curves $\textcolor{red}{\a_0}$ and $\textcolor{blue}{\gamma}$ are as follows:
\[
\begin{aligned}
\textcolor{red}{\DD_0} &=
\left[
\begin{tikzcd}[nodes={inner sep=2pt},column sep=14pt,ampersand replacement=\&, red]
\DotBred_0
\end{tikzcd}
\right]
\mathrel{\widehat{=}} \textcolor{red}{\a_0}
\\
\textcolor{blue}{\DD_1} &=
\left[
\begin{tikzcd}[nodes={inner sep=2pt},column sep=14pt,ampersand replacement=\&, blue]
\cdots
\arrow[r]
\&
\DotBblue_1
\arrow{r}{D}
\&
\DotBblue_2
\arrow{r}{S^2}
\&
\DotBblue_3
\arrow{r}{D}
\&
\DotBblue_4
\&
\DotBblue_5
\arrow[l,"S^2" above]
\&
\DotBblue_6
\arrow[l,"D" above]
\&
\cdots
\arrow[l]
\end{tikzcd}
\right]
\mathrel{\widehat{=}} \textcolor{blue}{\gamma}
\end{aligned}
\]
where the subscripts are used to simply label generators. According to \cite[Theorem~1.5]{KWZ}, the Lagrangian Floer homology between two curves is isomorphic to the homology of the morphism space between the corresponding complexes:
$$\operatorname{HF}(\textcolor{red}{\a_0},\textcolor{blue}{\gamma})\cong \Homology(\Mor(\textcolor{red}{\DD_0} ,\textcolor{blue}{\DD_1})) \subseteq \Khr(T(0))=\operatorname{HF}(\textcolor{red}{\a_0}, \textcolor{blue}{\Khr(T)})$$
Consider now a morphism consisting of a single arrow $ (\textcolor{red}{\DotBred_0} \xrightarrow{\id} \textcolor{blue}{\DotBblue_4}) \in \Homology(\Mor(\textcolor{red}{\DD_0},\textcolor{blue}{\DD_1}))$.
(This morphism corresponds to the single intersection in Figure~\ref{fig:interpicture}.) The basepoint action multiplies all the labels of a morphism by $D_{\smallDotB}$; see the discussion before \cite[Lemma~6.46]{KWZthinness} for a detailed explanation of the basepoint action in the context of morphism spaces of chain complexes. Thus the basepoint action sends the morphism $(\textcolor{red}{\DotBred_0} \xrightarrow{\id} \textcolor{blue}{\DotBblue_4})$ to the morphism $(\textcolor{red}{\DotBred_0} \xrightarrow{D} \textcolor{blue}{\DotBblue_4})$, which is null-homotopic; the null-homotopy is $(\textcolor{red}{\DotBred_0} \xrightarrow{\id} \textcolor{blue}{\DotBblue_3})$. Furthermore, the morphism $(\textcolor{red}{\DotBred_0} \xrightarrow{\id} \textcolor{blue}{\DotBblue_4})$ is not in the image of the basepoint action, because every morphism that is homotopic to the one in the image of the basepoint action cannot contain identity arrows $\xrightarrow{\id}$.
This is because both $\textcolor{red}{\DD_0}$ and $\textcolor{blue}{\DD_1}$ correspond to the pulled tight curves, and thus do not contain any $\xrightarrow{\id}$ in their differentials.
We conclude that the morphism $(\textcolor{red}{\DotBred_0} \xrightarrow{\id} \textcolor{blue}{\DotBblue_4})$ represents torsion in the basepoint action of \(\mathbb{F}[x]/(x^2)\) on \( \Khr(T(0))\). So this action is not free, contradicting the fact that \(T(0)\) is split.
\end{proof}
\begin{theorem}
Suppose $T$ is not horizontally split, and \(T(0)\) is a split link. Then the unoriented knots $\{K_n\}_{n\in \mathbb{Z}}$ are pairwise different.
\end{theorem}
\begin{proof}
Theorem~\ref{thm:split_link_closure_detection} implies that \(\Khr(T)\) only contains components of slope 0. By Theorem~\ref{thm:split_tangle_detection} we know that the tangle~\(T\) must contain special components of slope 0. Lemma~\ref{lem:only_zero_slopes_implies_GCCC} now proves the statement.
\end{proof}
\subsection{Towards split closure detection}
It is natural to wonder if the converse of Theorem~\ref{thm:split_link_closure_detection} also holds.
\begin{conjecture}[Split closure detection]\label{conj:split_closure_detection}
Given a Conway tangle \(T\), \(T(0)\) is a split link if and only if \(\Khr(T)\) only contains components of slope \(0\).
\end{conjecture}
In this direction, we can offer the following result:
\begin{theorem}
\(T(0)\) is a split link if \(\Khr(T)\) only contains components of slope 0 and no rational curves of length greater than 1.
\end{theorem}
\begin{proof}
If \(\Khr(T)\) only contains components of slope 0 then \(\Khr(T(0))\) is isomorphic to the Lagrangian Floer homology between \(\BN(\Lo)\) and the rational components of \(\Khr(T)\). If all rational components are (up to grading shift) equal to \(\r_1(0)\) this implies that the basepoint action on \(\Khr(T(0))\) is free. So by the main result of \cite{LS}, \(T(0)\) is a split link.
\end{proof}
To prove Conjecture~\ref{conj:split_closure_detection}, it remains to show that if $\Khr(T;\mathbb{F})$ only contains curves of slope 0 then all its rational components have length 1.
Note that it is important that we use coefficients in \(\mathbb{F}\), because this statement is false away from characteristic $2$.
However, over $\mathbb{F}$, we in fact expect this to be a more general property of the multicurve invariants \(\Khr\):
\begin{conjecture}
For any Conway tangle \(T\), the length of any rational component of \(\Khr(T;\mathbb{F})\) is equal to 1.
\end{conjecture}
\section{Review of the Khovanov multicurve invariants}\label{sec:background}
In this section, we review some properties of the immersed curve invariants \(\Khr\) and \(\BNr\) of pointed Conway tangles from~\cite{KWZ}.
We work exclusively over the field \(\mathbb{F}\) of two elements and only summarize those properties that we will need in this paper; more elaborate introductions highlighting different aspects of the invariants can be found in~\cite{KWZ-strong, KWZthinness}.
Let \(T\) be an oriented \emph{pointed} Conway tangle, that is a four-ended tangle in the three-ball \(B^3\) with a choice of distinguished tangle end, which we mark by~\(\ast\).
Denote by \(S^2_{4,\ast}\) the four-punctured sphere \(\partial B^3\smallsetminus \partial T\); the puncture marked by \(\ast\) will be called \emph{special}.
We associate with such a tangle \(T\) invariants \(\BNr(T)\) and \(\Khr(T)\) that take the form of multicurves on \(S^2_{4,\ast}\).
By multicurve, we mean a collection of immersed curves that carry certain extra data.
Broadly speaking, there are two kinds of such curves: compact and non-compact.
A compact immersed curve in \(S^2_{4,\ast}\) is an immersion of \(S^1\), considered up to regular homotopy, that (up to conjugation) defines a primitive element of \(\pi_1(S^2_{4,\ast})\), and
each of these curves is decorated with a local system, ie an invertible matrix over \(\mathbb{F}\)
considered up to matrix similarity.
A non-compact immersed curve in \(S^2_{4,\ast}\) is a non-null-homotopic immersion of an interval, with ends on the three non-special punctures of \(S^2_{4,\ast}\); see \cite[Definition~1.4]{KWZ}.
Non-compact curves do not carry local systems.
In addition, all curves are equipped with a bigrading; more on this in Section~\ref{subsec:background:bigrading} below.
\begin{remark}\label{rem:Khr:2m3pt}
We often draw \(S^2_{4,\ast}\) as the plane plus a point at infinity minus the four punctures. To help identify this abstract surface with \(\partial B^3\smallsetminus \partial T\), we then add two dotted gray arcs that parametrize the surface, see
Figures~\ref{fig:Kh:example:tangle}--\ref{fig:Kh:example:Curve:Downstairs}.
The blue curves in these figures show the multicurves \(\BNr(P_{2,-3})\) and \(\Khr(P_{2,-3})\) for the pretzel tangle \(P_{2,-3}\); cf~\cite[Example~6.7]{KWZ}.
All components of these curves carry the (unique) one-dimensional local system.
\end{remark}
\begin{figure}[b]
\centering
\begin{subfigure}{0.30\textwidth}
\centering
\(\pretzeltangleKh\)
\caption{The pretzel tangle \(P_{2,-3}\)}\label{fig:Kh:example:tangle}
\end{subfigure}
\begin{subfigure}{0.28\textwidth}
\centering
\(\pretzeltangleDownstairsBNr\)
\caption{\(\BNr(P_{2,-3})\)}\label{fig:BN:example:Curve:Downstairs}
\end{subfigure}
\begin{subfigure}{0.28\textwidth}
\centering
\(\pretzeltangleDownstairsKhr\)
\caption{\(\Khr(P_{2,-3})\)}\label{fig:Kh:example:Curve:Downstairs}
\end{subfigure}
\bigskip
\\
\begin{subfigure}{0.42\textwidth}
\centering
\(\pretzeltangleUpstairsBNr\)
\caption{A lift of \(\BNr(P_{2,-3})\) to \(\mathbb{R}^2\smallsetminus \mathbb{Z}^2\) }\label{fig:BN:example:Curve:Upstairs}
\end{subfigure}
\begin{subfigure}{0.55\textwidth}
\centering
\(\pretzeltangleUpstairsKhr\)
\caption{A lift of \(\Khr(P_{2,-3})\) to \(\mathbb{R}^2\smallsetminus \mathbb{Z}^2\) }\label{fig:Kh:example:Curve:Upstairs}
\end{subfigure}
\caption{The multicurve invariants for the pretzel tangle \(P_{2,-3}\).
Under the covering \(\mathbb{R}^2\smallsetminus \mathbb{Z}^2\rightarrowS^2_{4,\ast}\), the shaded regions in (b+c) correspond to the shaded regions in (d+e).
}\label{fig:Kh:example}
\end{figure}
\subsection{The construction of the multicurves}\label{subsec:background:construction} The starting point is the algebraic tangle invariant \(\KhTl{T}\) due to Bar-Natan. The invariant
\(\KhTl{T}\) is a chain complex over a certain cobordism category, whose objects are crossingless tangle diagrams \cite{BarNatanKhT}; we refer to~\cite[Section~2]{KWZ} for a detailed introduction to complexes over categories/algebras, and the equivalent viewpoint through type~D structures.
In \cite[Theorem~1.1]{KWZ}, it was shown that any such complex can be rewritten as a chain complex \(\DD(T)\) over the following category $\BNAlgH$, consisting of two objects and morphisms equal to paths in a quiver modulo relations:
\begin{equation*}\label{eq:B_quiver}
\BNAlgH
\coloneqq
\mathbb{F}\Big[
\begin{tikzcd}[row sep=2cm, column sep=1.5cm]
\DotB
\arrow[leftarrow,in=145, out=-145,looseness=5]{rl}[description]{D_{\bullet}}
\arrow[leftarrow,bend left]{r}[description]{S_{\circ}}
&
\DotC
\arrow[leftarrow,bend left]{l}[description]{S_{\bullet}}
\arrow[leftarrow,in=35, out=-35,looseness=5]{rl}[description]{D_{\circ}}
\end{tikzcd}
\Big]\Big/\Big(
\parbox[c]{90pt}{\footnotesize\centering
$D_{\bullet} \cdot S_{\circ}=0=S_{\circ}\cdot D_{\circ}$\\
$D_{\circ}\cdot S_{\bullet}=0=S_{\bullet}\cdot D_{\bullet}$
}\Big)
\end{equation*}
Here, the objects \(\DotB\) and \(\DotC\) correspond to the crossingless tangles \(\Lo\) and \(\Li\), respectively.
We will refer to $\BNAlgH$ as a (quiver) algebra, and to \(\DD(T)\) as a chain complex over the algebra $\BNAlgH$.
Defining \(D\coloneqq D_{\bullet} + D_{\circ}\) and \(S\coloneqq S_{\bullet} + S_{\circ}\) often allows us to drop the subscripts of the algebra elements of \(\BNAlgH\).
The chain homotopy type of \(\DD(T)\) is an invariant of the tangle~\(T\).
Moreover, using the central element
\[
H\coloneqq D+S^2 = D_{\bullet} + D_{\circ} + S_{\circ}S_{\bullet} + S_{\bullet}S_{\circ} ~ \in ~ \BNAlgH
\]
we define a chain complex \(\DD_1(T)\) as the mapping cone
\[\DD_1(T)\coloneqq \Big[q^{-1}h^{-1}\DD(T)\xrightarrow{H\cdot \id} q^{1}h^0\DD(T)\Big]\]
where $H\cdot \id$ is the endomorphism of $\DD(T)$ defined by $x\xrightarrow{H}x$ for all generators $x$ of $\DD(T)$.
The chain homotopy type of \(\DD_1(T)\) is also a tangle invariant.
\begin{figure}[b]
\centering
\begin{subfigure}[t]{0.29\textwidth}
\centering
$\PairingTrefoilLoopINTRO$
\caption{$\textcolor{red}{\gamma\mathrel{\widehat{=}}
[
\protect\begin{tikzcd}[nodes={inner sep=2pt}, column sep=23pt,ampersand replacement = \&]
\protect\DotCred
\protect\arrow{r}{D+S^2}
\protect\&
\protect\DotCred
\protect\end{tikzcd}
]}$}\label{fig:exa:classification:curves:loop}
\end{subfigure}
\begin{subfigure}[t]{0.29\textwidth}
\centering
$\PairingTrefoilArcINTRO$
\caption{$\textcolor{blue}{\gamma'\mathrel{\widehat{=}}
[
\protect\begin{tikzcd}[nodes={inner sep=2pt}, column sep=13pt,ampersand replacement = \&]
\protect\DotCblue
\protect\arrow{r}{S}
\protect\&
\protect\DotBblue
\protect\arrow{r}{D}
\protect\&
\protect\DotBblue
\protect\arrow{r}{S^2}
\protect\&
\protect\DotBblue
\protect\end{tikzcd}
]}$}\label{fig:exa:classification:curves:arc}
\end{subfigure}
\begin{subfigure}[t]{0.29\textwidth}
\centering
$\PairingBothINTRO$
\vspace{5pt}
\caption{\(\operatorname{HF}(\textcolor{red}{\gamma},\textcolor{blue}{\gamma'})\cong\mathbb{F}^3\)}\label{fig:exa:classification:curves:pairing}
\end{subfigure}
\caption{Two immersed curves and their corresponding chain complexes (a+b) and their Lagrangian Floer homology (c); cf \cite[Examples~1.6 and~1.7]{KWZ}
}\label{fig:exa:classification:curves}
\end{figure}
The multicurve invariants \(\BNr(T)\) and \(\Khr(T)\) are geometric interpretations of $\DD(T)$ and $\DD_1(T)$ respectively, made possible by the following classification result:
The homotopy equivalence classes of chain complexes over \(\BNAlgH\) are in one-to-one correspondence with multicurves on the four-punctured sphere \(S^2_{4,\ast}\) \cite[Theorem~1.5]{KWZ}.
In a little more detail, this correspondence (which we denote by $\mathrel{\widehat{=}}$) uses the parametrization of \(S^2_{4,\ast}\) given by the two dotted arcs described in Remark~\ref{rem:Khr:2m3pt}.
We will generally assume that the multicurves intersect these arcs minimally. Then, roughly speaking, the intersection points correspond to generators of the associated chain complexes and paths between those intersection points correspond to the differentials.
The two examples in Figure~\ref{fig:exa:classification:curves} should give the reader a general impression how this works.
\begin{example}\label{exa:rational_tangles}
For the trivial tangle \(Q_\infty=\Li\), the chain complex \(\DD(Q_\infty)\) consists of a single object \(\DotC\) and the differential vanishes. The corresponding multicurve \(\BNr(Q_{\infty})\) consists of a single vertical arc connecting the two non-special tangle ends. The chain complex \(\DD_1(Q_\infty)\) and the corresponding curve \(\Khr(Q_\infty)\) is shown in Figure~\ref{fig:exa:classification:curves:loop}. The local system on this curve is one-dimensional.
The tangle \(Q_{\nicefrac{1}{3}}=\ThreeTwistTangle\) is obtained from the trivial tangle \(Q_\infty\) by adding three twists to the two lower tangle ends. Its invariant \(\BNr(Q_{\nicefrac{1}{3}})\) is shown in Figure~\ref{fig:exa:classification:curves:arc}. Note that it agrees with the vertical arc \(\BNr(Q_{\infty})\) up to three twists. This is not a coincidence; one can show that adding twists to \emph{any} tangle (not just a rational tangle) corresponds to adding twists to the multicurves; see \cite[Theorem~1.13]{KWZ}.
Thus, the identification of \(\partial B^3\smallsetminus \partial T\) with the abstract surface~\(S^2_{4,\ast}\) containing the multicurves is natural.
\end{example}
\subsection{A gluing theorem}\label{subsec:background:gluing}
\begin{figure}[t]
\centering
\(
\tanglepairingI
\quad = \quad
\tanglepairingII
\)
\caption{Two tangle decompositions defining the link \(T_1\cup T_2\). The tangle \protect\reflectbox{\(T_2\)} is the result of rotating \(T_2\) around the vertical axis. By rotating the entire link on the right-hand side around the vertical axis, we can see that \(T_1\cup T_2=T_2\cup T_1\).}
\label{fig:tanglepairing}
\end{figure}
The multicurve invariants satisfy various gluing formulas \cite[Theorem~1.9]{KWZ}. The one that we will use in this paper is the following:
\begin{theorem}\label{thm:GlueingTheorem:Kh}
Let \(L=T_1\cup T_2\) be the result of gluing two oriented pointed Conway tangles as in Figure~\ref{fig:tanglepairing} such that the orientations match.
Let \(\mirror\)
be the map identifying the two four-punctured spheres.
Then
\[
\Khr(L)
\cong
\operatorname{HF}\left(\mirror(\Khr(T_1)),\BNr(T_2)\right)
\cong
\operatorname{HF}\left(\mirror(\BNr(T_1)),\Khr(T_2)\right)
\]
\end{theorem}
The Lagrangian Floer homology \(\operatorname{HF}(\textcolor{red}{\gamma},\textcolor{blue}{\gamma'})\) between two curves \(\textcolor{red}{\gamma}\) and \(\textcolor{blue}{\gamma'}\) is a vector space that can be computed as follows. First, we draw the curves in such a way that minimizes the number of intersection points between \(\textcolor{red}{\gamma}\) and \(\textcolor{blue}{\gamma'}\).
\(\operatorname{HF}(\textcolor{red}{\gamma},\textcolor{blue}{\gamma'})\) is then equal to a vector space freely generated by those intersection points, provided that the curves are not homotopic to each other \cite[Theorem~5.25]{KWZ}.
(We will always be able to make this assumption in this paper.)
For instance, with Example~\ref{exa:rational_tangles} and Figure~\ref{fig:exa:classification:curves} in mind, the Khovanov homology of the trefoil can be computed as follows:
$$\Khr(\Li \cup \ThreeTwistTangle) \cong \operatorname{HF}( \Khr(\Li), \BNr(\ThreeTwistTangle)=\mathbb{F}^3$$
Finally, the Lagrangian Floer homology between two multicurves is simply the direct sum of the Lagrangian Floer homologies between individual components.
\subsection{Gradings}\label{subsec:background:bigrading}
Khovanov homology is a \emph{bigraded} homology theory, and this bigrading is often what makes it a powerful invariant.
The multicurves \(\BNr(T)\) and \(\Khr(T)\) also carry bigradings. We now describe how the gradings work: first on the algebra $\BNAlgH$, then on chain complexes, then on multicurves, and then finally on Lagrangian Floer homology between multicurves.
Equip the algebra \(\BNAlgH\) with quantum grading \(q\), which is determined by
\[
q(D_{\bullet}) = q(D_{\circ}) = -2
\qquad
\text{and}
\qquad
q(S_{\bullet}) = q(S_{\circ}) = -1
\]
The homological grading is defined to be $0$ for all elements of \(\BNAlgH\).
Differentials of bigraded chain complexes over \(\BNAlgH\) are required to preserve quantum grading and increase the homological grading by $1$. Concretely this means that if a differential contains a morphism $x \xrightarrow{a}y$ (where $x$ and $y$ are generators of the complex and $a\in\BNAlgH$), then $q(a)+q(y)-q(x)=0$ and $h(a)+h(y)-h(x)=1$.
We often specify the quantum gradings \(\QGrad{q}\) of generators of such complexes via superscripts, like so: \(\GGzqh{x}{}{q}{}\).
The bigrading on a multicurve takes the form of a bigrading on the intersection points between the multicurve and the two parametrizing arcs, ie the generators of the corresponding chain complex over \(\BNAlgH\).
If \(T\) is an unoriented tangle, the bigradings on \(\BNr(T)\) and \(\Khr(T)\) are only well-defined as relative bigradings; to fix the overall shift, an orientation of \(T\) is required.
Let \(\textcolor{red}{\gamma}\) and \(\textcolor{blue}{\gamma'}\) be two bigraded multicurves and suppose \(\textcolor{red}{X\mathrel{\widehat{=}}\gamma}\) and \(\textcolor{blue}{X'\mathrel{\widehat{=}}\gamma'}\) are the corresponding bigraded chain complexes over \(\BNAlgH\).
Then \(\operatorname{HF}(\textcolor{red}{\gamma},\textcolor{blue}{\gamma'})\) also carries a bigrading.
It can be computed using the fact that this vector space is bigraded isomorphic to the homology of the morphism space \(\Mor(\textcolor{red}{X},\textcolor{blue}{X'})\) \cite[Theorem~1.5]{KWZ} (see the discussion before~\cite[Definition~2.4]{KWZ} for the definition of the differential on a morphism space between two complexes).
The quantum grading of a morphism
\(
\begin{tikzcd}[nodes={inner sep=2pt}, column sep=13pt,ampersand replacement = \&]
\GGzqh{x}{}{b}{}
\arrow{r}{\alpha}
\&
\GGzqh{y}{}{a}{}
\end{tikzcd}
\),
where \(\alpha\in\BNAlgH\), is computed using the formula \(\QGrad{a}-\QGrad{b}+q(\alpha)\); an analogous formula holds for the homological grading.
Each intersection point generating \(\operatorname{HF}(\textcolor{red}{\gamma},\textcolor{blue}{\gamma'})\) corresponds to a morphism from which we can read off the bigrading \cite[Section~7]{KWZ-strong}. For instance, the highlighted intersection point in Figure~\ref{fig:exa:classification:curves:pairing} corresponds to the morphism \(
\begin{tikzcd}[nodes={inner sep=2pt}, column sep=13pt,ampersand replacement = \&]
\DotCred
\arrow{r}{S}
\&
\DotBblue
\end{tikzcd}
\),
so the bigrading of this intersection point is equal to
\[
q(\DotB) = q(\DotBblue)-q(\DotCred)+q(S)
\quad
\text{ and }
\quad
h(\DotB) = h(\DotBblue)-h(\DotCred)
\]
\subsection{\texorpdfstring{Geography of components of \(\Khr\)}{Geography of components of Khr}}\label{subsec:background:geography}
We now recall some basic facts about $\Khr(T)$ and $\BNr(T)$ from \cite[Section~6]{KWZ}.
In this paper, we will focus only on tangles without closed components. For such tangles,
\(\BNr(T)\) consists of a single non-compact component and a (possibly zero) number of compact components.
In contrast, $\Khr(T)$ consists of compact components only.
Often, multicurves become easier to manage when considered in a certain covering space of~\(S^2_{4,\ast}\), namely the planar cover that factors through the toroidal two-fold cover:
$$
(\mathbb{R}^2 \smallsetminus \mathbb{Z}^2) \to (T^2 \smallsetminus 4{\text{pt}}) \to S^2_{4,\ast}
$$
This is illustrated in Figure~\ref{fig:Kh:example} for the multicurve invariants of the pretzel tangle \(P_{2,-3}\).
\begin{figure}[t]
\centering
\begin{subfigure}{0.3\textwidth}
\centering
\labellist
\footnotesize \color{blue}
\pinlabel $\mathbf{s}_{2}(0)$ at 65 115
\pinlabel $\mathbf{r}_{1}(0)$ at 65 28
\endlabellist
\includegraphics[scale=1]{figures/GeographyDownstairs1}
\caption{\(n=1\)}\label{fig:geography:Downstairs1}
\end{subfigure}
\begin{subfigure}{0.3\textwidth}
\centering
\labellist
\footnotesize \color{blue}
\pinlabel $\mathbf{s}_{4}(0)$ at 65 118
\pinlabel $\mathbf{r}_{2}(0)$ at 65 13
\endlabellist
\includegraphics[scale=1]{figures/GeographyDownstairs2}
\caption{\(n=2\)}\label{fig:geography:Downstairs2}
\end{subfigure}
\begin{subfigure}{0.3\textwidth}
\centering
\labellist
\footnotesize \color{blue}
\pinlabel $\mathbf{s}_{6}(0)$ at 65 121
\pinlabel $\mathbf{r}_{3}(0)$ at 65 10
\endlabellist
\includegraphics[scale=1]{figures/GeographyDownstairs3}
\caption{\(n=3\)}\label{fig:geography:Downstairs3}
\end{subfigure}
\\
\begin{subfigure}{0.9\textwidth}
\centering
\(\GeographyCovering\)
\caption{}\label{fig:geography:Upstairs}
\end{subfigure}
\caption{The curves \(\mathbf{r}_n(0)\) and \(\mathbf{s}_{2n}(0)\) (a--c) and their lifts to \(\mathbb{R}^2\smallsetminus \mathbb{Z}^2\) (d). While not visually apparent, the curves \(\mathbf{r}_n(0)\) are invariant under the Dehn twist interchanging the lower two punctures.
}\label{fig:geography}
\end{figure}
\begin{definition}
Given an immersed curve $c \looparrowright S^2_{4,\ast}$, denote by $\tilde{\c}$ a lift of $\c$ to the cover \(\mathbb{R}^2\smallsetminus \mathbb{Z}^2\).
For \(n\in\mathbb N\), let \(\mathbf{r}_{n}(0)\) and \(\mathbf{s}_{2n}(0)\) be the immersed curves in \(S^2_{4,\ast}\) that respectively admit lifts to the curves $\tilde{\r}_{n}(0)$ and $\tilde{\s}_{2n}(0)$ in Figure~\ref{fig:geography:Upstairs}; curves for $n=1,2,3$ are illustrated in Figures~\ref{fig:geography:Downstairs1}--\ref{fig:geography:Downstairs3}.
For every \(\nicefrac{p}{q}\in\operatorname{\mathbb{Q}P}^1\), we respectively define the curves $\mathbf{r}_n(\nicefrac{p}{q})$ and $\mathbf{s}_{2n}(\nicefrac{p}{q})$ as the images of \(\mathbf{r}_{n}(0)\) and \(\mathbf{s}_{2n}(0)\) under the action of
\[
\begin{bmatrix*}[c]
q & r \\
p & s
\end{bmatrix*}
\]
considered as an element the mapping class group fixing the special puncture $\Mod(S^2_{4,\ast}) \cong \PSL(2,\mathbb{Z})$, where \(qs-pr=1\).
(This transformation maps straight lines of slope 0 to straight lines of slope \(\nicefrac{p}{q}\).)
We call \(\mathbf{r}_{n}(\nicefrac{p}{q})\) a curve of \emph{rational type, slope \(\nicefrac{p}{q}\), and length \(n\)}.
We call \(\mathbf{s}_{2n}(\nicefrac{p}{q})\) a curve of \emph{special type, slope \(\nicefrac{p}{q}\), and length \(2n\)}.
The local systems on all these curves are defined to be trivial.
\end{definition}
The following classification result is \cite[Theorem~6.5]{KWZthinness}.
\begin{theorem}
For any pointed Conway tangle \(T\), every component of \(\Khr(T)\) is equal to \(\mathbf{r}_n(\nicefrac{p}{q})\) or \(\mathbf{s}_{2n}(\nicefrac{p}{q})\) for some \(n\in\mathbb N\) and \(\nicefrac{p}{q}\in\operatorname{\mathbb{Q}P}^1\), up to some bigrading shift.
In other words, components of \(\Khr(T)\) are completely classified by their type, slope, length, and bigrading.
\end{theorem}
As already mentioned in Example~\ref{exa:rational_tangles}, the multicurve invariants are natural with respect to adding twists; these twists generate $\Mod(S^2_{4,\ast})$.
Thus, the invariant \(\Khr(Q_{\nicefrac{p}{q}})\) of a \(\nicefrac{p}{q}\)-rational tangle \(Q_{\nicefrac{p}{q}}\) is equal to \(\mathbf{r}_1(\nicefrac{p}{q})\), justifying the terminology.
In fact, we have the following detection result \cite[Theorem~5.7]{KWZthinness}.
\begin{theorem}\label{thm:rational_tangle_detection}
A pointed Conway tangle \(T\) is rational if and only if \(\Khr(T)\) consists of a single component \(\mathbf{r}_1(\nicefrac{p}{q})\) for some \(\nicefrac{p}{q}\in\operatorname{\mathbb{Q}P}^1\).
\end{theorem}
Rational components can also occur in the invariants of non-rational tangles. In fact, if \(T\) has no closed component, we know that there is always at least one such component \cite[Corollary~6.42]{KWZthinness}. For example, the curve $\Khr(P_{2,-3})$ from Figure~\ref{fig:Kh:example:Curve:Downstairs} consists of the special component $\mathbf{s}_4(0)$ and the rational component $\r_1(\nicefrac{1}{2})$.
Such rational components detect how tangle ends are connected \cite[Theorem~6.41]{KWZthinness}:
\begin{theorem}\label{thm:detection:connectivity}
Suppose a pointed Conway tangle \(T\) has connectivity \(\No\). Then the slope \(\nicefrac{p}{q}\in\operatorname{\mathbb{Q}P}^1\) of any odd-length rational component of \(\Khr(T)\) satisfies \(p\equiv0\mod 2\).
\end{theorem}
\begin{remark}
Components of the invariant $\BNr(T)$, even compact ones, can be much more complicated than components of \(\Khr(T)\); for an example, see \cite[Figure~26]{KWZthinness}. However, the invariants of rational tangles are very simple: Any lift of \(\BNr(Q_{\nicefrac{p}{q}})\) to \(\mathbb{R}^2\smallsetminus \mathbb{Z}^2\) is a straight line segment of slope \(\nicefrac{p}{q}\) connecting the lifts of two non-special punctures.
\end{remark}
\subsection{A dimension formula}
\begin{definition}
Given two slopes \(\nicefrac{p}{q},\nicefrac{p'}{q'}\in\operatorname{\mathbb{Q}P}^1\), where \((p,q)\) and \((p',q')\) are pairs of mutually prime integers, define the distance between two slopes as
\[
\Delta(\nicefrac{p}{q},\nicefrac{p'}{q'})
\coloneqq
\left|
\det
\begin{bmatrix}
q & q' \\
p & p'
\end{bmatrix}
\right|
=
|qp'-pq'|
\]
\end{definition}
\begin{lemma}\label{lem:pairing_linear_curves:dimension_formula}
Let \(s,r\in\operatorname{\mathbb{Q}P}^1\) be two distinct slopes.
Let \(\a_{s}\coloneqq\BNr(Q_{s})\) and
let \(\gamma\) be a rational or special curve of length \(\ell\) and slope \(r\). Then
\[
\dim\operatorname{HF}(\a_{s},\gamma)
=
\ell\cdot\Delta(s,r)
\]
\end{lemma}
\begin{proof}
Write \(s=\nicefrac{p}{q}\) and \(r=\nicefrac{p'}{q'}\) for pairs \((p,q)\) and \((p',q')\) of mutually prime integers.
Observe that \(\dim\operatorname{HF}(\a_{s},\gamma)\) stays invariant under changing the parametrization of the four-punctured sphere.
So let us apply the linear transformation corresponding to the matrix
\[
\begin{bmatrix}
n & -m\\
p' & -q'
\end{bmatrix}
\]
where \(n,m\in\mathbb{Z}\) are such that \(mp'-nq'=1\).
This transformation maps \(\gamma\) to a curve of slope \(0\) and \(\a_s\) to a curve of slope \(\tfrac{p'q-q'p}{nq-mp}\). The distance between these curves remains the same.
This shows that if the formula holds for the case \(r=0\), then it also holds in general.
So suppose \(r=0\). In this case, \(\Delta(s,r)=|p|\), so we need to see that
\[
\dim\operatorname{HF}(\a_s,\gamma)
=
\ell\cdot|p|
\]
This can be easily checked in the covering space \(\mathbb{R}^2\smallsetminus \mathbb{Z}^2\).
If \(\tilde{\a}_s\) is a lift of \(\a_s\), \(\dim\operatorname{HF}(\a_s,\gamma)\) is equal to the number of times that \(\tilde{\a}_s\) intersects the preimage of \(\gamma\), which, by inspection, is indeed equal to \(\ell\cdot|p|\).
\end{proof}
\begin{remark}\label{rem_idea}
In the remainder of this paper we will frequently use the idea above without explicitly referencing it, namely that if two curves $\gamma_1,\gamma_2$ intersect minimally in $S^2_{4,\ast}$, then we can count these intersections by looking at intersections of the preimage of $\gamma_2$ and any lift of $\gamma_1$ in $\mathbb{R}^2 \smallsetminus \mathbb{Z}^2$.
\end{remark}
\section{Proof of Theorem~\ref{thm:intro:cosmeticrational}}\label{sec:ecsc}
The two-fold branched cover \(\Sigma(T)\) of $B^3$ branched over $T$ is the exterior of a knot \(K\subset S^3\).
Suppose, without loss of generality, that \(T(\infty)\) is the unknot.
Then, a curve of slope~\(\infty\) on the boundary of the three-ball containing~\(T\) lifts to a meridian of~\(K\). By adding the appropriate number of twists on the right of the tangle \(T\), we can further assume that \(T\) is parametrized such that a curve of slope~\(0\) lifts to a longitude of~\(K\). Suppose now that \(T(r)\cong T(r')\) as unoriented links. Then
\[
S^3_r(K)= \Sigma(T(r))\cong \Sigma(T(r'))=S^3_{r'}(K).
\]
According to \cite[Theorem~2]{HanselmanCSC}, this implies that \(r=-r'\) where
\[
r=\pm2
\text{ (case 1)} \quad \text{or} \quad
r=\nicefrac{\pm1}{n}
\text{ for some positive integer }n \text{ (case 2)}
\]
Moreover, as a consequence of our chosen parametrization, the connectivity of the tangle~\(T\) is \(\No\).
This can be seen as follows: First, observe that \(T\) has no closed component and the connectivity of the tangle is not \(\Ni\). Both follow from the fact that \(T(\infty)\) is the unknot.
Secondly, if \(T(0)\) were a knot, its two-fold branched cover would be a rational homology sphere.
This contradicts the fact that the two-fold branched cover is 0-surgery on the knot \(K\subset S^3\).
So \(T(0)\) is a two-component link and hence the connectivity of \(T\) is not \(\Nio\).
The strategy for the proof is to compare the reduced Khovanov homologies \(\Khr(T(r))\) and \(\Khr(T(r'))\).
We first equip \(T(r)\) and \(T(r')\) with orientations such that they agree as oriented links.
Then \(\Khr(T(r))\) and \(\Khr(T(r'))\) agree as absolutely bigraded groups.
We work with coefficients in~\(\mathbb{F}\), so that reduced Khovanov homology is independent of the reduction point \cite[Corollary~3.2.C]{Shumakovitch}.
We will compute \(\Khr(T(r))\) and \(\Khr(T(r'))\) by pairing the $\BNr$-invariants of the rational tangle fillings (arcs) with the multicurve \(\textcolor{blue}{C}\coloneqq\Khr(T)\).
A priori, the absolute bigrading on \(\textcolor{blue}{C}\) depends on the orientation of the tangle, but as we will see below, it is in fact orientation independent.
Since \(T(\infty)\) is the unknot, we know that \(\textcolor{blue}{C}\) has only one intersection with the vertical arc \(\a_\infty\coloneqq\BNr(\Li)\).
Special curves $\mathbf{s}_{2n}(s)$ intersect \(\a_\infty\) in more than one point, unless they have slope \(s=\infty\), in which case they are disjoint from \(\a_\infty\).
Similarly, a rational curve $\mathbf{r}_{n}(s)$ intersects \(\a_\infty\) in more than one point, unless $n=1$ and \(s\in \mathbb Z\), in which case there is a single intersection point.
Hence, we may write
\[
\textcolor{blue}{C}
\coloneqq
\textcolor{blue}{\gamma_1}
\cup \dots\cup
\textcolor{blue}{\gamma_m}
\cup
\textcolor{blue}{\rho},
\]
where
\(
\textcolor{blue}{\gamma_1},
\dots,
\textcolor{blue}{\gamma_m}
\)
are special components of slope \(\infty\) and \(\textcolor{blue}{\rho}=\mathbf{r}_1(s)\), of slope \(s\in\mathbb Z\). Since \(T\) is non-rational, \(m>0\) by Theorem~\ref{thm:rational_tangle_detection}.
Moreover, by Theorem~\ref{thm:detection:connectivity}, we know that \(s\) is an even integer.
We now consider the two cases separately. The arguments in both cases are essentially the same. We first show that the slope \(s\) of \(\textcolor{blue}{\rho}\) must be 0 for the total dimensions of \(\Khr(T(r))\) and \(\Khr(T(r'))\) to agree; then we compute the absolute quantum gradings and observe that they are different.
\medskip\noindent {\bf Case 1: \(\{r,r'\}=\{\pm2\}\). }
Since the connectivity of the tangle \(T\) is \(\No\), \(T(+2)=T(-2)\) is a link with two components.
Consider their linking number. If we choose the same orientation of the tangle \(T\), the linking numbers of \(T(+2)\) and \(T(-2)\) are different, since the crossings in the \(\pm2\)-twist tangles then have different signs.
So up to an overall orientation reversal (which does not affect the reduced Khovanov homology), we may assume that the orientations on \(T(+2)\) and \(T(-2)\) are as follows:
\[
T(+2)=
\Tpii
\qquad
T(-2)=
\Tmii
\]
Since the crossings in the \(\pm2\)-twist tangles are all positive, the linking number of the tangle \(T\) with the orientation as in \(T(+2)\) is the same as with the orientation as in \(T(-2)\). (The linking number of a tangle is defined in \cite[Definition~4.7]{KWZ}.) Since the two orientations are obtained by reversing one strand, these linking numbers also differ by a sign, so the linking number of $T$ is zero. Hence \(\textcolor{blue}{C}=\Khr(\textcolor{blue}{T})\) is independent of orientations; see for example~\cite[Proposition~4.8]{KWZ}.
Define two arcs
\[
\textcolor{red}{\a_+}\mathrel{\widehat{=}}
\left[
\begin{tikzcd}[nodes={inner sep=2pt},column sep=14pt, red]
\GGzqh{\DotCred}{}{-5}{}
\arrow{r}{D}
&
\GGzqz{\DotCred}{}{-3}{}
\arrow{r}{S}
&
\GGzqz{\DotBred}{}{-2}{}
\end{tikzcd}\right]
\qquad
\text{and}
\qquad
\textcolor{red}{\a_-}\mathrel{\widehat{=}}
\left[
\begin{tikzcd}[nodes={inner sep=2pt},column sep=14pt, red]
\GGzqh{\DotBred}{}{-4}{}
\arrow{r}{S}
&
\GGzqz{\DotCred}{}{-3}{}
\arrow{r}{D}
&
\GGzqz{\DotCred}{}{-1}{}
\end{tikzcd}\right]
\]
These are the arc invariants of the mirrors of the \(\pm2\)-twist tangles in \(T(+2)\) and \(T(-2)\), respectively. (For instance, this calculation follows from \cite[Example~4.27]{KWZ}, using (1) the relation $q=2(h+\delta)$ between the $\delta$-, homological, and quantum gradings in Khovanov theory, and (2) the formula from \cite[Proposition~4.8]{KWZ} for the grading shift induced by reversing the orientation of a tangle component.)
Then, by the pairing theorem,
\[
\Khr\Big(T(\pm2)\Big)
\cong
\operatorname{HF}(\textcolor{red}{\a_\pm},\textcolor{blue}{C})
=
\operatorname{HF}(\textcolor{red}{\a_\pm},\textcolor{blue}{\gamma_1})
\oplus
\dots
\oplus
\operatorname{HF}(\textcolor{red}{\a_\pm},\textcolor{blue}{\gamma_m})
\oplus
\operatorname{HF}(\textcolor{red}{\a_\pm},\textcolor{blue}{\rho})
\]
The total dimensions of the first \(m\) pairs of summands are identical.
By Lemma~\ref{lem:pairing_linear_curves:dimension_formula}, and because a figure eight and an arc of the same slope intersect minimally in two points, the dimensions of the final summands are
\[\operatorname{HF}(\textcolor{red}{\a_\pm},\textcolor{blue}{\rho}) =
\begin{cases}
3\mp1,& \text{ if } s=2; \\
3\pm1,& \text{ if } s=-2; \\
|s\mp 2|,& \text{ otherwise.}
\end{cases}\]
Our assumption that \(T(+2)\cong T(-2)\) implies that the dimensions of reduced Khovanov homology agree, so the slope \(s=0\).
We now consider the quantum gradings.
Recall that the grading on $\Khr(\Lo)$ is independent of the orientation on $\Lo$.
Since $T(\infty)$ is the unknot, the quantum gradings on $\textcolor{blue}{\rho}$ and $\Khr(\Lo)\mathrel{\widehat{=}}\Big[\GGzqz{\DotB}{}{-1}{}\xrightarrow{H}\GGzqz{\DotB}{}{+1}{}\Big]$ agree.
Thus, \(\operatorname{HF}(\textcolor{red}{\a_+},\textcolor{blue}{\rho})\) and \(\operatorname{HF}(\textcolor{red}{\a_-},\textcolor{blue}{\rho})\) are graded isomorphic, because both are graded isomorphic to the reduced Khovanov homology of the same oriented Hopf link.
Moreover, for all \(i=1,\dots,m\), the quantum grading is shifted such that
\[
\operatorname{HF}(\textcolor{red}{\a_-},\textcolor{blue}{\gamma_i})
\cong
q^{+2}\operatorname{HF}(\textcolor{red}{\a_+},\textcolor{blue}{\gamma_i})
\]
which together with the previous observation contradicts \(\Khr(T(+2))\cong\Khr(T(-2))\).
The grading shift for \(\operatorname{HF}(\textcolor{red}{\a_\pm},\textcolor{blue}{\gamma_i})\) can be seen as follows: After pulling the curves \(\textcolor{blue}{\gamma_i}\) sufficiently tight, their intersections with the arcs \(\textcolor{red}{\a_\pm}\) are all in a small neighbourhood of the intersection points of \(\textcolor{red}{\a_\pm}\) with the parametrizing arcs corresponding to the generators \(\DotBred\) as shown in Figure~\ref{fig:ecsc:case_i:specials} from the viewpoint of the planar cover of $S^2_{4,\ast}$.
The arcs \(\textcolor{red}{\a_\pm}\) are parallel in this region, so the grading difference is precisely the (negative of) the grading difference between these two generators.
\begin{figure}[b]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\(\ECSCIspecials\)
\caption{}\label{fig:ecsc:case_i:specials}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\centering
\(\ECSCIIspecials\)
\caption{}\label{fig:ecsc:case_ii:specials}
\end{subfigure}
\caption{Canonical generators of \(\operatorname{HF}(\textcolor{red}{\a_\pm},\textcolor{blue}{\gamma_i})\) in the proof of Theorem~\ref{thm:intro:cosmeticrational}, (a)~Case~1 and (b)~Case~2. The arrows indicate the corresponding morphisms used for computing the gradings, as in Section~\ref{subsec:background:bigrading}. }
\label{fig:ecsc}
\end{figure}
\medskip\noindent {\bf Case 2: \(\{r,r'\}=\{\nicefrac{\pm1}{n}\}\). }
Since the connectivity of the tangle \(T\) is \(\No\), \(T(\nicefrac{1}{n})=T(\nicefrac{-1}{n})\) is a knot and up to overall orientation reversal, the orientations on \(T(\nicefrac{1}{n})\) and \(T(\nicefrac{-1}{n})\) are as follows:
\[
T(\nicefrac{1}{n})=
\Tpn
\qquad
T(\nicefrac{-1}{n})=
\Tmn
\]
Define
\begin{align*}
\textcolor{red}{\a_+}
&\mathrel{\widehat{=}}
\left[
\begin{tikzcd}[nodes={inner sep=2pt},column sep=14pt,ampersand replacement=\&, red]
\GGzqh{\DotCred}{}{-2n}{}
\arrow{r}{S}
\&
\GGzqh{\DotBred}{}{1-2n}{}
\arrow{r}{D}
\&
\GGzqz{\DotBred}{}{3-2n}{}
\arrow[dotted]{r}{S^2}
\&
\quad
\arrow[dotted]{r}{S^2}
\&
\GGzqz{\DotBred}{}{4j-3-2n}{}
\arrow{r}{D}
\&
\GGzqz{\DotBred}{}{4j-1-2n}{}
\arrow[dotted]{r}{S^2}
\&
\quad
\arrow[dotted]{r}{}
\&
\GGzqz{\DotBred}{}{-1}{}
\end{tikzcd}\right]
\text{ and}
\\
\textcolor{red}{\a_-}
&\mathrel{\widehat{=}}
\left[
\begin{tikzcd}[nodes={inner sep=2pt},column sep=14pt,ampersand replacement=\&, red]
\GGzqz{\DotBred}{}{1}{}
\arrow[dotted]{r}{}
\&
\quad
\arrow[dotted]{r}{S^2}
\&
\GGzqz{\DotBred}{}{2n+1-4j}{}
\arrow{r}{D}
\&
\GGzqz{\DotBred}{}{2n+3-4j}{}
\arrow[dotted]{r}{S^2}
\&
\quad
\arrow[dotted]{r}{S^2}
\&
\GGzqz{\DotBred}{}{2n-3}{}
\arrow{r}{D}
\&
\GGzqz{\DotBred}{}{2n-1}{}
\arrow{r}{S}
\&
\GGzqz{\DotCred}{}{2n}{}
\end{tikzcd}\right]
\end{align*}
Note that \(j=1,\dots,\lfloor\tfrac{n}{2}\rfloor\).
The arcs \(\textcolor{red}{\a_+}\) and \(\textcolor{red}{\a_-}\) are the \(\BNr\) invariants of the mirrors of the \(\nicefrac{\pm1}{n}\)-twist tangles in \(T(\nicefrac{1}{n})\) and \(T(\nicefrac{-1}{n})\), respectively; see \cite[Example~6.2, Proposition~4.8]{KWZ}. So as in Case~1, the pairing theorem allows us to write
\[
\Khr\Big(T(\nicefrac{\pm1}{n})\Big)
\cong
\operatorname{HF}(\textcolor{red}{\a_\pm},\textcolor{blue}{C})
=
\operatorname{HF}(\textcolor{red}{\a_\pm},\textcolor{blue}{\gamma_1})
\oplus
\dots
\oplus
\operatorname{HF}(\textcolor{red}{\a_\pm},\textcolor{blue}{\gamma_m})
\oplus
\operatorname{HF}(\textcolor{red}{\a_\pm},\textcolor{blue}{\rho})
\]
The total dimensions of the first \(m\) pairs of summands are identical, regardless of the slope \(s\) of the rational component \(\textcolor{blue}{\rho}\).
Since the slope \(s\) is an even integer, it never agrees with \(\nicefrac{\pm1}{n}\), so the dimensions of the final summands \(\operatorname{HF}(\textcolor{red}{\a_\pm},\textcolor{blue}{\rho})\) are equal to \(|1\mp sn|\) by Lemma~\ref{lem:pairing_linear_curves:dimension_formula} and hence only agree if \(s=0\).
We now compute quantum gradings. First, \(\operatorname{HF}(\textcolor{red}{\a_\pm},\textcolor{blue}{\rho})\) agree as absolutely bigraded homology groups, since they compute the reduced Khovanov homology of an unknot, shifted in quantum grading by the same amount. To compute the grading shifts for the first \(m\) summands, we observe that after pulling the curves \(\textcolor{blue}{\gamma_i}\) sufficiently tight (see \cite[Definition~6.1]{KWZthinness} and the discussion afterwards), the intersection points between the arcs \(\textcolor{red}{\a_\pm}\) and \(\textcolor{blue}{\gamma_i}\) sit in a small neighbourhood of the vertical line through the special marked point.
If \(n\) is even, the relevant portions of the complexes \(\textcolor{red}{\a_+}\) and \(\textcolor{red}{\a_-}\) are, respectively,
\[
\left(
\begin{tikzcd}[nodes={inner sep=2pt},column sep=14pt,ampersand replacement=\&, red]
\GGzqz{\DotBred}{}{4j-3-2n}{}
\arrow{r}{D}
\&
\GGzqz{\DotBred}{}{4j-1-2n}{}
\end{tikzcd}
\right)
\qquad\text{and}\qquad
\left(
\begin{tikzcd}[nodes={inner sep=2pt},column sep=14pt,ampersand replacement=\&, red]
\GGzqz{\DotBred}{}{2n+1-4j}{}
\arrow{r}{D}
\&
\GGzqz{\DotBred}{}{2n+3-4j}{}
\end{tikzcd}
\right)
\]
where \(j=1,\dots,\lfloor\tfrac{n}{2}\rfloor\). The corresponding curve segments are illustrated in Figure~\ref{fig:ecsc:case_ii:specials}.
They are obviously parallel, so there is a one-to-one correspondence between generators \(x_+\in \operatorname{HF}(\textcolor{red}{\a_+},\textcolor{blue}{\gamma_i})\) and generators \(x_-\in \operatorname{HF}(\textcolor{red}{\a_-},\textcolor{blue}{\gamma_i})\)
such that the quantum gradings satisfy
\[
q(x_-)-q(x_+)=8j-4-4n<0
\]
for \(j=1,\dots,\lfloor\tfrac{n}{2}\rfloor\).
If \(n\) is odd, there are additional generators stemming from the generators
\(\GGzqz{\DotBred}{}{-1}{}\) and \(\GGzqz{\DotBred}{}{+1}{}\) of the complexes \(\textcolor{red}{\a_\pm}\).
The corresponding curve segments look as in Figure~\ref{fig:ecsc:case_i:specials}, except that the quantum gradings are different. The correspondence from the case that \(n\) is even extends to the case that \(n\) is odd so that the quantum gradings of the additional generators satisfy
\[
q(x_-)-q(x_+)=-2<0
\]
The grading shifts are strictly negative in all cases, contradicting \(\Khr(T(\nicefrac{1}{n}))\cong\Khr(T(\nicefrac{-1}{n}))\).\qed
\begin{remark}
This proof highlights the utility of the quantum gradings in Khovanov homology. However, it also suggests an alternate strategy that avoids gradings in these complexes altogether through the exact triangle. In general, by using the immersed curves reformulation of Khovanov homology, one is able to split up the skein exact triangle into several summands. There is one for each component of the immersed multicurve for the tangle complementary to the crossing where the exact triangle is being implemented. For each exact triangle, the dimensions of the three groups are simply computed by a count of intersections between the component and three rational curves in the four-punctured sphere of distance one. This gives much stronger constraints on the structure of the exact triangle. Frequently, as a result, the maps in the exact triangle can be computed as well, and additional grading structures can be deduced. A similar perspective on these exact sequences is seen in bordered Heegaard Floer homology for three-manifolds with torus boundary \cite[Section 11.2]{LOT}.
\end{remark}
|
\section{Introduction}
For quantum computation to lead to new discoveries \cite{nielsen2010quantum}, efficient means of solving for the ground state must be understood and implemented. Some near term algorithms that have been used in the era of noisy quantum devices have led to an increased interest in results obtained from variational quantum eigensolvers \cite{o2016scalable,kandala2017hardware,gard2020efficient,gokhale2020n,ollitrault2020quantum,klymko2021real,kremenetski2021simulation} which ultimately face noise and other limitations \cite{bittel2021training}. Looking to the long term capabilities of a quantum computer when error correction is available, other algorithms are useful to investigate.
By far the longest studied algorithms to obtain ground states is the implementation of real time evolution. In this algorithm, an initial Hamiltonian $\mathcal{H}_0$ is defined and an initial state $\Psi_0$ is prepared on the quantum computer. The time dependent Hamiltonian is
\begin{equation}\label{RTEham}
\mathcal{H}(t)=\mathcal{H}_0+\lambda(t)\mathcal{H}_I+\mathcal{C}
\end{equation}
with interaction term $\mathcal{H}_I$, constant $\mathcal{C}$, and time dependent coupling constant (or similar form) $\lambda(t)$ at time $t$. By adiabatically (slowly) increasing the interaction term in time, the wavefunction will eventually arrive at the ground state for the fully interacting Hamiltonian.
However, this solution strategy is known to be extremely slow for quantum chemical systems \cite{poulin2014trotter,lemieux2021resource}. In order to apply the Hamiltonian, a time evolution operator of the form $\exp(-i\mathcal{H}(t)\delta t)$ must be applied to the wavefunction. The Trotter-Suzuki decomposition of the time evolution operator must be decomposed into many terms, $\mathcal{O}(N^4)$ to capture the full electron-electron interaction term, although this can be reduced as $N\rightarrow\infty$ to $\mathcal{O}(N^2)$ for the case of local basis functions \cite{baker2020density}. However, since the time step $\delta t$ must be very small depending on the strength of correlation in the system, the resulting number of operations makes the time necessary to solve for even small molecules extremely long. This is true of other classical solution techniques such as Hartree-Fock \cite{gulania2021limitations}. This is expected based on the complexity of solving quantum chemistry systems \cite{schuch2009computational}.
The question of how to obtain excited states is one that has been investigated in recent papers \cite{higgott2019variational,nakanishi2019subspace,parrish2019quantum,mizuta2021deep,kuroiwa2021penalty,greene2021generalized,motta2020determining,jahangiri2020quantum,ollitrault2020quantum,roggero2020preparation,sim2018quantum,bauman2019quantum}, and a direct solution would provide a means to fully manipulate the wavefunction. If this can be accomplished without the use of time evolutions, then avoiding the small time step necessary for those methods could be possible.
In addition to ground state solvers, obtaining excited states is a highly valuable quantity for quantum chemistry systems. Historically, excited states have been more difficult to obtain. So, if a quantum algorithm could reliably obtain the excitations, then this would represent a major improvement over existing classical techniques \cite{elliott2011perspectives,khoromskaia2015tensor,helgaker2014molecular}.
Recently, Lanczos methods have been investigated in the context of solving quantum systems. One variety of Lanczos on the quantum computer, uses even vectors of a Krylov subspace and imaginary time evolution techniques to obtain the ground state and other quantities of interest \cite{motta2020determining,sun2021quantum}.
Another recently introduced variety of Lanczos algorithms \cite{baker2021lanczos} uses oblivious amplitude amplification (OAA) \cite{kothari2014efficient,berry2015simulating,childs2017quantum}--an algorithm similar to Grover's search but run on auxiliary qubits \cite{grover1996fast,grover2001schrodinger,brassard2002quantum}--to apply operators directly onto wavefunctions. Since OAA applies operators onto wavefunctions, all the necessary operations for Lanczos can be implemented directly. The full Lanczos recursion can be implemented to find the ground state or the continued fraction representation of the Green's function \cite{baker2021lanczos}. This second Lanczos technique called quantum Lanczos recursion (QLR) avoids the use of a time evolution operator and therefore bypasses the computational bottleneck in terms of the number of terms in the Trotter-Suzuki decomposition.
One other advantages of the QLR is that the traditional limitations of Lanczos on the classical computer are entirely circumvented \cite{cahill2000numerical}, and only one number of wavefunctions. The only limitation is how accurately the operations can be applied to the wavefunction in practice, and this has recently been investigated on a quantum computer \cite{jamet2021krylov}.
This paper uses similar techniques as QLR to show that the traditional Lanczos recursion can be replaced by a block Lanczos routine to resolve several excitations, leading to quantum block Lanczos recursion (QBLR), although ``block" can be replaced by ``banded" or some other name corresponding to a trivial change of gauge of the unitary matrices involved. While it is trivially demonstrated from QLR that excitations can be found, the performance of block Lanczos will allow for the resolving of degeneracies with greater ease. This was recently demonstrated in tensor network algorithms \cite{bakermultitarget,di2021efficient}.
The algorithm here in the can also be performed with the preparation of a single initialization of starting wavefunctions. A full collapse of the eigenstates is avoided, meaning that a wavefunction can be preserved for the next computation. This is accomplished by use of a state-preserving quantum counting algorithm \cite{baker2021lanczos}, referred to in some works as QMA-sampling \cite{marriott2005quantum,temme2011quantum}.
Some additional discussion on how errors in the Lanczos coefficients will affect the ground state energy and otherwise are also considered. The method of slowly introducing terms into the Hamiltonian is also discussed in the context of Lanczos to show that the OAA algorithm does not need to scale exponentially for limited additional terms.
\section{Excitations from quantum Lanczos recursion}
The Lanczos algorithm from Ref.~\cite{baker2021lanczos} can be used to find excitations of a given model by a simple alternative of coefficients. In this section, QLR will be reviewed and an extension to finding excited states will be shown.
A Lanczos recursion relation to find subsequent elements of the Krylov subspace, $\{\psi_0,\psi_1,\ldots,\psi_N\}$ is
\begin{equation}\label{scalarlanczos}
|\psi_{n+1}\rangle=\mathcal{H}|\psi_n\rangle-\alpha_n|\psi_n\rangle-\beta_n|\psi_{n-1}\rangle
\end{equation}
where the resulting Hamiltonian in the basis of the Krylov subspace forms a tridiagonal matrix. When diagonalized, the ground state is found to a high accuracy even if only a few $n$ are determined.
\paragraph{Applying operators to wavefunctions.--}
The only ingredient that is necessary for this form of the ground state solution is to ensure that the operators could be applied to the wavefunction. The method to do this was introduced in Refs.~\cite{berry2014exponential,kothari2014efficient} where an oblivious amplitude amplification (OAA) algorithm \cite{berry2015simulating,childs2017quantum} was used to improve on repeat-until-success methods which have the same goal \cite{guerreschi2019repeat,paetznick2013repeat}. Essentially, when an operator is applied to a wavefunction as in a repeat-until-success strategy, the operation is controlled on some auxiliary qubits that give 0 if the application of an operator is successful and 1 if it is not. OAA searches for a state on some auxiliary qubits corresponding to all 0s. By searching over the auxiliary qubit states, the algorithm remains oblivious to the wavefunction itself. This essentially guarantees the outcome from the related method of repeat-until-success to apply the operator in one step \cite{guerreschi2019repeat,paetznick2013repeat}. For how to construct the operators, see the discussion in Ref.~\cite{low2019hamiltonian}.
\paragraph{Minimal measurements of the wavefunctions.--}
In order to sample the coefficients $\alpha_n$ and $\beta_n$ without completely measuring (and therefore destroying) the wavefunction, a state-preserving quantum counting algorithm can be used. The algorithm applies a generic operator $\mathcal{A}$ as in the form
\begin{equation}
\mathcal{A}|\Psi\rangle=p|\Psi\rangle+p^{\perp}|\Psi^{\perp}\rangle
\end{equation}
which signifies a superposition of eigenstates. The operator $\mathcal{A}$ is normalized and represented as a unitary such that it can be represented on the qubits \cite{baker2020density,baker2021lanczos}. The resulting probability $p$ is the expectation value of $\mathcal{A}$. All states orthogonal to the original state are marked with a $\perp$ symbol. The basis of the superposition chosen here is that of the eigenbasis because this will be the natural basis to pick for quantum phase estimation (QPE) \cite{nielsen2010quantum,low2019hamiltonian,poulin2018quantum}.
The key to finding $p$ is to count the number of transitions from $\Psi$ to $\Psi$ (wavefunction to same wavefunction) after the application of $\mathcal{A}$ by OAA or some other algorithm. To verify that the same state is obtained after this procedure, the energy of the state can be computed with QPE at the start of the algorithm and stored on a register. After the operator is applied, the energy is then computed on a separate register. The two registers are compared and represented on a single qubit, called a pointer qubit. The pointer qubit is then measured. If the energies match, one value is returned and the original wavefunction is recovered. This is counted as a ``success" in the algorithm. The ratio of successes to the total times the algorithm is run is $p$. Once $p$ is found, the expectation value of $\mathcal{A}$ can be determined.
If the pointer qubit demonstrates that the wrong state was recovered, then a recovery procedure is used to find the original state \cite{baker2021lanczos,baker2020density,temme2011quantum,paetznick2013repeat}. In essence, the unitary of all operations applied onto the wrong wavefunction and the process above is repeated until the correct wavefunction is found \cite{temme2011quantum}.
\paragraph{Operators for Lanczos recursion.--}
To find the Lanczos coefficients, the following relations hold \cite{baker2021lanczos}
\begin{equation}
\alpha_n=\langle\psi_n|\mathcal{H}|\psi_n\rangle\quad\mathrm{and}\quad\beta_n=\langle\psi_{n-1}|\mathcal{H}|\psi_n\rangle
\end{equation}
and therefore provide a means to use state-preserving quantum counting to obtain the coefficients. The map between the Krylov states and the original ground state is crucial but can be summarized as
\begin{equation}
|\psi_n\rangle=\hat G_n|\Psi\rangle
\end{equation}
where some examples of these operators are shown in Ref.~\cite{baker2021lanczos}. The coefficient $\alpha_n$ can be regarded as applying quantum counting on the state $\Psi$ with the operator $(\hat G_n^\dagger\mathcal{H}\hat G_n)$. Each operator $\hat G_n$ depends on coefficients from $\{0,1,\ldots,n-1\}$ iterations and therefore the algorithm discovers the coefficients iteratively. To find $\beta_n$, the operator $(\hat G_{n-1}^\dagger\mathcal{H}\hat G_n)$ is used instead.
\paragraph{Energies from QLR.--}
Once the Lanczos coefficients are obtained, the Hamiltonain matrix in its tridiagonal form is then known. Diagonalizing this matrix and retaining both the energies and eigenvectors makes the new ground state in terms of the Krylov basis chosen.
Defining an operator $\hat Y^{(g)}$ for an excitation $g\in\mathbb{Z}^+$, the operator for the $g$th excitation would be defined as
\begin{equation}\label{Yops}
\hat Y^{(g)}=\sum_n\gamma^{(g)}_n\hat G_n
\end{equation}
in terms of coefficients $\gamma_n$ found from the diagonalization. The summation over $n$ is over as many elements in the Krylov basis as kept.
Thus, the QLR algorithm can access ground states ($g=0$) or excited states ($g\geq1$) by simply selecting different coefficients from the diagonalization of $\mathcal{H}$. At the end of this procedure, the ground state is recovered and can be used without fully re-preparing it.
\subsection{Reduced computational complexity of applying operators}\label{complexOAA}
\paragraph{First strategy: apply interaction incrementally.--}
When applying an operator to a wavefunction, several auxiliary qubits are used. OAA is a Grover search over auxuiliary qubits, so it will scale exponentially, $\mathcal{O}(\sqrt{2^q})$ for $q$ qubits, in terms of the number of auxiliary qubits. While Grover's algorithm promises a square root speedup over classical search methods, the size of the database on the quantum computer is exponentially large in terms of the number of qubits. This calls into question whether operators can actually be applied to wavefunctions in the manner prescribed earlier in total, but the particular details of the algorithm will show that this can be done efficiently. This clearly will not present an issue for finding quantities necessary for density functional theory \cite{baker2020density}, since the required operators are small ($q=2$ for coefficients of the one-body reduced density matrix).
By applying operators from the full Hamiltonian incrementally in groups of size $D$, the cost can be reduced to $\mathcal{O}(D2^{\frac{\lceil q/D\rceil}2})$. Thus, the OAA is feasible here since the Hamiltonian can be expressed as a linear combination of unitaries, this strategy applies to any system of interest \cite{berry2015simulating}. The minimal size of a group of operators is the same as a single term in the Hamiltonian. For many-body problems, this will be $q=$ 2 or 4 terms for the quadratic single-body terms or the quartic electron-interaction terms \cite{baker2020density}.
\paragraph{Second strategy: decrease strength of interaction.--}
Note that the operator applied onto the original wavefunction $\Psi$ does not need to be an operator whose eigenvalues include $\Psi$. This means that starting from some known state $\Psi$ with associated operator $\mathcal{H}_0$,
\begin{equation}
\mathcal{H}=\mathcal{H}_0+\lambda\mathcal{H}_I
\end{equation}
can be defined as a new Hamiltonian. This form of the Hamiltonian has a similar form as the the adiabatic evolution in Eq.~\eqref{RTEham}, except that the $\lambda$ coefficient can be much larger than the step size in time evolution.
The interaction term, $\mathcal{H}_I$, can be the full interaction term or one term of that interaction. By adding the terms incrementally, the final state is closer to the initial state provided and will allow for the algorithm to find the new ground state. It is crucial that each step of this procedure begins with an eigenvector so that the QPE can be applied correctly in the quantum counting procedure.
\begin{figure}
\includegraphics[width=\columnwidth]{small_convergence_Jz=1.png
\begin{picture}(0,0)
\put(-205,35){\includegraphics[height=0.28\columnwidth]{small_difference_Jz=1.png}}
\end{picture}
\put (-175,160) {\Large $\mathbf{J_{xy}=J_z}$}
\caption{\label{incrLanczos} Convergence in energy ($E$) while adding one interaction term $\hat S^z_i\cdot\hat S^z_{i+1}$ to an XY model on a 10 site lattice (9 interaction terms to add). Two solvers are used at each step. DMRG is shown as a solid blue line and a Lanczos solver is shown as a red dashed line. The initial wavefunction for the Lanczos algorithm is taken from the previous iteration. A black dashed line shows the energy of the full Heisenberg model. An inset shows the difference between the DMRG and Lanczos solutions. Only one Lanczos iteration was used for each added term. Inset: energy differences from the exact value, $\Delta E$.
}
\end{figure}
\subsection{Demonstration on model systems}\label{demoDMRG}
A full demonstration on real systems by implementing a linear combination of unitaries will be delayed for a future work, but it will be shown here taht Lanczos can be applied iteratively according to the suggestions in the previous paragraphs.
In order to demonstrate that only a limited number of Lanczos steps can be used to solve a model that includes a small number of extra terms from the starting wavefunction's Hamiltonian, a numerical study on a 10-site XXZ model of the form
\begin{equation}
\mathcal{H}=\sum_iJ_{xy}\left(\hat S^x_i\hat S^x_{i+1}+\hat S^y_i\hat S^y_{i+1}\right)+J_z\hat S^z_i\hat S^z_{i+1}
\end{equation}
where $J_{xy}=J_z$ is the Heisenberg Hamiltonian and spin matrices $\mathbf{S}=(\hat S^x,\hat S^y,\hat S^z)=\frac12\boldsymbol{\sigma}$ (with $\hbar=1$) for the vector of Pauli matrices $\boldsymbol{\sigma}$ \cite{townsend2000modern}. To start, an XY model,
\begin{equation}\label{XYmodel}
\mathcal{H}_\mathrm{XY}=J_{xy}\sum_i\hat S^x_i\hat S^x_{i+1}+\hat S^y_i\hat S^y_{i+1},
\end{equation}
will be solved for the initial wavefunction and a single interaction term of the form $\hat S^z_i\hat S^z_{i+1}$ will be added. In all, 9 terms will be added to the Hamiltonian. Computations were made with the DMRjulia library using the density matrix renormalization group (DMRG), Lanczos, and exact diagonalization (ED) routines \cite{bakerCJP21,*baker2019m,dmrjulia1,dmrjulia}.
There are three cases of study here: $J_{xy}=J_z$ (small perturbations), $J_{xy}\ll J_z$ (large perturbations), and when the initial $\psi_0$ is far from the starting state.
\begin{figure}
\includegraphics[width=\columnwidth]{large_convergence_Jz=100.png
\begin{picture}(0,0)
\put(-200,32){\includegraphics[height=0.28\columnwidth]{large_difference_Jz=100.png}}
\end{picture}
\put (-170,160) {\Large $\mathbf{J_{xy}\ll J_z}$}
\caption{\label{bigincrLanczos} Convergence in energy ($E$) while adding a large interaction term $\hat J_zS^z_i\cdot\hat S^z_{i+1}$ with $J_{xy}\ll J_z$. The parameters are the same as FIg.~\ref{incrLanczos}. By using 4 Lanczos steps instead of 2, the accuracy is greatly improved in this case. Decreasing the incremental interaction strength, $d\lambda$ (here, 0.5 with two rounds of $N_\mathrm{Lanczos}=2$), would require more than double-precision (see text).
}
\end{figure}
\subsubsection{Small perturbations ($J_{xy}=J_z$)}
Figure~\ref{incrLanczos} shows how the energy converges to the energy of the full Heisenberg model with increasing numbers of terms. In each case, using the initial wavefunction provided from the previous iteration, only one Lanczos iteration must be run to obtain a highly accurate energy with the new interaction term included. The small difference in energies between the DMRG solution and the Lanczos solution is shown in the inset of Fig.~\ref{incrLanczos}. There is also the possibility to add in a partial term by using a small term $d\lambda$ for each of $N$ times such that $Nd\lambda=\lambda$. This would require more applications of the Lanczos algorithm with new Hamiltonians and may be useful for long-range interactions or other cases.
In this one-dimensional example, the convergence is aided by only adding one term. In realistic systems, the same strategy of adding interaction terms piecemeal will both allow the Lanczos algorithm to converge quickly and also to reduce the time spent applying the operator with OAA. Still, this should be expected in general due to the rapid convergence of Lanczos techniques.
\subsubsection{Large perturbations ($J_{xy}\ll J_z$)}
In this case, $100J_{xy}=J_z$ to simulate a large perturbation. For the example here, increasing the number of Lanczos iterations to 4 for each new interaction term added allows for the energy to be obtained to a good accuracy here. This is a large improvement over using only 2 Lanczos iterations as shown by comparing the curves in Fig.~\ref{bigincrLanczos}.
It is possible to use only 2 Lanczos updates if one adds portions of the interaction term. In this case, $J_z/2$ is added to the Hamiltonian twice with 2 Lanczos steps each (total 4 Lanczos steps). The results could be improved if more Lanczos iterations were used here, but this would require greater precision in the coefficients than is available here with a double-precision classical implementation.
Note that tests with coefficients found on the classical computer will be subject to numerical instabilities that will eventually degrade the accuracy of the resulting ground state, so a full solution with this method will generate imprecise answers. For example, when increasing the number of Lanczos iterations to 10 for each step, this can occur in this example.
\begin{figure}
\includegraphics[width=\columnwidth]{rand_convergence_Jz=1.png
\begin{picture}(0,0)
\put(-110,55){\includegraphics[height=0.28\columnwidth]{rand_difference_Jz=1.png}}
\end{picture}
\put (-190,150) {\Large random $\psi_0$}
\caption{\label{randincrLanczos} Convergence in energy ($E$) similar to Fig.~\ref{incrLanczos} except that a random starting state was used. Tests on larger systems or with more Lanczos recursion steps quickly suffer from precision errors that will not be present on the quantum computer.
}
\end{figure}
\subsubsection{Arbitrary starting wavefunction}
So far, eigenstates for the starting wavefunction were used. The natural question is whether the starting states could be replaced by some arbitrary state and how the convergence of the algorithm is affected. When implementing this type of initialization on the quantum computer, the only key is to determine the energy of the wavefunction. If using QPE to do this, then the starting state must have an associated Hamiltonian.
For demonstration purposes, the initial state of alternating spins here is
\begin{equation}\label{spinvec}
|\Psi\rangle=|\psi_0\rangle=|\uparrow \uparrow \downarrow \downarrow \downarrow \uparrow \downarrow \downarrow \uparrow \uparrow\rangle
\end{equation}
is taken as the initial state with $J_{xy}=J_z$ in the Hamiltonian. While the Hamiltonian which has an eigenstate of Eq.~\eqref{spinvec} is not known and would therefore prevent the use of this state in the quantum algorithm, it is instructive to observe the convergence in this case where the initial state is not close to the final problem to solve.
Figure~\ref{randincrLanczos} shows the convergence for this starting wavefunction. The first Lanczos iteration will be the most difficult since it must change the wavefunction the most. Subsequent iterations converge more easily to the true ground-state just as in the previous cases.
These examples demonstrate that applying operators incrementally is possible and will yield a large cost reduction on the quantum computer. The examples here were on the ground-state, but higher excited states can be found similar to the discussion around Eq.~\eqref{Yops} once the Lanczos coefficients are known. In this example, the precision that can be found from the classical simulation is restricted \cite{cahill2000numerical}. The quantum computer is only limited is not a limitation on the quantum computer assuming perfect application of the operators.
\section{Quantum block Lanczos recursion}
In principle should be able to discover all excitations from QLR as described above \cite{chepiga2017excitation}, but the extension of this method to a block or banded Lanczos algorithm is known to resolve degeneracies to a higher degree and aid convergence generally \cite{bakermultitarget}. Block Lanczos is used in quantum chemistry \cite{bai2000templates} but also in physics, particularly for dynamical mean field theory computations \cite{senechal2008introduction}. The scalar coefficients of the Lanczos recursion are extended to a matrix of coefficients.
To formulate the problem, consider a set of wavefunctions grouped as a supervectors $\mathbf{\Psi}$ is a vector of $d$ excitations,
\begin{equation}
\mathbf{\Psi}_n=\Big(|\psi_n\rangle_1,|\psi_n\rangle_2,\ldots,|\psi_n\rangle_d\Big)
\end{equation}
and therefore the task is to find matrices $\mathbf{A}$ and $\mathbf{B}$ which are of dimension $d\times d$ block such that Lanczos can be performed. This means that $d$ registers, each with a wavefunction are also available.
Extending the 3-term recursion from Eq.~\ref{scalarlanczos} to a block or banded (or other) Lanczos scheme with more terms in the recursion would be possible to prepare more than one excitation at a time \cite{senechal2008introduction}. The expanded Lanczos recursion relation appears as \cite{weikert1996block}
\begin{equation}\label{blockLanczos}
\mathbf{\Psi}_{n+1}\mathbf{B}_{n+1}=\mathcal{H}\mathbf{\Psi}_n-\mathbf{\Psi}_n\mathbf{A}_n-\mathbf{\Psi}_{n-1}\mathbf{B}_{n}^\dagger
\end{equation}
for a vector of wavefunctions $\mathbf{\Psi}$ with matrices
\begin{equation}\label{Asubmat}
\mathbf{A}_n=\mathbf{\Psi}^\dagger_n\mathcal{H}\mathbf{\Psi}_n
\end{equation}
and $\mathbf{B}_n$ defined recursively from
\begin{equation}\label{Bsubmat}
\mathbf{B}_n=\mathbf{\Psi}_{n-1}^\dagger\mathcal{H}\mathbf{\Psi}_n=\mathbf{\Psi}_{n-1}^\dagger\mathcal{H}\mathbf{\Psi}_{n-1}=\mathbf{B}_n^\dagger
\end{equation}
as can be most immediately seen from the tensor network diagrams in Ref.~\cite{bakermultitarget}. It should be understood that when applying $\mathcal{H}$ onto the wavefunction supervector that the Hamiltonian is applied onto each wavefunction in the vector. The implication of Eqs.~\ref{Asubmat} and \ref{Bsubmat} is that there are simply $d^2$ applications of QLR to obtain QBLR.
Because the coefficients can be obtained by an operator acting on a ground state of an initial Hamiltonian, this implies that the coefficients of both $\mathbf{A}$ and $\mathbf{B}$ can be recovered by performing a state-preserving quantum counting operation on each term of the matrix. This was also true in the scalar case, and this establishes that the block Lanczos case is in the RWMP category of algorithms thus avoiding repeated wavefunction preparation \cite{baker2020density}. Each coefficient must be obtained individually: one quantum counting process for each operator. However, If operations can be performed in parallel on the quantum computer, then this could improve performance to find all coefficients of a block simultaneously.
On the classical computer, the blocks can be represented as a block diagonal super-matrix representing the Hamiltonian
\begin{equation}\label{blockHam}
\boldsymbol{\mathcal{H}}=\left(\begin{array}{cccccc}
\mathbf{A}_0 & \mathbf{B}_1^\dagger & \mathbf{0} & \cdots & \mathbf{0}\\
\mathbf{B}_1 & \mathbf{A}_1 & \mathbf{B}_2^\dagger & \cdots & \mathbf{0}\\
\mathbf{0} & \mathbf{B}_2 & \mathbf{A}_2 & \ddots & \mathbf{0}\\
\vdots & \vdots & \ddots & \ddots & \mathbf{B}_n^\dagger\\
\mathbf{0} & \mathbf{0} & \mathbf{0} & \mathbf{B}_n & \mathbf{A}_n\\
\end{array}\right)
\end{equation}
which can be diagonalized to find the energies of the excitations. From the coefficients of the eigenvectors of $\boldsymbol{\mathcal{H}}$ (denoted as $\gamma_{n}$ still here too), the excitation wavefunctions can also be obtained.
In order to understand how best to apply the operators on the quantum computer, the block Lanczos equation can be written as
\begin{align}
\sum_jB^{(n+1)}_{ij}|\psi_{n+1}\rangle_j=&\mathcal{H}|\psi_n\rangle_i-\sum_jA_{ij}^{(n)}|\psi_n\rangle_j\\
&-\sum_j(B^*)_{ij}^{(n)}|\psi_{n-1}\rangle_j\nonumber
\end{align}
for a single row $i$ of the left hand side. This form makes clear that the $d$ registers containing the $d$ excitations can be acted on with the appropriate operators (sum over $j$) and added with the wavefunction on other registers with the appropriate operators. Note that the elements of $\mathbf{A}$ and $\mathbf{B}$ are stored classically and therefore the inverse of each matrix can be found. If this is done, then apply $\mathbf{B}^{-1}_{n+1}$ once the coefficients are found from Eq.~\eqref{Bsubmat}.
The operators from Eq.~\eqref{Yops} will be denoted as $\mathbf{G}_n$ for a given level and are extended from the scalar Lanczos case as $\hat G_n\rightarrow\mathbf{G}_n$. Similarly, $\psi\rightarrow\mathbf{\Psi}$ for the wavefunction. One additional operation is included $\mathbf{G}$, compared with the operators listed in Ref.~\cite{baker2021lanczos}, that of the inverse $\mathbf{B}^{-1}$ operator. Since the operators can be determined on the classical computer, the operators can simply be prepared in a slightly different way. Again, the inverse is converted to a unitary as is standard for applying operators \cite{baker2021lanczos,baker2020density,berry2015simulating,kothari2014efficient,guerreschi2019repeat}.
In all, the algorithm costs a practical amount of $d^2$ over the scalar Lanczos case. Each element of the matrix equations is simply one application of a set of equations slightly modified from QLR. Note that the effort expended to apply operators onto wavefunctions is not wasted at the end with a single measurement. The state-preserving quantum counting algorithm can be used to obtain the correct expectation values without completely measuring the wavefunction.
\subsection{Algorithm summary}
The following summarizes the previous discussion to demonstrate how to implement block Lanczos on the quantum computer.
\textbf{Quantum block Lanczos recursion: Algorithm for multiple excited-states}
\begin{enumerate}
\item A set of starting eigenfunctions $\{|\Xi\rangle\}$ is prepared on $\nu$ registers (one for each wavefunction).
\item Start a counter at $n=0$.
\item Construct the operator $\mathbf{G}_n$ representing the appropriate step from Eq.~\eqref{blockLanczos}. This operator spans several excitations.
\item $\mathbf{G}_n$ is applied onto the current state and quantum counting is used to determine the Lanczos coefficients, given by Eq.~\eqref{Asubmat} and \eqref{Asubmat}.
\item With the new coefficient(s), the algorithm returns to step 2, increments $n$, and finds the next step of coefficients.
\item Storing the coefficients classically, the block diagonal Hamiltonian from Eq.~\eqref{blockHam} can be formed and diagonalized. A set of coefficients $\gamma_n$ can then be used to determine the energies.
\end{enumerate}
The scaling of the method is no worse than the application of the operators at each step, albeit the number of times this must be run for $d$ excitations is $\mathcal{O}(d^2)$. As pointed out in Sec.~\ref{complexOAA}, the incautious application of OAA will scale exponentially, but to aid convergence and avoid large computations of operations the interaction term can be applied in smaller pieces. This drastically reduces the number of Lanzcos steps required to obtain the next wavefunction and limits the time necessary to run the OAA algorithm. It is also possible to use a repeat-until-success strategy here \cite{baker2021lanczos,guerreschi2019repeat,paetznick2013repeat}, but the guaranteed result from application of OAA probably will save time in the long run.
It is not known beforehand the coefficients of the next step, so each step must be performed iteratively. Once the coefficients are obtained with the quantum counting process, then they can be stored classically and used without obtaining them again, although the $d^2$ coefficients of a given block can be parallelized. Note also that the algorithm can be restarted at any time since the coefficients are stored classically.
The ground state energy can be checked when diagonalizing Eq.~\eqref{blockHam} on the classical computer. If the ground state energy is converged within the accuracy needed for QPE, then there are sufficient Lanczos steps run to find an accurate ground state energy.
\subsection{Error analysis}
In this section, a careful analysis of the relevant parameters of the system will show how uncertainty in the Lanczos coefficients will influence the resulting energies.
\begin{figure}
\includegraphics[width=\columnwidth]{MAE_energies.png}
\caption{Mean average error for a system of block size $d=20$ for various numbers of iterations $\{4,5,6,\ldots,20\}$. All lines are nearly the same in magnitude. The general linear trend is mostly unchanging with the input parameters. \label{mae_energies}}
\end{figure}
To investigate the effects of noise on the resulting energies, a block Hamiltonian of a particular size $a$ is created for $b$ such blocks. When $a=1$, the QLR algorithm is used. When $a>1$, QBLR is represented. A set of matrices are generated with values between $[0,1]$ to mimic the values on the operators for the quantum computation. A noise term selected from a Gaussian distribution of width $\eta$ is selected for each element.
The energies with noise $E$ are computed with respect to the true energies without noise $E_0$ to find the mean absolute error (MAE) $\bar E$ defined as
\begin{equation}
\bar E=\frac1M\sum_{i=1}^M\left|E^{(i)}-E^{(i)}_0\right|
\end{equation}
for a number of eigenvalues $M$.
The MAE is shown in Fig.~\ref{mae_energies}. As expected, as the noise decreases in each of the parameters, the accuracy of the energy eigenstates increases. In both cases, the error decreases linearly with the noise applied to each term. This is a straightforward but useful result. If the quantum counting obtains coefficients to a precision $\delta$, then the resulting energy is also obtained to an error of roughly $\delta$. The change in this trend for other parameters that could have been picked (larger block sizes, more blocks, etc.) is very stable.
What this implies is that for a QPE to obtain the correct eigenvalue, the Lanczos coefficients must be sufficiently accurate ({\it i.e.}, $\delta$ is a threshold for how accurate the coefficients will be). This is a crucial aspect to know how much precision must be obtained from the quantum counting algorithm.
\subsection{Discussion}
The advantages of many excitations being discovered is chiefly one of stability. The ability to resolve degeneracies and generate orthogonal wavefunctions that is valuable in a variety of contexts and now available for the quantum computer. That this method is rapidly convergent \cite{ye1991convergence,cullum1996arnoldi}, resolves degeneracies, and avoids repeated wavefunction preparation means that this method could be a valuable alternative to time evolution methods.
Note that unlike in Ref.~\cite{bakermultitarget}, a singular value, QR, or LQ decomposition \cite{press1992numerical} was not used. This is primarily because it was not necessary here, but it has several advantages. First, many of these algorithms on the quantum computer require a quantum memory, which may be far off in terms of development \cite{chia2020sampling,tang2018quantum}, particularly with the growing consensus that probably approximately correct (PAC) learning on the quantum computer is not likely to be the best way forward for quantum machine learning \cite{baker2020density,huang2021information}. Although, expanding the definition has been suggested to lead to a logarithmic dependence on the number of measurements for tomography in terms of the total Hilbert space size \cite{aaronson2019shadow,huang2021provably}, finding the initial state for the tomography could be accomplished with QLR or QBLR.
The goal of QBLR is to retain the accuracy of the wavefunction, which is controlled directly through the gate fidelities and other quantities on the quantum computer. The quantum advantage sought here is for an exponential reduction in memory, which is a property that quantum computers exhibit under the extended Church-Turing thesis which asks if an algorithm can be computed with significantly less memory resources over classical computing \cite{aaronson2011computational}. Lanczos techniques do have this property when applied on a quantum computer theoretically if perfect application of gates is assumed and with sufficient time to sample the coefficients. The typical errors on a classical computer \cite{cahill2000numerical,paige1976error,bai1994error,chepiga2017excitation,aaronson2011computational} do not appear since values are retained to quantum precision.
The opportunity to obtain ground states from QLR places added emphasis on the development of state-preserving quantum counting and methods of applying operators to wavefunctions such as the linear combination of unitaries (and included subroutines such as OAA) to obtain both the Green's function and ground-state.
The possibility to discover ground states or excited states deserve more consideration since these methods are free of time steps and Trotter-Suzuki decompositions, a common feature in many other proposals. The Lanczos algorithms also do not need any notion of locality to aid convergence as would be expected for a matrix product state's relationship to local Hamiltonians \cite{dmrjulia1,verstraete2006matrix,hastings2004locality}, particularly for time-evolving block decimation (TEBD) \cite{bakerCJP21}. The only instance where locality plays any role is in the writing of the application of the operators to the wavefunction, which could play a role in near-term implementation. It is not clear how many Lanczos coefficients would need to be obtained to find accurate ground states from this method for phase-estimation to give accurate results. Further discussion would require specialization to a specific problem and is deferred to a future study.
Note that extensions to a fitting function for the remaining coefficients is possible in principle \cite{viswanath2008recursion}.
\section{Non-Hermitian operators}
The previous discussion was for the block Lanczos algorithm that is formulated for the problem of a Hermitian operator. The extension of the above methods to non-Hermitian operators is possible without introducing too much additional computational cost. Instead, only one additional set of matrices (or in the case of scalar Lanczos, one additional set of coefficients) should be found. The procedure will be define for the block-matrix case, but it can be reduced to the scalar case when the block size is of size 1.
Given a non-Hermitian operator $\tilde{\mathcal{H}}$ with eigenvalue $E$, a set of two eigenvalues can be defined. One is known as the left-eigenvectors (a transpose is used even if there are complex valued entries) \cite{bai1994error}
\begin{equation}
\mathbf\Psi_L^T\tilde{\mathcal{H}}=\mathbf\Psi_L^T E
\end{equation}
and the other is for the right-eigenvectors,
\begin{equation}
\tilde{\mathcal{H}}\mathbf\Psi_R=E\mathbf\Psi_R
\end{equation}
where block matrices $\mathbf{\Psi}_L$ and $\mathbf{\Psi}_R$ have the orthogonality relation \cite{bai1994error}
\begin{equation}
\mathbf{\Psi}_L^T\mathbf{\Psi}_R=\mathbb{I}_{d\times d}
\end{equation}
with an identity matrix $\mathbb{I}_{d\times d}$ of dimension $d\times d$.
The block operator represented analogous to Eq.~\eqref{blockHam} appears in this case as
\begin{equation}
\mathbf{\check T}_n=\left(\begin{array}{cccccc}
\mathbf{A}_0 & \mathbf{C}_1 & \mathbf{0} & \cdots & \mathbf{0}\\
\mathbf{B}_1 & \mathbf{A}_1 & \mathbf{C}_2 & \cdots & \mathbf{0}\\
\mathbf{0} & \mathbf{B}_2 & \mathbf{A}_2 & \ddots & \mathbf{0}\\
\vdots & \vdots & \ddots & \ddots & \mathbf{C}_n\\
\mathbf{0} & \mathbf{0} & \mathbf{0} & \mathbf{B}_n & \mathbf{A}_n\\
\end{array}\right).
\end{equation}
There are two recursion relations for each of the left and right eigenvectors. They take the form of \cite{bai1999able,gruning2011implementation}
\begin{align}
\vec{\mathbf{\Psi}}_{L,(n)}^T\tilde{\mathcal{H}}&=\mathbf{\check T}_n\vec{\mathbf{\Psi}}_{L,(n)}^T+\mathcal{I}_n\mathbf{C}_{n+1}\mathbf{\Psi}^T_{L,(n+1)}\\
\tilde{\mathcal{H}}\vec{\mathbf{\Psi}}_R&=\vec{\mathbf{\Psi}}_R\mathbf{\check T}_n+\mathbf{\Psi}_{R,(n+1)}\mathbf{B}_{n+1}\mathcal{I}_n^T
\end{align}
where
\begin{align}
\vec{\mathbf{\Psi}}_{R,(n)}&=\left(\begin{array}{ccccc}
\mathbf{\Psi}_R^{(0)}, & \mathbf{\Psi}_R^{(1)}, & \ldots, & \mathbf{\Psi}_R^{(n)}
\end{array}\right),\\
\vec{\mathbf{\Psi}}_{L,(n)}&=\left(\begin{array}{ccccc}
\mathbf{\Psi}_L^{(0)}, & \mathbf{\Psi}_L^{(1)}, & \ldots, & \mathbf{\Psi}_L^{(n)}
\end{array}\right),\\
\mathrm{and}\quad\mathcal{I}_n&=\left(\begin{array}{cccc}
\mathbf{0}_{d\times d},&\mathbf{0}_{d\times d},&\ldots,&\mathbb{I}_{d\times d}
\end{array}\right).
\end{align}
The final term is an identity matrix to ensure that the dimension of the resulting matrix is consistent with all terms in the recursion. Only the last block of the matrix is non-zero where blocks of zero matrices, $\mathbf{0}_{d\times d}$.
Writing out the explicit terms of the recursion relation in the style of the $\mathbf{G}$ operators from earlier is lengthy, but a single operator can be written to be applied on the starting wavefunction just as for QLR \cite{baker2021lanczos}. The only additional cost is the determination of an extra block matrix, $\mathbf{C}$, and double the number of registers to represent both the left- and right-eigenvectors.
The extension of the block Lanczos algorithm to the non-Hermitian case has therefore only introduced an extra matrix $\mathbf{C}_n$ which must also be found with the other two sets of matrices. However, this does not increase the overall computational cost of the algorithm presented here. The same forms for the operators connecting the original wavefunction provided to the algorithm can be derived similarly as for the Hermitian case.
\section{Conclusion}
Lanczos recursion methods on the quantum computer were extended to solve for many excitations simultaneously. This uses a block Lanczos technique that is good at resolving degeneracies in quantum states. This comes at only a cost of the number of excitations squared sought on the quantum computer. The use of the quantum counting algorithm here allows for the wavefunction to not be collapsed at each step, cutting out a major cost of many other algorithms. Further, interaction terms can be applied in small groups to aid convergence and keep the process of applying those operators to the wavefunction less than exponentially long. The error of diagonalizing the Hamiltonian in the Krylov basis was demonstrated to scale linearly with the noise of the coefficients. The method can also be applied onto non-Hermitian operators with a moderate additional cost. Due to the feasible cost of this algorithm and the rapid convergence of Lanczos techniques, it is expected that quantum block Lanczos recursion could be an alternative to existing methods to find the ground and excited states.
\section{Acknowledgements}
The author thanks Alexandre Foley, David Sénéchal, Anirban Chowdhury, and David Poulin for enlightening conversations and to Rex Godby for hosting.
The author is grateful to the US-UK Fulbright Commission for financial support under the Fulbright U.S. Scholarship programme as hosted by the University of York. This research was undertaken in part thanks to funding from the Bureau of Education and Cultural Affairs from the United States Department of State.
This project was undertaken on the Viking Cluster, which is a high performance compute facility provided by the University of York. T.E.B. is grateful for computational support from the University of York High Performance Computing service, Viking and the Research Computing team.
|
\section{Introduction}
Ensemble learning is a process by which multiple base learners are generated and combined into one composite learner \citep{dietterich2000ensemble}.
Bagging \citep{nowlan1991evaluation}, boosting \citep{freund1997decision} (including well-known Gradient Boosting \citep{friedman2002stochastic} and Adaboost \citep{freund1997decision}), and stacking \citep{kuncheva2004combining} are classic examples of ensemble algorithms.
Such techniques have been applied to a variety of machine learning domains including text mining \citep{williams2014predicting}, recommender systems \citep{amatriain2011data}, and many others. One popular and frequently used class of ensemble algorithms are random forests \citep{breiman2001random} which combine decision trees, as base classifiers, with bagging, to provide diversity among the base classifiers. More modern examples of ensembles include ensemble adversarial training \citep{tramer2017ensemble} (an ensemble technique proposed to defend adversarial attacks), and ensemble deep auto-encoders \citep{shao2018novel} (an ensemble method for fault diagnosis of rolling bearings).
A neural network ensemble \citep{li2018research} is an ensemble that combines the individual outputs of several neural networks, and has numerous machine learning applications including image classification \citep{giacinto2001design}, face recognition \citep{huang2000pose}, and intrusion detection \citep{amini2014effective}, to name but a few. Recent work by \citet{choi2018combining} combines Long Short-Term Memory (LSTM) network ensembles with adaptive weighting to improve time series forecasting, and \citet{nguyen2018acoustic} apply nearest neighbor filters to convolutional neural network ensembles to perform acoustic scene classifications.
\subsection{Related Work}
Over the years researchers have been investigating ways to build good ensembles that outperform their component learners \citep{liu1999ensemble, dietterich2002ensemble, zhang2012ensemble, sagi2018ensemble}. Two features of ensembles which have been shown to be key contributing factors to their performance are the individual accuracies of component learners and the diversity of the learners in the ensemble \citep{kuncheva2004combining}. In particular, there is no single widely accepted metric for measuring the diversity in a given ensemble, but intuitively one can think of a diverse ensemble as one where the failures of one classifier can be compensated for by the successes of the others \citep{sharkey1997combining}, that is, to have a diverse ensemble, the component learners need to avoid making coincident errors.
There also has been a vast amount of research and development on creating diversity in ensembles \citep{kuncheva2003measures}. In particular, \cite{melville2005creating} proposed a DECORATE algorithm that creates diversity in ensembles by adding randomly generated artificial training examples to the original training data (the details of this algorithm will be introduced later). We implemented their algorithm DECORATE on several non-image data sets, and the results are saved for making future comparisons with our method.
The theorems and numerical results in this paper are inspired by the recent paper \footnote{From December 2019} \citep{icmla_paper} where the authors of that paper demonstrated that there is a subtle trade-off between the individual accuracies and diversity of an ensemble, and that studying this trade-off can lead to both theoretical and practical improvements to machine learning algorithms. In this paper we extend both the theory and results in \citet{icmla_paper}
and derive a set of algorithms, armed with rigorous theoretical bounds, that improve on the performance of state-of-the-art deep neural networks for both binary and multi-label classification problems.
In many ways, the key results of this paper are quite easy to state. We merely observe that many approaches to constructing machine learning ensembles only achieve diversity in their base learners by happenstance. For example, classic random forests \citep{breiman2001random} achieve learner diversity merely by randomly performing bootstrap sampling of their data and randomly choosing subsets of their predictors for their constituent learners. Gradient boosting \citep{friedman2002stochastic}, and similar methods \citep{freund1997decision}, improve matters by taking the errors of the current set of learners into account when constructing a new learner. \emph{Our proposed algorithm goes even further by explicitly encouraging diversity between the set of learners in the ensemble as part of the training process}, and we use this approach to train ensembles of neural networks where the individual learners already have state-of-the-art performance. These ensembles then have classification error rates that are \emph{2-6 times lower} than the error rates of the state-of-the-art learners from which they are constructed.
\subsection{Our Contribution}
Herein we develop a new and deeper understanding of the balance between individual accuracies and ensemble diversity, building on \citet{icmla_paper}, by providing a new proof of one important theorem based upon the Cauchy-Schwarz inequality \citep{steele_2004}. This new proof provides valuable insights into connections between our work, coding theory \citep{irvine2001data}, and Welch bounds \citep{datta2012geometry}. In addition, we demonstrate how this theoretical work provides methods to assess and improve the optimality of any given classification ensemble (including random forests and ensembles of deep neural networks) by relating the simple majority vote accuracy of ensembles to the learner-learner correlations (which measure ensemble diversity) and truth-learner correlations (which measure individual learner accuracies). Furthermore, we propose an algorithm to train ensembles of deep neural networks and demonstrate the effectiveness of our approach with a variety of experiments on standard benchmark data sets. In particular, using our methods one is able to start from state-of-the-art neural networks and produce ensembles that improve their accuracy beyond their current capabilities. Moreover, our training algorithm is quite efficient in that we only need to train the pretrained base neural networks, as part of the ensemble, by a small number of additional epochs to achieve good performance.
In summary, in this paper, we make the following novel contributions:
\begin{itemize}
\item We provide a more concise and illuminating proof of a theorem from the literature, and extend its applicability from binary classification to multi-label classification.
\item We develop a new theorem that provides a closed-form formula of calculating the simple majority vote accuracy of ensembles based on the statistical correlations between pairs of learners and correlations between the learners and the ground truth.
\item This new theoretical work offers a clear understanding of how individual accuracies of component learners and diversity in an ensemble affect the performance of an ensemble, and thus inspires a metric to assess the optimality of given ensembles.
\item We propose a training algorithm for deep neural network ensembles that explicitly encourages ensemble diversity by the way of maximizing our proposed metric.
\item Our algorithm is demonstrated to be more effective than related algorithms by \cite{melville2005creating} in creating diversity in neural network ensembles and maintaining the overall accuracy in the meantime.
\item Our algorithm is shown to be general and efficient in improving state-of-the-art neural networks by experiments on both non-image and image classification problems with binary and multi-label data.
\end{itemize}
\section{Accuracy-Diversity Trade-off in Ensembles from the Perspective of Statistical Correlations}
There is a large extant literature \citep{brown2005managing, brown2004diversity, brown2010good, dai2017considering, asafuddoula2017incremental, hsu2017theoretical} that has studied the relationship between the individual accuracies of component learners and diversity in ensembles. One recent work by \citet{icmla_paper} showed the balance between these two features from the perspective of Pearson statistical correlations \citep{benesty2009pearson}. In particular, the term ``averaged learner-learner correlations" denoted by $r_{LL}^{(ave)}$ was defined as follows as a measure for the overall diversity in an ensemble \citep{icmla_paper}:
\begin{equation*}
r_{LL}^{(ave)}=\frac{1}{N(N-1)/2}\sum_{i=1}^{N}\sum_{j > i}^{N}r_{L_i, L_j},
\end{equation*}
where $N$ is the ensemble size (i.e. the number of learners in the ensemble), and $r_{L_i, L_j}$ is the pairwise Pearson correlation coefficient between the predicted labels of learners $L_i$ and $L_j$. Similarly the term ``averaged truth-learner correlations" noted by $r_{TL}^{(ave)}$ was defined as follows as a measure for the overall accuracy of learners in an ensemble \citep{icmla_paper}:
\begin{equation*}
r_{TL}^{(ave)}=\frac{1}{N}\sum_{i=1}^{N}r_{T, L_i},
\end{equation*}
where $N$ is again the ensemble size, and $r_{T, L_i}$ is the Pearson correlation coefficient between the labels of ground truth $T$ and the predicted labels of learner $L_i$.
To motivate our approach we show in Figure $\ref{correlation_matrix}$ a schematic of the correlation matrix between the ground truth $T$ and three learners $L_1$, $L_2$ and $L_3$. In this schematic one can see that $r_{TL}^{(ave)}$ and $r_{LL}^{(ave)}$ are calculated by taking the average of the blue and orange elements, respectively.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.25\linewidth]{figures/correlation_matrix.png}
\caption{Correlation matrix for the ground truth $T$ and three learners $L_1$, $L_2$ and $L_3$. The average of the blue elements is the averaged truth-learner correlations $r_{TL}^{(ave)}$, and the average of the orange elements is the averaged learner-learner correlations $r_{LL}^{(ave)}$.}
\label{correlation_matrix}
\end{figure}
As discussed earlier, the component learners in a diverse ensemble should not be too similar. Therefore small, or even negative, averaged learner-learner correlations are preferred. Hence a lower value of $r_{LL}^{(ave)}$ indicates a higher level of diversity in the ensemble. Similariy, since $r_{T, L_i}$ measures the correlation between a learner and the truth (which will be discussed in detail in Section 3), the higher the value of $r_{TL}^{(ave)}$ is, the higher the overall accuracy of the learners in the ensemble is.
Accordingly, one may be tempted to consider the left correlation matrix in Figure $\ref{examples_correlation_matrix}$ as optimal for the three-learner cases, as its $r_{LL}^{(ave)}$ has reached the lowest possible value $-1$ and its $r_{TL}^{(ave)}$ has reached the highest possible value $1$. However, \emph{no such correlation matrix can exist}, as it is not non-negative definite and thus is not a legitimate correlation matrix. However, the right correlation matrix in Figure $\ref{examples_correlation_matrix}$ is valid with $r_{TL}^{(ave)}=0.3$ and $r_{LL}^{(ave)}=-0.2$ since all of its eigenvalues are greater than or equal to $0$.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.53\linewidth]{figures/two_correlation_examples.png}
\caption{Two examples of a correlation matrix for the ground truth $T$ and three learners $L_1, L_2$ and $L_3$. The left one is not a valid correlation matrix, as it is not non-negative definite, while the right one is valid. These two examples show the fact that not all values between $-1$ and $1$ are possible for the averaged truth-learner correlations and the averaged learner-learner correlations.}
\label{examples_correlation_matrix}
\end{figure}
Generalizing these ideas by studying the eigenstructure of correlation matrices leads to the two theorems below, originally proven in \citet{icmla_paper}, which provide sharp bounds on what values of $r_{TL}^{(ave)}$ and $r_{LL}^{(ave)}$ are possible for ensembles of different sizes.
\begin{theorem}(Originally from \citet{icmla_paper} where it appears as Theorem 1)
\label{thm1}
For an ensemble with $N$ learners we have that
\begin{equation}
-\frac{1}{N-1}\leqslant r_{LL}^{(ave)} \leqslant 1.
\label{theorem1}
\end{equation}
\end{theorem}
\newpage
\begin{theorem}\label{thm2} (Originally from \citet{icmla_paper} where it appears as Theorem 2)
For an ensemble with $N$ learners we have that
\begin{equation}
-\sqrt{\frac{(N-1)\cdot r_{LL}^{(ave)}+1}{N}}\leqslant r_{TL}^{(ave)} \leqslant \sqrt{\frac{(N-1)\cdot r_{LL}^{(ave)}+1}{N}}.
\label{theorem2}
\end{equation}
\end{theorem}
Theorem $\ref{thm1}$ shows that there is a limitation to the extent of possible negative learner-learner correlations in an ensemble. In particular, as the ensemble size $N$ increases, the lower bound $-\frac{1}{N-1}$ in $\eqref{theorem1}$ approaches $0$ from the negative side, meaning that larger ensembles cannot have large negative values of $r_{LL}^{(ave)}$ as compared to smaller ensembles. Along these same lines, Theorem $\ref{thm2}$ provides the tight bounds for $r_{TL}^{(ave)}$ as functions of $r_{LL}^{(ave)}$.
The original proof in \citet{icmla_paper} to Theorem $\ref{thm2}$ followed the idea of Sylvester's criterion \citep{gilbert1991positive}. Herein we propose a new and shorter proof (in Appendix A.) by applying the Cauchy-Schwarz inequality. This new, more geometric proof, provides connections between these results and classic ideas in coding theory \citep{irvine2001data} such as Welch bounds \citep{datta2012geometry} and equiangular tight frames \citep{sustik2007existence}.
Perhaps more importantly for the work discussed here, one can combine the two theorems to visualize the relationship between $r_{LL}^{(ave)}$ and $r_{TL}^{(ave)}$. Figure $\ref{theorem}$ shows an example of this relationship for the cases of an ensemble of size $5$, and ensembles of different sizes will have the similar pattern. In particular, any ensemble will fall within the region bounded by the theoretical upper bound curve (colored in red) and the lower bound curve (colored in blue). An ensemble that is expected to show acceptable performance should at least fall above the black dotted line, where the averaged truth-learner correlations is $0$ (meaning that the component learners are making random guesses on the labels of data instances), as the Condorcet's jury theorem \citep{boland1989majority} states that the individual probabilities being correct must be greater than $1/2$ to ensure the success of the group decision under the rule of simple majority vote.
The red curve in Figure $\ref{theorem}$ can be thought of an optimal boundary, similar to a Pareto boundary \citep{sen1993markets}. In particular, any ensemble that is strictly below the red line can have either its individual accuracy or diversity improved at no cost to the other quantity.
For example, the two points on Figure $\ref{theorem}$ are examples of ensembles that are close to optimal, where the component learners in the ensemble at point A are very accurate but not diverse, while the learner in the ensemble at point B are very diverse but not that accurate.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.6\linewidth]{figures/theorem.png}
\caption{A theoretical plot (averaged truth-learner correlations versus averaged learner-learner correlations) for ensembles of size 5. Ensembles of different sizes have the similar pattern. Any ensemble will fall within the region bounded by the red and blue curve, and an acceptable ensemble should fall at least above the black dotted line. Ensembles that fall near and on the red curve are optimal, and points A and B are two examples of ensembles that are close to optimal.}
\label{theorem}
\end{figure}
\section{Theory for Assessing and Improving the Performance of Ensembles}
In Section 2, we discussed the balance between individual accuracies and diversity of learners in ensembles. In this section, we will demonstrate how exactly the two features affect the performance of ensembles for classification problems. In particular, the simple majority vote accuracy of ensembles will be used as the metric for assessing their performance, and we aim to develop a closed-form formula of calculating simple majority vote accuracy based on the averaged truth-learner correlations $r_{TL}^{(ave)}$ and the averaged learner-learner correlations $r_{LL}^{(ave)}$.
For now, we only derive the theory for simple majority vote accuracy for homogenous ensembles (which will be defined in Section 3.2). However, the lessons learned from the theory for homogeneous ensembles inspire us on developing approaches that are still effective for general ensembles, though without the same theoretical guarantees. In Section 4, we will demonstrate how such ideas lead to our training algorithm for general neural network ensembles.
\subsection{Relationship between the Accuracy of a Learner and its Correlation with the Truth}
Before investigating the theory for assessing the performance of ensembles, it is worthwhile to derive the mathematical formulation of the individual accuracy of a learner in terms of its correlation with the truth, as such formulation is critical to the formulation of simple majority vote accuracy of ensembles (which we will detail in Section 3.2).
For simplicity, consider a data set with two classes (the labels are indicated by $0$'s and $1$'s). The mathematical relationship for binary classification problems between the accuracy of a learner $L$ and its correlation with the true class labels $T$ can be characterized by the following theorem (Proof in Appendix B.). Note that extending the theorem to multi-label classification is one important direction for future work.
\begin{theorem}\label{thm3}
\begin{equation}
r_{T, L}=\frac{\beta p - \alpha (2\beta p - p+1-\alpha)}{\sqrt{\alpha(1-\alpha)(2\beta p - p+1-\alpha)(-2\beta p +p+\alpha))}}\label{theorem3},
\end{equation}
where
\begin{itemize}
\item the learner $L$ and the truth $T$ are both vectors of $0$'s and $1$'s,
\item $r_{T, L}$ is the Pearson correlation coefficient between $L$ and $T$,
\item $p$ is the accuracy of learner $L$ (i.e. the proportion of data points correctly classified by learner $L$),
\item $\alpha$ is the proportion of $1$'s in the truth class labels $T$ (note that $0<\alpha<1$ \footnote{Note that $\alpha$ cannot be 0 or 1, as correlation is not defined for constants (when true class labels are all $0$'s or all $1$'s).}, and for balanced data sets, $\alpha=\frac{1}{2}$),
\item $\beta$ is the ratio of accuracy in class $1$ to accuracy in both classes (i.e. $\beta=P(T=1\cap L=1)/(P(T=1\cap L=1)+P(T=0\cap L=0))=P(T=1\cap L=1)/p$, and $P(.)$ represents probability).\end{itemize}
Proof: In the Appendix.
\end{theorem}
The relationship between $p$ and $r_{T, L}$ appears to be complicated in \eqref{theorem3}. However one can solve $p$ from \eqref{theorem3} by writing out the Taylor series expansion \citep{dienes1957taylor} of $p$ at $\beta=\frac{1}{2}$ as shown in \eqref{taylorexpansion}:
\begin{align}
p= & 2\alpha(1-\alpha)(1+r_{T, L})-4(\alpha-2\alpha^2+2\alpha^3 \nonumber \\
& +2\alpha r_{T, L}-6\alpha^2 r_{T, L}+4\alpha^3 r_{T, L}+\alpha r_{T, L}^2-3\alpha^2 r_{T, L}^2 \nonumber \\
& +2\alpha^3 r_{T, L}^2)(\beta-\frac{1}{2})+O((\beta-\frac{1}{2})^2),\label{taylorexpansion}
\end{align}
\noindent where $O((\beta-\frac{1}{2})^2)$ represents all the higher-order (including the second-order) terms of $(\beta-\frac{1}{2})$.
\eqref{taylorexpansion} can be further simplified as
\begin{equation}
\begin{aligned}
p & =2\alpha(1-\alpha)(1+r_{T, L}) \indent (\text{Under the assumption of } \beta=1/2) \\
& =2\alpha(1-\alpha)+2\alpha(1-\alpha)r_{T, L}.
\end{aligned}
\label{linear}
\end{equation}
Based on \eqref{linear}, we can see that as long as the learner $L$ performs equally well on both classes (i.e., $\beta=1/2$) then the accuracy of the learner, $p$, is a linear function of its correlation with the truth $r_{T, L}$. Notice that the intercept and slope in \eqref{linear} share the same function, $2\alpha(1-\alpha)$. Since $0<\alpha<1$, the slope has $2\alpha(1-\alpha)>0$. Therefore, for any learner that is expected to perform equally well on both classes, the individual accuracy $p$ is positively linearly correlated with its correlation with the truth. Hence, the higher the value of correlation $r_{T, L}$ is, the more accurate the learner is.
For a balanced binary data set (i.e. $\alpha=1/2$), \eqref{linear} can be further reduced to
\begin{equation}
\label{linear_simpler}
p=\frac{1}{2}+\frac{1}{2}\cdot r_{T, L}. \indent (\text{Under assumptions of }\beta=1/2, \alpha=1/2)
\end{equation}
Figure \ref{taylor} visualizes the relationship between accuracy of a learner (the y-axis) and its correlation with the truth (the x-axis) for $2,000$ randomly generated pairs of true class labels $T$ and learner $L$, when no assumptions about $\alpha$ and $\beta$ are made. For each of the $2,000$ pairs, the correlation $r_{T, L}$ and accuracy $p$ are calculated, and marked as a blue dot in Figure \ref{taylor}. All the blue dots show a roughly linear trend (colored in red), where the accuracy is increasing when the correlation increases. There exists some fluctuation of the blue dots around the linear trend, and such fluctuation is due to the higher order terms brought by parameters $\alpha$ and $\beta$ to \eqref{theorem3}.
\begin{figure}[h]
\centerline{\includegraphics[width=0.6\textwidth]{figures/icdm_taylor_expansion.png}}
\caption{Accuracy of a Learner versus its Correlation with the Truth. Each of the 2000 blue points is generated by a pair of randomly generated binary classifier and true class labels of size 100 with two classes. All the blue points show a roughly linear trend as indicated by the red straight line.}
\label{taylor}
\end{figure}
In conclusion, the fact that a learner’s accuracy is linearly positively correlated with its correlation with the truth is rigorously true if the learner is assumed to perform equally well on predicting both class labels, and is empirically observed even when the assumption is not met. Hence it is reasonable for us to take the truth-learner correlation $r_{T, L}$ as a measure for a learner's individual accuracy. Extending this idea to ensembles, it is also reasonable to take the averaged truth-learner correlations $r_{TL}^{(ave)}$ as a measure for the overall accuracy level of learners in an ensemble.
\subsection{Simple Majority Vote Accuracy of Homogenous Ensembles}
In this section, we will derive the closed-form formula of calculating simple majority vote accuracy of ensembles. In particular, we focus on the concept of ``homogenous ensemble" which is an extension to ensemble learning of the jury design mentioned in \citet{kaniovski2011optimal}, where the authors defined ``homogeneous jury" as a jury in which each vote has an equal probability of being correct, and each pair of votes correlates with the same correlation coefficient. Similarly, we define the ``homogenous ensemble" as an ensemble whose component learners correlate with the true class labels with the same correlation coefficient, and all pairwise learner-learner correlations are equal.
Consider a homogenous jury under the simple majority vote decision rule, a closed-form formula for calculating the probability that the jury being correct (denoted by $M_n(p,c)$) is derived in \citet{kaniovski2011optimal} as follows
\begin{equation}
\label{eq-reference}
M_n(p,c)=\sum_{i=\frac{n+1}{2}}^{n}C_n^i p^i (1-p)^{n-i}+0.5c(n-1)(0.5-p) \frac{p^{\frac{n+1}{2}-1}(1-p)^{\frac{n+1}{2}-1}}{B(\frac{n+1}{2},\, \frac{n+1}{2})},
\end{equation}
where $n$ is the number of jurors in the jury, $p$ is the probability of each juror being correct, $c$ is the correlation coefficient of each pair of votes, and $B(.\ , .)$ is Euler's Beta function \citep{chaudhry1997extension}.
Thinking each juror in the homogenous jury as a learner in a homogenous ensemble, inspired by $\eqref{eq-reference}$, and replacing $p$ in $\eqref{eq-reference}$ with $2\alpha(1-\alpha)(1+r_{T, L})$ as derived in $\eqref{linear}$, we are able to derive the formula for calculating the simple majority vote accuracy of ensembles (denoted by $M_{N}(r_{T, L}, r_{L, L})$) as follows
\begin{equation}
\label{eq-majority-vote}
\begin{aligned}
\hspace{0.3in}M_{N}(r_{T, L}, r_{L, L}) & \\
& \hspace{-1.3in}=\sum_{i=\frac{N+1}{2}}^{N}C_N^i (2\alpha(1-\alpha)(1+r_{T, L}))^i (1-2\alpha(1-\alpha)(1+r_{T, L}))^{N-i}+0.5\cdot r_{L, L}(N-1) \\
& \hspace{-1.3in}\cdot(0.5-2\alpha(1-\alpha)(1+r_{T, L}))\cdot \frac{(2\alpha(1-\alpha)(1+r_{T, L}))^{\frac{N-1}{2}}\cdot (1-2\alpha(1-\alpha)(1+r_{T, L}))^{\frac{N-1}{2}}} {B(\frac{N+1}{2},\, \frac{N+1}{2})}, \\
& \hspace{-1.3in} (\text{Under the assumption of }\beta=1/2)
\end{aligned}
\end{equation}
where $N$ is the ensemble size, $r_{T, L}$ is the truth-learner correlation, $r_{L, L}$ is the Pearson correlation coefficient between each pair of learners, and $\alpha$ is the proportion of $1$’s in the truth class labels $T$ of $0$'s and $1$'s.
In a given homogenous ensemble with binary classifiers that are expected to perform equally well on predicting both classes (i.e. $\beta=1/2$), $\alpha$ can be easily calculated based on the given data set. In particular, $r_{T, L}$ and $r_{L, L}$ in $\eqref{eq-majority-vote}$ are computable from outputs of the base learners, hence one is able to estimate the simple majority vote accuracy of the ensemble through the formula given by $\eqref{eq-majority-vote}$. Even when the assumption $\beta=1/2$ is not met, $\eqref{eq-majority-vote}$ still offers a reference for evaluating the simple majority vote accuracy of homogenous ensembles.
For balanced binary data sets (i.e. $\alpha=1/2$), $\eqref{eq-majority-vote}$ can be further simplified as
\begin{equation}
\begin{aligned}
M_{N}(r_{T, L}, r_{L, L}) & =\sum_{i=\frac{N+1}{2}}^{N}C_N^i (0.5(1+r_{T, L}))^i (1-0.5(1+r_{T, L}))^{N-i} \\
& -0.25\cdot \frac{r_{T, L} r_{L, L}(N-1)}{B(\frac{N+1}{2},\, \frac{N+1}{2})}(0.5(1+r_{T, L}))^{\frac{N-1}{2}}\cdot (1-0.5(1+r_{T, L}))^{\frac{N-1}{2}}. \\
& (\text{Under the assumptions of }\alpha=1/2,\ \beta=1/2)
\end{aligned}
\label{eq-majority-vote-simplified}
\end{equation}
For fixed values of $r_{L, L}$ in a homogenous ensemble of size $N$, $\eqref{eq-majority-vote-simplified}$ is a function of the single variable $r_{T, L}$, therefore we can visualize the simple majority vote accuracy of the ensembles when $r_{T, L}$ is varying in $[0, 1]$. Figure $\ref{majority_vote}$ is one such example where ensemble size is 5, and the learner-learner correlation $r_{L, L}$ is set at six different levels (as shown in the legend box). Homogenous ensembles of different sizes have the similar pattern.
We can see that from Figure $\ref{majority_vote}$, for fixed level of truth-learner correlation $r_{T, L}$, homogenous ensembles with the lowest possible value of $r_{L, L}$ (indicated by the blue curve) show the highest simple majority accuracy. Moreover, the lower the value of $r_{L, L}$ (that is, the more diverse the ensembles), the higher the majority vote accuracy, which emphasizes the point that one should make the homogenous ensemble more diverse in order to achieve better simple majority vote accuracy, especially in cases that there is not much improvement that might be possible for the individual accuracies of the component learners. On the other hand, for a fixed level of $r_{L, L}$, the value of simple majority vote accuracy increases as the truth-learner correlation increases, meaning that in order to improve the simple majority vote accuracy of homogenous ensembles, one should try to make the component learners more accurate, in cases where the ensemble cannot be made more diverse.
\begin{figure}[h]
\centerline{\includegraphics[width=0.63\textwidth]{figures/kdd_majority_vote.png}}
\caption{Simple Majority Vote Accuracy for Homogenous Ensembles of Size 5. Homogenous ensembles of different sizes have the similar pattern. The learner-learner correlation is fixed at six different levels as shown in the legend box, while the truth-learner correlation is varying in $[0, 1]$.}
\label{majority_vote}
\end{figure}
In summary, one can assess the performance of any given ensemble by evaluating its simple majority vote accuracy. Although the formula $\eqref{eq-majority-vote}$ for calculating simple majority vote accuracy is only valid for homogenous ensembles with certain assumptions, the idea is still useful when we develop approaches for general ensembles in Section 4. More importantly, inspired by the insights obtained from Figure $\ref{majority_vote}$, one may improve the performance of any given ensemble by either making the component learners more accurate, or making them more diverse.
\section{Creating Diversity in Ensembles}
In this section, we will apply the theoretical analysis on ensembles from our earlier sections to practical classification problems. In particular, Theorem \ref{theorem1} and Theorem \ref{theorem2} are general in that they apply to any ensembles whatsoever, and Theorem \ref{theorem3} suggests that $r_{TL}^{(ave)}$ and ensemble accuracy are highly dependent. Accordingly, in this section we study the importance of and ways to create diversity in random forests and deep neural network ensembles from the perspective of statistical correlations ($r_{TL}^{(ave)}$ and $r_{LL}^{(ave)}$).
\cite{melville2005creating} proposed a DECORATE algorithm that creates diversity in ensembles by adding randomly generated artificial training examples to original training data, however, their algorithm is computationally expensive especially for large data sets. We thereafter propose a novel training algorithm following the idea of our theorems introduced earlier, and demonstrate the effectiveness and advantages of our approach in the case of deep neural network ensembles by experimental evaluations on both non-image and image classification problems.
\subsection{Creating Diversity in Random Forests}
Based on the analysis in earlier sections, we see that diversity plays an important role in the success of ensemble methods. This leads naturally to the question about how to introduce more diversity into ensembles. Herein, we start the analysis and experiments with random forests, as they represent a well-known ensemble method that uses decision trees as component classifiers and is easy to train.
One possible way to introduce diversity into random forests is varying the training parameters of the component decisions trees. Table \ref{random_forest} displays the three types of random forests we generated for the purpose of comparisons. In particular, an original random forest with $5$ trees is trained with default parameters, and is used as our baseline model. The second type is ``feature random forest", where each of the 5 component decision trees is trained with a random subset of features. And the third type is ``depth random forest", where the maximum depth of each tree is specified as a given value. One can imagine generating more types of random forests by varying the other training parameters.
\begin{table}[htbp]
\renewcommand{\arraystretch}{1.3}
\caption{Three Types of Random Forests}
\label{random_forest}
\centering
\begin{tabular}{cp{12cm}}
\hline
\hline
Original RF & Original random forest model with $5$ trees using bootstrap sampling of the data for diversity. This is our baseline random forest model for comparison. \\
\hline
Feature RF & Random forest with $5$ trees, but each tree is trained on a random subset of $m$ features. For the $5$ ``feature random forests" displayed in Figure $\ref{fig_random_forest}$, $m=1, 3, 5, 7, 20$, respectively. \\
\hline
Depth RF & Random forest with $5$ trees with a maximum depth of $d$. For the $5$ ``depth random forests" displayed in Figure $\ref{fig_random_forest}$, $d=3, 5, 7, 9, 11$, respectively. \\
\hline
\hline
\end{tabular}
\end{table}
We apply the three types of random forests on the train set of the widely studied CIFAR-10 image data \citep{krizhevsky2009learning} (note: the training set consists of $50,000$ $32\times32$ color images in $10$ classes including ``car", ``plane", ``ship", etc., with 5,000 images per class) to classify the images, and visualize their calculated metrics ($r_{TL}^{(ave)}$, $r_{LL}^{(ave)}$ and the simple majority vote accuracy) on Figure \ref{fig_random_forest}, where $1$ ``original rf", $5$ ``feature rf" and $5$ ``depth rf" are shown. The color bar is the simple majority vote accuracy, with the more yellow the points, the higher the accuracy of the forests.
\begin{figure}[htbp]
\centerline{\includegraphics[width=0.65\textwidth]{figures/aistats_random_forest.png}}
\caption{Three Types of Random Forests Applied on CIFAR-10 train set (check Table $\ref{random_forest}$ for details). There is one original random forest (square shaped), five ``feature rf" (triangle shaped), and five ``depth rf" (round shaped). The color bar is the simple majority vote accuracy, with the more yellow the points, the higher the accuracy of the forests.}
\label{fig_random_forest}
\end{figure}
We can see from Figure \ref{fig_random_forest}, the ``original rf" has the highest accuracy among all the random forests models, as it is the one closest to the upper bound curve. For ``feature rf" and ``depth rf", the closer they are to the upper bound curve, the higher their simple majority accuracy. ``feature rf" and ``depth rf" are created for the purpose of introducing diversity and outperforming the `'`original rf", but as we can see, the diversity provided by bootstrapping the data is of higher quality than the other types of diversity.
Based on the analysis on random forests, we can see that for ensembles consisting of traditional classification models, even though we can tune their training parameters, it is not guaranteed that such tuning will necessarily improve the performance of the ensembles. However, the theorems we provide here, and our measure of diversity, illuminates the underlying causes of the superior performance of some of the ensembles.
\subsection{Creating diversity in ensembles using artificial data} \label{decorate_section}
To introduce diversity intentionally to ensembles, \cite{melville2005creating} proposed an algorithm DECORATE (Diverse Ensemble Creation by Oppositional Relabeling of Artificial Training Examples), that creates diverse ensembles by adding randomly generated artificial training examples to the original training data. The ensemble is initialized with one existing ``strong" classifier, and then at each iteration a new classifier is trained on the combination of the original training set and a set of artificially generated data (the size of the artificial data is $r$ times the original training size, where $r \in [0,1]$) \citep{melville2005creating}. For numeric attributes, the new artificial data is randomly generated from a Gaussian distribution of which the mean and standard deviation are computed from the original training set; for nominal attributes, the distribution of the artificial data is obtained by applying Laplace smoothing on the probabilities of occurrences of all the distinct values \citep{melville2005creating}. The labels of the artificial training examples are then determined in a way such that the probability of the selection is inversely proportional to the predictions of the current ensemble \citep{melville2005creating}. To maintain training accuracy, at each iteration a new classifier will be rejected if adding it to the current ensemble will decrease its accuracy, and iterations can be repeated until an ensemble of desired size or the pre-fixed maximum number of iterations has been reached \citep{melville2005creating}.
The idea of the DECORATE algorithm is straightforward while attempting to increase diversity in ensembles, by forcing the new classifiers to differ from the current ensemble. However,
a few points with the algorithm may raise doubts on the effectiveness of this method. First of all, their assumptions on the distribution (Gaussian for numeric attributes) of new generated artificial data need further careful discussions. Secondly, at each iteration, a new artificial training set needs to be generated and the probabilities need to be inverted, which is time-consuming and computationally expensive, especially when the original training size is large. Last but not least, whether its success on training data can be generalized to testing data is questionable, as the ensemble diversity is created based on the combination of the artificial and original training data, not purely on the original training data.
In Section \ref{neural network ensemble}, we develop a different way of creating diversity in neural network ensembles, which shows better performance than the DECORATE algorithm.
\subsection{Creating diversity in Neural Network Ensembles} \label{neural network ensemble}
In this section, we will continue working on relating our theoretical results about accuracy and diversity to real-world ensembles. In particular, we propose a training algorithm for deep neural network ensembles, apply the algorithm to a variety of standard benchmark data sets, and demonstrate the effectiveness of the algorithm on both non-image and image classification problems.
\subsubsection{\textbf{Methodology}}
One advantage of neural network ensembles over other traditional ensembles (e.g. random forests) is that we can explicitly control the trade-off between base learner accuracy and diversity in the loss function we use to train the networks through the process of backpropagation \citep{hecht1992theory}. The loss function we proposed, inspired by \citet{icmla_paper}, is one that takes both the accuracy and diversity of neural network ensembles into consideration.
\begin{equation}
\label{eq-loss}
Loss=-(r_{TL}^{(ave)}-\lambda\cdot r_{LL}^{(ave)}),
\end{equation}
where $r_{TL}^{(ave)}$ and $r_{LL}^{(ave)}$ is taken as a measure for overall accuracy and diversity of ensembles, respectively.
Given the above loss function, we propose a training algorithm for neural network ensembles that works for both binary and multi-label classifications. The details of our proposed \textbf{training algorithm} are as follows\\
\newpage
\noindent \framebox[\linewidth][l]{
\begin{minipage}{\linewidth}
\begin{itemize}[leftmargin=0pt,label={}]
\item Input: $X\in R^{n\times q}$, $Y \in R^{n\times m}$, where $n$ is the number of instances, $q$ is the number of\\ features, $m$ ($m>1$) is the number of classes, $Y$ is the true class labels after one-hot \\encoding.
\item \setlength\itemindent{20pt} For epoch in range(num(epochs)):
\item optimizer.zero\_grad( )
\item $O$=[\ [ ] for $j$ in range (ensemble\_size)\ ]
\item \setlength\itemindent{40pt} for $j$ in range(ensemble\_size):
\item \setlength\itemindent{20pt} $O[j]$=torch.softmax(nets$[j]$($X$), dim=$1$)
\item $r_{TL}=0, r_{LL}=0$
\item \setlength\itemindent{40pt} for $k$ in range($m$):
\item \setlength\itemindent{60pt} for $j$ in range(ensemble\_size):
\item \setlength\itemindent{80pt} $r_{TL}+=Corr(Y[: , k], O[j][:, k])$
\item \setlength\itemindent{100pt} for $i$ in range(ensemble\_size):
\item \setlength\itemindent{110pt} if $i<j$:
\item \setlength\itemindent{20pt} $r_{LL}+=Corr(O[i][: ,k], O[j][:, k] )$
\item loss=$-(r_{TL}-\lambda\cdot r_{LL})$
\item loss.backward( )
\item optimizer.step( )
\end{itemize}
\end{minipage}
}\\
In particular, the major steps of the training algorithm for a data set with $m$ classes are:\\
\noindent \textbf{Step 1.} Obtain the one-hot encoded \citep{harris2010digital} true class labels $Y$. \\
\noindent \textbf{Step 2.} Obtain all the corresponding soft outputs $O$ produced from each of the neural networks in the ensemble of size $N$, where each column of $O$ represents the probability for instances to fall into that category, and each row sum of $O$ equals the total probability $1$.\\
\noindent \textbf{Step 3.} Compute all column-wise Pearson correlations between the network outputs $O$ and the true class labels $Y$, for each of the network in the ensemble. Take the overall sum of these column-wise correlations as a measure of truth-learner correlations $r_{TL}$ in the neural network ensemble. \\
\noindent \textbf{Step 4.} Similarly compute all column-wise Pearson correlations between the outputs $O$, for each pair of networks in the ensemble. Take the overall sum of these column-wise correlations as a measure of learner-learner correlations $r_{LL}$ in the neural network ensemble.\\
\noindent \textbf{Step 5.} The neural network ensemble now can be trained with the loss function
\begin{equation}
\label{new_loss}
Loss=-(r_{TL}-\lambda\cdot r_{LL}),
\end{equation}
where $r_{TL}$ (calculated by step 3) is a measure for the learner accuracy in the neural network ensemble (with the higher $r_{TL}$ is, the higher the learner accuracy is), and $r_{LL}$ (calculated by step 4) is a measure for the diversity (with the lower $r_{LL}$ is, the more diverse the ensemble is).
Notice that in the loss function $\eqref{new_loss}$, the outer-most minus sign is introduced, as we aim to minimize the training loss by maximizing $r_{TL}$ and in the meantime minimizing $r_{LL}$. Backpropagation can easily be performed with this loss function, as all the parts in $\eqref{new_loss}$ are smooth and differentiable. More importantly, the parameter $\lambda$ ($\lambda>0$) controls the diversity level imposed on the neural network ensemble, with the larger $\lambda$ is, the more diverse the networks in the ensemble will be.
\emph{The ability to explicitly control the level of diversity is a significant advantage of our training algorithm, as opposed to other current published work regarding neural network ensembles that merely allow diversity to occur by happenstance. Another advantage of our training algorithm is that it is generally applicable to neural network ensembles of any size and any network architecture, as all parts in the loss function $\eqref{new_loss}$ are computable with no restrictions on neural network architecture.}
\subsubsection{\textbf{Experiments on Non-Image Data}} \label{non-image}
In order to show the effectiveness of our proposed training algorithm on non-image data classification problems, perhaps more importantly, to show the advantage of our algorithm over the DECORATE algorithm \citep{melville2005creating} introduced in section \ref{decorate_section}, herein we run experiments on three UCI non-image data sets \citep{Dua:2019} used by \cite{melville2005creating}: Breast Cancer Wisconsin (Original) Data Set, Iris Data Set, and Image Segmentation Data Set. The detailed information of the three data sets is summarized in Table \ref{data_summary}.
We compared the performance of the neural network ensemble trained with our algorithm (with $\lambda=0.9$), the neural network ensemble trained with DECORATE, and the original J48-based DECORATE ensemble used in \cite{melville2005creating} \footnote{The Python code for the experiments in Section \ref{non-image} can be found on \url{https://bitbucket.org/wli5/neural-network-ensembles-theory-training-and-the-importance-of/src/master/}.}. Note that in \cite{melville2005creating} the DECORATE algorithm has not been applied to neural networks, therefore herein we apply it to neural network ensembles, which will be compared to the neural network ensembles trained with our algorithm. To have a fair comparison, the three ensemble methods share the same ensemble size $15$, and for DECORATE methods, the amount of artificial data generated is set to be equal to the size of the original training set. The prediction performance of each of the three ensemble methods are evaluated using $10$-fold cross validation \citep{hastie2005elements}, that is, each data set in Table \ref{data_summary} is randomly split into $10$ equal-size segments of which $9$ segments are used for training and the other one is used for testing. The prediction error rate is then averaged over the $10$ trials.
\begin{table}[htbp]
\renewcommand{\arraystretch}{1.3}
\caption{Summary of Non-Image Data Sets}
\label{data_summary}
\centering
\begin{tabular}{cccc}
\hline
\hline
Name & Instances & Classes & Attributes \\
\hline
Breast-w & 699 & 2 & 9 \\
Iris & 150 & 3 & 4 \\
Segment & 2310 & 7 & 19 \\
\hline
\hline
\end{tabular}
\end{table}
The $10$-fold cross validation error rates of the three ensemble methods on the three data sets are tabulated in Table \ref{DECORATE_comparison}. We can see that out of the three non-image datasets, for both Breast-w and Iris data, the neural network ensemble trained with our algorithm outperforms the DECORATE ensembles, and the error rate on Breast-w even gets reduced by half when using our algorithm compared to the original DECORATE ensemble. For Segment data, our method still shows competitive performance, and gives a lower error rate than the DECORATE neural network ensemble.
\begin{table}[htbp]
\renewcommand{\arraystretch}{1.3}
\caption{$10$-fold Cross Validation Error Rate \% (DECORATE Versus Our Algorithm)}
\label{DECORATE_comparison}
\centering
\begin{tabular}{p{9cm}|c|c|c}
\hline
\hline
& Breast-w & Iris & Segment \\
\hline
DECORATE in \cite{melville2005creating} & 3.69 & 5.33 & \color{red}1.97 \\
DECORATE on Neural Network Ensemble & 2.04 & 10.67 & 3.90 \\
Our Algorithm on Neural Network Ensemble & \color{red}1.74 & \color{red}4.00 & 3.33 \\
\hline
\hline
\end{tabular}
\end{table}
We can take a closer look at the training process of the neural network ensemble trained with DECORATE and our algorithm, respectively, by visualizing their computed averaged truth-learner correlations and averaged learner-learner correlations on the same graph. Figure \ref{decorate} shows such an example when the two neural network ensembles are trained on the Breast-w data set. Note that following the analysis in previous sections, our ensemble exhibits an overall higher diversity and higher accuracy than the DECORATE ensemble, and since our ensemble is closer to the optimal boundary, it is guaranteed by theory that our ensemble should show a better prediction performance, which is exactly the case as we look at the 10-fold cross validation error rates on the Breast-w data.
\begin{figure}[htbp]
\centerline{\includegraphics[width=0.5\textwidth]{figures/jmlr_decorate.png}}
\caption{Training on Breast-w Data. Our neural network ensemble (in orange) shows an overall higher diversity and higher accuracy than DECORATE neural network ensemble (in blue).}
\label{decorate}
\end{figure}
In summary, compared to the DECORATE algorithm, our algorithm does a better job on creating diversity in neural network ensembles and keeping the ensembles accurate in the meantime. In addition, the DECORATE algorithm is less efficient and much more expensive computationally, especially when the ensemble size is large as it needs to generate artificial training samples at each iteration.
\subsubsection{\textbf{Experiments on Image Data}} \label{image}
All our experimental studies included in this section are conducted using the library PyTorch 1.1.0 [22] in Python 3.6.9 [23]. We run the Python scripts on compute nodes with $1$ NVIDIA K20 GPU with $6$ Intel CPU 2.10 GHZ cores and $32$ GB memory each. \footnote{The Python code for the experiments in section \ref{image} can be found on \url{https://bitbucket.org/wli5/neural-network-ensembles-theory-training-and-the-importance-of/src/master/}.}
Herein our focus is on image classification of the benchmark image data sets CIFAR-10 and CIFAR-100 \citep{krizhevsky2009learning}, where we perform training on the train sets and testing on the test sets as provided by the PyTorch library. However, the theory and training algorithm provided in this paper are also applicable to many other machine learning tasks. To show the general effectiveness of our training algorithm on neural network ensembles, we conduct three major comparison studies (results are displayed in Table \ref{model comparison 1}, \ref{model comparison 2}, \ref{model comparison 3}), and compare the prediction performance of the neural network ensembles trained with our proposed approach and other state-of-the-art individual networks as well as ensembles of state-of-the-art networks.
Table \ref{model comparison 1} shows the comparison study on a binary subset (``plane" and ``car") of the CIFAR-10 image data. In particular, five single pretrained networks (ShuffleNet-v2 \citep{ma2018shufflenet}, GoogLeNet \citep{szegedy2015going}, MobileNet-v2 \citep{sandler2018mobilenetv2}, ResNet-18 \citep{he2016deep} and DenseNet-121 \citep{huang2017densely} included in ``torchvision.models") originally trained on ImageNet \citep{imagenet_cvpr09} are individually re-trained using the standard cross entropy loss function. The prediction error rate of the five single nets are listed in Table \ref{model comparison 1}. Based on the five single nets, two ensembles are built up: one (displayed as ``Model II" in Table \ref{model comparison 1}) is trained with standard ensemble-version of cross entropy loss (that is, the cross entropy loss of the ensemble is calculated between the averaged outputs from the component nets and the true class labels \citep{icmla_paper}), the other one (displayed as ``Our Model") is trained by our proposed algorithm with various specified $\lambda$ values. We can see that the neural network ensemble trained by our approach outperforms both those single networks and the ensemble trained with standard approach, in fact, our ensemble with $\lambda=0.3$ shows the lowest prediction error rate of $1.25\%$, which is over 2 times lower than the best single net MobileNet-v2, and 6 times lower than the standard ensemble. \emph{This comparison study successfully demonstrates the effectiveness of our training algorithm on binary image classification problems. }
\begin{table}[t]
\centering
\begin{threeparttable}
\renewcommand{\arraystretch}{1.3}
\caption{Pretrained Networks on CIFAR-10 (Binary Subset: ``Plane" \& ``Car") with Prediction Error Rate (\%)}
\begin{tabular}{m{2cm}m{6.8cm}c}
\hline \hline \noalign{\smallskip}
Model & Description & Prediction Error Rate (\%) \\
\hline
Model I & Five single re-trained neural networks trained with cross entropy loss until convergence. This is our baseline non-ensemble model for comparison. &
\begin{tabular}{cc}
Network & Error Rate \\
\hline
ShuffleNet-v2 & 2.80 \\
GoogLeNet & 3.30 \\
MobileNet-v2 & 2.70 \\
ResNet-18 & 5.25 \\
DenseNet-121 & 9.05 \\
\end{tabular} \\
\noalign{\smallskip}\noalign{\smallskip}
\hline
Model II & An ensemble (trained with the ensemble-version of cross entropy loss for $10$ epochs) of the five re-trained nets in Model I .
& 7.55 \\
\hline
\noalign{\smallskip}\noalign{\smallskip}
Our Model & An ensemble (trained with our algorithm for $10$ epochs) of the five re-trained nets in Model I. This model demonstrates the practical effectiveness of our approach on binary classification.
&
\begin{tabular}{cc}
$\lambda$ & Error Rate \\
\hline
0.1 & 1.35 \\
0.3 & \textbf{{\color{red}1.25}} \\
0.5 & 1.70 \\
0.7 & 1.90 \\
0.9 & 1.80 \\
\end{tabular} \\
\noalign{\smallskip}\hline\hline
\end{tabular}
\begin{tablenotes}
\small
\item Note: This table shows the prediction error rates for experiments on a binary subset (``plane" and ``car") of CIFAR-10. The neural network ensemble trained by our approach outperforms both those single networks and the ensemble trained with standard approach, demonstrating the effectiveness of our approach on binary classification.
\end{tablenotes}
\label{model comparison 1}
\end{threeparttable}
\end{table}
Table \ref{model comparison 2} shows the comparison study on the entire CIFAR-10 with all $10$ classes.
Similarly as the study shown in Table \ref{model comparison 1}, three single pretrained networks (ShuffleNet-v2, GoogLeNet and MobileNet-v2) are individually re-trained using the standard cross entropy loss, and two ensembles of size three are built up: one (displayed as ``Model II" in Table \ref{model comparison 2}) is trained with standard ensemble-version of cross entropy loss, the other one (displayed as ``Our Model") is trained by our proposed algorithm with various specified $\lambda$ values. In this case, our ensemble with $\lambda=0.1$ shows the lowest prediction error rate as $5.76\%$, which is much smaller than the best single net GoogLeNet and the standard ensemble. \emph{This comparison study demonstrates the effectiveness of our training algorithm on multi-label image classification problems.}
\begin{table}[t]
\centering
\begin{threeparttable}
\renewcommand{\arraystretch}{1.3}
\caption{Pretrained Networks on CIFAR-10 with Prediction Error Rate (\%)}
\begin{tabular}{m{2cm}m{6.8cm}c}
\hline \hline \noalign{\smallskip}
Model & Description & Prediction Error Rate (\%) \\
\hline
Model I & Three single re-trained neural networks trained with cross entropy loss using $40$ epochs. This is our baseline non-ensemble model for comparison. &
\begin{tabular}{cc}
Network & Error Rate \\
\hline
ShuffleNet-v2 & 8.50 \\
GoogLeNet & 6.43 \\
MobileNet-v2 & 7.11 \\
\end{tabular} \\
\noalign{\smallskip}\noalign{\smallskip}
\hline
Model II & An ensemble (trained with the ensemble-version of cross entropy loss function using $40$ epochs) of the three re-trained nets in Model I.
& 7.78 \\
\hline
\noalign{\smallskip}\noalign{\smallskip}
Our Model & An ensemble (trained with our algorithm using 40 epochs) of the three re-trained nets in Model I. This model demonstrates the practical effectiveness of our approach on multi-label classification. &
\begin{tabular}{cc}
$\lambda$ & Error Rate \\
\hline
0.1 & \textbf{\color{red}5.76} \\
0.2 & 5.91 \\
0.3 & 6.50 \\
\end{tabular} \\
\noalign{\smallskip}\hline\hline
\end{tabular}
\begin{tablenotes}
\small
\item Note: This table shows the prediction error rates for experiments on the entire CIFAR-10 with all 10 classes. The neural network ensemble trained by our approach outperforms both those single networks and the ensemble trained with standard approach, demonstrating the effectiveness of our approach on multi-label classification.
\end{tablenotes}
\label{model comparison 2}
\end{threeparttable}
\end{table}
A third comparison study is performed to show that with our theory and proposed algorithm, the prediction performance of any given neural network ensembles can be improved even further. Herein, we will take experiments on a recent state-of-art neural network EfficientNet \citep{tan2019efficientnet} as such an example. Table \ref{model comparison 3} shows the prediction error rates of experiments on CIFAR-10 and CIFAR-100 data sets. The baseline model for comparison is a single EfficientNet-b0 net trained with standard cross entropy loss function, while our ensemble model consisting of three copies of the trained EfficientNet-b0 net is trained with our algorithm. We can see that for both data sets, the neural network ensembles trained with our algorithm provide lower prediction error rates than the baseline neural network (which is already very accurate), meaning that \emph{our algorithm is able to make good neural networks even better by explicitly imposing diversity and learner accuracy on the loss function.}
\begin{table}[htbp]
\centering
\begin{threeparttable}
\caption{EfficientNet on CIFAR-10 and CIFAR-100 with Prediction Error Rate (\%)}
\begin{tabular}{m{2cm}m{7cm}m{2.1cm}m{2.2cm}}
\hline \hline \noalign{\smallskip}
Model & Description & CIFAR-10 & CIFAR-100 \\
\hline
Baseline & A single EfficientNet-b0 model trained with standard cross entropy loss by $40$ epochs. & 5.44\% & 17.28\% \\
\hline
Our Model & An ensemble of three copies of re-trained EfficientNet-b0 model trained with our algorithm by $40$ epochs. & \textbf{\color{red}4.65\%} \ ($\lambda=0.1$) & \textbf{\color{red}14.13\%} ($\lambda=0.1$) \\
\noalign{\smallskip}\hline\hline
\end{tabular}
\label{efficientnet}
\begin{tablenotes}
\small
\item Note: This table shows the prediction error rate for experiments on CIFAR-10 and CIFAR-100. For both data sets, the neural network ensembles trained with our algorithm provide lower prediction error rates than the baseline neural network (which is already very accurate), meaning that our algorithm is able to make good neural networks even better.
\end{tablenotes}
\label{model comparison 3}
\end{threeparttable}
\end{table}
Therefore, based on the three comparison studies shown in Table \ref{model comparison 1}, \ref{model comparison 2}, \ref{model comparison 3}, \emph{our training algorithm of neural network ensembles consistently achieves the lowest error rate on a variety of problems and improves the prediction performance of both state-of-the-art individual networks as well as ensembles of state-of-the-art networks.}
\section{CONCLUSIONS}
In this paper, we develop theory that offers a rigorous understanding of how individual accuracies of component learners and diversity affect the performance of an ensemble. These theoretical considerations inspires methodologies for assessing and improving the optimality of given ensembles. Just as importantly, we propose a training algorithm for deep neural network ensembles that explicitly encourages ensemble diversity, and this algorithm is shown to be generally effective in improving start-of-the-art neural networks by experiments on standard benchmark data sets. This training algorithm is also quite efficient in that one only needs to train the ensembles using a small number of epochs to achieve good performance, assuming that the individual learners are already trained outside of the ensemble.
Based upon the results in this paper, there are several interesting directions for future work. First, there appear to be interesting connections between the proof of Theorem \ref{thm2} and domains such as coding theory, and it would be quite useful to understand these connections better. Second, while our numerical results suggest a strong, and general, connection between $r_{TL}^{(ave)}$ and ensemble accuracy, our Theorem \ref{thm3} only address the case of homogeneous ensembles. Generalizing this result would likely lead to a deep understanding of the performance of ensembles. Third, and perhaps most importantly, our numerical experiments beg for application to many more possible machine learning tasks. There are a wide range of possibilities for taking already state-of-the-art methods and improving them even further by putting them into ensembles.
\newpage
|
\section{Introduction}\label{intro-section}
Let $S=K[t_1,\ldots,t_s]=\bigoplus_{d=0}^{\infty} S_d$ be a polynomial ring over
a field $K$ with the standard grading and let $I$ be a graded ideal of
$S$. A prime ideal $\mathfrak{p}$ of $S$ is an \textit{associated prime} of $S/I$ if
$(I\colon f)=\mathfrak{p}$ for some $f\in S_d$, where $(I\colon f)$ is
the set of all $g\in S$ such that $gf\in I$. The set of associated primes of $S/I$
is denoted by ${\rm Ass}(I)$ and set of maximal elements of ${\rm Ass}(I)$ with respect to inclusion is
denoted by ${\rm Max}(I)$. The v-{\em number} of $I$, denoted ${\rm
v}(I)$, is the following invariant of
$I$ that was introduced in \cite{min-dis-generalized} to study the asymptotic behavior
of the minimum distance of projective Reed--Muller-type codes
\cite[Corollary~4.7]{min-dis-generalized}:
$$
{\rm v}(I):=\min\{d\geq 0 \mid\exists\, f
\in S_d \mbox{ and }\mathfrak{p} \in {\rm Ass}(I) \mbox{ with } (I\colon f)
=\mathfrak{p}\}.
$$
\quad One can define the v-number of $I$ locally at each associated
prime $\mathfrak{p}$ of $I$\/:
$$
{\rm v}_{\mathfrak{p}}(I):=\mbox{min}\{d\geq 0\mid \exists\, f\in S_d \mbox{ with }(I\colon f)=\mathfrak{p}\}.
$$
\quad For a graded module $M\neq 0$, we define
$\alpha(M):=\min\{\deg(f) \mid f\in M\setminus\{0\}\}$. By convention,
we set $\alpha(0):=0$. Part (d) of the next result was shown in
\cite[Proposition~4.2]{min-dis-generalized} for unmixed graded
ideals. The next result gives a formula for the v-number of any graded
ideal.
\noindent \textbf{Theorem~\ref{vnumber-general}.}\textit{
Let $I\subset S$ be a graded ideal and let $\mathfrak{p}\in{\rm
Ass}(I)$. The following hold.
\begin{enumerate}
\item[(a)]
If $\mathcal{G}=\{\overline{g}_1,\ldots,\overline{g}_r\}$ is
a homogeneous minimal generating set of $(I\colon\mathfrak{p})/I$,
then
$$
{\rm v}_{\mathfrak{p}}(I)=
\min\{\deg(g_i)\mid 1\leq i\leq r\mbox{ and }(I\colon
g_i)=\mathfrak{p}\}.
$$
\item[(b)] ${\rm v}(I)=\min\{{\rm v}_{\mathfrak{q}}(I)\mid
\mathfrak{q}\in{\rm Ass}(I)\}$.
\item[(c)] ${\rm
v}_{\mathfrak{p}}(I)\geq\alpha((I\colon\mathfrak{p})/I)$ with equality
if $\mathfrak{p}\in{\rm Max}(I)$.
\item[(d)] If $I$ has no embedded primes, then $
{\rm v}(I)=\min\{\alpha\left((I\colon\mathfrak{q})/{I}\right)\vert\,
\mathfrak{q}\in{\rm Ass}(I)\}.
$
\end{enumerate}}
The formulas of parts (a) and (b) give an algorithm to compute the v-number
number using \textit{Macaulay}$2$ \cite{mac2} (Example~\ref{example1},
Procedure~\ref{procedure1}).
The v-number of non-graded ideals was used in \cite{dual} to compute the
regularity index of the minimum distance function of affine
Reed--Muller-type codes \cite[Proposition~6.2]{dual}. In this case,
one considers the vanishing ideal of a
set of affine points over a finite field.
For certain classes of graded ideals ${\rm v}(I)$ is a lower bound for
${\rm reg}(S/I)$, the regularity of the quotient ring $S/I$
(Definition~\ref{regularity-def}), see
\cite{min-dis-generalized,v-number,footprint-ci}. There are examples
of ideals where ${\rm v}(I)>{\rm reg}(S/I)$ \cite{v-number}. It is an open problem whether
${\rm v}(I)\leq {\rm reg}(S/I)+1$ holds for any squarefree
monomial ideal. Upper and lower bounds for the regularity of edge
ideals and their powers are given in
\cite{banerjee-etal,Beyarslan-etal,Dao-Huneke-schweig,Herzog-Hibi-ub,JS,edge-ideals,woodroofe-matchings},
see Section~\ref{prelim-section}.
Using the polarization technique of Fr\"oberg \cite{Fro1}, we give an
upper bound for the regularity of a monomial ideal $I$ in terms of the
dimension of $S/I$ and the exponents of the monomials that
generate $I$ (Proposition~\ref{reg-dim-pol}).
Let $G$ be a graph with vertex set $V(G)$ and edge
set $E(G)$. If $V(G)=\{t_1,\ldots,t_s\}$, we can regard each vertex $t_i$ as a
variable of the polynomial ring $S=K[t_1,\ldots,t_s]$ and think of each edge $\{t_i,t_j\}$ of
$G$ as the quadratic monomial
$t_it_j$ of $S$. The
\textit{edge ideal} of $G$ is the squarefree monomial ideal of $S$ defined as
$$I(G):=(t_{i}t_{j}\mid \{t_{i},t_{j}\}\in E(G)).$$
\quad This ideal, introduced in \cite{cm-graphs},
has been studied in the literature from
different perspectives, see
\cite{graphs-rings,Herzog-Hibi-book,unmixed-c-m,ITG,monalg-rev} and the
references therein. We use induced matchings of $G$ to compare the v-number of
$I(G)$ with the regularity of $S/I(G)$ for certain families of graphs.
A subset $C$ of $V(G)$ is a {\it vertex cover\/} of $G$ if every edge
of $G$ is incident with at least one vertex in $C$. A vertex cover $C$ of $G$
is {\it minimal\/} if each proper subset of $C$ is not a vertex cover
of $G$. A subset $A$ of $V(G)$ is called {\it stable\/}
if no two points in $A$ are
joined by an edge. Note that a set of vertices $A$ is a (maximal)
stable set of $G$ if and only if
$V(G)\setminus
A$ is a (minimal) vertex cover of $G$. The \textit{stability number}
of $G$, denoted by $\beta_0(G)$,
is the cardinality of a maximum stable set of $G$ and the
{\it covering number\/} of $G$, denoted $\alpha_0(G)$, is the
cardinality of a minimum vertex cover of $G$. For use below we introduce the following two families of stable
sets:
\begin{align*}
\mathcal{F}_G&:=\{A\mid A\mbox{ is a maximal stable set
of }G\},\mbox{ and}\\
\mathcal{A}_G&:=\{A\mid A\mbox{ is a stable set
of }\, G\mbox{ and }N_G(A)\mbox{ is a minimal vertex
cover of }G\}.
\end{align*}
According to \cite[Theorem~3.5]{v-number},
$\mathcal{F}_G\subset\mathcal{A}_G$ and the $\mathrm{v}$-number of
$I(G)$ is given by
$$
\mathrm{v}(I(G)) = \min\{\vert A \vert : A\in\mathcal{A}_G\}.
$$
\quad The v-number of $I(G)$ is a combinatorial invariant of $G$
that has been used to characterize the family of $W_2$-graphs (see
the discussion below before Corollary~\ref{w2-graph}). We can
define the v-number of a graph $G$ as ${\rm v}(G):={\rm v}(I(G))$
and study ${\rm v}(G)$ from the viewpoint of graph theory.
A set $P$ of pairwise disjoint edges of $G$ is called a
{\it matching\/}. A matching $P=\{e_1,\ldots,e_r\}$ is \textit{perfect} if
$V(G)=\bigcup_{i=1}^re_i$. An {\it induced matching\/} of a graph $G$ is a
matching $P=\{e_1,\ldots,e_r\}$ of $G$ such that the only edges of
$G$ contained in $\bigcup_{i=1}^re_i$ are $e_1,\ldots,e_r$.
The \textit{matching number} of $G$, denoted $\beta_1(G)$,
is the maximum cardinality of a matching of $G$ and the
{\it induced matching number\/} of $G$,
denoted ${\rm im}(G)$, is the number of edges in the largest
induced matching.
The graph $G$ is
{\it well-covered\/}
if every maximal stable set of $G$ is of the same size
and $G$ is \textit{very
well-covered} if $G$ is well-covered, has no isolated vertices,
and $|V(G)|=2\alpha_0(G)$. The class of very well-covered graphs
includes the bipartite well-covered graphs without isolated vertices \cite{ravindra,unmixed} and
the whisker graphs \cite[p.~392]{ITG} (Lemma~\ref{bipartite-whiskers}). A
graph without isolated vertices is very well-covered if and only
if $G$ is well-covered and $\beta_1(G)=\alpha_0(G)$
(Proposition~\ref{konig-vwc}). One of
the properties of very well-covered graphs that will be used
to show the following theorem is that they can be classified using combinatorial
properties of a perfect matching as was shown by Favaron \cite[Theorem~1.2]{favaron}
(Theorem~\ref{konig}, cf.~Theorem~\ref{lemma-Ivan}).
We come to one of our main results.
\noindent \textbf{Theorem~\ref{Domi-InduceMatch}.}\textit{
Let $G$ be a very well-covered graph and let $P=\{e_1, \ldots, e_r\}$
be a perfect matching of $G$.
Then, there is an induced submatching $P'$ of $P$ and $D \in {\mathcal A}_G$
such that $D \subset V(P')$ and $\vert e \bigcap D \vert = 1$ for
each $e\in P'$.
Furthermore ${\rm v}(I(G))\leq|P'|=|D|\leq{\rm im}(G)\leq{\rm
reg}(S/I(G))$.
}
Let $G$ be a graph and let $W_G$ be its whisker graph
(Section~\ref{prelim-section}).
As a consequence we recover a result of \cite{v-number} showing that
the v-number of
$I(W_G)$ is bounded
from above by the regularity of the quotient ring $K[V(W_G)]/I(W_G)$
(Corollary~\ref{sep29-21}). The \textit{independent
domination number} of $G$, denoted by $i(G)$, is the minimum size of
a maximal stable set \cite[Proposition~2]{Allan-Laskar}:
$$i(G):=\min\{|A|\colon
A\in\mathcal{F}_G\},$$
and $i(G)$ is equal to the v-number of the whisker graph $W_G$ of $G$
\cite[Theorem 3.19(a)]{v-number}.
A cycle of length $s$ is denoted by $C_s$.
The inequality ${\rm v}(I(G))\leq{\rm reg}(S/I(G))$ of
Theorem~\ref{Domi-InduceMatch}
is false if we only assume that $G$ is a well-covered graph,
since the cycle $C_5$ is a well-covered graph, but one has ${\rm im}(C_5)=1 <
2={\rm v}(I(C_5))$. We prove that $C_5$ is the only cycle
where the inequality ${\rm v}(I(C_s))\leq{\rm im}(C_s)$ fails.
\noindent \textbf{Theorem~\ref{cycles-indmat}.}\textit{
Let $C_s$ be an $s$-cycle and let $I(C_s)$ be its edge ideal. Then,
${\rm v}(I(C_s))\leq{\rm im}(C_s)$ if and only if $s \neq 5$.
}
If $v\in V(G)$, we denote the closed neighborhood of $v$ by $N_G[v]$.
A vertex $v$ of $G$ is called {\it simplicial} if the induced
subgraph $H=G[N_{G}[v]]$ on the vertex set $N_G[v]$ is a complete graph. A subgraph $H$ of $G$ is
called a \textit{simplex} if $H=G[N_{G}[v]]$ for some simplicial
vertex $v$. A graph $G$ is \textit{simplicial} if every vertex of $G$
is either simplicial
or is adjacent to a simplicial vertex of $G$.
If $A$ is a
stable set of a graph $G$, $H_i$ is a complete subgraph of $G$ for
$i=1,\ldots,r$ and $A\bigcup \{V(H_i)\}_{i=1}^r$ is a partition of
$V(G)$, then ${\rm reg}(S/I(G))\leq r$
\cite[Theorem~2]{woodroofe-matchings}. We consider a special type of
partitions of $V(G)$ that allow us to link $\mathcal{A}_G$ with induced
matchings of $G$. A graph $G$ has a \textit{simplicial partition} if $G$ has simplexes $H_1, \ldots, H_r$,
such that $\{V(H_i)\}_{i=1}^r$ is a partition of $V(G)$.
Our next
result shows that ${\rm v}(I(G))\leq{\rm im}(G)$ if $G$ has a
simplicial partition.
\noindent \textbf{Theorem~\ref{Domi-InduceMatch-simplex}.}\textit{
Let $G$ be a graph with simplexes $H_1, \ldots, H_r$,
such that $\{V(H_i)\}_{i=1}^r$ is a partition of $V(G)$.
If $G$ has no isolated vertices, then there
is $D=\{y_1,\ldots,y_k\} \in {\mathcal A}_G$,
and there are simplicial vertices $x_1,\ldots,x_k$ of $G$ and integers
$1\leq j_1< \cdots <j_k \leq r$ such that $P=\{\{x_i,y_i\}\}_{i=1}^k$ is an induced matching of $G$ and
$H_{j_i}$ is the induced subgraph $G[N_G[x_i]]$ on $N_G[x_i]$ for $i=1,\ldots,k$. Furthermore ${\rm
v}(I(G))\leq|D|=|P|\leq{\rm im}(G)\leq{\rm reg}(S/I(G))$.
}
As a consequence, using a result of Finbow, Hartnell and Nowakowski
that classifies the connected well-covered graphs without $4$- and
$5$-cycles \cite[Theorem 1.1]{Finbow2}
(Theorem~\ref{wellcovered-characterization1}), we show two more
families of graphs where the v-number is a lower bound for the
regularity.
\noindent \textbf{Corollary~\ref{simplicial-4-5-cycles}.}\textit{
Let $G$ be a well-covered graph and let $I(G)$ be its edge ideal. If $G$ is simplicial
or $G$ is connected and contain neither $4$- nor $5$-cycles, then
$${\rm v}(I(G))\leq{\rm im}(G)
\leq{\rm reg}(S/I(G))\leq\beta_0(G).$$
}
A vertex $v$ of a graph $G$ is called a \textit{shedding vertex} if
each stable set of $G\setminus N_{G}[v]$ is not a maximal stable set
of $G\setminus v$. We prove that every vertex of $G$ is a shedding
vertex if and only if ${\mathcal A}_{G}={\mathcal F}_{G}$
(Proposition~\ref{Shedding-stable}).
A graph $G$ belongs to class $W_2$ if
$|V(G)|\geq 2$ and any two disjoint stable sets $A_1,A_2$ are contained in two
disjoint maximum stable sets $B_1, B_2$ with $|B_i|=\beta_0(G)$ for
$i=1,2$.
A graph $G$ is in $W_2$ if and only if $G$ is well-covered,
$G\setminus v$ is well-covered for all $v\in
V(G)$ and $G$ has no isolated vertices \cite[Theorem~2.2]{Levit-Mandrescu}. A graph $G$
without isolated vertices is in $W_2$ if and only if ${\rm
v}(I(G))=\beta_0(G)$ \cite[Theorem~4.5]{v-number}. As an
application we recover the only if implication of this result
(Corollary~\ref{w2-graph}). Using that a graph $G$ without isolated vertices is in
$W_2$ if and only if $G$ is well-covered and ${\mathcal
A}_{G}={\mathcal F}_{G}$ \cite[Theorem 4.3]{v-number}, by
Proposition~\ref{Shedding-stable}, we recover the
fact that a graph $G$ without isolated vertices is in $W_2$ if and only if $G$ is well-covered and every
$v\in V(G)$ is a shedding vertex \cite[Theorem 3.9]{Levit-Mandrescu}. For other
characterizations of graphs in $W_2$ see
\cite{Levit-Mandrescu,Staples} and the references therein.
In Section~\ref{examples-section} we show examples illustrating some
of our results. In particular in Example~\ref{example2} we compute the
combinatorial and algebraic invariants of the well-covered graphs
$C_7$ and $T_{10}$ that are
depicted in Figure~\ref{C7-T10}. These two graphs occur in the
classification of connected well-covered graphs without $4$- and
$5$-cycles \cite[Theorem 1.1]{Finbow2}
(Theorem~\ref{wellcovered-characterization1}). A related result is the
characterization of
well-covered graphs of girth at least $5$ given in \cite{finbow1}.
For all unexplained terminology and additional information, we refer to
\cite{diestel,Har} for the theory of graphs and
\cite{graphs-rings,Herzog-Hibi-book,monalg-rev} for the theory of
edge ideals and monomial ideals.
\section{Preliminaries}\label{prelim-section}
In this section we give some definitions and present some well-known
results that will be used in the following sections.
To avoid repetitions, we continue to employ
the notations and
definitions used in Section~\ref{intro-section}.
\begin{definition}\cite{eisenbud-syzygies}\label{regularity-def} Let $I\subset S$ be a graded ideal and let
${\mathbf F}$ be the minimal graded free resolution of $S/I$ as an
$S$-module:
\[
{\mathbf F}:\ \ \ 0\rightarrow
\bigoplus_{j}S(-j)^{b_{g,j}}
\stackrel{}{\rightarrow} \cdots
\rightarrow\bigoplus_{j}
S(-j)^{b_{1,j}}\stackrel{}{\rightarrow} S
\rightarrow S/I \rightarrow 0.
\]
The {\it Castelnuovo--Mumford regularity\/} of $S/I$ ({\it
regularity} of $S/I$ for short) is defined as
$${\rm reg}(S/I):=\max\{j-i \mid b_{i,j}\neq 0\}.
$$
\quad The integer $g$, denoted ${\rm pd}(S/I)$, is the \textit{projective dimension} of $S/I$.
\end{definition}
Let $G$ be a graph with vertex set $V(G)$. Given $A\subset V(G)$,
the \textit{induced subgraph} on $A$, denoted $G[A]$, is the maximal
subgraph of $G$ with vertex set $A$. The edges of $G[A]$ are all the edges of $G$ that are
contained in $A$. The induced subgraph $G[V(G)\setminus A]$ of $G$ on the
vertex set $V(G)\setminus A$ is denoted by
$G\setminus A$.
If $v$ is a vertex of $G$, then we denote the neighborhood of
$v$ by $N_G(v)$ and the closed neighborhood $N_G(v)\bigcup\{v\}$
of $v$ by $N_G[v]$. Recall that $N_G(v)$ is the set of all vertices
of $G$ that are adjacent to $v$. If $A\subset V(G)$, we
set $N_G(A):=\bigcup_{a\in A} N_G(a)$.
\begin{theorem}\cite{Campbell}\label{Campbell-theo}
If a graph $G$ is well-covered and is not complete, then
$G_v:=G\setminus N_G[v]$ is well-covered
for all $v$ in $V(G)$. Moreover, $\beta_0(G_v) = \beta_0(G) - 1$.
\end{theorem}
If $G$ is a graph, then $\beta_1(G)\leq \alpha_0(G)$. We say that $G$
is a \textit{K\H{o}nig graph} if $\beta_1(G)=\alpha_0(G)$. This notion
can be used to classify very well-covered graphs
(Proposition~\ref{konig-vwc}).
\begin{theorem}\label{lemma-Ivan}{\rm
(\cite[Theorem 5]{disc-math}, \cite[Lemma~2.3]{susan-reyes-vila})}
Let $G$ be a graph without isolated vertices. If $G$ is a graph
without $3$-, $5$-, and $7$-cycles or $G$ is a K\H{o}nig graph, then $G$ is
well-covered if and only if $G$ is very well-covered.
\end{theorem}
\begin{definition}\label{p-def1}
A perfect matching $P$ of a
graph $G$
is said to have property {\bf(P)}
if for all $\{a,b\}$, $\{a^{\prime},b^{\prime}\}\in E(G)$, and
$\{b,b^{\prime}\}\in P$, one has $\{a,a^{\prime}\}\in E(G)$.
\end{definition}
\begin{remark}\label{sep26-19}
Let $P$ be a perfect matching of a graph $G$ with property {\bf(P)}.
Note that if $\{b,b'\}$ is in $P$ and $a\in V(G)$, then
$\{a,b\}$ and $\{a,b'\}$ cannot be both in $E(G)$ because $G$
has no loops. In other words $G$ has no triangle containing
an edge in $P$.
\end{remark}
\begin{theorem}\label{konig}{\rm \cite[Theorem 1.2]{favaron}}
The following conditions are equivalent for a graph $G$:
\begin{enumerate}
\item $G$ is very well-covered.
\item $G$ has a perfect matching with property {\bf{(P)}}.
\item $G$ has a perfect matching, and each perfect matching of $G$ has
property {\bf{(P)}}.
\end{enumerate}
\end{theorem}
Let $G$ be a graph with vertex set $V(G)=\{t_1,\ldots,t_s\}$ and let
$U=\{u_1,\ldots, u_s\}$ be a new set of vertices. The {\it whisker graph} or {\it
suspension\/} of $G$, denoted by
$W_G$, is the graph obtained from $G$ by attaching to each
vertex $t_i$ a new vertex $u_i$ and a new edge $\{t_i,u_i\}$.
The edge $\{t_i,u_i\}$ is called a {\it whisker} or \textit{pendant
edge}. The graph $W_G$ was
introduced in \cite{ITG} as a device to study the numerical invariants and
properties of graphs and edge ideals.
\begin{lemma}\label{bipartite-whiskers}
Let $G$ be a graph without isolated vertices. The following hold.
\begin{itemize}
\item[(a)] If $G$ is a bipartite well-covered graph,
then $G$ is very well-covered.
\item[(b)] The whisker graph $W_G$ of $G$ is very well-covered.
\end{itemize}
\end{lemma}
\begin{proof} (a) A bipartite well-covered graph without
isolated vertices has a perfect matching $P$ that satisfies
property {\bf(P)} \cite[Theorem~1.1]{unmixed}. Thus, by
Theorem~\ref{konig}, $G$ is very well-covered.
(b): The perfect matching $P=\{\{t_i,u_i\}\}_{i=1}^n$ of the whisker
graph $W_G$
satisfies property {\bf(P)} and, by
Theorem~\ref{konig}, $G$ is very well-covered.
\end{proof}
\begin{proposition}\cite[Lemma~17]{Ivan-Reyes}\label{konig-vwc} Let $G$ be a graph without
isolated vertices. Then, $G$ is a very well-covered graph if and only
if $G$ is well-covered and $\beta_1(G)=\alpha_0(G)$.
\end{proposition}
\begin{proof}
$\Rightarrow$) Assume that $G$ is very well-covered, then
$|V(G)|=2\alpha_0(G)$. It suffices to show that $\beta_1(G)=\alpha_0(G)$. In
general $\beta_1(G)\leq \alpha_0(G)$. By Theorem~\ref{konig}, $G$ has
a perfect matching $P=\{e_1,\ldots,e_r\}$. Then,
$|V(G)|=2r=2\alpha_0(G)$ and $r=\alpha_0(G)$. Thus,
$\alpha_0(G)=|P|\leq\beta_1(G)$, and one has $\alpha_0(G)=\beta_1(G)$.
$\Leftarrow$) Assume that $G$ is well-covered and
$\beta_1(G)=\alpha_0(G)$. Let $P=\{e_1,\ldots,e_r\}$ be a matching of $G$ with
$r=\beta_1(G)$. We need only show that $|V(G)|=2\alpha_0(G)$. Clearly
$|V(G)|$ is greater than or equal
to $2\alpha_0(G)$ because $\bigcup_{i=1}^re_i\subset V(G)$. We proceed by contradiction assuming that
$\bigcup_{i=1}^re_i\subsetneq V(G)$. Pick $v\in
V(G)\setminus\bigcup_{i=1}^re_i$. As $v$ is not an isolated vertex of
$G$, there is a minimal vertex cover $C$ of $G$ that contains $v$. As
$G$ is well-covered one has that $|C|=\alpha_0(G)=r$. Since $e_i\bigcap
C\neq\emptyset$ for $i=1,\ldots,r$ and $v\in C$, we get $|C|\geq r+1$,
a contradiction.
\end{proof}
We say that a graph $G$ is in the family $\mathcal{F}$ if there
exists $\{x_1,\ldots,x_k\}\subset V(G)$ where for each $i$, $x_i$ is
simplicial, $|N_G[x_i]|\leq 3$ and $\{N_G[x_i]\mid i=1,\ldots,k\}$ is
a partition of $V(G)$.
\begin{theorem}\label{wellcovered-characterization1}{\rm \cite[Theorem 1.1]{Finbow2}}
Let $G$ be a connected graph that contain neither $4$- nor $5$-cycles and let
$C_7$ and $T_{10}$ be the two graphs in Example~\ref{example2}. Then
$G$ is a well-covered graph if and only if
$G\in \{C_7,T_{10}\}$ or $G\in\mathcal{F}$.
\end{theorem}
\begin{theorem}\label{lower-bound-reg}
Let $G$ be a graph. The following hold.
\begin{enumerate}
\item [(a)] {\rm(\cite[Theorem 4.5]{Beyarslan-etal},
\cite{katzman1})}
$2(n-1)+{\rm im}(G)\leq{\rm
reg}(S/I(G)^n) $ for all $n\geq 1$.
\item[(b)] {\rm(\cite[Theorem 4.7]{Beyarslan-etal},
\cite{JS-very-well-covered})} If $G$ is a forest or
$G$ is very well-covered, then
$${\rm
reg}(S/I(G)^n)=2(n-1)+{\rm im}(G)\text{ for all } n\geq 1.
$$
\item[(c)] \cite[Theorem 1.3]{Mahmoudi-et-al} If $G$ is very
well-covered, then ${\rm reg}(S/I(G)) = {\rm im}(G)$.
\end{enumerate}
\end{theorem}
\section{The \text{v}-number of a graded ideal}\label{vnumber-section}
Let $S=K[t_1,\ldots,t_s]=\bigoplus_{d=0}^{\infty} S_d$ be a polynomial ring over
a field $K$ with the standard grading and let $I$ be a graded ideal of
$S$. In this section we show a formula for the v-number of $I$ that
can be used to compute this number using \textit{Macaulay}$2$
\cite{mac2}.
To avoid repetitions, we continue to employ
the notations and
definitions used in Sections~\ref{intro-section} and
\ref{prelim-section}.
\begin{lemma}\label{mingens-lemma}
Let $I\subset S$ be a graded ideal. If $(I\colon f)=\mathfrak{p}$ for
some prime ideal $\mathfrak{p}$ and some $f\in S_d$, $d\geq 0$, then
$I\subsetneq(I\colon\mathfrak{p})$ and there is a
minimal homogeneous generator $\overline{g}:=g+I$ of
$(I\colon\mathfrak{p})/I$
such that $\deg(f)\geq \deg(g)$ and $(I\colon g)=\mathfrak{p}$.
\end{lemma}
\begin{proof} The strict inclusion $I\subsetneq(I\colon\mathfrak{p})$
is clear because $f\in(I\colon\mathfrak{p})\setminus I$. Let
$\mathcal{G}=\{\overline{g}_1,\ldots,\overline{g}_r\}$ be a minimal
generating set of $(I\colon\mathfrak{p})/I$ such that
$g_i$ is a homogeneous polynomial for all $i$. As
$(I\colon f)=\mathfrak{p}$, one has $\overline{f}\neq\overline{0}$ and
$f\in(I\colon\mathfrak{p})$. Then, we can choose homogeneous
polynomials $h_1,\ldots,h_r$ in $S$, $p$ in $I$, such that
$f=\sum_{i=1}^rh_ig_i+p$ and $d=\deg(h_ig_i)$ for all $i$ with
$h_i\neq 0$. One
has the inclusion $\bigcap_{i=1}^r(I\colon g_ih_i)\subset(I\colon
f)$. Indeed, take $h$ in $\bigcap_{i=1}^r(I\colon g_ih_i)$, then
$hh_ig_i\in I$ for all $i$ and $hf=\sum_{i=1}^rhh_ig_i+hp$ is in $I$,
thus $h$ is in $(I\colon f)$.
Therefore, using that all $g_i$'s are in
$(I\colon\mathfrak{p})$, one has the inclusions
$$
\mathfrak{p}\ \mathlarger{\subset}\bigcap_{i=1}^r(I\colon g_i)\
\mathlarger{\subset}\bigcap_{i=1}^r(I\colon g_ih_i)\
\mathlarger{\subset}(I\colon f)=\mathfrak{p},
$$
and consequently $\mathfrak{p}=\bigcap_{i=1}^r(I\colon g_ih_i)$. Hence,
by \cite[p.~74, 2.1.48]{monalg-rev}, we get $(I\colon
g_ih_i)=\mathfrak{p}$ for some $1\leq i\leq r$. As $g_i$ is in
$(I\colon\mathfrak{p})$, we obtain
$$
\mathfrak{p}\subset(I\colon g_i)\subset(I\colon g_ih_i)=\mathfrak{p}.
$$
\quad Hence $\mathfrak{p}=(I\colon g_i)$ and
$d=\deg(f)=\deg(g_ih_i)\geq\deg(g_i)$.
\end{proof}
\begin{theorem}\label{vnumber-general}
Let $I\subset S$ be a graded ideal and let $\mathfrak{p}\in{\rm
Ass}(I)$. The following hold.
\begin{enumerate}
\item[(a)]
If $\mathcal{G}=\{\overline{g}_1,\ldots,\overline{g}_r\}$ is
a homogeneous minimal generating set of $(I\colon\mathfrak{p})/I$,
then
$$
{\rm v}_{\mathfrak{p}}(I)=
\min\{\deg(g_i)\mid 1\leq i\leq r\mbox{ and }(I\colon
g_i)=\mathfrak{p}\}.
$$
\item[(b)] ${\rm v}(I)=\min\{{\rm v}_{\mathfrak{q}}(I)\mid
\mathfrak{q}\in{\rm Ass}(I)\}$.
\item[(c)] ${\rm
v}_{\mathfrak{p}}(I)\geq\alpha((I\colon\mathfrak{p})/I)$ with equality
if $\mathfrak{p}\in{\rm Max}(I)$.
\item[(d)] If $I$ has no embedded primes, then $
{\rm v}(I)=\min\{\alpha\left((I\colon\mathfrak{q})/{I}\right)\vert\,
\mathfrak{q}\in{\rm Ass}(I)\}.
$
\end{enumerate}
\end{theorem}
\begin{proof} (a): Take any homogeneous polynomial $f$ in $S$ such that $(I\colon
f)=\mathfrak{p}$. Then, by Lemma~\ref{mingens-lemma}, there is
$g_j\in\mathcal{G}$ such that $\deg(f)\geq \deg(g_j)$ and $(I\colon
g_j)=\mathfrak{p}$. Thus, the set $\{g_i\mid (I\colon
g_i)=\mathfrak{p}\}$ is not empty and the inequality
$$
{\rm v}_{\mathfrak{p}}(I)\leq \min\{\deg(g_i)\mid 1\leq i\leq r\mbox{
and }(I\colon
g_i)=\mathfrak{p}\}
$$
follows by definition of ${\rm v}_{\mathfrak{p}}(I)$. Now, we can pick a homogeneous
polynomial $f$ in $S$ such
that $\deg(f)={\rm v}_\mathfrak{p}(I)$ and $(I\colon
f)=\mathfrak{p}$ . Then, by Lemma~\ref{mingens-lemma}, there is
$g_j\in\mathcal{G}$ such that $\deg(f)\geq \deg(g_j)$ and $(I\colon
g_j)=\mathfrak{p}$. Thus, $\deg(f)=\deg(g_j)$ and the inequality
``$\geq$'' holds.
(b): This follows at once from the definitions of ${\rm v}(I)$ and
${\rm v}_{\mathfrak{q}}(I)$.
(c): Pick a homogeneous polynomial $g$ in $S$ such that $\deg(g)={\rm
v}_{\mathfrak{p}}(I)$ and $(I\colon g)=\mathfrak{p}$. Then, $g\notin
I$ and $g\mathfrak{p}\subset I$, that is, $g$ is in
$(I\colon\mathfrak{p})\setminus I$. Thus ${\rm
v}_{\mathfrak{p}}(I)\geq\alpha((I\colon\mathfrak{p})/I)$. Now,
assume that $\mathfrak{p}$ is in ${\rm Max}(I)$. To show
the reverse inequality take any homogeneous polynomial $f$ in
$(I\colon\mathfrak{p})\setminus I$. Then $f\mathfrak{p}\subset I$ and
$\mathfrak{p}\subset(I\colon f)$. Since ${\rm Ass}(I\colon f)$ is
contained in ${\rm Ass}(I)$, there is $\mathfrak{q}\in{\rm Ass}(I)$
such that $\mathfrak{p}\subset (I\colon f)\subset\mathfrak{q}$. Hence,
$\mathfrak{p}=\mathfrak{q}$ and $\mathfrak{p}=(I\colon f)$. Thus
${\rm v}_{\mathfrak{p}}(I)\leq\deg(f)$ and
${\rm v}_{\mathfrak{p}}(I)\leq\alpha((I\colon\mathfrak{p})/I)$.
(d): This follows immediately from (b) and (c).
\end{proof}
We give a direct proof of the next result that in particular relates the v-number of
a Cohen--Macaulay monomial ideal $I\subset S$ with that of $(I,h)$,
where $h\in S_1$ and $(I\colon h)=I$.
\begin{corollary}\cite[Proposition~4.9]{v-number}
Let $I\subset S$ be a Cohen--Macaulay non-prime graded ideal whose associated primes are
generated by linear forms and let $h\in S_1$ be a regular element on
$S/I$. Then ${\rm v}(I,h)\leq{\rm v}(I)$.
\end{corollary}
\begin{proof} By Theorem~\ref{vnumber-general},
there are $\mathfrak{p}\in{\rm Ass}(I)$ and
$f\in(I\colon\mathfrak{p})\setminus I$ such that $\overline{f}=f+I$ is
a minimal generator of $M_\mathfrak{p}=(I\colon\mathfrak{p})/I$ and
$\deg(f)={\rm v}(I)$. The
associated primes of $(I\colon f)$ are contained in ${\rm Ass}(I)$,
thus there is $\mathfrak{q}\in{\rm Ass}(I)$
such that $\mathfrak{p}\subset(I \colon f)\subset\mathfrak{q}$. Hence,
$\mathfrak{p}=\mathfrak{q}$ because $I$ has no embedded associated
primes, and one has the equality $(I\colon f)=\mathfrak{p}$. We claim
that $f$ is not in $(I,h)$.
By contradiction assume that $f\in(I,h)$. Then we can write
$f=f_1+hf_2$, with $f_i$ a homogeneous polynomial for $i=1,2$, $f_1\in I$,
$f_2\in S$. Hence, one has
$$
\mathfrak{p}=(I\colon f)=(I\colon hf_2)=(I\colon f_2).
$$
\quad Therefore $f_2\in(I\colon\mathfrak{p})\setminus I$ and
$\overline{f}=\overline{h}\,\overline{f_2}$, a contradiction because
$\overline{f}$ is a minimal generator of $M_\mathfrak{p}$. This proves that $f\notin(I,h)$. Next we show
the equality $(\mathfrak{p},h)=((I,h)\colon f)$. The inclusion
``$\subset$'' is clear because $(I\colon f)=\mathfrak{p}$. Take an
associated prime $\mathfrak{p}'$ of $((I,h)\colon f)$. The height of
$\mathfrak{p}'$ is equal to ${\rm ht}(I)+1$ because $(I,h)$ is
Cohen--Macaulay and the associated primes of $((I,h)\colon f)$ are
contained in ${\rm Ass}(I,h)$. Then
$$
\mathfrak{p}=(I\colon f)\subset ((I,h)\colon f)\subset\mathfrak{p}',
$$
and consequently $(\mathfrak{p},h)\subset ((I,h)\colon
f)\subset\mathfrak{p}'$. Now, $(\mathfrak{p},h)$ is prime because
$\mathfrak{p}$ is generated by linear forms, and ${\rm
ht}(\mathfrak{p},h)={\rm ht}(\mathfrak{p})+1={\rm ht}(I)+1$ because
$I$ is Cohen--Macaulay and $h$ is a regular element on $S/I$. Thus,
$(\mathfrak{p},h)=\mathfrak{p}'$, $(\mathfrak{p},h)=((I,h)\colon
f)$, and ${\rm v}(I,h)\leq {\rm v}(I)$.
\end{proof}
\begin{proposition}\label{reg-dim-pol} Let $I\subset S$ be a monomial ideal minimally generated
by $G(I)$ and for each $t_i$ that occurs
in a monomial of $G(I)$ let $\gamma_i:=\max\{\deg_{t_i}(g)\vert\, g\in G(I)\}$.
Then
$$
{\rm reg}(S/I)\leq\dim(S/I)+\textstyle\sum_{i}(\gamma_i-1).
$$
\end{proposition}
\begin{proof} To show this inequality we use the polarization
technique due to Fr\"oberg
(see \cite{depth-monomial} and \cite[p.~203]{monalg-rev}).
To polarize $I$ we use the set of new variables
$$
T_I=\textstyle\bigcup_{i=1}^n\{t_{i,2},\ldots,
t_{i,\gamma_i}\},
$$
where $\{t_{i,2},\ldots,t_{i,\gamma_i}\}$ is empty if $\gamma_i=1$.
Note that $|T_I|=\sum_i(\gamma_i-1)$. We identify the variable $t_i$ with $t_{i,1}$ for all
$i$. A power $t_i^{c_i}$ of a variable $t_i$, $1\leq
c_i\leq\gamma_i$, polarizes
to $(t_i^{c_i})^{\rm pol}=t_i$ if $\gamma_i=1$, to
$(t_i^{c_i})^{\rm pol}=t_{i,2}\cdots t_{i,c_i+1}$ if $c_i<\gamma_i$, and
to $(t_i^{c_i})^{\rm pol}=t_{i,2}\cdots t_{i,\gamma_i}t_i$ if $c_i=\gamma_i$.
Setting $G(I)=\{g_1,\ldots,g_r\}$, the polarization $I^{\rm pol}$ of $I$ is the ideal
of $S[T_I]$ generated by $g_1^{\rm pol},\ldots,g_r^{\rm pol}$.
According to \cite[Corollary 1.6.3]{Herzog-Hibi-book} one has
$${\rm reg}(S/I)={\rm reg}(S[T_I]/I^{\rm pol})\mbox{ and }{\rm
ht}(I)={\rm ht}(I^{\rm pol}).$$
\quad As $I^{\rm pol}$ is squarefree, by \cite[Proposition 3.2]{v-number},
one has ${\rm reg}(S[T_I]/I^{\rm pol})\leq\dim(S[T_I]/I^{\rm pol})$.
Hence, we obtain
$$
{\rm reg}(S/I)={\rm reg}(S[T_I]/I^{\rm pol})\leq\dim(S[T_I]/I^{\rm
pol})=\dim(S[T_I])-{\rm
ht}(I).
$$
\quad To complete the proof notice that $\dim(S[T_I])-{\rm
ht}(I)=\dim(S/I)+|T_I|$.
\end{proof}
A result of Beintema \cite{beintema} shows that a zero-dimensional
monomial ideal is Gorenstein if and only if it is a complete
intersection. The next result classifies the complete intersection
property using the regularity.
\begin{proposition} Let $I$ be a monomial ideal of $S$ of dimension
zero minimally generated by
$G(I)=\{t_1^{d_1},\ldots,t_s^{d_s},t^{d_{s+1}},\ldots,t^{d_m}\}$, where
$d_i\geq 1$ for $i=1,\ldots,s$ and $d_i\in\mathbb{N}^s\setminus\{0\}$
for $i>s$. Then ${\rm reg}(S/I)\leq\sum_{i=1}^s(d_i-1)$, with
equality if and only if $I$ is a complete intersection.
\end{proposition}
\begin{proof} The inequality ${\rm reg}(S/I)\leq\sum_{i=1}^s(d_i-1)$
follows directly from Proposition~\ref{reg-dim-pol} because $\dim(S/I)=0$. If
$I$ is a complete intersection, then $I=(t_1^{d_1},\ldots,t_s^{d_s})$
and, by \cite[Lemma~3.5]{Chardin}, we get ${\rm
reg}(S/I)=\sum_{i=1}^s(d_i-1)$. Conversely assume that ${\rm
reg}(S/I)$ is equal to $\sum_{i=1}^s(d_i-1)$. We proceed by
contradiction assuming that $m>s$. Then the exponents of the monomial
$t^{d_m}=t_1^{c_1}\cdots t_s^{c_s}$ satisfy $c_i\leq d_i-1$ for
$i=1,\ldots,s$ because $t^{d_m}\in G(I)$. The regularity of $S/I$ is
the largest integer $d\geq 0$ such that $(S/I)_d\neq(0)$
\cite[Proposition~4.14]{eisenbud-syzygies}. Pick a monomial
$t^a=t_1^{a_1}\cdots t_s^{a_s}$ such that $t^a\in S_d\setminus I$ and
$d=\sum_{i=1}^s(d_i-1)$. Then, $a_i\leq d_i-1$ for $i=1,\ldots,s$
because $t^a$ is not in $I$, and consequently $a_i=d_i-1$ for
$i=1,\ldots,s$. Hence, $t^a=t^\delta t^{d_m}$ for some
$\delta\in\mathbb{N}^s$, a contradiction.
\end{proof}
\section{Induced matchings and the \text{v}-number}
In this section we show that the
induced matching number of a graph $G$ is an upper
bound for the v-number of $I(G)$ when $G$ is very well-covered, or $G$
has a simplicial partition, or $G$ is well-covered connected and
contain neither $4$- nor $5$-cycles. We classify when the upper bound holds when $G$ is a
cycle, and classify when all vertices of a graph $G$ are shedding
vertices, we use this to gain insight on $W_2$-graphs. To avoid repetitions, we continue to employ
the notations and definitions used in Sections~\ref{intro-section} and
\ref{prelim-section}.
\begin{theorem}\label{v(I)}{\rm \cite[Theorem~3.5]{v-number}}
If $I = I(G)$ is the edge ideal of a graph $G$, then
$\mathcal{F}_G\subset\mathcal{A}_G$ and the $\mathrm{v}$-number of $I$ is
$$
\mathrm{v}(I) = \min\{\vert A \vert : A\in\mathcal{A}_G\}.
$$
\end{theorem}
\begin{lemma}\label{stable}
Let $A$ be a stable set of a graph $G$. If $N_G(A)$ is a vertex cover
of $G$, then
$A \in {\mathcal A}_G$.
\end{lemma}
\begin{proof}
We take any $b \in N_G(A)$, then there is $e \in E(G)$ such that $e \subset A \bigcup \{b\}$.
Furthermore, $N_G(A) \bigcap A = \emptyset$, since $A$ is a stable set
of $G$.
Thus,
$$
e \textstyle\bigcap N_G(A) \subset (A\textstyle \bigcup
\{b\})\textstyle \bigcap N_G(A) \subset \{b\},
$$
and consequently $e\bigcap(N_G(A)\setminus\{b\})=\emptyset$.
Hence, $N_G(A) \setminus \{ b\}$ is not a vertex cover of $G$, since
$e \in E(G)$. Therefore $N_G(A)$ is a minimal vertex cover of $G$ and
$A \in {\mathcal A}_G$
\end{proof}
\begin{theorem}\label{Domi-InduceMatch}
Let $G$ be a very well-covered graph and let $P=\{e_1, \ldots, e_r\}$
be a perfect matching of $G$.
Then, there is an induced submatching $P'$ of $P$ and $D \in {\mathcal A}_G$
such that $D \subset V(P')$ and $\vert e \bigcap D \vert = 1$ for
each $e\in P'$.
Furthermore ${\rm v}(I(G))\leq|P'|=|D|\leq{\rm im}(G)\leq{\rm
reg}(S/I(G))$.
\end{theorem}
\begin{proof} To show the first part we use induction on $\vert P
\vert$. If $r=1$, we set $P'=P=\{e_1\}$ and $D=\{x_1\}$, where
$e_1=\{x_1,y_1\}$. Assume $r>1$. We set $e_r=\{x,x'\}$,
$G_1: = G\setminus\{x,x'\}$ and $P_1:= P \setminus \{e_r\}$.
By Theorem~\ref{konig}, $P$ satisfies the property {\bf{(P)}}.
Then, $P_1$ satisfies the property {\bf{(P)}} as well. Thus, by Theorem \ref{konig},
$G_1$ is very well-covered with a perfect matching $P_1$. Hence, by induction hypothesis, there is
an induced submatching $P'_1$ of $P_1$ and $D_1 \in {\mathcal A}_{G_1}$ such that $D_1 \subset V(P'_1)$
and $\vert e \bigcap D_1 \vert = 1$ for each $e \in P'_1$. Consequently, $N_{G_1}(D_1)$ is a minimal vertex
cover of $G_1$. We will consider
two cases: $e_r \bigcap N_G(D_1) \neq \emptyset$ and $e_r \bigcap N_G(D_1) = \emptyset$.
Case (I): Assume that $e_r \bigcap N_G(D_1) \neq \emptyset$. Thus, we
may assume that there is $\{x,d\} \in E(G)$
with $d \in D_1$. Then, $N_G(x')\subset N_G(d) \subset N_G(D_1)$, since $P$ satisfies
property {\bf{(P)}}. Hence, $N_G(D_1)$ is a vertex cover of $G$, since $N_{G_1}(D_1)$ is a vertex cover of
$G_1$ and $\{x\} \subset N_G(x') \subset N_G(D_1)$. Therefore, by
Lemma~\ref{stable},
$D_1 \in {\mathcal A}_G$, so this case follow by making $D=D_1$ and $P'=P'_1$.
Case (II): Assume that $e_r \bigcap N_G(D_1) = \emptyset$. We set
$D_2:=V(P'_1) \setminus D_1$, then $D_2$
is a stable set of $G_1$ and also of $G$, since $P'_1$ is an induced matching
of $G_1$ and also of $G$. One has the
inclusion
\begin{equation}\label{sep10-21-1}
V(P'_1)\textstyle \bigcap (N_G(x)\textstyle\bigcup N_G(x')) \subset
D_2,
\end{equation}
indeed take $z\in V(P'_1)\bigcap N_G(x)$ (the case $z\in
V(P'_1)\bigcap N_G(x')$ is similar). If $z\notin D_2$, then $z\in
D_1\bigcap N_G(x)$, $\{z,x\}\in E(G)$, and $x\in e_r\bigcap N_G(D_1)$, a
contradiction. We claim that $\vert e_r \bigcap N_G(D_2) \vert \leq
1$. By contradiction suppose
$x,x' \in N_G(D_2)$, then there are $d_1,d_2 \in D_2$ such that
$\{x,d_1\},\{x',d_2\} \in E(G)$. Then $\{d_1,d_2\} \in E(G)$, since $P$ satisfies
property {\bf{(P)}}, a contradiction, since $D_2$ is a stable set of
$G$.
Hence, $\vert e_r \bigcap N_G(D_2) \vert \leq 1$ and we may assume
\begin{equation}\label{sep10-21-2}
e_r \textstyle\bigcap N_G(D_2) \subset \{x\}.
\end{equation}
\quad Next we show that $V(P_1')\bigcap N_G(x')=\emptyset$. If the
intersection is non-empty, by Eq.~\eqref{sep10-21-1}, we can
pick $z$ in $D_2\bigcap N_G(x')$, then $\{z,x'\}\in E(G)$ and $x'\in
N_G(D_2)$, a contradiction to Eq.~\eqref{sep10-21-2}. Therefore, by
Eq.~\eqref{sep10-21-1}, we obtain the inclusion
$$
V(P'_1)\textstyle \bigcap (N_G(x)\textstyle\bigcup N_G(x')) \subset
D_2 \textstyle\bigcap N_G(x) =:A_2,
$$
\quad Thus, the edge set $Q:=\{e \in P'_1 \mid e \bigcap A_2 =
\emptyset \} \bigcup \{e_r\}$, is
an induced matching, since $P'_1$ is an induced matching.
Setting
\begin{equation}\label{oct2-21}
D_3:=\{ y \in D_1\mid \{y,y'\} \in P'_1 \mbox{ with } y'\notin A_2\}\textstyle \bigcup \{x\},
\end{equation}
i.e., $D_3=(D_1 \bigcap V(Q)) \bigcup \{x\}$, we get $\vert f \bigcap D_3 \vert = 1$ for each $f\in Q$,
since $\vert e \bigcap D_1 \vert = 1$ for each $e \in P'_1$.
Note that $D_3$ is a stable set of $G$, since $D_1$ is a stable set and
$\{x\} \bigcap N_G(D_1) = \emptyset$.
Now, take $ e\in E(G)$. We will prove that $e \bigcap N_G(D_3) \neq
\emptyset$. Clearly $N_G(x)\subset N_G(D_3)$ because $x\in D_3$.
If $x' \in e$, then $x' \in e \bigcap N_G(x) \subset e \bigcap N_G(D_3)$. Now, if
$x \in e$, then $e=\{x,y\}$ for some $y$ in $V(G)$, and $y \in e \bigcap N_G(x) \subset e \bigcap N_G(D_3)$.
So, we may assume $e \bigcap \{x,x'\}=\emptyset$, then $e \in E(G_1)$.
Thus, there is $z \in e \bigcap N_{G_1}(D_1)$, since $N_{G_1}(D_1)$
is a vertex cover of $G_1$. Then, there is $d \in D_1$, such that
$z \in N_{G_1}(d)$. If $d \in D_3$, then $z \in N_G(D_3) \bigcap e$.
Finally, if $d \notin D_3$, then by Eq.~\eqref{oct2-21} and the inclusion
$D_1\subset V(P_1')$, there is $d' \in A_2$ such that $\{d,d'\} \in P'_1$.
So, $\{x,d'\} \in E(G)$, since $d' \in A_2$ .
This implies, $ \{x,z\} \in E(G)$, since $\{d,z\}\in E(G)$,
$\{x,d'\}\in E(G)$, $\{d,d'\} \in P$, and
$P$ satisfies property {\bf{(P)}}. Thus, $z \in e \bigcap N_G(x) \subset e \bigcap N_G(D_3)$.
Hence, $N_G(D_3)$ is a vertex cover and, by Lemma~\ref{stable}, $D_3 \in {\mathcal A}_G$.
Therefore, this case follows by making $P'=Q$ and $D=D_3$. This
completes the induction process.
Next we show the equality $|P'|=|D|$. By the first part, we may assume that
$P'=\{e_1,\ldots,e_\ell\}$, $1\leq \ell\leq r$, $e_i=\{x_i,y_i\}$ for
$i=1,\ldots,\ell$, and $x_1,\ldots,x_\ell\in D$. Thus, $\ell=|P'|\leq
|D|$ and, since $D\subset V(P')$, we get $2|D|\leq 2|P'|$.
Then $|P'|=|D|$. The inequality ${\rm v}(I(G))\leq |D|$ follows by Theorem~\ref{v(I)}
and $|P'|\leq{\rm im}(G)$ is clear by definition of ${\rm
im}(G)$. Finally, the inequality ${\rm im}(G)\leq{\rm
reg}(S/I(G))$ follows directly from Theorem~\ref{lower-bound-reg}.
\end{proof}
\begin{corollary}\cite[Theorem 3.19(b)]{v-number}\label{sep29-21} Let $G$ be a graph and let $W_G$ be
its whisker
graph. Then
$$
{\rm v}(I(W_G))\leq{\rm reg}(K[V(W_G)]/I(W_G)).
$$
\end{corollary}
\begin{proof} By Lemma~\ref{bipartite-whiskers}, $W_G$ is very
well-covered. Thus, by Theorem~\ref{Domi-InduceMatch}, the v-number of
$I(W_G)$ is bounded from above by the regularity of
$K[V(W_G)]/I(W_G)$.
\end{proof}
\begin{lemma}\label{bound}
Let $\ell\geq 0$ and $s=4\ell+r$ be integers with $r \in \{0,1,2,3\}$.
If $s \geq 3$ and $s \neq 5$, then
$$
\left\lfloor \frac{s} {3} \right\rfloor \geq \ell \mbox { if } r=0 \mbox{ and } \left\lfloor \frac{s}{3} \right\rfloor \geq \ell+1 \mbox{ otherwise}.
$$
\end{lemma}
\begin{proof}
By the division algorithm, $ s\equiv r' \pmod{3}$, where $r' \in \{0,1,2\}$. Then
$$
\left\lfloor \frac{s} {3} \right\rfloor= \frac{4\ell+r-r'} {3}=\ell+ \frac{\ell+r-r'} {3} \in \mathbb{Z}.
$$
\quad Thus, $a:= \frac{\ell+r-r'} {3} \in \mathbb{Z}$. If $r=0$, then
$a \geq 0$. This follows using that $0\leq
r'\leq 2$ and $\ell\geq 0$. Hence, $ \left\lfloor \frac{s} {3}
\right\rfloor \geq \ell $. Now, assume
$r \in \{1, 2, 3\}$. We claim that $a \geq 1$. By contradiction assume
that $a \leq 0$, then $\ell+r \leq r'$. If $\ell=0$, then $s=r=3$,
since $s \geq 3$. A contradiction, since $3=\ell+r \leq r'$ and $r' \leq 2$.
Thus, $\ell \geq 1$ and we have $2 \leq \ell+1 \leq \ell +r \leq r' \leq 2$. This implies
$\ell=1=r$ and $r'=2$. Consequently $s=5$, a contradiction.
Therefore, $a \geq 1$ and $\left\lfloor \frac{s}{3} \right\rfloor \geq \ell+1$.
\end{proof}
\begin{theorem}\label{cycles-indmat}
Let $C_s$ be an $s$-cycle and let $I(C_s)$ be its edge ideal. Then,
${\rm v}(I(C_s))\leq{\rm im}(C_s)$ if and only if $s \neq 5$.
\end{theorem}
\begin{proof}
$\Rightarrow$) Assume that ${\rm v}(I(C_s))\leq{\rm im}(C_s)$. If $s=5$,
then ${\rm v}(I(C_s))=2$ and ${\rm im}(C_s)=1$, a contradiction. Thus
$s\neq 5$.
$\Leftarrow$) Assume that $s\neq 5$. We can write
$C_s=(t_1,e_1,t_2,\ldots, t_i,e_i,t_{i+1}, \ldots, t_s,e_s,t_1)$. The
matching $P=\{e_1,e_4, \ldots, e_{3q-2}\}$, where $q:=\lfloor \frac{s}{3} \rfloor$,
is an induced matching of $C_s$ and $\vert P \vert=q$.
Now we choose a stable set $A$ of $C_s$, for each one of the following cases:
Case $s=4\ell$. If $A=\{t_2, t_6, \ldots, t_{4\ell-2}\}$, then
$N_{C_s}(A)=\{t_1,t_3,t_5,t_7, \ldots, t_{s-3},t_{s-1}\}$ is a vertex
cover of $G$ and $\vert A \vert =\ell$.
Case $s=4\ell+1$. If $A=\{t_2, t_6, \ldots, t_{4\ell-2}\} \bigcup \{t_{4\ell}\}$, then
$N_{C_s}(A)=\{t_1,t_3, \ldots, t_{s-4},t_{s-2}\} \bigcup \{t_s\}$ is a
vertex cover of $G$ and
$\vert A \vert =\ell+1$.
Case $s=4\ell+2$. If $A=\{t_2, t_6, \ldots, t_{4\ell+2}\}$, then
$N_{C_s}(A)=\{t_1,t_3,t_5,t_7, \ldots, t_{s-3},t_{s-1}\}$ is a vertex
cover of $G$ and
$\vert A \vert =\ell+1$.
Case $s=4\ell+3$. If $A=\{t_2, t_6, \ldots, t_{4\ell+2}\}$, then
$N_{C_s}(A)=\{t_1,t_3,t_5,t_7, \ldots, t_{s-2},t_s\}$ is a vertex
cover of $G$ and
$\vert A \vert =\ell+1$.
In each case $N_{C_s}(A)=\{t_i \vert i \mbox{ is odd} \}$ and
$N_{C_s}(A)$ is a vertex cover of $G$.
So, by Lemma~\ref{stable}, $A \in {\mathcal A}_{C_s}$. Now, assume
$s=4\ell+r$, with $r \in \{0,1,2,3\}$ and $\ell\geq 0$ an integer.
Then, by Lemma~\ref{bound},
$\lfloor \frac{s} {3} \rfloor \geq \ell$
if $ r=0$ and $\lfloor \frac{s}{3} \rfloor \geq \ell+1$ otherwise.
Hence, $\vert P \vert = \lfloor \frac{s}{3} \rfloor \geq \vert A \vert$.
Therefore, ${\rm im}(C_s) \geq
{\rm v}(I(C_s))$, since ${\rm im}(C_s) \geq \vert P \vert$
and $\vert A \vert \geq{\rm v}(I(C_s))$.
\end{proof}
\begin{remark} The induced matching number of the cycle $C_s$ is equal to
$\lfloor\frac{s}{\scriptstyle 3} \rfloor$. The regularity of
$S/I(C_s)$ is equal to $\lfloor (s+1)/3\rfloor$
\cite[Proposition~10]{woodroofe-matchings}.
\end{remark}
\begin{lemma}\label{induced-lemma}
Let $G$ be a graph without isolated vertices and let $z_1,\ldots,z_m$
be vertices of $G$ such that $\{N_G[z_i]\}_{i=1}^m$ is a partition of
$V(G)$. If $G_1=G\setminus N_G[z_m]$, then
\begin{enumerate}
\item[(i)] $N_{G_1}[z_i]=N_G[z_i]$ for $i<m$, and
\item[(ii)] $G_1[N_{G_1}[z_i]]=G[N_G[z_i]]$ for $i<m$.
\end{enumerate}
\end{lemma}
\begin{proof} (i): Assume that $1\leq i\leq m-1$. Clearly
$N_{G_1}[z_i]\subset N_G[z_i]$ because $G_1$ is a subgraph of $G$. To
show the inclusion ``$\supset$'' take $z\in N_G[z_i]$. Then, $z=z_i$
or $\{z,z_i\}\in E(G)$. If $z\in
N_G[z_m]$, then $z\in N_G[z_m]\bigcap N_G[z_i]$,
a contradiction. Thus, $z\notin N_G[z_m]$ and, since $G_1$ is an
induced subgraph of $G$, we get $z=z_i$ or $\{z,z_i\}\in E(G_1)$.
Thus, $z\in N_{G_1}[z_i]$.
(ii): By part (i), one has
$N_G[z_i]=N_{G_1}[z_i]\subset V(G)\setminus N_G[z_m]=V(G_1)$. Then
\begin{align*}
E(G[N_G[z_i]])&=\{e\in E(G)\mid e\subset N_G[z_i]\}=\{e\in E(G)\mid
e\subset N_{G_1}[z_i]\} \\
&=\{e\in E(G_1)\mid e\subset N_{G_1}[z_i]\}=E(G_1[N_{G_1}[z_i]]).
\end{align*}
\quad Thus, $E(G[N_G[z_i]])=E(G_1[N_{G_1}[z_i]])$.
\end{proof}
\begin{theorem}\label{Domi-InduceMatch-simplex}
Let $G$ be a graph with simplexes $H_1, \ldots, H_r$,
such that $\{V(H_i)\}_{i=1}^r$ is a partition of $V(G)$.
If $G$ has no isolated vertices, then there
is $D=\{y_1,\ldots,y_k\} \in {\mathcal A}_G$,
and there are simplicial vertices $x_1,\ldots,x_k$ of $G$ and integers
$1\leq j_1< \cdots <j_k \leq r$ such that $P=\{\{x_i,y_i\}\}_{i=1}^k$ is an induced matching of $G$ and
$H_{j_i}$ is the induced subgraph $G[N_G[x_i]]$ on $N_G[x_i]$ for $i=1,\ldots,k$. Furthermore ${\rm
v}(I(G))\leq|D|=|P|\leq{\rm im}(G)\leq{\rm reg}(S/I(G))$.
\end{theorem}
\begin{proof} We proceed by induction on $r$. If $r=1$, then $V(H_1)=V(G)$ and there is a
simplicial vertex $x_1$ of $G$ such that $H_1=G[N_G[x_1]]$ is a
complete graph with at least two vertices. Picking $y_1\in
N_G[x_1]$, $y_1\neq x_1$, one has $\{x_1\}\in\mathcal{A}_G$ and
$\{x_1,y_1\}$ is an induced matching. Now assume that $r>1$.
We set $G_1: = G \setminus V(H_r)$. Note that
$H_1,\ldots,H_{r-1}$ are simplexes of $G_1$
(Lemma~\ref{induced-lemma}) and $\{V(H_i)\}_{i=1}^{r-1}$ is a
partition of $V(G_1)$. Then, by induction hypothesis, there is
$D_1 =\{y_1,\ldots, y_{k'}\} \in {\mathcal A}_{G_1}$, and there are
simplicial vertices $x_1,\ldots,x_{k'}$ of $G_1$ and integers
$1\leq j_1<\cdots <j_{k'} \leq r-1$, such that
$P_1=\{\{x_1,y_1\}, \ldots, \{x_{k'},y_{k'}\}\}$
is an induced matching of $G_1$ and
$H_{j_i}=G_1[N_{G_1}[x_i]]$ for $i=1,\ldots,k'$. By
Lemma~\ref{induced-lemma}, one has
$G_1[N_{G_1}[x_i]]=G[N_G[x_i]]$ for $i=1,\ldots,k'$. We can
write $H_r=G[N_G[x]]$ for some simplicial vertex $x$ of $G$.
Case (I): Assume that $V(H_r) \setminus \{x\} \subset N_G(D_1)$. Then, $N_G(D_1)$ is a
vertex cover of $G$. Indeed, take any edge $e$ of $G$. If $e\bigcap V(H_r)=\emptyset$, then
$e$ is an edge of $G_1$ and is covered by $N_{G_1}(D_1)$.
Assume that $e\cap V(H_r) \neq \emptyset$. If $x \notin e$, then there
is $z \in e$ with $z \in V(H_r)\setminus\{x\} \subset N_G(D_1)$. Now, if $x \in e$
then $e=\{x,z\}$ with $z \in N_G[x]\setminus\{x\}=V(H_r)\setminus\{x\} \subset N_G(D_1)$.
This proves that $N_G(D_1)$ is a vertex cover of
$G$. Hence,
by Lemma~\ref{stable}, $D_1 \in {\mathcal A}_G$ and, noticing that
$P_1$ is an induced matching of $G$, this case follows by making
$D=D_1$ and $P=P_1$.
Case (II): Assume that there is $y\in V(H_r) \setminus \{x\}$ such that $y \notin N_G(D_1)$.
Then, $D_2:=D_1 \bigcup \{y\}$ is a stable set of $G$. Also, $ N_G(D_2)$ is a vertex cover of $G$,
since $N_{G_1}(D_1)$ is a vertex cover of $G_1$, $H_r$ is a complete
subgraph of $G$, and $V(H_r) \subset N_G[y]$.
Thus, by Lemma~\ref{stable}, $D_2$ is in ${\mathcal A}_G$. We set $x_{k'+1}:=x$,
$y_{k'+1}:=y$ and $H_{j_{k'+1}}:=H_r$. Then, $\{x_{k'+1}, y_{k'+1}\} \in E(H_r)$ and
$P_2:=P_1 \bigcup \{\{x_{k'+1}, y_{k'+1}\} \}$ is an induced matching of $G$,
since $P_1$ is an induced matching of $G_1$, $y \in V(H_r)\setminus N_G(D_1)$
and $H_{j_i}=G[N_G[x_i]]$, for $i=1, \ldots, k'+1$ .
Therefore, this case follows by making $D=D_2$ and $P=P_2$.
The equality $|D|=|P|$ is clear.
The inequality ${\rm v}(I(G))\leq |D|$ follows from Theorem~\ref{v(I)}
and $|P|\leq{\rm im}(G)$ is clear by definition of ${\rm
im}(G)$. Finally, the inequality ${\rm im}(G)\leq{\rm
reg}(S/I(G))$ follows directly from Theorem~\ref{lower-bound-reg}.
\end{proof}
\begin{corollary}\label{simplicial-4-5-cycles}
Let $G$ be a well-covered graph and let $I(G)$ be its edge ideal. If $G$ is simplicial
or $G$ is connected and contain neither $4$- nor $5$-cycles, then
$${\rm v}(I(G))\leq{\rm im}(G)
\leq{\rm reg}(S/I(G))\leq\beta_0(G).$$
\end{corollary}
\begin{proof} Assume that $G$ is simplicial. Let $\{z_1,\ldots,z_\ell\}$ be
the set of all simplicial vertices of $G$. Then $V(G)=\bigcup_{i=1}^\ell
N_G[z_i]$. As $G$ is well-covered, by \cite[Lemma~2.4]{Finbow2}, for
$1\leq i<j\leq \ell$ either $N_G[z_i]=N_G[z_j]$ or
$N_G[z_i]\bigcap N_G[z_j]=\emptyset$. Thus there are simplicial
vertices $x_1,\ldots,x_k$ of $G$ such that $\{N_G[x_i]\}_{i=1}^k$ is a
partition of $V(G)$. Setting $H_i=G[N_G[x_i]]$ for $i=1,\ldots,k$ and
applying Theorem~\ref{Domi-InduceMatch-simplex}, we get that ${\rm
v}(I(G))\leq{\rm im}(G)\leq{\rm reg}(S/I(G))$. Noticing that
$\dim(S/I(G))=\beta_0(G)$, the inequality ${\rm
reg}(S/I(G))\leq\beta_0(G)$ follows from
Proposition~\ref{reg-dim-pol}.
Next assume that $G$ is connected and contain neither $4$- nor $5$-cycles. Then, by
Theorem~\ref{wellcovered-characterization1},
$G\in \{C_7,T_{10}\}$ or $G\in\mathcal{F}$.
The cases $G = C_7$ or $G=T_{10}$ are treated in
Example~\ref{example2} (cf. Theorem~\ref{cycles-indmat}).
If $G\in\mathcal{F}$, then there exists
$\{x_1,\ldots,x_k\}\subset V(G)$ where
for each $i$, $x_i$ is
simplicial, $|N_G[x_i]|\leq 3$ and $\{N_G[x_i]\mid i=1,\ldots,k\}$ is
a partition of $V(G)$. In particular $G$ is simplicial and the
asserted inequalities follow from the first part of the proof.
\end{proof}
\begin{proposition}\label{Shedding-stable}
Let $G$ be a graph. The following conditions are equivalent.
\begin{enumerate}
\item Every vertex of $G$ is a shedding vertex.
\item ${\mathcal A}_{G}={\mathcal F}_{G}$.
\end{enumerate}
\end{proposition}
\begin{proof}
(1) $\Rightarrow$ (2): The inclusion ${\mathcal
A}_{G}\supset{\mathcal F}_{G}$ follows from Theorem~\ref{v(I)}. To
show the inclusion ${\mathcal
A}_{G}\subset{\mathcal F}_{G}$ we proceed by contradiction, suppose
there is $D \in {\mathcal A}_{G} \setminus {\mathcal F}_{G}$.
Then, $D$ is a stable set of $G$ and $N_G(D)$ is a vertex cover of $G$.
Thus, $D \bigcap N_G(D) = \emptyset$. Furthermore, since $D \notin {\mathcal F}_{G}$,
there is $x \in V(G) \setminus D$ such that $D \bigcup \{x\}$ is a stable
set of $G$. Then, $x \notin N_G(D)$. But $N_G(D)$ is a vertex cover
of $G$, then
$N_G(x) \subset N_G(D)$ and $A:=V(G)\setminus N_G(D)$ is a stable set
of $G$.
So, $A \subset V(G)\setminus N_G(x)$ and $A':=A \setminus x$ is a stable set of $V(G)\setminus N_G[x]$.
Now, we prove $A'$ is a maximal stable set of $G \setminus x$.
By contradiction assume there is $a \in V(G \setminus x) \setminus A'$, such that
$A' \bigcup \{a\}$ is a stable set. Then, $a \in N_G(D)$, since $V(G)=A \bigcup N_G(D)$.
Also, $D\subset A'$, since $D \bigcap N_G(D) = \emptyset$ and
$x\notin D$, a contradiction, since $a \in N_G(D)$ and $A' \bigcup \{a\}$ is a stable set.
Hence, $A'$ is a maximal stable set of $G\setminus x$.
Therefore $x$ is not a shedding vertex of $G$, a contradiction.
(2) $\Rightarrow$ (1): By contradiction, suppose there is $x \in V(G)$
such that $x$ is not a shedding vertex.
Thus, there is a maximal stable set $A$ of $G \setminus x$ such that $A \subset V(G) \setminus N_G[x]$ .
Then, $C:=V(G\setminus x)\setminus A$ is a minimal vertex cover of $G\setminus x$
and $A \bigcup \{x\}$ is a stable set of $G$. So, $A \notin {\mathcal F}_G$.
Since $C$ is a minimal vertex cover of $G\setminus x$, we have that for each $z \in C$, there is
$z' \in V(G \setminus x) \setminus C = A$ such that $\{z,z'\} \in E(G)$. Consequently,
$C \subset N_G(A)$. Furthermore, if $a \in N_G(x)$, then $a \in G \setminus x$
and $a \notin A$. Thus, $a \in N_G(A)$, since $A$ is a maximal stable set of $G \setminus x$.
Hence, $N_G(x) \subset N_G(A)$. This implies, $N_G(A)$ is a vertex cover of $G$, since
$C \subset N_G(A)$. Therefore, by Lemma~\ref{stable},
$A \in {\mathcal A}_G$, a contradiction since $A \notin {\mathcal F}_G$.
\end{proof}
\begin{lemma}\cite[cf. Corollary
3.3]{Levit-Mandrescu}\label{shedding-w2} If $G\in W_2$, then
every $v\in V(G)$ is a shedding vertex.
\end{lemma}
\begin{proof} Let $v$ be a vertex of $G$. We may assume that $G$ is
not a complete graph. Let $A$ be a stable set of
$G_v:=G\setminus N_G[v]$. We proceed by contradiction assuming that
$A$ is a maximal stable set of $G\setminus v$. Then, as $G$ and
$G\setminus v$ are well-covered, we get
$$\beta_0(G)=\beta_0(G\setminus v)=|A|.$$
\quad According to \cite[Theorem~5]{Pinter-jgt}, the graph $G_v$ is in $W_2$ and
$\beta_0(G_v)=\beta_0(G)-1$. In particular $G_v$ is well-covered and
$\beta_0(G_v)=\beta_0(G)-1$ (cf. Theorem~\ref{Campbell-theo}).
But $A$ is a stable set of $G_v$ and $\vert A \vert =\beta_0(G)$, a contradiction.
\end{proof}
\begin{corollary}\cite[Theorem~4.5]{v-number}\label{w2-graph}
If $G$ is a $W_2$-graph and $I=I(G)$, then ${\rm v}(I)=\beta_{0}(G)$.
\end{corollary}
\begin{proof}
By Theorem \ref{v(I)}, there is $D \in {\mathcal A}_G$ such that
${\rm v}(I)=\vert D \vert$.
Since $G$ is a $W_2$-graph, by Lemma~\ref{shedding-w2}, every vertex of $G$ is
a shedding vertex.
Thus, by Proposition \ref{Shedding-stable}, $D \in {\mathcal F}_G$,
i.e., $D$ is a maximal
stable set of $G$. Furthermore, $G$ is well-covered, since $G$ is a $W_2$-graph. Hence,
$\vert D \vert=\beta_{0}(G)$. Therefore, ${\rm v}(I)=\beta_{0}(G)$.
\end{proof}
\section{Examples}\label{examples-section}
\begin{example}\label{example1}
Let $S=\mathbb{Q}[t_1,t_2,t_3]$ be a polynomial ring and
$I=(t_1^5,t_2^5,t_2^4t_3^5,t_1^4t_3^5)$. Then an irredundant
primary decomposition of $I$ is given by
$$
I=(t_1^4,t_2^4)\textstyle\bigcap(t_1^5,t_2^5,t_3^5).
$$
\quad The associated primes of $I$
are $\mathfrak{p}_1=(t_1,t_2)$ and $\mathfrak{p}_2=(t_1,t_2,t_3)$.
Setting $g_1=t_1^4t_2^4$, $g_2=t_1^3t_2^3t_3^5$, and
$g_3=t_1^4t_2^4t_3^4$, and using Procedure~\ref{procedure1}, we get that $(I\colon\mathfrak{p}_1)/I$ and
$(I\colon\mathfrak{p}_2)/I$ are minimally generated by
$\{\overline{g}_1,\overline{g}_2\}$
and $\{\overline{g}_3\}$, respectively. Using
Theorem~\ref{vnumber-general} and the equalities
$$
(I\colon g_1)=(t_1,t_2,t_3^5),\ (I\colon g_2)=\mathfrak{p}_1,\
(I\colon g_3)=\mathfrak{p}_2,
$$
we obtain that ${\rm v}(I)=11$. The regularity of the quotient ring
$S/I$ is equal to $12$.
\end{example}
\begin{example}\label{example-wc}
Let $S=\mathbb{Q}[t_1,\ldots,t_6]$ be a polynomial ring and let $I$ be
the edge ideal
$$
I=(t_1t_2,\, t_2t_3,\, t_3t_4,\, t_1t_4,\, t_1t_5,\, t_2t_5,\,
t_3t_5,\, t_4t_5,\, t_1t_6,\, t_2t_6,\, t_3t_6,\, t_4t_6).
$$
\quad The graph $G$ defined by the generators of this ideal is
well-covered and not very well-covered, $\alpha_0(G)=4$, and ${\rm v}(I)={\rm im}(G)={\rm
reg}(S/I)=1$.
\end{example}
\begin{example}\label{example2}
Let $C_7$ and $T_{10}$ be the well-covered graphs of
Figure~\ref{C7-T10}. Let $R$ and $S$ be polynomial rings over the
field $\mathbb{Q}$ in the
variables $\{t_1,\ldots,t_7\}$ and $\{t_1,\ldots,t_{10}\}$,
respectively.
Using \textit{Macaulay}$2$ \cite{mac2} and
Procedure~\ref{procedure1} we obtain ${\rm ht}(I(C_7))=\alpha_0(C_7)=4$, ${\rm
pd}(R/I(C_7))=5$ and
$$
{\rm v}(I(C_7))=2={\rm im}(C_7)={\rm
reg}(R/I(C_7))\leq\dim(R/I(C_7))=\beta_0(C_7)=3.
$$
\quad The neighbor set of $A=\{t_1,t_4\}$ in $C_7$ is
$N_{C_7}(A)=\{t_2,t_3,t_5,t_7\}$ and $N_{C_7}(A)$ is
a minimal vertex cover of $C_7$, that is, $A\in\mathcal{A}_{C_7}$.
Using \textit{Macaulay}$2$ \cite{mac2} and
Procedure~\ref{procedure1} we obtain ${\rm ht}(I(T_{10}))=\alpha_0(G)=6$, ${\rm
pd}(S/I(T_{10}))=7$ and
$$
{\rm v}(I(T_{10}))=2={\rm im}(T_{10})\leq{\rm
reg}(S/I(T_{10}))=3\leq\dim(S/I(T_{10}))=\beta_0(T_{10})=4.
$$
\quad The neighbor set of $A=\{t_1,t_4\}$ in $T_{10}$ is
$N_{T_{10}}(A)=\{t_2,t_3,t_5,t_7,t_8,t_{10}\}$ and $N_{T_{10}}(A)$ is
a minimal vertex cover of $T_{10}$, that is, $A\in\mathcal{A}_{T_{10}}$.
\begin{figure}[H]
\begin{tikzpicture}[line width=.5pt,scale=0.75]
\tikzstyle{every node}=[inner sep=1pt, minimum width=5.5pt]
\tiny{
\node (1) at (-11,1.5){$\bullet$};
\node (2) at (-9,0.8) {$\bullet$};
\node (3) at (-8.5,-0.5) {$\bullet$};
\node (4) at (-10,-1.5){$\bullet$};
\node (7) at (-13,0.8){$\bullet$};
\node (5) at (-12,-1.5) {$\bullet$};
\node (6) at (-13.5,-0.5) {$\bullet$};
\node at (-11,1.8){$t_{1}$};
\node at (-9,1.1) {$t_{2}$};
\node at (-8.2,-0.5){$t_{3}$};
\node at (-10,-1.8){$t_{4}$};
\node at (-13,1.1) {$t_{7}$};
\node at (-12,-1.8) {$t_{5}$};
\node at (-13.8,-0.5) {$t_{6}$};
\node at (-11,-2.5) {{\large $C_{7}$}};
\draw[-,line width=1pt] (1) to (2);
\draw[-,line width=1pt] (2) to (3);
\draw[-,line width=1pt] (3) -- (4);
\draw[-,line width=1pt] (4) -- (5);
\draw[-,line width=1pt] (5) -- (6);
\draw[-,line width=1pt] (6) -- (7);
\draw[-,line width=1pt] (7) -- (1);
}
\tiny{
\node (1) at (0,2){$\bullet$};
\node (2) at (2,0.8) {$\bullet$};
\node (3) at (2,-0.5) {$\bullet$};
\node (4) at (2,-2){$\bullet$};
\node (7) at (-2,0.8){$\bullet$};
\node (5) at (-2,-2) {$\bullet$};
\node (6) at (-2,-0.5) {$\bullet$};
\node (8) at (0,1){$\bullet$};
\node (9) at (0,0) {$\bullet$};
\node (10) at (0,-1) {$\bullet$};
\node at (-0,2.3){$t_{1}$};
\node at (2,1.1) {$t_{2}$};
\node at (2.3,-0.5){$t_{3}$};
\node at (2,-2.3){$t_{4}$};
\node at (-2,1.1) {$t_{7}$};
\node at (-2,-2.3) {$t_{5}$};
\node at (-2.3,-0.5) {$t_{6}$};
\node at (0.3,1) {$t_{8}$};
\node at (0.3,0) {$t_{9}$};
\node at (0,-1.3) {$t_{10}$};
\node at (-0,-2.8) {{\large $T_{10}$}};
\draw[-,line width=1pt] (1) to (2);
\draw[-,line width=1pt] (2) to (3);
\draw[-,line width=1pt] (3) -- (4);
\draw[-,line width=1pt] (4) -- (5);
\draw[-,line width=1pt] (5) -- (6);
\draw[-,line width=1pt] (6) -- (7);
\draw[-,line width=1pt] (7) -- (1);
\draw[-,line width=1pt] (1) -- (8);
\draw[-,line width=1pt] (8) -- (9);
\draw[-,line width=1pt] (9) -- (10);
\draw[-,line width=1pt] (5) -- (10);
\draw[-,line width=1pt] (4) -- (10);
}
\end{tikzpicture}
\caption{Two well-covered graphs with no $4$- or
$5$-cycles\quad\quad\quad\quad\quad}\label{C7-T10}
\end{figure}
\end{example}
\begin{example}\label{example-w2}
Let $G$ be the graph consisting of two disjoint $3$-cycles with
vertices $x_1,x_2,x_3$ and $y_1,y_2,y_3$. Take two disjoint
independent sets of $G$, say $A_1=\{x_1\}$ and $A_2=\{y_1\}$, to
verify that $G$ is a graph in $W_2$ note that $B_1=\{x_1,y_2\}$
and $B_2=\{y_1,x_2\}$ are maximum independent sets of $G$ containing
$A_1$ and $A_2$ and $|B_i|=\beta_0(G)=2$.
\end{example}
\begin{appendix}
\section{Procedures}\label{Appendix}
\begin{procedure}\label{procedure1}
Computing the v-number and other invariants of a graded ideal $I$
with \textit{Macaulay}$2$ \cite{mac2}. This procedure corresponds to
Example~\ref{example1}. One can compute other examples by changing the
polynomial ring $S$ and the generators of the ideal $I$.
\begin{verbatim}
S=QQ[t1,t2,t3]
I=ideal(t1^5,t2^5,t2^4*t3^5,t1^4*t3^5)
--This gives the dimension and the height of I
--If I=I(G), G a graph, this gives the stability
--number and the covering number of G
dim(I), codim I
--This gives the associated primes of I
--If I=I(G), this gives the minimal vertex covers of G
L=ass I
--This determines whether or not I has embedded primes
--If I=I(G), this determines whether or not G is well covered
apply(L,codim)
p=(n)->gens gb ideal(flatten mingens(quotient(I,L#n)/I))
--This computes a minimal generating set for (I:p)/I
MG=(n)->flatten entries p(n)
MG(0), MG(1)
--This gives the list of all minimal generators g of
--(I:p)/I such that (I: g)=p
F=(n)->apply(MG(n),x-> if not quotient(I,x)==L#n then 0
else x)-set{0}
F(0), F(1)
--This computes the v-number of a graded ideal I
vnumber=min flatten degrees ideal(flatten apply(0..#L-1,F))
M=coker gens gb I
regularity M
--This gives the projective dimension of S/I
pdim M
\end{verbatim}
\end{procedure}
\end{appendix}
\section*{Acknowledgments}
We used \textit{Macaulay}$2$ \cite{mac2}
to implement the algorithm to compute the v-number of
graded ideals and to compute other algebraic invariants.
\bibliographystyle{plain}
|
\section{Introduction}
Explicit (or first-order) gender bias was observed in coreference resolution systems by \citet{Zhao:2018:WinoBias}, by considering contrasting cases:
\begin{enumerate}
\item The doctor hired the secretary because he was overwhelmed. [he $\rightarrow$ doctor]
\item The doctor hired the secretary because she was overwhelmed. [she $\rightarrow$ doctor]
\item The doctor hired the secretary because she was highly qualified. [she $\rightarrow$ secretary]
\item The doctor hired the secretary because he was highly qualified. [he $\rightarrow$ secretary]
\end{enumerate}
Sentences 1 and 3 are pro-stereotypical examples because gender words align with a socially-held stereotype regarding the occupations. Sentences 2 and 4 are anti-stereotypical because the correct coreference resolution contradicts a stereotype.
It was observed that systems performed better on pro cases than anti cases, and the WinoBias test set was developed to quantify this disparity.
Here we make a new observation of gender-induced (or second-order) bias in coreference resolution systems, and provide the corresponding test set SoWinoBias. Consider cases:
\begin{enumerate}
\item The doctor liked the nurse because they were beautiful. [they $\rightarrow$ nurse]
\item The nurse dazzled the doctor because they were beautiful. [they $\rightarrow$ nurse]
\item The nurse admired the doctor because they were beautiful. [they $\rightarrow$ doctor]
\end{enumerate}
The examples do not contain any explicit gender cues at all, and yet we can observe that sentences 1 and 2 align with a gender-induced social stereotype, while sentence 3 opposes the stereotype. The induction occurs because ``nurse" is a female-coded occupation \citep{Tolga:2016, Zhao2018:GNglove}, and women are also more likely to be described based on physical appearance \citep{hoyle-etal-2019-unsupervised, Williams1975}. A coreference resolution system is gender-biased if correct predictions on sentences like 1 and 2 are more likely than on sentence 3.
The difference between first-order and second-order gender bias in a downstream application is especially interesting given current trends in debiasing static word embeddings. Early methods \citep{Tolga:2016, Zhao2018:GNglove} focused on eliminating direct bias from the embedding space, quantified as associations between gender-neutral words and an explicit gender vocabulary. In response to an influential critique paper by \citet{Gonen:2019}, the current trend is to focus on eliminating indirect bias from the embedding space, quantified either by gender-induced proximity among embeddings \citep{Kumar:2020:RAN} or by residual gender cues that could be learned by a classifier \citep{Ravfogel:2020:INLP, Davis:2020:D4}.
Indirect bias in the embedding space was viewed as an undesirable property a priori, but we do not yet have a good understanding of the effect on downstream applications. Here we test debiasing methods from both camps on SoWinoBias, and make a series of observations on sufficient and necessary conditions for mitigating the latent gender-biased coreference resolution.
Additionally, we consider the case that our coreference resolution model employs both static and contextual word embeddings, but debiasing methods are applied to the static word embeddings only.
Post-processing debiasing techniques applied to static word embeddings are computationally inexpensive, easy to concatenate, and have a longer development history. However contemporary models for downstream applications are likely to use some form of contextual embeddings as well. Therefore we might wonder whether previous work in debiasing static word embeddings remains relevant in this setting.
The WinoBias test set for instance was developed and tested using the ``end-to-end" coreference resolution model \citep{lee-etal-2017-end}, a state-of-the-art model at that time using only static word embeddings. Subsequent debiasing schemes reported results on WinoBias using the same model, just plugging in different debiased embeddings, for the sake of fair comparison. However this is becoming increasingly outdated given the progress in coreference resolution systems. A contribution of this work is to report WinoBias results for previous debiasing techniques using a more updated model, one that makes use of unaltered contextual embeddings in addition to the debiased static embeddings.
The remainder of the paper is organized as follows: In section 2, we further define the type of bias being measured by the SoWinoBias test set and discuss some limitations. In section 3, we review the 4 word embedding debiasing methods that we will analyze, in the context of how each method aims to alter the word embedding space. In section 4, we provide details of the experimental setup and report results on both coreference resolution test sets, the original WinoBias and the newly constructed SoWinoBias. In section 5, we discuss the results with respect to the geometric properties of the altered embedding spaces. In particular, we review whether mitigation of intrinsic measures of bias on the embedding space, quantified as direct bias and indirect bias by various definitions, are related to mitigation of the latent bias in a downstream application.
\section{Bias Statement}
Within the scope of this paper, bias is defined and quantified as the difference in performance of a coreference resolution system on test cases aligning with a socially-held stereotype vs. test cases opposing a socially-held stereotype. We observe that gender-biased systems perform significantly better in pro-stereotypical situations.
Such difference in performance creates representational harm by implying (for example) that occupations typically associated with one gender cannot have attributes typically associated with another.
Throughout this paper, the term ``second-order" is used interchangeably with ``latent". Characterizing the observed bias as ``second-order" follows from the observation of a gender-induced bias in the absence of gender-definitional vocabulary, resting on the definition of ``they" as a gender-neutral pronoun.
Therefore, a limitation in the test set construction is the possible semantic overloading of ``they". As discussed, the intention throughout this paper is to use the singular ``they" as a pronoun that does not carry any gender information (and could refer to someone of any gender). However, different contexts may choose to treat ``they" exclusively as a non-binary gender pronoun.
The gender stereotypes used throughout this paper are sourced from peer-reviewed academic journals written in English, which draw from the US Labor Force Statistics, as well as US-based crowd workers. Therefore a limitation may be that stereotypes used here are not common to all languages or cultures.
\section{Debiasing methods}
\subsection{Neutralization of static word embeddings}
\subsubsection{Methods addressing direct bias}
The first attempts to debias word embeddings focused on the mitigation of direct bias \citep{Tolga:2016}.
The definition of direct bias assumes the presence of a ``gender direction" $\vec{g}$; a subspace that mostly encodes the difference between the binary genders.
A non-zero projection of word $\vec{w}$ onto $\vec{g}$ implies that $\vec{w}$ is more similar to one gender over another. In the case of ideally gender-neutral words, this is an undesirable property.
Direct bias quantifies the extent of this uneven similarity\footnote{The original definition included a strictness exponent $c$, here set to 1 as has commonly been done in subsequent works.}:
\begin{equation}
DB(N) = \frac{1}{|N|} \sum_{\vec{w} \in N} |\text{cos}(\vec{w}, \vec{g})|
\end{equation}
The Hard Debias method \citep{Tolga:2016} is a post-processing technique that projects all gender-neutral words into the nullspace of $\vec{g}$. Therefore, the direct bias is made to be zero by definition. We measure the performance of \textbf{Hard-GloVe}\footnote{Hard debias: https://github.com/tolga-b/debiaswe. All base (undebiased) embeddings are GloVe trained on the 2017 January Wikipedia dump (vocab contains 322,636 tokens). Available at https://github.com/uclanlp/gnglove, based on the work of \citet{pennington2014glove}.} on the coreference resolution tasks.
A related retraining method used a modified version of GloVe's original objective function with additional incentives to reduce the direct bias for gender-neutral words, resulting in the GN-GloVe embeddings \citep{Zhao2018:GNglove}. Rather than allowing for gender information to be distributed across the entire embedding space, the method explicitly sequesters the protected gender attribute to the final component. Therefore the first $d - 1$ components are taken as the gender-neutral embeddings, denoted \textbf{GN-GloVe($w_a$)}\footnote{https://github.com/uclanlp/gnglove}.
\subsubsection{Methods addressing indirect bias}
The indirect bias
is less well defined, and loosely refers to the gender-induced similarity measure between gender-neutral words. For instance, semantically unrelated words such as ``sweetheart" and ``nurse" may appear quantitatively similar
due to a shared gender association.
One definition (first given in \citep{Tolga:2016}) measures the relative change in similarity after removing direct gender associations as
\begin{equation}
\beta(\vec{w}, \vec{v}) = \frac{1}{\vec{w} \cdot \vec{v}} \left(
\vec{w} \cdot \vec{v} - \frac{\vec{w}_\perp \cdot \vec{v}_\perp}{\lVert \vec{w}_\perp \rVert \lVert \vec{v}_\perp \rVert} \right),
\end{equation}
where $\vec{w}_\perp = \vec{w} - (\vec{w} \cdot \vec{g})\vec{g}$,
however this relies on a limited definition of the original gender association.
The Repulse-Attract-Neutralize (RAN) debiasing method attempts to repel undue gender proximities among gender-neutral words, while keeping word embeddings close to their original learned representations \citep{Kumar:2020:RAN}. This method quantifies indirect bias by incorporating $\beta$ into a graph-weighted holistic view of the embedding space (more on this later).
In this paper, we will measure the performance of \textbf{RAN-GloVe}\footnote{https://github.com/TimeTraveller-San/RAN-Debias} on the coreference resolution tasks.
A related notion of indirect bias is to measure whether gender associations can be predicted from the word representation. The Iterative Nullspace Linear Projection method (INLP) achieves linear guarding of the gender attribute by iteratively learning the most informative gender subspace for a classification task, and projecting all words to the orthogonal nullspace \citep{Ravfogel:2020:INLP}. After sufficient iteration, gender information cannot be recovered by a linear classifier. We will measure the performance of \textbf{INLP-GloVe}\footnote{https://github.com/shauli-ravfogel/nullspaceprojection}.
\subsection{Data augmentation}
In addition to debiasing methods applied to word embeddings, we measure the effect of simple data augmentation applied to the training data for our coreference resolution system. The goal is to determine whether data augmentation can complement the debiased word embeddings on this particular test set.
The training data is augmented using a simple gender-swapping protocol, such that binary gender words are replaced by their equivalent form of the opposite gender (e.g. ``he" $\leftrightarrow$ ``she", etc.).
\section{Detection of gender bias in coreference resolution: Experimental setup}
All systems were built using the ``Higher-order coreference resolution with coarse-to-fine inference" model \citep{lee-etal-2018-higher}\footnote{https://github.com/kentonl/e2e-coref}.
It is important to keep in mind that this model uses both static word embeddings and contextual word embeddings (specifically ELMo embeddings \citep{peters-etal-2018-deep}). Our experimental debiasing methods were applied to static word embeddings only, and contextual embeddings are left unaltered in all cases.
All systems were trained using the OntoNotes 5.0\footnote{https://catalog.ldc.upenn.edu/LDC2013T19} train and development sets, using the default hyperparameters\footnote{``best" configuration at https://github.com/kentonl/e2e-coref/blob/master/experiments.conf}, for approximately 350,000 steps until convergence. Baseline performance was tested using the OntoNotes 5.0 test set (results shown in Table 1).
Baseline performance is largely consistent across all models, indicating that neither debiased word embeddings nor gender-swapped training data significantly degrades the performance of the system overall.
\begin{table*}
\caption{Results on coreference resolution test sets. OntoNotes (F$_1$) performance provides a baseline for ``vanilla" coreference resolution ($n = 348$). WinoBias (F$_1$) measures explicit gender bias, observable as the diff. between pro ($n = 396$) and anti ($n = 396$) test sets. SoWinoBias (\% accuracy) measures second-order gender bias, likewise observable as the diff. between pro ($n = 4096$) and anti ($n = 4096$) test sets. Note: accuracy is the relevant metric to report on the SoWinoBias test set, rather than F$_1$, due to our assertion that ``they" is not a new entity mention.}
\begin{tabularx}{\textwidth}{l | c| c| cccc| cccc}
\hline
Embedding & Data Aug. & OntoNotes & \multicolumn{4}{c|}{WinoBias} & \multicolumn{4}{c}{SoWinoBias} \\
& & & pro & anti & avg. & diff. & pro & anti & avg. & diff. \\
\hline
GloVe & & 72.3 & 77.8 & 48.8 & 63.8 & 29.0 & 64.2 & 46.8 & 55.5 & 17.4 \\
GloVe & \checkmark & 72.0 & 67.0 & 59.0 & 63.0 & 8.0 & 62.8 & 56.5 & 59.7 & 6.4 \\
\hline
Hard-GloVe & & 72.2 & 66.5 & 59.1 & 62.8 & 7.4 & 63.6 & 49.2 & 56.4 & 14.3 \\
Hard-GloVe & \checkmark & 71.8 & 64.0 & 61.9 & 63.0 & \textbf{2.1} & 77.1 & 50.1 & 63.6 & 27.0 \\
\hline
GN-GloVe$(w_a)$ & & 72.2 & 63.4 & 61.1 & 62.3 & 2.3 & 68.0 & 49.7 & 58.9 & 18.3 \\
GN-GloVe$(w_a)$ & \checkmark & 71.4 & 59.0 & 66.0 & 62.5 & 7.0 & 72.1 & 69.7 & 70.9 & \textbf{2.4} \\
\hline
RAN-GloVe & & 72.4 & 72.8 & 53.2 & 63.0 & 19.6 & 70.2 & 60.0 & 65.1 & 10.2 \\
RAN-GloVe & \checkmark & 71.1 & 60.1 & 63.8 & 62.0 & 3.7 & 69.5 & 59.4 & 64.5 & 10.0 \\
\hline
INLP-GloVe & & 71.6 & 67.5 & 57.5 & 62.5 & 10.0 & 68.4 & 46.1 & 57.3 & 22.4 \\
INLP-GloVe & \checkmark & 72.1 & 66.2 & 59.1 & 62.7 & 7.1 & 73.4 & 65.1 & 69.3 & 8.3 \\
\hline
\end{tabularx}
\label{tab:Vanilla}
\end{table*}
\subsection{WinoBias}
The WinoBias test set was created by \citet{Zhao:2018:WinoBias}, and measures the performance of coreference systems on test cases containing explicit binary gender words. In particular, pro-stereotypical sentences contain coreferents where an explicit gender word (e.g. he, she) is paired with an occupation matching a socially held gender stereotype. Anti-stereotypical sentences use the same formulation but gender swap the explicit gender words such that coreferents now oppose a socially held gender stereotype. Gender bias is measured as the difference in performance on the pro. versus anti. test sets, each containing $n = 396$ sentences.
Recall that here we are reporting WinoBias results using a system incorporating unaltered contextual embeddings, in addition to the debiased static embeddings. Previously reported results on the ``end-to-end" coreference model \citep{lee-etal-2017-end}, using only debiased static word embeddings, are compiled in the Appendix for reference.
In this setting, we observe that debiasing methods addressing direct bias are more successful than those addressing indirect bias. In particular, without the additional resource of data augmentation, RAN-GloVe struggles to reduce the difference between pro and anti test sets (in contrast to RAN-GloVe's great success in the end-to-end model setting, as reported by \citet{Kumar:2020:RAN}). Data augmentation is found to be a complementary resource, providing further gains in most cases. Overall, Hard-GloVe with simple data augmentation successfully reduces the difference in $F_1$ from 29\% to 2.1\%, while not significantly degrading the average performance on WinoBias or baseline performance on OntoNotes. This suggests that debiasing the contextual word embeddings is not needed to mitigate the explicit gender bias in coreference resolution, as measured by this particular test set.
\subsection{SoWinoBias}
The SoWinoBias test set measures second-order, or latent, gender associations in the absence of explicit gender words. At present, we measure associations between male and female stereotyped occupations with female stereotyped adjectives, although this could easily be extended in the future. Adjectives with positive and negative polarities are represented evenly in the test set. We will denote the vocabularies of interest as
\begin{align}
M_{occ} &= \{\text{doctor}, \text{boss}, \text{developer}, ...\} \\
F_{occ} &= \{\text{nurse}, \text{nanny}, \text{maid}, ...\} \nonumber \\
F_{adj}^+ &= \{\text{lovely}, \text{beautiful}, \text{virtuous}, ...\} \nonumber \\
F_{adj}^- &= \{\text{hysterical}, \text{unmarried}, \text{prudish}, ...\}, \nonumber
\end{align}
where $|M_{occ}| = |F_{occ}| = |F_{adj}^+| = |F_{adj}^-| = 16$, and the full sets can be found in the appendix.
Stereotypical occupations were sourced from the original WinoBias vocabulary (drawing from the US labor occupational statistics), as well as the SemBias \citep{Zhao2018:GNglove} and Hard Debias analogy test sets (drawing from human-annotated judgements). Stereotypical adjectives with polarity were sourced from the latent gendered-language model of \citet{hoyle-etal-2019-unsupervised}, which was found to be consistent with the human-annotated corpus of \citet{Williams1975}.
SoWinoBias test sentences are constructed as ``The [\textbf{occ1}] (dis)liked the [\textbf{occ2}] because \textbf{they} were [adj]", where ``(dis)liked" is matched appropriately to the adjective polarity, such that ``they" always refers to ``occ2". Each sentence selects one occupation from $M_{occ}$, and the other from $F_{occ}$. In pro-stereotypical sentences, occ2 $\in F_{occ}$, such that the adjective describing the (they, occ2) entity matches a social stereotype. In anti-stereotypical sentences, occ2 $\in M_{occ}$, such that the adjective describing the (they, occ2) entity contradicts a social stereotype. Example sentences in the test set include:
\begin{enumerate}
\item The doctor liked the nurse because they were beautiful. (pro)
\item The nurse liked the doctor because they were beautiful. (anti)
\item The ceo disliked the maid because they were unmarried. (pro)
\item The maid disliked the lawyer because they were unmarried. (anti)
\end{enumerate}
In total, there are $n = 4096$ sentences in each of the pro and anti test sets.
Due to the simplicity of our constructed sentences, plus our desire to measure gendered associations, we further assert that ``they" should refer to one of the two potential occupations (i.e. ``they" cannot be predicted as a new entity mention). As with WinoBias, gender bias is observed as the difference in performance between the anti and pro test sets.
Firstly, we observe that the second-order gender bias is more difficult to the correct than the explicit bias, given access to the debiased embeddings alone. Methods that made good progress in reducing the WinoBias diff. make little to no progress on the SoWinoBias diff. However, even simple data augmentation was found to be a valuable resource. When combined with GN-GloVe($w_a$), the difference is reduced to 2.4\% while increasing average performance significantly. Again, we observe that good bias reduction can be achieved, even before incorporating methods to debias the contextual word embeddings. It is interesting that debiasing methods explicitly designed to address indirect bias in the embedding space do not do better at mitigating second-order bias in a downstream task. Further discussion in relation to the embedding space properties is provided in the following section.
\section{Relationship to embedding space properties}
\subsection{Single-attribute WEAT}
\begin{table*}
\caption{Single-Attribute WEAT association strength between gender and female-stereotyped adjectives with significance values. Lower association strength ($S$) values are better. Smaller significance values indicate that the observed association strength is meaningful with respect to gender. }
\begin{tabularx}{\textwidth}{l | ll|ll}
\hline
Embedding & $S(F_{adj}, F_{occ}, M_{occ})$ & Significance & $S(F_{adj}, F_{def}, M_{def})$ & Significance \\
\hline
GloVe & 0.0636 & 0.0001** & 0.0694 & 0.001** \\
Hard-GloVe & 0.0465 & 0.0001** & \textbf{-8.6889e-10} & 0.512 \\
GN-GloVe$(w_a)$ & 0.0664 & 0.0003** & \textbf{-0.0015} & 0.436 \\
RAN-GloVe & 0.0402 & 0.0003** & \textbf{0.0153} & 0.177 \\
INLP-GloVe & \textbf{0.0171} & 0.0251* & \textbf{0.0054} & 0.382 \\
\hline
\end{tabularx}
\label{tab:Vanilla}
\end{table*}
The Word Embedding Association Test (WEAT) measures the association strength between two concepts of interest (e.g. arts vs. science) relative to two defined attribute groups (e.g. female vs. male) \citep{Caliskan:2017:weat}. It was popularized as a means for detecting gender bias in word embeddings by showing that (arts, science), (arts, math), and (family, careers) produced significantly different association strengths relative to gender.
Here we adapt the original WEAT to measure relative association across genders given a single concept of interest. This provides a means to measure whether the set of female-stereotyped adjectives $F_{adj}$ are quantitatively gender-marked in the embedding space.
The relative association of a single word $t$ across attribute sets $A_1$, $A_2$ is given by
\begin{align}
s(t, A_1, A2) = & \frac{1}{|A_1|} \sum_{a_1 \in A_1}\text{cos}(t, a1)
\\& - \frac{1}{|A_2|} \sum_{a_2 \in A_2}\text{cos}(t, a2) \nonumber
\end{align}
where $s(t, A_1, A2) > 0$ indicates that $t$ is more closely related to attribute $A_1$ than $A_2$. The average relative association of concept $T$ is then
\begin{equation}
S(T, A_1, A_2) = \frac{1}{|T|} \sum_{t \in T} s(t, A_1, A2).
\end{equation}
The significance of a non-zero association strength can be assessed by a partition test. We randomly sample alternate attribute sets of equal size $A_1^*$ and $A_2^*$ from the union of the original attribute sets. The significance $p$ is defined as the proportion of samples to produce $S(T, A_1^*, A_2^*) > S(T, A_1, A_2)$. Small $p$ values indicate that the defined grouping of the attributes sets (here defined by gender) are meaningful compared to random groupings.
Table 2 shows the results of the single-attribute WEAT. We measure association strength of the female adjectives relative to gender in two ways: i) gender is defined using a ``definitional" vocabulary ($A_1 = F_{def} = \{ she, her, woman, ...\}$, $A_2 = M_{def} = \{he, him, man, ... \}$), and ii) gender is defined using a latent vocabulary $-$ the stereotypical occupations ($A_1 = F_{occs}$, $A_2 = M_{occs}$).
As shown, the $F_{adj}$ embeddings are strongly associated with the explicit gender vocabulary in the original GloVe space. However each of the four debiasing methods are successful in removing the explicit gender association, as expected. The Hard Debias method in particular asserts $S(F_{adj}, F_{def}, M_{def}) = 0$ by definition.
In contrast, the $F_{adj}$ embeddings are just as strongly associated with the latent gender vocabulary in the original GloVe space, but this is not undone by any of the debiasing methods.
This is somewhat of an unexpected result in the case of the RAN and INLP debiasing methods, as they promised to go beyond direct bias mitigation.
The INLP method makes the most progress in reducing the implicit association strength, however a significant non-zero association remains. Combined with the SoWinoBias test results, we can observe that the WEAT reduction achieved by INLP is not a sufficient condition for mitigating latent gender-biased coreference resolution. Inversely, we observe that reduction of the WEAT measure is not a necessary condition for mitigation when debiased embeddings are combined with data augmentation (demonstrated by GN-GloVe$(w_a)$).
\subsection{Clustering and Recoverability}
\begin{table*}
\caption{Clustering: (reported as accuracy and $v$-measure \citep{rosenberg:vmeas}) is performed by taking the $n = 1500$ most biased words in the original embedding space (excluding definitional gender words), and performing k-means clustering ($k = 2$) on the same words in the debiased space.
Recoverability: (reported as accuracy) is performed by taking the $n = 5000$ most biased words in the original embedding space, and training a classifier (linear SVM or rbf kernel SVM) on the same words in the debiased space. Smaller values are better (indicating less residual cues that can be used classify gender-neutral words).
GIPE: Smaller values are better (indicating less undue proximity bias in the embedding space).}
\begin{tabular}{l | cc|cc | ccc}
\hline
Embedding & Acc. & $v$-measure & linSVM & rbfSVM & GIPE($V_d$) & GIPE($V_{So}$) & Avg. $\eta(w_{So})$ \\
\hline
GloVe & 99.8 & 98.4 & 100 & 100 & 0.1153 & 0.1844 & 0.1373 \\
Hard-GloVe & 79.0 & 30.2 & 92.5 & 94.6 & 0.0701 & 0.1020 & 0.0894 \\
GN-GloVe$(w_a)$ & 85.3 & 49.7 & 99.1 & 99.4 & 0.1173 & 0.1650 & 0.1167 \\
RAN-GloVe & 80.4 & 41.9 & 95.3 & 96.0 & \textbf{0.0399} & \textbf{0.0827} & \textbf{0.0617} \\
INLP-GloVe & \textbf{57.1} & \textbf{1.52} & \textbf{52.9} & \textbf{74.8} & 0.0798 & 0.1265 & 0.0967 \\
\hline
\end{tabular}
\label{tab:Vanilla}
\end{table*}
Clustering and recoverability (C\&R) \citep{Gonen:2019} refer to a specific observation on the embedding space post debiasing; namely, that gender labels of words (assigned according to direct bias in the original embedding space) can be classified with a high degree of accuracy given only the debiased representations.
Here we follow the same experimental setup, and report results on an expanded set of embeddings (see Table 3).
In agreement with \citet{Gonen:2019}, we find that the Hard-GloVe and GN-GloVe embeddings retain nearly perfect recoverability of the original gender labels, indicating high levels of residual bias by this definition.
The INLP method was designed to guard against linear recoverability, and indeed we find that both C\&R by a linear SVM are reduced to near-random performance. Recoverability by an SVM with a non-linear kernel (rbf) achieves 75\% accuracy; much reduced compared to other debiasing methods, but still above the baseline of 50\%. This result is consistent with \citet{Ravfogel:2020:INLP}.
Of interest are the results obtained for the RAN-GloVe embeddings, which have not previously been reported. RAN was designed to mitigate undue proximity bias, conceptually similar to clustering. Despite this, C\&R are still possible with high accuracy given RAN-debiased embeddings. Given RAN's success on various gender bias assessment tasks (SemBias, and WinoBias using the end-to-end coreference model), this suggests that complete suppression of C\&R is unnecessary for many practical applications. Conversely, it may indicate that we have not yet developed any assessment tasks that probe the effect of indirect bias.
In reference to the SoWinoBias results, we can observe that linear attribute guarding (achieved by INLP) is not a sufficient condition for mitigating latent gender-biased coreference resolution.
However, even linear guarding is not a necessary condition for mitigating SoWinoBias when retraining with data augmentation is available.
\subsection{Gender-based Illicit Proximity Bias}
The gender-based illicit proximity bias (GIPE) was proposed by \citet{Kumar:2020:RAN} as a means to capture indirect bias on the embedding space as a well-defined metric, as opposed to the loosely defined idea of clustering and recoverabilty. Firstly, the gender-based proximity bias of a single word $w$, denoted $\eta(w)$, is defined as the proportion of $N$-nearest neighbours $\{n_i\}$ with indirect bias $\beta(n_i, w)$ above some threshold $\theta$. Intuitively, this is the proportion of words that are close by solely due to a shared gender association. The GIPE extends this word-level measure to a vocabulary-level measure using a weighted average over $\eta(w)$.
Table 3 shows the GIPE measure on the entire gender-neutral vocabulary $V_d$, the gender-neutral vocabulary used to construct SoWinoBias $V_{So} = F_{occ} \cup M_{occ} \cup {F_{adj}}$, and the simple (unweighted) average $\eta(w_{So})$ on the SoWinoBias vocabulary.
The RAN method mitigates indirect bias as measured by GIPE by design, and therefore achieves the lowest GIPE values as expected (followed by Hard-GloVe, somewhat unexpectedly). However, non-zero proximity bias persists, more so on the stereotyped sub-vocabulary than the total vocabulary.
Without extra help from data augmentation, RAN-GloVe achieves the best performance on the SoWinoBias (followed by Hard-GloVe). Therefore further reduction of GIPE may enable further mitigation of the latent gender-biased coreference resolution (cannot be ruled out as a sufficient condition at this time). However, RAN-GloVe does not benefit from the addition of data augmentation, unlike the majority of debiasing methods. Further investigation is needed to determine what conditions of the embedding properties allow for complementary data augmentation.
\section{Conclusion}
In this paper, we demonstrate the existence of observable latent gender bias in a downstream application, coreference resolution. We provide the first gender bias assessment test set not containing any explicit gender-definitional vocabulary. Although the present study is limited to binary gender, this construction should allow us to assess gender bias (or other demographic biases) in cases where explicit defining vocabulary is limited or unavailable. However, the construction does depend on knowledge of expected relationships or stereotypes (here occupations and adjectives). Therefore interdisciplinary work drawing from social sciences is encouraged as a future direction.
Our observations indicate that mitigation of indirect bias in the embedding space, according to our current understanding of such a notion, does not reduce the latent associations in the embedding space (as measured by WEAT), nor does it mitigate the downstream latent bias (as measured by SoWinoBias). Future work could seek bias assessment tasks in downstream applications that do depend on the reduction of gender-based proximity bias or non-linear recoverability. Currently the motivation for such reduction is unknown, despite being an active direction of debiasing research.
Finally, we do observe that an early debiasing method, GN-GloVe, combined with simple data augmentation, can mitigate the latent gender biased coreference resolution, even when contextual embeddings in the system remain unaltered. Future work could extend the idea of the SoWinoBias test set to more complicated sentences representative of real ``in the wild" cases, in order to determine if this result holds.
The SoWinoBias test set, all trained models presented in this paper, and code for reproducing the results are available at https://github.com/hillary-dawkins/SoWinoBias.
\section*{Acknowledgements}
We thank Daniel Gillis, Judi McCuaig, Stefan Kremer, Graham Taylor, and anonymous reviewers for their time and thoughtful discussion. This work is financially supported by the Government of Canada through the NSERC CGS-D program (CGSD3-518897-2018).
\bibliographystyle{acl_natbib}
|
\section{PRELIMINARIES} \label{sec:prelim}
We consider the problem of learning periodic manipulation skills from a single human demonstration. We assume that the human demonstrates a periodic task that is performed for at least 2 periods. Given a human demonstration $V_H = (I_H^{1}, \ldots, I_H^{T_H})$ as a sequence $V_H$ of $T_H$ RGB image frames $I_H^{t}$, the robot is allowed to execute a total of 50 trials in the environment to learn the demonstrated skill\edit{, after which the best execution is selected for evaluation}. In each trial, the robotic agent executes a trajectory \edit{$\tau_R = (\vc{x}_R^{1}, \ldots, \vc{x}_R^{T})$ where $\vc{x}^t$} denotes the robot end-effector position at timestep $t$. We denote the corresponding execution video of this executed trajectory as $V_R = (I_R^{1}, \ldots, I_R^{T})$, where $I_R^t$ denotes the camera image at timestep $t$.
We measure the similarity between the object motion in the human demonstration and in the video of the robot execution using consistent keypoints across the two videos (see Section~\ref{method:keypoint}).
In this work, we combine BO and DMPs to maximize this similarity score, and present our method in Section~\ref{sec:method}. Below we provide the technical background for BO and DMPs.
\subsection{Bayesian Optimization}
In Bayesian optimization (BO), an optimization problem is viewed as finding parameters $\mathbf{w}$ that optimize some objective function $f_{\text{BO}}(\mathbf{w}): f_{\text{BO}}(\mathbf{w}^*) = \max_\mathbf{w} f_{\text{BO}}(\mathbf{w})$. $f_{\text{BO}}$ is commonly modeled with a Gaussian process (GP): $f_{\text{BO}}(\mathbf{w}) \sim \mathcal{GP}(m(\mathbf{w}), k(\mathbf{w}_i, \mathbf{w}_j))$. At each trial, to select the next promising candidate $\mathbf{w}$, BO optimizes an acquisition function, e.g. the Upper Confidence Bound (UCB)~\cite{srinivas2009gaussian}, which explicitly balances exploration (high posterior uncertainty) \edit{vs.} exploitation (high posterior mean estimate): $UCB(\mathbf{w}) = m(\mathbf{w}) + \edit{\beta\cdot Std(\mathbf{w})}$.
The kernel defines a similarity function on the search space. The RBF kernel is a common choice: $k(\mathbf{w}_i, \mathbf{w}_j) = \sigma^2_k \exp(-\frac{1}{2}\lVert \mathbf{w}_i \text{-} \mathbf{w}_j \rVert_2^T \text{diag}(\mathbf{l})^{-2} \lVert \mathbf{w}_i \text{-} \mathbf{w}_j \rVert_2)$, where $\sigma^2_k$ and $\mathbf{l}$ are signal variance and a vector of length scales, respectively. In practice, $\sigma^2_k$ is a hyperparameter optimized automatically by maximizing marginal likelihood.
\begin{figure}
\vspace{6pt}
\centering
\includegraphics[width=0.44\textwidth]{figs/keypoint_learning.pdf}
\caption{Unsupervised keypoint learning from play data. The play data consists of unpaired, unlabeled and task-agnostic human and robot motion recorded from the same viewpoint. It is used to train a keypoint model similar to \cite{kulkarni2019unsupervised} that finds consistent keypoints across human and robot demonstrations. This allows to compute a keypoint-based distance between human and robot videos suitable for guiding the search for the best matching robot motion. Robot trajectories are included in the robot play data, but they are not used for keypoint learning purposes.}
\label{fig:keypoint}
\end{figure}
\subsection{Dynamic Movement Primitives (DMPs)}
DMPs are trajectory generators whose parameters can be learned from demonstrations of desired robot end-effector trajectories. They combine linear fixed-point attractors with non-linear function approximators to encode complex trajectories, while maintaining convergence guarantees. We refer readers to \cite{ijspeert2013dynamical} for a detailed overview.
The \textit{transformation system} of a DMP consists of a damped linear feedback term, and a forcing function $f$:
\begin{equation}\label{eq:base_dynamical_system}
\begin{aligned}
\tau \dot{z} &= \alpha_{z}\left(\beta_{z}(g-x)-z\right) + f \ ; \quad \tau \dot{x} = z \ ,
\end{aligned}
\end{equation}
where $x$ is position, $g$ is the goal, $\alpha_z$ and $\beta_z$ are constants, $\tau$ is a temporal scaling factor, $z$ is the scaled velocity, and the output of the transformation system is the scaled acceleration $\dot{z}$.
The second component of DMPs is a \textit{canonical system} which replaces time, and enables scaling the trajectory to different time lengths. The canonical system is different between rhythmic and discrete DMPs; in the discrete case it represents `time left' and goes to 0 at the end of the motion, while in the rhythmic case it represents the time from the start, and goes up linearly. Specifically, in rhythmic DMPs, the first-order canonical system $\tau \dot{\phi}=1$ encodes the phase $\phi$, increasing linearly as motion progresses.
In rhythmic DMPs, the forcing function $f$ is parameterized by $\phi$ and consists of cyclic basis functions:
\begin{equation}\label{eq:rdmp_def}
\begin{aligned}
f &=\frac{\sum_i \Psi_{i} w_{i}}{\sum_i \Psi_{i}} r\ ;
\
\Psi_{i} \!=\! \exp \left(h_{i}\left(\cos \left(\phi-c_{i}\right)-1\right)\right),
\end{aligned}
\end{equation}
where $\Psi$ is a function of the canonical system, and the weights $w_i$ are commonly learned using locally weighted regression~\cite{schaal1998constructive}. The cyclic nature of basis functions ensures that the transformation system yields cyclic motion, as the canonical system unrolls.
Typically, the goal $g$ of a rhythmic DMP is set to the mean of the demonstration trajectory and kept fixed. Discrete DMPs are shown to generalize well to changing goals. \cite{ijspeert2013dynamical} present ways to continuously change goals to new locations without causing a discontinuity in the acceleration $\dot{z}$. We adapt \cite{ijspeert2013dynamical} to smoothly move the goals of rhythmic DMPs between executions. This continuously modulates the mean point of the limit cycle of the DMP, allowing us to model motions that are mostly cyclic, but slightly shifting over time, e.g. as in wiping a surface.
\section{CONCLUSION}
We introduced Visual Periodic Task Learner (ViPTL), a method for representing periodic manipulation policies and efficiently learning them from a single human demonstration. We show that ViPTL~succeeds on three robot manipulation tasks that involve deformable and granular objects. This work demonstrates the benefit of leveraging the periodic structure of many tasks commonly seen in everyday life. In future work, we intend to extend our method to handle initial stages of tasks, such as grasping and other transient motions.
\edit{We also plan to extend our perception module so that it generalizes to unseen camera views and contexts.}
\section{EXPERIMENTS}
In our experiments, we aim to answer the following questions: (1) does our framework successfully learn to perform periodic tasks from a single human demonstration; (2) does our proposed framework perform better than methods that do not exploit periodicity in the target task; (3) is our proposed keypoint-based representation more suitable for learning from human demonstrations than other representations (e.g. latent vectors generated by 20BN classifiers)?
\begin{figure*}[h]
\vspace{6pt}
\centering
\begin{subfigure}{.21\textwidth}
\centering
\includegraphics[width=\linewidth]{figs/real_wiping.pdf}
\caption{\texttt{Table Wiping}}
\end{subfigure}
\begin{subfigure}{.21\textwidth}
\centering
\includegraphics[width=\linewidth]{figs/real_winding.pdf}
\caption{\texttt{Rope Winding}}
\end{subfigure}
\begin{subfigure}{.21\textwidth}
\centering
\includegraphics[width=\linewidth]{figs/real_stirring.pdf}
\caption{\texttt{Food Stirring}}
\end{subfigure}
\hspace{2pt}
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[width=\linewidth]{figs/variations.pdf}
\end{subfigure}
\caption{Results for the real robot experiments. Plots (a)-(c) show performance comparisons (for details of the methods see Section \ref{sec:baselines}, for evaluation metric details see Figure \ref{fig:sim_plots}). The right part shows skills learned with our method executing at different scales and numbers of repetitions without retraining.}
\label{fig:real_plots}
\end{figure*}
We consider 3 challenging manipulation tasks: (1) \texttt{Table Wiping}, where the objective is to wipe a rectangular table surface with a cloth using back-and-forth motions, shifting to cover all the visible the area of the table; (2) \texttt{Rope Winding}, where the objective is to wrap a rope around a fixed spool by repeating circular winding motion several times; (3) \texttt{Food Stirring}, where the objective is to stir granular objects in a tray with a spatula/spoon.
\paragraph*{Metric} The baselines and ablations are each optimizing a different objective function. Therefore, we define a performance metric that is comparable across the different approaches.
We prepare an exemplary robot trajectory $\tau_E = ( \edit{\vc{x}}_E^1, \ldots, \edit{\vc{x}}_E^{T_E})$ that, when executed, produces the same effect on the objects being manipulated as the human demo. During execution, the robot will produce a trajectory $\tau_R = (\edit{\vc{x}}_R^{1}, \ldots, \edit{\vc{x}}_R^{T})$, where $\edit{\vc{x}}^t$ denotes the robot end-effector position at timestep $t$. \edit{To evaluate the performance of the robot trajectory $\tau_R$, we compute the similarity of $\tau_R$ to $\tau_E$ through $\kappa(-\lVert \tau_E - \tau_{R'} \rVert_1)$,} where $\tau_{R'}$ is a sub-sampled trajectory of $\tau_R$ with length $T_E$ and $\kappa$ is a linear transformation that ensures a score between 0 and 1.
Note that $\tau_E$ is only used for evaluation purposes and not visible to any of the methods. Figures~\ref{fig:sim_plots} and~\ref{fig:real_plots} plot this metric for our simulation and hardware experiments.
The methods that use BO optimize different objective functions e.g. the keypoint-based objective for \textit{ViPTL~(our method)} or the cosine distance between latent features for the \textit{Twentybn Classifier}.
\subsection{Simulation Experiments}
\subsubsection{Baselines and ablations} \label{sec:baselines}
As appropriate baselines and ablations, we need methods that can imitate a single visual human demonstration on the robot. Standard image-based model-free and model-based RL methods \cite{hafner2020mastering, kalashnikov2018qt, ebert2018visual, chebotar2021actionable} cannot operate in this setup because images from human demos and robot execution are visually different. Thus, we use two baselines that use our keypoint-based visual representation as state and an ablation that does not use this visual representation to test whether our visual representation contributes to the final performance of our method.
First we have \textit{Direct Imitation}, which learns a function that maps keypoints at the current timestep to desired robot end-effector positions. This function is trained on robot play data, which contains both keypoints and robot trajectories. To imitate a human demonstration, we use keypoints from the demo video frames as input and output desired robot end-effector positions. The resulting robot trajectory is then executed by fitting a DMP to the predicted robot positions. This baseline studies the use of BO versus a trained neural network for optimizing DMP controllers.
Second we have the \textit{Twentybn Classifier}, an ablation in which the BO objective is based on the video activity classifier~\cite{wang2021tdn} trained on the 20bn dataset~\cite{goyal2017something}. Both the human demonstration and robot execution video are input to the classifier to obtain two feature vectors from the last hidden layer. The objective function is the cosine distance between these two features. We use this baseline to test if the keypoint-based visual representation is a suitable visual representation for the learning from human demo setup compared to alternatives.
Third we have \textit{MBIL}, a model-based imitation learning baseline that relies on a learned dynamics model of the keypoints. The model is trained on robot play data and updated every episode as new interaction data is collected. This model is then used to plan actions to imitate the human demo. At every timestep, we run single-step model-predictive control (MPC) by sampling 5,000 random actions and executing the action that leads to the smallest keypoint distance to the corresponding human demo frame. This baseline tests if our method outperforms model-based RL methods like \cite{manuelli2020keypoints} that do not model periodicity of the task.
\subsubsection{Experimental Setup}
For all our experiments, we use an image size of $512 \times 512$. In the distance function, the number of sub-sampled frames $N_s$ is set to $10\cdot n_H^{rep}$.
To create masks for the human hand, we use the MediaPipe library \cite{lugaresi2019mediapipe} to detect the hand skeleton from an image frame and use the color at the joints of the skeleton to construct a color range mask for the hand. We mask out the robot based on the depth readings (since the robot pose is known).
In BO, we optimize L=7 trajectory waypoints for wiping and winding, L=5 for stirring.
We use UCB~\cite{srinivas2009gaussian} with $\beta = 0.1$ in the acquisition function of BO. We use automatic hyperparameter optimization to find the appropriate length scales of the RBF kernel.
When constructing imagined trajectories, we use $10$ play data segments of length $T_s = 10$ each and use a distance threshold of $d_\text{seg} = \frac{1}{6}\lambda_\text{disp}$. We generate 5,000 imagined trajectories and select the top $N_{\text{imagined}} = 100$ trajectories as initial candidates for BO, and use 10 of these in the first 10 BO trials.
\subsubsection{Quantitative Results}
To evaluate the performance of our framework in comparison to competing methods, we run all methods and baselines for 50 trials in all tasks using 3 different random seeds. The performance of all the methods during BO trials is shown in Figure \ref{fig:sim_plots}. The \textit{Direct Imitation} and \textit{MBIL} baselines cannot imitate the human demo well in \texttt{Table Wiping} and \texttt{Rope Winding}, since modeling or capturing dynamics of the deformable objects in these tasks is difficult.
The suboptimal performance of these two baselines shows that our method outperforms methods that rely on single-step predictions or learning accurate dynamics models and do not exploit periodicity in the target task.
The \textit{Twentybn Classifier} baseline achieves limited performance in all three tasks as it lacks precision in the classifier-based distance metric used to compare the given human demo with robot executions. This shows that the keypoint-based cost function is a crucial component of our method that is more suitable for the learning from human demos setup. In contrast, our method (\textit{ViPTL}) \edit{achieves good performance in all three tasks, outperforming various baselines and ablations.}
We include qualitative results in the supplementary video.
\subsection{Real Robot Experiments}
\subsubsection{Hardware Setup}
Our hardware setup (Figure \ref{fig:real_plots}) includes a Kinova Gen3 robot arm with a Robotiq 2F-85 gripper and an Intel RealSense D435 camera.
The table workspace measures $50 \times 43~\text{cm}$, and the camera is mounted at the height of 68 cm at one side of the workspace.
The camera provides the RGB image data during experiments and is positioned to view the table surface. We use velocity control in the Cartesian (end-effector) space to execute the desired trajectories on the robot.
\subsubsection{Quantitative Results}
We select the \textit{Twentybn Classifier} ablation and the baseline with the most consistent performance across tasks in simulation experiments (\textit{Direct Imitation}) to compare with our method (see Figure~\ref{fig:real_plots}). Our method is able to quickly find high-scoring points due to an effective warm-start and fine-tune the generated motion, leading to consistently improving performance throughout the 50 trials. The baselines are unable to catch up with the performance of our method for similar reasons as mentioned in simulated experiments. \edit{In particular, the \textit{Direct Imitation} baseline performs worse here than in simulation since dynamics in the real world is more stochastic.}
\section{INTRODUCTION}
Periodic tasks such as wiping a table with a cloth, stirring food, winding cables, or tying ropes are ubiquitous in our daily lives (see Figure~\ref{fig:teaser}).
In this work, we address how robots can appropriately represent and learn periodic policies by watching humans. While prior works considered learning manipulation skills from human demonstrations~\cite{liu2018imitation, smith2019avid, sharma2019third, xiong2021learning, shao2020concept, chen2021learning}, less attention has been given to periodic tasks. These tasks repeat similar motion with only small differences between repetitions. If a robot was able to decompose the demonstration of a periodic task into periods, it could efficiently learn the underlying motion and repeat it as many times as necessary. Prior hierarchical learning approaches investigated learning compositional tasks either from demonstrations \cite{xu2018neural, Silver2019FewShotBI} or through reinforcement learning \cite{Nair2020HierarchicalFS, hundt2020good}. However, these approaches do not leverage the strong relationship \edit{across} repetitions and can be inefficient for learning periodic tasks.
In this work, we propose {\em Visual Periodic Task Learner\/} (ViPTL): a method with an explicit periodic policy representation that enables efficient learning of periodic robot manipulation skills from a single, visual human demonstration. \edit{For} policy representation\edit{,} we \edit{adopt} rhythmic {\em Dynamic Movement Primitives\/} (rDMPs) \cite{Ijspeert2002LearningRM} which model cyclic motion, with shift parameters that account for motion that translates over multiple periods. The benefits of using rhythmic DMPs are two-fold: (1) they succinctly represent periodic manipulation policies, enabling efficient learning -- the problem of learning a long-horizon periodic task is reduced to learning a single-period motion of the task; (2) the learned rhythmic DMP can be repeated any number of times and the speed and amplitude of the motion can be adjusted at run time by a simple change of a parameter.
Typically, DMPs are trained on robot trajectories demonstrated through kinesthetic teaching or teleoperation. However, this can be non-intuitive to non-expert users. Instead, we use human video demonstrations that do not contain trajectories but are easier to record. To train rDMPs from such videos, we learn a keypoint detector that identifies consistent keypoints across human demonstrations and robot executions. Based on this, we can now evaluate the similarity between a robot execution and the human demonstration -- a quantity we aim to maximize. Keypoint models do not assume rigid objects and are therefore well suited for challenging manipulation tasks considered in this paper that involve deformable objects and granular material. We learn the keypoint model from task-agnostic play data and leverage periodicity estimation from~\cite{dwibedi2020counting} to break down the demonstrated task into single-period components. We use Bayesian optimization (BO) to optimize similarity of the robot motion to the human demonstration.
To focus BO on promising regions, we create `imagined' trajectories from segments of robot play data that serve as initial candidates for BO.
\begin{figure}
\centering
\includegraphics[width=0.475\textwidth]{figs/teaser.pdf}
\caption{Examples of everyday periodic tasks. Our robot learns to imitate wiping, stirring and winding from human video demonstrations (top row).}
\label{fig:teaser}
\end{figure}
\begin{figure*}[ht]
\vspace{6pt}
\centering
\includegraphics[width=0.98\textwidth]{figs/method.pdf}
\caption{Overview of the proposed approach. Our method is composed of two modules: the unsupervised visual representation module and the parameter optimization module. The unsupervised module learns a model for keypoint correspondences between the motions of the objects in the human demonstration and the robot trials. The parameter optimization module uses active learning to adjust the parameters of the rhythmic DMP controllers.}
\label{fig:method_overview}
\end{figure*}
We quantitatively evaluate the proposed method in both simulation and on a real robot with three manipulation tasks: table wiping, rope winding, and food stirring (see Figure~\ref{fig:teaser}). We show that our approach can successfully learn challenging periodic manipulation tasks that involve deformable and granular objects from a single human demonstration within 50 robot trials. Our comparisons to existing approaches and ablations show how our perception and optimization modules contribute to the overall success of the method.
\edit{In summary, our contributions are as follows: (1) a framework that represents and learns periodic robot manipulation policies from human demonstrations; (2) a motion optimization algorithm that leverages BO for sample-efficient learning and an unsupervised keypoint model for score computation in BO; (3) empirical results on a range of periodic manipulation tasks that involve deformable and granular objects.}
\section{RELATED WORK}
\subsection{Modeling Periodic Motion}
Various methods have been proposed to model periodic motion. Early literature in robotics and neuroscience used limit cycles and central pattern generators to model periodic motions for locomotion~\cite{jalics1997pattern, kajita2002realtime, yang2004infant,2008ICRALudo}. Recently, pattern generators have also been used with robotic manipulators \cite{thor2019fast, oikonomou2020periodic}.
However, limit cycle formulations are not easily amenable to learning arbitrary periodic trajectories.
In such cases, dynamic movement primitives (DMPs)~\cite{Ijspeert2002LearningRM} can provide the needed flexibility, and ease of use with learning-based approaches. DMPs have been used for periodic manipulation tasks, writing and wiping being the most common~\cite{6651499, 10.1162/NECO_a_00393}. In comparison, our proposed work also learns winding and stirring tasks, but from visual demonstration that are not annotated with human hand poses. {\em Fourier movement primitives\/} (FMPs)~\cite{kulak2020fourier} are an extensions of DMPs using Fourier series as basis functions.
While our system is agnostic to the specific choice of periodic parameterization of the control policies, here we use rhythmic DMPs.
\subsection{Periodicity Estimation}
There has been a significant interest in estimating periodicity in the computer vision community. Prior works use Fourier analysis~\cite{azy2008segmentation, cutler2000robust, pogalin2008visual}, singular value decomposition~\cite{chetverikov2006motion}, or peak detection~\cite{thangali2005periodic} to detect repetition by converting the motion in videos to one-dimensional signals.
Recent works propose detecting non-stationary repetitive motion using wavelet transforms~\cite{runia2018real}, 3D convolution networks~\cite{zhang2020context}, and self-similarity between video frames~\cite{dwibedi2020counting}.
In this work, we use RepNet~\cite{dwibedi2020counting} for periodicity estimation. We find that once trained on the Countix dataset in \cite{dwibedi2020counting}, RepNet can successfully decompose human demonstrations of various manipulation tasks into single-period segments without further finetuning.
\subsection{Learning from Human Demonstrations}
Several works in learning from human video demonstrations propose using image-to-image translation to transform human demos to robot executions~\cite{liu2018imitation, smith2019avid, sharma2019third, xiong2021learning}. However, these require a large amount of training data. Recent works~\cite{shao2020concept, chen2021learning} leverage action recognition models, such as the action classifiers trained on the 20BN Something-Something dataset~\cite{goyal2017something}, to identify whether the robot is performing the desired task. However, while these classifiers are useful for identifying the class of motions for short interactions, we show that they do not retain enough information to analyze tasks with longer duration and multiple repetitions.
Our approach uses a small amount of task-agnostic, unpaired and unlabeled `play' data~\cite{lynch2020learning} to train a keypoint model that makes it possible to quantitatively compare human demos and robot executions. `Play' data is useful, because it can be collected without supervision and using a task-agnostic, randomized policy. However, unlike~\cite{lynch2020learning} that explores using hours of such data, we focus on a much more data-efficient alternative. We collect 10 minutes of human `play' data and 10 minutes of robot `play' data (unpaired), and then use a single human demonstration video to infer the appropriate parameters for the robot control policy.
\subsection{Sample-efficient Robot Learning}
We aim to imitate a visual human demo on a robot with high sample-efficiency. Prior works have explored model-based methods~\cite{hafner2020mastering, ebert2018visual, sekar2020planning} and self-supervised exploration algorithms~\cite{pathak2017curiosity, schmidhuber1991possibility, klyubin2005empowerment, eysenbach2018diversity, bellemare2016unifying, yarats2021reinforcement} to improve sample efficiency, but these methods often require much more data on the robot and do not directly generalize to visual imitation learning. Some works utilize large-scale training in simulation and transfer the learned policies to the real robot~\cite{tobin2017domain, james2019sim}. Since our tasks include hard-to-simulate deformable and granular objects, sim2real is not applicable in our setup. In contrast, we use Bayesian Optimization (BO) to optimize parameters of a rhythmic DMP. BO is capable of learning the demonstrated manipulation skills within 50 trials on the real robot.
\section{METHOD} \label{sec:method}
Our framework is composed of two parts: (1) a representation learning module, where a keypoint detection model is trained to extract consistent keypoints from independently collected and non-task-specific human and robot play data; (2) a parameter optimization module, where BO searches for a rhythmic DMP that \edit{would} produce a robot video that matches the human demo in terms of the detected keypoints. These modules are detailed in Sections \ref{method:keypoint} and \ref{method:bo}, respectively. Figure \ref{fig:method_overview} shows an overview.
\subsection{Unsupervised Keypoint Learning from Play Data}
\label{method:keypoint}
\begin{figure}[t]
\vspace{6pt}
\centering
\includegraphics[width=0.46\textwidth]{figs/bo.pdf}
\caption{Our Bayesian optimization pipeline.}
\label{fig:bo}
\end{figure}
To learn a manipulation skill from a human demo, we need a way to evaluate the similarity between the demo and robot execution. To learn such a similarity score, we assume that the agent has access to a small amount of human and robot play data. Play data is a dataset of self-guided, task-agnostic, and diverse interactions.
The human play data $\mathcal{D}_H = (I_{HP}^1, \ldots, I_{HP}^{T_{HP}})$ is a sequence of $T_{HP}$ unlabeled RGB image frames, while the robot play data is a sequence of $T_{RP}$ RGB image frames $\mathcal{D}_R = (I_{RP}^1, \ldots, I_{RP}^{T_{\edit{RP}}})$ accompanied by robot end-effector positions \edit{$\vc{x}_{RP}^{1}, \ldots, \vc{x}_{RP}^{T_{RP}} \in \mathbb{R}^3$}. Note that $\mathcal{D}_H$ and $\mathcal{D}_R$ are unpaired and independent.
To acquire a visual representation for the manipulated object that is invariant to change of agent between human and robot, we adopt a variation of the Transporter architecture \cite{kulkarni2019unsupervised} -- an unsupervised keypoint detection model that learns to generate temporally consistent keypoints $\Psi^*(I)$ on image input $I$. The learning process is illustrated in Figure \ref{fig:keypoint}. Humans and robots may move their hands very differently to generate the same object movement. To make sure that the keypoint model $\Psi^*$ allocates keypoints to the manipulated objects and not to the human and robot hand, we mask these areas in the reconstruction loss~\cite{kulkarni2019unsupervised} when training the keypoint model. This is achieved by using commonly available hand detectors~\cite{lugaresi2019mediapipe} and depth filtering, respectively, when computing the reconstruction loss that the Transporter is trained on. With this, the keypoint model is more likely to place keypoints on the objects, and be robust to different visual appearance of human and robot hands.
After the keypoint model is trained, we process the human demo $V_H$ and robot execution $V_R$ to produce sub-sampled videos $V_{H^{'}} = \{I_{H'}^i\}_{i=1}^{N_s}$ and $V_{R^{'}} = \{I_{R'}^i\}_{i=1}^{N_s}$ that both have length $N_s$. We then define the distance between the human demo and the robot execution as:
\vspace{-0.1cm}
\begin{align}
D(V_H, V_R) \!=\! \frac{1}{N_s N_k} \sum_{i=1}^{N_s} \lVert \Psi^*(I_{H^{'}}^i) - \Psi^*(I_{R^{'}}^i) \rVert_1,
\end{align}
where $\lVert \cdot \rVert_1$ denotes L$_1$ norm and $\Psi^*(I)$ denotes the locations of the $N_k$ detected keypoints of an image $I$ normalized to range $[0, 1]$. Note that by using the above distance function, we are optimizing robot trajectories to align with the given human demo (i.e. when the human is $p\%$ into the demo, the robot also aims to be roughly $p\%$ into the execution).
\subsection{Few-shot Motion Optimization with BO}
\label{method:bo}
\edit{With the learned keypoint representation, we can use the keypoint-based distance between a robot execution and a human demonstration as the objective for our optimization problem formulated in Section \ref{sec:prelim}.}
The remaining problem is how to efficiently find trajectories to be executed on the robot that best imitate the human demo. Training robot skills on large amounts of simulated data and then using sim2real techniques to transfer the skill to a real system is a common robot learning paradigm. However, deformable and granular objects are hard to simulate realistically and therefore pose a challenge for sim2real transfer. Thus, we propose a method to directly optimize the motion policy on the real robot.
\begin{figure}
\vspace{6pt}
\centering
\includegraphics[width=0.47\textwidth]{figs/imagined_trajectories.pdf}
\caption{Imagined trajectories as initial candidates for BO.}
\label{fig:imagined_trajectories}
\end{figure}
\subsubsection{Periodicity Estimation with RepNet}
To imitate periodic manipulation skills shown in the human demo, we need to first determine the periodicity of this demo.
We use RepNet \cite{dwibedi2020counting} -- an approach that can estimate when and how often a periodic task is repeated in a video to estimate periodicity. We observe that the RepNet model trained on the Countix dataset can reliably predict the periodicity of the human demos that we consider. So, we use the trained model (without any finetuning) to predict the number of periods $n^{rep}_H = \text{RepNet}(V_H)$ of the human demo $V_H$.
\begin{figure*}[h]
\vspace{6pt}
\centering
\begin{subfigure}{.28\textwidth}
\centering
\includegraphics[width=\linewidth]{figs/method_comparison.pdf}
\end{subfigure}
\hspace{1pt}
\begin{subfigure}{.21\textwidth}
\centering
\includegraphics[width=\linewidth]{figs/sim_wiping.pdf}
\caption{\texttt{Table Wiping}}
\end{subfigure}
\begin{subfigure}{.21\textwidth}
\centering
\includegraphics[width=\linewidth]{figs/sim_winding.pdf}
\caption{\texttt{Rope Winding}}
\end{subfigure}
\begin{subfigure}{.21\textwidth}
\centering
\includegraphics[width=\linewidth]{figs/sim_stirring.pdf}
\caption{\texttt{Food Stirring}}
\end{subfigure}
\caption{
The left part shows the objectives used in the methods we compare, then explains our evaluation metric -- `performance' in plots on the right. Plots (a)-(c) show the performance of the competing methods in the 3 tasks we consider, in simulation.
%
For the \textit{Direct Imitation} baseline, executions are fixed, no finetuning between trials.
We include two versions of this baseline: (1x) -- trained from the robot play data; (2x) -- trained from twice the amount of that data, so that the total size of training data exceeds the size of robot play data + the 50 trials of interactions. The performance saturates, showing no benefit from additional training data (red lines match).
%
For the other 3 methods, we execute 50 trials for each run. For every method, we do 3 runs using 3 random seeds. The solid lines denote the mean of performance across 3 seeds; the shaded areas denote the standard deviation of the performance values.
%
The MBIL line denotes model-based imitation learning baseline, which learns dynamics using keypoints as states and uses MPC for planning.
}
\label{fig:sim_plots}
\end{figure*}
\subsubsection{Motion Optimization with BO}
We propose to optimize single-period waypoints as BO parameters, then use rhythmic DMPs to fit a smooth trajectory and unroll it for multiple periods, as illustrated in Figure~\ref{fig:bo}.
Concretely, BO optimizes single-period waypoints: $\mathbf{w} = [\mathbf{v}_1, \mathbf{v}_2, \ldots, \mathbf{v}_L]$.
To execute a BO sample, we apply a cubic smoothing to the sampled waypoints, then fit a rhythmic DMP to this smooth trajectory and execute it for $n_H^{rep}$ periods, with the goal $g$ of the DMP shifting $\mathbf{v}_L - \mathbf{v}_1$ between two consecutive periods.
During conventional BO, the candidates for the first few trials are sampled at random. In the subsequent trials, an acquisition function samples $N$ candidates at random from the search space, then evaluates their posterior mean and variance to select the most promising next candidate. However, in high-dimensional spaces (above 10D) it is unlikely to sample a well-performing candidate randomly. Even with waypoints as the search space for BO, the space is very large. A leading BO method that recently reported success in high-dimensions~\cite{eriksson2019scalable} was not able to reliably succeed on our tasks within 50 trials. Hence, \edit{we need to} further improve the data efficiency of BO.
Our insight is that robot play data contains meaningful interactions that can help BO to focus on the promising regions of the search space.
To effectively use this data, we first generate a set of $N_s$ play data segments $\mathcal{S} = \{\tau_s^1, \tau_s^2, \ldots, \tau_s^{N_s}\}$, where each element $\tau_s^i = \edit{\vc{x}}_{RP}^{t_i:t_i + T_s}$ is a randomly sampled fixed length trajectory in the robot play data of length $T_s$. Then, we generate `imagined trajectories' by rejection sampling segment sequences $\tau_I = (\tau_s^{k_1}, \ldots, \tau_s^{k_m})$ such that the end position $\tau_s^{k_i, T_s}$ of each segment $\tau_s^{k_i}$ is less than $d_\text{seg}$ away from the start position $\tau_s^{k_{i+1}, 1}$ of the next segment $\tau_s^{k_{i+1}}$ in L2 distance.
Then, we can find the corresponding image frames of $\tau_I$ to construct an `imagined' video $V_I$,
and this video can be evaluated using the objective score function $D(V_H^*, V_I)$, where $V_H^*$ is a single-period demo trimmed from the original human demo according to RepNet period split. We then select the top $N_{\text{imagined}}$ trajectories with the highest estimated scores and construct a set of \textit{initial candidates}: $\{\mathbf{w}_i\}_{i=1...N_{\text{imagined}}}$. Each candidate $\mathbf{w}_i$ is represented by a set of waypoints sub-sampled from the imagined trajectory.
We illustrate this process in Figure~\ref{fig:imagined_trajectories}.
To warm-start BO, we sample the candidates for the first few trials from $\{\mathbf{w}_i\}_{i=1...N_{\text{imagined}}}$. By construction, these represent the waypoints of the trajectories that have a high alignment with the human demo for the 1st period. In the subsequent trials, we augment the pool of the candidates considered by the acquisition function by sampling in the regions close to these initial candidates. With that, the acquisition function can help us focus the search on the regions close to the initial candidates, but is not restricted to these regions. Hence, we avoid placing hard restrictions on the search space of BO based on prior information. As a result, our BO extension retains theoretical guarantees of BO, such as consistency and regret bounds. |
\section{Introduction}
Parametric computer-aided design (CAD) tools are at the core of the design process in mechanical engineering, aerospace, architecture, and many other disciplines.
These tools enable the specification of two- and three-dimensional structures in a way that empowers the user to explore parameterized variations on their designs, while also providing a structured representation for downstream tasks such as simulation and manufacturing.
In the context of CAD, \emph{parametric} refers to the dominant paradigm for professionals, in which the software tool essentially provides a graphical interface to the specification of a \emph{constraint program} which can then be solved to provide a geometric configuration.
This implicit programmatic workflow means that modifications to the design---alteration of dimensions, angles, etc.---propagate across the construction, even if it is an assembly with many hundreds or thousands of components.
\begin{figure}[b]
\vspace{-0.5cm}%
\centering%
\includegraphics[width=\linewidth]{figs/money.pdf}%
\vspace{-0.25cm}%
\caption[CAD sketch generation conditioned on a hand drawing]{\small CAD sketch generation conditioned on a hand drawing. Our model first conditions on a raster image of a hand-drawn sketch and generates sketch primitives. Given the primitives, the model generates constraints, and the resulting parametric sketch is imported and edited in CAD software, ultimately producing a 3D model.}%
\label{fig:money}%
\end{figure}
\begin{figure}
\centering%
\vspace{-0.1cm}%
\includegraphics[width=0.85\linewidth]{figs/end_to_end.pdf}%
\vspace*{0.3cm}%
\caption[Editing inferred sketches.]{\small Editing inferred sketches. We take raster images of hand-drawn sketches (left), infer the intended sketches using our image-conditional primitive model and constraint model (center), then solve the resultant sketches in CAD software, illustrating edit propagation arising from the constraints (right).} \vspace*{-0.6cm}
\label{fig:end2end}%
\end{figure}
At an operational level, parametric CAD starts with the specification of two-dimensional geometric representations, universally referred to as ``sketches'' in the CAD community \citep{shah1998designing, camba2016parametric, choi2002exchange}.
A sketch consists of a collection of geometric primitives (e.g., lines, circles), along with a set of \emph{constraints} that relate these primitives to one another.
Examples of sketch constraints include parallelism, tangency, coincidence, and rotational symmetry.
Constraints are central to the parametric CAD process as they reflect abstract design intent on the part of the user, making it possible for numerical modifications to coherently propagate into other parts of the design.
Figure~\ref{fig:end2end}, which demonstrates a capability made possible by our approach, illustrates how constraints are used to express design intent: the specific dimensions of the final model (shown on the right) can be modified while preserving the geometric relationships between the primitives.
These constraints are the basis of our view of parametric CAD as the specification of a program; indeed the actual file formats themselves are indistinguishable from conventional computer code.
This design paradigm, while powerful, is often a challenging and tedious process.
An engineer may execute similar sets of operations many dozens of times per design.
Furthermore, general motifs may be repeated across related parts, resulting in duplicated effort.
Learning to accurately predict such patterns has the potential to mitigate the inefficiencies involved in repetitive manual design.
In addition, engineers often begin visualizing a design by roughly drawing it by hand (\cref{fig:money}).
The automatic and reliable conversion of such hand drawings, or similarly noisy inputs such as 3D scans, to parametric, editable models remains a highly sought feature.
In this work, we introduce Vitruvion,
a generative model trained to synthesize coherent CAD sketches by autoregressively sampling geometric primitives and constraints (\cref{fig:model}).
The model employs self-attention \citep{Transformer} to flexibly propagate information about the current state of a sketch to a next-step prediction module.
This next-step prediction scheme is iterated until the selection of a stop token signals a completed sample.
The resultant geometric constraint graph is then handed to a solver that identifies the final configuration of the sketch primitives.
By generating sketches via constraint graphs, we allow for automatic edit propagation in standard CAD software.
Constraint supervision for the model is provided by the SketchGraphs dataset \cite{SketchGraphs}, which pairs millions of real-world sketches with their ground truth geometric constraint graphs.
In addition to evaluating the model via log-likelihood and distributional statistics, we demonstrate the model's ability to aid in three application scenarios when conditioned on specific kinds of context.
In \textbf{autocomplete}, we first prime the model with an incomplete sketch (e.g., with several primitives missing) and query for plausible completion(s).
In \textbf{autoconstrain}, the model is conditioned on a set of primitives (subject to position noise), and attempts to infer the intended constraints.
We also explore \textbf{image-conditional synthesis}, where the model is first exposed to a raster image of a hand-drawn sketch.
In this case, the model is tasked with inferring both the primitives and constraints corresponding to the sketch depicted in the image.
Overall, we find the proposed model is capable of synthesizing realistic CAD sketches and exhibits potential to aid the mechanical design workflow.
\begin{figure}[t]
\centering%
\vspace{-0.5cm}%
\includegraphics[width=0.8\linewidth]{figs/model_fig.pdf}%
\vspace{-0.00cm}%
\caption[Factorization of CAD sketch synthesis]{\small We factorize CAD sketch synthesis into two sequence modeling tasks: primitive generation (left) and constraint generation (right). The constraint model conditions on the present primitives and outputs a sequence of constraints, serving as edges in the corresponding geometric constraint graph. A separate solver (standard in CAD software) is used to adjust the primitive coordinates, if necessary.}%
\label{fig:model}%
\vspace{-0.1cm}%
\end{figure}
\section{Related Work}
When discussing related work, it is important to note that in computer-aided design, the word ``sketch'' is a term of art that refers to this combination of primitives and constraints.
In particular, it should not be confused with the colloquial use of the word ``sketch'' as has sometimes appeared in the machine learning literature, as in \citet{Sketch-RNN} discussed below.
\paragraph{CAD sketch generation}
The SketchGraphs dataset~\citep{SketchGraphs} was recently introduced to target the underlying relational structure and construction operations found in mechanical design software.
Specifically, the dataset provides ground truth geometric constraint graphs, and thus the original design steps, for millions of real-world CAD sketches created in Onshape\footnote{\url{https://www.onshape.com}}.
Initial models are trained in \citet{SketchGraphs} for constraint graph generation using message-passing networks, but without any learned attributes for the positions of the primitives; the sketch configuration is determined via constraints alone, limiting the sophistication of output samples.
Following SketchGraphs, several works concurrent to ours explore generative modeling of CAD sketches utilizing transformer-based architectures \citep{CurveGen, DMCAD, SketchGen, DeepCAD}.
In \citet{CurveGen}, modeling is limited to primitives, without any learned specification of geometric constraints;
Vitruvion, however, enables edit propagation in standard CAD software by outputting constraints.
Similar to our approach, both primitives and constraints are modeled in \citet{DMCAD, SketchGen}.
In contrast to our work, these do not condition on hand-drawn sketches nor consider noisy inputs when training on the constraint inference task (which proves crucial for generalization in \cref{sec:constraint_model_performance}).
Unlike the above works that solely target 2D sketches (including ours), \citet{DeepCAD} attempts to model sequences of both 2D and 3D CAD commands, but, similarly to \citep{CurveGen}, does not output sketch constraints.
\paragraph{Geometric program synthesis}
CAD sketch generation may be viewed via the broader lens of geometric program synthesis.
This comprises a practical subset of program synthesis generally concerned with inferring a set of instructions (a program) for reconstructing geometry from some input specification.
Recent examples include inferring TikZ drawing commands from images \citep{InferGraphics} and constructive solid geometry programs from voxel representations \citep{CSGNet}.
Related to these efforts is the highly-sought ability to accurately reverse engineer a mechanical part given underspecified and noisy observations of it (e.g., scans, photos, or drawings).
One of the conditional generation settings we explore in this work is training a version of our model that generates parametric sketches when given a raster hand drawing.
Note that previous work on geometric program synthesis have not generally considered constraints, which are critical to the CAD application.
\paragraph{Vector graphics generation}
Generative models have been successfully applied to vector graphics in recent years.
Sketch-RNN \citep{Sketch-RNN} is trained on sequences of human drawing strokes to learn to produce vector drawings of generic categories (dog, bus, etc.).
Note the common meaning of the word \emph{sketch} differs slightly from its usage as a term of art in mechanical design (as we use it in this work).
DeepSVG \citep{DeepSVG} models SVG icons, demonstrating flexible editing with latent space interpolation.
Less recent are tracing programs, which can convert raster drawings to vector graphics and have been widely available for decades (e.g., \citet{Portrace}).
None of this previous work considers constraints as a mechanism for specifying design intent as performed in CAD.
The image-conditional version of Vitruvion also takes as input a raster image (hand-drawn) and can output vector graphics.
Unlike tracing, which attempts to faithfully reproduce an input image as closely as possible, our model attempts to infer the sequence of geometric primitives and constraints intended by the user, learning to ignore noise in the input.
\paragraph{Graph-structured modeling}
This work is also related to modeling of graph structures, via our use of geometric constraint graphs.
Graphs constitute a natural representation of relational structure across a variety of domains, including chemistry, social networks, and robotics.
In recent years, message-passing networks \citep{MPN, ConvNetMol, GraphConv}, building off of \citet{Scarselli}, have been extensively applied to generative modeling of graph-structured data, such as molecular graphs \citep{JT-VAE, CG-VAE}.
These networks propagate information along edges, learning node- or graph-level representations. %
While message passing between adjacent nodes can be a useful inductive bias, it can be an impediment to effective learning when communication between distant nodes is necessary, e.g., in sparse graphs \citep{BodyCage}.
The constraint graphs in our work are sparse, as the constraints (edges) grow at most linearly in the number of primitives (nodes). %
Recent work leverages the flexible self-attention mechanism of transformers \citep{Transformer} to generate graph-structured data without the limitations of fixed, neighbor-to-neighbor communication paths, such as for mesh generation \citep{PolyGen}.
Our work similarly bypasses edge-specific message passing, using self-attention to train a generative model of geometric constraint graphs representing sketches.
\section{Method}
We aim to model a distribution over parametric CAD sketches.
Each sketch is comprised of both a sequence of primitives, $\mathcal{P}$, and a sequence of constraints, $\mathcal{C}$.
We note that these are sequences, as opposed to unordered sets, due to the ground truth orderings included in the data format~\citep{SketchGraphs}.
Primitives and constraints are represented similarly in that both have a categorical type (e.g., Line, Arc, Coincidence) as well as a set of additional parameters which determine their placement/behavior in a sketch.
Each constraint includes \emph{reference} parameters indicated precisely which primitive(s) (or components thereof) must adhere to it.
Rather than serving as static geometry, CAD sketches are intended to be dynamic---automatically updating in response to altered design parameters.
Likewise, equipping our model with the ability to explicitly constrain generated primitives serves to ensure the preservation of various geometric relationships during downstream editing.
We divide the sketch generation task into three subtasks: primitive generation, constraint generation, and constraint solving.
In our approach, the first two tasks utilize learned models, while the last step may be conducted by any standard geometric constraint solver.\footnote{We use D-Cubed (Siemens PLM) for geometric constraint solving via Onshape's public API.}
We find that independently trained models for primitives and constraints allows for a simpler implementation.
This setup is similar to that of \citet{PolyGen}, where distinct vertex and face models are trained for mesh generation.
The overall sketch generation proceeds as:
\begin{align} \label{eq:factorization}
p_\theta(\mathcal{P}, \mathcal{C} \mid \text{ctx}) &= p_\theta(\mathcal{C} \mid \mathcal{P})p_\theta(\mathcal{P} \mid \text{ctx}) &
\mathcal{S} &= \text{solve}(\mathcal{P}, \mathcal{C})
\end{align}
where:
\begin{itemize}
\item $\mathcal{P}$ is a sequence of primitives, including parameters specifying their positions;
\item $\mathcal{C}$ is a sequence of constraints, imposing mandatory relationships between primitives;
\item $\mathcal{S}$ is the resulting sketch, formed by invoking a solve routine on the $(\mathcal{P},\mathcal{C})$ pair; and
\item ctx is an optional context for conditioning, such as an image or prefix (primer).
\end{itemize}
This factorization assumes $\mathcal{C}$ is conditionally independent of the context given $\mathcal{P}$.
For example, in an image-conditional setting, access to the raster representation of a sketch is assumed to be superfluous for the constraint model given accurate identification of the portrayed primitives and their respective positions in the sketch.
Unconditional samples may be generated by providing a null context.
\subsection{Primitive Model} \label{sec:primitive_model}
The primitive model is tasked with autoregressive generation of a sequence of geometric primitives.
For each sketch, we factorize the distribution as a product of successive conditionals,
\begin{equation}
p_\theta(\mathcal{P} \mid \text{ctx}) = \prod_{i=1}^{N_\mathcal{P}}p_\theta(\mathcal{P}_i \mid \{\mathcal{P}_j\}_{j < i}, \text{ctx}),
\end{equation}
where $N_\mathcal{P}$ is the number of primitives in the sequence.
Each primitive is represented by a tuple of tokens indicating its type and its parameters.
The primitive type may be one of four possible
shapes (here either arc, circle, line, or point), and the associated parameters include both continuous positional coordinates and an \texttt{isConstruction} Boolean.\footnote{\emph{Construction} or \emph{virtual} geometry is employed in CAD software to aid in applying constraints to regular geometry.
\texttt{isConstruction} specifies whether the given primitive is to be physically realized (when false) or simply serve as a reference for constraints (when true).}
We use \emph{sequence} to emphasize that our model treats sketches as constructions arising from a step-by-step design process.
The model is trained in order to maximize the log-likelihood of $\theta$ with respect to the observed token sequences.
\begin{wraptable}{r}{0.55\linewidth}
\centering
\scalebox{0.75}{%
\begin{tabular}{c c c c}
\toprule
\texttt{Arc} & \texttt{Circle} & \texttt{Line} & \texttt{Point} \\
$(x_1, y_1, x_{\text{mid}}, y_{\text{mid}}, x_2, y_2)$ & $(x, y, r)$ & $(x_1, y_1, x_2, y_2)$ & $(x, y)$ \\
\midrule
\end{tabular}
}
\caption[Primitive parameterizations]{\small %
We convert the native Onshape numerical parameters to the above forms for modeling. Subscripts of $1$, $\text{mid}$, and $2$ indicate start, mid, and endpoints, respectively.}
\vspace{-0.4cm}
\label{table:prim_params}
\end{wraptable}
\paragraph{Ordering.}
Abstractly, the mapping from constructions sequences to final sketch geometries is clearly non-injective; it is typical for many construction sequences to result in the same final geometry.
Nevertheless, empirically, we know that the design routes employed by engineers often admit evident patterns, e.g., greater involvement of earlier ``anchor'' primitives in constraints~\citep{SketchGraphs}.
In our training data from SketchGraphs, the order of the design steps taken by the original sketch creators is preserved.
Thus, %
we expose the model to this ordering, training our model to autoregressively predict subsequent design steps conditioned on prefixes.
By following an explicit representation of this ordering, our model is amenable to applications in which conditioning on a prefix of design steps is required, e.g., autocompleting a sketch.
This approach also constrains the target distribution such that the model is not required to distribute its capacity among all orderings.
\paragraph{Parameterization.}
SketchGraphs uses the parameterization provided by Onshape.
In order to accommodate the interface to the Onshape constraint solver, these are often \emph{over-parameterizations}, e.g.\ line segments are defined by six continuous parameters rather than four.
We compress the original parameterizations into minimal canonical parameterizations for modeling (\cref{table:prim_params}) and inflate them back to the original encoding prior to constraint solving.
\paragraph{Normalization.} \label{sec:normalization}
The sketches in SketchGraphs exhibit widely varying scales in physical space, from millimeters to meters.
In addition, because Onshape users are free to place sketches anywhere in the coordinate plane, they are often not centered with respect to the origin in the coordinate space.
To reduce modeling ambiguity, we modify the primitive coordinates such that each sketch's square bounding box has a width of one meter and is centered at the origin.
\paragraph{Quantization.}
The sketch primitives are inherently described by continuous parameters such as $x, y$ Cartesian coordinates, radii, etc.
As is now a common approach in various continuous domains, we quantize these continuous variables and treat them as discrete.
This allows for more flexible, yet tractable, modeling of arbitrarily shaped distributions \citep{WaveNet}.
Following the sketch normalization procedure described in \cref{sec:normalization}, we apply 6-bit uniform quantization to the primitive parameters.
Lossy quantization is tolerable as the constraint model and solver are ultimately responsible for the final sketch configuration.
\paragraph{Tokenization.}
Each parameter (whether categorical or numeric) in the input primitive sequence is represented using a three-tuple comprised of a value, ID, and position token.
Value tokens play a dual role.
First, a small portion of the range for value tokens is reserved to enumerate the primitive types.
Second, the remainder of the range is treated separately as a set of (binned) values to specify the numerical value of some parameter, for example, the radius of a circle primitive.
ID tokens specify the type of parameter being described by each value token.
Position tokens specify the ordered index of the primitive that the given element belongs to.
See \cref{appendix:tokenization_details} for further details on the tokenization procedure.
\paragraph{Embeddings.}
We use learned embeddings for all tokens.
That is, each value, ID, and position token associated with a parameter is independently embedded into a fixed-length vector.
Then, we compute an element-wise sum across each of these three token embedding vectors to produce a single, fixed-length embedding to represent the parameter embedding.
The sequence of parameter embeddings are then fed as input to the decoder described below.
\paragraph{Architecture.}
The unconditional primitive model is based on a decoder-only transformer \citep{Transformer} operating on the flattened primitive sequences.
To respect the temporal dependencies of the sequence during parallel training, the transformer decoder leverages standard triangular masking during self-attention.
The last layer of the transformer outputs a final representation for each token that then passes through a linear layer to produce a logit for each possible value token.
Taking a softmax then gives the categorical distribution for the next-token prediction task.
We discuss conditional variants in \cref{sec:context}.
See the appendix for further details.
\subsection{Constraint Model} \label{sec:constraint_model}
Given a sequence of primitives, the constraint model is tasked with autoregressive generation of a sequence of constraints.
As with the primitive model, we factorize the constraint model as a product of successive conditionals,
\begin{equation}
p_\theta(\mathcal{C} \mid \mathcal{P}) = \prod_{i=1}^{N_\mathcal{C}} p_\theta(\mathcal{C}_i \mid \{\mathcal{C}_j\}_{j< i}, \mathcal{P}),
\end{equation}
where $N_\mathcal{C}$ is the number of constraints.
Each constraint is represented by a tuple of tokens indicating its type and parameters.
Like the primitive model, the constraint model is trained in order to maximize the log-likelihood of $\theta$ with respect to the observed token sequences.
Here, we focus our constraint modeling on handling all categorical constraints containing one or two reference parameters.
This omits constraints with numerical parameters (such as scale constraints) as well as ``hyperedge'' constraints that have more than two references.
\paragraph{Ordering.}
We canonicalize the order of constraints according to the order of the primitives they act upon, similar to \citet{SketchGraphs}.
That is, constraints are sorted according to their latest member primitive.
For each constraint type, we arbitrarily canonicalize the ordering of its tuple of parameters.
Constraint tuples specify the type of constraint and its reference parameters.
\paragraph{Tokenization.}
The constraint model employs a similar tokenization scheme to the primitive model in order to obtain a standardized sequence of integers as input.
For each parameter, a triple of value, ID, and position tokens indicate the parameter's value, what type of parameter it is, and which ordered constraint it belongs to, respectively.
\paragraph{Architecture.}
The constraint model employs an encoder-decoder transformer architecture which allows it to condition on the input sequence of primitives.
A challenge with the constraint model is that it must be able to specify references to the primitives that each constraint acts upon, but the number of primitives varies among sketches.
To that end, we utilize variable-length softmax distributions (via pointer networks \citep{vinyals2015pointer}) over the primitives in the sketch to specify the constraint references.
This is similar to \citep{PolyGen}, but distinct in that our constraint model is equipped to reference hierarchical entities (i.e., primitives or sub-primitives). %
\paragraph{Embeddings.}
Two embedding schemes are required for the constraint model.
Both the input primitive token sequence as well as the target output sequence (constraint tokens) must be embedded.
The input primitive tokens are embedded identically to the standalone primitive model, using lookup tables.
A transformer encoder then produces a sequence of final representations for the primitive tokens.
These vectors serve a dual purpose: conditioning the constraint model and embedding references in the constraint sequence.
The architecture of the encoder is similar to that of the primitive model, except it does not have the output linear layer and softmax.
From the tuple of representations for each primitive, we extract reference embeddings for both the primitive as a whole (e.g., a line segment) and each of its sub-primitives (e.g., a line endpoint) that may be involved in constraints.
\paragraph{Noise injection.}
The constraint model must account for potentially imperfect generations from the preceding primitive model.
Accordingly, the constraint model is conditioned on primitives whose parameters are subjected to independent Gaussian noise during training.
\subsection{Context Conditioning} \label{sec:context}
Our model may be optionally conditioned on a context.
As described below \cref{eq:factorization}, we directly expose the primitive model to the context while the constraint model is only implicitly conditioned on it via the primitive sequence.
Here, we consider two specific cases of application-relevant context: primers and images of hand-drawn sketches. %
\paragraph{Primer-conditional generation.}
Here, a primer is a sequence of primitives representing the prefix of an incomplete sketch.
Conditioning on a primer consists of presenting the corresponding prefix to a trained primitive model; the remainder of the primitives are sampled from the model until a stop token is sampled, indicating termination of the sequence.
This emulates a component of an ``autocomplete'' application, where CAD software interactively suggests next construction operations.%
\paragraph{Image-conditional generation.}
In the image-conditional setting, we are interested in accurately recovering a parametric sketch from a raster image of a hand-drawn sketch.
This setup is inspired by the fact that engineers often draw a design on paper or whiteboard from various orthogonal views prior to investing time building it in a CAD program.
Accurate inference of parametric CAD models from images or scans remains a highly-sought feature in CAD software, with the potential to dramatically reduce the effort of translating from rough paper scribble to CAD model.
When conditioning on an image, the primitive model is augmented with an image encoder.
We leverage an architecture similar to a vision transformer \citep{VisionTransformer}, first linearly projecting flattened local image patches to a sequence of initial embeddings, then using a transformer encoder to produce a learned sequence of context embeddings.
The primitive decoder cross-attends into the image context embedding.
The overall model is trained to maximize the probability of the ground truth primitives portrayed in the image; we condition on a sequence of patch embeddings.
This is similar to the image-conditioning route taken in \citet{PolyGen} for mesh generation, which also uses a sequence of context embeddings produced by residual networks \citep{ResBlocks}.
\paragraph{Hand-drawn simulation.}
We aim to enable generalization of the image-conditional model described above to hand-drawn sketches, which can potentially support a much wider array of applications than only conditioning on perfect renderings.
This requires a noise model to emulate the distortions introduced by imprecise hand drawing, as compared to the precise rendering of a parametric sketch using software.
Two reasonable noise injection targets are the parameters underlying sketch primitives and the raster rendering procedure.
Our noise model takes a hybrid approach, subjecting sketch primitives to random translations/rotations in sketch space, and augmenting the raster rendering with a Gaussian process model.
A full description is provided in \cref{appendix:hand-drawn-noise}.
\section{Experiments}
We evaluate several versions of our model in both primitive generation and constraint generation settings.
Quantitative assessment is conducted by measuring negative log-likelihood (NLL) and predictive accuracy on a held-out test set as well as via distributional statistics of generated sketches.
We also examine the model's performance on conditional synthesis tasks.
\subsection{Training Dataset}
Our models are trained on the SketchGraphs dataset \citep{SketchGraphs}, which consists of several million CAD sketches collected from Onshape.
We use the filtered version, which is restricted to sketches comprised of the four most common primitives (arcs, circles, lines, and points), a maximum of 16 primitives per sketch, and a non-empty set of constraints.
We further filter out any sketches with fewer than six primitives to eliminate most trivial sketches (e.g., a simple rectangle).
We randomly divide the filtered collection of sketches into a 92.5\% training, 2.5\% validation, and 5\% testing partition.
Training is performed on a server with four Nvidia V100 GPUs, and our models take between three and six hours to train.
See \cref{appendix:vitruv_train_details} for full details of the training procedure.
\paragraph{Deduplication.}
As components of larger parts and assemblies, CAD sketches may sometimes be reused across similar designs.
Onshape's native document-copying functionality supports this.
In addition, similar sketches may result from users of Onshape following tutorials, importing files from popular CAD collections, or by coincidentally designing for the same use case.
As a result, a portion of the examples in SketchGraphs may be considered to be duplicates, depending on the precise notion of duplicate adopted.
For example, two sketches may exhibit visually similar geometry yet differ in their ordering of construction operations, placement in the coordinate plane, or constraints.
First, we normalize each sketch via centering and uniform rescaling.
The numerical parameters in the primitive sequence are quantized to six bits, mapping their coordinates to a $64 \times 64$ grid.
Then, for any set of sketches with identical sequences of primitives (including their quantized parameters), we keep only one, resulting in a collection of 1.7 million unique sketches.
This procedure removes all sketches that have approximately equivalent geometry with the same ordering of construction operations.
\subsection{Baselines}
We evaluate several ablation and conditional variants of our model as well as four baselines:
1) A uniform sampler that selects tokens with equal probability at each time step.
2) An LZMA compression baseline (see \cref{appendix:compression}).
3) The models proposed in \citet{SketchGraphs} (see \cref{appendix:sg-baseline}).
4) The models proposed in \citet{CurveGen} (see \cref{appendix:qual-baselines}).
\subsection{Primitive Model Performance}
\paragraph{Unconditional model.}
The unconditional variant of our model is trained to approximate the distribution over primitive sequences without any context.
As shown in \cref{table:prim_model}, the unconditional primitive model achieves a substantially lower NLL than both the uniform and compression baselines.
We employ nucleus sampling \citep{NucleusSampling} with cumulative probability parameter of~${p=0.9}$ to remove low-probability tokens at each step. %
\cref{fig:unconditional_samples} displays a set of random samples.
\begin{figure}\BottomFloatBoxes
\begin{floatrow}%
{%
\ffigbox[8cm][][t]
{%
\caption[Random samples]{\small Random examples from the SketchGraphs dataset (left) and random samples from our unconditional primitive model (right).\\\hspace{\textwidth}}%
\label{fig:unconditional_samples}%
}%
{%
\includegraphics[width=\linewidth]{figs/sg_dataset_vs_unconditional.pdf}
}%
}%
\vbox%
{
\ffigbox[5.5cm][][t]
{%
\caption[Loss per primitive position]{\small Average NLL per primitive position for the unconditional primitive model. In general, the model achieves lower NLL for primitives later in a sequence, as ambiguity decreases. However, a staircase pattern of length four is also apparent, due to the high prevalence of rectangles.}
\label{fig:loss_per_primitive_position}
}%
{%
\includegraphics[width=\linewidth]{figs/loss_per_primitive_position_trimmed.pdf}
}
\vss
\capbtabbox
{%
\caption[Primitive model evaluation]{\small Primitive model evaluation. Negative test log-likelihood in bits/primitive and bits/sketch. Next-step prediction accuracy (per-token avg.) is also provided. Standard deviation across 5 replicates in parentheses.}%
\label{table:prim_model}%
}%
{%
\scalebox{0.65}{%
\setlength\tabcolsep{1.5pt}%
\begin{tabular}{lrrr} \toprule
& Bits/Prim & Bits/Sketch & Accuracy \\ \midrule
Uniform & 33.41 \hphantom{(0.00)} & 350.2 \hphantom{(0.0)} & 1.3 \hphantom{(0.0)} \% \\
Compression & 15.34 \hphantom{(0.00)} & 160.7 \hphantom{(0.0)} & --- \% \\
Unconditional & 6.35 (0.01) & 66.6 (0.1) & 71.7 (0.0)\% \\
Unconditional (perm.) & 8.15 (0.39) & 85.5 (4.1) & 60.3 (1.8)\% \\
Image-conditional & 3.50 (0.05) & 35.1 (0.5) & 82.2 (0.2)\% \\ \bottomrule
\end{tabular}%
}%
}}%
\end{floatrow}%
\vspace{-1cm}%
\end{figure}
To assess how the predictive performance of the model depends on %
the sequence order, %
we train an unconditional primitive model on sequences where the primitives are shuffled, removing the model's exposure to the true ordering.
As shown in \cref{table:prim_model}, this results in a decline on next-step prediction.
Importantly, the erasure of primitive ordering removes useful local hierarchical structure (e.g., high-level constructs like rectangles; see \cref{fig:loss_per_primitive_position}) as well as global temporal structure.
\paragraph{Image-conditional generation.}
\cref{table:prim_model} displays the performance of the basic image-conditional variant of the primitive model.
By taking a visual observation of the primitives as input, this model is able to place substantially higher probability on the target sequence.
In \cref{table:hand_model}, we evaluate the performance of several image-conditional models on a set of human-drawn sketches.
These sketches were hand-drawn on a tablet computer in a $128 \times 128$-pixel bounding box, where the drawer first eyeballs a test set sketch, and then attempts to reproduce it.
Likewise, each image is paired up with its ground truth primitive sequence, enabling log-likelihood evaluation.
We test three versions of the image-conditional model on the hand drawings, each trained on a different type of rendering: precise renderings, renderings from the hand-drawn simulator, and renderings with random affine augmentations of the hand-drawn simulator.
Both the hand-drawn simulation and augmentations substantially improve performance.
Despite never observing a hand drawing during training, the model is able to effectively interpret these.
\Cref{fig:conditional_samples} displays random samples from the primitive model when conditioned on images of hand-drawn sketches.
\begin{figure}
\centering
\vspace{-0.1cm}%
\includegraphics[width=0.9\linewidth]{figs/image_conditional_samples.pdf}%
\vspace{-0.3cm}%
\caption[Image-conditional samples]{\small Image-conditional samples from our primitive model. Raster images of real hand-drawn sketches (top) are input to the primitive model. Four independent samples are then generated for each input. The model makes occasional mistakes, but tends to largely recover the depicted sketch with only a few samples.}
\label{fig:conditional_samples}
\vspace{0.1cm}%
\end{figure}
\begin{figure}\BottomFloatBoxes
\vspace{-0.5cm}%
\begin{floatrow}%
\capbtabbox[\FBwidth]
{%
\caption[Evaluation of image-conditional primitive model on hand-drawn sketches]{\small Image-conditional primitive models evaluated on real hand-drawn sketches. %
NLL and predictive accuracy (per-token avg.) metrics are shown for three models trained with different data augmentation schemes.}
\label{table:hand_model}
}%
{%
\scalebox{0.67}{%
\begin{tabular}{@{}lrrrr@{}} \toprule
Training regimen & Bits/Prim & Bits/Sketch & Accuracy \\ \midrule
Precise rendering & 22.80 & 292.3 & 47.0\% \\
Hand-drawn augmentation & 10.59 & 135.5 & 65.5\% \\
Hand-drawn + affine & 6.14 & 78.6 & 75.9\% \\
\bottomrule
\end{tabular}
}
}
\capbtabbox[\FBwidth]
{%
\caption[Constraint model evaluation]{\small Constraint model evaluation with per-token perplexity and accuracy.
A model trained on noiseless data degrades substantially in the presence of noise.}
\label{table:constraint_model}
}%
{%
\scalebox{0.67}{%
\begin{tabular}{@{}lrrrr@{}} \toprule
& \multicolumn{2}{c}{Noiseless Testing} & \multicolumn{2}{c}{Noisy Testing} \\
\cmidrule(lr){2-3} \cmidrule(lr){4-5}
Model & Perplexity & Accuracy & Perplexity & Accuracy \\ \midrule
Uniform & 4.984 & 0.6\% & 4.984 & 0.6\% \\
SketchGraphs & 0.246 & --\% & 0.900 & --\% \\
Noiseless training & 0.184 & 91.5\% & 0.903 & 68.6\% \\
Noisy training & 0.198 & 90.3\% & 0.203 & 90.6\% \\ \bottomrule
\end{tabular}
}
}%
\end{floatrow}%
\vspace{-0.55cm}%
\end{figure}
\paragraph{Primer-conditional generation.}
We assess the model's ability to complete primitive sequences when only provided with an incomplete sketch.
Here, we randomly select a subset of test set sketches, deleting 40\% of the primitives from the suffix of the sequence.
The remaining sequence prefix then serves to prime the primitive model.
\cref{fig:autocomplete_prims} displays random examples of priming the primitive model with incomplete sketches.
Because just over half of the original primitives are in the primer, there is a wide array of plausible completions.
In some cases, despite only six completions for each input, the original sketch is recovered.
We envision this type of conditioning as part of an interactive application where the user may query for $k$ completions from a CAD program and then select one if it resembles their desired sketch.
\subsection{Constraint Model Performance} \label{sec:constraint_model_performance}
We train and test our constraint model with two different types of input primitive sequences: noiseless primitives and noise-injected primitives.
Noise-injection proves to be a crucial augmentation in order for the constraint model to generalize to imprecise primitive placements (such as in the image-conditional setting).
As shown in \cref{table:constraint_model,fig:constraints_noiseless_vs_noisy}, while both versions of the model perform similarly on a noiseless test set, the model trained according to the original primitive locations fails to generalize outside this distribution.
In contrast, exposing the model to primitive noise during training substantially improves performance.
\paragraph{Sketch editing.}
\Cref{fig:end2end} illustrates an end-to-end workflow enabled by our model.
Our model is first used to infer primitives and constraints from a hand-drawn sketch.
We then show how the resulting solved sketch remains in a coherent state due to edit propagation. %
\section{Conclusion and Future Work}
In this work, we have introduced a method for modeling and synthesis of parametric CAD sketches, the fundamental building blocks of modern mechanical design.
Adapting recent progress in modeling irregularly structured data, our model leverages the flexibility of self-attention to propagate sketch information during autoregressive generation.
We also demonstrate conditional synthesis capabilities including converting hand-drawings to parametric designs.
Avenues for future work include extending the generative modeling framework to capture the more general class of 3D inputs; including modeling 3D construction operations and synthesis conditioned on noisy 3D scans or voxel representations.
\section{Ethics Statement}
The goal of this work is to automate a tedious process and enable greater creativity in design, but we acknowledge the potential societal impact that could arise from any elimination of human labor.
Moreover, there is the potential risk in high-stakes applications for automatically synthesizing mechanical systems that do not meet human structural standards.
We view this work, however, as being part of a larger human-in-the-loop process which will still involve human expertise for validation and deployment.
\section{Reproducibility Statement}
In addition to the model descriptions in the main text, we have provided all details of the training procedure (optimizer, learning rate, hyperparameters etc.) in Appendix~\ref{appendix:experiment-details}.
For code and pre-trained models, see \url{https://lips.cs.princeton.edu/vitruvion}.
\subsubsection*{Acknowledgments}
The authors would like to thank Yaniv Ovadia and Jeffrey Cheng for early discussions of this project.
Thanks to all members of the Princeton Laboratory for Intelligent Probabilistic Systems for providing valuable feedback.
Additionally, we would like to thank Onshape for the API access as well as the many Onshape users who created the CAD sketches comprising the SketchGraphs training data.
This work was partially funded by the DataX Program at Princeton University through support from the Schmidt Futures Foundation and by the NSF Institute for Data-Driven Dynamical Design (NSF 2118201).
|
\section{Introduction}
We are now living in a fascinating era in which we know that ``other Suns'' are the centres of concentric systems of many worlds \citep[cf.][]{Bruno1584}.
Over the past few decades, astronomers have developed techniques that allow us to ``translate'' photons into worlds \citep[see e.g.][]{exoplanets2016,Struve1952}.
More than 4800\footnote{As of Aug 11, 2021, \url{http://exoplanet.eu}.} exoplanets discovered to date have shown that worlds are abundant and diverse.
Interestingly, most of these exoplanets have been discovered indirectly by observing planet-induced variations in their host stars.
The transit \citep[][]{Charbonneau2000,Henry1999} and radial velocity \citep[RV,][]{Mayor1995} methods are the most successful techniques today in terms of number of discovered exoplanets.
However, the current number of discovered and characterised exoplanets is dwarfed by the estimated number of exoplanets in the galaxy \citep[e.g., ][]{Batalha2014,Petigura2013}.
Present and future exoplanet search and characterisation instruments, such as \emph{TESS} \citep[][]{Ricker2015}, \emph{CHEOPS} \citep[][]{Broeg2013}, \emph{PLATO} \citep[][]{Rauer2014}, \emph{ESPRESSO} \citep[][]{Pepe2010}, and \emph{SPIRou} \citep[][]{Donati2017}, will provide us with a plethora of photometric and spectroscopic data with exoplanet induced signals waiting to be discovered.
Light curves and RVs are compared with models to infer planetary, orbital, and/or stellar parameters.
These analyses are generally performed numerically using diverse models combined with computational techniques.
{Fortunately, the wealth of exoplanetary data has spurred the development of a variety of exoplanet numerical tools.
To our knowledge, the codes that allow one to analyse jointly} RV and transit data are: \texttt{allesfitter} \citep[][]{allesfitter}, \texttt{EXOFAST} \citep[][]{exofast,exofast2}, \texttt{exoplanet} \citep[][]{exoplanet}, \texttt{Exo-Stricker} \citep[][]{exostricker}, \texttt{juliet} \citep[][]{juliet}, \texttt{MCMCI} \citep[][]{mcmci}, \texttt{PASTIS} \citep[][]{pastis,pastis2}, \texttt{PlanetPAck} \citep[][]{planetpack}, \texttt{TLCM} \citep[][]{tlcm}, and \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ \citep[][]{pyaneti}.
These software packages cover a wide range of programming languages and models, and have been used extensively in the literature.
One of the current challenges in exoplanet detection is related to stellar signals in our data. Particularly, RV variations caused by stellar activity jeopardise our ability to detect planet-induced Doppler signals \citep[e.g.,][]{Queloz2001,Rajpaul2016}.
Some methods try to remove stellar activity during the RV extraction to produce activity-free RV times-series \citep[e.g.,][]{CollierCameron2020,Cretignier2020,Rajpaul2020}.
Others attempt to filter the activity induced signals in the RV time-series \citep[e.g.,][]{Barragan2018b,Hatzes2010,Hatzes2011,Pepe2013}.
Gaussian Processes (GPs) have become a widely used tool to model activity induced RVs given their ability to describe stochastic variations \citep[e.g.,][]{radvel,Grunblatt2015,Haywood2014}.
However, the flexibility that GPs offer may also be their major drawback if not used carefully.
\citet[][]{Rajpaul2015} proposed a method of using spectroscopic activity-indicators together with RVs in order to constrain the activity induced signal in the RV time-series. This can be done by extending the GP approach to a multidimensional GP that exploits the correlations between the different time-series.
This method has proven useful in disentangling planetary and stellar induced signals in RV data with different levels of activity \citep[e.g.,][]{Barragan2019,Mayo2019}.
In this work we present a new version of the multi-planet modelling code \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\footnote{From the Italian word \emph{pianeti}, which means \emph{planets}.}\ \citep[][]{pyaneti}.
This updated version of the code focuses on multidimensional GP regression in order to model planetary and activity induced signals in spectroscopic times-series. This new version also allows one to model multi-band transits and single transit events, and features various performance improvements.
This new version of \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ has already been used in recent exoplanet characterisation works \citep[e.g.,][]{Carleo2020,Eisner2020b,Eisner2020a,Georgieva2021}.
This manuscript is part of a series of papers under the project
\emph{GPRV: Overcoming stellar activity in radial velocity planet searches} funded by the European Research Council (ERC, P.I.~S.~Aigrain).
The paper is organised as follows: for the sake of self-completeness, we provide a short recap on Gaussian processes in Section~\ref{sec:gps}. Section~\ref{sec:multigps} describes the multidimensional GPs, with a special emphasis on connecting the activity indicators and the RV time-series. We describe the new implementation of \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ in Section~\ref{sec:newpyaneti}. Section~\ref{sec:tests} describes the tests used to validate the code and we conclude in Section~\ref{sec:conlusions}.
\section{A brief overview of Gaussian Processes}
\label{sec:gps}
In this manuscript we do not provide a detailed description of Gaussian processes. Instead, we will provide the basics of GPs needed in order to apply them in data analysis, specifically, in the context of RV and light curve modelling. For further details, we advise the reader to consult specialist literature \citep[e.g.,][]{Rasmussen2006,Roberts2013}.
A \emph{stochastic~process} is a system which evolves in continuous space (in our case time) while undergoing fluctuations.
It is possible to describe the system as a finite set of random variables that are related by a given mathematical entity \citep[][]{Coleman1974}. If the mathematical object that describes the relation between the random variables is a multi-variate normal distribution, then the stochastic process is a \emph{Gaussian~process}.
{
Following \citet[][]{Tracey2018},
a GP assumes that the marginal joint distribution of function values at any finite set of input locations, $\bm{t}={t_{i,(i=1,\cdots,N)}}$, is given by a multi-variate Gaussian distribution }
\begin{equation}
P(\bm{t}) = \frac{1}{\sqrt{(2\pi)^N |\bm{K}|}}
\exp \left[
-\frac{1}{2} \left(\bm{t} - \bm{\mu}\right)^{\rm T} \bm{K}^{-1} (\bm{t} - \bm{\mu})
\right],
\label{eq:multivariate}
\end{equation}
\noindent
where $\bm{\mu}$ is a vector containing mean values, and $\bm{K}$ is a matrix containing the information about the correlation between the variables.
The only condition about the matrix $\bm{K}$ is that it has to be symmetric and positive semi-definite.
We note that for a given $\bm{\mu}$ and $\bm{K}$ we can draw an infinite number of curves as random samples of eq.~\eqref{eq:multivariate}. As these curves are not characterised by explicit sets of parameters, GPs are referred as non-parametric functions \citep[see e.g.,][ for more details]{Roberts2013}.
The important aspect is then to find a way to compute our $\bm{\mu}$ and $\bm{K}$ entities that describe a particular GP.
One advantage of GPs being defined over a continuous space is that the mean vector, $\bm{\mu}$, and covariance matrix, $\bm{K}$, can be computed from evaluations of continuous parametric functions at the positions $\bm{t}$.
Equation~\eqref{eq:multivariate} depends only on $\bm{\mu}$ and $\bm{K}$; therefore, a GP can be fully described with a mean and a covariance kernel function \citep[see][for more details]{Rasmussen2006}.
\subsection{Mean and covariance functions}
\label{sec:kernels}
Mean functions are the deterministic part of a GP. It can be any function $\mu(t;\bm{\phi})$ that depends on a set of parameters, $\bm{\phi}$, and the variable describing the continuous space, $t$. For example, a mean function can be a straight line, a sinusoid, a Keplerian, or a transit model.
A covariance (also called kernel) function $\gamma(t_i,t_j;\bm{\Phi})$ describes how two locations, $t_i$ and $t_j$, are related according to some parameters $\bm{\Phi}$.
Such kernel functions can be tuned in order to describe physical/instrumental signals, such as noise, periodicity, long-term evolution, etc.
We describe below some examples of covariance kernel functions widely used in astronomical literature.
One of the simplest covariance matrix is computed with the white noise kernel
\begin{equation}
\gamma_{\rm WN}(t_i,t_j) =
\sigma_i^2
\delta_{ij},
\end{equation}
\noindent
where $\sigma_i$ is the error associated with the datum $i$ and $\delta_{ij}$ is the Kronecker delta. This kernel creates a diagonal covariance matrix, and is used to take into account uncertainties in data.
Another widely used kernel is the squared exponential
\begin{equation}
\gamma_{\rm SE}({t_i,t_j}) = A^2 \exp \left(
- \frac{ \lvert t_i - t_j \rvert^2 }{2 \lambda^2}
\right)
,
\label{eq:se}
\end{equation}
\noindent
where $A$ is an amplitude that works as a scale factor that determines the typical deviation from the mean function, and $\lambda$ is the length scale, which can be interpreted as the characteristic distance for which two points are strongly correlated.
This kernel generates smooth functions with a typical length scale $\lambda$. Figure~\ref{fig:gp_examples} shows some examples of functions drawn using the $\gamma_{\rm SE}$ kernel and different mean functions.
Also widely used are the Mat\'ern family of kernels. They are based on the standard Gamma function and the modified Bessel function of second order \citep[see e.g.,][for more details]{Rasmussen2006}.
Two examples of the Mat\'ern kernels are the Mat\'ern 3/2 Kernel
\begin{equation}
\gamma_{\rm M32}({t_i,t_j}) = A^2
\left(
1 + t_{3/2}
\right)
\exp \left( - t_{3/2} \right)
,
\label{eq:m32}
\end{equation}
\noindent
with $t_{3/2} \equiv \sqrt{3} \lvert t_i - t_j \rvert \lambda^{-1}$, and the Mat\'ern 5/2 Kernel
\begin{equation}
\gamma_{\rm M52}({t_i,t_j}) = A^2
\left(
1 + t_{5/2}
+ \frac{ t_{5/2}^2}{3}
\right)
\exp \left( - t_{5/2} \right)
,
\end{equation}
\noindent
with $t_{5/2} \equiv \sqrt{5} \lvert t_i - t_j \rvert \lambda^{-1}$. The parameters $A$ and $\lambda$ have the same role as for the Squared Exponential kernel, but in these cases the resulting functions are less smooth.
Figure~\ref{fig:gp_examples} shows GP samples drawn using the $\gamma_{\rm M32}$ kernel.
A widely used kernel in astronomy, especially in exoplanet research, is the Quasi-Periodic (QP) kernel \citep[as defined by][]{Roberts2013}
\begin{equation}
\gamma_{\rm QP}(t_i,t_j) = A^2 \exp
\left\{
- \frac{\sin^2\left[\pi \left(t_i - t_j \right)/P_{\rm GP}\right]}{2 \lambda_{\rm p}^2}
- \frac{\left(t_i - t_j\right)^2}{2\lambda_{\rm e}^2}
\right\}
\label{eq:qp}
,
\end{equation}
\noindent
where $A$ has the same meaning as for the Squared Exponential kernel, $P_{\rm GP}$ is the characteristic period of the GP, $\lambda_{\rm p}$\ the inverse of the harmonic complexity (how complex variations are inside each period), and $\lambda_{\rm e}$\ is the long term evolution timescale (similar to the $\lambda$ for the squared exponential kernel).
We show some examples of functions created using the $\gamma_{\rm QP}$ kernel in Fig.~\ref{fig:gp_examples}.
Because the QP kernel generates stochastic periodic signal, this choice of covariance function is widely used to model stellar activity signals in both photometry and RVs \citep{Haywood2014,Rajpaul2015}.
In a general context the GP period, $P_{\rm GP}$, can be interpreted as the stellar rotation period, the long term evolution time-scale, $\lambda_{\rm e}$, can be associated with the active region lifetime on the stellar surface; and the inverse harmonic complexity, $\lambda_{\rm p}$, can be associated with the activity regions distribution on the stellar surface \citep[see e.g.,][]{Aigrain2015}.
We note that for the cases in which the GP has a relatively small evolution time scale ($\lambda_{\rm e}$\,$\lesssim P_{\rm GP}$), the periodicity of the GP is practically irrelevant \citep[as pointed out by][]{Rajpaul2015}. Therefore, special care has to be taken when dealing with signals in which the evolution time scale is smaller than the expected periodicity. It is better to use a QP kernel only in cases when $P_{\rm GP}<\,$$\lambda_{\rm e}$. If that is not case, the QP kernel might not be appropriate and some other kernels should be considered. In this work we ensure that when we use the QP kernel, $\lambda_{\rm e}$\,$> P_{\rm GP}$ is satisfied.
Figure \ref{fig:gp_examples} also shows an example of the non-parametric behaviour of the GPs \citep[see][for more details]{Rasmussen2006,Roberts2013}. While in a parametric deterministic model a given set of parameters will give always the same curve, in the non-parametric case, a given set of parameters can give different curves with the condition that the random variables satisfy their intrinsic correlation. Given that the parameters of GPs do not have the same interpretation as for parametric functions, they are often called \emph{hyper-parameters}.
\begin{figure*}
\centering
\includegraphics[width=0.99\textwidth]{images/gp_examples.pdf}
\caption{Example of functions generated by GPs with different mean and covariance functions. Left, middle, and right panels corresponds to GPs with mean functions $\mu=0$, $\mu = 1/2\, t$, and $\mu=\sin(t)$, respectively. Top panel shows a plot with the respective mean function.
From top to bottom, the second, third, and fourth panels show five GPs samples from square exponential (with $A=1, \lambda=1$), Mat\'ern 3/2 (with $A=1, \lambda=1$), and Quasi-Periodic kernels (with $A=1, \lambda_e=5$,$\lambda_p=0.5$,$P_{\rm GP}=2$), respectively. }
\label{fig:gp_examples}
\end{figure*}
\subsection{Gaussian process Regression}
\label{sec:gpr}
We can perform regression using GPs if we assume that our data (a finite set of variables $\bm{y}$, taken at times $\bm{t}$) are samples of a GP.
The mean function $\mu$ can be a physically motivated parametric model (e.g., Keplerian or transit curves), and the covariance kernel function, $\gamma(t_i,t_j)$, can encompass any intrinsic correlation in our data set (e.g., stellar activity and/or instrumental systematics).
Given that a finite set of variables drawn from a GP is described by a multi-variate normal distribution, we can use this property to write a logarithmic Gaussian likelihood to marginalise over variables as
\begin{equation}
\ln \mathcal{L}(\bm{\phi},\bm{\Phi}) =
- \frac{1}{2} \left(
N_{\rm obs} \ln 2\pi + \ln \lvert \bm{K} \rvert + \bm{r}^{\intercal} \bm{K}^{-1} \bm{r}
\right),
\label{eq:loglikelihood}
\end{equation}
\noindent
where $\bm{\phi}$ and $ \bm{\Phi}$ are the mean and covariance functions parameters, respectively; $\bm{r} = \bm{y - \mu}$ is the vector of residuals of the data $\bm{y}$ and the mean function $\bm{\mu}=\mu({\bm{t})}$ evaluated at the times $\bm{t}$, $\bm{K}$ is the covariance matrix for the observations $\bm{t}$ given a kernel function $\gamma(t_i,t_j)$, and $N_{\rm obs}$ is the number of observations.
Equation~\eqref{eq:loglikelihood} can be optimised or sampled with different numerical techniques in order to infer the parameters of the mean and covariance functions.
GP regression is the cornerstone of the updated version of \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}. Mean functions can be constructed easily with Keplerian and transit models, while any other intrinsic correlation can be absorbed by the covariance function. If we use a physically motivated kernel function, we can also learn about the underlying mechanism that gives rise to the correlation (e.g., stellar activity).
Section~\ref{sec:newpyaneti} describes how GP regression is included inside \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}.
\section{multidimensional Gaussian Processes}
\label{sec:multigps}
We have described how we can use a GP to describe non-parametric functions over a continuous space on $\mathbb{R}$ (in our case time).
It is possible to extend this idea and to assume that the same GP can describe correlations in multiple continuous spaces that may be related to each other.
multidimensional (also called multi-variate) GPs provide a solid and unified framework to make the prediction of GPs on $\mathbb{R}^{N}$, where $N$ is the number of dimensions \citep[for more details about the mathematical formalism of multidimensional GPs see e.g.,][]{Alvarez2011,Chen2020}.
An useful application of multidimensional GP is regression.
The general idea of multidimensional GP regression is to transform the multidimensional problem into a ``big'' one-dimensional GP regression.
This is done by vectorising the set of observations $[\bm{t},\bm{y}]_i$ (where $i = 1,\cdots, N$) as a big vector in $\mathbb{R}$. The residual vector $\bm{r}$ is created as a concatenation of residuals vectors $\bm{r}_i$, for each dimension $i$.
The covariance matrix $\bm{K}$ is constructed of small sub-matrices $\bm{k}^{l,m}$ that describe the correlations between the different dimensions $l$ and $m$.
This allows one to reformulate the multidimensional GP regression as a conventional GP regression (see Sect.~\ref{sec:gpr}).
In the remainder of this section we will describe how we can use multidimensional GPs to model activity induced signals in spectroscopic time-series.
\subsection{multidimensional GPs for spectroscopic time-series}
\label{sec:mgpforrv}
\citet[][]{Rajpaul2015} proposed a framework to model stellar activity in RV time-series simultaneously with the activity indicators using a multidimensional GP approach.
This approach assumes that the stellar induced signals in all observables can be described by the same latent GP and its time derivative.
\citet[][]{Jones2017} and \citet[][]{Gilbertson2020} expanded the work of \citet[][]{Rajpaul2015} by adding higher derivatives and generalising the work to a generic set of activity indicators. In this work we also generalise the work of \citet[][]{Rajpaul2015} to a generic set of activity indicators, but we maintain the approach of using only the first GP derivative.
\subsubsection{Physical motivation}
It has been shown that there is an intrinsic relation between the area covered by active regions on the stellar surface and the stellar-induced RV variations \citep[e.g.,][]{Aigrain2012,Boisse2009}. Following this idea, we can assume that we can describe a function, $G(t)$, which is a latent unobserved variable that represents the projected area of the visible stellar disc that is covered by active regions as function of time.
Such active regions affect photometric and spectroscopic observed parameters (e.g., RVs or activity indicators) in different ways.
Some of them are only affected by the projected area that is covered by active regions, i.e. they can be described as $G(t)$ with some scale factor; others are also affected by how these regions evolve in time on the stellar surface, i.e., they are described by $G(t)$ and its time derivatives ($\dot{G}(t)$, $\ddot{G}(t)$, etc.).
RVs are affected by the position of the active regions on the stellar surface and how these regions evolve in time \citep[see e.g.,][]{Dumusque2014}. Therefore, in our assumption that $G(t)$ describes the area covered by active regions of the visible stellar disc, activity-induced RV data can be described by $G(t)$ (e.g., to account for the convective blue-shift) and its time derivatives (to account for the evolution of the spots on the stellar surface).
Some activity indicators, such as $\log R'_{\rm HK}$ (flux of the \ion{Calcium}{II} H~\&~K lines relative to the bolometric flux) or $S_{\rm HK}$ (flux of the \ion{Calcium}{II} H~\&~K lines relative to the local continuum), are only affected by the fraction of the area that is covered by the stellar surface \citep[see e.g.,][]{Isaacson2010,Thompson2017}, i.e., they can be described only by $G(t)$.
Some other activity indicators, such as the bisector inverse slope (BIS), are also affected by how the active regions evolve on the stellar surface \citep[see e.g.,][]{Dumusque2014}, requiring higher time derivatives of $G(t)$ in order to describe them.
A set of contemporaneous time-series that contain activity-induced signals of a given star (RVs, $\log R'_{\rm HK}$, etc.) can then be modelled simultaneously assuming they are described by the same underlying function $G(t)$ and its derivatives.
We can assume that our function $G(t)$ is generated by a GP, given its flexibility to model stochastic signals and the property that any affine operator (including linear and/or derivative) applied to a GP yields another GP \citep[see][and references therein]{Rajpaul2015}.
We also note that, if we assume that our underlying function comes from a GP, we can use a multidimensional GP approach to exploit the correlation between observations in the different time-series, assuming each one is a different continuous space described by the same underlying GP-drawn function.
The advantage of this approach is that RV time-series contain stellar activity and planet-induced variations, while activity indicators are only sensitive to activity. Therefore, activity indicators can help to constrain the activity induced signal in RV time-series, thus allowing one to disentangle the planetary signals from the activity signals \citep[][]{Rajpaul2015}.
\subsubsection{Theoretical approach}
\label{sec:mgpr}
We follow the approach described by \citet[][]{Rajpaul2015} and we will assume that we have a set of $N$ time-series, $\mathcal{A}_{i=1,...,N}$, each one with $M$ points.
Although this is not necessary for the multidimensional GP approach, we make this assumption given that for spectroscopic time-series the RVs and activity indicators are computed from the same spectra.
A set of $N$ time-series that are characterised by the same GP-drawn function $G(t)$ and its derivative $\dot{G}(t)$ can be described as
\begin{equation}
\begin{matrix}
\mathcal{A}_1 = A_{1} G(t) + B_{1} \dot{G}(t) \\
\vdots \\
\mathcal{A}_N = A_{N} G(t) + B_{N} \dot{G}(t), \\
\end{matrix}
\label{eq:gps}
\end{equation}
\noindent
where the variables $A_{1}$, $B_{1}$, $\cdots$, $A_{N}$, $B_{N}$, are free parameters which relate the individual time-series to $G(t)$ and $\dot{G}(t)$.
We note that each $\mathcal{A}_i$ is a GP by itself, given the property that any derivative operator applied to a GP and the sum of two GPs is also a GP.
We can assume that the GP has zero mean because in the GP regression (see Sect.~\ref{sec:gpr}) we use the residuals vector to evaluate the likelihood, i.e., we remove the mean function from the observations.
To create the covariance matrix, we need to define how points are correlated between all time-series $\mathcal{A}_i$ and $\mathcal{A}_j$.
Following \citet[][]{Rajpaul2015}, the covariance between two observations at times $t_i$ and $t_j$ between the time-series $\mathcal{A}_l$ and $\mathcal{A}_m$ is given by
\begin{equation}
\begin{aligned}
k^{l,m}({i,j}) = & \, A_l A_m \gamma^{G,G}({i,j}) + B_l B_m \gamma^{dG,dG}({i,j}) \\
& + A_l B_m \gamma^{G,dG}({i,j}) + A_m B_l \gamma^{dG,G}({i,j}),
\label{eq:smallks}
\end{aligned}
\end{equation}
\noindent
where $\gamma^{G,G}({i,j})$ denotes the covariance between (non-derivative) observations of $G$ at times $t_i$ and $t_j$; $\gamma^{G,dG}({i,j})$ refers to the covariance between an observation of $G$ at time $t_i$ and an observation of $\dot{G}$ at time $t_j$; $\gamma^{dG,G}({i,j})$ refers to the covariance between an observation of $\dot{G}$ at time $t_i$ and an observation of ${G}$ at time $t_j$; and $\gamma^{dG,dG}({i,j)}$ denotes the covariance between two observations of $\dot{G}$ at times $t_i$ and $t_j$.
In appendix~\ref{ap:derivatives} we show the gamma terms ($\gamma^{G,G}$, $\gamma^{dG,G}$, $\gamma^{G,dG}$, and $\gamma^{dG,dG}$) for the squared exponential, Mat\'ern 5/2, and QP kernels.
The ``big'' covariance matrix $\bm{K}_{\rm big}$ that describes the covariance between all the $N$ time-series is
\begin{equation}
\bm{K}_{\rm big} =
\begin{pmatrix}
\bm{k}^{1,1} & \bm{k}^{1,2} & \cdots & \bm{k}^{1,N} \\
\bm{k}^{2,1} & \bm{k}^{2,2} & \cdots & \bm{k}^{2,N} \\
\vdots & \vdots & \ddots & \vdots \\
\bm{k}^{N,1} & \bm{k}^{N,2} & \cdots & \bm{k}^{N,N} \\
\end{pmatrix}
,
\label{eq:bigk}
\end{equation}
\noindent
where each $\bm{k}^{l,m}$ is computed using eq.~\eqref{eq:smallks} for any covariance function $\gamma$.
If $\gamma$ is a valid kernel function, the matrix $\bm{K}$ is a valid covariance matrix that we can use in GP regression. This matrix also has the property of being symmetric and positive definite, therefore, we need to compute only the upper triangle part of the matrix, while the lower panel matrices can be computed as $\bm{k}^{j,i} = (\bm{k}^{i,j})^{^{\intercal}}$.
At this point we have all the mathematical framework needed to perform multidimensional GP regression for RVs and activity indicators using the GP regression described in Sect.~\ref{sec:gpr}. We can create a residual vector $\bm{r}_i$ for each dimension $i$, e.g., the residuals corresponding to the RVs can be computed by subtracting Keplerian models, while the residuals of an activity indicator can be computed by subtracting constant offsets.
We then create the residual vector $\bm{r}$ by concatenating the $\bm{r}_i$ for each dimension.
We can create the ``big'' covariance matrix $\bm{K}_{\rm big}$ using eq.~\eqref{eq:bigk} with any valid kernel (e.g. the ones in Sect.~\ref{sec:kernels}). Once we have $\bm{r}$ and $\bm{K}_{\rm big}$ we can use the likelihood given by eq.~\eqref{eq:loglikelihood} to infer the parameters of our multidimensional GP using our preferred numerical method. Section~\ref{sec:newpyaneti} describes how this framework is included inside \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}.
\subsection{Comparison between multidimensional GP and other approaches}
\label{sec:gpcomparisons}
{
A common approach in the literature to describe stellar signals in RV time-series consists of modelling ancillary time-series (these can be light curves or activity indicators) with a GP to infer the hyper-parameters for a given kernel. The hyper-parameters inferred from one or more ancillary time series are then used to inform hyper-parameters priors for the GP modelling of the RV data using the same kernel \citep[see e.g.,][]{Grunblatt2015,Haywood2014}.
The process of retrieving kernel hyper-parameters for a GP modelling is known as \emph{training} a GP, and we therefore refer to this approach as the "training-GP" approach. Training the GP in this way ensures that the stellar activity signal in the RVs is modelled with the same characteristic features --such as periodicity, degree of smoothness, characteristic evolution timescale-- as the ancillary time-series. In this approach, the functions describing the ancillary time-series and the RVs share the similar covariance properties, but they are otherwise independent of each other, and their shapes are entirely unrelated.
A slight variation on the training-GP approach involves modelling the ancillary time-series and the RVs simultaneously, using independent GPs sharing the same covariance function \citep[see e.g.,][]{Osborn2021,Suarez2020}. In this case, both the RVs and the ancillary time-series are used to constrain the GP hyper-parameters, and the functions describing them share exactly the same covariance properties, but they are still independent of each other and have unrelated shapes.
These approaches make weaker assumptions about the relationship between the ancillary time-series and the RVs than those made in the multidimensional model developed by \citet{Rajpaul2015} and implemented in \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}. They result in more flexible models for the RV time-series, with the associated risk of over-fitting (where potential planetary signals can be absorbed or modified by the activity model). On the other hand, our assumptions are based on a fairly simplistic toy model, whose limitations will no doubt become apparent once the framework is applied to a large enough sample of high-precision datasets. Such failures should however be easy to diagnose, as they would lead to a poor fit to the data.
}
In Section~\ref{sec:toy2} we present a comparison of planetary signal recovering between the ``training-GP'' method described in this section and the multi-GP approach.
It is also important to note that, in our multidimensional GP model, the functions used to describe the activity signal in the RVs have different covariance properties from those used to model the activity indicators. The fact that the RVs and activity indicators are modelled as different linear combinations of the underlying GP and its time-derivative results in markedly different harmonic complexity, for example (see Sect.~\ref{sec:gpderivatives} for a more detailed discussion).
While we are not aware of examples in the literature, the "training-GP" approach could be generalised to take into account the derivatives of a GP to describe time-series. In Sect.~\ref{sec:run2} we show an example on how to train a GP to use the hyper-parameters to model RVs taking into account the derivative of the chosen kernel.
A more complete quantitative comparison between different methods to model stellar signals is given by \citet{Ahrer2021}.
\subsection{On the usefulness of the GP derivatives}
\label{sec:gpderivatives}
In this section we describe the importance of taking into account the derivatives of the GP to model RV data when assuming that our GP generates a function that describes the surface covered by active regions on the stellar surface.
We base our discussion on the QP kernel, but the conclusions can be extended to other kernels.
Let us suppose we have a 2-dimensional GP to describe two time-series, $S_1$ and $S_2$, that behave as
\begin{equation}
\begin{aligned}
S_1 = & G(t), \\
S_2 = & \dot{G}(t). \\
\end{aligned}
\label{eq:sgps}
\end{equation}
Equation~\eqref{eq:sgps} was computed from equation~\eqref{eq:gps} with $A_1=B_2=1$, and $A_2=B_1 = 0$.
Figure~\ref{fig:example_multigp} shows some samples of $S_1$ and $S_2$ time-series that were created using a QP kernel with $P_{\rm GP} = 1$, $\lambda_{\rm e}$\,$= 10$, and different values of $\lambda_{\rm p}$.
From the examples in the top panel of Fig.~\ref{fig:example_multigp} we can see that if the $S_1$ signal has a high harmonic complexity, then the contemporaneous $S_2$ would have an apparently higher harmonic complexity.
We can see that this behaviour is expected from the derivatives of the QP kernel (see Appendix~\ref{ap:derivatives}). From equations \eqref{eq:qpgdg} and \eqref{eq:qpdgdg} we can see that when $\lambda_{\rm p}$\,$\lesssim 1$ there are some ``$\tau$ terms'' (terms that include the $\tau$ parameter) that add extra ``wiggles'' to the behaviour of the $S_2$ curve.
This has a direct implication when training GPs with ancillary observations (that may behave as $S_1$) to model RVs (that may behave as $S_2$). Setting a prior on $\lambda_{\rm p}$\ for the $S_2$ signal based on our $S_1$ signal may lead to biased results.
The previous discussion has special implications when training GPs to model RVs using light curves. In general, light curves and RV data are not taken simultaneously, so the active regions on the stellar disc might not be the same between the two data sets \citep[see e.g.,][]{Aigrain2012,Barragan2021}. But even if they were, the harmonic complexity extracted from a light curve may not be the same as for the activity induced RV signal if modelled only with $G(t)$, i.e., without accounting for the GP derivatives.
There are examples of the importance of using the GP derivative when modelling RVs with high harmonic complexity \citep[e.g.,][]{Barragan2019,Barragan2021b}.
In the other extreme of low harmonic complexity ($\lambda_{\rm p}$$\gg 1$) we see that $S_1$ and $S_2$ behave as quasi-sinusoidal signals (lower panel of Fig.~\ref{fig:example_multigp}). If we consult equations \eqref{eq:qpgdg} and \eqref{eq:qpdgdg}, we can see that in the limit when $\lambda_{\rm p}$\,$\gg 1$, all $\tau$ terms are irrelevant and the behaviour of the GP and derivatives is quasi-sinusoidal with a long-term evolution regulated by $\lambda_{\rm e}$.
In this case using the derivative of the GP to model the RVs may not be crucial for modelling the activity induced signal. There are some examples in the literature of activity induced RV signals that behave similar to the activity indicators in the low harmonic complexity regime (e.g., Serrano et al., submitted.).
We also note that the fact that RVs depends on how active regions evolve in the stellar surface may cause an apparent asynchrony with the activity indicators.
This phenomenon of RV and activity indicators being out of phase has been reported in the literature \citep[e.g,][]{CollierCameron2019}.
If the stellar signal induced in the RVs depends on a combination of the kind $A G(t) + B \dot{G}$, this can generate curves that may seem similar to the one of the activity indicator, that may vary as $A G(t)$, but with an apparent phase shift.
This may be most applicable to the low harmonic complexity case in which signals tend to look more sinusoidal.
For example, in \cite{Georgieva2021} the stellar signal in the RV and activity indicators time-series are similar, but the RV curve seems to have a different phase (seemingly ahead) of the contemporary activity indicator time-series. In the multidimensional GP approach the time derivatives of $G(t)$ allow this behaviour to be accounted for.
We have discussed the importance of the GP derivatives based on the QP kernel.
However, any other kernel that has strong variations on short time scales may suffer from the same effect in which the derivatives of the GP are relevant to model RV time-series.
We note that this topic needs to be explored further (e.g., Nicholson et al., in prep.), but a detailed description of this issue is beyond the scope of this paper.
We mention it however, to apprise the reader of the usefulness of including the derivatives of the GP when modelling RV time-series with \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}.
\begin{figure*}
\centering
\includegraphics[width=0.99\textwidth]{images/multi_gp_examples1.pdf}\\
\includegraphics[width=0.99\textwidth]{images/multi_gp_examples3.pdf}\\
\caption{Example of samples from a multidimensional GP (two dimensions in this case) created with a QP kernel with $P_{\rm GP} = 1$, $\lambda_{\rm e}$$=10$ and high ($\lambda_{\rm p}$$=0.1$, top pannel) and low $\lambda_{\rm p}$$=10$, lower pannel) levels of harmonic complexity. Each panel shows two subplots with samples of $S_1$ and $S_2$ time-series (See eq.~\ref{eq:sgps}). Corresponding $S_1$ and $S_2$ draws from the same sample are shown with the same colour.}
\label{fig:example_multigp}
\end{figure*}
\subsection{Limitations of the multidimensional GP approach}
We warn the reader that the multidimensional GP approach should not be taken as a magic recipe that will always solve our stellar induced RV variation problems.
The first limitation is that this framework assumes a relation between all time-series via a function $G(t)$ and its time derivatives. This may work in many cases, but we warn that this is still only a first order approach to the problem. Stellar activity signals may be more complicated than the underlying model assumed for this framework. Therefore the multidimensional GP approach may lead to activity signals being fit imperfectly.
Problems may also arise if the data sampling is not optimal.
The main idea behind GP regression is to exploit the correlation between our observations.
This implies that if observations have time separations larger than the GP time-scales that we want to characterise, then the GP model may be poorly constrained. Therefore, special care has to be taken when using GPs to modelling data with sparse sampling.
In Sect.~\ref{sec:citlalatonac} we show and describe \texttt{citlalatonac}, a code that simulates spectroscopic-like time-series that can be helpful to plan RV observation campaigns for active stars.
It is also worth mentioning that a downside of GP regression is the computational cost of matrix inversion.
This is particularly relevant to the multidimensional GP approach where the dimension of the matrix to invert increases with the number of time-series to model. However, in the case of spectroscopic time-series, there are rarely more than several hundred observations per star. Therefore, the multidimensional GP regression to model RVs and activity indicators remains computationally treatable.
\section{The new \texttt{PYANETI}}
\label{sec:newpyaneti}
\citet[][]{pyaneti} describes the data analysis approach taken by \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}. Briefly, \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ uses a implementation of a Markov chain Monte Carlo (MCMC) sampler based on \texttt{emcee} \citep[][]{emcee} to create marginalised posterior distributions for exoplanet RV and transit parameters. The code uses a built-in Gaussian likelihood together with user-input priors for each sampled parameter.
The demanding computational routines are written in \texttt{FORTRAN} and the subroutines are wrapped into python using the \texttt{f2py} tool included in the \texttt{numpy} \citep[][]{numpy} package.
The new version of \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ is an extension of the original package presented in \citet[][]{pyaneti}.
The biggest update is the generalisation of the Gaussian likelihood that includes correlation using a GPR following eq.~\eqref{eq:loglikelihood}.
The code can also perform multi-band transit and single transit modelling.
The rest of this section describes in more detail the new additions to the code.
\subsection{Gaussian Processes}
\citet[][]{pyaneti} describes the functions used to model multi-planet signals in both RV and transit data sets. These equations are used as mean functions of the GP to compute the residual vector $\bm{r}$, allowing \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ to perform multi-planet fits with RV and/or transit data together with GP regression using eq.~\eqref{eq:loglikelihood} and user input priors.
The elements of the covariance matrix inside \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ are created as
\begin{equation}
K(t_i,t_j) = \gamma \left( t_i, t_j \right) +
\left(\sigma_i^2 + \sigma_{{\rm jitter}}^2\right)
\delta_{ij},
\label{eq:covarianceposterior}
\end{equation}
\noindent
where $\gamma(t_i,t_j)$ is any valid kernel function, $\delta_{ij}$ is the Kronecker delta, $\sigma_i$ the white noise associated with the datum $i$, and $\sigma_{{\rm jitter}}$ is a jitter term.
We note that the same jitter term can be shared by a collection of points with the same underlying systematics, e.g., a jitter term associated to a common instrument.
We note that in the case where no correlation is assumed in the data, i.e. $\gamma(t_i,t_j) = 0$, eq.~\eqref{eq:loglikelihood} reduces to the white noise Gaussian likelihood implemented in the previous version of \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ \citep{pyaneti}.
We have implemented in \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ all kernels described in Sect.~\ref{sec:kernels}, but more can be added easily if needed.
We also incorporated the multidimensional GP approach described in Sect.~\ref{sec:multigps} into \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ within the RV modelling routines.
The new version of \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ can reproduce the original approach given by \citet[][]{Rajpaul2015}, but also allows one to combine arbitrarily many time-series to use multiple activity indicators.
The residual vector for the RV data is computed subtracting Keplerian signals, and for the activity indicators with a constant offset. We note that \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ can deal with different instrumental offsets for the RV and activity indicators.
To create the ``big'' covariance matrix, eq.~\eqref{eq:bigk}, we define the sub-matrices $k^{l,m}_{\rm wn}({i,j})$ to account for white noise as
\begin{equation}
k^{l,m}_{\rm wn} (i,j) = k^{l,m}(i,j) + \left(\sigma_{i,l}^2 + \sigma_{{\rm jitter},l}^2\right)
\delta_{ij} \delta_{lm},
\label{eq:smallkswn}
\end{equation}
\noindent
where $k^{l,m}(i,j)$ is computed using eq.~\eqref{eq:smallks} and a valid covariance function, $\sigma_{i,l}$ and $\sigma_{{\rm jitter},l}$ are the white noise and jitter term associated to the dimension $l$; and $\delta_{ij}$ and $\delta_{lm}$ are Kronecker deltas. The $\delta_{lm}$ between two different dimensions, $l$ and $m$, ensures that the nominal errors are only added in the diagonal of the ``big'' matrix.
\href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ creates the $\bm{K}_{\rm big}$ covariance matrix using the sub-matrices computed with eq.~\eqref{eq:smallkswn}. The code has built-in routines to perform multidimensional GP regression using the squared exponential (eq.~\ref{eq:se}), Mat\'ern 5/2 (eq.~{\ref{eq:m32}}), and QP (eq.~\ref{eq:qp}) Kernels. Appendix~\ref{ap:derivatives} show the derivatives included into \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ to compute eq.~\eqref{eq:smallks} for these covariance functions.
\subsubsection{Dimensionality problem}
We note that \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ suffers from the high computational cost of GP regression, which in general entails matrix inversion.
The computational time of a matrix inversion scales as $\mathcal{O}(N^3)$, where $N$ is the number of data points.
This is generally not a problem for RV data sets, which include relative few observations (usually less than 1000). However, it becomes a problem when modelling light curve time-series with thousands of observations.
Fortunately, in the case of light curves, the time-scales of the planetary transits are small compared with those associated with stellar variability.
This makes it relatively easy to remove such trends from light curves. This \emph{detrending} is a common approach in the literature when one is only interested on modelling transit signals in flattened light curves \citep[see e.g.,][]{Hippke2019}.
A reason for fitting GPs to full light curves with transits would be to argue that the light curve might provide information on GP hyper-parameters that can also be used simultaneously with RV data.
Yet this may not be optimal because usually light curves and RVs are not observed simultaneously, and there is evidence that light curves and RVs may not be constrained by the same time scales, as discussed in Sect.~\ref{sec:gpderivatives}.
Since the GP implementation of \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ is mainly focused on the RV analysis, we did not explore GP computation acceleration for light curve analyses.
We note that \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ could still be used to model binned light curves in order to try to estimate hyper-parameters of the light curve, but a GP modelling of the light curve including transits is not advisable with \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}.
{
We note that progress has been made in fast matrix inversion for GP regression. We considered implementing the GP regression operations included in \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ using the \texttt{george} package \citep{george}; this would require explicitly coding the derivative kernels we use to model each time-series in \texttt{george}, a feasible but non-trivial endeavour. It has not been necessary to do it so far, given the typical size of the datasets we are modelling, but it is something to be considered for a future implementation of \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}. Another, even faster option for GP regression on large datasets is the \texttt{celerite} package \citep{celerite}, but that is not suitable for the present work as it is restricted to 1-dimensional datasets.}
\subsection{Multi-band fit}
Multi-band photometric follow-up of transiting planets has become common. This is because the number of ground and space-based instruments has increased, and more transiting planets are found around relatively bright stars.
For this reason, we have added multi-band transit modelling into \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}. The code solves for the same orbital parameters for all bands, but independently samples the wavelength dependent parameters, i.e., limb darkening coefficients, cadence and integration time, and scaled planet radius.
As in the previous version, the code uses the \citet[][]{Mandel2002} equations to model transits by assuming the star limb darkening can be modelled as a quadratic law.
The code samples for two limb darkening coefficients for each band following the $q_1$ and $q_2$ parametrization described in \citet[][]{Kipping2013}.
The code also allows for a different cadence and integration time for each band \citep[see][]{Kipping2010}.
Finally, the code also allows modelling for the same scaled planet radius $R_{\rm p}/R_\star$ for all bands, as well as an independent $R_{{\rm p},i}/R_\star$ for each band $i$. The latter can be useful to test false-positive scenarios \citep[e.g.,][]{Parviainen2019}, or to fit transit depths at different wavelengths as used in transmission spectroscopy \citep[e.g.,~][]{Charbonneau2002}.
\subsection{Single transit fit}
Single transit events can be caused by transiting planets with periods longer than the observational window. Fortunately, they can be detected by methods that do not rely on periodicity of transit-like events \citep[see e.g.][]{Eisner2020b,Osborn2016}.
The problem when dealing with mono-transits is that the period and the semi-major axis cannot be determined.
These parameters are important because they determine the velocity at which the planet moves during the transit.
In order to solve this, in the new version of \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\, we fix a period to a dummy value larger than our observing window, and we sample for a dummy scaled semi-major axis $a_{\rm dummy}$. While $a_{\rm dummy}$ does not have a physical sense, but it ensures that the transit shape is sampled.
Therefore, in order to model a single transit, \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ samples the time of mid-transit $T_0$, impact parameter $b$, scaled planet radius $R_{\rm p}$, $a_{\rm dummy}$, and limb darkening coefficients $q_1$ and $q_2$ \citep[][]{Kipping2013}.
With these parameters we can estimate the transit duration, and if we assume that the planetary orbit is circular, we can estimate the orbital period albeit with relatively large uncertainty \citep[see][]{Osborn2016}.
\subsection{\texttt{citlalicue} and \texttt{citlalatonac}}
\label{ap:citla}
We have created codes to create synthetic stellar photometric and spectroscopic time-series, called \texttt{citlalicue}\ and \texttt{citlalatonac}, respectively\footnote{In Aztec mythology, Citlalicue (goddess) and Citlalatonac (god) are the creators of the stars. The words root, \emph{Citlali}, is the Nahuatl word for star.}.
\subsubsection{\texttt{citlalicue}: the light curves creator}
\label{sec:citlalicue}
\texttt{citlalicue}\ is a \texttt{Python} module that allows one to create synthetic light curves. It is totally independent of \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}. It can be easily installed using \texttt{pip install citlalicue}.
The module has a class called \texttt{citlali} that contains all the attributes and methods needed to create a synthetic stellar light curves with transits, periodic modulation, and white noise.
The current version of the code uses a QP kernel (eq.~\ref{eq:qp}) to simulate stellar variability, and it allows one to create transits for any number of planets using \texttt{pytransit} \citep{pytransit}.
An example of how to create a light curve using \texttt{citlalicue}\ is given \href{https://github.com/oscaribv/citlalicue/blob/master/example_light_curves.ipynb}{here \faGithub}.
\texttt{citlalicue}\ also has a class called \texttt{detrend} that allows one to detrend light curves using GPs. \texttt{detrend} takes a plain-text file with light curve data containing time and flux (and errors as input if available).
The code allows one to mask out the transits from the data as well as to fit simultaneously for the transits and GP. The former is recommended given that it is faster.
The code uses \texttt{george} \citep[][]{george} to perform a fast GP regression that enables the modelling of the variability in the light curve.
The code allows for an iterative optimisation with a sigma clipping algorithm, where the threshold can be tuned by the user (the default is 5). Once the optimal model is found by the code, the inferred trend is removed from the light curve, creating a flattened signal with transits.
An example of how to detrend a light curve using \texttt{citlalicue}\ is available \href{https://github.com/oscaribv/citlalicue/blob/master/example_detrending.ipynb}{here \faGithub}.
Examples of the detrending capability of \texttt{citlalicue}\ can be found in e.g., \citet[][]{Barragan2021} and \citet{Georgieva2021}.
\subsubsection{\texttt{citlalatonac}: the spectroscopic time-series creator}
\label{sec:citlalatonac}
The \texttt{Python} package \texttt{citlalatonac}\ uses \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ in order to create synthetic spectroscopic (RVs and activity indicators-like) time-series.
This package comes together with \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ when the latter is cloned directly from its GitHub repository.
The code creates samples of a multidimensional GP following eq.~\eqref{eq:gps}. This simulates spectroscopic-like signals (RVs and activity indicators) of an active star, assuming they all are generated by the same underlying GP.
The main class of the package is named \texttt{citlali}. When \texttt{citlali} is called in \texttt{Python}, the user needs to specify the time range in which the synthetic data will be created (e.g., this range can be an observing season), the number of time-series to create, the amplitudes of the signals, following eq.~\eqref{eq:gps}, the kernel to use, and the kernel parameters.
By default the first time-series is called \texttt{rv} and it is always treated as RV-like, i.e., the planet-induced signals are added to this time-series only.
The class includes methods that allow one to add as many planets as needed, as well as white and red noise.
The package also includes the \texttt{create\_real\_times} function that allows one to create realistic sampling of targets at a given observatory using \texttt{astropy} \citep[][]{astropy1,astropy2}.
This utility can be useful for estimating the number of data points needed in order to measure the Doppler semi-amplitude of a given target, even if the star is active. Such numbers can be valuable while writing a telescope proposal.
A practical example of how to use \texttt{citlalatonac}\ to create synthetic time-series of a target observed at a given observatory can be found \href{https://github.com/oscaribv/pyaneti/blob/master/pyaneti_extras/synthetic_k2100.ipynb}{here \faGithub}.
\section{Tests}
\label{sec:tests}
\subsection{Recovering multi-GP hyper-parameters}
\label{sec:toy1}
We created a set of synthetic spectroscopic-like time-series using \texttt{citlalatonac}\ (see Appendix~\ref{ap:citla}) in order to test the ability of \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ to recover parameters using a multidimensional GP.
We assume that we have 3 time-series that are described by the same underlying function, $G(t)$, as
\begin{equation}
\begin{aligned}
S_1 & = & A_1 G(t) & + B_1 \dot{G}(t), \\
S_2 & = & A_2 G(t), \\
S_3 & = & A_3 G(t) & + B_3 \dot{G}(t).
\label{eq:toymodel}
\end{aligned}
\end{equation}
\noindent
We compute eq.~\eqref{eq:toymodel} using eq.~\eqref{eq:gps} with 3 time-series. We set the values for the amplitudes $A_1 = 0.005$\,${\rm km\,s^{-1}}$, $B_1= 0.05$\,${\rm km\,s^{-1}}$\,d, $A_2=0.02$\,${\rm km\,s^{-1}}$, $B_2 = 0$\,${\rm km\,s^{-1}}$\,d, $A_3=0.02$\,${\rm km\,s^{-1}}$, $B_3=-0.05$\,${\rm km\,s^{-1}}$\,d.
We assume a mean function of zero for all three time series, and we use a QP covariance function with hyper-parameters $\lambda_{\rm e}$$ = 30$ d, $\lambda_{\rm p}$$= 0.3$ and $P_{\rm GP} = 5$ d.
We created 50 simultaneous observations taken randomly in a window of 50 d.
We added white noise with standard deviation of $0.001$\,${\rm km\,s^{-1}}$\ for $S_1$, 0.005\,${\rm km\,s^{-1}}$\ for $S_2$ and 0.010\,${\rm km\,s^{-1}}$\ for $S_3$.
The synthetic time-series data are shown in Fig.~\ref{fig:toymodel}. The \texttt{Jupyter} notebook used to create the synthetic data is available \href{https://github.com/oscaribv/pyaneti/blob/master/inpy/example_toyp1/toy_model1.ipynb}{here \faGithub}.
We performed multidimensional GP modelling of the data set using \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}. Priors and parameters used are defined in Table~\ref{tab:parameterstoy}.
We perform an MCMC analysis with 100 Markov chains. We use the last 5000 iterations of converged chains, with a thin factor of 10, to create the posterior distributions.
{
We assume chains have converged when their \citet[][]{Gelman2004} criterion $\hat{R}$ is smaller than $1.02$ for all the sampled parameters \citep[for more details see][]{pyaneti,Gelman2004}.
}
The inferred parameters are shown in Table~\ref{tab:parameterstoy}, and the inferred models, together with the data, are shown in Fig.~\ref{fig:toymodel}.
We have made available the data and input file in \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ for this example;
it can be run as \texttt{./pyaneti.py example\_toyp1} from the main \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ directory.
From Table~\ref{tab:parameterstoy} we can see that the code is able to recover the injected amplitudes and kernel parameters within the error bars. Something to note is that the code is able to recover the value of $B_2 = 0$\,${\rm km\,s^{-1}}$\,d. This is important because it means that the analysis is able to differentiate between the pure $G(t)$ curves from those that depends on $\dot{G}(t)$. This has a practical application to understand the behaviour of the activity indicators that we use in our modelling.
\begin{figure*}
\centering
\includegraphics[width=0.99\textwidth]{images/example_toyp1_rv_timeseries.pdf}
\includegraphics[width=0.99\textwidth]{images/example_toyp1_timeseries2.pdf}
\includegraphics[width=0.99\textwidth]{images/example_toyp1_timeseries3.pdf}
\caption{
$S_1$, $S_2$, and $S_3$ time-series described in Sect.~\ref{sec:toy1}.
The green markers in each panel represent the synthetic observations with inferred offsets extracted.
The solid dark lines shows the inferred mean of the predictive distribution of our multidimensional GP, with dark and light shaded areas showing the one and two sigma credible intervals of the corresponding GP model, respectively. These plots are shown as they are provided by \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}. The code assumes that the first time-series is RV-like, and it may contain Keplerian signals that are represented with a red curve, in this case no Keplerian signals are included and the curve appears as a horizontal red line at 0.
}
\label{fig:toymodel}
\end{figure*}
\input{tables/table1}
\subsection{Recovering Keplerian signals with multi-instrument data}
\label{sec:toy2}
We perform another test similar to the one described in Sect.~\ref{sec:toy1}, but this time we added some more challenges to the test.
In this case we assume that we have contemporaneous observations of two time-series that behave as
\begin{equation}
\begin{aligned}
S_1 & = & A_1 G(t) & + B_1 \dot{G}(t), \\
S_2 & = & A_2 G(t), \\
\label{eq:toymodel2}
\end{aligned}
\end{equation}
\noindent
with $A_1 = 0.005$~${\rm km\,s^{-1}}$, $B_1= 0.05$~${\rm km\,s^{-1}}$\,d, $A_2=0.02$~${\rm km\,s^{-1}}$, and $B_2 = 0$~${\rm km\,s^{-1}}$\,d.
We use a QP covariance function with hyper-parameters $\lambda_{\rm e}$$ = 20$\,d, $\lambda_{\rm p}$$= 0.5$ and $P_{\rm GP}$$ = 5$\,d.
We assume that the spectroscopic data comes from two different instruments, $I_1$ and $I_2$, with an offset of zero for each time-series.
For instrument $I_1$ we created 20 random observations between 0 and 60 days, each datum with an error bar of 0.003~${\rm km\,s^{-1}}$.
For instrument $I_2$ we created 30 random observations in the same range, each one with an error bar of 0.005~${\rm km\,s^{-1}}$.
We included two Keplerian signals in the RV-like time-series ($S_1$). One signal is associated with a circular orbit \citep[$\sqrt{e} \sin \omega = 0$, and $\sqrt{e} \cos \omega = 0$, following][parametrisation]{Anderson2011} and the other one with an eccentricity of $0.3$ and angle of periastron of $\pi/3$ ($\sqrt{e} \sin \omega = 0.47$, and $\sqrt{e} \cos \omega = 0.27$). The amplitude of the Keplerian signals is significantly smaller than the amplitudes of the activity-like signal.
The parameters used to create both signals are listed in Table~\ref{tab:parameterstoy2}.
The \texttt{Jupyter} notebook used to create the synthetic data is available \href{https://github.com/oscaribv/pyaneti/blob/master/inpy/example_toyp2/toy_model2.ipynb}{here \faGithub}. We show the synthetic time-series in Fig.~\ref{fig:toymodel2}.
We perform four different analyses to the data using different techniques. For all the cases we assume that we know the ephemeris of the Keplerian signals and set Gaussian priors on the time of minimum conjunction, $T_0$, and period, $P$, for the two signals.
\subsubsection{Run 1}
\label{sec:run1}
We first perform a 1-dimensional GP modelling with \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ modelling only the RVs with a QP kernel.
We assume that the only information that we have of the GP hyper-parameters are the ranges where the true values lie.
Table~\ref{tab:parameterstoy2} shows the sampled parameters and priors we use for this run that we name as \emph{Run 1}.
We perform an MCMC sampling with 100 Markov chains.
We create the posterior distributions with the last 5000 iterations of converged chains with a thin factor of 10. This generates distributions with 50\,000 independent points per each sampled parameter.
The inferred parameters are shown in Table~\ref{tab:parameterstoy2}.
The first thing we note is that for this case the recovered $\lambda_{\rm e}$\ and $P_{\rm GP}$\ are consistent with the true values, but the value of $\lambda_{\rm p}$\ is smaller than the true vale used to create the time-series. This is expected given that we are modelling the data without taking into account the GP derivative (see discussion in Sect.~\ref{sec:gpderivatives}). However, the parameter true values used to create the Keplerian signals are recovered within the confidence interval.
\subsubsection{Run 2}
\label{sec:run2}
We then perform a \emph{Run 2} in which we train our GP based on our activity-indicator-like signal ($S_2$). As we mentioned in Sect.~\ref{sec:gpcomparisons}, this is a common approach in the literature.
We first do a 1-dimensional GP modelling of the $S_2$ signal using a QP kernel in order to obtain posterior distributions for the hyper-parameters $\lambda_{\rm e}$, $\lambda_{\rm p}$, and $P_{\rm GP}$.
We then model the RV data following the normal approach in the literature \citep[e.g.,][]{Grunblatt2015}, i.e., we use a QP kernel with Gaussian priors on $\lambda_{\rm e}$, $\lambda_{\rm p}$, and $P_{\rm GP}$\ based on our $S_2$ analysis.
The MCMC details are identical to the ones described in Sect.~\ref{sec:run1}.
Table~\ref{tab:parameterstoy2} shows priors and inferred values for all the sampled parameters.
From Table~\ref{tab:parameterstoy2} we can see that the results for this run are similar to the ones from the \emph{Run 1}. We note that despite the Gaussian prior that we set on $\lambda_{\rm p}$\, the recovered value for this parameter is significantly smaller than the Gaussian prior mean.
This is again expected given that we are modelling the RV data only with a QP kernel, without accounting for the time derivative (See Sect.~\ref{sec:gpderivatives}).
Nonetheless, the recovered values of the Doppler semi-amplitudes of the of the coherent signals are recovered with a significance similar to the ones in \emph{Run 1}.
\subsubsection{Run 3}
We explored another possibility on the modelling of the RV time-series training the GP. But this time for the RV modelling we are including the first time derivative of the GP. The GP training comes from the same analysis of the $S_2$ time-series described in Sect.~\ref{sec:run2}.
But, for the RV 1-dimensional GP regression we construct our covariance matrix with the kernel
\begin{equation}
\gamma(t_i,t_j) = A_1^2 \gamma_{\rm QP}^{G,G} + B_1^2 \gamma_{\rm QP}^{dG,dG},
\label{eq:newgammaqp}
\end{equation}
\noindent
that includes the derivative term of the QP kernel. Appendix~\ref{sec:gpderivatives} shows the full form of the $\gamma_{\rm QP}^{G,G}$ and $\gamma_{\rm QP}^{dG,dG}$ terms. We perform a 1-dimensional GP modelling of the RV data following the priors described in Table~\ref{tab:parameterstoy2} as Run 3.
The MCMC configuration follows the same parameters as in Sect.~\ref{sec:run1}.
Table~\ref{tab:parameterstoy2} show the recovered parameters for this run.
We note that for this case, the recovered value for $\lambda_{\rm p}$\ is consistent with the true value, as expected now that we are including the derivative of the GP.
However, we note that even if we include the derivative, the recovered values of the Doppler semi-amplitudes are not more precise that the values recovered in Runs 1 and 2.
This can be explained because even if we are using a better model (that we know agrees with the model used to create the synthetic data), this approach still does not constrain better the shape of the underlying function describing the stellar signal in the RV data.
\subsubsection{Run 4}
We then perform a two-dimensional GP modelling with \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ following the approach described in Sect.~\ref{sec:mgpforrv}. Table~\ref{tab:parameterstoy2} shows the sampled parameters, priors we use, and derived parameters.
again, the MCMC configuration is the same as the one described in Sect.~\ref{sec:run1}. This example can be reproduced by running \texttt{./pyaneti.py example\_toyp2} in the main \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ directory.
Figure~\ref{fig:toymodel2} shows the derived time-series and phase-folded models for this case.
We can see in Table~\ref{tab:parameterstoy2} that all derived parameters agree with the true values within the error bars. Specially the value of $\lambda_{\rm p}$\ agrees with the true value as in the \emph{Run 3}, as expected given that we are using the GP derivative in this case.
We note that in this case, the derived semi-amplitudes for both Keplerian signals are recovered with a relative higher precision than in the previous runs.
\begin{figure*}
\centering
\includegraphics[width=0.99\textwidth]{images/example_toyp2_rv_timeseries.pdf} \\
\includegraphics[width=0.99\textwidth]{images/example_toyp2_timeseries2.pdf} \\
\includegraphics[width=0.495\textwidth]{images/example_toyp2b_rv.pdf}
\includegraphics[width=0.495\textwidth]{images/example_toyp2c_rv.pdf}
\caption{ $S_1$, and $S_2$ time-series described in Sect.~\ref{sec:toy2}.
Top panel: The green (instrument $I_1$) and red markers (instrument $I_2$) in each panel represent the synthetic measurements with inferred offsets extracted. Solid dark line and shadow regions are as in Figure \ref{fig:toymodel}. The RV-like time-series also included the RV model for the two planets (red line).
Bottom panel: Synthetic RV-like data folded on the orbital period of each injected planet following the subtraction of the systemic velocities, GP signal, and the other planet. The plots also show the inferred RV model for each planet (solid black line).
These plots were generated automatically by \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}.}
\label{fig:toymodel2}
\end{figure*}
\input{tables/table2}
\subsubsection{Comparison between the different Runs}
Table~\ref{tab:parameterstoy2} shows the parameter value for all the four Runs described in this Section, and Figure~\ref{fig:kas} shows the recovered posterior distribution for the Keplerian signals b and c for each case.
From Table~\ref{tab:parameterstoy2} we can see that all Runs are able to recover the planetary-like signals within the error bars. This is relevant since we created this data set with activity-like amplitudes significantly larger than the Keplerian ones, with the intention to show that the code is able to recover coherent signals in the RV-like time-series.
This provides some tentative evidence that if RV observations are planned with suitable observing campaigns and with the right instruments, we may be able to find planetary signals even in cases with extreme activity.
From Table~\ref{tab:parameterstoy2} and Figure~\ref{fig:kas} we can also see that run that provides better precision on the detected Doppler semi-amplitudes is \emph{Run 4}. We can argue that \emph{Runs} 1 and 2 are not optimal to analyse this problem, because of the way we create the synthetic RV time-series.
But we may think that Run 3 and 4 are equivalent: the model of the $S_2$ data is created with a function draw using a QP kernel, while the $S_1$ time-series with a function created with a QP kernel and its first time derivative. And from Table~\ref{tab:parameterstoy2}, we see that the QP kernel hyper-parameters are fully consistent within the error bars for both runs.
However, we obtain better detection of the Keplerian signals on \emph{Run 4} where we use the multidimensional GP approach.
This is explained by the discussion in Sect.~\ref{sec:gpcomparisons}, where we mentioned that the multidimensional GP approach ensures that the underlying function $G(t)$ is the same for all time-series.
This result shows the advantage of using the activity indicators to model stellar signals within a multidimensional GP framework.
These tests also illustrate \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}'s ability to handle different instruments. However, we caution that this capability should be used with care. In particular, multiple instruments do not only have different offsets between them, but they may also observe in different wavelength ranges. This means that the stellar signal that each instrument observes may be different. Therefore, they cannot necessarily be treated as the same underlying signal that the multi-GP approach described in this manuscript assumes.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{images/kasb.pdf}\\
\includegraphics[width=0.47\textwidth]{images/kasc.pdf}
\caption{Probability distributions for the Doppler semi-amplitudes for the signals b (upper panel) and c (lower panel) as described in Sect.~\ref{sec:toy2}. Each colour corresponds to the different Runs described in the text. The true value of the parameter is shown with a vertical line in each case.}
\label{fig:kas}
\end{figure}
\subsection{Multi-band transit modelling}
We create a synthetic light curve using \texttt{citlalicue}\ (see Sect.~\ref{ap:citla}) in order to test the ability of \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ to model multi-band transit modelling.
We created data assuming we have a flattened light curve of a system with two transiting planets observed with two different instruments.
For the first instrument, named ``B1'', the data ranges from 0 to 15 days, with one data point every 5 min, with a precision of 500 ppm per datum.
The limb darkening coefficients for this fictitious observed star with instrument B1 are $u_1=0.25$ and $u_2=0$ \citep[$q1=0.06$ and $q2 =0.50$, following][]{Kipping2013}, following the quadratic law of \citet[][]{Mandel2002}.
For the second instrument, named ``B2'', data go from 20 to 30 days, with one observations every 5 min with a precision of 100 ppm.
The limb darkening coefficients for this fictitious star with instrument B2 are $u_1=0.50$ and $u_2=0.25$ \citep[$q1=0.56$ and $q2 =0.33$, following][]{Kipping2013}.
We injected two transiting planets with circular orbits into the light curves. We also assume that both planets are transiting a star with a density of $1.4\,{\rm g\,cm^{-3}}$. The time of transit $T_0$, orbital period $P$, impact parameter $b$, and scaled planet radius $r_{\rm p}$ for each planet are given in Table~\ref{tab:parameterstoytransits}.
The \texttt{Jupyter} notebook used to create this synthetic data set is provided \href{https://github.com/oscaribv/pyaneti/blob/master/inpy/example_multiband/multiband_transits.ipynb}{here \faGithub}, and
Figure \ref{fig:toymodel3} shows the synthetic light curves for both bands.
We perform a multi-band and multi-planet modelling with \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}. Table~\ref{tab:parameterstoytransits} shows the sampled parameters and priors we use. We note that we sample for a different limb darkening coefficients for each band. We also sample for the stellar density, and recover the scaled semi-major axis for each planet using Kepler's third law.
We sample the parameter space with 100 independent chains.
We create the posterior distributions for each sampled parameter with the last 5000 iterations of converged steps using a thin factor of 10.
This example can be reproduced in \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ by running \texttt{./pyaneti.py example\_multiband} in the main \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ directory.
Figure~\ref{fig:toymodel3} shows the phase-folded light curves for each transiting planet.
We show the inferred parameters in Table~\ref{tab:parameterstoytransits}.
We can see that the code is able to recover the orbital and planet parameters for both transiting signals. We also note that \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ can recover the band-dependent parameters for this example (limb darkening coefficients). In Sect.~\ref{sec:realdata} we describe some real stellar systems where the multi-band capabilities of the code have been used.
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{images/example_multiband_lightcurve.pdf} \\
\includegraphics[width=0.45\textwidth]{images/example_multibandb_tr.pdf}
\includegraphics[width=0.45\textwidth]{images/example_multibandc_tr.pdf}
\caption{Top panel: Synthetic light curve created with the fictitious B1 (green circles) and B2 (red squares) instruments. The light curve model with the two transiting signals is shown with a black thick line.
Bottom panel: Phase-folded light curves for injected planet signal $b$ (left) and $c$ (right). Each plot shows the data for instrument B1 (gray circles) and B2 (gray squares) separated by an offset. The plots also show the data in 10-min bins for each instrument (B1 as green and B2 as red circles) together with the inferred model (black line) for each case.
These plots are generated automatically by \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}.}
\label{fig:toymodel3}
\end{figure*}
\input{tables/table3}
\subsection{Single transit event}
We create a single transit event light curves using \texttt{citlalicue}\ (see Appendix~\ref{ap:citla}) in order to test the ability of \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ to model single transits.
We create the data assuming we have a flattened light curve of a system with one planetary transit.
The data range from 9 to 11 d, with one data point every 5 min, with a precision of 100 ppm per datum.
The limb darkening coefficients for this fictitious star are $u_1=0.25$ and $u_2=0$ \citep[$q1=0.06$ and $q2 =0.50$, following][]{Kipping2013}, following the quadratic law of \citet[][]{Mandel2002}.
We injected a transiting planet assuming a circular orbit around a star with a Sun-like density.
The time of transit is $T_0 = 10$ d, orbital period $P=30$ d, scaled semi-major axis $a/R_\star = 40.6$, impact parameter $b=0.5$ and scaled planet radius $r_{\rm p}=0.25$.
The code needed to create this synthetic data set is provided in this \href{https://github.com/oscaribv/pyaneti/blob/master/inpy/example_single/example_single.ipynb}{link \faGithub}.
We perform a single transit modelling with \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ by indicating \texttt{is\_single\_transit = True} in the input file for this system. Table~\ref{tab:parameterssingle} shows the sampled parameters and priors we use.
We sample the parameter space with 100 independent chains and create the posterior distributions with the last 5000 iterations of converged chains with a thin factor of 10.
This example can be reproduced in \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ by running \texttt{./pyaneti.py example\_single} within the main \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ directory.
Figure \ref{fig:toysingle} shows the inferred model for the single transit and Table~\ref{tab:parameterssingle} shows the inferred parameters. \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ is able to recover the injected values of $T_0$, $b$, $r_{\rm p}$, $q_1$, and $q_2$ within 1-sigma error bars.
We note that for single transit fits, the scaled semi-major axis and periods given by the code do not have the same meaning as for normal runs fitting multiple transits.
The scaled semi-major axis is a dummy value sampled to take into account the transit shape, while the orbital period is a derived parameter, i.e., we do not sample for it directly, we compute it with the other sampled parameters assuming the orbit is circular.
This capability of the code has been used before to estimate periods for transit signals detected by the \emph{TESS} mission \citep[e.g.,][]{Eisner2020b}.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{images/toy_singleb_tr.pdf} \\
\caption{Model for the single transit modelling test. Synthetic data are shown as green circles. The inferred model is shown with a black line.
This plot is shown as is provided automatically by \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}.}
\label{fig:toysingle}
\end{figure}
\input{tables/table4}
\subsection{Real planetary systems}
\label{sec:realdata}
We have shown how \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ is able to recover the injected planetary and orbital parameters in specific examples of synthetic spectroscopic-like and photometry-like time-series.
This demonstrates that if we believe that our RV and transit data behave as the models described in this paper, \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ will be able to provide reliable parameter estimates.
Fortunately, the new implementations of the code have already been applied to real data in peer-reviewed literature.
In this section we describe how \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ has been used in these analyses. We describe scenarios in which all the new additions of the code have been used combined.
It is not our intention to reproduce the analyses or plots published in the aforesaid manuscripts. We only describe how \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ was used in the relevant paper and we also provide examples to reproduce the analyses in such papers.
\subsubsection{K2-100}
\citet[][]{Barragan2019} published the RV detection of K2-100\,b, a transiting exoplanet orbiting a young active star in the Praesepe cluster.
In that manuscript we used the ability of the code to fit multi-band transit photometry simultaneously with a 3-dimensional GP approach for spectroscopic time-series including a Keplerian component.
The data were modelled using a multidimensional GP approach with a QP kernel, for which the harmonic complexity detected in the spectroscopic time-series was relatively high ($\lambda_{\rm p}$$= 0.6$).
Therefore, the use of the GP derivatives for the detection of the planetary signal was crucial in this case (see discussion in Sect.~\ref{sec:gpderivatives}).
Figure~\ref{fig:k2100} shows a 20 days subset of the RV and $\log R'_{\rm HK}$\ time-series of Figure 2 in \citet[][]{Barragan2019}.
This figure shows that the $\log R'_{\rm HK}$\ time-series behaves as the $S_1$ curve in the example in Sect~\ref{sec:gpderivatives}, and the RV data as the $S_2$ curve, i.e., the RV curve behaves as the time derivative of the $\log R'_{\rm HK}$\ curve.
We note that K2-100 is a fast rotating and spot-dominated young active star, and we expect that for young stars with similar characteristics, the GP derivative will be crucial to model the activity induced signals in the RV data.
\begin{figure*}
\centering
\includegraphics[width=0.99\textwidth]{images/k2100.pdf}
\caption{A 20 day subset of RV (upper panel) and $\log R'_{\rm HK}$\ (lower panel) time-series for K2-100. Both time-series have had the inferred offsets in \citet[][]{Barragan2019} subtracted.
Measurements are shown as filled symbols with error bars. For the RV time-series, we show the planet-induced signal alone (red line) and also the subtracted activity signal (black line). For the $\log R'_{\rm HK}$\ time-series, we show the activity model as a black line.}
\label{fig:k2100}
\end{figure*}
We made available different setups to reproduce the analysis of the K2-100 system from the \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ main directory.
To reproduce the multi-band analysis of the system, the relevant command is \texttt{./pyaneti.py example\_multiband\_k2100}. To model the RV together with activity indicators, the command is \texttt{./pyaneti.py example\_timeseries\_k2100}.
And to reproduce the full modelling including multi-band transits and RV together with activity indicators, type \texttt{./pyaneti.py example\_full\_k2100}.
\subsubsection{TOI-1260}
\citet[][]{Georgieva2021} report the detection and characterisation of two mini-Neptunes transiting TOI-1260. They use \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ to perform joint transit and RV modelling using a multi-planet, multi-band, and multidimensional GP configuration.
As for K2-100, the time-series were modelled using a multidimensional GP approach with a QP kernel. However, for this case, the harmonic complexity was moderate ($\lambda_{\rm p}$$= 1.4_{-0.5}^{+1.0}$). Even in this case of moderate harmonic complexity, the multidimensional GP approach improved the planetary detection when compared with other common approaches \citep[see][for more details]{Georgieva2021}.
\subsection{Execution time}
All examples presented in Sect.~\ref{sec:tests} were run on a personal laptop with 8$\times$1.90 GHz Intel\textregistered\ Core\texttrademark\ i7-8650U CPUs with Ubuntu 18.04.
For comparison, we also ran all the examples with the same setup in a cluster with 12$\times$3.40 GHz Intel\textregistered\ Xeon\textregistered\ CPUs.
We compiled the code with \texttt{GFortran} and ran all the examples in parallel.
Table~\ref{tab:times} show the execution time for the different setups presented in Section~\ref{sec:tests}.
We note that \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ is able to produce ready-to-publish results in a relative short time even on a personal laptop. However, the execution time can be longer, depending on several parameters, such as the MCMC configuration, number of planets being modelled, number of data, number of CPUs used, etc.
For example, the setup \texttt{example\_full\_k2100} is a complex problem that combines multi-band analysis and multidimensional GP regression. The total model samples 38 parameters simultaneously. Therefore, to reproduce the results by \citet[][]{Barragan2019} requires a relatively high computational power.
\begin{table}
\centering
\caption{Execution time for test runs in Sect.~\ref{sec:tests}. \label{tab:times}}
\begin{tabular}{lcc}
\hline
Setup & Time in laptop & Time in cluster \\
\hline
\noalign{\smallskip}
\texttt{example\_toyp1} & 29m47s & 11m21s \\
\texttt{example\_toyp2} & 23m10s & 8m35s \\
\texttt{example\_multiband} & 17m59s & 6m1s \\
\texttt{example\_single} & 1m03s & 35s \\
\texttt{example\_timeseries\_k2100} & 38m34s & 18m51s \\
\texttt{example\_multiband\_k2100} & 31m43s & 12m20s \\
\texttt{example\_full\_k2100} & 8h40m & 3h16m \\
\hline
\noalign{\smallskip}
\end{tabular}
\end{table}
\section{Conclusions}
\label{sec:conlusions}
In this manuscript we present a major update to the \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ code.
This new version of \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ allows one to perform GP regression, as well as multi-band and single transit fits.
The biggest advantage of this new version of \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ with respect to other codes is the multidimensional GP regression that is included within the RV modelling routines.
We present some tests to show how \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ can recover the model parameters in different data set configurations.
We will continue updating \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ in the future according to the needs of the exoplanet {community}.
We described how we expanded the likelihood to account for data correlation using a multidimensional GP. In this manuscript we use the bulit-in MCMC described in \citet[][]{pyaneti} to sample the parameter space. However, we note that this new likelihood can be used with other sampling techniques. In the near future we plan to add different sampling techniques in \href{https://github.com/oscaribv/pyaneti}{\texttt{pyaneti}\,\faGithub}\ with other MCMC \citep[e.g.,][]{emcee,zeus} sampling algorithms, as well as incorporate nested sampling \citep[e.g.][]{multinest,polychord,dynesty} methods.
Together with the code, we also provided some discussion on the use of GPs to model RV time-series.
We pointed out how training GPs with activity indicators or light curves may not be the best approach for all cases of stellar activity.
Even if the discussion is open, these aspects should be taken into account when modelling RV time-series using GPs.
We also presented \texttt{citlalicue}\ and \texttt{citlalatonac}. These numerical tools allow one to create synthetic photometric and spectroscopic time-series that may be useful to plan observing campaigns, especially in the cases in which the stellar signal is significantly larger than the planetary ones.
\section*{Acknowledgements}
{We thank the anonymous referee for their helpful suggestions that improved the quality of this paper.}
This publication is part of a project that has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (Grant agreement No. 865624).
NZ and SA acknowledge support from the UK Science and Technology Facilities Council (STFC) under Grant Code ST/N504233/1, studentship no. 1947725.
\section*{Data availability}
The data and software underlying this article are available as online supplementary material accessible following the links provided in the online version.
\bibliographystyle{mnras}
|
\section{Introduction}
\IEEEPARstart{I}{nterlaced} scanning mechanisms have been widely used in early TV broadcasting systems (\textit{e.g.}, NTSC, PAL, and SECAM). Different from the current progressive scanning method, the odd and even rows of pixels in interlaced frame are scanned from two different frames, namely the odd and the even fields (as shown in Fig. 1). In this way, the high frame rate and transmission bandwidth of the video display have been well balanced. Unfortunately, two fields captured at different time instances may cause displacement differences, which cannot be perfectly aligned in space. Hence, as illustrated in Fig. 2, obvious comb artifacts can be observed in the interlaced frames, especially in the case of large motion.
{\color{cblue} With the development of high-definition digital display technology, progressive scanning technology has become mainstream, and thus the researches on deinterlacing has received little attention. However, there are many precious early videos that can arouse the resonance of audiences. At the same time, the risk and cost of remaking classic early videos are lower than that of making new ones. Therefore, in recent years, the reconstruction of early videos has received more and more attentions. Unfortunately, early videos usually contain various artifacts during the processes of digitization, compression, transmission and storage. For example, interlaced artifacts in early interlaced videos are often mixed with various noises. Traditional de-interlacing methods are only designed for deinterlacing task, which cannot solve such complex artifacts. In summary, the main purpose of this paper is to design an effective joint enhancement model to reconstruct the real world early interlaced videos.
\begin{figure}[!t]
\setlength{\abovecaptionskip}{0.2cm}
\setlength{\belowcaptionskip}{0.2cm}
\centering
\includegraphics[height=8cm,width=0.48\textwidth]{Figure1.pdf}
\caption{Illustration of the interlaced scanning mechanism.}
\label{fig:1}
\end{figure}
\begin{figure}[htb]
\setlength{\abovecaptionskip}{0.2cm}
\setlength{\belowcaptionskip}{0.2cm}
\centering
\includegraphics[height=5cm,width=0.4\textwidth]{Figure2.pdf}
\caption{Interlaced frame may contain severe comb-teeth artifacts.}
\label{fig:2}
\end{figure}
The traditional interlaced scanning mechanism can be defined as $I=f_{interlace}(T_{odd},T_{even})$, where $I$ is an interlaced frame,$T_{odd}$ and $T_{even}$ denote the odd field and even field respectively, and $f_{interlace}$ represents the interlacing scanning process as shown in Fig. 1. Previous deinterlacing methods mainly solve the inverse process of $f_{interlace}$. Among them, the traditional methods can be categorized into temporal interpolation, spatial interpolation and motion-based methods. Temporal interpolation-based method utilizes adjacent frames to interpolate the lost lines, but it is sensitive to large motion and may produce severe motion artifacts. Spatial interpolation-based method uses only one field to estimate the whole frame to avoid interlacing aliasing brought by large motions. However, due to the missing information of another field, the quality of reconstructed details is significantly reduced. Motion-based method first estimates the motion between two frames and then performs temporal or spatial interpolation based on the estimated motion. With the rapid development of deep neural network (DNN) and video processing techniques, many effective DNN-based image and video restoration models \cite {zhang2017beyond, dong2015image, lim2017enhanced} have been proposed. But they cannot solve deinterlacing task directly, since interlacing is caused by specific manmade mechanisms. DICNN \cite{zhu2017real} first introduced shallow convolutional networks to the real-time video de-interlacing task. However, this model still ignores the mixture of various kinds of unnatural effects in real-world early interlaced videos. Besides, the shallow backbone of DICNN leads to limited learning ability, and thus cannot handle the complex artifacts in early videos well.
In this paper, the degradation model of early interlaced videos is defined as $I=f_{codec}(f_{interlace}(T_{odd},T_{even} ))+n$, where $f_{codec}$ denotes the video compression process and $n$ represents noises. Existing jointly deinterlacing and denoise methods \cite{zhao2020din,2020Deep} are based on single-frame reconstruction and cannot make full use of the temporal similarity between adjacent frames. However, interlaced frames are highly correlated with adjacent frames in time series, hence the performance of current single-frame de-interlacing networks \cite{zhu2017real,zhao2020din,2020Deep} still has a lot of room for improvement.}
In order to solve the above problems, we propose a multi-frame early video deinterlacing network (MFDIN), which aggregates information of multiple frames for reconstruction. The MFDIN is composed of three modules: , i.e., spatial feature interpolation module, temporal alignment fusion module and frame reconstruction module. To make fully use of the prior knowledge of the interlacing scanning mechanism, the input frames are firstly split into odd and even fields and then input to the first module for feature extraction and vertical pixel interpolation.Inspired by recent state-of-the-art (SOTA) super-resolution algorithm \cite{wang2019edvr} and deformable convolution (DCN) methods \cite{dai2017deformable, zhu2019deformable}, an efficient alignment fusion module was proposed to effectively perform temporal alignment and aggregation by learning more applicable offsets, which can reduce the negative impact of complex artifacts such as compression artifacts on inter-frame motion. {\color{cblue}In addition, the MFDIN can support the output of same frame rate or double frame rate, because each interlaced frame naturally contains two fields.}
The main contributions of this paper are summarized as follows:{\color{cblue}
\begin{enumerate}
\item This paper focuses on joint enhancement of early interlaced videos in real world, rather than traditional deinterlacing. An effective multi-frame deinterlacing network is proposed to fully use the temporal information, while current early video deinterlacing methods only focus on single frame.
\item Appropriate field vertical interpolation and lightweight alignment fusion module (AFM) have been specifically designed for this complex task. An efficient offset network is presented to perceive interframe motion accurately. Comparing to multiscale DCNs, the proposed AFM with single DCN can improve the stability and reduce the computational cost simultaneously.
\item Owing to the specifically designed multi-frame architecture, MFDIN can achieve convincing and visually appealing results on both synthetic datasets and real-world early interlaced videos.
\item Furthermore, extended experiments on different tasks, e.g., 1080i(25P)-4K(50P), demonstrate that the proposed model is suitable for the joint problem of deinterlacing, compression artifacts removal, video super-resolution and double frame rate.
\end{enumerate}
}
The rest of the paper is organized as follows. Related works are reviewed in Section II. Section III introduces the proposed method in detail. Experimental results are discussed in Section IV, and Section V concludes the paper.
\begin{figure*}[htb]
\setlength{\abovecaptionskip}{0.2cm}
\setlength{\belowcaptionskip}{0.2cm}
\centering
\includegraphics[height=6.2cm,width=0.9\textwidth]{Figure3.pdf}
\caption{{\color{cblue}Architecture of the proposed multi-frame deinterlacing network (MFDIN). Three input frames and two output frames are used as an illustrative example.}}
\label{fig:3}
\end{figure*}
\section{Related works}
\subsection{Video Deinterlacing.}
Traditional deinterlacing methods \cite{de1998deinterlacing, doyle1998interlaced} mainly focus on estimating the missing lines in each field, so as to realize the effect of progressive scanning. Related algorithms can be roughly divided into temporal interpolation, spatial interpolation, and motion-based methods. Temporal interpolation algorithm is a basic way of deinterlacing, which estimates missing lines by means of adjacent frames. However, this type of method may produce severe artifacts when there is a large motion. Spatial interpolation methods \cite{wang2012efficient, wang2013moving, jakhetiya2014fast, brox2014edge} estimate the missing lines via spatial interpolation of a single field. Because the interlacing artifacts tend to appear at the edges of objects, these methods focus on reducing the step effect at the image boundary. However, because only half a frame is used, the results are usually blur and less than satisfactory. To reproduce better results, motion-based methods \cite{mohammadi2012enhanced, jeon2009weighted, zhu2017motion} combine multiple fields in adjacent frames based on accurate motion adaptation or motion compensation. However, accurate motion is difficult to obtain and the computational complexity of these methods is usually too high. In recent years, DNN-based methods \cite{sajjadi2018frame, li2019feedback, zhang2018adversarial, zamir2017feedback, xiang2020zooming} have achieved much better performance than traditional methods in many low-level vision tasks. Zhu \textit{et al.} \cite{zhu2017real} first introduced a shallow deinterlacing convolutional neural network (DICNN) to video deinterlacing tasks and achieved impressive results. However, the learning ability of shallow networks is limited, and complex artifacts in real-world interlaced videos cannot be handled well in this manner. Zhao \textit{et al.} \cite{zhao2020din} proposed a deeper deinterlacing network (DIN) for early videos by vertically up-sampling and merging two fields. As mentioned before, these DNN-based methods are all based on a single frame, which cannot sufficiently utilize inter-frame redundancy information. To improve deinterlacing performance, we first propose a multi-frame-based method, which aggregates information from multiple fields through lightweight and efficient alignment, to restore high visual quality frames.
\subsection{Related Multi-Frame Video Restoration.}
{\color{cblue}Compared to single-frame restoration \cite{lim2017enhanced, zhang2018density, yu2018crafting, yu2019path}, video restoration tasks, such as video super resolution (VSR) and video compression artifact removal, usually utilize more temporal information to refine missing details.} But the ubiquitous occlusion and large motion in the frame sequence seriously affect the performance of the networks. Therefore, inter-frame information often requires temporal alignment or other processing to be fully utilized. The basic approach is to estimate the motion between frames by calculating the optical flow \cite{sun2018pwc, xue2019video, caballero2017real}. However, for the areas contain occlusion and large motion, it is difficult and high-computational-cost to obtain accurate flow. 3D convolution can also be used to achieve temporal aggregation \cite{kim20183dsrnet,jo2018deep,ying2020deformable}. For instance, Joe \textit{et al.} \cite{jo2018deep} used stacked 3D convolution layers for motion compensation to extract temporal information. Ying et al. \cite{ying2020deformable} adopted 3D deformable convolution to extract spatial and temporal information within a temporal sliding window for VSR and implemented this method over the entire video sequence. However, these methods lead to massive redundant computation, which limits the efficiency of the model. {\color{cblue}Some methods replace the above scheme by implicit alignment. For instance, TDAN \cite{tian2020tdan} and EDVR \cite{wang2019edvr} reconstructed high quality super-resolution frames through DCN-based alignment modules, which surpasses some flow-based methods. Niklaus et al. \cite{niklaus2017video} and zhou et al. \cite{zhou2019spatio} proposed adaptive convolution for video interpolation and video deblurring by implicitly utilizing motion cues and pixel information. In addition, recurrent neural networks (RNN) have also been introduced to video tasks to simplify sequence-to-sequence learning. For instance, a modified convolutional long short-term memory Network (LSTM) was introduced for video quality enhancement in \cite{yang2019quality}. Isobe et al. \cite{isobe2020revisiting} proposed a recursive time modeling method with residuals to stabilize the training of RNNs and improve the performance. In \cite{guan2019mfqe}, Bi-LSTM was used to distinguish high quality frames in the compressed sequence, so as to help enhance the reconstruction of the adjacent low-quality compressed frames.} For our task, the mixed interlacing and other artifacts increase the difficulty of explicit motion estimation. Inspired by \cite{wang2019edvr}, we propose an efficient alignment module, which implicitly aligns and fuses features from multiple fields by means of more accurate offset estimation.
\section{Proposed Method}
\subsection{Overview}
Given 3 consecutive interlaced frames $I_{[t-1:t+1]}$, our purpose of deinterlacing is to reconstruct a high-quality central frame $\hat{O}_{t}^{odd}$, which is consistent with the ground truth frame $O_{t}^{odd}$. Note that for the tasks that need to double the frame rate at the same time, \textit{e.g.} 25P-to-50P, the proposed model can also reconstruct two consecutive frames $\hat{O}_{t}^{odd}$and $\hat{O}_{t}^{even}$ for each central frame $I_t$. As illustrated in Fig. 3, the proposed MFDIN consists of three stages, \textit{i.e.}, feature extraction, alignment fusion, and reconstruction. Firstly, we divide each frame into odd and even fields for feature extraction and vertical magnification. In order to align and aggregate the features of multiple fields, the reconstructed features of adjacent frames and central frame are then gathered by the alignment fusion module. Finally, the aligned and fused multi-frame features pass through the reconstruction module, to further reduce artifacts and restore high-quality results. Each module will be introduced and analyzed in detail in the following.
\begin{figure}[htb]
\setlength{\abovecaptionskip}{0.2cm}
\setlength{\belowcaptionskip}{0.2cm}
\centering
\includegraphics[height=3cm,width=0.36\textwidth]{Figure4.pdf}
\caption{Illustration of the vertical pixel shuffle layer.}
\label{fig:4}
\end{figure}
\begin{figure}[htb]
\setlength{\abovecaptionskip}{0.2cm}
\setlength{\belowcaptionskip}{0.0cm}
\centering
\includegraphics[height=4.4cm,width=0.48\textwidth]{Figure5.pdf}
\caption{Architecture of the offset net.}
\label{fig:5}
\end{figure}
\begin{figure}[H]
\setlength{\abovecaptionskip}{0.2cm}
\setlength{\belowcaptionskip}{0.2cm}
\centering
\includegraphics[height=4.2cm,width=0.48\textwidth]{Figure6.pdf}
\caption{Alignment fusion module.}
\label{fig:6}
\end{figure}
\subsection{Feature extraction module}
The main purpose of the feature extraction module is to extract features from split fields, reproduce vertically interpolated frames, and reduce the combined artifacts preliminarily.
The classical ResNet \cite{he2016deep}, which is commonly used in related low-level vision tasks \cite{lim2017enhanced} is selected as the backbone. Motivated by DIN\cite{zhao2020din}, the odd and even fields of each frame are separately input for feature extraction, which can simulate the reverse process of manmade interlacing scanning mechanism. In this way, the network can perceive the difference among multiple fields, and further reduce the impact of interlacing artifacts during feature extraction. Note that the resolution of the input fields is half of the original frame, we interpolate the reconstructed field features in the vertical direction to restore the original resolution by means of a vertical pixel shuffle (VPS) layer \cite{shi2016real,zhao2020din}, which is illustrated in Fig. 4. Specifically, as shown in Fig. 3, the three input interlaced frames are firstly divided into six fields, and then these fields are separately input into the same feature extraction module to obtain six reconstructed frames with the original resolution. In this paper, the feature extraction module contains five residual blocks (RB) \cite{lim2017enhanced} and global skip connection is also utilized to improve the convergence of the algorithm.
After the feature extraction stage, interlacing artifacts in reconstructed frame features can be effectively reduced. However, many high frequency details and complex noises cannot be handled well because only single-field information is used. Hence, temporal information can be aggregated to obtain better reconstruct results.
\subsection{Alignment fusion module}
In multi-frame schemes, the neighbor frames need to be accurately aligned to avoid ghost artifacts. For deinterlacing task, common optical flow alignment or 3D convolution \cite{tran2015learning} methods may be suboptimal. Due to a variety of complex artifacts, it is difficult to learn accurate optical flow, and inaccurate optical flow seriously affects the performance. In addition, gradually fusing spatial-temporal information through multiple stacked 3D convolution layers creates a huge computational burden, which reduced the overall efficiency of the model. In this paper, we tend to find an efficient way to fuse temporal information with low computational complexity.
Inspired by SOTA multi-frame super-resolution method EDVR \cite{wang2019edvr}, we also use DCN-based implicit alignment instead of explicit optical-flow-based warping. In EDVR \cite{wang2019edvr}, multiple DCNs are used at multiple scales to enhance performance, but this scheme is not very suitable to our task. Firstly, although stacking multiple DCNs can enlarge the receptive field of DCN, the network fluctuates frequently in the training process, which often results in offset explosion in our de-interlacing experiment, making the network difficult to converge. Moreover, the original DCN structure is used directly for most video tasks, but the offset required by the original deformable convolution for target detection is different from that for temporal alignment. The motion flow representation we used has always been considered that is intuitively and empirically invariant with appearance. Especially in our task, the feature information contains a large number of complex artifacts, which further increases the difficulty of offset learning. Therefore, traditional DCN, which merely learns the offset with a 3$\times$3 convolution, should not be simply adopted. Hence, we design a lightweight offset network, which follows the multi-scale philosophy that commonly used in optical flow estimation \cite{dosovitskiy2015flownet}. The U-net \cite{ronneberger2015u} structure with large receptive fields and multiple scales is adopted to dynamically deal with different displacements, and to predict more appropriate offsets. Detailed architecture of the offset network is shown in Fig. 5. Different from the common optical flow alignment model, implicit alignment is directly performed on the feature level from the reconstructed multiple frame sequences, as illustrated in Fig. 6.
Given the six feature maps $F_{i} (i=1,2,3,4,5,6)$ reconstructed in the previous stage, our goal is to obtain aligned and fused features $F_{3}*$ and $F_{4}*$ of the central two fields. Taking $F_{3}^*$ as an example, each adjacent map is concatenated with $F_{3}$ and then corresponding offset $\Delta P$ is calculated as:
\begin{figure}[ht]
\centering
\includegraphics[height=4cm,width=0.35\textwidth]{Figure7.pdf}
\caption{{\color{cblue}Comparison of the loss curves of multi-DCN scheme and the proposed alignment fusion module.}}
\label{fig:7}
\end{figure}
\begin{equation}\label{eqn2}
\Delta P_{j}= f([F_{3}, F_{j}]), j=(1,2,4,5)
\end{equation}
{\color{cblue}where $f(\cdot)$ denotes the offset net, and [·,·] denotes the concatenation operation. Here, $\Delta P_j = \left\{\Delta p_{k} | k=1,2,...,K\right\}$ refer to the offsets of the K sampling positions of the convolution kernel. For instance, 9 sampling positions are corresponding to each $3 \times 3$ convolution kernel. With the learned offset $\Delta P_j$, dynamic sampling can be performed by means of DCN to achieve feature alignment:
\begin{equation}\label{eqn3}
A_{j}= f_{DCN}(F_{j},\Delta P_{j}), j=(1,2,4,5)
\end{equation}
where $f_{DCN}(\cdot)$ is the deformable convolution operation. More specifically, for each position $p_0$ on the aligned feature $A_{j}$, the dynamic sampling process can be defined as:
\begin{equation}\label{eqn1}
A_{j}(p_{0})=\sum_{k=1}^{K}w_{k}\cdot F_{j}(p_{0}+\Delta p_{k})
\end{equation}
where $w_k$ is the kernel weight, and the convolution will be operated on the irregular positions $p_{0}+\Delta p_{k}$. Then, the fusion features can be obtained:}
\begin{equation}\label{eqn4}
F_{3}* = g([A_{j},]) , j=(1,2,3,4,5)
\end{equation}
where $g(\cdot)$ denotes a $1\times1$ convolutional layer that is used to fuse aligned features. Similarly, $F_{4}^*$ is calculated in the same way as $F_{3}^*$ with the corresponding neighbor features $F_j (j=2,3,5,6)$.
{\color{cblue}Compared with the multiple-DCN scheme, the proposed alignment fusion module can estimate more applicable offsets, thus making the network more stable in the training phase. Fig. 7. shows the loss curves of these two schemes. It can be found that the proposed AFM performs better in terms of both training stability and convergence speed, which proves that the designed offset network structure can learn more accurate offsets and thus achieve better aligned and fused features in the case of complex degradations.}
\subsection{Frame Reconstruction module}
To further restore the details and refine the visual quality, the fused features from the previous stage are input into the reconstruction module, which is composed of 8 stacked RBs.
Motivated by the feature aggregation structure \cite{liu2020residual}, we introduce two residual aggregate blocks (RAB) to utilize the features in different levels. The structure of RAB is shown in Fig. 3, which concatenates the outputs of multiple RBs and then fuses them via a $1\times1$ convolutional layer. Finally, the features are reconstructed to one or two output frames by means of an output layer, which is a $3\times3$ convolutional layer.
To optimize our network, Charbonnier penalty function \cite{lai2017deep} is adopted:
\begin{equation}\label{eqn4}
Lr=\sqrt{||\hat{O}_{t}^{odd}-{O}_{t}^{odd}||^{2}+\varepsilon^{2}}
\end{equation}
When two frames are restored from one interlaced frame at the same time, the loss function becomes:
\begin{equation}\label{eqn4}
Lr=\sqrt{||\hat{O}_{t}^{odd}-{O}_{t}^{odd}||^{2}+||\hat{O}_{t}^{even}-{O}_{t}^{even}||^{2}+\varepsilon^{2}}
\end{equation}
where $\hat{O}_{t}^{odd}$ and $\hat{O}_{t}^{even}$ refer to the output frames of the network, ${O}_{t}^{odd}$ and ${O}_{t}^{even}$ refer to ground-truth HR video frame, and $\varepsilon$ is empirically set to $1\times10^{-3}$.
\section{EXPERIMENTS}
\subsection{Datasets}
Because high-quality ground truth of early interlaced video is difficult to be obtained, we build a synthetic training dataset based on YOUKU 2K video dataset \cite{YoukuVSRE2019}, and constructed three test sets to conduct objective quality assessment. Besides the synthetic datasets, several real-world early PAL/NTSC TV videos are used for comparison and subjective tests. In addition, extended experiments in different application scenarios are also implemented, such as 540i-to-1080p, and 1080i(25P)-to-4K(50P).
\textbf{Training Set.} {\color{cblue}YOUKU-2K \cite{YoukuVSRE2019} is an open-source video super-resolution dataset, which is composed of 1000 2K video sequences in RAW format. Note that this dataset contains various types of films and TV series, which are highly consistent with the contents of early interlaced TV videos.} For training, we collected 400 video sequences (excluding cartoons) from the dataset and then generated interlaced frames according to the traditional interlaced scanning mechanism, \textit{i.e.}, odd and even lines of an interlaced input frame are scanned from two adjacent ground truth frames. Note that two ground truth frames are thus corresponding to one interlaced frame. For the task that only restores one frame for an interlaced frame, the first ground truth frame is selected. Furthermore, to roughly simulate the mixed compression artifacts in early videos, we applied the FFMPEG tool to compress the synthesized interlaced videos with H.264 codec with random "-CRF" parameters from 30 to 38.
\textbf{Synthetic Testing Sets.} For testing, we conduct three synthetic testing sets. First, we selected 7 videos (70 groups of frames) from the remaining YOUKU-2K videos. Then, 40 groups of frames were also selected from SJTU-4K dataset \cite{song2013sjtu} to verify the reconstructed results for ultra-high-definition videos. {\color{cblue}Furthermore, 150 groups of frames were selected from the Vimeo90K dataset \cite{2019Video}, which is widely used in the video enhancement tasks. }The frames in these testing sets were degraded in the same way as the training set.
\begin{figure*}[h]
\centering
\includegraphics[height=6.2cm,width=\textwidth]{Figure8.pdf}
\caption{{\color{cblue}Results of different methods in ablation testing, (a)model with single frame as input, (b)model with multi frames as input, (c)model with split frame, (d)model with split frame and alignment.}}
\label{fig:8}
\end{figure*}
\begin{table}[htb]
\renewcommand{\arraystretch}{1.5}
\caption{{\color{cblue}Psnr Results of Ablation Testing on Synthetic Datasets}}
\label{table:1}
\centering
\tiny
\begin{tabular}{ccccc}
\hline
\textbf{Model} & \textbf{Multi Frames} & \textbf{Split Frame} & \textbf{Alignment Fusion} & \textbf{PSNR} \\ \hline
Baseline & & & & 29.50 \\
Multi Frames(entire) & $\surd$ & & & 31.44 \\ \hline
MFDIN(one frame with split) & & $\surd$ & $\surd$ & 32.19 \\
MFDIN(without split frame) & $\surd$ & & $\surd$ & 32.32 \\
MFDIN(without aligment) & $\surd$ & $\surd$ & & 32.18 \\
MFDIN & $\surd$ & $\surd$ & $\surd$ & \textbf{32.76} \\ \hline
\end{tabular}
\end{table}
\begin{table}[htb]
\renewcommand{\arraystretch}{1.5}
\caption{{\color{cblue}Psnr Results of MFDIN with Different number of input frames}}
\label{table:2}
\centering
\begin{tabular}{ccc}
\hline
Number of input frames & PSNR & FLOPS \\ \hline
1 frame(2fields) & 32.19 & \textbf{550G} \\
3 frames(5fields) & \underline{32.76} & \underline{1375G} \\
5 frames(7fields) & \textbf{32.88} & 1834G \\ \hline
\end{tabular}
\end{table}
\textbf{Real-world Early Video Testing Set.} A total of 8 PAL/NTSC TV videos, which contain severely interlaced artifacts, were collected to verify the generalization of the proposed method on real-world early videos. Note that there lacks ground truth for these real-world videos, hence, subjective evaluation is implemented by comparing the visual quality.
\textbf{Extended Training and Testing Sets.} In addition to the complex artifacts, early videos are often with low resolution and low frame rate. To verify the proposed method in different application scenarios that contain deinterlacing, compression artifacts removal, super-resolution and reproduce double frame rate, we conducted extended experiments on different application scenarios, \textit{i.e.}, 1080i-to-4K, 1080i(25P)-to-4K(50P), 540i-to-1080p, and 540i(25P)-to-1080p(50P). To build training and testing sets in these cases, we first used FFMPEG tool to perform $2\times$ down-sampling and frame rate adjustment, and then degraded these videos via the aforementioned interlacing and compression process.
\subsection{Implementational details}
In our experiment, we selected more than 28,000 groups of frames from the training set, and each group consists of three consecutive synthetic interlaced frames. These frames were further cropped at random position with the resolution of $128\times128$ in the training process. The number of channels in each RB is set to 64. Moreover, in the alignment fusion module, the number of channels in the offset network is set to 64 at all scales. For optimization, we use Adam optimizer \cite{kingma2014adam}, the learning rate is initially set as $4\times10^{-4}$, and then decayed to $1\times10^{-7}$ with a cosine annealing. The proposed MFDIN is implemented with the PyTorch framework and trained 100 epochs on a NVIDIA TATAN V GPU.
\subsection{Ablation studies}
Ablation testing is conducted to verify the effectiveness of different components in MFDIN. Firstly, we compare the reconstruction results of single frame input and multi-frame input. By comparing the results in Fig. 8(a) and (b), multi-frame method can restore better details than the single-frame method due to the more information is available from adjacent frames. By comparing Fig. 8(b) and (c), we can find that input split fields, which utilizes the prior knowledge of manmade interlaced scanning, can reproduce better results than input whole frames. As shown in Fig. 8(d), the proposed alignment module can effectively improve the utilization of temporal information and reduce the negative impact of large inter-frame motion on the reconstruction result, so as to restore clearer images. {\color{cblue}At the same time, the quality assessment results are listed in Table I. Compared to the complete model, each of the proposed strategies brings different degrees of improvement. Even only one frame is input, the proposed module can provide significant gain after applying field splitting and alignment. Because one interlaced frame is equivalent to two consecutive fields, and the effective alignment further improves the utilization and fusion of information.}
{\color{cblue}In addition, Table II lists the PSNR and FLOPS of the reconstructed results with different input frames, i.e., 2 fields (1 frame), 5 fields (3 frames) and 7 fields (5 frames). It can be found that multi-frame methods can achieve better performance by using more adjacent frames. Note that, although adding more neighbor frames can further refine the PSNR results, but it also significantly increases the computational cost. In our experiment, 5 neighbor fields (from 3 adjacent frames) are chosen as input to balance performance and cost.}
\begin{figure*}[htp]
\centering
\includegraphics[height=10cm,width=\textwidth]{Figure9.pdf}
\caption{Qualitative comparison on the \textbf{YOUKU 2K} synthetic testing set.}
\label{fig:9}
\end{figure*}
\begin{figure*}[htp]
\setlength{\abovecaptionskip}{0.2cm}
\setlength{\belowcaptionskip}{0.0cm}
\centering
\includegraphics[height=5.6cm,width=\textwidth]{Figure10.pdf}
\caption{Qualitative comparison on the \textbf{SJTU 4K} synthetic testing set.}
\label{fig:10}
\end{figure*}
\begin{figure*}[htp]
\setlength{\abovecaptionskip}{0.2cm}
\setlength{\belowcaptionskip}{0.0cm}
\centering
\includegraphics[height=5.6cm,width=\textwidth]{Figure11.pdf}
\caption{{\color{cblue}Qualitative comparison on the \textbf{Vimeo 90K} synthetic testing set.}}
\label{fig:11}
\end{figure*}
\begin{table*}[htp]
\renewcommand{\arraystretch}{1.5}
\caption{{\color{cblue}Psnr/ssim Results of Different Methods on the Synthetic Datasets}}
\label{table:3}
\centering
\begin{tabular}{c|cccccccccccccc}
\hline
& \multicolumn{2}{c}{Input} & \multicolumn{2}{c}{SFI} & \multicolumn{2}{c}{DICNN[4]} & \multicolumn{2}{c}{EDSR[3](retrain)} & \multicolumn{2}{c}{EDVR[6](retrain)} & \multicolumn{2}{c}{DIN[5]} & \multicolumn{2}{c}{MFDIN} \\
\multirow{-2}{*}{} & PSNR & SSIM & PSNR & SSIM & PSNR & SSIM & PSNR & SSIM & PSNR & SSIM & PSNR & SSIM & PSNR & SSIM \\ \hline
YOUKU 2K & 24.25 & 0.7382 & 27.86 & 0.8439 & 29.16 & 0.8764 & 29.50 & 0.8750 & 31.90 & 0.9036 & {\color[HTML]{00B0F0} \underline{ 32.01}} & {\color[HTML]{00B0F0} \underline{ 0.9071}} & {\color[HTML]{FF0000} \textbf{32.76}} & {\color[HTML]{FF0000} \textbf{0.9132}} \\ \hline
SJTU 4K & 21.75 & 0.6971 & 28.16 & 0.8341 & 29.30 & 0.8679 & 30.56 & 0.8887 & 32.42 & 0.8963 & {\color[HTML]{00B0F0} \underline{ 32.52}} & {\color[HTML]{00B0F0} \underline{ 0.8995}} & {\color[HTML]{FF0000} \textbf{33.10}} & {\color[HTML]{FF0000} \textbf{0.9051}} \\ \hline
Vimeo 90K & 24.45 & 0.7343 & 28.80 & 0.8381 & 29.87 & 0.8581 & 30.74 & 0.8629 & 31.80 & 0.8781 & {\color[HTML]{00B0F0} \underline{ 32.27}} & {\color[HTML]{00B0F0} \underline{ 0.8813}} & {\color[HTML]{FF0000} \textbf{32.82}} & {\color[HTML]{FF0000} \textbf{0.8880}} \\ \hline
Parameters & \multicolumn{2}{c}{\textbackslash{}} & \multicolumn{2}{c}{\textbackslash{}} & \multicolumn{2}{c}{0.05M} & \multicolumn{2}{c}{1.42M} & \multicolumn{2}{c}{2.30M} & \multicolumn{2}{c}{2.22M} & \multicolumn{2}{c}{1.88M} \\ \hline
Inference time & \multicolumn{2}{c}{\textbackslash{}} & \multicolumn{2}{c}{\textbackslash{}} & \multicolumn{2}{c}{0.0164s} & \multicolumn{2}{c}{0.0391s} & \multicolumn{2}{c}{0.0986s} & \multicolumn{2}{c}{0.0451s} & \multicolumn{2}{c}{0.0835s} \\ \hline
Flops & \multicolumn{2}{c}{\textbackslash{}} & \multicolumn{2}{c}{\textbackslash{}} & \multicolumn{2}{c}{8.05G} & \multicolumn{2}{c}{397G} & \multicolumn{2}{c}{1288G} & \multicolumn{2}{c}{665G} & \multicolumn{2}{c}{1375G} \\ \hline
\end{tabular}
\end{table*}
\begin{table*}[htp]
\renewcommand{\arraystretch}{1.5}
\caption{Quantitative Comparison of Extended Experiments}
\label{table:4}
\centering
\setlength{\tabcolsep}{5.5mm}{
\begin{tabular}{c|cccccccc}
\hline
& \multicolumn{2}{c}{Youku2K} & \multicolumn{2}{c}{SJTU} & \multicolumn{2}{c}{Youku2K} & \multicolumn{2}{c}{SJTU} \\
& \multicolumn{2}{c}{540i-to-1080P} & \multicolumn{2}{c}{1080i-to-4K} & \multicolumn{2}{c}{540i(25P)-to-1080p(50P)} & \multicolumn{2}{c}{1080i(25P)-to-4K(50P)} \\
& PSNR & SSIM & PSNR & SSIM & PSNR & SSIM & PSNR & SSIM \\ \hline
VI+EDSR[3] & 36.58 & 0.9520 & 34.51 & 0.9366 & 36.58 & 0.9522 & 34.51 & 0.9370 \\ \hline
DICNN[4]+EDSR[3] & 40.29 & 0.9670 & 39.05 & 0.9585 & \color[HTML]{00B0F0} \underline{40.30} & \color[HTML]{00B0F0} \underline{0.9670} & \color[HTML]{00B0F0} \underline{39.04} & \color[HTML]{00B0F0} \underline{0.9585} \\ \hline
DIN[5]+SloMo[41] & \color[HTML]{00B0F0} \underline{40.94} & \color[HTML]{00B0F0} \underline{0.9693} & \color[HTML]{00B0F0} \underline{39.71} & \color[HTML]{00B0F0} \underline{0.9617} & 35.73 & 0.9309 & 33.95 & 0.9282 \\ \hline
MFDIN & \textbf{\color[HTML]{FE0000} 43.18} & \textbf{\color[HTML]{FE0000} 0.9808} & \textbf{\color[HTML]{FE0000} 42.28} & \textbf{\color[HTML]{FE0000} 0.9747} & \textbf{\color[HTML]{FE0000} 42.04} & \textbf{\color[HTML]{FE0000} 0.9765} & \textbf{\color[HTML]{FE0000} 41.09} & \textbf{\color[HTML]{FE0000} 0.9722} \\ \hline
\end{tabular}}
\end{table*}
\begin{figure}[htb]
\centering
\includegraphics[height=4cm,width=0.32\textwidth]{Figure12.pdf}
\caption{Average SSIS scores of subjective testing on real-world early videos.}
\label{fig:12}
\end{figure}
\subsection{Results on the synthetic testing sets}
The proposed MFDIN is compared with traditional single field vertical interpolation (SFI) and recent DNN-based deinterlacing methods, \textit{i.e.}, DICNN \cite{zhu2017real} and DIN\cite{zhao2020din}. In addition, we also retrained a deeper single-frame model EDSR \cite{lim2017enhanced} with 16RBs and a SOTA multi-frame model EDVR \cite{wang2019edvr}. Note that their pixel-shuffle (PS) magnification module is removed for deinterlacing task. As in DICNN, these methods take the whole frame as input directly. For fair comparison, all networks are retrained on the same training set. {\color{cblue}Reconstructed results of interlaced and compressed frames on three synthetic datasets are shown in Fig. 9, Fig. 10. and Fig. 11. respectively.} It can be seen that simple SFI method can reduce the comb-teeth effect, but it also enlarges various mixed artifacts. DICNN \cite{zhu2017real} is a lightweight shallow network with few parameters. However, it is only designed for real-time deinterlacing tasks, and thus cannot handle the complex noise in early videos. Because the existing mixed artifacts break the simple inter-field relationship of the interlacing scanning mechanism, the deinterlacing performance of the retrained EDSR is greatly reduced, and various complex noises also cannot be removed well. DIN\cite{zhao2020din} is a deinterlacing method specially designed for early video, which can effectively remove the interlacing comb-teeth and suppress the noise. However, only the information of a single input frame is used, many details cannot be restored totally. In addition, when there is a large motion between frames, as shown in Fig. 10, the unnatural ghost artifacts cannot be eliminated well. Due to the introduction of learning of temporal information, the retrained EDVR can restore more details than the single-frame method. However, under the influence of complex interlacing effects and severe compression noise, inter-frame motion can not be well learned, and ghost artifacts still exist in the restored image. Compared with these SOTA methods, the proposed MFDIN performs much better. On the one hand, fields split and vertical interpolation effectively eliminate the interlaced comb artifacts. On the other hand, alignment and fusion of multi-frames can suppress the ghosting artifacts caused by large-scale motion, and reduce mixed compression artifacts significantly.
{\color{cblue}In addition, Table III lists the quantitative assessments of these methods, which shows that the proposed MFDIN can achieve higher PSNR and SSIM values on three test sets. The number of parameters, the amount of computation and the inference time are also list in Table III, note that the computation and inference time are obtained at a resolution of 720$\times$540.}
\begin{figure*}[htp]
\centering
\includegraphics[height=22cm,width=\textwidth]{Figure13.pdf}
\caption{Results on real-world early videos.}
\label{fig:13}
\end{figure*}
\begin{figure*}[htp]
\centering
\includegraphics[height=10.2cm,width=\textwidth]{Figure14.pdf}
\caption{Results of extended experiments on 540i-to-1080p and 1080p-to-4K tasks.}
\label{fig:14}
\end{figure*}
\subsection{Results on real-world early videos}
Fig. 13 shows the reconstruction results of some early videos. It can be found that there are severe and complex noises such as interlaced comb-teeth, aliasing, noise, and compression artifacts in these early videos. Single-field SFI may amplify various noises in vertical direction. DICNN can reproduce better results than traditional SFI, but it still cannot handle the complex compression noise well. Note that DICNN directly assembles the low-quality input field and reconstructed output field via interlaced scanning. This strategy is not very suitable for early videos because the input frames often contain various noises. DIN can reconstruct high quality results, but it still cannot totally remove severe noise. In contrast, the proposed MFDIN can reconstruct a cleaner and smoother results without obvious artifacts.
Furthermore, subjective single-stimulus-impairment-scale (SSIS) test has been implemented to compare the visual quality of reconstructed early videos. The display condition and test session details were set according to Rec. ITU-R BT.500 \cite{ITU_R_BT500}. A total of 20 observers were invited to score the impairment scales of 5 groups of early video frames, with the rating of 0-5 (5 denotes the highest quality level). The average SSIS scores are illustrated in Fig. 12. It can be found that the MFDIN obtains the highest subjective scores.
\subsection{Extended experiments}
To verify the scalability and generalization performance of the model, we performed additional experiments on several different application scenarios, such as 540i-to-1080p and 1080i-to-4K. These tasks require deinterlacing, denoising, $2\times$ spatial and temporal interpolation. The proposed MFDIN can be directly applied to these joint tasks by merely adding a PS layer before the last output convolutional layer. In particular, the SFI and DICNN methods do not have the $2\times$ super-resolution ability, while the DIN method does not have the frame rate up-conversion ability. For comparison, we used super-resolution network EDSR \cite{lim2017enhanced} to magnify the results of SFI and DICNN, and utilized video interpolation method Super-SloMo \cite{jiang2018super} to double the frame rate of DIN results.
Quantitative results are shown in Table IV, from which we can find that the proposed multi-frame and multi-stage structure can uniformly deal with the joint and complex tasks, and thus achieves much better performance on these tasks than simple combination of deinterlacing, super-resolution, and frame interpolation methods. In the 540i-to-1080p task, DIN performed better than the two-stage DICNN+EDSR, but the combination of DIN+SloMo did not perform well in the 540i(25P)-to-1080p(50P) task. By contrast, our one-stage method achieves the best performance on both tasks.
The results of different methods on 540i-to-1080p and 1080i-to-4K tasks are shown in Fig. 14. These results illustrate that the MFDIN significantly outperforms other simply combined methods, which can totally remove the interlaced and compression artifacts, restore clearer and more accurate details by means of inter-frame information.
There are some reasons that the designed architecture is suitable for the joint tasks for deinterlacing, compression artifacts removal, super-resolution and frame interpolation. First, the field split and reconstruction strategy can effectively separate the deinterlacing task by using the prior knowledge of interlaced scanning. Second, the backbone of each module is motivated by SOTA super-resolution methods \cite{lim2017enhanced,wang2019edvr,liu2020residual}, which have been proved to be effective for super-resolution and deblocking problems. Third, one interlaced frame contains two fields indeed, which is naturally suitable to double the frame rate. As a result, the MFDIN can even handle these difficult joint problems with a well-designed, efficient and relatively lighter architecture.
{\color{cblue}At last, the shortcomings of the proposed method are discussed. To recover the real-world early videos, MFDIN is designed to suppress the complex interlacing artifacts mixed with other artifacts. However, when there is no other degradations in the videos, the MFDIN is not suitable because some lighter networks can handle normal deinterlacing task, such as DICNN.}
\section{Conclusion}
This paper proposed a multi-frame network (MFDIN) for joint enhancement of early interlacing videos, which can effectively remove complex degradations such as interlacing and compression artifacts. In MFDIN, multiple interlaced frames are firstly split into separate fields, and then feature extraction and vertical interpolation are carried out by a feature extraction module. A temporal alignment module based on deformable convolution was then presented, which can effectively align and fuse the information from multiple fields. Finally, an efficient feature aggregation and reconstruction module was used to further refine details. Experimental results show that the proposed method can reproduce high quality results from early interlaced videos containing severe and complicated unnatural artifacts.
\bibliographystyle{IEEEtran}
|
\section{Introduction}\label{intro}
WAVES was designed to build on the AGES observations of the Virgo Cluster \citep{2012MNRAS.423..787T,2013MNRAS.428..459T} and to take advantage of Arecibo's great brightness temperature sensitivity as a filled aperture telescope. Over a 10 km\,s$^{-1}$\ velocity width, AGES reached a $1\sigma$ column-density sensitivity of $1.5\times 10^{17} {\rm\ cm}^{-2}$ \citep{2016MNRAS.456..951K} while the shallower ALFALFA survey with Arecibo reached $5\times 10^{17} {\rm\ cm}^{-2}$ \citep{2008AA...487..161G}. The Arecibo surveys are highly complementary to the VLA, where the VIVA survey of Virgo reached a relatively shallow $1-1.7\times 10^{19} {\rm\ cm}^{-2}$ \citep{2007ApJ...659L.115C} but had a 15\arcsec\ resolution compared to Arecibo's 3.5\arcmin\ beam.
AGES discovered a large number of optically-dark H\,{\sc i}\ clouds, including some with anomalously high velocity widths that could not be easily explained as products of harassment \citep{2012MNRAS.423..787T,2017MNRAS.467.3648T}. The AGES data also demonstrated that deep surveys with good sensitivity to low column-density gas could detect previously undiscovered ram pressure stripped H\,{\sc i}\ tails: ten new tails were confidently detected, almost doubling the number of known tails in the cluster, along with sixteen possible detections \citep{2020AJ....159..218T}.
WAVES has the same integration time and observing method as AGES, and was intended to extend this column-density sensitivity to a large fraction of the Virgo cluster, either discovering new dark clouds and tails or putting strong upper limits on their population. The survey strategy was to complete small chunks of the survey to full sensitivity before moving on to the next section, thus 30 of a planned 55 square degrees was completed to full sensitivity before the Arecibo Telescope's collapse in 2020. A further 2.5 square degrees reached 2/3 of the planned integration time, i.e. 80\% of the planned sensitivity.
\section{Results}\label{results}
Initial results from the first five square degrees of WAVES were published in \citet{2019AJ....158..121M}, including the discovery of around 150\% more mass in H\,{\sc i}\ in the ALFALFA Virgo 7 Complex \citep{2007ApJ...665L..15K} and of a much more extended ram-pressure stripped tail on NGC 4522 than was seen by \citet{2007ApJ...659L.115C}. More recent analysis of the known dark features in the rest of the WAVES area are presented in Figure \ref{fig1}. These are:
\begin{itemize}
\item{VirgoHI 4 \citep{2004MNRAS.349..922D}, where the WAVES observations reveal around 50\% more H\,{\sc i}\ than the WSRT observations of \citet{2005AA...437L..19O} and showing that although the H\,{\sc i}\ plume does extend further to the north east, where the WSRT observations reached the limits of the primary beam, this extension is only on the order of a few arcminutes (approximately the size of the Arecibo beam).}
\item{ALFALFA Virgo Cloud 3 \citep{2007ApJ...665L..15K}, which was noted as unresolved in the ALFALFA data but with the greater sensitivity of AGES is seen to be elongated (with a position angle of around $62.5\deg$) and to show a clear velocity gradient.}
\item{H\,{\sc i}\ tail on NGC 4424 originally seen by VIVA \citep{2007ApJ...659L.115C} and with a large extension detected by \citet{2017MNRAS.464..530S}. While they have a combined WSRT and KAT-7 dataset, both with similar column density sensitivities, this extension is only seen in their KAT-7 data (see their Figure 5 and associated discussion). WAVES reaches the same column-density sensitivity as their combined map and does not detect this extension, while follow-up single-pixel observations with the L-Band Wide receiver at Arecibo similarly failed to detect it. We therefore conclude that this was probably an artefact in the KAT-7 data and is not real.}
\item{KW Cloud detected by \citet{2017MNRAS.464..530S} near NGC 4451. This is not detected by WAVES despite being well above the sensitivity limit, nor was it seen in the ALFALFA data by \citet{2007ApJ...665L..15K} where (based on the stated mass and velocity width) it should have fallen above their $6.5\sigma$ confident detection threshold. As with the extension to NGC 4424, we were also unable to detect this source in observations with the L-Band Wide receiver at Arecibo, and thus we conclude that it is also not real. While \citet{2017MNRAS.464..530S} do not state whether this was detected in both the WSRT and KAT-7, the H\,{\sc i}\ morphology shown in their Figure A2 is similar to that of the extension on the H\,{\sc i}\ tail of NGC 4424, leading us to speculate that this is also an artefact in the KAT-7 data.}
\end{itemize}
\begin{figure*}
\plotone{WAVES_RNAASfig_mod.png}
\caption{WAVES (blue) and AGES (red) coverage of the Virgo cluster, with final integration times indicated by weight. Known dark H\,{\sc i}\ structures, clockwise from the top right: VirgoHI 4 \citep{2004MNRAS.349..922D} showing H\,{\sc i}\ contours from WAVES and the outer contour from the WSRT \citep[][blue]{2005AA...437L..19O} overlaid on the DSS. ALFALFA Virgo Cloud 3 \citep{2007ApJ...665L..15K} showing elongation in the moment map (bottom) and a velocity gradient in a cut along this elongation (top). The tail on NGC 4424 with H\,{\sc i}\ contours from WAVES (blue) and VIVA (red) overlaid on the map of \citet{2017MNRAS.464..530S}. The KW Cloud seen by \citet{2017MNRAS.464..530S}, with the WAVES data (blue) overlaid on their spectrum.}
\label{fig1}
\end{figure*}
\section{Conclusions}
WAVES is revealing new details about H\,{\sc i}\ in the Virgo cluster and verifying (and sometimes falsifying) earlier detections of optically-dark H\,{\sc i}\ clouds. We expect full analysis of the data set to reveal new dark objects and ram pressure tails, as with the AGES Virgo cluster fields. It is able to do this thanks to the combination of high brightness-temperature sensitivity, zero-spacing information, high mapping speed and relatively high spatial resolution that was offered by the Arecibo telescope. The loss of that telescope leaves the US community with limited options for such surveys that are sensitive to low column-density gas.
\begin{acknowledgments}
This research was conducted in part at the SOFIA Science Center, which is operated by the Universities Space Research Association under contract NNA17BF53C with the National Aeronautics and Space Administration. RT and BD acknowledge the support of the Czech Science Foundation grant 19-18647S and the institutional project RVO 67985815. The Arecibo Observatory is a facility of the National Science Foundation operated under cooperative agreement by the University of Central Florida and in alliance with Universidad Ana G. Mendez, and Yang Enterprises, Inc.
\end{acknowledgments}
|
\section{Introduction}
Considerable research efforts have been spent within the last few years towards approaching classical control problems with modern statistical and optimization tools from the Machine Learning framework, envisaging practical applications, see for example \cite{Dean2018}, \cite{mania2019}, \cite{dean2020}, \cite{furieri2020}.
The starting point of the aforementioned research efforts has been the classical LQG control problem, which deals with partially observed linear and time-invariant dynamical systems driven by Gaussian noise and where the problem is finding the optimal output feedback law that minimizes the expected value of a quadratic cost.
In this paper an end-to-end sample-complexity bound on learning LQG controllers that stabilize the true system with high probability is established by incorporating recent advances in finite time (non-asymptotic) system identification (\cite{Dahleh2020}). The contribution resides in the development of a tractable robust control synthesis procedure, that allows for bounded additive model uncertainty on the coprime factors of the model of the plant, thus circumventing both the need for state feedback and the restrictive assumption on the plant's open loop stability. The resulted sub-optimality gap is bounded as a function of the level of the model uncertainty. The end-to-end sample complexity bound for learning robust LQG controllers is $\mathcal{O}(\sqrt{{logT}/{T}})$, where $T$ is the time horizon for learning. For open-loop stable systems, \cite{furieri2020} recently proved that the performance for LQG controllers deteriorates linearly with the model estimation error, starting from the original analysis of \cite{Dean2018} from the case of learning fully observed LQR controllers. The robust control synthesis proposed here achieves the same scaling for the sub-optimality gap as \cite{Dean2018}, namely $\mathcal{O}(\gamma^2)$, where $\gamma$ is the model uncertainty level.
\subsection{The Linear Quadratic Gaussian Problem}
Within the last few years, modern statistical and algorithmic methods led to new solutions for classical control problems, such as the Linear Quadratic Gaussian problem. For a discrete-time {\bf LTI} (Linear and Time Invariant) systems driven by Gaussian process and sensor noise:
\useshortskip
\begin{equation}\label{stateeq}
\begin{aligned}
x_{k+1} & = {A} x_k + {B} u_k + {\delta x}_k,\\ y_k & = {C} x_k + {D} u_k + \nu _k,
\end{aligned}
\end{equation}
\noindent
where $x_k \in \mathbb{R}^n$ is the state of the system, $u_k \in \mathbb{R}^m$ is the control input and $y_k \in \mathbb{R}^p$ is the measurement output with ${\delta x}_k \in \mathbb{R}^m$, $\nu_k \in \mathbb{R}^p$ are Gaussian noise with zero mean, covariance $\sigma_{\delta x}^2 I$ and $\sigma_\nu^2 I$
respectively, the classical LQG control problem is defined as:
\noindent
\useshortskip
\begin{equation}\label{LQGCostOriginal}
\begin{aligned}
\min_{u_0, u_1, ...} \quad \lim\limits_{T \rightarrow \infty} & \mathbb{E} \bigg[ \dfrac{1}{T} \sum \limits_{t=0}^{T} (y^T_t P_1 y_t + u^T_t P_2 u_t) \bigg] \\
\textrm{subject to} \quad & \eqref{stateeq},
\end{aligned}
\end{equation}
where, $P_1, P_2$ is positive definite. Without loss of generality, it is assumed that $P_1 = I_p$, $P_2 = I_m$, $\sigma_{\delta x} = 1$, $\sigma_\nu = 1$.
In a nutshell, the problem can be stated as {\em learning} with high probability and in finite time the model of an unknown LTI system and subsequently designing its optimal LQG controller, while accounting for the inherent model uncertainty incurred at the {\em learning} stage.
\subsection{The Main Technical Ingredient}
Identifying LTI models from input-output data has been the focus of time-domain identification.
Using coprime factors instead of the state space realization of the system has a great advantage as it ensures to work on unstable system identification. The Transfer Function Matrix (\textbf{TFM}) of the plant is written as:
\begin{equation}\label{plantequation}
\textbf{G}(z) = \tilde{\textbf{M}}^{-1}(z) {\bf \tilde{N}}(z) = \textbf{N}(z)\textbf{M}^{-1}(z)
\end{equation}
\noindent
where $\tilde{\textbf{M}}(z)$, $\tilde{\textbf{N}}(z)$, $\textbf{M}(z)$ and $\textbf{N}(z)$ are stable TFM.
Doubly coprime factorization of the TFM of a LTI system plays a key role in many sectors of the factorization approach to filter synthesis and multi-variable control systems analysis. A doubly coprime factorization of a given LTI plant is closely related to the Parameterization of all stabilizing controllers for this plant.
With a realization of the TFM, various formulas to compute doubly coprime factorizations over the ring of stable and proper Rational Matrix Functions (\textbf{RMF}) have been proposed both for standard (proper) and for generalized (improper, singular, or descriptor) systems.
The formulas are expressed either in terms of a stabilizable and detectable realization of the underlying TFM or make additional use of a realization of a full or reduced order observer based stabilizing controller.
\subsection{Contributions}
Recently, LQG control has been studied in a model based Reinforcement Learning framework (\cite{furieri2020}) and the sub-optimality performance degradation of the robust LQG controller was proved to scale as a function of the modeling error. However, the results in \cite{furieri2020} are valid only for open-loop stable systems, thus excluding many situations of practical interest. This paper shows how to remove the stability assumption on the unknown system, while at the same time streamlining the equivalent optimization problem, by reducing the size of the subsequent linear constraints. The proposed algorithm is consistent with previous results, while allowing for a much stronger description of the modeling error as bounded additive uncertainty on the coprime factors of the model of the plant (without restriction on the McMillan degree of the true plant or on its number of unstable poles). As expected, the presence of additive, norm-bounded factors on the coprimes of the plant renders the cost functional non-convex, therefore the derivation of an upper-bound on the cost functional is needed. This is subsequently exploited to derive a quasi-convex approximation of the robust LQG problem. An inner approximation of the quasi convex problems via FIR truncation is employed. Previous results (\cite{mania2019}, \cite{furieri2020}) show that indeed the {\em certainty equivalent controller} may achieve superior sub-optimality scaling than our result, but only for the fully observed LQR settings (\cite{mania2019}), in the setup of a stricter requirement on admissible uncertainty. Given the lack of prior gain margin for the optimal LQG controller, which is known to be notoriously fragile, even under small model uncertainty the stabilizability of the resulted controller may be lost, thus the availability of a more general framework for modelling of uncertainty is important.
Existing non-asymptotic identification methods (\cite{Dahleh2020}) have been adapted in order to yield a comparable end-to-end sample complexity. The identification of the unstable plant is performed in closed loop, directly on its the coprime factors via the dual Youla Parameterization (\cite{Anderson1998}). The algorithm employed for system identification doesn't require the knowledge of the model's order (\cite{Oymak1}), which is the common scenario in many applications.
Pursuing the identification of the plant a $\mathcal{H}_2$ bound for Hankel matrix estimation with high probability is derived, followed by a $\mathcal{H}_{\infty}$ bound on the uncertainty on the coprime factors, which quantifies the modeling error. The robust controller design is recast as convex optimization for estimated nominal model within a {\em worst case scenario} on the uncertainty. For the output feedback of potentially unstable plants, the resulted sample complexity result is matched to the same level as that obtained in recent papers (\cite{Boczar2018}, \cite{dean2020}, \cite{furieri2020}), where the robust so-called {\em SLP} or {\em IOP} procedures (\cite{Zheng2020a}) are used for design.
The paper is organized as follows: the general setup is given in \hyperref[generalsetupandpreliminaries]{Section II}. The robust controller synthesis with uncertainty on the coprime factors is included in \hyperref[robustcontrollersynthesis]{Section III}. The sub-optimality guarantees are discussed in \hyperref[endtoendanalysis]{Section IV}. A brief discussion on the closed-loop system identification is provided in \hyperref[clsysid]{Section V} with end-to-end sample complexity results. Conclusions and future possible directions are given in \hyperref[conclusion]{Section V}. All the proofs are postponed to the \hyperref[appendix]{Appendices}, where literature review, mathematical preliminaries and few remarks also have been discussed briefly.
\section{General Setup and Technical Preliminaries}\label{generalsetupandpreliminaries}
The notation used in this paper is fairly common in control systems.
Upper and lower case boldface letters (e.g. ${\bf z} $ and ${\bf G}$) are used to denote signals and transfer function matrices, and lower and upper case letters (e.g. $z$ and $A$) are used to denote vectors and matrices. The enclosed results are valid for discrete-time linear systems, therefore $z$ denotes the complex variable associated with the $\mathbf{Z}$-transform for discrete-time systems.
A LTI system is {\em stable} if all the poles of its TFM are situated
inside the unit circle for discrete time systems. The TFM of a LTI system is called {\em unimodular} if it is square, stable and has a stable inverse. For the sake of brevity the $z$ argument after a transfer function may be omitted. Some frequently used notation is listed in the next page.
\begin {table}[h!]
\begin{tabular}{ >{\centering\arraybackslash}m{0.8in} >{\arraybackslash}m{5in} }
\toprule[1.25pt]
& {\bf Nomenclature of Basic Notation} \\
\midrule
LTI & Linear and Time Invariant \\
TFM & Transfer Function Matrix \\
DCF & Doubly Coprime Factorization \\
LCF & Left Coprime Factorization\\
RCF & Right Coprime Factorization\\
$x\overset{def}{=} y$ & $x$ is by definition equal to $y$ \\
$\mathbb{R}(z) $ & Set of all real--rational transfer functions \\
$\mathbb{R}(z)^{p \times m} $ & Set of $p \times m$ matrices having all entries in $\mathbb{R} (z)$ \\
${\bf T}^{ \ell \varepsilon}$ & The TFM of the (closed-loop) map having $\varepsilon$ as input and $\ell$ as output \\
${\bf T}^{ \ell \varepsilon}_{\bf Q}$ & The TFM of the (closed-loop) map from the exogenous signal $\varepsilon$ to the signal $\ell$ inside the feedback loop, as a function of the Youla parameter ${\bf Q}$\\
$\| {\bf G} \|_{{F}}$ & Frobenius norm, Schur norm or $l_2$ norm of ${\bf G} \in \mathbb{R}(z)$, defined as $|tr(\bf GG^*)|^{1/2}$\\
$\|{\bf G}\|_{\infty}$ & $\mathcal{H}_{\infty}$-norm of ${\bf G} \in \mathbb{R}(z)$, defined as
$\sup_{\omega} \sigma_{\max} \big({\bf G}(e^{j \omega})\big)$\\
$\| {\bf G} \|_{\mathcal{H}_{2}}$ & $\mathcal{H}_{2}$-norm of ${\bf G} \in \mathbb{R}(z)$, defined as
$ {{{\dfrac{1}{2\pi} \Big( \bigintsss \limits_{-\pi}^{\pi} tr \big({\bf G}^* (e^{j\omega}){\bf G}(e^{j\omega})\big) d\omega }}\Big)}^{1/2}$\\
$\texttt{pt}$ & Notations for true plant (e.g. ${\bf G}^{\mathtt{pt}}$, ${\bf K}^{\mathtt{pt}}$) \\
$\texttt{md}$ & Notations for nominal/estimated model (e.g. ${\bf G}^{\mathtt{md}}$, ${\bf K}^{\mathtt{md}}$)\\
${\bf G}^{\mathtt{pt}}$, ${\bf K}^{\mathtt{opt}}$ & True Plant, Optimal Controller\\
${\bf G}^{\mathtt{md}}$, ${\bf K}^{\mathtt{md}}_{\bf Q}$ & Estimated Model, Nominal stabilizing controller for any stable Youla parameter ${\bf Q}$\\
\bottomrule[1.25pt]
\end {tabular}
\end {table}
\subsection{Standard Unity Feedback}
A standard unity feedback configuration is depicted in Figure~\ref{2Block}, where ${\bf G}\in \mathbb{R}(z)^{p \times m}$ is a multi-variable LTI plant and ${\bf K}\in \mathbb{R}(z)^{m \times p}$ is an LTI controller. Here $w$, $\nu$ and $r$ are the input disturbance, sensor noise and reference signal respectively while $u$, $z$ and $y$ are the controls, regulated signals and measurements vectors, respectively.
If all the closed--loop maps from the exogenous signals $\displaystyle [ r^T\; \: w ^T \; \: \nu^T \;]^T$ to
any point inside the feedback loop of Figure~\ref{2Block} are stable, then ${\bf K}$ is said to be an (internally) stabilizing controller of ${\bf G}$ or equivalently that ${\bf K}$ stabilizes ${\bf G}$.
\begin{figure}[h]
\begin{tikzpicture}[scale=0.4]
\draw[xshift=0.1cm, >=stealth ] [->] (0,0) -- (3.5,0);
\draw[ xshift=0.1cm ] (4,0) circle(0.5);
\draw[xshift=0.1cm] (3,1) node {\bf{+}} (1,0.8) node {$r$};
\draw [xshift=0.1cm](6,0.8) node {$z$} ;
\draw[ xshift=0.1cm, >=stealth] [->] (4.5,0) -- (8,0);
\draw[ thick, xshift=0.1cm] (8,-1.5) rectangle +(3,3);
\draw [xshift=0.1cm](9.5,0) node {\Large{${\bf K}$}} ;
\draw[ xshift=0.1cm, >=stealth] [->] (11,0) -- (15.5,0);
\draw[ xshift=0.1cm ] (16,0) circle(0.5cm);
\draw [xshift=0.1cm](12.6,0.8) node {$u$} ;
\draw [xshift=0.1cm](18.6,0.8) node {$v$} ;
\draw [xshift=0.1cm] (14.8,0.7) node {\bf{+}};
\draw[ xshift=0.1cm, >=stealth] [->] (16,3.7) -- (16,0.5);
\draw [xshift=0.1cm] (16,3.6) node[anchor=south] {$w$} (15.3,1.5) node {\bf{+}};
\draw[ xshift=0.1cm, >=stealth] [->] (16.5,0) -- (21,0);
\draw[ thick, xshift=0.1cm ] (21,-1.5) rectangle +(3,3) ;
\draw [xshift=0.1cm] (22.5,0) node {\Large{${\bf G}$}} ;
\draw[ xshift=0.1cm, >=stealth] [->] (24,0) -- (28.5,0);
\draw[ xshift=0.1cm ] (29,0) circle(0.5);
\draw [xshift=0.1cm] (29,3.6) node[anchor=south] {$\nu$} (28.5,1.5) node {\bf{+}};
\draw [xshift=0.1cm] (28,0.7) node {\bf{+}};
\draw [xshift=0.1cm] (30.6,0.7) node {$y$};
\draw[ xshift=0.1cm, >=stealth] [->] (29.5,0) -- (33,0);
\draw[ xshift=0.1cm, >=stealth] [->] (29,3.7) -- (29,0.5);
\draw[ xshift=0.1cm, >=stealth] [->] (29,-0.5) -- (29,-3) -- (4,-3)-- (4, -0.5);
\draw [xshift=0.1cm] (3.3,-1.3) node {\bf{--}};
\useasboundingbox (0,0.1);
\end{tikzpicture}
\caption{Standard unity feedback loop of the plant $\bf G$ with the controller $\bf K$}
\label{2Block}
\end{figure}
\noindent
The notation ${\bf T}^{ \ell \varepsilon}$ is used to indicate the
mapping from signal $\varepsilon$ to signal $\ell$ after combining
all the ways in which $\ell$ is a function of $\varepsilon$ and solving any
feedback loops that may exist. For example, ${\bf T}^{zw}$ in Figure~\ref{2Block} is the mapping from the disturbances $w$ to the regulated measurements $z$.
\subsection{The Youla-Ku\c{c}era Parameterization}\label{2}
\begin{prop} \label{14Martie2019}
Given a TFM ${\bf K} \in \mathbb{R}(z)^{m \times p}$, a fractional representation of the form ${\bf K}={\bf R}^{-1}{\bf S}$ with ${\bf R}\in \mathbb{R}(z)^{m \times m}$, ${\bf S} \in \mathbb{R}(z)^{m \times p}$ is called a {\em left factorization} of {\bf K}. If ${\bf K}={\bf Y}^{-1}{\bf X}$ is a left factorization of ${\bf K}$ then any other left factorization of ${\bf K}$ such as ${\bf K}={\bf R}^{-1}{\bf S}$ is of the form ${\bf R}={\bf UY}$, ${\bf S}={\bf UX}$, for some invertible TFM ${\bf U}$.
\end{prop}
\noindent
Given a plant ${\bf G} \in \mathbb{R}(z)^{p \times m}$, a left coprime factorization of ${\bf G}$ is defined by ${\bf G} = \tilde{\bf M}^{-1}\tilde{\bf N}$, with $\tilde{\bf N}\in \mathbb{R}(z)^{p \times m}$, $ \tilde {\bf M} \in \mathbb{R}(z)^{p \times p} $ both stable and satisfying $\tilde{\bf M} \tilde{\bf Y} + \tilde{\bf N}\tilde{\bf X} =I_p$, for certain stable TFMs $\tilde{ \bf X} \in \mathbb{R}(z)^{m \times p}$, $\tilde {\bf Y} \in \mathbb{R}(z)^{p \times p}$. \\
\noindent
Analogously, a right coprime factorization of ${\bf G}$ is defined by ${\bf G} = {\bf NM}^{-1}$ with both factors ${\bf N}\in \mathbb{R}(z)^{p \times m}$, ${\bf M} \in \mathbb{R}(z)^{m \times m}$ being stable and for which there exist ${\bf X}\in \mathbb{R}(z)^{m \times p}$, ${\bf Y} \in \mathbb{R}(z)^{m \times m} $ also stable, satisfying ${\bf YM} + {\bf XN} = I_m$ \cite[Ch.~4, Corollary~17]{vidyasagar1985}, with $I_m$ being the identity matrix.\\
\begin{defn} \cite[Ch.4, Remark pp. 79]{vidyasagar1985}
A collection of eight stable TFMs $\big({\bf M}, {\bf N}$, $\tilde {\bf M}, \tilde {\bf N}$, ${\bf X}, {\bf Y}$, $\tilde {\bf X}, \tilde {\bf Y}\big)$ is called a {\em doubly coprime factorization} of ${\bf G}$ if $\tilde {\bf M}$ and ${\bf M}$ are invertible, yield the
factorizations
${\bf G}=\tilde{\bf M}^{-1}\tilde{\bf N}={\bf NM}^{-1}$, and satisfy the following equality (B\'{e}zout's identity):
\begin{equation}\label{bezoutidentity}
\left[ \begin{array}{cc} \tilde {\bf M} & \tilde {\bf N} \\ - {\bf X} & {\bf Y} \end{array} \right]
\left[ \begin{array}{cc} \tilde {\bf Y} & -{\bf N} \\ \tilde {\bf X} & {\bf M} \end{array} \right] = I_{p+m},
\left[ \begin{array}{cc} \tilde {\bf Y} & -{\bf N} \\ \tilde {\bf X} & {\bf M} \end{array} \right]
\left[ \begin{array}{cc} \tilde {\bf M} & \tilde {\bf N} \\ - {\bf X} & {\bf Y} \end{array} \right]
= I_{p+m}.
\end{equation}
\end{defn}
\vspace{3pt}
\begin{thm} \label{Youlaaa}
{\bf (Youla-Ku\u{c}era)} \cite[Ch.5, theorem 1]{vidyasagar1985} Let $\big({\bf M}, {\bf N}$, $\tilde {\bf M}, \tilde {\bf N}$, ${\bf X}, {\bf Y}$, $\tilde {\bf X}, \tilde {\bf Y}\big)$ be a doubly coprime factorization of ${\bf G}$. Any controller ${\bf K}_{\bf Q}$ stabilizing the plant ${\bf G}$, in the feedback interconnection of Figure~\ref{2Block}, can be written as
\begin{equation}
\label{YoulaEq}
{\bf K_Q}={\bf Y}_{\bf Q}^{-1}{\bf X_Q} = {\bf \tilde{X}}_{\bf Q} {\bf \tilde{Y}}_{\bf Q}^{-1},
\end{equation}
where
${\bf X_Q}$, ${\bf \tilde{X}_Q}$, ${\bf Y_Q}$ and ${\bf \tilde{Y}_Q}$ are defined as:
\begin{equation} \label{5}
\begin{split}
{\bf X_Q} \overset{def}{=} {\bf X}+{\bf Q} \tilde{\bf M}, \quad
&{\bf \tilde{X}_Q} \overset{def}{=} \tilde{\bf X}+{\bf MQ}, \quad \\
{\bf Y_Q} \overset{def}{=} {\bf Y} - {\bf Q} \tilde{\bf N}, \quad
&{\bf \tilde{Y}_Q} \overset{def}{=} \tilde{\bf Y}-{\bf NQ}
\end{split}
\end{equation}
for some stable ${\bf Q}$ in $\mathbb{R}(z)^{m\times p}$. It also holds that ${\bf K_Q}$ from (\ref{YoulaEq}) stabilizes ${\bf G}$, for any stable ${\bf Q}$ in $\mathbb{R}(z)^{m\times p}$.
\end{thm}
\begin{prop} \label{2018April29}
\label{rem:232pm27feb2012}
Starting from any doubly coprime factorization (\ref{bezoutidentity}), the following identity
\begin{equation}\label{dcrelQ}
\! \left[ \begin{array}{cl} \! {\bf U}_1 \tilde {\bf M} & {\bf U}_1 \tilde {\bf N} \\ - {\bf U}_2 {\bf X_Q}& {\bf U}_2 {\bf Y_Q} \! \end{array} \right] \!
\left[ \begin{array}{lr} \! \tilde {\bf Y}_{\bf Q} {\bf U}^{-1}_1 & -{\bf N} {\bf U}^{-1}_2 \\ \tilde {\bf X}_{\bf Q} {\bf U}^{-1}_1 & {\bf M} {\bf U}^{-1}_2 \! \end{array} \right] = I_{p+m}
\end{equation}
provides the class of {\em all} doubly coprime factorizations of ${\bf G}$, where ${\bf Q}$ is stable in $\mathbb{R}(z)^{m\times p}$ and ${\bf U}_1 \in \mathbb{R}(z)^{p\times p}$, ${\bf U}_2 \in \mathbb{R}(z)^{m \times m}$ are both unimodular.
\end{prop}
\begin{defn} \label{klap} $\mathbb{H}({\bf G}, {\bf K}_{{\bf Q}})$ denotes the TFM whose entries are the closed-loop maps from $\displaystyle [ r^T\; \: w ^T \; \: \nu^T \; ]^T$ to $\displaystyle [y^T\;\; u^T \;\; z^T \;\; v^T ]^T$ achievable via the stabilizing controllers (\ref{YoulaEq}), namely equation (\ref{klap1}) her
\begin{equation*} \label{klap3}
\left[ \begin{array}{c} y\\u\\z\\v\\ \end{array} \right] = \mathbb{H}({\bf G}, {\bf K}_{{\bf Q}}) \left[ \begin{array}{c} r\\w\\ \nu\\ \end{array} \right],
\end{equation*}
{\begin{equation} \label{klap1}
\mathbb{H}({\bf G}, {\bf K}_{{\bf Q}})\overset{def}{=}\left[ \begin{array}{ccc} (I_p+{\bf GK_Q})^{-1}{\bf GK_Q} & (I_p+{\bf GK_Q})^{-1}{\bf G} &(I_p+{\bf GK_Q})^{-1} \\
(I_m+{\bf K_Q}{\bf G})^{-1}{\bf K_Q} & -(I_m+{\bf K_QG})^{-1}{\bf K_QG} &-(I_m+{\bf K_Q}{\bf G})^{-1}{\bf K_Q} \\
(I_p+{\bf GK_Q})^{-1} & -(I_p+{\bf GK_Q})^{-1}{\bf G} & -(I_p+{\bf GK_Q})^{-1} \\
(I_m+{\bf K_QG})^{-1} {\bf K_Q}& (I_m+{\bf K_QG})^{-1} & - (I_m+{\bf K_QG})^{-1}{\bf K_Q} \\
\end{array} \right]
\end{equation}
}
\end{defn}
\begin{prop} \label{afinecloop} \cite[(7)/~pp.101]{vidyasagar1985} ${\bf T}^{ \ell \varepsilon}_{\bf Q}$ denotes the dependency on the Youla parameter ${\bf Q}$ of the closed loop map from the exogenous signal $\varepsilon$ to the signal $\ell$ inside the feedback loop. The set of {\em all} closed loop maps (\ref{klap1}) achievable via stabilizing controllers (\ref{YoulaEq}) depends on the plant ${\bf G}$ alone and not on the particular doubly coprime factorization (\ref{bezoutidentity}) in which the Youla parameterization is formulated.
Furthermore, the parameterization of the closed loop maps (\ref{klap1}) is affine in the Youla parameter $\bf{Q}$.
\begin{equation} \label{finallly2}
\small
\begin{tabular}{|c | c| c| c|} %
\hline
${} $ & ${r}$ & ${w}$ & ${v}$ \\
\hline
& & & \\[-2.35ex]
${y} $ & $I_p- \tilde {\bf Y}_{\bf Q} \tilde {\bf M}$ & $\tilde {\bf Y}_{\bf Q} \tilde {\bf N}$ & $\tilde {\bf Y}_{\bf Q} \tilde {\bf M}$ \\ [0.75ex]
\hline
& & & \\[-2.35ex]
${ u} $ & $\tilde {\bf X}_{\bf Q} \tilde {\bf M}$ & $-\tilde {\bf X}_{\bf Q} \tilde {\bf N}$ & $-\tilde {\bf X}_{\bf Q} \tilde {\bf M}$ \\[0.75ex]
\hline
& & & \\[-2.35ex]
${z} $ & $ \tilde {\bf Y}_{\bf Q} \tilde {\bf M}$ & $- \tilde {\bf Y}_{\bf Q} \tilde {\bf N}$ &$- \tilde {\bf Y}_{\bf Q} \tilde {\bf M}$ \\[0.75ex]
\hline
& & & \\[-2.35ex]
${v} $ & $\tilde {\bf X}_{\bf Q} \tilde {\bf M}$ & $I_m -\tilde {\bf X}_{\bf Q} \tilde {\bf N}$ & $ {-\tilde {\bf X}_{\bf Q} \tilde {\bf M}}$ \\[0.75ex]
\hline
\end{tabular}\textcolor{white}{,}
\begin{tabular}{|c | c| c| c|} %
\hline
${} $ & ${r}$ & ${w}$ & ${v}$ \\
\hline
& & & \\[-2ex]
${y} $ & ${{\bf N}{\bf X}_{\bf Q}}$ & $ {{\bf N}{\bf Y}_{\bf Q}} $ & $I_p- {\bf N}{\bf X}_{\bf Q} $ \\ [0.75ex]
\hline
& & & \\[-2ex]
${ u} $ & $ {{\bf M} {\bf X}_{\bf Q}} $ & $ {- (I_m - {\bf M}{\bf Y}_{\bf Q})} $ & $- {\bf M} {\bf X}_{\bf Q} $ \\[0.75ex]
\hline
& & & \\[-2ex]
${z} $ & $ {I_p - {\bf N}{\bf X}_{\bf Q}} $ & $- {\bf N}{\bf Y}_{\bf Q} $ &$-( I_p - {\bf N}{\bf X}_{\bf Q} ) $ \\[0.75ex]
\hline
& & & \\[-2ex]
${v} $ & $ {\bf M}{\bf X}_{\bf Q} $ & $ {{\bf M}{\bf Y}_{\bf Q}} $ & $ {- {\bf M}{\bf X}_{\bf Q} } $ \\[0.75ex]
\hline
\end{tabular}
\end{equation}
\end{prop}
\subsection{Dual Youla-Ku\c{c}era Parameterization}
\begin{thm} \label{DualYoulaaa}
{\bf (Dual Youla-Ku\c{c}era)}(\cite{VanHof1992}) Let $\big({\bf M}, {\bf N}$, $\tilde {\bf M}, \tilde {\bf N}$, ${\bf X}, {\bf Y}$, $\tilde {\bf X}, \tilde {\bf Y}\big)$ be a doubly coprime factorization of ${\bf G}$. Any plant ${\bf G}_R$ stabilized by a fixed controller ${\bf K}$, can be written as
\begin{equation}
\label{YoulaEqR}
{\bf G_R} = {\bf \tilde{M}_R}^{-1}{\bf \tilde{N}_R} = {\bf N_R}{\bf M_R}^{-1},
\end{equation}
where
${\bf M_R}$, ${\bf \tilde{M}_R}$, ${\bf N_R}$ and ${\bf \tilde{N}_R}$ are defined as:
\begin{equation} \label{DualEqYoula4}
\begin{split}
{\bf M_R} \overset{def}{=} {\bf M}- \tilde{\bf X}{\bf R}, \quad
&{\bf \tilde{M}_R} \overset{def}{=} \tilde{\bf M}-{\bf RX}, \quad \\
{\bf N_R} \overset{def}{=} {\bf N} + \tilde{\bf Y}{\bf R}, \quad
&{\bf \tilde{N}_R} \overset{def}{=} \tilde{\bf N}+{\bf RY}
\end{split}
\end{equation}
for some stable ${\bf R}$ in $\mathbb{R}(z)^{p\times m}$.
\end{thm}
\begin{prop}\label{dualyoula2}(\cite{VanHof1992})
Let
$\bf{G}$ with LCF
${\bf \tilde{M}}^{-1}{\bf \tilde{N}}$ be any system that is stabilized by a controller $\bf{K}$ with LCF ${\bf Y}^{-1}{\bf X}$.
Then, the plant $\bf{G_R}$ is stabilized by controller $\bf{K}$ iff there exists a stable ${\bf R} \in \mathbb{R}(z)^{p\times m}$, such that $\bf{G_R}$ = $(\tilde{\bf M}-{\bf RX})^{-1}(\tilde{\bf N}+{\bf RY})$. Similarly, for RCF of controller ${\bf K}$ = ${\bf \tilde{X}}{\bf \tilde{Y}}^{-1}$ and plant $\bf{G}$ = ${\bf N}{\bf M}^{-1}$, the plant $\bf{G}$ is stabilized by controller $\bf{K}$ iff there exists a stable ${\bf R} \in \mathbb{R}(z)^{p\times m}$, such that $\bf{G_R}$ = $( {\bf N} + \tilde{\bf Y}{\bf R}) ({\bf M}- \tilde{\bf X}{\bf R})^{-1}$.
\end{prop}
\section{Robust Controller Synthesis}\label{robustcontrollersynthesis}
Given a DCF of the nominal model of the plant ${\bf G}^{\mathtt{md}} = (\tilde{\bf M}^{\mathtt{md}})^{-1} \tilde{\bf N}^{\mathtt{md}} = {\bf N}^{\mathtt{md}} ({\bf M}^{\mathtt{md}})^{-1}$, we can write the Bezout identity that incorporates the corresponding Youla parameterization of all stabilizing controller for the nominal model ${{\bf K}^{\mathtt{md}}_{\bf Q}} = ({\bf Y}^{\mathtt{md}}_{\bf Q})^{-1} {\bf X}^{\mathtt{md}}_{\bf Q} = {\bf {\tilde{X}}}^{\mathtt{md}}_{\bf Q} ({\bf \tilde{Y}}^{\mathtt{md}}_{\bf Q})^{-1}$ as:
\begin{equation}\label{bezout2}
\left[ \begin{array}{cc} {\tilde {\bf M}}^{\mathtt{md}} & {\tilde {\bf N}}^{\mathtt{md}} \\ - {{\bf X}}^{\mathtt{md}}_{\bf Q} & {{\bf Y}}^{\mathtt{md}}_{\bf Q} \end{array} \right]
\left[ \begin{array}{cc} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q} & -{{\bf N}}^{\mathtt{md}} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q} & {{\bf M}}^{\mathtt{md}} \end{array} \right] = \left[ \begin{array}{cc} { I_p} & {0} \\ {0} & { I_m} \end{array} \right],
\end{equation}
where {\bf Q} denotes as usually the Youla parameter.
\begin{defn}[Model Uncertainty Set] \label{modeluncertaintyset}
The $\gamma$-radius {\em model uncertainty} set (for the nominal plant ${\bf G}^{\mathtt{md}}$ \textrm{with} \;
$\Delta_{\bf \tilde{M}}$, $\Delta_{\bf \tilde{N}}$ {both stable}) is defined as:
\begin{equation} \label{Youlamd}
\mathcal{G}_\gamma \overset{\mathrm{def}}{=} \{ {\bf G} = {\tilde {\bf M}}^{-1} {\tilde {\bf N}} \hspace{2pt} \big | \hspace{2pt} {\tilde {\bf M}} = ({\tilde {\bf M}^{\mathtt{md}}}+\Delta_{\bf {\tilde{M}}}), {\tilde {\bf N}} = ({\tilde {\bf N}}^{\mathtt{md}}+\Delta_{\bf \tilde{N}}); \; \; \; \Big\| \left[ \begin{array}{cc} \Delta_{\bf \tilde{M}} & \Delta_{\bf \tilde{N}} \end{array} \right] \Big\|_{\infty} < \gamma \}
\end{equation}
\end{defn}
\begin{defn}[$\gamma$-Robustly Stabilizing]\label{perturbedplant} A stabilizing controller ${\bf K}^{\mathtt{md}}$ of the nominal plant
is said to be $\gamma$-robustly stabilizing iff ${\bf K}^{\mathtt{md}}$ stabilizes not only ${\bf G}^{\mathtt{md}}$ but also all plants ${\bf G} \in \mathcal{G}_\gamma$.
\end{defn}
\begin{assumption}
It is assumed that the true plant, denoted by ${\bf G}^{\mathtt{pt}}$, belongs to the model uncertainty set introduced in \defref{modeluncertaintyset}, i.e. that there exist stable $\Delta_{\bf \tilde{M}}$, $\Delta_{\bf \tilde{N}}$ with $ \Big\| \left[ \begin{array}{cc} \Delta_{\bf \tilde{M}} & \Delta_{\bf \tilde{N}} \end{array} \right] \Big\|_{\infty} < \gamma$ for which ${\bf G}^{\mathtt{pt}} = ({\tilde {\bf M}}^{\mathtt{md}}+\Delta_{\bf \tilde{M}})^{-1} ({\tilde {\bf N}}^{\mathtt{md}}+\Delta_{\bf \tilde{N}})$.
\end{assumption}
\noindent In the presence of additive uncertainty on the coprime factors the Bezout identity in \eqref{bezout2} no longer holds, however, the following holds for {\em certain stable} ${\Delta_{\bf M}}$, ${\Delta_{\bf N}}$ {\em factors}:
\begin{equation}\label{PhiMatrix}
\left[ \begin{array}{cc} ({\tilde {\bf M}}^{\mathtt{md}}+\Delta_{\bf \tilde{M}}) & ({\tilde {\bf N}}^{\mathtt{md}}+\Delta_{\bf \tilde{N}}) \\ - {{\bf X}}^{\mathtt{md}}_{\bf Q} & {{\bf Y}}^{\mathtt{md}}_{\bf Q} \end{array} \right]
\left[ \begin{array}{cc} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q} & -({{\bf N}}^{\mathtt{md}}+\Delta_{\bf {N}}) \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q} & ({{\bf M}}^{\mathtt{md}}+\Delta_{\bf {M}}) \end{array} \right] = \left[ \begin{array}{cc} {\bf \Phi}_{11} & O \\ O & {\bf \Phi}_{22} \end{array} \right].
\end{equation}
The block diagonal structure of the right hand side term in \eqref{PhiMatrix} is due to the fact that ${\bf G}^{\mathtt{pt}} = ({\tilde {\bf M}}^{\mathtt{md}}+\Delta_{\bf \tilde{M}})^{-1} ({\tilde {\bf N}}^{\mathtt{md}}+\Delta_{\bf \tilde{N}}) =({{\bf N}}^{\mathtt{md}}+\Delta_{\bf {N}}) ({{\bf M}}^{\mathtt{md}}+\Delta_{\bf {M}})^{-1} $ for the aforementioned {\em certain stable} ${\Delta_{\bf M}}$, ${\Delta_{\bf N}}$ {\em factors}.
\begin{lem}\label{phi11phi22}
A stabilizing controller of the nominal plant ${{\bf K}^{\mathtt{md}}_{\bf Q}} = ({\bf Y}^{\mathtt{md}}_{\bf Q})^{-1} {\bf X}^{\mathtt{md}}_{\bf Q} = {\bf {\tilde{X}}}^{\mathtt{md}}_{\bf Q} ({\bf \tilde{Y}}^{\mathtt{md}}_{\bf Q})^{-1}$ is $\gamma$-robustly stabilizing iff for any stable model perturbations $\Delta_{\bf \tilde{M}}, \Delta_{\bf \tilde{N}}$ with $ \Big\| \left[ \begin{array}{@{}cc@{}} \Delta_{\bf \tilde{M}} & \Delta_{\bf \tilde{N}} \end{array} \right] \Big\|_{\infty} < \gamma$ the TFM
\begin{equation} \label{August10_7am}
{\bf \Phi}_{11} = { I_p} + \left[ \begin{array}{cc} \Delta_{\bf \tilde{M}} & \Delta_{\bf \tilde{N}} \end{array} \right] \left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q} \end{array} \right]
\end{equation}
is unimodular ({\em i.e.} it is square, stable and has an inverse ${\bf \Phi}_{11}^{-1}$ which is also stable) from \eqref{PhiMatrix}. A similar condition for $\gamma$-robust stabilizability can be formulated in terms of ${\bf \Phi}_{22}$ TFM, whereas
\begin{equation}
{\bf \Phi}_{22} = { I_m} + \left[ \begin{array}{cc} {\bf X}^{\mathtt{md}}_{\bf Q} & {\bf Y}^{\mathtt{md}}_{\bf Q} \end{array} \right] \left[ \begin{array}{c} \Delta_{\bf {N}} \\ \Delta_{\bf {M}} \end{array} \right].
\end{equation}
\end{lem}
Since $ {\bf \Phi}_{11}$ in \eqref{August10_7am} clearly depends on the Youla parameter (via the right coprime factors of the controller), the condition for the $\gamma$-robust stabilizability of the controller can be recast in the following particular form, which will be instrumental in the sequel:
\begin{thm}\label{oarasthm1} The Youla parameterization yields a $\gamma$-robustly stabilizing controller iff its corresponding Youla parameter satisfies
$\Bigg\| \left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q} \end{array} \right] \Bigg\|_{\infty} \leq \dfrac{1}{\gamma}$
\end{thm}
\noindent
The proofs for \lemref{phi11phi22} and \thmref{oarasthm1} are given on \hyperref[appendixC]{Appendix C}.
As an intermediary result, by employing \thmref{oarasthm1} and the standard inequality \eqref{submultiplicativeinfty} from \hyperref[appendixB]{Appendix B} it is concluded that:
\begin{center}
$\Bigg\| \left[ \begin{array}{cc} -\Delta_{\bf \tilde{M}} & -\Delta_{\bf \tilde{N}} \end{array} \right]
\left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q} \end{array} \right] \Bigg\|_{\infty}\leq \Bigg\| \left[ \begin{array}{cc} \Delta_{\bf \tilde{M}} & \Delta_{\bf \tilde{N}} \end{array} \right] \Bigg\|_{\infty} \Bigg\|
\left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q} \end{array} \right] \Bigg\|_{\infty} < \gamma \times \dfrac{1}{
\gamma} = 1$.
\end{center}
\noindent
\noindent
\begin{prop}\label{closedloopresponses11}
The square root of the LQG cost function from \eqref{LQGCostOriginal} has the form:
\begin{equation}\label{cost2}
\begin{split}
\mathcal{H}({\bf G}^{\mathtt{pt}}, {{\bf K}^{\mathtt{md}}_{\bf Q}}) \overset{def}{=} & \Bigg\| \left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q} \end{array} \right] {\bf \Phi}_{11}^{-1} \left[ \begin{array}{cc} ({\tilde {\bf M}^{\mathtt{md}}}+\Delta_{\bf \tilde{M}}) & ({\tilde {\bf N}^{\mathtt{md}}}+\Delta_{\bf \tilde{N}}) \end{array} \right] \Bigg\|_{\mathcal{H}_{2}}\\
\end{split}
\end{equation}
representing the following closed loop responses:
\begin{equation}\label{closedloopresponses1}
\left[ \begin{array}{cc} {{\bf T}^{y \nu}_{\bf Q}} & {{\bf T}^{yw}_{\bf Q}} \\ {{\bf T}^{u \nu}_{\bf Q}} & {{\bf T}^{u w}_{\bf Q}} \end{array} \right]
= \left[ \begin{array}{cc} {(I_p+{\bf G}^{\mathtt{pt}} {{\bf K}^{\mathtt{md}}_{\bf Q}})^{-1}} & { {(I_p+{\bf G}^{\mathtt{pt}} {{\bf K}^{\mathtt{md}}_{\bf Q}})^{-1}}{{\bf G}^{\mathtt{pt}}}} \\
{{{\bf K}^{\mathtt{md}}_{\bf Q}}{(I_p+{\bf G}^{\mathtt{pt}} {{\bf K}^{\mathtt{md}}_{\bf Q}})^{-1}}} & {{{\bf K}^{\mathtt{md}}_{\bf Q}}{(I_p+{\bf G}^{\mathtt{pt}} {{\bf K}^{\mathtt{md}}_{\bf Q}})^{-1}}}{{\bf G}^{\mathtt{pt}}} \end{array} \right]
\end{equation}
\end{prop}
\noindent
The proofs for the \propref{closedloopresponses11} and \thmref{theoremAcost} are given on \hyperref[appendixC]{Appendix C}.
\begin{thm}\label{theoremAcost}
The Robust LQG Control Problem is defined as :
\begin{equation} \label{theoremA}
\begin{aligned}
\min_{\bf Q\: \text{stable}} \quad & \max_{\Big\| \left[ \begin{array}{cc} \Delta_{\bf \tilde{M}} & \Delta_{\bf \tilde{N}} \end{array} \right] \Big\|_\infty < \gamma } \mathcal{H}({\bf G}^{\mathtt{pt}}, {{\bf K}^{\mathtt{md}}_{\bf Q}}) \\[5pt]
\textrm{s.t.} \quad
\quad & \Bigg\|
\left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q} \end{array} \right] \Bigg\|_{\infty} \leq \dfrac{1}{
\gamma}.
\end{aligned}
\end{equation}
whose solution, obtained for the optimal Youla parameter ${\bf Q_*}$ in \eqref{theoremA} will be denoted by ${\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q_*}$, ${\tilde {\bf X}}^{\mathtt{md}}_{\bf Q_*}$ such that the optimal, robust controller reads ${{\bf K}^{\mathtt{md}}_{\bf Q_*}} = {\bf {\tilde{X}}}^{\mathtt{md}}_{\bf Q_*} ({\bf \tilde{Y}}^{\mathtt{md}}_{\bf Q_*})^{-1}$.
\end{thm}
\noindent
Canonical min-max formulation caused by the additive uncertainty on the coprime factors of the robust LQG controller synthesis renders the problem non-convex.
In order to circumvent this, an upper bound on the $\mathcal{H}({\bf G}^{\mathtt{pt}}, {{\bf K}^{\mathtt{md}}_{\bf Q}})$ cost functional will be derived, much in the spirit of \cite{Dean2018} and \cite{furieri2020}. This bound will further be exploited to derive a quasi-convex approximation for the robust LQG control problem in the next subsection.
\subsection{Quasi-convex formulation}
\begin{prop}\label{LQGcostupperbound2}
Given any $\gamma$-robustly stabilizing controller satisfying $\Bigg\|
\left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q} \end{array} \right] \Bigg\|_{\infty} < \dfrac{1}{
\gamma}$ then for any additive model perturbations $\Big\| \left[ \begin{array}{cc} \Delta_{\bf \tilde{M}} & \Delta_{\bf \tilde{N}} \end{array} \right] \Big\|_\infty < \gamma$,
the cost functional of the robust LQG problem from \eqref{theoremA} admits the upper bound:
\begin{equation} \label{LQGcostupperbound3}
\mathcal{H}({\bf G}^{\mathtt{pt}}, {{\bf K}^{\mathtt{md}}_{\bf Q}}) \leq \dfrac{1}{1 - \gamma \bigg\| \left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q} \end{array} \right] \bigg\|_{\infty}} {\Bigg[ {h \Big( \gamma, \bigg\| \left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q} \end{array} \right] \bigg\|_{\infty} \Big)} \Bigg\| \left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q} \end{array} \right] \bigg\|_{\mathcal{H}_{2}} \Bigg] }
\end{equation}
\noindent
where ${h \bigg( \gamma, \bigg\| \left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q} \end{array} \right] \bigg\|_{\infty} \bigg)} \overset{def}{=} \bigg(1 + {\gamma \bigg\| \left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q} \end{array} \right] \bigg\|_{\infty} }\bigg) \bigg( \bigg\| \left[ \begin{array}{cc} {\tilde {\bf M}^{\mathtt{md}}} & {\tilde {\bf N}^{\mathtt{md}}} \end{array} \right] \bigg\|_{\infty} + \gamma \bigg)$.
\end{prop}
\noindent
Next one of the main results is stated that provides an aptly designed approximation of the robust LQG control problem from Definition~\ref{theoremAcost} by means of the LQG cost upper bound from \eqref{LQGcostupperbound3}. The proofs for \propref{LQGcostupperbound2} and \thmref{theoremB} are given in \hyperref[appendixE]{Appendix E}.
\begin{thm}\label{theoremB}
For the true plant, ${\bf G}^{\mathtt{pt}} \in \mathcal{G}_\gamma$
and $(\forall) \alpha >0$, the robust LQG control problem in \eqref{theoremA} admits the following upper bound:
\begin{equation} \label{theoremB2}
\begin{aligned}
\min_{\delta \in [0,1/\gamma)} & \dfrac{1}{1 - \gamma \delta} \min_{\bf Q\: \text{stable}} {\bigg( {h \big(\gamma, \alpha\big)} \Bigg\| \left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q} \end{array} \right] \bigg\|_{\mathcal{H}_{2}} \bigg)}\\ \vspace{5pt}
\textrm{s.t.} & \Bigg\| \left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q}\end{array} \right] \Bigg\|_{\infty} \leq \min \{\delta, \alpha\},
\end{aligned}
\end{equation}
\noindent where as before, ${h \big( \gamma, \alpha \big)}$ = $ (1 + \gamma \alpha) \bigg( \bigg\| \left[ \begin{array}{cc} {\tilde {\bf M}^{\mathtt{md}}} & {\tilde {\bf N}^{\mathtt{md}}} \end{array} \right] \bigg\|_{\infty} + \gamma\bigg) $.\\
\end{thm}
\begin{rem} \label{numerical} (Numerical Computation) For each fixed $\delta$, the inner optimization is convex but the dimension of ${\bf Q}(z)$ remains infinite.For numerical computation, a (FIR) truncation is considered on ${\bf Q}(z)$, such that an equivalent Semi-Definite Program (SDP) can be formulated for the inner optimization problem. Details of this is discussed in \hyperref[appendixE]{Appendix E}.
\end{rem}
\begin{rem} (Feasibility)
Since the subsequent quasi-convex optimization problem in Theorem-\ref{theoremB} must include an additional constraint $\bigg\| \left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q} \end{array} \right] \bigg\|_{\infty} < \alpha$ proportional to $\alpha$, this $\alpha$ should be chosen as small as possible.
However, $\alpha$ can't be made arbitrarily small and therefore the feasibility of the quasi-convex program from Theorem-\ref{theoremB} cannot be guaranteed. This is caused by the fact that $\alpha$ must be assimilated to the norm of a LCF of a stabilizing controller for {\em the true plant}, whose $H_\infty$ attenuation is simultaneously greater or equal to the one of the optimal $H_\infty$ controller {\em for the nominal model} (see Theorem~\ref{oarasthm1}). This reflects the fact that the feasibility of the closed loop "learning" problem depends inherently on the performance (with respect to the model uncertainty of the plant) of the initially chosen stabilizing controller (the one with which the closed loop "learning" is being performed).
\end{rem}
\section{Analysis of End-to-End Performance}\label{endtoendanalysis}
\subsection{Sub-optimality guarantee}
\noindent
If we denote by ${{\bf K}^{\mathtt{opt}}}$ the optimal $\mathcal{H}_2$ controller for the true plant, then by Assumption~1 there exist stable additive factors such that ${\bf G}^{\mathtt{pt}}$ = $(\tilde{\bf M}^{\mathtt{md}} + \Delta_{\bf \tilde{M}})^{-1} (\tilde{\bf N}^{\mathtt{md}} + \Delta_{\bf \tilde{N}})$ = $({\bf N}^{\mathtt{md}}+\Delta_{\bf {N}})({\bf M}^{\mathtt{md}}+\Delta_{\bf {M}})^{-1}$ and furthermore, there always exists a Bezout identity of the true plant that features the optimal controller ${{\bf K}^{\mathtt{opt}}} = ({\bf Y}^{\mathtt{opt}})^{-1} {\bf X}^{\mathtt{opt}} = {\bf {\tilde{X}}}^{\mathtt{opt}} ({\bf \tilde{Y}}^{\mathtt{opt}})^{-1}$ as its ``central controller'', thus reading:
\begin{equation}\label{RealPlantOptimalController1}
\left[ \begin{array}{cc} ({\tilde {\bf M}}^{\mathtt{md}}+\Delta_{\bf \tilde{M}}) & ({\tilde {\bf N}}^{\mathtt{md}}+\Delta_{\bf \tilde{N}}) \\ - {{\bf X}}^{\mathtt{opt}} & {{\bf Y}}^{\mathtt{opt}} \end{array} \right]
\left[ \begin{array}{cc} {\tilde {\bf Y}}^{\mathtt{opt}} & -({{\bf N}}^{\mathtt{pt}}+\Delta_{\bf {N}}) \\ {\tilde {\bf X}}^{\mathtt{opt}} & ({{\bf M}}^{\mathtt{pt}}+\Delta_{\bf {M}}) \end{array} \right] = \left[ \begin{array}{cc} { I_p} & {{0}} \\ {0} & {I_m} \end{array} \right]
\end{equation}
\noindent
\noindent
Consequently, the square root of the LQG cost functional for optimal controller is given by:
\begin{equation}\label{cost3}
\mathcal{H}({\bf G}^{\mathtt{pt}}, {\bf K}^{\mathtt{opt}}) \overset{def}{=} \Bigg\| \left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{opt}} \\ {\tilde {\bf X}}^{\mathtt{opt}} \end{array} \right] \left[ \begin{array}{cc} ({\tilde {\bf M}}^{\mathtt{md}}+\Delta_{\bf \tilde{M}}) & ({\tilde {\bf N}}^{\mathtt{md}}+\Delta_{\bf \tilde{N}}) \end{array} \right] \Bigg\|_{\mathcal{H}_{2}}
\end{equation}
\noindent
Next the main result on the sub-optimality guarantee for the performance of the robust controller with model uncertainty of radius $\gamma$ is stated. The proof for \thmref{suboptimalitygurantee} is given in \hyperref[appendixF]{Appendix F}.
\begin{thm}\label{suboptimalitygurantee}
Let ${{\bf K}^{\mathtt{opt}}}$ be the optimal LQG controller and ${{\bf G}^{\mathtt{pt}}}$ be the model of the true plant, with modeling error uncertainty satisfying ${\Big\| \left[ \begin{array}{cc} \Delta_{\bf \tilde{M}} & \Delta_{\bf \tilde{N}} \end{array} \right] \Big\|_\infty < \gamma }$. Furthermore, let ${\bf Q_*}$ and $\delta_*$ denote the solution to \eqref{theoremB2}.
Then, when applying the resulting controller ${{\bf K}^{\mathtt{md}}_{\bf Q_*}}$ in feedback interconnection with the true plant ${{\bf G}^{\mathtt{pt}}}$, the relative error in the LQG cost is upper bounded by:
\begin{equation}\label{relativeerror}
\dfrac{ \mathcal{H}({\bf G}^{\mathtt{pt}}, {{\bf K}^{\mathtt{md}}_{\bf Q_*}})^2 - \mathcal{H}({\bf G}^{\mathtt{pt}}, {\bf K}^{\mathtt{opt}})^2 }{ \mathcal{H}({\bf G}^{\mathtt{pt}}, {\bf K}^{\mathtt{opt}})^2 } \leq \dfrac{1}{\Big(1-\gamma \bigg\| \left[ \begin{array}{c} {\bf \tilde{Y}}^{\mathtt{md}}_{\bf Q_*} \\ {\bf \tilde{X}}^{\mathtt{md}}_{\bf Q_*} \end{array} \right] \bigg\|_{\infty} \Big)^2} \times {g \bigg( \gamma, \bigg\| \left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q_*} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q_*} \end{array} \right] \bigg\|_{\infty} \bigg)}^2 -1,
\end{equation}
\begin{equation*}
\textrm{where} \hspace{4pt}
\small {g \bigg( \gamma, \bigg\| \left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q_*} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q_*} \end{array} \right] \bigg\|_{\infty} \bigg)} \overset{def}{=} \bigg\| \left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q_*} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q_*} \end{array} \right] \bigg\|_{\infty} \bigg(1 + {\gamma \bigg\| \left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q_*} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q_*} \end{array} \right] \bigg\|_{\infty} }\bigg) \bigg( \bigg\| \left[ \begin{array}{cc} {\tilde {\bf M}^{\mathtt{md}}} & {\tilde {\bf N}^{\mathtt{md}}} \end{array} \right] \bigg\|_{\infty} + \gamma \bigg) .
\end{equation*}
\end{thm}
\noindent
\begin{rem} (Optimality vs Robustness) If $\gamma \bigg\| \left[ \begin{array}{c} {\bf \tilde{Y}}^{\mathtt{md}}_{\bf Q_*} \\ {\bf \tilde{X}}^{\mathtt{md}}_{\bf Q_*} \end{array} \right] \bigg\|_{\infty} = \eta$, it's easy to observe that $\eta \in (0,1)$. Then it's immediate to see that the upper bound of the relative error in the LQG cost increases as a function of $\eta$. The price of obtaining a faster rate
is that the controller becomes less robust to model uncertainty as pointed out in \cite{mania2019}, \cite{furieri2020}. It holds for this case too as shown in \thmref{suboptimalitygurantee}. In practice, using a relatively large value for $\eta$ forces a trade-off of optimality for robustness in the controller design procedure. In general, optimality stands i.e. better controller performance is guaranteed as $\eta$ goes closer to $0$ and better robustness performance is guaranteed as $\eta$ goes closer to $1$ with the upper bound \eqref{relativeerror} of relative error in LQG cost might be large. This is shown with an example below.
\noindent
Let's set $\eta = \dfrac{1}{5}$.
Then by \thmref{suboptimalitygurantee} relative error in the LQG cost is
\begin{equation}
\dfrac{ \mathcal{H}({\bf G}^{\mathtt{pt}}, {{\bf K}^{\mathtt{md}}_{\bf Q_*}})^2 - \mathcal{H}({\bf G}^{\mathtt{pt}}, {\bf K}^{\mathtt{opt}})^2 }{ \mathcal{H}({\bf G}^{\mathtt{pt}}, {\bf K}^{\mathtt{opt}})^2 } \leq 2 \times {g \bigg( \gamma, \bigg\| \left[ \begin{array}{c} {\tilde {\bf Y}}^{\mathtt{md}}_{\bf Q_*} \\ {\tilde {\bf X}}^{\mathtt{md}}_{\bf Q_*} \end{array} \right] \bigg\|_{\infty} \bigg)} \bigg]^2 -1.
\end{equation}
Hence, the the relative error in the LQG cost grows as $\mathcal{O}(\gamma^2)$ as long as $\gamma \bigg\| \left[ \begin{array}{c} {\bf \tilde{Y}}^{\mathtt{md}}_{\bf Q_*} \\ {\bf \tilde{X}}^{\mathtt{md}}_{\bf Q_*} \end{array} \right] \bigg\|_{\infty} < \dfrac{1}{5}$.
\end{rem}
\section{Closed Loop Identification Scheme} \label{clsysid
\noindent
Figure~2 at the top of next page depicts the closed-loop identification setup of a potentially unstable {\em noise contaminated plant} $\textbf{G}^\mathtt{md}$ with control input $u$, noise $\nu$ (taken $w$ = $0$) and output measurement $y$ (where $u$ and $\nu$ are assumed independent and stationary), provided that some initial stabilizing controller ${{\bf K}^{\mathtt{md}}}$ is available beforehand.
\begin{figure}[ht]
\centering
\includegraphics[width=15.6cm]{NoiseContaminatedPlant4.pdf}
\caption{Closed loop Identification for Noise Contaminated Plant}
\label{clsysid2}
\end{figure}
\noindent
If any plant $\mathbf{G}^\mathtt{pt}$ is LTI and stable, for open-loop identification it can be written that:
\begin{equation}
y = \textbf{G}^\mathtt{pt} u + \nu
\end{equation}
\noindent
Here $u$ and $y$ are available through measurements, then the model of the plant can be estimated as $\mathbf{G}^\mathtt{md}$ (\cite{Anderson1998}).
\noindent
Suppose that the controller ${{\bf K}^{\mathtt{md}}} = ({\bf Y}^{\mathtt{md}})^{-1} {\bf X}^{\mathtt{md}} = {\bf {\tilde{X}}}^{\mathtt{md}} ({\bf \tilde{Y}}^{\mathtt{md}})^{-1}$, where ${\bf X}^{\mathtt{md}}$, ${\bf Y}^{\mathtt{md}}$ are left coprimes and ${\bf \tilde{X}}^{\mathtt{md}}$, ${\bf \tilde{Y}}^{\mathtt{md}}$ are right coprimes of ${{\bf K}^{\mathtt{md}}}$. Then by the coprimeness there exists $ {{\bf M}}^{\mathtt{md}}$, $ { {\bf N}}^{\mathtt{md}}$, $ {\tilde {\bf M}}^{\mathtt{md}}$, $ {\tilde {\bf N}}^{\mathtt{md}}$ with
\begin{equation}\label{coprime1}
{\bf {X}}^{\mathtt{md}}{\bf {N}}^{\mathtt{md}} + {\bf {Y}}^{\mathtt{md}}{\bf {M}}^{\mathtt{md}} = {I}_m, \hspace{2pt} {\bf \tilde{M}}^{\mathtt{md}}{\bf \tilde{Y}}^{\mathtt{md}} + {\bf \tilde{N}}^{\mathtt{md}}{\bf \tilde{X}}^{\mathtt{md}} = {I}_p.
\end{equation}
\noindent
This means that ${\bf G}^{\mathtt{md}}$ = ${\bf {N}}^{\mathtt{md}} {({\bf {M}}^{\mathtt{md}})}^{-1}$ = ${({\bf \tilde{M}}^{\mathtt{md}})}^{-1}{\bf \tilde{N}}^{\mathtt{md}}$ is the nominal model stabilized by ${\bf K}^{\mathtt{md}}$. Alternatively, it can be started with a RCF representation ${\bf {N}}^{\mathtt{md}} ({\bf {M}}^{\mathtt{md}})^{-1}$ of the nominal model ${\bf G}^{\mathtt{md}}$ and then choose that particular representation ${\bf {\tilde{X}}}^{\mathtt{md}} ({\bf \tilde{Y}}^{\mathtt{md}})^{-1}$ of the known stabilizing ${\bf K}^{\mathtt{md}}$ so that equation \eqref{coprime1} holds. The set of all plants ${\bf G}_{\bf R}^{\mathtt{md}}$ stabilized by ${\bf K}^{\mathtt{md}}$ is given by Dual Youla-Ku\c{c}era Parameterization in \propref{dualyoula2}:
\begin{equation}\label{coprime2}
{\bf G}_{\bf R}^{\mathtt{md}} = {({\bf {N}}^{\mathtt{md}}+{\bf {\tilde{Y}}}^{\mathtt{md}}\mathbf{R}^\mathtt{md})({\bf {M}}^{\mathtt{md}} -{\bf {\tilde{X}}}^{\mathtt{md}}\mathbf{R}^\mathtt{md})^{-1}}
\end{equation}
for some stable $\mathbf{R}^\mathtt{md} \in \mathbb{R}^{p \times m}$ and ${\bf {\tilde{X}}}^{\mathtt{md}} ({\bf \tilde{Y}}^{\mathtt{md}})^{-1}$ is the right coprime factorization of ${\bf K}^{\mathtt{md}}$.
\noindent
Suppose that ${({\bf \tilde{M}}^{\mathtt{md}})}^{-1}{\bf \tilde{N}}^{\mathtt{md}}$ is a left coprime fractional description of the nominal model ${\bf G}^{\mathtt{md}}$ such that
\begin{equation}
\begin{bmatrix} {\mathbf{Y}}^{\mathtt{md}} & {\mathbf X}^{\mathtt{md}}\\ -{\mathbf {\tilde{N}}}^{\mathtt{md}} & {\mathbf {\tilde{M}}}^{\mathtt{md}} \end{bmatrix}
\begin{bmatrix} {\mathbf{M}}^{\mathtt{md}} & - {\mathbf{\tilde{X}}}^{\mathtt{md}} \\ {\mathbf {N}}^{\mathtt{md}} & {\mathbf {\tilde{Y}}}^{\mathtt{md}} \end{bmatrix} = \begin{bmatrix} {I}_m & 0 \\ 0 & {I}_p \end{bmatrix}
\end{equation}
\noindent
\begin{lem}\label{noisecontaminatedplant}
The noise contaminated plant set-up of \figref{clsysid2} is identical to \figref{2Block} with input $u$, noise $\nu$ and output $y$, it implies
\begin{equation}\label{GwithR}
y = {\bf G}_{\bf R}^{\mathtt{md}} u + \nu
\end{equation}
\noindent
where ${\bf G}_{\bf R}^{\mathtt{md}}$ is given by \eqref{coprime2}.
\end{lem}
\noindent
The key idea dating back to \cite{Anderson1998} is to identify the {\em stable} dual-Youla parameter ${\bf R}^{\mathtt{md}}$ from Theorem~\ref{DualYoulaaa} rather than ${\bf G}^{\mathtt{md}}$, thus recasting the problem in a standard, open-loop identification form. More specifically, direct inspection of Figure~2 shows that
\begin{equation}\label{zblock2}
e_2 = {\bf R}^{\mathtt{md}} e_1 + ({\bf \tilde{M}}^{\mathtt{md}} - {\bf R}^{\mathtt{md}}{\bf X}^{\mathtt{md}})\nu = {\bf R}^{\mathtt{md}}(e_1 - {\bf X}^{\mathtt{md}} \nu ) + {\bf \tilde{M}}^{\mathtt{md}}\nu.
\end{equation}
In \eqref{zblock2} we have the knowledge of ${\bf X}^{\mathtt{md}}$; $e_1$ and $e_2$ are available from measurements with noise $\nu$. The recent algorithm from \cite{Dahleh2020} given below is employed toward identifying the dual-Youla parameter.\\
Details of this idea and proof for \lemref{noisecontaminatedplant} is given in \hyperref[appendixG1]{Appendix G.1}, also choice of noise is discussed in \hyperref[appendixG1]{Appendix G.2}.
\subsection{Identification Algorithms}
\noindent
Identification algorithms from \cite{Dahleh2020} has been employed here. Preliminaries of system identification is given in \hyperref[appendixG]{Appendix G.3}.
The state space representation of ${\bf R}^{\mathtt{md}}$ in $e_2 = {\bf R}^{\mathtt{md}} u + {\bf \tilde{M}}^{\mathtt{md}}\nu$ with $u$ = $e_1 - {\bf X}^{\mathtt{md}} \nu$ is:
\begin{equation}\label{system}
\begin{aligned}
l_{t+1} &= A_R l_t + B_R u_t + \eta_{t+1}\\
z_t &= C_R l_t + {\bf \tilde{M}}^{\mathtt{md}} \nu_t
\end{aligned}
\end{equation}
\noindent
Here two important assumptions is stated as below:
\begin{assumption} \label{noiseassumption}
The noise process $\{\eta_{t}\}_{t=1}^{\infty}$ in the dynamics of ${\bf R}^{\mathtt{md}}$ are i.i.d., and isotropic with sub-gaussian parameter 1.
The noise process $\{r_{t}\}_{t=1}^{\infty}$, $\{w_{t}\}_{t=1}^{\infty}$ and $\{\nu_{t}\}_{t=1}^{\infty}$ are Gaussian processes with mean function $m_r(t) = m_w(t) = m_{\nu}(t)$ $0$, and their spectral density ${\bf \phi}_r(\omega)$, ${\bf \phi}_w(\omega)$ and ${\bf \phi}_{\nu}(\omega)$ satisfy the constraint in \hyperref[appendixG]{Appendix G}.
\end{assumption}
\begin{assumption} \label{existenceofR}
There exists constants $\beta, R \geq 1$ s.t.
$\|\mathcal{T}_{0,\infty}\|_2 < \beta$ and $\dfrac{\|\mathcal{TO}_{k,d}\|_2}{\|\mathcal{T}_{0,\infty}\|_2} \leq \mathcal{R}$.\\
$\beta$ exists since ${\bf R}^{\mathtt{md}}$ is stable.
\end{assumption}
\noindent
The Algorithms
for system identification is illustrated in \hyperref[appendixG]{Appendix G}.
\noindent
\subsubsection{Probabilistic guarantees} \label{probabilisticgurantee}
\noindent
Let's define, $T_*(\delta)$ = $\inf \{ T | d_*(T, \delta) \in \mathcal{D}(T), d_*(T,\delta) \leq 2 d_*(\frac{T}{256}, \delta) \}$
where, $d_*(T,\delta)$ = $\inf \{ d|16 \beta \mathcal{R} \alpha(d) $ $ \geq \Big\| \mathcal{\hat{H}}_{0,\hat{d},\hat{d}} - \mathcal{\hat{H}}_{0,\infty,\infty}\Big\|_2 \}$,
with $\mathcal{\hat{H}}_{p,q,r}$ is the $(p, q, r)$ - dimensional estimated Hankel matrix.
Whenever $T \geq T_*(\delta)$ for the failure probability $\delta$, then it follows with the probability at least $(1-\delta)$ that
\begin{equation}\label{probabilitybound}
\Big\| \mathcal{\hat{H}}_{0,\hat{d},\hat{d}} - \mathcal{\hat{H}}_{0,\infty,\infty}\Big\|_2 \leq 12c\beta \mathcal{R} \bigg(\sqrt{\dfrac{m\hat{d}+p\hat{d}^2+ \hat{d}log(T/\delta)}{T}}\bigg) .
\end{equation}
\subsection{Sample Complexity}
\begin{lem}\label{HRbound} The norm of the identification error incurred by the proposed scheme after Algorithm-\ref{identificationalgorithm} and \ref{identificationalgorithm2} is bounded by:
$\Big\| {\bf{R}}^\mathtt{md} - {\bf R}^\mathtt{pt} \Big\|_{\infty} \leq \Big\| \mathcal{\hat{H}}_{0,\hat{d},\hat{d}} - \mathcal{\hat{H}}_{0,\infty,\infty}\Big\|_2 \leq 12c\beta \mathcal{R} \bigg(\sqrt{\dfrac{m\hat{d}+p\hat{d}^2+ \hat{d}log(T/\delta)}{T}}\bigg) $
\end{lem}
\noindent
Finally, the error on the model uncertainty can be directly checked using \lemref{HRbound} as:\\
\noindent
\begin{equation*}
\small
\begin{aligned}
\Big\| \left[ \begin{array}{cc} \Delta_{\bf \Tilde{M}} & \Delta_{\bf \Tilde{N}} \end{array} \right]\Big\|_{\infty} & = \Big\| \left[ \begin{array}{cc} (\Tilde{\bf M}^\mathtt{md} + {\bf X}^\mathtt{md} {\bf R}^\mathtt{md}) & ({\Tilde{\bf N}}^\mathtt{md} - {\bf Y}^\mathtt{md} {\bf R}^\mathtt{md}) \end{array} \right] - \left[ \begin{array}{cc} (\Tilde{\bf M}^{\mathtt{md}} + {\bf X}^{\mathtt{md}} {\bf R}^{\mathtt{pt}}) & ({\Tilde{\bf N}}^{\mathtt{md}} - {\bf Y}^{\mathtt{md}} {\bf R}^{\mathtt{pt}}) \end{array} \right] \Big\|_{\infty} \\ & = \Big\| \left[ \begin{array}{cc} {\bf X}^{\mathtt{md}} & {\bf Y}^{\mathtt{md}} \end{array} \right] ({\bf R}^\mathtt{md} - {\bf R}^\mathtt{pt}) \Big\|_{\infty} \leq \Big\|\left[ \begin{array}{cc} {\bf X}^\mathtt{md} & {\bf Y}^\mathtt{md} \end{array} \right] \Big\|_{\infty} \Big\| {\bf R}^\mathtt{md} - {\bf{R}}^\mathtt{pt}\Big\|_{\infty}
\end{aligned}
\end{equation*}
\noindent
\noindent
Consequently, the uncertainty level on the LCF of the model satisfies $\Big\| \left[ \begin{array}{cc} \Delta_{\bf \Tilde{M}} & \Delta_{\bf \Tilde{N}} \end{array} \right]\Big\|_{\infty} < \gamma $.
\begin{thm}\label{finaltheorem}
Define ${s}$ = $144 \Big\|\left[ \begin{array}{cc} {\bf X}^\mathtt{md} & {\bf Y}^\mathtt{md} \end{array} \right] \Big\|_{\infty}^2 c^2 \beta^2 \mathcal{R}^2$.
Then, the robust controller will achieve the relative cost within the bound with probability $(1 - \delta)$ provided $T \geq \max\{T_s, T_*({\delta}) \}$.
Here, $T_s$ is the right most zero of $g(T)$ = $\gamma^2 T - s\hat{d}log(T/\delta) - s(m\hat{d}+p\hat{d}^2)$.
\noindent
If $g(T)$ doesn't have any zero for $T>0$, then define $T_s = 0$ and
$T_*(\delta)$ = $\inf \{ T | d_*(T, \delta) \in \mathcal{D}(T), d_*(T,\delta) \leq 2d_* (\frac{T}{256}, \delta) \}$,
\noindent
where, $d_*(T,\delta)$ = $\inf \{ d | 16 \beta \mathcal{R} \alpha(d) \geq \Big\| \mathcal{\hat{H}}_{0,d,d} - \mathcal{\hat{H}}_{0,\infty,\infty}\Big\|_2 \}$,
\noindent
$\mathcal{D}(T) = \{ d \in \mathbb{N} | d \leq \dfrac{T}{cm^2 log^3(Tm/\delta)} \} $ and $\alpha(h) = \sqrt{h}.\big(\sqrt{\dfrac{m+hp+log(T/\delta)}{T}}\big)$.
\end{thm}
\noindent
Combining \thmref{finaltheorem} with \thmref{suboptimalitygurantee}, it follows that with high probability the suboptimality gap behaves as
\begin{equation*}
\dfrac{ \mathcal{H}({\bf G}^{\mathtt{pt}}, {{\bf K}^{\mathtt{md}}_{\bf Q_*}})^2 - \mathcal{H}({\bf G}^{\mathtt{pt}}, {\bf K}^{\mathtt{opt}})^2 }{ \mathcal{H}({\bf G}^{\mathtt{pt}}, {\bf K}^{\mathtt{opt}})^2 } \sim \mathcal{O}\Big(\sqrt{\dfrac{logT}{T}}\Big)
\end{equation*}
\noindent
Finally, we note here that the resulted sample complexity is on par with the existing methods from \cite{furieri2020} and \cite{Dean2018}.
\section{Conclusion and Future work}\label{conclusion}
\noindent
In this paper, we have provided the sample complexity bounds for a robust controller synthesis procedure for LQG problems with unknown dynamics, able to cope with unstable plants. We combined finite-time, non-parametric LTI system identification with the Youla parameterization for robust stabilization under uncertainty on the coprime factors of the plant.
One exciting avenue for future research is the online learning LQG control problem under the same type of model uncertainty. Another direction is to work out the sample complexity of learning the optimal state feedback (LQR) controller in tandem with the optimal state-observer (Kalman Filter (\cite{matni2020})) for a potentially unstable system. Combining these two results, should yield precisely the optimal LQG controller discussed above and reveal the {\em separation principle} within this framework.
\vskip 0.2in
|
\section{Introduction}
\label{Sec:intro}
Cancer is a generic definition of a disease that, among
its typical features, is driven by dynamic alterations in the
genome~\cite{hanahan2011hallmarks}.
These microscopic changes not only give
birth to a variety of different
types of cancer at the macroscopic scale, but can also lead to heterogeneity within the same
cancer tissue: tumour phenotypes undergo clonal expansion and
genetic diversification, promoting natural selection mechanisms that
favor cell clones
with advantageous characteristics~\cite{tabassum2015tumorigenesis, gerlinger2010darwinian}.
Alterations in the DNA of the cell, such as inclusions of copy number
aberrations and point mutations, occur early during the neoplastic
transformation and usually before any possible clinical
detection~\cite{sottoriva2015big}. The step-wise accumulation of
driver mutations may confer survival advantages in relation to the
particular environment in which they are embedded and may be accelerated by
so-called selective sweeps~\cite{rubben2017cancer}. Furthermore,
although the immune system
routinely recognises and kills any dangerous host including cancer,
mutations can provide cancer cells with the ability to avoid detection or
immuno-suppress the environment, advantaging tumour progression or
preventing eradication~\cite{hanahan2011hallmarks}.
Processes involving mutations, cell
growth and immune surveillance cumulatively result in the emergence of different
cancer populations integrated in an environment made up of healthy tissue,
immune cells and stroma~\cite{pacheco2014ecology, hillen2014mathematical}.
Understanding how these complex interactions shape and influence each
other is one of the greatest challenges in current medical
biosciences. For example, morphology is known to be
strongly sensitive to tumour adaptation to the environment (e.g. the lack of
nutrients, oxygen, space) and by the combined action of immune response
and existing anti-cancer therapies~\cite{anderson2006tumor} such as
chemotherapy, radiotherapy, immune-boosting and so on. In the last quarter of century, a number of diverse
contributions have been proposed from the biomathematical community to
shed light on some of these complex interaction mechanisms.
Several
mathematical models have been advanced using the framework of population dynamics,
with tumour immune interactions considered, for example, in Ref.~\cite{de2006mixed,
wilson2012mathematical} and cancer mutations in
Ref.~\cite{asatryan2016evolution}. Other works have involved a discrete
Cellular Potts approach~\cite{sottoriva2011modeling} or different
degrees of hybrid
modelling~\cite{anderson2008integrative, jeon2010off}, with particular
focus on tumour
shape~\cite{anderson2006tumor,jiao2012diversity}. The effects of
some of the currently available anti-tumour therapies have also been analysed in the context of
evolutionary dynamics~\cite{gatenby2009adaptive, gatenby2009lessons},
with immunotherapy~\cite{eladdadi2014modeling, frascoli2014dynamical,
piretto2018combination, bunimovich2008mathematical} and, recently, using agent-based modelling in the context of virotherapies~\cite{jenner2018modelling}. A number of reviews
detailing the evolution and the contribution of these and
other models also exist in the
literature~\cite{bellomo2008, wilkie2013review,
eftimie2011interactions, eladdadi2014mathematical}.
The focus of the present work is on metastatic secondary solid
lesions, with particular emphasis on the role of the immune system and
mutations. Scope of the this work is the study of the effects of
different combination therapies on secondary lesions in order to better
understand the dynamics involved and the role of mutations on
treatments' effectiveness. The rest of the paper is organised as
follows. In the
``\nameref{Sec:model}'' section, a description of the mathematical approach used to
describe tumours, immune responses and anti-cancer therapies is given. Findings obtained via
computational analysis are illustrated and analysed in the
``\nameref{sec:result}'' and ``\nameref{sec:discuss}'' sections. Finally, the
``\nameref{sec:conclusion}'' section terminates the paper.
\section{Model}
\label{Sec:model}
Let us consider the biological setting under study as follows: a
primary, clinically detected cancer is present in a patient and it is
scheduled to be treated with different therapeutic approaches, in an
effort to improve the patients' clinical outlook. A
secondary lesion is also growing, undetected and located away from the primary
site, due to previous metastatic events and migration of tumour
cells belonging to the first lesion. We are interested in
understanding how the secondary lesion is affected by strategies aimed
at reducing the primary one. {\it Our approach is based on an existing
mathematical model for tumour-immune interaction
\cite{kim2012modeling}, which has been validated previously both from
the point of view of biological appropriateness and sensitivity to
model parameters. The phenomena at hand are inherently complex and
there is a number of unknowns that still characterise these
processes. Our work is thus focussed on understanding the major trends
and the typical outcomes that can emerge in treating secondary
lesions, providing some quantitative data that can be tested
experimentally.}
The dynamics between a heterogeneous, small, solid
cancer lesion and the immune system is formulated using an hybrid agent-based model
(ABM) coupled with a delay differential equation (DDE) system. An
immune response to cancer cells that grow and mutate is simulated using a population
of cytotoxic T lymphocytes (CTLs), which mature in a tumour-draining lymph node. The overall
approach rests on an existing framework, originally
discussing tumour cells endowed with only a unique, single
phenotype. The novelty of the present formulation lies in considering
more than one clone, with mutation processes strongly influencing and
shaping tumour growth dynamics.
For a full analysis and description
of the model we refer the reader to Ref.~\cite{kim2012modeling}, and
only discuss the equations briefly in the following.
The system describing immune activation is given by:
\begin{equation}
\begin{cases}
A'_0(t) =& s_A-d_0A_o(t)-\alpha T(t)A_0(t), \\
A'_1(t) =& V_{\mathrm{ratio}}\alpha T(t)A_0(t)-d_1A_1(t),\\
C'_0(t) =& r_C \left( 1-\frac{C_0(t)}{K} \right) C_0(t)-\mu A_1(t) C_0(t), \\
C'_1(t) =& 2^m \mu A_1(t-\sigma)C_0(t-\sigma)-\mu A_1(t)C_1(t) + \\
& + 2 \mu A_1(t-\rho) C_1(t-\rho)-\delta_1 C_1(t)-f C_1(t),\\
C'_2(t) =& \frac{f C_1(t)}{V_{\mathrm{ratio}}}-\delta_1C_2(t),
\end{cases}
\label{eq:dde}
\end{equation}
where $T$ is the total cancer cell population and $A_0$, $A_1$, $C_0$,
$C_1$, $C_2$ are the concentrations of antigen
presenting cells (APC), mature APCs, memory CTLs, effector CTLs and
CTLs, respectively. A sketch of the dynamics captured by the above
equations is depicted in Fig.\ref{au:figimmuno}.
\begin{figure}[htbp!]
\centering
\includegraphics[width=1\textwidth]{linfo.png}
\caption[Immune system activation cycle]{{\bf Immune system activation cycle} as described by the system of equation \eqref{eq:dde}. The behaviour of the immune system is modeled as ~\cite{kim2012modeling}}
\label{au:figimmuno}
\end{figure}
The first two equations describe the transition from immature APCs
circulating in the periphery to mature ones migrating to the lymph
node as a response to tumour antigens.
The population of immature
APCs is generated and dies at constant rates $s_A$ and
$d_0$, with the maximum value of $A_0$
corresponding to the equilibrium level $s_A/d_0$. When tumour antigens
are presented, $A_0$ decreases proportionally
to the antigenicity value $\alpha$ and
mature APCs ($A_1$) begin entering the lymph node, with some dying at
natural death rate $d_1$.
The presence of mature, tumour-antigen-bearing APCs in the lymph node
causes memory CTLs to activate and mature into effector CTLs, with a
certain delay. Consequently, the effector CTLs proliferate and
migrate to the tumour site where the anti-tumour immune response
starts. This process is captured as follows. The third equation represents the stimulation by the APCs of
the memory CTLs ($C_0$), with a logistic growth rate independent
of the external stimuli and a stimulation by
mature APCs that follows a mass action law. In the fourth equation,
memory CTLs develop a minimal
division process, characterised by $m$ times divisions, and
evolve in effector CTLs ($C_1$) with a time delay $\sigma$. Then, effector CTLs divide
again in a time $\rho$ and flow away of the lymph node or naturally
die with a rate $\delta_1$. The last equation represents the concentration of CTLs ($C_2$) in
the periphery around the tumour and provides the concentration $C_2$
used by the ABM component of the model to generate the
boundary conditions for the tumour-site domain. Table
\ref{tab:param_dde} shows parameter values used in this system of equations and their
meanings. Subsection \ref{sec:parsen} reports the parameter estimation and the related sensitivity analysis, which have been mostly performed in previous works ~\cite{kim2012modeling}.
\begin{table}[!ht]
\begin{adjustwidth}{-2cm}{-2cm}
\begin{center}
\caption[Parameters used in the DDE component of the model]{{\bf Parameters used in the DDE component of the model. For more details about parameter estimation refer to ~\cite{kim2012modeling}}}\label{tab:param_dde}
\begin{tabular}[c]{|p{1cm}|p{4cm}|p{3.5cm}|p{0.75cm}|p{6cm}|
\hline
{\bf Par.} & {\bf Description } & {\bf Value (range) } & {\bf Ref. } & {\bf Note } \\ \hline \hline
$A_0(0)$ & Initial concentration of immature APCs & 0.01 $\mathrm{k}/\mathrm{mm}^3$ & \cite{catron2004visualizing} & same order of magnitude as the APCs in the lymph node \\ \hline
$d_0$ & Death/turnover rate of immature APCs & 0.03 $\mathrm{day}^{-1}$ & \cite{mohri2001increased} & similar to those of na\"ive T cells \\ \hline
$s_A$ & Supply rate of immature APCs & $0.3 ~\mathrm{k}/\mathrm{mm}^3 \mathrm{day}^{-1}$ & \cite{mohri2001increased} & $s_A=A_0(0)d_0$ \\ \hline
$d_1$ & Death/turnover rate of mature APCs & 0.8 $\mathrm{day}^{-1}$ & \cite{belz2007killer} & using a half-life of 20 h: $d_1=(\ln{2})/20 \; h^{-1}$ \\ \hline
$K$ & Equilibrium concentration of memory CTLs & $2\%\cdot 200 ~\mathrm{k}/\mathrm{mm}^3$ & \cite{catron2004visualizing} & $2 \%$ of the T-cells in a lymph node of radius 1 mm \\ \hline
$r$ & Logistic growth rate of memory CTLs
& $\log 2~\mathrm{day}^{-1}$ & \cite{kim2012modeling} & minimum doubling time of 1 day \\ \hline
$m$ & Minimal number of CTL divisions & 10 & \cite{wodarz2005effect} & range from 7 to 17 cell divisions \cite{de2001recruitment,kaech2001memory} \\ \hline
$\delta_1$ & Death/turnover rate of effector CTLS & 0.4 $\mathrm{day}^{-1}$ & \cite{de2003different} & half-life during T-cell contraction of 41 h: $\delta_1=(\ln{2})/41 \; h^{-1}$ \\ \hline
$\mu$ & Mass-action coefficient & 20 ($\mathrm{k}/\mathrm{mm}^3$)$^{-1}\mathrm{day}^{-1}$ & \cite{catron2004visualizing} & $\mu = 0.5 \mu_0$ with $\mu_0=4.8 cell^{-1}day^{-1}$ and $V_{\mathrm{lymph \; node}}=8.4\cdot10^{-3}mm^3$\\ \hline
$\rho$ & Duration of one CTL division & 8 h & \cite{janeway1996immunobiology} & T-cell doubles every 8 hours during expansion \cite{de2003different} \\ \hline
$\sigma$ & Duration of CTL division program & $1+(m-1)\rho$ & \cite{veiga2000response} & first division does not occur until 24 hours after stimulation\\ \hline
$\alpha$ & Antigenicity of the tumour & $10^{-9}$ ($\mathrm{k}/\mathrm{mm}^3$)$^{-1}\mathrm{day}^{-1}$ & \cite{kim2012modeling} & reciprocal of the rate of encountering of antigen from a tumor cell by APC \\ \hline
$f$ & CTLs flow rate out of lymph node to tissue & $0.7 \; \mathrm{day}^{-1}$ & \cite{kim2012modeling} & effector CTLs emigrate at a half-life of 1 day: $f=\ln{2} day^{-1}$ \\ \hline
$V_{\mathrm{ratio}}$ & Ratio of volume of tissue to the lymph node & 1000 & \cite{ying2009three} & lymph node compartment is $\sim 1 ml$ and the breast tissue $\sim 1 l$ \\
\hline
\end{tabular
\end{center}
\end{adjustwidth}
\end{table}
\subsubsection*{Cancer dynamics and immune cells response}\label{subsec:ABM}
The ABM controls tumour growth dynamics and the
interaction between tumour cells and CTLs at the tumour site, which follow
specific algorithmic rules. Our model does not consider healthy
tissue around tumours and other structures such as the stroma or the
cells part of a vascular network: it is assumed that tumour's
surroundings are effectively healthy cells being ``pushed away'' by the
growing tumour. Note also that no vascularisation is present due to the limited
size of the secondary tumour lesion, which is considered to be small, solid and with no
necrotic core. Furthermore, other motility of metastatic processes from the secondary lesion are neglected. The overall assumption is that the secondary tumour is trying to colonize the site and is in its early stages of proliferation.
All cells partaking the dynamics are represented as spheres of radius $r$ in 3D space, with no overlap.
ABM is updated in discrete timesteps $\Delta t$.
{\bf CTL agents}. The rules that govern CTLs cells via the ABM are
three: motion around the tumour, recruitment of other immune system
cells and killing of tumour cells. As mentioned, CTLs cells appear at a
concentration $C_2(t)$ at the border of
the spherical domain representing the region of interest where the tumour is
growing. They then move into that region performing Brownian motion in
3D space until they either collide with a cell or leave the
domain. At each time step, the position of the
cells are given by independent random variables with normal distribution
$\mathcal{N}(0,\sigma^2 \Delta t)$, where the variance is such that
$\sigma^2=2D$, with $D$ being the diffusion rate of the CTLs. When an
immune cell comes into contact with a cancer cells three possibilities
exist:
\begin{itemize}
\item A CTL clone can be recruited with a probability $1-e^{-\Delta
t/C_{\mathrm{recruit}}}$, with $C_{\mathrm{recruit}}$ being
the average recruitment time. Mathematically, CTL recruitment is modeled similarly to Mallet et al. \cite{mallet2006cellular} with cellular automata, and it is biologically validated as in \cite{soiffer2003vaccination, soiffer1998vaccination}. When the first CTL cell
engages a cancer cell and starts recruiting another CTL clone, a
second cell appears at a position adjacent to the first
cell. The direction of the new clone is chosen randomly among
all directions available.
\item A cancer cell is not recognised with a probability $1 -
P_{i,\mathrm{recog}}\cdot\Delta t$, where $P_{i,\mathrm{recog}}$ is
the probability of the i-th cancer phenotype (see below) to be
recognised by the immune system. The parameter
$P_{i,\mathrm{recog}}$ has a value of one for cancer agents, expressing antigens completely matching with the T-cell receptors and thus, that are always
detected by the immune system. A value of zero indicates that
the antigens of a phenotype are completely unrecognised. If the
cancer cell is not
recognised, the CTL starts to move again choosing a new random direction and
accelerating up to the maximum
unit standard deviation $\sigma_{\mathrm{max}}$. If
$C_{\mathrm{acc}}$ is the time necessary to accelerate from the
stationary to the maximum diffusion rate, the CTL acceleration
is computed as: $\sigma(t)=\sigma_{\mathrm{max}}\cdot
\mathrm{min} \left( t/C_{\mathrm{acc}},1 \right)$.
This approach aims at approximating CTL chemotaxis along a chemokine
gradient \cite{mackay1996chemokine, maurer2004macrophage}.
\item A cancer cell is recognised and killed with a probability
$1-e^{-\Delta t/C_{\mathrm{kill}}}$, with
$C_{\mathrm{kill}}$ being the average time for a CTL to eliminate
a cancer cell. The killing process is obtained
by removing the agent. After the agent is removed, the
immune cell starts to move again as described above.
\end{itemize}
If CTLs die naturally, then they are removed from the system. An
explanation of the ABM-parameters is reported in Table
\ref{tab:param_abm} whereas parameter estimation and sensitivity analysis is discussed in Subsection \ref{sec:parsen}.
\begin{table}[!ht]
\begin{adjustwidth}{-.6in}{-.6in}
\begin{center}
\caption[Parameters used in the ABM component of the model]{{\bf Parameters used in the ABM component of the model. For more details about parameter estimation refer to ~\cite{kim2012modeling}}}\label{tab:param_abm}
\begin{tabular}[c]{|p{1.1cm}|p{5cm}|p{2cm}|p{0.75cm}|p{6.5cm}|
\hline
{\bf Par.} & {\bf Description } & {\bf Value (range) } & {\bf Ref. } & {\bf Note } \\ \hline \hline
$\Delta t$ & Time step & 1 min & \cite{kim2012modeling} & timescale of the fastest dynamic simulated in the model \\ \hline
$r$ & Radius of cells & 5 $\mu \mathrm{m}$ & \cite{lin2004model} & \cite{mallet2006cellular, alarcon2003cellular, catron2004visualizing} \\ \hline
$T_{\mathrm{div},i}$ & Avg. division time of i-th cancer phenotype & 1-39 day & \cite{kirschner1998modeling} & \cite{mallet2006cellular, kuroishi1990tumor, michaelson2003estimates} \\ \hline
$T_{o}$ & Avg. division time of {\it original} phenotype & 7 day & \cite{kirschner1998modeling} & \cite{mallet2006cellular, kuroishi1990tumor, michaelson2003estimates} \\ \hline
$\sigma_{\mathrm{max}}$ & Max unit standard deviation of CTL diffusion & 12 $\mu \mathrm{m} ~\mathrm{min}^{-1}$ & \cite{catron2004visualizing} & \cite{friedl2001interaction} \\ \hline
$C_{\mathrm{acc}}$ & CTL acceleration time from 0 to $\sigma_{\mathrm{max}}$ & 5 h & \cite{kim2012modeling} & \\ \hline
$C_{\mathrm{death}}$ & Avg. CTL lifespan & 41 h & \cite{de2003different} & \\ \hline
$C_{\mathrm{recruit}}$ & Avg. time fro CTL recruitment & 22 h & \cite{kim2012modeling} & \cite{soiffer2003vaccination, soiffer1998vaccination} \\ \hline
$C_{\mathrm{kill}}$ & Avg. time fro CTL to kill tumour cell & 24 h & \cite{kim2012modeling} & killing target cells may require a long recovery period \\ \hline
$R$ & Radius of region of interest & 620.4 $\mu \mathrm{m}$ & \cite{kim2012modeling} & \\ \hline
$h$ & Thickness of CTL cloud & $3\sigma_{\mathrm{max}}\sqrt{\Delta t}$ & \cite{kim2012modeling} & probability that a CTL could pass from outside into the region of interest is 0.001 \\ \hline
$P_{\mathrm{mut}}$ & Probability of mutation & 0.01 $\mathrm{min}^{-1/2}$ & \cite{wood2007genomic} & \\ \hline
$P_{\mathrm{recog},i}$ & Probability of recognition of i-th cancer phenotype & 0-1 & & Span the entire probability range \\ \hlin
$P_{o}$ & Probability of recognition of {\it original} phenotype & 1 & & the APC cell can always recognize the antigen released \\
\hline
\end{tabular}
\end{center}
\end{adjustwidth}
\end{table}
{\bf Cancer agents}. Tumour cells can
proliferate, mutate or die, killed by the immune system, and no migration
is considered. This approximation is motivated by the scope of the
study, which is focused on the solid, growing secondary lesion after
the colonisation of a new tissue. In this early stage of implantation
most of the cells are assumed to be in a proliferation state and
migration can be neglected \cite{hatzikirou2012go}.
Cellular division occurs with a probability $1-e^{-\Delta t/T_{i,
\mathrm{div}}}$, $i=1,...,5$, where $T_{i, \mathrm{div}}$ is the
average division time of the i-th tumour phenotype. When a tumour cell
divides, the position of a new cell is chosen randomly on the mother
cell's perimeter, such that the daughter cell is tangent. If no space
is available in the chosen position, the division process
fails and no new agent is created, mimicking the contact-inhibition
mechanism occurring in the early stages of metastasis implantation \cite{mendonsa2018cadherin}.
To analyse the effect of mutations on cancer development and immune
response, we use five different cancer phenotypes that may emerge from
the mutation of an {\it original} clone, identified by different values of
characteristic parameters $T_{\mathrm{div},o}=T_o$ and $P_{\mathrm{recog},o}=P_o=1$.
Mutations can occur during cell duplication,
with a probability $P_{\mathrm{mut}}\cdot \Delta t$ that aims to
capture the genetic instability of the system.
Each mutated cell is then identified by indices
representing the level of expression of the two characteristic quantities
$T_{\mathrm{div}}$ and $P_{\mathrm{recog}}$. These values effectively classify the
mutated clones and the following mutated phenotypes.
Modeling few phenotypes of mutated cells is a simplification justified
by several works showing that only a limited number of phenotypes are
predominant in a tumour, see for example \cite{anderson2006tumor}. For the
scope of our study, the five mutated clones are prototypical of a wide
range of similar mutations.
In Table \ref{tab:clones} cancer clonal composition is considered. One of the assumptions is that only one class of CTLs is modeled and it is not antigen specific. Although different types of CTLs could take part in an immune response and act differently depending on the clone, our immune attacks are regulated only via $P_{recog,i}$.
\begin{table}[!ht]
\begin{adjustwidth}{-.6in}{-.6in}
\begin{center}
\caption[Cancer clonal composition]{{\bf Cancer clonal composition}}\label{tab:clones}
\begin{tabular}{|c|l|c|c|}
\hline
{\bf Name} & {\bf Description } & {\bf $P_{\mathrm{recog}}$} & {\bf $T_{\mathrm{div}}$ } \\ \hline \hline
{\it original} & \begin{tabular}{@{}l@{}}
First metastatic breast cancer clone \\
that colonises the new tissue, with \\
evolutionary potential of phenotypic mutations. \end{tabular}
& 1 $\mathrm{min}^{-1}$ & 7 day \\
\hline
(0.5,0.5) & \begin{tabular}{@{}l@{}}
Clone proliferates at
the same rate of the {\it original} \\ clone,
but has an increased ability to hide from \\
the immune system. \end{tabular}
& 0.5 $\mathrm{min}^{-1}$ & 7 day \\
\hline
(0,1) & \begin{tabular}{@{}l@{}}
Clone is not recognised by CTLs, but \\
the evolutionary cost of its ability is\\
paid in term of proliferation: this \\
phenotype is the slowest to reproduce. \end{tabular}
& 0 $\mathrm{min}^{-1}$ & 13 day\\
\hline
(0.25,0.75) & \begin{tabular}{@{}l@{}}
Clone has intermediate properties: \\
strong ability to hide and slow proliferation \end{tabular}
& 0.25 $\mathrm{min}^{-1}$ & 10 day \\
\hline
(0.75,0.25) & \begin{tabular}{@{}l@{}}
Clone has intermediate properties: \\
weak ability to hide and fast proliferation \end{tabular}
& 0.75 $\mathrm{min}^{-1}$ & 4 day \\
\hline
(1,0) & \begin{tabular}{@{}l@{}}
Clone has the ability to reach \\
high number of cellular duplication, but is \\
always recognised by the immune system. \end{tabular}
& 1 $\mathrm{min}^{-1}$ & 1 day \\
\hline
\end{tabular}
\end{center}
\end{adjustwidth}
\end{table}
Using these different types of phenotypes, as we will see shortly,
helps us to shed light on the role of mutations in determining the
effectiveness of immune response and anti-cancer therapies.
Different clonal compositions and reproductive and immunoediting
advantages dramatically influence the outcomes of anti-cancer therapies.
\subsection*{Modeling therapies: chemotherapy, immune boosting and radiotherapy} \label{sec:fitland}
One of the typical features of secondary lesions is that they usually
show cells with mutated
functional characteristics respect to the original tumour, due to the genetic instability typical of
metastatic masses they originate from.
We reiterate that there is no analysis of the fate
of the global cancer disease but only on such secondary lesions, which
can show different dimensions, compositions, structures and biological
characteristics from the primary neoplasia. Chemotherapy, immune
boosting and radiotherapy are the strategies our modelling focuses on.
{\bf Chemotherapy.} This treatment consists of cytotoxic drugs
targeting a specific cellular
phase of the cell cycle to induce cell death. The procedure acts
against rapidly proliferating cells, independently from their nature
~\cite{brunton2011therapeutics}. This means that healthy cells and
immune system cells are usually damaged along with cancer cells, and
this leads to well-known side effects for the patients. In this work,
only the primary killing effect against cancer cells and no
direct effects on the immune system is assumed. This simplification is
motivated by two main points. First, the average CTL lifespan is 41
hours, whereas the tumour division rate is greater and the tumour death
rate due to the therapy is slower. CTL cells are rapidly affected by
the reduction of the tumour mass and no new CTL is recruited: the
``old" cells tend to naturally die. Second, if on one side
chemotherapy affects the immune cells, on the other specific T-cell
response is reinforced \cite{zitvogel2008immunological}, and
the investigation of these secondary effects is not in the scope of
the future present work.
During a cycle of chemotherapy of duration $Ch_{\mathrm{time}}$, the
i-th cancer phenotype can go through cellular death with probability
$1-e^{-{\Delta t}/{Ch_{\mathrm{kill},i}}}$, where the average time for
the drugs to induce cellular death is $Ch_{\mathrm{kill},i}$ and
depends from the proliferation potential of the phenotype. $Ch_{\mathrm{time}}$ takes into account a single cycle of three injections every three days and represents the global time duration of the chemotherapy's effects. The drug remains two days above a certain percentage level such that the cytotoxic effects on tumour cells can be considered constant.
Different values of $Ch_{\mathrm{time}}$ have been explored as reported in Table \ref{tab:paramTh}, supposing that the same total dose is inoculated in continuous cycles of low metronomic doses. The effect of different $Ch_{\mathrm{time}}$ with the same total dose is a faster or slower decrease of the cancer population with similar qualitative dynamics.
In particular, for clone $(0,1)$ (refer to Table \ref{tab:clones} for notation), i.e. the phenotype that grows slowly
but is poorly immunogenic,
$Ch_{\mathrm{kill},\mathrm{(0,1)}}=Ch_{\mathrm{time}}$, namely a
$(0,1)$-death is very rare. Clone $(0+0.25i,1-0.25i)$, with
$i=1,...,4$, has
$Ch_{\mathrm{kill},\mathrm{(0+0.25i,1-0.25i)}}=Ch_{\mathrm{time}}-i\cdot
Ch_{\mathrm{eff}}\cdot Ch_{\mathrm{time}}$, so that the tumour with
higher proliferation rate has very high probability to die due to the
effect of the drug.
{\bf Immune boosting.}
We use this generic term to capture the number of clinically available strategies that
potentiate an immune response. For example, a
treatment that is increasingly used for cancer patients is the so-called
adoptive cell transfer (ACT), where patients' own immune cells are
stimulated and modified to treat their tumour. There are several types of
ACTs that go under different acronyms depending on the boosting
strategy employed, with the most used ones nowadays being
TIL (Tumour infiltrating lymphocytes), TCR (Tumour cell receptors)
T-cell and CAR (Chimeric antigen receptors) T-cells
treatments~\cite{rosenberg2008adoptive}.
We concentrate in particular on TIL therapy, where T-cells are
extracted from the patient's tumour, grown {\it in vitro} to boost
their numbers and injected back into the patient to contrast cancer
progression. This strategy appears to be, for example, one of the most effective treatment
against metastatic melanoma~\cite{rosenberg2008adoptive}. In our
approach, TIL is modeled as a continuous increase of the CTLs concentration in
the cloud, depending on the value of $C_2$ at the starting time for
the therapy. The net increase is modeled by a number of $Bo_{\mathrm{eff}}$
cells for a short time $Bo_{\mathrm{time}}$. For simplicity, in
the following we refer to this treatment as immune boosting or
simply boost. The parameters used to model boost and
chemotherapy are explained and collated in Table~\ref{tab:paramTh}.
\begin{table}[!ht]
\begin{adjustwidth}{-.6in}{-.6in}
\begin{center}
\caption[Parameters used to model therapies]{{\bf Parameters used to model therapies}. Ranges indicate that different therapies (single and combined) are simulated with different values.}\label{tab:paramTh}
\begin{tabular}{|c|l|c|}
\hline
{\bf Parameter} & {\bf Description } & {\bf Value (range) } \\ \hline \hline
$Ch_{\mathrm{time}}$ & Duration of a chemotherapy cycle & 10-50 day \\
$Ch_{\mathrm{eff}}$ & Effect of chemotherapy & $0-1/4$ \\
$Bo_{\mathrm{time}}$ & Persistence time of boosting (TIL) & 3 day \\
$Bo_{\mathrm{eff}}$ & Number of CTL cells injected & 500-1000 \\
\hline
\end{tabular}
\end{center}
\end{adjustwidth}
\end{table}
{\bf Radiotherapy.} Radiotherapy (RT) uses ionising radiation to
induce cell death in a localised area under treatment. This therapy has
several positive and negative feedbacks on the immune system, modulating
different compartments of the tumour microenvironment. In particular,
tumour-specific antigens and immune-stimulatory signals are released by
the dying cancer cells.
Because of its contributing
primarily to the original, metastatic neoplasia, the effect of RT is
here modelled as an indirect effect
on the secondary lesion and is accounted for as
as a restoring factor in the ability of CTL cells to recognise and kill various cancer phenotypes.
\subsection{Parameter estimation and sensitivity analysis} \label{sec:parsen}
The biological significance of parameters and processes that underpin
the present model has been discussed at length elsewhere
\cite{kim2012modeling, Frascoli2017215}. In some cases, such as, for
example, parameters used for tumour
division time or cell radius, well-established values in the literature
have been used \cite{lin2004model, alarcon2003cellular,
kirschner1998modeling, mallet2006cellular}. In other cases,
estimations from the available experimental and theoretical data have
been carried out.
A sensitivity analysis has originally also been carried out
for eight parameters of the model: $T_{\rm div},~\sigma_{\rm
max},~C_{\rm acc},~C_{\rm recruit},~C_{\rm kill},~K,~\mu, ~m$ and
$\alpha$. Other parameters have not been considered because their role is
known to be marginal. For instance, the replenishment rate for memory
CTLs is known to be irrelevant, since only a very tiny fraction of
memory CTLs ($~ 1\%$) is known to be affected by the tumour. Similarly,
the duration of CTL division (time delay parameter $\rho$ in the DDE)
is too small to impact the CTL division program as a whole and does
not influence final outcomes. Using Spearman's rank-order
correlations, tumour populations' values and extinction times, Kim et al.
have concluded \cite{kim2012modeling} that tumour division times $T_{\rm div}$, antigenicity
$\alpha$ and the number of divisions of memory CTLs upon activation
$m$ are the most sensitive parameters.
In this work, we use the same parameters proposed in the original paper, with the
only difference of $T_{\rm div}$, still chosen in the proposed
interval but capturing a more aggressive tumour (i.e. $T_{\rm
div} = 1-39$ days). The effect on simulations is to shorten the
proliferating phase, which occurs at a larger growth rate and allows
for a quicker immune response.
The new probability coefficients introduced here, i.e. $P_{\rm mut}$ and
$P_{\rm recog,~i}$, have different effects. By using $5$ different
simulations with different initial random seeds and $10$ different
values of the parameters, we conclude that $P_{\rm mut}$ has no effect
on the final outcomes of the system, but only accelerates or delays
the identical dynamics shown by the model.
$P_{\rm recog,~i}$ instead has a notable effect on the system. When
$P_{\rm recog,~i} = 0$ for a given $i$-th clone, the immune system is
unable to eradicate that particular phenotype and, if no external
therapy is present, the tumour endlessly grows.
As far as the values for therapies' parameters are concerned,
i.e. $Ch_{\rm time}$, $~Ch_{\rm eff}$, $~Bo_{\rm time}$, $~Bo_{\rm
eff}$, they are chosen so that the dynamics
between tumour, immune system and therapies display interesting
behaviours and does not result in an immediate negative or positive
outcome.
In particular, $Ch_{\rm eff}$ and $Bo_{\rm time}$ have been
varied in a number of different instantiations of the model, with only
the cases $Ch_{\rm eff} = 0.25$, $Bo_{\rm time} = 1000$ cells used in
the discussion of results. Variations of those parameters do not alter
in a significant way the prototypical dynamics that we will discuss
shortly. Note that $Ch_{\rm eff} = 0.25$ has been chosen so that the
cytotoxic drug targets fast proliferating cells.
\subsection{Morphological and complexity measures} \label{sec:measures}
Three indices that capture the shape and cellular compositions of the
tumour mass are
introduced and monitored in our computational experiments. {\it Note that
these indices can guide the evaluation of collective
properties of the evolving tumours. They are useful to discriminate
between different evolutions of the cancer masses and have also been
validated in some in vitro experiments, as shown by other authors in
previous works \cite{sottoriva2010cancer, jeon2010off}.}
{\bf Roughness.} Although random proliferation of a group of
cells leads to an almost smooth and spherical object, a
tumoural mass with diverse clonal families under the action of the
immune system can present itself as a rough
aggregate. To account for this, a measure of
roughness $M$ is introduced, as the ratio
between the surface $S$ and the volume $V$ of the aggregate~\cite{sottoriva2010cancer}.
The minimum ratio is represented by a sphere $S_s/V_s=(4\pi
R_s^2)/(4/3\pi R_s^3)=3/R_s$, where the value has been
non-dimensionalised as follows:
$M_{\mathrm{min}}=\sqrt{S_s}/\sqrt[3]{V_s}=\sqrt{4\pi}/\sqrt[3]{(4/3)\pi}$. The
roughness index $M$, expressed in terms of the minimal ratio for a
sphere, is given by:
\begin{equation}
M=\frac{\sqrt{S}}{\sqrt[3]{V}}\cdot\frac{1}{M_{\mathrm{min}}}=\frac{\sqrt{4\pi S}}{\sqrt[3]{3(4\pi)^2V}}.
\label{eq:rough}
\end{equation}
A compact, non-infiltrated, almost spherical tumour mass has an index
$M$ close to unity while a tumour with highly irregular borders, for instance a solid
tumour with fingers and clusters of invasive cells or a mass highly
infiltrated by the immune system, displays a higher value.
{\bf Radius of gyration.} This value represents the radius of a
sphere that contains the whole tumour aggregate and reads:
\begin{equation}
R_g=\sqrt{\frac{\displaystyle\sum_{i=1}^{N_c} (\mathbf{r}_i-\mathbf{r}_{cm})^2 }{N_c}},
\label{eq:Rg}
\end{equation}
where $N_c$ is the total number of cancer cells and $\mathbf{r}_i$ is
the distance of each clone from the center-of-mass of the tumour
($\mathbf{r}_{cm}$) that can vary during the tumour progression.
{\bf Shannon Index.} This indicator is introduced to account
for the presence of different phenotypes within a tumour, with regards
to tumour heterogeneity and
relative frequency of each clonal family ($p_i$). The Shannon index
$H$ is thus defined as:
\begin{equation}
H=-\frac{\displaystyle\sum_{i=1}^{s}p_i \ln(p_i)}{\ln(s)}
\label{eq:shannon}
\end{equation}
where $p_i$ is the relative abundance of the phenotype $i$ and $s$ is
the total number of different phenotypes (in our case $s=6$). For
simplicity, $H$ is then normalised to the interval $[0,1],$ where zero
indicates a homogeneous population with only one clonal family and unity
represents a fully heterogeneous population where all phenotypes
are equally present.
\\
\section{Results}\label{sec:result}
The model outlined in the previous sections is the basis for {\it in
silico} experiments, where a different number of therapies and their
combinations are tried out for significant values of the parameter
set. Depending on the initial conditions, the system exhibits
three typical behaviours, namely eradication, sustained (irregular) oscillations
or exponential, uncontrolled growth when an immune response to a
growing tumour is present. We consider parameter values where the effect of clinical therapies are relevant. Cases where the tumour grows too fast or too slow, making the effects of therapies not noticeable, are excluded from our analysis. Stationary behaviour has
never been observed. Outcomes also
depend on the characteristics of cellular phenotypes present in the
growing mass, strongly influencing its speed of growth, its ability to
counteract the action of T-cells with immunoediting and its
morphological qualities, which can hinder the ability of the immune
system to effectively erode the cancer.
Considering the dynamics observed in a number of computational
experiments performed at biologically meaningful parameter values, tumour
growth generally appears as exponential, with a consequent
linear increase in the radius of gyration $R_g$ with time{\it , as
previously observed \cite{Bru19984008}. The main reason, as
explained in Ref.~\cite{jeon2010off}, is that the growth is driven by those cells that
reside at the periphery of the mass.} The nearly
spherical shape of the tumour when only a single clone is present
changes significantly in the presence of mutations. The
greatest contribution to asymmetry occurs when a new population with
a faster proliferation rate than neighbouring cells is generated. In that case,
this new population forms an evolutionary niche that can alter the
sphericity of the tumour, until the new clones have proliferated enough
to surround the slower cells and recreate a spherical appearance, as
shown in Fig. \ref{fig0}. Note that, in our model, cells acquire a new phenotype upon mutation in a purely stochastic way and there is equal probability to mutate from the \textit{original} phenotypes to all the others.
\begin{figure}[htbp!]
\centering\noindent
\makebox[\textwidth]{\begin{subfigure}{0.42\textwidth}
\includegraphics[width=1\textwidth]{fig0a.eps}
\caption{}
\end{subfigure}
\begin{subfigure}{0.42\textwidth}
\includegraphics[width=1\textwidth]{fig0b.eps}
\caption{}
\end{subfigure}
\begin{subfigure}{0.42\textwidth}
\includegraphics[width=1\textwidth]{fig0c.eps}
\caption{}
\end{subfigure}}\\
\makebox[\textwidth]{\begin{subfigure}{0.42\textwidth}
\includegraphics[width=1\textwidth]{fig0d.eps}
\caption{}
\end{subfigure}
\begin{subfigure}{0.42\textwidth}
\includegraphics[width=1\textwidth]{fig0e.eps}
\caption{}
\end{subfigure}
\begin{subfigure}{0.42\textwidth}
\includegraphics[width=1\textwidth]{fig0f.eps}
\caption{}
\end{subfigure}}
\caption[Tumor growth dynamic]{{\bf Tumour growth dynamics}
Panel (A) $R_g$ as function of time, slow initial trend followed by a
fast linear growth. Panel (B) $H$ as function of time, increases with
mutations and decreases when the fittest clonal population outnumbers
the others.
Panel (C) $M$ as function of time. The roughness index increases with
mutations,
is higher when a new population with fast proliferation arises and
reaches an almost stationary level when $H$ is near zero.
Panel (D), (E), (F) 3D-views of the tumour, respectively, just after
the birth of the fittest clone (t=30 days), when the fittest
population starts to invade (t=36 days), when the tumour returns to
growing almost spherically with low grade of heterogeneity (t=50
days).}
\label{fig0}
\end{figure}
The heterogeneity of the mass increases with mutations
until the faster clones are not outnumbering the other
phenotypes. If this occurs, then the Shannon index
$H$ rapidly decreases with time in a way that it is inversely proportional to
the growth of the more proliferating clones, i.e. the faster they grow
the faster $H$ decreases.
The action of the immune system usually tends to favor homogeneity
over heterogeneity, rebalancing the distribution of phenotypes as long
as the immune response is active. As T-cells erode the tumour,
natural selection leads to an evolutionary bottleneck
characterised by low $H$. It is interesting to note that the roughness
of the tumour tends to be in the interval $1 < M \leq 1.5$, with signs
of superficial infiltration by the
immune system. The limited life span of the CTLs used in our model tends
to promote attacks that occur on the periphery and rarely result in
deep infiltration, which, as we will show shortly, is instead present when therapies are activated.
\subsection*{The effect of therapies on tumours}
In the following, all the parameters have been set as
in Ref.~\cite{kim2012modeling} and are reported in Tables
\ref{tab:param_dde} and \ref{tab:param_abm}. The tumour mutation rate has been chosen following
the principle that positive mutations, i.e. mutations that lead to an
evolutionary benefit to the cells over therapies and immune response,
are rare. As it is expected, the dominant phenotype
population usually appears after a clonal expansion of few
mutations. The range of variation of the proliferating, aggressive tumour has been
set to ensure a biological meaning and a rate of growth
that allows the cancer mass to escape the control of the immune
system in a limited range of time. The main reason for this choice is that we are interested in modeling the impact of different therapies on cancers that will not
be eradicated in the absence of anti-tumour therapies and that, at the same time, do not show growth rates that are unrealistic. Thanks to the probabilistic structure of the system, simulations can generate different outcomes also when parameters are kept fixed for the particular cancer studied.
Among the different experiments, three paradigmatic dynamics
emerge, which bear particular relevance and help understanding the
typical scenarios that our model predicts. They are the result of stochastic variations on the evolution of initially identical tumours. These cases respectively correspond to a tumour mass
with an initial slow growth and high heterogeneity (case A), and two fast
growing tumours with either initial low (case B) or high (case C)
heterogeneity.
\subsubsection*{First single-therapy strategy: chemotherapy}\label{res:chemo}
Our first choice is to simulate a cytotoxic chemotherapy that acts with more efficiency against the
cancer cells that have the largest growth rate, starting at
day $60$ after the first tumour cell colonizes the site and for a total
duration of $10$ days.
The probability of a cell to be killed by chemotherapy, with a total dose of drug fixed, is set independently from the time duration
of the protocol in an effort to maximize the efficiency of the
therapy.
Fig.~\ref{au:fig1} shows the evolution of the tumour cell population as a function
of time, according to different
phenotypic compositions. In the following, {\it time is evaluated
starting from the instant at which the original clone starts to
colonize the new organ. This time does not refer to the primary
tumour or the history of his evolution.}. In
panel (A), the effect of chemotherapy on tumour case A, which is representative of those
cancers with lower rates of proliferation but higher
propensities to mutate.
A too early start of the treatment results
in a completely ineffective strategy, with a negative outcome. This is
because chemotherapy reduces the more proliferating cells (in pink) at day $60$
when those cells are still scarce and the tumour is too small to
benefit from the action of the cytotoxic drug. Once the treatment is
over, the remaining cells restart to mutate and proliferate, with an
exponential growth that the immune system alone cannot contain. As
shown in the inset of panel (A), the number of cells belonging to the original phenotype (in black) remain almost
constant throughout the procedure and do not change significantly for the duration of the
experiment.
Panel (B) of Fig.~\ref{au:fig1} instead shows a complete eradication of
case B, where tumour cells have initially a low heterogeneity but are
reproducing fast. The effect of the therapy
is in this case to eliminate every cell belonging to the dominating,
fast-reproducing phenotype before it is over, i.e. approximately
around day $7$ of its $10$ days duration. Also, all cells
of the original phenotype are eradicated by the end of the treatment,
with the tumour completely cleared out by the effect of the cytotoxic
drug and the immune response.
An initially fast reproducing tumour with high heterogeneity can instead
lead to uncontrollable rebounds, with an overall negative outcome for
the patient. In panel (C), the action of the chemotherapy is not
sufficient to eliminate every single cell belonging to the mutated
phenotype. According to our choice of parameters, it is enough that one original
cancer cell or one of the more proliferating clone survives after the
chemotherapy that a fast, uncontrollable rebound can be expected.
Interestingly, these last two cases (i.e. B and C) do not show different evolutions
of the radius of gyration $R_g$ {\it (not shown in the Figure)}
during the action of chemotherapy, since the treatment acts homogeneously
on the cancer mass as a whole. This is associated directly to the
limited dimension of the tumour, leading to the drug acting
on the aggregate with a high strength from all spatial directions.
Roughness $M$ instead show significant
changes from case A and cases B and C. Tumour case A
remains spherical and compact during the
experiment, essentially because the treatment has a very limited
effect on the mass due to its premature start. Tumour case B and C, instead, reach high
level of roughness during the treatment, showing, for the cases
reported in the Figure~\ref{au:fig1}, a maximal $M$ of $3.77$ and $3.45$,
respectively at day $64$ and $65$ for cases B and C. This clearly
indicates that the tumour
loses density and becomes morphologically inhomogeneous at around half
of the treatment duration and is infiltrated to a relevant degree by
the T-cells taking part in the immune response. The overall
consideration from these results is that the correct timing of
treatment, {\it here intended as the ideal treatment
starting time and therapy duration to achieve optimal patient’s benefit},
is a major variable for the outcome of the treatment and
it is also strongly affected by phenotypical compositions.
\begin{figure}[htbp!]
\centering\noindent
\makebox[\textwidth]{
\begin{subfigure}[]{0.68\textwidth}
\includegraphics[width=1\textwidth]{fig1a.eps}
\caption{Case A}
\end{subfigure}
\begin{subfigure}[]{0.68\textwidth}
\includegraphics[width=1\textwidth]{fig1b.eps}
\caption{Case B}
\end{subfigure}}\\
\begin{subfigure}[]{0.68\textwidth}
\includegraphics[width=1\textwidth]{fig1c.eps}
\caption{Case C}
\end{subfigure}
\caption[Chemotherapy effect]{{\bf Chemotherapy effect}
Number of cancer cells and phenotypical composition as functions of time. Chemotherapy starts at
day $60$ and lasts for $10$ days. Panel (A) shows an initial phenotype
(in black) with slow reproductive rate (case A), eventually overtaken by a new
phenotype that reproduces very fast (in pink). The inset shows
cellular distribution during the last part of the treatment. Panel (B)
shows an initially fast reproducing tumour with low heterogeneity (case B),
with an inset of the last two days before complete
eradication. Panel (C) depicts an initially fast reproducing tumour
with high heterogeneity (case C), with the inset focusing on the dynamics
at the end of the treatment and in the first days of the rebound phase.}
\label{au:fig1}
\end{figure}
\subsubsection*{Second single-therapy strategy: immune boosting}\label{res:immuno}
The overall effect of immune boosting is to increase the number of CTL cells circulating
around the tumour site, which we simulate as an injection of cells
starting at day $50$
and occurring for a duration of $3$ days. Erosion of cancer cells by the immune system proceeds
from the periphery towards the center of the tumour mass, and is
usually characterised by a linear decrease of $R_g$ during the first
few days. Another typical characteristics of the dynamics that
follows boosting is a clonal expansion of the CTL
population shortly after treatment. For the prototypical three
cases A, B and C introduced above, all of our computational experiments
indicate that boosting alone is not able to
eradicate cancer: after an initial decrease in the tumour mass, two types of evolutions have been
observed, both resulting in negative outcomes. Of particular relevance
is case C, which, although not treatable by chemotherapy alone, shows a somewhat unexpected and
complicated morphology when
subject to immune boosting. In fact,
after an initial clonal selection of the less immunogenic phenotype,
case C displays a clear deviation from sphericity in the mass, with a
nonlocal spread of the tumour population in islands of different sizes,
as reported in Fig.~\ref{au:fig2}. After a decreasing phase
due to an immune response that does not result in a complete
eradication, the tumour population is eventually subject to
a faster, uncontrollable increase driven by disconnected, smaller masses.
Overall, a selection of 5 parameter sets and 10 trials for different seeds give qualitatively similar results.
\begin{figure}[htbp!]
\centering\noindent
\makebox[\textwidth]{
\begin{subfigure}[]{0.55\textwidth}
\includegraphics[width=1\textwidth]{fig2a.eps}
\caption{}
\end{subfigure}
\begin{subfigure}[$t=170 \mathrm{days}$]{0.59\textwidth}
\includegraphics[width=1\textwidth]{fig2c.eps}
\caption{$t=170 \mathrm{days}$}
\end{subfigure}}
\caption[Immunotherapy effect.]{{\bf Immunotherapy effect.}
Results for the prototypical case C, a fast growing cancer with high
phenotypical heterogeneity. Boost immunotherapy starts at day $50$ and
lasts a total of $3$ days. Panel (A): time series of each cell
population, panel (B): snapshot of the tumour mass at time
$t = 170$ days.}
\label{au:fig2}
\end{figure}
Panel (A) in Fig.~\ref{au:fig2}, shows the number of cells for each
phenotype as a function of time: an initially exponential growth
is firstly slowed down then halted by immunotherapy, with a maximum cell population
occurring about $20$ days after the beginning of the
treatment. Around day $100$, the evolution of the cell population
changes. Frequent, local maxima in both tumour and T-cell curves represent
the failed attempts made by the immune system to completely erode the tumour due to the increasing sparseness of the
cancer. This behaviour seems to occur for a protracted period of time
of about $60$ days.
As the cells in the island's sizes begin to proliferate faster than
the rate of killing of the T-cells, a rebound phase with a higher
speed of growth than the \textit{original} unbroken mass appears at day $180$. Panel (B) provides an image
of the scattered status of the tumour immediately before its exponential
rebound. Let us remark that the model does not allow for migration of cancer cells and this picture is the result of the
infiltration of T-cells coming from boosting and immune
response.
As expected, morphology immediately before the rebound phase
is characterised by a high value of roughness, with
$\max(M_{\mathrm{Case C}})=2.16$ at day $162$. Also, there is almost a
twofold increase in $R_g$ compared to the value for identical number
of cells in the first growth phase. For example, for $10^4$ cells, we have $R_{g_{\mathrm{Case
C}}} = 125$ at day $47$
and $R_{g_{\mathrm{Case C}}} = 240$ at day 160.
\subsubsection*{First example of synergistic therapy: chemotherapy and
boosting}\label{res:combi}
Cancer heterogeneity has been invoked to explain one of the major
aspects of cancer development, namely acquired drug resistance, by
which phases of remission are often followed by a rapid growth of
tumour cells~\cite{gerlinger2010darwinian}.
One of the ways to overcome resistance is, for instance, to find more ``evolutionarily
enlightened'' strategies that places malignant cells in an
``evolutionary double bind''~\cite{gatenby2009lessons}. In cancer, a double blind could be
obtained using the immune system as natural biological
predator~\cite{basanta2012exploiting}. Clinical evidence shows that
immunotherapy or oncolytic viruses alone are not effective, despite
the possible theoretical advantages. Therefore, cancer treatment is
adopting a multistep approach that combines biological and
chemical$/$radioactive therapies using cytotoxic effects on one side
and subsequent adaptation on the other side to limit tumour adaptive
resistance~\cite{ramakrishnan2010chemotherapy,
antonia2006combination}.
Guided by the poor outcomes displayed by immune boosting alone in the
prototypical cases, we now consider the
combination of chemotherapy and immune boosting, with the aim of
discussing the major factors that maximize positive outcomes. The
prototypical cases have been subjected to a protocol of an immune boosting
injection at day $50$ lasting three days, followed by a chemotherapy
session at day $60$. Results are displayed in Fig~\ref{au:fig3}, with the
insets displaying phenotypical composition over time.
For cases A and B,
the complete temporal range is shown, whereas for C the last $20$
days are reported. Timing for these therapies has been chosen
arbitrarily. {\it For cases B and C the second lesion grows up
to numbers of tumour cells that are close to
the detectability threshold.}
\begin{figure}[htbp!]
\centering\noindent
\makebox[\textwidth]{
\begin{subfigure}[Case A]{0.65\textwidth}
\includegraphics[width=1\textwidth]{fig3a.eps}
\caption{Case A}
\end{subfigure}
\begin{subfigure}[Case B]{0.65\textwidth}
\includegraphics[width=1\textwidth]{fig3b.eps}
\caption{Case B}
\end{subfigure}}\\
\begin{subfigure}[Case C]{0.65\textwidth}
\includegraphics[width=1\textwidth]{fig3c.eps}
\caption{Case C}
\end{subfigure}
\caption[First kind of synergistic therapy: chemotherapy and
immune boosting.]{{\bf First kind of synergistic therapy: chemotherapy and
immune boosting.}
Plots represent the time evolution of cellular populations, with
insets showing phenotypical composition of the tumours with time.
Immune boosting
starts at day $50$ and lasts $3$ days, whereas chemotherapy starts at
day $60$ and has a duration of $10$ days. Panel (A), (B) and (C)
respectively show cases A (slowly growing tumour), B (quickly growing tumour, low
heterogeneity) and C (quickly growing tumour, high
heterogeneity). {\it Note that different scales have been used to allow for
greater details of the dynamics. The
inset in Panel (C) represents a close-up of the time range $90-110$
days. The unit on the $y-$ axes of every plot in the Figure is cells' numbers.}}
\label{au:fig3}
\end{figure}
The effects of this synergistic therapy in cases A and C are
similar: chemotherapy preferentially kills those cells that are
fast to reproduce, leaving the slowest reproducing phenotype unaffected. As a
result, rebounds occur once therapies end, with case A showing a
negative outcome within the simulated time window and case C
displaying a still moderate but uncontrollable growth at the end of
the simulation. In other words, the effect of chemotherapy is
to create an evolutionary bottleneck
that selects the poorly immunogenic clones. {\it In particular, case C (see
the inset of Panel (C)) shows a surviving tumour composed by only two
clones: the clone $(0,1)$ and the clone $(0.25,0.75)$: these are the two
families that are the slowest in proliferating and have the lowest
immunogenicity.}
These clones have a strong
immunoediting ability and remain unnoticed by the immune system,
resulting in unending growth: {\it because of their ability to be
elusive to both the immune response and chemtherapy, outcome C
appears the worst of all.} Note that, for this reason, this phase
is of a different nature than those previously reported for individual therapies
(i.e. Figs.~\ref{au:fig1} and ~\ref{au:fig2}). Also, because of the low
heterogeneity of case B, this combined therapy is instead successful
in fully eradicating the tumour, which is eliminated during the
administration of the cytotoxic drug. On the other hand, case A, differently from the
results obrained for chemotherapy alone, displays a selection of the
poorly differentiated immunogenic clones.
From the morphological perspective, masses emerging from this
synergistic intervention appear to be low in roughness when the
reproductive rate is slow, with
$\max(M_{\mathrm{Case A}}) = 1.62$ at $t = 81$ days. If the rate is
instead fast, the level of heterogeneity usually determines the level
of roughness, with
low heterogeneity contributing high roughness during the chemotherapy
phase, i.e. $\max(M_{\mathrm{Case B}}) = 3.05$ at
$t=64$ days and $\max(M_{\mathrm{Case C}})=2.64$ at $t=63$ days.
Contrary to case B, case C shows a Shannon index of $H>0.5$ for most
of the simulation, which results in an unsuccessful eradication.
Other time protocols and order of administration between boosting and
chemotherapy are possible, and have been tested to some degree
(results not
shown here). Although a study of optimisation of protocols is not
within the scope of the present work, the overall insight from the
simulations is that heterogeneity always plays an important role in the
outcomes. For this combination therapy, high values of $H$ are
consistently associated with negative prognosis \cite{greaves2012clonal}.
\subsection*{Second example of synergistic therapy: radiotherapy,
boosting and ``abscopal'' effect}\label{res:abs}
A second example of synergistic therapies that is currently used in
clinical practice is the combination of an initial cycle of
radiotherapy with an immune boosting protocol. Besides a better
understanding of the parameters that can trigger a positive outcome,
our interest in testing such a combination resides in the occurrence
of a somewhat rare and poorly understood
event, which is named ``abscopal'' effect. There are a number of clinical cases discussed in the medical
literature where a reduction of a
secondary tumour or an existing metastasis outside the primary, radiated
lesion has been observed \cite{reits2006radiation, lugade2005local,
finkelstein2012clinical, finkelstein2012combination,
finkelstein2011confluence, vatner2014combinations}. Differently from
chemotherapy, radiotherapy has a localised action on the region
irradiated and this makes the phenomenon, to some extent,
counterintuitive.
Sometimes, the effect appears to affect a secondary lesion very distant from the region treated.
The
complications inherent to the stimulation of such an event are due
to the immune action apparently being as crucial as radiotherapy in triggering such a response.
The model allows us to test some hypotheses on the nature and causes
of the effect of protocols introduced by Demaria et al. in
Ref.~\cite{demaria2004ionizing}, who have reported some interesting
and positive outcomes. In particular, they have treated mice with a
syngeneic mammary carcinoma in both flanks with immunotherapy and only
one of the two tumours with radiotherapy.
They use the non-irradiated lesion to monitor the insurgence of the abscopal effect.
Biologically, a tumour-specific T-cells activation occurs after
inflammatory signals are introduced in the system as a consequence of
the therapy. Dying cancer cells
release tumour-specific antigens and immune-stimulatory signals that seem to
induce an increased
recognition of cancer cells with the same phenotypical
characteristics in other areas of the body. Further, radiation modulates different
compartments of the tumour microenvironment, resulting in
exclusion-inhibition of effector T-cell and induction of de novo
anti-tumour immune responses~\cite{demaria2015role}. The protocol that
we simulate is a radiotherapy (RT) on the primary tumour (not
simulated or showed here) at day 1,
followed by an immune boosting that lasts 10 days.
As anticipated, RT is considered a restoring factor in the ability
of CTL cells to recognise and kill all cancer phenotypes, with no
exceptions.
The secondary lesion is composed of $5$ x $10^4$
heterogeneous cells (the same number of cells injected in mice in the
experiment by Demaria et al.),
generated randomly with each clonal family representing at least $10\%$ of the total population.
We compare the
action of two single therapies (RT and immune boosting),
with a combination of the two and a control case where the second
lesion remains untreated. Results are presented in
Fig~\ref{au:fig4}:
{\it each panel represents the typical outcome from
a single simulation. For each case, i.e. control (no treatment),
combination, immunotherapy, and radiotherapy, we have performed three
different simulation runs, with different initial conditions. The
outcome of each simulation for any configuration consistently gives
comparable results. Variation due to stochasticity are minimal and
do not affect the outcomes.}
Firstly, no treatment or RT alone result in similar negative outcomes for the
secondary mass, not directly treated by RT, both
from the perspective of surviving cancer cells (panel (A)) and the
response of the CTLs of the immune system (panel (B)). An initial RT
with no follow-up has the only effect of delaying an exponential
rebound, not dissimilar to the behaviour of an untreated
mass. Boosting alone does not impact the mass as much as when
we combine boosting and RT, with the former giving rise to a tumour
that after $30$ days has less than half the mass than in the case of
chemotherapy alone. Stimulated by the release of the antigens of
the dead tumour cells of the primary lesion, both therapies show a
maximum in the number of active CTLs, which occurs around day $7$ and is then
followed by a characteristic drop around day $10$. Qualitatively the results of the model reproduce the experimental data in Ref.~\cite{demaria2015role}, with indications of a start of a remission for the secondary tumour.
Different strategies on the secondary mass also lead to different
clonal compositions, which have an effect on the final outcome..One of
the keys for the success of the strategy is to have no phenotype
prevailing over the others, as shown in panel (C) for the combination
of RT and immune boosting and, to a lesser
extent, in panel (D) for immune boosting alone. {\it Note that the width of
the coloured regions in panels (C)-(F) indicates the number of cells
that belong to a specific clone population: the larger the width, the
larger the population.}
For example, combination therapy
provides a very high Shannon index, $H>0.99$, throughout the whole
duration of the experiment. RT and no treatment show instead
reduced indices, with values at day $30$
of $0.43$ and $0.55$ respectively. Interestingly, the RT case appears to be less
heterogeneous than the control case. Overall, it is important to stress
that, for the case of the secondary lesion, high heterogeneity is not
per se associated to a worse prognosis.
The reason is that a successful
action on the secondary mass
reduces the fitness advantage of the phenotypes and makes the immune
system able to recognise each clones equally. {\it Note that Panel (E)
refers to the radiotherapy case made on primary tumour, showing the
evolution of the secondary mass and the fact that the immunogenic
effects induced by the treatment are not sufficient to contrast cancer
growth. Panel (F) represents the control case, where no treatments
are administered and the tumour is growing unchallenged as an
aggressive breast cancer.}
\begin{figure}[htbp]
\centering\noindent
\makebox[\textwidth]{
\begin{subfigure}[]{0.50\textwidth}
\includegraphics[width=1\textwidth]{fig4a.eps}
\caption{Therapy's action on tumour}
\end{subfigure}
\begin{subfigure}[]{0.50\textwidth}
\includegraphics[width=1\textwidth]{fig4b.eps}
\caption{Therapy's action on CTL}
\end{subfigure}}\\
\centering\noindent
\makebox[\textwidth]{\begin{subfigure}[]{0.50\textwidth}
\includegraphics[width=1\textwidth]{fig4c.eps}
\caption{Combination}
\end{subfigure}
\begin{subfigure}[]{0.50\textwidth}
\includegraphics[width=1\textwidth]{fig4d.eps}
\caption{Immunotherapy}
\end{subfigure}}\\
\centering\noindent
\makebox[\textwidth]{\begin{subfigure}[]{0.50\textwidth}
\includegraphics[width=1\textwidth]{fig4e.eps}
\caption{RT}
\end{subfigure}
\begin{subfigure}[]{0.50\textwidth}
\includegraphics[width=1\textwidth]{fig4f.eps}
\caption{Control case}
\end{subfigure}}
\caption[Effects of anti-cancer treatments on a secondary tumour.]{{\bf Effects of anti-cancer treatments on a secondary tumour.}
Panel (A and B): Cellular populations of a
secondary tumour as functions of time, with cancer populations depicted
by a continuous line (panel (A)) and CTLs population by a dashed line
(panel (B)). Each color corresponds to a different treatment. {\it These curves represent
a typical simulation run. Variability due to stochasticity is
minimal.} Other panels show
phenotypical composition of the secondary tumour for the {\it curves
shown in Panels (A)-(B)) and for the} following
strategies applied the primary tumour: combination therapy (panel (C)),
immune boosting (panel (D)), radiotherapy (panel (E)), no treatment
(panel (F)). Each color indicates a different cancer clone
family. {\it Legends show labels for such families, whose properties
are described in
detail in Table~\ref{tab:clones}.}}
\label{au:fig4}
\end{figure}
We find that the relation between the phenotypical composition of the primary and
secondary lesions plays a very important role in the dynamics of the
so-called abscopal effect. The previous examples refer to
a secondary lesion that is antigenically related to the first tumour,
but this is not always the case in practice. Results vary considerably
if the phenotypical compositions differ and this is important to stress.
For example, if the secondary tumour is characterised
by clones that are not antigenically related to the first lesion, the
final outcome of combination therapy cannot be as positive as in the
previously discussed cases. Given that radiotherapy affects
phenotypes that are distributed in different ways in the first and
second tumours, the immune system is not capable of recognising specific tumour
cells in the same successful way as in the previous examples. As a
result, the therapy shows a worse outcome, as can be seen in
Fig.~\ref{au:fig5}(a). Furthermore, if the second tumour mass is instead
implanted in an immune suppressed host (mathematically obtained by setting $\alpha=10^{-15}$),
where the level of CTLs circulating around the tumour is lower than ordinary
levels, the outcome is negative. As reported in
Fig.~\ref{au:fig5}(b), a reduced fitness of the immune system causes
one phenotype to prevail over the others and proliferate quickly out
of control. It is reasonable to suppose that, if more cycles of therapies are repeated, the effectiveness of treatments is likely to be also reduced.
\\
\begin{figure}[htbp!]
\centering\noindent
\makebox[\textwidth]{
\begin{subfigure}[]{0.55\textwidth}
\includegraphics[width=1\textwidth]{fig5b.eps}
\caption{Antigenical unrelated tumour}
\end{subfigure}
\begin{subfigure}[]{0.55\textwidth}
\includegraphics[width=1\textwidth]{fig5c.eps}
\caption{Immune suppressed host}
\end{subfigure}}
\caption[Immune mediation on a secondary lesion.]{{\bf Immune mediation on a secondary lesion.}
Tumour size and composition as functions of time for three different
secondary tumours, subjected to different anti-tumour strategies on the primary
tumour. Panel (A) shows results for a secondary tumour antigenically unrelated to the
primary and panel (B) shows the outcome for a secondary tumour that is antigenically related to the
primary tumour but implanted in an immune-suppressed host.
{\it
Legends show labels for the tumour subpopulations emerging from the
“original” population, via secondary mutations caused by the
therapies. Properties of each clone family are described in detail
in Table~\ref{tab:clones}.}}
\label{au:fig5}
\end{figure}
\subsubsection*{Mutation rates and eradication of secondary tumours} \label{res:abs_mut}
A complete eradication of a secondary tumour as
an indirect result of an anti-cancer therapy on a primary lesion is a
relatively rare occurrence. Also, it appears to be associated mainly with certain types of cancer,
namely melanoma or breast cancer. This
might be linked to the fact that generic metastatic events are
characterised by a high genetic instability, often
making secondary lesions phenotypically unrelated to the first
tumour. From this point of view, a possible speculation could be that
the so-called abscopal effect is not a rare event per se,
but it is an effect limited to secondary tumours that have a
phenotypical clone composition that is not too dissimilar from the
originally treated lesion, and thus the effect only seldom changes the
prognosis for secondary lesions. Indeed, extensive genetic and
phenotypic variation are known to strongly influence therapeutic
outcomes \cite{burrell2013causes}.
To investigate how the rate
of mutation of cancer cells affects outcomes on secondary lesions,
we generate a tumour of $5$ x $10^4$ cells with a full mutation capability and apply a combination therapy
(RT and immune boosting) as per the previously introduced
protocols. In other words, the complete secondary lesion before the
start of treatments is composed by the \textit{original} clone.
Typical results are reported in Fig~\ref{au:fig6}(a) and show that outcomes
do not linearly depend on the rate of mutation. Interestingly, tumour reductions at
day $30$ are larger when the mutation rate is lower, but become
negligible when the rate of mutation is approximately larger than
$P_{\rm mut} = 0.6$, with no relevant change in the overall outcome
for higher rates (see purple, yellow and cyan lines). Also, for rates lower than $P_{\rm mut} = 0.1$ (see
blue and black lines),
different dynamics of eradication can be present, with tumours having
different cell counts after the treatment is administered, although the final result
at day $30$ appears almost identical.
\begin{figure}[htbp!]
\centering\noindent
\makebox[\textwidth]{
\begin{subfigure}[]{0.65\textwidth}
\includegraphics[width=1\textwidth]{fig6a.eps}
\caption{}
\end{subfigure}}\\
\centering\noindent
\makebox[\textwidth]{
\begin{subfigure}[]{0.55\textwidth}
\includegraphics[width=1\textwidth]{fig6b.eps}
\caption{$\max(H(\mathrm{P_{mut}}))$}
\end{subfigure}
\begin{subfigure}[]{0.55\textwidth}
\includegraphics[width=1\textwidth]{fig6c.eps}
\caption{$\max(M(\mathrm{P_{mut}}))$}
\end{subfigure}}
\caption[Dependence of eradication of the secondary lesion from its mutation rate.]{{\bf Dependence of eradication of the secondary lesion from its mutation rate.}
Results for secondary lesions with different mutation rates, when a
combination therapy is administered to the first lesion.
Panel (A): Total cancer population as function of time for different
$P_{\mathrm{mut}}$. At day $0$, the secondary tumour mass is
phenotypically identical
to the primary lesion. Panel (B) and (C): Maximum values of the Shannon index $H$ and of
the roughness $M$ as functions of the mutation rate. {\it Points in the
Figure represent results from single simulations. Maxima in panels (B)
and (C) have been taken over time for each rate of mutation for the
clones.}}
\label{au:fig6}
\end{figure}
Heterogeneity tends to increase with the mutation rate, but
its maximum value during therapy is not directly proportional to how
fast the system can mutate. After a given rate of mutation, which is
approximately $P_{\rm mut} \approx 0.5-0.6$,
mutations do not increase the heterogeneity of the mass. This is
because the fittest
clone usually becomes dominant, its cells outnumber the other
phenotypes and heterogeneity reduces. Fig.~\ref{au:fig6}(b) shows that
the maximum values of $H$ are a concave function of mutation
probability. Also, the overall
distribution of phenotypes in the secondary mass is, as discussed
previously for the case of a single cancer, strongly affected by the
morphology.
Different rates induce different roughness as shown in
Fig.~\ref{au:fig6}(c), where the maximum value of
$M$ reached by the tumour mass is plotted as a function of
mutation rate. This value increases for small rates and reaches a
maximum for a rate $P_{\rm mut} = 0.2$, followed by a sharp
decrease. For rates larger than $P_{\rm mut} = 0.5$, the value does
not change significantly. Tumour morphology of the secondary lesion
influences how the immune system progresses in its attack. In our
simulations, we observe different dynamics of attack carried out by
the immune system, with the tumour being eroded in different ways and
often not in a homogeneous fashion. Nonetheless, a rougher tumour always appears more vulnerable to
immune system attacks because of the degree of infiltration by CTLs it
allows. Even in ``abscopal'' positive outcomes, infiltration plays a major role in the dynamics of erosion and
high $M$ correlates with better results. At the highest mutation rates, roughness is low
because the immune system is not able to recognise
phenotypes that are different from those of the primary mass and kill
them. This results in a fast, unbounded growth of one or two phenotypes that
increase the sphericity of the mass and quickly lower the roughness
value approximately to unity, which corresponds to a spherical
object. This is reflected in the plateau observed in Fig.~\ref{au:fig6} for $P_{\rm mut} > 0.5$.
\section{Discussion}\label{sec:discuss}
Cancer and immune cells are complex systems with different
characteristics that also depend on internal and external evolutionary
pressures. In the last decades, improvements on the general
knowledge of these processes have stimulated new
therapeutic strategies which take into account to patients'
particularities to some degree. The detailed model of
immune interaction described here focuses on the
salient traits of the dynamics and is able to reproduce the
major features of a number of therapeutic interventions.
An analysis of the effect of different drugs on three prototypical
secondary masses arising from a metastatic breast cancer (not modeled) has been proposed, showing a faithful representation of
some of the main mechanisms of tumour-immune interactions present in
literature~\cite{piretto2018combination, de2006mixed}. In particular,
we note a significant dependence of the outcomes on the
heterogeneity of the tumour, with higher heterogeneity generally
correlated with negative outcomes confirming biological evidence suggested in Ref.~\cite{burrell2013causes}.
Indeed, therapies targeting heterogeneous cancer micro environments often show large
rebounds of more resistant tumour cells, that are able to counteract
the action of drugs or boosting in a consistent way. In particular,
one original result from our modeling is that
chemotherapy appears more efficient in a less phenotypically differentiated secondary lesion
independently of the rate of growth or the dimension of the mass.
For the reproductive rates considered in this study,
immune boosting alone is not sufficient to produce full eradication, but rather can trigger
the spread of the more aggressive cells in the body by making the
existing mass more sparse. A well-timed intervention with a combination of
boosting and chemotherapy seems to be the safest of the protocols,
allowing for a relevant reduction of the mass and preventing the
unbounded growth of the most proliferating cells. Nevertheless, timing of intervention on the secondary lesion can be critical.
One further result of this work is to uncover the importance
of tumour morphology in evolution and fate of secondary lesions.
The shape of secondary masses conveys important information
that could be an indicator of successful eradication. For instance, during or
immediately after administration of chemotherapy, our modeling shows
that a high infiltrated tumour is associated with the best outcome. At the same time, a
harsh environment or a high selective pressure tend to generate a tumour that has a
greater roughness and the tendency to spread, as previously noted
~\cite{anderson2006tumor}. This often occurs and persists for many
days after administration of
chemotherapy. In particular, findings for case A suggest that clonal composition of
surviving cells that originated from the beginning tumour colony, and
are later influenced
by the selected therapies, strongly affect the final outcomes of the
metastasis. Similar dynamics is reported to be present in some types
of tumours that are known to be particularly resistant to therapies~\cite{Kusoglu201980}.
Here we consider one cycle of treatment, but some extra {\it care} should be
taken when multiple cycles are considered, since the immune system
could be further weakened and respond less efficiently.
On the other hand, there is evidence in the medical literature that a
combination of radiotherapy and immunotherapy can provide a positive
effect not only in the area affected by the radiation but also in
other areas of the body. This seems to be due to the release of autologous neoantigens to the
immune system \cite{demaria2015role}, with the overall result of what
it appears to be an individualised tumour vaccine. Our model
captures the effects of this cascading action on a secondary
metastatic mass and confirms that the immune system can act
as a mediator for secondary attacks. In particular, there is evidence
in our findings
that immune-suppressed hosts or secondary lesions
antigenically unrelated to the treatment area do not show any abscopal
effect, as experimentally noted.
There is a growing discussion in the community about the causes of
this rare, positive occurrence on secondary masses. Currently, it
seems that this is the result of a fragile balance between positive
and negative signals activated with the radiation, dependent on the
pre-existing environment and the immunogenicity of the
tumour~\cite{demaria2015role}. These biological elements, alongside
a critical dependence on the dose and the interval between
radiation fractions, contribute to the low occurrence of this effect.
Furthermore and interestingly, a
dependence of the effect on the mutation rate of the cancer clones in
the secondary site is apparent, suggesting that the role of genetic instability at that
site should be investigated more.
Overall, our findings emphasises that the morphology of
the secondary lesions before, during and after the treatment, bears some indications of the rate of success for the treatment.
For the lesions under the detectability threshold, this work suggests that heterogeneity and roughness
are both important quantities. Negative prognosis is linked to the selection of a
poorly immunogenic clone and has been shown to lead to a large,
unbounded regrowth of the tumour.
It is thus vital to design a protocol that can minimize the immunoediting ability
of cells surviving from therapies or improve the immune system ability to recognise
and attacks such clones especially when we cannot detect individual,
isolated lesions but only total tumour burden.
In regards to the latter, the activation of an
``abscopal-like'' response seems to be a strategy in re-calibrating the immune reaction to such
cells. Results are still in their infancy and it is unknown whether such a response can be
elicited and how. One of the ideas
we suggest is to carefully consider the best tumour target to be irradiated:
when possible, it could be advantageous to target a secondary
metastasis antigenically related to the more common lesions in the
body, with a low grade of hypoxia and with a good grade of
immunogenicity.
\section{Conclusions}
\label{sec:conclusion}
Cellular mutation constitutes one of the causes of negative
outcomes in therapeutic strategies against cancer. Morphology, growth rates and the clonal composition of a tumour mass can, to some extent, be used as predictors of tumour
resistance to a range of anti-cancer therapies and be analysed to combine
treatments to maximize their impacts. In most of the commonly used protocols, the action of the immune system is
crucial. Modern therapies elicit and enhance patient's immune
response, also because of its ability to adapt to, change and modify the
tumour microenvironment.
The mathematical model we have presented tries to capture the
complexity of tumour-immune dynamics and discuss how
therapies with different scopes, doses and protocols can influence
prognosis for small, solid, secondary tumoural lesions. {\it Even if
these lesions can be small and not yet detectable, their role can be
catastrophic for the patient if they are untreated or, as in some
cases we have shown, the effect of therapies on the primary tumour
can lead to the selection of more aggressive and resilient clones in
the secondary lesions.} Ideally,
individual, evidence-based modelling
might provide a fast, reliable and patient-centred way to test and find optimum control of
protocols {\it in vivo}.
Our findings suggest that the success of synergistic therapies is
strongly influenced by the phenotypical composition of all the lesions,
alongside their mutation rates and immunogenic properties.
Effective strategies that can ``normalise'' the
microenvironment \cite{jain2005normalization} and will try to
limit tumour clonal mutation could be trialled to improve prognosis.
Therapies that target slowly proliferating and undifferentiated cells
can also become viable in the future.
\section{Acknowledgements}
EP, PSK and FF gratefully acknowledge support for this work through
the Australian Research Council Discovery Project DP180101512,
“Dynamical systems theory and mathematical modelling of viral
infections”.
\section{Code and Data availability}
The code used and the produced data for the simulations discussed in this work are available upon request.
\bibliographystyle{model1-num-names}
|
\section{Byzantine Causal Consistency} \label{section:cc}
Causal consistency variants in the literature~\cite{
CM:DC95,CC:PPoPP16,VCC:POPL17,Jiang:SRDS20}
are defined based on the \emph{happens-before} relation
among events~\cite{Lamport:CACM79}.
However, they are not applicable to systems that allows Byzantine nodes,
particularly Byzantine clients.
We now adapt the happens-before relation in Byzantine-tolerant systems,
and define Byzantine Causal Consistency (Byz-CC) as follows.
For two events $e$ and $f$,
we say that $e$ happens before $f$, denoted $e \leadsto f$,
if and only if one of the following three rules holds:
\begin{itemize}
\item \emph{Session-order}.
Events $e$ and $f$ are two operation requests
issued by the same \emph{correct} client,
and $e$ is issued before $f$.
We denote it by $e \rel{{\sf so}} f$.
We do \emph{not} require session order among operations
issued by Byzantine clients.
\item \emph{Read-from relation}.
Event $e$ is a \put{} request issued by some client
and $f$ is a \textsc{get}{} request issued by a \emph{correct} client,
and $f$ reads the value updated by $e$.
We denote it by $e \rel{{\sf rf}} f$.
Since a \textsc{get}{} of Byzantine clients may return an arbitrary value,
we do \emph{not} require read-from relation induced by it.
\item \emph{Transitivity}.
There is another operation request $g$
such that $e \leadsto g$ and $g \leadsto f$.
\end{itemize}
If $e \leadsto f$, we also say that $f$ causally depends on $e$
and $e$ is a causal dependency of $f$.
A version $\mathit{vv}$ of a key $\mathit{k}$ causally depends on
version $\mathit{vv}'$ of key $\mathit{k}'$,
if the update of $\mathit{vv}$ causally depends on that of $\mathit{vv}'$.
A key-value store satisfies Byz-CC{} if,
when a certain version of a key is visible to a client,
then so are all of its causal dependencies.
\section{Conclusion} \label{section:conclusion}
We present Byz-GentleRain, the first
causal consistency protocol which tolerates up to $f$ Byzantine
servers among $3f + 1$ servers in each partition
\emph{and} any number of Byzantine clients.
The preliminary experiments show that
Byz-GentleRain{} is quite efficient on typical workloads.
Yet, more extensive large-scale experiments on more benchmarks are needed.
We will also explore optimizations of our synchronization protocol
in Algorithm~\ref{alg:cgst} in future work.
\section{Evaluation} \label{section:evaluation}
We evaluate Byz-GentleRain{} in terms of
performance, throughput, and latency of remote update visibility.
We also compare Byz-GentleRain{} to Byz-RCM.
\subsection{Implementation and Setup} \label{ss:impl-setup}
We implement both Byz-GentleRain{} and Byz-RCM{} in Java
and use Google's Protocol Buffers~\footnote{
Protocol Buffers: \url{https://developers.google.com/protocol-buffers}.}
for message serialization.
We implement the key-value stores as \textsf{HashMap},
where each key is associated with a linked list of versioned values.
The key-value stores hold 300 keys in main memory,
with each key of size 8 bytes and each value of size 64 bytes.
We run all experiments on 4 Aliyun~\footnote{
Alibaba Cloud: \url{https://www.alibabacloud.com/}.} instances running Ubuntu 16.04.
Each instance is configured as a data center,
with 1 virtual CPU core, 300 MB memory, and 1G SSD storage.
All keys are shared into 3 partitions within each data center,
according to their hash values.
\subsection{Evaluation Results} \label{ss:results}
\begin{figure}[t]
\centering
\begin{subfigure}[c]{0.45\textwidth}
\centering
\includegraphics[width = \textwidth]{figs/1-throughput.pdf}
\caption{Throughput}
\label{fig:ff-throughput}
\end{subfigure}
\hfill
\begin{subfigure}[c]{0.45\textwidth}
\centering
\includegraphics[width = \textwidth]{figs/3-get-latency.pdf}
\caption{\textsc{get}{} latency}
\label{fig:ff-get-latency}
\end{subfigure}
\hfill
\begin{subfigure}[c]{0.45\textwidth}
\centering
\includegraphics[width = \textwidth]{figs/2-put-latency.pdf}
\caption{\put{} latency}
\label{fig:ff-put-latency}
\end{subfigure}
\hfill
\begin{subfigure}[c]{0.45\textwidth}
\centering
\includegraphics[width = \textwidth]{figs/6-visibility.pdf}
\caption{\put{} visibility}
\label{fig:ff-visibility}
\end{subfigure}
\caption{Evaluation of Byz-GentleRain{} and Byz-RCM{} in failure-free scenarios.}
\label{fig:ff-comparison}
\end{figure}
Figure~\ref{fig:ff-comparison} shows the system throughput
and the latency of \textsc{get}{} and \put{} operations
of both Byz-GentleRain{} and Byz-RCM{} in failure-free scenarios.
We vary the $\textsc{get}:\put$ ratios of workloads.
First, Byz-RCM{} performs better than Byz-GentleRain,
especially with low $\textsc{get}:\put$ ratios.
This is because Byz-RCM{} assumes Byzantine fault-free clients
and is \emph{signature-free}.
In contrast, Byz-GentleRain{} requires clients sign each $\textsc{put\_req}$ request.
Second, it demonstrates that Byz-GentleRain{} is quite efficient
on typical workloads, especially for read-heavy workloads.
Third, the performance of Byz-GentleRain{}
is closely comparable to that of Byz-RCM,
if digital signatures are omitted deliberately from Byz-GentleRain;
see Figures~\ref{fig:ff-get-latency} and \ref{fig:ff-put-latency}.
Finally, Figure~\ref{fig:ff-visibility} shows the latency of \put{} visibility,
which gets higher and higher with more and more \put{} operations.
\begin{figure}[t]
\begin{subfigure}[c]{0.45\textwidth}
\centering
\includegraphics[width = \textwidth]{figs/4-byzantine-faults-of-replicas.pdf}
\caption{}
\label{fig:byz-replicas}
\end{subfigure}
\hfill
\begin{subfigure}[c]{0.45\textwidth}
\centering
\includegraphics[width = \textwidth]{figs/5-byzantine-faults-of-clients.pdf}
\caption{}
\label{fig:ff-clients}
\end{subfigure}
\caption{Evaluation of Byz-GentleRain{} in Byzantine scenarios.}
\label{fig:byz}
\end{figure}
We also evaluate Byz-GentleRain{} in several typical Byzantine scenarios.
Generally, both Byzantine clients and replicas
may fail by crash or send arbitrary messages.
Particularly, we consider
\begin{inparaenum}[(1)]
\item Byzantine clients that may send \textsc{get\_req}{} and/or \textsc{put\_req}{}
requests with incorrect timestamps
(line~\code{\ref{alg:client}}{\ref{line:get-send-getreq}}
and line~\code{\ref{alg:client}}{\ref{line:put-send-putreq}}), and
\item Byzantine replicas that may broadcast
different global stable time $\mathit{cgst}$ to replicas in different partitions
(line~\code{\ref{alg:metadata}}{\ref{line:bc-send-newcgst}}).
\end{inparaenum}
Figure~\ref{fig:byz} demonstrates the impacts of
various Byzantine failures on the system throughput of Byz-GentleRain.
On the one hand, the Byzantine failures of types (1) and (2) above
has little impact on throughput.
On the other hand, frequently sending arbitrary messages,
such as \textsc{new\_cgst}{} or \textsc{propose}{} messages, does hurt throughput.
This is probably due to the signatures carried by these messages.
\section{Evaluation}
We evaluate Byz-GentleRain in terms of local latencies, through\_put, and remote update visibility latency,
by varying the workloadss.
We compare Byz-GentleRain to Byz-RCM.
\subsection{Implementation and Setup}
We implement Byz-GentleRain in Java and use Google’s Protocol Buffers for message serialization.
We also implement Byz-RCM in the same code base for performance comparison.
The data store is a key-value store. The data set used
in the experiments contains one thousand key-value pairs,
with the key size being eight bytes and the value
size 64 bytes. The key-value store keeps all key-value pairs
in main memory.
A key points to a linked list that contains
different versions of the same item.
A PUT operation inserts a new version to the version
chain of the updated item and
adds a record to the operation log.
We use Date class to keep physical clocks synchronized. We
configure it to change clock frequency, so that our physical
clocks always move forward, a requirement for correctness
in Byz-GentleRain.
We run all experiments on Alibaba Cloud
instances running Ubuntu 12.04.
Each server has $x$ virtual
CPU cores, $x$ GB memory, and $x$ 16GB SSD storage.
We replicate each dataitem at three replicas.
\subsection{Throughput}
\subsection{Update Visibility Latency}
\begin{figure}
\includegraphics[width=\columnwidth]{figs/latency.png}
\includegraphics[width=\columnwidth]{figs/get_latency.png}
\includegraphics[width=\columnwidth]{figs/FGST_update.png}
\end{figure}
\section{Introduction} \label{section:intro}
For high availability and low latency
even under network partitions,
distributed systems often partition and replicate data
among multiple nodes~\cite{RDT:POPL14}.
Due to the CAP theorem~\cite{CAP:PODC00}
many distributed systems choose to sacrifice strong consistency
and to implement weak ones.
Causal consistency~\cite{CM:DC95} is one of
the most widely used consistency model in distributed systems.
There are several variants of causal consistency
in the literature~\cite{CM:DC95, CC:PPoPP16, VCC:POPL17, Jiang:SRDS20}.
They all guarantee that
\emph{an update does not become visible until all its causality are visible.}
We informally explain it in the ``Lost-Ring'' example~\cite{Chapar:POPL16}.
Alice first posts that she has lost her ring.
After a while, she posts that she has found it.
Bob sees Alice's two posts, and comments that ``Glad to hear it''.
We say that there is a \emph{read-from} dependency
from Alice's second post to Bob's get operation,
and a \emph{session} dependency
from Bob's get operation to his own comment.
By \emph{transitivity}, Bob's comment causally depend on Alice's second post.
Thus, when Carol, a friend of Alice and Bob, sees Bob's comment,
she should also see Alice's second post.
If she saw only Alice's first post,
she would mistakenly think that Bob is glad to hear that Alice has lost her ring.
There are plenty of research prototypes
and industrial deployments of causally consistent
distributed systems (e.g., COPS~\cite{COPS:SOSP11},
Eiger~\cite{Eiger:NSDI13},
GentleRain~\cite{GentleRain:SoCC14},
Cure~\cite{Cure:ICDCS16},
MongoDB~\cite{MongoDB:SIGMOD19}, and
Byz-RCM~\cite{ByzRCM:NAC19}).
GentleRain{} uses a stabilization mechanism
to make updates visible while respecting causal consistency.
It timestamps all updates with the physical clock value
of the server where they originate.
Each server $\mathit{s}$ periodically computes a global stable time ${\sf gst}$,
which is a lower bound on the physical clocks of all servers.
This ensures that no updates with timestamps $\le {\sf gst}$ will be generated.
Thus, it is safe to make the updates with timestamps $\le {\sf gst}$
at $\mathit{s}$ visible to clients.
A get operation with dependency time $\mathit{dt}$ issued to $\mathit{s}$
will wait until ${\sf gst} \ge \mathit{dt}$
and then obtain the latest version before ${\sf gst}$.
However, none of these causal consistency protocols/systems
consider Byzantine faults,
except Byz-RCM (Byzantine Resilient Causal Memory) in~\cite{ByzRCM:NAC19}.
Byz-RCM achieves causal consistency in the client-server model with $3f + 1$ servers
where up to $f$ servers may suffer Byzantine faults,
and any number of clients may crash.
Byz-RCM has also been shown optimal in terms of failure resilience.
However, Byz-RCM did not tolerate Byzantine clients,
and thus it could rely on clients' requests
to identify bogus requests from Byzantine servers~\cite{ByzRCM:NAC19}.
In this work, we present Byz-GentleRain,
the first Byzantine-tolerant causal consistency protocol
which tolerates up to $f$ Byzantine servers among $3f + 1$ servers in each partition
\emph{and} any number of Byzantine clients.
It uses PBFT~\cite{PBFT:Thesis00} to reach agreement among servers on
a total order of client requests.
The major challenge Byz-GentleRain faces is to
ensure that the agreement is consistent with the causal order.
To this end, Byz-GentleRain should prevent causality violations
caused by Byzantine clients or servers:
Byzantine clients may violate the session order
by fooling some servers that a request happened before another
that was issued earlier.
Byzantine servers may forge causal dependencies
by attaching arbitrary metadata for causality tracking
to the forward messages.
To migrate the potential damages of Byzantine servers,
we let clients assign totally ordered timestamps to updates in Byz-GentleRain.
Utilizing the digital signatures mechanism,
Byzantine servers cannot forge causal dependencies.
\begin{figure}[t]
\centering
\includegraphics[width = 0.70\textwidth]{figs/example-gst}
\caption{Why the servers in Byz-GentleRain{} need to synchronize global stable times.}
\label{fig:example-gst}
\end{figure}
To preserve causality,
Byz-GentleRain{} uses the stabilization mechanism of GentleRain.
As explained above, the timestamps in Byz-GentleRain{} are generated by clients.
However, it is unrealistic to compute a lower bound on physical clock values
of an arbitrary number of clients.
Therefore, each server $\mathit{s}$ in Byz-GentleRain{} maintains
and periodically computes a global stable time ${\sf gst}$
which is a lower bound on physical clock values of the clients it is aware of.
In the following, we argue that simply refusing any updates
with timestamps $\le {\sf gst}$ on each server as GentleRain{} does
may lead to causality violations.
Consider a system of four servers which are replicas
all maintaining a single key $\mathit{k}$,
as shown in Figure~\ref{fig:example-gst}.
Due to asynchrony, these four servers may have different values of ${\sf gst}$.
Without loss of generality, we assume that
${\sf gst}_{1} < {\sf gst}_{2} = {\sf gst}_{3} < {\sf gst}_{4}$,
as indicated by vertical lines.
Now suppose that a new update $\mathit{u}: \mathit{k} \gets 5$
with timestamp between ${\sf gst}_{3}$ and ${\sf gst}_{4}$ arrives,
and we want to install it on $\ge 3$ servers,
using quorum mechanism.
In this scenario, if each server refuses any updates with timestamps
smaller than or equal to its ${\sf gst}$,
the update $\mathit{u}$ can only be accepted
by the first 3 servers, indicated by dashed boxes.
Suppose that server 3 is a Byzantine server,
which may expose or hide the update $\mathit{u}$ as it will.
Consequently, later read operations which read from $\ge 3$ servers
may or may not see this update $\mathit{u}$.
That is, the Byzantine server $3$ may cause causality violations.
To cope with this problem,
we need to synchronize the global stable times of servers.
When a server periodically computes its ${\sf gst}$,
it checks whether no larger global stable time
has been or is being synchronized.
If so, the server will try to synchronize its ${\sf gst}$
among all servers, by running PBFT independently in each partition.
For each partition, the PBFT leader is also responsible for
collecting updates with timestamps $\le {\sf gst}$ from $2f + 1$ servers,
and synchronizing them on all servers.
Once successfully synchronized,
a global stable time becomes a \emph{common} global stable time,
denoted ${\sf cgst}$, and in each partition
the updates with timestamps $\le {\sf cgst}$ on all correct servers are the same.
Therefore, each server can safely refuse any updates with timestamps
smaller than or equal to its ${\sf cgst}$.
Still, the classic PBFT is insufficient to guarantee causality,
since a Byzantine leader of each partition
may propose an arbitrary set of updates.
To avoid this, in Byz-GentleRain{} the PBFT leader will also
include the sets of updates it collects from $2f+1$ servers
in its \textsc{propose}{} message.
A server will reject the \textsc{propose}{} message
if it finds the contents of this message have been manipulated
by checking hash and signatures.
Thus, we make the following contributions:
\begin{itemize}
\item We define Byzantine Causal Consistency (Byz-CC),
which is a causal consistency variant in the presence of Byzantine faults
(Section~\ref{section:cc}).
\item We present Byz-GentleRain,
the first Byzantine-tolerant causal consistency protocol.
It tolerates up to $f$ Byzantine servers among $3f + 1$ ones
and any number of Byzantine clients
(Section~\ref{section:protocol}).
All reads and updates complete in one round-trip.
\item We evaluate Byz-GentleRain{} on Aliyun.
The preliminary results show that Byz-GentleRain{} is efficient
on typical workloads (Section~\ref{section:evaluation}).
\end{itemize}
Section~\ref{section:model} describes the system model and failure model.
Section~\ref{section:related-work} discusses related work.
Section~\ref{section:conclusion} concludes the paper.
The proofs can be found in Appendix~\ref{section:proof}.
\section{Model} \label{section:model}
\input{tables/notations}
We adopt the client/server architecture~\cite{Liskov:ICDCS16, ByzRCM:NAC19},
in which each client or server has its unique id.
Table~\ref{table:notations} summarizes the notations used in this paper.
\subsection{System Model} \label{ss:model}
We consider a distributed multi-version key-value store,
which maintains keys in the set ${\sf Key}$ (ranged over by $\mathit{k}$)
with values in the set ${\sf Val}$ (ranged over by $\mathit{v}$).
Each value is associated with a unique version,
consisting of the timestamp of the update which creates this version
and the id of the client which issues this update.
We denote by ${\sf VVal}$ (ranged over by $\mathit{vv}$) the set of versioned values.
The distributed key-value store runs at $\D$ data centers,
each of which has a full copy of data.
In each data center, the full data is sharded into $\P$ partitions.
For a key $\mathit{k} \in {\sf Key}$,
we use $\textsc{partition}(\mathit{k})$ to denote the partition that holds $\mathit{k}$.
Each partition is replicated across $D$ data centers.
For a partition $\mathit{p}$,
we use $\textsc{replicas}(\mathit{p})$ to denote the set of replicas of $\mathit{p}$.
For a data center $\mathit{d}$,
we use $\textsc{datacenter}(\mathit{d})$ to denote the set of servers in $\mathit{d}$.
We denote by $\replica{\mathit{p}}{\mathit{d}}$
the replica of partition $\mathit{p}$ at data center $\mathit{d}$.
We denote by $\mathcal{S}$ the set of all $\D \times \P$ servers in the key-value store.
For each individual partition $\mathit{p}$,
we call a set $\Q$ of $2f + 1$ replicas in $\textsc{replicas}(\mathit{p})$ a quorum
and denote it by ${\sf quorum}(\Q)$.
For convenience,
we model the key-value store at replica $\replica{\mathit{p}}{\mathit{d}}$,
denoted ${\sf store}^{\mathit{p}}_{\mathit{d}}$, by a set of (unique) versioned values.
That is, ${\sf store}^{\mathit{p}}_{\mathit{d}} \subseteq {\sf VVal}$.
We denote by $\textsc{Store}$ the union of stores at all replicas.
That is, $\textsc{Store} \triangleq \bigcup_{1 \le i \le \D, 1 \le j \le \P}
{\sf store}^{\jvar}_{\ivar}$.
The distributed key-value store offers two operations to clients:
\begin{itemize}
\item $\textsc{get}(\mathit{k})$.
A get operation which returns the value of some version of key $\mathit{k}$.
\item $\put(\mathit{k}, \mathit{v})$.
A put operation which updates key $\mathit{k}$ with value $\mathit{v}$.
This creates a new version of $\mathit{k}$.
\end{itemize}
We assume that each client or server is equipped with a physical clock,
which is monotonically increasing.
Clocks at different clients are loosely synchronized by a protocol
such as NTP~\footnote{NTP: The Network Time Protocol. \url{http://www.ntp.org/}.}.
The correctness of Byz-GentleRain{} does not
depend on the precision of clock synchronization,
but large clock drifts may negatively impact its performance.
\subsection{Failure Model} \label{ss:failure-model}
Clients and servers are either correct or faulty.
Correct clients and servers obey their protocols,
while faulty ones may exhibit Byzantine behaviors~\cite{PBFT:Thesis00},
by deviating arbitrarily from their protocols.
We assume asynchronous point-to-point communication channels
among clients and servers.
Messages may be delayed, duplicated, corrupted, or delivered out of order.
We do not assume known bounds on message delays.
The communication network is fully connected.
We require that if the two ends of a channel are both correct
and the sender keeps retransmitting a message,
then the message can eventually be delivered.
We also assume the channels are authenticated.
Clients and servers can sign messages using digital signatures when needed.
A message $\mathit{m}$ signed by a client $\mathit{c}$ or a replica $\replica{\mathit{p}}{\mathit{d}}$
is denoted by $\sign{\mathit{m}}{}{\mathit{c}}$ or $\sign{\mathit{m}}{\mathit{p}}{\mathit{d}}$, respectively.
We denote by ${\sf valid}(\mathit{m})$ that $m$ is valid in signatures.
We also use a cryptographic hash function ${\sf hash}()$,
which is assumed to be collision-resistant:
the probability of an adversary producing inputs $\mathit{m}$ and $\mathit{m}'$
such that ${\sf hash}(\mathit{m}) = {\sf hash}(\mathit{m}')$
is negligible~\cite{PBFT:Thesis00,ByzLive:DISC20}.
\section{Correctness of Byz-GentleRain} \label{section:proof}
We show that Byz-GentleRain{} satisfies Byz-CC.
We assume that single-shot PBFT is correct
and refer its detailed correctness proof to~\cite{ByzLive:DISC20}.
Table~\ref{table:predicates} gives the definitions
of the predicates ${\sf ValidNewLeader}$ and ${\sf safe\_propose}$
used in Algorithm~\ref{alg:cgst},
which are also adapted from~\cite{ByzLive:DISC20}.
\begin{remark} \label{remark:notations}
In the following, we use $R$ and $W$ to denote the set of
\textsc{get}{} and \put{} operations, respectively.
We also define $\O \triangleq R \cup W$
to denote the set of all operations.
For a variable, e.g., ${\sf clock}_{\mathit{c}}$ at client $\mathit{c}$,
we refer to its value at time $\sigma$
by, e.g., $({\sf clock}_{\mathit{c}})_{\sigma}$.
\end{remark}
According to the description of Algorithms~\ref{alg:client}
and~\ref{alg:replica},
\begin{lemma} \label{lemma:rules}
\emph{\textsc{Rule}~\ref{rule:get}--\textsc{Rule}~\ref{rule:get-ts}}
are maintained by \emph{Byz-GentleRain}.
\end{lemma}
\begin{lemma} \label{lemma:invs}
\emph{\textsc{Inv}~\ref{inv:cgst-c-put}--\textsc{Inv}~\ref{inv:cgst-updates}}
are maintained by \emph{Byz-GentleRain}.
\end{lemma}
\begin{proof} \label{proof:invs}
\textsc{Inv}~\ref{inv:cgst-c-put} holds due to
line~\code{\ref{alg:client}}{\ref{line:put-wait-clock}}.
\textsc{Inv}~\ref{inv:cgst-replica} holds due to
the read rule at line~\code{\ref{alg:replica}}{\ref{line:putreq-pre}}.
By the correctness of single-shot PBFT~\cite{ByzLive:DISC20},
\textsc{Inv}~\ref{inv:cgst-updates} holds.
\end{proof}
\begin{definition}[Timestamps] \label{def:ts}
We use $\textsl{ts}(\mathit{o})$ to denote the timestamp of operation $\mathit{o}$,
which is defined as follows:
\begin{itemize}
\item For a \textsc{get}{} operation $\mathit{o}$,
$\textsl{ts}(\mathit{o})$ refers to the value of ``$\mathit{ts}$''
at line~\code{\ref{alg:client}}{\ref{line:get-ts}}.
\item For a \put{} operation $\mathit{o}$,
$\textsl{ts}(\mathit{o})$ refers to the value of ``$\mathit{cl}$''
at line~\code{\ref{alg:client}}{\ref{line:put-send-putreq}}.
\end{itemize}
\end{definition}
\begin{lemma} \label{lemma:so-w-w}
\[
(\mathit{w} \rel{{\sf so}} \mathit{w}' \land \mathit{w} \in W \land \mathit{w}' \in W) \implies
\textsl{ts}(\mathit{w}') > \textsl{ts}(\mathit{w}).
\]
\end{lemma}
\begin{proof} \label{proof:so-w-w}
Suppose that $\mathit{w}'$ is issued by client $\mathit{c}$
at time $\sigma'$.
\[
\textsl{ts}(\mathit{w}) \le ({\sf dt}_{c})_{\sigma'}
< ({\sf clock}_{\mathit{c}})_{\sigma'} = \textsl{ts}(\mathit{w}').
\]
\end{proof}
\begin{lemma} \label{lemma:so-r-w}
\[
(\mathit{r} \rel{{\sf so}} \mathit{w} \land \mathit{r} \in R \land \mathit{w} \in W) \implies
\textsl{ts}(\mathit{w}) > \textsl{ts}(\mathit{r}).
\]
\end{lemma}
\begin{proof} \label{proof:so-r-w}
Suppose that $\mathit{r}$ and $\mathit{w}$ are issued by correct client $\mathit{c}$
at time $\sigma_{\mathit{r}}$ and $\sigma_{\mathit{w}}$, respectively.
By line~\code{\ref{alg:client}}{\ref{line:get-ts}},
\[
\textsl{ts}(\mathit{r}) = \max\set{({\sf cgst}_{\mathit{c}})_{\sigma_{\mathit{r}}},
({\sf dt}_{\mathit{c}})_{\sigma_{\mathit{r}}}}.
\]
By \textsc{Inv}~\ref{inv:cgst-c-put},
\[
\textsl{ts}(\mathit{w}) > ({\sf cgst}_{\mathit{c}})_{\sigma_{\mathit{w}}}
> ({\sf cgst}_{\mathit{c}})_{\sigma_{\mathit{r}}}.
\]
Moreover,
\[
\textsl{ts}(\mathit{w}) > ({\sf dt}_{\mathit{c}})_{\sigma_{\mathit{w}}}
\ge ({\sf dt}_{\mathit{c}})_{\sigma_{\mathit{r}}}.
\]
Putting it together yields
\[
\textsl{ts}(\mathit{w}) > \textsl{ts}(\mathit{r}).
\]
\end{proof}
\begin{lemma} \label{lemma:so-w-r}
\[
(\mathit{w} \rel{{\sf so}} \mathit{r} \land \mathit{w} \in W \land \mathit{r} \in R) \implies
\textsl{ts}(\mathit{r}) \ge \textsl{ts}(\mathit{w}).
\]
\end{lemma}
\begin{proof} \label{proof:so-w-r}
Suppose that $\mathit{r}$ are issued by correct client $\mathit{c}$ at time $\sigma$.
By line~\code{\ref{alg:client}}{\ref{line:get-ts}},
\[
\textsl{ts}(\mathit{r}) = \max\set{({\sf cgst}_{\mathit{c}})_{\sigma},
({\sf dt}_{\mathit{c}})_{\sigma}}
\ge ({\sf dt}_{\mathit{c}})_{\sigma}.
\]
By line~\code{\ref{alg:client}}{\ref{line:put-dt}},
\[
({\sf dt}_{\mathit{c}})_{\sigma} \ge \textsl{ts}(\mathit{w}).
\]
Thus,
\[
\textsl{ts}(\mathit{r}) \ge \textsl{ts}(\mathit{w}).
\]
\end{proof}
\begin{lemma} \label{lemma:so-r-r}
\[
(\mathit{r} \rel{{\sf so}} \mathit{r}' \land \mathit{r} \in R \land \mathit{r}' \in R) \implies
\textsl{ts}(\mathit{r}') \ge \textsl{ts}(\mathit{r}).
\]
\end{lemma}
\begin{proof} \label{proof:so-r-r}
Suppose that $\mathit{r}$ and $\mathit{r}'$ are issued by correct client $\mathit{c}$
at time $\sigma$ and $\sigma'$, respectively.
By line~\code{\ref{alg:client}}{\ref{line:get-ts}},
\[
\textsl{ts}(\mathit{r}) = \max\set{({\sf cgst}_{\mathit{c}})_{\sigma},
({\sf dt}_{\mathit{c}})_{\sigma}},
\]
and
\[
\textsl{ts}(\mathit{r}') = \max\set{({\sf cgst}_{\mathit{c}})_{\sigma'},
({\sf dt}_{\mathit{c}})_{\sigma'}}.
\]
Moreover,
\[
({\sf cgst}_{\mathit{c}})_{\sigma'} \ge ({\sf cgst}_{\mathit{c}})_{\sigma}
\land ({\sf dt}_{\mathit{c}})_{\sigma'} \ge ({\sf dt}_{\mathit{c}})_{\sigma}.
\]
Thus,
\[
\textsl{ts}(\mathit{r}') \ge \textsl{ts}(\mathit{r}).
\]
\end{proof}
\begin{lemma} \label{lemma:so}
\[
\mathit{o} \rel{{\sf so}} \mathit{o}' \land \mathit{o} \in \O \land \mathit{o}' \in \O
\implies \textsl{ts}(\mathit{o}') \ge \textsl{ts}(\mathit{o}).
\]
\end{lemma}
\begin{proof} \label{proof:so}
By Lemmas~\ref{lemma:so-w-w}--\ref{lemma:so-r-r}.
\end{proof}
\begin{lemma} \label{lemma:read-from}
\[
\mathit{w} \rel{{\sf rf}} \mathit{r} \land \mathit{w} \in W \land \mathit{r} \in R
\implies \textsl{ts}(\mathit{r}) \ge \textsl{ts}(\mathit{w}).
\]
\end{lemma}
\begin{proof} \label{proof:read-from}
By the read rule at line~\code{\ref{alg:replica}}{\ref{line:get-val}}.
\end{proof}
\begin{lemma} \label{lemma:hb-ts}
\[
\mathit{o} \leadsto \mathit{o}' \land o \in \O \land o' \in \O
\implies \textsl{ts}(\mathit{o}') \ge \textsl{ts}(\mathit{o}).
\]
\end{lemma}
\begin{proof} \label{proof:hb-ts}
By Lemmas~\ref{lemma:so} and~\ref{lemma:read-from}.
\end{proof}
\begin{lemma} \label{lemma:local-cc}
Consider $\mathit{r} \in R$ and $\mathit{w} \in W$.
Suppose $\mathit{r}$ reads from some value
at a correct replica $\replica{\mathit{p}}{\mathit{d}}$
at time $\sigma$
(line~\code{\ref{alg:replica}}{\ref{line:getreq-value}}).
If $\mathit{w}$ would be added to ${\sf store}^{\mathit{p}}_{\mathit{d}}$
at a later time than $\sigma$
(line~\code{\ref{alg:replica}}{\ref{line:putreq-store}})
then $\textsl{ts}(\mathit{w}) > \textsl{ts}(\mathit{r})$.
\end{lemma}
\begin{proof} \label{proof:local-cc}
By \textsc{Rule}~\ref{rule:get},
\[
({\sf cgst}^{\mathit{p}}_{\mathit{d}})_{\sigma} \ge \textsl{ts}(\mathit{r}).
\]
By \textsc{Inv}~\ref{rule:put},
\[
\textsl{ts}(\mathit{w}) > ({\sf cgst}^{\mathit{p}}_{\mathit{d}})_{\sigma}.
\]
Thus,
\[
\textsl{ts}(\mathit{w}) > \textsl{ts}(\mathit{r}).
\]
\end{proof}
\begin{lemma} \label{lemma:partition-cc}
Consider $\mathit{r} \in R$ and $\mathit{w} \in W$.
Suppose the successful $\mathit{r}$ returns at time $\sigma$
(line~\code{\ref{alg:replica}}{\ref{line:getreq-value}})
and the successful
$\mathit{w}$ starts at a later time than $\sigma$ in partition $\mathit{p}$.
Then $\lnot(\mathit{w} \leadsto \mathit{r}$).
\end{lemma}
\begin{proof} \label{proof:partition-causality}
By lines~\ref{alg:client}{\ref{line:get-wait-receive-getack}}
and \ref{alg:client}{\ref{line:put-wait-receive-putack}},
there is a \emph{correct} replica at which
$\mathit{r}$ obtains its value
(line~\code{\ref{alg:replica}}{\ref{line:getreq-value}})
\emph{before} $\mathit{w}$ is added to the store
(line~\code{\ref{alg:replica}}{\ref{line:putreq-store}}).
By Lemma~\ref{lemma:local-cc},
\[
\textsl{ts}(\mathit{w}) > \textsl{ts}(\mathit{r}).
\]
By Lemma~\ref{lemma:hb-ts},
\[
\lnot(\mathit{w} \leadsto \mathit{r}).
\]
\end{proof}
\begin{theorem} \label{thm:byzcc}
\emph{Byz-GentleRain} satisfies Byz-CC{}.
That is, when a certain \put{} operation is visible to a client,
then so are all of its causal dependencies.
\end{theorem}
\begin{proof} \label{proof:byzcc}
By Lemmas~\ref{lemma:partition-cc}
and \ref{lemma:cross-partition-cc}.
\end{proof}
\begin{lemma} \label{lemma:w-successful}
Suppose a \put{} operation $\mathit{w}$ successfully
returns in partition $\mathit{p}$ at time $\sigma$
(line~\code{\ref{alg:client}}{\ref{line:put-wait-receive-putack}}).
Then, it will eventually be in ${\sf store}^{\mathit{p}}_{\ivar}$
for each correct data center $\ivar$.
\end{lemma}
\begin{proof} \label{proof:w-successful}
By line~\code{\ref{alg:client}}{\ref{line:put-wait-receive-putack}}
and line~\code{\ref{alg:cgst}}{\ref{line:newleader-wait-receive-collectack}},
there is a \emph{correct} replica in partition $\mathit{p}$ at which
$\mathit{w}$ is added to the store
(line~\code{\ref{alg:replica}}{\ref{line:putreq-store}})
before it is sent to the PBFT leader in the $\textsc{collect\_ack}$ message
(line~\code{\ref{alg:cgst}}{\ref{line:collect-send-collectack}}).
By the correctness of single-shot PBFT~\cite{ByzLive:DISC20},
it will eventually be in ${\sf store}^{\mathit{p}}_{\ivar}$
for each correct data center $\ivar$.
\end{proof}
\begin{lemma} \label{lemma:update-visible}
Let $\mathit{w}$ be a successful \put{} operation.
Then, eventually for each correct replica
$\replica{\jvar}{\ivar}$ ($1 \le \ivar \le \D, 1 \le \jvar \le \P$),
${\sf cgst}^{\jvar}_{\ivar} \ge \textsl{ts}(\mathit{w})$.
\end{lemma}
\begin{proof}
Suppose $\mathit{w}$ successfully returns in partition $\mathit{p}$.
Then it is added to the stores of
at least $f+1$ \emph{correct} replicas in $\textsc{replicas}(\mathit{p})$.
By Algorithm~\ref{alg:metadata},
eventually for each correct replica $\replica{\jvar}{\ivar}$,
${\sf lst}^{\jvar}_{\ivar} \ge \textsl{ts}(\mathit{w})$ and
${\sf gst}^{\jvar}_{\ivar} \ge \textsl{ts}(\mathit{w})$.
By Algorithm~\ref{alg:cgst},
${\sf cgst}^{\jvar}_{\ivar} \ge \textsl{ts}(\mathit{w})$.
\end{proof}
\begin{lemma}
Suppose a correct replica sends a $\textsc{new\_cgst}(\mathit{gst})$ message
(line~\code{\ref{alg:metadata}}{\ref{line:bc-send-newcgst}}).
Then, there is a $\mathit{cgst} \ge \mathit{gst}$ such that
eventually for each correct replica $\replica{\jvar}{\ivar}$
($1 \le \ivar \le \D, 1 \le \jvar \le \P$),
${\sf cgst}^{\jvar}_{\ivar} \ge \mathit{cgst}$.
\end{lemma}
\begin{proof}
A byzantine replica can propose $gst^{\mathit{p}}_{\mathit{d}}$
to replica $\replica{\mathit{p}}{\mathit{d}}$,
but it will take effect only when
$gst^{\mathit{p}}_{\mathit{d}} < lst^{\mathit{p}}_{\mathit{d}}$
maintained by $\replica{\mathit{p}}{\mathit{d}}$.
If $gst^{\mathit{p}}_{\mathit{d}}$ is smaller than
a $gst'$ broadcast by a correct replica, then it will be overwritten at $\replica{\mathit{p}}{\mathit{d}}$.
And eventually, there must be a $gst$ version $\ge \replica{\mathit{p}}{\mathit{d}}$ broadcast by a correct replica.
Thus it will eventually by overwritten by a correct $gst$ version.
\end{proof}
\begin{lemma} \label{lemma:cross-partition-cc}
Consider $\mathit{r} \in R$ and $\mathit{w} \in W$.
Suppose the successful $\mathit{r}$ returns at time $\sigma$
(line~\code{\ref{alg:replica}}{\ref{line:getreq-value}})
and the successful
$\mathit{w}$ starts at a later time than $\sigma$
in any partition $\jvar \neq \mathit{p}$.
Then $\lnot(\mathit{w} \leadsto \mathit{r}$).
\end{lemma}
\begin{proof} \label{proof:across-partition-cc}
By Algorithm~\ref{alg:replica}\ref{line:putreq-pre}, $\textsl{ts}{\mathit{w}} > {\sf lst}$
at each replca who accept $\mathit{w}$.
By Algorithm~\ref{alg:replica}\ref{line:getreq-wait-until},
$\textsl{ts}{\mathit{r}} \le {\sf cgst}$ at each replica who reply it.
Since ${\sf lst}$ at any correct correct server is an upbound of
all the ${\sf cgst}$ at all correct replicas, $tsof(\mathit{w}) > {\sf lst} > {\sf cgst} > \textsl{ts}(\mathit{r})$.
So $\lnot(\mathit{w} \leadsto \mathit{r})$.
\end{proof}
\section{The Byz-GentleRain Protocol} \label{section:protocol}
As discussed in Section~\ref{section:intro},
it is the clients in Byz-GentleRain{} that are responsible
for generating totally ordered timestamps for updates.
Specifically, when a client issues an update,
it assigns to the update a timestamp
consisting of its current clock and identifier.
As in GentleRain, we also distinguish between the updates
that have been received by a server
and those that have been made visible to clients.
Byz-GentleRain{} guarantees that
an update can be made visible to clients
only if so are all its causal dependencies.
The pseudocode in Algorithms~\ref{alg:client}--\ref{alg:metadata}
dealing with Byzantine faults is underlined.
\input{sections/key-design}
\input{sections/client-op}
\input{sections/replica-op}
\input{sections/metadata-op}
\input{sections/cgst-op}
\section{Related Work} \label{section:related-work}
As far as we know, Byz-RCM~\cite{ByzRCM:NAC19}
is the only causal consistency protocol
that considers Byzantine faults.
It achieves causal consistency in the client-server model with $3f + 1$ servers
where up to $f$ servers may suffer Byzantine faults,
and any number of clients may crash.
Byz-RCM has also been shown optimal in terms of failure resilience.
However, Byz-RCM did not tolerate \emph{Byzantine clients},
and thus it could rely on clients' requests
to identify bogus requests from Byzantine servers~\cite{ByzRCM:NAC19}.
Linde \emph{et. al.}~\cite{ClientSide:VLDB20}
consider \emph{peer-to-peer} architecture.
A centralized server maintains the application data,
while clients replicate a subset of data
and can directly communicate with each other.
They analyze the possible attacks of clients to causal consistency
(the centralized server is assumed correct),
derive a secure form of causal consistency,
and propose practical protocols for implementing it.
Liskov and Rodrigues extend the notion of linearizability~\cite{Lin:TOPLAS90}
and define BFT-linearizability in the presence of
Byzantine servers and clients~\cite{Liskov:ICDCS16}.
They also design protocols that achieve BFT-linearizability
despite Byzantine clients.
The protocols require $3f + 1$ replicas
of which up to $f$ replicas may be Byzantine.
They are quite efficient for linearizable systems:
Writes complete in two or three phases,
while reads complete in one or two phases.
Auvolat \emph{et. al.}~\cite{BTCB:TCS21}
defines a Byzantine-tolerant Causal Order broadcast
(BCO-broadcast) abstraction
and proposes an implementation for it.
However, as a communication primitive for replicas,
BCO-broadcast does not capture the get/put semantics
from the perspective of clients.
Thus, it does not prevent Byzantine clients
from violating causality.
\section{Acknowledgements}
This work was partially supported by the CCF-Tencent Open Fund
(CCF-Tencent RAGR20200124) and
the National Natural Science Foundation of China (No. 61772258).
\subsection{Key Designs} \label{ss:design}
In Byz-GentleRain{}, both clients and servers
maintain a \emph{common global stable time} ${\sf cgst}$.
We denote the ${\sf cgst}$ at client $\mathit{c}$ by ${\sf cgst}_{\mathit{c}}$
and that at replica $\replica{\mathit{p}}{\mathit{d}}$ by
${\sf cgst}^{\mathit{p}}_{\mathit{d}}$.
We maintain the following invariants
that are key to the correctness of Byz-GentleRain:
\begin{enumerate}[\textsc{Inv}~(I):]
\item \label{inv:cgst-c-put}
Consider ${\sf cgst}_{\mathit{c}}$ at any time $\sigma$.
All updates issued by correct client $\mathit{c}$ after time $\sigma$
have a timestamp $> {\sf cgst}_{\mathit{c}}$.
\item \label{inv:cgst-replica}
Consider ${\sf cgst}^{\mathit{p}}_{\mathit{d}}$ at any time $\sigma$.
No updates with timestamps $\le {\sf cgst}^{\mathit{p}}_{\mathit{d}}$
will be successfully executed at $> f$ correct replicas
in $\textsc{replicas}(\mathit{p})$ after time $\sigma$.
\item \label{inv:cgst-updates}
Consider a $\mathit{cgst}$ value.
For any two correct replicas $\replica{\mathit{p}}{\mathit{d}}$
and $\replica{\mathit{p}}{\ivar}$ (where $i \neq d$) of partition $\mathit{p}$,
if ${\sf cgst}^{\mathit{p}}_{\mathit{d}} \ge \mathit{cgst}$
and ${\sf cgst}^{\mathit{p}}_{\ivar} \ge \mathit{cgst}$,
then the updates with timestamps $\le \mathit{cgst}$
in ${\sf store}^{\mathit{p}}_{\mathit{d}}$ and ${\sf store}^{\mathit{p}}_{\ivar}$
are the same.
\end{enumerate}
Byz-GentleRain{} further enforces the following rules
for reads and updates:
\begin{enumerate}[\textsc{Rule}~(I):]
\item \label{rule:get}
For a correct replica $\replica{\mathit{p}}{\mathit{d}}$,
any updates with timestamps $> {\sf cgst}^{\mathit{p}}_{\mathit{d}}$
in ${\sf store}^{\mathit{p}}_{\mathit{d}}$ are invisible to any clients.
\item \label{rule:put}
Any correct replica $\replica{\mathit{p}}{\mathit{d}}$
will reject any updates with timestamps $\le {\sf cgst}^{\mathit{p}}_{\mathit{d}}$.
\item \label{rule:get-ts}
For a read operation with timestamp $\mathit{ts}$
issued by client $\mathit{c}$,
any correct replica $\replica{\mathit{p}}{\mathit{d}}$
that receives this operation
must wait until ${\sf cgst}^{\mathit{p}}_{\mathit{d}} \ge \mathit{ts}$
before it returns a value to client $\mathit{c}$.
\end{enumerate}
In the following sections,
we explain how these invariants and rules are implemented
and why they are important to the correctness.
\subsection{Sign checking} \label{sign checking}
\begin{definition}
Safe_
\end{definition}
\subsection{Client Operations} \label{ss:client}
\input{algs/client}
To capture the session order,
each client maintains a dependency time ${\sf dt}$,
which is the clock value of its last put operation.
When a client $\mathit{c}$ issues a $\textsc{get}$ operation on key $\mathit{k}$,
it first takes as $\mathit{ts}$ the minimum of its dependency time ${\sf dt}_{\mathit{c}}$
and common global stable time ${\sf cgst}_{\mathit{c}}$
(line~\code{\ref{alg:client}}{\ref{line:get-ts}}).
Then it sends a $\textsc{get\_req}$ request with $\mathit{ts}$ to $\textsc{replicas}(\mathit{p})$
(line~\code{\ref{alg:client}}{\ref{line:get-send-getreq}}),
where $\mathit{p}$ is the partition holding $\mathit{k}$
(line~\code{\ref{alg:client}}{\ref{line:get-partition}}).
Next, the client waits to receive a set $\M$ of
$\textsc{get\_ack}$ responses from a quorum $Q$ of $\textsc{replicas}(\mathit{p})$
(line~\code{\ref{alg:client}}{\ref{line:get-wait-receive-getack}}).
For each replica $\replica{\mathit{p}}{\ivar} \in Q$,
the $\textsc{get\_ack}$ response from $\replica{\mathit{p}}{\ivar}$
carries a value $\mathit{v}_{i}$ of key $\mathit{k}$
and its common global stable time $\mathit{cgst}_{i}$
when $\replica{\mathit{p}}{\ivar}$ computes the value $\mathit{v}_{i}$ to return
(line~\code{\ref{alg:replica}}{\ref{line:getreq-value}},
discussed in Section~\ref{ss:replica}).
The client takes the minimum $\mathit{cgst}_{i}$ for $\replica{\mathit{p}}{\ivar} \in Q$,
and uses it to update ${\sf cgst}_{\mathit{c}}$ if the latter is smaller
(line~\code{\ref{alg:client}}{\ref{line:get-cgst}}).
Since there are at most $f$ Byzantine replicas in a partition,
at least $f + 1$ \textsc{get\_ack}{} responses are from correct replicas.
By \textsc{Rule}~(\ref{rule:get-ts}) and \textsc{Inv}~(\ref{inv:cgst-updates}),
these responses from correct replicas
contain the same value, denoted $\mathit{v}$.
Hence, $\mathit{v}$ is the majority $\mathit{v}_{i}$ in $\M$
(line~\code{\ref{alg:client}}{\ref{line:get-val}}).
Finally, the client returns $\mathit{v}$
(line~\code{\ref{alg:client}}{\ref{line:get-return}}).
When a client $\mathit{c}$ issues a \put{} operation on key $\mathit{k}$ with value $\mathit{v}$,
it sends a \textsc{put\_req}{} request carrying its ${\sf clock}_{\mathit{c}}$
and id $\mathit{c}$ to $\textsc{replicas}(\mathit{p})$
(line~\code{\ref{alg:client}}{\ref{line:put-send-putreq}}),
where $\mathit{p}$ is the partition holding $\mathit{k}$
(line~\code{\ref{alg:client}}{\ref{line:put-partition}}).
Next the client waits to receive a set $\M$ of \textsc{put\_ack}{} responses
from a quorum $Q$ of $\textsc{replicas}(\mathit{p})$
(line~\code{\ref{alg:client}}{\ref{line:put-wait-receive-putack}}).
For each replica $\replica{\mathit{p}}{\ivar} \in Q$,
the \textsc{put\_ack}{} response from $\replica{\mathit{p}}{\ivar}$
carries its common global stable time $\mathit{cgst}_{i}$.
Then, the client takes the minimum $\mathit{cgst}_{i}$
for $\replica{\mathit{p}}{\ivar} \in Q$,
and uses it to update ${\sf cgst}$ if the latter is smaller
(line~\code{\ref{alg:client}}{\ref{line:put-cgst}}).
Finally, ${\sf dt}_{\mathit{c}}$ is set to the current ${\sf clock}_{\mathit{c}}$
(line~\code{\ref{alg:client}}{\ref{line:put-dt}}).
\subsection{Operation Executions at Replicas} \label{ss:replica}
\input{algs/replica}
When a replica $\replica{\mathit{p}}{\mathit{d}}$
receives a $\textsc{get\_req}(\mathit{k}, \mathit{ts})$ request from some client $\mathit{c}$,
it first waits until ${\sf cgst} \ge \mathit{ts}$
(line~\code{\ref{alg:replica}}{\ref{line:getreq-wait-until}})
where $\mathit{ts} \triangleq \max\set{{\sf dt}_{\mathit{c}}, {\sf cgst}_{\mathit{c}}}$
(line~\code{\ref{alg:client}}{\ref{line:get-ts}}).
This implements \textsc{Rule}~\ref{rule:get-ts},
and is used to ensure the session guarantee on client $\mathit{c}$
and eventual visibility of updates to $\mathit{c}$.
Then the replica obtains the value $\mathit{v}$ of key $\mathit{k}$
in ${\sf store}$ which has the largest timestamp before $\mathit{ts}$,
breaking ties with client ids
(line~\code{\ref{alg:replica}}{\ref{line:getreq-value}}).
Finally, it sends a signed $\textsc{get\_ack}$ response,
along with the value $\mathit{v}$ and its current ${\sf cgst}$, to client $\mathit{c}$
(line~\code{\ref{alg:replica}}{\ref{line:getreq-send-getack}}).
When a replica $\replica{\mathit{p}}{\mathit{d}}$
receives a $\textsc{put\_req}(\mathit{k}, \mathit{v}, \mathit{cl}, \mathit{c})$ request from client $\mathit{c}$,
it first checks the precondition $\mathit{cl} \ge {\sf curr\_cgst}^{\mathit{p}}_{\mathit{d}}$
(line~\code{\ref{alg:replica}}{\ref{line:putreq-pre}}).
This enforces \textsc{Rule}~\ref{rule:put},
and prevents fabricated updates with timestamps $\le {\sf cgst}$
at ${\sf store}^{\mathit{p}}_{\mathit{d}}$ from now on.
If the precondition holds,
the replica adds the new versioned version
$\mathit{vv} \triangleq \sign{\mathit{k}, \mathit{v}, \mathit{cl}, \mathit{c}}{}{\mathit{c}}$ signed by $\mathit{c}$
(line~\code{\ref{alg:replica}}{\ref{line:putreq-vv}})
to ${\sf store}^{\mathit{p}}_{\mathit{d}}$
(line~\code{\ref{alg:replica}}{\ref{line:putreq-store}}).
Then, the replica sends a signed $\textsc{put\_ack}$ response,
with its ${\sf cgst}^{\mathit{p}}_{\mathit{d}}$, to client $\mathit{c}$
(line~\code{\ref{alg:replica}}{\ref{line:putreq-send-putack}}).
Finally, it broadcasts a signed $\textsc{replicate}$ message with $\mathit{vv}$
to other replicas in partition $\mathit{p}$
(line~\code{\ref{alg:replica}}{\ref{line:putreq-send-replicate}}).
\subsection{Metadata} \label{ss:metadata}
\subsubsection{Replica States} \label{sss:replica-states}
Each replica $\replica{\mathit{p}}{\mathit{d}}$ maintains
a \emph{version vector} ${\sf VV}^{\mathit{p}}_{\mathit{d}}$ of size $\D$,
with each entry per data center.
For data center $\mathit{d}$, ${\sf VV}^{\mathit{p}}_{\mathit{d}}[\mathit{d}]$
is the timestamp of the last update that happens at $\replica{\mathit{p}}{\mathit{d}}$.
For data center $\ivar \neq \mathit{d}$,
${\sf VV}^{\mathit{p}}_{\mathit{d}}[\ivar]$ is the largest timestamp
of the updates that happened at replica $\replica{\mathit{p}}{\ivar}$
and have been propagated to $\replica{\mathit{p}}{\mathit{d}}$.
For fault-tolerance, we compute the \emph{local stable time} ${\sf lst}^{\mathit{p}}_{\mathit{d}}$
at replica $\replica{\mathit{p}}{\mathit{d}}$
as the \emph{$(f+1)$-st minimum} element of its ${\sf VV}^{\mathit{p}}_{\mathit{d}}$.
Each replica $\replica{\mathit{p}}{\mathit{d}}$ also maintains a
\emph{lst vector} ${\sf LV}^{\mathit{p}}_{\mathit{d}}$ of size $\P$,
with each entry per partition.
For partition $1 \le \jvar \le \P$,
${\sf LV}^{\mathit{p}}_{\mathit{d}}[\jvar]$ is the largest $\mathit{lst}$
of replica $\replica{\jvar}{\mathit{d}}$
of which $\replica{\mathit{p}}{\mathit{d}}$ is aware.
We compute the \emph{global stable time} ${\sf gst}^{\mathit{p}}_{\mathit{d}}$
at replica $\replica{\mathit{p}}{\mathit{d}}$
as the minimum element of its ${\sf LV}^{\mathit{p}}_{\mathit{d}}$.
That is, ${\sf gst}^{\mathit{p}}_{\mathit{d}} \triangleq
\min_{1 \le \jvar \le P} {\sf LV}^{\mathit{p}}_{\mathit{d}}[\jvar]$.
Each replica $\replica{\mathit{p}}{\mathit{d}}$
periodically synchronizes their $\mathit{gst}^{\mathit{p}}_{\mathit{d}}$
with others via PBFT,
and maintains a \emph{common global stable time} ${\sf cgst}^{\mathit{p}}_{\mathit{d}}$.
We discuss it in Section~\ref{ss:cgst}.
\subsubsection{Propagation} \label{sss:propagation}
\input{algs/metadata}
As in GentleRain,
Byz-GentleRain{} propagates and updates metadata in the background.
Once a new version $\mathit{vv}$ is created at replica $\replica{\mathit{p}}{\mathit{d}}$,
the replica sends a signed $\textsc{replicate}(\mathit{vv})$ message
to other replicas of partition $\mathit{p}$
(line~\code{\ref{alg:replica}}{\ref{line:putreq-send-replicate}}).
When replica $\replica{\mathit{p}}{\mathit{d}}$ receives a $\textsc{replicate}(\mathit{vv})$ message
from another replica $\replica{\mathit{p}}{i}$ in data center $i \neq d$,
it stores $\mathit{vv}$ in its ${\sf store}^{\mathit{p}}_{\mathit{d}}$
(line~\code{\ref{alg:metadata}}{\ref{line:replicate-store}}),
and updates ${\sf VV}^{\mathit{p}}_{\mathit{d}}[\ivar]$ to $\mathit{vv}.\mathit{cl}$
if the latter is larger
(line~\code{\ref{alg:metadata}}{\ref{line:replicate-vv}}).
Each replica $\replica{\mathit{p}}{\mathit{d}}$ periodically
computes its ${\sf lst}^{\mathit{p}}_{\mathit{d}}$
(line~\code{\ref{alg:metadata}}{\ref{line:broadcast-lst}})
and sends a signed $\textsc{broadcast}({\sf lst}^{\mathit{p}}_{\mathit{d}})$ message
to $\textsc{datacenter}(\mathit{d})$, all the servers in data center $\mathit{d}$
(line~\code{\ref{alg:metadata}}{\ref{line:broadcast-send-bc}}).
When replica $\replica{\mathit{p}}{\mathit{d}}$ receives
a $\textsc{broadcast}(\mathit{lst})$ message from another replica $\replica{\jvar}{\mathit{d}}$
in data center $d$, it updates ${\sf LV}^{\mathit{p}}_{\mathit{d}}$
and ${\sf gst}^{\mathit{p}}_{\mathit{d}}$ accordingly
(lines~\code{\ref{alg:metadata}}{\ref{line:bc-lv}}
and~\code{\ref{alg:metadata}}{\ref{line:bc-gst}}).
If the new ${\sf gst}^{\mathit{p}}_{\mathit{d}}$
is larger than ${\sf cgst}^{\mathit{p}}_{\mathit{d}}$
(line~\code{\ref{alg:metadata}}{\ref{line:bc-gst-if}}),
the replica $\replica{\mathit{p}}{\mathit{d}}$ sends a
signed $\textsc{new\_cgst}({\sf gst}^{\mathit{p}}_{\mathit{d}})$ message
to all servers $\mathcal{S}$ of the key-value store
(line~\code{\ref{alg:metadata}}{\ref{line:bc-send-newcgst}}).
To ensure liveness, a replica $\replica{\mathit{p}}{\mathit{d}}$
periodically (e.g., at time interval $\Delta$;
line~\code{\ref{alg:metadata}}{\ref{line:heartbeat-pre}})
sends a signed $\textsc{hb}({\sf clock}^{\mathit{p}}_{\mathit{d}})$ heartbeat
to $\textsc{replicas}(\mathit{p})$
(line~\code{\ref{alg:metadata}}{\ref{line:heartbeat-send-hb}}).
When replica $\replica{\mathit{p}}{\mathit{d}}$ receives
a heartbeat $\textsc{hb}(\mathit{clock})$ message
from replica $\replica{\mathit{p}}{\ivar}$,
it updates its ${\sf VV}^{\mathit{p}}_{\mathit{d}}[\ivar]$ to $\mathit{clock}$
if the latter is larger
(line~\code{\ref{alg:metadata}}{\ref{line:hb-vv}}).
\subsection{Synchronization of Global Stable Time} \label{ss:cgst}
\input{algs/cgst}
Each individual partition $\mathit{p}$ independently
runs PBFT~\cite{PBFT:Thesis00} to reach agreement on
a common global stable time $\mathit{cgst}$
\emph{and} the same set of updates before $\mathit{cgst}$
across $\textsc{replicas}(\mathit{p})$ (Algorithm~\ref{alg:cgst}).
We follow the pseudocode of single-shot PBFT
described in~\cite{ByzLive:DISC20},
and refer its detailed description
and correctness proof to~\cite{ByzLive:DISC20}.
In the following, we elaborate the parts
specific to synchronization of global stable time;
see the pseudocode underlined in Algorithm~\ref{alg:cgst}.
When replica $\replica{\mathit{p}}{\mathit{d}}$ receives
a $\textsc{new\_cgst}(\mathit{gst})$ message,
it first checks whether $\mathit{gst} \le {\sf lst}^{\mathit{p}}_{\mathit{d}}$ as expected
and $\mathit{gst} > {\sf curr\_cgst}^{\mathit{p}}_{\mathit{d}}$
which means that no smaller global stable time has been
or is being synchronized
(line~\code{\ref{alg:cgst}}{\ref{line:newcgst-pre}}).
If so, it sets ${\sf curr\_cgst}$ to $\mathit{gst}$
(line~\code{\ref{alg:cgst}}{\ref{line:newcgst-currcgst}}).
Now the replica stops accepting updates
with timestamps $< {\sf curr\_cgst}^{\mathit{p}}_{\mathit{d}}$
(line~\code{\ref{alg:replica}}{\ref{line:putreq-pre}}).
Then it triggers a $\textsc{newview}$ action
with a $\mathit{view}$ larger than ${\sf curr\_view}^{\mathit{p}}_{\mathit{d}}$
(line~\code{\ref{alg:cgst}}{\ref{line:newcgst-call-newview}}).
As in classic PBFT~\cite{PBFT:Thesis00,ByzLive:DISC20},
the $\textsc{newview}(\mathit{view})$ action can also be triggered spontaneously,
due to timeout, or by failure detectors.
When it is triggered at a replica $\replica{\mathit{p}}{\mathit{d}}$,
the replica will send a signed $\textsc{newleader}$ message
to the leader ${\sf leader}(\mathit{view})$ of $\mathit{view}$ in $\textsc{replicas}(\mathit{p})$
(line~\code{\ref{alg:cgst}}{\ref{line:newview-send-newleader}}).
The $\textsc{newleader}$ message carries both ${\sf curr\_cgst}^{\mathit{p}}_{\mathit{d}}$
and ${\sf prepared\_store}^{\mathit{p}}_{\mathit{d}}$
which is the set of updates collected in ${\sf prepared\_view}^{\mathit{p}}_{\mathit{d}}$.
When replica $\replica{\mathit{p}}{\mathit{d}}$ receives
a set $\M$ of $\textsc{newleader}$ messages
from a quorum $Q$ of $\textsc{replicas}(\mathit{p})$,
it selects as its proposal from $\M$
the set $\mathit{store}_{\jvar}$ of collected updates
that is prepared in the highest view, say $\mathit{view}_{\ivar}$
(line~\code{\ref{alg:cgst}}{\ref{line:newleader-send-propose-if}}),
or, if there are no such $\mathit{store}_{\jvar}$, its own proposal.
In the latter case, the replica sets its ${\sf curr\_cgst}^{\mathit{p}}_{\mathit{d}}$
to the maximum of $\mathit{cgst}_{\ivar}$ in $Q$
that are $\le {\sf lst}^{\mathit{p}}_{\mathit{d}}$
(line~\code{\ref{alg:cgst}}{\ref{line:newleader-currcgst}}).
Then, it sends a signed $\textsc{collect}({\sf curr\_cgst}^{\mathit{p}}_{\mathit{d}})$
message to $\textsc{replicas}(\mathit{p})$
(line~\code{\ref{alg:metadata}}{\ref{line:newleader-send-collect}}),
and waits to receive enough $\textsc{collect\_ack}$ messages.
When replica $\replica{\mathit{p}}{\mathit{d}}$ receives
a $\textsc{collect}(\mathit{cgst})$ message from replica $\replica{\mathit{p}}{\ivar}$
and $\mathit{cgst}$ passes the precondition
(line~\code{\ref{alg:cgst}}{\ref{line:collect-pre}}),
it first sets its ${\sf curr\_cgst}^{\mathit{p}}_{\mathit{d}}$ to $\mathit{cgst}$
(line~\code{\ref{alg:cgst}}{\ref{line:collect-currcgst}}).
Now the replica stops accepting updates
with timestamps $< {\sf curr\_cgst}^{\mathit{p}}_{\mathit{d}}$
(line~\code{\ref{alg:replica}}{\ref{line:putreq-pre}}).
Then it sends a signed $\textsc{collect\_ack}(\mathit{cgst}, \mathit{st})$ message
back to $\replica{\mathit{p}}{\ivar}$
(line~\code{\ref{alg:cgst}}{\ref{line:collect-send-collectack}}),
where $\mathit{st}$ is the set of updates
with timestamps $\le {\sf curr\_cgst}^{\mathit{p}}_{\mathit{d}}$
in its ${\sf store}^{\mathit{p}}_{\mathit{d}}$
(line~\code{\ref{alg:cgst}}{\ref{line:collect-stvar}}).
The replica $\replica{\mathit{p}}{\mathit{d}}$ waits to receive
a set, denoted $\mathit{store}$, of $\textsc{collect\_ack}$ messages
from a quorum $Q'$ of $\textsc{replicas}(\mathit{p})$.
We require the messages in $\mathit{store}$ carry the same
${\sf curr\_cgst}$ as in the corresponding $\textsc{collect}$ message
and the signatures of all the collected updates be valid
(i.e., ${\sf safe\_collect}(\mathit{store})$ holds).
Then, it sends a signed $\textsc{propose}$ message
with $\mathit{store}$ as its proposal to $\textsc{replicas}(\mathit{p})$
(line~\code{\ref{alg:cgst}}{\ref{line:newleader-send-propose-else}}).
When replica $\replica{\mathit{p}}{\mathit{d}}$ receives
a $\textsc{propose}$ message from replica $\replica{\mathit{p}}{\ivar}$,
it also checks the predicate ${\sf safe\_collect}(\mathit{store})$
(line~\code{\ref{alg:cgst}}{\ref{line:propose-pre}}).
After setting ${\sf curr\_cgst}^{\mathit{p}}_{\mathit{d}}$ to $\mathit{store}$,
it sends a signed $\textsc{prepared}$ message to $\textsc{replicas}(\mathit{p})$.
When replica $\replica{\mathit{p}}{\mathit{d}}$ receives
a set $\C$ of $\textsc{prepared}$ messages
from a quorum $Q$ of $\textsc{replicas}(\mathit{p})$,
both its ${\sf curr\_view}^{\mathit{p}}_{\mathit{d}}$
and ${\sf curr\_store}^{\mathit{p}}_{\mathit{d}}$ are prepared
(lines~\code{\ref{alg:cgst}}{\ref{line:prepared-preparedview}}
and~\code{\ref{alg:cgst}}{\ref{line:prepared-preparedstore}}).
The certification $\C$ is also remembered in ${\sf cert}^{\mathit{p}}_{\mathit{d}}$
(line~\code{\ref{alg:cgst}}{\ref{line:prepared-cert}}).
They will be sent to new leaders in view changes
to ensure agreement across views
(line~\code{\ref{alg:cgst}}{\ref{line:newview-send-newleader}}).
Then the replica sends a signed $\textsc{commit}$ message to $\textsc{replicas}(\mathit{p})$
(line~\code{\ref{alg:cgst}}{\ref{line:prepared-send-commit}}).
When replica $\replica{\mathit{p}}{\mathit{d}}$ receives
a set of $\textsc{commit}$ message from a quorum $Q$ of $\textsc{replicas}(\mathit{p})$,
it computes $\mathit{store}$ as the union of
the sets of updates $\mathit{st}_{i}$ collected
from each $\replica{\mathit{p}}{\ivar}$ in ${\sf curr\_store}^{\mathit{p}}_{\mathit{d}}$
(line~\code{\ref{alg:cgst}}{\ref{line:commit-storevar}}).
If ${\sf cgst}^{\mathit{p}}_{\mathit{d}}$ is smaller than the $\mathit{cgst}$
in ${\sf curr\_store}^{\mathit{p}}_{\mathit{d}}$,
the replica sets ${\sf cgst}^{\mathit{p}}_{\mathit{d}}$ to this $\mathit{cgst}$
(line~\code{\ref{alg:cgst}}{\ref{line:commit-cgst-if}}),
and replaces the set of updates with timestamps
$\le {\sf curr\_cgst}^{\mathit{p}}_{\mathit{d}}$ in ${\sf store}^{\mathit{p}}_{\mathit{d}}$
with the new $\mathit{store}$
(line~\code{\ref{alg:cgst}}{\ref{line:commit-store}}).
\subsection{Discussions} \label{ss:discussions}
\subsubsection{Optimization} \label{sss:optimization}
If the physical clock of a client is much later
than the physical clocks in other processes,
its updates may be refused and it has to reset the physical clock.
Replicas will refuse these updates
because byzantine clients may update a value
with an old timestamp to break the causality.
In this case, a correct client needs to reset its physical clock
and update the value again.
Since the physical clock increases all time
and the $FGST$ that replicas maintain increases by update events,
this case seldom take place in a correct client,
unless its physical clock is broken.
\subsubsection{Liveness} \label{sss:liveness}
Let $ml$ denote the maximum network delay between the system,
$lt_r$ denote the lower bound of the physical time on replicas,
and $lt_c$ denote the lower bound of the physical time on clients.
If $lt_c - ml > lt_r$,
then all the updates sent by a correct client will be valid,
which means correct clients will not propose an update
with a timestamp older than $FGST$.
Ideally, all the processes share the same physical clock
and the network latency tends to be zero.
In this case $lt_c - ml = lt_r$,
so the property ensures that all the updates from correct clients will be valid.
Now let us relax the requirement for clock synchronization and network latency.
If a new update form $c$ has a too old timestamp $t_c$,
then its physical clock is too slow or the network latency $l_{cr}$ is too long,
which means $t_c - l_{cr}$ is too small.
Though we cannot control $t_c$ and $l_{cr}$,
we can set a reasonable lower bound.
Here we choose $lt_r$(the lower bound of current physical time on replicas),
because clients with a correct physical clock will finally satisfy the inequality.
If the physical clock of client $c$ works correctly,
then $t_c - lt_r > k$, where $k$ is a constant.
By adding an offset to the physical clock,
it will finally satisfy that $t_c - lt_r > l_{cr}$,
and then the client can keep proposing its update.
If the whole system satisfying $lt_c - ml > lt_r$,
then all the updates sent by a correct client
will be valid by the third property 'reasonable'.
Otherwise, a client will be told to reset its physical time and retry its update.
By resetting physical clock and giving an offset to $t_c$,
it will finally propose updates successfully.
|
\section{Introduction}
Let $R$ denote a polynomial ring over an algebraic closed field $K$, say $R=K[x_1,\cdots,x_N]$. Let $I$ be an ideal of $R$ generated by $r$ homogeneous polynomials of degree at most $d$. Stillman's Conjecture states that there is a bound $M(r,d)$ on the projective dimension of $I$, which depends only on $r$ and $d$.\\
As part of their proof, Ananyan and Hochster \cite{Ananyan-Hochster-small} introduced the term strength $k$ or $k$-strong, which is equivalent to the following definition:
\begin{defn}\label{strength} Let $f$ be a homogeneous element of $R$. The \textbf{strength} of $f$ is the minimal integer $k\geq 0$ for which there is a decomposition $f =\sum_{i=1}^{k+1} g_ih_i$ with $g_i$ and $h_i$ homogeneous elements of $R$ of positive degree, or $\infty$ if no such decomposition exists. We take the convention that the strength of any constant, including 0, is -1. The \textbf{collective strength} of a set of homogeneous elements $\{f_i\}_{i\in I}$ of $R$ is the minimal strength of a non-trivial homogeneous $k$-linear combination.\\
\end{defn}
Ananyan and Hochster \cite{Ananyan-Hochster-strength} proved that there is a bound on the projective dimension of $R/$I that depends only on $n$, and not on $N$. They also showed that there are functions $^\eta A(n, d)$ with the following property: in a graded $n$-dimensional $K$-vector subspace $V$ of $R$ spanned by forms of degree at most $d$, if no nonzero form in $V$ is in an ideal generated by $^\eta A(n, d)$ forms of strictly lower degree, which is the strength condition, then any homogeneous basis for $V$ is an regular sequence.\\
In later work, \cite{Ananyan-Hochster-strength}, they obtained the function $^\eta A(n,d)$ in degree 2, 3 and 4. In degree 2, they obtained an explicit value for $^\eta B(n,2)$ that gives the Stillman’s conjecture for quadrics when there is no restriction on $n$. In particular, for an ideal $I$ generated by $n$ quadrics, the projective dimension $R/I$ is at most $2^{n+1}(n-2) + 4$.\\
On the other hand, Erman, Sam and Snowden \cite{Erman-Sam-Snowden} used an alternative approach, i.e., proving that the limiting rings are polynomial rings and using these polynomiality results to show the existence of small subalgebras $^\eta \mathcal{B}(n,d)$ in \cite{Ananyan-Hochster-small} and Stillman's conjecture.\\
In fact, in \cite{Ananyan-Hochster} when proving that there is a bound on the projective dimension of $R/I$ where $R=(x_1,\cdots,x_N)$ and $I$ is an ideal of $R$ generated by $n$ polynomials of degree at most 2, that depends only on $n$, not on $N$, Ananyan and Hochster show that there exists a positive integer $C(n)$ for positive integer $n$ such that these $n$ homogeneous polynomials is a regular sequence when their collective strength $\leq C(n)$ and $C(n)$ approaches to $2n^{2n}$.\\
With all these as the motivation, we are interested in finding the explicit bound $N=N(r,d)$. i.e., For homogeneous polynomials $f_1,\cdots,f_r$ of degree $\leq d$, if $N=N(r,d)$ is the bound for the collective strength such that $\{f_i\}$ is a regular sequence, we would like to know that if $N$ depends on (increase with) $d$ ?\\
More specifically, we study the explicit bound $N=N(r,d)$ on collective strength of regular sequence of three homogeneous polynomials, $f_1,f_2,f_3$. Note that $r=3$ is the first interesting case. In fact, we will give a simple proof that $N$ does not depend on $d$ when $r=1$ or 2. Our main results will show that $N(3,2)=2$ and $N(3,3)>2$.\\
\section{Notation and Background}
Here are some basic foundations, notations and background that are used in this paper:
\begin{defn} A sequence of elements $x_1,\cdots,x_d$ in a ring $R$ is called a \textbf{regular sequence} on $R$ if the ideal $\langle x_1,\cdots,x_i\rangle$ is proper and for each $i$ with $i=1,\cdots,d-1$, the image of $x_{i+1}$ is a nonzerodivisor in $R/\langle x_1,\cdots,x_i\rangle$.
\end{defn}
\begin{thm}\textbf{Principal Ideal Theorem}\label{PIT} \cite[Theorem 10.1]{Eisenbud} If $x\in R$, and $P$ is the minimum among primes of $R$ containing $(x)$, then codim $P\leq 1$.
\end{thm}
\begin{defn} \cite{Miller-Stephenson} \label{quadratic} A \textbf{quadratic form} is a polynomial in several variables with coefficients in an algebraically closed filed $\mathbf{C}$, all of whose terms have degree 2. If the variables in question are $x,y,$ and $z$, examples of quadratic forms include $x^2, xy-z^2$ and $2x^2+3y^2-5z^2$. To every quadratic form $q$ there is a nonnegative integer called the \textbf{rank} of $q$. For example, the quadratic form $x^2$ has rank 1. Quadratic forms of the form $\alpha x^2+\beta y^2$, where $\alpha,\beta\in\mathbf{C}^*$, have rank 2. In fact, every nonzero quadratic form $q$ can be written in the form $X_1^2+\cdots+X_k^2$ for some positive integer $k$ and some linearly independent set $\{X_1,\cdots,X_k\}$ of linear forms in the original variables, and $k$ is precisely the rank of $q$.
\end{defn}
\begin{rmk}\label{rank_strength} By changing the coordinate, one can see that any quadratic form of rank $k$ with $k$ even has strength $\frac{1}{2}k-1$ since $X_1^2+\cdots+X_k^2$ can be written as $X_1X_2+\cdots+X_{k-1}X_{k}$ that has $\frac{1}{2}k$ terms.
\end{rmk}
\begin{lemma}\label{dim} Let $(R,m)$ be a local ring and let $x\in m$ and assume that $x$ is not a zero divisor. Then dim $(R/(x))=$ dim $R-1$.
\end{lemma}
\begin{proof} Recall that dimension of $R/(x)$ is the maximum length of the chain $$(x)\subsetneq P_1\subsetneq P_2\subsetneq\cdots$$
since the set of zero divisors is the union of the associated primes $\cup_{\textbf{Assoc}(R)}P$. So if $x$ is not a zero divisor, $(x)$ does not belong to any associated prime. In particular, it does not belong to any minimum prime. So if we take a chain of prime ideals in $R$ containing $x$, say, taking a maximum chain, the smallest element we called the minimum prime. So if $x$ is not a zero divisor, $(x)$ is not in this chain. That means $(x)$ can never be the minimum prime in a chain. Hence, if $x$ is not a zero divisor, the chain containing $(x)$ has to be strictly shorter which means dim $R/(x)<$ dim $R$.\\
Now, Theorem \ref{PIT} (Principal Ideal Theorem) gives dim $R/(x)\geq$ dim$R$ - 1, hence we have the equality. In another words, adding the hypothesis of $x$ being a nonzero divisor drops dimension exactly one, i.e. $$\text{dim } R/(x) = \text{dim }R - 1$$
\end{proof}
\begin{prop}\label{reg} In a polynomial ring, $f_1,\cdots,f_r$ are regular sequence if and only if $I=\langle f_1,\cdots,f_r\rangle$ has codim $(I)=r$.
\end{prop}
\begin{proof}
If $f_1,\cdots,f_c$ regular sequence, then by applying Lemma \ref{dim} $c$ times, we get
$$\text{dim} R/(f_1,\cdots,f_c) = \text{dim }R - c$$
which means
$$\text{codim } (f_1,\cdots,f_c)=c$$
About the other direction, note that Cohen-Macaulay ring is a ring where every system of parameters is automatically a regular sequence. So if codim $(f_1,\cdots,f_c)=c$, then $f_1,\cdots,f_c$ is a regular sequence.
\end{proof}
\section{The Bound on Collective Strength when $r=1$ and $r=2$}
We want to examine the explicit bound $N=N(r,d)$ for the collective strength of homogeneous polynomials $f_1,\cdots,f_r$ of degree $\leq d$ such that $\{f_i\}$ is a regular sequence. We would like to know whether $N(r,d)$ increases with $r$ fixed and $d\to\infty$.\\
When $r=1$, the answer is no since one polynomial is a regular sequence if and only if it is non-zero. In particular, $N(1,d)=0$. If $r=2$, we prove that $N(2,d)=1$ as below.
\begin{lemma}\label{notreq} $f_1$ and $f_2$ are not a regular sequence if and only if gcd$(f_1,f_2)\neq 1$.
\end{lemma}
\begin{proof}
Let $I=(f_1,f_2)$. If $f_1$ and $f_2$ are not a regular sequence, by Theorem \ref{PIT} Principal Ideal Theorem, we have codim $(I)<2$. If codim $(I)=0$, there's only one ideal that has codimension zero which is the zero ideal. But this would mean that $f_1=f_2=0$. So codim $(I)=0$ is the same as being the zero ideal. So if two polynomial which are the zero ideal, they both have to be zero. If codim $(I)=1$, by Primary Decomposition Theorem, the ideal $I$ can be decomposed into the intersection of primary ideals. And there exists codimension-one pieces. These codimension-one pieces will be principal and are generated by single element, called $(a)$. i.e. $(f_1,f_2)=(a)\cap\cdots$. What is the relationship between $a$ and $f_i$? The claim is $a\mid f_1$ and $a\mid f_2$. So we want to prove that $a$ is a common factor of $f_1$ and $f_2$. But according to property of intersection of ideals and definition of Primary Decomposition, this $a$ is actually the common factor of $f_1$ and $f_2$. This shows that $a$ divides the gcd$(f_1,f_2)$ and therefore gcd $(f_1,f_2)\neq 1$.\\
On the other hand, if $f_1=gh_1$ and $f_2=gh_2$, need to show that $f_2$ is a zero divisor in $R/(f_1)=R/(gh_1)$. Since $h_1f_2=h_1(gh_2)=(h_1g)h_2=0$ on $R/(f_1)$. Hence $f_2$ is a zero divisor in $R/(f_1)$, so $f_1,f_2$ are not a regular sequence.
\end{proof}
\begin{cor} If strength$(f_1,f_2)\geq 1$, then $f_1, f_2$ are regular sequence.
\end{cor}
\begin{proof}
Note that an element is irreducible if and only if its strength is at least 1. By Lemma \ref{notreq}, if $f_1$, $f_2$ fail to be a regular sequence, then both elements must be reducible and have a common factor. So their collective strength is $\leq 0$. Thus, if either $f_1$ or $f_2$ is irreducible, then they will automatically form a regular sequence.
\end{proof}
Therefore, we conclude that $N(2,d)=1$, i.e. $N$ does not depend on $d$.\\
\section{Minors of the $(n+1) \times n$ Matrix}
Let $\Phi$ be a $(n+1)\times n$ matrix $(x_{i,j})$ where $(x_{i,j})$ are variables in a polynomial rings $\C\C[x_{1,1},\cdots,x_{n+1,n}]$. Let $f_i$ be the $n\times n$ minor obtained by dropping the $i^{th}$ row. Then $\langle f_1,f_2,\cdots,f_{n+1}\rangle$ has codim 2 by the Hilbert-Burch Theorem \cite[Theorem 20.15]{Eisenbud}. Note that every $n\times n$ minor has strength $\leq n-1$ by Laplace expansion. \\
Hence, if $f_1,f_2,f_3$ are $n\times n$ minors of the $(n+1) \times n$ matrix, by Proposition \ref{reg} they are not a regular sequence because the maximal minors of this matrix have codimension 2. i.e. $\langle f_1,f_2,f_3\rangle$ has codimension 2.\\
This is a natural way to build a collection of three homogeneous polynomials of degree $n$ that are not a regular sequence. To prove that $N(3,n)>n-1$, it is enough to show that their collective strength is $n-1$. Thus, there exists homogeneous polynomials $f_1,f_2,f_3$ of degree $n$ with collective strength $n-1$ that is not a regular sequence. i.e. $N(3,n)>n-1$.\\
\section{The Bound on Collective Strength when $r=3$ and $d=2$}
While the main goal of this paper is to prove that $N(3,3)>3$, this section will focus on the case $r=3$ and $d=2$. We obtain a stronger result: $N(3,2) = 2$.
\begin{lemma}\label{prime} Let $f_1$, $f_2\in R$ be homogeneous polynomials. If the singular locus of $V(f_1,f_2)$ has codimension $>4$, then the ideal $\langle f_1,f_2\rangle$ is prime.
\end{lemma}
\begin{proof}
Let $X=V(f_1,f_2)$ be the vanishing locus of the ideal $\langle
f_1,f_2\rangle$. If the ideal is not prime, the corresponding vanishing locus has multiple components.
Assume for contradiction that the ideal $\langle f_1,f_2\rangle$ is not prime. Then X will have multiple irreducible components, $X_1$ and $X_2$, each of which has codimension at two. Since $X_1$ and $X_2$ are subsets of projective space, they will intersect in codimension at most 4, and this intersection will belong to the singular locus of $X$. It follows that the singular locus of $X$ has codimension at most 4 inside of projective space. Hence if the singular locus of $X$ has codimension bigger than 4, then we have a contradiction, implying that the ideal $\langle f_1,f_2\rangle$ has to be prime.
\end{proof}
\begin{defn} For any two quadrics $f_1, f_2$ in $k[x_1,\cdots,x_n]$, define \textbf{minrank$(f_1,f_2)$} to be the lowest rank of a nontrivial $k$-linear combination of $f_1$ and $f_2$. Note that \textbf{minrank$(f_1,f_2)=0$} if and only if $f_1$ and $f_2$ are $k$-linear dependent.
\end{defn}
\begin{thm} \label{n32} If $f_1,f_2,f_3\in k[x_1,\cdots,x_n]$ are polynomials of degree 2 and collective strength $\geq 2$, then they form a regular sequence. i.e. $N(3,2)\leq 2$.
\end{thm}
\begin{proof}
Since we are over an algebraically closed field of characteristic not equal to 2, \cite{Harris} Lemma 22.42 implies that we can simultaneously write $f_1, f_2$ in $k[x_1,\cdots,x_n]$ as $$f_1=a_1x_1^2+a_2x_2^2+\cdots+a_nx_n^2$$ $$f_2=b_1x_1^2+b_2x_2^2+\cdots+b_nx_n^2$$
It is enough to show that $\langle f_1,f_2\rangle$ is prime. This is because, if $\langle f_1,f_2\rangle$ is prime then $k[x_1, \cdots, x_n]/\langle f_1,f_2\rangle$ will have no zero divisors, and so $f_3$ will be a non-zero divisor modulo $\langle f_1,f_2\rangle$. Hence, $f_1, f_2, f_3$ would be a regular sequence. Thus, our goal will be to show that the hypothesis on collective strength implies that $\langle f_1,f_2\rangle$ is a prime ideal.\\
Recall that the Jacobian ideal of $\langle f_1,f_2\rangle$ is generated by the $2\times 2$ minors of the Jacobian matrix of $\langle f_1,f_2\rangle$. It describes the singular locus. Now, if $a_i=b_i=0$ for some $i$, then we can simply reduce the number of variables. So we can assume that either $a_i$ or $b_i$ is nonzero for all $i$. Moreover, by replacing $f_1$ by an appropriate $k$-linear combination of $f_1$ and $f_2$, we can further assume that $a_i \neq 0$ for all $i$. And finally, since we are over an algebraically closed field, we can rescale the $x_i$-coordinates so that $f_1=x_1^2+\cdots+x_n^2$ and $f_2=b_1x_1^2+\cdots+b_nx_n^2$. The Jacobian matrix is
\[\begin{bmatrix}
2x_1 & 2x_2 & \cdots & 2x_n\\
2b_1x_1& 2b_2x_2 & \cdots & 2b_nx_n\\
\end{bmatrix}.
\]
Let Jac$(f_1,f_2)=\langle 2\times 2 \text{ minors of Jacobian matrix} \rangle$ denote the Jacobian ideal of $f_1, f_2$. So $(i,j)$ minor $=(4b_j-4b_i)x_ix_j$. For simplicity, we denote $J=$ Jac $(f_1,f_2)$.\\
So we write
\begin{eqnarray}
J &=& \langle 2\times 2 \text{ minors of Jacobian matrix}\rangle\nonumber\\
&=&\langle (b_2-b_1)x_1x_2,\cdots, (b_n-b_1)x_1x_n,\cdots, (b_n-b_{n-1})x_{n-1}x_n\rangle\nonumber\\
&=& \bigcap_{i=1}^n\langle x_1,x_2,\cdots,x_{i-1},x_{i+1},\cdots,x_n\rangle\qquad\cdots (5.1)\nonumber
\end{eqnarray}
Our goal is now to compute the codimension of $J$ and relate this to minrank$(f_1,f_2)$. We will compute the codimension of $J$ by first producing an explicit primary decomposition of $J$.\\
Hence, to complete the proof of this theorem, it suffices to prove two auxiliary results. First, we will show that codim $J$ equals minrank$(f_1,f_2)$, in Lemma \ref{minrank}. Second, in Lemma \ref{coll_strength}, we will show that minrank$(f_1,f_2) \geq 4$ if the collective strength of $f_1, f_2, f_3$ is at least 2. Then, by applying Lemma \ref{prime}, this will complete the proof.
\end{proof}
\begin{lemma}
\label{minrank} With notation as in Theorem \ref{n32}, we claim that codim $J$ = minrank$(f_1,f_2)$.
\end{lemma}
\begin{proof}
The ideal $J$ depends on the number of pairs $b_i, b_j$ where $b_i = b_j$. To streamline the computations about $J$, it will therefore be convenient to change notation in a way that focus on which coefficients of $f_2$ are the same. So we suppose that coefficients of $f_2$ are $\{\alpha_1,\cdots, \alpha_r\}$ where the $\alpha_i$ are all distinct and $r\leq n$. We let $\lambda_i$ denote the number of times that $\alpha_i$ appears as a coefficient of $f_2$. And we let $\mu_i = \sum_{j=1}^i \lambda_j$. After permuting the variables, we can then rewrite
$$f_2=\alpha_1(\sum_{j=1}^{\mu_1}x_j^2)+\alpha_2(\sum_{j=\mu_1+1}^{\mu_2}x_j^2)+\cdots+\alpha_r(\sum_{j=\mu_{r-1}+1}^{\mu_r}x_j^2).$$\\
It will be further convenient to relabel our variables in accordance with these $\lambda_i$. So for each $i = 1,\cdots, r$, we introduce variables $z_{i, j}$ for $1 \leq j \leq \lambda_i$.\\
With this convention, we can now express the Jacobian matrix as:\\
\[\begin{bmatrix}
z_{1,1} & \cdots & z_{1,\lambda_1} & \vline & z_{2,1} & \cdots & z_{2,\lambda_2} & \vline & \cdots & \vline & z_{k,1} & \cdots & z_{k,\lambda_k}\\
\alpha_1z_{1,1} & \cdots & \alpha_1z_{1,\lambda_1} & \vline & \alpha_2z_{2,1} & \cdots & \alpha_2z_{2,\lambda_2} & \vline & \cdots & \vline & \alpha_kz_{k,1} & \cdots & \alpha_kz_{k,\lambda_k}\\
\end{bmatrix}
\]\\
By (5.1), we now see that the Jacobian ideal $J$ can be expressed as:
$$J=\langle z_{i,m}z_{j,l} \mid i\neq j, 1\leq m\leq\lambda_i, 1\leq l\leq \lambda_j\rangle$$
Now we are ready to compute a primary decomposition of J. We let $I_t$ be the prime ideal $I_t= \langle z_{a ,b} | a\neq t \rangle$. And we let $$I = \bigcap_{t=1}^r I_t$$\\
We claim that $I=J$, first we prove $J\subseteq I$. Without loss of generality, prove that $z_{i,1}z_{j,1}\in I_t$ for all $t$.\\
Case 1: $t=i$\\
When $t=i$, we have $I_t=\langle z_{a,b}\mid a\neq i\rangle$ which means $z_{j,1}\in I_t$. Hence the product $z_{i,1}z_{j,1}$ is in $I_t$ for all $n$.\\
Case 2: $t=j$\\
When $t=j$, we have $I_t=\langle z_{a,b}\mid a\neq j\rangle$ which means $z_{i,1}\in I_t$. Hence the product $z_{i,1}z_{j,1}$ is in $I_t$ for all $t$.\\
Case 3: $t\neq i,j$\\
This is clear because both $z_{i,1}$ and $z_{j,1}$ are in $I_t$, so is their product.\\
Thus, we have proved that $J\subseteq I$.\\
Now we prove $I\subseteq J$. Indeed, if a monomial $m$ does not lie $J$, then $m$ must not divisible by any pair of variables $z_{i,a}$ and $z_{j,b}$ with $i\neq j$. In other words, for some $i$, m must be a monomial only involving the variables $z_{i,1}, \cdots, z_{i,\lambda_i}$. This means that $m \in \langle z_{i,1},\cdots, z_{i,\lambda_i}\rangle$. Then we have $m\not\in I_t=\langle z_{a,b}\mid a\neq t\rangle$, which implies that $m\not\in I$ as desired. Therefore, we have that $I=J$.\\
Hence, we have codim $J$ = codim $I=\min_t\{\text{codim }I_t\}$. Finally, we need to compare minrank$(f_1,f_2)$ to codim $I$. Given our special form for $f_1$ and $f_2$, the smallest rank of a quadric will be obtained by a $k$-linear combination of $f_1$ and $f_2$ which has the most coefficients equal to zero. In other words, minrank$(f_1,f_2)$ is $n - \max {\lambda_t , 1\leq t \leq r}$. But we note that $n - \lambda_t$ is precisely codim$(I_t)$, and we conclude that codim$(J) =$ minrank$(f_1,f_2)$ as desired.
\end{proof}
\begin{lemma}
\label{coll_strength} If collective strength of $f_1,f_2,f_3$ is at least 2 then minrank$(f_1,f_2)$ is at least 4.
\end{lemma}
\begin{proof}
This follows from Remark \ref{rank_strength}.
\end{proof}
\begin{prop} \label{nn32}There exist homogeneous polynomials $f_1,f_2,f_3$ with collective strength 1 that do not form a regular sequence. This implies $N(3,2)\geq 2$.
\end{prop}
\begin{proof}
Consider the $3\times 2$ matrix
\[\begin{bmatrix}
x_{11} & x_{12}\\
x_{21} & x_{22}\\
x_{31} & x_{32}\\
\end{bmatrix}
\]\\
The minors of the matrix $$f_1=x_{21}x_{32}-x_{31}x_{22},\quad f_2=x_{11}x_{32}-x_{31}x_{12},\quad f_3=x_{11}x_{22}-x_{21}x_{12}.$$ do not form a regular sequence by the discussion in Section 4. So to complete the proof, we need to show that $f_1, f_2, f_3$ have collective strength 1.\\
Let $a(x_{21}x_{32}-x_{31}x_{22})+b(x_{11}x_{32}-x_{31}x_{12})+c(x_{11}x_{22}-x_{21}x_{12})$ be the linear combination of $f_1,f_2,f_3$ for some $a,b,c$ in a field $k$. If $(a,b,c)=(0,0,0)$, there's nothing to prove. So we can assume that at least one of $a,b,c$ is nonzero. By symmetry, without loss of generality we can assume that $a\neq 0$.\\
By definition, the collective strength of $f_1,f_2,f_3$ is the minimal strength of a non-trivial linear combination of $f_1,f_2,f_3$, so is obviously $\leq 1$. Want to prove that it is precisely 1, hence $N(3,2)\geq 2$.\\
The matrix of this quadratic polynomial is
\[\begin{bmatrix}
0 & 0 & 0 & 0 & a & b\\
0 & 0 & 0 & -a & 0 & c\\
0 & 0 & 0 & -b & -c & 0\\
0 &-a & -b & 0 & 0 & 0\\
a & 0 & -c & 0 & 0 & 0\\
b & c & 0 & 0 & 0 & 0\\
\end{bmatrix}.
\]\\
Since we have $a\neq 0$. Reducing to echelon form, we get
\[\begin{bmatrix}
1 & 0 & -\frac{c}{a} & 0 & 0 & 0\\
0 & 1 & \frac{b}{a} & 0 & 0 & 0\\
0 & 0 & 0 & 1 & 0 & -\frac{c}{a}\\
0 & 0 & 0 & 0 & 1 & \frac{b}{a}\\
0 & 0 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 0 & 0 & 0\\
\end{bmatrix}.
\]\\
So the rank is $\geq 4$, i.e. the rank of the linear combination of $f_1,f_2,f_3$ has rank $\geq 4$, which means its strength is $\geq 1$ by Remark \ref{rank_strength}. Therefore, it is exactly 1. Hence, $N(3,2)\geq 2$.\\
\end{proof}
Finally, combining Theorem \ref{n32} and Proposition \ref{nn32} proves that $N(3,2)=2$.\\
\section{Homogeneous Polynomials and Column Grading}
We introduce column grading in this section. And in Lemma \ref{homog} we prove that a cubic homogeneous polynomial of strength 1 can be written as a linear combination of homogeneous polynomials of the same strength in column grading.\\
\begin{defn}
We imagine the variables of the $\C\C[x_{11},\cdots , x_{43}]$ as corresponding to entries in a generic $4\times 3$ matrix:
\[M=\begin{bmatrix}
x_{11} & x_{12} & x_{13}\\
x_{21} & x_{22} & x_{23}\\
x_{31} & x_{32} & x_{33}\\
x_{41} & x_{42} & x_{43}\\
\end{bmatrix}
\]
We give this ring a $\Z^3$ grading by its columns. Namely, $deg(x_{i1})=(1,0,0)$ for $1\leq i \leq 4$ and similarly $deg(x_{i2})=(0,1,0)$ and $deg(x_{i3})=(0,0,1)$.
\end{defn}
\begin{eg}
$(x_{11}+x_{23})$ is not homogeneous because $x_{11}$ and $x_{23}$ have different degree. They are $\deg(x_{11})=(1,0,0)$ for $1\leq i \leq 4$ and $\deg(x_{23})=(0,0,1)$.
\end{eg}
\begin{eg}
If we choose degree $(1,1,0)$ then $S_{(1,1,0)}$ has a basis consisting of products $x_{i1}x_{j2}$ for any $1\leq i \leq 4$ and any $1\leq j \leq 4$.
\end{eg}
In the following section, we will be interested in studying the strength of $3\times 3$ minors of the above matrix. We note that any such determinant has degree $(1,1,1)$ in this grading.\\
Next, we turn our attention to analyzing ideals generated by a pair of linear forms, up to appropriate symmetries, as this will be used in Section 7.
\begin{lemma}
Up to the natural $GL_4\times GL_3$ -action on $\C[x_{11}, … , x_{43}]$, any ideal I which is homogeneous in the $\Z^3$ grading, and which is generated by a pair of linearly independent linear forms, is equivalent to one of the following:
\begin{itemize}
\item $(x_{11},x_{21})$
\item $(x_{11},x_{12})$
\item $(x_{11},x_{22})$
\end{itemize}
\end{lemma}
\begin{proof}
We start with an ideal generated by two linearly independent linear forms $\ell_1, \ell_2$, each of which is homogeneous in the column grading. Without loss of generality, we can assume that $ \ell_1 = x_{11}$. If $\deg(\ell_2) = \deg(\ell_1)$ then we can assume that $\ell_2 = x_{21}$, which is the first case listed in the lemma. If $deg(\ell_2) \ne \deg(\ell_1)$ then we can assume that $\ell_2 = x_{i2}$ for some i. Again there are two cases: if $i=1$ then we are in one the second case listed in the lemma. If $i\ne 1$, then we can assume that $i =2$ and our forms are $x_{11}$ and $x_{22}$ and so we are in the third case of the lemma.\\
Hence, when fixing $x_{11}$ there are three equivalent classes with the following representatives: $$\{(x_{11},x_{21}), (x_{11}, x_{22}), (x_{11},x_{12})\}.$$
\end{proof}
This lemma will be used in proving $N(3,3)>2$ where we apply column grading.\\
\section{The Bound on Collective Strength when $r=3$ and $d=3$}
In this section, we prove that there exist cubic homogeneous polynomials $f_1,f_2,f_3$ with collective strength 2 that is not a regular sequence. i.e. $N(3,3)>2$.\\
As mentioned in Section 4, the $n\times n$ minors of the $(n+1) \times n$ matrix are not a regular sequence. So let
\[M=\begin{bmatrix}
x_{11} & x_{12} & x_{13}\\
x_{21} & x_{22} & x_{23}\\
x_{31} & x_{32} & x_{33}\\
x_{41} & x_{42} & x_{43}\\
\end{bmatrix}
\] be a $4\times 3$ matrix.\\
And let
\[\Delta_1=\begin{bmatrix}
x_{21} & x_{22} & x_{23}\\
x_{31} & x_{32} & x_{33}\\
x_{41} & x_{42} & x_{43}\\
\end{bmatrix},\quad
\Delta_2=\begin{bmatrix}
x_{11} & x_{12} & x_{13}\\
x_{31} & x_{32} & x_{33}\\
x_{41} & x_{42} & x_{43}\\
\end{bmatrix}
\]\\
\[\Delta_3=\begin{bmatrix}
x_{11} & x_{12} & x_{13}\\
x_{21} & x_{22} & x_{23}\\
x_{41} & x_{42} & x_{43}\\
\end{bmatrix},\quad
\Delta_4=\begin{bmatrix}
x_{11} & x_{12} & x_{13}\\
x_{21} & x_{22} & x_{23}\\
x_{31} & x_{32} & x_{33}\\
\end{bmatrix}
\]
be $3\times 3$ submatrices of $M$ obtained by deleting $i^{th}$ row. Then $f_1=\text{det}\Delta_1, f_2=\text{det}\Delta_2, f_3=\text{det}\Delta_3$ are $3\times 3$ minors and $f_1,f_2,f_3$ are homogeneous polynomials of degree 3 that is not a regular sequence.\\
By using Laplace expansion, we have:
\[
\Delta_1 = x_{21}(x_{32}x_{43} - x_{33}x_{42}) - x_{22}(x_{31}x_{43} - x_{33}x_{41}) + x_{23}(x_{31}x_{42} - x_{32}x_{41}).
\]
It follows that the strength of $\Delta_1$ is at most $2$ and thus the collective strength of $f_1,f_2$, and $f_3$ is also at most $2$. We want to prove that the collective strength is not 1, hence it is 2. Suppose it is 1, then $\Delta$ can be written as $\Delta=ab+cd$, where $a,c$ are linear and $b,d$ are quadrics. Note that each of $f_1,f_2,f_3$ is a cubic homogeneous polynomial of degree $\begin{pmatrix}
1\\
1\\
1\\
\end{pmatrix}$ in column grading, so is their linear combination $\Delta$.\\
Since $a$ and $c$ are linear, they are linear combination of degree $\begin{pmatrix}
1\\
0\\
0\\
\end{pmatrix}$, $\begin{pmatrix}
0\\
1\\
0\\
\end{pmatrix}$ and $\begin{pmatrix}
0\\
0\\
1\\
\end{pmatrix}$ in column grading. So $a$ can be written as $a=a_{100}+a_{010}+a_{001}$ where $a_{100}$ denotes the combination of terms from column 1, $a_{010}$ denotes the combination of terms from column 2, $a_{001}$ denotes the combination of terms from column 3. Similarly, $c=c_{100}+c_{010}+c_{001}$.\\
Likewise, since $b,d$ are quadric, they are linear combination of degree $\begin{pmatrix}
2\\
0\\
0\\
\end{pmatrix}$, $\begin{pmatrix}
0\\
2\\
0\\
\end{pmatrix}$, $\begin{pmatrix}
0\\
0\\
2\\
\end{pmatrix}$, $\begin{pmatrix}
1\\
1\\
0\\
\end{pmatrix}$, and $\begin{pmatrix}
0\\
1\\
1\\
\end{pmatrix}$, $\begin{pmatrix}
1\\
0\\
1\\
\end{pmatrix}$ in column grading, denoted by $b_{200},b_{020},b_{002},b_{110},b_{101}, b_{011}$. Hence, $b$ can be written as $b=b_{200}+b_{020}+b_{002}+b_{110}+b_{101}+b_{011}$. Similarly, $d=d_{200}+d_{020}+d_{002}+d_{110}+d_{101}+d_{011}$.\\
Furthermore, since $\Delta=ab+cd$ can be seen as an ideal generated by two linear elements $a, c$, where $a=a_{100}+a_{010}+a_{001}, c=c_{100}+c_{010}+c_{001}$ i.e. $\Delta=\langle a,c\rangle$. By multiplying a suitable invertible $2\times 2$ matrix, $\Delta$ can be re-written as $\Delta=\langle a,a+c\rangle$ for example. Hence, without loss of generality, we can assume that $a_{100}\neq c_{100}$ and both nonzero. In fact, we can apply to other degrees, such as $a_{010}, c_{010}$ and $a_{001}, c_{001}$. That means, fixing $a_{100}$ and $c_{100}$ with $a_{100}\neq c_{100}$ and both nonzero, through a suitable linear transformations we can assume that the other two pairs $(a_{010}, c_{010})$, $(a_{001},c_{001})$ in $a$ and $c$ are either both zero or both nonzero and non-equal.\\
\begin{eg} If $a=a_{100}+a_{010}$, $c=c_{100}$, where $a_{100}\neq c_{100}$. Then we can make $(a,c)$ to be nonzero on degrees $\begin{pmatrix}
1\\
0\\
0\\
\end{pmatrix}$ and $\begin{pmatrix}
0\\
1\\
0\\
\end{pmatrix}$ for both $a$ and $c$ by multiplying an invertible matrix
\[
\begin{pmatrix}
1 & 0\\
q & 1\\
\end{pmatrix}
\]
where a general choice of $q$ will have the desired properties as below:
i.e. $$
\begin{pmatrix}
1 & 0\\
q & 1\\
\end{pmatrix}
\cdot
\begin{pmatrix}
a_{100}+a_{010}\\
c_{100}\\
\end{pmatrix}
=
\begin{pmatrix}
a_{100}+a_{010}\\
q(a_{100}+a_{010})+c_{100}\\
\end{pmatrix}
$$
\end{eg}
\begin{eg} If $a=a_{100}+a_{001}$, $c=c_{100}+c_{010}$, where $a_{100}\neq c_{100}$. Then we can make $(a,c)$ nonzero in all degrees by multiplying by a general $2\times 2$ matrix to obtain the desired effect.
Thus, if at least one of $a$ and $c$ is nonzero in some certain degree, say $\begin{pmatrix}
1\\
0\\
0\\
\end{pmatrix}$, we can always make both $a$ and $c$ nonzero for that degree via some matrix transformation. Obviously, this won't work if both $a$ and $c$ are zero in that degree. Therefore, for each pair of column degree: $(a_{100}, c_{100})$, $(a_{010}, c_{010})$, $(a_{001},c_{001})$ we can assume that they are either both zero, or both nonzero and non-equal.
\end{eg}
\begin{lemma}\label{homog} Let $\Delta$ be a form of degree $\begin{pmatrix}
1\\
1\\
1\\
\end{pmatrix}$ in column grading. Assume that $\Delta$ has strength one, that is, we can write $\Delta=ab+cd$, where $a,c$ are linear (in the standard grading) and $b,d$ are quadric (in the standard grading). Then we can find $a^\prime,b^\prime,c^\prime,d^\prime$ all homogeneous in column grading, such that $\Delta=a^\prime b^\prime +c^\prime d^\prime$.
\end{lemma}
\begin{proof}
Since $a,c$ are linear and $b,d$ quadric, by discussion above we can write $$a=a_{100}+a_{010}+a_{001}$$ $$c=c_{100}+c_{010}+c_{001}$$ $$b=b_{200}+b_{020}+b_{002}+b_{110}+b_{101}+b_{011}$$ $$d=d_{200}+d_{020}+d_{002}+d_{110}+d_{101}+d_{011}$$ Since $\Delta$ has degree $(1,1,1)$ in the column grading, we have that:
\begin{eqnarray}
\Delta &=& ab+cd\nonumber\\
&=&(a_{100}+a_{010}+a_{001})(b_{200}+b_{020}+b_{002}+b_{110}+b_{101}+b_{011})\nonumber\\
&+& (c_{100}+c_{010}+c_{001})(d_{200}+d_{020}+d_{002}+d_{110}+d_{101}+d_{011})\nonumber\\
&=& a_{100}b_{011}+a_{010}b_{101}+a_{001}b_{110}+c_{100}d_{011}+c_{010}d_{101}+c_{001}d_{110}\nonumber
\end{eqnarray}
with the following equations
$$a_{100}b_{200}+c_{100}d_{200}=0\quad\cdots (1)$$
$$a_{100}b_{020}+a_{010}b_{110}+c_{100}d_{020}+c_{010}d_{110}=0\quad\cdots (2)$$
$$a_{100}b_{002}+a_{001}b_{101}+c_{100}d_{002}+c_{001}d_{101}=0\quad\cdots (3)$$
$$a_{010}b_{200}+a_{100}b_{110}+c_{010}d_{200}+c_{100}d_{110}=0\quad\cdots (4)$$
$$a_{010}b_{020}+c_{010}d_{020}=0\quad\cdots (5)$$
$$a_{010}b_{002}+a_{001}b_{011}+c_{010}d_{002}+c_{001}d_{011}=0\quad\cdots (6)$$
$$a_{001}b_{200}+a_{100}b_{101}+c_{001}d_{200}+c_{100}d_{101}=0\quad\cdots (7)$$
$$a_{001}b_{002}+c_{001}d_{002}=0\quad\cdots (8)$$
$$a_{001}b_{020}+a_{010}b_{011}+c_{001}d_{020}+c_{010}d_{011}=0\quad\cdots (9)$$\
We separate the proof into two major cases: Case 1 is where $a_{100}$, $c_{100}$ are linearly independent and Case 2 is where $a_{100}$, $c_{100}$ are linearly dependent.\\
\textbf{Case 1}: Assume that $a_{100}$, $c_{100}$ are linearly independent. Then WLOG $a_{100}= x_{11}$ and $c_{100} = x_{21}$. We will prove the lemma in this case by first proving a series of subclaims.\\
\textit{Claim 1.1}: There exists $B$ such that $b_{200} = x_{21}B$ and $d_{200} = -x_{11}B$.\\
Proof of Claim 1.1\\
Let $a_{100}=x_{11}, c_{100}=x_{21}$, then (1) becomes $$x_{11}b_{200}+x_{21}d_{200}=0$$ If $b_{200}=0$, then $d_{200}=0$ or vice versa, then there's nothing to prove and we move to Case 2. So we assume that $b_{200}\neq 0$ and $d_{200}\neq 0$. Thus, from (1) $b_{200}$ must have a factor $x_{21}$, i.e. $b_{200}=x_{21}B$ for some linear term $B$ from the first column. Similarly, $d_{200}=x_{11}D$ for some $D$ from the first column. So we can write
\begin{eqnarray}
0&=& x_{11}x_{21}B+x_{21}x_{11}D=0\quad\cdots (1)\nonumber\\
&=& x_{11}x_{21}(B+D)\nonumber
\end{eqnarray}
which implies that $B+D=0$, or $B=-D$. Hence, we proved that there exists $B$ such that $b_{200} = x_{21}B$ and $d_{200} = -x_{11}B$.\\
\textit{Claim 1.2} This further implies $b_{110}=c_{010}B$ and $d_{110}=-a_{010}B$ for some homogeneous $c_{010}$ and $a_{010}$ of degree $(0,1,0)$.\\
Proof of Claim 1.2\\
Substitute $b_{200}=x_{21}B$ and $d_{200}=x_{11}D$ into (4)\\
We have
\begin{eqnarray}
0&=& a_{010}b_{200}+a_{100}b_{110}+c_{010}d_{200}+c_{100}d_{110}\nonumber\\
&=& a_{010}x_{21}B+x_{11}b_{110}+c_{010}x_{11}D+x_{21}d_{110}\nonumber\\
&=& x_{11}(b_{110}+c_{010}D)+x_{21}(d_{110}+a_{010}B)\nonumber
\end{eqnarray}
If two terms are canceled with each other, $b_{110}+c_{010}D$ must have the factor $x_{21}$ and $a_{010}B+d_{110}$ has factor $x_{11}$. This implies that $D=x_{21}$ and $B=x_{11}$ which is a contradiction since $B+D\neq 0$. So we must have $$b_{110}=c_{010}B,\quad d_{110}=-a_{010}B$$
\textit{Claim 1.3} This further implies $b_{020}=d_{020}=0$.\\
Proof of Claim 1.3\\
Plugging the result of Claim 1.2 into (2) we have
\begin{eqnarray}
0&=& a_{100}b_{020}+a_{010}b_{110}+c_{100}d_{020}+c_{010}d_{110}\nonumber\\
&=& x_{11}b_{020}-a_{010}c_{010}D+x_{21}d_{020}-c_{010}a_{010}B\nonumber\\
&=& x_{11}b_{020}+x_{21}d_{020}-a_{010}c_{010}(D+B)\nonumber\\
&=& x_{11}b_{020}+x_{21}d_{020}\nonumber
\end{eqnarray}
So we must have $b_{020}=d_{020}=0$.\\
$a_{010}=c_{010}=0$\\
This means that $d_{020}$ must be a multiple of $x_{11}$ but that is impossible because of the column grading. Thus $d_{020}=0$ and similarly $b_{020}= 0$.\\
\textit{Claim 1.4} This further implies that $b_{101}=c_{001}B$ and $d_{101}=-a_{001}B$ for some homogeneous $c_{001}$ and $a_{001}$ of degree $(0,0,1)$ and that $b_{002}=c_{002}=0$.\\
Proof of Claim 1.4\\
The same arguments as in the proofs of Claim 1.2 and 1.3 apply here.\\
Thus, from our initial assumption that $a_{100}$, $c_{100}$ were linearly independent, Claims 1.2 and 1.4 allow us to reduce to the case where $a=a_{100}=x_{11}$ and $c=c_{100}=x_{21}$. Making this substitution into the expression for $\Delta$ in terms of the $a_{ijk}$ and so on, we then have $$\Delta= a_{100}b_{011}+c_{100}d_{011}$$ and this yields the desired expression for $\Delta$ which is homogeneous in the column grading.\\
We have thus completely handled Case 1, which is where $a_{100}$ and $c_{100}$ are linearly independent. By symmetry we have also handled the cases where $a_{010}$ and $c_{010}$ are linearly independent or when $a_{001}$ and $c_{001}$. So we have reduced to the case where each of these pairs is linearly dependent.\\
\textbf{Case 2} We have that $a_{100}$, $c_{100}$ are linearly dependent, and similarly for $a_{010}$, $c_{010}$ and $a_{001}$, $c_{001}$. By choosing a different basis for the vector space spanned by $a$ and $c$ (and altering $b$ and $d$ accordingly), we can assume that there exist scalars $\lambda_1$, $\lambda_ 2$, $\lambda_3$ such that $$c_{100}=\lambda_1 a_{100}\text{ and }c_{010}=\lambda_2 a_{010}\text{ and } c_{001}=\lambda_3 a_{001}.$$
Moreover, by replacing $c$ by $c-\lambda_1a$ (and again altering $b$ and $d$ accordingly), we can also assume that $\lambda_1=0$ and thus $c_{100}=0$.\\
Substituting these expressions into our equations for $\Delta$ and equation (2) from above, we get $$\Delta = a_{100}b_{011}+a_{010}(b_{101}+\lambda_2 d_{101})+a_{001}(b_{11}+\lambda_3 d_{110})$$
$$0=a_{100}b_{020}+a_{010}(b_{110} +\lambda_2d_{110})$$ If any of $a_{100}$, $a_{010}$ or $(b_{110}+\lambda_2d_{110})$ is zero, then substituting this into the expression for $\Delta$ gives an expression $\Delta= a'b' + c'd'$ of the desired form. If these are all nonzero, then we can write $$(b_{110} + \lambda_2d_{110}) = a_{100}f$$ for some $f$ of degree $(0,1,0)$. Substituting this into the expression for $\Delta$, we have
\begin{eqnarray}
\Delta &=& a_{100}b_{011}+a_{010}(b_{101}+\lambda_2d_{101})
+a_{001}(b_{110}+\lambda_3d_{110})\nonumber\\
&=& a_{100}b_{011}+a_{010}(b_{101}+\lambda_2d_{101})+a_{001}a_{100}f\nonumber\\
&=& a_{100}(b_{011}+a_{001}f)+a_{010}(b_{101}+\lambda_2d_{101})\nonumber
\end{eqnarray}
which is an expression $\Delta= a'b'+c'd'$ of the desired form.
\end{proof}
\begin{thm} There exist cubic homogeneous polynomials $f_1,f_2,f_3$ with collective strength 2 that do not form a regular sequence, i.e. $N(3,3)>2$.
\end{thm}
\begin{proof}
Let $f_i=\det\Delta_i$ for $i=1,2,3$. By the arguments in Section 4, we know that $f_1, f_2, f_3$ do not form a regular sequence, and that they have collective strength at most 2. So to prove this theorem, it suffices to prove that the collective strength of $f_1, f_2$, and $f_3$ is at least 2.\\
Let
$\Delta=c_1\Delta_1+c_2\Delta_2+c_3\Delta_3=
c_1\Big (x_{21}(x_{32}x_{43}-x_{33}x_{42})-x_{22}(x_{31}x_{43}-x_{41}x_{33})+x_{23}(x_{31}x_{42}-x_{32}x_{41})\Big )+c_2\Big (x_{11}(x_{32}x_{43}-x_{33}x_{42})-x_{12}(x_{31}x_{43}-x_{33}x_{41})+x_{13}(x_{31}x_{42}-x_{32}x_{41})\Big )+c_3\Big (x_{11}(x_{22}x_{43}-x_{42}x_{23})-x_{12}(x_{21}x_{43}-x_{41}x_{23})+x_{13}(x_{21}x_{42}-x_{41}x_{22})\Big )$\\
Suppose that this collective strength is 1. Then there is some $\mathbb C$-linear combination $\Delta = c_1f_1 + c_2f_2 + c_3f_3$ which has strength 1, i.e. $\Delta = ab+cd$, with $a,c$ linear. By Lemma 7.3 we can assume that $a$ and $c$ are homogeneous in the column grading. By Lemma 6.4, we can reduce to the cases where $a = x_{11}$ and $c = x_{22}$ or $x_{12}$ or $x_{21}$. We now perform an explicit computation to show that this is impossible.\\
Recall that\\
$\Delta=c_1\Delta_1+c_2\Delta_2+c_3\Delta_3=
c_1\Big (x_{21}(x_{32}x_{43}-x_{33}x_{42})-x_{22}(x_{31}x_{43}-x_{41}x_{33})+x_{23}(x_{31}x_{42}-x_{32}x_{41})\Big )+c_2\Big (x_{11}(x_{32}x_{43}-x_{33}x_{42})-x_{12}(x_{31}x_{43}-x_{33}x_{41})+x_{13}(x_{31}x_{42}-x_{32}x_{41})\Big )+c_3\Big (x_{11}(x_{22}x_{43}-x_{42}x_{23})-x_{12}(x_{21}x_{43}-x_{41}x_{23})+x_{13}(x_{21}x_{42}-x_{41}x_{22})\Big )$\\
\textbf{Case 1} $I_1=(x_{11},x_{12})$\\
We consider $\Delta$ modulo $I_1$. Imposing the relations $x_{11}=x_{12}=0$, $\Delta$ reduces to\\
$c_1\Big (x_{21}(x_{32}x_{43}-x_{33}x_{42})-x_{22}(x_{31}x_{43}-x_{41}x_{33})+x_{23}(x_{31}x_{42}-x_{32}x_{41})\Big )+c_2\Big (x_{13}(x_{31}x_{42}-x_{32}x_{41})\Big )+c_3\Big (x_{13}(x_{21}x_{42}-x_{41}x_{22})\Big )=0$.\\
This can be re-written as\\
$\Delta/I_1=c_1 x_{21}x_{32}x_{43}-c_1x_{32}x_{33}x_{42}-c_1x_{22}x_{31}x_{43}+c_1x_{22}x_{41}x_{33}+c_1x_{23}x_{31}x_{42}-c_1x_{23}x_{32}x_{41}+c_2x_{13}x_{31}x_{42}-c_2x_{13}x_{32}x_{41}+c_3x_{13}x_{21}x_{42}-c_3x_{13}x_{41}x_{22}=0$.\\
Note that these 10 terms: $x_{21}x_{32}x_{43}, x_{32}x_{33}x_{42}, x_{22}x_{31}x_{43}, \cdots$ are linearly independent, so their coefficients $c_1, c_2$ and $c_3$ must be zero. This is a contradiction because $\Delta=c_1\Delta_1+c_2\Delta_2+c_3\Delta_3\neq 0$. Therefore, we conclude that $\Delta/I_1\neq 0$, which means, $\Delta\not\in I_1=(x_{11},x_{12})$.\\
\textbf{Case 2} $I_2=(x_{11},x_{21})$\\
We consider $\Delta$ modulo $I_2$. Imposing the relations $x_{11}=x_{21}=0$, $\Delta$ reduces to\\
$c_1\Big (-x_{22}(x_{31}x_{43}-x_{41}x_{33})+x_{23}(x_{31}x_{42}-x_{32}x_{41})\Big )+c_2\Big (-x_{12}(x_{31}x_{43}-x_{33}x_{41})+x_{13}(x_{31}x_{42}-x_{32}x_{41})\Big )+c_3\Big (-x_{12}(-x_{41}x_{23})+x_{13}(-x_{41}x_{22})\Big )=0$\\
This can be re-written as\\
$-c_1x_{22}x_{31}x_{43}+c_1x_{22}x_{41}x_{33}+c_1x_{23}x_{31}x_{42}-c_1x_{23}x_{32}x_{41}-c_2x_{12}x_{31}x_{43}+c_2x_{12}x_{33}x_{41}+c_2x_{13}x_{31}x_{42}-c_2x_{13}x_{32}x_{41}+c_3x_{12}x_{41}x_{23}-c_3x_{13}x_{41}x_{22}=0$.\\
Note that these 10 terms: $x_{22}x_{31}x_{43}, x_{22}x_{41}x_{33}x, x_{23}x_{31}x_{42}, \cdots$ are linearly independent, so their coefficients $c_1, c_2$ and $c_3$ must be zero. This is a contradiction because $\Delta=c_1\Delta_1+c_2\Delta_2+c_3\Delta_3\neq 0$. Therefore, we conclude that $\Delta/I_2\neq 0$, which means, $\Delta\not\in I_2=(x_{11},x_{21})$.\\
\textbf{Case 3} $I_3=(x_{11},x_{22})$\\
We consider $\Delta$ modulo $I_3$. Imposing the relations $x_{11}=x_{22}=0$, $\Delta$ reduces to\\
$c_1\Big (x_{21}(x_{32}x_{43}-x_{33}x_{42})+x_{23}(x_{31}x_{42}-x_{32}x_{41})\Big )+c_2\Big (-x_{12}(x_{31}x_{43}-x_{33}x_{41})+x_{13}(x_{31}x_{42}-x_{32}x_{41})\Big )+c_3\Big (-x_{12}(x_{21}x_{43}-x_{41}x_{23})+x_{13}(x_{21}x_{42})\Big )=0$\\
This can be re-written as\\
$c_1x_{21}x_{32}x_{43}-c_1x_{21}x_{33}x_{42}+c_1x_{23}x_{31}x_{42}-c_1x_{23}x_{32}x_{41}-c_2x_{12}x_{31}x_{43}+c_2x_{12}x_{33}x_{41}+c_2x_{13}x_{31}x_{42}-c_2x_{13}x_{32}x_{41}-c_3x_{12}x_{21}x_{43}+c_3x_{12}x_{41}x_{23}+c_3x_{13}x_{21}x_{42}=0$.\\
Note that these 10 terms: $x_{21}x_{32}x_{43},x_{21}x_{33}x_{42},x_{23}x_{31}x_{42}, \cdots$ are linearly independent, so their coefficients $c_1, c_2$ and $c_3$ must be zero. This is a contradiction because $\Delta=c_1\Delta_1+c_2\Delta_2+c_3\Delta_3\neq 0$. Therefore, we conclude that $\Delta/I_3\neq 0$, which means, $\Delta\not\in I_3=(x_{11},x_{22})$.\\
Thus, the collective strength of $f_1,f_2,f_3$ is not 1, hence is 2. Therefore, we proved that there exists cubic homogeneous polynomials $f_1,f_2,f_3$ with collective strength 2 that is not a regular sequence. i.e. $N(3,3)>2$.
\end{proof}
\newpage
\begin{bibdiv}
\begin{biblist}
\bib{Ananyan-Hochster}{article}{
author={Ananyan, Tigran},
author={Hochster, Melvin},
title={Ideals generated by quadratic polynomials},
series={Math. Research Letters},
volume={19},
publisher={Math. Research Letters},
date={2012},
pages={233--244},
}
\bib{Ananyan-Hochster-small}{article}{
author={Ananyan, Tigran},
author={Hochster, Melvin},
title={Small subalgebras of polynomial rings and Stillman's conjecture},
series ={arXiv:1610.09268},
publisher={preprint},
date={2016},
}
\bib{Ananyan-Hochster-strength}{article}{
author={Ananyan, Tigran},
author={Hochster, Melvin},
title={Strength conditions, small subalgebras, and Stillman bounds in degree},
series={arXiv:1810.00413},
publisher={Math. AC},
date={2018},
}
\bib{Erman-Sam-Snowden}{article}{
author={Erman, Daniel},
author={Sam, Steven},
author = {Snowden, Andrew},
title={Big polynomial rings and Stillman's conjecture},
series ={arXiv:1801.09852v2},
publisher={math. AC},
date={2018},
}
\bib{Miller-Stephenson}{article}{
author={Miller, Nina},
author={Stephenson, Darin},
author = {Wells, Andrew},
author = {Erika, Wittenborn},
title={Counting Quadratic Forms of Rank 1 and 2},
publisher={Pi Mu Epsilon Journal Vol. 12},
date={2009},
}
\bib{Eisenbud}{book}{
author={Eisenbud, David},
title={Commutative Algebra with a view toward Algebraic Geometry},
series ={Graduate Texts in Mathematics},
publisher={Springer},
date={2004},
}
\bib{Harris}{book}{
author={Harris, Joe},
title={Algebraic Geometry},
series ={Graduate Texts in Mathematics},
publisher={Springer},
date={1992},
}
\end{biblist}
\end{bibdiv}
\end{document} |
\section{Introduction}
Neuroscience has become one of the most significant areas in biology, as scientists pursue the understanding of the functionality of perception and the brain, and investigate therapeutics for neurological diseases and injuries (see \cite{ribar2020neuromorphic, 7402491}). Neurological disorders, such as Parkinson's disease, spinal cord injuries, and Alzheimer's disease, result from a loss of function of neurons \cite{maccioni2001molecular}, \cite{dauer2003parkinson}, \cite{liu1997neuronal}. Particular medical therapeutics, i.e. Chondroitinase ABC (ChABC), may restore the functionality of neurons (see \cite{karimi2010synergistic, bradbury2011manipulating}) by manipulating the extracellular matrix (ECM), the area surrounding the neuron which contains various macromolecules and minerals that facilitate cell processes \cite{frantz2010extracellular}.
Neurons are cells that are specialized to obtain perception by transmitting electrical signals along their axon
. These signals enter from the dendrites, travel through the axon, and transmit through the growth cone to another neuron as shown in Figure \ref{fig:my_label}. The presence of chemical cues surrounding the growth cone attract or repel the axon toward another neuron \cite{DIEHL2014194}. When the chemical cues attract the growth cone, tubulin proteins assemble from free tubulin dimers and create microtubules which extend the axon toward the other neuron \cite{julien1999neurofilament}.
The formation of microtubules is determined by the size and dynamics of tubulin concentration in the neuron and is supported by ECM \cite{DENT2003209, barros2011extracellular}. The dynamics of tubulin depend on the tubulin production rate in the soma, the degradation rate, the assembly rate, and the transportation process \cite{DIEHL2014194}. According to recent research, it is also possible to control axon elongation, namely tubulin concentration, by manipulating ECM \cite{burnside2014manipulating}.
The dynamics of axon growth have been described by several different models using various mathematical tools. One of the pioneer axonal growth models has been proposed in \cite{van1994neuritic} by considering the transportation of tubulin as diffusion.
A continuum model of axon growth dynamics has been proposed in \cite{mclean2004continuum}, and the stability analysis for the proposed model is studied in \cite{mclean2006stability}. In another work, the tubulin concentration is modelled by a PDE with a moving boundary \cite{graham2006mathematical, diehl2014one}. \cite{diehl2016efficient} provides a numerical solution for this moving boundary PDE model.
While PDEs have been utilized for the computational modeling of axon growth, stabilization of axon growth by means of control theory has not been studied so far. Boundary control of PDEs has been intensively developed by the method of ``backstepping" in the last few decades for various kinds of systems \cite{krstic2008boundary}. One of the initial contributions was achieved in \cite{smyshlyaev2004closed} by applying a Volterra type of transformation to parabolic PDEs by utilizing the method of successive approximations. Following the procedure, the class of the system has been extended to a cascade and coupled PDE-ODE system, see \cite{krstic2009compensating, susto2010control, tang2011state}. While most literature on backstepping design of PDEs has focused on a system with a constant size of domain in time, recently the method has been successfully applied to the Stefan problem, which is a special class of a parabolic PDE with a moving boundary, see \cite{koga2018control, krstic2020materials}. For the model related to axon growth mentioned earlier, the backstepping method for the Stefan problem has been designed and applied to a screw extrusion process of a polymer 3-D printing \cite{koga2020stabilization}. The results mentioned above have been proven to achieve global stability by virtue of the monotonicity of the moving boundary.
Several researchers have tackled stability analysis for nonlinear PDE systems under the backstepping design of a boundary control by restricting the region of attraction in a local sense. For instance, \cite{coron2013local} proved a local exponential stability of a $2 \times 2$ quasi-linear hyperbolic PDE under the backstepping design by analyzing a Lyapunov function of the spatial $H_2$-norm. As a class of PDEs with a moving boundary, \cite{buisson2018control} designed a backstepping control law for $2 \times 2$ hyperbolic PDEs with a moving boundary governed by an ODE modeling a piston movement, and proved a local exponential stability in the spatial $H_1$-norm by applying the Lyapunov method to the target system. A similar approach has been done for a moving boundary hyperbolic PDE modeling a shock-wave arising in traffic congestion \cite{yu2020bilateral}. However, those results for local stability analysis have been achieved only for hyperbolic PDE systems, even though the axon growth model proposed in \cite{diehl2014one} is a nonlinear parabolic PDE system.
In this paper, we develop a boundary control for a coupled PDE-ODE system with a moving boundary which models the dynamics of tubulin concentration and axon growth. First, we present a steady-state solution of the tubulin concentration for a given constant axon length, and obtain a reference error system to be stabilized at zero states. Next, we apply linearization to the reference error system to deal with the algebraic nonlinearity.
Then, a backstepping transformation is employed to the linearized reference error dynamics. By solving the gain kernel equations that are derived from backstepping transformation, the control law is obtained. Finally, we prove local exponential stability by applying the Lyapunov method to the nonlinear target system, which ensures the local stability of the original PDE-ODE system of the axon growth model.
This paper is structured as follows. Section \ref{sec:model} introduces the PDE-ODE model of axon growth and tubulin concentration with the steady-state solution. Section \ref{sec:control} presents the control design by the method of backstepping, and the stability result and its proof are given in Section \ref{sec:stability}. The paper ends with the conclusion in Section \ref{sec:conclusion}.
\begin{figure}
\centering
\includegraphics[scale=0.4]{Neuron3.jpg}
\caption{Neuron Structure and PDE domain}
\label{fig:my_label}
\end{figure}
\section{Modeling of Axon Growth} \label{sec:model}
In this section, we present a mathematical model of axon growth governed by a moving boundary PDE, derive a steady-state solution for a given set point of the axon length, and provide a reference error system to be stabilized.
\subsection{Axon growth model by a moving boundary PDE}
Tubulin is a group of proteins which is responsible for the growth of a newly created axon. Two assumptions can be described to model this responsibility. First, tubulin proteins are modeled as a homogeneous continuum because free tubulin molecules are very small. Then, tubulin molecules are assumed to be the only factor responsible for axonal growth.
With these assumptions, as proposed in \cite{diehl2014one, diehl2016efficient}, the axonal growth of a newborn axon by tubulin can be modelled as
\begin{align}
\label{sys1}
c_t (x,t) =& D c_{xx} (x,t) - a c_x (x,t) - g c(x,t) , \\
\label{sys2} c_x(0,t) = & - q_{\rm s}(t), \\
\label{sys3} c(l(t),t) =& c_{\rm c} (t), \\
\label{sys4} l_{\rm c} \dot{c}_{\rm c}(t) = & (a-gl_{\rm c}) c_{\rm c}(t) - D c_x (l(t), t) \notag\\
& - (r_{\rm g} c_{\rm c}(t) + \tilde{r}_{\rm g} l_{\rm c} )(c_{\rm c}(t) - c_{\infty}), \\
\label{sys5} \dot{l}(t) =& r_{\rm g} (c_c(t)-c_{\infty}),
\end{align}
where the tubulin concentration in the axon is denoted as $c(x,t)$, and the variables with subscripts $t$ and $x$ denote the partial derivatives with respect to the subscripts. The variable $q$ denotes the concentration flux. Subscript ``s" is used for the soma of the neuron, and subscript ``c" is used for the cone of the neuron.
Namely, $q_{\rm s}(t)$ is the concentration flux of tubulin in the soma, and $c_{\rm c}(t)$ is the concentration of tubulin in the cone. The length of the axon in $x$-coordinate is $l(t)$. As time passes, tubulin in the neuron degrade at the constant rate, $g$. $D$ is the diffusivity constant, and $a$ is the velocity constant of tubulin proteins in \eqref{sys1}. $l_{\rm c}$ is the growth ratio of the cone, and $\tilde{r}_{\rm g}$ is the reaction rate to create microtubules. $\tilde{s}_{\rm g}$ is the disassemble rate which means it is the transformation rate from microbules to tubulin dimers and $r_{\rm g}$ is a lumped parameter. The equilibrium of the tubulin concentration in the cone, which is denoted as $c_{\infty}$, causes the axonal growth to stop. The control problem to be solved in this paper is presented in the following statement.
\textbf{Problem:} Develop a feedback control law of $q_{\rm s}(t)$ so that $l(t)$ converges to $l_{\rm s}$ for a given desired length of the axon $l_{\rm s}>0$, subject to the governing equations \eqref{sys1}--\eqref{sys5}.
\subsection{Steady-state solution}
To tackle the problem stated above, we first solve a steady-state solution of the concentration profile for a given axon length $l_{\rm s}$. By setting the time derivatives in \eqref{sys1}, \eqref{sys4}, and \eqref{sys5} to zero, one can derive the steady-state solution of \eqref{sys1}-\eqref{sys5} as
\begin{align}
\label{ceq}
c_{\rm eq}(x) = c_{\infty} \left( K_{+} e^{\lambda_+ (x - l_{\rm s})} + K_- e^{\lambda_{-} (x - l_{\rm s}) } \right),
\end{align}
where
\begin{small}
\begin{align}
\lambda_+ =& \frac{a + \sqrt{a^2 + 4 D g}}{2 D}, \quad \lambda_- = \frac{a - \sqrt{a^2 + 4 D g}}{2 D}, \\
K_+ = & \frac{1}{2} + \frac{a - 2 g l_{\rm c} }{2 \sqrt{a^2 + 4 D g}}, K_- = \frac{1}{2} - \frac{a - 2 g l_{\rm c} }{2 \sqrt{a^2 + 4 D g}}.
\end{align}
\end{small}
The steady-state input for the concentration flux in the soma is obtained as
\begin{align}
q_{\rm s}^* = - c_{\infty} \left( K_{+} \lambda_+ e^{ - \lambda_+ l_{\rm s}} + K_- \lambda_- e^{ - \lambda_{-} l_{\rm s} } \right).
\end{align}
\subsection{Reference error system}
Let $u(x,t)$, $z_1(t)$, and $z_2(t)$ be the reference error states, and $U(t)$ be the reference error input, defined as
\begin{align}
u(x,t) =& c(x,t) - c_{\rm eq}(x), \\
z_{1}(t) =& c_{\rm c}(t) - c_{\infty}, \\
z_2(t) =& l(t) - l_{\rm s}, \\
U(t) = & - ( q_{\rm s}(t) - q_{\rm s}^*).
\end{align}
By substituting the steady-state solution in \eqref{ceq} from the governing equations \eqref{sys1}-\eqref{sys5}, the reference error system is obtained as in \cite{diehl2014one, krstic2020materials}.
Let $X \in {\mathbb R}^2$ be an ODE state vector for the reference error states $z_1(t)$ and $z_2(t)$, defined by
\begin{align} \label{xdef}
X(t)=[ z_1(t) \quad z_2(t)]^\top .
\end{align}
Applying the linearization of $X(t)$ around zero states leads to the following linearized reference error system (see Section 12-2 in \cite{krstic2020materials} for the detailed derivation):
\begin{align}
\label{ulin-PDE}
u_t (x,t) =& D u_{xx} (x,t) - a u_x (x,t) - g u(x,t) , \\
u_x(0,t) = & U(t), \label{ulin-BC1} \\
\label{linreferr3}u(l(t),t) =&C^\top X(t) , \\
\dot{X}(t) = & A X(t) + B u_x (l(t), t), \label{ulin-ODE}
\end{align}
where
\begin{align} \label{AB-def}
A = &\left[
\begin{array}{cc}
\tilde a & 0 \\
r_{\rm g} & 0
\end{array}
\right] , \quad B = \left[
\begin{array}{c}
- \beta \\
0
\end{array}
\right], \\
C = &\left[1 \quad - \frac{(a-gl_{\rm c}) c_{\infty}}{D}\right]^\top . \label{C-def}
\end{align}
\section{Control Design} \label{sec:control}
The control design in this paper is based on a backstepping transformation \cite{krstic2008boundary}, which maps the reference error system $(u,X)$ to a target system $(w,X)$. The backstepping transformation and the target system in this paper are given in the remainder of this section.
\subsection{Backstepping transformation and target system}
Following the procedure in \cite{koga2018control} for the Stefan problem, we consider the following backstepping transformation:
\begin{small}
\begin{align}
\label{bkst}
w(x,t) = & u(x,t) - \int_x^{l(t)} k(x,y) u(y,t) dy
- \phi(x - l(t))^\top X(t),
\end{align}
\end{small}where $k(x,y) \in {\mathbb R} $ and $\phi(x-l(t)) \in {\mathbb R}^2$ are the gain kernel functions to be determined. We suppose the desired target system as
\begin{align}
\label{tar-PDE} w_t (x,t) =& D w_{xx} (x,t) - a w_x (x,t) - g w(x,t) \notag\\
&- \dot l(t) F(x,X(t)) , \\
\label{tar-BC1} w_x(0,t) = & \gamma w(0,t),\\
\label{tar-BC2}w(l(t),t) =&0 , \\
\label{tar-ODE} \dot{X}(t) = & (A + BK^\top) X(t) + B w_x (l(t), t),
\end{align}
where $K \in {\mathbb R}^2$ is a feedback control gain vector chosen to make $A +BK$ Hurwitz which means that $(A,B)$ is controllable. In detail, by setting
\begin{align} \label{K-def}
K = [ k_1 \quad k_2], \quad k_1 > \frac{\tilde a}{\beta} , \quad k_2 > 0 ,
\end{align}
one can observe that the conditions for $(k_1, k_2)$ makes $A + BK$ Hurwitz.
Also, the redundant nonlinear term $F(x,X(t)) \in {\mathbb R}$ in \eqref{tar-PDE} caused by the moving boundary is described as
\begin{align} \label{F-def}
F(x,X(t))= \left(\phi'(x-l(t))^T-k(x, l(t)) C^T \right) X(t) .
\end{align}
\subsection{Gain kernel solution}
Taking the time and spatial derivatives of \eqref{bkst} together with the solution of \eqref{ulin-PDE}-\eqref{ulin-ODE}, and substituting $x = l(t)$ in both the transformation \eqref{bkst} and its spatial derivative, and by matching
with the target system \eqref{tar-PDE}--\eqref{tar-ODE}, we have the following PDE and an ODE for gain kernels.
\begin{align}
&k_{xx}(x,y)-k_{yy}(x,y)=\frac{a}{D}\left(k_x(x,y)+k_y(x,y)\right),\label{kernel1}\\
&k_x(x,x)+k_y(x,x)=0,\label{kernel2}\\
&k(x, l(t))=-\frac{1}{D}\phi(x-l(t))^\top B,\label{kernel3}\\
&D\phi''(x-l(t))^\top-a\phi'(x-l(t))^\top-\phi(x-l(t))^T\left[gI+A\right]\notag\\
&\ \ \ \ \ \ \ \ \ \ \ -Dk_y(x,l(t))C^\top+ak(x,l(t))C^\top=0, \label{phiODE1}\\
&\phi(0)=C,\quad \phi'(0)=k(l(t),l(t))C^\top+K^\top. \label{phiODE3}
\end{align}
By the conditions \eqref{kernel1}--\eqref{kernel3}, the solution of $k(x,y)$ is uniquely given by
\begin{align}
k(x,y)= -\frac{1}{D}\phi(x-y)^\top B.
\label{kstar}
\end{align}
Substituting \eqref{kstar} into \eqref{phiODE1}--\eqref{phiODE3}, the ODE of $\phi(\cdot)$ becomes
\begin{small}
\begin{align}
&D\phi''(x-l(t))^\top-\phi'(x-l(t))^\top \left(BC^\top+aI\right)\notag\\&\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ -\phi(x-l(t))^\top\left[gI+A+\frac{a}{D}BC^\top\right]=0,\label{phiup1}\\
&\phi(0)=C, \quad \phi'(0)^\top=-\frac{1}{D}C^\top BC^\top+K^\top. \label{phiup3}
\end{align}
\end{small}
The solution to \eqref{phiup1}-\eqref{phiup3} is given by (see \cite{tang2011state})
\begin{align}
\phi(x)^\top=\begin{bmatrix}C^\top & K^\top-\frac{1}{D}C^\top BC^\top\end{bmatrix}e^{N_1x}\begin{bmatrix} I \\ 0
\end{bmatrix},
\label{phix}
\end{align}
where
the matrix $N_1 \in {\mathbb R}^{4 \times 4}$ is defined as
\begin{align}
N_1=\begin{bmatrix}0 & \frac{1}{D}\left(gI+A+\frac{a}{D}BC^\top\right)\\ I &\frac{1}{D}\left(BC^\top+aI\right)\end{bmatrix}.
\end{align}
In order to check invertibility of \eqref{bkst}, we define the inverse Volterra Transformation
\begin{small}
\begin{align}
u(x,t)=&w(x,t)+\int_{x}^{l(t)}q(x,y)w(y,t)dy+\varphi(x-l(t))^\top X(t) .
\end{align}
\end{small}
Then, we apply the same strategy to obtain the inverse gain kernel functions, $q(x,y)\in \mathbb{R}$, and $\varphi(x-l(t))\in \mathbb{R}^2$. The PDE and ODE for the inverse gain kernels are
\begin{align}
&q_{xx}(x,y)-q_{yy}(x,y)=\frac{a}{D}\left(q_x(x,y)+q_y(x,y)\right)\label{eqn:inv-ker-1}\\
&q_x(x,x)+q_y(y,y)=0\\
&q(x,l(t))=-\frac{1}{D}\varphi(x-l(t))^\top B \\
&D\varphi^{''}(x-l(t))^\top+a\varphi^{'}(x-l(t))^\top\nonumber \\
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ +\left(gI+A+BK^\top\right)\varphi(x-l(t))^\top=0 \\
&\varphi(0)=C, \quad \varphi'(0)=K
\label{eqn:inv-ker-6}
\end{align}
which is well-posed, so one can obtain the solution of \eqref{eqn:inv-ker-1}-\eqref{eqn:inv-ker-6} by applying the same procedure as the one we applied to the forward kernel equations.
\subsection{Control law}
The control design is derived from the boundary condition \eqref{tar-BC1} of the target system at $x = 0$.
Substituting $x = 0$ into the transformation \eqref{bkst} and its spatial derivative, and using boundary conditions for both the target system and the error system, the control input is described as follows
\begin{align}
U(t)= & \left(\gamma-\frac{\beta}{D}\right)u(0,t) -\frac{1}{D}\int_0^{l(t)}p(x)Bu(x,t)dx\notag\\
&+p(l(t))X(t),
\label{real-input}
\end{align}
where
$p(x) = \phi'(-x)^\top-\gamma\phi(-x)^\top$.
Plugging the system matrices $(A,B,C,K)$ into the gain kernel function \eqref{phix}, and calculating the transition matrix, one can explicitly derive the function $p(x) \in {\mathbb R}^2$.
\section{Proof of Main Result} \label{sec:stability}
This section presents the main result of the paper, which is a local stability of the closed-loop system, with providing its proof by considering the $H_1$-norm
\begin{align}
Z(t) = || u(\cdot, t) ||^2 + || u_x(\cdot, t) ||^2 + X^\top X.
\end{align}
We present our main theorem below.
\begin{theorem}
Consider the closed-loop system consisting of the plant \eqref{ulin-PDE}--\eqref{ulin-ODE} with the control law \eqref{real-input} and \eqref{phix}. Suppose the control parameter $\gamma >0$ is chosen to satisfy $\gamma \geq \frac{a}{D}$. Then, there exist positive parameters $ \bar M>0$, $c>0$, and $\kappa>0$, such that if $Z(0)< \bar M$ then the following norm estimate holds
\begin{align}
Z(t)\leq c Z(0) \exp( - \kappa t)
\end{align}
for all $ t\geq 0$, which guarantees the local exponential stability of the closed-loop system.
\end{theorem}
\subsection{Basic idea of proof of Theorem 1}
Since the backstepping transformation \eqref{bkst} is invertible, the stability property of the target system \eqref{tar-PDE}-\eqref{tar-ODE} is equivalent to the stability property of the closed-loop system consisting of the plant \eqref{ulin-PDE}-\eqref{ulin-ODE} with the control law \eqref{real-input}.
The local stability of the target system is studied under the assumption that the following two conditions hold:
\begin{align} \label{ineq-l}
0 < l(t) \leq \bar l, \\
|\dot l(t) | \leq \bar v, \label{ineq-ldot}
\end{align}
for some $\bar l>l_{\rm s} >0$ and $\bar v>0$. The restricted initial state will be given later in order to satisfy these conditions for all $t \geq 0$.
\subsection{Useful inequalities}
The following inequalities are used in Lyapunov analysis. Under the condition \eqref{ineq-l}, and by boundary conditions \eqref{tar-BC1} and \eqref{tar-BC2}, Poincare's inequality is provided as
\begin{align} \label{poincare-1}
|| w ||^2 \leq & 4 \bar{l}^2 ||w_{x}||^2, \quad
|| w_{x} ||^2 \leq & 2 \bar l \gamma^2 w(0,t)^2 + 4 \bar{l}^2 ||w_{xx} ||^2,
\end{align}
and the Agmon's inequality is given as
\begin{align}
\label{Agmon}
w_x(l(t),t)^2 \leq & 2 \gamma^2 w(0,t)^2 + 4 \bar l ||w_{xx}||^2 .
\end{align}
\subsection{Proof of Lyapunov stability}
We consider the Lyapunov function of the target system as
\begin{align}
V = d_1 V_1 + V_2 + \frac{\gamma}{2} w(0,t)^2 + d_2 V_3,
\label{Vtotal}
\end{align}
where $d_1>0$ and $d_2 >0$, and each Lyapunov function is
\begin{align}
V_1 =& \frac{1}{2} ||w||^2 := \frac{1}{2} \int_0^{l(t)} w(x,t)^2 dx, \\
\label{V2-def} V_2 =& \frac{1}{2} ||w_x||^2 := \frac{1}{2} \int_0^{l(t)} w_x(x,t)^2 dx, \\
V_3=&X(t)^\top PX(t),
\end{align}
where positive definite matrix $P$ is the solution of the following Lyapunov equation
\begin{align}
(A + BK^\top )^\top P + P (A + BK^\top ) = - Q,
\end{align}
for some positive definite matrix $Q$. Since $(A+BK^\top)$ is Hurwitz, positive definite matrices $P$ and $Q$ exist. Due to the positive definiteness, it holds that
\begin{align} \label{ineq-XPX}
\lambda_{\rm min}(P) X^\top X \leq X^\top P X \leq \lambda_{\rm max}(P) X^\top X,
\end{align}
where $\lambda_{\rm min}(P) >0$ and $\lambda_{\rm max}(P)>0$ are the smallest and the largest eigenvalues of $P>0$. Then we have the following lemma.
\begin{lemma} Assume that \eqref{ineq-l}--\eqref{ineq-ldot} are satisfied with \begin{align}
\bar{v}=\min\left\{\frac{g }{4\gamma}, \frac{D}{8\bar{l}}\right\},
\end{align}
for all $ t \geq 0$. Then, for sufficiently large $d_1>0$ and small $d_2>0$, there exists a positive constant $\beta>0$ such the following norm estimate holds for all $t \geq 0$:
\begin{align}
\dot{V}\leq -\alpha V+\beta V^{3/2},
\label{vdotbound}
\end{align}
where $\alpha=\min\left\{2g+\frac{D}{4\bar{l}},\frac{4g+d_1D}{2},\frac{\lambda_{\rm min}(Q)}{2\lambda_{\rm max}(P)}, \frac{d_2\left(2d_1D+g\right)}{4}\right\}$.
\end{lemma}
\begin{proof}
Taking the time derivative of the Lyapunov functions along the target system \eqref{tar-PDE}--\eqref{tar-ODE}, we have
\begin{align}
\dot V_1 =& - D || w_{x}||^2 - g ||w||^2 - \left(\gamma D - \frac{a}{2} \right) w(0,t)^2 \notag\\
&+ \dot l(t) \int_0^{l(t)} F(x,X(t)) w(x,t) dx. \label{V1dot-2}\\
\dot{V}_2 =& - D || w_{xx}||^2 + a \int_0^{l(t)} w_{xx}(x,t) w_{x}(x,t) dx \notag\\&- g || w_{x}||^2
- \gamma w(0,t) w_{t}(0,t) -\frac{1}{2}\dot{l}(t)w_x(l(t),t)^2\notag\\
& - \dot{l}(t) (F(l(t),X(t)) w_x(l(t),t)- \gamma F(0,X(t)) w(0,t)) \notag\\
- &\gamma g w(0,t)^2- \dot{l}(t)\int_0^{l(t)} F_x(x,X(t)) w_{x}(x,t) dx \label{A-3-2:V2dot2-1} \\
\dot{V}_3=&-X(t)^\top QX(t)+ 2 w_x(l(t), t)B^\top PX(t) . \label{V3dot}
\end{align}
Applying the Agmon's inequality \eqref{Agmon} and Young's inequality to $\dot{V}_2$ in \eqref{A-3-2:V2dot2-1} and $\dot{V}_3$ in \eqref{V3dot} leads to
\begin{align}
\dot{V}_2 \leq & - \frac{D}{4} || w_{xx}||^2 - \frac{\gamma g}{2} w(0,t)^2 - \left(g - \frac{a^2}{D} \right) || w_{x}||^2 \notag\\
& + \dot{l}(t) \gamma F(0,X(t)) w(0,t)\notag\\&+\dot{l}(t) \int_0^{l(t)} F_x(x,X(t)) w_{x}(x,t) dx \notag\\
&+\frac{\big|\dot{l}(t)\big|}{2}F(l(t),X(t))^2- \gamma w(0,t) w_{t}(0,t), \label{A-3-2:V2dot2-4}\\
\dot{V}_3 \leq& -\frac{\lambda_{\rm min}(Q)}{2}X^\top X \notag\\
& + \frac{2\big|B^\top P \big|^2}{\lambda_{\rm min}(Q)}\left(2 \gamma^2 w(0,t)^2 + 4 \bar l ||w_{xx}||^2\right).
\label{dotV3last}
\end{align}
By using \eqref{V1dot-2}, \eqref{A-3-2:V2dot2-4} and \eqref{dotV3last}, recalling $\gamma \geq \frac{a}{D}$, and choosing the constants $d_1$ and $d_2$ to satisfy
\begin{align}
d_1\geq \frac{2a^2}{D^2}, \
d_2\leq \min\left\{\frac{D\lambda_{\rm min}(Q)}{64\bar{l}\big|B^TP\big|^2},\frac{g D\lambda_{\rm min}(Q)}{16a\left|B^TP\right|^2}\right\},
\end{align}
the time derivative of Lyapunov function \eqref{Vtotal} for the target system is shown to satisfy the following inequality
\begin{small}
\begin{align}
\dot V &
\leq -\alpha V +\frac{\big|\dot{l}(t)\big|}{2}F(l(t),X(t))^2\notag+\dot{l}(t) \gamma F(0,X(t)) w(0,t)\notag\\
& + \dot{l}(t) \int_0^{l(t)} F_x(x,X(t)) w_{x}(x,t) + d_1 F(x,X(t)) w(x,t) dx.
\label{dotVwalpha}
\end{align}
\end{small}
Taking the square of \eqref{F-def}, it follows that the redundant nonlinear terms in \eqref{dotVwalpha} can be bounded by a quadratic norm of the ODE state. Namely, there exist positive constants $L_1>0$, $L_2>0$, and $L_3$ such that $F(0,X(t))^2 \leq L_1 X^\top X$, $F(l(t), X(t))^2 \leq L_1 X^\top X$, $ \int_0^{l(t)} F_x(x,X(t))^2 dx \leq L_2 X^\top X$, $\int_0^{l(t)} F(x,X(t))^2 dx \leq L_3 X^\top X$. In addition,
$\dot{l}(t)$ can be rewritten as $\dot{l}(t)=r_{\rm g} e_1 X$.
By using these relations and applying Cauchy-Schwarz inequality and Young's inequality, one can show that
\begin{small}
\begin{align}
\dot V &\leq -\alpha V +r_{\rm g}\left( \frac{L_1(1+\gamma)+1}{2d_2\lambda_{\rm min}(P)}+(1+L_3+d_1L_2)\right)V^{3/2}.
\end{align}
\end{small}
Thus, there exists $\beta>0$ such that Lemma 1 holds.
\end{proof}
\subsection{Guaranteeing the conditions for all time}
Next, we prove important lemmas to conclude with Theorem 1 ensuring the local stability of the closed-loop system.
\begin{lemma}
There exists a positive constant $M_1>0$ such that in the region $\Omega_1 := \{(w,X) \in H_1 \times {\mathbb R}^2 | V(t) < M_1\}$ the conditions \eqref{ineq-l} and \eqref{ineq-ldot} are satisfied.
\end{lemma}
\begin{proof} By using \eqref{xdef} and \eqref{sys5}, $X(t)$ can be described as
$X(t)=\left[\begin{array}{cc}
\frac{\dot l(t)}{r_{\rm g}}&
l(t)-l_{\rm s}
\end{array}\right]^\top$.
Thus, for some $r>0$, if $| X|<r$ then both the following two inequalities hold:
\begin{align}
\left|\frac{\dot{l}(t)}{r_{\rm g}}\right|< r, \ \ \big|l(t)-l_{\rm s}\big|< r.
\end{align}
The former inequality tells that if $r < \frac{\bar v}{r_{\rm g}}$ then the condition \eqref{ineq-l} holds. Moreover, the latter inequality yields $- r + l_{\rm s} < l(t) < r + l_{\rm s}$, and thus if both $r < l_{\rm s}$ and $r < \bar l - l_{\rm s}$ hold, then the condition \eqref{ineq-ldot} holds.
Therefore, the constant $r>0$ is chosen as
$
r=\min \left\{\frac{\bar{v}}{r_{\rm g}}, l_{\rm s}, \bar{l}-l_{\rm s}\right\}.
$
Since $|X|^2 \leq \frac{1}{\lambda_{\rm min}(P)} X^\top P X \leq \frac{d_2}{\lambda_{\rm min}(P)} V$, we deduce that by setting $M_1 = \frac{\lambda_{\rm min}(P)}{d_2} r^2$, if $V(t) < M_1 $ holds then $| X| < r$ and thus the conditions \eqref{ineq-l} and \eqref{ineq-ldot} are satisfied, by which we can conclude Lemma 2.
\end{proof}
\begin{lemma}
There exists a positive constant $M>0$ such that if $V(0) < M$ then the conditions \eqref{ineq-l} and \eqref{ineq-ldot} are satisfied and the following norm estimate holds:
\begin{align} \label{V-decay}
V(t) \leq V(0) \exp\left(- \frac{\alpha}{2} t \right).
\end{align}
\end{lemma}
\begin{proof}
For a positive constant $M>0$, let $\Omega := \{ (w,X) \in H_1 \times {\mathbb R}^2 | V(t) < M\}$. By Lemma 2, it is easily shown that if $M \leq M_1$ then $\Omega \subset \Omega_1$, and thus the conditions \eqref{ineq-l} and \eqref{ineq-ldot} are satisfied in the region $\Omega$. Thus, by Lemma 1, the norm estimate \eqref{vdotbound} holds. Moreover, by setting $M \leq \frac{\alpha^2}{4 \beta^2}$, we can see that applying $V(t) < M$ to \eqref{vdotbound} leads to
\begin{align}
\dot V \leq - \frac{\alpha}{2} V,
\end{align}
by which the norm estimate \eqref{V-decay} is deduced. Since \eqref{V-decay} is a monotonically decreasing function in time, by setting $M = \min\{M_1, \frac{\alpha^2}{4 \beta^2}\} $, the region $\Omega$ is shown to be an invariant set. Thus, if $V(0) < M$, then $V(t) < M$ for all $t \geq 0$, and one can conclude with Lemma 3.
\end{proof}
Finally, owing to Lemma 3, and the equivalent norm estimate in the $H_1$-norm between the target and the closed-loop system, the local stability of the closed-loop system is proved, which completes the proof of Theorem 1.
\section{Numerical Simulation}
\begin{figure}[t]
\centering
\subfloat[\textbf{The axon length converges to the desired length. }\label{subfig-2}]{%
\includegraphics[width=0.89\linewidth]{lt.png}}
\vfill
\subfloat[\textbf{The tubulin concentration converges to the steady-state. }\label{subfig-1}]{ \includegraphics[width=0.90\linewidth]{conc.png}
}
\caption{Closed-loop response under the designed control law. }
\end{figure}
\begin{table}[t]
\hfill
\caption{\label{tab:initial}Biological constants and control parameters}
\centering
\begin{tabular}{cccc}
\hline
Parameter & Value & Parameter & Value \\
\hline
$D$ & $10\times10^{-6} m^2/s$& $\gamma$ & $10^4$\\
$a$& $1\times 10^{-8} m/s$ & $l_{\rm c}$ & $4\mu m$\\
$g$& $5\times 10^{-7} \ s^{-1}$ & $l_s$ & $12\mu m$\\
$r_{\rm g}$& $1.783\times 10^{-5} \ m^4/(mol s)$ & $l_0$& $1\mu m$ \\
$c_{\infty}$ & $0.0119 \ mol/m^3$ & $k_1$& $-0.1$\\
$\tilde{r}_{\rm g}$ & $0.053$ & $k_2$ & $10^{13}$ \\
\hline
\end{tabular}
\end{table}
Simulation is performed for the axon growth dynamics \eqref{sys1}-\eqref{sys5} under the designed control law \eqref{real-input}. We use the biological constants given in \cite{diehl2014one}, and choose the control parameters, as shown in Table \ref{tab:initial}. The initial tubulin concentration is set as a constant profile $c_0(x) =2c_{\infty}$, and the initial axon length is set as $l_0=1 \mu m$.
Fig. \ref{subfig-2} shows that the axon length converges to the desired length $l_{\rm s}$. In addition, Fig. \ref{subfig-1} depicts that the tubulin concentration converges to the steady-state solution. Hence, we observe that the simulation result is consistent with our theoretical results.
\section{Conclusion} \label{sec:conclusion}
In this paper, a boundary feedback control for an axonal growth model governed by a coupled PDE-ODE with a moving boundary is studied. The backstepping transformation is utilized for the original system to convert it to the target system which has a stable system matrix in ODE. The gain kernel solutions in the transformation are obtained, and the boundary feedback control law is designed explicitly. Using Lyapunov's method, we first prove that the target system is locally exponentially stable in the spatial $H_1$-norm and then we prove local stability of
the original coupled PDE-ODE system of axonal growth model. Showing the local stability without applying linearization will be studied in future.
\bibliographystyle{IEEEtran}
|
\section{Introduction}\label{intro}
Epidemiological studies, particularly based on randomized trials, often aim to estimate the average treatment effect (ATE), or another causal parameter of interest, to understand the effect of a health intervention or exposure on an outcome of interest. Most commonly, in observational studies, inverse probability of treatment weighted (IPTW) estimation and its variants have been used for this purpose ~\cite{HorvitzGSRF1952, ipw_hajek_1971, AronowEACED2017}. Alternative estimators for causal inference include substitution (or direct) estimators based on G-computation~\cite{RobinsNACIJ1986,YuCCGFD2002, DanielGECED2011, WangGATEJ2017}, those based on the approach of estimating equations (EE) \cite{RobinsMSMS2000,LaanUMCL2003}, including IPTW and its augmented variant (A-IPTW), and substitution estimators developed within the framework of targeted learning (TL) (we also refer to targeted maximum likelihood estimator, TMLE, a product of this framework \cite{LaanTLCI2011}). The latter of these has seen increasing use in recent years, both in biostatistical methodological research and applied public health and medical research \cite{PetersenAIUTJ2017,SkeemCPSOS2017,RoseRMLVO2018,PlattTERCJ2018,NeafseyGDPEN2015}. In Table~\ref{tab:lit_summary}, we provide a list of studies that have examined the relative performance of TL-based and competing estimators (mainly against EE-based methods), including a summary of whether the results suggested superior, neutral, or poorer relative performance of TL-based estimators in comparison to other estimators (the ``Pro/Con'' column). Thus, while this work is contextualized within dozens of previous studies, few such studies performed ``realistic'' simulations, and even fewer compared several variants of TL estimators alongside corresponding EE approaches. For example, in Zivich and Breskin's paper \cite{ZivichMLCIM2021}, the authors compared G-computation, IPTW, A-IPTW, TMLE and double cross-fit estimators with data generated from predefined parametric models. Exceptions are efforts that used the proposed realistic bootstrap~\cite{PetersenDRVPF2012} to evaluate the performance for data-generating distributions modeled semiparametrically (using ensemble machine learning) from an existing data set. These include a study of estimating variable importance under positivity violations using collaborative targeted maximum likelihood estimation (C-TMLE)~\cite{PirracchioCTMLJ2018}. In this paper, we use an augmentation of this proposed methodology to examine the relative performance of several versions of both TL and EE estimators in ten realistic data simulations, each based on data collected as part of the Knowledge Integration (KI) database from the Bill \& Melinda Gates Foundation ~\cite{MertensCCCGJ2020}. In so doing, we provide a realistic survey, across both different data-generating distributions and different study designs, of the relative performance of estimators of causal parameters.
\input{tablenplots/table_lit}
\section{Background}\label{background}
As large and complex data sets have become increasingly more commonplace, the habitual use of parametric approaches is encountering more data science research for which they are ill-suited. . This has led to machine learning (ML) taking a more central role in deriving estimators of causal impacts in very big statistical models (semiparametric). The theory for the use of ML in the estimators discussed herein has been continuously refined, from developing double robust estimators (both A-IPTW and TMLE substitution estimators) to augmentations of these estimators that are more robust to the overfitting potentially introduced by flexible ML fits. The latter modifications to the original estimators are the cross-validated TMLE (or CV-TMLE, chapter 27 in van der Laan\cite{LaanTLCI2011} and Zheng \cite{ZhengATCTN2010}), and subsequently the proposal for an analogous modification to estimating equation approaches (double machine learning or cross-fitting \cite{ChernozhukovDDMLF2018}).
While simulation studies have investigated all of these estimators, they have yet to be analyzed together in a single series of realistic simulation studies. Here, we seek to determine how well these estimators perform in realistic settings, under which conditions they perform best, which augmentations provide the most robustness, and whether or not the results support more general recommendations. In addition, there exist other choices of target parameter when the one being analyzed fails to have adequate performance for any of the competing estimators, such as realistic rules \cite{BembomPIIR2007}. A recently developed machine learning algorithm (the highly adaptive lasso; HAL \cite{BenkeserHALE2016}), is potentially an important improvement in estimating realistic DGDs for simulation studies such as ours. It can be optimally undersmoothed to dependably generate realistic estimates of the actual data generating distributions. HAL is particularly well suited to these types of simulations, as it uses a very large nonparametric model and can be tuned to be as flexible as the data support. In this paper, we explore the use of HAL as a basis in conducting realistic data-inspired simulations. The results suggest the proposed use of HAL for realistic data-generating simulations could provide a general method for choosing between machine-learning-based estimators for a particular parameter and data set.
We first introduce the data sets that were selected to motivate our realistic simulations, describe the steps taken for simulating data, including a short description of the estimators tested, and discuss the results. The simulations suggest a general recommendation for the use of an additional layer of cross-validation (CV-TMLE and double machine learning) to ensure robust inference in finite samples.
\section{Methods}\label{method}
\subsection{Study Selection}
We utilized data from ten nutrition intervention trials conducted in Africa and South Asia. In all studies, the measured outcome was a height-to-age Z-score for children from birth to 24 months, which was calculated using World Health Organization (WHO) 2006 child growth standards~\cite{WHOGrowth2006}. Details about the resulting composite data, study design and data processing, can be found in companion technical reports \cite{MertensCCCGJ2020,MertensCWCSJ2020,Benjamin-ChungECLGJ2020}. All interventions were nutrition-based, and for the purposes of this analysis, multi-level interventions were simplified to a binary treatment variable (e.g. nutrition intervention - yes/no). Although different baseline covariates were measured among these studies, there was significant overlap. The sample size of each study is shown in Table \ref{tab_study}. We anonymized the study IDs and removed the location information due to confidentiality concerns. Details on each study can be found in the shuffled list in Section B of the Appendix.
\input{tablenplots/table_study}
\subsection{Data Processing}
Data from each study was cleaned and processed for this analysis. Our goal for defining the analysis data used to simulate is different from the goals of the original studies and thus our data processing might differ from that used in the resulting publications of the study results. We note that the data are used to motivate the simulations, but, since we define the true DGD to be one that we estimate for each study, and at that point differences with the original study become irrelevant to our comparisons of estimators. Data was filtered down to the last height-to-age Z-score measurement taken at the end of each study for each subject. Subjects were dropped if either their treatment assignment ($A$) or outcome measurement ($Y$) were missing. For covariates ($W$) that were missing, those that were continuous and discrete were imputed using the median and mode, respectively. In both cases, missingness indicator variables were added to the data set for each covariate with missing rows. As mentioned above, the treatment assignment variable ($A$) was binarized if it consisted of more than two treatment arms. The control and treatment groups were originally assigned in each study as described in Section B of the Appendix.
\subsection{Simulation with Undersmoothed Highly Adaptive Lasso}
\subsubsection{Undersmoothed Highly Adaptive Lasso}
We adopted the undersmoothed highly adaptive lasso (HAL) method to generate data for simulations in a (nearly) nonparametric model, with as few assumptions as possible. HAL is a nonparametric regression estimator, which neither relies upon local smoothness assumptions nor is constructed using local smoothing techniques \cite{BenkeserHALE2016}. HAL has been shown to have competitive finite-sample performance relative to many other popular machine learning algorithms.
With the assumption that the target parameter $\psi$ falls in the Donsker class of all cadlag functions (right-hand continuous, with left-hand limits) with finite variation norm, we have the following representation\cite{GillIEBS1995}:
$$
\psi(\bx) = \psi(0) + \sum_{s\subset \{1,2,...,p\}} \int_{0_s}^{\bx_s}\psi_s(du)
$$
where $\bx \in \mathbb{R}^p$ and $s$ denotes the indices of sections of $\psi$.
Then let us further denote $\bx_s = (x_k: k \in s)$ as the subvector with support of $s$, and $\tilde{\bx}_{s,i}$ as the values of the subvector for the $i$\textsuperscript{th} observation. Now $\psi$ can be approximated by $\psi_m$ such that\cite{BenkeserHALE2016}:
$$
\psi_m(\bx) = \psi(0) + \sum_{s\subset \{1,2,...,p\}}\sum_{i=1}^{n} I(\tilde{\bx}_{s,i} \leq \bx_s)d\psi_{m,s,i}
$$
Now if we consider a model with the basis functions $\phi_{s,i} = I(\tilde{x}_{i,s} \leq x_s)$ as predictors and $d\psi_{m,s,i}$ as coefficients, we have\cite{BenkeserHALE2016}:
$$
\psi_{\beta} = \beta_0 + \sum_{s\subset \{1,2,...,p\}}\sum_{i=1}^{n} \beta_{s,i} \phi_{s,i}
$$
By the assumption of finite sectional variation norm (an entropy assumption required of all but two of the estimators), cross-validated TMLE and double-robust EE) we have the corresponding subspace $ \Psi_{n,M} = \{\psi_{\beta}: \beta_0 + \sum_{s\subset \{1,2,...,p\}}\sum_{i=1}^{n} \beta_{s,i} < M \}$\cite{BenkeserHALE2016}.
The HAL estimator starts with a very large number (at most $n*(2^p-1)$) of basis functions that are indicator functions with support at the observed data values. In practice, when some covariates are categorical or binary, the number of unique basis functions will be much fewer than the upper bound. Moreover,
to avoid overfitting, one can define a subspace of the linear model such that: $\sum_{s\subset \{1,2,...,p\}}\sum_{i=1}^{n} |\beta_{s,i}| \le C$, for submodels where the $L_1$-norm is bounded by $C$. The dimension of basis functions can be restricted. For example, one can consider only main-term indicators for each of the original predictors as well as all second order tensor products (interaction terms involving the main effect terms). One can use cross-validated selection of $C$ to optimize the fit of the model to future observations from the data-generating distribution (DGD).
It has been recently shown that undersmoothing HAL (using a $C$ that is larger than that selected by cross-validation) can yield an asymptotically efficient estimators for functionals of the relevant portions of the DGD while preserving the same rate of convergence, and also solving the efficient score equation for any desired path-wise differentiable target feature of the data distribution~\cite{vanderLaanEEPDA2019}. As a consequence, an undersmoothed HAL results in an efficient plug-in estimator of the desired estimand, and moreover, it will also be efficient for any other smooth estimands of the data distribution~\cite{vanderLaanEEPDA2019}. This could serve as the basis for using HAL in our settings; that is, to estimate the DGD by HAL in a way that optimally preserves the relevant functionals. More intuitively, HAL, with the properly chosen $C$, will result in a DGD for simulations that is as close as one can get nonparametrically to the true DGD, without blowing up the variance of estimation. So, it creates a comparison that is as faithful as possible to the DGD of interest, itself represented by a single data set (experiment). Thus, we argue that it can serve as the basis of a simulation where one wishes to compare estimators for the data in hand. We provide more rigorous justification below.
In our study, we only use the undersmoothed HAL to generate data without pre-specifying any parameter of interest. The stopping criterion for this undersmoothing process is to increase the initial bound $C_{cv}$ until the score equations formed by the product of basis functions and residuals are solved at the rate of $\frac{\sigma_n}{\sqrt{n} \log(n)}$ \cite{Laan_underHAL_2019}. Namely, for all ``non-trivial directions'' (combinations of $s,i$ with non-zero coefficients selected by the initial fit) we need:
\begin{align}
|P_n \big(\phi_{s,i}(Y-\bar{Q}_{n,C})\big)| \leq \frac{\sigma_n}{\sqrt{n} \log(n)}
\end{align}
where $P_n$ is the empirical average function and $\sigma_n^2 = Var\big(\phi_{s,i}(Y-\bar{Q}_{n,C_{cv}})\big)$. Following the convention of notation in \cite{LaanTLCI2011}, we define $\bar{Q}_0 = \EE_{\PP_0}(Y|A,W)$ and $\bar{Q}_n$ as its estimate. Also, we use $Q_{W,0} = \PP_0(W), Q_0 = (\bar{Q}_0, Q_{W,0})$, and $Q$ denoting the possible value of true $Q_0$.\\
To justify this criterion, first fix $(s,i)$ and consider the target parameter $\Psi_{s,i}(Q_0) = P_0(\phi_{s,i}\bar{Q}_0) = \EE_{\PP_0}(\phi_{s,i}\bar{Q}_0) =
\EE_{\PP_0}(\phi_{s,i}Y)$. (Notice that for this target parameter, we can treat $A$ as a member of $W$ so that $Q_0$ actually contains $\bar{Q}_0$ and $Q_{W',0} = \PP_0(A,W)$).
The last equality is true since $\EE_{\PP_0}(\phi_{s,i}Y) = \EE_{\PP_0}[\EE_{\PP_0}(\phi_{s,i}Y | A,W)] =
\EE_{\PP_0}[\phi_{s,i}\EE_{\PP_0}(Y|A,W)]$.
We claim that $\phi_{s,i}(Y - \bar{Q}_0)$ is a component of the efficient influence curve (EIC) of $\Psi_{s,i}({Q}_0)$, where we denote the EIC as $D^*_{s,i}({Q}_0)$. To prove this, we can start with the empirical estimator $\frac{1}{n}\sum_{k=1}^{n}\phi_{s,i}y_k$. Observe that
\begin{align*}
\frac{1}{n}\sum_{k=1}^{n}\phi_{s,i}y_k - \Psi_{s,i}({Q}_0) =
\frac{1}{n}\sum_{k=1}^{n}(\phi_{s,i}y_k - \Psi_{s,i}({Q}_0))
\end{align*}
Thus the influence curve of the empirical estimator is $\phi_{s,i}Y - \Psi_{s,i}({Q}_0)$, denote it as $D^0_{s,i}({Q}_0)$. With it, we can obtain $D^*_{s,i}({Q}_0)$ by projecting $D^0_{s,i}({Q}_0)$ onto the tangent space $T(\PP_0)$~\cite{LaanTLCI2011}.
In addition, since $\PP(O) = \PP(Y, A, W) = \PP(Y|A,W)\PP(A,W)$, the tangent space $T(\PP_0)$ can be decomposed as: $T(\PP_0) = T_{Q}(\PP_0) = T_{Q_Y}(\PP_0) \oplus T_{Q_{A,W}}(\PP_0)$~\cite{LaanTLCI2011}. So the projection of $D^0_{s,i}({Q}_0)$ on $T(\PP_0)$ is equal to the sum of the projections of $D^0_{s,i}({Q}_0)$ on $T_{Q_Y}(\PP_0)$ and $T_{Q_{A,W}}(\PP_0)$, namely, $\sqcap(D^0_{s,i}({Q}_0) | T(\PP_0)) = \sqcap(D^0_{s,i}({Q}_0) | T_{Q_Y}(\PP_0)) + \sqcap(D^0_{s,i}({Q}_0) | T_{Q_{A,W}}(\PP_0))$. Thereby $D^*_{s,i}({Q}_0) = \sqcap(D^0_{s,i}({Q}_0) | T_{Q_Y}(\PP_0)) + \sqcap(D^0_{s,i}({Q}_0) | T_{Q_{A,W}}(\PP_0))$. Then
\begin{align*}
\sqcap(D^0_{s,i}({Q}_0) | T_{Q_Y}(\PP_0))
=& D^0_{s,i}({Q}_0) - \EE_{\PP_0}(D^0_{s,i}({Q}_0)|A,W)~\cite{LaanTLCI2011}\\
=& \phi_{s,i}Y - \Psi_{s,i}({Q}_0) - \EE_{\PP_0}(\phi_{s,i}Y -
\Psi_{s,i}({Q}_0)|A,W)\\
=& \phi_{s,i}Y - \Psi_{s,i}({Q}_0) - \EE_{\PP_0}(\phi_{s,i}Y|A,W) +
\Psi_{s,i}({Q}_0)\\
=& \phi_{s,i}(Y - \EE_{\PP_0}(Y|A,W)) \\
=& \phi_{s,i}(Y - \bar{Q}_0)\\
\mbox{and}\\
\sqcap(D^0_{s,i}({Q}_0) | T_{Q_{A,W}}(\PP_0)) =& \EE_{\PP_0}(D^0_{s,i}({Q}_0)|A,W) ~\cite{LaanTLCI2011}\\
=& \EE_{\PP_0}(\phi_{s,i}Y - \Psi_{s,i}({Q}_0)|A,W) \\
=& \EE_{\PP_0}(\phi_{s,i}Y |A,W) -
\Psi_{s,i}({Q}_0)
\end{align*}
So, $D^*_{s,i}({Q}_0) = \phi_{s,i}(Y - \bar{Q}_0) + \EE_{\PP_0}(\phi_{s,i}Y |A,W) -
\Psi_{s,i}({Q}_0)$.\\
Now we have proved that $\phi_{s,i}(Y - \bar{Q}_0)$ is a component of $D^*_{s,i}({Q}_0)$. Another observation from the calculation above is that $D^0_{s,i}({Q}_0) = D^*_{s,i}({Q}_0)$.\\
For different pairs of $(s,i)$, each $D^*_{s,i}(Q_{n,C})$ corresponds with an EIC for a particular target parameter $\Psi_{s,i}(Q_0)$.
For each plug-in estimator $\Psi_{s,i}({Q}_{n,C})$ being asymptotically linear we want at minimal $P_n D^*_{s,i}(Q_{n,C})=o_P(n^{-\frac{1}{2}})$ for every $(s,i)$, which is guaranteed by our choice of criterion. By doing so, we will also be solving $P_n (\sum_{s,i}\alpha(s,i)\phi_{s,i})(Y-\bar{Q}_{n,C})$ for any $\alpha$ vector with finite $L_1$-norm, which enables us to approximate any function of $(A,W)$ with rate approximately equal to $n^{-\frac{1}{3}}$~\cite{vanderLaanEEPDA2019}. So in this way we are rich enough to guarantee to solve any EIC that can be written as $f(A,W)(Y-\bar{Q}_0)$, thereby cover all EIC of features of Q. So the undersmoothing process essentially yields an estimator that is efficient for any target feature of Q that is pathwise differentiable. Combined with the fact that when the bias of an estimator is smaller than $\frac{se}{\max(10, \log{n})}$ then it has minimal impact on coverage, we choose (1) as the stopping criterion based on the proof above.
The specific procedure is stated as follows:
Step 1. Fit the HAL with $L_1$-norm, obtain the set of basis functions and a starting value of $\lambda$, denote it as $\lambda_{cv}$. This $\lambda_{cv}$ is a CV-optimal value of the penalty parameter returned by the hal9001 package \cite{coyle2021hal9001-rpkg, HejaziHSHAS2020}, which is essentially from the ``cv.glmnet'' function with 10-fold cross-validation.
Step 2. Calculate the absolute value of the normalized score equations for each direction:
$$
\big|\frac{\frac{1}{n}\sum_{k=1}^{n}[(y_{k} - \hat{y}_{\lambda})\phi_{s,i}]}
{\sqrt{\frac{1}{n}\sum_{k=1}^{n}[(y_{k} - \hat{y}_{\lambda_{cv}})\phi_{s,i} -
\frac{1}{n}\sum_{k=1}^{n}(y_{k} - \hat{y}_{\lambda_{cv}})\phi_{s,i}]^2}}\big|
$$
Step 3. Take the maximum of this value from all subsets. Compare the max with $\frac{1}{\sqrt{n} \log(n)}$. If the max is larger, then increase the bound $C$ (i.e. decrease the value of the penalty parameter $\lambda$) and refit the HAL.
Step 4. Repeat 1,2,3 until the stopping criterion is satisfied.
In addition, we speed up the algorithm by controlling the number of basis functions in the initial HAL fits. First, we set the maximum interaction degree to $\mathbb{I}(p \geq 20)*2 + \mathbb{I}(p < 20)*3$, where $p$ is the number of covariates. Second, we use binning method to restrict the maximum number of knots to $\sqrt{n}/(2^{d-1})$ for the $d^{th}$ degree basis functions. These hyperparameters can be set through the \texttt{hal9001} package~\cite{coyle2021hal9001-rpkg, HejaziHSHAS2020}. We make the decisions on hyperparameters based on two factors: they can help form a rich model with complex interaction terms and the computing time is acceptable. To make it more rigorous, a cross-validation-based tuning procedure can be considered in future practice.
In the Appendix, we provide a list showing the variables included in the $Q$ models after undersmoothing (Table \ref{tab:variable_list}).
\subsubsection{Data Generating Process}
The DGD for each study was based upon the following structural causal model (SCM):
\begin{eqnarray*}
W &=& f_W(U_W) \\
A &=& f_A(W,U_A) \\
Y &=& f_Y(W,A,U_Y),
\end{eqnarray*}
where $W$, $A$, and $Y$ are, in time ordering, the confounders, the binary intervention of interest and the outcome, respectively, with the $U$ exogenous independent errors and deterministic functions, $f_\cdot$.
Specifically, the following steps were taken:
\begin{enumerate}
\item Covariates $W$ were sampled with replacement from the study data sets with sample size $n$, where $n$ is the size of the original data set.
\item The undersmoothed HAL fit was then used to predict $\mathbb{P}(A = 1 | W)$. The intervention $A$ was then sampled using a binomial distribution with the predicted $\mathbb{P}(A = 1 | W)$.
\item The outcome $Y$ was then simulated with the undersmoothed HAL fit, using the sampled $W$ and simulated $A$ as input. A mean zero, normal random error was added to the simulated $Y$, using a variance based upon the residual variance of the predicted $Y$ (Namely, $\frac{1}{n}\sum_{k=1}^n(\hat{y}_{\lambda} - y_{k})^2$).
\end{enumerate}
Note, we could have used other ways of estimating the error distribution in step 4, including density estimation using HAL, but we left this for future studies.
Steps 1 through 3 were repeated 500 times to generate the data sets for each simulation. For each of the study data (Table \ref{tab_study}), we repeated these steps and analyzed the performance of the competing estimators separately by study.
\subsubsection{Target parameter}
Our treatment variable $A$ is binary, and our outcome $Y$ is continuous, indicating a height-to-age Z-score. $W$ represents the measured covariates in each study. The data structure is defined as: $O = (W, A, Y) \sim{\mathbb{P}_{0}} \in \mathcal{M}$ with $n$ independent and identically distributed (i.i.d.) observations $O_1,...,O_n$, where $\mathcal{M}$ denotes the set of possible probability distributions of $\mathbb{P}_{0}$. The target parameter is a feature of $\mathbb{P}_{0}$ that is our quantity of interest \cite{SchulerTMLEJ2017}. We selected as our target parameter the average treatment effect (ATE), or $\Psi^F(\mathbb{P}_{U,X}) = \mathbb{E}_{{U,X}}(Y(1)-Y(0))$, $\mathbb{P}_{U,X} \in \mathcal{M}^F$; where $\mathcal{M}^F$ denotes the collection of possible distributions of $(U,X)$ as described by the SCM, and $Y(a)$ is the outcome for a subject if, possibly contrary to fact, they received nutrition intervention $A=a$. Given we simulated the data based upon on our causal model, under randomization assumption and positivity assumption we can show that this causal parameter is identified by the following statistical estimand \cite{PearlC2009}:
$$\Psi(\mathbb{P}_{0}) = \mathbb{E}_{W,0}[\mathbb{E}_{0}(Y|A = 1, W) - \mathbb{E}_{0}(Y|A = 0, W)]$$
We calculate the true ATE value for each study by first randomly drawing a large number of observations ($N=50000$) from the empirical of $W$ and using:
$$
\psi_{0} =
\frac{1}{N}\sum_{i=1}^{N}[\mathbb{E}_{0}(Y|A = 1, W) - \mathbb{E}_{0}(Y|A = 0, W)]
$$
where we define the $\mathbb{E}_{0}(Y|A = 1, W)$ and $\mathbb{E}_{0}(Y|A = 0, W)$ term using the fitted undersmooth HAL model. Note that our simulation process insures the randomization assumption is true and there is no asymptotic violation of the positivity assumption. However, there can be practical violations of positivity (close to 0 or 1 estimated probabilities of getting treatment for some observations given the $W$) which can deferentially impact estimator performance.
\subsection{The Estimation Problem}
The target parameter depends on the true DGD, $\mathbb{P}_{0}$, through the conditional mean $\bar{Q}_{0}(A,W) = \mathbb{E}_{0}(Y|A,W)$ and the marginal distribution $Q_{W,0} = \mathbb{P}_0(W)$ of $W$, so we can write $\Psi(Q_{0})$, where $Q_{0}=(\bar{Q}_{0},Q_{W,0})$. Our targeted learning estimation procedure begins with estimating the relevant part $Q_{0}$ of the data-generating distribution $\mathbb{P}_{0}$ needed for evaluating the target parameter \cite{LaanTMLLD2006}.
The two general methods we compare are substitution and estimating equation estimators. Depending on the specific estimator, they can depend on estimators of of the propensity score, $g_0(W) = \mathbb{P}(A=1,W)$, the outcome model, $Q_0(A,W)$, and sometimes both. We use consistent settings when modeling the outcome and the propensity score via Super Learner (see section \ref{computesection} below for details).
The estimators we compare are not exhaustive and new methods will be developed, so such studies will continue to be important sources of information for deciding what to do in practice. We quickly describe the particular estimators compared in our study below.
\subsection{Inverse Probability of Treatment Weighting Estimator}
The inverse probability of treatment weighting (IPTW) is a method that relies on estimates of the conditional probability of treatment given covariates $g(W) = \mathbb{P}(A=1|W)$, referred to as the propensity score \cite{ROSENBAUMCRPSA1983}.
After it is estimated, the propensity score is used to weight observations such that a simple weighted average is a consistent estimate of the particular causal parameter if the propensity score model is consistent \cite{SchulerTMLEJ2017}. For the ATE (if $g$ were known) the weight is $\frac{A}{g(W)} + \frac{1-A}{1-g(W)}$.
The average treatment effect is then estimated by \cite{AustinMBPWD2015}:
$$\psi_{IPTW,n}= \frac{1}{n}\sum_{i=1}^{n}\Big( \frac{A_{i}}{g_n(W_i)}* Y_{i}\Big) - \frac{1}{n}\sum_{i=1}^{n}\Big( \frac{1-A_{i}}{(1-
g_n(W_i))}*Y_{i}\Big)$$
where $g_n(W)$ is the estimate of the true propensity score ($g_0(W)$). IPTW is not a double robust estimator, in that its consistency depends on consistent estimation of the propensity score \cite{LaanUMCL2003}. As it is not a substitution estimator, it is not as robust to sparsity \cite{SchulerTMLEJ2017}. However, it is a commonly used estimator of the ATE, and its form and relationship to well-known inverse probability methods in the analysis of survey data make it relatively popular.
We derived statistical inference using a conservative standard error which assumes that $g$ is known (there is an extensive literature on IPTW estimators, but \cite{LaanUMCL2003} is a good reference for technical details). Specifically, the standard error for this estimator was constructed by multiplying $1/\sqrt{n}$ by the standard deviation of the plug-in resulting influence curve:
$$
Y\Big[\frac{A}{g_n(W)} - \frac{1-A}{1-g_n(W)}\Big] - {\psi}_{IPTW,n}
$$
Since IPTW estimator has many problems such as not
invariant to location transformation of the outcome and suffering from the extreme predictions of $g(W)$ (close to 0 or 1), we use the Hajek/stablized IPTW\cite{ipw_hajek_1971} by normalizing the weights of $Y$ as follows:
$$\psi_{IPTW-Hajek,n} = \frac{\sum_{i=1}^{n}\Big(\frac{A_{i}}{g_n(W_i)}* Y_{i}\Big)}{\sum_{i=1}^{n}\Big(\frac{A_{i}}{g_n(W_i)}\Big)} - \frac{\sum_{i=1}^{n}\Big(\frac{1-A_{i}}{1-g_n(W_i)}*Y_{i}\Big)}{\sum_{i=1}^{n}\Big( \frac{1-A_{i}}{1-g_n(W_i)}\Big)} $$
\subsubsection{Cross-Validated Inverse Probability of Treatment Weighting (CV-IPTW) Estimator}
To avoid problems that arise when $g(W)$ is overfit, we also implemented the CV-IPTW estimator by adding another layer of cross-validation when estimating the propensity score \cite{ChernozhukovDDMLF2018}. Specifically, the same SL fitting procedure was implemented on training sets. Then, we use this estimate of $g$ on the corresponding validation sets; as such, we employ a nested cross-validation. In practice, we used the ``Split Sequential SL'' method, an approximation to the nested cross-validation proposed by Coyle \cite{CoyleCCTL2017}, to speed up the estimation while obtaining similar results to standard nested cross-validation. More details on the implementation can be found in section~\ref{computesection} below.
\subsection{Augmented Inverse Probability of Treatment Weighted (A-IPTW) Estimator}
The other estimating equation method included in our study is an augmented version of the IPTW estimator, aptly named the augmented inverse probability of treatment weighted (A-IPTW) estimator \cite{GlynnIAIP2010}. It is a double robust estimator that is consistent for the ATE as long as either the propensity score model ($g_0(W)$) or the outcome regression ($Q_0(A,W)$) is correctly specified. When compared with the IPTW estimator in a Monte Carlo simulation, A-IPTW typically outperformed IPTW with a lower mean squared error when either the propensity score or outcome model was misspecified \cite{GlynnIAIP2010}.
Intuitively, the A-IPTW improves upon IPTW by fully utilizing the information in the conditioning set of covariates $W$, which contains both information about the probability of treatment and information about the outcome variable \cite{GlynnIAIP2010}. More formal justification comes from the fact that the A-IPTW estimator arises as the solution to the efficient influence curve (a key quantity in semiparametric theory), and thus is locally efficient if both $Q$ and $g$ are correctly specified.
For the ATE, A-IPTW estimator solves the mean of the empirical efficient influence curve and can be expressed explicitly for the average treatment effect as follows:
$$\psi_{A-IPTW}= \frac{1}{n}\sum_{i=1}^{n}\Big(\Big[\frac{A_iY_i}{g(W_i)}-\frac{(1-A_i)Y_i}{1-g(W_i)}\Big]-\frac{(A_i-g(W_i))}{g(W_i)(1-g(W_i))}\Big[(1-g(W_i)){\mathbb{E}}(Y_i|A_i=1,W_i)+g(W_i){\mathbb{E}}(Y_i|A_i=0,W_i)\Big]\Big).$$
The standard error for this estimator was constructed by multiplying $1/\sqrt{n}$ by the standard deviation of the plug-in efficient influence curve:
$$
(Y-\bar{Q}_n(A,W))\Big[\frac{A}{g_n(W)} - \frac{1-A}{1-g_n(W)}\Big] + (\bar{Q}_n(1,W) - \bar{Q}_n(0,W)) - {\psi}_{A-IPTW,n}.
$$
\subsubsection{Cross-Validated Augmented Inverse Probability of Treatment Weighted (CV-A-IPTW) Estimator}
Similar to CV-IPTW, to avoid overfitting of the outcome model ($Q$) or propensity score model ($g$), we implemented the CV-A-IPTW estimator by adding another layer of cross-validation when estimating the $Q$ and $g$. In practice, as discussed above for the IPTW estimator, we used the ``Split Sequential SL'' method proposed by Coyle \cite{CoyleCCTL2017} to speed up the estimation (for more details, see section~\ref{computesection} below).
\subsection{Targeted Maximum Likelihood Estimator (TMLE)}
The targeted maximum likelihood estimator (TMLE) is an augmented substitution estimator that, in context of the ATE, adds a targeting step to the original outcome model fit to optimize the bias-variance trade-off for the parameter of interest \cite{LaanTMLLD2006}. Similar to A-IPTW, TMLE is doubly robust, producing unbiased estimates if either $\bar{Q}_{0}(A,W)$ (i.e. $\mathbb{E}_0(Y|A,W)$) or $g_0(W)$ (i.e. $\mathbb{P}_0(A=1|W)$) is correctly specified. It is asymptotically efficient when both quantities are consistently estimated. As it is a substitution estimator, it is typically more robust to outliers and sparsity than EE estimators \cite{SchulerTMLEJ2017}. A finite sample advantage over estimating equation methods comes from the fact that the estimator respects constraints on the parameter bound, such as ensuring that an estimated probability in the $[0,1]$ range \cite{LaanTMLLD2006}.
The TMLE, like the A-IPTW estimator, requires preliminary estimates of both $g$ and $Q$. The first step in TMLE is finding an initial estimate of the relevant part $Q_{0}$ of data-generating distribution $\mathbb{P}_{0}$. For all estimators, we use an ensemble machine learning algorithm, the Super Learner (SL) algorithm. This avoids arbitrarily using a single algorithm and ensures that the corresponding fit will be optimal (with respect to the true risk) relative to the candidate algorithms used in the estimation. Once this initial estimate has been found, TMLE updates the initial fit to make an optimal bias-variance trade-off for the target parameter \cite{LaanTMLLD2006}.
For the ATE, the TMLE first requires $\bar{Q}_{n}(A,W)$, the estimate of the conditional expectation of the outcome given the treatment and covariates $\bar{Q}_{0}(A,W)$ \cite{SchulerTMLEJ2017}. Next is the targeting step for optimizing the bias-variance trade-off for the parameter of interest. The propensity score ($g_0$) can also be estimated with a flexible algorithm like the Super Learner \cite{vanderLaanSL2007}, and these fits are used to predict the conditional probability of treatment and no treatment for each subject ($g_n(W),1-g_n(W)$). These probabilities are used for updating the initial estimate of the outcome model. This updated estimate is then used to generate potential outcomes for when $A=1$ and $A=0$. Like the G-computation estimator, the TMLE estimate of the ATE is calculated as the mean difference between these pairs \cite{SchulerTMLEJ2017}.
With the ATE as our target parameter, the Super Learner substitution estimator is \cite{LaanTLCI2011}:
$$\psi_{MLE,n}=\Psi(Q_{n}) = \frac{1}{n}\sum_{i=1}^{n}[\bar{Q}_{n}^{0}(1, W_{i}) - \bar{Q}_{n}^{0}(0, W_{i})]$$
where $Q_n$ is the estimate of $Q_0$ and $\bar{Q}_{n}^{0}(\cdot, W)$ the initial estimate of $\bar{Q}_{0}(\cdot, W)$.
The next step is to update the estimator above toward the parameter of interest. The targeting process uses $g_{n}$ in a so-called clever covariate to define a one-dimensional model for fluctuating the initial estimator. The clever covariate is defined as:
$$H_{n}^{*}(A,W) = \Big(\frac{I(A=1)}{g_{n}(1|W)}-\frac{I(A=0)}{g_{n}(0|W)}\Big)$$
A simple, one-variable logistic regression is then run for the outcome $Y$ on the clever covariate, using $logit\bar{Q}_{n}^{0}(A,W)$ as the offset to estimate the fluctuation parameter $\epsilon$. This is used for updating the initial estimate $\bar{Q}_{n}^{0}$ into a new estimate $\bar{Q}_{n}^{1}$ as follows:
$$logit\bar{Q}_{n}^{1}(A,W) = logit\bar{Q}_{n}^{0}(A,W) + \epsilon_{n}H_{n}^{*}(A,W)$$
where $\epsilon_{n}$ is the estimate of $\epsilon$.
The updated fit is used to calculate the expected outcome under $A=1$ ($\bar{Q}_{n}^{1}(1,W)$) and $A=0$ ($\bar{Q}_{n}^{1}(0,W)$) for all subjects. These estimates are then plugged into the following equation for the final TMLE estimate of the ATE:
$$\psi_{TMLE,n}=\Psi(Q_{n}^{*}) = \frac{1}{n}\sum_{i=1}^{n}[\bar{Q}_{n}^{1}(1, W_{i}) - \bar{Q}_{n}^{1}(0, W_{i})]$$
The fitting of both the $Q$ and $g$ models to the entire data set for the substitution estimator requires entropy assumptions on the fits and underlying true models. It is possible to violate this assumption by an overfit of the models of the DGD, and this can occur even when cross-validation is used to choose the resulting fits (though, this helps tremendously). One can generalize both the estimating equation approach and TMLE to estimators that do not need these entropy assumptions by inclusion of an additional layer of cross-validation. This has also been described as double-machine learning in the context of estimating equations \cite{ChernozhukovDDMLF2018}, though it had previously been proposed as a way of robustifying the TMLE \cite{ZhengATCTN2010,LaanTLCI2011}.
The standard error estimate for TMLE can be constructed by multiplying $1/\sqrt{n}$ by the standard deviation of the plug-in efficient influence curve:
$$
(Y-\bar{Q}_n(A,W))\Big[\frac{A}{g_n(W)} - \frac{1-A}{1-g_n(W)}\Big] + (\bar{Q}_n(1,W) - \bar{Q}_n(0,W)) - {\psi}_{TMLE, n}
$$
\subsubsection{Cross-Validated Targeted Maximum Likelihood Estimation (CV-TMLE)}
Though TMLE is a doubly robust and efficient estimator, its performance suffers when the initial estimator is too adaptive \cite{LaanTLCI2011}. Intuitively, if the initial estimator of $Q$ is overfit, there is not realistic residual variation left for the targeting step and the update is unable to reduce residual bias.
To address these shortcomings of TMLE, cross-validated targeted maximum likelihood estimation (CV-TMLE) was developed \cite{ZhengATCTN2010}. This modified implementation of TMLE utilizes 10-fold cross-validation for the initial estimator to make TMLE more robust in its bias reduction step. The result is that one has greater leeway to use adaptive methods to estimate components of the DGD while keeping realistic residual variation in the validation sample.
Whereas CV-TMLE can add robustness by making the estimator consistent in a larger statistical model, there is still another way for finite sample performance issues to enter estimation. Specifically, if the data suffers from a lack of experimentation such that $g_n(W)$ gets too close to 0 or 1, then the estimator can begin to suffer from the unstable inverse weighting in the targeting step, a violation ``positivity''. There are simple methods to avoid this, by choosing a fixed truncation point, such as truncating the estimate of $g$: $g^*_n=max(min(1-\delta,g_n),\delta)$, for some small $\delta$ (typical value is $\delta=0.025)$. However, there exists a more sophisticated method that does a type of model selection in estimating the $g$ model which prevents the update from hurting the fit of the $Q$ model. This is an area of active research and several collaborative-TMLE (C-TMLE) estimators have been proposed, including adaptive selection of the truncation level $\delta$~\cite{LaanTMLLD2006,JuAPSTJ2019a}.
\subsubsection{Collaborative Targeted Maximum Likelihood Estimation (C-TMLE)}
Collaborative Targeted Maximum Likelihood Estimation (C-TMLE) is an extension of TMLE. In the version used for estimation in this study, it applies variable/model selection for nuisance parameter (e.g. the propensity score) estimation in a ``collaborative'' way, by directly optimizing the empirical metric on the causal estimator~\cite{vanderLaanCDRTM2010}. In this case, we used the
original C-TMLE proposed by van der Laan and Gruber~\cite{vanderLaanCDRTM2010}, which is also called ``the greedy C-TMLE algorithm''. It consists of two major steps: first, a sequence of candidate estimators of the nuisance parameter is constructed from a greedy forward stepwise selection procedure; second, cross-validation is used to select the candidate from this sequence which minimizes a criterion that incorporates a measure of bias and variance with respect to the targeted parameter~\cite{vanderLaanCDRTM2010}. More recent development on C-TMLE includes scalable variable-selection C-TMLE~\cite{JuSCTLF2019} and glmnet-C-TMLE algorithm~\cite{JuCLCPA2019}, which might have improved computational efficiency in high-dimensional setting.
\subsection{Computation}
\label{computesection}
Our simulation study was coded in the statistical programming language R\cite{R_general}. We used \texttt{hal9001}\cite{coyle2021hal9001-rpkg, HejaziHSHAS2020} and \texttt{glmnet}\cite{FriedmanRPGL2010} packages to generate the data via undersmoothed HAL. We used \texttt{sl3}\cite{coyle2021sl3}, \texttt{tmle3}\cite{coyle2021tmle3-rpkg} and \texttt{ctmle}\cite{JuGruber_ctmle} packages to implement each of the estimators described above. To estimate the propensity score and the conditional expectation of the outcome, linear models, mean, \texttt{GAM} (general additive models)~\cite{HastieGAMA1986}, \texttt{ranger} (random forest)\cite{WrightRFIRM2017}, \texttt{glmnet} (lasso), and \texttt{XGBoost}\cite{ChenXSTBA2016} with different tuning parameters were used to form the SL library. For ``Study 9'', we dropped \texttt{GAM} and \texttt{ranger} from the learner library to improve the computational efficiency. Ten-fold cross-validation was chosen by default of \texttt{sl3} package for every SL fit. We used logistic regression meta-learner for propensity scores, and non-negative least squares meta-learner for estimating conditional expectation of the outcome. We truncated the propensity score estimates $g_n(W)$ between $[0.025, 0.975]$ for all estimators.
Theoretically, when constructing CV-TMLE, CV-IPTW and CV-A-IPTW estimators, we need to implement nested SL by adding one more layer of cross-validation. Namely, we first split the data, then fit the SL model (which itself uses a cross-validation) on the training set and make predictions on the validation set. Then we rotate the roles of the validation set and finally obtain a vector of cross-validated predictions of propensity scores and conditional expectations. As discussed above, we used the ``Split Sequential SL'' method proposed by Coyle \cite{CoyleCCTL2017}.
After we estimated the relevant parts of the DGD separately for each of the data study data using undersmoothed HAL, the resulting fits were used to simulate data 500 times for each of the 10 studies. Details of the implementation, including the code, can be found in the GitHub repository:
\url{https://github.com/HaodongL/realistic_simu.git}
\section{Results}\label{results}
\subsection{Undersmoothed HAL Models and The True Average Treatment Effect}
We implemented undersmoothed HAL on the real data and used the fitted model to generate sample for each simulation. Details of each model and the resulting true ATE values are presented in Table~\ref{tab_hal}.
\input{tablenplots/table_hal}
For Study 7, 8 and 10, the initial HAL fits of $g$ models contain no variables, so one $A$ is randomized as in a clinical trial. Thereby, the undersoomthing process for $g$ model was omitted for these three studies, and the initial HAL models were used instead. This is not surprising since all ten studies were randomized controlled trials (RCT). Grouping categorical intervention variables into binary variables at data cleaning step might preserve or change the randomization. The remainder of the studies included basis functions in $W$ and so are more akin to observational studies. For only these studies, we also compare the performance of the estimators above with the standard difference-in-means estimates, which is also provides consistent estimators for the ATE for these three data-generating distributions. On the other hand, the counts of non-zero coefficients (``Num.coef.'' in Table~\ref{tab_hal}) in the undersmoothed $Q$ models are large for the remaining studies, and so, regardless of the original treatment mechanism that underlied these studies, these ones do not come from a simple treatment randomization model. The details on the variables included after undersmoothing can be found in Table \ref{tab:variable_list} in the Appendix.
\subsection{Estimators' Performance}
The results are shown in Figure \ref{fig_perf} and Table \ref{tab_perf}. Variance dominates bias for all estimators and so contributes overwhelmingly to the mean squared error (MSE) and the relative MSE (rMSE), where rMSE was relative to the IPTW estimator's MSE. Putting aside Study 1 for now, the MSE/rMSE results suggest that the A-IPTW generally is more efficient than the other estimators, the TMLE, CV-TMLE, CV-A-IPTW and C-TMLE with similar MSE to each other, and the IPTW and CV-IPTW having more erratic performance. The bar plots of the main performance metrics in Table \ref{tab_perf} can be found in the Appendix (see Figure \ref{Figure 3} - \ref{Figure 10})
The 95\% confidence interval (CI) coverage, however, shows different relative performance (Figure \ref{fig_perf}, Table \ref{tab_perf}). Taking 92.5\% as the lower bound defining consistent coverage, then we can observe that: The CV-A-IPTW consistent coverage for all studies. The CV-TMLE and C-TMLE had consistent coverage for all studies except study 1. The TMLE and A-IPTW had coverage ranging from 90\% to 95\% for most studies. IPTW and CV-IPTW estimates of CI had very conservative coverage (close to 100\%) for most studies.
To examine more closely issues of CI coverage, we removed the bias introduced by the estimation procedure for the standard error by using the true sample variance of each estimator (i.e. the sample variance of the estimator across 500 simulations) to derive the standard error (``Coverage2'' in Table \ref{tab_perf}). The coverage of this CI is the oracle coverage one would obtain if one is given the true variance. For this measurement, both CV-TMLE and CV-A-IPTW achieved 95\% coverage in all studies, followed by TMLE, C-TMLE, IPTW and CV-IPTW with 95\% coverage for nine studies. A-IPTW had 95\% coverage for eight studies.
\begin{figure}[!ht]
\centering\includegraphics[
width=17cm,
height=14cm,
keepaspectratio
] {tablenplots/plot_dots}
\raggedright{\scriptsize ~~~~~~~~~~
\textsuperscript{*} The black dots represent the estimator-specific medians across all ten studies.}
{\scriptsize ~~~~~~~~~~
\textsuperscript{*} The reference line is 1 for rMSE, 0 for bias and 0.95 for coverage.}
\raggedright{\scriptsize ~~~~~~~~~~
\textsuperscript{*} The original rMSE value (14.698) of TMLE estimator for Study 1 was truncated at 6.3.}
\caption{\label{fig_perf} \text{Dot plot of the main metrics of performance}}
\end{figure}
\input{tablenplots/table_perf}
The simulations suggest, across 10 realistic data-generating distributions, that CV-A-IPTW, CV-TMLE and C-TMLE has overall relatively good performance in terms of MSE and reliable 95\% coverage. The A-IPTW estimator had superior MSE-based performance, though the confidence interval coverage was sometimes between 90\% and 95\%. However, plugging in the true standard deviation of the A-IPTW estimator instead of the plug-in influence-curve based one typically used resulted in good coverage. This suggests more robust standard error (SE) estimators could make it a more compelling choice than the empirical performance in these simulations. In addition, CV-A-IPTW can improve the coverage of A-IPTW in most cases, but, due to the estimator being consistent in a bigger model, will have bigger MSE. The results at least show that the CV-TMLE as implemented in the tmle3 package \cite{coyle2021tmle3-rpkg} can provide robust inferences, suggesting using it ``off the shelf'' provides reliable results. In next section, we will discuss situations where even the CV-TMLE under-performed, potentially because of small sample size and related empirical positivity violations \cite{PetersenDRVPF2012}.
\subsection{Exploration on Positivity Violation}
We now consider Study 1, where the TMLE and CV-TMLE had significantly anti-conservative coverage. In this case, certainly one cause appears to insufficient experimentation of treatment within some covariate groups.
Specifically, consider Figure \ref{fig3}, which shows the distributions of the adjustment variable, \emph{\text{W\_perdiar24}} in Study 1. As one can see, there are large differences in the marginal distribution of this covariate; in fact, a fit $g_n$ without smoothing would result in a perfect positivity violation. However, given the variance-bias trade-off resulting in the estimators, it is possible that these empirical violations are smoothed over. A potential consequence of this positivity violation is that the resulting estimator, for the parameter which requires support in the data, will be unstable and biased.
\begin{figure}[!ht]
\centering\includegraphics[
width=9cm,
height=4.5cm,
] {tablenplots/plot_sas_perdiar}
\caption{\label{fig3} \text{Distributions of \emph{W\_perdiar24} in Study 1 by intervention group}}
\end{figure}
Table~\ref{tab_dropvar} shows the performance of estimators before and after dropping the variable \emph{\text{W\_perdiar24}} in Study 1. We can observe that all estimators can benefit from removing the problematic variable in terms of higher coverage or lower MSE.
\input{tablenplots/table_drop}
\subsection{Estimators' Efficiency in Randomized Experiment Setting}
As mentioned in earlier section, the initial HAL models for propensity score include no variables for Study 7, 8 and 10, which leads to randomized experiments in the corresponding simulations. In these cases, we add the ``difference-in-means'' estimator (i.e. $\frac{1}{n_1}\sum_{i=1}^n A_iY_i - \frac{1}{n_0}\sum_{i=1}^n (1-A_i)Y_i$) with its variance estimator proposed by Neyman in 1923~\cite{Splawa-NeymanAPTAN1990}. Table~\ref{table_diffmean} shows that the CV-TMLE and CV-A-IPTW estimators still gain efficiency in the randomized experiments setting. This is consistent with proposals for using doubly robust estimators of the ATE in randomized trials if there are informative covariates that can increase efficiency over simple, unadjusted estimates \cite{MooreCARTJ2009,tsiatis2008covariate}.
\input{tablenplots/table_dim}
\section{Conclusion}\label{conclu}
The ultimate goal of studies, such as ours, is to move incrementally towards algorithms that can take information on the design, causal model and known constrains in order to produce a data-adaptively optimized estimator without relying on arbitrary model assumptions. Asymptotic theory can provide guidance on some of the choices, but asymptotic efficiency is not a guarantee for superior performance in finite samples. Thus, simulation studies that are based on realistic DGD's are invaluable for both evaluating estimators and modifying them to increase finite-sample robustness. We provided results supporting the use of a strategically undersmoothed HAL for estimating the relevant components of the DGD in data-driven simulations. Though much remains unresolved, such an approach could be an approach for generating synthetic data \cite{mannino2019real}.
Our results suggest that if accurate inferences are the highest priority, then the CV-A-IPTW, CV-TMLE, and C-TMLE are good choices for providing robust inferences. Specifically, the results suggest that CV-TMLE might serve as an ``off the shelf'' algorithm given that 1) it is an asymptotically linear estimator; 2) it is consistent in a large class of statistical models; 3) it allows for the use of aggressive ensemble learning, while protecting the final performance of the estimator with an outer layer of cross-validation; 4) its influence-curve-based standard error combined with the well-behaved (normal) distribution of the estimator results in near perfect coverage for all but one of the studies used. Our results also suggest that modifications to the algorithms for other estimators (such as improving the SE estimator for the A-IPTW) would result in an estimator with acceptable CI coverage and relatively low MSE. We also suggest one basis for deciding which estimator to use for particular data is to perform a similar simulation study for the data based upon fitting the undersmoothed HAL to derive the DGD. Then, one could choose to report the results from the estimator that provided the most reliable performance in such a simulation study. Of course, this is itself a form of over-fitting, since it uses the data both for estimator selection and for reporting the results of that estimator applied to the original data. However, it seems better than applying an arbitrary estimator and hoping that the advertised asymptotic performance matches the performance on the data of interest. Finally, our results support the observations that careful use of covariate information can be used to gain efficiency in the randomized experiment setting.
\section*{Acknowledgments}
This research was financially supported by a global development grant (OPP1165144) from the Bill \& Melinda Gates Foundation to the University of California, Berkeley, CA, USA. We would also like to thank the following collaborators on the included cohorts and trials for their contributions to study planning, data collection, and analysis: Muhammad Sharif, Sajjad Kerio, Ms. Urosa, Ms. Alveen, Shahneel Hussain,
Vikas Paudel (Mother and Infant Research Activities), Anthony Costello (University College London), Noel Rouamba, Jean-Bosco Ouédraogo, Leah Prince, Stephen A Vosti, Benjamin Torun, Lindsey M Locks, Christine M McDonald, Roland Kupka, Ronald J Bosch, Rodrick Kisenge, Said Aboud, Molin Wang, Azaduzzaman, Abu Ahmed Shamim, Rezaul Haque, Rolf Klemm, Sucheta Mehra, Maithilee Mitra, Kerry Schulze, Sunita Taneja, Brinda Nayyar, Vandana Suri, Poonam Khokhar, Brinda Nayyar, Poonam Khokhar, Jon E Rohde, Tivendra Kumar, Jose Martines, Maharaj K Bhan, and all other members of the study staffs and field teams. We would also like to thank all study participants and their families for their important contributions. We are grateful to the LCNI5 and iLiNS research teams, participants and people of Lungwena, Namwera, Mangochi and Malindi, our research assistants for their positive attitude, support, and help in all stages of the studies.
In addition, this research used the Savio computational cluster resource provided by the Berkeley Research Computing program at the University of California, Berkeley (supported by the UC Berkeley Chancellor, Vice Chancellor for Research, and Chief Information Officer). The authors would like to further thank the university and the Savio group for providing computational resources.
\subsection*{Author contributions}
Conceptualization: A.E.H., H.L., S.R. \\
Funding Acquisition: J.M.C., A.E.H., M.J.V., B.F.A. \\
Data curation: A.M., J.B., J.C. \\
Formal analyses: H.L., S.R.\\
Methodology: H.L., S.R., A.E.H., M.J.V. \\
Visualization: H.L., S.R., J.C. \\
Writing – Original Draft Preparation: H.L., S.R. \\
Writing – Review \& Editing: A.E.H., H.L., R.V.P., N.H., I.M., B.F.A., J.B.
\subsection*{Financial disclosure}
None reported.
\subsection*{Conflict of interest}
The authors declare no potential conflict of interests.
\subsection*{Data availability}
The data used in this analysis was held by Bill \& Melinda Gates Foundation in a repository. The sensitive information contained in the data was still considered theoretically identifiable and can not be released to the public at this point, with the exception of the WASH Benefits trials. We provide the data from "WASH Benefits Bangladesh"\cite{LubyEWQSM2018} (Study 2) and "WASH Benefits Kenya"\cite{StewartEWQSA2018} (Study 3) as example data sets in the GitHub repository:
\url{https://github.com/HaodongL/realistic_simu.git}
\newpage
|
\section{Introduction} \label{sec:intro}
Weak gravitational lensing is a powerful and increasingly utilized tool for measuring how mass is distributed throughout the Universe \citep[e.g.][]{mandelbaum18}. Applications include use of cosmic shear to probe the mass density and the growth of structure in the Universe (e.g. \citealt{desy1, desy3, hikage19}), galaxy-galaxy lensing to characterize halo masses of galaxies in different stellar-mass regimes and perform $3 \times 2$ correlation studies (e.g. \citealt{alexie12, krause17}), and cluster lensing profiles to measure cluster-scale halo mass profiles and halo shapes \citep[e.g.][]{umetsu18, mandelbaum06}. These science goals have motivated large-scale weak-lensing surveys with new instruments, telescopes, and even satellite missions. Major projects include the Hyper Suprime Cam \citep[HSC,][]{aihara18}, the Kilo Degree Survey \citep[KiDS,][]{kuijken15}, the Legacy Survey of Space and Time \citep[LSST,][]{lsst}, \textit{Euclid} \citep{laureijs11}, the Nancy Grace Roman Space Telescope \citep{wfirst}, and the Dark Energy Survey \citep[DES,][]{krause17, desy1, desy3}.
Weak-lensing surveys seek to measure distortions in a galaxy's shape caused by lensing shear, a geometric effect caused by the gravitational lensing of a background source by mass in the foreground. To first order, this distortion results in background galaxies becoming elongated in the direction tangential to the surface-density gradient of the foreground mass distribution.
A number of factors complicate lensing measurements in conventional surveys that use imaging to measure galaxy shapes. Weak lensing distortions only induce a ${<}1\%$ change in observed ellipticity. Measurements at this level of precision are typically achieved by statistically stacking results from multiple sources, but our uncertain knowledge of the intrinsic or ``pre-lensed'' galaxy ellipticity distribution (``shape noise'') typically necessitates thousands of stacked sources to detect a lensing signal.
Potential biases become a greater concern with added complications, such as intrinsic alignments of source galaxies with near-foreground dark matter structure (e.g., filaments) and systematic errors in shape measurements resulting from instrumentation \citep[e.g.,][]{troxel15,mandelbaum18}.
Very large samples are therefore required to drive down statistical errors and test for systematics that can otherwise overwhelm the shear signal.
For instance, the CLASH survey \citep{umetsu14} used ${\sim}10^4$ background galaxies around each of 20 foreground Abell clusters to recover mass density profiles precise enough to constrain dark matter halo mass profiles. Meanwhile, cosmic shear surveys like the DES \citep{desy1} use ${\sim}10^7$ galaxies to sufficiently characterize their signal, but even these large samples remain susceptible to systematic biases. All of these efforts require deep imaging over wide areas, typically carried out on 2 meter- to 8 meter-class telescopes.g
The burgeoning subfield of kinematic weak lensing (KWL), also referred to as ``precision weak lensing'' in the literature \citep{gurri20, gurri21}, provides an additional means of inferring lensing shear by measuring distortions in the projected velocity field of source galaxies. Although KWL requires more expensive spectroscopic observations, many fewer galaxies are needed to detect a signal. This is because the velocity field encodes a trace of the galaxy's original, pre-lensed coordinate space. The induced lensing distortions can be fit directly, eliminating or at least dramatically reducing the shape noise. If future weak-lensing analyses could make use of KWL, they could enable higher spatial resolution maps of foreground mass, more sensitive mass measurements, and independent checks on results from conventional imaging surveys \citep{huff20}.
The literature investigating applications of KWL has fallen into two regimes: high-spatial-resolution kinematic measurements on small samples, and lower-precision kinematic measurements derived from large samples. The idea of KWL was first put forward by \citet{blain02}, who determined that weak lensing shear would change the symmetry of a galaxy's rotational velocity as measured in an azimuthal ring of constant radius. The idea was further developed by \citet{morales06}, who proposed measuring the angle between the kinematic major and minor axes, which are no longer perpendicular in a lensed galaxy. A similar idea was developed by \citet{dbd15}, who suggested searching for lensing-induced asymmetries in the reflection symmetries of galaxy kinematic data. All these authors emphasize the potential of making a shear measurement with a single galaxy and obtaining a measurement that would be independent of shape noise. However, none of these studies reported a detection due to the small size of lensing effects at nearby redshifts where the high-spatial-resolution spectroscopic measurements required for precise measurements are most readily available.
\citet{huff20} instead proposes implementing KWL on a survey scale, considering large samples where individual measurements with less kinematic information are statistically stacked.
Building on the work of \cite{huff13}, they propose targeted measurements of lensing-induced differences in the projected velocity along the major and minor axes.
\cite{wittman21} performs a Fisher Matrix analysis of this technique in a hypothetical DES-scale survey to derive theoretical limits on the covariances of the derived lensing parameters.
In the first reported detection of kinematic weak lensing, \cite{gurri20} stake out a middle ground in sample size versus per-galaxy information content.
They collect and analyze 2D velocity fields of ${\sim} 20$ galaxies at $z < 0.15$, and report a positive mean shear amplitude detected at $2.5\sigma$. They forward-model each source galaxy with a rotating thin-disk model that allows shear to vary and apply their technique to selected source galaxies likely to be sheared by foreground halos. The detection reported by \cite{gurri20} is an exciting development in the young field of KWL, but also highlights upcoming observational challenges. For instance, they find discrepancies between their theoretical and observed shear magnitudes, which they say likely arise from a combination of intrinsic kinematic irregularities in their sample galaxies (``dynamical shape noise") and the scatter in the stellar-halo mass relation \citep[see][]{gurri21}. It will be difficult to obtain the kinds of highly-sampled and high-S/N velocity fields in \cite{gurri20} for source galaxies at $z \lesssim 1$, but it is at these redshifts where probes of structure formation are most needed and the lensing kernel more favorable \citep{weinberg13}.
With this challenge in mind, we build on the techniques of \cite{gurri20} by introducing additional constraints from galaxy shape measurements into the KWL formalism using a Bayesian forward model.
We show that by including measurements of the major-axis position angle of an ellipse fit to the surface brightness profile and comparing it to the position angle derived for the sheared velocity field, we can improve the per-galaxy shear S/N by several times in many cases. This motivates us to explore the potential use case of measuring the mass and shape of galaxy cluster halos at $z \sim 0.3$ \citep{bartelmann17} with IFU observations of 50--100 galaxies at $z \sim 0.7$. Immediately valuable for studies of individual clusters, a future survey program could aid mass calibrations required for cluster cosmology \citep[e.g.][]{sptsz19} and opens new possibilities for wide-scale KWL surveys at $z \lesssim 1$ with a dedicated survey.
The paper is structured as follows: Section \ref{sec:conventional} gives background on relevant weak-lensing theory and introduces our formalism for how lensing affects the shape of a galaxy. Section \ref{sec:kwl} studies the impact of lensing shear on the shape of a rotating galaxy and compares the merits of different techniques for measuring kinematic shear observables. Section \ref{sec:phot} develops our modeling framework for including offsets between the kinematic axes and axes derived from imaging for a given galaxy and characterizes the improvement of the precision of lensing measurements and the important increase of S/N that results. Section \ref{sec:conclusions} gives a summary and looks forward to future applications of the technique.
Throughout this paper, we assume a Planck 2018 cosmology \citep{planck18} as implemented in the \texttt{Colossus} cosmology package from \citet{diemer18}, with $H_0 = 67.36$ km s$^{-1}$ Mpc$^{-1}$ and $\Omega_m = 0.3111$ at $z=0$.
\section{Weak Lensing Effects on Imaging} \label{sec:conventional}
To understand the benefits image position angle measurements can have for KWL, we first review key aspects of how lensing geometry changes the shape of galaxies. Imaging-based lensing surveys typically model background galaxies as ellipses with a measured on-sky position angle and ellipticity that are used to derive lensing amplitude. This serves both as context for the current state of the field and as a basis for the techniques developed in Section \ref{sec:phot}. For a more detailed treatment of the subject, see \cite{bartelmann17}.
\subsection{General Lensing Theory}
Gravitational lensing is a well-established result of general relativity and its effects have been well-characterized \citep{me91}. Because photons travel along geodesics, when they pass through regions of spacetime that have been distorted by a mass distribution, the geodesic is deflected, curving their paths. So the observed image of a background galaxy will be distorted by any foreground mass along the line of sight.
If we assume that the characteristic scale of the foreground mass distribution is small compared to the other relevant distance scales in the system --- the angular diameter distance from the observer to the foreground mass, $D_L$, the distance from the observer to the background source, $D_S$, and the distance from the foreground mass to the background source, $D_{LS}$ --- we may adequately describe the system using the thin lens approximation. This stipulates that the deflection of the photons in the system happens in an infinitely thin plane located at the lens position and that the photon travels in a straight line everywhere else. The magnitude of the angular deflection relates to the projected surface density $\Sigma$ of the 3D foreground mass distribution $\rho$, where
\begin{equation}
\Sigma(\bm{\xi}) = \int\rho(\bm{\xi, z}) d\bm{z},
\end{equation}
\noindent for $\bm{\xi} = \sqrt{\bm{x}^2+\bm{y}^2}$ with $\bm{x}$ and $\bm{y}$ being the physical coordinates of the mass distribution on-sky and $\bm{z}$ being the line of sight vector from the observer. The angular deflection of a photon is given by
\begin{equation}
\bm{\alpha}(\bm{\xi}) = \frac{4G}{c^2}\int \Sigma(\bm{\xi}') \frac{\bm{\xi-\xi'}}{|\bm{\xi-\xi'}|^2}\, d^2\bm{\xi}'.
\end{equation}
\noindent The deflection angle can also be given in terms of the lensing potential $\psi$ as
\begin{equation}
\bm{\alpha}(\theta) = \frac{D_S}{D_{LS}} \nabla_\theta \psi,
\end{equation}
\noindent where $\theta$ is the angular separation of the source as seen from the perspective of the observer: $\theta = |\bm{\xi}|/D_L$.
For a point mass lens in the weak lensing regime,
$\psi$ can be written as
\begin{equation} \label{psi}
\psi = \frac{4GM}{c^2}\frac{D_{LS}}{D_L D_S} \ln |\bm{\xi}|.
\end{equation}
The magnitude of the lensing effect is often expressed in terms of the critical surface mass density, which defines the characteristic angular scale of a lensing system:
\begin{equation} \label{sigcrit}
\Sigma_\textrm{crit} = \frac{c}{4\pi G}\frac{D_S}{D_L D_{LS}}.
\end{equation}
\noindent The distortion manifests itself in two ways. The first is the convergence $\kappa$, which has magnitude $\kappa = \Sigma/\Sigma_\textrm{crit}$ and magnifies the background source. However, for the purposes of this paper, we will be more interested in the lensing shear.
The shear, as its name implies, distorts the coordinate system of the source, changing its observed shape. This distortion can be broken into two separate components defined by their relation to the derivatives of the lensing potential:
\begin{equation}
\gamma_+ = \frac{1}{2}\left(\frac{\partial^2\psi}{\partial\theta_x^2} - \frac{\partial^2\psi}{\partial\theta_y^2}\right), \ \ \ \gamma_{\times} = \frac{\partial^2\psi}{\partial\theta_x \partial\theta_y}.
\end{equation}
\noindent where $\theta_x$ and $\theta_y$ are perpendicular angular coordinates in the arbitrarily defined coordinate system projected onto the lens plane. Following the formalism of \citet{huff13}, the shear aligned with the coordinate axes is called $\gamma_+$, while the cross term $\gamma_{\times}$ is aligned with the diagonals of the coordinate system. For a simplified case of an axisymmetric lens, the shear magnitude is given by
\begin{equation} \label{shear}
|\gamma| = \frac{\Delta \Sigma}{\Sigma_\textrm{crit}},
\end{equation}
\noindent where $\Delta \Sigma$ is the differential projected surface mass density, the difference between the value of $\Sigma$ at a given radius $r_0$ and the mean $\Sigma$ within that radius:
\begin{equation}
\Delta \Sigma = \bar{\Sigma}(r < r_0) - \Sigma(r_0).
\end{equation}
\noindent In the simplified axisymmetric case, the two shear components can also be expressed as
\begin{equation} \label{sheardefs}
\gamma_+ = |\gamma|\cos 2\theta\ \ \ \gamma_\times = |\gamma|\sin 2\theta,
\end{equation}
\noindent with $\theta = \tan^{-1} (\theta_y/\theta_x)$. Positive and negative values for the shear result in distortions in opposite directions. We assume $|\gamma| < 0.1$ in this paper to remain in the weak-lensing regime where the magnitude of the distortions are small and the small angle approximation is still valid.
\subsection{Image Distortions} \label{sec:phottheory}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{demoplot.pdf}
\caption{Representation of how shear from gravitational lensing affects the positions of major and minor axes derived from kinematics and imaging/photometry. The original orientation of the galaxy is shown as a dashed elliptical isophote, with the dashed lines marking the positions of the unlensed major and minor axes (photometric and kinematic are the same). A lens to the upper left of the galaxy applies a shear, distorting the galaxy's intrinsic coordinate system and velocity field contours according to the transformation matrix $A$ (Equation \ref{Amatrix}). The solid ellipse represents how a new elliptical isophote would be fit to the sheared galaxy, along with its new imaged major and minor axes. The colored contours represent the observed velocity field, with red and blue indicating receding and approaching sides of the galaxy, respectively, and darker colors indicating higher velocity magnitude. The point along the lensed isophote that is measured as the lensed imaged major axis (filled black circle) was not on the major axis of the unlensed galaxy (unfilled black circle), leading to a discrepancy between the lensed axis observed in imaging and the kinematic axis (filled white circle). A similar effect happens with the minor axis. The galaxy shown has an axis ratio $q = 0.8$ and the shear applied is $\gamma_\times = 0.1$.}
\label{example}
\end{figure}
In the weak-lensing regime, the recognizable features of strong lensing, like arcs and multiple images, give way to tiny distortions in background sources that are usually only detectable through statistical methods applied to large samples. Weak lensing effects are well-parameterized by a single transformation matrix that can be applied to the coordinate system of the background source:
\begin{equation} \label{originalAmatrix}
A=
\begin{pmatrix}
1-\kappa+\gamma_+ & \gamma_\times \\
\gamma_\times & 1-\kappa-\gamma_+
\end{pmatrix}
\end{equation}
\noindent If we do not care about the size and luminosity of the source, we may disregard the convergence $\kappa$ since it is small in the weak-lensing regime, so we may simplify $A$ to be:
\begin{equation}\label{Amatrix}
A\approx
\begin{pmatrix}
1+\gamma_+ & \gamma_\times \\
\gamma_\times & 1-\gamma_+
\end{pmatrix}.
\end{equation}
\noindent The effect of $\gamma_+$ is to symmetrically elongate the source along the tangential axis\footnote{Note that this is distinct from $\kappa$, which elongates the galaxy in all directions equally, causing a perceived magnification.}, meaning the on-sky orientation of the source is not substantially affected. However, $\gamma_\times$ shears the source, causing an effective rotation.
We can apply this transformation to a toy model of a galaxy, a typical background source in a weak-lensing survey. We may model the galaxy as a circle inclined at some angle relative to the observer, so it is seen simply as a series of concentric elliptical isophotes, each satisfying
\begin{equation}
1 = q^2x^2 + y^2,
\end{equation}
\noindent where $q$ is the axis ratio of the projected ellipse. Applying the shear transformation matrix to this object, we find that, post-lensing, the coordinate system has become $\bm{x}' = A^{-1}\bm{x}$ for $\bm{x} = (x,y)^T$, so
\begin{equation} \label{transform}
1 \approx q^2x'^2(1-2\gamma_+) + y'^2(1+2\gamma_\times) - 2x'y' \gamma_\times (1+q^2),
\end{equation}
\noindent where we have dropped higher order terms in $\gamma_+$ and $\gamma_\times$ since we expect the magnitude of these terms to be small in the weak lensing regime.
We then treat the deformation of the elliptical isophote as a rotation rather than a shear. The general equation of an ellipse rotated by angle $\alpha$ relative to the origin is
\begin{multline}
1=\left(\frac{\cos^2\alpha}{a^2} + \frac{\sin^2\alpha}{b^2}\right)x^2 + \left(\frac{\sin^2\alpha}{a^2} + \frac{\cos^2\alpha}{b^2}\right)y^2 \\
- 2 \cos\alpha \sin\alpha \left(\frac{1}{b^2} - \frac{1}{a^2}\right),
\end{multline}
\noindent where $a$ and $b$ are the lengths of the major and minor axes, respectively. We may greatly simplify this equation if we restrict ourselves to the weak-lensing regime. By assuming $\alpha$ to be small, we can apply the small angle approximation and ignore higher order terms in $\alpha$. If we then multiply through by $b^2$ to put everything in terms of the axis ratio $q = b/a$ and set $b=1$ in arbitrary units, we have
\begin{equation} \label{rotate}
1 \approx q^2 x^2 + y^2 - 2\alpha xy\left(1-q^2\right).
\end{equation}
Equations \ref{transform} and \ref{rotate} are identical except for the cross term. If we assume that the shear can indeed be thought of as a rotation, we can set these two equations equal to each other to find
\begin{equation}\label{angphot}
\alpha (1-q^2) \approx \gamma_\times (1+q^2)\ \ \ \Rightarrow\ \ \ \alpha \approx \gamma_\times \frac{1+q^2}{1-q^2}.
\end{equation}
\noindent So the effect of the shear on an isophote can be well-approximated in the weak-lensing regime as a rotation, as seen in Figure \ref{example}. The rotation angle is determined by $\gamma_{\times}$, which depends on the magnitude of the overall shear, the relative positions of the foreground lens and the background source, and the axis ratio of the background source, which are in turn determined by Equation \ref{sheardefs} and the inclination of the galaxy. The magnitude of $\alpha$ is maximized when the background source has a low inclination, has a position angle that is misaligned with the radial vector by $45^\circ$, and when the foreground lens is more massive.
The axis ratio $q$ changes as a result of lensing as well, as seen by the effect of the transformation matrix $A$ on a point at $(a,b)$:
\begin{multline}
\begin{pmatrix}
1+\gamma_+ & \gamma_\times \\
\gamma_\times & 1-\gamma_+
\end{pmatrix}
\begin{pmatrix}
a \\ b
\end{pmatrix}
\\ =
\begin{pmatrix}
(1 + \gamma_{\times} + \gamma_+)\, a \\
(1 + \gamma_{\times} - \gamma_+)\, b
\end{pmatrix}
=
\begin{pmatrix}
a' \\ b'
\end{pmatrix},
\end{multline}
\noindent where we have defined $(a', b')$ as the major and minor axis lengths of the lensed galaxy. Thus, the lensed axis ratio $q'$ is
\begin{equation}
q' = \frac{b'}{a'} = \frac{(1 + \gamma_{\times} - \gamma_+)}{(1 + \gamma_{\times} + \gamma_+)} \frac{b}{a} \\
\approx \left(1- \frac{2 \gamma_+}{1 + \gamma_{\times}}\right) q,
\end{equation}
\noindent where we have again exploited a Taylor series expansion to make a first order approximation. What observers see is the post-lensing axis ratio $q'$ rather than $q$, so we again approximate to write this as
\begin{equation} \label{qdistortion}
q \approx \frac{q'}{1-2\gamma_+}.
\end{equation}
\noindent We see that for small $\gamma_+$, $q \approx q'$, and because we will be concerned mostly with cases where this is true, we will ignore the distinction between pre- and post-lensing axis ratios for the remainder of the paper.
However, neither the angular distortion nor the axis ratio distortion can be directly measured in images of any single galaxy because nothing is known about its original shape. Before the galaxy is lensed, it has some intrinsic position angle and ellipticity that provide the baseline for any lensing distortions. With conventional imaging surveys, the only way we can perceive changes is by looking at the statistical properties of the sample in aggregate. A large sample of galaxies under the same shear will show a net alignment perpendicular to the induced shear that deviates from a random distribution of position angles. Measuring this net alignment is a primary goal of imaging-based weak lensing surveys.
\section{Weak Lensing Effects on Kinematics} \label{sec:kwl}
While it is impossible to tell the difference between an elliptical isophote that has been rotated and elongated due to lensing shear and one that was never sheared, the same is not true for projected kinematics of a rotating disk. By taking spectroscopic measurements across the face of a rotating galaxy, the relative velocity of different parts of the galaxy can be determined from the Doppler shift of spectral features compared to the systemic velocity of the galaxy.\footnote{In this paper, we assume an infinitely thin, rotating disk. This assumption may cause problems with highly inclined disks.} The kinematic measurements that make up the velocity field are associated with specific coordinates in the galaxy's intrinsic plane, a relationship that is not broken by gravitational lensing. This probe of the galaxy's intrinsic geometry enables much more precise per-galaxy lensing measurements.
In what follows, our analysis is motivated by reducing the final statistical errors on a galaxy's shear measurement, as limited by the quality of the data. We do not treat here the per-galaxy error that stems from intrinsic irregularities in the physical structure of the galaxy. Features that deviate from the model used in the fit will bias the results regardless of the quality of the data. However, this dynamical shape noise \citep{gurri20} is expected to be randomly distributed across a sample and should beat down as the sample size increases. We will return to the role of statistical and systematic errors in Section \ref{sec:gurri}.
\subsection{Kinematic Axis Distortion} \label{sec:radon}
Unlike the photometric measurements made by fitting isophotes, the transformation of a galaxy's velocity field cannot be modeled simply as a rotation of the projected field. In order to illustrate the effect of shear on the velocity field, we examine the behavior of two reference locations: a point on the major axis and one on the minor axis.
A point on the major axis of the unlensed galaxy (unfilled white circle on Figure \ref{example}) can be described by the coordinates $(x,0)$ for a rectilinear coordinate system aligned with the major and minor axes. We define the on-sky angle of the major axis in a particular reference frame to be the kinematic position angle (PA). Applying the transformation matrix $A$ to this coordinate location, we find that this point gets moved to ${((1+\gamma_+)\,x,\ \gamma_\times x)}$ (filled white circle), so the angular displacement of the lensed major axis from the original major axis is
\begin{equation} \label{majkin}
\theta_{maj} = \tan^{-1} \left(\frac{\gamma_\times}{1+\gamma_+}\right) \approx \frac{\gamma_\times}{1+\gamma_+},
\end{equation}
\noindent where we have performed a Taylor series expansion of the inverse tangent and taken just the first order since we expect the shear to be small in the weak-lensing regime.
We can do a similar transformation to a point on the unlensed minor axis $(0,y)$ to find
\begin{equation} \label{minkin}
\theta_{min} = \tan^{-1}\left(\frac{1-\gamma_+}{\gamma_\times}\right) \approx \frac{\pi}{2} - \frac{\gamma_\times}{1-\gamma_+},
\end{equation}
\noindent where we have again performed a first-order Taylor series expansion.
We can again exploit Taylor series to expand Equations \ref{majkin} and \ref{minkin} with respect to $\gamma_+$:
\begin{equation} \label{angkin}
\theta_{maj} \approx \gamma_{\times}(1 - \gamma_+);\ \ \ \ \ \theta_{min} \approx \frac{\pi}{2} - \gamma_{\times}(1 + \gamma_+).
\end{equation}
\noindent The dependence of the angular differences on $\gamma_+$ is on the order of $|\gamma|^2$ (assuming that $\gamma_+ \approx \gamma_{\times}$, which is true for most galaxy-lens orientations). This means that we can neglect the $\gamma_+$ term\footnote{The same is true for the $\kappa$ term that we have already neglected. If we had kept it, it would have shown up in the denominator here as well and could be neglected with the same logic.} and just write
\begin{equation} \label{angkinsimp}
\theta_{maj} \approx \gamma_{\times}; \ \ \ \ \ \theta_{min} \approx \frac{\pi}{2} - \gamma_{\times}.
\end{equation}
\noindent These simple relations allow us to estimate the shear by measuring how much the axes differ from being orthogonal, as was proposed by \cite{morales06}. In an ideal velocity field, the angle between the major and minor kinematic axes should be
\begin{equation} \label{perp}
\theta_\perp \approx \frac{\pi}{2} \pm 2\gamma_{\times}.
\end{equation}
\noindent If we could determine the position angles of the velocity maximum, minimum, and zeros and see how they differ from perpendicularity using a technique like the Radon transform \citep{stark18} or kinemetry \citep{kinemetry}, we could use the above result to gain an estimate of the induced shear.
\subsection{Sheared Velocity Field Fitting} \label{sec:data}
We can produce a more complete picture of the shear of a velocity field if we model the distortion of the entire galaxy rather than just its major and minor axes. We construct a Bayesian forward model that allows us to utilize all of the velocity measurements to obtain a best-fit sheared velocity field.
For the purpose of comparing different KWL techniques in our regime of interest, we create a simulated observation of a galaxy. We consider a source galaxy at $z = 0.7$ behind a $7 \times 10^{14}\ M_\odot$ foreground cluster (approximately equal in mass to the Coma cluster, \citealt{coma}) at $z \sim 0.3$. We assume the mass profile of the cluster follows a Navarro-Frenk-White profile \citep[NFW,][]{nfw97}. Two-dimensional emission-line kinematics are feasible at this redshift with 8m-class ground-based optical telescopes (e.g., \citealt{contini16}). We assume an impact parameter of 0.3 Mpc in the plane of the lens corresponding to a 1 arcminute separation from the cluster center, which is the field of view in the MUSE wide-field configuration \citep{muse}. This system yields a shear magnitude of $|\gamma| = 0.0589$. We assume that the kinematic major axis of the background galaxy is at a 45$^\circ$ angle to the radial vector to the lens, meaning $\gamma_{\times} = 0.0589$ and $\gamma_+ = 0$. With this shear magnitude in mind, we explore what types of measurements are possible.
\begin{figure}
\centering
\includegraphics[width=.9\linewidth]{samplevf.pdf}
\caption{
Left: A mock intrinsic velocity field to test our KWL shear extraction. This galaxy was generated using the model in Equation \ref{vfmodel} with parameters $v_{max} = 220$ km/s , $i = 45^\circ$, $\phi=2$ rad, $h=2$ kpc, and $v_{sys}=0$ km/s. A shear of $\gamma_{\times}=0.0589$ has been applied to it. The overlaid contours show the error on the velocity measurements in km/s, which are normalized to a central value of 5 km/s. The inverse variance on the velocity measurements is assumed to be proportional to surface brightness, which is modeled as a S\'ersic profile of the same galaxy with $n=1$ and $R_e = 2$ kpc. The velocity inverse variance has been normalized to 0.05 (km/s)$^{-2}$, typical of a MaNGA galaxy. Right: The velocity field from the left panel but smeared and sampled coarsely to emulate an observation at $z \sim 0.7$ with a FWHM of 0.7" on a MUSE-like instrument. See details in Section \ref{sec:data}.}
\label{samplevf}
\end{figure}
We assume that the inverse variances on the velocity measurements are proportional to surface brightness, which we set as an $n=1$ S\'ersic profile. The scale, inclination, and position angle of the S\'ersic profile are defined by the same parameters that define the mock velocity field. normalized to a peak value of 0.05 (km/s)$^{-2}$, corresponding to a velocity error of about $4.5$ km/s at the center of the galaxy. This value was chosen because it was the most common value for galaxies in MaNGA \citep{bundy15} and should capture some of the errors from approximating rotation as a thin disk as well.
For the velocity field, we first generate an idealized toy model of a rotating galaxy using a simple model:
\begin{equation} \label{vfmodel}
v(r,\theta) = \frac{2}{\pi} v_{max} \arctan\left(\frac{r}{h}\right) \sin i\, \cos(\theta-\phi) + v_{sys}.
\end{equation}
\noindent Here $r$ and $\theta$ are polar coordinates of the spectroscopic measurements, $v_{max}$ is the asymptotic rotation speed, $h$ is the characteristic scale radius, $i$ is the inclination, $\phi$ is the kinematic PA, and $v_{sys}$ is the systemic velocity of the galaxy. The positions of the primary kinematic axes, the part of the velocity field model that the shear measurement primarily relies on, are not sensitive to the specific shape of the rotation curve model. We compared results from this model with the more complex empirically-derived Universal Rotation Curve from \cite{persic96} and found negligible difference for the results of this paper, as did \cite{wittman21} in their analysis. We then apply the transformation matrix (Equation \ref{Amatrix}) to the data to shear the velocity field, an effect more easily seen in Figure \ref{example}. For the remainder of this paper, we will use this model to both simulate mock velocity fields and to fit our mock data, allowing the model complete freedom to describe the mock galaxy.
To simulate a real ground-based observation, we first convolve the velocity field with a Moffat point-spread function (PSF) with a full-width half maximum (FWHM) of 0.7" and $\beta = 2.9$ to the velocity field and inverse variance to model atmospheric distortions, a very good night at a mountaintop observatory. We weigh the PSF convolution using the observed surface brightness profile of the galaxy. We then assume a spatial sampling of $0.2 \times 0.2$ arcsec$^2$, the same as the wide field mode for MUSE, corresponding to ${\sim} 1$ kpc spatial resolution at redshift $z = 0.7$. We generate a velocity field with a characteristic scale radius $h = 2$ kpc and apply a shear of $\gamma_{\times} \approx 0.06$ to its spatial coordinates. We assume measurements extending out to ${\sim} 2 R_e$ for an $n=1$ S\'ersic profile also with $R_e = 2$ kpc\footnote{For all other mock galaxies generated in this paper, we will always make the simplification that the characteristic scale radius of the velocity field $h$ and the effective radius of its surface brightness profile $R_e$ are equal.}, meaning we should have a grid of $8 \times 8$ spatial samples across the field of the galaxy. We also apply this same PSF convolution step when fitting this model to data in order to more accurately recover the input parameters.
We also apply random Gaussian errors in the measured velocity according to the assumed inverse variance profile, perturbing the results from the ideal velocity field. The resulting measurements, shown in Figure \ref{samplevf}, qualitatively resemble actual data taken by MUSE (see \citealt{contini16}, Fig. 5). Because we generate our mock data from ideal rotation curve models, we do not expect per-galaxy systematic errors from dynamical shape noise to be present, leaving only statistical errors. We do not apply this step when fitting the model to the data, instead using the ideal unperturbed model.
To find the best velocity field parameters for each mock galaxy, we use the Markov Chain Monte Carlo (MCMC) package \texttt{emcee} \citep{emcee}, feeding in the velocity measurements across the face of a galaxy. We use the same rotation curve model as was used to generate the velocity field, but we add shear by simply applying the inverse of the transformation matrix given in Equation \ref{Amatrix} to the underlying coordinates of the measurements and letting $\gamma_{\times}$ be a free parameter in the fit. Because our velocity field model is sheared directly by the transformation matrix, it is not affected by the approximations we made in Section \ref{sec:radon}. This, however, comes at the price of analytic simplicity of the model.
\begin{figure*}
\centering
\includegraphics[width=.9\linewidth]{vfcorner.pdf}
\caption{Black: The resulting posteriors from applying the Bayesian velocity field model from Equation \ref{vfmodel} with a free shear parameter to the simulated sheared galaxy described in Section \ref{sec:data}. Most posteriors are centered near their true values (shown as blue lines), but the uncertainty is inflated by injected Gaussian errors, blurring from the PSF, and low spatial resolution. This is especially apparent in inclination and shear. Since the magnitude of the introduced errors greatly outweighs the signal from weak lensing, the posterior for $\gamma_{\times}$ has very poor precision and the statistical error dominates the measurement. Red: The performance of the model is greatly improved when image PA is included as described in Section \ref{sec:phot}. With a 3 degree error on the image PA, the posteriors are noticeably more constrained than the kinematics-only model, especially the shear, which has a factor of 5 reduction in spread, and the position angle $\phi$. The degeneracy between $\phi$ and $\gamma_\times$ is also resolved.}
\label{vfcorner}
\end{figure*}
We apply uniform priors to inclination and position angle to allow for random variation in orientation. We use a Gaussian prior on asymptotic velocity with mean 200 km/s and standard deviation 100 km/s and is truncated at 0, parameters that are loosely based off of the MaNGA sample \citep{bundy15}. We allow the rotation scale to vary uniformly up to 4 arcseconds. For $\gamma_{\times}$, we apply a uniform prior bounded at $\pm 0.5$ to allow for a reasonable amount of variation, but not so much as to allow the model to fit any irregularities it sees in the data with unphysical amounts of shear. We use a standard Gaussian likelihood function for comparing the model to the mock data at each iteration. This model produces shear magnitudes and statistical errors that are similar to those measured by \cite{gurri20} when applied to the same data.
With these assumptions, we apply our Bayesian velocity-field model to the simulated data and recover the posteriors shown in black in Figure \ref{vfcorner}. We can see that the posteriors include the true values for most of the model parameters, although only some appear near the median of the posterior distributions. Two primary factors in the model precision are the effects of the PSF and the added Gaussian noise. Inclination relies on the specific shapes of isovelocity contours, but much of their variation with inclination is masked by the PSF. In addition, the width of the shear posterior $\sigma_\gamma$ is hampered by the relatively shallow velocity gradient, limited spatial resolution, and a significant degeneracy with the kinematic position angle. So while we can use this method to extract some of the parameters of the velocity field, we cannot fit our key parameter of interest with much precision. More information on the shear is needed if we want to lower statistical errors enough to produce a successful fit for data of this spatial resolution.
\section{Combining Imaging and Kinematics} \label{sec:phot}
\subsection{Kinematic and Photometric Position Angle Offset}
To better constrain the lensing distortion of the velocity field, we incorporate the image distortion we explored in Section \ref{sec:phottheory}. Comparing the kinematic axis differences derived in Equations \ref{angkin} and \ref{angkinsimp} to the photometric angular difference from Equation \ref{angphot}, we can see (as in Figure \ref{example}) that there is a difference between the angle measured by fitting an isophote and the angle from the velocity field for the major and minor axes:
\begin{multline}\label{dthmaj}
\Delta\theta_{maj} = \gamma_\times\frac{1+q^2}{1-q^2} - \gamma_\times\\
= \gamma_\times\left(\frac{1+q^2}{1-q^2} - 1\right) = \frac{2 \gamma_{\times} q^2}{1-q^2}
\end{multline}
\begin{multline} \label{dthmin}
\Delta\theta_{min} = \left(\gamma_\times\frac{1+q^2}{1-q^2} + \frac{\pi}{2} \right) - \left(\frac{\pi}{2} - \gamma_\times\right) \\
= \gamma_\times\left(\frac{1+q^2}{1-q^2} + 1\right) = \frac{2\gamma_{\times}}{1-q^2}.
\end{multline}
\noindent The minor axis deviates more because the velocity field shearing effect goes in the opposite direction of the apparent rotation of the image. A comparison of the dependencies of $\Delta \theta$ on $\gamma_{\times}$ and $q$ for the major and minor axes is seen in Figure \ref{shearvsdth}. More face-on galaxies have larger apparent rotation angles because their shape is easy to distort. A slight elongation of an elliptical isophote in a given direction will have much more effect on the orientation of the isophote if it starts off as relatively round rather than relatively extended.
\begin{figure}
\plotone{shearvsdth.pdf}
\caption{The angular difference between the imaged and kinematic axis for the major axis (top) and minor axis (bottom) as a function of the shear $\gamma_{\times}$ and the axis ratio $q=b/a$. The magnitude of the rotation of the photometric axis increases more quickly for more face-on galaxies because less elongated ellipses are more easily distorted in an arbitrary direction than more elongated ellipses. The major axis shows slightly less angular difference than the minor axis because the imaged and kinematic axes are being shifted in the same direction, while for the minor axis, they are shifted in opposite directions. This is also why the edge-on $q=0.2$ line for the minor axis still shows some difference.}
\label{shearvsdth}
\end{figure}
We can also see that these angular differences mainly depend on $\gamma_\times$ and $q$. We can then solve for the shear $\gamma_\times$ in terms of the two observables: $\Delta\theta$ (the difference between measured kinematic axis and imaged axis) and the axis ratio $q$. The relationship between $\gamma_\times$ and the deflection of the major and minor axis position angles differ by a factor of $q^2$:
\begin{equation}\label{shearmaj}
\gamma_{\times,maj} \approx \frac{\Delta\theta_{maj}(1-q^2)}{2q^2}
\end{equation}
\begin{equation}\label{shearmin}
\gamma_{\times,min} \approx \frac{\Delta \theta_{min}(1-q^2)}{2}.
\end{equation}
\noindent The dependencies we recover here agree well with \cite{wittman21} and we will use them in the next section to derive an improved lensing inference based on a combination of imaging and kinematic measurements. By combining the imaging measurements of position angle and axis ratio with velocity measurements of the \textit{kinematic} position angle, we can gain access to a constraint on the gravitational shear induced on the galaxy.
\subsection{KWL Models with Imaging Information}
We rerun the previously described Bayesian velocity field fitting models but with added constraints on the difference between the kinematic and image PAs. We feed the model a mock image position angle, setting the value by perturbing the expected value derived using Equation \ref{dthmaj} with a Gaussian error based on the assumed level of photometric uncertainty. Commonly-used photometric codes like \textsc{Galfit} \citep{galfit} often drastically under-report the magnitude of their PA errors. Accurate error accounting must be done by comparing intrinsic PAs in simulated galaxies with the values recovered by photometric fitting. \cite{haussler07} suggests that photometric PA measurements have average errors between 1 and 6 degrees depending on image depth, so we will largely restrict our analyses to that range.
We then allow the model to fit the angular difference between the kinematic PA and the perturbed image PA with its shear parameter, incorporating the result as a new Gaussian term in the likelihood function. With this added information, we see a significant reduction in the width of the posteriors $\sigma_\gamma$, as shown by the red posteriors in Figure \ref{vfcorner}. Adding in the image PA allows for the degeneracy between the kinematic PA and the shear strength to be broken much more effectively, resulting in a factor of 5 reduction in $\sigma_\gamma$ for an image PA with an error of 3 degrees. This vast reduction in statistical errors has large implications for regimes where the measurement error would be dominated by statistical error.
These improvements persist even for larger PA errors, as shown in Figure \ref{photerrs}. Within the range of expected image PA errors (shown as the shaded blue region), we see a reduction in $\sigma_\gamma$ by a factor of 2--6, meaning that a KWL technique utilizing imaging shape information is more sensitive. We also see that for large image PA errors, $\sigma_\gamma$ approaches its value from before imaging information was added. As expected, as uncertainty in the image PA measurement increases, the statistical error in the measured shear tends towards what is obtained in a fit without image PA constraints.
\begin{figure}
\centering
\includegraphics[width=.9\linewidth]{photerrs.pdf}
\caption{
The gains in precision seen after adding information on image position angle to our Bayesian KWL model for a mock galaxy like the one in Figure \ref{samplevf}, including the applied shear of $\gamma_\times \approx 0.06$. The model that fits both the kinematic and imaging distortion (blue line) has an error on the shear posterior $\sigma_\gamma$ that is 3-6 times smaller than the model with only kinematics (dashed line) was able to achieve within the region of expected errors on image PA in real data (blue shaded area). If borne out in real observations, this could lead to significant decreases in necessary sample sizes and exposure times to obtain a given lensing S/N. For very large errors in image PA, $\sigma_\gamma$ tends back towards the value obtained with no imaging information.}
\label{photerrs}
\end{figure}
In Figure \ref{psfgrid30}, we explore differing combinations of image PA error and velocity error for a relatively face-on source galaxy (30 degree inclination). Larger errors in image PA or velocity measurement may come from shallower exposures, poor seeing or angular resolution, or irregularities in the galaxy. At this low inclination, improvements in image PA precision improve the KWL shear measurements as much or more than improvements to the velocity precision, reducing the need for higher S/N or higher resolution spectroscopy. Even with moderate image PA uncertainties on the order of several degrees, the precision of KWL measurements can reach levels comparable to those of velocity-field-only fits relying on kinematic measurements at a fraction of the velocity error.
\begin{figure}
\centering
\includegraphics[width=.9\linewidth]{psfphotnoise_30.pdf}
\caption{The change in the error on the shear measurement $\sigma_\gamma$ for a mock velocity field with simulated seeing with different amounts of error on the image PA and different central velocity errors. The input shear is $\gamma_\times \approx 0.06$ and the model velocity field is identical to those used in previous figures except that it has a $30^\circ$ inclination with respect to the observer. The blue shaded region represents the expected range for image PA errors in real observations. Smaller errors on the image PA lead to higher precision in shear, as is expected. For larger image PA errors, the precision approaches the value derived from just the velocity field, but the value it levels off at depends on the errors in the velocity field itself and the inclination of the galaxy. For image PA errors expected in real data, even poor-quality velocity fields can match the precision reached by very good velocity field-only KWL measurements, presenting opportunities for exposure time savings.}
\label{psfgrid30}
\end{figure}
These gains are much larger for lower-inclination galaxies since their imaged axes are much more distorted, allowing them to contribute more to the shear measurement. Higher-inclination galaxies, as shown in Figure \ref{psfgrid4560}, gain less from their photometric information because the magnitude of the angular deflection is greatly reduced, forcing the shear measurement to rely more on kinematic distortion. The distribution of inclinations within a population of randomly-oriented disks is uniform over $\sin i$ \citep{law09}, so the inclination distribution of a random sample from a survey will be weighted towards the high end, with a mean inclination angle of 1 radian (about $57^\circ$) or a mean axis ratio of $q \approx 0.54$. So unless a survey sample is deliberately designed to sample low-inclination galaxies, most galaxies will see improvements on par with Figure \ref{psfgrid4560} rather than \ref{psfgrid30}.
Still, adding imaging information to KWL measurements for these higher-inclination galaxies that would make up the majority of a random sample would cut required exposure times dramatically for spectroscopic observations. If we assume that S/N scales with the square root of exposure time, then even a factor of two improvement in S/N from including the image PA can cut exposure time by a factor of four.
\subsubsection{Comparison to Previous Work} \label{sec:gurri}
As mentioned in Section \ref{sec:data}, our kinematics-only model delivered similar statistical errors as that of \cite{gurri20} when tested on their data, so to quantify the benefits of our kinematics + imaging technique on presently-available data sets, we again benchmark against their approach by modifying our mock data to better match the \cite{gurri20} data. We move the source galaxy to $z=0.15$ and the lens to $z=0.03$, increase the FWHM of the PSF to 1.5", and increase the spatial element size to 0.5" to roughly mimic their data set. We add mock image PA measurements, as we have done for the rest of our mock data, and we assume an error on the image PA of 3 degrees, which is in the middle of the expected error range.
Applying our kinematics + imaging algorithm to these simulated galaxies results in significantly improved statistical uncertainties. Our estimate on the resulting $\sigma_\gamma$ values are lower by a factor of ${\sim}6$ on a per-galaxy basis, consistent with the results from higher redshift mock data in Figure \ref{photerrs}.
This result demonstrates that even for existing data sets at lower redshifts, this technique can increase lensing precision.
However, our inclusion of imaging data does not address the dynamical shape noise term.
In order to account for this error, \cite{gurri20} estimate the magnitude of the error using a different sample of unsheared galaxies, and find an amplitude that is similar to the statistical error on a per-galaxy basis. This dynamical shape noise term is added in quadrature to determine the final error for each galaxy in their sample. As a result, our estimated factor of 6 improvement in the statistical error, when applied to the full sample in \citealt{gurri20}, only increases their overall 2.5$\sigma$ shear detection to 3$\sigma$.
\begin{figure}
\centering
\includegraphics[width=.9\linewidth]{psfphotnoise_4560.pdf}
\caption{The same as Figure \ref{psfgrid30} but for a $45^\circ$ and $60^\circ$ galaxies. Higher-inclination galaxies have a much smaller image distortion than lower-inclination galaxies, meaning the model cannot rely on this information as much when constraining the shear. This creates a much larger stratification in velocity error than for lower-inclination galaxies. Overall, $\sigma_\gamma$ values are larger, particularly in the $60^\circ$ galaxy, but improvements to the image PA observation have little effect.}
\label{psfgrid4560}
\end{figure}
\subsection{Survey Design Considerations} \label{sec:survey}
While statistical error is small relative to systematic error for low redshift samples, the same is not true at higher redshifts. Resolved spectroscopic measurements become more difficult at higher redshifts due to lower spatial resolution and surface brightness, resulting in higher statistical errors. So the gains made by our imaging + kinematics model will be especially salient at increased redshifts, where there are the added benefits of greater lensing magnitudes (thanks to more favorable lensing kernels) and higher on-sky source densities, allowing for easier collection of larger sample sizes.
If we observe a $1 \times 1$ arcmin$^2$ field near a galaxy cluster (e.g. using MUSE), we will catch tens of galaxies with velocity fields that are well-defined enough to perform our KWL analysis. A multiplexed fiber IFU instrument like the proposed FOBOS \citep{fobos} would be able to patrol a significantly wider field and collect velocity data from only the most promising galaxies, enabling a larger sample in less overall exposure time. We can stack the lensing information from the multiple galaxies to obtain higher precision and reduce the random errors introduced by noise or galaxy irregularities either by a weighted average of shear results (as was done in \citealt{gurri20}) or a hierarchical Bayesian model that simultaneously fits all galaxies in a given spatial bin. We expect these individual galaxy ``dynamical shape noise" errors to average out to zero in a large sample because they will be randomly and symmetrically distributed \citep{gurri20}.
Figures \ref{psfgrid30} and \ref{psfgrid4560} show that, with attainable observational errors and low inclinations, it is possible to obtain a $\sigma_\gamma$ value comparable in magnitude to the applied shear, resulting in a ${\sim}1\sigma$ KWL measurement per galaxy for the input shear of $\gamma_\times \approx 0.06$. However, if we stack multiple such measurements within a radial bin defined by a selection of foreground lenses
and assume that the source measurements are statistically independent samples of the ensemble mass density profile of the lenses, then the error on this galaxy--galaxy shear measurement should scale as $n^{-1/2}$. So for a sample of only $n=9$ galaxies in a spatial bin, $\sigma_\gamma$ should be lowered by a factor of 3, raising a 1$\sigma$ detection to 3$\sigma$. This scaling should apply to the systematic error as well since we expect individual galaxy errors to average to zero over a large sample.
\subsubsection{Source Redshift}
Depending on the instrumentation, if we want to build a larger sample size, higher source density on-sky may be beneficial if the number of independent pointings can be minimized. Source density increases with redshift, as does the magnitude of lensing effects according to Equations \ref{sigcrit} and \ref{shear}, but there is a trade-off because surface brightness and physical size decrease with redshift. To explore this trade-off, we run our KWL model that incorporates imaging information on the same mock galaxy placed at a range of redshifts. We ignore complications from galaxy evolution like mass/radius growth or dynamical changes. We simulate the observations of different inclination source galaxies at the varying distances as if they are behind a lens at $z=0.3$. We hold the rest of the observational parameters constant, on-sky spatial resolution, image PA error, central velocity error, and PSF width (unless specified). Maintaining constant errors would require deeper exposures for both imaging and spectroscopy, which we do not account for in this paper.
\begin{figure}
\centering
\includegraphics[width=.9\linewidth]{fig8alt10.pdf}
\caption{Top: The KWL S/N per galaxy for source galaxies at varying inclinations with different redshifts behind a massive halo at $z=0.3$. We maintain constant on-sky spatial resolution, imaging and velocity errors, and PSF width, ignoring the effects of surface brightness dimming and galaxy evolution. Error bars are based on the random variation from running 10 identical velocity fields with distinct randomized errors. Middle: The $45^\circ$ inclination galaxy from the top panel but with FWHM of the PSF varied to simulate different observing conditions, still holding on-sky instrumental resolution constant. Realistic observing conditions provide measurements that are a factor of 2--3 lower in S/N compared to an ideal observation without PSF smearing. Improving the PSF by a factor of 2 with something like a ground-layer adaptive optics system to 0.35" leads to S/N gains of about 50\%. Bottom: The overall magnitude of the lensing shear, shown in the red dashed line, increases with redshift, making the effects of KWL more noticeable. The fraction of a square FWHM covered by a square kpc in the on-sky plane of the source galaxy, shown in blue solid line, indicates that the PSF becomes very large compared to the spatial scales of the source galaxy, worsening the quality of the fit. S/N largely tracks with shear magnitude.}
\label{snz_zl3}
\end{figure}
Although closer source galaxies have smaller $\sigma_\gamma$ uncertainties because of their greater number of spatial resolution elements, Figure \ref{snz_zl3} shows that they deliver a poorer S/N per galaxy because they are simply not sheared as much as more distant galaxies.
The top panel shows a high degree of S/N stratification in inclination because the change in image PA is much greater for lower inclinations. We find that there is modest gain in S/N for all inclinations with redshift before leveling off around $z\sim0.7-1$, the redshift range assumed in this paper, before keeping a relatively constant value at higher $z$. This mirrors the trend seen in the overall shear magnitude seen as the red dashed line in the bottom panel of Figure \ref{snz_zl3}.
Another factor that influences precision is the FWHM of the spatial PSF of the resolved kinematic field. The blue line in the bottom panel of \ref{snz_zl3} shows that as the source galaxy redshift increases, the size of the galaxy on-sky becomes smaller relative to the size of the PSF, here represented as the ratio of the square FWHM compared to the size of a square kpc in the plane of the galaxy.
Reductions to the FWHM result in higher surface brightness in many cases and more independence of the spatial resolution elements of an IFU observation, which would improve the precision of the velocity field fit even if the spatial sampling is not changed.
These types of gains could be obtained in the real world using a ground-layer adaptive optics system \citep[e.g.][]{chun18, hartke20}, which would lead to an improvement in seeing by a factor of 1.5 to 2 over a several arcminute-wide field. In the middle panel of Figure \ref{snz_zl3}, we vary the FWHM to simulate different observing conditions, from an ideal observation without seeing to a FWHM of 2". We find modest gains in S/N with better seeing, especially at moderate and higher redshifts. However, even fits with poor seeing are able to locate the kinematic axes with acceptable precision, so the effects of better seeing are limited. Still, the increase in surface brightness is valuable for driving down exposure time.
\subsubsection{Lens Redshift}
\begin{figure}
\centering
\includegraphics[width=.9\linewidth]{snz_zs1.pdf}
\caption{The same as Figure \ref{snz_zl3} but holding the source galaxy fixed at $z=1$ while moving the lens in front of it. The S/N again largely follows the shear magnitude (bottom), which peaks at $z \sim 0.25-0.3$ because that is the configuration of angular diameter distances that minimizes the critical radius (Equation \ref{sigcrit}). FWHM$^2$ per kpc$^2$ is not shown because it is constant due to the stationary source galaxy.}
\label{snz_zs1}
\end{figure}
To test KWL's ability to probe lenses at various redshifts, we also simulate systems where the source is held at a constant $z=1$ for varying redshift lenses (Figure \ref{snz_zs1}). We find similar dependencies on inclination and FWHM as in Figure \ref{snz_zl3}. However, since shear magnitude scales with the angular diameter distances to the source and the lens (Equation \ref{sigcrit}), shear magnitude peaks at $z\sim0.3$, falling off sharply from there. In fact, due to how angular diameter distance is defined, for source galaxies further than $z\sim0.7$, shear magnitude will always be largest for lenses at $0.2 < z < 0.3$. So KWL, like any other lensing technique, will be most sensitive to lenses at these moderate redshifts.
\subsubsection{General Strategy}
An ideal strategy for a KWL survey would likely be to target low-inclination galaxies that benefit more from the gains imaging can provide. They should be regularly-rotating blue spiral galaxies that are likely to have a high emission line flux for ease of kinematic measurements.
The targets should have regular, symmetric isophotes and imaging sufficient to measure image position angles with 1--3 degree uncertainties.
If the source galaxy population reaches $z \sim 1$, the ideal lens sample would have a redshift range $z = 0.2$--0.7 to ensure shear magnitudes near the peak of the lensing kernel.
Ideally, the major-axis position angles of selected source galaxies should be misaligned with respect to the on-sky direction to the lens center. A $45^\circ$ offset maximizes the cross-term in the shear.
A full optimization of a KWL survey design would take into account the lens and source redshift distributions, the intrinsic variation in galaxy shape, systematic errors in individual galaxies due to dynamical shape noise, and observational constraints such as seeing and exposure time as it relates to brightness dimming, distance, and size.
\section{Conclusion} \label{sec:conclusions}
We have demonstrated a new formalism for combining imaging and kinematic information that significantly improves the S/N of kinematic weak-lensing observations. Our Bayesian models fit both the kinematic distortion and the photometric offset caused by lensing shear in a given galaxy, leading to decreases in per-galaxy statistical error by a factor of 2 to 6 compared to kinematics alone.
If borne out in real observations, this approach opens the door for future studies to more effectively utilize kinematics-based lensing observations of lens samples at moderate redshifts.
Even with existing instruments, the methodology appears promising for enabling probes of galaxy cluster halo profiles with greater spatial resolution and S/N
while mitigating many of the systematics that affect conventional lensing measurements, like shape noise and intrinsic alignments. Ultimately, this will allow
for individual, total mass measurements in the weak-lensing regime for a greater number of clusters.
A purpose-designed survey on a new or existing instrument could provide the observations necessary for a successful KWL measurement. Further study is needed to determine the appropriate redshift range to balance shear magnitude and source density with spatial resolution and surface brightness, while more detailed galaxy rotation models are needed to quantify the systematic effects of kinematic irregularities on KWL measurements.
\acknowledgments
The authors would like to acknowledge the anonymous reviewer for their helpful comments in improving this paper. We woudl also like to acknowledge Eric Huff and Tim Eifler for their early guidance on the theory of kinematic weak lensing and Kevin McKinnon for early statistical guidance.
|
\section{Introduction}
With the rise of the Internet, connections among people has become easier than
ever; so has been for the availability of information and its accessibility. As
such, the Internet is also the source of unprecedented collective phenomena,
some of which, however, cast shadows on our contemporary
society^^>\cite{Bak-Colemane2025764118}.
Indeed, the dissemination of heavily biased, or worse, downright false
information, once relatively moderate in size, and limited possibly to the
class of hoaxes and scams, exploded in the last decade, creating the broader
category of^^>\emph{fake news}. The urgent need for models that can describe
the increasing spread of fake news has been highlighted by the current COVID-19
pandemic. Governments in many countries have found themselves in enormous
difficulty because of the slowdown in vaccination campaigns due to the spread
of false information and the inability of individuals to discern the
authenticity of such information^^>\cite{fake1,fake2}.
Let us first briefly recall some of the main challenges in modeling fake news.
First of all, one of the priorities is to introduce a definition of
fake news with a consensus that is wide enough to make research works
relatable. In this direction, one of the most accepted (though not universally
so) traits for fake news to be labeled so is
\emph{purpose}: fake news is
intentionally false
news^^>\cite{Zhang2020,Shu2017FakeND,Allcott,Gelfert201884}. The concept of
purpose seems to be useful when
differentiating between theories that are focused on the content rather than on
the conveyor^^>\cite{Zhang2020}. In this sense, a piece of information that is
accidentally false (e.g., by inaccuracy) is substantially different, both
semantically and stylistically, from a maliciously fabricated one.
Next, the challenge is to detect fake news. The majority of recent lines of
research in the direction of automatic detection moves toward the aid of
\emph{big data} and artificial intelligence
tools^^>\cite{Shin2018278,Conroy20151,Ruchansky2017797,Vargo20182028}. An
alternative strategy is instead \emph{component-based} and
focuses on the analysis of the multiple parts involved in the diffusion of the
fake news, that is, both on the side of the creator and on the side of the
user, but also on the linguistics and semantics of the actual content and on
its style, and finally on the social context of the information
(see^^>\cite{Zhang2020} and references therein).
Different lines are possible, though. Information theory, for instance, has
been used to model fake news: in^^>\cite{fninfoth}, fake news are defined as
time series with an inherent bias, that is, its expectation is nonzero,
involved in a stochastic process of which the user tries to judge the
likelihood of the truth, together with noise.
Finally, epidemiological theory has been proving for long to be fertile ground
for modeling of fake news^^>\cite{Daley19641118,daley_gani_1999},
especially in the somewhat
broader category of rumor-spreading dynamics.
The analogy between rumors and epidemics has often proved
fruitful: Daley and Kendall^^>\cite{Daley19641118} took inspiration by the
classical works of Kermack and Mckendrick^^>\cite{KermackACT,Hethcote2000TheMO}
to propose a
SIR-like model involving ignorant, spreader and stifler agents who played the
role of the susceptible, infectious and recovered ones in^^>\cite{KermackACT}.
Since the seminal paper^^>\cite{Daley19641118}, rumor-spreading dynamics has
taken ideas from epidemiological models to improve their prediction accuracy.
Recently, networks theory
delved in this direction, too^^>\cite{Cheng,networks,zhao2019fake,trammell}.
Substantial research has merged networks and
epidemiology through for instance classical compartmental
models like the SIR (both epidemiological
in^^>\cite{PhysRevLett.86.3200} and
rumor-oriented^^>\cite{PIQUEIRA2020123406}),
the SIS^^>\cite{PhysRevLett.89.108701} and the
SIRS^^>\cite{PhysRevLett.86.2909,IEEE}. Moreover, more sophisticated
epidemiological models have been developed, like the SEIZ
model^^>\cite{BETTENCOURT2006513} describing the evolution in time of the
compartments of susceptible, exposed, infectious and skeptic agents, which has
been adapted to the analysis of
fake news dissemination (see,
e.g.,^^>\cite{maleki2021using,rumorsontwitter}). In
particular, the tendency seems to be to define the skeptic
agents like the ones who are aware of the information but do not actively
spread it^^>\cite{rumorsontwitter}. In a symmetric fashion, spreaders need not
to believe a piece of
information to be able to spread it (this is especially useful when thinking
that bots are often encountered in social networks^^>\cite{datarepository},
both for legitimate and malicious purposes). This descriptions are also
sensible in terms of matching the model with data available.
In this paper we follow this pathway: borrowing ideas from kinetic
theory^^>\cite{PhysRevE.102.022303,intermultiagent}, we combine a classical
compartmental approach inspired by
epidemiology^^>\cite{Hethcote2000TheMO,KermackACT} with a kinetic description
of the effects of competence^^>\cite{PARESCHI2017201,wealthPareschiToscani}.
We refer also to the recent work^^>\cite{rey} concerning evolutionary models
for knowledge.
In fact, the first wave of initiatives addressing fake news focused on news
production by trying to limit citizen exposure to fake news. This can be done
by fact-checking, labeling stories as fake, and eliminating them before they
spread. Unfortunately, this strategy has already been proven not to work, it is
indeed unrealistic to expect that only high quality, reliable information will
survive. As a result, governments, international organizations, and social
media companies have turned their attention to digital news consumers, and
particularly children and young adults. From national campaigns in several
countries to the OECD, there is a wave of action to develop new curricula,
online learning tools, and resources that foster the ability to \lq\lq spot
fake news\rq\rq^^>\cite{PIAAC19}.
It is therefore of paramount importance to build models capable of describing
the interplay between the dissemination of fake news and the creation of
competence among the population. To this end, the approach we have followed in
this paper falls within the recent socio-economic modeling described by kinetic
equations (see^^>\cite{intermultiagent} for a recent monograph on the subject).
More precisely, we adapted the competence model introduced
in^^>\cite{PARESCHI2017201,wealthPareschiToscani} to a compartmental model
describing fake news dissemination. Such a model allows not only to introduce
competence as a static feature of the dynamics but as an evolutionary component
both taking into account learning by interactions between agents and possible
interventions aimed at educating individuals in the ability to identify fake
news. Furthermore, in our modeling approach agents may have memory of fake news
and as such be permanently immune to it once it has been detected, or fake news
may not have any inherent peculiarities that would make it memorable enough for
the population to immunize themselves against it in the future. The approach
can be easily adapted to other compartmental models present in the literature,
like the ones previously discussed^^>\cite{BETTENCOURT2006513, maleki2021using,
PIQUEIRA2020123406}.
The rest of the manuscript is organized as follows. In Section^^>\ref{sec:2} we
introduce the structured kinetic model describing the spread of fake news in
presence of different competence levels among individuals. The main properties
of the resulting kinetic models are also analyzed. Next, Section^^>\ref{sec:3}
is devoted to study the Fokker-Planck approximation of the kinetic model and to
derive the corresponding stationary states in terms of competence. Several
numerical results are then presented in Section^^>\ref{sec:4} that illustrate
the theoretical findings and the capability of the model to describe transition
effects in the spread of fake news due to the interaction between
epidemiological and competence parameters. Some concluding remarks are reported
in the last Section together with details on the
theoretical results and the numerical methods in two separate appendices.
\section{Fake news spreading in a socially structured population}\label{sec:2}
In this section, we introduce a structured model for the dissemination of fake
news in presence of different levels of skills among individuals in detecting
the actual veracity of information, by combining a compartmental model in
epidemiology and rumor-spreading
analysis^^>\cite{Hethcote2000TheMO, daley_gani_1999} with the kinetic
model of competence evolution proposed in^^>\cite{PARESCHI2017201}.
We consider a population of individuals divided into four classes. The
oblivious
ones, still not aware of the news; the reflecting ones, who are aware of the
news and
are evaluating how to act; the spreader ones, who
actively disseminate the news and the silent ones, who have recognized the fake
news and do not contribute to its spread.
Terminology, when describing this compartmental models, is not fully
established; however, the dominant one, inspired by epidemiology, refers to the
definitions provided by
Daley^^>\cite{daley_gani_1999} of a population composed of ignorant, spreader
and stifler individuals. The class of reflecting agents can be referred to as
a group that has a time-delay before taking a decision and enter an active
compartment^^>\cite{BETTENCOURT2006513, maleki2021using}.
Notation, i.e., the choice of letters to represent the compartments, is even
more scattered and somewhat confusing. In
Table^^>\ref{tab:notations} for readers' convenience we have summarized some of
the different possible
choices of letters and terminology found in literature.
Given the widespread use of epidemiological models compared to fake news
models, in order to make the analogies easier to understand, we chose to align
with notations conventionally used in epidemiology. Therefore, in
the rest of the paper we will describe the population in terms of susceptible
agents (S), who are the oblivious ones; exposed agents (E), who are in the
time-delay compartment after exposure and before shifting into an active class;
infectious agents (I), who are the spreader ones and finally removed agents (R)
who are aware of the news but not actively engaging in its spread.
Note that this subdivision of the population does not take into account actual
beliefs of agents about the truth of the news, so that removed agents, for
instance, need not be actually skeptic, nor the spreaders need to actually
believe the news. To simplify the mathematical treatment, as in the original
works by Daley and Kendall^^>\cite{Daley19641118,daley_gani_1999}, we ignored
the possible \lq active\rq\ effects of the population of removed individuals by
interacting with other compartments and producing immunization among
susceptible (the role of skeptic individuals
in^^>\cite{BETTENCOURT2006513,maleki2021using}) and remission among spreaders
(the role of stiflers in^^>\cite{PIQUEIRA2020123406}). Of course, the model
easily generalizes to include these additional dynamics.
The main novelty in our approach is to consider an additional structure on the
population based on the concept
of competence of the agents, here understood as the ability to assess and
evaluate information.
\begin{table}
\begin{center}
\begin{tabular}{lllll}
\toprule
SEIR (this paper)
& DK^^>\cite{Daley19641118,daley_gani_1999} & ISR^^>\cite{PIQUEIRA2020123406} &
SEIZR^^>\cite{BETTENCOURT2006513}
& SEIZ^^>\cite{maleki2021using}\\
\toprule
\multicolumn{5}{c}{\bf Category name}\\
\midrule
Susceptible
& Ignorant & Ignorant & Susceptible
& Susceptible\\
Exposed
& - & - & Idea incubator
& Exposed\\
Infectious
& Spreader & Spreader & Idea adopter
& Infectious\\
Removed
& Stifler & Stifler & Skeptic/Recovered
& Skeptic\\
\midrule
\multicolumn{5}{c}{\bf Variable notation}\\
\midrule
S & X & I & S & S\\
E & - & - & E & E\\
I & Y & S & I & I\\
R & Z & R & Z/R & Z\\
\bottomrule
\end{tabular}
\end{center}
\caption{Different compartments and notations for some of the models found in
literature.}
\label{tab:notations}
\end{table}
Let us suppose that agents in the system are completely characterized by their
competence \mbox{$x \in X \contenutoin \R^+$}, measured in a suitable unit. We
denote
by \mbox{$f_S = f_S(x, t)$}, \mbox{$f_E = f_E(x,t)$}, \mbox{$f_I = f_I(x,t)$},
\mbox{$f_R = f_R(x,t)$}, the competence
distribution at time^^>\mbox{$t > 0$} of susceptible, exposed, infectious and
removed individuals, respectively. Aside from natality or
mortality concerns (i.e., the social network is a closed system---nobody enters
or leaves it during the diffusion of the fake news, which is a common
assumption, based on the average lifespan of fake news) we therefore have:
\[
\int_X \bigl(f_S(x,t) + f_E(x,t) + f_I(x,t) + f_R(x,t) \bigr)\, \de x = 1,\quad
t > 0,
\]
which implies that we will refer to
\begin{align*}
S(t) &= \int_X f_S(x,t)\, \de x, &
E(t) &= \int_X f_E(x,t)\, \de x,\\
I(t) &= \int_X f_I(x,t)\, \de x, &
R(t) &= \int_X f_R(x,t)\, \de x
\end{align*}
as the fractions of the population that are susceptible, exposed, infected, or
recovered
respectively. We also denote the relative mean competences as
\begin{align*}
m_S(t) &= \int_X x f_S(x,t)\, \de x, &
m_E(t) &= \int_X x f_E(x,t)\, \de x, \\
m_I(t) &= \int_X x f_I(x,t)\, \de x, &
m_R(t) &= \int_X x f_R(x,t)\, \de x.
\end{align*}
\subsection{A SEIR model describing fake news dynamics}
The fake news dynamics proceeds as follows: a susceptible agent gets to know it
by a spreader. At this point, the now-exposed agent evaluates the piece of
information---the reflecting, or delay, stage---and decide whether to
share it with other individuals (and turning into a spreader themselves) or to
keep silent, removing themselves by the dissemination process.
\begin{figure}
\centering
\includegraphics[scale=1.5]{SEIR-diagram.pdf}
\caption{SEIR diagram with transition rates.}
\label{fig:SEIRdiagram}
\end{figure}
When the dynamic is independent from the knowledge of individuals, the model
can be expressed by
the following system of ODEs
\begin{equation}
\left\lbrace
\begin{aligned}
\der St &= -\beta SI + (1 - \alpha)\gamma I\\
\der Et &= \beta SI - \delta E \\
\der It &= (1-\eta)\delta E - \gamma I,\\
\der Rt &= \eta\delta E + \alpha\gamma I,
\end{aligned}
\right.
\label{eq:seirnormale}
\end{equation}
with $S+E+I+R=1$ and where $\beta$ is the contact rate between the class of the
susceptible and the
class of infectious, $\delta$ is the rate at which agents make their decision
about spreading the news or not, $1 - \eta$ is the portion of agents who become
infectious and $\gamma$ is the rate at which spreaders remove themselves from
the compartment, due, e.g., to loss of interest in sharing the news or
forgetfulness. Finally, $\alpha$ is related to the specificity of the fake news
and the probability of individulas to remember it. A probability of^^>$0$
means that the fake news has not
any inherent peculiarity (e.g., in terms of content, structure, style, \ldots)
that can make it memorable enough for the population to \lq immunize\rq\
against it in the future, while a probability of^^>$1$ allows for the agents to
have the full ability to not fall for that fake news a second time. The various
parameters have been summarized in Table \ref{tab:parameters}. The diagram of
the SEIR model^^>\eqref{eq:seirnormale} is shown in
Figure^^>\ref{fig:SEIRdiagram}.
\begin{table}
\begin{center}
\begin{tabular}{cl}
\toprule
Parameter & Definition\\
\toprule
$\beta$ & contact rate between susceptible and infected individuals\\
$1/\delta$ & average decision time on whether or not to spread fake news\\
$\eta$ & probability of deciding not to spread fake news \\
$1/\gamma$ & average duration of a fake news\\
$\alpha$ & probability of remembering fake news\\
\bottomrule
\end{tabular}
\end{center}
\caption{Parameters definition in the SEIR model \eqref{eq:seirnormale}.}
\label{tab:parameters}
\end{table}
It is straightforward to notice that when^^>$\alpha$ and $\eta$ are zero,
system^^>\eqref{eq:seirnormale} specializes in a classic SEIS epidemiological
model. This is consistent with treating the dissemination of non-specific fake
news in a population as the spread of a disease with multiple strains, for
which a durable immunization is never attained. In this case
system^^>\eqref{eq:seirnormale} has two equilibrium states: a
disease-free equilibrium state^^>$(1,0,0)$ and an endemic equilibrium
state^^>\mbox{$\tilde P = (\tilde S, \tilde E, \tilde I)$} where
\begin{equation}\label{eq:endemicequilibria}
\tilde S = \frac1{R_0}, \qquad
\tilde E = \frac{\gamma}{\gamma + \delta}\biggl(1 - \frac1{R_0}\biggr),\qquad
\tilde I = \frac{\delta}{\gamma + \delta}\biggl(1 - \frac1{R_0}\biggr),
\end{equation}
and $R_0 = \beta/\gamma$ is the basic reproduction number. It is
known^^>\cite{korobeinikov} that if
\mbox{$R_0 > 1$} the endemic equilibrium state^^>$\tilde P$ of
system^^>\eqref{eq:seirnormale} is globally asymptotically stable.
If instead $\alpha > 0$ or $\eta > 0$, there also is the possibility to
permanently immunize against fake news with those traits; moreover, both
infectious and exposed agents eventually vanish, leaving only the
susceptible and removed compartments populated. In the case of maximum
specificity of the fake news, i.e., \mbox{$\alpha = 1$}, the stationary
equilibrium state has the form
\begin{equation}\label{eq:orsmequilibria}
S(t) \to S^\infty, \quad E(t) \to 0, \quad I(t) \to 0, \quad R(t) \to R^\infty
= 1-S^\infty,
\end{equation}
where $S^\infty$ is solution of the nonlinear equation
\begin{equation}
\log\frac{S_0}{S^\infty} = \frac\beta\gamma(1 - \eta)(1 - S^\infty),
\end{equation}
in which $S_0$ is the initial datum $S(t = 0)$.
We refer to^^>\cite{BETTENCOURT2006513,maleki2021using,PIQUEIRA2020123406} for
the inclusion of additional interaction dynamics, taking into account
counter-information effects due to the removed population interacting against
susceptible and infectious, and the relative analysis of the resulting
equilibrium states.
\subsection{The interplay with competence and learning}
In the following, we combine the evolution of the densities according to the
SEIR model \eqref{eq:seirnormale} with the competence dynamics proposed
in^^>\cite{PARESCHI2017201}.
We refer to the degree of competence
that an individual can gain or loose in a single interaction from
the background as \mbox{$z \in \R^+$}; in what follows we denote by^^>$C(z)$
the bounded-mean distribution of $z$, satisfying
\[
\int_{\R^+} C(z)\, \de z = 1, \quad \int_{\R^+} zC(z)\, \de z = m_B.
\]
Assuming a susceptible agent has a competence level $x$ and interacts with
another one belonging to the various compartments in the population and having
a competence level^^>$x\*$, their levels after the interaction will be given by
\begin{equation}
\begin{aligned}
\left\lbrace
\begin{aligned}
x' &= (1 - \lambda_S(x))x + \lambda_{CJ}(x)x\* + \lambda_{BS}(x)z
+ \kappa_{SJ} x\\
x\*' &= (1 - \lambda_J(x\*))x\* + \lambda_{CS}(x\*)x + \lambda_{BJ}(x\*)z
+ \tilde\kappa_{SJ} x\*,
\end{aligned}
\right.
&& J \in \ins{S, E, I, R}
\end{aligned}
\label{eq:competencebinaryS}
\end{equation}
where $\lambda_S(\cdot)$ and $\lambda_{BS}(\cdot)$ quantify the
amount of competence lost by susceptible individuals by the natural process of
forgetfulness and the amount gained by susceptible individuals from the
background, respectively. $\lambda_{CJ}$, instead, models the competence
gained through the interaction with members of the
class^^>$J$, with \mbox{$J \in \ins{S, E, I, R}$}; a possible choice
for^^>$\lambda_{CJ}(x)$ is^^>\mbox{$\lambda_{CJ}(x) = \lambda_{CJ}\chi(x \ge
\bar x)$}, where $\chi(\cdot)$ is the characteristic function and \mbox{$\bar x
\in X$} a minimum level of competence required to the agents for increasing
their own skills by interactions. Finally, $\kappa_{SJ}$ and
$\tilde\kappa_{SJ}$ are independent and identically distributed zero-mean
random variables with the same variance^^>$\sigma(t)$ to consider the
non-deterministic nature of the competence acquisition process.
The binary interactions involving the exposed agents can be similarly defined
\begin{equation}
\begin{aligned}
\left\lbrace
\begin{aligned}
x' &= (1 - \lambda_E(x))x + \lambda_{CJ}(x)x\* + \lambda_{BE}(x)z
+ \kappa_{EJ} x\\
x\*' &= (1 - \lambda_J(x\*))x\* + \lambda_{CE}(x\*)x + \lambda_{BJ}(x\*)z
+ \tilde\kappa_{EJ} x\*,
\end{aligned}
\right.
&& J \in \ins{S, E, I, R}
\end{aligned}
\label{eq:competencebinaryE}
\end{equation}
the same holds for the interactions concerning the infectious fraction of
the population
\begin{equation}
\begin{aligned}
\left\lbrace
\begin{aligned}
x' &= (1 - \lambda_I(x))x + \lambda_{CJ}(x)x\* + \lambda_{BI}(x)z
+ \kappa_{IJ} x\\
x\*' &= (1 - \lambda_J(x\*))x\* + \lambda_{CI}(x\*)x + \lambda_{BJ}(x\*)z
+ \tilde\kappa_{IJ} x\*,
\end{aligned}
\right.
&& J \in \ins{S, E, I, R}
\end{aligned}
\label{eq:competencebinaryI}
\end{equation}
and finally we have the interactions regarding the removed agents
\begin{equation}
\begin{aligned}
\left\lbrace
\begin{aligned}
x' &= (1 - \lambda_R(x))x + \lambda_{CJ}(x)x\* + \lambda_{BR}(x)z
+ \kappa_{RJ} x\\
x\*' &= (1 - \lambda_J(x\*))x\* + \lambda_{CR}(x\*)x + \lambda_{BJ}(x\*)z
+ \tilde\kappa_{RJ} x\*,
\end{aligned}
\right.
&& J \in \ins{S, E, I, R}.
\end{aligned}
\label{eq:competencebinaryR}
\end{equation}
It is reasonable to assume that both the processes of gain and loss of
competence from the interaction with other agents or with the background
in^^>\eqref{eq:competencebinaryS}--\eqref{eq:competencebinaryR} are bounded by
zero. Therefore we suppose that if \mbox{$J, H \in \ins{S, E, I, R}$}, and if
\mbox{$\lambda_J \in [\lambda_J^-, \lambda_J^+]$}, with \mbox{$\lambda_J^- > 0$}
and \mbox{$\lambda_J^+ < 1$}, and \mbox{$\lambda_{CJ}(x),\lambda_{BJ}(x)
\in [0, 1]$} then $\kappa_{HJ}$ may, for example, be uniformly distributed
in^^>\mbox{$[-1 + \lambda_J^+, 1 - \lambda_J^+]$}.
In order to combine the compartmental model SEIR with the evolution
of the competence levels given by
equations^^>^^>\eqref{eq:competencebinaryS}--\eqref{eq:competencebinaryR} we
introduce the interaction operator^^>$Q_{HJ}(\cdot, \cdot)$ following the
standard Boltzmann-type theory^^>\cite{intermultiagent}. As earlier, we will
denote with^^>$J$ a suitable compartment of
the population, i.e., \mbox{$H, J \in \ins{S, E, I, R}$}, and we will use the
brackets^^>$\expvalue\cdot$ to indicate the expectation with respect to the
random variable^^>$\kappa_{HJ}$. Thus, if $\psi(x)$ is an observable function,
then the action of^^>$Q_{HJ}(f_H,f_J)(x,t)$ on^^>$\psi(x)$ is given by
\begin{equation}\label{eq:wk1}
\int_{\R^+} Q_{SJ}(f_S,f_J) \psi(x)\, \de x = \expvalue*{
\int_{\R^2_+} f_S(x,t) f_J(x\*,t)
\bigl(\psi(x') - \psi(x)\bigr)\, \de x\*\de x
},
\end{equation}
with $x'$ defined by^^>\eqref{eq:competencebinaryS}
\begin{equation}\label{eq:wk2}
\int_{\R^+} Q_{EJ}(f_E,f_J) \psi(x)\, \de x = \expvalue*{
\int_{\R^2_+} f_E(x,t) f_J(x\*,t)
\bigl(\psi(x') - \psi(x)\bigr)\, \de x\*\de x
},
\end{equation}
with $x'$ defined by^^>\eqref{eq:competencebinaryE}
\begin{equation}\label{eq:wk3}
\int_{\R^+} Q_{IJ}(f_I,f_J) \psi(x)\, \de x = \expvalue*{
\int_{\R^2_+} f_I(x,t) f_J(x\*,t)
\bigl(\psi(x') - \psi(x)\bigr)\, \de x\*\de x
},
\end{equation}
with $x'$ defined by^^>\eqref{eq:competencebinaryI},
\begin{equation}\label{eq:wk4}
\int_{\R^+} Q_{RJ}(f_R,f_J) \psi(x)\, \de x = \expvalue*{
\int_{\R^2_+} f_R(x,t) f_J(x\*,t)
\bigl(\psi(x') - \psi(x)\bigr)\, \de x\*\de x
},
\end{equation}
with $x'$ defined by^^>\eqref{eq:competencebinaryR}. All the above operators
preserve the total number of agents as the unique interaction invariant,
corresponding to $\psi(\cdot) \equiv 1$.
The system then reads:
\begin{equation}
\left\lbrace
\begin{aligned}
\pd{f_S(x,t)}{t} &= -K(x,t)f_S(x,t) + (1-\alpha(x))\gamma(x)f_I(x,t)
+ \sum_{\mathclap{J \in \ins{S,E,I,R}}}
Q_{SJ}(f_S,f_J)(x,t),\\
\pd{f_E(x,t)}{t} &= K(x,t)f_S(x,t) - \delta(x) f_E(x,t)
+ \sum_{\mathclap{J \in \ins{S,E,I,R}}}
Q_{EJ}(f_E, f_J)(x,t),\\
\pd{f_I(x,t)}{t} &= \delta(x)(1 - \eta(x)) f_E(x,t) - \gamma(x)f_I(x,t)
+ \sum_{\mathclap{J \in \ins{S,E,I,R}}}
Q_{IJ}(f_I, f_J)(x,t),\\
\pd{f_R(x,t)}{t} &= \delta(x)\eta(x) f_E(x,t) + \alpha(x)\gamma(x) f_I(x,t)
+ \sum_{\mathclap{J \in \ins{S,E,I,R}}}
Q_{RJ}(f_R, f_J)(x,t),
\end{aligned}
\right.
\label{eq:seiscompetenza1}
\end{equation}
where the function
\[
K(x,t) = \int_{\R^+} \beta(x,x\*) f_I(x\*,t)\, \de x\*
\]
is responsible for the contagion, $\beta(x,x\*)$ being the contact rate between
agents with competence levels^^>$x$ and^^>$x\*$. In the above formulation we
also assumed $\beta$,
$\gamma$, $\delta$, $\eta$ and^^>$\alpha$ functions of^^>$x$. Note that,
clearly, the most important parameters influenced by individuals' competence
are $\beta(x,x\*)$, since individuals have the highest rates of contact with
people belonging to the same social class, and thus with a similar level of
competence, $\delta(x)$ as individuals with greater competence invest more time
in checking the authenticity of information,
and $\eta(x)$, which characterizes individuals' decision to spread fake news.
On the other hand, the values of $\gamma$ and $\alpha$
we may assume to be less influenced by the level of expertise of individuals.
\subsection{Properties of the kinetic SEIR model with competence}
In this section we analyze some of the properties of the Boltzmann
system^^>\eqref{eq:seiscompetenza1}. First let us consider the reproducing
ratio in presence of knowledge.
By integrating system \eqref{eq:seiscompetenza1} against $x$, and considering
only the compartments of individuals which may disseminate the fake news we have
\begin{equation}
\left\lbrace
\begin{aligned}
\der {E(t)}t &= \int_X K(x,t)f_S(x,t)\,dx - \int_X \delta(x) f_E(x,t)\,dx,\\
\der {I(t)}t &= \int_X \delta(x)(1 - \eta(x)) f_E(x,t)\,dx - \int_X
\gamma(x)f_I(x,t)\,dx.\\
\end{aligned}
\right.
\label{eq:seisintegrated}
\end{equation}
In the above derivation we used the fact that the Boltzmann interaction terms
describing knowledge evolution among agents preserve the total number of
individuals and therefore vanish. Following the analysis in
\cite{BertagliaPareschi}, and omitting the details for brevity, we obtain a
reproduction number generalizing the classical one
\begin{equation}\label{eq:reproductionnumber}
R_0(t) = \frac{\int_X K(x,t)f_S(x,t)\,dx}{\int_X \gamma(x)f_I(x,t)\,dx}.
\end{equation}
Next, following \cite{PhysRevE.102.022303}, we can prove uniqueness of the
solution of^^>\eqref{eq:seiscompetenza1} in the simplified case of constant
parameters: \mbox{$\beta(x,x\*) = \beta > 0$}, \mbox{$\gamma(x) = \gamma > 0$},
\mbox{$\delta(x) = \delta > 0$}, \mbox{$\eta(x) = \eta \in [0,1]$},
\mbox{$\alpha(x) = \alpha \in [0,1]$}. In this
case, exploiting the fact that the interaction operator^^>$Q(\cdot,\cdot)$ has a
natural connection with the Fourier transform by choosing its
kernel^^>\mbox{$\e^{-ix\xi}$} as test function, we can analyze the
system^^>\eqref{eq:seiscompetenza1} with the Fourier transforms of the
densities as unknowns.
Indeed, given a function $f(x) \in L_1(\R^+)$, its Fourier transform is defined
as
\[
\ft f(\xi) = \int_\R \e^{-ix\xi} f(x)\, \de x.
\]
The system^^>\eqref{eq:seiscompetenza1} becomes
\begin{equation}
\left\lbrace
\begin{aligned}
\pd{\ft f_S(\xi,t)}{t} &= -\beta I(t) \ft f_S(\xi,t) + (1-\alpha) \gamma
\ft f_I(\xi,t)
+ \sum_{\mathclap{J \in \ins{S,E,I,R}}}
\ft Q_{SJ}(\ft f_S, \ft f_J)(\xi,t),\\
\pd{\ft f_E(\xi,t)}{t} &= \beta I(t) \ft f_S(\xi,t) - \delta \ft f_E(\xi,t)
+ \sum_{\mathclap{J \in \ins{S,E,I,R}}}
\ft Q_{EJ}(\ft f_E, \ft f_J)(\xi,t),\\
\pd{\ft f_I(\xi,t)}{t} &= \delta(1 - \eta) \ft f_E(\xi,t)
- \gamma \ft f_I(\xi,t)
+ \sum_{\mathclap{J \in \ins{S,E,I,R}}}
\ft Q_{IJ}(\ft f_I, \ft f_J)(\xi,t),\\
\pd{\ft f_R(\xi,t)}{t} &= \delta\eta \ft f_E(\xi,t)
+ \alpha\gamma \ft f_I(\xi,t)
+ \sum_{\mathclap{J \in \ins{S,E,I,R}}}
\ft Q_{RJ}(\ft f_R, \ft f_J)(\xi,t),
\end{aligned}
\right.
\label{eq:seiscompetenzaFourier}
\end{equation}
where the operators $\ft Q_{HJ}(\ft f_H, \ft f_J)$ are defined in terms of the
Fourier transforms of their arguments for^^>\mbox{$J \in \ins{S, E, I, R}$}, so
that
\[
\ft Q_{HJ}(\ft f_H, \ft f_J) = \expvalue{\ft f_H(A_{HJ} \xi -\lambda_{BH}z, t)}
\ft f_J(\lambda_{CJ}\xi,t) - \ft f_H(\xi,t) J(t),
\]
where $A_{HJ}$, with $H,J\in \ins{S, E, I, R}$ is defined as
\begin{equation}\label{eq:defAHJ}
A_{HJ} = 1 - \lambda_H +\kappa_H.
\end{equation}
We suppose that the parameters satisfy the condition
\begin{equation}\label{eq:nucondition}
\nu = \max_{H,J \in \ins{S, E, I, R}} [\lambda_{CJ}^2 + \expvalue{A_{JH}^2}] <
1,
\end{equation}
which will prove useful in the proof.
As in^^>\cite{PhysRevE.102.022303} we recall a class of metrics which is of
natural use in bilinear Boltzmann equations^^>\cite{intermultiagent}. Let $f$
and^^>$g$ be probability densities. Then, for^^>\mbox{$s > 0$} we define
\begin{equation}\label{eq:defd2}
d_s(f,g) = \sup_{\xi \in \R}\frac{\abs{\ft f(\xi) - \ft g(\xi)}}{\abs \xi ^s},
\end{equation}
which is finite whenever $f$ and $g$ have equal moments up to the integer part
of^^>$s$ or to^^>\mbox{$s-1$} if $s$ is an integer.
We have the following result.
\begin{theorem}\label{teo:fouriermetric}
Let $J \in \ins{S, E, I, R}$, and let $f_J(x,t)$ and^^>$g_J(x,t)$ be two
solutions
of the system^^>\eqref{eq:seiscompetenza1} with initial values $f_J(x,0)$
and^^>$g_J(x,0)$ such that $d_2(f_J(x,0), g_J(x,0)$ is finite. Then,
condition^^>\eqref{eq:nucondition} implies that the Fourier based
distance^^>$d_2(f_J(x,t), g_J(x,t)$ decays exponentially (in time) to zero, so
that
\[
\sum_{\mathclap{J \in \ins{S,E,I,R}}} d_2(f_J(x,t), g_J(x,t) \le
\sum_{\mathclap{J \in \ins{S,E,I,R}}} d_2(f_J(x,0), g_J(x,0)\e^{-(1 - \nu)t}.
\]
\end{theorem}
For the details of the proof we refer to Appendix^^>\ref{appendix:A}.
\section{Mean-field approximation}\label{sec:3}
A highly useful tool to obtain information analytically on the large-time
behavior of Boltzmann-type models are scaling techniques; in particular the
so-called \textit{quasi-invariant} limit^^>\cite{intermultiagent}, which allows to derive the
corresponding mean-field description of the kinetic
model \eqref{eq:seiscompetenza1}.
Indeed, let us consider the case in which the interactions between agents
produce small variations of the competence. We scale the quantities involved
in the binary interactions \eqref{eq:competencebinaryS}-\eqref{eq:competencebinaryR} accordingly
\begin{equation}\label{eq:quasiinvariant1}
\begin{aligned}
\lambda_{CJ} \to \epsilon\lambda_{CJ}, &&
\lambda_{BJ} \to \epsilon\lambda_{BJ}, &&
\lambda_J \to \epsilon\lambda_J, &&
\sigma \to \epsilon\sigma,
\end{aligned}
\end{equation}
where $J\in \ins{S, E, I, R}$ and the functions involved in the dissemination
of the fake news, as well
\begin{equation}\label{eq:quasiinvariant2}
\begin{aligned}
\beta(x,x\*) \to \epsilon\beta(x,x\*), &&
\gamma(x) \to \epsilon\gamma(x), &&
\delta(x) \to \epsilon\delta(x), &&
\eta(x) \to \epsilon\eta(x).
\end{aligned}
\end{equation}
We denote by $Q^\epsilon_{HJ}(\cdot, \cdot)$ the scaled interaction terms.
Omitting
the dependence on time on mean values and re-scaling time as
\mbox{$t \to t/\epsilon$}, we obtain up to $\mathcal O(\epsilon)$
\[
\begin{aligned}
\frac1\epsilon \int_{\R^+} Q_{SJ}^{\epsilon}(f_S, f_J) \psi(x)\, \de x
&\approx \int_{\R^+}
\biggl[
-\psi(x)'(\lambda_S xJ - \lambda_{CJ} m_J - \lambda_{BS} m_B J)
+ \frac\sigma2 \psi(x)'' x^2 J
\biggr]
f_S(x,t)\, \de x\\
\frac1\epsilon \int_{\R^+} Q_{EJ}^{\epsilon}(f_E, f_J) \psi(x)\, \de x
&\approx \int_{\R^+}
\biggl[
-\psi(x)'(\lambda_E xJ - \lambda_{CJ} m_J - \lambda_{BE} m_B J)
+ \frac\sigma2 \psi(x)'' x^2 J
\biggr]
f_E(x,t)\, \de x\\
\frac1\epsilon \int_{\R^+} Q_{IJ}^{\epsilon}(f_I, f_J) \psi(x)\, \de x
&\approx \int_{\R^+}
\biggl[
-\psi(x)'(\lambda_I xJ - \lambda_{CJ} m_J - \lambda_{BI} m_B J)
+ \frac\sigma2 \psi(x)'' x^2 J
\biggr]
f_I(x,t)\, \de x\\
\frac1\epsilon \int_{\R^+} Q_{RJ}^{\epsilon}(f_R, f_J) \psi(x)\, \de x
&\approx \int_{\R^+}
\biggl[
-\psi(x)'(\lambda_R xJ - \lambda_{CJ} m_J - \lambda_{BR} m_B J)
+ \frac\sigma2 \psi(x)'' x^2 J
\biggr]
f_R(x,t)\, \de x,
\end{aligned}
\]
where we used a Taylor expansion for small values of $\varepsilon$ of
\[
\psi(x')=\psi(x)+(x'-x)\psi'(x)+\frac{(x'-x)^2}{2}\psi''(x)+\mathcal
O(\epsilon^2)
\]
in \eqref{eq:wk1}-\eqref{eq:wk4} and the scaled interaction rules
\eqref{eq:competencebinaryS}-\eqref{eq:competencebinaryR}.
\subsection{Stationary solutions of Fokker-Planck SEIR models}
Let us impose that $\epsilon \to 0$, following^^>\cite{intermultiagent} from the computations of the previous section we formally obtain the Fokker-Planck
system
\begin{align}
\pd{f_S(x,t)}{t} &= -K(x,t)f_S(x,t) + (1-\alpha(x))\gamma(x)f_I(x,t)
+ \pd{}{x}[(x\lambda_S -\overline{m}(t)
-\lambda_{BS}m_B)f_S(x,t)]\notag\\
&\phantom{=} {}+ \frac\sigma2
\pd{^2}{x^2}(x^2f_S(x,t))\label{eq:fokkerplanck4}\\
\pd{f_E(x,t)}{t} &= K(x,t)f_S(x,t) -\delta(x)f_E(x,t)
+ \pd{}{x}[(x\lambda_E -\overline{m}(t)
-\lambda_{BE}m_B)f_E(x,t)]\notag\\
&\phantom{=} {}+ \frac\sigma2
\pd{^2}{x^2}(x^2f_E(x,t))\label{eq:fokkerplanck5}\\
\pd{f_I(x,t)}{t} &= \delta(x)(1 - \eta(x))f_E(x,t) - \gamma(x)f_I(x,t)
+ \pd{}{x}[(x\lambda_I -\overline{m}(t)
-\lambda_{BI}m_B)f_I(x,t)]\notag\\
&\phantom{=} {}+ \frac\sigma2
\pd{^2}{x^2}(x^2f_I(x,t))\label{eq:fokkerplanck6}\\
\pd{f_R(x,t)}{t} &= \delta(x)\eta(x)f_E(x,t) + \alpha(x)\gamma(x)f_I(x,t)
+ \pd{}{x}[(x\lambda_R -\overline{m}(t)
-\lambda_{BR}m_B)f_R(x,t)]\notag\\
&\phantom{=} {}+ \frac\sigma2
\pd{^2}{x^2}(x^2f_R(x,t))\label{eq:fokkerplanck7}
\end{align}
where now
\[
\overline{m}(t) = \lambda_{CS}m_S(t) + \lambda_{CE}m_E(t) +
\lambda_{CI}m_I(t) + \lambda_{CR}m_R(t).
\]
We can consider the mean values system associated
to^^>\eqref{eq:fokkerplanck4}--\eqref{eq:fokkerplanck7} in the case of constant
epidemiological parameters
\begin{align}
\der{m_S(t)}{t} &= -\beta I(t)m_S(t) + (1-\alpha)\gamma m_I(t)
+ \lambda S(t)(m(t) - m_B)/2-\lambda
m_S(t)\label{eq:fokkerplancmeank4}\\
\der{m_E(t)}{t} &= \beta I(t)m_S(t) -\delta m_E(t)
+ \lambda E(t)(m(t) - m_B)/2-\lambda
m_E(t)\label{eq:fokkerplanckmean5}\\
\der{m_I(t)}{t} &= \delta(1 - \eta)m_E(t) - \gamma m_I(t)
+ \lambda I(t)(m(t) - m_B)/2-\lambda
m_I(t)\label{eq:fokkerplanckmean6}\\
\der{m_R(t)}{t} &= \delta\eta m_E(t) + \alpha\gamma m_I(t)
+ \lambda R(t)(m(t) - m_B)/2-\lambda
m_R(t)\label{eq:fokkerplanckmean7},
\end{align}
with
\[
m(t)=m_S(t) + m_E(t) + m_I(t) + m_R(t).
\]
In the case $\alpha > 0$ or $\eta > 0$, we know that $E(t) \to 0$, $I(t) \to
0$, $S(t) \to S^\infty$ and $R(t) \to R^\infty = 1 - S^\infty$ due to mass
conservation, so that $m_E(t) \to 0$ and $m_I(t) \to 0$ as well. Thus, adding
all the equations together leads us to
\begin{align}\label{eq:meanconvergence2}
\frac\lambda2(m_S^\infty + m_R^\infty) + \frac\lambda2 m_B
&= \lambda(m_S^\infty + m_R^\infty),
\end{align}
i.e., $m_S^\infty + m_R^\infty = m_B$. At this point, adding together
equations^^>\eqref{eq:fokkerplanck4} to^^>\eqref{eq:fokkerplanck7} gives us
\[
0 = \lambda\pd{}{x}(x - m_B)
\sum_{\mathclap{J \in \ins{S,R}}}f_J^\infty(x)
+ \frac\sigma2 \pd{^2}{x^2}\biggl(x^2
\sum_{\mathclap{\quad J \in \ins{S,R}}}f_J^\infty(x) \biggr),
\]
which has as solution an inverse Gamma density
\begin{equation}\label{eq:seirFPfinal}
f^\infty(x) = f_S^\infty(x) + f_R^\infty(x) =
\frac{k^\mu}{\Gamma(\mu)}\frac{\e^{-k/x}}{x^{1 + \mu}},
\end{equation}
\[
\begin{aligned}
\mu = 1 + \frac{2\lambda}\sigma, &&\quad k = (\mu - 1)m_B.
\end{aligned}
\]
It is straightforward to see that the scaled Gamma densities
\[
f_S^\infty (x) = S^\infty \frac{k^\mu}{\Gamma(\mu)}
\frac{\e^{-k/x}}{x^{1 + \mu}}
\quad
f_R^\infty (x) = (1-S^\infty) \frac{k^\mu}{\Gamma(\mu)}
\frac{\e^{-k/x}}{x^{1 + \mu}}
\]
are solutions of the
system^^>\eqref{eq:fokkerplanck4}--\eqref{eq:fokkerplanck7}.
If, instead, $\alpha = \eta = 0$, we find again the same solution
as^^>\eqref{eq:seirFPfinal}, but in this case $J \to \tilde J$, where $\tilde
J$ are defined as in^^>\eqref{eq:endemicequilibria}.
In Figure^^>\ref{fig:stationarysols0125075} we report two examples of
the stationary solutions where we chose the competence variable^^>$z$ to be
uniformly distributed in $[0,1]$: in the first case (left) we considered
$\alpha = \eta = 0$, while in the second case (right) we set $\alpha = 0.2$ and
$\eta = 0.1$.
\begin{figure}
\centering
\includegraphics[width=0.5\columnwidth]{seis-0125-075-1-eps-converted-to.pdf}%
\hskip-.1cm
\includegraphics[width=0.5\columnwidth]{seir-0125-075-1-eps-converted-to.pdf}
\caption{Exact solutions for competence distributions at the end of
epidemic\eqref{eq:seirFPfinal} for $\lambda = 0.1$, $\mu = 5$, $m_B = 0.5$
(solid), $m_B = 3$ (dash-dotted), $\tilde S = 0.5$, $\tilde E = 0.1$, $\tilde I
= 0.4$ (left) and $S^\infty = 0.6$, $R^\infty= 0.4$ (right).
Left: case $\alpha=\eta=0$; right: case $\alpha=0.2$, $\eta = 0.1$.}
\label{fig:stationarysols0125075}
\end{figure}
\section{Numerical examples}\label{sec:4}
In this section we present some numerical tests to show the characteristics of
the model in describing the dynamics of fake news dissemination in a population
with a competence-based structure.
To begin with, we validate the Fokker-Planck model obtained as the
\emph{quasi-invariant limit} of the Boltzmann equation: we will do so through a Monte
Carlo method for the competence distribution^^>(see \cite{intermultiagent}, Chapter 5 for more details).
Next, we approximate the
Fokker-Planck systems^^>\eqref{eq:fokkerplanck4}--\eqref{eq:fokkerplanck7} by
generalizing the structure-preserving
numerical scheme^^>\cite{sscp} to explore the interplay between competence and
disseminating dynamics in the more realistic case of epidemiological parameters
dependent on the competence level (see Appendix^^>\ref{appendix:B}). Lastly, we
investigate how the fake news'
diffusion would impact differently on different classes of the population
defined in terms of their capabilities of interacting with information.
\subsection{Test 1: Numerical quasi-invariant limit}\label{test:1}
In this test we show that the mean-field Fokker-Planck
system^^>\eqref{eq:fokkerplanck4}--\eqref{eq:fokkerplanck7} obtained under the
quasi-invariant scaling^^>\eqref{eq:quasiinvariant1}
and^^>\eqref{eq:quasiinvariant2} is a good approximation of the Boltzmann
models^^>\eqref{eq:seiscompetenza1} when $\epsilon \ll 1$. We do so by using a
Monte Carlo method with $N = 10^4$ particles, starting with a uniform
distribution of competence $f_0(x) = \frac12\chi(x\in[0,2])$, where
$\chi(\cdot)$ is the indicator function, and performing various iterations
until the stationary state was reached; next, the distributions were averaged
over the next 500 iterations.
We considered constant competence-related parameters $\lambda_{CJ} =
\lambda_{BJ}$ and $\lambda_J = \lambda_{CJ} + \lambda_{BJ}$ as well as a
constant variance $\sigma$ for the random variables $\eta_{HJ}$.
In Figure^^>\ref{fig:FPvalidation}, we plotted the results for
$(\lambda,\sigma) = (0.075,0.150)$ (circle-solid, teal) and for
$(\lambda,\sigma) = (0.001,0.002)$ (square-solid, ochre): those choices
correspond to a scaling regime of $\epsilon=0.075$ and $\epsilon =0.001$,
respectively, with $\mu = 2$.
Finally, we assumed that $m_B = 0.75$ (left) and $m_B = 1$ (right).
\begin{figure}
\includegraphics[width=0.5\columnwidth]{seir-validation-epsilon00750001-mb075-eps-converted-to.pdf}%
\hskip -.1cm
\includegraphics[width=0.5\columnwidth]{seir-validation-epsilon00750001-mb1-eps-converted-to.pdf}
\caption{Test 1. Comparison of the competence distributions at the
end of epidemics for system^^>\eqref{eq:seiscompetenza1} with the explicit
Fokker-Planck solution^^>\eqref{eq:seirFPfinal} with scaling parameters
$\epsilon=0.075,0.001$. We considered the case $m_B = 0.75$ (left) and $m_B =
1$ (right).}
\label{fig:FPvalidation}
\end{figure}
Directly comparing the Boltzmann dynamics equilibrium with the explicit
analytic solution of the Fokker-Planck regime shows that if $\epsilon$ is small
enough, Fokker-Planck asymptotics provide a consistent approximation of the
steady states of the kinetic distributions.
\subsection{Test 2: Learning dynamics and fake news
dissemination}\label{test:2}
For this test, we applied the structure-preserving scheme to
system \eqref{eq:fokkerplanck4}--\eqref{eq:fokkerplanck7} in a more realistic
scenario featuring an interaction term dependent on the competence level of
the agents, as well as a competence-dependent delay during which agents
evaluate the information and decide how to act.
In this setting, we refer to the recent Survey of Adult
Skills (SAS) made by the OECD^^>\cite{PIAAC19}: in particular, we focus on
competence understood as a set of information-processing skills, especially
through the lens of literacy, defined^^>\cite{PIAAC19} as \lq\lq the ability to
understand, evaluate, use and engage with written texts in order to participate
in society\rq\rq. One of the peculiarities that makes the SAS, which is an
international, multiple-year spanning effort in the framework of the PIAAC
(Programme for the International Assessment of Adult Competencies) by the
OECD, interesting in our case is that it was administered digitally to more than
70\% of the respondents. Digital devices are arguably the most important
vehicle for information diffusion in OECD countries, so that helps to keep
consistency.
Literacy proficiency was defined through $6$ increasing levels;
we therefore consider a population partitioned in $6$ classes based on the
competence level of their occupants, equated to the score of the literacy
proficiency test of the SAS, normalized. Thus, we chose a log-normal-like
distribution
\[
f(x) = \frac{1}{(\tilde\xi - x)\tilde\sigma\sqrt{2\pi}}
\cdot \e^{- \frac{(\log(\tilde\xi-x)-\tilde\mu)^2}{2\tilde\sigma^2}},
\]
where $\tilde\xi=5$, $\tilde \mu \approx 0.85$ and $\tilde \sigma \approx 0.22$
to make $f(x)$
agree with the empirical findings in^^>\cite{PIAAC19}. The computational domain is restricted
to $x\in[0,5]$ and stationary boundary conditions have been applied as
described in Appendix^^>\ref{appendix:B}.
Initial distributions for the epidemiological compartments were set as
\
f_S(x,0) = \rho_S f(x), \quad f_E(x,0) = \rho_E f(x), \quad f_I(x,0) = \rho_I f(x),\quad
f_R(x,0) = \rho_R f(x),
\]
with $\rho_I = 10^{-2}$, $\rho_S= 1 - \rho_I$ and $\rho_E = \rho_R = 0$.
The contact rate $\beta(x,x_*)$ was set as
\begin{equation}\label{eq:beta}
\beta(x,x_*) = \frac{\beta_0}{(1 + x^2)(1 + x_*^2)}\chi(\abs{x - x_*}\le \Delta)
\end{equation}
with $\Delta =2$
on the hypothesis that interactions occur more frequently among people with a
similar competence level and are higher for people with lower competence levels.
The rate $\delta$ at which the information is evaluated by the agents, who
therefore exits the exposed class, was set to be
\begin{equation}
\delta(x) = \delta_R+(\delta_L-\delta_R) \frac1{1 + \e^{a(b - x)}},
\end{equation}
with $\delta_L = 1$, $\delta_R=5$, $a = 2$ and $b=2.5$. Here,
we are taking into
account that people with higher efficacy at identifying fake news spend
significantly more time on conducting their evaluations than people with lower
efficacy^^>\cite{LEEDER2019100967}. In this specific test case the time range for the
evaluation of the information spans
between 1 day and about 5 hours. The values were purposely chosen rather
large compared to realistic values in order to highlight also the behavior of the exposed
compartment. Finally, we set $\gamma = 0.2$, which correspond to an average fake news duration of 5 days,
and $\alpha = 0.2$, so that individuals have a moderate possibility to remember the
fake news and become immune to it, and assume $\eta=0.1$,
namely in this test we do not relate the decision to spread or not fake news to
the level of competence.
We investigate the relation between the
dissemination-related component of the model and the competence-related one,
which entails that agents can learn, i.e., increase their competence level,
both from the background and from direct binary interactions. Under the
assumption that $\lambda_{CJ} + \lambda_{BJ} = \lambda_J$, which is a
conservative choice: the expected value of the competence gained through
interactions cancels out the one lost due to forgetfulness, in this latter
process two main parameters are involved: $\lambda = \lambda_J$ (i.e., all
compartments have the same learning rate) and $m_B$,
which is the mean of the background competence variable^^>$z$.
For what
concerns the dissemination-related component, instead, the main factor is the
reproduction number $R_0$^^>\eqref{eq:reproductionnumber}. Hence, we measured
the differences on the spread of fake news varying these three parameters. In
Figure^^>\ref{fig:interplay} (left) we show the highest portion of spreaders in
relation to the mean of background^^>$m_B$ and to the reproduction
number^^>$R_0$; in the right image $\lambda$ is opposed to $R_0$.
\begin{figure}
\centering
\includegraphics[width=0.5\columnwidth]{interplay-mBdiv4-60-eps-converted-to.pdf}\hskip
-.1cm
\includegraphics[width=0.5\columnwidth]{interplay-lambda-60-eps-converted-to.pdf}
\caption{Test 2. Interplay between competence levels and dissemination
dynamics. Contour plots of the highest number of infected in relation to the
reproduction number^^>\eqref{eq:reproductionnumber} $R_0 \in [1.1,10]$ and the
background competence mean $m_B \in [0.03125, 0.25]$ (left) and learning rate
$\lambda \in [0.0125, 0.125]$ (right).}
\label{fig:interplay}
\end{figure}
To perform the test, we leveraged the structure-preserving numerical
scheme^^>\cite{sscp} whose details are presented for
convenience in Appendix^^>\ref{appendix:B}.
In both images of Figure^^>\ref{fig:interplay} we see transition effects: the
learning process triggered by the competence dynamics is capable of slowing
down the dissemination of fake news in the population, even to the point of
preventing it to take place. In the first case, the mean of the background
$m_B$, i.e., the mean of the distribution of the background competence variable
$z$, which we assumed uniformly distributed, varies between $0.03125$ and
$0.25$, while the reproduction number $R_0$ varies between $1.1$ and $10$. In
the second case, we left untouched $R_0$, while $\lambda$ varies between
$0.0125$ and $0.125$ with a background mean $m_B = 0.125$. We can see that the
mean of the background has a more pronounced impact on the slowing of the
diffusion of fake news, with a steeper transition effect.
\subsection{Test 3: Impact of the different competence levels}
In this final test we considered how much of an impact the competence level can
have on the dissemination of fake news in the population. We simulated the
mean-field model^^>\eqref{eq:fokkerplanck4}--\eqref{eq:fokkerplanck7} assuming the same
competence-dependent contact rate $\beta(x,x_*)$ of Test 2, in
this case with $\beta_0=4$, as well the same delay rate $\delta(x)$ and the
same $\gamma$, but we additionally assume that the decision to spread or not a
fake news is affected by the level of competence. This is somewhat controversial
in the literature since other factors also affect this behavior like the age
of individuals (tests carried out on young people have shown independence from the
competence in the decision to share a fake news in contrast to what happens in
in older people, see \cite{PIAAC19,LEEDER2019100967}). To emphasize this effect we assume
\begin{equation}
\eta(x)=1-e^{-kx^2},
\end{equation}
with $k=0.1$. Thus individuals with high level of competence rarely decide to spread
fake news.
In Figure^^>\ref{fig:seir-piaac-evolution-01} we
report the time evolution of the distributions of susceptible (top left),
exposed (top right), infected (bottom left) and removed (bottom right) agents
with competence parameters of $\lambda_{BJ} = \lambda_{CJ} = 0.125$,
$\lambda_J= \lambda_{BJ} + \lambda_{CJ}$ and $m_B = 0.125$, in the case $\alpha
= 0.1$. In Figure^^>\ref{fig:seir-piaac-evolution-09}, instead, we show the
evolution with the same parameters except for a larger probability $\alpha=0.9$
of remembering the fake news.
\begin{figure}
\centering
\includegraphics[width=0.495\columnwidth]{susceptible-piaac-eta-01-eps-converted-to.pdf}
\includegraphics[width=0.495\columnwidth]{exposed-piaac-eta-01-eps-converted-to.pdf}
\includegraphics[width=0.495\columnwidth]{infected-piaac-eta-01-eps-converted-to.pdf}
\includegraphics[width=0.495\columnwidth]{removed-piaac-eta-01-eps-converted-to.pdf}
\caption{Test 3. Time evolution of the competence distribution for the kinetic
model^^>\eqref{eq:seiscompetenza1} with competence parameters $\lambda_{BJ} =
\lambda_{CJ} = 0.125$, $\lambda_J= \lambda_{BJ} + \lambda_{CJ}$, with a
background competence mean of $m_B$ = 0.125. We considered $\alpha=0.1$,
$\beta_0 = 4$, $\gamma =0.2$. Top left: susceptible; top right: exposed; bottom
left: infected; bottom right: removed.}\label{fig:seir-piaac-evolution-01}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.495\columnwidth]{susceptible-piaac-eta-09-eps-converted-to.pdf}
\includegraphics[width=0.495\columnwidth]{exposed-piaac-eta-09-eps-converted-to.pdf}
\includegraphics[width=0.495\columnwidth]{infected-piaac-eta-09-eps-converted-to.pdf}
\includegraphics[width=0.495\columnwidth]{removed-piaac-eta-09-eps-converted-to.pdf}
\caption{Test 3. Time evolution of the competence distribution for the kinetic
model^^>\eqref{eq:seiscompetenza1} with competence parameters $\lambda_{BJ} =
\lambda_{CJ} = 0.125$, $\lambda_J= \lambda_{BJ} + \lambda_{CJ}$, with a
background competence mean of $m_B$ = 0.125. We considered $\alpha=0.9$,
$\beta_0 = 4$, $\gamma =0.2$. Top left: susceptible; top right: exposed; bottom
left: infected; bottom right: removed.}\label{fig:seir-piaac-evolution-09}
\end{figure}
In Figure^^>\ref{fig:seir-piaac-relative-numbers} are shown the relative
numbers of susceptible, exposed, infected and removed agents, on the left for
$\alpha = 0.1$ and on the right for $\alpha = 0.9$.
\begin{figure}
\centering\includegraphics[width=0.495\columnwidth]{densities-piaac-eta-01-eps-converted-to.pdf}
\centering\includegraphics[width=0.495\columnwidth]{densities-piaac-eta-09-eps-converted-to.pdf}
\caption{Test 3. Evolution in time of the densities for susceptible, exposed,
infectious and removed agents. Left: $\alpha = 0.1$; right: $\alpha = 0.9$.}
\label{fig:seir-piaac-relative-numbers}
\end{figure}
To measure the effects of the competence, we considered the curve of the
infected agents depending on their levels accordingly to^^>\cite{PIAAC19} for
$x\in [0,5]$:
\begin{itemize}
\item below level 1: scoring less then 175/500, ($x < 1.75$);
\item level 1: scoring between 176/500 and 225/500, ($x > 1.75$ and $x < 2.25$);
\item level 2: scoring between 226/500 and 275/500, ($x > 2.25$ and $x < 2.75$);
\item level 3: scoring between 276/500 and 325/500, ($x > 2.75$ and $x < 3.25$);
\item level 4: scoring between 326/500 and 375/500, ($x > 3.25$ and $x < 3.75$);
\item level 5: scoring more than 375/500, ($x > 3.75$).
\end{itemize}
\begin{figure}
\centering
\includegraphics[width=0.495\columnwidth]{levels-piaac-eta-01-eps-converted-to.pdf}
\includegraphics[width=0.495\columnwidth]{levels-piaac-eta-09-eps-converted-to.pdf}
\caption{Test 3. Evolution of the fraction of infectious agents with different
levels of competence as defined in^^>\cite{PIAAC19}. Left: $\alpha = 0.1$;
right: $\alpha = 0.9$.}
\label{fig:seir-piaac-levels}
\end{figure}
Figure^^>\ref{fig:seir-piaac-levels} shows clearly that the more
competent the individual, the lesser they contribute to the spread of fake
news, in perfect agreement with the transition effects observed in
Test^^>\ref{test:2} due to the interplay between competence and the
dissemination dynamics.
Moreover, we can see how the probability $\alpha$ of detecting fake news
influences its dissemination in the population: a lower probability implies a
higher peak of infected agents for each competence level, as well as a slower
spread overall.
\section{Concluding remarks}\label{sec:conclusions}
In this paper, we introduced a compartmental model for fake news dissemination
that also considers the competence of individuals. In the model, the concept of
competence is not introduced as a static feature of the dynamic, but as an
evolutionary component that takes into account both learning through
interactions between agents and interventions aimed at educating individuals in
the ability to detect fake news.
From a mathematical viewpoint the competence dynamics has been introduced as a
Boltzmann interaction term in the corresponding system of differential
equations. A suitable scaling limit, permits to recover the corresponding
Fokker-Planck models and then the resulting stationary states in terms of
competence. These, in agreement with \cite{PARESCHI2017201}, are given
explicitly by Gamma distributions.
The numerical results demonstrate the model's ability to correctly describe the
interplay between fake news dissemination and individuals' level of competence,
highlighting transition phenomena at the level of expertise that allow fake
news
to spread more rapidly.
Future developments of the model will be considered in particular in the case
of networks, in order to describe the spread of fake news on social networks
and present plausible scenarios useful to limit the spread of false
information. This can be done by following an approach similar to that of
kinetic models for opinion formation on networks^^>\cite{network1}.
Another challenging aspect concerns the matching of the model with realistic
data that requires the introduction of quantitative aspects not always easy to
identify^^>\cite{maleki2021using,datarepository}. One of the main
applications will be related to combating misinformation in the vaccination campaign
against COVID-19.
\section*{Acknowledgments}
This work was partially supported by MIUR (Ministero dell’Istruzione,
dell’Università e della
Ricerca) PRIN 2017, project “\textit{Innovative numerical methods for
evolutionary partial differential equations and applications}”, code 2017KKJP4X.
|
\section{Introduction}
\label{sec:intro}
\blfootnote{This work was made with the support of the French National Research Agency, in the framework of the project LEAUDS Learning to understand audio scenes (ANR-18-CE23-0020), the project CPS4EU Cyber Physical Systems for Europe (Grant Agreement number: 826276) and the French region Grand-Est. Experiments presented in this paper were carried out using the Grid5000 testbed, supported by a scientific interest group hosted by Inria and including CNRS, RENATER and several Universities as well as other organizations (see \href{https://www.grid5000}{https://www.grid5000}).}
The main goal of ambient sound and scene analysis is to automatically extract information from sounds that surround us and analyze them for different purposes and applications.
Between the different area of interest, ambient sound analysis have a considerable impact on applications such as noise monitoring in smart cities \cite{bello2018sonyc, bello2018sound}, domestic applications such as smart homes and home security solutions \cite{serizel2018large, debes2016monitoring}, health monitoring systems \cite{zigel2009method}, multimedia information retrieval \cite{jin2012event} and bioacoustics domain \cite{morfi2021deep}.
Sound Event Detection (SED) aims to identify the onset and offset of the sound events present in a soundscape and to correctly classify them, labeling the events according to the target sound classes that they belong to. Nowadays, deep learning is the main method used to approach the problem. However, one of the main limitations of deep learning models is the requirement of large amounts of labeled training data to reach good performance. The process of labeling data is time-consuming and bias-prone mainly due to human errors and disagreement given the subjectivity in the perception of some sound event onsets and offsets \cite{Turpault2019_DCASE}. To overcome these limitations, recent works are investigating alternatives to train deep neural networks with a small amount of labeled data together with a bigger set of unlabeled data \cite{serizel2018large, serizel2019sound, shah2018closer, Turpault2019_DCASE, mcfee2018adaptive}. Among them, Detection and Classification Acoustic Scenes and Events Challenge (DCASE) 2021 Task 4 uses an heterogeneous dataset that includes both recorded and synthetic soundscapes \cite{Turpault2019_DCASE}.
This latter soundscapes provide a cheap way to obtain strongly labeled data. Until recently, synthesized soundscapes were generated considering only target sound events. However, recorded soundscapes also contain a considerable amount of non-target events that might influence the performance of the system.
The purpose of this paper is to focus on the impact on the system's performance when non-target events are included in the synthetic soundscapes of the training dataset. The study has been mainly divided into three stages.
Firstly, we investigate to what extent using non-target events alternatively during training or validation helps the system to correctly detect the target sound events. Mainly motivated from the results of the first experiment, in the second part of the study, we focus on understanding to what extend adjusting the target to non-target signal-to-noise ratio (TNTSNR) at training improves the sound event detection performance. Results regarding a preliminary study on the evaluation of the system using clips containing only non-target events are also reported, opening questions for future studies on possible acoustic similarity between target and non-target sound events which might confuse the SED system. \footnote{To promote reproducibility, the code, \url{https://github.com/DCASE-REPO/DESED_task},
and pre-trained models \url{https://zenodo.org/record/5529692}, are made available under an open-source license.}.
\section{PROBLEM DEFINITION AND DATASET GENERATION}
\label{sec:format}
\subsection{Problem definition}
\label{ssec:probeldef}
The primary goal of the DCASE 2021 Challenge Task 4 is the development of a semi-supervised system for SED, exploiting an heterogeneous and unbalanced training dataset.
The goal of the system is to correctly classify the sound event classes and to localize the different target sound events present in an audio clip in terms of timing. Each audio recording can contain more than one event. Some of those could also be overlapped.
The use of a larger amount of unlabeled recorded clips is motivated by the limitations related to annotating a SED dataset (human-error-prone and time-consuming).
Alternatively, synthesized soundscapes are an easy way to have strongly annotated data. In fact, the user can easily generate the soundscapes starting from isolated sound events.
On the other hand, in most of the recorded soundscapes the target sound classes are almost never present alone.
For this reason, one of the main novelties of the DCASE 2021 Challenge Task 4 is the introduction of non-target isolated events in the synthetic soundscapes\footnote{\url{http://dcase.community/challenge2021}}.
This paper explores the impact of the non-target sound events on the baseline system performance, with the final goal of understanding and highlighting how to correctly exploit them to generate realistic soundscapes.
\subsection{Dataset generation}
\label{ssec:datasetgen}
The dataset used in this paper is the DESED dataset\footnote{\href{https://project.inria.fr/desed/}{https://project.inria.fr/desed/}} \cite{serizel:hal-02355573, turpault:hal-02160855}, which is the same provided for the DCASE 2021 Challenge Task 4.
It is composed of 10 seconds length audio clips either recorded in a domestic environment or synthesized to reproduce such an environment\footnote{For a detailed description of the DESED dataset and how it is generated the reader is referred to the original DESED article~\cite{turpault:hal-02160855} and DCASE 2021 task 4 webpage: \url{http://dcase.community/challenge2021}}.
The synthetic part of the dataset is generated with Scaper \cite{salamon2017scaper}, a Python library for soundscape synthesis and augmentation,
which allows to control audio parameters.
The recorded soundscapes are taken from AudioSet \cite{gemmeke2017audio}. The foreground events (both target and non-target) are obtained from the Freesound Dataset (FSD50k) \cite{fonseca2020fsd50k}, while the background sounds are obtained from the SINS dataset (activity class “other”) \cite{dekkers2017sins} and TUT scenes 2016 development dataset \cite{mesaros2016tut}. In particular, non-target events are the intersection of FUSS dataset \cite{wisdom2021s} and FSD50k dataset in order to have compatibilty with the source separation baseline system.
In this article, we modify only the synthetic subset of the dataset. Starting from the synthetic part of the DESED dataset, we generated different versions of it in order to investigate how non-target events impact the system performance and to what extent their relationship with the target events affects the training phase of the system.
The following subsections describe the different subsets used for the experiments, which have been generated using Scaper.
\subsubsection{Synthetic training set}
\label{sssec:trainset}
The synthetic training set is the same set of data released for the DCASE 2021 Challenge Task 4. It includes 10000 audio clips where both target and non-target sound events could be present in each clip. The distribution of the sound events among the files have been determined considering the co-occurrences between the different sound events. The co-occurrences have been calculate considering the strong annotations released for the AudioSet dataset~\cite{hershey2021benefit}\footnote{The co-occurrences distribution and the code used to compute them will be distributed.}. A second version of this dataset has been generated where only target events are present.
The datasets will be hereafter referred as \textbf{synth\_tg\_ntg} (used by the official baseline system) and \textbf{synth\_tg} for the synthetic subset including target and non-target events and the synthetic subset including only target events, respectively.
\subsubsection{Synthetic validation set}
\label{sssec:valset}
The synthetic validation set is the same as the synthetic validation dataset supplied for the DCASE 2021 Challenge Task 4. It includes 3000 audio clips including target and non-target events, which distribution has been defined calculating the co-occurrences between sound events. We generated a second version of the dataset containing only target events. The datasets will be referred to as \textbf{synth\_tg\_ntg\_val} (used by the baseline system) and \textbf{synth\_tg\_val} (only target sound events).
\subsubsection{Synthetic evaluation set}
\label{sssec:synth21eval}
The synthetic 2021 evaluation set is composed by 1000 audio clips.
In the context of the challenge, this subset is used for analysis purposes. We will refer to it as \textbf{synth\_tg\_ntg\_eval}. It contains target and non-target events distributed between the different audio clips according to the pre-calculated co-occurrences. Two different versions of the \textbf{synth\_tg\_ntg\_eval} set have been generated, \textbf{synth\_tg\_eval} (only target sound events) and \textbf{synth\_ntg\_eval} (only non-target sound events).
\subsubsection{Varying TNTSNR training and validation set}
\label{sssec:TNTSNRset}
With the aim of studying what would be the impact of varying the TNTSNR on the system performance, different versions of \textbf{synth\_tg\_ntg} and \textbf{synth\_tg\_ntg\_val} have been generated. In particular, for each of them, three versions have been created. The SNR of the non-target events have been decreased by 5 dB, 10 dB and 15 dB compared to their original value. The original SNR of the sound events is randomly selected between 6 dB and 30 dB, so the more we decrease the SNR, the less the sound will be audible, with some of the events that will not be audible at all. These subsets will be subsequently referred to as \textbf{synth\_5dB}, \textbf{synth\_10dB}, \textbf{synth\_15dB} for the training subsets and \textbf{synth\_5dB\_val}, \textbf{synth\_10dB\_val}, \textbf{synth\_15dB\_val} for the validation subsets.
\subsubsection{Public evaluation set}
\label{sssec:publicset}
The public evaluation set is composed of recorded audio clips extracted from Youtube videos that are under creative common licenses.
This is part of the evaluation dataset released for the evaluation phase of the DCASE 2021 Challenge Task 4 and considered for ranking.
The set will be referred to as \textbf{public}.
\section{EXPERIMENTS TASK SETUP}
\label{sec:experiments}
In order to compare the results with the official baseline, we used the same SED
mean-teacher system released for this year challenge. More information regarding the system can be found at Turpault et al.~\cite{Turpault2019_DCASE} and on the official webpage of the DCASE Challenge Task 4.
All the different models have been trained 5 times. This paper reports the average of the scores and the confidence intervals related to those. Only for the baseline model we do no report the confidence intervals because we have considered the results using the checkpoint made available for it \footnote{\url{https://zenodo.org/record/4639817}}. The metrics considered for the study are the two polyphonic sound detection score (PSDS)~\cite{bilen2020framework} scenarios defined for the DCASE 2021 Challenge Task 4, since these are the official metrics used in the challenge.
The scope of these experiments is twofold: understand the impact of non-target events on the system performance and investigate to what extend the TNTSNR helps the network to correctly predict the sound events in both matched and mismatched conditions. In order to do so, we divided the experiment into three stages. The first part of the study is focused on understanding the influence of training the system with non-target events. This experiment is described and discussed in Section ~\ref{sec:tgntgexp}. Section ~\ref{sec:varyingTNT} reports the results and the relative discussion of the second part of the experiment where we investigate if a mismatch in terms of TNTSNR between datasets could have an impact on the output of the system. Section ~\ref{sec:notgeval} reports preliminary results of the last stage of the experiment, regarding the evaluation of the system on the \textbf{synth\_ntg\_eval} dataset, formed by only non-target sound events, in order to investigate if some classes could get acoustically confused at training, having a negative impact on the performance. The last stage has been motivated by the results of the second part of the experiment.
\section{Using TARGET/NON-TARGET at training}
\label{sec:tgntgexp}
\begin{table}[t]
\centering
\begin{tabular}{c|c|c|c}
\toprule
\multicolumn{2}{c|}{\textbf{Non-target}} & \multirow{2}{*}{\textbf{PSDS1}} & \multirow{2}{*}{\textbf{PSDS2}} \\
\cline{0-1}
Train & Val&& \\
\midrule
\checkmark & & 33.81 (0.36) & 52.62 (0.19) \\
& \checkmark & 35.92 (0.49) & 54.85 (0.29) \\
& & 34.90 (0.82) & 53.07 (1.22) \\
\checkmark & \checkmark & \textbf{36.40} & \textbf{58.00} \\
\bottomrule
\end{tabular}
\caption{Evaluation results for the \textbf{public} set, considering the different combinations of using target and non-target sound events at training and validation.}
\label{tab:results_public}
\end{table}
In the first experiment we concentrate on training the system with different combinations of the training dataset.
Table ~\ref{tab:results_public} reports the results of the experiment evaluating the system on the \textbf{public} set. We check-marked the columns NT Train or/and NT Val according to if the non-target sound events are present or not in the synthetic sounscapes. From the results it is possible to observe that using non-target sound events during training and validation improves the performance by a large margin with relaxed segmentation constraints (PSDS2) but only marginally with strict segmentation constraints (PSDS1). In this latter case what matters the most is the use of non-target sound events during the validation. A possible explanation is that synthetic soundscapes with non-target sound events are actually too difficult and confuse the systems when used during the training but they still help reducing the mismatch with recorded soundscapes during model selection (validation).
\begin{table}[!t]
\begin{tabularx}{\columnwidth}{c@{\hskip0.10in}|c@{\hskip0.10in}|c@{\hskip 0.10in}|c@{\hskip0.10in}|c@{\hskip0.10in}}
\toprule
\multicolumn{2}{c|}{\textbf{Non-target}} & \multirow{2}*{\textbf{Eval set}} & \multirow{2}*{\textbf{PSDS1}} & \multirow{2}*{\textbf{PSDS2}} \\
\cline{0-1}
Train & Val&& \\
\midrule
\checkmark & & synth\_tg\_ntg\_eval & 23.22 (1.33) & 36.44 (2.62) \\
& \checkmark & synth\_tg\_ntg\_eval & 20.08 (0.39) & 31.33 (1.29) \\
& & synth\_tg\_ntg\_eval & 20.13 (0.35) & 30.99 (1.07) \\
\checkmark & \checkmark & synth\_tg\_ntg\_eval & \textbf{25.14} & \textbf{40.12} \\
\hline
\hline
\checkmark & & synth\_tg\_eval & 42.82 (2.42) & 58.26 (2.08) \\
& \checkmark & synth\_tg\_eval & 46.92 (1.02) & \textbf{62.79 (0.55)} \\
& & synth\_tg\_eval & \textbf{47.73 (0.33)} & 62.54 (1.00) \\
\checkmark & \checkmark & synth\_tg\_eval & 43.22 & 61.09 \\
\bottomrule
\end{tabularx}
\caption{Evaluation results for the \textbf{synth\_tg\_ntg\_eval} set and \textbf{synth\_tg\_eval} set, considering the different combination of using target and non-target sound events at training and validation.}
\label{tab:results_tgeval}
\end{table}
Table ~\ref{tab:results_tgeval} reports the results considering the \textbf{synth\_tg\_ntg\_eval} and \textbf{synth\_tg\_eval} evaluation sets. In all cases the best performance is obtained in matched training/evaluation conditions. The performance obtained on \textbf{synth\_tg\_ntg\_eval} are lower than the performance obtained on \textbf{synth\_tg\_eval} even in matched conditions. Not surprisingly, this confirm that including non-target sound events makes the SED task more difficult. Interestingly, as opposed to the previous experiment, the most important here is to have matched conditions during training and to a lesser extent during validation. In order to verify the low impact of non-target sound events at training when evaluating on recorded soundscapes, in the next experiment we investigate a possible mismatch in terms in TNTSNR.
\section{Varying TNTSNR at training}
\label{sec:varyingTNT}
\begin{table}[!t]
\centering
\begin{tabular}{c|c|c|c}
\toprule
\multicolumn{2}{c|}{\textbf{Non-target}} & \multirow{2}*{\textbf{PSDS1}} & \multirow{2}*{\textbf{PSDS2}} \\
\cline{0-1}
Train & Val&& \\
\midrule
Original & 5~dB & 35.57 (0.28) & 56.68 (1.77) \\
5~dB & Original & 36.25 (1.26) & 57.53 (1.06) \\
5~dB & 5~dB & 35.46 (0.46) & \textbf{58.09} (0.74) \\
Original & Original & \textbf{36.40} & 58.00 \\
\bottomrule
\end{tabular}
\caption{Evaluation results for the second part of the experiment, varying TNTSNR by 5 dB (\textbf{synth\_5dB} and \textbf{synth\_5dB\_val}). Evaluating with \textbf{public} set.}
\label{tab:results_5db}
\end{table}
\begin{table}[!t]
\centering
\begin{tabular}{c|c|c|c}
\toprule
\multicolumn{2}{c|}{\textbf{Non-target}} & \multirow{2}*{\textbf{PSDS1}} & \multirow{2}*{\textbf{PSDS2}} \\
\cline{0-1}
Train & Val &&\\
\midrule
Original & 10~db & 36.23 (1.11) & 57.82 (1.37) \\
10~db & Original & \textbf{36.42 (0.77)} & \textbf{58.94 (0.89)} \\
10~db & 10~db & 36.20 (1.14) & 57.92 (1.04) \\
Original & Original & 36.40 & 58.00 \\
\bottomrule
\end{tabular}
\caption{Evaluation results for the second part of the experiment, varying TNTSNR by 10 dB (\textbf{synth\_10dB} and \textbf{synth\_10dB\_val}). Evaluating with \textbf{public} set.}
\label{tab:results_10db}
\end{table}
\begin{table}[!t]
\centering
\begin{tabular}{c|c|c|c}
\toprule
\multicolumn{2}{c|}{\textbf{Non-target}} & \multirow{2}*{\textbf{PSDS1}} & \multirow{2}*{\textbf{PSDS2}} \\
\cline{0-1}
Train & Val \\
\midrule
Original & 15~dB & 36.08 (1.13) & 57.78 (1.33) \\
15~dB & Original & \textbf{37.37 (0.70)} & \textbf{58.64 (1.34)} \\
15~dB & 15~dB & 36.10 (0.50) & 57.36 (0.89) \\
Original & Original & 36.40 & 58.00 \\
\bottomrule
\end{tabular}
\caption{Evaluation results for the second part of the experiment, varying TNTSNR by 15 dB (\textbf{synth\_15dB} and \textbf{synth\_15dB\_val}). Evaluating with \textbf{public} set.}
\label{tab:results_15db}
\end{table}
\begin{table}[!t]
\centering
\begin{tabular}{c|c|c}
\toprule
Validation set & PSDS1 & PSDS2 \\
\midrule
synth\_5dB\_val & 38.68 (1.07) & 60.57 (0.78) \\
synth\_10dB\_val & \textbf{39.07 (0.75)} & \textbf{60.75 (0.80)} \\
synth\_15dB\_val & 37.95 (0.53) & 59.99 (1.14) \\
\bottomrule
\end{tabular}
\caption{Evaluation results of the SED system, training with \textbf{synth\_tg}, validating with varying TNTNSNR set and evaluating with \textbf{public} set.}
\label{tab:results_tgSNR}
\end{table}
The second part of the study focuses on understanding the impact of varying the TNTSNR at training and validation aiming at finding a TNTSNR condition that could match better the recorded soundscapes.
For each TNTSNR, we use similar combinations
as the ones used in Section ~\ref{sec:tgntgexp}, replacing the set without non-target sound events by a set with adjusted TNTSNR.
For example, considering the 5 dB case, the combinations considered would be:
\begin{itemize}
\item training using the \textbf{synth\_tg\_ntg} set and validating with \textbf{synth\_5dB\_val};
\item training with \textbf{synth\_5dB} and validating with \textbf{synth\_tg\_ntg\_val};
\item training and validating with \textbf{synth\_5dB} and \textbf{synth\_5dB\_val}.
\end{itemize}
The fourth combination is the official DCASE Task 4 baseline.
Repeating the experiment with all the varying TNTSNR, allow us to analyse to what extend the loudness of the non-target events helps matching the evaluation conditions on recorded clips.
Table ~\ref{tab:results_5db}, ~\ref{tab:results_10db} and ~\ref{tab:results_15db} report the performance on the \textbf{public} set when using a TNTSNR of 5~dB, 10~dB and 15~dB, respectively. When the TNTSNR is 5~dB or 10~dB, the performance changes only marginally between configurations. Increasing the TNTSNR to 15~dB leads to a behaviour more similar to the one obtained in Table~\ref{tab:results_public}. The best performance is obtained when training with TNTSNR is 15~dB and validating on \textbf{synth\_tg\_ntg\_val}. This could be explained by the fact TNTSNR 15~dB is a condition closer to that of the recorded soundscapes and the fact that it allows for selecting models that will be more robust towards non-target events at test time.
In the last experiment, we investigate the impact of varying the TNTSNR during validation phase, while using the \textbf{synt\_tg} for training.
Results are reported on Table \ref{tab:results_tgSNR}, where it is possible to observe that all of them overcome the baseline or are comparable with it, with the best performance obtained for 10~dB TNTSNR. These experiments could indicate that recorded soundscapes in \textbf{public} in general have a TNTSNR of about 10 -- 15~dB which should be confirmed by complementary experiments.
\section{EVALUATING ON NON-TARGET EVENTS ONLY}
\label{sec:notgeval}
Based on the previous experiments, TNTSNR could be one reason of mismatch between the synthetic soundscapes and the recorded soundscapes. But this could not explain all the performance differences observed here. In particular why in general having lower TNTSNR during training is decreasing the performance regardless of the validation. One possibility is that the system gets acoustically confused by a possible similarity in sound between events when soundscapes tend to be less dominated by target events. So we evaluated the system using the \textbf{synth\_ntg\_eval}, where only non-target events are considered, to see for which classes the system would output false positives. We evaluated the system on the \textbf{public} set; considering the systems trained for the first experiment (see Table~\ref{tab:results_public}). Results show that some sound events are detected more than others. For some classes as Speech, this could be explained by the original event distribution (indicated in the first column) but for some other classes as Dishes there is a discrepancy between the original distribution and the amount of false alarms. Interestingly the amount of false alarms is decreased sensibly for most of the classes when including non-target sound events during training.
\begin{table}[!t]
\centering
\begin{tabular}{l||c||c|c|c|c}
\toprule
&\multirow{2}{*}{Nref }&\multicolumn{4}{c}{Nsys}\\
Classes& & A & B & C & Base \\
\midrule
Dog & 197 & 135 & 126 & 146 & 79 \\
Vacuum\_cleaner & 127 & 31 & 42 & 44 & 47 \\
Alarm\_bell & 191 & 47 & 50 & 52 & 59 \\
Running\_water & 116 & 34 & 41 & 61 & 30 \\
Dishes & 405 & 1478 & 395 & 1270 & 305 \\
Blender & 100 & 63 & 32 & 55 & 19\\
Frying & 156 & 70 & 41 & 60 & 33\\
Speech & 1686 & 206 & 181 & 180 & 201 \\
Cat & 141 & 99 & 103 & 98 & 73 \\
Electric\_shaver & 103 & 21 & 18 & 18 & 7\\
\bottomrule
\end{tabular}
\caption{Preliminary evaluation results by classes, evaluating the system with \textbf{synth\_ntg\_eval}. Nsys (A): training with \textbf{synth\_tg}, validating with \textbf{synth\_tg\_val}; Nsys (B): training with \textbf{synth\_tg\_ntg}, validating with \textbf{synth\_tg\_val}; Nsys (C): training with \textbf{synth\_tg}, validating with \textbf{synth\_tg\_ntg\_val}; Base: baseline using target and non-target events for training and validation.}
\label{tab:results_last}
\end{table}
\section{CONCLUSIONS AND FUTURE WORK}
\label{sec:conclusion}
This paper analyzes the impact of including non-target sound events in the synthetic soundscapes of the training dataset for SED systems trained on heterogeneous dataset. In particular, the experiments are divided into three stages: in the first part, we explore to what extend using non-target sound events at training has an impact on the system's performance, secondly we investigate the impact of varying TNTSNR and we conclude the study by analyzing a possible confusion of the SED model in case of false alarms triggered by non-target sound events.
From the results reported on this paper, we can conclude that using non-target sound events can help the SED system to better detect the target sound events, but it is not clear to what extend and what would be the best way to generate the soundscapes. Results show that the final SED performance could depend on mismatches between synthetic and recorded soundscapes, part of which could be due to the TNTSNR but not only. Results on the last experiment show that using non-target events at training decreases the amount of false alarms at test but from this experiment it is not possible to conclude on the impact of non-target sound events on the confusion between the target sound events. This is a first track for future investigation on the topic. Additionally, the impact of the non-target sound events at training on the ability of the system to better segment the target sound events in noisy soundscapes would have to be investigated. A final open question is the impact of the per class distribution of the sound events (both target and non-target) and their co-occurrence distribution on the SED performance.
\section{ACKNOWLEDGEMENTS}
\label{sec:acknowledgements}
We would like to thank all the other organizers of DCASE 2021 Challenge Task
4. In particular, we thank Eduardo Fonseca and Daniel P. W. Ellis for their help with the strong labels of the AudioSet dataset used to compute the events co-occurrences, and Justin Salamon and Prem Seetharaman for their help with Scaper.
\bibliographystyle{IEEEtran}
|
\section{Credits}
This document has been adapted by Roberto Navigli
from the instructions for earlier ACL, NAACL and EMNLP proceedings, including those for
EMNLP 2020 by Yulan He,
ACL 2020 by Steven Bethard, Ryan Cotterrell and Rui Yan,
ACL 2019 by Douwe Kiela and Ivan Vuli\'{c},
NAACL 2019 by Stephanie Lukin and Alla Roskovskaya,
ACL 2018 by Shay Cohen, Kevin Gimpel, and Wei Lu,
NAACL 2018 by Margaret Michell and Stephanie Lukin,
2017/2018 (NA)ACL bibtex suggestions from Jason Eisner,
ACL 2017 by Dan Gildea and Min-Yen Kan,
NAACL 2017 by Margaret Mitchell,
ACL 2012 by Maggie Li and Michael White,
ACL 2010 by Jing-Shing Chang and Philipp Koehn,
ACL 2008 by Johanna D. Moore, Simone Teufel, James Allan, and Sadaoki Furui,
ACL 2005 by Hwee Tou Ng and Kemal Oflazer,
ACL 2002 by Eugene Charniak and Dekang Lin,
and earlier ACL and EACL formats written by several people, including
John Chen, Henry S. Thompson and Donald Walker.
Additional elements were taken from the formatting instructions of the \emph{International Joint Conference on Artificial Intelligence} and the \emph{Conference on Computer Vision and Pattern Recognition}.
\section{Introduction}
The following instructions are directed to authors of papers submitted to ACL-IJCNLP 2021 or accepted for publication in its proceedings.
All authors are required to adhere to these specifications.
Authors are required to provide a Portable Document Format (PDF) version of their papers.
\textbf{The proceedings are designed for printing on A4 paper.}
\section{Electronically-available resources}
ACL provides this description and accompanying style files at
\begin{quote}
\url{https://2021.aclweb.org/files/acl-ijcnlp2021-templates.zip}
\end{quote}
We strongly recommend the use of these style files, which have been appropriately tailored for the ACL-IJCNLP 2021 proceedings. You can also start from the Overleaf template linked from the conference website.
\paragraph{\LaTeX-specific details:}
The templates include the \LaTeX2e{} source (\texttt{\small acl2021.tex}),
the \LaTeX2e{} style file used to format it (\texttt{\small acl2021.sty}),
an ACL bibliography style (\texttt{\small acl\_natbib.bst}),
an example bibliography (\texttt{\small acl2021.bib}),
and the bibliography for the ACL Anthology (\texttt{\small anthology.bib}).
\section{Length of Submission}
\label{sec:length}
The conference accepts submissions of long papers and short papers. Long papers may consist of up to eight (8) pages of content plus unlimited pages for references and ethics/broader impact statement. Short papers may consist of up to four (4) pages of content, plus unlimited pages for references and ethics/broader impact statement. Upon acceptance, final versions of both long and short papers will be given one additional content page to address the reviewers' comments.
For both long and short papers, all illustrations and tables that are part of the main text must be accommodated within these page limits, observing the formatting instructions given in the present document.
\textbf{Papers that do not conform to the specified length and formatting requirements are subject to be rejected without review.}
The conference encourages the submission of additional material that is relevant to the reviewers but not an integral part of the paper.
There are two such types of material: appendices, which can be read, and non-readable supplementary materials, often data or code.
Additional material must be submitted as separate files, and must adhere to the same anonymity guidelines as the main paper. Appendices should be no longer than 4 pages.
The paper must be self-contained: it is optional for reviewers to look at the supplementary material.
The paper should not refer, for further detail, to documents, code or data resources that are not available to the reviewers.
Refer to Section~\ref{sec:supplementary} for further information.
Workshop chairs may have different rules for allowed length and whether supplemental material is welcome.
As always, the respective call for papers is the authoritative source.
\section{Anonymity}
As reviewing will be double-blind, papers submitted for review should not include any author information (such as names or affiliations). Furthermore, self-references that reveal the author's identity, \emph{e.g.},
\begin{quote}
We previously showed \citep{Gusfield:97} \ldots
\end{quote}
should be avoided. Instead, use citations such as
\begin{quote}
\citet{Gusfield:97} previously showed\ldots
\end{quote}
Please do not use anonymous citations and do not include acknowledgements.
\textbf{Papers that do not conform to these requirements may be rejected without review.}
Any preliminary non-archival versions of submitted papers should be listed in the submission form but not in the review version of the paper.
Reviewers are generally aware that authors may present preliminary versions of their work in other venues, but will not be provided the list of previous presentations from the submission form.
Please do not include github links that reveal the authors' identities in your submission. If you feel it is important to include your source code, you can zip the code and submit it to softconf.
Once a paper has been accepted to the conference, the camera-ready version of the paper should include the author's names and affiliations, and is allowed to use self-references and provide the related github link.
\paragraph{\LaTeX-specific details:}
For an anonymized submission, ensure that {\small\verb|\aclfinalcopy|} at the top of this document is commented out, and that you have filled in the paper ID number (assigned during the submission process on softconf) where {\small\verb|***|} appears in the {\small\verb|\def\aclpaperid{***}|} definition at the top of this document.
For a camera-ready submission, ensure that {\small\verb|\aclfinalcopy|} at the top of this document is not commented out.
\section{Multiple Submission Policy}
ACL-IJCNLP 2021 will not consider any paper that is under review in a journal or another conference at the time of submission, and submitted papers must not be submitted elsewhere during the ACL-IJCNLP 2021 review period. This policy covers all refereed and archival conferences and workshops (e.g., NAACL). The only exception is that a paper can be dual-submitted to both ACL-IJCNLP 2021 and an ACL-IJCNLP workshop.
In addition, we will not consider any paper that overlaps significantly in content or results with papers that will be (or have been) published elsewhere.
Authors submitting more than one paper to ACL-IJCNLP 2021 must ensure that their submissions do not overlap significantly ($>25$\%) with each other in content or results.
\section{Formatting Instructions}
Manuscripts must be in two-column format.
Exceptions to the two-column format include the title, authors' names and complete addresses, which must be centered at the top of the first page, and any full-width figures or tables (see the guidelines in Section~\ref{ssec:title-authors}).
\textbf{Type single-spaced.}
Start all pages directly under the top margin.
The manuscript should be printed single-sided and its length should not exceed the maximum page limit described in Section~\ref{sec:length}.
Pages should be numbered in the version submitted for review, but \textbf{pages should not be numbered in the camera-ready version}.
\paragraph{\LaTeX-specific details:}
The style files will generate page numbers when {\small\verb|\aclfinalcopy|} is commented out, and remove them otherwise.
\subsection{File Format}
\label{sect:pdf}
For the production of the electronic manuscript you must use Adobe's Portable Document Format (PDF).
Please make sure that your PDF file includes all the necessary fonts (especially tree diagrams, symbols, and fonts with Asian characters).
When you print or create the PDF file, there is usually an option in your printer setup to include none, all or just non-standard fonts.
Please make sure that you select the option of including ALL the fonts.
\textbf{Before sending it, test your PDF by printing it from a computer different from the one where it was created.}
Moreover, some word processors may generate very large PDF files, where each page is rendered as an image.
Such images may reproduce poorly.
In this case, try alternative ways to obtain the PDF.
One way on some systems is to install a driver for a postscript printer, send your document to the printer specifying ``Output to a file'', then convert the file to PDF.
It is of utmost importance to specify the \textbf{A4 format} (21 cm x 29.7 cm) when formatting the paper.
Print-outs of the PDF file on A4 paper should be identical to the hardcopy version.
If you cannot meet the above requirements about the production of your electronic submission, please contact the publication chairs as soon as possible.
\paragraph{\LaTeX-specific details:}
PDF files are usually produced from \LaTeX{} using the \texttt{\small pdflatex} command.
If your version of \LaTeX{} produces Postscript files, \texttt{\small ps2pdf} or \texttt{\small dvipdf} can convert these to PDF.
To ensure A4 format in \LaTeX, use the command {\small\verb|\special{papersize=210mm,297mm}|}
in the \LaTeX{} preamble (below the {\small\verb|\usepackage|} commands) and use \texttt{\small dvipdf} and/or \texttt{\small pdflatex}; or specify \texttt{\small -t a4} when working with \texttt{\small dvips}.
\subsection{Layout}
\label{ssec:layout}
Format manuscripts two columns to a page, in the manner these
instructions are formatted.
The exact dimensions for a page on A4 paper are:
\begin{itemize}
\item Left and right margins: 2.5 cm
\item Top margin: 2.5 cm
\item Bottom margin: 2.5 cm
\item Column width: 7.7 cm
\item Column height: 24.7 cm
\item Gap between columns: 0.6 cm
\end{itemize}
\noindent Papers should not be submitted on any other paper size.
If you cannot meet the above requirements about the production of your electronic submission, please contact the publication chairs above as soon as possible.
\subsection{Fonts}
For reasons of uniformity, Adobe's \textbf{Times Roman} font should be used.
If Times Roman is unavailable, you may use Times New Roman or \textbf{Computer Modern Roman}.
Table~\ref{font-table} specifies what font sizes and styles must be used for each type of text in the manuscript.
\begin{table}
\centering
\begin{tabular}{lrl}
\hline \textbf{Type of Text} & \textbf{Font Size} & \textbf{Style} \\ \hline
paper title & 15 pt & bold \\
author names & 12 pt & bold \\
author affiliation & 12 pt & \\
the word ``Abstract'' & 12 pt & bold \\
section titles & 12 pt & bold \\
subsection titles & 11 pt & bold \\
document text & 11 pt &\\
captions & 10 pt & \\
abstract text & 10 pt & \\
bibliography & 10 pt & \\
footnotes & 9 pt & \\
\hline
\end{tabular}
\caption{\label{font-table} Font guide. }
\end{table}
\paragraph{\LaTeX-specific details:}
To use Times Roman in \LaTeX2e{}, put the following in the preamble:
\begin{quote}
\small
\begin{verbatim}
\usepackage{times}
\usepackage{latexsym}
\end{verbatim}
\end{quote}
\subsection{Ruler}
A printed ruler (line numbers in the left and right margins of the article) should be presented in the version submitted for review, so that reviewers may comment on particular lines in the paper without circumlocution.
The presence or absence of the ruler should not change the appearance of any other content on the page.
The camera ready copy should not contain a ruler.
\paragraph{Reviewers:}
note that the ruler measurements may not align well with lines in the paper -- this turns out to be very difficult to do well when the paper contains many figures and equations, and, when done, looks ugly.
In most cases one would expect that the approximate location will be adequate, although you can also use fractional references (\emph{e.g.}, this line ends at mark $295.5$).
\paragraph{\LaTeX-specific details:}
The style files will generate the ruler when {\small\verb|\aclfinalcopy|} is commented out, and remove it otherwise.
\subsection{Title and Authors}
\label{ssec:title-authors}
Center the title, author's name(s) and affiliation(s) across both columns.
Do not use footnotes for affiliations.
Place the title centered at the top of the first page, in a 15-point bold font.
Long titles should be typed on two lines without a blank line intervening.
Put the title 2.5 cm from the top of the page, followed by a blank line, then the author's names(s), and the affiliation on the following line.
Do not use only initials for given names (middle initials are allowed).
Do not format surnames in all capitals (\emph{e.g.}, use ``Mitchell'' not ``MITCHELL'').
Do not format title and section headings in all capitals except for proper names (such as ``BLEU'') that are
conventionally in all capitals.
The affiliation should contain the author's complete address, and if possible, an electronic mail address.
The title, author names and addresses should be completely identical to those entered to the electronical paper submission website in order to maintain the consistency of author information among all publications of the conference.
If they are different, the publication chairs may resolve the difference without consulting with you; so it is in your own interest to double-check that the information is consistent.
Start the body of the first page 7.5 cm from the top of the page.
\textbf{Even in the anonymous version of the paper, you should maintain space for names and addresses so that they will fit in the final (accepted) version.}
\subsection{Abstract}
Use two-column format when you begin the abstract.
Type the abstract at the beginning of the first column.
The width of the abstract text should be smaller than the
width of the columns for the text in the body of the paper by 0.6 cm on each side.
Center the word \textbf{Abstract} in a 12 point bold font above the body of the abstract.
The abstract should be a concise summary of the general thesis and conclusions of the paper.
It should be no longer than 200 words.
The abstract text should be in 10 point font.
\subsection{Text}
Begin typing the main body of the text immediately after the abstract, observing the two-column format as shown in the present document.
Indent 0.4 cm when starting a new paragraph.
\subsection{Sections}
Format section and subsection headings in the style shown on the present document.
Use numbered sections (Arabic numerals) to facilitate cross references.
Number subsections with the section number and the subsection number separated by a dot, in Arabic numerals.
\subsection{Footnotes}
Put footnotes at the bottom of the page and use 9 point font.
They may be numbered or referred to by asterisks or other symbols.\footnote{This is how a footnote should appear.}
Footnotes should be separated from the text by a line.\footnote{Note the line separating the footnotes from the text.}
\subsection{Graphics}
Place figures, tables, and photographs in the paper near where they are first discussed, rather than at the end, if possible.
Wide illustrations may run across both columns.
Color is allowed, but adhere to Section~\ref{ssec:accessibility}'s guidelines on accessibility.
\paragraph{Captions:}
Provide a caption for every illustration; number each one sequentially in the form:
``Figure 1. Caption of the Figure.''
``Table 1. Caption of the Table.''
Type the captions of the figures and tables below the body, using 10 point text.
Captions should be placed below illustrations.
Captions that are one line are centered (see Table~\ref{font-table}).
Captions longer than one line are left-aligned (see Table~\ref{tab:accents}).
\begin{table}
\centering
\begin{tabular}{lc}
\hline
\textbf{Command} & \textbf{Output}\\
\hline
\verb|{\"a}| & {\"a} \\
\verb|{\^e}| & {\^e} \\
\verb|{\`i}| & {\`i} \\
\verb|{\.I}| & {\.I} \\
\verb|{\o}| & {\o} \\
\verb|{\'u}| & {\'u} \\
\verb|{\aa}| & {\aa} \\\hline
\end{tabular}
\begin{tabular}{lc}
\hline
\textbf{Command} & \textbf{Output}\\
\hline
\verb|{\c c}| & {\c c} \\
\verb|{\u g}| & {\u g} \\
\verb|{\l}| & {\l} \\
\verb|{\~n}| & {\~n} \\
\verb|{\H o}| & {\H o} \\
\verb|{\v r}| & {\v r} \\
\verb|{\ss}| & {\ss} \\
\hline
\end{tabular}
\caption{Example commands for accented characters, to be used in, \emph{e.g.}, \BibTeX\ names.}\label{tab:accents}
\end{table}
\paragraph{\LaTeX-specific details:}
The style files are compatible with the caption and subcaption packages; do not add optional arguments.
\textbf{Do not override the default caption sizes.}
\subsection{Hyperlinks}
Within-document and external hyperlinks are indicated with Dark Blue text, Color Hex \#000099.
\subsection{Citations}
Citations within the text appear in parentheses as~\citep{Gusfield:97} or, if the author's name appears in the text itself, as \citet{Gusfield:97}.
Append lowercase letters to the year in cases of ambiguities.
Treat double authors as in~\citep{Aho:72}, but write as in~\citep{Chandra:81} when more than two authors are involved. Collapse multiple citations as in~\citep{Gusfield:97,Aho:72}.
Refrain from using full citations as sentence constituents.
Instead of
\begin{quote}
``\citep{Gusfield:97} showed that ...''
\end{quote}
write
\begin{quote}
``\citet{Gusfield:97} showed that ...''
\end{quote}
\begin{table*}
\centering
\begin{tabular}{lll}
\hline
\textbf{Output} & \textbf{natbib command} & \textbf{Old ACL-style command}\\
\hline
\citep{Gusfield:97} & \small\verb|\citep| & \small\verb|\cite| \\
\citealp{Gusfield:97} & \small\verb|\citealp| & no equivalent \\
\citet{Gusfield:97} & \small\verb|\citet| & \small\verb|\newcite| \\
\citeyearpar{Gusfield:97} & \small\verb|\citeyearpar| & \small\verb|\shortcite| \\
\hline
\end{tabular}
\caption{\label{citation-guide}
Citation commands supported by the style file.
The style is based on the natbib package and supports all natbib citation commands.
It also supports commands defined in previous ACL style files for compatibility.
}
\end{table*}
\paragraph{\LaTeX-specific details:}
Table~\ref{citation-guide} shows the syntax supported by the style files.
We encourage you to use the natbib styles.
You can use the command {\small\verb|\citet|} (cite in text) to get ``author (year)'' citations as in \citet{Gusfield:97}.
You can use the command {\small\verb|\citep|} (cite in parentheses) to get ``(author, year)'' citations as in \citep{Gusfield:97}.
You can use the command {\small\verb|\citealp|} (alternative cite without parentheses) to get ``author year'' citations (which is useful for using citations within parentheses, as in \citealp{Gusfield:97}).
\subsection{References}
Gather the full set of references together under the heading \textbf{References}; place the section before any Appendices.
Arrange the references alphabetically by first author, rather than by order of occurrence in the text.
Provide as complete a citation as possible, using a consistent format, such as the one for \emph{Computational Linguistics\/} or the one in the \emph{Publication Manual of the American
Psychological Association\/}~\citep{APA:83}.
Use full names for authors, not just initials.
Submissions should accurately reference prior and related work, including code and data.
If a piece of prior work appeared in multiple venues, the version that appeared in a refereed, archival venue should be referenced.
If multiple versions of a piece of prior work exist, the one used by the authors should be referenced.
Authors should not rely on automated citation indices to provide accurate references for prior and related work.
The following text cites various types of articles so that the references section of the present document will include them.
\begin{itemize}
\item Example article in journal: \citep{Ando2005}.
\item Example article in proceedings, with location: \citep{borschinger-johnson-2011-particle}.
\item Example article in proceedings, without location: \citep{andrew2007scalable}.
\item Example arxiv paper: \citep{rasooli-tetrault-2015}.
\end{itemize}
\paragraph{\LaTeX-specific details:}
The \LaTeX{} and Bib\TeX{} style files provided roughly follow the American Psychological Association format.
If your own bib file is named \texttt{\small acl2021.bib}, then placing the following before any appendices in your \LaTeX{} file will generate the references section for you:
\begin{quote}\small
\verb|\bibliographystyle{acl_natbib}|\\
\verb|
\section{Introduction}
Reviewing is at the center of the scientific publication process, and the quality of publications is dependent on it. In many scientific fields, including natural language processing and machine learning, submissions for publication are reviewed using a peer review system. Recently, these fields are seeing increasing volumes of submissions each year, especially in high reputation venues. This has created an issue of over-burdening of reviewers, which is not only a problem for the quality of life of scientists, but also consequently affects the quality of the reviews.
With ever increasing volume of new results in these fields, submissions for publication are expected to multiply still, and the problem is only expected to deepen, which is raising concerns in the scientific community \cite{rogers2020can}.
One avenue for ameliorating this problem is relying on artificial intelligence to assist with the process, in order to remove some of the burden from the human reviewers. A possibility would be to generate reviews or article summaries automatically, in order to speed up the human's understanding of the paper, or to assist with parts of the review writing, e.g., a few sentences summary.
Text generation has seen impressive improvements in recent years, being one of the most active fields in NLP, with the highest leaps in performance of newly published models. Models such as BERT \cite{devlin2019bert}, GPT-3 \cite{brown2020language} have shown impressive results for text generation, as well as for other tasks, acting as language models which can generalize for a wide range of tasks in NLP with only little fine-tuning.
Text summarization is a problem of text generation. Depending on the approach, summarization can be extractive \cite{zheng-lapata-2019-sentence} or abstractive \cite{see-etal-2017-get,nallapati-etal-2016-abstractive}. Extractive summarization is performed by selecting key sentences from the original text, while abstractive summarization tackles the more difficult problem of generating novel text that summarizes a given input---the problem we are interested in and explore in this paper.
As for text generation in general, state-of-the-art models for summarization are generally neural and transformer-based such as PEGASUS \cite{zhang2020pegasus} and Prophet \cite{qi2020prophetnet}. These models have been used for text summarization for different domains, including news \cite{desai2020compressive} and scientific texts. For scientific text summarization, \citet{zhang2020pegasus} have obtained best results in existing literature, based on evaluation on a dataset of articles published on arXiv and PubMed using papers' abstracts as ground truth.
Scientific texts pose specific problems for summarization, given their particular structure and way of organizing information. This is why the problem of scientific text summarization has been approached separately from general summarization systems.
The problem of scientific text summarization has been approached before \cite{yasunaga2019scisummnet,altmami2020automatic,ju2020scisummpip,Cohan2017ContextualizingCF,qazvinian2010citation}. Top conferences in NLP have organized workshops on scholarly document processing, including shared tasks specifically focused on scientific document summarization \cite{chandrasekaran2019overview}. Most approaches for scientific text summarization use an extractive \cite{saggion2000selective,saggion2011learning,yang2016amplifying,slamet2018automated,agrawal2019scalable,hoang2010towards} or citation-based approach \cite{Cohan2017ContextualizingCF,qazvinian2010citation,ronzano2016empirical}, with a few exceptions attempting abstractive summarization on scientific texts \cite{lloret2013compendium}. Notably, \citet{ju2020scisummpip} use a combined extractive and abstractive approach based on BERT. \citet{sun2018summarization} propose an approach based on semantic link networks for summarizing scientific texts. A recently published survey \cite{altmami2020automatic} contains a more exhaustive overview of previous attempts at summarizing scientific papers.
Given the excellent results of recent text generation models, it is promising to consider new applications in fields where they have not been leveraged in practice before.
We propose that one such task is {\em scientific review summary generation}. We evaluate in this paper the feasibility of automatically generating review summaries for scientific papers. We use state-of-the-art models for text summarization, and apply them to our problem. We release a dataset of articles and reviews from NeurIPS, which we use to assess the performance of automatic summarization models for the problem of review summary generation.
\section{Dataset}
We build a dataset of articles and associated reviews by scraping NeurIPS's conference website,\footnote{https://papers.neurips.cc} and collecting all articles published in NeurIPS between 2013 and 2020, along with their reviews. To obtain the full text of the papers, we downloaded the PDFs from the website and extracted the text using Grobid.\footnote{https://github.com/kermitt2/grobid} Reviews were extracted directly from the HTML content of the web pages, and, where needed, heuristics were used in order to exclude the texts of the author's responses. Each article can have several reviews. Table \ref{tab:dataset} summarizes statistics about the dataset.
\begin{table}[]
\centering
\small
\begin{tabular}{r|l}
\textbf{Articles} & 5,950 \\
\textbf{Reviews} & 18,926 \\
\textbf{Avg review len (words)} & 399 \\
\textbf{Avg review len (sentences)} & 21 \\
\textbf{Avg abstract len (words)} & 159 \\
\textbf{Avg abstract len (sentences)} & 7 \\
\end{tabular}
\caption{Dataset statistics.}
\vspace{-3mm}
\label{tab:dataset}
\end{table}
\section{Summarization Experiments}
Reviews of scientific articles are usually comprised of a short summary, followed by the comments comprising the reviewer's evaluation of the article, mentioning its strengths and its weaknesses. The initial summary of the paper is usually a short objective description of its contents, so in theory it could be inferred solely based on the article's content. Based on this premise, we formulate the problem of automatic review generation as a text summarization problem.
\begin{table*}[ht]
\centering
\small
\begin{tabular}{r|c|c|c|c}
& \textbf{R-1} & \textbf{R-2} & \textbf{R-L} & \textbf{BERTScore} \\
\hline
\hline
\textbf{vs. arXiv abstracts \cite{zhang2020pegasus}} & .447 & .173 & .258 & - \\
\hline
\textbf{vs. abstract (NeurIPS)} & .236 & .046 & .151 & .793 \\
\hline
\textbf{vs. review summaries (individual whole)} & .169 & .023 & .117 & .789 \\
\textbf{vs. review summaries (concatenated whole)} & .206 & .033 & .127 & .784
\end{tabular}
\caption{Performance of pretrained model}
\label{tab:eval_pretrain}
\end{table*}
\textbf{Pre-processing.} We aim to separate the two different parts of each review: the initial part containing a short summary of the paper, from the following comments and evaluation of the paper. A manual inspection of extracted reviews in our dataset for papers up to 2019 shows that many reviews include replies to author responses from the rebuttal phase of the review, and these can be found either in the beginning or end of the review, without a consistent pattern, sometimes separated from the main review by ASCII separators (strings of "-"/"="/"~"). We then rely on heuristics in order to correctly extract the summary part of the review, by searching the review text for keywords such as "rebuttal" or "response": if these are found at the beginning of the review, we then look for ASCII separator characters, and consider the original review to begin after the separator; otherwise, we assume the summary is found at the beginning of the review. For papers from NeurIPS 2020, the different sections of the review are clearly marked (\textit{summary}, \textit{strengths}, \textit{weaknesses}, \textit{clarity} and \textit{correctness}), so this pre-processing step was not needed. After this step, we split the obtained text into sentences and select the first $k$ sentences as the summary. Our motivation in doing so was driven by several works on extreme classification \cite{DBLP:journals/corr/abs-1907-08722,narayan-etal-2018-dont} aimed at generating short, one-sentence news summary to answer the question: {``}What is the article about?{''}.
\begin{table*}[ht]
\centering
\small
\begin{tabular}{r|c|c|c|c}
& \textbf{R-1} & \textbf{R-2} & \textbf{R-L} & \textbf{BERTScore} \\
\hline
\hline
\textbf{vs. abstract (NeurIPS)} & .261 & .034 & .141 & .812 \\
\hline
\textbf{vs. review summaries (indiviual whole)} & .230 & .031 & .148 & .817\\
\textbf{vs. review summaries (concatenated whole)} & .254 & .046 & .145 & .806 \\
\hline
\textbf{vs. review summaries (concatenated 5 sents)} & .273 & .047 & .155 & .808 \\
\textbf{vs. review summaries (concatenated 4 sents)} & .279 & .046 & .158 & .810 \\
\textbf{vs. review summaries (concatenated 3 sents)} & .287 & .045 & .164 & .813\\
\textbf{vs. review summaries (concatenated 2 sents)} & .290 & .042 & .170 & .817 \\
\textbf{vs. review summaries (concatenated 1 sent)} & .246 & .032 & .160 & .821\\
\hline
\textbf{vs. review summaries (individual 5 sents)} & .227 & .030 & .149 & .818 \\
\textbf{vs. review summaries (individual 4 sents)} & .220 & .028 & .147 & .819 \\
\textbf{vs. review summaries (individual 3 sents)} & .207 & .026 & .117 & .819 \\
\textbf{vs. review summaries (individual 2 sents)} & .176 & .022 & .127 & .820 \\
\textbf{vs. review summaries (individual 1 sent)} & ,114 & .053 & .091 & .822 \\
\end{tabular}
\caption{Performance of fine-tuned model on abstract and review summary}
\label{tab:eval_finetune}
\end{table*}
\textbf{Model.} Language modeling in NLP has recently seen great advancements, and is one of the most active areas of research in NLP, with new results being published every few months. The best performing models are based on neural architectures, among which transformers play an important role. Text summarization in particular is a type of text generation problem; the current state of the art in text generation is PEGASUS \cite{zhang2020pegasus}, which is a transformers-based model trained to generate summaries by masking important sentences in a source text. PEGASUS obtained state-of-the-art results in text summarization across 12 different datasets in different domains, including scientific texts.
We experiment with using PEGASUS in order to generate summaries of scientific articles in our dataset, and assess its performance compared to the collected reviews.
\textbf{Model pre-trained on abstracts.} We first experiment with a pre-trained version of PEGASUS for scientific text summarization, which was trained to generate abstracts of scientific texts based on a dataset of arXiv articles \cite{cohan2018discourse}.
In order to ensure no overlap between the test set used for evaluation in our experiments and the articles in the arXiv database used in pre-training of the model, we select as our test set only the articles in our dataset published in 2020 (the arXiv dataset was published in 2018) - we use 1000 of these articles as our test set and keep the rest of 898 as a validation set. The 2020 reviews are also the highest-quality of our dataset, since the summary section of the review is clearly marked and used as is for evaluation (as opposed to extracted based on heuristics).
\textbf{Model fine-tuned on reviews.} Second, we attempt to generate paper summaries which best approximate a review. For this purpose, we fine-tune the pre-trained model used in the previous experiment on our own data, using as targets the reviews in our dataset.
As a training set, we use the articles and reviews in our dataset published before 2020. While our dataset is smaller than the arXiv dataset used for the pre-trained model, it is expected to be similar to the original training data. For each article, one review is selected at random and used as ground truth for training the summarization model. The training set contains 4,052 papers and their reviews.
\textbf{Evaluation.} We evaluate the models using the ROUGE metric, and compare the generated summaries both to the abstract and the reviews. We report ROUGE-1, ROUGE-2 and ROUGE-L, as well as BERTScore, using the RoBERTa-large model\footnote{\tt roberta-large\_L17\_no-idf\_version=0.3.9 (hug\_trans=4.2.2)} \cite{zhang2019bertscore}. Our setup
can be evaluated on multiple labels for the same input text: in our test set, one paper can have several reviews. We evaluate our models with multiple labels: first by considering them separately as independent examples, and second by concatenating all reviews for a given input article into one single reference text, and evaluating against it.
We show examples of generated reviews using our model, along with the original reviews for the same article, in the Appendix.
\textbf{Results.} We report separately the results of the pre-trained and the fine-tuned model. We compare different setups, using as target texts both the abstracts and the reviews. In the case of the reviews, we consider separately as a target test the whole review or only the summary section, varying the number of extracted sentences from 1 to 5, and experiment with the two evaluation setups: concatenating the different reviews corresponding to one article, or considering them as separate test examples.
Tables \ref{tab:eval_pretrain} and \ref{tab:eval_finetune} and show the results for all setups.
The pre-trained model obtains better results when evaluated against abstracts than against reviews, across configurations and metrics.
Although the pre-trained model was trained to generate abstracts, the fine-tuned model still obtains slightly better results compared to abstracts, suggesting it might solve a relevant domain adaptation aspect.
The fine-tuned model also shows improved results for review summary generation.
In terms of ROUGE scores, the optimal number of sentences of the summary extracted from the review summary seems to be 2 in the concatenated setup, while in the individual setup, the performance increases with the number of sentences. BERTScore strictly decreases with the number of sentences for both setups. Especially in the concatenated setup, using the first 1-2 sentences in the review summary as labels out-performs evaluating against the full review summary, suggesting that the generated summaries generally contain information present in the beginning of the review.
\begin{table}[bt]
\centering
\small
\resizebox{.45\textwidth}{!}{%
\begin{tabular}{r|c|c|c|c}
& \textbf{R-1} & \textbf{R-2} & \textbf{R-L} & \textbf{BERT} \\
& & & & \textbf{Score} \\
\hline
\hline
\textbf{vs. full review (concat)} & .152 & .036 & .092 & .803 \\
\textbf{vs. full review (individual)} & .241 & .040 & .139 & .806 \\
\hline
\textbf{vs. strenghts (concat)} & .270 & .039 & .159 & .815 \\
\textbf{vs. strengths (individual) }& .200 & .038 & .135 & .820 \\
\textbf{vs. weaknesses (concat)} & .232 & .028 & .134 & .803 \\
\textbf{ vs. weaknesses (individual) }& .212 & .027 & .134 & .808 \\
\end{tabular}
}
\caption{Performance of fine-tuned model on full review and other review sections}
\label{tab:eval_finetune_full}
\end{table}
\subsection{Feasibility of Generating Full Reviews}
The fine-tuned model is better at generating review summaries than the pre-trained model, across setups.
The generation of a full review, including critical interpretations from the reviewers, is a much more challenging problem than generating paper summaries. In order to assess how well a summarization model can approximate a full review, including not only the summary, but also the critical comments sections, we separately evaluate our model using the full reviews as targets, as well as against the separate sections (we consider the \textit{Strengths} and \textit{Weaknesses} sections), as show in Table \ref{tab:eval_finetune_full}. We notice that the performance is generally lower than for the review summary, but still comparable. The \textit{Strengths} section seems to have the most in common with the review summary according the better results.
\section{Conclusions}
We have formulated the problem of scientific text review generation, as a novel task in NLP with practical applications for the scientific community. Review generation is related to the text summarization task, but has its own specific features, which is what makes it a difficult problem to solve. We have taken the first steps towards building an automatic system for review generation; and have collected and are releasing a dataset of scientific articles and reviews which can be used for future experimentation into the topic.
We conclude that scientific review generation is a difficult problem, with current performance considerably below that of state-of-the-art text generation models on scientific abstracts. Nevertheless, the small improvements in performance we obtain through fine-tuning the model suggest that the problem might be approachable, and encourage us to continue to study it. We propose that more training data could be useful to obtain better results, as would a more accurate extraction of the summary section of the review.
In the future, we would like to explore a more complex training strategy in order to improve performance, such as multi-task learning (to jointly train the model to generate reviews and abstracts), or conditional text generation, in order to constrain the model to generate review-like texts, while keeping the content relevant to the source article.
\section{Ethical Considerations}
Our dataset poses no privacy issues.
With regards to the task of paper review generation, it is unclear if generating reviews entirely automatically is desirable from a practical as well as ethical perspective. Instead, we approach the problem summary generation for reviews, in view of a possible computer-assisted process for review generation, which would not exclude humans. We think a computational tool for assisting with the ever-growing burden of reviewing can help the community and eventually lead to higher quality reviews, and hope our paper can encourage discussion on the topic. We leave open to discussion the question of how such a tool could best be integrated in the current review system.
\bibliographystyle{acl-ijcnlp2021-templates/acl_natbib}
|
\section{Introduction} \label{sec:intro}
The majority of galaxies in the Universe exist in groups, where the dark matter halos cover mass ranges of $10^{12} \la \rm{M_{halo}} \la 10^{14.5}$~$\rm{M_{\odot}}$ \citep{Tully1987}. The diffuse, hot gas gravitationally bound to the group is commonly referred to as the intragroup medium (IGrM) and may constitute a significant entry into the missing baryon problem \citep{Persic1992, Fukugita2006, Spergel2007}. The effect on galaxy evolution of the IGrM and the halos of groups remains uncertain. The gas in galaxy group halos can be characterized through X-rays, the Sunyaev-Zel'dovich (SZ) effect, and through UV absorption lines from background quasars (QSOs).
Early IGrM\ detections were based on \emph{ROSAT} observations of high mass, elliptical rich groups \citep{Mulchaey1996Xray, Helsdon2000, Mulchaey2000}. These groups were believed to be more massive than spiral rich groups and hence more luminous in X-rays. From these observations, initial scaling relations \citep{Helsdon2000} were derived and the mass of the hot gas was determined to be comparable to the stellar mass of the galaxies \citep{Mulchaey1996Xray}. More recently, \cite{Bregman07} studied \ion{O}{7} absorption to distinguish it's origin as from the Milky Way's galactic halo or from the Local Group's IGrM. In searching for the IGrM, they found that the Milky Way halo models were preferred, but a contribution to the \ion{O}{7} absorption from the IGrM could not be conclusively ruled out.
The thermal SZ effect, where cosmic microwave background (CMB) photons are scattered as a result of energetic, free electrons, provides an alternative means of observing the diffuse gas bound to dark matter halos. While the SZ effect is typically used to analyze galaxy clusters, recent studies using large stacks have led to detections around galaxy groups and individual galaxies \citep[and references therein]{Greco2015, Vikram2017, Bregman2018, Pratt2020, Tanimura2020}. As noted in \cite{LeBrun2015, Tumlinson2017} and \cite{Tanimura2020}, the gas content of galaxy halos down to $10^{11}$$\rm{M_{\odot}}$\ comes into tension with existing X-ray observations as the self-similar scaling relations appear to fail.
However, \cite{LeBrun2015} proposes that the discrepancy may result from the low resolution of the Plank SZ map and therefore might not be as robust at low radii ($r \la r_{500}$) when compared to X-ray observations. This effect was reproduced using X-ray simulations were convolved with the Plank beam \citep{LeBrun2015}. Cosmological ``zoom in'' simulations by \cite{VandeVoort2016} find that hot gas near the virial temperature causes more consistent X-ray luminosity scaling relations for halos with $\rm{M_{halo}} \ga 10^{13}$~$\rm{M_{\odot}}$, while less massive halos show X-ray luminosities that are more strongly affected by star formation feedback.
Ultraviolet (UV) absorption lines observed in the spectra of background QSOs remain one of the most robust methods to probe gas at intermediate temperatures, where the gas is not hot enough for X-ray emission. QSO absorption lines (QALs) have shown that a significant amount of baryons lie in the diffuse gas that makes up the intergalactic medium (IGM) \citep{Rauch1998, Shull2012}. This provides means of probing the composition of the IGrM since the large majority of galaxy groups are lower in mass and do not have the temperature and density necessary for X-ray emission. At the virial temperature of typical galaxy groups, \cite{Mulchaey1996} predicted the existence of broad, shallow \ion{O}{6}\ absorption with Lyman series transitions without lower ions such as \ion{C}{4} and \ion{N}{5}\ based upon collisional ionization equilibrium (CIE) models. In this scenario, \ion{C}{4} and \ion{N}{5}\ are present at levels not currently detectable with current instruments such as the Cosmic Origins Spectrograph (COS; \cite{Green2012}) aboard the \emph{Hubble Space Telescope}.
With this background, studies by \cite{Tripp2000, TrippSavage2000} and \cite{Stocke2006} used background quasars to search for \ion{O}{6}, but the data were inconclusive in correlating \ion{O}{6}\ absorption with the IGrM. \cite{Stocke2014} conducted redshift surveys around 14 previously detected broad Ly$\alpha$\ and \ion{O}{6}\ detections, which were indicative of gas above 10$^5$~K. They found galaxy groups around these QSO sightlines and concluded with 2$\sigma$ confidence that these absorbers were due to the group environment and not the nearest galaxy to the sightline. The possibility that the \ion{O}{6}\ detections were tracing cooler clouds rather than the hot component of the IGrM\ was still a hypothesis and it lacked any direct observational confirmation. Other \ion{O}{6}\ studies by \cite{Pointon2017} and \cite{Stocke2017} compared the detections in group environments to the circumgalactic medium (CGM) of isolated galaxies. These studies found that group environments contained \ion{O}{6}\ absorption that could be modeled with broader components than isolated systems and concluded that \ion{O}{6}\ was characteristic of the boundary between cooler CGM gas and the hotter IGrM.
Studies by \cite{Tripp2008,Savage10,Savage12,Savage14} and \cite{Rosenwasser2018} detected \ion{O}{6}\ absorption features that are consistent with multiphase gas at both cooler and hotter temperatures that could be produced by photoionization or collisional ionization, respectively.
However, from these studies, it is difficult to distinguish the origin of \ion{O}{6}\ as being due to the boundary of multiphase gas in the IGrM\ or resulting from the CGM of member galaxies. The COS-Halos Survey \citep{Tumlinson2011,Tumlinson2013,Peeples2014,Werk2014,Werk16} analyzed spectra of 44 QSO-galaxy pairs and found \ion{O}{6}\ in addition to a significant amount of metals in the halos of isolated galaxies. The COS-Halos Survey found a strong correlation of \ion{O}{6}\ detections in the inner CGM of star forming galaxies leading to the idea that it may originate from large streams of cooling gas or from the hotter component of the CGM if a temperature gradient is assumed as opposed to a uniform halo at the virial temperature \citep{Werk2014, McQuinn2018}.
\cite{Heckman2002} and \cite{Bordoloi2017} (and references therein) show that \ion{O}{6}\ observed in the intergalactic medium, CGM of galaxies, and the Milky Way halo can be explained by radiative cooling models. These models agree with observations and simulations showing complex, multiphase structures at the interfaces between hot and cold gas \citep{Oppenheimer2009,Churchill2012,Pachat2016, Narayanan2018,Ahoranta2020}.
Recently, \cite{Stocke2019} carried out a survey of 12 galaxy groups paired with background QSOs to look for \ion{O}{6}\ associated with galaxy groups. They find that \ion{O}{6}\ was not uniformly detected within the sample, leading to the idea that CGM-like clouds can escape individual galaxies and can be observed within the group. They do not find evidence that these clouds can easily escape the group, which means that galaxy groups might be ``closed-boxes'' for galaxy evolution. Lastly, they conclude that the gas traced through \ion{O}{6}\ is not volume filling and that a hotter component is necessary for a complete baryon census in the group environment.
Here we present the COS-IGrM survey, designed to probe the IGrM of lower mass groups than those probed by \cite{Stocke2019}, where \ion{O}{6}\ could be a better tracer of the IGrM\ due to lower virial temperatures. The COS-IGrM sample consists of 18 galaxy groups paired with background UV bright quasars (QSOs) and was selected without bias towards predefined sightlines with \ion{O}{6}\ detections. This is the largest sample of low redshift ($z_{gp} \le 0.2$) galaxy groups ever probed for \ion{O}{6}\ associated with the IGrM.
This paper is organized as follows: in \S\ref{sec:sample} we describe the COS-IGrM\ sample, \S\ref{sec:observations} details the HST/COS observations along with the data reduction and analysis, \S\ref{sec:results} presents the results of the survey, \S\ref{sec:discussion} discusses the overall significance of our results and \S\ref{sec:conclusion} presents the conclusions of our survey.
\input{Table1.tex}
\begin{figure}
\centering
\plotone{./Fig1.pdf}
\caption{Environment of galaxy group J0841+1406. The color of the points represent the velocity of the member galaxies relative to the center of the group. The thick, dashed line represents the virial radius of the group and the thin, dotted lines represent the virial radii of the member galaxies.}
\label{fig:env}
\end{figure}
\section{Sample} \label{sec:sample}
The COS-IGrM sample\footnote{The data is available at MAST: \dataset[10.17909/t9-wqg9-9043]{\doi{10.17909/t9-wqg9-9043}}} is composed of 18 galaxy groups\footnote{In the HST proposal, there were 19 sightlines; however, one sightline had an insufficient signal to noise to use for this analysis.} each with a background QSO with a GALEX far ultraviolet (FUV) magnitude brighter than 19. The sample was created by cross referencing the \cite{Tago2010} galaxy group catalog with the catalog of unique GALEX Data Release 5 QSOs \citep{Bianchi2011}. Four additional criteria were implemented to create a robust sample from the \cite{Tago2010} group catalog:
\begin{enumerate}
\item The groups must have at least three spectroscopically confirmed members;
\vspace{-0.25cm}
\item The group redshifts must be between $0.075 \le z_{gp} \le 0.2$ for \ion{O}{6}\ to be within the COS bandpass;
\vspace{-0.25cm}
\item The QSO redshift, $z_{QSO} > z_{gp} + 0.1$ to eliminate confusion between absorption features from the group and from the QSO;
\vspace{-0.25cm}
\item The QSO impact parameter must be less than 1.5 times the group's virial radius ($1.5\rm{R_{vir}}$).
\end{enumerate}
In order to ensure that the groups in the COS-IGrM sample are physical groups, the \cite{Yang2007} galaxy group catalog was used to look for confirmed groups in the same location and with similar halo mass as provided by the \cite{Tago2010} catalog. This additional check aided in confidently identifying galaxy groups with as little as three spectroscopically confirmed members. The environment of one group in our sample is shown in Figure \ref{fig:env} along with the location of the background QSO. The remaining 17 group environments are listed in Appendix \ref{ap:env}. Since galaxy groups were identified from Sloan Digital Sky Survey (SDSS) spectroscopic group catalogs, our sample is biased towards groups with luminous galaxies, $L \ga L_*$.
We also include sightlines from \cite{Stocke2019} in our analysis to extend the sample to larger halo masses. In order to make sample parameters consistent with those from \cite{Stocke2019}, we re-defined the group parameters such as the group halo mass, virial radius, and velocity dispersion through the following relations from their paper:
\begin{equation}
M_{gp} = 310\times \left( \frac{L_{gp}}{L_*}\right) \times 10^{10} ~M_{\odot
\end{equation}
Where $L_{gp}$ is total the r-band luminosity of the group members calculated via the r-band magnitudes from the \cite{Tago2010} catalog.
\begin{equation}
R_{vir} = 957\times \left(\frac{M_{gp}}{10^{14}}\right)^{1/3} \rm{kpc}
\end{equation}
and
\begin{equation}
\sigma _{gp} = 387\times \left( \frac{M_{gp}}{10^{14}M_{\odot} } \right)^{1/3}~\rm{kms^{-1}}
\end{equation}
In Equation 2, the virial radius is defined as the limit where the overdensity of the medium is equal to 200$\rho_{crit}$ as described in \cite{Shull2012} and \cite{Stocke2019}. The full properties of our galaxy group sample and each corresponding background QSO are listed in Table \ref{tab:sample} along with the adopted values for the halo mass, virial radius, and velocity dispersion.
While each sightline probes the IGrM, some also pass through within the CGM of member galaxies. Therefore, we divide our sample into two sub-samples - one with sightlines passing through the CGM of member galaxies and the other where the sightline is at impact parameters larger than the viral radius of the member galaxies (assuming an isolated halo). This assumption may over-estimate the size of the CGM of group members; however, it remains the most reliable radius estimate without requiring extensive cosmological simulations. The first group contains six sightlines and are referred to as the CGM~+~IGrM. The remaining 12 sightlines fall in the latter category. In the absence of a deeper spectroscopic survey, our limiting magnitude allows us to claim that the ``pure" IGrM sightlines do not go through the CGM of any $\sim$L$_*$ galaxies. While it is possible that occasionally a much smaller galaxy could be close to the QSO sightline, care was taken to identify any possible galaxy candidates at the same redshift near the QSO sightline. Statistically, we do not find a significant number of possible member galaxies. This was followed up by recent multi object spectroscopy of two of the fields with the MMT and the Gemini Observatory, which confirmed this result , i.e., very few new galaxies were detected to be part of the groups and none very close to the QSO. These results will be discussed further in a future paper (McCabe et al. in prep).
In Table \ref{tab:sample}, we refer to the sub-grouping for each sightline/group as either IGrM\ or CGM~+~IGrM depending on the location of the QSO sightline.
\section{HST/COS Observations} \label{sec:observations}
The 18 QSOs in our sample were observed with the G130M grating of COS aboard the \textit{Hubble Space Telescope}. Data for 16 of the QSO sightlines were obtained under program 13314, while the remaining were obtained through archival data. The observations were designed to achieved a signal to noise (S/N) greater than 10 per resolution element for each sightline. This resolution is necessary in order to observe broad and shallow absorption lines that are expected to be associated with hot media. The spectra covered a observed frame wavelength from 1070$-$1465 \AA\ corresponding to a rest-frame wavelength range of 946$-$1295 \AA\ for the median redshift of the sample of $z$=0.1311.
The QSO spectra were created by coadding the individual exposures. The spectra were binned by three pixels, corresponding to half the resolution element, to enhance the S/N before any analysis was performed. As the first step, we identified the absorption lines associated with the groups. To do so, absorption lines in the entire spectra were identified. This is critical to ensure that the absorption lines associated with the target system are not blended with intervening or Milky Way absorbers. Special care has been given to the identification of metal-line species, in particular to \ion{O}{6}. Fortunately, the redshift range of our sample resulted in the observed wavelength of \ion{O}{6}\ shortward of 1215~$\rm \AA$, thus eliminating the possibility of any contamination from weak Ly$\alpha$\ absorbers in the IGM.
Continua were fit through an automated pipeline created and described by \cite{Tumlinson2011} and \cite{Werk2012}. A few sightlines exhibited complicated continua where the automated system failed and, as a result, those data were reduced individually by the authors following the prescriptions in \cite{Sembach1992} and \cite{Sembach2004}. The continuum fitting was done in a way consistent with the automated system.
Our data covered absorption lines from species such as Ly$\alpha$\ $\lambda$1215, Ly$\beta$\ $\lambda$1025, \ion{C}{2}\ $\lambda$1036, \ion{N}{5}\ $\lambda$, \ion{Si}{2}\ $\lambda$1190 $\lambda$1193 $\lambda$1260, \ion{Si}{3}\ $\lambda$1206 and \ion{O}{6}\ $\lambda\lambda$1031 $\lambda\lambda$1037, which trace gas from $10^{2-6}$~K. Absorption features were searched within $\pm$800 $\rm{kms^{-1}}$ of the group's systematic redshift. Features beyond this range were not considered to be physically related to the group.
Features with an equivalent width greater than 3$\sigma$ were considered detections; otherwise, a 3$\sigma$ upper limit was estimated. The uncertainty corresponding to each equivalent width measurement was determined through the RMS noise of the data within the measurement window. Each feature was fit with a Voigt profile in order to determine the column density, doppler `$b$' parameter, and velocity centroid. For sightlines with multiple absorption systems, we determined the total column density by linearly adding up the components. Unless stated otherwise, the column density represented in the figures refers to the total column density along the line of sight.
\begin{figure}[]
\centering
\plotone{Fig2.pdf}
\caption{Ly$\alpha$\ column density as a function of group halo mass. The COS-IGrM\ sightlines are shown as filled circles and the data from \cite{Stocke2019} are are grey squares. The COS-IGrM\ sample is further split into sightlines that are expected to probe the CGM and the IGrM of galaxy groups (blue data points) and those that probe just the IGrM\ of the group (green data points).}
\label{fig:lyamass}
\end{figure}
\begin{figure*}[]
\centering
\plottwo{Fig3a.pdf}{Fig3b.pdf}
\caption{Ly$\alpha$\ column densities as a function of projected QSO impact parameter (left) and the QSO impact parameter normalized by the group's virial radius (right). The colors are the same as for Figure \ref{fig:lyamass}. }
\label{fig:lyarho}
\end{figure*}
\section{Results} \label{sec:results}
We detected Ly$\alpha$, Ly$\beta$, \ion{C}{2}, \ion{N}{5}, \ion{Si}{2}, \ion{Si}{3}\ and \ion{O}{6}\ throughout the 18 sightlines. The detection rate of Ly$\alpha$\ is the highest at 67$\pm$5\% (12/18) followed by \ion{O}{6}\ at 44$\pm$5\% (8/18). We also detected low ionization species such as \ion{Si}{2}\ and \ion{C}{2}\ at detection rates of 6$\pm$5\% (1/18) and 28$\pm$5\% (5/18) respectively; the intermediate species \ion{Si}{3}\ at 28$\pm$5\% (5/18); and and high ionization \ion{N}{5}\ at 11$\pm$5\% (2/18).
Table \ref{tab:lines} presents measurements for each of these species for the COS-IGrM\ sample\footnote{Table 3 is published in machine-readable format in the online journal.}. In the following subsections, we discuss and analyze the properties and distribution of each species. In our analysis, we also include data from the \cite{Stocke2019} IGrM survey, which covered higher mass groups. This allows us to search for trends over a larger range of halo masses and group sizes. One significant difference to note between the two surveys is that unlike the \citeauthor{Stocke2019} sample, we do not eliminate QSO sightlines that fall within 0.25~$\rm{R_{gp}}$ in our sample selection.
\subsection{\ion{H}{1} Ly$\alpha$\ Absorption} \label{sec:lya}
We detected Ly$\alpha$\ absorption features in 12 of our 18 galaxy groups, with four groups having accompanying Ly$\beta$. Figure \ref{fig:lyamass} presents our \ion{H}{1} column density measurements as a function of the halo mass of the group. We found that lower mass halos exhibit a slightly narrower range of Ly$\alpha$\ column densities compared to higher mass halos. We see no evidence of varying column densities of Ly$\alpha$\ absorption between CGM~+~IGrM and IGrM sightlines. We find that there appears to be two main populations of data points: one with moderate column densities, log[N(\ion{H}{1})]$\sim$ 14.5-15, and another set clustered around log[N(\ion{H}{1})]$\sim$ 13. These groupings may indicate that the QSO sightlines are passing through patchy, non-uniform \ion{H}{1} clouds as opposed to a continuous distribution with a decreasing density gradient.
One sightline, J1127+2654 (Figure 19), was seen to have saturated Ly$\alpha$\ and Ly$\beta$\ absorption. The column density of this absorption feature should be treated as a lower limit due to the absorption line occupying the flat regime of the curve of growth. This sightline probes the IGrM as well as the CGM of the closest galaxy to the sightline, which is at $\sim$70~$\rm{kms^{-1}}$\ from the group's systematic velocity and an impact parameter of $\sim$119~kpc. This saturated \ion{H}{1} feature is composed of three components centered at -59, 32, and 116~$\rm{kms^{-1}}$, respectively from the systemic velocity of the group, with the middle component being the strongest. The low impact parameter of $\sim$119~kpc from the closest galaxy to the sightline suggests that we are likely probing the CGM of this galaxy. The location of the QSO with respect to the group member in the full environment plot in Appendix \ref{ap:env}.
Figure \ref{fig:lyarho} shows the distribution of Ly$\alpha$\ column density as a function of impact parameter ($\rho_{\rm QSO}$) in the left panel and as a function of normalized impact parameter in the right panel ($\rho_ {\rm QSO}/\rm R_{vir}$). We overplot the \cite{Stocke2019} sample as grey squares in Figure \ref{fig:lyarho}. We find no statistically significant correlation between the column density of Ly$\alpha$\ absorbers and the QSO impact parameter using the Kendall's Tau correlation test provided in the \textsc{Astronomy SURvival Analysis} (ASURV) package \citep{ASURV1,ASURV2,ASURV3}. Using the ASURV Kendall's Tau test, we observed no correlation between the column density of Ly$\alpha$\ absorption and IGrM or CGM~+~IGrM sightlines. Most of the stronger Ly$\alpha$\ absorbers (log[N(\ion{H}{1})]$\sim$ 14.5-15) are seen in sightlines that pass through only the IGrM.
The origins of cooler, partially neutral gas are not well understood. Possible scenarios include remnants of tidally stripped structures \citep{Davis1997,Bekki2009, Borthakur10, Nestor11, Gauthier13,Fossati2019, chen19, Peroux19}, in-situ condensation \citep{Voit19}, outflowing material from star forming galaxies \citep{Veilleux2005, Tripp2011, Nielsen18, frye19} and/or cold gas accretion from the intergalactic medium \citep{Keres2005, vogt15, Bielby17, borthakur19}. These processes are all capable of producing strong Ly$\alpha$\ absorption. On the other hand, lower column density absorbers and non-detections are seen in sightlines irrespective of whether they probe the CGM or just the IGrM. The presence of weak Ly$\alpha$\ absorbers ($\le 10^{14}~\rm{cm}^2$), including several non-detections, indicates that the sightlines pass through an ionized medium.
In those sightlines, we do not find \ion{O}{6} or \ion{N}{5}, indicating that the medium must be at temperatures greater than 10$^6$~K, assuming collisional ionization equilibrium.
This phenomena might be related to the inability for galaxies in groups to continue the gas accretion necessary to fuel star formation. This has been observed in galaxy clusters \citep{Yoon2013, Gim21}, and when scaled to the group environment, could indicate the beginning of the preprocessing and quenching processes \citep{Zabludoff1998,McGee2009,Wetzel2013,Schawinski2014,Crossett2017,kacprzak21}.\\
\subsection{Low and intermediate-ionization tracing tracing cool/warm gas} \label{sec:low-ionization}
Apart from Ly$\alpha$, we also observe other transitions like \ion{Si}{2}\, \ion{C}{2}\ and \ion{Si}{3}\ tracing gas up to the ionization potentials of 33.5~eV. \ion{C}{2}\ and \ion{Si}{3}\ are the most commonly detected low and intermediate-ionization species that are seen in five of the eighteen sightlines. This is consistent with other studies of the CGM and IGM \citep{collins09, shull09, lehner12, lehner15, richter16, borthakur16}. All of these absorbers are associated with strong, most likely saturated Ly$\alpha$\ absorbers
\begin{figure}
\centering
\plotone{Fig4.pdf}
\caption{\ion{O}{6}\ column density detections and 3$\sigma$ upper limits as a function of group halo mass for the COS-IGrM\ sample. The blue points indicate QSO sightlines that pass within the virial radius of individual group members, while the green points show sightlines that we expect to only probe IGrM. The gray points are from \cite{Stocke2019}, which systematically selects higher mass groups.}
\label{fig:ovimass}
\end{figure}
\subsection{\ion{O}{6}\ and \ion{N}{5} absorption tracing highly ionized gas} \label{sec:ovi}
We detect \ion{O}{6}\ absorbers in 8 and \ion{N}{5} absorbers in 2 out of our 18 sightlines. Each \ion{N}{5}\ absorption feature was also present with \ion{O}{6}\ absorption. Of the eight detections, five of the sightlines were ``pure'' IGrM\ sightlines, while three sightlines were CGM~+~IGrM. We detected both the transitions of the \ion{O}{6}\ doublet for two sightlines. Five sightlines showed the stronger of the two transition at \ion{O}{6}\ $\lambda$1031~$\rm\AA$, while one sightline showed absorption at \ion{O}{6}\ $\lambda$1037~$\rm\AA$ with an intervening absorption line at the expected position of \ion{O}{6}\ $\lambda$1031~$\rm\AA$. As noted earlier, the redshift range places the \ion{O}{6}\ doublet at observed wavelengths lower than 1215$ \rm \AA$ and hence we do not expect any misidentification of lower redshift Ly$\alpha$\ absorbers as \ion{O}{6}. Figure \ref{fig:ovimass} shows the column density of \ion{O}{6}\ absorbers from the COS-IGrM\ survey as well the survey by \cite{Stocke2019}. Over the entire halo mass range, the \ion{O}{6}\ and \ion{N}{5}\ detection rates are 44$\pm$5\% and 11$\pm$5\%, respectively.
\begin{figure*}
\centering
\plottwo{Fig5a.pdf}{Fig5b.pdf}
\caption{\ion{O}{6}\ column density detections and 3$\sigma$ upper limits as a function of projected QSO impact parameter (left) and the QSO impact parameter normalized by the group's virial radius (right). The colors are the same as in Figure \ref{fig:ovimass}.}
\label{fig:ovirho}
\end{figure*}
Figure \ref{fig:ovirho} shows the \ion{O}{6}\ detections as a function of QSO impact parameter from the center of the group. We observe a flat distribution of detections from $0.1-1.5~\rm{R_{vir}}$, which suggests that the sightlines may be probing gas that is not at the virial temperature. Since X-ray studies \citep{helsdon_ponman00, Mulchaey2000,Robson2020} show temperature gradients in galaxy groups, the observed flat distribution of \ion{O}{6}\ detections provides evidence that \ion{O}{6}\ is not tracing the bulk component of the IGrM.
Another indication that the OVI absorbers in our sample is tracing a mix of hot and cool gas is the fact that while all the systems that show \ion{O}{6}\ also show Ly$\alpha$, but the kinematics can be quite different. For example, six of the eight sightlines with \ion{O}{6}\ detections show Ly$\alpha$\ absorption with the same (or slightly offset) velocity centroid, while the remaining two sightlines have Ly$\alpha$\ at a much larger ($\ga 200$~$\rm{kms^{-1}}$). A single ionization process could not produce both Ly$\alpha$\ and \ion{O}{6}\ at the levels detected in some cases. Therefore, for these two different species to be observed within a close velocity offset, multiple clouds must be present, which indicates hot and cool gas in close proximity.
\subsection{Absorber Kinematics} \label{sec:kinematics}
In this section, we use the kinematics of the absorbers to explore further the nature and distribution of gas as traced by absorption.
First, we use the velocity spread of the absorbers to ascertain if the absorbing gas is bound to the group. Figure \ref{fig:escapevel} (left) shows the absorption lines detected in the COS spectra at velocity relative to the systemic velocity of the group\footnote{This velocity offset is only from the line of sight velocity. As a result, if the other two velocity components were known, then the fraction of unbound absorbers could increase.}, which is depicted by the dashed line at $v-v_{sys}=0$. Each species is color coded with the dominant absorption feature indicated by a larger halo around the data point. In total, there are 70 absorbers depicted in the plot. 29 and 12 of those are Ly$\alpha$\ and \ion{O}{6}, respectively, while the remaining represent the other species discussed above.
The solid lines show the escape velocity as a function of halo mass and virial radius. The right panel of Figure \ref{fig:escapevel} includes the data from \cite{Stocke2019}, which extends the dynamic range of halo masses.
The vast majority of the absorption features are bound to the gravitational potential of the groups. There are 9 out of 70 absorbers from five sightlines (J0841+1406, J1017+4702, J1020+1003, J1216+0712, and J1339+5355) that have sufficient velocities, relative to the group, to escape the gravitational potential. These are composed of 5 Ly$\alpha$, 2 \ion{O}{6}, 1 \ion{C}{2}, and 1 \ion{Si}{3}\ absorbers. The same trend is observed with the data from \cite{Stocke2019} as only two absorption features (1 Ly$\alpha$\ and 1 \ion{O}{6}) are observed at high enough velocity offsets to escape the group.
Among the 12 \ion{O}{6}\ absorbers detected in 8 sightlines, 83\% (10/12) are gravitationally bound to their group halo and only two absorbers show velocities greater than the escape velocity. One of the unbound \ion{O}{6}\ absorbers is seen in the sightline towards J0841+1406 passes through the CGM of a 4.4~L$_*$ galaxy at 145~kpc. The velocity offset between the \ion{O}{6}\ absorber and this is $\sim 650$~$\rm{kms^{-1}}$.
Interestingly, the \ion{O}{6}\ absorber does not have a corresponding Ly$\alpha$\ absorber. The Ly$\alpha$\ absorber is seen in this group is more than 600~$\rm{kms^{-1}}$\ offset from the \ion{O}{6}\ absorption feature and is offset by $\sim 150$~$\rm{kms^{-1}}$\ to the closest member galaxy. Hence, the \ion{O}{6}\ absorber could be tracing infalling or outflowing gas. The large doppler width of the \ion{O}{6}\ absorber of $b=81.6$~$\rm{kms^{-1}}$\ suggests that this absorber is tracing WHIM-like material \citep{Cen1999,Dave2001}. While this doppler width is consistent with WHIM-like material, we cannot conclusively rule out the possibility that this absorber is not related to the overall gas phase. As stated in \cite[and references therein]{Oppenheimer2009}, \ion{O}{6}\ with these doppler widths cannot be only a result of thermal broadening, but also requires a kinematic origin. This leaves some uncertainty as to the exact gas phase due to the lack of other metal-line transitions.
\begin{figure*}
\centering
\plottwo{Fig6a.pdf}{Fig6b.pdf}
\caption{\emph{Left: }Observed line of sight velocity relative to the group center for all detections compared to the halo mass of each group. The solid line represents the escape velocity of the group's gravitational potential. The absorption features with the largest column density are marked by a lighter halo over the data point to indicate the dominant component of the transition.
\emph{Right: }Same as the left panel, but only for Ly$\alpha$\ and \ion{O}{6}\ including the data from \cite{Stocke2019}. }
\label{fig:escapevel}
\end{figure*}
The second \ion{O}{6}\ absorber with a large velocity offset relative to the group is in the sightline towards J1017+4702 (Figure 2.2). The velocity offset of this \ion{O}{6}\ absorber is sufficient to escape the gravitational potential of the group. This sightline also exhibits a saturated Ly$\alpha$\ profile with a column density, logN(HI)$>$15.2. Since Ly$\beta$\ is blended with an intervening absorber, we cannot utilize it to help constrain the column density. Interestingly, in this case, the sightline does not pass within the virial radius of any spectroscopically confirmed L$_*$ galaxy and is at 926~kpc ($\equiv \rm 1.6~R_{vir}$) from the group center. However, there is one galaxy with matching photometric redshift at 90~kpc from the sightline, which might be the host of this saturated absorption system. It also has a neighbor with similar photometric redshift at an impact parameter from the QSO sightline of 212~kpc. Future spectroscopic redshift measurements are needed to confirmation the association between this neighboring galaxy and the absorption features present in this sightline. If the photometric redshifts are confined, then these galaxies would most likely be part of the group.
\begin{figure*}
\centering
\plottwo{Fig7a.pdf}{Fig7b.pdf}
\caption{\emph{Left: }Absolute velocity offset of \ion{H}{1} (black) and \ion{O}{6}\ (red) absorbers from the center the group. The fraction of absorbers within 200, 300, and 600~$\rm{kms^{-1}}$\ for both ions are displayed for all 29 Ly$\alpha$\ and 12 \ion{O}{6}\ absorbers.
\emph{Right: }Absolute velocity offset of \ion{H}{1} (black) and \ion{O}{6}\ (red) absorbers from the nearest galaxy to the sightline. The fraction of absorbers within 200, 300, and 600~$\rm{kms^{-1}}$\ for both ions are displayed. For both ions, more than 50\% are within 200~$\rm{kms^{-1}}$\ of the closest galaxy to the sightline, which suggests that some of these absorbers may originate from galactic outflows.}
\label{fig:abshist}
\end{figure*}
Figure \ref{fig:abshist} (left) shows a histogram of all of the \ion{H}{1} and \ion{O}{6}\ absorbers as a function of absolute velocity offset from the group center. This histogram quantitatively shows that the majority of both \ion{H}{1} and \ion{O}{6}\ absorbers are gravitationally bound to the group, while $\sim$10\% are observed to have velocities high enough to escape the group potential. These could either indicate infalling clouds or outflows. Therefore, we conclude that the large majority of the absorbers are tracing the cooler, gravitationally bound gas, which is centered well within the group's escape velocity. A similar conclusion was made by the \citet{Stocke2019} study, which concluded that galaxy groups primarily act as ``closed boxes'' for galactic evolution at low redshifts. However, the IGrM should still experience ``outside-in'' enrichment from the IGM \citep{Tegmark1993,Scanniapieco2002,Oppenheimer2012}. While the source of the initial IGM enrichment at early epochs is model dependent, each model of ``outside-in'' enrichment predicts that structures can regain metals that were expelled at earlier times.
We also investigate the velocity of the absorbers with respect to the nearest, spectroscopically confirmed member galaxy in projection. This is illustrated in the right panel of Figure \ref{fig:abshist} where \ion{H}{1} and \ion{O}{6}\ absorption features are shown by a histogram as a function of absolute velocity from the nearest galaxy, which could range between 0 and 1600~$\rm{kms^{-1}}$. For both \ion{H}{1} and \ion{O}{6}\ absorbers, we see that more than 50\% are within 200~$\rm{kms^{-1}}$\ of the closest galaxy to the QSO sightline. The fraction increases to 70\% within 300~$\rm{kms^{-1}}$, which is less than the escape velocity of an L$_*$ galaxy.
This indicates that these absorption features may originate from gas in the CGM of member galaxies in the group. On the other hand, the absorbers at higher velocity offsets are most likely tracing patchy components of the IGrM, or inflows/outflows from individual group members.
\subsection{Nature of IGrM} \label{sec:multiphase}
\begin{figure}
\centering
\plotone{Fig8.pdf}
\caption{Ratio of \ion{O}{6}\ to \ion{H}{1} column density as a function of normalized impact parameter for each of the 18 sightlines in the COS-IGrM sample. }
\label{fig:ovi_lya_ratio}
\end{figure}
In order to look at the overall ionization state of the IGrM, the ratio of \ion{O}{6}\ to \ion{H}{1} was examined for each of the 18 sightlines in the COS-IGrM sample (Figure \ref{fig:ovi_lya_ratio}). For these ratios, only the components of \ion{H}{1} found at the same velocity as \ion{O}{6}\ were used. If there were no \ion{O}{6}\ or \ion{H}{1} detections, then an upper limit was used. Sightlines with no Ly$\alpha$\ or \ion{O}{6}\ absorption were not included in this analysis. The sightlines with \ion{O}{6}\ column densities greater than the \ion{H}{1} columns shows that there is highly ionized gas throughout the IGrM, while the lack of a correlation between the ratio of column densities and impact parameter shows that there is no significant dependence of ionization state on normalized impact parameter.
Out of the 12 sightlines that show Ly$\alpha$\ absorption, 9 of those sightlines show evidence of multiple metal-line species detected in absorption that allows us to model the ionization state of the gas. Of these 9 sightlines, 7 clearly depict multiphase gas, where various metal-line species are present in varying levels suggesting that the components have very different ionization states.
The presence of these multiple components in most of the sightlines indicates that the absorption is associate with pockets of gas that maybe cooler than the rest of the media (and possibly more dense if they are in pressure equilibrium). Therefore, we believe that our data is primarily tracing a complex multiphase media, which cannot be described by a single ionization process. In Table \ref{tab:ionization}, we present the probable ionization process for each group, based upon the observed spectra.
For ionization modeling, the primary interest was to determine if any of the absorption lines from the COS-IGrM sample are consistent with photoionization, collisional ionization, or inconsistent with either process. For CIE modeling, the ratio of Ly$\alpha$\ to \ion{O}{6}\ absorption at 50\% solar was examined over a range of temperatures. If the observed column density ratio was consistent with CIE predictions, then it was noted that the absorption features were consistent with CIE. Since CIE predicts broad, shallow \ion{O}{6}\ without the presence of lower ionization state transitions, only those sightlines that had Ly$\alpha$\ and \ion{O}{6}\ were examined for consistency with CIE models.
The photoionization modeling was inherently less certain due to unresolved, blended components and a lack of multiple metal line species in the majority of sightlines. We used CLOUDY \citep{Ferland2013} with a Haart-Madau background and a total hydrogen density grid (log[n(h)] from [-5,-2]~particles~$cm^{-3}$) in 0.5 dex increments. The total neutral column density was fixed to the observed Ly$\alpha$\ column density. If a point in the grid existed where the column density ratio each metal species was consistent within the same density grid point, then we stated that the absorption components were consistent with photoionization.
\input{Table2.tex}
Some sightlines show absorbers that match the ratios and strengths predicted by CIE \citep{Gnat2007} for a hot $\approx \rm 10^{5.5}~K$ medium. For example, the sightline J1343+2538 (Figure 23) passing through a group at an impact parameter of 346~kpc ($\equiv \rm 0.9~R_{vir}$), shows broad Ly$\alpha$\ ($b_{Ly\alpha}$=47~$\rm{kms^{-1}}$) along with \ion{O}{6}\ suggestive of hot media \citep{Richter06}. The ratio of \ion{O}{6}\ to Ly$\alpha$\ column density of 0.65 dex is consistent with temperatures of $\rm 10^{5.3-5.4}$~K or 10$^{5.9}$~K for 0.5$-$[1~Z/H]$_{\odot}$ with temperature inverse proportional to the metallicity for the same column density ratio. The choice of this metallicity range is based on measurement from X-ray studies for groups of galaxies that typically find the average metallicities of the X-ray bright IGrM to be 0.4$-$0.6~[Z/H]$_{\odot}$ \citep{helsdon_ponman00}. We do not have strong metallicity constrains for non-X-ray bright groups, so we adopt the metallicities seen in X-ray studies.
Another example of collisional ionized gas are seen in the sightline towards group J1301+2819 (Figure 21).
In addition to tracing hot gas, this sightline shows a mix of multiple ionization states at slightly different velocities possibly tracing a multi-phase medium.
The Ly$\alpha$\ feature shows three components - two strong components with associated Ly$\beta$\ and one weak component with log~N(HI)=13.05. The strong components are seen in both \ion{N}{5}\ and \ion{O}{6}\ (the components are blended in \ion{O}{6}), whereas the weakest component is most prominent in \ion{O}{6}. This indicates that the different components trace different ionization states. For weakest component the ratios of \ion{O}{6}\ and Ly$\alpha$\ are in agreement with collision ionization equilibrium model. The ratio of column densities, $\rm log N(OVI) - log N(HI) =0.60$, corresponds to gas at 10$^{5.3}$~K or 10$^{5.9-6}$~K at 50\% solar metallicity. At lower metallicities, the observed ratio of column densities between \ion{O}{6}\ and Ly$\alpha$\ would indicate a slightly higher temperature.
On the other hand, the stronger components are quite puzzling. If photoionization was responsible for the observed ionization states, there should be other low-ionization transitions detected besides \ion{H}{1}\ such as \ion{Si}{2}, \ion{C}{2}\ and \ion{Si}{3}. Despite these transitions not being present at detectable levels in the spectra, the \ion{N}{5}\ to \ion{O}{6}\ ratio is consistent with photoionization (Table \ref{tab:ionization}). Therefore, we are unable to conclusively state the process behind the observed column densities, it is most likely a mixture of multiple ionization processes.
A similar case of multiphase media is seen in J1424+4214 (Figure 25), which shows two distinct ionization states with a velocity separation of about 45~$\rm{kms^{-1}}$: a less ionized system at $\sim$~-55~$\rm{kms^{-1}}$\ and a highly ionized state at $\sim$~-100~$\rm{kms^{-1}}$. One component is seen in lower ionization transitions like Ly$\alpha$, \ion{C}{2}, and \ion{Si}{3}, while the second components is seen in higher ionization transitions like \ion{N}{5}, \ion{O}{6}, as well as \ion{Si}{3}\ that show a weak feature suggesting \ion{Si}{3}\ is not the dominant ionization state of silicon.
The ratio of these lines indicate that the two components are at very different ionization states, thus suggesting that the IGrM\ is multiphase and cannot be described by a single ionization state. While the ionization processes for each of the two components cannot definitively be determined based upon the data at hand, the component centered at $\sim$~-100~$\rm{kms^{-1}}$\ is consistent with CIE at $\sim$10$^{5.2-5.3}$~K, while the component at $\sim$~-55~$\rm{kms^{-1}}$\ is consistent with photoionization based upon the \ion{C}{2}\ to \ion{Si}{3}\ line ratio.
Another sightline of interest is towards the group J1127+2654 (Figure 19). This sightline exhibits a saturated Ly$\alpha$\ profile with column density, logN(HI)$>$18.3, which makes it a Lyman-limit system \citep[LLS,][and references therein]{Lanzetta95}. The absorber complex shows multiple components commonly associated with extended disk \citep{lehner09}, inner CGM \citep{Werk2014,Armillotta2017,Fielding2020}, and/or tidal structures \citep{frye19}. This QSO passes within $\sim$119~kpc from a known group member; however, higher resolution spectroscopy and a rotation curve is necessary to confirm the connection between this LLS and the member galaxy. Similarly, due to the blending of \ion{O}{6}\ components in this QSO spectra, we cannot rule out photoionization as the primary ionization mechanism for these absorption lines.
Lastly, the sightline towards J1017+4702 (Figure 16) shows that Ly$\alpha$\ is saturated at the same position as \ion{C}{2}, \ion{Si}{3}, and broad, shallow \ion{O}{6}. Photoionization alone cannot produce broad, shallow \ion{O}{6}\ and CIE does not predict the existence of saturated Ly$\alpha$\ and broad \ion{O}{6}\ at a single temperature. Since this is an IGrM\ sightline, the broad \ion{O}{6}\ may be tracing a hotter component; however, the lower transitions show evidence of cooler gas at the same velocity. Photoionization modeling with CLOUDY \citep{Ferland2013} showed that the lower velocity components of \ion{C}{2}\ and \ion{Si}{3}\ are consistent with photoionization, while the higher velocity components are inconsistent with photoionization. Overall, it is clear that the ionization states of these groups are complex and the ionization processes behind the multiphase gas cannot always be explained by either photoionization or CIE. Future studies with better modeling, higher resolution observations, and broader wavelength coverage can help shed insight into these ionization processes.
\subsection{Origin of \ion{O}{6}\ Absorbers} \label{sec:coshalos}
\begin{figure}
\centering
\plotone{Fig9.pdf}
\caption{\ion{O}{6}\ column density as a function of the specific star formation rate of the nearest member galaxy to the QSO sightline. The blue (CGM~+~IGrM) and green (IGrM) points show the COS-IGrM sample and the size of the points correspond to the impact parameter from the closest, spectroscopically confirmed member galaxy. The cyan and orange data points show the results from the COS-Halos survey \citep{Tumlinson2011}, which robustly observed an increased amount of \ion{O}{6}\ in the CGM of star forming galaxies.}
\label{fig:coshalos}
\end{figure}
We differentiate between CGM and IGrM\ absorption in galaxy groups by comparing our \ion{O}{6}\ detections to those detected in the COS-Halos survey \citep{Tumlinson2011}. The COS-Halos survey discovered that a strong correlation between \ion{O}{6}\ in the CGM and the star formation rate of galaxies existed. In order to compare the our data with the COS-Halos sample, we determined the galaxy closest to the QSO sightline and then matched the galaxy to the star formation rate from the MPA-JHU DR7\footnote{https://wwwmpa.mpa-garching.mpg.de/SDSS/DR7/} galaxy catalog \citep{Brinchmann2004}.
Figure \ref{fig:coshalos} shows our data along with those from the COS-Halos survey. The CGM+IGrM sightlines, shown as deep blue circles, have a clear host galaxy as the sightline passes within the viral radius (assuming an isolated halo) of a member galaxy. The pure IGrM sightlines do not pass through the CGM of the nearest galaxy (shown in green circles). Therefore, they are not applicable for comparison with the COS-Halos sample; nevertheless, we show them on the plot for comparison with the CGM+IGrM sub-sample. It is worth noting that even the blue points that do probe the CGM pass through the outer CGM ($\rho >$110~kpc) and not the inner CGM like COS-Halos sample.
Overall, we do not see a trend of higher \ion{O}{6}\ levels as a function of the specific star formation rate (sSFR) or the star formation rate of the nearest galaxy. This is not surprising considering the impact parameters. However, it does suggest that the origin of our \ion{O}{6}\ absorbers are probably not related to the star formation activity of individual galaxies and therefore, we are most likely not tracing the CGM gas physics as seen in the COS-Halos survey, but instead, a more group-related phenomena.
\begin{figure}[t]
\centering
\plotone{Fig10.pdf}
\caption{BPT diagram showing the locations of the COS-IGrM sample where reliable flux measurements from the MPA-JHU DR7 catalog were present. The division between phase space pertaining to either star forming or AGN galaxies is marked by the dashed line \citep{Kauffmann2003}. The colored circles represent sightlines with \ion{O}{6}\ detections vs. non-detections and the size of the circles represents the projected impact parameter to the closest galaxy to the sightline. }
\label{fig:bpt}
\end{figure}
While we can confidently rule out the CGM of L$_{\star}$\ galaxies as the source of \ion{O}{6}\ absorbers, there could potentially be smaller galaxies that may be present closer to the sightline. A much deeper redshift survey of galaxies in the vicinity of the QSO sightlines would enable us to quantify the presence of low-mass galaxies. Nevertheless, sub-L$_{\star}$\ galaxies are not expected to have significant metal reservoirs beyond their inner CGM ($\rho \rm > 0.5~R_{vir}$) \citep{bordoloi14}. Hence, it is not likely that the CGM of sub-L$_{\star}$\ galaxies could dominate the \ion{O}{6}\ detected in our sample. On the other hand, material spread out by tidal interactions can have a large cross-section on the sky and may survive as faint diffuse partially neutral gas in the IGrM for hundreds of millions of years \citep{Borthakur10,borthakur15}.
\begin{figure*}[t]
\centering
\includegraphics[width=2.25in]{Fig11a.pdf}
\includegraphics[width=2.25in]{Fig11b.pdf}
\vspace{-.50cm}
\caption{Stacked spectra for Ly$\alpha$\ (left) and \ion{O}{6}\ (right) for each of the 18 sightlines in the COS-IGrM\ sample (top). The sightlines were also divided into the IGrM\ and CGM sightlines and stacks of each were created (middle and bottom respectively). The stacks are centered on the center of mass velocity of the group and all intervening absorption features were removed. }
\label{fig:stacks}
\end{figure*}
Our sample shows a larger fraction of green valley galaxies than typically observed in the Universe. \cite{Jian2020} finds that on average, 20\% of galaxies populate the green valley and the majority of those are field galaxies and not those found in more dense environments. Observing $\sim$33\% of the closest galaxies to the QSO sightline in our sample to be in the green valley reinforces the idea that galaxy group environments may act as important sites where the process of quenching is active \citep{Wetzel2012, Wetzel2013}. The role of the IGrM or the CGM in turning these galaxies green is still unclear.
Another possibility for the origin of \ion{O}{6}\ in the IGrM could be due to AGN activity. In order to address this, used the emission line ratios from the MPA-JHU\footnote{\url{https://www.sdss.org/dr14/spectro/galaxy_mpajhu/}} DR7 catalog to construct a Baldwin, Phillips \& Terlevich (BPT) diagram \citep{Baldwin1981} so that star forming galaxies could be separated from AGN using the demarcation as defined by \citet{Kauffmann2003}. The locations of the COS-IGrM\ sample compared to the SDSS DR7 sample from the MPA-JHU catalog are shown in Figure \ref{fig:bpt} as colored circles. 15 out of the 18 groups in our sample had emission line measurements for the closest galaxy to the QSO sightline and therefore, could be included in the BPT diagram. The color of the circle represents \ion{O}{6}\ detections (orange) vs. non-detections (magenta). The size of the symbol represents the impact parameter of the sightline, where larger sizes indicate small impact parameter. We do not find any systematic over-density of \ion{O}{6}\ detection or non-detection in sightlines with or without AGN. Therefore, we conclude that AGN activity is not the primarily contributor of \ion{O}{6}\ in the IGrM.
\subsection{Stacked Spectra} \label{stacks}
In order to look for fainter gas associated with the IGrM, we stacked sightlines centered around the group systemic velocity for Ly$\alpha$\ and \ion{O}{6}. For each species, stacks were created using all 18 sightlines as well as subsets of CGM $+$ IGrM\ or IGrM\ only sightlines. These stacks are shown in Figure \ref{fig:stacks} along with the number of sightlines going into each subset. The equivalent widths were measured for velocities within $\pm400$ $\rm{kms^{-1}}$\ and $\pm800$ $\rm{kms^{-1}}$\ from the group's systematic redshift. These values are listed as $\rm{W_{800}}$ and $\rm{W_{1600}}$ respectively.
The Ly$\alpha$\ stacks show net absorption centered around zero velocity for the CGM $+$ IGrM\ subset, and absorption corresponding to higher velocity offsets in the IGrM\ stack. When we stack the full COS-IGrM\ sample, we observe a combination of the two subsets meaning that the IGrM\ in our sample is traced by two distinct regions: gas at the systematic velocity of the group as well as gas that is at larger velocities than the group's systemic velocity. This could perhaps be a result of warmer gas condensing in the outskirts and falling back towards the center of the group.
The \ion{O}{6}\ stacks show weak net absorption throughout all the sightlines. However, there is absorption in the IGrM\ sightlines. Both the CGM $+$ IGrM\ stack and the pure IGrM stack show that the majority of the absorption is within of the central 800 $\rm{kms^{-1}}$\ of the group. The covering fraction of the IGrM stacks is more uniform than the CGM stacks as there is net absorption throughout the $\pm800$ $\rm{kms^{-1}}$. Non-detection of \ion{O}{6}\ in the full stack indicates that there is not a volume filling phase of the IGrM, but instead, an \ion{O}{6}\ traced IGrM is a more transient phenomena.
\begin{figure*}
\centering
\plottwo{Fig12a.pdf}{Fig12b.pdf}
\caption{(Left:) Observed \ion{O}{6}\ column densities normalized by the QSO path length through the group as a function of group virial temperature (N(\ion{H}{1})$_{model}$). The blue and green data points are from the COS-IGrM\ sample, while the gray squares show the results from \cite{Stocke2019}. The solid, black line represents theoretical predictions based on collisional ionization equilibrium models from \cite{Gnat2007} assuming 50\% solar metallicity, assuming a total hydrogen density of $10^{-3}~\rm{cm^{-3}}$ (Right:) \ion{O}{6}\ column density as a function of the detected \ion{O}{6}\ line width ($\Delta v$). The dotted, dot-dashed, and dashed lines show radiative cooling models from \cite{Bordoloi2017} at $10^{5.5}$~K, $10^{5.8}$~K, and $10^{6}$~K, respectively. }
\label{fig:temps}
\end{figure*}
\section{Discussion} \label{sec:discussion}
From the COS-IGrM\ survey, we have observed no significant trends between the column density of Ly$\alpha$\ or \ion{O}{6}\ and the physical parameters of the group such as virial radius, impact parameter, and halo mass. This may be an indication that we are not observing a hot, volume filling IGrM; instead we are detecting cooler pockets of gas that are perhaps in pressure confinement within the IGrM. This would line up more closely with what was concluded in \cite{Stocke2017, Pointon2017} and \cite{Stocke2019} for more massive groups. If this is indeed correct, X-ray spectroscopy of \ion{O}{7} and \ion{O}{8} would be required to observe the hotter component of the IGrM, even for lower mass groups ($10^{12.8}-10^{13.7}$~$\rm{M_{\odot}}$).
This idea is further reinforced by looking at the virial temperatures of the groups compared to the predicted \ion{O}{6}\ column densities from collisional ionization equilibrium models \citep{Gnat2007}. Figure \ref{fig:temps} (left) shows the predicted and observed column densities of \ion{O}{6}\ normalized by the total hydrogen column density through the group as a function of virial temperature (denoted by N(\ion{H}{1})$_{model}$). The column density of hydrogen was estimated by using the IGrM gas density of $n=10^{-3}$~cm$^{-3}$, and multiplying it by the total path length through each group in our sample, which is approximated by a sphere of radius, 2R$_{vir}$. The gas density was selected as a conservative estimate based upon electron density profiles from X-ray data of galaxy groups \citep{Sun2003, Khosroshahi2004} and from density measurements of the IGrM from double bent radio jets \citep{Freeland2011}. From this figure, it is evident that we are primarily observing cooler gas than what would be at the group's virial temperature based upon the amount of \ion{O}{6}\ observed, which provides more support to our previous statements.
To investigate the theory that the observed \ion{O}{6}\ is due to cooler gas than the hotter IGrM, we looked at the relationship between the \ion{O}{6}\ column density and the \ion{O}{6}\ line width for our sample and other samples from various environments (right panel of Figure \ref{fig:temps}). \cite{Heckman2002} demonstrated that \ion{O}{6}\ absorption lines in various environments such as the Milky Way, high velocity clouds, Magellanic Clouds, starburst galaxies, and the intergalactic medium all can be described by radiatively cooling gas through the relationship between column density and the Doppler '$b$' parameter. \cite{Bordoloi2017} revisited these models to show that the line width, $\Delta v=3b_D/\sqrt{2}$, is a more appropriate tracer of the flow velocity than the Doppler '$b$' parameter in describing the radiatively cooling \ion{O}{6}. We show data from \cite{Bowen2008}, \cite{Burchett2015}, and \cite{Stocke2019} along with the COS-IGrM survey \ion{O}{6}\ detections\footnote{Since the line width is related to the Doppler `$b$' parameter and therefore a Voigt profile fit, there is no physical upper limit for non-detections.} to show that the trends observed from \ion{O}{6}\ in the Milky Way and the IGM also largely agree with \ion{O}{6}\ detected in the IGrM, respectively. On average, the COS-IGrM data can be described by radiatively cooling gas between $10^{5.8}$~K, and $10^{6}$~K. This may be indicating that the \ion{O}{6}\ detected in our sample originates from gas falling towards member galaxies and cools radiatively as it passes through the CGM of a group member or passes through cooler pockets within the hotter IGrM.
The cooling models described in \cite{Bordoloi2017} predict that \ion{N}{5}\ column densities should be about an order of magnitude lower than those predicted for \ion{O}{6}. This prediction is consistent with our three \ion{N}{5}\ detections as well as our upper limits in this sample based upon the $10^{5.27}$~K cooling curves in \cite{Bordoloi2017}. Since many of the \ion{O}{6}\ detections are relatively close to the detection limit, the lack of \ion{N}{5}\ detections is not unexpected due to this prediction.
Lastly, we can make an estimate as to the total amount of oxygen in these galaxy groups. Following equation 1 in \cite{Tumlinson2011}, we can calculate the minimum mass of oxygen in galaxy group halos by:
\begin{equation}
M_O = 5\pi \langle R_{vir} \rangle^2 \langle N_{OVI} \rangle m_O f_{hit} (\frac{0.2}{f_{OVI}})
\label{eq:mass}
\end{equation}
where $f_{OVI}$ is the fraction of oxygen that is in \ion{O}{6}\ based upon CIE models \cite{Gnat2007}. Using both the mean and median values of the \ion{O}{6}\ column densities and group virial radii, we can determine the minimum amount of oxygen mass in our galaxy groups. This can be compared to the amount of oxygen in the member galaxies by assuming $M_O \sim 0.065M_*$ \citep{Peeples2014,Tumlinson2017}. This difference (gray shaded region) is shown in Figure \ref{fig:oxygenmass} for both the mean (dashed lines) and median (solid lines) values of the stellar masses of group members. Based upon the virial temperature of these galaxy groups in Figure \ref{fig:temps}, the corresponding fraction of oxygen in \ion{O}{6}\ is $<10^{-4}$. Therefore, we can estimate that over the narrow temperature range corresponding to $f_{OVI}$ of $10^{-4} - 10^{-5}$ ($10^{6.55}-10^{6.75}$~K), there is upwards of $10^{11.6}~M_{\odot}$ of oxygen in the IGrM.
\begin{figure}[t]
\centering
\plotone{Fig13.pdf}
\caption{The minimum mass of oxygen as calculated by Equation \ref{eq:mass} compared to the mass of oxygen in the member galaxies. The solid lines show the masses determined by median values while the dashed lines show the masses determined through the mean of the stellar mass, virial radii, and \ion{O}{6}\ column densities. The gray shaded region shows the mass of oxygen that can be attributed to the IGrM at various values of $f_{OVI}$, the fraction of oxygen that is in \ion{O}{6}.}
\label{fig:oxygenmass}
\end{figure}
\subsection{Future Outlook}
In order to accurately and completely characterize the IGrM, higher ionization species should be targeted in future studies. From the COS-IGrM\ survey, it is clear that \ion{O}{6}\ is not an ideal tracer of the IGrM. Since \ion{O}{6}\ is only observed in 8 out of our 18 groups, the predominant, volume filling component of the IGrM\ should exist at a hotter temperature for galaxy groups at halo masses between 12.~$\rm{M_{\odot}}$ $<$ log[M$_{\rm halo}$] $<$ 14.7~$\rm{M_{\odot}}$. We can rule out a pervasive media of the IGrM at cooler temperatures due to the weak low and medium ionization potential lines observed in our data. Therefore, to observe the dominant phase of the IGrM, future studies should look to \ion{O}{7}, \ion{O}{8}, \ion{Ne}{8} and \ion{Mg}{10}, which are stronger transitions at temperatures of $10^{6.5}-10^{7.5}$~K.
Once the pervasive phase of the IGrM\ is observed, it can be combined with other studies to fully characterize the IGrM\ of galaxy groups. Simulations by \cite{Dave2002, LeBrun2017,Farahi2018} have made substantial progress in determining consistent scaling relations for lower mass halos that are consistent with observational programs such as those by \cite{Sun2003,Eckmiller2011,Babyk2018,Lovisari2020}. Additionally, the thermal SZ effect is being utilized in order to determine the baryonic content of lower mass galaxy clusters and groups \citep[and references therein]{Vikram2017, Henden2019,Pratt2020}. By the combination of these results, these hot halos can be fully characterized.
\section{Conclusions} \label{sec:conclusion}
We present the results of the COS-IGrM\ survey, where 18 QSO sightlines passing through galaxy groups (0.2R$_{vir}\leq \rho \leq$1.6R$_{vir}$) were studied in an effort to characterize the IGrM. Our conclusions are as follows:
\begin{enumerate}
\item We detect Ly$\alpha$\ absorption in 12 of the 18 galaxy groups, with 4 of those groups also having corresponding Ly$\beta$\ absorption. However, we detect no statistically significant trend between Ly$\alpha$\ column density and halo mass or QSO impact parameter.
\item 8 of the 18 groups show the presence of \ion{O}{6}\, thus the covering fraction of \ion{O}{6}\ is 44$\pm$5\%. The lack of \ion{O}{6}\ absorption in over 50\% of our sample indicates that the volume filling IGrM\ at (or near) the virial temperature of galaxy groups is not primarily traced by \ion{O}{6}. We also find no correlation between column density of \ion{O}{6}\ and halo mass or QSO impact parameter.
\item \ion{C}{2}, \ion{Si}{2}, \ion{Si}{3}, and \ion{N}{5}\ absorption was detected in 5, 1, 5, and 2 groups, respectively. These lead to covering fractions of 28$\pm$5\%, 6$\pm$5\%, 28$\pm$5\%, and 11$\pm$5\% for \ion{C}{2}, \ion{Si}{2}, \ion{Si}{3}, and \ion{N}{5}, respectively. These data suggest that the low-ionization transitions are primarily due to photoionization or other non-equilibrium processes.
\item We find evidence that the IGrM\ is multiphase and has a complex structure. While higher resolution spectra and coverage of more intermediate ionization transitions are necessary for complete ionization modeling, we find five instances where CIE explains the observed spectra and four instances where photoionization is consistent with the transitions present.
\item We find that 9 out of 70 absorbers (13$\pm$1\%) have sufficient velocities, relative to the group, to escape the group's gravitational potential. Therefore, we conclude that galaxy groups are primarily ``closed boxes'' for galaxy evolution at low redshifts (0.1 $\le z \le$0.2).
\item We show that the \ion{O}{6}\ absorbers can be described by radiatively cooling gas between $10^{5.8}$~K, and $10^{6}$~K. This might indicate that the \ion{O}{6}\ detected in our sample originates from pockets of gas cooling within the hotter component of the IGrM.
\item We do not find evidence of AGN activity having an impact on whether or not \ion{O}{6}\ is detected within a group or not. Similarly, we do not observe the star formation of the nearest spectroscopically confirmed neighbor to be a driver for \ion{O}{6}.
\item We observe some \ion{O}{6}\ absorption in our stacked data. This shows evidence of \ion{O}{6}\ traced IGrM\ throughout our sample. Despite \ion{O}{6}\ not being the dominate form of oxygen at the virial temperature of these galaxy groups, we see evidence that we are observing gas cooler than the hot, volume filling component of the IGrM that could be observed in X-rays via \ion{O}{7}, \ion{O}{8}, \ion{Na}{12}, or extreme-UV lines such as \ion{Ne}{8}.
\end{enumerate}
Since the \ion{O}{6}\ detections are determined to be primarily tracers of cooler pockets of gas and not the IGrM\ at the virial temperature of the group,
full accounting for the amount of baryonic matter in these groups cannot be accurately measured with the data in hand. In order to complete the baryon census for galaxy groups, future studies should try and observe higher ionization states such as \ion{O}{7} and \ion{O}{8}, which will trace gas closer to the virial temperature of galaxy groups.
\acknowledgements
The authors would like to thank the anonymous referee for their feedback and suggestions to improve this manuscript. TM, SB, TH are supported by HST-GO-13314 from STScI. The authors would also like to thank Mark Voit along with Chris Dupuis, Mansi Padave, Ed Buie II, Hansung Gim, Jacqueline Monkiewicz, and Martin Flores from the STARs Lab at ASU for thoughtful discussion and input throughout this project. The Arizona State University authors acknowledge the twenty two Native Nations that have inhabited this land for centuries. Arizona State University’s four campuses are located in the Salt River Valley on ancestral territories of Indigenous peoples, including the Akimel O’odham (Pima) and Pee Posh (Maricopa) Indian Communities, whose care and keeping of these lands allows us to be here today. We acknowledge the sovereignty of these nations and seek to foster an environment of success and possibility for Native American students and patrons.
|
\section{Introduction}
Gaussian process (GP) regression has become ubiquitous as a statistical tool in many applied sciences including astrophysics, environmental sciences, molecular dynamics, financial economics, and social sciences~\cite{banerjee1, dfm1, bda, gonzalvez1, stein2, cressie1, bartok1, luger1}.
In most Gaussian process regression problems a user has observed data $\{(x_i, y_i)\}$ where $x_1,...,x_N$ are independent variables that belong to some set $D \subseteq \mathbb{R}^d$ and $y_i \in \mathbb{R}$ are dependent variables. The observed data $y_1,...,y_N$ are assumed to be observations of the form
\begin{align}
y_i = f(x_i) + \epsilon_i
\end{align}
where $\epsilon_i$ is iid Gaussian noise, and $f: D \to \mathbb{R}$ is given a Gaussian process prior $f(x) \sim \mathcal{GP}(m(x), k(x,x'))$ where $m$ is a user-specified mean function and $k$ is a covariance function~\cite{rasmus1}.
The primary computational limitation of GP regression as a practical tool is the prohibitive cost of evaluating the likelihood function of a Gaussian process
\begin{align}\label{63}
p(\vct{y}) \propto \frac{1}{|\mtx{K} + \sigma^2 \mtx{I} |^{1/2}} \exp^{-\frac{1}{2} \vct{y}^t (\mtx{K} + \sigma^2\mtx{I})^{-1} \vct{y}}
\end{align}
where $\mtx{K}$ is the $N \times N$ matrix such that $\mtx{K}_{i,j} = k(x_i, x_j)$ and $\sigma^2$ is the variance of the iid observation noise. For a general symmetric matrix, direct evaluation of the quadratic form in (\ref{63}) requires $O(N^3)$ operations, which for many modern problems is far too costly. A large body of literature has emerged over the last couple of decades on efficient schemes for evaluating that quadratic form when $N$ is large (i.e. greater than around 10,000). Often, these computational methods rely on taking advantage of the particular structure of covariance matrices~\cite{oneil1, minden1, dfm1}, approximating the covariance matrix with a low rank matrix (reduced-rank methods)~\cite{candela1, solin1, greengard2}, spectral methods~\cite{hensman1, lazaro1, rahimi1}, or many more strategies.
In this paper we introduce numerical algorithms for representing a zero-mean Gaussian process $f$ with a translation invariant covariance kernel $k$ as a random expansion of the form
\begin{align}\label{23}
f(x) \sim \gamma_1 \alpha_1 \cos(\xi_1 x) + \beta_1 \gamma_1 \sin(\xi_1 x)
+ ... + \alpha_m \gamma_m \cos(\xi_n x) + \beta_m \gamma_m \sin(\xi_m x)
\end{align}
where for $i=1,...,m$ the frequencies $\xi_i \in \mathbb{R}$ are fixed, and $\alpha_i$ and $\beta_i$ are the iid Gaussians
\begin{align}
\alpha_i, \beta_i \sim \mathcal{N} \big(0, 1 \big).
\end{align}
The coefficients $\gamma_i$ of (\ref{23}) are defined by
\begin{align}
\gamma_i = \sqrt{2 w_i \hat{k}(\xi_i)}
\end{align}
where $\hat{k}$ denotes the Fourier transform of the covariance kernel and $w_i \in \mathbb{R}$. The numerical work in constructing expansion (\ref{23}) involves finding the frequencies $\xi_i$ and weights $w_i$ such that expansion (\ref{23}) has an effective covariance kernel that approximates the desired kernel (or family of kernels) to high accuracy. We provide numerical results for constructing an expansion of the form (\ref{23}) that is valid for a family of Mat\'ern kernels. The Mat\'ern kernel in one dimension, $k$, is defined by
\begin{equation}\label{131}
k(r) = \frac{2^{1-\nu}}{\Gamma(\nu)}
\bigg( \sqrt{2\nu}\frac{r}{\rho}\bigg)^{\nu} K_{\nu}\bigg(\sqrt{2\nu}\frac{r}{\rho}\bigg)
\end{equation}
where $\rho, \nu > 0$. The Fourier transform of $k$, which we denote $\hat{k}$, is given by
\begin{align}
\hat{k}(\xi) = \frac{2 \pi^{1/2} \, \Gamma(\nu + 1/2) (2\nu)^{\nu}}{\Gamma(\nu) \rho^{2\nu}}
\bigg(\frac{2\nu}{\rho^2} + 4\pi^2 \xi^2 \bigg)^{-(\nu + 1/2)},
\end{align}
for all $\xi \in \mathbb{R}$ where $\Gamma$ is the gamma function.
We demonstrate the algorithm of this paper on finding $\xi_1,...,\xi_{m}$ and $w_1,...,w_{m}$ such that expansion (\ref{23}) accurately approximates all GPs defined on $[-1, 1]$ with Mat\'ern kernels where $\rho \in [0.1, 0.5]$, and $\nu \in [3/2, 7/2]$.
There are several advantages to using expansion (\ref{23}) as a representation of a family of GPs. One computational benefit stems from the so-called function-space approach to Gaussian process regression (see e.g. \cite{rasmus1}, \cite{greengard2}, \cite{filip2019random}) in which the posterior density is defined over the coefficients of a function space expansion -- in this paper expansion (\ref{23}). Specifically, given data $\{(x_i, y_i)\}$ and a covariance kernel $k$, function space Gaussian process regression is the $L^2$-regularized linear regression
\begin{equation}\label{72}
\begin{aligned}
\vct{y} & \sim \mtx{X}\vct{\beta} + \vct{\epsilon} \\
\vct{\epsilon} & \sim \mathcal{N}(0, \sigma^2 \mtx{I}) \\
\vct{\beta} & \sim \mathcal{N}(0, \mtx{I})
\end{aligned}
\end{equation}
where $\vct{\beta} \in \mathbb{R}^{2m}$ is the vector of coefficients and $\mtx{X}$ is the $N \times 2m$ matrix
\[
\mtx{X} =
\renewcommand*{\arraystretch}{1.3}
\begin{bmatrix}
\gamma_1 \cos(\xi_1 x_1) & \dots & \gamma_m \cos(\xi_m x_1) & \gamma_1 \sin(\xi_1 x_1) & \dots & \gamma_m \sin(\xi_m x_1) \\
\gamma_1 \cos(\xi_1 x_2) & \dots & \gamma_m \cos(\xi_m x_2) & \gamma_1 \sin(\xi_1 x_2) & \dots & \gamma_1 \sin(\xi_m x_2) \\
\vdots & & \vdots & \vdots & & \vdots \\
\gamma_1\cos(\xi_1 x_n) & \dots & \gamma_m\cos(\xi_m x_N) & \gamma_1\sin(\xi_1 x_n) & \dots & \gamma_m\sin(\xi_m x_n)
\end{bmatrix}.
\]
The posterior density corresponding to (\ref{72}) is defined over $2m + 1$ dimensions ($\sigma$ and the $2m$ coefficients $\vct{\beta}$) and the conditional mean of $\vct{\beta}$ is the function
\begin{align}
\sum_{i=1}^{k} \vct{\bar{\beta}}_{1,i} \gamma_i \cos(\xi_i x) + \vct{\bar{\beta}}_{2,i} \gamma_i \sin(\xi_i x)
\end{align}
where $[\vct{\bar{\beta}}_1 \, \vct{\bar{\beta}}_2]^t$ is the solution to the linear system of equations
\begin{align}\label{203}
(\mtx{X^{\mathsf{T}} X} + \sigma^2 \mtx{I}) x = \mtx{X^{\mathsf{T}}}\vct{y}.
\end{align}
Representing a GP with Fourier expansion (\ref{23}) has the advantage that GP regression via linear system (\ref{203}) can be solved in $O(N\log{N} + m^3)$ operations. For a general $N \times 2m$ matrix, $\mtx{X}$, solving linear system (\ref{203}) requires $O(Nm^2)$ operations, which can be prohibitively expensive for big data problems. However since we use Fourier expansions, the $2m \times 2m$ matrix $\mtx{X^{\mathsf{T}} X}$ can be formed in $O(N\log{N})$ operations using a non-uniform fast Fourier transform (FFT)~\cite{dutt1, lgreengard1}. The efficient formation of $\mtx{X^{\mathsf{T}} X^t}$ reduces the total computational cost of solving the linear system from $O(Nm^2)$ operations to $O(N \log{N} + m^3)$. In general, $m$ is sufficiently small that $m^3$ operations is easily affordable on a laptop. Furthermore, for GP problems that involve fitting hyperparameters, the cost of solving linear system (\ref{203}) is $O(m^3)$ operations for all hyperparameters after a precomputation of $O(N\log{N})$ operations.
The numerical work in constructing $\xi_i$ and $w_i$ of expansion (\ref{23}) reduces to constructing a quadrature rule (see, e.g. \cite{ma1}) for computing the inverse Fourier transform of $\hat{k}$, the Fourier transform of the covariance kernel. Specifically, if $k$ is an integrable translation-invariant covariance kernel with Fourier transform $\hat{k}$, then $k$ satisfies
\begin{align}
k(d) = \int_{\mathbb{R}} \hat{k}(\xi) e^{i \xi d}.
\end{align}
It turns out that we can construct expansion (\ref{23}) by evaluating $\xi_i \in \mathbb{R}$ and $w_i \in \mathbb{R}^+$ such that $k(d)$ is well approximated by the sum
\begin{align}
\sum_{j=1}^n w_j \hat{k}(\xi_j) e^{i\xi_jd}
\end{align}
for a family of covariance kernels where $d$ is in some region of interest.
In this paper, we find $\xi_i$ and $w_i$ by using algorithms for constructing generalized Gaussian quadratures. The theory associated with generalized Gaussian quadratures was originally introduced in 1966~\cite{karlin1} and more recently, efficient numerical algorithms have made constructing generalized Gaussian quadratures practical in many modern problems\cite{ma1, bremer1}. The use of such quadratures is now widespread in several environments including in computational physics for the solution of integral equations with singular kernels, and in the numerical solution of certain partial differential equations (e.g.~\cite{ma1, hoskins1, bremer2}).
The GP basis function representations of this paper are closely related to those of~\cite{greengard2}. In~\cite{greengard2} a Gaussian process on an interval $[a, b]$ is approximated with a so-called Karhunen-Lo\`eve (KL) expansion, an expansion of the form
\begin{align}
\alpha_1 f_1(x) + ... + \alpha_n f_n(x)
\end{align}
where $\alpha_i \sim \mathcal{N}(0, 1)$ are iid and $f_i$ are eigenfunctions of the integral operator
\begin{align}
\int_{a}^{b} k(x, y) f(y) dy
\end{align}
where $k:[a, b] \times [a, b] \rightarrow \mathbb{R}$ is the covariance kernel.
The primary advantage of KL-expansions is that for any expansion of length $m$, the KL-expansion is optimal in the sense that its effective covariance kernel approximates the exact covariance kernel in $L^2$ better than other other expansion of length $m$.
The methods of this paper, however, have several significant advantages over those of~\cite{greengard2}.
\begin{itemize}
\item
The Fourier expansions of this paper are valid over families of covariance kernels, whereas those of~\cite{greengard2} require recomputing basis functions for kernels with different hyperparameter values.
\item Trigonometric basis functions are amenable to fast algorithms for solving the linear systems of GP regression. We capitalize on this fact to obtain $O(N\log{N} + m^3)$ computational complexity for GP regression, a significant improvement over the $O(Nm^2)$ complexity of~\cite{greengard2}.
\item
Adaptation of hyperparameters is greatly simplified. For all hyperparameter values in the domain of the Fourier expansion, GP regression is performed in $O(m^3)$ operations after a precomputation of $O(N\log{N})$ operations. Furthermore, for Bayesian methods the set of basis functions is unchanged for all hyperparamters, facilitating straightforward use of normal-normal algorithms for inference. For maximum likelihood estimation, evaluation of gradients requires $O(m^2)$ operations after the $O(m^3)$ operations of GP regression.
\end{itemize}
The remainder of this paper is structured as follows. In the following section, we introduce theoretical and numerical tools for representing GPs as Fourier expansions. In Section \ref{s20} we describe fast algorithms for GP regression. We provide the numerical results of the algorithms of this paper in Section \ref{s120}. We conclude with a brief discussion of future directions of research in Section \ref{s125}.
\section{Spectral representation of GPs}\label{s15}
Translation invariant (or stationary) covariance kernels are commonly used in practice and include kernel families such as Mat\'ern, squared-exponential, rational quadratic, periodic, and many more~\cite{rasmus1}. A translation invariant kernel is a function $k(x, y) : \mathbb{R}^d \times \mathbb{R}^d \rightarrow \mathbb{R}$ that's a function of $\|x - y\|$, or in one dimension, kernels $k(x, y)$ that only depend on $|x-y|$. In a slight abuse of notation we refer to translation invariant kernels $k(x, y)$ as functions of one variable, $k(d)$ for $d\in \mathbb{R}$.
Since translation invariant kernels are symmetric, that is $k(d) = k(-d)$, their Fourier transforms are real-valued and symmetric. However, not all symmetric functions are valid covariance kernels. Translation invariant kernels have a particular property -- $k$ is a valid kernel if and only if its Fourier transform is non-negatively valued. This property is known as Bochner's theorem~\cite{rasmus1}.
Clearly an integrable translation-invariant kernel $k$ can be expressed as the inverse Fourier transform
\begin{align}
k(d) = \int_{-\infty}^{\infty} \hat{k}(\xi) e^{2\pi i\xi d} d\xi,
\end{align}
where $\hat{k}$ denotes the Fourier transform of $k$. Or equivalently, since $\hat{k}$ is even, \begin{align}\label{107}
k(d) = \int_{0}^{\infty} 2 \hat{k}(\xi) \cos(2\pi \xi d) d\xi.
\end{align}
A discretized version of integral (\ref{107}), or a quadrature rule for approximating the integral is a sum of the form
\begin{align}\label{110}
k(d) = \sum_{j=1}^n 2 w_j \hat{k}(\xi_j) \cos(2\pi \xi_j d).
\end{align}
where $\xi_j > 0$ and $w_j \in \mathbb{R}$.
The following theorem shows that any sum of the form (\ref{110}) corresponds to a function space representation of a Gaussian process with a covariance kernel that approximates $k$ with the accuracy of discretization (\ref{110}).
\begin{theorem}\label{120}
Let $f$ be the random expansion defined by
\begin{align}\label{56}
f(x) \sim \sum_{i=1}^{m} \alpha_i \gamma_i \cos(\xi_i x) + \beta_i \gamma_i \sin(\xi_i x)
\end{align}
where for all $i,j = 1,...,m$ the random coefficients $\alpha_i$ and $\beta_j$ are iid and
\begin{align}
\alpha_i, \beta_j \sim \mathcal{N} (0, 1)
\end{align}
and where $\gamma_i$ are defined by
\begin{align}
\gamma_i = \sqrt{2 w_i \hat{k}(\xi_i)}
\end{align}
where $\hat{k}$ denotes the Fourier transform of the covariance kernel $k$ and $\xi_i, w_i > 0$.
Then $f$ is a Gaussian process with covariance kernel $k'$ defined by the formula
\begin{align}
k'(d) = \sum_{i=1}^m 2 w_j \hat{k}(\xi_j) \cos(2\pi \xi_j d).
\end{align}
\end{theorem}
\begin{proof}
Clearly,
\begin{equation}\label{111}
\begin{aligned}
E[f(x) f(y)] & = \sum_{j=1}^m 2w_j \hat{k}(\xi_j) \cos(2\pi\xi_j x)\cos(2\pi\xi_j y)
+ \sum_{j=1}^m 2w_j \hat{k}(\xi_j) \sin(2\pi\xi_j x)\sin(2\pi\xi_j y) \\
& = \sum_{j=1}^m 2 w_i \hat{k}(\xi_j)
\bigg( \cos(2\pi\xi_j x)\cos(2\pi\xi_j y) + \sin(2\pi\xi_j x)\sin(2\pi\xi_j y) \bigg)
\end{aligned}
\end{equation}
Applying standard trigonometric properties to (\ref{111}), we obtain
\begin{align}
E[f(x) f(y)]
& = \sum_{j=1}^m 2 w_j \hat{k}(\xi_j) \cos(2\pi\xi_j (x-y)).
\end{align}
\end{proof}
An immediate consequence of Theorem \ref{120} is that an $m$-point quadrature rule for the evaluation of the inverse Fourier transform of a covariance kernel provides a function-space representation of that Gaussian process. Moreover the accuracy of the quadrature rule is exactly the accuracy of the effective covariance kernel. This leads to a natural tradeoff: the more terms in an expansion, the greater the accuracy of the effective covariance kernel but the less compressed the GP representation. We now describe the numerical procedure we use for constructing the quadrature rule.
\subsection{Gaussian quadratures for covariance kernels}
In the numerical scheme of~\cite{bremer1} for constructing Gaussian quadratures, the user inputs functions $\phi_1,...,\phi_{n} : [a, b] \rightarrow \mathbb{R}$ and is returned the nodes $x_1,...,x_m \in [a, b]$ and weights $w_1,...,w_m \in \mathbb{R}^+$ such that
\begin{align}\label{121}
\bigg| \int_{a}^{b} \phi_j(x) dx
- \sum_{i=1}^m w_i \phi_j(x_i) \bigg| < \epsilon
\end{align}
for some user-specified $\epsilon > 0$ for all $j = 1,...,n$. The algorithm involves an optimization scheme to find the minimum $m$ such that condition (\ref{121}) is satisfied.
In this paper, we are concerned with constructing a particular class of quadrature rules. If $f$ is a Gaussian process defined on $[a, b]$ with covariance kernel $k$, then we seek to approximate integrals of the form
\begin{align}
\int_{0}^{\infty} 2 \hat{k}(\xi) \cos(2\pi \xi t) d\xi
\end{align}
for all $t \in [0, b-a]$. Therefore, we use the numerical scheme in ~\cite{bremer1} to construct quadrature rules for the set of functions $\phi_1,...,\phi_n$ where $\phi_j$ is defined by
\begin{align}
\phi_j(\xi) = 2\hat{k}(\xi) \cos(2\pi \xi t_j)
\end{align}
for all $\xi \in [0, \infty)$. Since the integrals
\begin{align}
\int_{0}^{\infty} 2 \hat{k}(\xi) \cos(2\pi \xi t) d\xi
\end{align}
are smooth functions in $t$, it is sufficient to choose $t_1,...,t_n$ as, for example, order-$n$ Chebyshev nodes on $[0, b-a]$~\cite{trefethen}, provided that $n$ is sufficiently large. Specifically, suppose that we construct a quadrature rule with nodes $\xi_1,...,\xi_m > 0$ and $w_1,...,w_m > 0$ such that
\begin{align}
\bigg| \int_{0}^{\infty} 2 \hat{k}(\xi) \cos(2\pi \xi t_j) d\xi
- \sum_{i=1}^m 2 w_i \hat{k}(\xi_i) \cos(2\pi \xi_i t_j) \bigg| < \epsilon
\end{align}
for some user-specified $\epsilon > 0$ and $j=1,...,n$. Then for sufficiently large $n$,
\begin{align}
\bigg| \int_{0}^{\infty} 2 \hat{k}(\xi) \cos(2\pi \xi t) d\xi
- \sum_{i=1}^m 2 w_i \hat{k}(\xi_i) \cos(2\pi \xi_i t) \bigg| < \epsilon
\end{align}
for all $t \in [0, b-a]$.
Similarly, we can use the same strategy to discretize a family of covariance functions over ranges of hyperparameters, provided that $\hat{k}$ is a smooth function of those hyperparameters. For example, suppose that $f$ is a GP defined on $[-1, 1]$ with a Mat\'ern kernel. Now consider the set of covariance kernels $ k_{\nu_i, \rho_j} $ for $i,j = 1,...,p$ where $\nu_1,...,\nu_p$ are the order-$p$ Chebyshev nodes on $[3/2, 7/2]$ and $\rho_1,...,\rho_p$ are the order-$p$ Chebyshev nodes defined on $[0.1, 0.5]$. Then $\hat{k}$ is a smooth function of $\nu, \rho$ over their domain and we can use~\cite{bremer1} to construct quadrature rules for the family of integrals
\begin{align}
\int_{0}^{\infty} 2 \hat{k}_{\nu_i, \rho_j} (\xi) \cos(2\pi \xi t_k) d\xi
\end{align}
for all $i, j \in \{1,...,p\}$ and $k \in \{1,...,n\}$ up to some tolerance $\epsilon$.
For large enough $p$ and $n$, the resulting quadrature rules satisfy
\begin{align}
\bigg| \int_{0}^{\infty} 2 \hat{k}(\xi) \cos(2\pi \xi t) d\xi
- \sum_{i=1}^m 2 w_i \hat{k}(\xi_i) \cos(2\pi \xi_i t) \bigg| < \epsilon
\end{align}
for all $t \in [0, 2]$, $\rho \in [0.1, 0.5]$, $\nu \in [3/2, 7/2]$.
We used the implementation of~\cite{serkh1} for this particular computation with $p = 100$ and $n = 200$. The code took $88$ seconds to run on a laptop and returned $86$ nodes $\xi_1,...,\xi_{86}$ and positive weights $w_1,...,w_{86}$.
The results of numerical experiments using this quadrature rule are included in Section \ref{s120} and the nodes and weights are listed in Table \ref{7813}. Figure \ref{1025} provides plots of the first few basis functions obtained from this procedure.
We now describe an algorithm for constructing Fourier representations of Gaussian processes for a range of hyperparameter values.
\begin{algorithm}\label{a10}[Construction of Fourier representations]
\begin{enumerate}
\item Set the interval on which the Gaussian process is defined, $[a, b]$, in addition to the intervals where the hyperparameters $\nu \in [\nu_0, \nu_1]$ and $\rho \in [\rho_0, \rho_1]$ are defined. Additionally, set the error tolerance $\epsilon$ for the accuracy of the quadrature, or equivalently accuracy of the effective covariance kernel.
\item Construct a quadrature rule over the region of interest. Specifically, find nodes $\xi_1,...,\xi_m \in \mathbb{R}$ and weights $w_1,...,w_m \in \mathbb{R}$ such that
\begin{align}\label{113}
\bigg| k(d) - \sum_{i=1}^m 2 w_i \hat{k}(\xi_j) \cos(\xi_j d) \bigg| < \epsilon
\end{align}
for all $\nu \in [\nu_0, \nu_1]$ , $\rho \in [\rho_0, \rho_1]$, and $d \in [0, b-a]$.
\item Define $f: [a, b] \rightarrow \mathbb{R}$ to be the random expansion
\begin{align}\label{122}
f(x) \sim \gamma_1 \alpha_1 \cos(\xi_1 x) + \beta_1 \gamma_1 \sin(\xi_1 x)
+ ... + \alpha_m \gamma_m \cos(\xi_n x) + \beta_m \gamma_m \sin(\xi_m x)
\end{align}
where for $i=1,...,m$ the $\alpha_i$ and $\beta_i$ are the iid Gaussians
\begin{align}
\alpha_i, \beta_i \sim \mathcal{N} \big(0, 1 \big),
\end{align}
and where $\gamma_i$ are defined by
\begin{align}
\gamma_i = \sqrt{2 w_i \hat{k}(\xi_i)}
\end{align}
where $\hat{k}$ denotes the Fourier transform of the covariance kernel $k$ and $w_i \in \mathbb{R}$.
Then $f$ is a Gaussian process with effective covariance kernel $k'$ defined by
\begin{align}
k'(d) = \sum_{i=1}^n 2w_i \hat{k}(\xi_j) \cos(\xi_j d).
\end{align}
\end{enumerate}
\end{algorithm}
\section{Regression}\label{s20}
Typically, GP regression is used in the following environment. An applied scientist has observed data $\{(x_i, y_i)_{i=1,...,N}\}$ where $x_i$ are independent variables that belong to some interval $[a, b]\in \mathbb{R}$ (for $1$-dimensional problems) and $y_i \in \mathbb{R}$ are dependent variables. The observed data $y_1,...,y_N$ are assumed to be observations of the form
\begin{align}
y_i = f(x_i) + \epsilon_i
\end{align}
where $\epsilon_i$ is iid Gaussian noise and $f: [a, b] \to \mathbb{R}$ is an unknown function, which is given a Gaussian process prior with covariance function $k$. Assumptions about $k$ are critical for statistical inference and typically arise from domain expertise or physical knowledge about the data-generating process. There is a large body of literature on the selection of suitable covariance kernels (e.g.~\cite{vehtari1, stein2, duvenaud1}). In many applied settings $k$ is not known a priori, but is assumed to belong to some parametric family of functions (e.g. Mat\'ern, squared-exponential, etc.) that depends on hyperparameters that are fit to the data.
The goal of GP regression is to perform statistical inference on the unknown function $f$ at a set of points $\tilde{x} \in [a, b]$ or to understand certain properties of the data-generating process. Inference typically involves evaluating the mean and covariance of the density, conditional on observed data $\vct{x}, \vct{y} \in \mathbb{R}^{N}$. The conditional distribution of $f$ at any $\tilde{x} \in \mathbb{R}$ is the Gaussian
\begin{align}
f(\tilde{x}) \, | \, \vct{x}, \vct{y} \sim \mathcal{N}(\tilde{\mu}, \tilde{\sigma}^2)
\end{align}
with
\begin{equation}\label{84}
\begin{aligned}
& \tilde{\mu} = \vct{k}(\tilde{x}, \vct{x}) (\mtx{K} + \sigma^2 \mtx{I})^{-1} \vct{y} \\
& \tilde{\sigma}^2 = k(\tilde{x}, \tilde{x}) - \vct{k}(\tilde{x}, \vct{x}) (\mtx{K} + \sigma^2 \mtx{I})^{-1} \vct{k}(\vct{x}, \tilde{x}),
\end{aligned}
\end{equation}
where $\sigma$ is the standard deviation of $\epsilon_i$ (also called the nugget), the matrix $\mtx{K}$ is the $N \times N$ covariance matrix $\mtx{K}_{i,j} = k(x_i, x_j)$, the vector $\vct{k}(\vct{x}, x') \in \mathbb{R}^N$ is the column vector such that $\vct{k}(\tilde{x}, \vct{x})_i = f(\tilde{x}, \vct{x}_i)$, and $\vct{k}(\vct{x}, \tilde{x}) = \vct{k}(\tilde{x}, \vct{x})^t$.
Since $\mtx{K} + \sigma^2\mtx{I}$ is an $N \times N$ matrix, direct inversion is computationally intractable for large $N$. However, using the Fourier representations of this paper admits a natural low-rank approximation to $\mtx{K}$. Specifically, $\mtx{K}$ is well-approximated by $\mtx{XX^{\mathsf{T}}}$ where $\mtx{X}$ is the $N \times 2m$ matrix
\begin{equation}\label{205}
\mtx{X} =
\renewcommand*{\arraystretch}{1.3}
\begin{bmatrix}
\gamma_1 \cos(\xi_1 x_1) & \dots & \gamma_k \cos(\xi_m x_1) & \gamma_1 \sin(\xi_1 x_1) & \dots & \gamma_k \sin(\xi_m x_1) \\
\gamma_1 \cos(\xi_1 x_2) & \dots & \gamma_k \cos(\xi_m x_2) & \gamma_1 \sin(\xi_1 x_2) & \dots & \gamma_1 \sin(\xi_m x_2) \\
\vdots & & \vdots & \vdots & & \vdots \\
\gamma_1\cos(\xi_1 x_n) & \dots & \gamma_k\cos(\xi_m x_N) & \gamma_1\sin(\xi_1 x_n) & \dots & \gamma_k\sin(\xi_m x_n)
\end{bmatrix}.
\end{equation}
In particular, $\mtx{K}_{i, j} = k(x_i, x_j)$ and $\mtx{XX^{\mathsf{T}}}_{i, j}$ is the quadrature rule approximation to $k(x_i, x_j)$ given by
\begin{align}
\sum_{\ell=1}^m 2\hat{k}(\xi_{\ell}) w_{\ell} \cos(2\pi \xi_{\ell} (x_i - x_j)).
\end{align}
Using $\mtx{X X^{\mathsf{T}}}$ as the reduced-rank approximation of $\mtx{K}$, the linear systems that appear in $\tilde{\mu}$ and $\tilde{\sigma}^2$,
\begin{align}\label{138}
(\mtx{K} + \sigma^2 \mtx{I}) \vct{x} = \vct{y} \quad \text{and} \quad (\mtx{K} + \sigma^2 \mtx{I}) \vct{x} = \vct{k}(\vct{x}, \tilde{x}),
\end{align}
can be approximated in $O(Nm^2)$ operations using standard direct methods.
For problems where $Nm^2$ operations is computationally infeasible, the conjugate gradient method (see, e.g., \cite{dahlquist1}) is an efficient $O(N\log{N})$ alternative to direct methods, provided $\mtx{XX^{\mathsf{T}}} + \sigma^2 \mtx{I}$ is well-conditioned. For these problems conjugate gradient has two significant advantages. First, convergence is rapid for well conditioned systems. Specifically, if $x$ is the exact solution to the linear system and $\hat{x}_n$ is the approximate solution after $n$ iterations of conjugate gradient then
\begin{align}\label{85}
\| \vct{\hat{x}}_n - \vct{x} \| = O\bigg( \bigg(\frac{\sqrt{\kappa} - 1}{\sqrt{\kappa} + 1} \bigg)^n \bigg)
\end{align}
where $\kappa$ is the condition number of the matrix (see e.g.~\cite{dahlquist1}).
The second advantage of conjugate gradient in this context is that the matrix $\mtx{XX^{\mathsf{T}}} + \sigma^2 \mtx{I}$ of the linear system can be applied in $O(N\log{N})$ time using a type 3 non-uniform FFT~\cite{barnett1}. More precisely, the type 3 non-uniform FFT computes, in $O(N\log{N})$ operations, the sums $f_{\ell}$ defined by
\begin{align}
f_{\ell} = \sum_{j=1}^N c_j e^{i \xi_{\ell} x_j}
\end{align}
for $\ell = 1,...,m$ for any $c_1,...,c_N \in \mathbb{C}$, $\xi_1,...,\xi_m \in \mathbb{R}$ and $x_1,...,x_N \in \mathbb{R}$. Contrary to the standard FFT, in the non-uniform FFT the $\xi_i$ and $x_i$ need not be uniformly spaced.
\subsection{Function-space inference}
In the function space approach to inference, the posterior density is defined over the coefficients of the basis function expansion -- in this paper a Fourier expansion. Specifically, given data $\{(x_i, y_i)\}$ and a covariance kernel $k$, Gaussian process regression becomes
\begin{equation}\label{721}
\begin{aligned}
\vct{y} & \sim \mtx{X}\vct{\beta} + \vct{\epsilon} \\
\vct{\epsilon} & \sim \mathcal{N}(0, \sigma^2 \mtx{I}) \\
\vct{\beta} & \sim \mathcal{N}(0, \mtx{I})
\end{aligned}
\end{equation}
where $\vct{\beta} \in \mathbb{R}^{2m}$ is the vector of coefficients and $\mtx{X}$ is the $N \times 2m$ matrix
(\ref{205}). The posterior density corresponding to (\ref{721}) is defined over $2m + 1$ dimensions ($2m$ coefficients $\vct{\beta}$ and $\sigma$). The conditional mean of $\vct{\beta}$ (conditioning on $\sigma, \vct{x}, \vct{y}$) is the function
\begin{align}
\sum_{i=1}^{k} \vct{\bar{\beta}}_{1,i} \cos(\xi_i x) + \vct{\bar{\beta}}_{2,i} \sin(\xi_i x)
\end{align}
where $[\vct{\bar{\beta}}_1 \, \vct{\bar{\beta}}_2]^t$ is the solution to the linear system of equations
\begin{align}
(\mtx{X^{\mathsf{T}} X} + \sigma^2 \mtx{I}) x = \mtx{X^{\mathsf{T}}}\vct{y}.
\end{align}
We now describe a numerical implementation of a solver for this system of equations that requires $O(N\log{N} + m^3)$ operations by taking advantage of the Fourier representation of GPs.
\subsection{Numerical Implementation}
For a general $N \times 2m$ matrix $\mtx{X}$, the computational complexity of the numerical solution of the linear system of equations
\begin{align}\label{91}
(\mtx{X^{\mathsf{T}} X} + \sigma^2 \mtx{I} ) \vct{x} = \mtx{X^{\mathsf{T}}}\vct{y}
\end{align}
is $O(Nm^2)$. For GP regression tasks with large amounts of data, $O(Nm^2)$ can be prohibitively expensive. However for the linear system that appears in the case of Fourier representations, we take advantage of the structure of $\mtx{X}$ to reduce evaluation to $O(N\log(N) + m^3)$ operations. We do this by constructing the $2m \times 2m$ matrix $\mtx{X^{\mathsf{T}} X}$ in $O(N \log{N})$ operations using a non-uniform fast Fourier transform (FFT)~\cite{dutt1, greengard2} whereas constructing $\mtx{X^{\mathsf{T}} X}$ ordinarily requires $O(Nm^2)$ operations for general matrices.
The $N \times 2m$ matrix $\mtx{X}$ has the form
\[
\mtx{X} =
\renewcommand*{\arraystretch}{1.3}
\begin{bmatrix}
\gamma_1 \cos(\xi_1 x_1) & \dots & \gamma_m \cos(\xi_m x_1) & \gamma_1 \sin(\xi_1 x_1) & \dots & \gamma_m \sin(\xi_m x_1) \\
\gamma_1 \cos(\xi_1 x_2) & \dots & \gamma_m \cos(\xi_m x_2) & \gamma_1 \sin(\xi_1 x_2) & \dots & \gamma_m \sin(\xi_m x_2) \\
\vdots & & \vdots & \vdots & & \vdots \\
\gamma_1\cos(\xi_1 x_n) & \dots & \gamma_m\cos(\xi_m x_N) & \gamma_1\sin(\xi_1 x_n) & \dots & \gamma_m\sin(\xi_m x_n)
\end{bmatrix}.
\]
where $\gamma_i$ is defined by
\begin{align}
\gamma_i = \sqrt{2 \hat{k}(\xi_i) w_i}.
\end{align}
for $i=1,...,m$ where $\hat{k}$ is the Fourier transform of the covariance function, and $w_i, \xi_i > 0$ are the nodes and weights of (\ref{113}).
We observe that $\mtx{X^{\mathsf{T}} X}$ can be factorized as
\begin{align}\label{97}
\mtx{X^{\mathsf{T}} X} = \mtx{B^{\mathsf{T}} D \mtx{X'}^T \mtx{X'} D B}
\end{align}
where $\mtx{X'}$ is defined by
\begin{equation}
\mtx{X'} =
\renewcommand*{\arraystretch}{1.2}
\begin{bmatrix}
e^{i\xi_1 x_1} & \dots & e^{i\xi_m x_1} & e^{i (-\xi_1) x_1} & \dots & e^{i (-\xi_m) x_1} \\
e^{i\xi_1 x_2} & \dots & e^{i\xi_m x_2} & e^{i (-\xi_1) x_2} & \dots & e^{i (-\xi_m) x_2} \\
\vdots & & \vdots & \vdots & & \vdots \\
e^{i\xi_1 x_N} & \dots & e^{i\xi_m x_N} & e^{i (-\xi_1) x_N} & \dots & e^{i (-\xi_m) x_N} \\
\end{bmatrix},
\end{equation}
where
$\mtx{D}$ is the diagonal $2m \times 2m$ matrix
\begin{equation}\label{117}
\mtx{D} =
\renewcommand*{\arraystretch}{1.2}
\begin{bmatrix}
\gamma_1 & &&&&&\\
& \ddots &&&&&\\
& & \gamma_m&&&&\\
& & & \gamma_1 &&&\\
& & & & \ddots & &\\
& & & & & \gamma_m &\\
\end{bmatrix},
\end{equation}
and $\mtx{B}$ is $2m \times 2m$ block matrix
\begin{equation}\label{118}
\mtx{B} =
\renewcommand*{\arraystretch}{1.5}
\begin{bmatrix}
\frac{1}{2} \mtx{I}_{m} & \frac{1}{2i}\mtx{I}_{m} \\
\frac{1}{2} \mtx{I}_{m} & -\frac{1}{2i}\mtx{I}_{m}
\end{bmatrix}
\end{equation}
where $\mtx{I}_{m}$ denotes the $m \times m$ identity matrix.
The matrix $\mtx{X'}^{\mathsf{T}}\mtx{X'}$ is a symmetric matrix where entry ${i, \ell}$ is given by
\begin{align}\label{223}
(\mtx{X'}^{\mathsf{T}} \mtx{X'})_{i, \ell} = \sum_{j=1}^{N} e^{i x_j (\xi_i + \xi_{\ell})},
\end{align}
for all $i, \ell \in \{1,...,2m\}$, where we denote $-\xi_i$ with $\xi_{m + i}$. The sums of (\ref{223}) can be evaluated with a type 3 non-uniform FFT, a calculation that requires $O(N \log{N})$ operations.
More precisely, we use the type 3 non-uniform FFT of~\cite{barnett1} to compute the sums
\begin{align}
\sum_{j=1}^N e^{ix_j \omega_{\ell}}
\end{align}
where $\ell = 1,...,2m^2+m$ and $\omega_{\ell} = \xi_p + \xi_q$ for $p, q \in \{1,...,2m\}$ with $q \geq p$. We also use the non-uniform FFT to compute $\mtx{X^{\mathsf{T}}}\vct{y}$, the right hand side of (\ref{91}).
Since matrix multiplications of (\ref{97}) other than $\mtx{X'}^{\mathsf{T}} \mtx{X'}$ can be applied in $O(m^2)$ operations, the total complexity of forming matrix $\mtx{X^{\mathsf{T}} X}$ using a non-uniform FFT is $O(N\log{N} + m^2)$. Once $\mtx{X^{\mathsf{T}} X}$ is formed, the resulting linear system can be solved in $O(m^3)$ operations using standard direct methods. In the following, we provide an algorithm for solving linear system (\ref{91}) in $O(N\log{N} + m^3)$ operations.
\begin{algorithm}\label{a20}[GP regression solver]
\begin{enumerate}
\item
Use Algorithm \ref{a10} to construct nodes $\xi_1,...,\xi_m$ and weights $w_1,...,w_m$ of a Fourier expansion for a certain covariance kernel or family of kernels.
\item\label{213}
Use the non-uniform FFT to compute the matrix-vector product
\begin{align}
\mtx{X^{\mathsf{T}}}\vct{y}
\end{align}
that appears in the right hand side of (\ref{91}).
\item\label{215}
Use the non-uniform FFT to compute the sums
\begin{align}
\sum_{j=1}^N e^{ix_j \omega_{\ell}}
\end{align}
where $\ell = 1,...,2m^2+m$ and $\omega_{\ell} = \xi_p + \xi_q$ for $p, q \in \{1,...,2m\}$ with $q \geq p$.
\item
Construct $\mtx{X^{\mathsf{T}} X}$ via
\begin{align}
\mtx{X^{\mathsf{T}} X} = \mtx{B}^t \mtx{D} \mtx{X'}^{\mathsf{T}} \mtx{X'} \mtx{D} \mtx{B}
\end{align}
where $\mtx{D}$ is defined in (\ref{117}) and $\mtx{B}$ is defined in (\ref{118}).
\item\label{217}
Compute the eigendecomposition of $\mtx{X^{\mathsf{T}} X}$. That is, find the $2m \times 2m$ orthogonal matrix $\mtx{U}$ and the diagonal $2m \times 2m$ matrix $\mtx{S}$ such that
\begin{align}
\mtx{X^{\mathsf{T}} X} = \mtx{USU^{\mathsf{T}}}
\end{align}
\item
Compute the solution to linear system (\ref{91}) via
\begin{align}
\vct{x} = \mtx{U} (\mtx{S} + \sigma^2\mtx{I})^{-1} \mtx{U}^t \mtx{X^{\mathsf{T}}}\vct{y}.
\end{align}
\end{enumerate}
\end{algorithm}
In Table \ref{7802} we provide computation times for solving the linear system of equations (\ref{91}) as well as forming $\mtx{X^{\mathsf{T}} X}$ and the matrix vector multiply $\mtx{X^t}\vct{y}$. Additionally, in Figure \ref{2200} we plot the total computation times for GP regression for various amounts of data.
\subsection{Adaptation of hyperparameters}
In certain GP problems, the covariance function, $k$ is known a priori, however in most environments, $k$ is assumed to belong to a certain parametric family of functions (e.g. Mat\'ern, squared-exponential, etc.) that depends on hyperparameters that are fit to the data. Methods for fitting hyperparameters to data involve performing GP regression and solving the linear system of equations (\ref{91}) for many hyperparameter values.
For those problems, the methods of this paper have several advantages:
\begin{itemize}
\item
Algorithm \ref{a10} can be used to generate a Fourier expansion that is valid over the posterior domain of hyperparameters. That expansion depends on hyperparameters by a rescaling of basis functions (or equivalently by rescaling the standard deviation of the coefficients).
\item
Using Algorithm \ref{a20}, GP regression is performed in $O(m^3)$ operations for all hyperparameters after a precomputation of $O(N\log{N})$ operations. The precomputation involves the applications of non-uniform FFTs (steps \ref{213} and \ref{215} of Algorithm \ref{a20}).
\item
Since both the observation noise and the priors on regression coefficients are Gaussian, the posterior density in the coefficients $\beta$ is also Gaussian. As a result, efficient numerical methods can be used for evaluating Bayesian posterior moments~\cite{lindley1, greeng1}, in particular the algorithm of \cite{greengard2} for fitting hyperparameters of GP models.
\item
Evaluation of gradients for maximum likelihood estimates can be performed in $O(m^2)$ operations after inversion of the posterior covariance matrix (step \ref{217} in Algorithm \ref{a20}). The component-wise formula for the gradient of the log-likelihood of the posterior density is given by
\begin{align}\label{219}
\frac{\partial}{\partial \theta_j} \log(p(\vct{y} | \vct{\theta})) = \frac{1}{2}\vct{y^{\mathsf{T}}} \mtx{C^{-1}} \frac{\partial \mtx{C}}{\partial \theta_j}
\mtx{C^{-1}} \vct{y} - \frac{1}{2}\text{tr}\bigg(\mtx{C^{-1}}\frac{\partial \mtx{C}}{\partial \theta_j}\bigg)
\end{align}
where $\theta_j \in \mathbb{R}$ is the $j$-th hyperparameter (see e.g. \cite{rasmus1}), and $\mtx{C} = (\mtx{X^{\mathsf{T}}}\mtx{X} + \sigma^2 \mtx{I})$.
\end{itemize}
\section{Numerical Experiments}\label{s120}
In this section we demonstrate the performance of the algorithms of this paper on randomly generated data. We implemented Algorithm \ref{a10} and Algorithm \ref{a20} in Fortran with the GFortran compiler on a 2.6 GHz 6-Core Intel Core i7 MacBook Pro. All examples were run in double precision arithmetic.
We used Algorithm \ref{a10} to construct Fourier expansions for the Gaussian processes defined on $[-1, 1]$ with Mat\'ern kernel, $k_{\nu, \rho}$ (see (\ref{131})) for all $\nu \in [3/2, 7/2]$ and $\rho \in [0.1, 0.5]$. We used the implementation of \cite{serkh1} for the generalized Gaussian quadrature with $\epsilon = 10^{-5} $ in Algorithm \ref{a10}. The total run time for generating the quadrature was $88$ seconds.
The output of the code was $86$ total nodes and weights and all weights were positive.
Figure \ref{2270} is a plot of the locations of the nodes obtained from this procedure. In Table \ref{7813} we list those nodes and the corresponding weights.
In Table \ref{7801} we provide the $L^2$ error of the effective covariance kernel for various values of $\nu$ and $\rho$. The $L^2$ error of the effective kernel is defined by the formula
\begin{align}\label{133}
\bigg(
\int_{-1}^{1}
\int_{-1}^{1}
(k'(x, y) - k(x, y))^2
dx \, dy
\bigg)^{1/2}
\end{align}
where $k'$ denotes the effective covariance kernel and $k$ denotes the true kernel.
Table \ref{7802} contains the results of numerical experiments for Gaussian process regression using Algorithms \ref{a10} and \ref{a20}. The data, $\{(x_i, y_i)\}$ was randomly generated on the interval $[-1, 1]$ via
\begin{align}
y_i = \cos(3e^{x_i}) + \epsilon_i
\end{align}
where
\begin{align}
\epsilon_i \sim \mathcal{N}(0, 0.5).
\end{align}
The $x_i$ are equispaced on $[-1, 1]$. The nodes and weights of the Fourier expansions used are given by Table \ref{7801}. We used Algorithm \ref{a20} to compute the conditional mean and covariance and report timings in Table \ref{7802} where $N$ denotes the number of data points, $\nu,\rho$ denote the hyperparameters of the Mat\'ern kernel, and the column ``$\sigma^2$" denotes the residual variance (or nugget). The $L^2$ error of the effective covariance kernel (see (\ref{133})) is provided in the column ``$L^2$ error". We include timings for solving the regression problem via Algorithm \ref{a20}. In column ``FFT time (s)" we provide the total time for formation of the matrix $\mtx{X^{\mathsf{T}} X}$ and the matrix vector multiply $\mtx{X^{\mathsf{T}}} \vct{y}$. The column ``solve time (s)" denotes the remaining time in the regression solution, which involves inversion of $(\mtx{X^{\mathsf{T}} X} + \sigma^2 \mtx{I})$.
\begin{figure}[!ht]
\centering
\begin{tikzpicture}
\begin{axis}[
width=17cm,
height=3cm,
xlabel={$\xi$},
xmin = -1,
xmax=51,
ymin=-0.5,
ymax=0.5,
ytick=\empty,
]
\addplot[only marks, mark=*]
coordinates
{
( 0.0960748783232733, 0.0000000000000000)
( 0.2949311558758212, 0.0000000000000000)
( 0.5121811831688609, 0.0000000000000000)
( 0.7553750530381257, 0.0000000000000000)
( 1.0272554775524241, 0.0000000000000000)
( 1.3267469591800769, 0.0000000000000000)
( 1.6509109783729810, 0.0000000000000000)
( 1.9964656597608781, 0.0000000000000000)
( 2.3604331716760738, 0.0000000000000000)
( 2.7402749478447399, 0.0000000000000000)
( 3.1338448880682450, 0.0000000000000000)
( 3.5393213277133500, 0.0000000000000000)
( 3.9551382871240710, 0.0000000000000000)
( 4.3799500415428971, 0.0000000000000000)
( 4.8125843418627330, 0.0000000000000000)
( 5.2520252807076444, 0.0000000000000000)
( 5.6973962780821781, 0.0000000000000000)
( 6.1479398698978196, 0.0000000000000000)
( 6.6029703189280022, 0.0000000000000000)
( 7.0619657471053916, 0.0000000000000000)
( 7.5244050646589304, 0.0000000000000000)
( 7.9898781249339503, 0.0000000000000000)
( 8.4580188390522402, 0.0000000000000000)
( 8.9284697412154603, 0.0000000000000000)
( 9.4010606000280621, 0.0000000000000000)
( 9.8755021184860947, 0.0000000000000000)
( 10.3515866576509996, 0.0000000000000000)
( 10.8290813405759305, 0.0000000000000000)
( 11.3078856905533307, 0.0000000000000000)
( 11.7878547186769005, 0.0000000000000000)
( 12.2690444362196605, 0.0000000000000000)
( 12.7507923286660407, 0.0000000000000000)
( 13.2340308273640499, 0.0000000000000000)
( 13.7168180215943405, 0.0000000000000000)
( 14.2023546685620694, 0.0000000000000000)
( 14.6852339737999902, 0.0000000000000000)
( 15.1736671929864606, 0.0000000000000000)
( 15.6561205601530808, 0.0000000000000000)
( 16.1465532564475396, 0.0000000000000000)
( 16.6287678133340897, 0.0000000000000000)
( 17.1217168205720291, 0.0000000000000000)
( 17.6027784484422583, 0.0000000000000000)
( 18.0975219052295202, 0.0000000000000000)
( 18.5783668916359304, 0.0000000000000000)
( 19.0739289334232396, 0.0000000000000000)
( 19.5535187961381212, 0.0000000000000000)
( 20.0540157049340202, 0.0000000000000000)
( 20.5303087220011804, 0.0000000000000000)
( 21.0308285956013101, 0.0000000000000000)
( 21.5033438976054896, 0.0000000000000000)
( 22.0147025165826093, 0.0000000000000000)
( 22.4786227659423616, 0.0000000000000000)
( 22.9890428278831500, 0.0000000000000000)
( 23.4620311240316504, 0.0000000000000000)
( 23.9659051179655798, 0.0000000000000000)
( 24.4374955606510014, 0.0000000000000000)
( 24.9268582817982995, 0.0000000000000000)
( 25.4355862458376691, 0.0000000000000000)
( 25.8908822711706499, 0.0000000000000000)
( 26.4135765170375585, 0.0000000000000000)
( 26.8244665781450813, 0.0000000000000000)
( 27.3929596371526216, 0.0000000000000000)
( 27.8810764015625807, 0.0000000000000000)
( 28.2872085273741511, 0.0000000000000000)
( 28.9186103516193818, 0.0000000000000000)
( 29.3361902609053793, 0.0000000000000000)
( 29.7475402852299098, 0.0000000000000000)
( 30.2919596593750207, 0.0000000000000000)
( 31.1707837877708087, 0.0000000000000000)
( 31.7921709344750916, 0.0000000000000000)
( 32.0844373549515112, 0.0000000000000000)
( 32.8290834819148998, 0.0000000000000000)
( 33.6870452221527970, 0.0000000000000000)
( 34.6556177080743311, 0.0000000000000000)
( 35.5973697159031417, 0.0000000000000000)
( 36.0608079398967192, 0.0000000000000000)
( 37.4828497993489194, 0.0000000000000000)
( 38.1056387575873927, 0.0000000000000000)
( 38.4560335475650206, 0.0000000000000000)
( 39.9230354160448471, 0.0000000000000000)
( 41.8408663755605872, 0.0000000000000000)
( 43.7213283268615385, 0.0000000000000000)
( 44.2767860363890975, 0.0000000000000000)
( 45.6320859437675992, 0.0000000000000000)
( 47.5466594857420191, 0.0000000000000000)
( 49.4591701554423580, 0.0000000000000000)
};
\end{axis}
\end{tikzpicture}
\caption{\em Location of the $86$ nodes for GPs defined on $[-1, 1]$ with Mat\'ern kernels with $\nu \in [1.5, 3.5]$, and $\rho \in [0.1, 0.5]$.}
\label{2270}
\end{figure}
\begin{table}[!ht]
\centering
\resizebox{!}{8cm}{%
\begin{tabular}{cccccc}
$i$ & nodes & weights & $i$ & nodes & weights \\
\hline
$ 1$&$ 0.0960748783232733$&$ 0.1933002284075283$& $ 44$&$ 18.5783668916359304 $ & $ 0.4901095181946152$ \\
$ 2$&$ 0.2949311558758212$&$ 0.2064005047360611$&$ 45$&$ 19.0739289334232396$&$ 0.4936700008508603$ \\
$ 3$&$ 0.5121811831688609$&$ 0.2293690006634405$&$ 46$&$ 19.5535187961381212$&$ 0.4806551264127251$ \\
$ 4$&$ 0.7553750530381257$&$ 0.2574584689846101$&$ 47$&$ 20.0540157049340202$&$ 0.4919980518793062$ \\
$ 5$&$ 1.0272554775524241$&$ 0.2860863008101612$&$ 48$&$ 20.5303087220011804$&$ 0.4886725084331118$ \\
$ 6$&$ 1.3267469591800769$&$ 0.3123806483750251$&$ 49$&$ 21.0308285956013101$&$ 0.4985884540483315$ \\
$ 7$&$ 1.6509109783729810$&$ 0.3353906370922179$&$ 50$&$ 21.5033438976054896$&$ 0.4525339724563319$ \\
$ 8$&$ 1.9964656597608781$&$ 0.3552203181130367$&$ 51$&$ 22.0147025165826093$&$ 0.5300911113825829$ \\
$ 9$&$ 2.3604331716760738$&$ 0.3722929241325456$&$ 52$&$ 22.4786227659423616$&$ 0.4703295964417554$ \\
$ 10$&$ 2.7402749478447399$&$ 0.3870336516793261$&$ 53$&$ 22.9890428278831500$&$ 0.5031381396430530$ \\
$ 11$&$ 3.1338448880682450$&$ 0.3998022537083804$&$ 54$&$ 23.4620311240316504$&$ 0.4385610519729455$ \\
$ 12$&$ 3.5393213277133500$&$ 0.4108896687408796$&$ 55$&$ 23.9659051179655798$&$ 0.5529348248583776$ \\
$ 13$&$ 3.9551382871240710$&$ 0.4205230182176792$&$ 56$&$ 24.4374955606510014$&$ 0.4879736648158969$ \\
$ 14$&$ 4.3799500415428971$&$ 0.4289035535544789$&$ 57$&$ 24.9268582817982995$&$ 0.4320150910275336$ \\
$ 15$&$ 4.8125843418627330$&$ 0.4361897874230700$&$ 58$&$ 25.4355862458376691$&$ 0.5030088994396542$ \\
$ 16$&$ 5.2520252807076444$&$ 0.4425587482527165$&$ 59$&$ 25.8908822711706499$&$ 0.4986129817647325$ \\
$ 17$&$ 5.6973962780821781$&$ 0.4480785412063698$&$ 60$&$ 26.4135765170375585$&$ 0.5698981287926346$ \\
$ 18$&$ 6.1479398698978196$&$ 0.4528922343761104$&$ 61$&$ 26.8244665781450813$&$ 0.3580010222446956$ \\
$ 19$&$ 6.6029703189280022$&$ 0.4571127652712647$&$ 62$&$ 27.3929596371526216$&$ 0.4477770521368305$ \\
$ 20$&$ 7.0619657471053916$&$ 0.4607536593597344$&$ 63$&$ 27.8810764015625807$&$ 0.6306236380217451$ \\
$ 21$&$ 7.5244050646589304$&$ 0.4640030341808996$&$ 64$&$ 28.2872085273741511$&$ 0.5684848926018280$ \\
$ 22$&$ 7.9898781249339503$&$ 0.4668744890692324$&$ 65$&$ 28.9186103516193818$&$ 0.1661641302432533$ \\
$ 23$&$ 8.4580188390522402$&$ 0.4693377032983527$&$ 66$&$ 29.3361902609053793$&$ 0.7145700506865926$ \\
$ 24$&$ 8.9284697412154603$&$ 0.4715695985713800$&$ 67$&$ 29.7475402852299098$&$ 0.3653264402765759$ \\
$ 25$&$ 9.4010606000280621$&$ 0.4735626358176499$&$ 68$&$ 30.2919596593750207$&$ 0.7987670620900847$ \\
$ 26$&$ 9.8755021184860947$&$ 0.4753261760015748$&$ 69$&$ 31.1707837877708087$&$ 0.6496969625503436$ \\
$ 27$&$ 10.3515866576509996$&$ 0.4768203234940871$&$ 70$&$ 31.7921709344750916$&$ 0.6374198048803309$ \\
$ 28$&$ 10.8290813405759305$&$ 0.4782250403991462$&$ 71$&$ 32.0844373549515112$&$ 0.4525776393523478$ \\
$ 29$&$ 11.3078856905533307$&$ 0.4795013762603529$&$ 72$&$ 32.8290834819148998$&$ 0.5792967675329964$ \\
$ 30$&$ 11.7878547186769005$&$ 0.4806019700535451$&$ 73$&$ 33.6870452221527970$&$ 1.2316989151794200$ \\
$ 31$&$ 12.2690444362196605$&$ 0.4816049378992831$&$ 74$&$ 34.6556177080743311$&$ 0.6653181217090052$ \\
$ 32$&$ 12.7507923286660407$&$ 0.4826033499503408$&$ 75$&$ 35.5973697159031417$&$ 0.7979748203948971$ \\
$ 33$&$ 13.2340308273640499$&$ 0.4830779751411056$&$ 76$&$ 36.0608079398967192$&$ 0.9871538295211217$ \\
$ 34$&$ 13.7168180215943405$&$ 0.4845298908539934$&$ 77$&$ 37.4828497993489194$&$ 1.1429690155529550$ \\
$ 35$&$ 14.2023546685620694$&$ 0.4833485049199899$&$ 78$&$ 38.1056387575873927$&$ 0.3983778654241495$ \\
$ 36$&$ 14.6852339737999902$&$ 0.4872979567198063$&$ 79$&$ 38.4560335475650206$&$ 0.7492963615598504$ \\
$ 37$&$ 15.1736671929864606$&$ 0.4828723127156595$&$ 80$&$ 39.9230354160448471$&$ 1.7911442981045280$ \\
$ 38$&$ 15.6561205601530808$&$ 0.4883987918117061$&$ 81$&$ 41.8408663755605872$&$ 1.9661520413352620$ \\
$ 39$&$ 16.1465532564475396$&$ 0.4852092322682524$&$ 82$&$ 43.7213283268615385$&$ 1.6715931870761731$ \\
$ 40$&$ 16.6287678133340897$&$ 0.4863731168868852$&$ 83$&$ 44.2767860363890975$&$ 0.2685526601061519$ \\
$ 41$&$ 17.1217168205720291$&$ 0.4871456130687457$&$ 84$&$ 45.6320859437675992$&$ 1.7725506245674350$ \\
$ 42$&$ 17.6027784484422583$&$ 0.4894715943022971$&$ 85$&$ 47.5466594857420191$&$ 1.9715488370126710$ \\
$ 43$&$ 18.0975219052295202$&$ 0.4843933188273860$&$ 86$&$ 49.4591701554423580$&$ 1.5256479816263220$ \\
\end{tabular}
}
\caption{\em Nodes and weights for GPs defined on $[-1, 1]$ with Mat\'ern kernels with $\nu \in [0.5, 3.5]$, and $\rho \in [0.1, 0.5]$ }
\label{7813}
\end{table}
\begin{table}[!ht]
\centering
\begin{tabular}{ccc}
$\nu$ & $\rho$ & $L^2$ error \\
\hline
$ 1.5 $ & $ 0.1 $ & $ 0.780 \times 10^{-4}$ \\
$ 1.5 $ & $ 0.3 $ & $ 0.295 \times 10^{-5}$ \\
$ 1.5 $ & $ 0.5 $ & $ 0.140 \times 10^{-5}$ \\
$ 2.0 $ & $ 0.1 $ & $ 0.141 \times 10^{-4}$ \\
$ 2.0 $ & $ 0.3 $ & $ 0.611 \times 10^{-6}$ \\
$ 2.0 $ & $ 0.5 $ & $ 0.118 \times 10^{-4}$ \\
$ 2.5 $ & $ 0.1 $ & $ 0.326 \times 10^{-5}$ \\
$ 2.5 $ & $ 0.3 $ & $ 0.608 \times 10^{-6}$ \\
$ 2.5 $ & $ 0.5 $ & $ 0.445 \times 10^{-6}$ \\
$ 3.0 $ & $ 0.1 $ & $ 0.113 \times 10^{-5}$ \\
$ 3.0 $ & $ 0.3 $ & $ 0.577 \times 10^{-6}$ \\
$ 3.0 $ & $ 0.5 $ & $ 0.239 \times 10^{-6}$ \\
$ 3.5 $ & $ 0.1 $ & $ 0.693 \times 10^{-6}$ \\
$ 3.5 $ & $ 0.3 $ & $ 0.630 \times 10^{-6}$ \\
$ 3.5 $ & $ 0.5 $ & $ 0.222 \times 10^{-6}$ \\
\end{tabular}
\caption{\em Accuracy of Fourier expansion with nodes and weights of Table \ref{7813} for Mat\'ern kernels with various hyperparameter values.}
\label{7801}
\end{table}
\begin{table}[!ht]
\centering
\begin{tabular}{lccccccc}
$N$ & $\nu$ & $\rho$ & $\sigma^2 $ & $L^2$ error & FFT time (s) & solve time (s) & total time (s) \\
\hline
$ 10^5 $ & $ 3.0 $ & $ 0.1 $ & $0.5$ & $0.113 \times 10^{-5}$ & 0.03 & 0.004 & 0.035 \\
$ 10^6 $ & $ 2.0 $ & $ 0.5 $ & $0.6$ & $0.630 \times 10^{-6}$ & 0.11 & 0.005 & 0.12 \\
$ 10^7 $ & $ 1.5 $ & $ 0.1 $ & $3.0$ & $0.630 \times 10^{-6}$ & 1.1 & 0.004 & 1.1 \\
$ 10^8 $ & $ 3.5 $ & $ 0.3 $ & $0.95$ & $0.630 \times 10^{-6}$ & 12.0 & 0.005 & 12.0
\end{tabular}
\caption{\em Computation times and accuracy of effective kernel for GP regression nodes and weights of Table \ref{7813}. The GP is defined on $[-1, 1]$ with Mat\'ern kernels with $\nu \in [0.5, 3.5]$, and $\rho \in [0.1, 0.5]$. }
\label{7802}
\end{table}
\begin{figure}[!ht]
\centering
\begin{tikzpicture}[scale=0.9]
\centering
\begin{axis}[
xmode=log,
ymode=log,
xmin=50000, xmax=200000000,
ymin=0.002, ymax=30,
xtick={0.0,100000,1000000,10000000,100000000},
xlabel=$N$,
ytick={0, 0.1, 1.0, 10.0},
ylabel=time (s),
legend pos= north west
]
\addplot[line width=0.5mm, mark=*, color=blue]
coordinates
{
( 100000, 0.035)
( 1000000, 0.13)
( 10000000, 1.1)
( 100000000, 12.0)
};
\addplot[line width=0.5mm, mark=*, color=red, dashed]
coordinates
{
( 100000, 0.005)
( 1000000, 0.06)
( 10000000, .70)
( 100000000, 8.0)
};
\legend{Algorithm \ref{a20}, $N\log{N}$}
\end{axis}
\end{tikzpicture}
\caption{Scaling times for evaluation of conditional mean for varying amounts of data. We include a plot proportional to $N \log{N}$ for comparison.}
\label{2200}
\end{figure}
\begin{figure}[!ht]
\centering
\begin{subfigure}{0.45\linewidth}
\begin{tikzpicture}[scale=0.9]
\centering
\begin{axis}[
xmin=-1, xmax=1,
ymin=-3, ymax=3,
xtick={-1,-0.5,0.0,0.5,1},
xlabel=$x$,
ytick={-3,-2,-1,0,1,2,3},
ylabel=$\cos(2\pi\xi_i x)$,
legend pos= south west
]
\addplot[dotted, color=purple, line width=0.3mm]
coordinates
{
( -1.0000000000000000, 0.8232656209879091)
( -0.8947368421052632, 0.8576503217889460)
( -0.7894736842105263, 0.8885732675455514)
( -0.6842105263157895, 0.9159096432127798)
( -0.5789473684210527, 0.9395491103065678)
( -0.4736842105263158, 0.9593962522656663)
( -0.3684210526315790, 0.9753709595838473)
( -0.2631578947368421, 0.9874087531578716)
( -0.1578947368421053, 0.9954610445460746)
( -0.0526315789473685, 0.9994953320870869)
( 0.0526315789473684, 0.9994953320870869)
( 0.1578947368421053, 0.9954610445460746)
( 0.2631578947368420, 0.9874087531578716)
( 0.3684210526315788, 0.9753709595838473)
( 0.4736842105263157, 0.9593962522656663)
( 0.5789473684210527, 0.9395491103065678)
( 0.6842105263157894, 0.9159096432127798)
( 0.7894736842105261, 0.8885732675455514)
( 0.8947368421052631, 0.8576503217889461)
( 1.0000000000000000, 0.8232656209879091)
};
\addplot[dashed, color=red, line width=0.3mm]
coordinates
{
( -1.0000000000000000, -0.2785756949406452)
( -0.9487179487179487, -0.1861867991666161)
( -0.8974358974358975, -0.0921177310123569)
( -0.8461538461538461, 0.0027826187798550)
( -0.7948717948717949, 0.0976578578808709)
( -0.7435897435897436, 0.1916518205634818)
( -0.6923076923076923, 0.2839162938463456)
( -0.6410256410256410, 0.3736186718713654)
( -0.5897435897435898, 0.4599494694404864)
( -0.5384615384615385, 0.5421296269088355)
( -0.4871794871794872, 0.6194175405139585)
( -0.4358974358974359, 0.6911157546986072)
( -0.3846153846153846, 0.7565772560347405)
( -0.3333333333333334, 0.8152113119515976)
( -0.2820512820512820, 0.8664888015784548)
( -0.2307692307692308, 0.9099469905958851)
( -0.1794871794871795, 0.9451937070066866)
( -0.1282051282051282, 0.9719108801438068)
( -0.0769230769230770, 0.9898574109788304)
( -0.0256410256410257, 0.9988713478290127)
( 0.0256410256410255, 0.9988713478290127)
( 0.0769230769230769, 0.9898574109788305)
( 0.1282051282051282, 0.9719108801438068)
( 0.1794871794871795, 0.9451937070066866)
( 0.2307692307692308, 0.9099469905958851)
( 0.2820512820512819, 0.8664888015784549)
( 0.3333333333333333, 0.8152113119515977)
( 0.3846153846153846, 0.7565772560347405)
( 0.4358974358974359, 0.6911157546986072)
( 0.4871794871794872, 0.6194175405139585)
( 0.5384615384615383, 0.5421296269088359)
( 0.5897435897435896, 0.4599494694404866)
( 0.6410256410256410, 0.3736186718713654)
( 0.6923076923076923, 0.2839162938463456)
( 0.7435897435897436, 0.1916518205634818)
( 0.7948717948717947, 0.0976578578808713)
( 0.8461538461538460, 0.0027826187798552)
( 0.8974358974358974, -0.0921177310123567)
( 0.9487179487179487, -0.1861867991666161)
( 1.0000000000000000, -0.2785756949406452)
};
\addplot[
dash dot, color=black, line width=0.3mm
]
coordinates
{
( -1.0000000000000000, -0.9970725015397894)
( -0.9487179487179487, -0.9960868156945284)
( -0.8974358974358975, -0.9680335747831501)
( -0.8461538461538461, -0.9136750945307376)
( -0.7948717948717949, -0.8344885063824219)
( -0.7435897435897436, -0.7326256180937423)
( -0.6923076923076923, -0.6108544407401442)
( -0.6410256410256410, -0.4724839710843112)
( -0.5897435897435898, -0.3212742732571218)
( -0.5384615384615385, -0.1613343031828786)
( -0.4871794871794872, 0.0029897477432044)
( -0.4358974358974359, 0.1672325555884659)
( -0.3846153846153846, 0.3269310041109083)
( -0.3333333333333334, 0.4777454649957640)
( -0.2820512820512820, 0.6155777224445229)
( -0.2307692307692308, 0.7366823376464313)
( -0.1794871794871795, 0.8377684269264030)
( -0.1282051282051282, 0.9160890878609811)
( -0.0769230769230770, 0.9695160433067891)
( -0.0256410256410257, 0.9965974749727319)
( 0.0256410256410255, 0.9965974749727319)
( 0.0769230769230769, 0.9695160433067892)
( 0.1282051282051282, 0.9160890878609811)
( 0.1794871794871795, 0.8377684269264030)
( 0.2307692307692308, 0.7366823376464313)
( 0.2820512820512819, 0.6155777224445232)
( 0.3333333333333333, 0.4777454649957643)
( 0.3846153846153846, 0.3269310041109083)
( 0.4358974358974359, 0.1672325555884659)
( 0.4871794871794872, 0.0029897477432044)
( 0.5384615384615383, -0.1613343031828779)
( 0.5897435897435896, -0.3212742732571214)
( 0.6410256410256410, -0.4724839710843112)
( 0.6923076923076923, -0.6108544407401442)
( 0.7435897435897436, -0.7326256180937423)
( 0.7948717948717947, -0.8344885063824213)
( 0.8461538461538460, -0.9136750945307375)
( 0.8974358974358974, -0.9680335747831501)
( 0.9487179487179487, -0.9960868156945284)
( 1.0000000000000000, -0.9970725015397894)
};
\addplot[color=blue, line width=0.3mm
]
coordinates
{
( -1.0000000000000000, 0.0337660346170957)
( -0.9487179487179487, -0.2080886632398184)
( -0.8974358974358975, -0.4376769029983298)
( -0.8461538461538461, -0.6414648646077155)
( -0.7948717948717949, -0.8074396087126023)
( -0.7435897435897436, -0.9258172181487528)
( -0.6923076923076923, -0.9896195425290694)
( -0.6410256410256410, -0.9950855478481102)
( -0.5897435897435898, -0.9418930227745723)
( -0.5384615384615385, -0.8331775724385186)
( -0.4871794871794872, -0.6753477800632417)
( -0.4358974358974359, -0.4777074323361474)
( -0.3846153846153846, -0.2519070776638734)
( -0.3333333333333334, -0.0112572469810134)
( -0.2820512820512820, 0.2300561784707598)
( -0.2307692307692308, 0.4578081997136078)
( -0.1794871794871795, 0.6585732385257039)
( -0.1282051282051282, 0.8205165514958970)
( -0.0769230769230770, 0.9340918671842293)
( -0.0256410256410257, 0.9926041219134548)
( 0.0256410256410255, 0.9926041219134549)
( 0.0769230769230769, 0.9340918671842295)
( 0.1282051282051282, 0.8205165514958970)
( 0.1794871794871795, 0.6585732385257039)
( 0.2307692307692308, 0.4578081997136078)
( 0.2820512820512819, 0.2300561784707604)
( 0.3333333333333333, -0.0112572469810128)
( 0.3846153846153846, -0.2519070776638734)
( 0.4358974358974359, -0.4777074323361474)
( 0.4871794871794872, -0.6753477800632417)
( 0.5384615384615383, -0.8331775724385182)
( 0.5897435897435896, -0.9418930227745721)
( 0.6410256410256410, -0.9950855478481102)
( 0.6923076923076923, -0.9896195425290694)
( 0.7435897435897436, -0.9258172181487528)
( 0.7948717948717947, -0.8074396087126029)
( 0.8461538461538460, -0.6414648646077155)
( 0.8974358974358974, -0.4376769029983306)
( 0.9487179487179487, -0.2080886632398184)
( 1.0000000000000000, 0.0337660346170957)
};
\end{axis}
\end{tikzpicture}
\label{2201}
\end{subfigure}
\quad
\begin{subfigure}{.45\linewidth}
\centering
\begin{tikzpicture}[scale=0.9]
\begin{axis}[
xmin=-1, xmax=1,
ymin=-3, ymax=3,
xtick={-1,-0.5,0.0,0.5,1},
xlabel=$x$,
ytick={-3,-2,-1,0,1,2,3},
ylabel=$\sin(2\pi \xi_i x)$,
]
\addplot[dotted, color=purple, line width=0.3mm]
coordinates
{
( -1.0000000000000000, -0.5676563373198544)
( -0.9487179487179487, -0.5419028186482554)
( -0.8974358974358975, -0.5156300255592384)
( -0.8461538461538461, -0.4888631337648602)
( -0.7948717948717949, -0.4616277924436535)
( -0.7435897435897436, -0.4339500996624873)
( -0.6923076923076923, -0.4058565773682860)
( -0.6410256410256410, -0.3773741459735676)
( -0.5897435897435898, -0.3485300985601544)
( -0.5384615384615385, -0.3193520747257771)
( -0.4871794871794872, -0.2898680340986320)
( -0.4358974358974359, -0.2601062295452721)
( -0.3846153846153846, -0.2300951800975030)
( -0.3333333333333334, -0.1998636436242303)
( -0.2820512820512820, -0.1694405892744405)
( -0.2307692307692308, -0.1388551697177270)
( -0.1794871794871795, -0.1081366932089560)
( -0.1282051282051282, -0.0773145955038463)
( -0.0769230769230770, -0.0464184116523714)
( -0.0256410256410257, -0.0154777476970128)
( 0.0256410256410255, 0.0154777476970128)
( 0.0769230769230769, 0.0464184116523713)
( 0.1282051282051282, 0.0773145955038463)
( 0.1794871794871795, 0.1081366932089560)
( 0.2307692307692308, 0.1388551697177270)
( 0.2820512820512819, 0.1694405892744405)
( 0.3333333333333333, 0.1998636436242302)
( 0.3846153846153846, 0.2300951800975030)
( 0.4358974358974359, 0.2601062295452721)
( 0.4871794871794872, 0.2898680340986320)
( 0.5384615384615383, 0.3193520747257770)
( 0.5897435897435896, 0.3485300985601543)
( 0.6410256410256410, 0.3773741459735676)
( 0.6923076923076923, 0.4058565773682860)
( 0.7435897435897436, 0.4339500996624873)
( 0.7948717948717947, 0.4616277924436534)
( 0.8461538461538460, 0.4888631337648602)
( 0.8974358974358974, 0.5156300255592383)
( 0.9487179487179487, 0.5419028186482554)
( 1.0000000000000000, 0.5676563373198544)
};
\addplot[dashed, color=red, line width=0.3mm]
coordinates
{
( -1.0000000000000000, -0.9604142763351327)
( -0.9487179487179487, -0.9825143641779952)
( -0.8974358974358975, -0.9957481225857948)
( -0.8461538461538461, -0.9999961285088688)
( -0.7948717948717949, -0.9952200474237441)
( -0.7435897435897436, -0.9814629792685524)
( -0.6923076923076923, -0.9588490695039317)
( -0.6410256410256410, -0.9275823888092513)
( -0.5897435897435898, -0.8879450915239157)
( -0.5384615384615385, -0.8402948694521981)
( -0.4871794871794872, -0.7850617240087805)
( -0.4358974358974359, -0.7227440858335505)
( -0.3846153846153846, -0.6539043168927261)
( -0.3333333333333334, -0.5791636356558957)
( -0.2820512820512820, -0.4991965111447928)
( -0.2307692307692308, -0.4147245764425977)
( -0.1794871794871795, -0.3265101165889931)
( -0.1282051282051282, -0.2353491896270112)
( -0.0769230769230770, -0.1420644428781768)
( -0.0256410256410257, -0.0474976892727601)
( 0.0256410256410255, 0.0474976892727599)
( 0.0769230769230769, 0.1420644428781766)
( 0.1282051282051282, 0.2353491896270112)
( 0.1794871794871795, 0.3265101165889931)
( 0.2307692307692308, 0.4147245764425977)
( 0.2820512820512819, 0.4991965111447926)
( 0.3333333333333333, 0.5791636356558956)
( 0.3846153846153846, 0.6539043168927261)
( 0.4358974358974359, 0.7227440858335505)
( 0.4871794871794872, 0.7850617240087805)
( 0.5384615384615383, 0.8402948694521979)
( 0.5897435897435896, 0.8879450915239155)
( 0.6410256410256410, 0.9275823888092513)
( 0.6923076923076923, 0.9588490695039317)
( 0.7435897435897436, 0.9814629792685524)
( 0.7948717948717947, 0.9952200474237441)
( 0.8461538461538460, 0.9999961285088688)
( 0.8974358974358974, 0.9957481225857948)
( 0.9487179487179487, 0.9825143641779952)
( 1.0000000000000000, 0.9604142763351327)
};
\addplot[
dash dot, color=black, line width=0.3mm
]
coordinates
{
( -1.0000000000000000, 0.0764619295675085)
( -0.9487179487179487, -0.0883801765077133)
( -0.8974358974358975, -0.2508206492547120)
( -0.8461538461538461, -0.4064453488899187)
( -0.7948717948717949, -0.5510253467088739)
( -0.7435897435897436, -0.6806318415360555)
( -0.6923076923076923, -0.7917429205417916)
( -0.6410256410256410, -0.8813392633194097)
( -0.5897435897435898, -0.9469861885704079)
( -0.5384615384615385, -0.9868998138699262)
( -0.4871794871794872, -0.9999955306942286)
( -0.4358974358974359, -0.9859174774550610)
( -0.3846153846153846, -0.9450482096438432)
( -0.3333333333333334, -0.8784983043102481)
( -0.2820512820512820, -0.7880761813619379)
( -0.2307692307692308, -0.6762389617581860)
( -0.1794871794871795, -0.5460256979714969)
( -0.1282051282051282, -0.4009747911054207)
( -0.0769230769230770, -0.2450278387668393)
( -0.0256410256410257, -0.0824225265202121)
( 0.0256410256410255, 0.0824225265202117)
( 0.0769230769230769, 0.2450278387668390)
( 0.1282051282051282, 0.4009747911054207)
( 0.1794871794871795, 0.5460256979714969)
( 0.2307692307692308, 0.6762389617581860)
( 0.2820512820512819, 0.7880761813619376)
( 0.3333333333333333, 0.8784983043102479)
( 0.3846153846153846, 0.9450482096438432)
( 0.4358974358974359, 0.9859174774550610)
( 0.4871794871794872, 0.9999955306942286)
( 0.5384615384615383, 0.9868998138699263)
( 0.5897435897435896, 0.9469861885704081)
( 0.6410256410256410, 0.8813392633194097)
( 0.6923076923076923, 0.7917429205417916)
( 0.7435897435897436, 0.6806318415360555)
( 0.7948717948717947, 0.5510253467088747)
( 0.8461538461538460, 0.4064453488899191)
( 0.8974358974358974, 0.2508206492547120)
( 0.9487179487179487, 0.0883801765077133)
( 1.0000000000000000, -0.0764619295675085)
};
\addplot[color=blue, line width=0.3mm
]
coordinates
{
( -1.0000000000000000, 0.9994297648690663)
( -0.9487179487179487, 0.9781099673508421)
( -0.8974358974358975, 0.8991323198405174)
( -0.8461538461538461, 0.7671524147611121)
( -0.7948717948717949, 0.5899502337333545)
( -0.7435897435897436, 0.3779715314402721)
( -0.6923076923076923, 0.1437120768918033)
( -0.6410256410256410, -0.0990189500238535)
( -0.5897435897435898, -0.3359129852336449)
( -0.5384615384615385, -0.5530055449861755)
( -0.4871794871794872, -0.7374994074327459)
( -0.4358974358974359, -0.8785189861868696)
( -0.3846153846153846, -0.9677514268771952)
( -0.3333333333333334, -0.9999366351876545)
( -0.2820512820512820, -0.9731773500998828)
( -0.2307692307692308, -0.8890509840695220)
( -0.1794871794871795, -0.7525166373561227)
( -0.1282051282051282, -0.5716227678471887)
( -0.0769230769230770, -0.3570327487224108)
( -0.0256410256410257, -0.1213962814933773)
( 0.0256410256410255, 0.1213962814933767)
( 0.0769230769230769, 0.3570327487224103)
( 0.1282051282051282, 0.5716227678471887)
( 0.1794871794871795, 0.7525166373561227)
( 0.2307692307692308, 0.8890509840695220)
( 0.2820512820512819, 0.9731773500998827)
( 0.3333333333333333, 0.9999366351876545)
( 0.3846153846153846, 0.9677514268771952)
( 0.4358974358974359, 0.8785189861868696)
( 0.4871794871794872, 0.7374994074327459)
( 0.5384615384615383, 0.5530055449861763)
( 0.5897435897435896, 0.3359129852336457)
( 0.6410256410256410, 0.0990189500238535)
( 0.6923076923076923, -0.1437120768918033)
( 0.7435897435897436, -0.3779715314402721)
( 0.7948717948717947, -0.5899502337333539)
( 0.8461538461538460, -0.7671524147611121)
( 0.8974358974358974, -0.8991323198405171)
( 0.9487179487179487, -0.9781099673508421)
( 1.0000000000000000, -0.9994297648690663)
};
\end{axis}
\end{tikzpicture}
\label{2202}
\end{subfigure}
\caption{The first $4$ basis functions $\cos(2\pi \xi_i x)$ and $\sin(2\pi \xi_i x)$ for the Mat\'ern kernel with $\nu \in [1.5, 3.5]$, $\rho \in [0.1, 0.5]$, and nodes and weights of Table \ref{7813}. }
\label{1025}
\end{figure}
\section{Generalizations and Conclusions}\label{s125}
In this paper we introduce algorithms for representing and computing with Gaussian processes in $1$-dimension. In Algorithm \ref{a10}, we describe a numerical scheme for representing families of Gaussian processes as Fourier expansions of the form
\begin{align}\label{137}
f(x) \sim \gamma_1 \alpha_1 \cos(\xi_1 x) + \beta_1 \gamma_1 \sin(\xi_1 x)
+ ... + \alpha_m \gamma_m \cos(\xi_n x) + \beta_m \gamma_m \sin(\xi_m x)
\end{align}
where for $i=1,...,m$ the coefficients $\xi_i \in \mathbb{R}$ are fixed, and $\alpha_i$ and $\beta_i$ are the iid Gaussians
\begin{align}
\alpha_i, \beta_i \sim \mathcal{N} \big(0, 1 \big),
\end{align}
and where $\gamma_i$, are defined by
\begin{align}
\gamma_i = \sqrt{2 w_i \hat{k}(\xi_i)}
\end{align}
where $\hat{k}$ denotes the Fourier transform of the covariance kernel $k$ and $w_i \in \mathbb{R}$.
where $\alpha_i$ and $\beta_i$ are iid standard normal random variables. These expansions are constructed in such a way that they are valid over families of covariance kernels. Representing a GP as expansion (\ref{137}), allows the use Algorithm \ref{a20} to perform GP regression in $O(N\log{N} + m^3)$ time where $N$ is the number of data points and $m$ the size of the expansion.
While this paper is focused on GPs in $1$-dimensions, much of the theory and numerical machinery extends naturally to higher dimensions. In particular, for GPs over $\mathbb{R}^2$, the generalization of the $1$-dimensional Fourier expansion is a Fourier-Bessel expansion of the form
\begin{align}
f(x) \sim \sum_{n=-N_0}^{N_0} \sum_{j=1}^{p} \alpha_{j} J_n(\xi_j |x|) \cos(n\theta_x)
+ \beta_{j} J_n(\xi_j |x|) \sin(n\theta_x)
\end{align}
where $J_n$ is the Bessel function of the first kind, $\alpha_{j}, \beta_{j}$ are iid Gaussians for all $j$, and $(|x|, \theta_x)$ is the representation in polar coordinates of $x \in \mathbb{R}^2$. As in the $1$-dimensional case, the linear system of equations in the $2$-dimensional GP regression contains structure that's amenable to fast algorithms. For example, the algorithm of~\cite{oneil2} allows for the matrix of the GP regression linear system to be applied in $O(N\log{N})$ operations.
We plan to address the higher dimensional extension of the tools of this paper in subsequent publications.
\newpage
\bibliographystyle{abbrv}
|
\section{Introduction \label{sec:Introduction-1}}
The pseudo-marginal Metropolis-Hastings (PMMH) \citep{Andrieu:2009} approach is a standard method for conducting Bayesian inference in a wide range of challenging statistical models having an intractable likelihood which
can be estimated unbiasedly. Our article develops a PMMH sampling method for estimating parameters of the complex and high-dimensional state space models that is efficient
and scalable (relative to current methods) in the number of observations and the latent states and can handle state transition densities that cannot be expressed in closed form; for example,
many dynamic stochastic general equilibrium (DSGE) models, which are popular class of macroeconomic time series state space models, do not have closed form transition densities.
A key issue in efficiently estimating statistical models using a PMMH approach is that the variance of the log of
the estimated likelihood grows with the number of observations and the dimension of the latent states \citep{Deligiannidis2018}. \citet{Pitt:2012} show that to obtain a
balance between computational time and the mixing of the Markov chain Monte Carlo
(MCMC) chain, the
number of particles used in the particle filter should be such that the variance
of the log of the estimated likelihood is in the range 1 to 3, depending on the efficiency of the proposal for $\theta$, and that the efficiency
of PMMH schemes deteriorates exponentially as that variance increases.
In many complex statistical applications, it is computationally very expensive ensuring that the variance of the log of the estimated likelihood is within the required range.
\citet{Deligiannidis2018} propose a more
efficient PMMH scheme, called the correlated pseudo-marginal (CPM) method,
which correlates the random numbers used in constructing the estimated
likelihoods at the current and proposed values of the parameters. This
correlation induces a correlation between
the estimated likelihoods and reduces the variance of the difference
in the logs of the estimated likelihoods which appears in the Metropolis-Hastings
(MH) acceptance ratio. They show that the CPM scales up with
the number of observations compared to the standard pseudo marginal
method of \citet{Andrieu:2010} if the state dimension
is moderate.
\citet{Tran2016} propose an alternative approach,
called the block pseudo marginal (BPM) method, which divides the random
numbers into blocks and then updates the parameters jointly with one randomly chosen
block of the random numbers in each MCMC iteration; this
induces a positive correlation between the numerator and denominator
of the MH acceptance ratio, similarly to the CPM. They showed that for large samples the
correlation of the logs of the estimated likelihoods at the current
and proposed values is close to $1-1/G$, where $G$ is the number
of blocks in the blocking approach.
Our paper proposes a new PMMH sampler, which we call the mixed PMMH algorithm (MPM);
it builds on, and extends, the
CPM method of \citet{Deligiannidis2018} and the BPM method
of \citet{Tran2016}. The important innovations of the MPM sampler
compared to block and correlated PMMH are that: (a) the unbiased
likelihood estimator is an average of the likelihood estimators obtained
by multiple independent particle filters (PFs) which lowers its variance; these PFs can be run in parallel. (b)
The unknown parameters, and only the random numbers used in one of
the PFs, are updated jointly. Section \ref{subsec:Multiple-Particle-Filter} provides
the implementation details. Sections \ref{subsec:Linear-Gaussian-State Space Model} and \ref{subsec:Non-linear-RBC} show that
the MPM sampler is able to maintain the correlation between the logs
of the estimated likelihoods for relatively high dimensional state space models
compared to the BPM and CPM. (c) In many applications of state space models, such as DSGE models, it is difficult to estimate
the likelihood efficiently because the bootstrap filter is very inefficient while it is difficult to construct an auxiliary particle filter because the
state transition density does not have a closed form. We follow \citet{Murray2013b} and \citet{Hall2014} and use the disturbance particle filter (DPF), which works on the disturbances rather than the states. An improved disturbance PF (IDPF) algorithm is proposed to improve the performance of the standard disturbance particle filter. Section \ref{subsec:Disturbance-Particle-Filter} discusses this further. (d) The delayed-acceptance version \citep{Christen2005} of the
MPM algorithm is also developed to speed up the computation for
non-linear state space models. The motivation for the delayed acceptance
algorithm is to minimize the expensive computation of the likelihood or its estimate when it appears likely that the proposal will be rejected. See
Section \ref{SS: delayed accept} for details.
We illustrate the MPM sampler empirically, and compare its performance to the BPM and CPM methods,
using a standard linear Gaussian
state space model and two Dynamic Stochastic General Equilibrium (DSGE) models,
using simulated and real data.
We also compare the performance of the improved disturbance particle filter (IDPF) to the tempered particle
filter (TPF) proposed by \citet{Herbst2019}; however, we note that \citet{Herbst2019} only apply the TPF to linear (first order) DSGE models. Sections \ref{subsec:Linear-Gaussian-State Space Model} to \ref{subsec:SecondOrderSmallScale} show that: (1) the IDPF is much more efficient than the standard bootstrap particle filter (BPF) and the tempered particle filter (TPF); (2) the MPM, which runs multiple independent particle filters and only updates the random numbers used in one particle filter, maintains the correlation between the logs of the estimated likelihoods much better than both CPM and BPM. The BPM method of \citet{Tran2016}, which runs a single particle filter and updates a block of random numbers for all time periods in the particle filter algorithm, is unsuitable for time-series state space model; (3) the delayed acceptance version of the MPM sampler is much more efficient than the standard MPM sampler.
Our work in estimating non-linear DSGE models is also related to \citet{FVVRR2007} and \citet{Hall2014} who use standard PMMH methods. We note that the MPM sampler will also be useful for other complex statistical models, such as partially observed diffusions
models, panel data models, and stochastic differential equation mixed
effects models.
The rest of the article is organised as follows. Section \ref{sec:statespacemodel}
introduces the state space model and gives some examples. Section
\ref{sec:Bayesian-Inference} discusses the MPM sampler. Section
\ref{sec:Examples} presents results from both simulated and real
data. Section \ref{sec:Conclusions} concludes with a discussion of our major results and findings.
The paper has an online supplement
containing some further technical and empirical results.
\section{State Space Models \label{sec:statespacemodel}}
\subsection{General State Space Models\label{SS:general SS models}}
We use the colon notation for collections of variables, i.e., $a_{t}^{r:s}:=\left(a_{t}^{r},...,a_{t}^{s}\right)$
and for $t\leq u$, $a_{t:u}^{r:s}:=\left(a_{t}^{r:s},...,a_{u}^{r:s}\right)$.
Consider the stochastic process $\left\{ \left(Z_{t},Y_{t}\right)\right\} $,
with parameter $\theta$, where the $Y_{t}$ are the observations and
the $Z_{t}$ are the latent state vectors; random variables are denoted by capital letters and their realizations by lower case letters. We consider the state space model
with $p\left(z_{0}|\theta\right)$ the density of $Z_{0},$
$p\left(z_{t}|z_{t-1},\theta\right)$ the density
of $Z_{t}=z_t$ given $Z_{0:t-1}=z_{0:t-1}$
for $t\geq1$, and $p\left(y_{t}|z_{t},\theta\right)$
is the density of $Y_{t}=y_t$ given $Z_{0:t}=z_{0:t}$,
$Y_{1:t-1}=y_{1:t-1}$.
The stochastic volatility model is an example of a popular non-linear and non-Gaussian state space model.
It is described by
\begin{align*}
y_{t} & = \exp\left(z_{t}/2\right)\eta_{t}\;\; \left(t\geq1\right), \\
z_{t+1} & = \mu+\phi\left(z_{t}-\mu\right)+\tau\epsilon_{t+1}\;\left(t\geq0\right),
\;\; \text{with} \;\;
z_{0} \sim N\left(\mu,\frac{\tau^{2}}{1-\phi^{2}}\right);
\end{align*}
$\{z_t\}$ is the latent volatility process, $\theta:=\left(\mu,\tau,\phi\right)$ is the vector of unknown parameters, with the persistence parameter $\phi$
satisfying $|\phi|<1$ to ensure stationarity; the sequence
$\left(\eta_{t}, \epsilon_{t}\right)^\top\sim N\left(0,I\right)$
is independent for all $t$.
\begin{comment}
Hence, the observation and state transition
densities are,
\begin{align}
p\left(y_{t}|z_{t},\theta\right) & = N\left(0,\exp\left(z_{t}\right)\right)\left(t\geq1\right), \\
p\left(z_{0}|\theta\right) & \sim N\left(\mu,\frac{\tau^{2}}{1-\phi^{2}}\right) \; \text{and}\;
p\left(z_{t+1}|z_{t},\theta\right) \sim N\left(\mu+\phi\left(z_{t}-\mu\right),\tau^{2}\right),\left(t\geq0\right).
\label{SVtransitiondensity}
\end{align}
\end{comment}
\subsection{State space representations of DSGE models \label{subsec:Dynamic-Stochastic-GeneralModel}}
Our article illustrates the proposed methods using Dynamic Stochastic General Equilibrium (DSGE) models; however, they
apply more generally to nonlinear state space models.
We start by describing the state space representation of DSGE models and then highlight the specific source of nonlinearity we tackle in our applications.
In the article we deal with DSGE models having a state transition equation of the form
\begin{align}
z_{t+1} & =F(z_{t},\epsilon_{t+1} ,\zeta; \theta ),\label{eq: DSGE state transition}
\end{align}
such that $\{\epsilon_{t} \}$ is an independent $N(0,\Sigma_\epsilon)$ sequence;
$\zeta$ is a perturbation parameter and $\theta$ is a vector of the unknown parameters of the DSGE model.
For most applications, the function $F(\cdot)$ in Eq. \eqref{eq: DSGE state transition} is analytically intractable and
is approximated using local solution techniques. We use first and second order Taylor series approximations around the \textit{deterministic} steady state $z^s$ with $\zeta=0$ to approximate $F$; $z_s$ satisfies $z^s=F(z^s,0,0; \theta)$.
A first order-accurate approximation, around the \textit{deterministic} steady state, to Eq. \eqref{eq: DSGE state transition} is
\begin{equation}\label{equation:dsge3paper1}
z_{t+1}^d = F_1 (\theta)z_t^d
+ F_2(\theta) \epsilon_{t+1}
\end{equation}
where
$z_t^d=( z_{t}- z^s)$. For Eq. \eqref{equation:dsge3paper1} to be stable, it is necessary that all the eigenvalues of $F_1(\theta)$ are less than 1 in absolute value.
The initial value $z_0^d=0$ because we work with approximations around the \textit{deterministic} steady state. For a given set of parameters $\theta$, the matrices $F_1(\theta)$ and $F_2(\theta)$ can be solved using existing software. Our applications use Dynare.
The second order accurate approximation (around the \textit{deterministic} steady state) to Eq. \eqref{eq: DSGE state transition} is
\begin{equation}\label{equation:dsge2paper1}
z_{t+1}^d=F_{0}(\theta) \zeta^{2}+F_{1}(\theta) z_{t}^d+F_2(\theta) \epsilon_{t+1}+F_{11}(\theta) P(z_t^d)+F_{12}(\theta)(z_t^d\otimes\epsilon_{t+1})+F_{22}(\theta) P(\epsilon_{t+1});
\end{equation}
where $z_{t}^d=( z_{t}- z^s)$ and for any vector $x:=(x_1, \dots, x_m)^\top$, we define $P(x) := \text{\rm vech}(xx^\top),$
where $\text{\rm vech}(xx^\top)$ is the strict upper triangle and diagonal of $xx^\top$,
such that \newline $\text{\rm vech}(xx^\top):=(x_1^2, x_1x_2 , x_2^2, x_1 x_3, ... , x_m^2)^\top$. The term $F_{0}(\theta) \zeta^{2}$ captures the level correction due to uncertainty that arises from taking a second-order approximation. For the scope of our analysis we normalize the perturbation parameter $\zeta$ to 1. As before, for a given set of parameters $\theta$ the matrices $F_1(\theta), F_2(\theta),F_{11}(\theta), F_{12}(\theta),F_{22}(\theta)$ can be solved using Dynare.
The measurement (observation) equation for the DSGE model and its approximations is
\begin{align}
y_{t} & = Hz_{t}+\eta_{t},\;\eta_{t}\sim N\left(0,\Sigma_{\eta}\right), \label{eq:measurementeqn}
\end{align}
where $H$ is a known matrix;
$\{\eta_{t}\}$ is an independent $N\left(0,\Sigma_{\eta}\right)$ and sequence and it is also independent of the $\{\epsilon_t \}$ sequence. The matrix $\Sigma_{\eta}$ is usually unknown and estimated from the data.
Although it is possible to apply the bootstrap filter for the second order solution, the resulting estimated likelihood is too variable to be useful in a PMMH scheme. Instead, we use an auxiliary disturbance particle filter because the transition density of the second order DSGE models is usually intractable.
We do not use Eq. \eqref{equation:dsge2paper1} as the state space transition equation because it can generate explosive solutions for the variables in the model. Instead, this problem is solved
by modifying the state transition equation using the
pruning approach of \citet{Kim2008}, as in Section~\ref{DescriptionOfPrunedDSGEModel} of the supplement.
Such a pruned representation can substantially increase the size of the state vector; if $z_t$ in
Eq. \eqref{equation:dsge2paper} is $d$ dimensional, then the pruned state vector is $2d + d(d+1)/2$ dimensional. The pruned state vector is then also likely to be substantially greater than the dimension of the disturbances $\{\epsilon_t\}, $ thus giving a further motivation for using the disturbance filter.
\section{Bayesian Inference\label{sec:Bayesian-Inference}}
\subsection{Preliminaries\label{subsec:Preliminaries-1}}
The objective in Bayesian inference is to obtain the posterior distributions
of the model parameters $\theta$ and the latent states $z_{0:T}$, given
the observations $y_{1:T}$ and a prior distribution $p\left(\theta\right)$;
i.e.,
\begin{eqnarray}
p\left(\theta,z_{0:T}|y_{1:T}\right) & = & p\left(y_{1:T}|\theta,z_{0:T}\right)p\left(z_{0:T}|\theta\right)p\left(\theta\right)/p\left(y_{1:T}\right),\label{eq:posteriorwithstates}
\end{eqnarray}
where
\begin{equation}
p\left(y_{1:T}\right)=\int\int p\left(y_{1:T}|\theta,z_{0:T}\right)p\left(z_{0:T}|\theta\right)p\left(\theta\right)dz_{0:T}d\theta
\end{equation}
is the marginal likelihood. The likelihood
\[
p\left(y_{1:T}|\theta\right)=\int p\left(y_{1:T}|\theta,z_{0:T}\right)p\left(z_{0:T}|\theta\right)p\left(\theta\right)dz_{0:T}
\]
can be calculated exactly using the Kalman filter for the linear Gaussian
state space model (LGSS) and hence for the linear (first order) approximation. The posterior samples can then
be obtained using a MCMC sampling
scheme. However, the likelihood cannot be computed exactly if
a higher-order approximation is used, such as Eq. \eqref{equation:dsge4},
for the evolution of the state and other non-linear and non-Gaussian
state space models. In this case, the likelihood can only be estimated.
The bootstrap particle filter of \citet{Gordon:1993} provides an unbiased
estimator of the likelihood for a general state space model. \citet{Andrieu:2009}
and \citet{Andrieu:2010} show that it is possible to use this unbiased
estimator of the likelihood to carry out full Bayesian inference for
the parameters of the general state space model. They call this MCMC approach pseudo
marginal Metropolis-Hastings (PMMH).
The non-linear (second-order) DSGE models we consider lie within the class
of general non-linear state space models whose state transition density is difficult to work with or cannot be expressed in closed form. In such cases, it is useful to express
the model in terms of its latent disturbance variables as they can be expressed in closed form. The posterior in Eq. \eqref{eq:posteriorwithstates}
becomes
\begin{equation}
p\left(\theta,\epsilon_{1:T}|y_{1:T}\right)\propto
\prod_{t=1}^{T}p\left(y_{t}|\epsilon_{1:t},\theta\right)p\left(\epsilon_{t}\right)p\left(\theta\right),
\end{equation}
which \citet{Murray2013b} call the disturbance state-space model.
The standard state space model can be recovered from the disturbance state space model by using
the deterministic function $F\left(z_{t-1},\epsilon_{t};\theta\right)\rightarrow z_{t}$.
This gives us a state trajectory $z_{0:T}$ from any sample $\left(\epsilon_{1:T},z_{0}\right)$, where $z_0=0$.
In the disturbance state-space model the target becomes the posterior
distribution over the parameters $\theta$ and the latent noise variables
$\epsilon_{1:T}$, rather than $\left(\theta,z_{0:T}\right)$.
\subsection{Standard Pseudo Marginal Metropolis-Hastings \label{subsec:Standard-Pseudo-Marginal}}
This section outlines the standard PMMH scheme
which carries out MCMC
on an expanded space using an unbiased estimate of the likelihood.
Our paper focuses on estimating the posterior density of the parameters $\theta$,
but not of the states.
Let $u$ consist of all the random variables required to compute the unbiased likelihood estimate
$\widehat{p}_{N}\left(y|\theta,u\right)$,
with $p(u)$ the density of $u$; let $p\left(\theta\right)$ be the prior of $\theta$.
The joint posterior density $\theta$ and $u$ is
\begin{equation*}
p\left(\theta,u|y_{1:T}\right)=\widehat{p}_{N}\left(y_{1:T}|\theta,u\right)p\left(\theta\right)p\left(u\right)/p\left(y_{1:T}\right),
\end{equation*}
so that
\[
p\left(\theta|y_{1:T}\right)=\int p\left(\theta,u|y_{1:T}\right)du=p\left(y_{1:T}|\theta\right)p\left(\theta\right)/p\left(y_{1:T}\right)
\]
is the posterior of $\theta$ and $\intop\widehat{p}_{N}\left(y_{1:T}|\theta,u\right)p\left(u\right)du=p\left(y_{1:T}|\theta\right)$
because the likelihood estimate is unbiased. We can therefore sample from the posterior
density $p\left(\theta|y_{1:T}\right)$ by sampling $\theta$ and
$u$ from $p\left(\theta,u|y_{1:T}\right)$. The subscript $N$ indicates the number of particles used to estimate likelihood.
Let $q\left(\theta^{'}|\theta\right)$ be the proposal density for $\theta^{\prime}$
with the current value $\theta$ and $q\left(u^{'}|u\right)$ the proposal
density for $u^{'}$ given $u$. We always assume that the $q\left(u^{'}|u\right)$
satisfies the reversibility condition
\begin{equation}
q\left(u^{'}|u\right)p\left(u\right)=q\left(u|u^{'}\right)p\left(u^{'}\right);
\end{equation}
it is clearly satisfied by the standard PMMH where $q\left(u^{'}|u\right)=p\left(u^{'}\right)$.
We generate a proposal $\theta^{'}$ from $q\left(\theta^{'}|\theta\right)$
and $u^{'}$ from $q\left(u^{'}|u\right)$, and accept both proposals
with probability
\begin{eqnarray}
\alpha\left(\theta,u;\theta^{'},u^{'}\right) & = & \min\left(1,\frac{\widehat{p}_{N}\left(y|\theta^{'},u^{'}\right)p\left(\theta^{'}\right)p\left(u^{'}\right)q\left(\theta|\theta^{'}\right)q\left(u|u^{'}\right)}{\widehat{p}_{N}\left(y|\theta,u\right)p\left(\theta\right)p\left(u\right)q\left(\theta^{'}|\theta\right)q\left(u^{'}|u\right)}\right)\nonumber \\
& = & \min\left(1,\frac{\widehat{p}_{N}\left(y|\theta^{'},u^{'}\right)p\left(\theta^{'}\right)q\left(\theta|\theta^{'}\right)}{\widehat{p}_{N}\left(y|\theta,u\right)p\left(\theta\right)q\left(\theta^{'}|\theta\right)}\right).\label{eq:acceptancePMMH}
\end{eqnarray}
The expression in Eq. \eqref{eq:acceptancePMMH} is identical
to a standard Metropolis-Hastings algorithm except that estimates
of the likelihood at the current and proposed parameters are used.
\citet{Andrieu:2009} show that the resulting
PMMH algorithm has the correct invariant distribution regardless of
the variance of the estimated likelihoods. However, the performance
of the PMMH approach crucially depends on the number of particles $N$ used to
estimate the likelihood. The variance
of the log of the estimated likelihood should be between 1 and 3 depending on the quality
of the proposal
for $\theta$; see
\citet{Sherlock2015} and \citet{Doucet2015} for further details. In many
applications of the non-linear state space models considered in this
paper, it is computationally very expensive to ensure that the variance
of the log of the estimated likelihood is within the required range.
Section \ref{subsec:Multiple-Particle-Filter} discusses the new PMMH sampler, which we call the mixed PMMH algorithm (MPM), that builds on and extends, the
CPM of \citet{Deligiannidis2018} and BPM
of \citet{Tran2016}.
\subsection{Mixed PMMH (MPM) \label{subsec:Multiple-Particle-Filter}}
This section discusses the proposed mixed
PMMH (MPM) method that uses multiple particle filters to obtain
the unbiased estimate of the likelihood, similarly
to the approach of \citet{Sherlock2017},
and uses a combination of the BPM and CPM methods to induce a
high correlation between successive logs of the estimated likelihoods.
Suppose that $G$ particle filters are run in parallel. Let $\widehat{p}_{N}\left(y|\theta,u_{g}\right)$
be the unbiased estimate of the likelihood obtained from the $g$th particle filter, for
$g=1,...,G$. We now define the joint target density of $\theta$
and $\widetilde{u}=\left(u_{1},...,u_{G}\right)$ as
\begin{equation}
p\left(\theta,\widetilde{u}|y_{1:T}\right)\propto
\overline{\widehat{p}}_{N}\left(y|\theta,\widetilde{u}\right)
p\left(\theta\right)\prod_{g=1}^{G}p\left(u_{g}\right),
\end{equation}
where
$\overline{\widehat{p}}_{N}\left(y|\theta,u\right)=
\sum_{g=1}^{G}\widehat{p}_{N}\left(y|\theta,u_{g}\right)/G$
is the average of the $G$ unbiased likelihood estimates and hence it is
also unbiased \citep{Sherlock2017}. We update the parameters $\theta$
jointly with a randomly selected block $u_{g}$ in each MCMC iteration,
with $\Pr\left(G=g\right)=1/G$ for any $g=1,...,G$. The selected block $u_{g}$ is updated using
$u_{g}^{'}=\rho_{u}u_{g}+\sqrt{1-\rho_{u}^{2}}\eta_{u}$, where $\rho_{u}$
is the non-negative correlation between the random numbers $u_{g}$
and $u_{g}^{'}$ and $\eta_{u}$ is a standard normal vector of the
same length as $u_{g}$. Using this scheme, the acceptance probability
is
\begin{equation}
\alpha\left(\theta,\widetilde{u};\theta^{'},\widetilde{u}^{'}\right)=\min\left(1,\frac{\overline{\widehat{p}}_{N}\left(y|\theta^{'},\widetilde{u}^{'}=\left(u_{1},...,u_{g-1},u_{g}^{'},u_{g+1},...,u_{G}\right)\right)p\left(\theta^{'}\right)q\left(\theta|\theta^{'}\right)}{\overline{\widehat{p}}_{N}\left(y|\theta,\widetilde{u}=\left(u_{1},...,u_{g-1},u_{g},u_{g+1},...,u_{G}\right)\right)p\left(\theta\right)q\left(\theta^{'}|\theta\right)}\right).\label{MHacceptance}
\end{equation}
It is important that the
logs of the likelihood estimates evaluated at the current and proposed
values of $\theta$ and $\widetilde{u}$ are highly correlated to reduce
the variance of $\log\overline{\widehat{p}}_{N}\left(y|\theta^{'},\widetilde{u}^{'}\right)-\log\overline{\widehat{p}}_{N}\left(y|\theta,\widetilde{u}\right)$ which helps the Markov chain to mix well. However, the
resampling step in the particle filter introduces discontinuities
even when $\theta$ and $\theta^{'}$ are close, where $\theta$ is
the current value and $\theta^{'}$ is the proposed value of the parameters.
Section \ref{subsec:Disturbance-Particle-Filter} discusses this discontinuity issue and the proposed correlated disturbance particle filter algorithm, which helps to maintain the correlation. Algorithm \ref{alg:MPM-PMMH} in Section~\ref{MPM-PMMH algorithm} of the Supplement gives the MPM algorithm.
\subsection{Delayed Acceptance PMMH\label{SS: delayed accept}}
We propose the delayed acceptance version of the MPM algorithm, which we call delayed acceptance mixed PMMH (DA-MPM), to speed up the computation.
The motivation for the
delayed acceptance sampler \citep{Christen2005} is to avoid the computation of the expensive likelihood estimate
if it is likely that the proposed draw will ultimately
be rejected. A first accept-reject stage uses a cheap
(or deterministic) approximation to the likelihood instead of the expensive likelihood estimate
in the MH acceptance ratio. Then, the particle filter is used to estimate the likelihood
only for a proposal that is
accepted in the first stage;
a second accept-reject stage ensures that detailed
balance is satisfied with respect to the true posterior. We use the likelihood obtained from the central difference Kalman filter (CDKF) proposed by \citet{Norgaard2000} in the first accept-reject stage of the delayed acceptance scheme. We chose the CDKF for the first staqe rather than
the extended Kalman Filter (EKF) because it
frequently outperforms the EKF for
general non-linear and non-Gaussian state-space models \citep{Andreasen2013}. Section \ref{subsec:Delayed-Acceptance-Multiple} of the Supplement gives further details.
\subsection{The Disturbance Particle Filter \label{subsec:Disturbance-Particle-Filter} }
This section discusses the disturbance particle filter we use to
obtain the unbiased estimates of the likelihood in the MPM sampler described in
Section~\ref{subsec:Multiple-Particle-Filter}.
Suppose that $z_{1}=\Phi\left(z_{0},\epsilon_{1};\theta\right)$,
where $z_0$ is the initial state vector; for $t \geq 2 ,$ $z_{t}=F\left(z_{t-1},\epsilon_{t};\theta\right)$,
where $\epsilon_{t}$ is a $n_{e}\times1$ vector of normally distributed
latent noise with density $p\left(\epsilon_{t}\right). $
\citet{Murray2013b}
express the standard state-space model in terms of the latent noise
variables $\epsilon_{1:T}$, and call
\[
\epsilon_{t}\sim p\left(\epsilon_{t}\right),y_{t}|\epsilon_{1:t}\sim p\left(y_{t}|\epsilon_{1:t},x_{0};\theta\right)=p\left(y_{t}|z_{t};\theta\right),t=1,...,T,
\]
the disturbance state-space model. We note that the conditional distribution of $y_{t}$
depends on all the latent error variables $\epsilon_{1:t}.$ The disturbance particle filter provides an unbiased estimate of the likelihood.
For the MPM sampler described in Section \ref{subsec:Multiple-Particle-Filter} to work efficiently, the
logs of the likelihood estimates evaluated at the current and proposed
values of $(\theta, \widetilde{u})$ need to be highly correlated. However, the standard
resampling step in the (disturbance) particle filter introduces discontinuities and breaks down the correlation between the logs of the likelihood estimates at the current and proposed values
even when the current parameters $\theta$ and the proposed parameters $\theta^{'}$ are close.
Sorting the particles from smallest to largest helps to preserve the correlation between the logs of the likelihood estimates at the current and proposed values
\citep{Deligiannidis2018}.
However, such sorting is unavailable for multidimensional state particles.
\citet{Deligiannidis2018} use the Hilbert sorting method of \citet{Skilling:2004} to order multidimensional state particles. Our article uses a simpler and faster method proposed by \citet{Choppala2016} and given in Section~\ref{Multidimensional Sorting} of the Supplement. Algorithm \ref{alg:The-Disturbance-Particle filter} in Section~\ref{Disturbance particle filter} of the Supplement outlines the correlated disturbance particle filter algorithm used by the MPM algorithm described in Section \ref{subsec:Multiple-Particle-Filter}.
\subsection*{Improved Disturbance Particle Filter
(IDPF) \label{subsec:Proposals-Distribution}}
Our article uses the mixture proposal density
\begin{equation}
m\left(\epsilon_{t}|u_{\epsilon,t},\theta\right)=\pi p\left(\epsilon_{t}|\theta\right)+(1-\pi) q\left(\epsilon_{t}|\theta,y_{1:T}\right);
\label{eqnprop}
\end{equation}
$0 \leq \pi \leq 1.$
If $\pi = 1$, then $m\left(\epsilon_{t}\right) = p\left(\epsilon_{t}\right)$
is the bootstrap disturbance particle filter. However, the empirical performance of this bootstrap filter is usually poor because the resulting likelihood estimate is too variable. In practice, we take $0< \pi << 1$ and
$q\left(\epsilon_{t}|\theta,y_{1:T}\right)
=N\left(\epsilon_{t}|\widehat{\mu}_{t},\widehat{\Sigma}_{t}\right)$.
The methods to obtain $\widehat{\mu}_{t}$ and the covariance matrix
$\widehat{\Sigma}_{t}$ are now discussed. First, the $G$ disturbance particle filters are run in parallel. Second, the ancestral tracing algorithm given
in Section~\ref{sec:AncestralTracing} of the Supplement is used to obtain $G$ trajectories of $\epsilon_{g,1:T}$. Third, the mean $\widehat{\mu}_{t}$ and the covariance matrix $\widehat{\Sigma}_{t}$ are set as the mean and the covariance matrix of these $G$ trajectories of $\epsilon_{g,1:T}$ at each time $t$. Applying additional ancestral tracing step for each disturbance particle filter gives the particle trajectories from the smoothing distribution $p(\epsilon_{1:T}|\theta,y_{1:T})$, that is the conditional distribution of disturbances given the parameters $\theta$ and all available data up to time $T$. The proposal defined in
Eq. \eqref{eqnprop} is generated without much additional computational cost because the $G$ disturbance particle filters are needed to be run at each MCMC iteration to compute the unbiased estimate of the likelihood and the ancestral tracing method is fast to compute. The proposal density in Eq. \eqref{eqnprop} is in the form of the defensive mixture approach of \citet{Hesterberg1995} which guarantees the boundedness of the weights in the disturbance particle filter algorithm defined in Eq. \eqref{importanceweights} of Section~\ref{Disturbance particle filter} of the Supplement, given that the observation density $p\left(y_{t}|z_{t},\theta\right)$ is bounded which is satisfied for all models defined in Section~\ref{subsec:Dynamic-Stochastic-GeneralModel}.
We set $\pi=0.05$ in all our examples in Section~\ref{sec:Examples}.
\section{Examples\label{sec:Examples}}
Section \ref{SS: preliminaries} discusses the inefficiency measures we use
to compare the performance of different particle filters or PMMH samplers. Section \ref{subsec:Linear-Gaussian-State Space Model} investigates empirically the ability of the proposed mixed PMMH (MPM) sampler to maintain the correlation between successive log-likelihood estimates using a high-dimensional linear Gaussian state space model. Section \ref{subsec:Small-Scale-DSGE} discusses the linear small-scale DSGE example used by \citet{Herbst2019}. Section \ref{subsec:Non-linear-RBC} discusses a non-linear Real Business Cycle (RBC) using simulated datasets. Section \ref{subsec:SecondOrderSmallScale} applies the MPM sampler to estimate a non-linear small scale DSGE model.
\subsection{Definitions of Inefficiency\label{SS: preliminaries}}
We define the time normalised variance (TNV) of a particle filter method
\begin{equation*}
\textrm{TNV}_{PF}:=\widehat{\textrm{V}}
\left(\log\widehat{p}\left(y|\theta\right)\right)\times\textrm{CT},
\end{equation*}
as the measure of inefficiency of the method
that takes computing time into account;
$\textrm{CT}$ is the computing time to obtain a single log of the estimated likelihood in seconds,
and $\widehat{\textrm{V}}\left(\log\widehat{p}\left(y|\theta\right)\right)$
is the estimated variance of the log of the likelihood estimate.
The relative time normalised variance (RTNV) of a particle filter method is defined as $\textrm{RTNV}_{PF}:= \textrm{TNV}_{PF}/ \textrm{TNV}_{IDPF}$.
We use the inefficiency factor (IF)
\[
\textrm{IF}_{\psi}:=1+2\sum_{t=1}^{\infty}\rho_{\psi}\left(t\right),
\]
to measure the inefficiency of a PMMH sampler at estimating the posterior expectation
of a univariate function $\psi\left(\theta\right)$ of $\theta$;
here, $\rho_{\psi}\left(t\right)$ is the $j$th autocorrelation of
the iterates $\psi\left(\theta\right)$ in the MCMC chain after it
has converged to its stationary distribution. We estimate the IACT
using the CODA R package of \citet{Plummer2006}. A low value of the IACT
estimate suggests that the Markov chain mixes well. Our measure of the inefficiency
of a PMMH sampler that takes computing time into account for a given parameter $\theta$ based on $\textrm{IF}_\psi$ is
the time normalised inefficiency factor (TNIF) is defined as
$
\textrm{TNIF}_\psi := \textrm{IF}_\psi\times\textrm{CT}.
$
The estimate of TNIF is the estimate of the IF times the computing time (CT).
For a given sampler, let $\textrm{IACT}_{\textrm{MAX}}$ and $\textrm{IACT}_{\textrm{MEAN}}$
be the maximum and mean of the IACT values over all the
parameters in the model.
\subsection{Performance for the Linear Gaussian State Space Model \label{subsec:Linear-Gaussian-State Space Model}}
This section examines empirically the ability of the following methods to maintain the correlation between successive values of the log of the estimated likelihood for a linear Gaussian state space model: (1) the block PMMH (BPM) of \citet{Tran2016}, (2) the correlated PMMH (CPM) of \citet{Deligiannidis2018},
and (3) the mixed PMMH (MPM).
We consider the model discussed in \citet{Guarniero2017}
and \citet{Deligiannidis2018}, where $\left\{ X_{t};t\geq1\right\} $
and $\left\{ Y_{t};t\geq1\right\} $ are ${\cal R}^{d}$ valued with
\begin{eqnarray*}
Y_{t} & = & X_{t}+W_{t},\\
X_{t+1} & = & A_{\theta}X_{t}+V_{t+1},
\end{eqnarray*}
with $X_{1}\sim{\cal N}\left(0_{d},I_{d}\right)$, $V_{t}\sim N\left(0_{d},I_{d}\right)$,
$W_{t}\sim N\left(0_{d},I_{d}\right)$, and $A_{\theta}^{i,j}=\theta^{|i-j|+1}$;
the true value of $\theta$ is $0.4$. We use the state based bootstrap
filter so that the proposal density is the state transition density of the
particle filter for all methods. The experimental setup is
now discussed. The simulated data is generated from the model above
with $T=100$ and $T=1000$ time periods and $d=5,10,20,50,100$ dimensions.
The correlation parameter in the CPM is set to
$\rho_{u}=\left(0.5,0.75,0.8,0.9,0.99\right)$ and the number of blocks
in the BPM is set to $G=\left(2,4,5,10,100\right)$.
We use the same number of blocks for the
MPM as in the BPM and set the correlation coefficients
$\rho_{u}=0$ and $\rho_{u}=0.99$.
There are important implementation differences between the BPM of \citet{Tran2016}, the CPM of \citet{Deligiannidis2018}
and the proposed MPM method. The BPM updates a block of random numbers for all time periods in
the particle filter algorithm and runs a single particle filter; e.g. suppose $N=100$ particles, one block of the particles is updated for all time periods, with the other $99$ block of particles unchanged for all time periods. The CPM method correlates all the random numbers used in the particle filter using $u^{'}=\rho_{u}u+\sqrt{1-\rho_{u}^{2}}\eta_{u}$, where $\rho_{u}$
is the non-negative correlation between the random numbers $u$ and $u^{'}$ and $\eta_{u}$ is a standard normal vector of the
same length as $u$, and runs a single particle filter. The MPM method updates the random numbers in one particle filter, keeping the random numbers in the other $G-1$ particles filters unchanged. The selected block $u_{g}$ is updated using the correlated approach $u_{g}^{'}=\rho_{u}u_{g}+\sqrt{1-\rho_{u}^{2}}\eta_{u}$. Given those random numbers, the MPM then runs $G$ particle filters in parallel.
We ran the CPM, BPM, and MPM approaches for $100$ iterations
holding the parameter $\theta$ fixed at the true value. At each
iteration we generated the $u^{\left(s\right)}$ and $u^{\left(s\right)'}$
and obtained $\log\widehat{p}_{N}\left(y|\theta,u^{\left(s\right)}\right)$
and $\log\widehat{p}_{N}\left(y|\theta,u^{'\left(s\right)}\right)$, for $s=1,...,100$,
for the BPM and CPM; we then computed their sample correlation.
Similarly, we generated $\widetilde{u}$, where $\widetilde{u}=\left(u_{1},...,u_{G}\right)$
and $\widetilde{u}^{'}$ and obtained
$\log\overline{\widehat{p}}_{N}\left(y|\theta,\widetilde{u}^{\left(s\right)}\right)$
and $\log\overline{\widehat{p}}_{N}\left(y|\theta^{'},\widetilde{u}^{\left(s\right)'}\right)$
for the MPM approach and computed their sample correlation.
Figures~\ref{fig:The-estimated-correlation T100} and \ref{fig:The-estimated-correlation T1000}
report the correlation estimates of $\log\widehat{p}_{N}\left(y|\theta,u\right)$
and $\log\widehat{p}_{N}\left(y|\theta,u^{'}\right)$ for the BPM and CPM approaches and $\log\overline{\widehat{p}}_{N}\left(y|\theta,\widetilde{u}^{\left(s\right)}\right)$
and $\log\overline{\widehat{p}}_{N}\left(y|\theta^{'},\widetilde{u}^{\left(s\right)'}\right)$
for the MPM approach for $T=100$ and $1000$, respectively. The figures
show that: (1) the correlation estimates for the BPM approach
are close to zero for all the reported number of blocks and dimensions,
supporting the observation made by \citet{Tran2016} that the BPM approach is unsuitable for
time series state space models;
(2) when the correlation between random numbers $u$ and $u^{'}$ is set to
0.99 in the CPM approach, the estimated correlations between the log of the
estimated likelihoods
are 0.73, 0.59, 0.59, 0.55, and 0.47 for $T=100$ and 0.66, 0.51,
0.52, 0.51, and 0.53 for $T=1000$ for $d=5,10,20,50,100$, respectively.
The estimated correlations seem to be slightly worse when $T=1000$.
The higher the dimension of the latent state variables, the lower is
the estimated correlation between the log of estimated likelihoods.
Section~\ref{subsec:Non-linear-RBC} shows that the estimated correlation between the log of the estimated likelihoods gets worse for the non-linear model;
(3) the MPM method
maintains a high correlation between the logs of the estimated likelihoods
even for $d=100$ dimensions and $T=1000$ time periods. The estimated
correlations between logs of the estimated likelihoods when the number
of blocks $G$ is 100 and $d=100$ dimensions are 0.99 for both $T=100$
and $1000$. This
supports the result in \citet{Tran2016} that shows that the
correlation of the logs of the estimated likelihood at the current
and proposed values is close to $1-1/G=1-1/100=0.99$.
In summary, this example suggests that: (1) the
BPM of \citet{Tran2016} is ineffective for time series
state space models because it is unable to maintain the correlation between
the logs of the estimated likelihoods in successive iterates; (2) the correlation estimates
between the logs of the estimated likelihoods obtained from the CPM deteriorate as the dimension of the state variables and the length of time series increase; (3) the MPM
approach maintains a high correlation even when the dimension
of the state variables is 100. There is no noticeable difference between setting $\rho_u=0$ or $0.99$ for the MPM approach in this example. It is therefore sufficient to set the correlation $\rho_u=0$.
\begin{figure}
\caption{The estimated correlation of log of the estimated likelihoods
obtained using the four approaches: (1) Block PMMH (BPM), (2) Correlated PMMH (CPM), (3) Mixed PMMH (MPM) with $\rho_u=0$, and (4) Mixed PMMH (MPM) with $\rho_u=0.99$ for $T=100$ and $d=5,10,20,50,100$ dimensions \label{fig:The-estimated-correlation T100}}
\centering{}\includegraphics[width=15cm,height=6cm]{corr_sim_LGSS_T100}
\end{figure}
\begin{figure}
\caption{The estimated correlation of log of the estimated likelihoods
obtained using the four approaches: (1) Block PMMH (BPM), (2) Correlated PMMH (CPM), (3) Mixed PMMH (MPM) with $\rho_u=0$, and (4) Mixed PMMH (MPM) with $\rho_u=0.99$ for $T=1000$ and $d=5,10,20,50,100$ dimensions \label{fig:The-estimated-correlation T1000}}
\centering{}\includegraphics[width=15cm,height=6cm]{corr_sim_LGSS_T1000}
\end{figure}
\subsection{Linear Small Scale DSGE Model\label{subsec:Small-Scale-DSGE}}
This section investigates empirically the performance of the proposed improved disturbance particle filter (IDPF), the bootstrap particle filter
(BPF), and the tempered particle filter (TPF) of \citet{Herbst2019} based on the variance of the log of the estimated likelihood for each estimator. Thus, an estimator is better than a second estimator if the variance of the log of the first estimator is smaller than that of the second. We compare the variances using the linear small scale DSGE model considered by \citet{Herbst2019}, and described in Section~\ref{Description of SmallScale Model} of the Supplement. A linearised DSGE model (with normally distributed innovations)
leads to a linear Gaussian state-space representation, making it feasible to compute the likelihood using the Kalman filter instead of estimating the likelihood by the particle filter. The likelihood estimated using Kalman filter is treated as the ground truth for comparing the accuracy of different particle filter methods.
The small scale DSGE model has three observables (output growth, inflation,
and the federal funds rate). \citet{Herbst2019} augment the observation
equations by independent measurement errors $\eta_t$, whose standard deviations
are fixed at 20\% of the standard deviations of the observables $y_t$.
We augment similarly, but set the standard deviations
at 10\% of the standard deviation of the observables,
making the estimation more challenging as the signal to noise is higher.
The real dataset used in this section is obtained from \citet{Herbst2019}, using data from 1983Q1
to 2002Q4, a total of 80 observations for each series\footnote{The standard
deviations of the measurement errors are 0.057 for output growth, 0.147
for inflation, and 0.223 for the interest rates. The DSGE model in
this section is solved using the GENSYS software.}.
The variance of the log of the likelihood estimate is obtained by
computing the sample variance of the log of the estimated
likelihood for each particle filter method using $N_{run}=100$ independent
runs holding the parameters fixed at the values given in Table~\ref{tab:Small-Scale-DSGE parameter values} in Section~\ref{additionalTablesFigures} of the Supplement.
The fixed parameter values are those used
by \citet{Herbst2019}. The proposal for the IDPF is estimated from the collection of particles generated by the $G$ particle filters running in parallel. These particle filters are run once and the same proposal is kept for different numbers of particles $N$. The tuning parameters of the TPF are set to the default values given in \citet{Herbst2019}. All the particle filters are implemented in Matlab. The TPF Matlab code is obtained from the website\footnote{https://web.sas.upenn.edu/schorf/publications/}
of Frank Schorfheide.
Table~\ref{tab:The-variance-of log-likelihood-firstorder} reports the
variance of the log of the estimated likelihoods using the BPF, TPF,
and IDPF for the linear small scale DSGE model with $T=80$ time
periods and $N=10$, $20$, $50$, $100$, $250$, $500$, $1000$, and
$2000$ particles evaluated at the parameter values given in
Table~\ref{tab:Small-Scale-DSGE parameter values} of Section~\ref{additionalTablesFigures}
of the Supplement.
Section~\ref{subsec:Proposals-Distribution}
shows that there is no additional computational cost in constructing the
IDPF proposal because it is obtained from the collection of particles
generated by the $G$ particle filters running in parallel at each MCMC
iteration, but the IDPF computing time in Table~\ref{tab:The-variance-of log-likelihood-firstorder}
is set to twice of the computing time to obtain a single
estimate of the likelihood to account for the time spent constructing the proposal.
In this example, the exact log-likelihood obtained from the Kalman
filter is $-296.45$. In general, the IDPF is much more efficient
than BPF and TPF for all $N$. For example, the IDPF is 350
times more efficient than BPF and 2855 times more efficient than
TPF for $N=500$. The log of the likelihood estimate obtained from the IDPF
is very close to the exact value even with
$N=20$. The BPF fails to provide reasonable values of the of the likelihood estimate
even with $N=2000$ and the TPF requires $N=500$ particles to obtain the value of the
log of the likelihood estimate that is close to the exact value. In addition, in Table 1, we note that for all three estimators $\log\overline{\widehat{p}}_{N}\left(y|\theta,u\right)$ increases as $N$ increases and seems to converge to the log of the likelihood obtained by the Kalman filter. This can be explained heuristically as follows. Suppose that $E(X) =a, \sigma_X^2 = var(X)$ and $|X-a|/|a|$ is small. Then $$E(\log X) \approx \log a- \frac{\sigma_X^2}{a^2} \leq \log a$$ and
$\log X $ monotonically increases to $\log a $ as $\sigma_X^2$ tends to 0.
\begin{sidewaystable}
\caption{The variance of the log of the estimated likelihoods obtained using various
particle filter methods: (1) the bootstrap particle filter (BPF), (2) the tempered particle filter (TPF), and (3) the improved disturbance particle filter (IDPF) for the small scale linear DSGE model
with $T=80$ periods and $N$ is the of particles. The
computation time (CT) in seconds to obtain a single log of the estimated likelihood
is in seconds. The $\log\overline{\widehat{p}}_{N}\left(y|\theta,u\right)$
is the average of the log of the estimated likelihoods over 100 replications. The TNV and RTNV are the time normalised variance and the relative time normalised variance of a particle filter method relative to the IDPF method, respectively, and are defined in Section \ref{SS: preliminaries}.
\label{tab:The-variance-of log-likelihood-firstorder}}
\centering{}{\scriptsize{}}%
\begin{tabular}{cccccccccccccccc}
\hline
& \multicolumn{5}{c}{{\scriptsize{}BPF}} & \multicolumn{5}{c}{{\scriptsize{}TPF}} & \multicolumn{5}{c}{{\scriptsize{}IDPF}}\tabularnewline
{\scriptsize{}N} & {\scriptsize{}Var} & {\scriptsize{}CT} & {\scriptsize{}TNV} & {\scriptsize{}RTNV} & {\scriptsize{}$\log\overline{\widehat{p}}_{N}\left(y|\theta,u\right)$} & {\scriptsize{}Var} & {\scriptsize{}CT} & {\scriptsize{}TNV} & {\scriptsize{}RTNV} & {\scriptsize{}$\log\overline{\widehat{p}}_{N}\left(y|\theta,u\right)$} & {\scriptsize{}Var} & {\scriptsize{}CT} & {\scriptsize{}TNV} & {\scriptsize{}RTNV} & {\scriptsize{}$\log\overline{\widehat{p}}_{N}\left(y|\theta,u\right)$}\tabularnewline
\hline
\hline
{\scriptsize{}10} & {\scriptsize{}$373665.65$} & {\scriptsize{}$0.01$} & {\scriptsize{}$3736.60$} & {\scriptsize{}$54.04$} & {\scriptsize{}$-2855.24$} & {\scriptsize{}$14335.67$} & {\scriptsize{}$9.60$} & {\scriptsize{}$137622.43$} & {\scriptsize{}$1990.20$} & {\scriptsize{}$-594.7452$} & {\scriptsize{}$1728.84$} & {\scriptsize{}$0.04$} & {\scriptsize{}$69.15$} & {\scriptsize{}$1$} & {\scriptsize{}$-301.58$}\tabularnewline
{\scriptsize{}20} & {\scriptsize{}$131366.07$} & {\scriptsize{}$0.01$} & {\scriptsize{}$1313.66$} & {\scriptsize{}$57.24$} & {\scriptsize{}$-2016.73$} & {\scriptsize{}$1346.30$} & {\scriptsize{}$13.31$} & {\scriptsize{}$17919.25$} & {\scriptsize{}$780.80$} & {\scriptsize{}$-388.1918$} & {\scriptsize{}$573.66$} & {\scriptsize{}$0.04$} & {\scriptsize{}$22.95$} & {\scriptsize{}$1$} & {\scriptsize{}$-298.93$}\tabularnewline
{\scriptsize{}50} & {\scriptsize{}$30537.59$} & {\scriptsize{}$0.01$} & {\scriptsize{}$305.37$} & {\scriptsize{}$136.94$} & {\scriptsize{}$-1113.85$} & {\scriptsize{}$213.04$} & {\scriptsize{}$18.74$} & {\scriptsize{}$3992.37$} & {\scriptsize{}$1790.30$} & {\scriptsize{}$-318.1254$} & {\scriptsize{}$55.77$} & {\scriptsize{}$0.04$} & {\scriptsize{}$2.23$} & {\scriptsize{}$1$} & {\scriptsize{}$-297.82$}\tabularnewline
{\scriptsize{}100} & {\scriptsize{}$20879.66$} & {\scriptsize{}$0.02$} & {\scriptsize{}$417.59$} & {\scriptsize{}$707.78$} & {\scriptsize{}$-729.90$} & {\scriptsize{}$97.64$} & {\scriptsize{}$21.51$} & {\scriptsize{}$2100.24$} & {\scriptsize{}$3559.73$} & {\scriptsize{}$-302.4716$} & {\scriptsize{}$14.68$} & {\scriptsize{}$0.04$} & {\scriptsize{}$0.59$} & {\scriptsize{}$1$} & {\scriptsize{}$-296.64$}\tabularnewline
{\scriptsize{}250} & {\scriptsize{}$7246.88$} & {\scriptsize{}$0.02$} & {\scriptsize{}$144.94$} & {\scriptsize{}$630.17$} & {\scriptsize{}$-444.62$} & {\scriptsize{}$29.99$} & {\scriptsize{}$29.04$} & {\scriptsize{}$870.91$} & {\scriptsize{}$3786.57$} & {\scriptsize{}$-302.4189$} & {\scriptsize{}$5.66$} & {\scriptsize{}$0.04$} & {\scriptsize{}$0.23$} & {\scriptsize{}$1$} & {\scriptsize{}$-296.28$}\tabularnewline
{\scriptsize{}500} & {\scriptsize{}$3682.07$} & {\scriptsize{}$0.02$} & {\scriptsize{}$73.64$} & {\scriptsize{}$350.67$} & {\scriptsize{}$-420.32$} & {\scriptsize{}$15.41$} & {\scriptsize{}$38.91$} & {\scriptsize{}$599.60$} & {\scriptsize{}$2855.24$} & {\scriptsize{}$-296.7779$} & {\scriptsize{}$2.57$} & {\scriptsize{}$0.08$} & {\scriptsize{}$0.21$} & {\scriptsize{}$1$} & {\scriptsize{}$-296.79$}\tabularnewline
{\scriptsize{}1000} & {\scriptsize{}$2610.05$} & {\scriptsize{}$0.03$} & {\scriptsize{}$78.30$} & {\scriptsize{}$652.50$} & {\scriptsize{}$-368.68$} & {\scriptsize{}$6.98$} & {\scriptsize{}$56.14$} & {\scriptsize{}$391.86$} & {\scriptsize{}$3265.50$} & {\scriptsize{}$-296.5825$} & {\scriptsize{}$1.22$} & {\scriptsize{}$0.10$} & {\scriptsize{}$0.12$} & {\scriptsize{}$1$} & {\scriptsize{}$-296.48$}\tabularnewline
{\scriptsize{}2000} & {\scriptsize{}$1558.63$} & {\scriptsize{}$0.05$} & {\scriptsize{}$77.93$} & {\scriptsize{}$519.53$} & {\scriptsize{}$-328.80$} & {\scriptsize{}$3.78$} & {\scriptsize{}$96.09$} & {\scriptsize{}$363.22$} & {\scriptsize{}$2421.47$} & {\scriptsize{}$-297.0468$} & {\scriptsize{}$0.75$} & {\scriptsize{}$0.20$} & {\scriptsize{}$0.15$} & {\scriptsize{}$1$} & {\scriptsize{}$-296.32$}\tabularnewline
\hline
\end{tabular}{\scriptsize\par}
\end{sidewaystable}
\subsection{Non-Linear Real Business Cycle Models \label{subsec:Non-linear-RBC}}
This section reports the results of three studies using data generated
from the non-linear Real Business Cycle (RBC) model
described in Section~\ref{descriptionRBC} of the Supplement.
The true values of the parameters are set to $\beta=0.95$, $\delta=0.025$, $\alpha=0.90$, $\sigma=2$, $\rho_{a}=\rho_{e}=0.8$,
$\sigma_{e}=\sigma_{a}=0.5$. The measurement error covariance matrix
is set to $$\Sigma_{\eta}=\left[\begin{array}{cc}
0.0135 & 0\\
0 & 0.0137
\end{array}\right],$$ which corresponds to a large signal to noise ratio (SNR). We simulated
three data series of $T=50$, $100$, and $200$ observations. These numbers
of observations are chosen because they are roughly the same size
as standard ($T=50$ and $100$) and large $(T=200)$ macroeconomic time series. The particle filters and the parameter samplers are implemented in Matlab. All the computations are done on a single desktop computer with 6-CPU cores.
The {\bf first} study compares variance of the log of the estimated likelihoods for the
improved disturbance particle filter (IDPF), the block particle filter (BPF) and the tempered particle filter (TPF) using $N_{run}=100$ independent
runs holding the parameters fixed at the true values.
The proposal for the IDPF is estimated from the collection of particles generated by the $G$ particle filters running in parallel. These particle filters are run once and the same proposal is kept for different numbers of particles $N$. The tuning parameters of the TPF are set to the default values given in \citet{Herbst2019}.
Table~\ref{tab:The-variance-of log-likelihood} reports the variance
of the log of the estimated likelihood for the BPF, TPF, and IDPF methods
for the non-linear RBC model with $T=50, 100, 200$ time periods and
$N=100$, $250$, $500$, $1000$, and $2000$ particles. In general,
the IDPF is much more efficient than BPF and TPF for all $T$ and
$N$, for the non-linear RBC model. For example, the IDPF is
$5373$ times more efficient than BPF and $294020$ times more efficient
than TPF when the number of particles $N=500$ and $T=200$ time periods.
The {\bf second} study compares the ability of the CPM of \citet{Deligiannidis2018},
and the MPM methods to maintain the correlation between successive log-likelihood estimates
for the non-linear RBC model; the TPF is not used in the CPM method because it is not straightforward to correlate
all the random numbers, especially with the random walk Metropolis-Hastings mutation
steps and the adaptive tempering iterations. The correlation parameter in the CPM approach is set to $\rho_u=(0.5,0.75,0.8,0.9,0.99,0.9999,0.999999)$. The number of blocks in the MPM approach is set to $G=(2,4,5,10,100)$ and the correlation coefficient $\rho_u$ is set to $0.99$. Two different sorting methods are considered. The first sorts the state particles and the second sorts the disturbance particles. We ran the CPM and the MPM approaches for $N_{run}=100$ iterations holding the parameters fixed at the true values.
Figure~\ref{fig:The-estimated-correlation RBC Corr}
reports the correlation estimates between successive values of the log of the estimated likelihood
for the CPM method with disturbance sorting and state sorting. It shows
that: (a) the correlation estimates are quite low even when the correlation between the
random numbers are set to be very close to 1; (b) the correlation
estimates with the states sorted are 0.49, 0.28, 0.29 when the correlation between
random numbers is set to 0.999999, for $T=50$, $100$, $200$, respectively;
(c) sorting the state particles gives similar results to sorting the disturbance
particles.
Figure~\ref{fig:The-estimated-correlation RBC MPM} reports
the correlation estimates between the log of the estimated likelihoods
for four different methods: (i) the MPM with the BPF and disturbance sorting, (ii) the MPM with the IDPF and disturbance sorting, (iii) the MPM with the BPF and state sorting, and (iv) the MPM with the IDPF and state sorting.
The figure shows that: (a) the correlation estimates
are 0.98, 0.99, 0.98 when the number of blocks
is set to 100, for $T=50$, $100$, $200$, respectively, when sorting
the state particles and the IDPF are used. This suggests that the
proposed MPM sampler is able to maintain high correlations between
the logs of the estimated likelihoods in scenarios where the other methods
fail. (b) Sorting the disturbances gives bigger correlation estimates than sorting the states, especially when $T=200$. This is because the dimension of the disturbances is smaller than that of the states.
\begin{sidewaystable}
\caption{The variance of the log of the estimated likelihood obtained using three particle filter
methods: (1) the bootstrap particle filter (BPF), (2) the tempered particle filter (TPF), and (3) the improved disturbance particle filter (IDPF) for the non-linear RBC model with $T=50$, $100$, and $200$ time periods and $N=100,250,500,1000,2000$ particles. We also report the time normalised variance (TNV), the relative time normalised variance (RTNV), and the computation time (CT) per iteration in seconds as defined in Section \ref{SS: preliminaries}. \label{tab:The-variance-of log-likelihood} }
\centering{}%
\begin{tabular}{cccccccccccccc}
\hline
& & \multicolumn{4}{c}{{\footnotesize{}BF}} & \multicolumn{4}{c}{{\footnotesize{}TPF}} & \multicolumn{4}{c}{{\footnotesize{}IDPF}}\tabularnewline
{\footnotesize{}$T$} & {\footnotesize{}$N$} & {\footnotesize{}Var} & {\footnotesize{}CT} & {\footnotesize{}TNV} & {\footnotesize{}RTNV} & {\footnotesize{}Var} & {\footnotesize{}CT} & {\footnotesize{}TNV} & {\footnotesize{}RTNV} & {\footnotesize{}Var} & {\footnotesize{}CT} & {\footnotesize{}TNV} & {\footnotesize{}RTNV}\tabularnewline
\hline
\hline
{\footnotesize{}50} & {\footnotesize{}100} & {\footnotesize{}36.81} & {\footnotesize{}0.015} & {\footnotesize{}0.552} & {\footnotesize{}17.25} & {\footnotesize{}9.23} & {\footnotesize{}3.171} & {\footnotesize{}29.268} & {\footnotesize{}914.625} & {\footnotesize{}0.85} & {\footnotesize{}0.038} & {\footnotesize{}0.032} & {\footnotesize{}1}\tabularnewline
& {\footnotesize{}250} & {\footnotesize{}9.60} & {\footnotesize{}0.017} & {\footnotesize{}0.163} & {\footnotesize{}11.64} & {\footnotesize{}3.64} & {\footnotesize{}8.082} & {\footnotesize{}29.418} & {\footnotesize{}2101.290} & {\footnotesize{}0.33} & {\footnotesize{}0.044} & {\footnotesize{}0.014} & {\footnotesize{}1}\tabularnewline
& {\footnotesize{}500} & {\footnotesize{}4.69} & {\footnotesize{}0.029} & {\footnotesize{}0.136} & {\footnotesize{}11.33} & {\footnotesize{}1.60} & {\footnotesize{}15.843} & {\footnotesize{}25.349} & {\footnotesize{}2112.41} & {\footnotesize{}0.16} & {\footnotesize{}0.078} & {\footnotesize{}0.012} & {\footnotesize{}1}\tabularnewline
& {\footnotesize{}1000} & {\footnotesize{}2.22} & {\footnotesize{}0.050} & {\footnotesize{}0.111} & {\footnotesize{}11.10} & {\footnotesize{}0.84} & {\footnotesize{}31.532} & {\footnotesize{}26.487} & {\footnotesize{}2648.70} & {\footnotesize{}0.08} & {\footnotesize{}0.136} & {\footnotesize{}0.010} & {\footnotesize{}1}\tabularnewline
& {\footnotesize{}2000} & {\footnotesize{}1.03} & {\footnotesize{}0.080} & {\footnotesize{}0.082} & {\footnotesize{}10.25} & {\footnotesize{}0.42} & {\footnotesize{}63.942} & {\footnotesize{}26.856} & {\footnotesize{}3357.00} & {\footnotesize{}0.04} & {\footnotesize{}0.196} & {\footnotesize{}0.008} & {\footnotesize{}1}\tabularnewline
\hline
{\footnotesize{}100} & {\footnotesize{}100} & {\footnotesize{}1424.20} & {\footnotesize{}0.029} & {\footnotesize{}41.302} & {\footnotesize{}0.83} & {\footnotesize{}421.48} & {\footnotesize{}8.701} & {\footnotesize{}3667.297} & {\footnotesize{}73.48} & {\footnotesize{}779.84} & {\footnotesize{}0.064} & {\footnotesize{}49.910} & {\footnotesize{}1}\tabularnewline
& {\footnotesize{}250} & {\footnotesize{}760.33} & {\footnotesize{}0.030} & {\footnotesize{}22.810} & {\footnotesize{}3.29} & {\footnotesize{}375.70} & {\footnotesize{}20.346} & {\footnotesize{}7643.992} & {\footnotesize{}1101.44} & {\footnotesize{}86.75} & {\footnotesize{}0.080} & {\footnotesize{}6.940} & {\footnotesize{}1}\tabularnewline
& {\footnotesize{}500} & {\footnotesize{}596.48} & {\footnotesize{}0.052} & {\footnotesize{}31.017} & {\footnotesize{}22.22} & {\footnotesize{}214.52} & {\footnotesize{}37.74} & {\footnotesize{}8095.984} & {\footnotesize{}5799.41} & {\footnotesize{}10.27} & {\footnotesize{}0.128} & {\footnotesize{}1.396} & {\footnotesize{}1}\tabularnewline
& {\footnotesize{}1000} & {\footnotesize{}265.98} & {\footnotesize{}0.087} & {\footnotesize{}23.140} & {\footnotesize{}71.42} & {\footnotesize{}89.02} & {\footnotesize{}76.947} & {\footnotesize{}6849.822} & {\footnotesize{}21141.42} & {\footnotesize{}1.36} & {\footnotesize{}0.238} & {\footnotesize{}0.324} & {\footnotesize{}1}\tabularnewline
& {\footnotesize{}2000} & {\footnotesize{}50.54} & {\footnotesize{}0.149} & {\footnotesize{}7.531} & {\footnotesize{}36.21} & {\footnotesize{}16.28} & {\footnotesize{}146.671} & {\footnotesize{}2387.804} & {\footnotesize{}11489.83} & {\footnotesize{}0.52} & {\footnotesize{}0.398} & {\footnotesize{}0.208} & {\footnotesize{}1}\tabularnewline
\hline
{\footnotesize{}200} & {\footnotesize{}100} & {\footnotesize{}89615.69} & {\footnotesize{}0.035} & {\footnotesize{}3136.549} & {\footnotesize{}2.96} & {\footnotesize{}18178.00} & {\footnotesize{}19.732} & {\footnotesize{}358688.300} & {\footnotesize{}338.56} & {\footnotesize{}9459.50} & {\footnotesize{}0.112} & {\footnotesize{}1059.460} & {\footnotesize{}1}\tabularnewline
& {\footnotesize{}250} & {\footnotesize{}41655.73} & {\footnotesize{}0.053} & {\footnotesize{}2207.754} & {\footnotesize{}1760.57} & {\footnotesize{}9704.34} & {\footnotesize{}37.996} & {\footnotesize{}368726.100} & {\footnotesize{}294039.95} & {\footnotesize{}8.25} & {\footnotesize{}0.152} & {\footnotesize{}1.254} & {\footnotesize{}1}\tabularnewline
& {\footnotesize{}500} & {\footnotesize{}26864.78} & {\footnotesize{}0.098} & {\footnotesize{}2632.748} & {\footnotesize{}5372.95} & {\footnotesize{}2088.47} & {\footnotesize{}73.720} & {\footnotesize{}153962.010} & {\footnotesize{}314208.18} & {\footnotesize{}1.91} & {\footnotesize{}0.256} & {\footnotesize{}0.490} & {\footnotesize{}1}\tabularnewline
& {\footnotesize{}1000} & {\footnotesize{}6590.42} & {\footnotesize{}0.175} & {\footnotesize{}1153.324} & {\footnotesize{}2270.32} & {\footnotesize{}29.47} & {\footnotesize{}147.365} & {\footnotesize{}4342.850} & {\footnotesize{}8548.92} & {\footnotesize{}1.08} & {\footnotesize{}0.470} & {\footnotesize{}0.508} & {\footnotesize{}1}\tabularnewline
& {\footnotesize{}2000} & {\footnotesize{}919.93} & {\footnotesize{}0.296} & {\footnotesize{}272.299} & {\footnotesize{}677.36} & {\footnotesize{}9.49} & {\footnotesize{}294.720} & {\footnotesize{}2796.890} & {\footnotesize{}6957.44} & {\footnotesize{}0.52} & {\footnotesize{}0.774} & {\footnotesize{}0.402} & {\footnotesize{}1}\tabularnewline
\hline
\end{tabular}
\end{sidewaystable}
\begin{figure}
\caption{The estimated correlation of log of the estimated likelihoods $\left(\log\widehat{p}_{N}\left(y|\theta,u\right),\log\widehat{p}_{N}\left(y|\theta,u^{'}\right)\right)$
obtained using the two approaches: (1) CPM with disturbance sorting, (2) CPM with state sorting for $T=50,100,200$ \label{fig:The-estimated-correlation RBC Corr}}
\centering{}\includegraphics[width=15cm,height=6cm]{corr_sim_RBC}
\end{figure}
\begin{figure}
\caption{The estimated correlation of the log of the estimated likelihoods, $\log\overline{\widehat{p}}_{N}^{\left(s\right)}\left(y|\theta,\widetilde{u}\right)$
and $\log\overline{\widehat{p}}_{N}^{\left(s\right)}\left(y|\theta^{'},\widetilde{u}^{'}\right)$
obtained using 4 approaches: (1) MPM with the BPF and disturbance sorting, (2) MPM with the IDPF and disturbance sorting, (3) MPM with the BPF and state sorting, and (4) MPM with the IDPF and state sorting, for $T=50,100,200$. \label{fig:The-estimated-correlation RBC MPM}}
\centering{}\includegraphics[width=15cm,height=6cm]{ParPMMH_sim_RBC}
\end{figure}
The {\bf third} study compares the performance of the following PMMH samplers in estimating the parameters of the non-linear RBC model for $T=200$: (i) The delayed acceptance mixed PMMH (DA-MPM) with the improved disturbance particle filter (IDPF), disturbance sorting, and the correlation between random numbers used in estimating the log of estimated likelihoods set to $\rho_u=0.99$; (ii) the DA-MPM with the IDPF, state sorting, and the $\rho_u=0.99$; (iii) the DA-MPM with the bootstrap particle filter (BPF), disturbance sorting, and the $\rho_u=0.99$; (iv) the DA-MPM with the IDPF, disturbance sorting, and $\rho_u=0$, (v) the MPM with the IDPF, disturbance sorting, and the $\rho_u=0.99$; and (vi) the delayed acceptance correlated PMMH (DA-CPM) with $\rho_u=0.99$ and the BPF.
The samplers ran for 25000
iterations, with the initial 5000 iterations discarded as burn-in. All the samplers are initialized at the
posterior means from a second order approximation of the model obtained
from the central difference Kalman Filter (CDKF) proposed by \citet{Norgaard2000}. We chose this initialisation method because the variance
of the log of the estimated likelihood increased significantly in
some areas of the support of $\theta$ away from the true values,
making it difficult for the MCMC sampler to converge. We use the adaptive random
walk proposal of \citet{Roberts:2009} for $q\left(\theta^{'}|\theta\right)$ and use the adaptive scaling approach of \citet{Garthwaite:2015} to tune the Metropolis-Hastings acceptance probability to $20\%$. The covariance matrix of the proposals is fixed to the estimates from the preliminary run of the MCMC.
We fix $\beta$, $\delta$, $\alpha$, and $\sigma$ and estimate the four parameters $\rho_{a},\rho_{e},\sigma_{e},\sigma_{a}$. This gives a smaller number of parameters so that we can focus on the comparisons of different PMMH samplers. All the samplers ran on a single desktop computer with 6-CPU cores.
Table \ref{tab:Posterior-mean-estimates RBC} reports the inefficiency factors for each parameter and the relative time normalised inefficiency factor (RTNIV) of a PMMH sampler relative to the DA-MPM with the IDPF, disturbance sorting, and $\rho_u=0.99$ for the non-linear
RBC models with $T=200$ time periods. The computing time reported in the table is the time to run a single particle filter for the CPM and $G$ particle filters for the MPM approach using a single desktop computer with $6$ cores.
The table shows that: (a) the
DA-MPM sampler with the IDPF, disturbance sorting, $\rho_u=0.99$, and $N=100$ particles is much more
efficient than the CPM with $2000$ particles in terms of
$\widehat{\textrm{RTNIF}}_{\textrm{MAX}}$ and $\widehat{\textrm{RTNIF}}_{\textrm{MEAN}}$;
(b) the time taken for running the MPM method with $G=100$ particles filters each with $N=100$ particles
is similar to the CPM with $N=1000$ particles. The MPM method can be much faster than the CPM method if it was run using high-performance computing with a large number of cores. This also shows that the multidimensional sorting algorithm is quite expensive for
a large number of particles and long time periods; (c) the MPM sampler
allows us to use of a much smaller number of particles for each independent
PF and these multiple PFs can be run in parallel; (d) the performance of DA-MPM with IDPF, disturbance sorting, and $\rho_u=0.99$ is $4.49$ and $2.66$ times more efficient than the DA-MPM method with state sorting method when $N=250$ and it has similar performance to the DA-MPM approach with IDPF, disturbance sorting, and $\rho_u=0$; (e) the delayed acceptance MPM is much more efficient than the standard MPM approach. The delayed acceptance algorithm is 5 times faster on average because the target Metropolis-Hastings acceptance probability is set to $20\%$ using the \citet{Garthwaite:2015} approach.
\begin{sidewaystable}
\caption{Real Business Cycle (RBC) models using simulated dataset with $T=200$.
Sampler I: DA-MPM ($\rho_{u}=0.99$, disturbance sorting, and IDPF), Sampler II: DA-MPM ($\rho_{u}=0.99$, state sorting, and
IDPF) , Sampler III: DA-MPM ($\rho_{u}=0.99$, disturbance sorting,
and BPF), Sampler IV: DA-MPM ($\rho_{u}=0$, disturbance sorting,
and IDPF) Sampler V: MPM ($\rho_{u}=0.99$, disturbance sorting,
and IDPF), and Sampler VI: DA-CPM ($\rho_{u}=0.99$ and BPF). The table reports
the time normalised inefficiency factor (TNIV), the relative
time normalised inefficiency factor (RTNIV), and the computation time
(CT) per iteration (in seconds) as defined in Section \ref{SS: preliminaries}.
The symbol NA means the Markov chain gets stuck and does not converge. \label{tab:Posterior-mean-estimates RBC}}
\centering{}%
\begin{tabular}{c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
{\footnotesize{}Param} & \multicolumn{2}{c|}{{\footnotesize{}I}} & \multicolumn{2}{c|}{{\footnotesize{}II}} & \multicolumn{1}{c|}{{\footnotesize{}III}} & \multicolumn{2}{c|}{{\footnotesize{}IV}} & \multicolumn{2}{c|}{{\footnotesize{}V}} & \multicolumn{3}{c|}{{\footnotesize{}V1}}\tabularnewline
\hline
{\footnotesize{}$N$} & {\footnotesize{}100} & {\footnotesize{}250} & {\footnotesize{}100} & {\footnotesize{}250} & {\footnotesize{}250} & {\footnotesize{}100} & {\footnotesize{}250} & {\footnotesize{}100} & {\footnotesize{}250} & {\footnotesize{}500} & {\footnotesize{}1000} & {\footnotesize{}2000}\tabularnewline
\hline
{\footnotesize{}$\rho_{a}$} & {\footnotesize{}53.81} & {\footnotesize{}25.02} & {\footnotesize{}150.31} & {\footnotesize{}108.08} & {\footnotesize{} NA} & {\footnotesize{}60.81} & {\footnotesize{}27.69} & {\footnotesize{}112.66} & {\footnotesize{}15.34} & {\footnotesize{}NA} & {\footnotesize{}2495.09} & {\footnotesize{}439.28}\tabularnewline
{\footnotesize{}$\rho_{e}$} & {\footnotesize{}53.40} & {\footnotesize{}26.40} & {\footnotesize{}122.10} & {\footnotesize{}39.62} & {\footnotesize{} NA} & {\footnotesize{}43.70} & {\footnotesize{}24.50} & {\footnotesize{}39.12} & {\footnotesize{}15.88} & {\footnotesize{}NA} & {\footnotesize{}5040.98} & {\footnotesize{}874.69}\tabularnewline
{\footnotesize{}$\sigma_{a}$} & {\footnotesize{}72.82} & {\footnotesize{}24.07} & {\footnotesize{}116.61} & {\footnotesize{}34.45} & {\footnotesize{} NA} & {\footnotesize{}65.98} & {\footnotesize{}27.64} & {\footnotesize{}38.68} & {\footnotesize{}15.44} & {\footnotesize{}NA} & {\footnotesize{}4790.05} & {\footnotesize{}544.69}\tabularnewline
{\footnotesize{}$\sigma_{e}$} & {\footnotesize{}71.87} & {\footnotesize{}31.02} & {\footnotesize{}107.30} & {\footnotesize{}37.85} & {\footnotesize{} NA} & {\footnotesize{}40.32} & {\footnotesize{}31.51} & {\footnotesize{}34.91} & {\footnotesize{}17.65} & {\footnotesize{}NA} & {\footnotesize{}4901.89} & {\footnotesize{}943.21}\tabularnewline
\hline
{\footnotesize{}$\widehat{\textrm{IACT}}_{\textrm{MAX}}$} & {\footnotesize{}72.82} & {\footnotesize{}31.03} & {\footnotesize{}150.31} & {\footnotesize{}108.08} & {\footnotesize{} NA} & {\footnotesize{}65.99} & {\footnotesize{}31.51} & {\footnotesize{}112.66} & {\footnotesize{}17.65} & {\footnotesize{}NA} & {\footnotesize{}5040.98} & {\footnotesize{}943.22}\tabularnewline
{\footnotesize{}$\widehat{\textrm{TNIF}}_{\textrm{MAX}}$} & {\footnotesize{}327.69} & {\footnotesize{}347.85} & {\footnotesize{}858.27} & {\footnotesize{}1560.68} & {\footnotesize{} NA} & {\footnotesize{}296.95} & {\footnotesize{}353.23} & {\footnotesize{}2534.85} & {\footnotesize{}989.28} & {\footnotesize{}NA} & {\footnotesize{}22029.08} & {\footnotesize{}13535.21} \tabularnewline
{\footnotesize{}$\widehat{\textrm{RTNIF}}_{\textrm{MAX}}$} & {\footnotesize{}0.94} & {\footnotesize{}1} & {\footnotesize{}2.47} & {\footnotesize{}4.49} & {\footnotesize{} NA} & {\footnotesize{}0.85} & {\footnotesize{}1.02} & {\footnotesize{}7.29} & {\footnotesize{}2.84} & {\footnotesize{}NA} & {\footnotesize{}63.33} & {\footnotesize{}38.91} \tabularnewline
{\footnotesize{}$\widehat{\textrm{IACT}}_{\textrm{MEAN}}$} & {\footnotesize{}62.98} & {\footnotesize{}26.63} & {\footnotesize{}124.08} & {\footnotesize{}55.00} & {\footnotesize{} NA}& {\footnotesize{}52.71} & {\footnotesize{}27.84} & {\footnotesize{}56.34} & {\footnotesize{}16.08} & {\footnotesize{}NA} & {\footnotesize{}4307.00} & {\footnotesize{}700.47}\tabularnewline
{\footnotesize{}$\widehat{\textrm{TNIF}}_{\textrm{MEAN}}$} & {\footnotesize{}283.41} & {\footnotesize{}298.52} & {\footnotesize{}1791.72} & {\footnotesize{}794.20} & {\footnotesize{} NA} & {\footnotesize{}237.19} & {\footnotesize{}312.09} & {\footnotesize{}1267.65} & {\footnotesize{}901.28} & {\footnotesize{}NA} & {\footnotesize{}18821.59} & {\footnotesize{}10051.74} \tabularnewline
{\footnotesize{}$\widehat{\textrm{RTNIF}}_{\textrm{MEAN}}$} & {\footnotesize{}0.95} & {\footnotesize{}1} & {\footnotesize{}6.00} & {\footnotesize{}2.66} & {\footnotesize{} NA} & {\footnotesize{}0.79} & {\footnotesize{}1.05} & {\footnotesize{}4.25} & {\footnotesize{}3.02} & {\footnotesize{}NA} & {\footnotesize{}63.05} & {\footnotesize{}33.67} \tabularnewline
\hline
{\footnotesize{}CT} & {\footnotesize{}4.50} & {\footnotesize{}11.21} & {\footnotesize{}5.71} & {\footnotesize{}14.44} & {\footnotesize{}13.04} & {\footnotesize{}4.50} & {\footnotesize{}11.21} & {\footnotesize{}22.5} & {\footnotesize{}56.05} & {\footnotesize{}1.54} & {\footnotesize{} 4.37} & {\footnotesize{}14.35}\tabularnewline
\hline
\end{tabular}
\end{sidewaystable}
\subsection{Nonlinear Small Scale DSGE Model \label{subsec:SecondOrderSmallScale}}
This section reports on how well a number of PMMH samplers of interest estimate the parameters of the
nonlinear small scale DSGE model described in Section~\ref{Description of SmallScale Model} of the Supplement, using the same data as in Section~\ref{subsec:Small-Scale-DSGE}. The PMMH samplers considered are: (a) the MPM sampler with the IDPF, disturbance sorting, and $\rho_u=0.99$; (b) the MPM sampler with the IDPF, disturbance sorting, and $\rho_u=0$; and (c) the CPM sampler with the BPF, disturbance sorting, and $\rho_u=0.99$. Each sampler ran for $55000$ iterations, with the initial $5000$ iterations discarded as burn-in. All the samplers are initialized at the
posterior means from a second order approximation of the model obtained
from the central difference Kalman Filter (CDKF) proposed by \citet{Norgaard2000}. We use the adaptive random walk proposal of \citet{Roberts:2009} for $q\left(\theta^{'}|\theta\right)$ and the adaptive scaling approach of \citet{Garthwaite:2015} to tune the Metropolis-Hastings acceptance probability to $20\%$. The covariance matrix of the proposals is fixed to the estimates from the preliminary run of the MCMC. Section \ref{Description of SmallScale Model} of the supplement gives details on model specifications and model parameters.
Table \ref{tab:Posterior-mean-estimates MPM} reports the inefficiency factors for each parameter and the relative time normalised inefficiency factor (RTNIV) of a sampler relative to the DA-MPM with the IDPF, disturbance sorting, and $\rho_u=0.99$ with $T=80$ time periods. The computing time reported in the table is the time to run a single particle filter for the CPM and $G$ particle filters for the MPM approach using a single desktop computer with $6$ cores.
The table shows that: (a) the
DA-MPM sampler with the IDPF, disturbance sorting, $\rho_u=0.99$, and $N=250$ particles has similar performance to the DA-MPM approach with IDPF, disturbance sorting, and $\rho_u=0$; (b) the time taken to run MPM with $G=100$ particles filters each with $N=250$ particles
is similar to the CPM with $N=2000$ particles. The MPM method can be much faster than the CPM method if it runs using high-performance computing with a large number of cores. In addition, the multidimensional sorting algorithm is quite expensive for
a large number of particles; (c) Figure \ref{fig:The-traceplotscorrpmmhsmallscale} shows the trace plot of some of the parameters for the last $10000$ iterations estimated using the CPM method with $N=2000,3000$ particles. The CPM still gets stuck and does not converge even with $N=3000$ particles. Section~\ref{FurtherResultsOfSmallScale} of the supplement provides
further analysis comparing the linear and non-linear small scale DSGE models.
\begin{table}
\caption{Small scale DSGE model using a real dataset with $T=80$. Sampler I:
DA-MPM ($\rho_{u}=0.99$, disturbance sorting, and IDPF), Sampler
II: DA-MPM ($\rho_{u}=0$, disturbance sorting, and IDPF), Sampler
III: DA-CPM ($\rho_{u}=0.99$, disturbance sorting, and BPF). We also report the time normalised inefficiency factor (TNIV),
the relative time normalised inefficiency factor (RTNIV), and the
computation time (CT) per iteration in second, defined in Section~\ref{SS: preliminaries}. The symbol NA means the Markov chain got stuck and did not converge.
$\sigma^{m}_{r}$, $\sigma^{m}_{g}$, and $\sigma^{m}_{m}$ are the measurement error variances.
\label{tab:Posterior-mean-estimates MPM}}
\centering{}%
\begin{tabular}{c|c|c|c|c|}
\hline
{\footnotesize{}Param} & \multicolumn{1}{c|}{{\footnotesize{}I}} & \multicolumn{1}{c|}{{\footnotesize{}II}} & \multicolumn{2}{c|}{{\footnotesize{}III}}\tabularnewline
\hline
{\footnotesize{}$N$} & {\footnotesize{}250} & {\footnotesize{}250} & {\footnotesize{}2000} & {\footnotesize{}3000}\tabularnewline
\hline
{\footnotesize{}$\pi^{\left(A\right)}$} & {\footnotesize{}$139.24$} & {\footnotesize{}$157.92$} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$\tau$} & {\footnotesize{}$206.79$} & {\footnotesize{}$244.24$} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$\psi_{1}$} & {\footnotesize{}$169.58$} & {\footnotesize{}$169.89$} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$\psi_{2}$} & {\footnotesize{}$269.15$} & {\footnotesize{}$214.37$} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$\gamma^{\left(Q\right)}$} & {\footnotesize{}$158.96$} & {\footnotesize{}$148.30$} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$r^{\left(A\right)}$} & {\footnotesize{}$171.73$} & {\footnotesize{}$156.14$} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$\rho_{r}$} & {\footnotesize{}$385.22$} & {\footnotesize{}$275.34$} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$\rho_{g}$} & {\footnotesize{}$203.93$} & {\footnotesize{}$239.14$} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$\rho_{m}$} & {\footnotesize{}$334.20$} & {\footnotesize{}$516.61$} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$\sigma_{r}$} & {\footnotesize{}$292.06$} & {\footnotesize{}$206.95$} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$\sigma_{g}$} & {\footnotesize{}$318.61$} & {\footnotesize{}$433.34$} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$\sigma_{m}$} & {\footnotesize{}$177.47$} & {\footnotesize{}$172.45$} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$\sigma^{m}_{r}$} & {\footnotesize{}$193.66$} & {\footnotesize{}$202.54$} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$\sigma^{m}_{g}$} & {\footnotesize{}$190.19$} & {\footnotesize{}$197.82$} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$\sigma^{m}_{m}$} & {\footnotesize{}$187.61$} & {\footnotesize{}$222.94$} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
\hline
{\footnotesize{}$\widehat{\textrm{IACT}}_{\textrm{MAX}}$} & {\footnotesize{}385.22} & {\footnotesize{}516.61} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$\widehat{\textrm{TNIF}}_{\textrm{MAX}}$} & {\footnotesize{}2037.81} & {\footnotesize{}2732.87} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$\widehat{\textrm{RTNIF}}_{\textrm{MAX}}$} & {\footnotesize{}1} & {\footnotesize{}1.34} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$\widehat{\textrm{IACT}}_{\textrm{MEAN}}$} & {\footnotesize{}226.56} & {\footnotesize{}237.20} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$\widehat{\textrm{TNIF}}_{\textrm{MEAN}}$} & {\footnotesize{}1198.50} & {\footnotesize{}1254.79} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
{\footnotesize{}$\widehat{\textrm{RTNIF}}_{\textrm{MEAN}}$} & {\footnotesize{}1} & {\footnotesize{}1.05} & {\footnotesize{}NA} & {\footnotesize{}NA}\tabularnewline
\hline
{\footnotesize{}Time} & {\footnotesize{}5.29} & {\footnotesize{}5.29} & {\footnotesize{}6.25} & {\footnotesize{}14.04} \tabularnewline
\hline
\end{tabular}
\end{table}
\begin{figure}
\caption{The trace plots of some of the parameters of the non-linear (second order) small scale DSGE models estimated using DA-CPM with disturbance sorting, $\rho_u=0.99$, $N=2000$ particles (top) and $N=3000$ particles (bottom). \label{fig:The-traceplotscorrpmmhsmallscale}}
\centering{}\includegraphics[width=15cm,height=6cm]{TracePlots_CorrPMMH_NonLinearSmallScale}
\end{figure}
\section{Summary and conclusions \label{sec:Conclusions}}
Our article proposes a general particle MCMC approach (MPM) for estimating the posterior density of the parameters of complex and high-dimensional state-space models. It is especially useful when the bootstrap filter is inefficient, while the auxiliary particle filter cannot be used because the state transition density is computationally intractable.
The MPM method is a general extension of the PMMH method and consists of four parts; (a) it is based on an average of unbiased likelihood estimators; (b)~it combines
block and correlated PMMH methods; (c) a delayed acceptance proposal is used to speed up the computation, when estimating the likelihood is expensive; (d) an auxiliary disturbance particle filter sampler is proposed to estimate the likelihood. The MPM methodology is then applied to complex DSGE models with many latent state variables.
Our empirical results suggest that: (i) the improved disturbance particle filter (IDPF) is much more efficient than the standard bootstrap particle filter (BPF) and the tempered particle filter (TPF); (ii) the MPM maintains the correlation between logs of the estimated likelihoods in successive iterates much better than the CPM and BPM; (iii) the delayed acceptance version of the MPM sampler is much more efficient than the standard MPM sampler; (iv) the MPM with disturbance sorting is more efficient than the MPM with state sorting; (v) the performance of the MPM just using block sampling is as efficient as that using both block sampling and correlated sampling.
Finally, we believe that the methods in the paper will be very useful for many other models
where particle alternative sophisticated methods, such as the particle Gibbs, are either
inefficient or impossible to use; e.g., partially observed diffusions and large panel data models. Future research will also consider developing
better proposals for the parameters $\theta$.
\begin{comment}
\section*{COMMENT}
In the last paragraph, maybe
\begin{itemize}
\item
we should take out \lq and complex DSGE models \rq{} as we deal with such models in the paper.
\item
\lq partially observed diffusions\rq{}
maybe we should tell reader, what parts of the methods will be useful? Is the disturbance part useful, is delayed acceptance useful?
\item
\lq We note that this methodology is also particularly beneficial\rq{}
maybe to \lq We note that all, or parts, of this methodology will also particularly beneficial\rq{}
\end{itemize}
\section*{END COMMENT}
\end{comment}
\bibliographystyle{apalike}
|
\section{Introduction}
\label{sec:intro}
Dense matter inside neutron stars (NSs) could contain unbound quarks that retain some vestige of the forces described by the fundamental theory of the strong interaction, Quantum Chromodynamics (QCD).
Lacking exact methods for its solution, numerous models have been constructed to investigate the hadron-to-quark transition in the region of high baryon density and zero temperature. On the other hand, the phase diagram of QCD at low density (zero or small baryon chemical potential) and high temperature is amenable to precision numerical studies which clearly point to a crossover with no clear phase boundary between the hadron resonance gas and the quark-gluon plasma~\cite{Borsanyi:2010cj}.
Recently, Kapusta and Welle~\cite{Kapusta:2021ney} (KW hereafter)
have proposed a crossover model for the hadron-to-quark transition in NSs to mimic the crossover feature of finite temperature lattice studies. The key trait of this model is an analytic mixing or switching function that accounts for the partial pressure of each component as a function of a single parameter - the baryon chemical potential. They found that NSs as massive as $\sim2.2\,{\rm M}_{\odot}$ could be supported by their crossover equation of state (EOS).
As hadrons/nucleons and quarks both appear explicitly as separate degrees of freedom in the KW description, it is straightforward to keep track of their individual contributions to the total pressure. KW report that within their model between $1-10\%$ of the total pressure could be contributed by quark matter in the core.
In treatments in which hadron and quark interactions are intermingled, the individual contributions from hadrons and quarks to the total pressure may not be possible to disentangle.
Given that EOSs with first-order phase transitions treated using the Maxwell construction face some challenges in obtaining the stable $\sim 2\,{\rm M}_{\odot}$ NSs that have been observed, it is worthwhile to investigate alternatives such as the crossover model for hybrid stars.
Our objective in this paper is two-fold: first, we extend the pure neutron matter (PNM) model of KW~\cite{Kapusta:2021ney}
to include $\beta$-equilibrium, as well as a crust for the hybrid star. We also include vector interactions among quarks. These modifications enable a direct comparison to other approaches in the literature~\cite{Han:2019bub}. Second, we investigate $g$-mode~oscillations, a potentially observable signature of the hadron-to-quark transition.
A $g$-mode~is a specific fluid oscillation where a parcel of fluid is displaced against the background of a stratified environment inside a neutron star. While pressure equilibrium is rapidly restored via sound waves, chemical equilibrium can take longer, causing buoyancy forces to oppose the displacement. Since cold NSs are not convective, the opposing force sets up stable oscillations, with a typical frequency, called the (local) Brunt--V\"ais\"al\"a~frequency. The kind of core $g$-modes~we study here were introduced in~\cite{RG92,RG94,Lai:1993di,Lai:1998yc} and in a recent work, which we shall refer to as Paper I~\cite{Jaikumar:2021jbw}, we showed that the $g$-mode~frequency rises steeply with the onset of quarks due to a rapid change in the equilibrium sound speed (see also Ref.~\cite{Wei:2018tts}). Since $g$-mode~oscillations couple to tidal forces, they may be excited during the merger of two NSs and provide information on the interior composition, specifically quarks here.
In Paper I, we chose a Gibbs construction for the mixed phase, which yields NS properties that are more compatible with astrophysical constraints than a Maxwell construction.\footnote{See Table VII in Ref.~\cite{Han:2019bub} for a comprehensive comparison of Gibbs, Maxwell and certain crossover models.} In this work, we present the systematics of the $g$-mode~frequency for hybrid stars in a crossover scenario, adopting the generalized KW model as a representative of this class.
It is worth noting that Maxwell-constructed first-order phase transitions cannot generate $g$-modes~in the transition region, because the equilibrium and adiabatic sound speeds both become zero (due to frozen pressure and composition over the phase coexistence region) there.
In the first phase of this work, we will restrict ourselves to zero temperature without the effects of superfluidity in both nucleons and quarks.
To the best of our knowledge, this is the first study of $g$-mode~for hybrid stars in a crossover model. References to earlier work in which hadronic EOSs with and without superfluidity were used to investigate $g$-mode~frequencies can be found in Paper I (see Refs. [21]-[35] therein).
Models akin to the crossover model of KW, but with some differences, have been considered earlier in the literature. Examples include the smooth crossover model of Ref.~\cite{Dexheimer:2014pea}, interpolated EOSs considered in Refs.~\cite{Baym:2017whm,Masuda:2012ed,Fukushima:2015bda,Kojo:2014rca}, and quarkyonic models of Refs.~\cite{McLerran:2018hbz,Zhao:2020dvu,Jeong:2019lhv,Sen:2020qcd}, et cetera.
In the chiral model of Refs.~\cite{Papazoglou:1998vr,Dexheimer:2008ax,Dexheimer:2009hi}, a scalar field $\Phi$, acting as an order parameter, is responsible for the deconfinement phase transition which can be first-order or crossover depending upon temperature and baryon chemical potential. Depending on the specific EOS models used in the hadronic and quark phases, chemical potential and pressure equilibrium between the two phases - of either the Maxwell or the Gibbs sort - may not be realized~\cite{Baym:2017whm}. In such cases, several interpolation procedures have been adopted to connect the two phases on the premise that at high supra-nuclear densities, a purely hadronic phase is untenable.
In quarkyonic models with a momentum shell structure~\cite{McLerran:2018hbz,Zhao:2020dvu},
quarks emerge at relatively low (but still
supra-saturation) densities but remain bound by strong interactions below the Fermi surface: while hadrons and quarks are separated in momentum space, they coexist in configuration space.
A key parameter that enters the calculation of the $g$-mode, is
the squared adiabatic speed of sound $c_{\rm ad}^2 =(\partial P/\partial \varepsilon)|_{y_i,\beta}$, where $P$ and $\varepsilon$ are the total pressure and energy density, respectively, and $y_{i,\beta}$ refer to the partial fractions of each component in beta-equilibrium.\footnote{$c_{\rm ad}^2$ is different from the squared equilibrium speed of sound $c_{\rm eq}^2 = dP/d\varepsilon$ commonly defined as $c_s^2$ in the literature for static EOS models.}
We find that in
the quarkyonic ``shell'' models of Refs.~\cite{McLerran:2018hbz,Zhao:2020dvu}, $c_{\rm ad}^2$ becomes discontinuous
with respect to density at the shell boundary.
To address $g$-modes~in this specific category of quarkyonic shell models, and to
extend the study of NS oscillations to span the various ways in which quarks can affect the dense matter EOS and the properties of NSs, such discontinuities must be smoothed.
As such a task is outside the scope of the present work, we do not consider these quarkyonic models in this paper.
The organization of this paper is as follows. In Sec.~\ref{sec:EOS}, we describe the EOSs for pure hadronic matter, pure quark matter and leptons used in the construction of the
KW crossover model of Ref.~\cite{Kapusta:2021ney}
by extending it to include $\beta$-equilibrium and interactions between quarks. The rationale for our parameter choices and the basic features of this model are also highlighted here for orientation.
Sec.~\ref{sec:multi_compt}
reviews
the thermodynamics of a multicomponent system as pertinent to the KW description of crossover matter generalized here to describe neutron-star matter (NSM).
In Sec.~\ref{sec:KWG}, the KW model formulation of crossover matter is described followed by the procedure to render the unconstrained system\footnote{The unconstrained system here refers to matter in which baryon number conservation, charge neutrality and $\beta$-equilibrium are not imposed.} in $\beta$-equilibrium. This section also contains
a comparison of the extended KW model with those of selected quarkyonic shell models.
In Sec.~\ref{sec:soundspeed}, the calculation of the equilibrium and adiabatic sound speeds in the crossover and Gibbs approaches are outlined.
In Sec.~\ref{sec:Results}, we present results for the chosen EOSs and their associated NS structural properties as well as the two speeds of sound in the crossover model, and discuss emergent differences from other models that include a phase transition. This section also contains results for the sound speed difference, the Brunt--V\"ais\"al\"a~frequency, and the $g$-mode~frequency in hybrid stars along with their interpretation. Our conclusions and outlook are in Sec.~\ref{sec:Concs}.
\section{Equation of State}
\label{sec:EOS}
In the KW description of the crossover transition, EOSs in the pure hadronic and quark phases are combined using a mixing or switch function that depends on the baryon chemical potential (to be described below).
We therefore begin by discussing EOS models that we use in each of these two sectors.
As we extend the KW model to include leptons, the EOS in the leptonic sector is also provided. These EOSs are first set forth without reference to baryon number conservation, charge neutrality, as well as chemical equilibrium (i.e., unconstrained), which are imposed at the appropriate junctures.
\subsection{Pure hadronic matter}
To describe nucleons, we use the Zhao-Lattimer (ZL) \cite{Zhao:2020dvu} parametrization of the EOS of neutron-star matter. With reasonable adjustments of its parameters, this EOS can be made consistent with laboratory data at nuclear saturation density $n_{\rm sat}\simeq 0.16~{\rm fm}^{-3}$ as well as recent chiral effective field theory calculations
of Ref.~\cite{Drischler:2020hwi,Drischler:2020fvz} in which error quantifications up to $\sim2.0\,n_{\rm sat}$ were made.
The high-density behavior can be controlled by varying the slope of the symmetry energy parameter, $L$, at $n_{\rm sat}$ within the range established from analyses of nuclear and observational data,
see, e.g., Ref.~\cite{Lattimer:2012xj}, and a power-law index. Consistency with astrophysical data on known masses and radii of NSs is also attainable with this EOS.
To begin with, no constraints are placed on the multi-particle system and therefore the independent variables are the baryon density $n_{\rm B}$ and the individual particle fractions $y_n$, $y_p$. The total energy density of nucleons with a common mass $m_H=939.5$ MeV is described by the ZL functional
\begin{eqnarray}
\varepsilon_H &=& \frac{1}{8\pi^2\hbar^3}\sum_{h=n,p}\left\{k_{Fh}(k_{Fh}^2 + m_H^2)^{1/2}(2k_{Fh}^2 + m_H^2)\right. \nonumber \\
&-& \left. m_H^4\ln\left[\frac{k_{Fh}+(k_{Fh}^2 + m_H^2)^{1/2}}{m_H}\right]\right\} \nonumber \\
&+& 4 n_{\rm B}^2 y_n y_p \left\{\frac{a_0}{n_{\rm sat}}
+\frac{b_0}{n_{\rm sat}^{\gamma}} [n_{\rm B}(y_n + y_p)]^{\gamma - 1}\right\} \nonumber \\
&+& n_{\rm B}^2 (y_n - y_p)^2\left\{\frac{a_1}{n_{\rm sat}}
+ \frac{b_1}{n_{\rm sat}^{\gamma_1}}[n_{\rm B}(y_n + y_p)]^{\gamma_1-1}\right\} \nonumber \\
\end{eqnarray}
The Fermi momentum of nucleon species $h$\footnote{While the symbols ``$H$'' and ``$h$'' connote hadrons, they actually refer to nucleons in the context of this paper.} is given by $k_{Fh} = (3\pi^2\hbar^3n_{\rm B} y_h)^{1/3}$. The constants $a_0, b_0$ and $\gamma$ refer to isospin symmetric matter, whereas $a_1, b_1$ and $\gamma_1$ to that of isospin asymmetric matter. The chosen values of these constants are listed in a later section.
The corresponding chemical potentials are
\begin{eqnarray}
\mu_n &=& (k_{Fn}^2 + m_H^2)^{1/2} \nonumber \\
&+& 4 n_{\rm B} y_p \left\{\frac{a_0}{n_{\rm sat}}+ \frac{b_0}{n_{\rm sat}^{\gamma}}[n_{\rm B}(y_n + y_p)]^{\gamma-1}\right\}
\nonumber \\
&+& 4 n_{\rm B}^2 y_p y_n \frac{b_0}{n_{\rm sat}^{\gamma}}(\gamma-1)[n_{\rm B}(y_n + y_p)]^{\gamma-2}
\nonumber \\
&+& 2 n_{\rm B} (y_n-y_p)\left\{\frac{a_1}{n_{\rm sat}} + \frac{b_1}{n_{\rm sat}^{\gamma_1}}[n_{\rm B}(y_n+y_p)]^{\gamma_1-1}\right\} \nonumber \\
&+& n_{\rm B}^2 (y_n-y_p)^2 \frac{b_1}{n_{\rm sat}^{\gamma_1}}(\gamma_1-1)[n_{\rm B}(y_n + y_p)]^{\gamma_1-2} \, , \\
\mu_p &=& (k_{Fp}^2 + m_H^2)^{1/2} \nonumber \\
&+& 4 n_{\rm B} y_n \left\{\frac{a_0}{n_{\rm sat}}+ \frac{b_0}{n_{\rm sat}^{\gamma}}[n_{\rm B}(y_n + y_p)]^{\gamma-1}\right\}
\nonumber \\
&+& 4 n_{\rm B}^2 y_p y_n \frac{b_0}{n_{\rm sat}^{\gamma}}(\gamma-1)[n_{\rm B}(y_n + y_p)]^{\gamma-2}
\nonumber \\
&-& 2 n_{\rm B} (y_n-y_p)\left\{\frac{a_1}{n_{\rm sat}} + \frac{b_1}{n_{\rm sat}^{\gamma_1}}[n_{\rm B}(y_n+y_p)]^{\gamma_1-1}\right\} \nonumber \\
&+& n_{\rm B}^2 (y_n-y_p)^2 \frac{b_1}{n_{\rm sat}^{\gamma_1}}(\gamma_1-1)[n_{\rm B}(y_n + y_p)]^{\gamma_1-2} \,.
\end{eqnarray}
Note the opposite $y_h$ and signs in the second and fourth terms of $\mu_n$ and $\mu_p$, respectively.
The pressure is obtained from the thermodynamic identity
\begin{equation}
P_H = n_{\rm B}\sum_{h=n,p}\mu_h y_h - \varepsilon_H
\end{equation}
and the equilibrium speed of sound from
\begin{equation}
\left(\frac{
c_{\rm eq}}{c}\right)^2 = \frac {d P_H}{d \varepsilon_H} \,.
\end{equation}
The adiabatic speed of sound is obtained by taking partial derivatives of the pressure and the total energy density with respect to baryon density while keeping all particle fractions fixed
\begin{equation}
\left(\frac{
c_{\rm ad}}{c}\right)^2 = \left.\frac{\partial P_H}{\partial n_{\rm B}}\right|_{y_h}
\left(\left.\frac{\partial \varepsilon_H}{\partial n_{\rm B}}\right|_{y_h}\right)^{-1}~.
\end{equation}
This is made particularly convenient by the choice of starting with a completely unconstrained system.
\subsection{Pure quark matter}
For the calculation of the quark EOS, we use the vMIT bag model~\cite{Gomes:2018eiv, Klahn:2015mfa}. The Lagrangian density of this model is
\begin{equation}
\mathcal{L}=\sum_{q=u,d,s}\left[\bar{\psi}_{q}\left(i \slashed{\partial} -m_{q}-B\right) \psi_{q}+\mathcal{L}_{\mathrm{vec}}\right] \Theta \,,
\end{equation}
where $\mathcal{L}_{\mathrm{vec}}$ describes repulsive interactions between quarks of mass $m_{q}$ confined within a bag as denoted by the $\Theta$ function:
\begin{equation}
\mathcal{L}_{\text {vec }}=-G_{v} \sum_{q} \bar{\psi} \gamma_{\mu} V^{\mu} \psi+\left(m_{V}^{2} / 2\right) V_{\mu} V^{\mu} \,.
\end{equation}
$B$ is a constant which reflects the cost of confining the quarks inside the bag, and the $m_{q}$ are the current quark masses.
Perturbative contributions~\cite{Kurkela:2009gj,Kurkela:2014vha}
are not included in vMIT because these become relevant at densities well above those achievable in the cores of the most massive neutron stars.
The state functions, energy density, chemical potential, and pressure, corresponding to the above Lagrangian (\textit{before} any constraints of baryon number conservation, charge neutrality, and chemical equilibrium are applied) for matter containing $u,d$ and $s$ quarks are
\begin{eqnarray}
\varepsilon_Q &=& \sum_{q=u,d,s}\varepsilon_q + \frac{1}{2}a~\hbar~[n_{\rm B}(y_u+y_d+y_s)]^2
+ \frac{B}{\hbar^3} \\
\varepsilon_q &=& \frac{3}{8\pi^2\hbar^3}\left\{k_{Fq} (k_{Fq}^2 + m_q^2)^{1/2}(2 k_{Fq}^2 + m_q^2)
\right. \nonumber \\
&-& \left. m_q^4\ln\left[\frac{k_{Fq}+(k_{Fq}^2 + m_q^2)^{1/2}}{m_q}\right]\right\} \\
\mu_q &=& (k_{Fq}^2 + m_q^2)^{1/2} + a~\hbar~n_{\rm B} (y_u + y_d + y_s) \\
P_Q &=& n_{\rm B} \sum_{q=u,d,s} \mu_q y_q - \varepsilon_Q \,,
\end{eqnarray}
where $a \equiv (G_v/m_V)^2$ and $k_{Fq} = (\pi^2 \hbar^3 n_{\rm B} y_q)^{1/3}$. The value of the vector interaction parameter $a$ is varied in the range $(0.1-0.3)~ \rm{fm}^{-2}$ to obtain different stiffness in the quark sector.
\subsection{Leptons}
Owing to the smallness of the electromagnetic fine structure constant $\alpha \simeq 1/137$, leptons are treated as non-interacting, relativistic particles for which
\begin{eqnarray}
\varepsilon_L &=&\frac{1}{8\pi^2\hbar^3}\sum_l\left\{k_{Fl}(k_{Fl}^2+m_l^2)^{1/2}(2k_{Fl}^2 + m_l^2)\right. \nonumber \\
&-& \left. m_l^4\ln\left[\frac{k_{Fl}+(k_{Fl}^2 + m_l^2)^{1/2}}{m_l}\right]\right\} \\
\mu_l &=& (k_{Fl}^2 + m_l^2)^{1/2} \\
P_L &=& n_{\rm B} \sum_l y_l \mu_l - \varepsilon_L \\
k_{Fl} &=& (3\pi^2\hbar^3 n_{\rm B} y_l)^{1/3} ~~;~~ l=e,\mu~.
\end{eqnarray}
At low baryon densities only electrons are present in the system, with muons appearing at a density $n_{\rm B}$
such that $\mu_e -m_{\mu} = 0$. Depending on the parametrization choice, this condition also gives the density at which muons vanish.
\section{Thermodynamics of multicomponent systems}
\label{sec:multi_compt}
The original formulation of the KW approach deals with one nucleon, the neutron, and three massless quarks
(the latter are, operationally, a single species with multiplicity 3).
Before the KW approach can be applied to more realistic neutron-star matter with the EOSs of the previous section, it must be generalized to include several particle species. To that end, we begin with a brief review of multicomponent thermodynamics to introduce the fundamental result from which the aforementioned generalization will be
performed in the next section. The relations laid out below are particularly helpful in highlighting the role of the (baryon chemical potential dependent) switch function $S$ which is instrumental in realizing a crossover transition.
The number density of a single-component system in the grand-canonical ensemble is given by the total derivative of the pressure with respect to the chemical potential,
\begin{equation}
n=\frac{dP}{d\mu}~.
\end{equation}
The equivalent expression for a multicomponent system is obtained from the grand potential
\begin{equation}
\Phi(T,V,\mu_i)=U-TS-\sum_i N_i\mu_i
\end{equation}
or, in units of energy-density,
\begin{equation}
\phi=\varepsilon-Ts-\sum_i n_i\mu_i ~.
\end{equation}
The differential of $\phi$ is
\begin{equation}
d\phi=d\varepsilon-sdT-\sum_i n_i d\mu_i
\end{equation}
which implies that the number density of particle species $i$ is given by
\begin{equation}
n_i=-\left.\frac{\partial \phi}{\partial \mu_i}\right|_{T,\mu_j}~.
\end{equation}
The thermodynamic identity
\begin{equation}
\varepsilon = Ts-P + \sum_i n_i\mu_i
\label{TI}
\end{equation}
means that $P=-\phi$ and therefore
\begin{equation}
n_i=\left.\frac{\partial P}{\partial \mu_i}\right|_{\mu_j} \label{eqni}
\end{equation}
where the temperature $T$ has been suppressed as, in what follows, only cold matter is considered.
\Eqn{eqni} is central to the subsequent discussion where we show the manner in which the individual number densities of the various nucleonic and quark species are modified by the switch function $S$ of the KW machinery.
\section{Unconstrained and beta-equilibrated crossover matter}
\label{sec:KWG}
In this section, we start with the crossover EOS where
baryon number conservation, charge neutrality, and weak interaction equilibrium are not imposed, i.e., ``unconstrained'' matter. Working with unconstrained quantities enables us to calculate the various partial derivatives required in the determination of the squared adiabatic speed of sound $c_{\rm ad}^2$ (see Sec.~\ref{sec:soundspeed}) prior to the imposition of the conditions mentioned above.
In the KW description of crossover matter, the pressure is given by
\begin{equation}
P_B = (1-S)P_H + S \, P_Q \label{eqPB}
\end{equation}
where $P_H$ and $P_Q$ are the hadron and quark pure-phase pressures respectively, and the switch function
\begin{equation}
S = \exp\left[-\left(\frac{\mu_0}{\mu}\right)^4\right]
\end{equation}
gives the fraction of quark matter to the total baryonic matter in the crossover setting, with $\mu$ being the average hadronic chemical potential
\begin{equation}
\mu = \frac{n_n \mu_n + n_p \mu_p}{n_n + n_p}~\,,
\end{equation}
and $\mu_0$ a typical energy scale for the crossover. This choice for $\mu$ will be justified in the next section.
Applying \Eqn{eqni} to hadrons leads to
\begin{eqnarray}
n_h^* &=& (1-S)\frac{\partial P_H}{\partial \mu_h}+S\frac{\partial P_Q}{\partial \mu_h}+(P_Q-P_H)\frac{\partial S}{\partial \mu_h} \nonumber \\
&=& (1-S)n_h+0+(P_Q-P_H) \frac{4\mu_0^4S}{\mu^5}\frac{\partial \mu}{\partial \mu_h} \nonumber \\
&=& (1-S)n_h+(P_Q-P_H) \frac{4\mu_0^4S}{\mu^5}\frac{n_h}{n_n+n_p} \nonumber \\
&=& n_h\left[1-S\left(1-\frac{4\mu_0^4}{\mu^5}\frac{P_Q-P_H}{n_n+n_p}\right)\right] \,,
\label{eqnh}
\end{eqnarray}
where $n_i^*=n_{\rm B} y_i^*$ refers to a crossover-matter density and $n_i = n_{\rm B} y_i$ to a pure-phase density. Thus, in the present context, the starred fractions are the physical quantities whereas the unstarred ones are merely bookkeeping devices. For leptons this distinction is irrelevant.
For quarks, one obtains
\begin{eqnarray}
n_q^* &=& (1-S)\frac{\partial P_H}{\partial \mu_q}+S\frac{\partial P_Q}{\partial \mu_q}+(P_Q-P_H)\frac{\partial S}{\partial \mu_q} \nonumber \\
&=& 0+S\,n_q+0 = S\,n_q \,.
\label{qno}
\end{eqnarray}
Finally, the energy density $\varepsilon$ is obtained from \Eqn{TI} using \Eqn{eqPB} for the pressure, Eqs.~(\ref{eqnh})-(\ref{qno}) for the number densities of hadrons and quarks, respectively, and the pure-phase chemical potentials defined in Sec.~\ref{sec:EOS}.
\subsection{Beta-equilibrium}
We turn now to the discussion of neutron-star matter that consists of nucleons, leptons and quarks.
Initially, the system is entirely unconstrained with $n_{\rm B}$ and $y_i$ ($i=n,p,u,d,s,e,\mu$) as the free variables. Then, strong equilibrium
\begin{equation}
\mu_n = 2\mu_d + \mu_u ~~;~~\mu_p = 2\mu_u + \mu_d
\end{equation}
and weak equilibrium
\begin{equation}
\mu_n = \mu_p + \mu_e ~~;~~\mu_e = \mu_\mu ~~;~~ \mu_d = \mu_s
\end{equation}
are enforced, as well as charge neutrality
\begin{equation}
n_p^*+(2n_u^*-n_d^*-n_s^*)/3-(n_e+n_{\mu})=0
\end{equation}
and baryon number conservation
\begin{equation}
n_n^*+n_p^*+(n_u^*+n_d^*+n_s^*)/3-n_{\rm B}=0 ~.
\label{barno}
\end{equation}
These conditions eliminate the particle fractions in favor of the total baryon density:
\begin{equation}
y_i \rightarrow y_{i,\beta}(n_{\rm B}) ~~;~~ i=n,p,u,d,s,e,\mu
\end{equation}
\subsection{Comparison of Kapusta-Welle (KW) with McLerran-Reddy (MR) and Zhao-Lattimer (ZL) EOSs}
In this subsection, we briefly discuss interesting similarities and differences between the crossover model of KW~\cite{Kapusta:2021ney} and recently proposed quarkyonic ``shell'' models of MR~\cite{McLerran:2018hbz} and ZL~\cite{Zhao:2020dvu}, and explain the reason why the latter is not suitable for $g$-mode~calculations in its present form.
The baryon number densities in the quarkyonic matter descriptions of MR~\cite{McLerran:2018hbz} and ZL~\cite{Zhao:2020dvu} are
\begin{eqnarray}
n_h^* &=& \frac{k_{Fh}^3-k_{0h}^3}{3\pi^2\hbar^3}
=\frac{k_{Fh}^3}{3\pi^2\hbar^3}\left(1-\frac{k_{0h}^3}{k_{Fh}^3}\right)
= n_h\left(1-\frac{k_{0h}^3}{k_{Fh}^3}\right) \,, \nonumber \\
\end{eqnarray}
where $k_{0h}$ are the minimum momenta of hadrons or nucleons
in quarkyonic matter, which depend on the corresponding Fermi momenta $k_{Fh}$ and thus the baryon number density. The precise way in which $k_{Fh} - k_{0h}$ depends on a chosen momentum scale $\Lambda$ and a common transition density $n_t$ is detailed in Eq.~(17) of Ref.~\cite{Zhao:2020dvu}.
Comparing the above expression to \Eqn{eqnh} from the previous section, it becomes clear that the presence of the hadron shell in the MR and ZL approaches forces hadrons to higher-momentum states much like $S$ does in the KW scheme:
\begin{equation}
S\left(1-\frac{4\mu_0^4}{\mu^5}\frac{P_Q-P_H}{n_n+n_p}\right) \stackrel{\wedge}{=} \frac{k_{0h}^3}{k_{Fh}^3} \,.
\end{equation}
This means that any particle species participating in $S(\mu)$ (that is, a species $i$ for which $\partial \mu/\partial \mu_i\neq0$) will invariably inherit a shell-like term in its crossover-matter number density.
In quarkyonic matter realizations, such a term is desirable for baryons but not for quarks and therefore $\mu$ must be a function of baryonic chemical potentials only.
Note that in the case of KW, the quark densities in crossover-matter $n_q^*$ are the product of the corresponding pure-phase densities and the quark-to-baryon fraction $S$ which is an \textit{a priori} assumption.
On the other hand, for both MR and ZL models, the densities and fractions of baryons and quarks in the quarkyonic phase are determined by the solution of the equilibrium equations.
Here, we should point out that in the MR and ZL implementations of the quarkyonic matter scenario, the nucleonic Fermi momenta are weakly dependent on baryon density when the latter exceeds the transition density, $n_t$; that is,
\begin{equation}
|k_{\infty,i}-k_{Fi,\beta}|/k_{\infty,i}\ll 1 ~~\text{for all}~ n_{\rm B}>n_t \,,
\end{equation}
where $k_{\infty,i}\equiv k_{Fi,\beta}(n_{\rm B} \rightarrow \infty)$.
On the other hand, the nucleonic chemical potentials and, by extension, the pressure change very rapidly with $k_{Fi}$ for $n_{\rm B}>n_t$ due to the presence of denominators $\propto(1-K_i)$ (see Eqs.~(19)-(20) in \cite{Zhao:2020dvu}) in their kinetic parts, where
\begin{eqnarray}
K_i^{\rm MR} &=& \left(\frac{k_{0i}}{k_{Fi}}\right)^2 \left(1 + \frac{2\Lambda^3}{k_{Fi}^3}\right)\\
K_i^{\rm ZL} &=& \left(\frac{k_{0i}}{k_{Fi}}\right)^2 \left(1 + \frac{\Lambda^2}{k_{Fi}^2}\right)~.
\end{eqnarray}
These terms remain close to zero [i.e. $(1-K_i)^{-1}\rightarrow \infty$], in equilibrium matter as a result of the aforementioned behavior of the nucleonic Fermi momenta, throughout the quarkyonic regime.
These two features of MR and ZL are responsible for divergent pressure derivatives with respect to $k_{Fi}$ which, in turn, lead to superluminal adiabatic sound speeds. Thus MR and ZL, in their current formulations, are unsuitable for our purposes and we do not apply them in $g$-mode~calculations for crossover matter.
\section{Sound speeds}
\label{sec:soundspeed}
In this section, we describe how calculations of the squared adiabatic and equilibrium sound speeds, $c_{\rm ad}^2$ and $c_{\rm eq}^2$, required in the calculation
of $g$-mode~frequencies, are performed. As one of our objectives is to provide contrasts between $g$-mode~frequencies in crossover matter and the case of a first-order transition treated via the Gibbs construction, both cases are considered below.
\subsection{Sound speeds in crossover matter}
Within the KW framework, the total pressure and energy density in the crossover region are
\begin{eqnarray}
P &=& P_B + P_e + P_{\mu} \\
\varepsilon &=& \varepsilon_B + \varepsilon_e + \varepsilon_{\mu} \\
\varepsilon_B &=& -P_B + \sum_{i=n,p,u,d,s}n_i^*\mu_i \,.
\end{eqnarray}
Using these, the adiabatic speed of sound is obtained by first calculating the expression
\begin{equation}
c_{\rm ad}^2(n_{\rm B},y_i) = \left.\frac{\partial P}{\partial n_{\rm B}}\right|_{y_i}
\left(\left.\frac{\partial \varepsilon}{\partial n_{\rm B}}\right|_{y_i}\right)^{-1}
\end{equation}
and then evaluating it in $\beta$-equilibrium
\begin{equation}
c_{\rm{ad},\beta}^2(n_{\rm B}) = c_{\rm ad}^2[n_{\rm B},y_{i,\beta}(n_{\rm B})]~.
\end{equation}
The equilibrium sound speed is given by the total derivatives of the pressure and the energy density with respect to the baryon density after the enforcement of $\beta$-equilibrium,
\begin{equation}
c_{\rm eq}^2 = \frac{dP_{\beta}}{dn_{\rm B}}\left(\frac{d\varepsilon_{\beta}}{dn_{\rm B}}\right)^{-1}~.
\end{equation}
\subsection{Sound speeds with Gibbs construction}
As in the crossover matter case, all thermodynamic quantities are expressed in terms of functions of the total baryon density $n_{\rm B}$, and the individual particle fractions $y_n$,~$y_p$,~$y_e$,~$y_{\mu}$,~$y_u$,~$y_d$,~$y_s$ which are, at this point, independent variables. That is,
\begin{eqnarray}
\varepsilon_H &=& \varepsilon_H(n_{\rm B},y_n,y_p) ~;~ P_H = P_H(n_{\rm B},y_n,y_p) ~; \nonumber \\
\mu_h &=& \mu_h(n_{\rm B},y_n,y_p) \\ \nonumber \\
\varepsilon_Q &=& \varepsilon_Q(n_{\rm B},y_u,y_d,y_s) ~;~ P_Q = P_Q(n_{\rm B},y_u,y_d,y_s) ~; \nonumber \\
\mu_q &=& \mu_q(n_{\rm B},y_q) \\ \nonumber \\
\varepsilon_L &=& \varepsilon_L(n_{\rm B},y_e,y_{\mu}) ~;~ P_L = P_L(n_{\rm B},y_e,y_{\mu}) ~; \nonumber \\
\mu_l &=& \mu_l(n_{\rm B},y_l) \,.
\end{eqnarray}
The conditions for weak equilibrium, charge neutrality, and baryon number conservation are applied afterwards.
These introduce another independent variable, $\chi$, which is the volume fraction of quarks in the mixed phase of Gibbs construction:
\begin{eqnarray}
&&P_H = P_Q ~;~ \mu_n = 2\mu_d + \mu_u ~;~ \mu_p = 2\mu_u + \mu_d \\ \nonumber \\
&&\mu_n = \mu_p + \mu_e ~;~ \mu_e = \mu_\mu ~;~\mu_d = \mu_s \\ \nonumber \\
&&3 (1-\chi) y_p+\chi(2y_u-y_d-y_s)-3(y_e+y_{\mu}) = 0 \\ \nonumber \\
&&3 (1-\chi)(y_n+y_p)+\chi(y_u+y_d+y_s)-3 = 0 \,.
\end{eqnarray}
Solving these equations eliminates the $y_i$ and $\chi$ in favor of $n_{\rm B}$. Thus the state variables become functions of only $n_{\rm B}$ according to the rule
\[ Q(n_{\rm B},y_i,y_j,...) \rightarrow Q[n_{\rm B},y_i(n_{\rm B}),y_j(n_{\rm B}),...] = Q(n_{\rm B})~.\] \\
Then, the thermodynamics of the mixed $(^*)$ phase are:
\begin{eqnarray}
\varepsilon^* &=& (1-\chi)\varepsilon_H + \chi \varepsilon_Q + \varepsilon_L \\
P^* &=& P_H+ P_L = P_Q+ P_L \nonumber \\
&=& (1-\chi) P_H + \chi P_Q + P_L \\
\mu_h^* &=& \mu_h ~~;~~ \mu_q^* = \mu_q \\
y_h^* &=& (1-\chi)y_h ~~;~~ y_q^* = \chi y_q \,.
\label{qnoG}
\end{eqnarray}
Quantities corresponding to leptons are not affected by the ratio of the two baryonic components in the mixed phase.
\begin{table}[h]
\caption{Parameter sets for the EOSs used in this work. }
\begin{center}
\begin{tabular}{ccrrrc}
\hline
\hline
Model & Parameter & XOA & XOB & XOC & Units \\ \hline
& $a_0$ & -96.64 & -90.39 & -96.64 & MeV \\
& $b_0$ & 58.85 & 52.60 & 58.85 & MeV \\
ZL & $\gamma$ & 1.40 & 1.446 & 1.40 & \\
& $a_1$ & -26.06 & -232.78 & -28.15 & MeV \\
& $b_1$ & 7.34 & 212.46 & 7.83 & MeV \\
& $\gamma_1$ & 2.45 & 1.1 & 3.5 & \\
\hline
& $m_u$ & 5.0 & 5.0 & 5.0 & MeV \\
& $m_d$ & 7.0 & 7.0 & 7.0 & MeV \\
vMIT & $m_s$ & 150.0 & 150.0 & 150.0 & MeV \\
& $a$ & 0.20 & 0.23 & 0.15 & fm$^2$ \\
& $B^{1/4}$ & 180.0 & 180.0 & 180.0 & MeV \\
\hline
KW & $\mu_0$ & 1.8 & 1.8 & 1.8 & GeV \\
\hline \hline
\end{tabular}
\end{center}
\label{tab:Parameters}
\end{table}
The mixed phase extends over those densities $n_{\rm B}$ for which $0\le \chi(n_{\rm B}) \le 1$. In contrast to the crossover case where $S$ operates at \textit{all} densities, $\chi$ is active only when the above condition is satisfied.
Moreover,
in the case of a
Gibbs construction of the first-order phase transition scenario, $\chi$ and $y_i$ are treated on an equal footing with no prior assumptions regarding their density dependence, whereas in the crossover
scenario, $S$ has a definitive functional form which the particle fractions must be adjusted to fit.
As a consequence, even though
both $\chi$ and $S$
describe the quark-to-baryon fraction, the former is a quantity for which we solve while the latter acts as a constraint replacing the Gibbs condition for mechanical equilibrium.
The sound speeds are obtained following the prescription outlined in the previous subsection. Alternatively, the adiabatic sound speed in the mixed phase can be calculated from the corresponding ones in the pure hadronic and quark phases separately according to Ref.~\cite{Jaikumar:2021jbw}
\begin{equation}
\frac{1}{c_{\rm ad}^{*2}} = \frac{1-\chi}{c_{{\rm{ad}},H}^2} + \frac{\chi}{c_{{\rm{ad}},Q}^2}~.
\end{equation}
\section{Results}
\label{sec:Results}
We turn now to present results based on calculations of the crossover EOS, associated NS properties, the two sound speeds and the resulting $g$-mode~frequencies. For contrast, results corresponding to pure hadronic matter and those for a first-order phase transition treated using the Gibbs construction are also presented.
\subsection{EOS and structural properties of NSs}
To construct crossover models, we have chosen the parameter values shown in Table~\ref{tab:Parameters}, labeled as XOA, XOB and XOC, for the parametrization of the EOSs used in this work.~\footnote{The method to determine the constants for the ZL parametrization is described in Refs.~\cite{Zhao:2020dvu} and \cite{Jaikumar:2021jbw}.}
These sets of parameters correspond to the nuclear and neutron-star properties shown in Table~\ref{tab:Properties}.
\begin{table}[h]
\caption{Nuclear and neutron-star properties corresponding to the parametrizations shown in Table~\ref{tab:Parameters}.
The symbols refer to
$n_{\rm sat}$: nuclear saturation density,
$E_0$: energy per particle at $n_{\rm sat}$,
$K_0$: compression modulus of symmetric nuclear matter (SNM) at $n_{\rm sat}$,
$S_v$: symmetry energy at $n_{\rm sat}$,
$L$: slope of $S(n)$ at $n_{\rm sat}$,
$n_{\mu,\text{on}}^{\beta}$: onset density of muons,
$n_{\mu,\text{off}}^{\beta}$: turnoff density of muons, and $n_{p,\text{off}}^{\beta}$: turnoff density of protons.
Quantities related to neutron stars are
$R$: radius, $M$: mass, $\beta=GM/Rc^2$: compactness,
$n_c$: central density, $p_c$: central pressure,
$\varepsilon_c$: central energy density,
$\Lambda$: dimensionless tidal deformability,
$c_{\rm eq}^2$: squared equilibrium sound speed,
and $c_{\rm ad}^2$: squared adiabatic sound speed.
The subscripts $1.4$ and ${\rm max}$ denote the masses of stars in ${\rm M}_{\odot}$.}
\begin{center}
\begin{tabular}{crrrrrc}
\hline
\hline
Property & XOA & XOB & XOC & ZL & Gibbs & Units \\ \hline
$n_{\rm sat}$ & 0.16 & 0.16 & 0.16 & 0.16 & 0.16 & fm$^{-3}$ \\
$E_0$ & -16.0 & -16.0 & -16.0 & -16.0 & -16.0 & MeV \\
$K_0$ & 250.0 & 260.0 & 250.0 & 250.0 & 250.0 & MeV \\
$S_v$ & 31.6 & 30.0 & 30.0 & 31.6 & 31.6 & MeV \\
$L$ & 43.0 & 70.0 & 65.0 & 43.0 & 43.0 & MeV \\
$n_{\mu,\text{on}}^{\beta}$ & 0.13 & 0.15 & 0.14 & 0.13 & 0.13 & fm$^{-3}$ \\
$n_{\mu,\text{off}}^{\beta}$ & 1.39 & 0.77 & 1.67 & N/A & 1.32 &fm$^{-3}$ \\
$n_{p,\text{off}}^{\beta}$ & N/A & 0.88 & N/A & N/A & N/A &fm$^{-3}$ \\
\hline
$R_{1.4}$ & 12.4 & 12.4 & 13.8 & 12.4 & 12.4 &{\rm km} \\
$\beta_{1.4}$ & 0.167 & 0.166 & 0.150 & 0.166 & 0.166 & \\
$n_{c,1.4}/n_{\rm sat}$ & 2.64 & 3.16 & 1.96 & 2.64 & 2.68 & \\
$p_{c,1.4}$ & 60.2 & 73.6 & 41.7 & 60.0 & 69.8 & ${\rm MeV~fm^{-3}}$ \\
$\varepsilon_{c,1.4}$ & 424.7 & 518.4 & 316.7 & 424.9 & 436.3 & ${\rm MeV~fm^{-3}}$ \\
$\Lambda_{1.4}$ & 428.9 & 426.7 & 841.3 & 430.1 & 421.2 & \\
$(c_{\rm eq}^2)_{c,1.4}$ & 0.398 & 0.331 & 0.294 & 0.397 & 0.406 & $c^2$ \\
$(c_{\rm ad}^2)_{c,1.4}$ & 0.429 & 0.331 & 0.497 & 0.426 & 0.528 & $c^2$ \\
\hline
$\ensuremath{R_{\Mmax}}$ & 11.5 & 10.4 & 13.0 & 11.1 & 11.2 &{\rm km} \\
$\ensuremath{M_{\rm max}}$ & 2.11 & 2.04 & 2.13 & 2.23 & 2.08 &${\rm M}_{\odot}$ \\
$\beta_{\rm max}$ & 0.270 & 0.289 & 0.242 & 0.295 & 0.275 & \\
$n_{\rm c,max}/n_{\rm sat}$ & 5.83 & 7.38 & 4.70 & 6.14 & 6.32 & \\
$p_{\rm c,max}$ & 362.7 & 696.1 & 213.4 & 577.0 & 457.7 & ${\rm MeV~fm^{-3}}$ \\
$\varepsilon_{\rm c,max}$ & 1142.9 & 1549.5 & 886.8 & 1202.4 & 1264.9 & ${\rm MeV~fm^{-3}}$ \\
$\Lambda_{\rm max}$ & 13.8 & 6.4 & 30.1 & 6.2 & 11.1 & \\
$(c_{\rm eq}^2)_{\rm c,max}$ & 0.426 & 0.566 & 0.316 & 0.767 & 0.576 & $c^2$ \\
$(c_{\rm ad}^2)_{\rm c,max}$ & 0.507 & 0.818 & 0.353 & 0.889 & 0.653 & $c^2$ \\
\hline \hline
\end{tabular}
\end{center}
\label{tab:Properties}
\end{table}
For XOA, all values are within 1-$\sigma$ deviations of empirical/observational constraints discussed below. While XOB and XOC do not perform as well, they are used here to illustrate some important physics related to the behavior of the $g$-mode~frequency.
Specifically, with XOB we investigate $g$-mode~frequency features corresponding to a sound-speed peak due to proton disappearance, whereas in XOC the peak in the speed of sound is not related to a change in the number of degrees of freedom.
Models labeled ZL (nucleons only) and Gibbs (nucleons plus quarks with a Gibbs construction) use the appropriate parameters of XOA.
\vspace{3mm}
\begin{figure}[h!]
\includegraphics[scale=0.32]{plotMR.pdf}
\caption{Neutron-star $M$-$R$ curves for the various EOS models used in this work. The black, dashed lines represent the 90\%-confidence level constraints extracted from recent radio, x-ray, and gravitational-wave observations by Legred et al.~\cite{Legred:2021hdx}.
Models corresponding to the ``A'' parameter set fit these constraints well with differences between the three depending on the order of the transition to quark matter or the lack thereof.
While stars with $M\leq 1.8\,{\rm M}_{\odot}$ using model XOB are within the constraints, with model XOC only stars close to the maximum mass satisfy the Legred et al. constraints.
}
\label{fig:MR}
\end{figure}
\begin{figure*}[htb]
\parbox{0.5\hsize}{
\includegraphics[width=\hsize]{plotPvN.pdf}\\[-4ex]
}\parbox{0.5\hsize}{
\includegraphics[width=\hsize]{plotEvN2.pdf}\\[-4ex]
}
\caption{Left panel: pressure versus baryon density as obtained by the assorted EOSs used herein compared with the astrophysical constraints of Legred et al.~\cite{Legred:2021hdx} (in black, dashed lines).
All parametrizations meet these constraints successfully with the exception of XOC which fails to stay in the allowed region around $2.0\,n_{\rm sat}$, where $n_{\rm sat}=0.16~{\rm fm}^{-3}$.
Right panel: energy-per-particle versus baryon density of beta-equilibrium matter for the various EOSs used in this work compared with the 1-$\sigma$ and 2-$\sigma$ constraints
from Drischler et al.~\cite{Drischler:2020fvz} obtained in a chiral EFT framework. Results corresponding to Gibbs and ZL are not shown because, over the range of densities displayed, they are identical to XOA.
Only XOA remains within the 2-$\sigma$ constraints of \cite{Drischler:2020fvz} up to $\sim2.0\,n_{\rm sat}$.
}
\label{fig:PN_EN_nB}
\end{figure*}
We wish to note that the values of the symmetry energy $S_v$ and and its slope $L$ at $n_{\rm sat}$ used in our work (see Table~\ref{tab:Properties}) lie in the range $\simeq 31 \pm 2$ MeV and $\simeq 51\pm 11$ MeV, respectively, recommended in Ref.~\cite{Lattimer:2012xj}. These values led to the bounds on the radius of a $1.4\,{\rm M}_{\odot}$ star to be $R_{1.4} \simeq 12 \pm 1$ km.
\vspace{3mm}
Interpretations of the recent PREX-II experiment carried out at the Jefferson Lab (JLab) measuring the neutron skin thickness of $^{208}$Pb~\cite{Adhikari:2021phr}, $R_\mathrm{skin}^{^{208}\mathrm{Pb}} = 0.283 \pm 0.071$~fm, however, widely vary in their inferences of the appropriate values of $S_v$ and $L$ to be used.
For example, Reed et al.~\cite{Reed:2021nqk}, using relativistic mean-field theory (RMFT) calculations to analyze the JLab data, conclude that $S_v=38.1\pm 4.7$ MeV and $L=106\pm 37$ MeV, values that are significantly higher than those deduced in earlier works. Furthermore, the bound $R_{1.4} > 13.25$~km was found there.
Reinhard et al.~\cite{Reinhard:2021utv}, use covariant RMFT (with density-dependent couplings) and nonrelativistic energy functionals to analyze the PREX-II data and
combine it with the dipole polarizability data of $^{208}$Pb
to arrive at $S_v = 32\pm 1$ MeV and $L=54\pm 8$ MeV. In addition, these authors obtain $R_\mathrm{skin}^{^{208}\mathrm{Pb}} = 0.19\pm 0.02$~fm in accord with earlier deductions.
Similar results are obtained by Essick et al.~\cite{Essick:2021kjb} who report $S_v = 34\pm 3$ MeV, $L=58\pm 18$ MeV and $R_\mathrm{skin}^{^{208}\mathrm{Pb}} = 0.19~^{+0.03}_{-0.04} $~fm from a non-parametric EOS coupled with Gaussian processes.
Combining recent mass and radius measurement from radio and x-ray data from \textit{NICER}, Biswas~\cite{Biswas:2021yge} finds $R_\mathrm{skin}^{^{208}\mathrm{Pb}} = 0.20\pm 0.05$~fm and $R_{1.4} = 12.75~^{+0.42}_{-0.54}$ km using nuclear EOSs with piecewise polytrope parametrization.
Given the fluid state of theoretical inferences from the analysis of JLab data, we have opted to stick with the values used in Table~\ref{tab:Properties}.
\vspace{3mm}
Figure~\ref{fig:MR} shows mass versus radius ($M$-$R$) curves for all the models considered along with the
recent constraints obtained by Legred et al.~\cite{Legred:2021hdx}, which combined available observations including the radio mass measurements of PSR J0348+0432 and J0470+6620~\cite{Fonseca:2021wxt,Cromartie:2019kug,Antoniadis:2013pzd}, the mass and tidal deformability measurements of GW170817 and GW190425~\cite{Abbott:2018wiz,LIGO:2017qsa,Abbott:2020uma}, and the x-ray mass and radius constraints from latest \textit{NICER} measurements of J0030+0451 and J0470+6620~\cite{Miller:2019cac,Riley:2019yda,Miller:2021qha,Riley:2021pdl}.
The constraints of Legred et al. were obtained by using the hierarchical inference~\cite{Loredo:2004nn} and a nonparametric survey through Gaussian Processes (GPs) conditioned on existing EOS models in the literature~\cite{Landry:2020vaw,Landry:2018prl,Essick:2019ldf}.
\vspace{3mm}
The left panel of Fig.~\ref{fig:PN_EN_nB} displays results of the pressure versus baryon number density ($P-n_{\rm B}$) relation for the various models used in the present work, in contrast to those inferred from Ref.~\cite{Legred:2021hdx} mentioned above (the black dashed boundaries, adapted from their Fig.~4).
To provide a comparison, results of energy versus density $E$ vs $n_{\rm B}$ of $\beta$-equilibrated neutron-star matter (NSM) are shown in the right panel of Fig.~\ref{fig:PN_EN_nB},
together with those for NSM from the chiral effective theory calculations of Ref.~\cite{Drischler:2020fvz} where 1-$\sigma$ and 2-$\sigma$ error estimates up to $\sim2.0\,n_{\rm sat}$ were provided.
Although not shown, we also find that results of the crossover, ZL and Gibbs models for $P-n_{\rm B}$ and $E$ versus $n_{\rm B}$ are consistent with microscopic Greens' function calculations of Gandolfi et al.~\cite{Gandolfi:2011xu}.
\vspace{3mm}
The squared adiabatic and equilibrium sound speeds $c_{\rm ad}^2$ and $c_{\rm eq}^2$ versus baryon density $n_{\rm B}$ are shown in the left and right panels of Fig.~\ref{fig:cs_ad_eq}, respectively. Both $c_{\rm ad}^2$ and $c_{\rm eq}^2$ increase monotonically with $n_{\rm B}$ for the ZL model in which nucleons are the only baryons. The non-monotonic behaviors of the other curves are due to admixtures of nucleons and quarks in the baryon sector. The $c_{\rm eq}^2 (n_{\rm B})$ for the Gibbs model suddenly drops (rises) at the onset (end) of the mixed phase (the latter not shown in the figure), whereas $c_{\rm ad}^2 (n_{\rm B})$ varies smoothly. Results for the crossover models XOA and XOC are similar in structure, whereas those for XOB show more structure at large $n_{\rm B}$ due to the disappearance of protons. With the exception of model XOB, $c_{\rm ad}^2 > c_{\rm eq}^2$ for all other models at all $n_{\rm B}$.
\vspace{3mm}
\begin{figure*}[htb]
\parbox{0.5\hsize}{
\includegraphics[width=\hsize]{plotCAD.pdf}\\[-4ex]
}\parbox{0.5\hsize}{
\includegraphics[width=\hsize]{plotCEQ.pdf}\\[-4ex]
}
\caption{Left panel: the squared adiabatic sound speed $c_{\rm ad}^2$ as a function of the baryon density $n_{\rm B}$.
Right panel: the squared equilibrium sound speed $c_{\rm eq}^2$ as a function of the baryon density $n_{\rm B}$.
}
\label{fig:cs_ad_eq}
\end{figure*}
\begin{figure}[h!]
\includegraphics[scale=0.38]{plot1overC2.pdf}
\caption{Difference between the inverse-squared sound speeds versus the baryon number density.
The peaks in the vicinity of $n_{\rm sat}$ correspond to muon appearance and are present in all models.
The peak at $\sim3\,n_{\rm sat}$ for Gibbs occurs at the onset of the mixed phase which extends beyond the densities shown here.
The peak around $5.5\,n_{\rm sat}$ for XOB is the combined effect of muon and proton disappearance in this model.
On the other hand, the peak at $\sim2.5\,n_{\rm sat}$ for XOC results from inflection points in the quark and neutron fractions.}
\label{fig:OOC}
\end{figure}
\begin{figure}[h!]
\includegraphics[scale=0.32]{plotyiA.pdf}
\caption{Particle fractions of the crossover model XOA versus baryon density. Quarks are still present below 0.3 fm$^{-3}$ but at two orders of magnitude less than what is shown here.}
\label{fig:YIA}
\end{figure}
\begin{figure}[h!]
\includegraphics[scale=0.32]{plotyiB.pdf}
\caption{Particle fractions of model XOB versus baryon density. Muons drop out of the system at 0.77 fm$^{-3}$ and protons at 0.88 fm$^{-3}$ leading to a sharp peak in the adiabatic sound speed (cf. Fig.~\ref{fig:OOC}). Electrons are still present above 0.85 fm$^{-3}$ but at two orders of magnitude less than what is shown here.}
\label{fig:YIB}
\end{figure}
\begin{figure}[h!]
\includegraphics[scale=0.32]{plotyiC.pdf}
\caption{Particle fractions of model XOC versus baryon density. The inflectionary behavior of the quark and neutron fractions around 0.4 fm$^{-3}$ is responsible for the broad peak behavior of XOC occurring in Fig.~\ref{fig:OOC}.}
\label{fig:YIC}
\end{figure}
Figure~\ref{fig:OOC} shows the difference of the inverses of the adiabatic and equilibrium sound speeds, $\Delta(c^{-2})\equiv 1/c_{\rm eq}^2-1/c_{\rm ad}^2$ as a function of the baryon density. This quantity is particularly important in the context of $g$-modes~because it enters directly in the calculation of the Brunt--V\"ais\"al\"a~frequency (discussed in more detail in the next section).
\vspace{3mm}
A comparison between this figure and Figs.~\ref{fig:YIA}--\ref{fig:YIC} which show the particle fractions corresponding to the three crossover models used in this work, reveals a direct correlation between sharp maxima in the former and particle appearance/disappearance in the latter.
Smooth maxima, such as those exhibited by XOA around 0.7 fm$^{-3}$ and XOC around 0.4 fm$^{-3}$, reflect non-monotonic behaviors in the slopes of the quark and the neutron fractions. Also worth noting is that, in the present framework and with the chosen parametrizations, quarks are never the dominant contributors to the baryon density for densities relevant to neutron stars.
\vspace{3mm}
\begin{table}[h]
\caption{Various representations of the quark content of $2\,{\rm M}_{\odot}$ NSs corresponding to the Gibbs and the three crossover models in this paper. The symbols are
$M_{\rm B}$: the total baryonic mass,
$Y_Q^{\rm{bar}}$: contribution of quarks to the total baryon number/total baryon number,
$Y_Q^{\rm{part}}$: quark particle number/total particle number,
$Y_Q^{\rm{nuc}}$: quark particle number/nucleon number,
${M_Q^{\rm{B}}}/{M_{\rm B}}$: contribution of quarks to the total baryonic mass/total baryonic mass, and
${M_Q^{\rm{G}}}/{M_{\rm G}}$: contribution of quarks to the total gravitational mass/total gravitational mass.
}
\begin{center}
\begin{tabular}{ccccccc}
\hline
\hline
Model & $M_{\rm B}$ & $Y_Q^{\rm{bar}}$ & $Y_Q^{\rm{part}}$ & $Y_Q^{\rm{nuc}}$ & ${M_Q^{\rm{B}}}/{M_{\rm B}}$ & ${M_Q^{\rm{G}}}/{M_{\rm G}}$ \\
& $({\rm M}_{\odot})$ & ($\times 10^{-2}$) & ($\times 10^{-2}$) &
($\times 10^{-2}$) & ($\times 10^{-2}$) & ($\times 10^{-2}$) \\ \hline
XOA & 2.31 & 0.35 & 1.03 & 1.04 & 0.35 & 0.50 \\
XOB & 2.33 & 2.15 & 6.18 & 6.58 & 2.15 & 3.04 \\
XOC & 2.29 & 0.06 & 0.19 & 0.19 & 0.06 & 0.09 \\
Gibbs & 2.35 & 1.91 & 5.53 & 5.85 & 1.91 & 2.89 \\
\hline \hline
\end{tabular}
\end{center}
\label{tab:quarkfrac}
\end{table}
Consequently, the contributions of quarks to the total baryon number as well as the total mass of the star (both baryonic and gravitational) are rather small in the models considered, as shown in Table~\ref{tab:quarkfrac} for stars with gravitational mass $M=2\,{\rm M}_{\odot}$. These are straightforwardly calculated as follows: The solution of the TOV equations gives, among other things, the baryon density as a function of the star's radius, $n_{\rm B}(r)$. The quark particle fractions as functions of the baryon density are obtained from the $\beta$-equilibrated equivalents of Eqs. (\ref{qno}) and (\ref{qnoG}) for the KW and Gibbs cases, respectively. Explicitly, $y_q (r) = y_{q,\beta}^*[n_{\rm B}(r)]$. For the quark baryon fraction at each radius we divide this by 3. The quark particle densities are then $n_q(r) = y_q(r)n_{\rm B}(r)$. The baryon number due to quarks of all species $q=u,d,s$ in the star is given by the integral
\begin{equation}
N_Q = 4\pi \int_0^R dr r^2 \frac{\sum_q n_q(r)/3}{[1-2GM(r)/r]^{1/2}}.
\end{equation}
Here, $M(r)$ is the total gravitational mass of the star as a function of its radius $r$, also given by the solution of the TOV equations. Therefore, the amount of baryonic mass in the star provided by quarks is $M_Q^{\rm{B}} = m_H N_Q$~\cite{glendenning2012compact}. For the quark gravitational mass we begin by calculating the quark energy density as a function of the radius according to $\varepsilon_Q(r) = \varepsilon_{Q,\beta}^*[n_{\rm B}(r)]$ where $\varepsilon_Q^* = -S P_Q + \sum_q n_q^*\mu_q$. Afterwards, we perform the integral $M_Q^{\rm{G}} = 4\pi \int_0^R dr r^2 \varepsilon_Q(r)$.
\vspace{3mm}
Specifically, in model XOA, quarks contribute less than 0.5\% of the total baryon number, about 1\% of particles in its $2\,{\rm M}_{\odot}$ stars are quarks and they are responsible for around 0.35\% (0.5\%) of the total baryonic (gravitational) mass of the star. For XOB and Gibbs models, quarks contribute about $2\%$ of the total baryon number, over 5\% of the total particle number, and 2\% (3\%) of the total baryonic (gravitational) mass of the respective $2\,{\rm M}_{\odot}$ stars. In model XOC, quarks are, for all intents and purposes, irrelevant.
As noted before, such a clean separation may not be possible in treatments that intermingle hadron/nucleon and quark interactions.
\subsection{Sound speeds and the Brunt--V\"ais\"al\"a~frequency}
Having outlined the thermodynamics of the multi-component system and the specific models employed in our study of the smooth crossover transition in neutron stars, we turn now to the calculation of the star's $g$-mode~frequencies.
Our main goal is to compare the behavior of the $g$-mode~frequencies in the crossover model with those in the Gibbs mixed phase.
The $g$-mode~frequencies ($\nu_g$=$\omega/(2\pi)$) and normalized amplitudes for the radial and tangential parts of the fluid perturbation ($\xi_r$ and $\xi_h$ respectively) are estimated within the relativistic Cowling approximation (see below) by computing numerical solutions to the following equations of motion for fluid variables $U,V$~\cite{Jaikumar:2021jbw}
\begin{eqnarray}
\label{eq:uv}
\frac{dU}{dr}&=&\frac{g}{c_{\rm ad}^2}U+{\rm e}^{\lambda/2}\left[\frac{l(l+1){\rm e}^{\nu}}{\omega^2}-\frac{r^2}{c_{\rm ad}^2}\right]V \nonumber \\
\frac{dV}{dr}&=&{\rm e}^{\lambda/2-\nu}\frac{\omega^2-N^2}{r^2}U+g\Delta(c^{-2})V \,,
\end{eqnarray}
which are simplified forms of the original perturbation equations~\cite{McD83,RG92,Kantor:2014lja}.
In \Eqn{eq:uv}, $U$ = $r^2{\rm e}^{\lambda/2}\,\xi_r$, $V$ = $\omega^2 r\, \xi_h$ = $\delta P/(\varepsilon+P)$, $\Delta(c^{-2})=c_{\rm eq}^{-2}-c_{\rm ad}^{-2}$, $\lambda$ and $\nu$ are metric functions. The scale of the mode frequency is set by the Brunt--V\"ais\"al\"a~frequency
\begin{eqnarray}
N^2 = g^{2}\Delta(c^{-2}){\rm e}^{\nu-\lambda},
\end{eqnarray}
where $g=-\nabla P/(\varepsilon+P)$.
The relativistic Cowling approximation neglects the back reaction of the gravitational potential by excluding metric perturbations that must accompany matter perturbations in a general relativistic treatment~\cite{Thorne:1967a,Thorne:1967b,Lindblom:1983,Detweiler:1985,Finn:1987,Andersson:1995wu}.
It reduces the number and complexity of the equations we have to solve, while providing results for $g$-mode~frequencies that are accurate at the few \% level~\cite{gregorian2015nonradial}.
Details on the solution methods for~\Eqn{eq:uv} and relevant boundary conditions are provided in Ref.~\cite{Jaikumar:2021jbw}.
\vspace{3mm}
\begin{figure}[h!]
\includegraphics[scale=0.32]{plotBV.pdf}\\[-3.0ex]
\caption{The Brunt--V\"ais\"al\"a~frequency in a hybrid star of mass $2.0\,{\rm M}_{\odot}$ for the Gibbs and crossover models.
The Brunt--V\"ais\"al\"a~frequency for the ``good'' crossover model XOA is very similar to the nucleonic ZL EOS (which includes muons), whereas the Gibbs model shows a distinct peak corresponding to the rapid onset of quark matter. Parameters for the nuclear and quark EOSs are as in Table~\ref{tab:Parameters}.
}
\label{fig:BV-EOS}
\end{figure}
Figure~\ref{fig:BV-EOS} is a comparison of the Brunt--V\"ais\"al\"a~frequency in the three models considered in this work. The crossover model, where quarks are always present in the core EOS, albeit in minuscule fractions, resembles the purely nucleonic ZL EOS in this respect, while the sudden onset of quarks in the Gibbs model is clearly imprinted in the form of a sharp peak.
Quarks enter at a density $n_{\rm B} \simeq$ 0.514 fm$^{-3}$ corresponding to $r/R$ = 0.473 ($r=0$ at the center) in the Gibbs model. As a consequence of the difference of sound speeds being negative in distinct density regimes for XOB (Fig.~\ref{fig:OOC}), the corresponding Brunt--V\"ais\"al\"a~frequency is imaginary, implying an instability to convection.\footnote{This feature could be an artifact of the Cowling approximation, but will likely be absent in the solution of the full general relativistic treatment~\cite{Thorne:1967a,Thorne:1967b,Lindblom:1983,Detweiler:1985,Finn:1987,Andersson:1995wu}
of the $g$-mode~in which $c_{\rm eq}^2$ does not enter explicitly.}
However, convection is absent at zero temperature; therefore these regions are unphysical and can play no role in the global $g$-mode~spectrum. Accordingly, XOB is omitted from Figs.~\ref{fig:gmode-EOS}, \ref{fig:eigenfunctions} and \ref{fig:energy}.
\begin{figure}[h!]
\includegraphics[scale=0.32]{plotGMnoB.pdf}\\[-3.0ex]
\caption{The $g$-mode~frequency as a function of the stellar mass in the Gibbs, crossover, and ZL models.
Parameters for the nuclear and quark EOSs are as in Table.~\ref{tab:Parameters}.
The g-modes corresponding to XOB are unstable and therefore, this model, is excluded from the present and the next two figures.}
\label{fig:gmode-EOS}
\end{figure}
\begin{figure}[]
\includegraphics[scale=0.28]{plotXI.pdf}\\[-3.0ex]
\caption{Amplitudes of the radial ($r$) and transverse ($h$) components of the $g$-mode~displacement (eigenfunctions) as a function of distance from the center for the ZL, XOA, XOC and Gibbs EOSs for a $\sim2\,{\rm M}_{\odot}$ star. The order of the mode (g1 is fundamental, g2 is overtone) is indicated in the legend in the panels.
}
\label{fig:eigenfunctions}
\end{figure}
Figure~\ref{fig:gmode-EOS} compares the $g$-mode~frequency, $\nu_g=\omega/(2\pi)$, for the crossover and Gibbs models. While XOA and XOC are very similar to the nucleonic ZL EOS (which includes muons), the Gibbs model shows a distinctly rising spectrum corresponding to the rapid onset of quark matter. These findings are consistent with the result for the Brunt--V\"ais\"al\"a~frequency in Fig.~\ref{fig:BV-EOS} and the conclusions in Paper I.
In the crossover model XOB, protons disappear above some critical density. In contrast to the smooth behavior of $g$-mode~frequencies in XOA and XOC, the sudden disappearance of protons in XOB produces a sharp rise in the spectrum akin to the Gibbs case but renders $g$-modes~to become unstable (not shown in Fig.~\ref{fig:gmode-EOS}), confirming that dramatic changes in the $g$-mode~frequency require {\it appearance or disappearance of a (strongly interacting) particle species, not merely a smooth change in composition}.
This is why, except for extreme parameter choices, crossover models will not show the $g$-mode~feature resulting from the presence of quarks that Gibbs models do.
The panels in Fig.~\ref{fig:eigenfunctions} show the comparison of the core $g$-mode~amplitude between the three chosen models (ZL, crossover and Gibbs). The radial component $\xi_r$ of the fundamental mode (labeled ``g1'' in the panels) has no nodes in the core, while the radial part of the first overtone (labeled ``g2'') has one, as expected. The horizontal component $\xi_h$ has one more node than the corresponding radial component of the same order.
The larger amplitude of $\xi_h$ relative to $\xi_r$ indicates that the $g$-mode~is dominated by transverse motion of the perturbed fluid. While there is little difference between the ZL and crossover models in the profile of these eigenfunctions, the Gibbs case is markedly different. Its amplitude relative to the other two is larger, and it changes abruptly upon the onset of quark matter in the core.
\begin{figure}[h!]
\includegraphics[scale=0.32]{plotDEDR.pdf}\\[-3.0ex]
\caption{Energy/unit distance of the fundamental $g$-mode~as a function of distance from the center for the ZL, XOA, XOC, and Gibbs EOS for a $\sim2\,{\rm M}_{\odot}$ star. Note that the mode energies in the ZL and crossover cases are scaled up by a factor of 10 for the purpose of comparison.}
\label{fig:energy}
\end{figure}
The energy per unit radial distance $dE_T/dr$ contained in the oscillatory motion corresponding to a frequency $\omega$ is given in terms of the amplitude as~\cite{McD83}
\begin{equation}
\frac{dE_T}{dr} =\frac{\omega^2\,r^2}{2}(\varepsilon+P){\rm e}^{(\lambda-\nu)/2}\,\left[\xi_r^2{\rm e}^{\lambda}+l(l+1)\xi_h^2\right] \,.
\end{equation}
Figure~\ref{fig:energy} shows the comparison of the core $g$-mode~energy/unit distance for the three chosen models (ZL, crossover and Gibbs) for a $\sim2\,{\rm M}_{\odot}$ star. The typical scale of the energy/unit distance deep in the core is approximately $10^{50}\, \rm{ergs/km}$ for the ZL and crossover models, while it is of order $10^{51}\, \rm{ergs/km}$ for the Gibbs case.
While the profiles are similar for the first two, the mode energy in the Gibbs case is overwhelmingly larger in the core, once quark matter appears. Thus, both the frequency and the amplitude (and hence the energy) of the core $g$-mode~is strongly amplified in quark matter in comparison to nucleonic matter or weaker forms of the phase transition. This can have bearing on the gravitational wave detection of $g$-modes~excited in neutron star mergers as discussed below.
\subsection{Discussion}
\label{sec.Discs}
The results in the previous subsection raise some points that are noteworthy. In particular, they highlight the relation between the behaviors of the speeds of sound and of the particle concentrations, and the spectrum of the $g$-mode~signal. Thus, the latter becomes a diagnostic which distinguishes nucleonic and hybrid matter \textit{as well as} the Gibbs and crossover transitions.
Peaks in the sound-speed difference $\delta c \equiv c_{\rm ad}^2 - c_{\rm eq}^2$, and the difference of the inverses $\Delta (c^{-2}) \equiv 1/c_{\rm eq}^2 - 1/c_{\rm ad}^2$ occur when particles appear or disappear [i.e. when the number of degrees of freedom (DOF) of the system changes] and when the particle concentrations $y_i(n_{\rm B})$ are not monotonic or have inflection points (i.e., their first or second derivatives with respect to $n_{\rm B}$ change sign).
The introduction or removal of a particle species from the system occurs when the relevant chemical potential either exceeds or falls below its rest mass threshold while maintaining charge neutrality.
Whereas the functional form of the $y_i$'s depends on the parametrization of the EOS, a change in the number of DOFs also leads to non-monotonic or inflectional behavior to the concentrations of particles already present in the system. Thus, in some sense, specific choices of the EOS parameters can mimic aspects of the emergence of new particles that are relevant to the sound speeds.
Signals of $g$-mode~with a characteristic fast rise in its frequency (such as
those corresponding to Gibbs shown in Fig.~\ref{fig:gmode-EOS} and XOB for which the $g$-mode~becomes unstable)
can occur only when particle species enter or leave the system. In such cases, the peaks in $\Delta (c^{-2})$ are sharp and asymmetric: vertical rise and quasi-lorentzian decay for appearance (see Fig.~\ref{fig:OOC}, Gibbs), and quasi-lorentzian rise and vertical drop for disappearance (XOB); whereas peaks due to parametrizations resemble symmetric Gaussians (XOC). Therefore, this kind of signal cannot be produced by quarks in matter with a smooth crossover because quarks are always present: their concentrations are vanishingly small at lower densities but never identically zero.
It is, however, possible to parametrize the hadronic EOS such that, in $\beta$-equilibrated crossover matter, protons exit the system. The magnitude of the peak of $\Delta (c^{-2})$ \textit{appears} to be proportional to the number of remaining DOFs (smallest in PNM; largest in $n$-$uds$-$e$ matter). This produces $g$-mode~frequency spectra similar to those found in matter with a Gibbs construction. So, although quarks are not directly responsible for this effect, they do serve the purpose of amplifying it.
For the specific case of crossover model XOB, a rather extreme parametrization ($K=260$ MeV, $S_v=30$ MeV, $L=70$ MeV, $\gamma_1=1.1$) was required for the proton disappearance while also meeting neutron-star constraints and having a peak at low-enough densities to be relevant. Consequently, the tentative conclusion is that hyperons or a first-order transition into quark matter through Gibbs construction are more likely to cause a distinctive peak in the Brunt--V\"ais\"al\"a~frequency than a crossover transition; nevertheless, the latter remains a viable, if improbable, option.
It is pertinent to mention that we have not performed calculations for first-order transitions with a Maxwell construction, which assume a sufficiently large surface tension between the pure hadronic and quark phases in bulk separated by a sharp boundary.
It has been shown that in such cases, $g$-modes~always vanish when the perturbed fluid element adjusts instantaneously (i.e. a very rapid conversion) to maintain thermodynamic equilibrium, but can arise if the microscopic phase conversion at the interface is slow enough, leading to distinctive features on the extended hybrid branch of NSs~\cite{Pereira:2017rmp,Tonetto:2020bie}. Such $g$-modes~associated with a discontinuity in density (``\textit{discontinuity} $g$-modes'', different from the ``\textit{compositional} $g$-modes'' that we consider)
has been widely studied in the literature; see e.g., extensive discussions in Refs.~\cite{Sotani:2001bb,Flores:2013yqa,Ranea-Sandoval:2018bgu,Rodriguez:2020fhf,Lau:2020bfq}.
\section{Summary and Conclusions}
\label{sec:Concs}
The main objectives of this work were to examine $g$-mode~frequencies in a smooth crossover scenario of the hadron-to-quark transition, and to compare them with those of a first-order transition treated using Gibbs and Maxwell constructions in Paper I~\cite{Jaikumar:2021jbw}.
For the crossover model, we chose the recent approach adopted by Kapusta and Welle~\cite{Kapusta:2021ney}, who constructed an EOS for PNM that resembled the smooth crossover observed in lattice calculations at finite temperatures.
We have generalized their approach to $\beta$-equilibrated NSM so that comparisons with the results of Paper I could be made. To describe nucleons, we used the ZL parametrization~\cite{Zhao:2020dvu} which reproduces near-saturation laboratory data as well as results of chiral EFT calculations~\cite{Drischler:2020hwi,Drischler:2020fvz} up to $\sim$ $2.0\,n_{\rm sat}$.
For quark matter we used the vMIT model~\cite{Gomes:2018eiv} with repulsive interactions. Results of our crossover EOSs tally with observational findings of the radii of $\sim\,1.4$ and $\sim\,2.0\,{\rm M}_{\odot}$ NSs~\cite{Miller:2019cac,Riley:2019yda,Miller:2021qha,Riley:2021pdl}.
Calculations of the equilibrium and adiabatic speeds of sound were performed following the procedures developed in Paper I. Our work here is focused on zero temperature and does not consider superfluidity in either nucleons or quarks. Inclusion of these effects will be taken up in a future work.
The results of the amplitudes of the $g$-mode~frequencies and their associated amplitudes of the gravitational energy radiated for the chosen hadron-to-quark crossover models lie between those of the first-order phase transitions that employ Maxwell and Gibbs constructions.
For the case of the Maxwell construction, the transition region is devoid of $g$-mode~frequencies as the equilibrium and adiabatic sound speeds both vanish. Consequently, $g$-mode~oscillations are permitted only for the pure nucleonic and quark phases. Their amplitudes in these regions are, however, rather small.
In contrast, the mixed phase in the case of the Gibbs construction yields amplitudes of $g$-mode~frequencies and the energy radiated that are significantly larger than those of the crossover model.
We note that $g$-mode~frequencies can be exceptionally large in the presence of superfluidity ($\approx$ 750 Hz for a hyperonic star~\cite{Dommes:2015wul} and $\approx$ 450 Hz for a nucleonic star~\cite{Gusakov:2013eoa,Passamonti:2015oia}), similar to the results for the Gibbs mixed phase. However, the reason for the enhancement in the two cases is different. In the case of superfluidity, the temperature-dependence of the entrainment terms serves to increase the $g$-mode~frequency at typical neutron star temperatures, while in the Gibbs case, the enhancement is purely composition-dependent.
A study of resonant excitations of such superfluid modes in coalescing neutron star binaries~\cite{Yu:2016ltf} suggests that the amplitude of these modes is weaker by a factor of 20 or so, compared to modes from the normal fluid. Note that the $g$-modes~we discuss here are also higher in frequency compared to the low-frequency $g$-modes~($\sim 50$ Hz) that might strain the neutron star crust to breaking point and lead to precursor flares in gamma-ray bursts~\cite{Kuan:2021jmk}.
The relatively larger frequency, amplitude and energy of the $g$-mode~in the Gibbs case inferred from
Figs.~\ref{fig:gmode-EOS}-\ref{fig:energy}
have observational implications for gravitational waves from neutron star mergers. It has been established that $g$-modes~can couple to tidal forces and draw energy and angular momentum from the binary to the neutron star, leading to an accelerated merger and a concomitant phase shift in the gravitational waveform~\cite{Lai:1993di}. This coupling will be largest for the Gibbs case, with its higher energy at resonance, and also because higher resonance frequencies are excited later in the inspiral,
when tidal forces are strongest. Estimates of the resulting phase shift were presented in Paper I (Eq. (89)) and found to be comparable to that from $g$-modes~in ordinary neutron stars (due to longer merger times) within uncertainties arising from the value of the tidal coupling. As these uncertainties are reduced through improved theoretical calculations, the case of a hybrid star may be distinguished from an ordinary neutron star.
We also infer from our results that were high-frequency $g$-modes~to be detected in upgraded LIGO and Virgo observatories, it would indicate a first-order phase transition akin to a Gibbs construction. In light of data from GW170817, lower bounds on the excitation of non-radial oscillations in binary mergers~\cite{Pratten:2019sed} affirm that a third generation network with its improved sensitivity and larger bandwidth can shed new light on the composition of the neutron star core.
\vspace{4mm}
\begin{acknowledgments}
C.C. acknowledges support from the European Union's Horizon 2020 research and innovation programme under the Marie Sk\l{}odowska-Curie grant agreement No. 754496 (H2020-MSCA-COFUND-2016 FELLINI).
S.H. is supported by the National Science Foundation, Grant PHY-1630782, and the Heising-Simons Foundation, Grant 2017-228.
P.J. is supported by the U.S. National Science Foundation Grant No. PHY-1913693.
M.P.'s research was supported by the Department of Energy, Grant No. DE-FG02-93ER40756.
\end{acknowledgments}
\section{Introduction}
\label{sec:intro}
Dense matter inside neutron stars (NSs) could contain unbound quarks that retain some vestige of the forces described by the fundamental theory of the strong interaction, Quantum Chromodynamics (QCD).
Lacking exact methods for its solution, numerous models have been constructed to investigate the hadron-to-quark transition in the region of high baryon density and zero temperature. On the other hand, the phase diagram of QCD at low density (zero or small baryon chemical potential) and high temperature is amenable to precision numerical studies which clearly point to a crossover with no clear phase boundary between the hadron resonance gas and the quark-gluon plasma~\cite{Borsanyi:2010cj}.
Recently, Kapusta and Welle~\cite{Kapusta:2021ney} (KW hereafter)
have proposed a crossover model for the hadron-to-quark transition in NSs to mimic the crossover feature of finite temperature lattice studies. The key trait of this model is an analytic mixing or switching function that accounts for the partial pressure of each component as a function of a single parameter - the baryon chemical potential. They found that NSs as massive as $\sim2.2\,{\rm M}_{\odot}$ could be supported by their crossover equation of state (EOS).
As hadrons/nucleons and quarks both appear explicitly as separate degrees of freedom in the KW description, it is straightforward to keep track of their individual contributions to the total pressure. KW report that within their model between $1-10\%$ of the total pressure could be contributed by quark matter in the core.
In treatments in which hadron and quark interactions are intermingled, the individual contributions from hadrons and quarks to the total pressure may not be possible to disentangle.
Given that EOSs with first-order phase transitions treated using the Maxwell construction face some challenges in obtaining the stable $\sim 2\,{\rm M}_{\odot}$ NSs that have been observed, it is worthwhile to investigate alternatives such as the crossover model for hybrid stars.
Our objective in this paper is two-fold: first, we extend the pure neutron matter (PNM) model of KW~\cite{Kapusta:2021ney}
to include $\beta$-equilibrium, as well as a crust for the hybrid star. We also include vector interactions among quarks. These modifications enable a direct comparison to other approaches in the literature~\cite{Han:2019bub}. Second, we investigate $g$-mode~oscillations, a potentially observable signature of the hadron-to-quark transition.
A $g$-mode~is a specific fluid oscillation where a parcel of fluid is displaced against the background of a stratified environment inside a neutron star. While pressure equilibrium is rapidly restored via sound waves, chemical equilibrium can take longer, causing buoyancy forces to oppose the displacement. Since cold NSs are not convective, the opposing force sets up stable oscillations, with a typical frequency, called the (local) Brunt--V\"ais\"al\"a~frequency. The kind of core $g$-modes~we study here were introduced in~\cite{RG92,RG94,Lai:1993di,Lai:1998yc} and in a recent work, which we shall refer to as Paper I~\cite{Jaikumar:2021jbw}, we showed that the $g$-mode~frequency rises steeply with the onset of quarks due to a rapid change in the equilibrium sound speed (see also Ref.~\cite{Wei:2018tts}). Since $g$-mode~oscillations couple to tidal forces, they may be excited during the merger of two NSs and provide information on the interior composition, specifically quarks here.
In Paper I, we chose a Gibbs construction for the mixed phase, which yields NS properties that are more compatible with astrophysical constraints than a Maxwell construction.\footnote{See Table VII in Ref.~\cite{Han:2019bub} for a comprehensive comparison of Gibbs, Maxwell and certain crossover models.} In this work, we present the systematics of the $g$-mode~frequency for hybrid stars in a crossover scenario, adopting the generalized KW model as a representative of this class.
It is worth noting that Maxwell-constructed first-order phase transitions cannot generate $g$-modes~in the transition region, because the equilibrium and adiabatic sound speeds both become zero (due to frozen pressure and composition over the phase coexistence region) there.
In the first phase of this work, we will restrict ourselves to zero temperature without the effects of superfluidity in both nucleons and quarks.
To the best of our knowledge, this is the first study of $g$-mode~for hybrid stars in a crossover model. References to earlier work in which hadronic EOSs with and without superfluidity were used to investigate $g$-mode~frequencies can be found in Paper I (see Refs. [21]-[35] therein).
Models akin to the crossover model of KW, but with some differences, have been considered earlier in the literature. Examples include the smooth crossover model of Ref.~\cite{Dexheimer:2014pea}, interpolated EOSs considered in Refs.~\cite{Baym:2017whm,Masuda:2012ed,Fukushima:2015bda,Kojo:2014rca}, and quarkyonic models of Refs.~\cite{McLerran:2018hbz,Zhao:2020dvu,Jeong:2019lhv,Sen:2020qcd}, et cetera.
In the chiral model of Refs.~\cite{Papazoglou:1998vr,Dexheimer:2008ax,Dexheimer:2009hi}, a scalar field $\Phi$, acting as an order parameter, is responsible for the deconfinement phase transition which can be first-order or crossover depending upon temperature and baryon chemical potential. Depending on the specific EOS models used in the hadronic and quark phases, chemical potential and pressure equilibrium between the two phases - of either the Maxwell or the Gibbs sort - may not be realized~\cite{Baym:2017whm}. In such cases, several interpolation procedures have been adopted to connect the two phases on the premise that at high supra-nuclear densities, a purely hadronic phase is untenable.
In quarkyonic models with a momentum shell structure~\cite{McLerran:2018hbz,Zhao:2020dvu},
quarks emerge at relatively low (but still
supra-saturation) densities but remain bound by strong interactions below the Fermi surface: while hadrons and quarks are separated in momentum space, they coexist in configuration space.
A key parameter that enters the calculation of the $g$-mode, is
the squared adiabatic speed of sound $c_{\rm ad}^2 =(\partial P/\partial \varepsilon)|_{y_i,\beta}$, where $P$ and $\varepsilon$ are the total pressure and energy density, respectively, and $y_{i,\beta}$ refer to the partial fractions of each component in beta-equilibrium.\footnote{$c_{\rm ad}^2$ is different from the squared equilibrium speed of sound $c_{\rm eq}^2 = dP/d\varepsilon$ commonly defined as $c_s^2$ in the literature for static EOS models.}
We find that in
the quarkyonic ``shell'' models of Refs.~\cite{McLerran:2018hbz,Zhao:2020dvu}, $c_{\rm ad}^2$ becomes discontinuous
with respect to density at the shell boundary.
To address $g$-modes~in this specific category of quarkyonic shell models, and to
extend the study of NS oscillations to span the various ways in which quarks can affect the dense matter EOS and the properties of NSs, such discontinuities must be smoothed.
As such a task is outside the scope of the present work, we do not consider these quarkyonic models in this paper.
The organization of this paper is as follows. In Sec.~\ref{sec:EOS}, we describe the EOSs for pure hadronic matter, pure quark matter and leptons used in the construction of the
KW crossover model of Ref.~\cite{Kapusta:2021ney}
by extending it to include $\beta$-equilibrium and interactions between quarks. The rationale for our parameter choices and the basic features of this model are also highlighted here for orientation.
Sec.~\ref{sec:multi_compt}
reviews
the thermodynamics of a multicomponent system as pertinent to the KW description of crossover matter generalized here to describe neutron-star matter (NSM).
In Sec.~\ref{sec:KWG}, the KW model formulation of crossover matter is described followed by the procedure to render the unconstrained system\footnote{The unconstrained system here refers to matter in which baryon number conservation, charge neutrality and $\beta$-equilibrium are not imposed.} in $\beta$-equilibrium. This section also contains
a comparison of the extended KW model with those of selected quarkyonic shell models.
In Sec.~\ref{sec:soundspeed}, the calculation of the equilibrium and adiabatic sound speeds in the crossover and Gibbs approaches are outlined.
In Sec.~\ref{sec:Results}, we present results for the chosen EOSs and their associated NS structural properties as well as the two speeds of sound in the crossover model, and discuss emergent differences from other models that include a phase transition. This section also contains results for the sound speed difference, the Brunt--V\"ais\"al\"a~frequency, and the $g$-mode~frequency in hybrid stars along with their interpretation. Our conclusions and outlook are in Sec.~\ref{sec:Concs}.
\section{Equation of State}
\label{sec:EOS}
In the KW description of the crossover transition, EOSs in the pure hadronic and quark phases are combined using a mixing or switch function that depends on the baryon chemical potential (to be described below).
We therefore begin by discussing EOS models that we use in each of these two sectors.
As we extend the KW model to include leptons, the EOS in the leptonic sector is also provided. These EOSs are first set forth without reference to baryon number conservation, charge neutrality, as well as chemical equilibrium (i.e., unconstrained), which are imposed at the appropriate junctures.
\subsection{Pure hadronic matter}
To describe nucleons, we use the Zhao-Lattimer (ZL) \cite{Zhao:2020dvu} parametrization of the EOS of neutron-star matter. With reasonable adjustments of its parameters, this EOS can be made consistent with laboratory data at nuclear saturation density $n_{\rm sat}\simeq 0.16~{\rm fm}^{-3}$ as well as recent chiral effective field theory calculations
of Ref.~\cite{Drischler:2020hwi,Drischler:2020fvz} in which error quantifications up to $\sim2.0\,n_{\rm sat}$ were made.
The high-density behavior can be controlled by varying the slope of the symmetry energy parameter, $L$, at $n_{\rm sat}$ within the range established from analyses of nuclear and observational data,
see, e.g., Ref.~\cite{Lattimer:2012xj}, and a power-law index. Consistency with astrophysical data on known masses and radii of NSs is also attainable with this EOS.
To begin with, no constraints are placed on the multi-particle system and therefore the independent variables are the baryon density $n_{\rm B}$ and the individual particle fractions $y_n$, $y_p$. The total energy density of nucleons with a common mass $m_H=939.5$ MeV is described by the ZL functional
\begin{eqnarray}
\varepsilon_H &=& \frac{1}{8\pi^2\hbar^3}\sum_{h=n,p}\left\{k_{Fh}(k_{Fh}^2 + m_H^2)^{1/2}(2k_{Fh}^2 + m_H^2)\right. \nonumber \\
&-& \left. m_H^4\ln\left[\frac{k_{Fh}+(k_{Fh}^2 + m_H^2)^{1/2}}{m_H}\right]\right\} \nonumber \\
&+& 4 n_{\rm B}^2 y_n y_p \left\{\frac{a_0}{n_{\rm sat}}
+\frac{b_0}{n_{\rm sat}^{\gamma}} [n_{\rm B}(y_n + y_p)]^{\gamma - 1}\right\} \nonumber \\
&+& n_{\rm B}^2 (y_n - y_p)^2\left\{\frac{a_1}{n_{\rm sat}}
+ \frac{b_1}{n_{\rm sat}^{\gamma_1}}[n_{\rm B}(y_n + y_p)]^{\gamma_1-1}\right\} \nonumber \\
\end{eqnarray}
The Fermi momentum of nucleon species $h$\footnote{While the symbols ``$H$'' and ``$h$'' connote hadrons, they actually refer to nucleons in the context of this paper.} is given by $k_{Fh} = (3\pi^2\hbar^3n_{\rm B} y_h)^{1/3}$. The constants $a_0, b_0$ and $\gamma$ refer to isospin symmetric matter, whereas $a_1, b_1$ and $\gamma_1$ to that of isospin asymmetric matter. The chosen values of these constants are listed in a later section.
The corresponding chemical potentials are
\begin{eqnarray}
\mu_n &=& (k_{Fn}^2 + m_H^2)^{1/2} \nonumber \\
&+& 4 n_{\rm B} y_p \left\{\frac{a_0}{n_{\rm sat}}+ \frac{b_0}{n_{\rm sat}^{\gamma}}[n_{\rm B}(y_n + y_p)]^{\gamma-1}\right\}
\nonumber \\
&+& 4 n_{\rm B}^2 y_p y_n \frac{b_0}{n_{\rm sat}^{\gamma}}(\gamma-1)[n_{\rm B}(y_n + y_p)]^{\gamma-2}
\nonumber \\
&+& 2 n_{\rm B} (y_n-y_p)\left\{\frac{a_1}{n_{\rm sat}} + \frac{b_1}{n_{\rm sat}^{\gamma_1}}[n_{\rm B}(y_n+y_p)]^{\gamma_1-1}\right\} \nonumber \\
&+& n_{\rm B}^2 (y_n-y_p)^2 \frac{b_1}{n_{\rm sat}^{\gamma_1}}(\gamma_1-1)[n_{\rm B}(y_n + y_p)]^{\gamma_1-2} \, , \\
\mu_p &=& (k_{Fp}^2 + m_H^2)^{1/2} \nonumber \\
&+& 4 n_{\rm B} y_n \left\{\frac{a_0}{n_{\rm sat}}+ \frac{b_0}{n_{\rm sat}^{\gamma}}[n_{\rm B}(y_n + y_p)]^{\gamma-1}\right\}
\nonumber \\
&+& 4 n_{\rm B}^2 y_p y_n \frac{b_0}{n_{\rm sat}^{\gamma}}(\gamma-1)[n_{\rm B}(y_n + y_p)]^{\gamma-2}
\nonumber \\
&-& 2 n_{\rm B} (y_n-y_p)\left\{\frac{a_1}{n_{\rm sat}} + \frac{b_1}{n_{\rm sat}^{\gamma_1}}[n_{\rm B}(y_n+y_p)]^{\gamma_1-1}\right\} \nonumber \\
&+& n_{\rm B}^2 (y_n-y_p)^2 \frac{b_1}{n_{\rm sat}^{\gamma_1}}(\gamma_1-1)[n_{\rm B}(y_n + y_p)]^{\gamma_1-2} \,.
\end{eqnarray}
Note the opposite $y_h$ and signs in the second and fourth terms of $\mu_n$ and $\mu_p$, respectively.
The pressure is obtained from the thermodynamic identity
\begin{equation}
P_H = n_{\rm B}\sum_{h=n,p}\mu_h y_h - \varepsilon_H
\end{equation}
and the equilibrium speed of sound from
\begin{equation}
\left(\frac{
c_{\rm eq}}{c}\right)^2 = \frac {d P_H}{d \varepsilon_H} \,.
\end{equation}
The adiabatic speed of sound is obtained by taking partial derivatives of the pressure and the total energy density with respect to baryon density while keeping all particle fractions fixed
\begin{equation}
\left(\frac{
c_{\rm ad}}{c}\right)^2 = \left.\frac{\partial P_H}{\partial n_{\rm B}}\right|_{y_h}
\left(\left.\frac{\partial \varepsilon_H}{\partial n_{\rm B}}\right|_{y_h}\right)^{-1}~.
\end{equation}
This is made particularly convenient by the choice of starting with a completely unconstrained system.
\subsection{Pure quark matter}
For the calculation of the quark EOS, we use the vMIT bag model~\cite{Gomes:2018eiv, Klahn:2015mfa}. The Lagrangian density of this model is
\begin{equation}
\mathcal{L}=\sum_{q=u,d,s}\left[\bar{\psi}_{q}\left(i \slashed{\partial} -m_{q}-B\right) \psi_{q}+\mathcal{L}_{\mathrm{vec}}\right] \Theta \,,
\end{equation}
where $\mathcal{L}_{\mathrm{vec}}$ describes repulsive interactions between quarks of mass $m_{q}$ confined within a bag as denoted by the $\Theta$ function:
\begin{equation}
\mathcal{L}_{\text {vec }}=-G_{v} \sum_{q} \bar{\psi} \gamma_{\mu} V^{\mu} \psi+\left(m_{V}^{2} / 2\right) V_{\mu} V^{\mu} \,.
\end{equation}
$B$ is a constant which reflects the cost of confining the quarks inside the bag, and the $m_{q}$ are the current quark masses.
Perturbative contributions~\cite{Kurkela:2009gj,Kurkela:2014vha}
are not included in vMIT because these become relevant at densities well above those achievable in the cores of the most massive neutron stars.
The state functions, energy density, chemical potential, and pressure, corresponding to the above Lagrangian (\textit{before} any constraints of baryon number conservation, charge neutrality, and chemical equilibrium are applied) for matter containing $u,d$ and $s$ quarks are
\begin{eqnarray}
\varepsilon_Q &=& \sum_{q=u,d,s}\varepsilon_q + \frac{1}{2}a~\hbar~[n_{\rm B}(y_u+y_d+y_s)]^2
+ \frac{B}{\hbar^3} \\
\varepsilon_q &=& \frac{3}{8\pi^2\hbar^3}\left\{k_{Fq} (k_{Fq}^2 + m_q^2)^{1/2}(2 k_{Fq}^2 + m_q^2)
\right. \nonumber \\
&-& \left. m_q^4\ln\left[\frac{k_{Fq}+(k_{Fq}^2 + m_q^2)^{1/2}}{m_q}\right]\right\} \\
\mu_q &=& (k_{Fq}^2 + m_q^2)^{1/2} + a~\hbar~n_{\rm B} (y_u + y_d + y_s) \\
P_Q &=& n_{\rm B} \sum_{q=u,d,s} \mu_q y_q - \varepsilon_Q \,,
\end{eqnarray}
where $a \equiv (G_v/m_V)^2$ and $k_{Fq} = (\pi^2 \hbar^3 n_{\rm B} y_q)^{1/3}$. The value of the vector interaction parameter $a$ is varied in the range $(0.1-0.3)~ \rm{fm}^{-2}$ to obtain different stiffness in the quark sector.
\subsection{Leptons}
Owing to the smallness of the electromagnetic fine structure constant $\alpha \simeq 1/137$, leptons are treated as non-interacting, relativistic particles for which
\begin{eqnarray}
\varepsilon_L &=&\frac{1}{8\pi^2\hbar^3}\sum_l\left\{k_{Fl}(k_{Fl}^2+m_l^2)^{1/2}(2k_{Fl}^2 + m_l^2)\right. \nonumber \\
&-& \left. m_l^4\ln\left[\frac{k_{Fl}+(k_{Fl}^2 + m_l^2)^{1/2}}{m_l}\right]\right\} \\
\mu_l &=& (k_{Fl}^2 + m_l^2)^{1/2} \\
P_L &=& n_{\rm B} \sum_l y_l \mu_l - \varepsilon_L \\
k_{Fl} &=& (3\pi^2\hbar^3 n_{\rm B} y_l)^{1/3} ~~;~~ l=e,\mu~.
\end{eqnarray}
At low baryon densities only electrons are present in the system, with muons appearing at a density $n_{\rm B}$
such that $\mu_e -m_{\mu} = 0$. Depending on the parametrization choice, this condition also gives the density at which muons vanish.
\section{Thermodynamics of multicomponent systems}
\label{sec:multi_compt}
The original formulation of the KW approach deals with one nucleon, the neutron, and three massless quarks
(the latter are, operationally, a single species with multiplicity 3).
Before the KW approach can be applied to more realistic neutron-star matter with the EOSs of the previous section, it must be generalized to include several particle species. To that end, we begin with a brief review of multicomponent thermodynamics to introduce the fundamental result from which the aforementioned generalization will be
performed in the next section. The relations laid out below are particularly helpful in highlighting the role of the (baryon chemical potential dependent) switch function $S$ which is instrumental in realizing a crossover transition.
The number density of a single-component system in the grand-canonical ensemble is given by the total derivative of the pressure with respect to the chemical potential,
\begin{equation}
n=\frac{dP}{d\mu}~.
\end{equation}
The equivalent expression for a multicomponent system is obtained from the grand potential
\begin{equation}
\Phi(T,V,\mu_i)=U-TS-\sum_i N_i\mu_i
\end{equation}
or, in units of energy-density,
\begin{equation}
\phi=\varepsilon-Ts-\sum_i n_i\mu_i ~.
\end{equation}
The differential of $\phi$ is
\begin{equation}
d\phi=d\varepsilon-sdT-\sum_i n_i d\mu_i
\end{equation}
which implies that the number density of particle species $i$ is given by
\begin{equation}
n_i=-\left.\frac{\partial \phi}{\partial \mu_i}\right|_{T,\mu_j}~.
\end{equation}
The thermodynamic identity
\begin{equation}
\varepsilon = Ts-P + \sum_i n_i\mu_i
\label{TI}
\end{equation}
means that $P=-\phi$ and therefore
\begin{equation}
n_i=\left.\frac{\partial P}{\partial \mu_i}\right|_{\mu_j} \label{eqni}
\end{equation}
where the temperature $T$ has been suppressed as, in what follows, only cold matter is considered.
\Eqn{eqni} is central to the subsequent discussion where we show the manner in which the individual number densities of the various nucleonic and quark species are modified by the switch function $S$ of the KW machinery.
\section{Unconstrained and beta-equilibrated crossover matter}
\label{sec:KWG}
In this section, we start with the crossover EOS where
baryon number conservation, charge neutrality, and weak interaction equilibrium are not imposed, i.e., ``unconstrained'' matter. Working with unconstrained quantities enables us to calculate the various partial derivatives required in the determination of the squared adiabatic speed of sound $c_{\rm ad}^2$ (see Sec.~\ref{sec:soundspeed}) prior to the imposition of the conditions mentioned above.
In the KW description of crossover matter, the pressure is given by
\begin{equation}
P_B = (1-S)P_H + S \, P_Q \label{eqPB}
\end{equation}
where $P_H$ and $P_Q$ are the hadron and quark pure-phase pressures respectively, and the switch function
\begin{equation}
S = \exp\left[-\left(\frac{\mu_0}{\mu}\right)^4\right]
\end{equation}
gives the fraction of quark matter to the total baryonic matter in the crossover setting, with $\mu$ being the average hadronic chemical potential
\begin{equation}
\mu = \frac{n_n \mu_n + n_p \mu_p}{n_n + n_p}~\,,
\end{equation}
and $\mu_0$ a typical energy scale for the crossover. This choice for $\mu$ will be justified in the next section.
Applying \Eqn{eqni} to hadrons leads to
\begin{eqnarray}
n_h^* &=& (1-S)\frac{\partial P_H}{\partial \mu_h}+S\frac{\partial P_Q}{\partial \mu_h}+(P_Q-P_H)\frac{\partial S}{\partial \mu_h} \nonumber \\
&=& (1-S)n_h+0+(P_Q-P_H) \frac{4\mu_0^4S}{\mu^5}\frac{\partial \mu}{\partial \mu_h} \nonumber \\
&=& (1-S)n_h+(P_Q-P_H) \frac{4\mu_0^4S}{\mu^5}\frac{n_h}{n_n+n_p} \nonumber \\
&=& n_h\left[1-S\left(1-\frac{4\mu_0^4}{\mu^5}\frac{P_Q-P_H}{n_n+n_p}\right)\right] \,,
\label{eqnh}
\end{eqnarray}
where $n_i^*=n_{\rm B} y_i^*$ refers to a crossover-matter density and $n_i = n_{\rm B} y_i$ to a pure-phase density. Thus, in the present context, the starred fractions are the physical quantities whereas the unstarred ones are merely bookkeeping devices. For leptons this distinction is irrelevant.
For quarks, one obtains
\begin{eqnarray}
n_q^* &=& (1-S)\frac{\partial P_H}{\partial \mu_q}+S\frac{\partial P_Q}{\partial \mu_q}+(P_Q-P_H)\frac{\partial S}{\partial \mu_q} \nonumber \\
&=& 0+S\,n_q+0 = S\,n_q \,.
\label{qno}
\end{eqnarray}
Finally, the energy density $\varepsilon$ is obtained from \Eqn{TI} using \Eqn{eqPB} for the pressure, Eqs.~(\ref{eqnh})-(\ref{qno}) for the number densities of hadrons and quarks, respectively, and the pure-phase chemical potentials defined in Sec.~\ref{sec:EOS}.
\subsection{Beta-equilibrium}
We turn now to the discussion of neutron-star matter that consists of nucleons, leptons and quarks.
Initially, the system is entirely unconstrained with $n_{\rm B}$ and $y_i$ ($i=n,p,u,d,s,e,\mu$) as the free variables. Then, strong equilibrium
\begin{equation}
\mu_n = 2\mu_d + \mu_u ~~;~~\mu_p = 2\mu_u + \mu_d
\end{equation}
and weak equilibrium
\begin{equation}
\mu_n = \mu_p + \mu_e ~~;~~\mu_e = \mu_\mu ~~;~~ \mu_d = \mu_s
\end{equation}
are enforced, as well as charge neutrality
\begin{equation}
n_p^*+(2n_u^*-n_d^*-n_s^*)/3-(n_e+n_{\mu})=0
\end{equation}
and baryon number conservation
\begin{equation}
n_n^*+n_p^*+(n_u^*+n_d^*+n_s^*)/3-n_{\rm B}=0 ~.
\label{barno}
\end{equation}
These conditions eliminate the particle fractions in favor of the total baryon density:
\begin{equation}
y_i \rightarrow y_{i,\beta}(n_{\rm B}) ~~;~~ i=n,p,u,d,s,e,\mu
\end{equation}
\subsection{Comparison of Kapusta-Welle (KW) with McLerran-Reddy (MR) and Zhao-Lattimer (ZL) EOSs}
In this subsection, we briefly discuss interesting similarities and differences between the crossover model of KW~\cite{Kapusta:2021ney} and recently proposed quarkyonic ``shell'' models of MR~\cite{McLerran:2018hbz} and ZL~\cite{Zhao:2020dvu}, and explain the reason why the latter is not suitable for $g$-mode~calculations in its present form.
The baryon number densities in the quarkyonic matter descriptions of MR~\cite{McLerran:2018hbz} and ZL~\cite{Zhao:2020dvu} are
\begin{eqnarray}
n_h^* &=& \frac{k_{Fh}^3-k_{0h}^3}{3\pi^2\hbar^3}
=\frac{k_{Fh}^3}{3\pi^2\hbar^3}\left(1-\frac{k_{0h}^3}{k_{Fh}^3}\right)
= n_h\left(1-\frac{k_{0h}^3}{k_{Fh}^3}\right) \,, \nonumber \\
\end{eqnarray}
where $k_{0h}$ are the minimum momenta of hadrons or nucleons
in quarkyonic matter, which depend on the corresponding Fermi momenta $k_{Fh}$ and thus the baryon number density. The precise way in which $k_{Fh} - k_{0h}$ depends on a chosen momentum scale $\Lambda$ and a common transition density $n_t$ is detailed in Eq.~(17) of Ref.~\cite{Zhao:2020dvu}.
Comparing the above expression to \Eqn{eqnh} from the previous section, it becomes clear that the presence of the hadron shell in the MR and ZL approaches forces hadrons to higher-momentum states much like $S$ does in the KW scheme:
\begin{equation}
S\left(1-\frac{4\mu_0^4}{\mu^5}\frac{P_Q-P_H}{n_n+n_p}\right) \stackrel{\wedge}{=} \frac{k_{0h}^3}{k_{Fh}^3} \,.
\end{equation}
This means that any particle species participating in $S(\mu)$ (that is, a species $i$ for which $\partial \mu/\partial \mu_i\neq0$) will invariably inherit a shell-like term in its crossover-matter number density.
In quarkyonic matter realizations, such a term is desirable for baryons but not for quarks and therefore $\mu$ must be a function of baryonic chemical potentials only.
Note that in the case of KW, the quark densities in crossover-matter $n_q^*$ are the product of the corresponding pure-phase densities and the quark-to-baryon fraction $S$ which is an \textit{a priori} assumption.
On the other hand, for both MR and ZL models, the densities and fractions of baryons and quarks in the quarkyonic phase are determined by the solution of the equilibrium equations.
Here, we should point out that in the MR and ZL implementations of the quarkyonic matter scenario, the nucleonic Fermi momenta are weakly dependent on baryon density when the latter exceeds the transition density, $n_t$; that is,
\begin{equation}
|k_{\infty,i}-k_{Fi,\beta}|/k_{\infty,i}\ll 1 ~~\text{for all}~ n_{\rm B}>n_t \,,
\end{equation}
where $k_{\infty,i}\equiv k_{Fi,\beta}(n_{\rm B} \rightarrow \infty)$.
On the other hand, the nucleonic chemical potentials and, by extension, the pressure change very rapidly with $k_{Fi}$ for $n_{\rm B}>n_t$ due to the presence of denominators $\propto(1-K_i)$ (see Eqs.~(19)-(20) in \cite{Zhao:2020dvu}) in their kinetic parts, where
\begin{eqnarray}
K_i^{\rm MR} &=& \left(\frac{k_{0i}}{k_{Fi}}\right)^2 \left(1 + \frac{2\Lambda^3}{k_{Fi}^3}\right)\\
K_i^{\rm ZL} &=& \left(\frac{k_{0i}}{k_{Fi}}\right)^2 \left(1 + \frac{\Lambda^2}{k_{Fi}^2}\right)~.
\end{eqnarray}
These terms remain close to zero [i.e. $(1-K_i)^{-1}\rightarrow \infty$], in equilibrium matter as a result of the aforementioned behavior of the nucleonic Fermi momenta, throughout the quarkyonic regime.
These two features of MR and ZL are responsible for divergent pressure derivatives with respect to $k_{Fi}$ which, in turn, lead to superluminal adiabatic sound speeds. Thus MR and ZL, in their current formulations, are unsuitable for our purposes and we do not apply them in $g$-mode~calculations for crossover matter.
\section{Sound speeds}
\label{sec:soundspeed}
In this section, we describe how calculations of the squared adiabatic and equilibrium sound speeds, $c_{\rm ad}^2$ and $c_{\rm eq}^2$, required in the calculation
of $g$-mode~frequencies, are performed. As one of our objectives is to provide contrasts between $g$-mode~frequencies in crossover matter and the case of a first-order transition treated via the Gibbs construction, both cases are considered below.
\subsection{Sound speeds in crossover matter}
Within the KW framework, the total pressure and energy density in the crossover region are
\begin{eqnarray}
P &=& P_B + P_e + P_{\mu} \\
\varepsilon &=& \varepsilon_B + \varepsilon_e + \varepsilon_{\mu} \\
\varepsilon_B &=& -P_B + \sum_{i=n,p,u,d,s}n_i^*\mu_i \,.
\end{eqnarray}
Using these, the adiabatic speed of sound is obtained by first calculating the expression
\begin{equation}
c_{\rm ad}^2(n_{\rm B},y_i) = \left.\frac{\partial P}{\partial n_{\rm B}}\right|_{y_i}
\left(\left.\frac{\partial \varepsilon}{\partial n_{\rm B}}\right|_{y_i}\right)^{-1}
\end{equation}
and then evaluating it in $\beta$-equilibrium
\begin{equation}
c_{ad,\beta}^2(n_{\rm B}) = c_{\rm ad}^2[n_{\rm B},y_{i,\beta}(n_{\rm B})]~.
\end{equation}
The equilibrium sound speed is given by the total derivatives of the pressure and the energy density with respect to the baryon density after the enforcement of $\beta$-equilibrium,
\begin{equation}
c_{\rm eq}^2 = \frac{dP_{\beta}}{dn_{\rm B}}\left(\frac{d\varepsilon_{\beta}}{dn_{\rm B}}\right)^{-1}~.
\end{equation}
\subsection{Sound speeds with Gibbs construction}
As in the crossover matter case, all thermodynamic quantities are expressed in terms of functions of the total baryon density $n_{\rm B}$, and the individual particle fractions $y_n$,~$y_p$,~$y_e$,~$y_{\mu}$,~$y_u$,~$y_d$,~$y_s$ which are, at this point, independent variables. That is,
\begin{eqnarray}
\varepsilon_H &=& \varepsilon_H(n_{\rm B},y_n,y_p) ~;~ P_H = P_H(n_{\rm B},y_n,y_p) ~; \nonumber \\
\mu_h &=& \mu_h(n_{\rm B},y_n,y_p) \\ \nonumber \\
\varepsilon_Q &=& \varepsilon_Q(n_{\rm B},y_u,y_d,y_s) ~;~ P_Q = P_Q(n_{\rm B},y_u,y_d,y_s) ~; \nonumber \\
\mu_q &=& \mu_q(n_{\rm B},y_q) \\ \nonumber \\
\varepsilon_L &=& \varepsilon_L(n_{\rm B},y_e,y_{\mu}) ~;~ P_L = P_L(n_{\rm B},y_e,y_{\mu}) ~; \nonumber \\
\mu_l &=& \mu_l(n_{\rm B},y_l) \,.
\end{eqnarray}
The conditions for weak equilibrium, charge neutrality, and baryon number conservation are applied afterwards.
These introduce another independent variable, $\chi$, which is the volume fraction of quarks in the mixed phase of Gibbs construction:
\begin{eqnarray}
&&P_H = P_Q ~;~ \mu_n = 2\mu_d + \mu_u ~;~ \mu_p = 2\mu_u + \mu_d \\ \nonumber \\
&&\mu_n = \mu_p + \mu_e ~;~ \mu_e = \mu_\mu ~;~\mu_d = \mu_s \\ \nonumber \\
&&3 (1-\chi) y_p+\chi(2y_u-y_d-y_s)-3(y_e+y_{\mu}) = 0 \\ \nonumber \\
&&3 (1-\chi)(y_n+y_p)+\chi(y_u+y_d+y_s)-3 = 0 \,.
\end{eqnarray}
Solving these equations eliminates the $y_i$ and $\chi$ in favor of $n_{\rm B}$. Thus the state variables become functions of only $n_{\rm B}$ according to the rule
\[ Q(n_{\rm B},y_i,y_j,...) \rightarrow Q[n_{\rm B},y_i(n_{\rm B}),y_j(n_{\rm B}),...] = Q(n_{\rm B})~.\] \\
Then, the thermodynamics of the mixed $(^*)$ phase are:
\begin{eqnarray}
\varepsilon^* &=& (1-\chi)\varepsilon_H + \chi \varepsilon_Q + \varepsilon_L \\
P^* &=& P_H+ P_L = P_Q+ P_L \nonumber \\
&=& (1-\chi) P_H + \chi P_Q + P_L \\
\mu_h^* &=& \mu_h ~~;~~ \mu_q^* = \mu_q \\
y_h^* &=& (1-\chi)y_h ~~;~~ y_q^* = \chi y_q \,.
\label{qnoG}
\end{eqnarray}
Quantities corresponding to leptons are not affected by the ratio of the two baryonic components in the mixed phase.
\begin{table}[h]
\caption{Parameter sets for the EOSs used in this work. }
\begin{center}
\begin{tabular}{ccrrrc}
\hline
\hline
Model & Parameter & XOA & XOB & XOC & Units \\ \hline
& $a_0$ & -96.64 & -90.39 & -96.64 & MeV \\
& $b_0$ & 58.85 & 52.60 & 58.85 & MeV \\
ZL & $\gamma$ & 1.40 & 1.446 & 1.40 & \\
& $a_1$ & -26.06 & -232.78 & -28.15 & MeV \\
& $b_1$ & 7.34 & 212.46 & 7.83 & MeV \\
& $\gamma_1$ & 2.45 & 1.1 & 3.5 & \\
\hline
& $m_u$ & 5.0 & 5.0 & 5.0 & MeV \\
& $m_d$ & 7.0 & 7.0 & 7.0 & MeV \\
vMIT & $m_s$ & 150.0 & 150.0 & 150.0 & MeV \\
& $a$ & 0.20 & 0.23 & 0.15 & fm$^2$ \\
& $B^{1/4}$ & 180.0 & 180.0 & 180.0 & MeV \\
\hline
KW & $\mu_0$ & 1.8 & 1.8 & 1.8 & GeV \\
\hline \hline
\end{tabular}
\end{center}
\label{tab:Parameters}
\end{table}
The mixed phase extends over those densities $n_{\rm B}$ for which $0\le \chi(n_{\rm B}) \le 1$. In contrast to the crossover case where $S$ operates at \textit{all} densities, $\chi$ is active only when the above condition is satisfied.
Moreover,
in the case of a
Gibbs construction of the first-order phase transition scenario, $\chi$ and $y_i$ are treated on an equal footing with no prior assumptions regarding their density dependence, whereas in the crossover
scenario, $S$ has a definitive functional form which the particle fractions must be adjusted to fit.
As a consequence, even though
both $\chi$ and $S$
describe the quark-to-baryon fraction, the former is a quantity for which we solve while the latter acts as a constraint replacing the Gibbs condition for mechanical equilibrium.
The sound speeds are obtained following the prescription outlined in the previous subsection. Alternatively, the adiabatic sound speed in the mixed phase can be calculated from the corresponding ones in the pure hadronic and quark phases separately according to Ref.~\cite{Jaikumar:2021jbw}
\begin{equation}
\frac{1}{c_{\rm ad}^{*2}} = \frac{1-\chi}{c_{{\rm{ad}},H}^2} + \frac{\chi}{c_{{\rm{ad}},Q}^2}~.
\end{equation}
\section{Results}
\label{sec:Results}
We turn now to present results based on calculations of the crossover EOS, associated NS properties, the two sound speeds and the resulting $g$-mode~frequencies. For contrast, results corresponding to pure hadronic matter and those for a first-order phase transition treated using the Gibbs construction are also presented.
\subsection{EOS and structural properties of NSs}
To construct crossover models, we have chosen the parameter values shown in Table~\ref{tab:Parameters}, labeled as XOA, XOB and XOC, for the parametrization of the EOSs used in this work.~\footnote{The method to determine the constants for the ZL parametrization is described in Refs.~\cite{Zhao:2020dvu} and \cite{Jaikumar:2021jbw}.}
These sets of parameters correspond to the nuclear and neutron-star properties shown in Table~\ref{tab:Properties}.
\begin{table}[h]
\caption{Nuclear and neutron-star properties corresponding to the parametrizations shown in Table~\ref{tab:Parameters}.
The symbols refer to
$n_{\rm sat}$: nuclear saturation density,
$E_0$: energy per particle at $n_{\rm sat}$,
$K_0$: compression modulus of symmetric nuclear matter (SNM) at $n_{\rm sat}$,
$S_v$: symmetry energy at $n_{\rm sat}$,
$L$: slope of $S(n)$ at $n_{\rm sat}$,
$n_{\mu,\text{on}}^{\beta}$: onset density of muons,
$n_{\mu,\text{off}}^{\beta}$: turnoff density of muons, and $n_{p,\text{off}}^{\beta}$: turnoff density of protons.
Quantities related to neutron stars are
$R$: radius, $M$: mass, $\beta=GM/Rc^2$: compactness,
$n_c$: central density, $p_c$: central pressure,
$\varepsilon_c$: central energy density,
$\Lambda$: dimensionless tidal deformability,
$c_{\rm eq}^2$: squared equilibrium sound speed,
and $c_{\rm ad}^2$: squared adiabatic sound speed.
The subscripts $1.4$ and ${\rm max}$ denote the masses of stars in ${\rm M}_{\odot}$.}
\begin{center}
\begin{tabular}{crrrrrc}
\hline
\hline
Property & XOA & XOB & XOC & ZL & Gibbs & Units \\ \hline
$n_{\rm sat}$ & 0.16 & 0.16 & 0.16 & 0.16 & 0.16 & fm$^{-3}$ \\
$E_0$ & -16.0 & -16.0 & -16.0 & -16.0 & -16.0 & MeV \\
$K_0$ & 250.0 & 260.0 & 250.0 & 250.0 & 250.0 & MeV \\
$S_v$ & 31.6 & 30.0 & 30.0 & 31.6 & 31.6 & MeV \\
$L$ & 43.0 & 70.0 & 65.0 & 43.0 & 43.0 & MeV \\
$n_{\mu,\text{on}}^{\beta}$ & 0.13 & 0.15 & 0.14 & 0.13 & 0.13 & fm$^{-3}$ \\
$n_{\mu,\text{off}}^{\beta}$ & 1.39 & 0.77 & 1.67 & N/A & 1.32 &fm$^{-3}$ \\
$n_{p,\text{off}}^{\beta}$ & N/A & 0.88 & N/A & N/A & N/A &fm$^{-3}$ \\
\hline
$R_{1.4}$ & 12.4 & 12.4 & 13.8 & 12.4 & 12.4 &{\rm km} \\
$\beta_{1.4}$ & 0.167 & 0.166 & 0.150 & 0.166 & 0.166 & \\
$n_{c,1.4}/n_{\rm sat}$ & 2.64 & 3.16 & 1.96 & 2.64 & 2.68 & \\
$p_{c,1.4}$ & 60.2 & 73.6 & 41.7 & 60.0 & 69.8 & ${\rm MeV~fm^{-3}}$ \\
$\varepsilon_{c,1.4}$ & 424.7 & 518.4 & 316.7 & 424.9 & 436.3 & ${\rm MeV~fm^{-3}}$ \\
$\Lambda_{1.4}$ & 428.9 & 426.7 & 841.3 & 430.1 & 421.2 & \\
$(c_{\rm eq}^2)_{c,1.4}$ & 0.398 & 0.331 & 0.294 & 0.397 & 0.406 & $c^2$ \\
$(c_{\rm ad}^2)_{c,1.4}$ & 0.429 & 0.331 & 0.497 & 0.426 & 0.528 & $c^2$ \\
\hline
$\ensuremath{R_{\Mmax}}$ & 11.5 & 10.4 & 13.0 & 11.1 & 11.2 &{\rm km} \\
$\ensuremath{M_{\rm max}}$ & 2.11 & 2.04 & 2.13 & 2.23 & 2.08 &${\rm M}_{\odot}$ \\
$\beta_{\rm max}$ & 0.270 & 0.289 & 0.242 & 0.295 & 0.275 & \\
$n_{\rm c,max}/n_{\rm sat}$ & 5.83 & 7.38 & 4.70 & 6.14 & 6.32 & \\
$p_{\rm c,max}$ & 362.7 & 696.1 & 213.4 & 577.0 & 457.7 & ${\rm MeV~fm^{-3}}$ \\
$\varepsilon_{\rm c,max}$ & 1142.9 & 1549.5 & 886.8 & 1202.4 & 1264.9 & ${\rm MeV~fm^{-3}}$ \\
$\Lambda_{\rm max}$ & 13.8 & 6.4 & 30.1 & 6.2 & 11.1 & \\
$(c_{\rm eq}^2)_{\rm c,max}$ & 0.426 & 0.566 & 0.316 & 0.767 & 0.576 & $c^2$ \\
$(c_{\rm ad}^2)_{\rm c,max}$ & 0.507 & 0.818 & 0.353 & 0.889 & 0.653 & $c^2$ \\
\hline \hline
\end{tabular}
\end{center}
\label{tab:Properties}
\end{table}
For XOA, all values are within 1-$\sigma$ deviations of empirical/observational constraints discussed below. While XOB and XOC do not perform as well, they are used here to illustrate some important physics related to the behavior of the $g$-mode~frequency.
Specifically, with XOB we investigate $g$-mode~frequency features corresponding to a sound-speed peak due to proton disappearance, whereas in XOC the peak in the speed of sound is not related to a change in the number of degrees of freedom.
Models labeled ZL (nucleons only) and Gibbs (nucleons plus quarks with a Gibbs construction) use the appropriate parameters of XOA.
\vspace{3mm}
\begin{figure}[h!]
\includegraphics[scale=0.32]{plotMR.pdf}
\caption{Neutron-star $M$-$R$ curves for the various EOS models used in this work. The black, dashed lines represent the 90\%-confidence level constraints extracted from recent radio, x-ray, and gravitational-wave observations by Legred et al.~\cite{Legred:2021hdx}.
Models corresponding to the ``A'' parameter set fit these constraints well with differences between the three depending on the order of the transition to quark matter or the lack thereof.
While stars with $M\leq 1.8\,{\rm M}_{\odot}$ using model XOB are within the constraints, with model XOC only stars close to the maximum mass satisfy the Legred et al. constraints.
}
\label{fig:MR}
\end{figure}
\begin{figure*}[htb]
\parbox{0.5\hsize}{
\includegraphics[width=\hsize]{plotPvN.pdf}\\[-4ex]
}\parbox{0.5\hsize}{
\includegraphics[width=\hsize]{plotEvN2.pdf}\\[-4ex]
}
\caption{Left panel: pressure versus baryon density as obtained by the assorted EOSs used herein compared with the astrophysical constraints of Legred et al.~\cite{Legred:2021hdx} (in black, dashed lines).
All parametrizations meet these constraints successfully with the exception of XOC which fails to stay in the allowed region around $2.0\,n_{\rm sat}$, where $n_{\rm sat}=0.16~{\rm fm}^{-3}$.
Right panel: energy-per-particle versus baryon density of beta-equilibrium matter for the various EOSs used in this work compared with the 1-$\sigma$ and 2-$\sigma$ constraints
from Drischler et al.~\cite{Drischler:2020fvz} obtained in a chiral EFT framework. Results corresponding to Gibbs and ZL are not shown because, over the range of densities displayed, they are identical to XOA.
Only XOA remains within the 2-$\sigma$ constraints of \cite{Drischler:2020fvz} up to $\sim2.0\,n_{\rm sat}$.
}
\label{fig:PN_EN_nB}
\end{figure*}
We wish to note that the values of the symmetry energy $S_v$ and and its slope $L$ at $n_{\rm sat}$ used in our work (see Table~\ref{tab:Properties}) lie in the range $\simeq 31 \pm 2$ MeV and $\simeq 51\pm 11$ MeV, respectively, recommended in Ref.~\cite{Lattimer:2012xj}. These values led to the bounds on the radius of a $1.4\,{\rm M}_{\odot}$ star to be $R_{1.4} \simeq 12 \pm 1$ km.
\vspace{3mm}
Interpretations of the recent PREX-II experiment carried out at the Jefferson Lab (JLab) measuring the neutron skin thickness of $^{208}$Pb~\cite{Adhikari:2021phr}, $R_\mathrm{skin}^{^{208}\mathrm{Pb}} = 0.283 \pm 0.071$~fm, however, widely vary in their inferences of the appropriate values of $S_v$ and $L$ to be used.
For example, Reed et al.~\cite{Reed:2021nqk}, using relativistic mean-field theory (RMFT) calculations to analyze the JLab data, conclude that $S_v=38.1\pm 4.7$ MeV and $L=106\pm 37$ MeV, values that are significantly higher than those deduced in earlier works. Furthermore, the bound $R_{1.4} > 13.25$~km was found there.
Reinhard et al.~\cite{Reinhard:2021utv}, use covariant RMFT (with density-dependent couplings) and nonrelativistic energy functionals to analyze the PREX-II data and
combine it with the dipole polarizability data of $^{208}$Pb
to arrive at $S_v = 32\pm 1$ MeV and $L=54\pm 8$ MeV. In addition, these authors obtain $R_\mathrm{skin}^{^{208}\mathrm{Pb}} = 0.19\pm 0.02$~fm in accord with earlier deductions.
Similar results are obtained by Essick et al.~\cite{Essick:2021kjb} who report $S_v = 34\pm 3$ MeV, $L=58\pm 18$ MeV and $R_\mathrm{skin}^{^{208}\mathrm{Pb}} = 0.19~^{+0.03}_{-0.04} $~fm from a non-parametric EOS coupled with Gaussian processes.
Combining recent mass and radius measurement from radio and x-ray data from \textit{NICER}, Biswas~\cite{Biswas:2021yge} finds $R_\mathrm{skin}^{^{208}\mathrm{Pb}} = 0.20\pm 0.05$~fm and $R_{1.4} = 12.75~^{+0.42}_{-0.54}$ km using nuclear EOSs with piecewise polytrope parametrization.
Given the fluid state of theoretical inferences from the analysis of JLab data, we have opted to stick with the values used in Table~\ref{tab:Properties}.
\vspace{3mm}
Figure~\ref{fig:MR} shows mass versus radius ($M$-$R$) curves for all the models considered along with the
recent constraints obtained by Legred et al.~\cite{Legred:2021hdx}, which combined available observations including the radio mass measurements of PSR J0348+0432 and J0470+6620~\cite{Fonseca:2021wxt,Cromartie:2019kug,Antoniadis:2013pzd}, the mass and tidal deformability measurements of GW170817 and GW190425~\cite{Abbott:2018wiz,LIGO:2017qsa,Abbott:2020uma}, and the x-ray mass and radius constraints from latest \textit{NICER} measurements of J0030+0451 and J0470+6620~\cite{Miller:2019cac,Riley:2019yda,Miller:2021qha,Riley:2021pdl}.
The constraints of Legred et al. were obtained by using the hierarchical inference~\cite{Loredo:2004nn} and a nonparametric survey through Gaussian Processes (GPs) conditioned on existing EOS models in the literature~\cite{Landry:2020vaw,Landry:2018prl,Essick:2019ldf}.
\vspace{3mm}
The left panel of Fig.~\ref{fig:PN_EN_nB} displays results of the pressure versus baryon number density ($P-n_{\rm B}$) relation for the various models used in the present work, in contrast to those inferred from Ref.~\cite{Legred:2021hdx} mentioned above (the black dashed boundaries, adapted from their Fig.~4).
To provide a comparison, results of energy versus density $E$ vs $n_{\rm B}$ of $\beta$-equilibrated neutron-star matter (NSM) are shown in the right panel of Fig.~\ref{fig:PN_EN_nB},
together with those for NSM from the chiral effective theory calculations of Ref.~\cite{Drischler:2020fvz} where 1-$\sigma$ and 2-$\sigma$ error estimates up to $\sim2.0\,n_{\rm sat}$ were provided.
Although not shown, we also find that results of the crossover, ZL and Gibbs models for $P-n_{\rm B}$ and $E$ versus $n_{\rm B}$ are consistent with microscopic Greens' function calculations of Gandolfi et al.~\cite{Gandolfi:2011xu}.
\vspace{3mm}
The squared adiabatic and equilibrium sound speeds $c_{\rm ad}^2$ and $c_{\rm eq}^2$ versus baryon density $n_{\rm B}$ are shown in the left and right panels of Fig.~\ref{fig:cs_ad_eq}, respectively. Both $c_{\rm ad}^2$ and $c_{\rm eq}^2$ increase monotonically with $n_{\rm B}$ for the ZL model in which nucleons are the only baryons. The non-monotonic behaviors of the other curves are due to admixtures of nucleons and quarks in the baryon sector. The $c_{\rm eq}^2 (n_{\rm B})$ for the Gibbs model suddenly drops (rises) at the onset (end) of the mixed phase (the latter not shown in the figure), whereas $c_{\rm ad}^2 (n_{\rm B})$ varies smoothly. Results for the crossover models XOA and XOC are similar in structure, whereas those for XOB show more structure at large $n_{\rm B}$ due to the disappearance of protons. With the exception of model XOB, $c_{\rm ad}^2 > c_{\rm eq}^2$ for all other models at all $n_{\rm B}$.
\vspace{3mm}
\begin{figure*}[htb]
\parbox{0.5\hsize}{
\includegraphics[width=\hsize]{plotCAD.pdf}\\[-4ex]
}\parbox{0.5\hsize}{
\includegraphics[width=\hsize]{plotCEQ.pdf}\\[-4ex]
}
\caption{Left panel: the squared adiabatic sound speed $c_{\rm ad}^2$ as a function of the baryon density $n_{\rm B}$.
Right panel: the squared equilibrium sound speed $c_{\rm eq}^2$ as a function of the baryon density $n_{\rm B}$.
}
\label{fig:cs_ad_eq}
\end{figure*}
\begin{figure}[h!]
\includegraphics[scale=0.38]{plot1overC2.pdf}
\caption{Difference between the inverse-squared sound speeds versus the baryon number density.
The peaks in the vicinity of $n_{\rm sat}$ correspond to muon appearance and are present in all models.
The peak at $\sim3\,n_{\rm sat}$ for Gibbs occurs at the onset of the mixed phase which extends beyond the densities shown here.
The peak around $5.5\,n_{\rm sat}$ for XOB is the combined effect of muon and proton disappearance in this model.
On the other hand, the peak at $\sim2.5\,n_{\rm sat}$ for XOC results from inflection points in the quark and neutron fractions.}
\label{fig:OOC}
\end{figure}
\begin{figure}[h!]
\includegraphics[scale=0.32]{plotyiA.pdf}
\caption{Particle fractions of the crossover model XOA versus baryon density. Quarks are still present below 0.3 fm$^{-3}$ but at two orders of magnitude less than what is shown here.}
\label{fig:YIA}
\end{figure}
\begin{figure}[h!]
\includegraphics[scale=0.32]{plotyiB.pdf}
\caption{Particle fractions of model XOB versus baryon density. Muons drop out of the system at 0.77 fm$^{-3}$ and protons at 0.88 fm$^{-3}$ leading to a sharp peak in the adiabatic sound speed (cf. Fig.~\ref{fig:OOC}). Electrons are still present above 0.85 fm$^{-3}$ but at two orders of magnitude less than what is shown here.}
\label{fig:YIB}
\end{figure}
\begin{figure}[h!]
\includegraphics[scale=0.32]{plotyiC.pdf}
\caption{Particle fractions of model XOC versus baryon density. The inflectionary behavior of the quark and neutron fractions around 0.4 fm$^{-3}$ is responsible for the broad peak behavior of XOC occurring in Fig.~\ref{fig:OOC}.}
\label{fig:YIC}
\end{figure}
Figure~\ref{fig:OOC} shows the difference of the inverses of the adiabatic and equilibrium sound speeds, $\Delta(c^{-2})\equiv 1/c_{\rm eq}^2-1/c_{\rm ad}^2$ as a function of the baryon density. This quantity is particularly important in the context of $g$-modes~because it enters directly in the calculation of the Brunt--V\"ais\"al\"a~frequency (discussed in more detail in the next section).
\vspace{3mm}
A comparison between this figure and Figs.~\ref{fig:YIA}--\ref{fig:YIC} which show the particle fractions corresponding to the three crossover models used in this work, reveals a direct correlation between sharp maxima in the former and particle appearance/disappearance in the latter.
Smooth maxima, such as those exhibited by XOA around 0.7 fm$^{-3}$ and XOC around 0.4 fm$^{-3}$, reflect non-monotonic behaviors in the slopes of the quark and the neutron fractions. Also worth noting is that, in the present framework and with the chosen parametrizations, quarks are never the dominant contributors to the baryon density for densities relevant to neutron stars.
\vspace{3mm}
\begin{table}[h]
\caption{Various representations of the quark content of $2\,{\rm M}_{\odot}$ NSs corresponding to the Gibbs and the three crossover models in this paper. The symbols are
$M_{\rm B}$: the total baryonic mass,
$Y_Q^{\rm{bar}}$: contribution of quarks to the total baryon number/total baryon number,
$Y_Q^{\rm{part}}$: quark particle number/total particle number,
$Y_Q^{\rm{nuc}}$: quark particle number/nucleon number,
${M_Q^{\rm{B}}}/{M_{\rm B}}$: contribution of quarks to the total baryonic mass/total baryonic mass, and
${M_Q^{\rm{G}}}/{M_{\rm G}}$: contribution of quarks to the total gravitational mass/total gravitational mass.
}
\begin{center}
\begin{tabular}{ccccccc}
\hline
\hline
Model & $M_{\rm B}$ & $Y_Q^{\rm{bar}}$ & $Y_Q^{\rm{part}}$ & $Y_Q^{\rm{nuc}}$ & ${M_Q^{\rm{B}}}/{M_{\rm B}}$ & ${M_Q^{\rm{G}}}/{M_{\rm G}}$ \\
& $({\rm M}_{\odot})$ & ($\times 10^{-2}$) & ($\times 10^{-2}$) &
($\times 10^{-2}$) & ($\times 10^{-2}$) & ($\times 10^{-2}$) \\ \hline
XOA & 2.31 & 0.35 & 1.03 & 1.04 & 0.35 & 0.50 \\
XOB & 2.33 & 2.15 & 6.18 & 6.58 & 2.15 & 3.04 \\
XOC & 2.29 & 0.06 & 0.19 & 0.19 & 0.06 & 0.09 \\
Gibbs & 2.35 & 1.91 & 5.53 & 5.85 & 1.91 & 2.89 \\
\hline \hline
\end{tabular}
\end{center}
\label{tab:quarkfrac}
\end{table}
Consequently, the contributions of quarks to the total baryon number as well as the total mass of the star (both baryonic and gravitational) are rather small in the models considered, as shown in Table~\ref{tab:quarkfrac} for stars with gravitational mass $M=2\,{\rm M}_{\odot}$. These are straightforwardly calculated as follows: The solution of the TOV equations gives, among other things, the baryon density as a function of the star's radius, $n_{\rm B}(r)$. The quark particle fractions as functions of the baryon density are obtained from the $\beta$-equilibrated equivalents of Eqs. (\ref{qno}) and (\ref{qnoG}) for the KW and Gibbs cases, respectively. Explicitly, $y_q (r) = y_{q,\beta}^*[n_{\rm B}(r)]$. For the quark baryon fraction at each radius we divide this by 3. The quark particle densities are then $n_q(r) = y_q(r)n_{\rm B}(r)$.The baryon number due to quarks of all species $q=u,d,s$ in the star is given by the integral
\begin{equation}
N_Q = 4\pi \int_0^R dr r^2 \frac{\sum_q n_q(r)/3}{[1-2GM(r)/r]^{1/2}}.
\end{equation}
Here, $M(r)$ is the total gravitational mass of the star as a function of its radius $r$, also given by the solution of the TOV equations. Therefore, the amount of baryonic mass in the star provided by quarks is $M_Q^{\rm{B}} = m_H N_Q$~\cite{glendenning2012compact}. For the quark gravitational mass we begin by calculating the quark energy density as a function of the radius according to $\varepsilon_Q(r) = \varepsilon_{Q,\beta}^*[n_{\rm B}(r)]$ where $\varepsilon_Q^* = -S P_Q + \sum_q n_q^*\mu_q$. Afterwards, we perform the integral $M_Q^{\rm{G}} = 4\pi \int_0^R dr r^2 \varepsilon_Q(r)$.
\vspace{3mm}
Specifically, in model XOA, quarks contribute less than 0.5\% of the total baryon number, about 1\% of particles in its $2\,{\rm M}_{\odot}$ stars are quarks and they are responsible for around 0.35\% (0.5\%) of the total baryonic (gravitational) mass of the star. For XOB and Gibbs models, quarks contribute about $2\%$ of the total baryon number, over 5\% of the total particle number, and 2\% (3\%) of the total baryonic (gravitational) mass of the respective $2\,{\rm M}_{\odot}$ stars. In model XOC, quarks are, for all intents and purposes, irrelevant.
As noted before, such a clean separation may not be possible in treatments that intermingle hadron/nucleon and quark interactions.
\subsection{Sound speeds and the Brunt--V\"ais\"al\"a~frequency}
Having outlined the thermodynamics of the multi-component system and the specific models employed in our study of the smooth crossover transition in neutron stars, we turn now to the calculation of the star's $g$-mode~frequencies.
Our main goal is to compare the behavior of the $g$-mode~frequencies in the crossover model with those in the Gibbs mixed phase.
The $g$-mode~frequencies ($\nu_g$=$\omega/(2\pi)$) and normalized amplitudes for the radial and tangential parts of the fluid perturbation ($\xi_r$ and $\xi_h$ respectively) are estimated within the relativistic Cowling approximation (see below) by computing numerical solutions to the following equations of motion for fluid variables $U,V$~\cite{Jaikumar:2021jbw}
\begin{eqnarray}
\label{eq:uv}
\frac{dU}{dr}&=&\frac{g}{c_{\rm ad}^2}U+{\rm e}^{\lambda/2}\left[\frac{l(l+1){\rm e}^{\nu}}{\omega^2}-\frac{r^2}{c_{\rm ad}^2}\right]V \nonumber \\
\frac{dV}{dr}&=&{\rm e}^{\lambda/2-\nu}\frac{\omega^2-N^2}{r^2}U+g\Delta(c^{-2})V \,,
\end{eqnarray}
which are simplified forms of the original perturbation equations~\cite{McD83,RG92,Kantor:2014lja}.
In \Eqn{eq:uv}, $U$ = $r^2{\rm e}^{\lambda/2}\,\xi_r$, $V$ = $\omega^2 r\, \xi_h$ = $\delta P/(\varepsilon+P)$, $\Delta(c^{-2})=c_{\rm eq}^{-2}-c_{\rm ad}^{-2}$, $\lambda$ and $\nu$ are metric functions. The scale of the mode frequency is set by the Brunt--V\"ais\"al\"a~frequency
\begin{eqnarray}
N^2 = g^{2}\Delta(c^{-2}){\rm e}^{\nu-\lambda},
\end{eqnarray}
where $g=-\nabla P/(\varepsilon+P)$.
The relativistic Cowling approximation neglects the back reaction of the gravitational potential by excluding metric perturbations that must accompany matter perturbations in a general relativistic treatment~\cite{Thorne:1967a,Thorne:1967b,Lindblom:1983,Detweiler:1985,Finn:1987,Andersson:1995wu}.
It reduces the number and complexity of the equations we have to solve, while providing results for $g$-mode~frequencies that are accurate at the few \% level~\cite{gregorian2015nonradial}.
Details on the solution methods for~\Eqn{eq:uv} and relevant boundary conditions are provided in Ref.~\cite{Jaikumar:2021jbw}.
\vspace{3mm}
\begin{figure}[h!]
\includegraphics[scale=0.32]{plotBV.pdf}\\[-3.0ex]
\caption{The Brunt--V\"ais\"al\"a~frequency in a hybrid star of mass $2.0\,{\rm M}_{\odot}$ for the Gibbs and crossover models.
The Brunt--V\"ais\"al\"a~frequency for the ``good'' crossover model XOA is very similar to the nucleonic ZL EOS (which includes muons), whereas the Gibbs model shows a distinct peak corresponding to the rapid onset of quark matter. Parameters for the nuclear and quark EOSs are as in Table~\ref{tab:Parameters}.
}
\label{fig:BV-EOS}
\end{figure}
Figure~\ref{fig:BV-EOS} is a comparison of the Brunt--V\"ais\"al\"a~frequency in the three models considered in this work. The crossover model, where quarks are always present in the core EOS, albeit in minuscule fractions, resembles the purely nucleonic ZL EOS in this respect, while the sudden onset of quarks in the Gibbs model is clearly imprinted in the form of a sharp peak.
Quarks enter at a density $n_{\rm B} \simeq$ 0.514 fm$^{-3}$ corresponding to $r/R$ = 0.473 ($r=0$ at the center) in the Gibbs model. As a consequence of the difference of sound speeds being negative in distinct density regimes for XOB (Fig.~\ref{fig:OOC}), the corresponding Brunt--V\"ais\"al\"a~frequency is imaginary, implying an instability to convection.\footnote{This feature could be an artifact of the Cowling approximation, but will likely be absent in the solution of the full general relativistic treatment~\cite{Thorne:1967a,Thorne:1967b,Lindblom:1983,Detweiler:1985,Finn:1987,Andersson:1995wu}
of the $g$-mode~in which $c_{\rm eq}^2$ does not enter explicitly.}
However, convection is absent at zero temperature; therefore these regions are unphysical and can play no role in the global $g$-mode~spectrum. Accordingly, XOB is omitted from Figs.~\ref{fig:gmode-EOS}, \ref{fig:eigenfunctions} and \ref{fig:energy}.
\begin{figure}[h!]
\includegraphics[scale=0.32]{plotGMnoB.pdf}\\[-3.0ex]
\caption{The $g$-mode~frequency as a function of the stellar mass in the Gibbs, crossover, and ZL models.
Parameters for the nuclear and quark EOSs are as in Table.~\ref{tab:Parameters}.
The g-modes corresponding to XOB are unstable and therefore, this model, is excluded from the present and the next two figures.}
\label{fig:gmode-EOS}
\end{figure}
\begin{figure}[]
\includegraphics[scale=0.28]{plotXI.pdf}\\[-3.0ex]
\caption{Amplitudes of the radial ($r$) and transverse ($h$) components of the $g$-mode~displacement (eigenfunctions) as a function of distance from the center for the ZL, XOA, XOC and Gibbs EOSs for a $\sim2\,{\rm M}_{\odot}$ star. The order of the mode (g1 is fundamental, g2 is overtone) is indicated in the legend in the panels.
}
\label{fig:eigenfunctions}
\end{figure}
Figure~\ref{fig:gmode-EOS} compares the $g$-mode~frequency, $\nu_g=\omega/(2\pi)$, for the crossover and Gibbs models. While XOA and XOC are very similar to the nucleonic ZL EOS (which includes muons), the Gibbs model shows a distinctly rising spectrum corresponding to the rapid onset of quark matter. These findings are consistent with the result for the Brunt--V\"ais\"al\"a~frequency in Fig.~\ref{fig:BV-EOS} and the conclusions in Paper I.
In the crossover model XOB, protons disappear above some critical density. In contrast to the smooth behavior of $g$-mode~frequencies in XOA and XOC, the sudden disappearance of protons in XOB produces a sharp rise in the spectrum akin to the Gibbs case but renders $g$-modes~to become unstable (not shown in Fig.~\ref{fig:gmode-EOS}), confirming that dramatic changes in the $g$-mode~frequency require {\it appearance or disappearance of a (strongly interacting) particle species, not merely a smooth change in composition}.
This is why, except for extreme parameter choices, crossover models will not show the $g$-mode~feature resulting from the presence of quarks that Gibbs models do.
The panels in Fig.~\ref{fig:eigenfunctions} show the comparison of the core $g$-mode~amplitude between the three chosen models (ZL, crossover and Gibbs). The radial component $\xi_r$ of the fundamental mode (labeled ``g1'' in the panels) has no nodes in the core, while the radial part of the first overtone (labeled ``g2'') has one, as expected. The horizontal component $\xi_h$ has one more node than the corresponding radial component of the same order.
The larger amplitude of $\xi_h$ relative to $\xi_r$ indicates that the $g$-mode~is dominated by transverse motion of the perturbed fluid. While there is little difference between the ZL and crossover models in the profile of these eigenfunctions, the Gibbs case is markedly different. Its amplitude relative to the other two is larger, and it changes abruptly upon the onset of quark matter in the core.
\begin{figure}[h!]
\includegraphics[scale=0.32]{plotDEDR.pdf}\\[-3.0ex]
\caption{Energy/unit distance of the fundamental $g$-mode~as a function of distance from the center for the ZL, XOA, XOC, and Gibbs EOS for a $\sim2\,{\rm M}_{\odot}$ star. Note that the mode energies in the ZL and crossover cases are scaled up by a factor of 10 for the purpose of comparison.}
\label{fig:energy}
\end{figure}
The energy per unit radial distance $dE_T/dr$ contained in the oscillatory motion corresponding to a frequency $\omega$ is given in terms of the amplitude as~\cite{McD83}
\begin{equation}
\frac{dE_T}{dr} =\frac{\omega^2\,r^2}{2}(\varepsilon+P){\rm e}^{(\lambda-\nu)/2}\,\left[\xi_r^2{\rm e}^{\lambda}+l(l+1)\xi_h^2\right] \,.
\end{equation}
Figure~\ref{fig:energy} shows the comparison of the core $g$-mode~energy/unit distance for the three chosen models (ZL, crossover and Gibbs) for a $\sim2\,{\rm M}_{\odot}$ star. The typical scale of the energy/unit distance deep in the core is approximately $10^{50}\, \rm{ergs/km}$ for the ZL and crossover models, while it is of order $10^{51}\, \rm{ergs/km}$ for the Gibbs case.
While the profiles are similar for the first two, the mode energy in the Gibbs case is overwhelmingly larger in the core, once quark matter appears. Thus, both the frequency and the amplitude (and hence the energy) of the core $g$-mode~is strongly amplified in quark matter in comparison to nucleonic matter or weaker forms of the phase transition. This can have bearing on the gravitational wave detection of $g$-modes~excited in neutron star mergers as discussed below.
\subsection{Discussion}
\label{sec.Discs}
The results in the previous subsection raise some points that are noteworthy. In particular, they highlight the relation between the behaviors of the speeds of sound and of the particle concentrations, and the spectrum of the $g$-mode~signal. Thus, the latter becomes a diagnostic which distinguishes nucleonic and hybrid matter \textit{as well as} the Gibbs and crossover transitions.
Peaks in the sound-speed difference $\delta c \equiv c_{\rm ad}^2 - c_{\rm eq}^2$, and the difference of the inverses $\Delta (c^{-2}) \equiv 1/c_{\rm eq}^2 - 1/c_{\rm ad}^2$ occur when particles appear or disappear [i.e. when the number of degrees of freedom (DOF) of the system changes] and when the particle concentrations $y_i(n_{\rm B})$ are not monotonic or have inflection points (i.e., their first or second derivatives with respect to $n_{\rm B}$ change sign).
The introduction or removal of a particle species from the system occurs when the relevant chemical potential either exceeds or falls below its rest mass threshold while maintaining charge neutrality.
Whereas the functional form of the $y_i$'s depends on the parametrization of the EOS, a change in the number of DOFs also leads to non-monotonic or inflectional behavior to the concentrations of particles already present in the system. Thus, in some sense, specific choices of the EOS parameters can mimic aspects of the emergence of new particles that are relevant to the sound speeds.
Signals of $g$-mode~with a characteristic fast rise in its frequency (such as
those corresponding to Gibbs shown in Fig.~\ref{fig:gmode-EOS} and XOB for which the $g$-mode~becomes unstable)
can occur only when particle species enter or leave the system. In such cases, the peaks in $\Delta (c^{-2})$ are sharp and asymmetric: vertical rise and quasi-lorentzian decay for appearance (see Fig.~\ref{fig:OOC}, Gibbs), and quasi-lorentzian rise and vertical drop for disappearance (XOB); whereas peaks due to parametrizations resemble symmetric Gaussians (XOC). Therefore, this kind of signal cannot be produced by quarks in matter with a smooth crossover because quarks are always present: their concentrations are vanishingly small at lower densities but never identically zero.
It is, however, possible to parametrize the hadronic EOS such that, in $\beta$-equilibrated crossover matter, protons exit the system. The magnitude of the peak of $\Delta (c^{-2})$ \textit{appears} to be proportional to the number of remaining DOFs (smallest in PNM; largest in $n$-$uds$-$e$ matter). This produces $g$-mode~frequency spectra similar to those found in matter with a Gibbs construction. So, although quarks are not directly responsible for this effect, they do serve the purpose of amplifying it.
For the specific case of crossover model XOB, a rather extreme parametrization ($K=260$ MeV, $S_v=30$ MeV, $L=70$ MeV, $\gamma_1=1.1$) was required for the proton disappearance while also meeting neutron-star constraints and having a peak at low-enough densities to be relevant. Consequently, the tentative conclusion is that hyperons or a first-order transition into quark matter through Gibbs construction are more likely to cause a distinctive peak in the Brunt--V\"ais\"al\"a~frequency than a crossover transition; nevertheless, the latter remains a viable, if improbable, option.
It is pertinent to mention that we have not performed calculations for first-order transitions with a Maxwell construction, which assume a sufficiently large surface tension between the pure hadronic and quark phases in bulk separated by a sharp boundary.
It has been shown that in such cases, $g$-modes~always vanish when the perturbed fluid element adjusts instantaneously (i.e. a very rapid conversion) to maintain thermodynamic equilibrium, but can arise if the microscopic phase conversion at the interface is slow enough, leading to distinctive features on the extended hybrid branch of NSs~\cite{Pereira:2017rmp,Tonetto:2020bie}. Such $g$-modes~associated with a discontinuity in density (``\textit{discontinuity} $g$-modes'', different from the ``\textit{compositional} $g$-modes'' that we consider)
has been widely studied in the literature; see e.g., extensive discussions in Refs.~\cite{Sotani:2001bb,Flores:2013yqa,Ranea-Sandoval:2018bgu,Rodriguez:2020fhf,Lau:2020bfq}.
\section{Summary and Conclusions}
\label{sec:Concs}
The main objectives of this work were to examine $g$-mode~frequencies in a smooth crossover scenario of the hadron-to-quark transition, and to compare them with those of a first-order transition treated using Gibbs and Maxwell constructions in Paper I~\cite{Jaikumar:2021jbw}.
For the crossover model, we chose the recent approach adopted by Kapusta and Welle~\cite{Kapusta:2021ney}, who constructed an EOS for PNM that resembled the smooth crossover observed in lattice calculations at finite temperatures.
We have generalized their approach to $\beta$-equilibrated NSM so that comparisons with the results of Paper I could be made. To describe nucleons, we used the ZL parametrization~\cite{Zhao:2020dvu} which reproduces near-saturation laboratory data as well as results of chiral EFT calculations~\cite{Drischler:2020hwi,Drischler:2020fvz} up to $\sim$ $2.0\,n_{\rm sat}$.
For quark matter we used the vMIT model~\cite{Gomes:2018eiv} with repulsive interactions. Results of our crossover EOSs tally with observational findings of the radii of $\sim\,1.4$ and $\sim\,2.0\,{\rm M}_{\odot}$ NSs~\cite{Miller:2019cac,Riley:2019yda,Miller:2021qha,Riley:2021pdl}.
Calculations of the equilibrium and adiabatic speeds of sound were performed following the procedures developed in Paper I. Our work here is focused on zero temperature and does not consider superfluidity in either nucleons or quarks. Inclusion of these effects will be taken up in a future work.
The results of the amplitudes of the $g$-mode~frequencies and their associated amplitudes of the gravitational energy radiated for the chosen hadron-to-quark crossover models lie between those of the first-order phase transitions that employ Maxwell and Gibbs constructions.
For the case of the Maxwell construction, the transition region is devoid of $g$-mode~frequencies as the equilibrium and adiabatic sound speeds both vanish. Consequently, $g$-mode~oscillations are permitted only for the pure nucleonic and quark phases. Their amplitudes in these regions are, however, rather small.
In contrast, the mixed phase in the case of the Gibbs construction yields amplitudes of $g$-mode~frequencies and the energy radiated that are significantly larger than those of the crossover model.
We note that $g$-mode~frequencies can be exceptionally large in the presence of superfluidity ($\approx$ 750 Hz for a hyperonic star~\cite{Dommes:2015wul} and $\approx$ 450 Hz for a nucleonic star~\cite{Gusakov:2013eoa,Passamonti:2015oia}), similar to the results for the Gibbs mixed phase. However, the reason for the enhancement in the two cases is different. In the case of superfluidity, the temperature-dependence of the entrainment terms serves to increase the $g$-mode~frequency at typical neutron star temperatures, while in the Gibbs case, the enhancement is purely composition-dependent.
A study of resonant excitations of such superfluid modes in coalescing neutron star binaries~\cite{Yu:2016ltf} suggests that the amplitude of these modes is weaker by a factor of 20 or so, compared to modes from the normal fluid. Note that the $g$-modes~we discuss here are also higher in frequency compared to the low-frequency $g$-modes~($\sim 50$ Hz) that might strain the neutron star crust to breaking point and lead to precursor flares in gamma-ray bursts~\cite{Kuan:2021jmk}.
The relatively larger frequency, amplitude and energy of the $g$-mode~in the Gibbs case inferred from
Figs.~\ref{fig:gmode-EOS}-\ref{fig:energy}
have observational implications for gravitational waves from neutron star mergers. It has been established that $g$-modes~can couple to tidal forces and draw energy and angular momentum from the binary to the neutron star, leading to an accelerated merger and a concomitant phase shift in the gravitational waveform~\cite{Lai:1993di}. This coupling will be largest for the Gibbs case, with its higher energy at resonance, and also because higher resonance frequencies are excited later in the inspiral,
when tidal forces are strongest. Estimates of the resulting phase shift were presented in Paper I (Eq. (89)) and found to be comparable to that from $g$-modes~in ordinary neutron stars (due to longer merger times) within uncertainties arising from the value of the tidal coupling. As these uncertainties are reduced through improved theoretical calculations, the case of a hybrid star may be distinguished from an ordinary neutron star.
We also infer from our results that were high-frequency $g$-modes~to be detected in upgraded LIGO and Virgo observatories, it would indicate a first-order phase transition akin to a Gibbs construction. In light of data from GW170817, lower bounds on the excitation of non-radial oscillations in binary mergers~\cite{Pratten:2019sed} affirm that a third generation network with its improved sensitivity and larger bandwidth can shed new light on the composition of the neutron star core.
\vspace{4mm}
\begin{acknowledgments}
C.C. acknowledges support from the European Union's Horizon 2020 research and innovation programme under the Marie Sk\l{}odowska-Curie grant agreement No. 754496 (H2020-MSCA-COFUND-2016 FELLINI).
S.H. is supported by the National Science Foundation, Grant PHY-1630782, and the Heising-Simons Foundation, Grant 2017-228.
P.J. is supported by the U.S. National Science Foundation Grant No. PHY-1913693.
M.P.'s research was supported by the Department of Energy, Grant No. DE-FG02-93ER40756.
\end{acknowledgments}
|
\section{Introduction}
In this paper we consider the following nonlinear iteration with window size $m$,
\begin{equation}\label{eq:AA-iteration}
x_{k+1}= q(x_k) + \sum_{i=1}^{\min(k,m)}\beta_{i}^{(k)}(q(x_k)-q(x_{k-i})), \qquad k=0,1,2,\ldots,
\end{equation}
which aims to accelerate fixed-point (FP) iterations of the type
\begin{equation}\label{eq:fixed-point}
x_{k+1}=q(x_{k}), \quad x_k\in\mathbb{R}^n, k=0,1,2,\ldots,
\end{equation}
with fixed point $x^*=q(x^*)$.
The coefficients $\beta_{i}^{(k)}$ in \cref{eq:AA-iteration} are determined by solving an optimization
problem in each step $k$ that minimizes a linearized residual in the new iterate $x_{k+1}$. Method
\cref{eq:AA-iteration} is known as \emph{Anderson acceleration (AA)} \cite{anderson1965iterative}.
Specifically, AA($m$), with window size $m$, solves in every iteration the optimization problem
\begin{equation}\label{eq:Andersonbetas}
\min_{\{\beta_i^{(k)} \}} \bigg\| r(x_k) + \sum_{i=1}^{\min(k,m)} \beta_i^{(k)} ( r(x_k) - r(x_{k-i})) \bigg\|^2
\end{equation}
with up to $m$ variables. Here, the residuals $r(x)$ of the fixed-point iteration are defined by
\begin{equation}\label{eq:resid}
r(x)=x-q(x).
\end{equation}
\begin{remark}
The $2$-norm is normally used in the optimization problem \cref{eq:Andersonbetas}. When $m=0$, \cref{eq:AA-iteration} is reduced to the fixed-point iteration \cref{eq:fixed-point}.
\end{remark}
We will also consider a version of AA($m$) iteration \cref{eq:AA-iteration} with more general initial guess:
\begin{equation}\label{eq:general-guess-AAm}
x_{k+1}= q(x_k) + \sum_{i=1}^{m}\beta_{i}^{(k)}(q(x_k)-q(x_{k-i})), \qquad k=m,m+1,\ldots,
\end{equation}
with initial guess $\{x_0,x_1,\ldots, x_m\}$.
Given $\{x_0,x_1,\ldots, x_m\}$, we denote $x_{m+1}$ generated by \cref{eq:general-guess-AAm} as
\begin{equation}\label{eq:x-m+1-from-AAj}
x_{m+1} =\mathcal{F}_m(\{x_j\}_{j=0}^m).
\end{equation}
Note that for AA($m$) in \cref{eq:AA-iteration} the initial guess $\{x_k\}_{k=0}^{m}$ is recursively generated by AA($j$), $j=0,1,\ldots,m-1$: given $x_0$,
\begin{equation*}
x_{k+1} =\mathcal{F}_k(\{x_j\}_{j=0}^k),\quad k=0,1,\ldots, m-1.
\end{equation*}
So we will sometimes refer to iteration \cref{eq:AA-iteration} as AA($m$) with recursive initial guess.
We will also consider AA($m$) with the initial guess fully determined by the fixed-point method:
\begin{equation*}
x_{k+1} =q(x_{k}),\quad k=0,1,\ldots,m-1.
\end{equation*}
This is less widely used, but we discuss it for completeness.
Assume that $k>m$. Define $r_k =x_k-q(x_k)$ and
\begin{equation}\label{eq:beta-vector-form}
\boldsymbol{\beta}^{(k)} =\begin{bmatrix} \beta_1^{(k)} \\ \vdots \\ \beta_m^{(k)} \end{bmatrix}, \quad
R_k = \begin{bmatrix} r_k-r_{k-1} & r_k-r_{k-2} & \ldots & r_k-r_{k-m} \end{bmatrix}.
\end{equation}
Then, using the 2-norm in \cref{eq:Andersonbetas}, the solution of the least-squares problem \cref{eq:Andersonbetas} is given by
\begin{equation}\label{eq:AAm-beta-form}
\boldsymbol{\beta}^{(k)} = -(R_k^TR_k)^{-1} R_k^Tr_k,
\end{equation}
if $ R_{k}^TR_k$ is invertible. Otherwise, we can take
\begin{equation}\label{eq:AAm-beta-form-pseudo}
\boldsymbol{\beta}^{(k)} = -R_k^{\dag}r_k,
\end{equation}
where $R_k^{\dag}$ is the pseudo-inverse of $R_k$, and $ \boldsymbol{\beta}^{(k)}$ corresponds to the minimum-norm solution of the least-squares problem. We note that
$R_k^{\dagger}= \big(R_k^TR_k\big)^{\dagger} R_k^T$.
Particularly, when $m=1$ and $r_k\neq r_{k-1}$,
\begin{equation}\label{eq:AA-1-step-beta}
\beta_1^{(k)} = \frac{-r_k^T(r_k-r_{k-1})}{\|r_k-r_{k-1}\|^2}=:\beta_k.
\end{equation}
When $r_k= r_{k-1}$, we take $\beta_k=0$ according to \cref{eq:AAm-beta-form-pseudo}.
Anderson acceleration dates back to the 1960s \cite{anderson1965iterative} and is widely used in computational science to accelerate iteration methods which converge slowly or do not converge. It has gained significant new interest over the past decade both in terms of theoretical developments and applications. Early on, most research focused on the application of AA without theoretical convergence analysis. For example, \cite{lott2012accelerated} examines the effectiveness of AA applied to modified Picard iteration for nonlinear problems arising in variably saturated flow modeling. Similarly, \cite{ho2017accelerating} applies AA to the Uzawa algorithm for the solution of saddle-point problems, and \cite{ni2009anderson} investigates the self-consistent field method accelerated by AA for electronic structure computations. For more applications, we refer to \cite{an2017anderson,brune2015composing,fang2009two,lipnikov2013anderson}. Anderson acceleration is also closely related to the nonlinear GMRES method from \cite{oosterlee2000krylov,sterck2012nonlinear,sterck2013steepest}.
It is only recently that the first results have been obtained on the convergence of AA. In \cite{walker2011anderson} it was shown that when $m>k$, AA($m$) applied to linear problems is essentially equivalent to GMRES. In \cite{toth2015} it was shown that AA($m$) is
locally $r$-linearly convergent under the assumptions that $q(x)$ is contractive and the AA coefficients remain bounded. However, this work does not prove that AA actually improves the convergence speed. Further progress was made in \cite{evans2020proof}, showing that, to first order, the convergence gain provided by AA in
step $k$ is quantified by a factor $\theta_k \leq 1$ that is the ratio of the square root of the
optimal value defined in \cref{eq:Andersonbetas} to $\| r(x_k)\|_2$.
However, it is not clear how this result may be used to quantify the asymptotic gain in convergence speed,
since $\theta_k$ does not appear to observe an upper bound $<1$ as $k$ increases.
In \cite{desterck2020,wang2020}, the authors consider stationary versions of the AA($m$) iteration where constant iteration coefficients $\beta_k$ are chosen in such a way that they minimize the $r$-linear convergence factor of the stationary AA($m$) iteration, given knowledge of $q'(x^*)$.
This provides insight into how the optimal stationary AA($m$) iteration improves the asymptotic convergence of the FP method by reducing the spectral radius of $q'(x^*)$, where $q(x)$ can be interpreted as a nonlinear preconditioner for AA($m$), see also \cite{brune2015composing}.
In this paper, we focus on AA($m$) applied to the linear case, that is,
\begin{equation}\label{eq:linear-fixed-point}
q(x) = Mx + b,
\end{equation}
where the fixed point satisfies $Ax^*=b$ with $A =I-M$. We will assume that $A$ is nonsingular and we exclude the trivial case that $A=I$ and $M=0$.
We are interested in exploring polynomial update formulas for the residual of AA($m$). Recall that the order-s Krylov subspace generated by a matrix $T$ and a vector $v$ is the linear subspace spanned by the images of $v$ under the first $s$ powers of $T$, that is
\begin{equation*}
\mathcal{K}_s (T,v) =\Big\{v, Tv, \ldots, T^{s-1} v\Big\}.
\end{equation*}
The specific case of AA(1) in \cref{eq:AA-iteration} reads
\begin{equation}\label{eq:anderson-1-step}
x_{k+1} = (1+\beta_k) q(x_k) -\beta_k q(x_{k-1}).
\end{equation}
In previous work, \cite{kindermann2021optimal,liu2018parametrized,niu2020momentum} have interpreted Nesterov acceleration, which is a form of AA(1) with a prescribed sequence of acceleration coefficients $\beta_k$, as a Krylov method. Inspired by this, we investigate in this paper how AA($m$) for linear problems, with $\boldsymbol{\beta}^{(k)}$ given by \cref{eq:AAm-beta-form}, relates to Krylov methods. Following \cite{kindermann2021optimal,liu2018parametrized,niu2020momentum}, this is easy to see for AA(1) applied to \cref{eq:linear-fixed-point}, as we now explain. Given $x_0$, let $ x_1=q(x_0)$. The residual $r_{k+1}$ generated by AA(1) iteration \cref{eq:anderson-1-step} satisfies
\begin{align}
r_{k+1}&= x_{k+1}-(Mx_{k+1}+b)= Ax_{k+1}-b,\nonumber\\
& = A\big((1+\beta_k) (Mx_k+b) -\beta_k(Mx_{k-1}+b)\big)-b,\nonumber\\
&= (1+ \beta_k)AMx_k- \beta_kAMx_{k-1}+Ab-b,\nonumber \\
&=(1+ \beta_k)M(Ax_k-b) - \beta_k M(Ax_{k-1}-b)+Mb+Ab-b,\nonumber\\
&=(1+ \beta_k)Mr_k - \beta_k M r_{k-1}.\label{pro:AA1-rk+1=Mg_k}
\end{align}
We list the following expressions for the first few $r_k$:
\begin{align*}
r_1&=M r_0,\\
r_2&=\big((1+\beta_1)M^2-\beta_1 M\big) r_0,\\
r_3&=\big((1+\beta_2)(1+\beta_1)M^3-((1+\beta_2)\beta_1+\beta_2) M^2\big)r_0,\\
r_4&=\big((1+\beta_3)(1+\beta_2)(1+\beta_1)M^4-((1+\beta_3)(1+\beta_2)\beta_1\\
&\quad +(1+\beta_3)\beta_2+\beta_3(1+\beta_1)) M^3+\beta_3\beta_1M^2\big)r_0.
\end{align*}
Clearly, AA(1) is a Krylov space method. We will investigate in this paper how this extends to AA($m$) and allows to derive new properties of the AA($m$) iteration.
The rest of this paper is organized as follows. In \Cref{sec:AAm-krylov}, we derive polynomial residual update formulas for AA($m$), showing that AA($m$) is a Krylov method or a multi-Krylov method, depending on the initial condition $\{x_0, \ldots, x_m\}$. We derive several further results based on these polynomial residual update formulas, including orthogonality relations. In \Cref{sec:AA(1)-Krylov}, we focus specifically on AA(1). We obtain a lower bound on the AA(1) acceleration coefficient and explicit nonlinear recursions for the AA(1) residuals and residual polynomials. Moreover, we explore the influence of the initial guess on the convergence speed of AA(1). Some numerical results are presented in \Cref{sec:Numerical-result}. Finally, we draw conclusions in \Cref{sec:con}.
\section{AA($m$) as a multi-Krylov space method}\label{sec:AAm-krylov}
In this section, we consider AA($m$) with $m>1$ applied to linear problems and establish links with Krylov methods. We will discuss some specific aspects of the case that $m=1$ in \Cref{sec:AA(1)-Krylov}. It has been known for a long time that AA($m$) is essentially equivalent to GMRES when the window size $m$ is taken as infinite \cite{oosterlee2000krylov,walker2011anderson}. When this is not the case, AA($m$) is
essentially a windowed version of restarted GMRES($m$). This section establishes results on AA($m$) with finite $m$ viewed as a Krylov space method.
The AA($m$) iteration with general initial guess from \cref{eq:general-guess-AAm} can be rewritten as
\begin{equation}\label{eq:AA(m)-rewrite-form}
x_{k+1}=\Big(1+ \sum_{i=1}^{m}\beta_i^{(k)}\Big)q(x_k)- \sum_{i=1}^{m}\beta_i^{(k)}q(x_{k-i}),\quad k=m,m+1,\ldots\,\, .
\end{equation}
In the linear case, we obtain
\begin{equation}\label{eq:xk-matrix-form-AAm}
x_{k+1}=\Big(1+ \sum_{i=1}^{m}\beta_i^{(k)}\Big)Mx_k- \sum_{i=1}^{m}\beta_i^{(k)}Mx_{k-i}+b,\quad k=m,m+1,\ldots\,\, .
\end{equation}
\subsection{ AA($m$) with general initial guess}\label{subsec:AAm-random-guesses}
In this subsection, we consider AA($m$) iteration \cref{eq:general-guess-AAm} with general initial guess $\{x_k\}_{k=0}^{m}$.
\begin{proposition}\label{prop:rk-recursive-AAm-random-guess}
The residuals $r_{k+1}$ generated by AA($m$) iteration \cref{eq:general-guess-AAm} with general initial guess $\{x_k\}_{k=0}^{m}$ applied to linear iteration \cref{eq:linear-fixed-point} satisfy
\begin{equation}\label{eq:m+1-term-rk-linear-random-guess}
r_{k+1} = \Big(1+ \sum_{i=1}^{m}\beta_i^{(k)}\Big)M r_k- \sum_{i=1}^{m}\beta_i^{(k)} M r_{k-i},\quad k\geq m.
\end{equation}
\end{proposition}
\begin{proof}
From \cref{eq:xk-matrix-form-AAm} and $AM=MA$ we have, for $k\geq m$,
\begin{align*}
Ax_{k+1} &= \Big(1+ \sum_{i=1}^{m}\beta_i^{(k)}\Big)AMx_k- \sum_{i=1}^{m}\beta_i^{(k)}AMx_{k-i}+Ab,\\
&=\Big(1+ \sum_{i=1}^{m}\beta_i^{(k)}\Big)MAx_k- \sum_{i=1}^{m}\beta_i^{(k)}MAx_{k-i}+Ab,\\
&=\Big(1+ \sum_{i=1}^{m}\beta_i^{(k)}\Big)M(Ax_k-b) - \sum_{i=1}^{m}\beta_i^{(k)}M(Ax_{k-i}-b)+Mb+Ab,\\
&=\Big(1+ \sum_{i=1}^{m}\beta_i^{(k)}\Big)Mr_k - \sum_{i=1}^{m}\beta_i^{(k)}M r_{k-i}+ b.
\end{align*}
%
Thus, $r_{k+1}=Ax_{k+1}-b=\displaystyle \Big(1+ \sum_{i=1}^{m}\beta_i^{(k)}\Big)Mr_k - \sum_{i=1}^{m}\beta_i^{(k)}Mr_{k-i}$.
\end{proof}
Based on \cref{prop:rk-recursive-AAm-random-guess}, we can write the residual $r_k$ in terms of the initial residuals $\{r_j\}_{j=0}^m$:
\begin{proposition}\label{pro:polynomial-form-AAm-random-guess}
AA($m$) iteration \cref{eq:general-guess-AAm} with general initial guess $\{x_j\}_{j=0}^{m}$ applied to linear iteration \cref{eq:linear-fixed-point} is a multi-Krylov method. That is, the residual can be expressed as
\begin{equation}\label{eq:rk-polynomial-form-M-AAm-random-guess}
r_{k+1} = \sum_{j=0}^m p_{k-m+1,j}(M)\,r_j,\quad k\geq m,
\end{equation}
where the $p_{k-m+1,j}(\lambda)$ are polynomials of degree at most $k-m+1$ satisfying the following relations:
\begin{align}
p_{1,j}(\lambda)&=-\beta^{(m)}_{m-j}\lambda, \quad j=0,\ldots,m-1; \qquad p_{1,m}(\lambda)=\Big(1+\sum_{i=1}^{m}\beta_i^{(m)}\Big)\lambda; \label{AAm-AAj-eq1} \\
p_{k-m+1,j}(\lambda)&=\lambda \left(\Big(1+\sum_{i=1}^{m}\beta_i^{(k)}\Big)p_{k-m,j} - \sum_{i=1}^{m}\beta_i^{(k)} p_{k-m-i,j}\right), \quad k-m+1>1, j=0,\ldots,m, \label{AAm-AAj-eq3}
\end{align}
where for $i=1,\ldots,m,$ and $j=0,\ldots, m$,
\begin{equation*}
p_{1-i,j}(\lambda) = \begin{cases}
1 & \text{if}\,\, i=j+1-m,\\
0 & \text{otherwise}.
\end{cases}
\end{equation*}
\end{proposition}
\begin{proof}
The results of \cref{AAm-AAj-eq1} are obvious from \cref{eq:m+1-term-rk-linear-random-guess}. For \cref{AAm-AAj-eq3}, when $k\geq 2m+1$, from \cref{eq:m+1-term-rk-linear-random-guess}, $r_{k+1}$ is a linear combination of $\{r_{k+1-i}\}_{i=1}^{m+1}$ where the smallest subscript index in the residual is $k-m\geq m+1$. Thus, every term $r_{k+1-j}$ can be rewritten as a linear combination of $\{r_j\}_{j=0}^m$. Then, \cref{AAm-AAj-eq3} can be validated easily. As to $k< 2m+1$, since some terms in $\{r_{k+1-i}\}_{i=1}^{m+1}$ do not contain all $\{r_j\}_{j=0}^m$, we require that $p_{1-i,j}(\lambda)=0$ or 1 for \cref{AAm-AAj-eq3} to hold.
\end{proof}
\begin{remark}\label{rmk:multilevel-krylov-method}
Expression \cref{eq:rk-polynomial-form-M-AAm-random-guess} indicates that the residual $r_{k+1}$ with $k\geq m$ of AA($m$) can be decomposed as a sum of $m+1$ vectors which are in $m+1$ Krylov spaces, $\{ \mathcal{K}_s(M,r_j)\}_{j=0}^m$ or $\{ \mathcal{K}_s(A,r_j)\}_{j=0}^m$, where $s=k-m+2$. Therefore, we refer to AA($m$) with general initial guess as a {\em{multi-Krylov space method}}. Note that if each initial guess $r_j\in \{r_j\}_{j=1}^m$ can be expressed as a polynomial in $M$ or $A$ applied to $r_0$, then AA($m$) is a Krylov space method, that is, $r_{k+1}\in \mathcal{K}_s (M,r_0)$, as discussed in the next subsections.
\end{remark}
\begin{remark}
In \cref{pro:polynomial-form-AAm-random-guess}, we can also rewrite the residual in terms of polynomials in the matrix $A$ as
\begin{equation}\label{eq:rk-polynomial-form-A-AAm-random-guess}
r_{k+1} = \sum_{j=0}^m\widetilde{ p}_{k-m+1,j}(A)\,r_j,\quad k\geq m,
\end{equation}
%
where the $\widetilde{p}_{k-m+1,j}(\lambda)$ satisfy the relations:
\begin{align*}
\widetilde{p}_{1,j}(\lambda)&=-\sum_{i=1}^{m}\beta_i^{(m)}(1-\lambda), \,\, j=0,\ldots,m-1; \qquad \widetilde{p}_{1,m}(\lambda)=\Big(1+\sum_{i=1}^{m}\beta_i^{(m)}\Big)(1-\lambda); \\
\widetilde{p}_{k-m+1,j}(\lambda)&=(1-\lambda)\left(\Big(1+\sum_{i=1}^{m}\beta_i^{(k)}\Big) \widetilde{p}_{k-m,j} - \sum_{i=1}^{m}\beta_i^{(k)} \widetilde{p}_{k-m-i,j}\right), \quad k-m+1>1, j=0,\ldots, m,
\end{align*}
where for $i=1,\ldots,m, j=0,\ldots, m$,
\begin{equation*}
\widetilde{p}_{1-i,j}(\lambda) = \begin{cases}
1 & \text{if}\,\, i=j+1-m,\\
0 & \text{otherwise}.
\end{cases}
\end{equation*}
\end{remark}
Using \cref{pro:polynomial-form-AAm-random-guess}, it can be shown easily that there is a periodic pattern with period $m+1$ in the AA($m$) polynomials:
\begin{proposition}\label{pro:AAm-polynomial-arbitrary-initial}
The residuals generated by AA($m$) iteration \cref{eq:general-guess-AAm} with general initial guess $\{x_j\}_{j=0}^{m}$ applied to linear iteration \cref{eq:linear-fixed-point} can be written as
\begin{equation}\label{eq:rk-polynomial-form-M-AAm-random-guess-power}
r_{s(m+1)+i} =M^{s} \sum_{j=0}^m g_{s(m+1)+i-m-s,j}(M)\,r_j,\quad s\geq 1, i=0,\ldots, m,
\end{equation}
where $g_{s(m+1)+i-m-s,j}(\lambda)$ is a polynomial of degree at most $s(m+1)+i-m-s$ and $\lambda^s g_{s(m+1)+i-m-s,j}(\lambda)=p_{s(m+1)+i-m,j}(\lambda)$ from \cref{eq:rk-polynomial-form-M-AAm-random-guess}.
\end{proposition}
Expression \cref{eq:rk-polynomial-form-M-AAm-random-guess-power} indicates that every $m+1$ iterations, the power of $M$ in the right-hand side of \cref{eq:rk-polynomial-form-M-AAm-random-guess-power} increases by 1.
We refer to this property as the AA($m$) iterations possessing a periodic \emph{memory effect}.
Expression \cref{eq:rk-polynomial-form-M-AAm-random-guess-power}, thus, reveals that AA($m$) provides acceleration as a result of two multiplicative effects: damping of error modes by $M^{s}$ as in the FP method, augmented by polynomial acceleration.
As a result of the windowing in AA($m$), the effect of the FP iteration is, thus, retained in an accumulative fashion as $k$ increases.
This is in contrast to restarted GMRES($m$), where there is no such cumulative damping since the iteration is fully restarted every $m$ steps.
Next, we show an orthogonality property of the residual $r_{k+1}$ with respect to $R_k$ of \cref{eq:beta-vector-form}.
\begin{theorem}\label{AAm-rk-M-orthogonal}
Assume that $M$ is invertible. Then the residuals $r_{k+1}$ generated by AA($m$) iteration \cref{eq:general-guess-AAm} with general initial guess $\{x_j\}_{j=0}^{m}$ applied to linear iteration \cref{eq:linear-fixed-point} satisfy
\begin{equation*}
R_k^T M^{-1}r_{k+1}=0,
\end{equation*}
where $R_k$ is defined in \cref{eq:beta-vector-form}.
\end{theorem}
\begin{proof}
Recall $\boldsymbol{\beta}^{(k)}$ in \cref{eq:AAm-beta-form-pseudo}. We have
\begin{equation*}
(R_k^TR_k)\boldsymbol{\beta}^{(k)} + R_k^Tr_k=0,
\end{equation*}
which leads to
\begin{equation*}
R_k^T(R_k\boldsymbol{\beta}^{(k)}+r_k)=0,
\end{equation*}
or, by \cref{eq:beta-vector-form},
\begin{equation}\label{eq:rk-Rk-dot-product}
R_k^T \left( \Big(1+ \sum_{i=1}^{m}\beta_i^{(k)}\Big)r_k - \sum_{i=1}^{m}\beta_i^{(k)}r_{k-i}\right )=0.
\end{equation}
Using \cref{prop:rk-recursive-AAm-random-guess} this gives
\begin{equation*}
R_k^T M^{-1}M\left(\Big(1+ \sum_{i=1}^{m}\beta_i^{(k)}\Big)r_k - \sum_{i=1}^{m}\beta_i^{(k)}r_{k-i}\right)=R_k^T M^{-1} r_{k+1}=0,
\end{equation*}
which completes the proof.
\end{proof}
Next we show how the residuals $r_{k+1}$ are related to the projector operator $R_kR_k^{\dag}$, as expected in the context of least-squares problem \cref{eq:Andersonbetas}.
\begin{proposition}\label{pros-AAm-residual-A-random}
The residuals $r_{k+1}$ generated by AA($m$) iteration \cref{eq:general-guess-AAm} with general initial guess $\{x_k\}_{k=0}^{m}$ applied to linear iteration \cref{eq:linear-fixed-point} satisfy the following:
If $R_k^T R_k$ is invertible, then
\begin{equation*}
r_{k+1} =M\left(I-R_k(R_k^T R_k)^{-1}R_k^T\right)r_k, \, k\geq m.
\end{equation*}
%
More generally,
\begin{equation}\label{eq:prop-rk-based-on-projection}
r_{k+1} =M(I-R_kR_k^{\dag})r_k, \, k\geq m.
\end{equation}
Furthermore,
\begin{equation}\label{eq:prop-orthognal-Rk}
R_k^T(I-R_kR_k^{\dag})r_k =0, \, k\geq m.
\end{equation}
\end{proposition}
\begin{proof}
From \cref{eq:AAm-beta-form-pseudo} and \cref{eq:m+1-term-rk-linear-random-guess} we have
\begin{align*}
r_{k+1} & = M r_k + \sum_{i=1}^{m}\beta_i^{(k)} M (r_k-r_{k-i}),\\
&= Mr_k + MR_k \boldsymbol{\beta}^{(k)}, \\
& = M(I-R_kR_k^{\dag})r_k.
\end{align*}
Expression \cref{eq:prop-orthognal-Rk} is then obtained from \cref{eq:prop-rk-based-on-projection} using \cref{AAm-rk-M-orthogonal}.
\end{proof}
\subsection{AA($m$) with initial guess generated by AA($j$)}\label{subsec:AAm-recursive-guess}
In this subsection, we consider the classical AA($m$) iteration \cref{eq:AA-iteration} where the $m+1$ initial guesses are generated by AA($j$), that is, given $x_0$, $\{x_{j+1}=\mathcal{F}_j(\{x_i\}_{i=0}^j)\}_{j=0}^{m-1}$, see \cref{eq:x-m+1-from-AAj}. For this choice of recursive initial guess,
residual expressions in \Cref{subsec:AAm-random-guesses} can be written as a function of $r_0$ only.
We start with specializing \cref{prop:rk-recursive-AAm-random-guess} and \cref{pro:polynomial-form-AAm-random-guess} for this recursive initial guess.
\begin{proposition}\label{pro:polynomial-form-AAm-AAj}
AA(m) iteration \cref{eq:AA-iteration} with recursive initial guess $\{x_{j+1}=\mathcal{F}_j(\{x_i\}_{i=0}^j)\}_{j=0}^{m-1}$ applied to linear iteration \cref{eq:linear-fixed-point} is a Krylov method, with the residuals given by
\begin{equation}\label{eq:rk-polynomial-form-M-AAm-AAj}
r_{k+1} =p_{k+1}(M)\,r_0,\quad k\geq 0,
\end{equation}
where $p_{k+1}(\lambda)$ is a polynomial of degree at most $k+1$, satisfying the following recurrence relations:
\noindent When k=0,
\begin{equation*}
p_1(\lambda) =\lambda.
\end{equation*}
When $1\leq k< m$,
\begin{equation}\label{polynomial-sequence-AAj}
p_{k+1}(\lambda)=\Big(1+\sum_{i=1}^{k}\beta_i^{(k)}\Big)\lambda p_{k}(\lambda) - \sum_{i=1}^{k}\beta_i^{(k)} \lambda p_{k-i}(\lambda) \,\,\,
\text{where}\,\,\, p_0(\lambda)=1.
\end{equation}
When $k\geq m$,
\begin{equation*}
p_{k+1}(\lambda)=\Big(1+\sum_{i=1}^{m}\beta_i^{(k)}\Big)\lambda p_{k}(\lambda) - \sum_{i=1}^{m}\beta_i^{(k)} \lambda p_{k-i}(\lambda).
\end{equation*}
Moreover, $p_k(1)=1$ and $p_k(0)= 0, k=1,2,\ldots \,\, .$
\end{proposition}
Similar to \cref{pro:AAm-polynomial-arbitrary-initial}, we can further factor the above result as follows, specializing the above-mentioned memory effect to the case of AA($m$) with recursive initial guess.
\begin{proposition}\label{pro:AAm-polynomial-power-initial-AAj}
The residuals of AA(m) iteration \cref{eq:AA-iteration} with recursive initial guess $\{x_{j+1}=\mathcal{F}_j(\{x_i\}_{i=0}^j)\}_{j=0}^{m-1}$ applied to linear iteration \cref{eq:linear-fixed-point} satisfy
\begin{equation*}
r_{s(m+1)+i}=M^{s+1} g_{s(m+1)+i-(s+1)}(M)\,r_0, \quad s=0,1,2,\ldots, \quad i =1,\ldots, m+1,
\end{equation*}
where $g_{s(m+1)+i-(s+1)}(\lambda)$ is a polynomial of degree at most $s(m+1)+i-(s+1)$ and $\lambda^{s+1} g_{s(m+1)+i-(s+1)}(\lambda)=p_{s(m+1)+i}(\lambda)$ from \cref{eq:rk-polynomial-form-M-AAm-AAj}.
\end{proposition}
Next, the following result follows directly from \cref{eq:rk-polynomial-form-M-AAm-AAj} in \cref{pro:polynomial-form-AAm-AAj}, since GMRES determines the optimal degree-$k$ polynomial with $p_k(1)=1$.
\begin{proposition}\label{pro:k-AAm-worse-GMRES}
When applied to linear iteration \cref{eq:linear-fixed-point},
$k$ steps of AA($m$) cannot produce a residual that is smaller in the 2-norm than the residual obtained by GMRES($k$) applied to
the corresponding linear system.
\end{proposition}
\begin{remark}
\cref{pro:k-AAm-worse-GMRES} does not imply anything about the relative convergence speed of (windowed) AA($m$) versus restarted GMRES($m$) in the linear case. To our knowledge, no general theoretical results exist on this topic. Note that it can be expected that AA($m$) would often converge faster than GMRES($m$), since AA($m$) uses $m+1$ previous iterates in each step, while GMRES($m$) uses about half that number per step on average. Some numerical results in the literature confirm this, see, e.g., Figure 9 in \cite{LinearacAA}.
\end{remark}
\subsection{AA($m$) with initial guess generated by fixed-point iteration}\label{AA(m)-properties-fixed-initial-guess}
In this subsection, we consider the version of AA($m$) where the initial $m$ iterates are generated by the fixed-point method.
Note that, if we set $\beta_i^{(k)}=0$ in \cref{polynomial-sequence-AAj}, that is, the initial guesses are $\{x_{k+1}=q(x_{k})\}_{k=0}^{m-1}$, then \cref{pro:polynomial-form-AAm-AAj} leads to the following result.
\begin{proposition}\label{pro:polynomial-form-AAm}
AA($m$) iteration \cref{eq:general-guess-AAm} with fixed-point initial guess $\{x_{k+1}=q(x_{k})\}_{k=0}^{m-1}$ applied to linear iteration \cref{eq:linear-fixed-point} is a Krylov method. That is, the residuals satisfy
\begin{equation}\label{eq:rk-polynomial-form-AAm}
r_{k+1} = \widetilde{p}_{k+1}(M)\,r_0,
\end{equation}
where $\widetilde{p}_{k+1}(\lambda)$ is a polynomial of degree at most $k+1$ satisfying the recurrence relation:
When $0 \leq k < m$, $\widetilde{p}_{k+1}(\lambda) =\lambda^{k+1}$.
When $k \geq m$,
\begin{equation}\label{eq:AAm-poly-three-M}
\widetilde{p}_{k+1}(\lambda)=\Big(1+\sum_{i=1}^{m}\beta_i^{(k)}\Big)\lambda\widetilde{p}_k(\lambda) - \sum_{i=1}^{m}\beta_i^{(k)} \lambda \widetilde{p}_{k-i}(\lambda), \quad \widetilde{p}_0(\lambda)=1.
\end{equation}
Moreover, $\widetilde{p}_k(1)=1$ and $\widetilde{p}_k(0)= 0$.
\end{proposition}
Similar to \cref{pro:AAm-polynomial-arbitrary-initial}, we can further factor the result in \cref{pro:polynomial-form-AAm} as follows.
\begin{proposition}\label{pro:polynomial-form2-AAm}
The residuals of AA($m$) iteration \cref{eq:general-guess-AAm} with fixed-point initial guess $\{x_{k+1}=q(x_{k})\}_{k=0}^{m-1}$ applied to linear iteration \cref{eq:linear-fixed-point} satisfy
\begin{equation*}
r_{s(m+1)+i}=M^{s+1} \widehat{p}_{s(m+1)+i-(s+1)}(M)\,r_0, \quad s=0,1,2,\ldots, \quad i =1,\ldots, m+1,
\end{equation*}
where $\widehat{p}_{s(m+1)+i-(s+1)}(\lambda)$ is a polynomial with degree at most $s(m+1)+i-(s+1)$ and $\lambda^{s+1} \widehat{p}_{s(m+1)+i-(s+1)}(\lambda)=\widetilde{p}_{s(m+1)+i}(\lambda)$ from \cref{eq:rk-polynomial-form-AAm}.
\end{proposition}
\section{Specific results for AA(1)}\label{sec:AA(1)-Krylov}
In this section, we consider AA(1) with different choices for initial guess.
Building on the properties from \Cref{sec:AAm-krylov}, we obtain some additional properties that are special for AA(1).
\subsection{AA(1) with general initial guess}
We first give a result for the coefficients $\beta_k$ in AA(1).
\begin{theorem}\label{thm:lower-boound-betak}
Let $x_0$ and $x_1$ be initial guesses with $||r_1||\leq ||r_0||$. For AA(1) iteration \cref{eq:anderson-1-step}
applied to linear fixed-point function $q(x)=Mx+b$ with $\|M\|\leq 1$,
the coefficients $\beta_k$ satisfy
\begin{equation*}
\beta_k>-1, \quad k=1,2,\ldots\,\, .
\end{equation*}
\end{theorem}
\begin{proof}
Recall that
\begin{equation*}
\beta_k = \frac{-r_k^T(r_k-r_{k-1})}{(r_k-r_{k-1})^T(r_k-r_{k-1})},
\end{equation*}
if $r_k\neq r_{k-1}$, and $\beta_k=0$ otherwise.
Let $w_k= r_k-r_{k-1}$. Then, $\beta_k ||w_k||^2 =-r_k^Tw_k$, and
\begin{equation}\label{rk-wk-relation}
r_{k-1}+w_k =r_k.
\end{equation}
It follows that
\begin{equation*}
-\beta_k||w_k||^2 =(r_k,w_k)= (r_{k-1},w_k)+(w_k,w_k),
\end{equation*}
which means that
\begin{equation}\label{eq:pf-inner-product}
(r_{k-1}, w_k) =(-1-\beta_k)||w_k||^2.
\end{equation}
Next, we use induction to complete the proof. Consider $k=1$ in \cref{eq:pf-inner-product}. We have $\|r_1\|\leq \|r_0\|$ and it follows from \cref{rk-wk-relation} that
\begin{align*}
(r_0, w_1) &=r_0^T(r_1-r_0) = ||r_0|| \,||r_1||\cos\theta-||r_0||^2=||r_0||\big(||r_1||\cos\theta-||r_0||\big).
\end{align*}
Thus, $(r_{0}, w_1)<0$ if $r_0\neq r_1$, and $(r_0,w_1)=0$ otherwise. If $r_0\neq r_1$, \cref{eq:pf-inner-product} shows that $-1-\beta_1<0$, i.e., $\beta_1>-1$. This also holds when $r_0=r_1$, since then $\beta_1=0$.
We will now show that $\beta_{n-1}>-1$ implies $\beta_n>-1$. According to the definition of the least-squares problem \cref{eq:Andersonbetas} for AA(1), $\beta_{n-1}$ satisfies
\begin{equation*}
\|r_{n-1} + \beta_{n-1}(r_{n-1}-r_{n-2})\|\leq \|r_{n-1}\|.
\end{equation*}
For the linear case where $q(x)=Mx+b$, we have from \cref{pro:AA1-rk+1=Mg_k}
\begin{equation*}
r_n = M\big(r_{n-1} + \beta_{n-1}(r_{n-1}-r_{n-2})\big).
\end{equation*}
Thus, we have $||r_n||\leq ||r_{n-1}||$ if $\|M\|\leq 1$. Now, if we consider $k=n$ in \cref{eq:pf-inner-product}, from \cref{rk-wk-relation} and $||r_n||\leq ||r_{n-1}||$ it follows as above that $(r_{n-1}, w_n)<0$ if $r_{n-1}\neq r_n$ and $(r_{n-1},w_n)=0$ otherwise. If $r_{n-1}\neq r_n$, \cref{eq:pf-inner-product} shows that $-1-\beta_n<0$. If $r_{n-1}=r_n$, $\beta_n=0$. Therefore, we have $\beta_n>-1$.
\end{proof}
\begin{remark}
It is easy to see that \cref{thm:lower-boound-betak} applies to the traditional AA(1) iteration \cref{eq:AA-iteration} with recursive initial guess, since then $x_1=Mx_0+b$ and $r_1=Mr_0$, so $\|r_1\|\leq \|r_0\|$ if $\|M\|\leq 1$.
\end{remark}
The following result
establishes an explicit nonlinear three-term recurrence for the AA(1) residuals that does not include $\beta_k$.
\begin{proposition}\label{AA1-rk-simple-form-randomx0x1}
The residuals $r_{k+1}$ generated by AA(1) iteration \cref{eq:general-guess-AAm} with general initial guess $\{x_j\}_{j=0}^{1}$ applied to linear iteration \cref{eq:linear-fixed-point} satisfy
\begin{itemize}
\item if $r_k\neq r_{k-1}$,
\begin{equation*}
r_{k+1} =\frac{1}{(r_k-r_{k-1})^T(r_k-r_{k-1})} M\left(-r_k r_{k-1}^T + (r_k r_{k-1}^T)^T\right)(r_k-r_{k-1}), \, k\geq 1.
\end{equation*}
%
\item if $r_k=r_{k-1}$, $r_{k+1}=Mr_k$.
\end{itemize}
\end{proposition}
\begin{proof}
We first consider the case that $r_k\neq r_{k-1}$, so
\begin{equation*}
\beta_k = \frac{-r_k^T(r_k- r_{k-1})}{(r_k-r_{k-1})^T(r_k-r_{k-1})}.
\end{equation*}
From \cref{eq:m+1-term-rk-linear-random-guess} with $m=1$, we have
\begin{align*}
r_{k+1} &=(1+\beta_k) Mr_k -\beta_k Mr_{k-1}, \\
&= \frac{-r_{k-1}^T(r_k- r_{k-1})}{(r_k-r_{k-1})^T(r_k-r_{k-1})} Mr_k+\frac{r_k^T(r_k- r_{k-1})}{(r_k-r_{k-1})^T(r_k-r_{k-1})} Mr_{k-1},\\
&=\frac{1}{(r_k-r_{k-1})^T(r_k-r_{k-1})}M(-r_k r_{k-1}^T + r_{k-1}r_k^T)(r_k-r_{k-1}),\\
&=\frac{1}{(r_k-r_{k-1})^T(r_k-r_{k-1})}M\big(-r_k r_{k-1}^T + (r_{k}r_{k-1}^T)^T\big)(r_k-r_{k-1}).
\end{align*}
If $r_{k}=r_{k-1}$, then $\beta_k=0$. So from \cref{eq:m+1-term-rk-linear-random-guess} we have $r_{k+1}=Mr_k$.
\end{proof}
\subsection{AA(1) with $x_1=q(x_0)$}
For AA(1), the case of recursive initial guess discussed in \Cref{subsec:AAm-recursive-guess} coincides with the case of fixed-point initial guess discussed in \Cref{AA(m)-properties-fixed-initial-guess} since, in both cases, $x_1=q(x_0)$.
For this case, it is useful to specialize some results from \Cref{subsec:AAm-recursive-guess} for future use.
First, \cref{pro:polynomial-form-AAm-AAj} for AA(1) is specialized as follows.
\begin{proposition}\label{pro:polynomial-form2}
AA(1) iteration \cref{eq:general-guess-AAm} with initial guess $x_1=q(x_0)$ applied to linear iteration \cref{eq:linear-fixed-point} is a Krylov method, with the residuals given by
\begin{equation}\label{eq:rk-polynomial-form}
r_{k+1} = p_{k+1}(M)\,r_0,
\end{equation}
where the residual polynomials satisfy the recurrence relation
\begin{equation}\label{eq:AA1-three-term-M}
p_{k+1}(\lambda)=(1+\beta_k)\lambda p_k(\lambda) - \beta_k \lambda p_{k-1}(\lambda), \quad k\geq 1,
\end{equation}
and $p_0(\lambda) =1, p_1(\lambda) =\lambda$. Moreover, $p_k(\lambda)$ is a polynomial with degree at most $k$ and $p_k(1)=1$ and $p_k(0)= 0$ for $k\geq 1$.
\end{proposition}
Next, \cref{pro:AAm-polynomial-power-initial-AAj} is reduced to the following result for AA(1).
\begin{proposition}\label{pro:polynomial-form3}
The residuals of AA(1) iteration \cref{eq:general-guess-AAm} with initial guess $x_1=q(x_0)$ applied to linear iteration \cref{eq:linear-fixed-point} satisfy
\begin{equation*}
r_{2s+1} =M^{s+1}\widehat{p}_{s}(M)\,r_0, \quad r_{2s+2}=M^{s+1} \widehat{p}_{s+1}(M)\,r_0,
\end{equation*}
where $\widehat{p}_{s+i-1}(\lambda), i=1,2$ is a polynomial with degree at most $s+i-1$.
\end{proposition}
\cref{pro:polynomial-form3} shows that the AA(1) residual polynomial gains a power of $M$ every two iterations.
The following property is special for AA(1), where the residuals are expressed not in terms of $\beta_k$ but directly using multivariate matrix polynomials $L_k$ in $M$, $M^T$ and $R_0=r_0r_0^T$, and with the $L_k$ satisfying a nonlinear three-term recursion.
\begin{proposition}
Let $r_k$ be the residual of AA(1) iteration \cref{eq:general-guess-AAm} with initial guess $x_1=q(x_0)$ applied to linear iteration \cref{eq:linear-fixed-point}. Assume that $r_k\neq r_{k-1}$ for all $k$. Then
\begin{equation}\label{eq:rk-Lk-form}
r_{k} =L_k r_0, \quad k=0, 1, \ldots
\end{equation}
with $L_0=I, L_1= M$ and
\begin{equation}\label{eq:Lk-form}
L_{k+1} = \frac{1}{\|(L_k-L_{k-1})r_0\|^2} M\left(-L_kR_0 L_{k-1}^T+ (L_kR_0 L_{k-1}^T)^T \right) (L_k-L_{k-1}), \quad k\geq 1,
\end{equation}
where $R_0= r_0r_0^T$.
\end{proposition}
\begin{proof}
Assume that when $k\leq n$, \cref{eq:rk-Lk-form} is true with $L_k$ defined by \cref{eq:Lk-form}. We show that \cref{eq:rk-Lk-form} remains valid for $k=n+1$ and $L_{k+1}$ given by \cref{eq:Lk-form}. Using the equation for $L_{k+1}$ in \cref{eq:Lk-form}, we have
\begin{align*}
L_{n+1}r_0 &= \frac{1}{\|(L_n-L_{n-1})r_0\|^2} M\left(-L_nR_0 L_{n-1}^T+ (L_nR_0 L_{n-1}^T)^T \right) (L_n-L_{n-1})r_0, \\
&= \frac{1}{\|r_n-r_{n-1}\|^2} M\left(-r_n r_{n-1}^T+ (r_n r_{n-1}^T)^T \right) (r_n-r_{n-1}),\\
& =r_{n+1},
\end{align*}
by \cref{AA1-rk-simple-form-randomx0x1}. This completes the proof.
\end{proof}
We list the first few $L_k$ in the following:
\begin{itemize}
\item $L_0 =I$.
\item $L_1= M$.
\item $L_2 = \displaystyle \frac{1}{\|(M-I)r_0\|^2} M\left(-MR_0+ (MR_0)^T\right)(M-I)$.
\item Let $T_1=L_2 R_0L_1^T=\displaystyle \frac{1}{\|(M-I)r_0\|^2} M\left(-MR_0+ (MR_0)^T\right)(M-I)R_0 M^T$ and
$T_2 =L_2-L_1=\displaystyle \frac{1}{\|(M-I)r_0\|^2} M\left(-MR_0+ (MR_0)^T\right)(M-I)-M$.
Then,
\begin{align*}
L_3 & = \frac{1}{\|T_2r_0\|^2}M(-T_1+T_1^T)T_2.
\end{align*}
\end{itemize}
\subsection{Influence of initial guess on convergence speed for AA(1)}
We now apply some of the AA(1) results of this section to investigate asymptotic convergence behavior of AA(1). We first discuss a scaling invariance property of the initial guess $x_0$ for the AA(1) method with $x_1=q(x_0)$ in the linear case. Since solving $Ax=b$ is equivalent to solving $Ay=0$ with $y=x-A^{-1}b$, we formulate these scaling properties for the case of a homogenous system.
\begin{proposition}\label{scaling-initial-guess-polynomial-invariant}
Consider solving $Ax=0$. Given any nonzero initial guess $x_0$, let $x_1=q(x_0)$. Consider the AA(1) polynomials in \cref{eq:AA1-three-term-M}, which depend on $x_0$ through the $\beta_k$. We have the following properties: for any nonzero scalar $\alpha$,
\begin{equation}\label{beta-invariant-scaling-guess}
\beta_k(x_0) =\beta_k(\alpha x_0),
\end{equation}
and, therefore,
\begin{equation}\label{pk-invariant-scaling-guess}
p_{k}(\lambda,x_0) =p_{k}(\lambda,\alpha x_0),
\end{equation}
where we explicitly indicate the dependence of $\beta_k$ and $p_k(\lambda)$ on the initial condition.
\end{proposition}
\begin{proof}
We use induction. Note that
\begin{align*}
r_0 &= Ax_0, \\
r_1&= Mr_0=MAx_0, \\
\beta_1(x_0)&= \frac{-r_1^T(r_1-r_0)}{(r_1-r_0)^T(r_1-r_0)}=\frac{-(MAx_0)^T(MA-A)x_0}{\left((MA-A)x_0\right)^T (MA-A)x_0}.
\end{align*}
Thus, when $k=1$, we have $\beta_1(x_0) =\beta_1(\alpha x_0)$. Since $p_1(\lambda)=\lambda$, it is obvious that $p_{1}(\lambda,x_0) =p_{1}(\lambda,\alpha x_0)$.
Assume that for $k\leq n$, $p_{k}(\lambda,x_0) =p_{k}(\lambda,\alpha x_0)$. Note that for $k\leq n$,
\begin{align*}
\beta_k(\alpha x_0) &=\frac{-r_k^T(r_k-r_{k-1})}{\|r_k-r_{k-1}\|^2},\\
&=\frac{- \left(p_k(M,\alpha x_0) A\alpha x_0\right)^T \left(p_k(M,\alpha x_0)-p_{k-1}(M,\alpha x_0)\right)A\alpha x_0 }{\|\big(p_k(M,\alpha x_0)-p_{k-1}(M,\alpha x_0)\big)A\alpha x_0\|^2}, \\
&=\beta_k(x_0).
\end{align*}
For $k=n+1$, we then have
\begin{align*}
p_{n+1}(\lambda,\alpha x_0)&=(1+\beta_n(\alpha x_0))\lambda p_n(\lambda,\alpha x_0) - \beta_n(\alpha x_0) \lambda p_{n-1}(\lambda,\alpha x_0),\\
& = (1+\beta_n(x_0))\lambda p_n(\lambda, x_0) - \beta_n( x_0) \lambda p_{n-1}(\lambda, x_0),\\
&= p_{n+1}(\lambda, x_0),
\end{align*}
which completes the proof.
\end{proof}
The following result shows the convergence of the root-linear asymptotic convergence factor under scaling of the initial condition.
\begin{proposition}\label{pro:scaling-invariant-limit}
Consider solving $Ax=0$. Given $x_0$, let $x_1=q(x_0)$. Then the AA(1) residuals defined in \cref{eq:rk-polynomial-form} have the following property:
\begin{equation}
\lim_{k\rightarrow \infty }\|r_k(x_0)\|^{\frac{1}{k}} = \lim_{k\rightarrow \infty } \|r_k( \alpha x_0)\|^{\frac{1}{k}},
\end{equation}
where we explicitly indicate the dependence of $r_k$ on the initial condition.
\end{proposition}
\begin{proof}
For initial guess $\alpha x_0$, we have $r_0(\alpha x_0)= A\alpha x_0=\alpha Ax_0=\alpha r_0(x_0)$. From \cref{scaling-initial-guess-polynomial-invariant}, we know that $p_k(\lambda,x_0)=p_k(\lambda,\alpha x_0)$. Furthermore,
\begin{align*}
\lim_{k\rightarrow \infty } \|r_k( \alpha x_0)\|^{\frac{1}{k}} &=\lim_{k\rightarrow \infty } \|p_k(M,\alpha x_0)r_0(\alpha x_0))\|^{\frac{1}{k}},\\
&= \lim_{k\rightarrow \infty } (\alpha)^{\frac{1}{k}}\|p_k(M,x_0)r_0(x_0))\|^{\frac{1}{k}},\\
&=\lim_{k\rightarrow \infty }\|p_k(M,x_0)r_0(x_0))\|^{\frac{1}{k}},\\
&= \lim_{k\rightarrow \infty } \|r_k( x_0)\|^{\frac{1}{k}}.
\end{align*}
\end{proof}
Next, we give a result on the number of iterations in which AA(1) converges exactly for a special choice of initial guess.
\begin{theorem}\label{AA1-2-iterations-eigevector}
Consider solving $Ax =b$ with AA(1). For initial guess $x_0=v+ x^*$, where $v$ is any eigenvector of $A$ and $Ax^*=b$, AA(1) with $x_1=q(x_0)$ converges to the true solution at iteration 2, i.e., $x_2=x^*$.
\end{theorem}
\begin{proof}
Assume that $A v =\mu v$ with $v\neq 0$. Note that $r_0 = Ax_0-b=A(v+x^*)-b=Av=\mu v$ and $r_1=Mr_0= \mu Mv$. Since
\begin{equation*}
r_1r_0^T = \mu M v (\mu v)^T = \mu^2(I-A)v v^T =\mu^2(1-\mu)v v^T,
\end{equation*}
we have
\begin{equation*}
-r_1r_0^T + (r_1r_0^T)^T =0.
\end{equation*}
From \cref{AA1-rk-simple-form-randomx0x1} we then have $r_2=0$, which means that $x_2=x^*$.
\end{proof}
\section{Numerical Results}\label{sec:Numerical-result}
In this section, we illustrate numerically how the theoretical results of this paper help in understanding AA(1) convergence. We first investigate the influence of the initial guess on the convergence of AA(1) applied to a linear problem. Then we compare the polynomials associated with the residuals of AA(1) with the polynomials of the fixed-point method. Finally, we broaden our view and
investigate the influence of the initial guess on the convergence of AA(1) applied to a nonlinear problem.
We first introduce some convergence terminology for our discussion, see, e.g., \cite{LinearacAA}.
\begin{definition}[$r$-linear convergence]
Let $\{x_k\}$ be any sequence that converges to $x^*$. Define
\begin{equation*}
\rho_{\{x_k\}} = \limsup\limits_{k\rightarrow \infty}\|x_k-x^*\|^{\frac{1}{k}}.
\end{equation*}
We say $\{x_k\}$ converges $r$-linearly with $r$-linear convergence factor $\rho_{\{x_k\}}$ if $\rho_{\{x_k\}}\in(0,1)$ and $r$-superlinearly if $\rho_{\{x_k\}}=0$. The ``r-'' prefix stands for ``root''.
\end{definition}
\begin{definition}[r-linear convergence of a fixed-point iteration]\label{def:rho-method}
Consider fixed-point iteration $x_{k+1}=q(x_k)$. We define the set of iteration sequences that converge to a given fixed point $x^*$ as
\begin{equation*}
C(q, x^*)=\Big\{ \{x_k\}_{k=0}^{\infty} | \quad x_{k+1} = q(x_k) \textrm{ for } k=0,1,\ldots, \textrm{ and } \lim_{k \rightarrow \infty}x_k=x^*\Big\},
\end{equation*}
and the worst-case r-linear convergence factor over $C(q, x^*)$ is defined as
\begin{equation}\label{eq:r-factor-defi}
\rho_{q,x^*} = \sup \Big\{ \rho_{\{x_k\}} |\quad \{x_k\}\in C(q, x^*) \Big\}.
\end{equation}
We say that the FP method converges r-linearly to $x^*$ with r-linear convergence factor $\rho_{q,x^*}$ if $\rho_{q,x^*} \in(0,1)$.
\end{definition}
We define the root-averaged error sequence of $\{x_k\}$ converging to $x^*$ as
\begin{equation}\label{eq:def_numer-r-convergence}
\sigma_k = \|x_k-x^*\|_2^{\frac{1}{k}}.
\end{equation}
The two numerical tests we consider in this section were previously discussed in \cite{LinearacAA}, which is a companion paper to this paper that discusses continuity and differentiability of the iteration function of AA($m$) viewed as a fixed-point method. Paper \cite{LinearacAA} identifies and sheds interesting light on AA($m$) convergence patterns for these problems, including oscillating behavior of $\beta_k$ as $k\rightarrow \infty$, and strong dependence of the root-linear convergence factors of AA($m$) sequences $\{x_k\}$ on the initial guess. Here we further explain some of these observations using the theoretical results from this paper.
\subsection{AA(1) for a linear system}\label{subsec:linear-problem}
In our numerical tests, we first consider a linear example:
\begin{problem}\label{prob:linear2x2-upper-tri}
Consider linear FP iteration function
\begin{equation}\label{eq:q-linear-2x2}
q(x) =Mx=\begin{bmatrix}
\frac{2}{3} & \frac{1}{4}\\
0 & \frac{1}{3}
\end{bmatrix}
x,
\end{equation}
with fixed point $x^* = (0,0)^T$ (see also \cite{LinearacAA}).
\end{problem}
\cref{AA-Linear-plot} shows $\sigma_k$ for FP and AA(1) with $x_1=q(x_0)$, where we use three different initial guesses, $(0.2,0.3)^T, \ (0.2,1)^T$, and $(0.2,-0.2)^T$ for both FP and AA(1). Note that for FP, $\lim_{k \rightarrow\infty }\sigma_k =\rho_{q,x^*}=\rho(q'(x^*))=\frac{2}{3}$. However, for AA(1), $\lim_{k \rightarrow \infty }\sigma_k =\rho_{\{x_k\}}$ depends on the initial condition.
\begin{figure}[h]
\centering
\includegraphics[width=.6\textwidth]{Ex2-Linear-p4.png}
\caption{\cref{prob:linear2x2-upper-tri} (linear). Root-averaged error $\sigma_k$ as a function of iteration number $k$ for different initial guesses.} \label{AA-Linear-plot}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=.49\textwidth]{fig1_prob04_mc_rconv.pdf}
\includegraphics[width=.49\textwidth]{fig1_prob04_mc_betas.pdf}
\caption{Monte Carlo tests for \cref{prob:linear2x2-upper-tri} (linear). (Results from \cite{LinearacAA}.)} \label{prob2-mc}
\end{figure}
Monte Carlo results with a large number of random initial guesses for \cref{prob:linear2x2-upper-tri} with FP and AA(1) iterations are shown in \cref{prob2-mc} (see also \cite{LinearacAA}). \cref{prob2-mc} indicates that AA(1) sequences $\{x_k\}$ converge $r$-linearly. However, the $r$-linear convergence factors $\rho_{\{x_k\}}$ strongly depend on the initial guess on a set of nonzero measure (see \cite{LinearacAA} for more discussion). There seems to be a least upper bound $\rho_{AA(1),x^*}$ for $\rho_{\{x_k\}}$ for the AA(1) iteration \cref{eq:anderson-1-step} that is smaller than the $r$-linear convergence factor $\rho_{q,x^*}=2/3$ of fixed-point iteration \cref{eq:fixed-point} by itself, but as far as we know, there are no theoretical results that allow us to compute $\rho_{AA(1),x^*}$ and to show it is smaller than $\rho_{q,x^*}$. Furthermore, we note that the $\beta_k$ sequences oscillate for this linear problem when the AA(1) iteration approaches $x^*$, and $\beta_k>-1$, which validates our \cref{thm:lower-boound-betak}.
\begin{figure}[h]
\centering
\includegraphics[width=.7\textwidth]{fig13_prob41_mc_rhoinitcont.pdf}
\includegraphics[width=.7\textwidth]{fig13_prob41_mc_rhoinitsurf.pdf}
\caption{Asymptotic convergence factor ($\sigma_k$ for large $k$) as a function of initial condition for \cref{prob:linear2x2-upper-tri} (linear).} \label{prob2-grid}
\end{figure}
In \cref{prob2-grid} we show the convergence factor of AA(1) applied to the linear \cref{prob:linear2x2-upper-tri} for different initial guesses $x_0$ on a regular grid with 50 by 50 points and $x_1=q(x_0)$. Again, we see that a least upper bound $\rho_{AA(1),x^*}$ for $\rho_{\{x_k\}}$ exists for the AA(1) iteration \cref{eq:anderson-1-step}. We see, however, a clear pattern in the numerically determined asymptotic convergence factors for the AA(1) sequences $\{x_k\}$, with radial invariance that we investigate further below. We also see in \cref{prob2-grid} that the apparent gap in the convergence factor spectrum in \cref{prob2-mc} corresponds to a steep, possibly discontinuous jump in the convergence factor surface in the initial condition plane.
The radial invariance in \cref{prob2-grid} validates our \cref{scaling-initial-guess-polynomial-invariant,pro:scaling-invariant-limit}.
To investigate the radial dependence in \cref{prob2-grid} further, we take $x_0=(\cos(\theta),\sin(\theta))^T$ and $x_1=q(x_0)$ for AA(1), and plot in \cref{prob2-theta} the numerical convergence factor for large $k$ as a function of $\theta$.
\begin{figure}[h]
\centering
\includegraphics[width=.6\textwidth]{rho_theta_prob4.pdf}
\caption{Asymptotic convergence factor as a function of initial condition angle $\theta$ for \cref{prob:linear2x2-upper-tri} (linear).} \label{prob2-theta}
\end{figure}
Note that, for \cref{prob:linear2x2-upper-tri}, $A$ in \cref{AA1-2-iterations-eigevector} is given by
\begin{equation*}
A = \begin{bmatrix}
\frac{1}{3} & -\frac{1}{4}\\
0 & \frac{2}{3}
\end{bmatrix}.
\end{equation*}
The largest eigenvalue of $A$ is $\frac{2}{3}$ with eigenvector $v_{\max}=(1,-4/3)^T$, and $v_{\min}=(1,0)^T$ corresponds to the smallest eigenvalue $\frac{1}{3}$. \cref{prob2-grid,prob2-theta} also validate \cref{AA1-2-iterations-eigevector}: the convergence factors decay to 0 for initial conditions in the direction of the eigenvectors of $A$, corresponding to exact convergence in two steps.
\cref{prob2-grid} suggests that the greatest convergence factor occurs for initial conditions in the direction perpendicular to the largest eigenvector of $A$. This suggests the following conjecture:
\begin{conj}\label{conj:worse-initial-guess}
Consider AA(1) for $q(x)=Mx+b$, with $A=I-M$ and $A\in\mathbb{R}^{2\times 2}$. Assume $A$ has two different eigenvalues with non-orthogonal eigenvectors. Let $v_{\max}$ be an eigenvector of $A$ corresponding to the largest eigenvalue of $A$. In AA(1), if the initial guess $x_0$ satisfies $x_0 \bot v_{\max}$, then the r-linear convergence factor for AA(1) defined in \cref{eq:r-factor-defi} is achieved by the sequence generated by $x_0$.
\end{conj}
We now further investigate this conjecture for the problem of \cref{prob2-grid,prob2-theta}. If we take $x_{0}=(\eta,\frac{3}{4}\eta)^T$ with $\eta\neq 0$, then $x_{0}\bot v_{\max}$ and $x_{0}$ is not an eigenvector of $A$. Denote by $\theta^*$ the polar angle of $x_{0}$, i.e., $\theta^*\approx0.6435$ or $3.7851$. We verify numerically that the worst convergence factor indeed arises around $\theta\approx0.6095$ and $\theta\approx 3.7511$ in \cref{prob2-theta}, consistent with \cref{conj:worse-initial-guess}. We have further validated \cref{conj:worse-initial-guess} for several other 2$\times$2 matrices $A$, but a proof is not yet available.
Finally, we compare in more detail two choices of initial guesses for $x_0$ and $x_1$: the first choice is $x_0=(\cos \theta_1, \sin \theta_1)^T$ and $x_1= \alpha \, (\cos \theta_2, \sin \theta_2)^T$, and the second choice is $x_0=(\cos \theta_1, \sin \theta_1)^T$ and $x_1=q(x_0)$, as for \cref{prob2-grid,prob2-theta}. We present the histogram of the convergence factor for these two choices in \cref{prob2-x0-x1}.
\begin{figure}[h]
\centering
\includegraphics[width=.49\textwidth]{x0_x1_hist.pdf}
\includegraphics[width=.49\textwidth]{x0_hist.pdf}
\caption{\cref{prob:linear2x2-upper-tri} (linear). Left: Histogram of asymptotic convergence factors for $50^3=125,000$ initial condition pairs
$x_0=(\cos \theta_1, \sin \theta_1)^T$ and $x_1= \alpha \, (\cos \theta_2, \sin \theta_2)^T$, where $\theta_1$ and $\theta_2$ take on 50 equally spaced values $2k\pi/50$, $k=0, 1,\ldots,49$, and $\alpha$ takes on 50 equally spaced values
$10k/50$, $k=1, \ldots,50$.
The largest convergence factor value is $\approx$ 0.440, which is slightly greater than the largest value, 0.410, in Fig.\ \ref{prob2-theta} where $x_0=(\cos \theta, \sin \theta)$ and $x_1=Mx_0$. Right: Histogram of asymptotic convergence factors for 5,000 initial condition
$x_0=(\cos \theta_1, \sin \theta_1)$ and $x_1=q(x_0)$, as in \cref{prob2-mc,prob2-grid,prob2-theta}.} \label{prob2-x0-x1}
\end{figure}
For the choice where $x_0=(\cos \theta_1, \sin \theta_1)^T$ and $x_1= \alpha \, (\cos \theta_2, \sin \theta_2)^T$ with 125,000 initial condition pairs, the average convergence factor is 0.164131496689178.
For the choice where $x_0=(\cos \theta_1, \sin \theta_1)^T$ and $x_1= q(x_0)$ with 5,000 initial guesses, the average convergence factor is 0.273090075503153. Note that increasing the number of initial condition points does not have a big influence on these results. It is clear that the first choice gives a better average convergence factor. So while $x_1=q(x_0)$ appears to give a slightly better worst-case performance, the average convergence factor appears to be better when $x_1$ is chosen randomly as in the experiment, independent of $x_0$.
\begin{figure}[h]
\centering
\includegraphics[width=.49\textwidth]{p5_polynomial.pdf}
\includegraphics[width=.49\textwidth]{p10_polynomial.pdf}
\caption{Residual polynomials of AA(1) for \cref{prob:linear2x2-upper-tri} (linear).} \label{prob2-poly}
\end{figure}
Next, we investigate the polynomials \cref{eq:AA1-three-term-M} for the residual of AA(1) with $x_1=q(x_0)$ applied to \cref{prob:linear2x2-upper-tri}. It is easy to show that the residual $r_k$ for fixed-point iteration $x_{k+1}=q(x_k)$ in the linear case is given by $r_k=M^k r_0$. We take initial guess $x_0=(0.2,0.3)^T$ and $x_1=q(x_0)$ for AA(1). \cref{prob2-poly} presents the AA(1) residual polynomials from \cref{eq:rk-polynomial-form}, and $\lambda^k$ for the fixed-point method. The filled circles are the two eigenvalues $\lambda=\frac{1}{3}, \frac{2}{3}$ of matrix $M$. We can see that $p_5(\lambda)$ of AA(1) is a better polynomial than $\lambda^5$ at the eigenvalues of $M$, and $p_{10}(\lambda)$ of AA(1) is a better polynomial than $\lambda^{10}$ at the eigenvalues of $M$, illustrating the mechanism by which AA(1) accelerates the convergence of the FP method.
\begin{figure}[h]
\centering
\includegraphics[width=.49\textwidth]{fig2_prob02_mc_rconv.pdf}
\includegraphics[width=.49\textwidth]{fig2_prob02_mc_betas.pdf}
\caption{Monte Carlo tests for \cref{prob:nonlinear2x2} (nonlinear). (Results from \cite{LinearacAA}.)} \label{prob3-mc}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=.7\textwidth]{fig14_prob201_mc_rhoinitcont.pdf}
\includegraphics[width=.7\textwidth]{fig14_prob201_mc_rhoinitsurf.pdf}
\caption{Asymptotic convergence factor as a function of initial condition for \cref{prob:nonlinear2x2} (nonlinear).} \label{prob3-grid}
\end{figure}
\subsection{AA(1) for a nonlinear system}
Although our analysis in this paper is for linear systems, it is interesting to explore how the convergence of AA(1) depends on the initial guess for a nonlinear system.
\begin{problem}\label{prob:nonlinear2x2}
Consider the nonlinear system
\begin{align}
x_2= x_1^2 \label{eq:exm1-1}\\
x_1+(x_1-1)^2 +x_2^2 =1 \label{eq:exm1-2}
\end{align}
with solution $(x_1^*,x_2^*) = (0,0)$.
Let $x=[x_1 \ x_2]^T$ and define the FP iteration function
\begin{equation*}
q(x) = \begin{bmatrix} \displaystyle \frac{1}{2}(x_1+x_1^2+x_2^2) \\ \\ \displaystyle \frac{1}{2}(x_2+x_1^2) \end{bmatrix},
\end{equation*}
with Jacobian matrix
\begin{equation*}
q'(x)=
\begin{bmatrix}
x_1+\displaystyle \frac{1}{2}& x_2\\
x_1 & \displaystyle \frac{1}{2}
\end{bmatrix}.
\end{equation*}
We have
\begin{equation*}
q'(x^*) = \begin{bmatrix}
\displaystyle \frac{1}{2}& 0 \\
0 & \displaystyle \frac{1}{2}
\end{bmatrix}, \ \textrm{and} \quad
\rho(q'(x^*)) =\displaystyle \frac{1}{2}<1.
\end{equation*}
\end{problem}
Monte Carlo results with a large number of random initial guesses for the nonlinear \cref{prob:nonlinear2x2}, with FP and AA(1) with $x_1=q(x_0)$, are shown in \cref{prob3-mc}. From \cref{prob3-mc}, we see that the nonlinear convergence behavior is qualitatively similar to that of the linear case shown in \cref{prob2-mc}, that is, the AA(1) sequences $\{x_k\}$ converge $r$-linearly, while the $r$-linear convergence factors $\rho_{\{x_k\}}$ depend strongly on the initial guess. An upper bound $\rho_{AA(1),x^*}$ seems to exist for $\rho_{\{x_k\}}$ for the AA(1) iteration which is smaller than $\rho_{q,x^*}=1/2$ of fixed-point iteration \cref{eq:fixed-point} by itself. It is clear that the $\beta_k$ sequences oscillate for this nonlinear problem as the AA(1) iteration approaches $x^*$, and $\beta_k>-1$ even though we do not have theoretical results on this for the nonlinear case.
In \cref{prob3-grid} we show the convergence factor of AA(1) with $x_1=q(x_0)$ applied to the nonlinear \cref{prob:nonlinear2x2} for different initial guesses, where we take $x_0$ on a uniform grid with 101 by 101 points.
It is interesting to see, for the nonlinear problem of \cref{prob3-grid}, that there are also preferred directions with fast convergence for the initial condition near the solution $(0,0)^T$, as in the linear case. Future work will investigate this further, possibly using eigenvectors obtained after linearization.
\section{Conclusions}\label{sec:con}
In this paper, we have examined polynomial residual update formulas for Andersen acceleration with window size $m$ in the linear case.
First we have shown that, for window sizes $m=1$ and $m>1$, AA($m$) is a (multi-) Krylov method.
Specifically, we have derived polynomial residual update formulas, orthogonality relations, a lower bound on the AA(1) acceleration coefficient $\beta_k$, and explicit nonlinear recursions for the AA(1) residuals and residual polynomials that do not include the acceleration coefficient $\beta_k$.
We have further studied the influence of the initial guess on the asymptotic convergence factor of AA(1). We find that there are preferred directions for the initial condition of AA(1) that result in much faster convergence, related to the eigenvectors of $A$.
Note that for the case of AA($m$) with stationary coefficients $\beta_i$ in \cref{eq:AA-iteration}, \cite{desterck2020,wang2020} provide insight in the convergence improvement that results from the optimal stationary AA($m$) iteration, based on how it improves the spectral radius of $q'(x^*)$. However, \cite{desterck2020,wang2020}, as well as the results presented in this paper, leave open the question of determining the $r$-linear convergence factor $\rho_{ AA(m),x^*}$ for the non-stationary AA($m$) that is widely used in science and engineering applications. It is, hence, not known how to compute by how much $\rho_{ AA(m),x^*}$ improves upon $\rho_{q,x^*}$, even for the linear case with 2$\times$2 matrices $M=I-A$.
\bibliographystyle{siamplain}
|
\section{Introduction}
As computers and network connectivity have become an ubiquitous part of society, the size and scope of distributed systems has grown. It is now commonplace for these systems to contain hundreds or even thousands of computers spread across the globe connected through the Internet. To better facilitate common operations for applications, like routing and searching, many distributed systems are built using \emph{overlay networks}, where connections occur over logical links that consist of zero or more physical links. Overlay networks allow nodes to embed a predictable topology onto their (usually fixed) physical topology, selecting the best network for the application's particular needs.
Complicating the use of these overlay networks, however, is the reality that systems composed of such a wide variety and distribution of devices are more prone to failures caused from problems with the devices or physical links. For instance, fiber optic cables can be severed, power outages can cause machines to disconnect without warning, and even intentional user actions like joining or leaving the system on a predictable schedule can result in an incorrectly-configured overlay network causing the client application to fail.
One approach for managing these faults and preventing failures is to design protocols which are resilient to a targeted set of specific system faults, such as nodes joining or leaving the system. However, the unpredictable nature of these distributed systems makes it difficult to identify and control for every possible fault. It is for this reason that researchers have turned to \emph{self-stabilizing overlay networks}. A self-stabilizing overlay network guarantees that after \emph{any} transient fault, a correct topology will eventually be restored. This type of network can ensure autonomous operation of distributed systems even in the face of a variety of unforeseen transient faults.
\subsubsection{The Problem} Our focus is on building robust self-stabilizing overlay networks efficiently. More specifically, we are interested in creating \emph{efficient} algorithms that add and delete logical edges in the network to transform an arbitrary weakly-connected initial topology into a correct \emph{robust} topology. By \emph{efficient}, we mean these algorithms have a time \emph{and} space complexity which is polylogarithmic in the number of nodes in the network. By \emph{robust}, we mean topologies where the failure of a few nodes is insufficient to disconnect the network.
\subsubsection{Main Results and Significance}
With this paper, we present an efficient self-stabilizing overlay network with desirable practical properties like robustness and low diameter. In particular, we present a self-stabilizing algorithm for the creation of a \textsc{Chord} network which has expected polylogarithmic space and time requirements. Note that this is the first work to present an efficient (in terms of time and space) self-stabilizing overlay network for a \emph{robust} topology. Note that, while our algorithm is deterministic, it depends upon the prior work of Berns~\cite{berns_avatar-2016}, which was randomized, and therefore our results are in expectation.
Our second result is the explicit identification of a ``design pattern'' we call \emph{network scaffolding} for creating self-stabilizing overlay networks. This pattern has been used in several other works, and the success of this approach, both previously as well as in this work for building \textsc{Chord}, leads us to believe it can be used for many other topologies as well. Our work is a first step towards fully defining and analyzing this design pattern. Our goal is that explicit identification of this design pattern can be useful to other researchers and practitioners in the design and implementation of other self-stabilizing overlay networks.
A preliminary version of this work appeared as a brief announcement at SPAA 2021~\cite{berns_chord_spaa_2021}. In this version, we provide a more detailed discussion of the \textsc{Avatar} background, an improved analysis showing a better bound on the degree expansion, and also provide an extended discussion about the identified design pattern.
\subsubsection{Related Work and Comparisons}
\label{section:related_work}
The past few decades have seen tremendous growth in both the theory and practice of overlay networks. Some of this work has focused on \emph{unstructured} overlay networks where connections need not satisfy any particular property and there are no constraints on what is considered a ``legal'' topology, such as Napster and Gnutella~\cite{saroiu_napster_2003}. There are also \emph{semi-structured} overlay networks, where the network topology must satisfy a particular constraint or has a particular property such that there are several possible correct topologies. For instance, an overlay may be built to mimic a ``small world'' graph~\cite{hui_swop_2004}.
Our work focuses on \emph{structured} overlay networks, where there is exactly one correct topology for any given set of nodes. While constructing and maintaining the correct topology adds additional work for the algorithm designer, common operations such as routing and searching are much more efficient with these structured networks. There are many examples of such overlays, including \textsc{Chord}~\cite{stoica_chord_01}, \textsc{Pastry}~\cite{rowstron_pastry_2001}, and \textsc{Tapestry}~\cite{zhao_tapestry_2002}. These early examples of structured networks, however, provided very limited fault tolerance.
To this end, much previous work has focused on improving the fault tolerance of overlay networks. One approach has looked at \emph{self-healing} networks, where the network can maintain certain properties while a limited number of faults occur during a fixed time period. Examples of this work include the Forgiving Tree~\cite{hayes_forgiving-tree_2008}, the Forgiving Graph~\cite{hayes_forgiving-graph_2009}, and DEX~\cite{pandurangan-dex-2016}. Many of these approaches also use virtual nodes~\cite{trehan_virtual-2012} as done in our work. More recently, Gilbert et al. presented \textsc{DConstructor}~\cite{gilbert_dconstructor_2020} which is able to build a correct topology from any initial topology and maintain this in the face of some joins and leaves. G\"{o}tte et al.~\cite{gotte_time-optimal_podc21} also presented an algorithm for transforming a constant-degree network into a tree in $\mathcal{O}(\log n)$ rounds. The key difference between these works and ours, however, is we use a stronger fault model, requiring our algorithm to build the correct configuration regardless of the initial topology \emph{or} the initial state of the nodes. This is a paradigm called self-stabilization, which we discuss next.
\emph{Self-stabilizing overlay networks} are those that guarantee a legal configuration will be automatically restored by program actions after \emph{any} transient fault so long as the network is not disconnected. This is often modeled as the ability for the network to form a correct topology when starting from an arbitrary weakly-connected state. One of the first such examples of a self-stabilizing overlay network was for the simple structured \textsc{Linear} topology~\cite{onus_linear_07} where nodes were arranged in a ``sorted list''. Since then, there have been several other self-stabilizing structured overlay networks created, including \textsc{Skip+}~\cite{jacob_skipplus_09} and \textsc{Re-Chord}~\cite{kniesburges_rechord_11}. Unlike the simple \textsc{Linear} topology, \textsc{Skip+} and \textsc{Re-Chord} maintain several desirable properties for client applications, including low node degree and low diameter. Unfortunately, their worst-case time (in the case of \textsc{Re-Chord}) or space (in the case of \textsc{Skip+}) complexity is linear in the number of nodes.
To date, we are aware of only one self-stabilizing overlay network that is efficient in terms of both time and space. Berns presented the \textsc{Avatar} network framework as a mechanism for ensuring a faulty configuration is detectable for a wide variety of networks, and also gave a self-stabilizing algorithm for the construction of a binary search tree~\cite{berns_avatar-2016}. Our current work builds upon this \textsc{Avatar} network.
A goal of our current work is to identify a general ``design pattern'' which can be used for building self-stabilizing overlay networks. There has been little work done on identifying these general patterns for overlay network construction. One exception to this is the \emph{Transitive Closure Framework}~\cite{berns_tcf_11} (TCF), which provides a way to build any locally-checkable topology by detecting a fault, forming a clique, and then deleting those edges which are not required in the correct configuration. While TCF can create any locally-checkable topology quickly, it requires node degrees to grow to $\mathcal{O}(n)$ during convergence and is therefore not practical for large networks.
While we are interested in structured topologies, there has been work in semi-structured networks. For instance, self-stabilizing semi-structured networks such as a \emph{small world network}~\cite{kniesburges_small-world_2012} or a \emph{power law network}~\cite{alsulaiman_power-law_2015} have been proposed. G{\"o}tte et al~\cite{gotte_sirocco19} recently presented an algorithm for quickly converting certain topologies into ones with a logarithmic diameter.
\section{Preliminaries}
\label{section:prelims}
\subsection{Model of Computation}
We model our distributed system as an undirected graph $G = (V, E)$, with processes being the $n$ nodes of $V$ and the communication links being the edges $E$. Each node $u$ has a unique identifier $u.id \in \mathbb{N}$, which is stored as immutable data in $u$. Where clear from the context, we will use $u$ to represent the identifier of $u$.
Each node $u \in V$ has a \emph{local state} consisting of a set of variables and their values, along with its immutable identifier $u.id$. A node may execute \emph{actions} from its \emph{program} to modify the values of the variables in its local state. All nodes execute the same program. Besides modifying its local state, a node can also communicate with its neighbors. We use the \emph{synchronous message passing} model of computation with bounded communication channels, where computation proceeds in synchronous rounds. During each round, a node may receive messages sent to it in the previous round from any of its neighbors, execute program actions to update its state, and send messages to any node in its neighborhood $N(u) = \{v \in V:(u,v) \in E\}$. We assume the communication channels are reliable with bounded delay so that a message is received in some round $i$ if and only if it was sent in round $i - 1$.
In the overlay network model, nodes communicate over logical links that are part of a node's state, meaning a node may execute actions to create or delete edges in $G$. In particular, in any round a node may delete any edge incident upon it, as well as create any edge to a node $v$ which has been ``introduced'' to it from some neighbor $w$, such that $(u,w)$ and $(w, v)$ are both in $E$. Said in another way, in a particular round a node may connect its neighbors to one another by direct logical links.
The goal for our computation is for nodes to execute actions to update their state (including modifying the topology by adding and deleting edges) until a legal configuration is reached. A \emph{legal configuration} can be represented as a predicate over the state of the nodes in the system, and as links are part of a node's state for overlay networks, a legal configuration is defined at least in part by the network topology. The \emph{self-stabilizing overlay network problem} is to design an algorithm $\mathcal{A}$ such that when executing $\mathcal{A}$ on a connected network with nodes in an arbitrary state, eventually a legal configuration is reached. This means that a self-stabilizing overlay network will always automatically restore a legal configuration after any transient fault so long as the network remains connected.
\subsection{Performance Metrics}
We analyze the performance of our self-stabilizing overlay network algorithms in terms of both time and space. For time, we are interested in how quickly the network will be able to recover from a transient fault. Specifically, we measure the maximum number of (synchronous) rounds that may be required in the worst case to take any set of $n$ nodes from an arbitrary connected configuration to a legal configuration. This is called the \emph{convergence time}.
The space complexity measure of interest for us is related to the maximum number of neighbors a node might have during convergence that were not present in the initial configuration and are not required in the final configuration. Said in another way, we are interested in the number of ``extra'' neighbors a node may acquire during convergence. More specifically, we use the \emph{degree expansion} metric~\cite{berns_avatar-2016}, which is the ratio of the maximum node degree of any node during convergence over the maximum node degree from the initial or final configuration. The degree expansion helps us to judge how much a node's degree increases as a result of program actions and not from the requirements of the final topology or from a high-degree initial configuration.
Finally we note that, as with many distributed algorithms, we consider an efficient algorithm one which keeps these measures polylogarithmic in the number of nodes in the network.
\section{\textsc{Avatar}}
\label{section:avatar}
Our efficient algorithm for creating the \textsc{Chord} network builds upon previous work on the \textsc{Avatar} overlay network framework. While a full discussion of this network can be found in the work by Berns~
\cite{berns_avatar-2016}, we present a brief summary and discussion below to provide the necessary background to understand the new contributions of our own work.
\subsection{The \textsc{Avatar} Overlay Network}
The \textsc{Avatar} overlay network framework can be used to define a variation of any particular network topology. The general idea behind the \textsc{Avatar} framework is to create a dilation-1 embedding of a particular $N$ node \emph{guest network} (with node identifiers in the range of $[0, N)$) onto the $n$ node \emph{host network} (with $n \leq N$). More specifically, in \textsc{Avatar} each node $u \in V$ from the host network (except the two nodes with the smallest and largest identifiers) simulates or ``hosts'' all nodes from the guest network with identifiers in $u$'s \emph{responsible range}, defined as the range of $[u.id, v.id)$, where $v.id$ is the smallest identifier greater than $u.id$ taken from all nodes in $V$, which we call the \emph{successor} of $u$. The node $u_0$ with the smallest identifier has a responsible range of $[0, v.id)$ (where again $v$ is the successor of $u_0$), while the node with the largest identifier $u_{n-1}$ has a responsible range of $[u_{n-1}.id, N)$. To ensure a dilation-1 embedding, for every edge $(a, b)$ in the guest network there exists an edge between the host nodes of $a$ and $b$ in the host network, or the host node for $a$ and $b$ is the same -- that is, either both $a.id$ and $b.id$ are in the responsible range of the same host node, or there exists the edge $(host(a), host(b))$ in the host network such that $a.id$ is in the responsible range of host node $host(a)$ and $b.id$ is in the responsible range of host node $host(b)$. The definition of some $N$ node guest network $\textsc{Guest}(N)$ along with the constraints on the corresponding edges in the host network define the legal $\textsc{Avatar}(\textsc{Guest}(N))$ network.
The use of a guest and host network provides two advantages. First, the requirement that there is exactly one correct configuration for any given $N$ (meaning the guest network uses nodes $[0, N)$), along with the fact that the successor relationship used in the host network to define the responsible ranges can easily be determined from a node's local state, ensures that any topology is \emph{locally checkable}. Second, we can design our algorithms (for both stabilization and end-user applications) to execute on the guest network, which has a single predictable configuration for a given $N$, regardless of the node set $V$. This simplifies both the design and analysis of our algorithms.
As a final comment, we note that the use of \textsc{Avatar} does require all nodes to know $N$, the upper bound on the number of nodes in the network, and that our analysis of convergence time and degree expansion is in terms of this $N$. Given that all of our algorithms have polylogarithmic time and space requirements, in practice one could easily select an $N$ which was large enough to accommodate any possible node additions while still having a time and space complexity less than many existing algorithms which have complexity at least linear in the actual number of nodes $n$. Said in another way, even when $N$ is much larger than $n$, our efficient algorithms may still require fewer resources if $\log N \ll n$. If we consider IPv6, for instance, $\log N$ would be only $128$.
\subsection{\textsc{Avatar}(\textsc{CBT})}
Beyond defining the \textsc{Avatar} framework, Berns also defined the \textsc{Cbt} guest network and a self-stabilizing algorithm for building the $\textsc{Avatar}(\textsc{Cbt})$ network in a polylogarithmic number of rounds with a polylogarithmic degree expansion (both in expectation). The \textsc{Cbt} topology is simply a complete binary search tree of the specified $N$ nodes.
\subsubsection{Stabilization}
Our work building the \textsc{Chord} network depends upon the existence of a ``scaffold'' \textsc{Cbt} network. To be self-stabilizing, we need a way to build this \textsc{Cbt} in a self-stabilizing manner. This is exactly what is provided by the earlier \textsc{Avatar} work of Berns. The full description of the self-stabilizing algorithm for $\textsc{Avatar}(\textsc{Cbt})$ can be found in the original work~\cite{berns_avatar-2016}. We present a short informal summary of the algorithm's operation here to assist in understanding and verifying the correctness of our approach.
The general idea for the $\textsc{Avatar}(\textsc{Cbt})$ algorithm can be described using three components:
\begin{enumerate}
\item \emph{Clustering}: The first step in the algorithm is for nodes to form clusters. These clusters begin as a single host node hosting a full $N$ virtual node \textsc{Cbt} network. In the initial configuration, nodes may not be a part of a cluster, but since \textsc{Avatar} is locally checkable, all faulty configurations of $\textsc{Avatar}(\textsc{Cbt})$ contain at least one node which detects the faulty configuration and will begin forming the single-node clusters. This fault detection and cluster creation will propagate through the network until eventually all nodes are members of $N$ virtual node \textsc{Cbt} clusters.
\item \emph{Matching}: The second step of the algorithm is to match together clusters so that they may merge together. To do this, the root node of the binary tree repeatedly polls the nodes of its cluster, asking them to either find neighboring clusters that are looking for merge partners (called the leader role), or to look at neighboring clusters that can assign them a merge partner (called the follower role). The role of leader or follower is randomly selected. Leader clusters will match together all of their followers for merging, creating a matching between clusters that may not be direct neighbors. This ability to create edges to match non-neighboring clusters allows more matches to occur, and thus more merges, and thus a faster convergence time.
\item \emph{Merging}: The algorithm then deals with the merging of matched clusters. To prevent degrees from growing too large, a cluster is only allowed to merge with at most one other cluster at a time. Once two clusters have matched from the previous step, the roots of the clusters connect as ``partners'' and update their successor pointers based upon the identifier of the host of the root of the other cluster. One node will have its responsible range become smaller, and this node will send all guest nodes that were in its old responsible range to its partner in the other cluster. The children of the root nodes are connected, and then they repeat the process of updating successor pointers and passing along guest nodes outside their new responsible range. Eventually this process reaches the leaves, at which point all nodes in both clusters have updated their responsible ranges and now form a new legal \textsc{Cbt} cluster.
\end{enumerate}
This process of matching and merging continues until eventually only a single cluster is left, which is the correct $\textsc{Avatar}(\textsc{Cbt})$ network. We restate the following theorem from the original work and offer a brief sketch of the proof's intuition.
\begin{theorem}
The self-stabilizing algorithm for $\textsc{Avatar}(\textsc{Cbt})$ by Berns~\cite{berns_avatar-2016} has a convergence time of $\mathcal{O}(\log^2 N)$ rounds in expectation, and a degree expansion of $\mathcal{O}(\log^2 N)$ in expectation.
\end{theorem}
\noindent
\emph{Intuition:} A cluster has a constant probability of being matched and merged with another cluster in $\mathcal{O}(\log N)$ rounds, meaning the number of clusters is reduced by a constant fraction every $\mathcal{O}(\log N)$ rounds in expectation. This matching and merging only needs to happen $\mathcal{O}(\log N)$ times until we have a single cluster, giving us a time complexity of $\mathcal{O}(\log^2 N)$ rounds in expectation. The degree of a node can grow during a merge or during the matching process. However, during a merge a node's degree will grow to at most $\mathcal{O}(\log^2 N)$, and the node's degree will increase by only a constant amount during each match and there are only $\mathcal{O}(\log N)$ such matches in expectation, meaning the degree expansion of the algorithm is also $\mathcal{O}(\log^2 N)$.
\subsubsection{Communication}
The original work on \textsc{Avatar}(\textsc{Cbt}) also defined a communication mechanism to execute on the guest \textsc{Cbt} network. We will also use this mechanism in our algorithm to ensure edges are added systematically and thus limiting unnecessary degree growth. In particular, we will use a variant of a \emph{propagation of information with feedback} (PIF) algorithm~\cite{pif_jpdc_2010} which will execute on the (guest) nodes of \textsc{Cbt}. While the original work was snap-stabilizing, this would not be a requirement in our work. We are instead simply interested in an organized way to communicate information in waves in a tree.
In PIF, communication happens in waves that are initiated by the root of the binary tree. First, the root executes a \emph{propagate} action which sends information down the tree level by level. Once the propagated information reaches the leaves, the leaves begin a \emph{feedback} action, performing some operation and then signalling to their parent that the message has been received and acted upon by all descendants in the tree. Once the parent of a node has received the message, the node may \emph{clean} and prepare for the next wave. Once the root receives the feedback wave, it knows the message was successfully received and acted upon by all nodes in the tree, and the root may continue with further PIF waves if necessary.
We will use this communication mechanism to add edges to our network to build \textsc{Chord}. As the PIF process itself is previously defined, we only need to provide the actions each node will perform for each part, as well as any data that is sent. In particular, we will say that a tree $T$ executes a $PIF(X)$ wave, meaning the root of tree $T$ will signal to its children that a propagation wave has begun with the $PIF(X)$ message. Furthermore, we will specify the \emph{propagate} action of $a$, which is what each non-root node $a$ should do when it receives the propagation message $PIF(X)$. We also specify the \emph{feedback} action of $a$, which is the actions each node $a$ should take when it receives acknowledgements from its children that the most-recent propagation wave has completed and the corresponding feedback wave is underway.
\section{\textsc{Avatar}(\textsc{Chord})}
\label{section:algorithm}
In this section we discuss how we can use the existing \textsc{Avatar}(\textsc{Cbt}) self-stabilizing overlay network as the starting point for the efficient creation of a variant of the \textsc{Chord} overlay network.
\subsection{Overview of Our Approach}
Arguably one of the major barriers to the practical implementation for self-stabilizing overlay networks is the complexity that must be managed when designing and analyzing these networks, particularly when we desire efficient self-stabilization. For instance, TCF~\cite{berns_tcf_11} is simple and works with any locally-checkable topology, but it requires $\Theta(n)$ space. One could imagine a simple ``design pattern'' which simply suggests that in every round, a node computes their ideal neighborhood given the information available to them from their state and the state of their neighbors, and then add and delete edges to form this ideal neighborhood. Unfortunately, analyzing this algorithm in terms of \emph{both} correctness and efficiency is quite difficult as one must consider the implications of a variety of actions on a variety of possible initial configurations.
One approach to managing complexity is to start by building smaller or simpler structures, and then using these to continue towards the final goal. Consider, for instance, the construction of a large building. One common approach is to erect a simple scaffold and use this scaffold to build the more complex permanent structure. As another example, consider the prior work on using \emph{convergence stairs} for analyzing general self-stabilizing algorithms. In this technique, one first must show the system converges to some weaker predicate $A_0$ from an arbitrary initial configuration, then show it converges to $A_1$ provided it is in $A_0$, then show it converges to $A_2$ provided it is in $A_1$, and so on until you have reached the correct configuration. These patterns of design and analysis are similar in that they take a complex set of required actions, decompose them into smaller distinct steps, and then rely on prior solutions to the smaller steps to move to the next ones.
In the remainder of this section, we discuss our approach for efficiently creating a self-stabilizing version of the \textsc{Chord} network based upon this idea of scaffolding. In particular, we shall define the \textsc{Chord} topology, and then discuss how we can use $\textsc{Avatar}(\textsc{Cbt})$ as a starting point for constructing \textsc{Chord}. We then show how nodes can determine in a short amount of time whether they should be building the ``scaffold'' (\textsc{Cbt}) or the target topology (\textsc{Chord}).
\subsection{\textsc{Chord}(N)}
\label{section:target}
Our target network aims to resolve the lack of robustness of the \textsc{Cbt} scaffold network. In particular, our target network is an $N$-node \textsc{Chord} network defined as follows:
\begin{definition}
For any $N \in \mathbb{N}$, let $\textsc{Chord}(N)$ be a graph with nodes $[N]$ and edge set defined as follows. For every node $i$, $0 \leq i < N$, add to the edge set $(i, j)$, where $j = (i + 2^k) \mod N$, $0 \leq k < \log N - 1$. When $j = (i + 2^k)\mod N$, we say that $j$ is the $k$-th finger of $i$.
\end{definition}
It is worth again noting that our use of the \textsc{Avatar} framework results in a locally-checkable version of the \textsc{Chord} network. \textsc{Chord} as defined on an arbitrary set of nodes is actually \emph{not} locally checkable, particularly because of the ``ring'' edges (in a legal configuration, exactly one node should have two immediate neighbors with smaller identifiers, but which node this should be cannot be determined if the node set is arbitrary). Unlike prior approaches, then, our stabilizing \textsc{Chord} network is \emph{silent}, meaning no messages or ``probes'' need to be continuously exchanged between nodes in a legal configuration.
Our goal, then, is to use the $N$-node topology of \textsc{Cbt} to add edges to the guest nodes (and to the corresponding host network as required to maintain a dilation-1 embedding) until we have formed the correct $N$-node \textsc{Chord} network.
\subsection{Building \textsc{Chord} from \textsc{Cbt}}
Figure~\ref{algo:chord_scaffold} elaborates on the algorithm which uses our guest \textsc{Cbt} network as a scaffold for creating the guest \textsc{Chord} network. The algorithm uses the fact that \textsc{Chord} edges can be created inductively. That is, assuming all fingers from $0$ to $k$ are present, the $k+1$ finger can be created in a single round. Specifically, if node $b$ is the $(i-1)$ finger of $c_0$, and $c_1$ is the $(i-1)$ finger of $b$, the $i$th finger of $c_0$ is $c_1$. The algorithm begins by correctly building finger $0$, then recursively adds the first finger, then the second, and so on. This adding of edges is done in a metered fashion, however, to prevent unnecessary degree growth from faulty initial configurations.
Once the scaffold network has been built, we can begin the process of constructing our final target topology. We design our algorithm to execute on the $N$ guest nodes of \textsc{Cbt}, with the goal being to add edges to the nodes of \textsc{Cbt} until they have formed the $N$ guest node \textsc{Chord} network. For now, we shall assume that the network is in the legal \textsc{Cbt} configuration. We will relax this assumption and consider an arbitrary initial configuration shortly.
The algorithm begins with the root of $\textsc{Cbt}$ initiating a $\mathit{PIF}$ wave which connects each guest node with its $0$th finger. Notice that, with the exception of one node, the edges in the host network realizing every guest node's $0$th finger are already present. For any guest node $b \neq N-1$, the $0$th finger of $b$ is either (i) a guest node with the same host as $b$, or (ii) a guest node which is hosted by the successor of $\mathit{host}_b$. Edges to guest nodes $0$ and $N-1$ are forwarded up the tree during the feedback wave, allowing the root of the tree to connect them at the completion of the wave, thus forming the base ring and completing every guest node's $0$th finger. The root then executes $\log N-1$ additional $\mathit{PIF}$ waves, with wave $k$ correctly adding the $k$th finger for all guest nodes. After $\mathcal{O}(\log^2 N)$ rounds, we have built the correct $\textsc{Avatar}(\textsc{Chord})$ network.
\begin{figure}
\begin{tabbing}
........\=....\=....\=....\=....\=....\=....\kill
\>\textit{// Execute when $\mathit{phase}_u = \mathit{CHORD}$; If $\mathit{phase}_u = \mathit{CBT}$, then execute} \\
\>\textit{// the original $\textsc{Avatar}(\textsc{Cbt})$ algorithm~\cite{berns_avatar-2016}}.\\
\>\textit{// As part of each round, nodes exchange their local state, including}\\
\>\textit{// $LastWave$, and check for faulty configurations as described in Section \ref{sect:phase}.}\\
1.\>Tree $T$ executes a $\mathit{PIF}(\mathit{MakeFinger}(0))$ wave:\\
2.\>\>\textbf{Propagate Action for $a$:} $\mathit{LastWave}_a = 0$\\
3.\>\>\textbf{Feedback Action for $a$:}\\
\>\>\>\textit{// Let $b$ be the $0$th finger of $a$.}\\
4.\>\>\>\textbf{if} $\mathit{LastWave}_a = \mathit{LastWave}_b = 0$ \textbf{then}\\
5.\>\>\>\>Create the edge $(a,b)$\\
6.\>\>\>\>Forward an edge to node $0$ \\
\>\>\>\>\>\>or $N-1$ (if present) to parent\\
7.\>\>\>\textbf{else} $\mathit{phase}_u = \mathit{CBT}$ (where $u$ is $\mathit{host}_a$) \textbf{fi}\\
8.\>\textbf{for} $k = 1,2,\ldots,\log N - 1$ \textbf{do}\\
9.\>\>Tree $T$ executes a $\mathit{PIF}(\mathit{MakeFinger}(k))$ wave:\\
10.\>\>\>\textbf{Propagate Action for $a$:} $\mathit{LastWave}_a = k$\\
11.\>\>\>\textbf{Feedback Action for $a$:}\\
\>\>\>\>\textit{// Let $b_0, b_1$ be the $k-1$ fingers of $a$.}\\
12.\>\>\>\>\textbf{if} $\mathit{LastWave}_a = \mathit{LastWave}_{b_0} = \mathit{LastWave}_{b_1} = k$ \textbf{then}\\
13.\>\>\>\>\>Create edge $(b_0,b_1)$, the $k$th finger of $b_0$.\\
14.\>\>\>\>\textbf{else} $\mathit{phase}_u = \mathit{CBT}$ (where $u$ is $\mathit{host}_a$) \textbf{fi}\\
15.\>\textbf{od}
\end{tabbing}
\caption{Algorithm 1: PIF for \textsc{Chord} Target from \textsc{Cbt} Scaffold}
\label{algo:chord_scaffold}
\end{figure}
\subsection{Phase Selection}
\label{sect:phase}
The final piece for our self-stabilizing \textsc{Chord} network is to create a mechanism by which nodes can know which algorithm they should be executing: either executing the steps required to build the $\textsc{Avatar}(\textsc{Cbt})$ network, or the steps required to build the \textsc{Chord} target network from an existing \textsc{Cbt} network). We assume each host node $u$ maintains a phase variable $\mathit{phase}_u$ whose value is from the set $\{CBT, CHORD, DONE\}$. When $\mathit{phase}_u = \mathit{CBT}$, a node is executing the algorithm for the $\textsc{Avatar}(\textsc{Cbt})$ network. If $\mathit{phase}_u = \mathit{CHORD}$, then the PIF waves in Algorithm \ref{algo:chord_scaffold} are executed. If $\mathit{phase}_u = \mathit{DONE}$, then a node will take no actions provided its local neighborhood is consistent with a legal $\textsc{Avatar}(\textsc{Chord})$ network.
Determining which algorithm to execute requires a node be able to determine if the configuration they are in now is either completely correct or consistent with one reached by building \textsc{Chord} from \textsc{Cbt}. We define a subset of states under which Algorithm \ref{algo:chord_scaffold} will converge, and then define a predicate which nodes can use to determine if the network is in one of these states.
\begin{definition}
A graph $G$ with node set $V$ is in a \emph{scaffolded \textsc{Chord} configuration} if $G$ is reachable by executing the PIF waves defined by Algorithm \ref{algo:chord_scaffold} on a correct $\textsc{Avatar}(\textsc{Cbt})$ network
\end{definition}
Thanks to the predictability of the \textsc{Cbt} scaffold network, nodes can determine if their state is consistent with that of a scaffolded \textsc{Chord} configuration. Informally, each guest node can determine this by simply checking to see if its neighborhood is a superset of $\textsc{Cbt}$ but a subset of $\textsc{Chord}$, with the first $k$ fingers from $\textsc{Chord}$ present, for some $k \in [0, \log N)$. We define the predicate a node can use for this operation below.
\begin{definition}
Let $\mathit{scaffolded}_b$ be a predicate defined over the local state of a guest node $b$, as well as the state of nodes $b' \in N(b)$. The value of $\mathit{scaffolded}_b$ is the conjunction of the following conditions.
\begin{enumerate}
\item Node $b$ has all neighbors from $\textsc{Cbt}$, each with the proper host and tree identifier (a value set as part of a legal \textsc{Cbt} scaffold network).
\item Node $b$ has last executed the $k$th feedback wave of a\\ $\mathit{PIF}(\mathit{MakeFinger}(k),\bot)$ wave, for some $0 \leq k < \log N$
\item All neighbors of $b$ have either all $k$ fingers present, or $k+1$ fingers (if a child has just processed a feedback wave), or $k-1$ (if parent has not yet processed the current feedback wave), where $k$ is the last feedback wave $b$ has executed
\item Node $b$'s parent has last executed the $k$th feedback wave, and has the first $k$ $\textsc{Chord}$ fingers, or $k-1$ fingers if $b$ has just completed the feedback transition and $b$'s parent has not.
\end{enumerate}
\end{definition}
In every round of computation, all nodes are checking their local state and the state of their neighbors to determine if a faulty configuration is found. This check for faults, along with the $\mathit{scaffolded}_b$ predicate, is used to set the $\mathit{phase}_u$ variable as follows. If a fault is detected and $\mathit{scaffolded}_b = \mathit{false}$, then $u = \mathit{host}_b$ sets $\mathit{phase}_u = \mathit{CBT}$. Furthermore, if any neighbor $v$ has a different value for $\mathit{phase}_v$, then $\mathit{phase}_u = \mathit{CBT}$. We will show in a moment that this procedure is sufficient to ensure the correct algorithm is executed within a short amount of time (i.e. if the configuration is not a scaffolded \textsc{Chord} configuration, then all nodes begin executing the $\textsc{Avatar}(\textsc{Cbt})$ algorithm quickly). Notice that once the correct configuration is built, nodes can execute a final $\mathit{PIF}$ wave to set $\mathit{phase}_u = \mathit{DONE}$. If any node detects \emph{any} fault during this process, it simply sets $\mathit{phase}_u = \mathit{CBT}$. Since $\textsc{Avatar}(\textsc{Chord})$ is locally checkable, at least one node will not set $\mathit{phase}_u = \mathit{DONE}$ during the final $\mathit{PIF}$ wave, and the $\textsc{Avatar}(\textsc{Cbt})$ algorithm will begin.
\section{Analysis}
\label{section:analysis}
We sketch the proofs for our main results below. The full proofs of convergence and degree expansion are contained in the appendix.
\begin{theorem}
Algorithm \ref{algo:chord_scaffold}, when combined with the self-stabilizing algorithm for $\textsc{Avatar}(\textsc{Cbt})$ from Berns~\cite{berns_avatar-2016}, is a self-stabilizing algorithm for the network $\textsc{Avatar}(\textsc{Chord})$ with convergence time $\mathcal{O}(\log^2 N)$ in expectation.
\end{theorem}
\begin{proofsketch}
To prove the convergence time of our algorithm, we first show that if the configuration is not a scaffolded \textsc{Chord} configuration, within $\mathcal{O}(\log N)$ rounds, all nodes are executing the algorithm to build the scaffold \textsc{Cbt} network. We then show that nodes will have built the correct \textsc{Cbt} network within an additional $\mathcal{O}(\log^2 N)$ rounds in expectation, at which point all nodes begin building the target \textsc{Chord} network. We will then show that this process succeeds in $\mathcal{O}(\log^2 N)$ rounds. Putting these together, we get an overall convergence time of $\mathcal{O}(\log^2 N)$ in expectation.
\end{proofsketch}
\begin{theorem}
Algorithm \ref{algo:chord_scaffold}, when combined with the self-stabilizing algorithm for $\textsc{Avatar}(\textsc{Cbt})$ from Berns~\cite{berns_avatar-2016}, is a self-stabilizing algorithm for the network $\textsc{Avatar}(\textsc{Chord})$ with degree expansion of $\mathcal{O}(\log^2 N)$ in expectation.
\end{theorem}
\begin{proofsketch}
By design, any edge that is added to the network when building \textsc{Chord} from \textsc{Cbt} is an edge that will remain in the final correct configuration and therefore does not affect the degree expansion. Furthermore, we know from the original \textsc{Avatar} paper that the expected degree expansion is $\mathcal{O}(\log^2 N)$ when all nodes are executing the \textsc{Cbt} algorithm.
The only new piece we need to consider, then, is to analyze the actions nodes might take when they incorrectly believe, based on their local state, that they are building the \textsc{Chord} network from the \textsc{Cbt} scaffold (a ``false \textsc{Chord}'' phase), which we show can only happen for $\mathcal{O}(\log N)$ rounds. Since adding \textsc{Chord} edges is coordinated with a PIF wave, each guest node $b$ can only increase its degree by one during this time. At most, then, a node may increase its degree by a factor of 2 during this time, leading to the initial degree growth of $2$ during the ``false \textsc{Chord}'' phase.
\end{proofsketch}
\section{Generalizing Our Approach}
\label{section:overview}
Above we have provided an algorithm for using one self-stabilizing overlay network to create another self-stabilizing overlay network. While we are not the first to use this general idea in the construction of overlay networks, we are the first to explicitly define and discuss this approach, which we call \emph{network scaffolding}. To use the network scaffolding approach, one must define several components. In particular, we must define:
\begin{itemize}
\item The \emph{scaffold network}, an intermediate topology which we can construct from any initial configuration.
\item The \emph{target network}, the network topology that we wish to build for use with our final application.
\item A self-stabilizing algorithm for constructing the scaffold network.
\item An algorithm for building the target network when starting from the correct scaffold network.
\item A local predicate allowing nodes to determine whether they should be building the scaffold network or the target network.
\end{itemize}
Our self-stabilizing algorithm from above used $\textsc{Avatar}(\textsc{Cbt})$ as the scaffold network to build a $\textsc{Avatar}(\textsc{Chord})$ target network. To do was relatively straightforward: we defined a way to build \textsc{Chord} from \textsc{Cbt}, and then proved nodes would quickly determine which network they were building.
This network scaffolding approach has been used in some form by other previous work, and we hope it will be extended in future work as well. Our approach heavily depends upon the scaffold network selected. The \textsc{Cbt} network has many desirable properties for a scaffold network when compared to other examples in prior work. These properties include:
\noindent
\textbf{Efficient self-stabilization:} If the scaffold itself is inefficient to build, we cannot expect the target topology to be built efficiently. TCF~\cite{berns_tcf_11} can be thought of as an inefficient scaffold network that requires $\mathcal{O}(n)$ space. \textsc{Avatar}(\textsc{Cbt}) is a logical choice as, prior to this work, it is the only self-stabilizing overlay network we are aware of with efficient stabilization in terms of both time and space.
\noindent
\textbf{Low node degree:} Unlike a real scaffold, we maintain the scaffold edges after the target network is built. Therefore, the scaffold network must have low degree if we wish our final configuration to be so. Again, the suitability of \textsc{Avatar}(\textsc{Cbt}) is apparent, as it requires only a few edges per virtual node (and a logarithmic number of edges per real node).
\noindent
\textbf{Low diameter:} Low diameter allows (relatively) fast communication for adding the target network's edges one at a time. A previous work, \textsc{Re-Chord}~\cite{kniesburges_rechord_11}, used a ``scaffold'' of the \textsc{Linear} network, whose $\mathcal{O}(n)$ diameter contributed to the $\mathcal{O}(n \log n)$ convergence time of their algorithm.
\noindent
\textbf{Predictable routing:} The predictable routing, particularly for communication, allows us to add edges in a metered and checkable fashion. This predictability helps with both design and analysis. It would be interesting to see if a semi-structured overlay network could be used as a scaffold, as semi-structured overlays may be easier to build. To date, little work has been done on self-stabilizing semi-structured overlay networks, but there are several examples of efficient creation of semi-structured networks in non-self-stabilizing settings~\cite{gotte_time-optimal_podc21} which may be interesting starting points for future work.
\noindent
\textbf{Local checkability:} To be able to determine which phase of the algorithm should be executed quickly, without ``wasting'' time and resources adding edges from a faulty configuration, the scaffold should ideally be locally checkable. Some previous overlay networks have used a ``probing'' approach where messages were circulated continuously to try and detect faulty configurations. The risk of this approach in network scaffolding is that nodes may spend too long adding edges from an incorrect scaffold, or take too long to detect a faulty configuration.
\section{Concluding Thoughts}
\label{section:concluding_thoughts}
In this paper, we have presented the first time- and space-efficient algorithm for building a \textsc{Chord} network using a technique we call \emph{network scaffolding}. We discussed considerations for expanding this technique, in particular pointing out considerations and implications for various properties of the scaffold network.
An obvious extension to our work would be to consider building other target topologies using \textsc{Avatar}(\textsc{Cbt}) as a scaffold network. For instance, networks with good load balancing properties or with high resilience to churn could be converted into self-stabilizing variants using \textsc{Avatar} to define the network topology and the \textsc{Cbt} scaffold to build this correct topology. It would also be interesting to investigate the correctness and complexity of this approach when using a more realistic asynchronous communication model.
\bibliographystyle{splncs04}
|
\section{Introduction}
Cirrus clouds are wispy filamentary structures observed at high latitudes of our Galaxy. They were discovered in the far-infrared (FIR) based on IRAS observations in the early 1980s~\citep{Low_etal1984}. However, much earlier they were also identified at optical~\citep{1955Obs....75..129D,1960Obs....80..106D,1972VA.....14..163D,1976AJ.....81..954S,1979A&A....78..253M,1985A&A...145L...7D,Roman_etal2020} and later at ultraviolet wavelengths~\citep{1995ApJ...443L..33H,Gillmon_Shull2006,2015A&A...579A..29B}. It was also demonstrated that the position of the cirrus clouds is well correlated with the position of some molecular clouds~\citep{Weiland_etal1986,deVries_etal1987,Gillmon_Shull2006}.
\par
The cirrus clouds and molecular clouds, as well as the clouds of neutral hydrogen, are well-known to have a rather complex geometry. They possess hierarchy and self-similarity \citep{Bazell_Desert1988,Dickman_etal1990,Falgarone_etal1991,Hetem_Lepine1993,Vogelaar_Wakker1994,Elmegreen_Falgarone1996,Stutzki_etal1998,Sanchez_etal2005,Sanchez_etal2007,Elia_etal2018,Juvela_etal2018}, that is, they can be described as fractals~\citep{Mandelbrot_1983} or even multifractals~\citep{Chappell_Scalo2001,Elia_etal2018,Beattie_etal2019a,Beattie_etal2019b}. Naturally, an important property of a fractal is its fractal dimension, the value of which characterises how a cloud fills the volume. If it is close to $3$, then the cloud fills the volume like a simple 3D object and if, for example, the cloud consists mainly of linear filaments, the dimension should have a value closer to 1. A number of other approaches to characterise the structure of the clouds was also proposed in the literature, such as the power spectrum analysis~\citep{Stutzki_etal1998,Miville-Deschenes_etal16}, the multi-fractal spectrum analysis~\citep{Chappell_Scalo2001}, the statistical analysis based on the probability distribution function~\citep{Donkov_etal2017} and correlation integral value~\citep{Sanchez_etal2005}. In the present work, we focus mainly on the fractal dimension and do not discuss other approaches further.
\par
As the fractal dimension characterises the cloud density distribution, its value is determined by the physical processes that govern the cloud evolution. Quite the number of studies is dedicated to exploring the connection between the value of the fractal dimension and the physical parameters of the clouds~\citep{Sanchez_etal2006} including various hydrodynamic~\citep{Federrath_etal2009,Beattie_etal2019a,Beattie_etal2019b} and magneto-hydrodynamic studies~\citep{Kritsuk_etal2007,Kowal_etal2007,Kritsuk_etal2013}. In general, it was found that turbulent flows structure the cloud in such a way that the fractal dimension spans the whole range from $2.0$ to $2.9$ depending on the Mach number value and how the turbulence itself is implemented in simulations~\citep{Kowal_etal2007,Federrath_etal2009, Konstandin_etal2016,Beattie_etal2019a,Beattie_etal2019b}.
\par
For real clouds observed in astronomical images, a measurement of the fractal dimension is complicated by the fact that the observer usually deals with a 2D intensity distribution with a finite resolution produced by a 3D object with a complicated geometry~\citep{Sanchez_etal2005}. Therefore, strictly speaking, an observational data allows one to only measure the fractal dimension of the projection of a cloud $D_\mathrm{p}$ and not the fractal dimension $D$ of the cloud itself. This problem was thoroughly addressed by~\cite{Sanchez_etal2005} where modelled clouds were considered. An important conclusion of their work is that for a fixed value of the $3D$ fractal dimension $D$, one can obtain different values of $D_\mathrm{p}$ depending on image resolution. \cite{Sanchez_etal2005} also clearly showed that $D$ does not need to be equal to $D_\mathrm{p}+1$, but $D$ tends to be slightly greater than this, e.g. for $D_\mathrm{p}\approx1.35$ it was found that $D$ should be in a range from $2.5$ to $2.7$. It was also found that $D_\mathrm{p}$ should decrease with an increase of $D$ (see their figure 8). In a subsequent work, \citet{Sanchez_etal2007} obtained that some other observational parameters, such as noise level, can also influence the obtained values of the fractal dimension. In recent studies by~\citet{Beattie_etal2019a} and \citet{Beattie_etal2019b}, the authors used hydrodynamic simulations to study how the calculated values of the projection fractal dimension $D_\mathrm{p}$ depend on the volumetric fractal dimension $D$ \textit{and} the Mach number of the flow. It was found that $D$ should be in a range from $D_\mathrm{p}+1/2$ to $D_\mathrm{p}+1$ for high and low Mach number limits, respectively. In contradiction to the results of~\cite{Sanchez_etal2005}, \citet{Beattie_etal2019a} and \citet{Beattie_etal2019b} found that $D$ should increase with an increase of $D_\mathrm{p}$ (see their figure 6). Here we should note these authors calculated the fractal dimension using a mass-length fractal dimension method that differs from the perimeter-area method which is commonly used for measuring the fractal dimension of actual astronomical clouds. Nevertheless, we admit that the simulation setup of~\citet{Beattie_etal2019b} with real clouds and physical processes, which govern their evolution, is much more appealing from a physical point of view than experiments with static geometric fractals with no underlying physics which were considered by~\citet{Sanchez_etal2005}.
\par
For our convenience, hereinafter we abandon the lower index in $D_\mathrm{p}$ and write it simply as $D$.
\par
Most of previous studies on fractal dimension of the cirrus and molecular clouds were carried out based on infrared~IRAS and \textit{Herschel} data~\citep{Bazell_Desert1988,Dickman_etal1990,Vogelaar_Wakker1994,Juvela_etal2018,Beattie_etal2019b}, although a number of other observations were also utilised~\citep{Falgarone_etal1991, Hetem_Lepine1993, Vogelaar_Wakker1994,Sanchez_etal2005}.
\par
The general approach to measure the fractal dimension, used in those studies, is as follows. First, the observer chooses some contours with a fixed level of intensity (or some set of them), calculates the perimeter $P$ and area $A$ of the structure enclosed by a contour, and approximates the perimeter and area dependence by a power function of the following form: $P \propto A^{D/2}$. This approach turned out to be rather fruitful. In general, it was shown that the structures, measured in such a way, indeed form an almost straight line on the $(\log P, \log A)$ plane with a resultant error on $D$ of about several hundredths and smaller~\citep{Bazell_Desert1988,Dickman_etal1990,Falgarone_etal1991,Vogelaar_Wakker1994,Sanchez_etal2005}. As for the actual values of the fractal dimension, it was found to span a range from 1.1 to 1.7 with preferred values of about $1.3-1.4$ if we account for all results presented in the literature, see Sect.~\ref{sec:comparison}.
\par
One of the problems with the interpretation of fractal dimension measurements is that different data can have different image resolution (pixel scales), as well as some other observation specifications, such as different point spread functions (PSFs), etc. As was already mentioned, \citet{Sanchez_etal2005} showed that differences in resolution can affect the obtained values of the fractal dimension and, in some cases, a low resolution can even lead to a significant underestimation of the fractal dimension (see figure~7 in~\citealt{Sanchez_etal2005}). In terms of this problem, it is important to examine how the previous results hold if we use some new data with better observational specifications.
\par
In recent work by~\cite{Roman_etal2020}, the authors distinguished a number of cirrus clouds in several fields of the Sloan Digital Sky Survey (SDSS) Stripe82 \citep{Abazajian_etal2009,Fliri_Trukillo2016} after some additional image processing, which included creating mosaics, masking, and subtracting the sky background and instrumental scattered light. In the present article, we use the same fields as presented in~\cite{Roman_etal2020}, where the cirrus clouds have already been distinguished, to calculate their fractal dimension. In this we pursue two goals.
\par
The first goal is to actually calculate the fractal dimension of the clouds in the optical. To our knowledge, measuring the fractal dimension in the visible has not been done before. At the same time, most of previous investigations used IR data as a source material with much worse resolution. As mentioned above, these factors are important for fractal dimension measurements. We should also note that, from a physical point of view, it is important to explore whether the results of fractal dimension measurements depend on wavelength. And if they do, that means that the upcoming physical simulations of the clouds should take this into account when comparing the simulated data with actual observational data. In this study, we also compare the obtained results for the optical with those obtained for the IR for the same fields and, in general, with the results of previous works.
\par
The second goal of this paper is rather methodological. Optical images have better resolution and, hence, there are many small-scale features present on such images that cannot be detected in the IR. Thus, optical data can be used to directly measure to what degree the various effects, such as image and angular resolution (PSF), can change the fractal dimension values. This is also important for future studies of the geometric properties of dust clouds.
\par
On the other hand, although the use of optical images has many advantages, bright sources in the optical are much more numerous as compared to the IR. Masking such sources often introduces empty areas which can shred up the clouds one tries to measure. For this reason, to reliably measure the fractal dimension, one should study how the masking itself affects the measurement. We pay special attention to this matter in course of the present article and prepare several additional simple experiments to estimate the effect of masking.
\par
The structure of the present paper is as follows.
In Section~\ref{sec:data}, we describe the optical and IR data we used to calculate the fractal dimension of the clouds.
In Section~\ref{sec:methods}, we provide a thorough description of our algorithm to calculate the fractal dimension and give a description of a Monte-Carlo simulation setup which we used to estimate the errors for our fractal dimension measurement, as well as to estimate how various observational specifications affect the results.
In Section~\ref{sec:results}, we present results of our fractal dimension measurements for the optical and IR data and compare the obtained results with those from the literature.
In Section~\ref{sec:effects_that_change_D}, we discuss the influence of various effects on our fractal dimension measurements, including angular resolution and masking.
In Section~\ref{sec:discussion}, we discuss the physical reasons for the observed differences of the fractal dimensions in the optical and IR and compare our results with those obtained in previous studies.
We summarise our results in Section~\ref{sec:sum}.
\section{Data}
\label{sec:data}
\begin{figure}
\includegraphics[width=0.95\columnwidth]{imgs/sky.png}
\caption{Position of the fields under consideration in the plane of the sky. The dashed lines depict the Stripe82 borders. The frames of the fields are colour-coded for a better distinction of adjacent fields.}
\label{fig:skypos}
\end{figure}
\subsection{Optical data}
For optical data, we use 16 different fields in the $g,\,r,\,i,$ and $z$ bands which were prepared by~\cite{Roman_etal2020} to specifically distinguish and study cirrus clouds. A detailed description of each field is summarised in table~2 of \cite{Roman_etal2020}. Here we briefly provide some important details on the data reduction pipeline, which was designed to carefully probe cirrus clouds in the Stripe82 fields, and on our further processing.
\par
The original raw fields were taken from the SDSS Stripe82 database~\citep{Abazajian_etal2009}. The Stipe82 data itself is obtained with the 2.5-meter telescope of the Apache Point Observatory and covers a stripe in the sky with an angular area of 275 square degrees within $-50^\circ < \mathrm{RA} < 60^\circ$ and $-1.25^\circ<\mathrm{Dec.}<1.25^\circ$. The exposure time was about one hour for each field. The original pixel size of the images was 0.396 arcsec per pixel.
In addition to the standard SDSS data reduction, the fields were stacked by~\cite{Fliri_Trukillo2016}, who tried to carefully preserve the characteristics of the background which represents a sum of several diffuse light components. Then, the residuals of the co-adding process were removed in~\cite{Roman_Trujillo2018}. In~\cite{Roman_etal2020}, the fields were further processed to account for the instrumental scattered light produced by the extended PSF wings of the stars. Also, in~\cite{Roman_etal2020}, accurate masking of all sources that contaminate the cirrus emission was carried out and the fields were rebinned from the original SDSS pixel size to 6 arcsec per pixel. The relative location of all fields in the sky is shown in Fig.~\ref{fig:skypos}. As can be seen from this figure, the fields are distributed all over Stripe82 and do not lie in one area or direction. The fields span the galactic latitude $b$ between $-35^\circ$ and $-61^\circ$ for Field\#1 and \#5, respectively.
\par
One important difference from~\cite{Roman_etal2020} is that we have 8 large fields instead of their smaller 16. The reason is that we decided to compose the overlapping fields into a single mosaic because the total image area can be an important factor for measuring the fractal dimension of a cloud. Specifically, we merged two sets of the fields, namely Fields\#6-10 and \#11-15. These fields form a sequence of tiles that are located next to each other in the original Stripe82 data. We refer to the resulted composed fields as `\#join1' and `\#join2' in our further discussion. The notation of the remaining Fields\#1-5 and \#16 stays unchanged. The reason why they were cut in the first place is that \cite{Roman_etal2020} aimed to have more points with different densities of dust to better sample the correlation between density and optical colours.
There is one major problem with the large fields which stems from the SDSS data reduction pipeline: it can cause over-subtraction of the diffuse light on a scale of several acrminutes, thus not allowing to correctly investigate cirrus clouds of similar sizes. More about this issue can be found in the original paper by~\cite{Roman_etal2020} and later in Sect.~\ref{sec:discussion}.
\subsection{IR data}
Since most of previous studies on the fractal dimension utilised IR data, it is interesting to compare the fractal dimension of the clouds that we observe in the optical with their counterparts in the IR. To this aim, we use the Improved
Reprocessing of the IRAS Survey (IRIS, \citealt{IRIS}) and \textit{Herschel} \citep{Viero_etal2014} data. IRIS provides reprocessed IRAS data with a slightly improved angular resolution (4.3 arscmin at 100~$\mu$m), better calibration and zodiacal light subtraction. For all of the optical fields we extracted their IR counterparts from the IRIS 100~$\mu$m database using special {\small IDL} routines, which were designed to produce IRIS mosaics\footnote{Available at \url{https://www.cita.utoronto.ca/~mamd/IRIS/IrisDownload.html}}. For Field\#5, we also analyse the~\textit{Herschel} 250~$\mu$m data from~\cite{Roman_etal2020} where the \textit{Herschel} Stripe82 Survey (HerS, \citealt{Viero_etal2014}) was used to identify whether the diffuse emission, observed in the optical, is due to the Galactic dust or there are some other sources responsible for it. This field has its own mask to filter out all sources, which are not related to dust, see fig.~7 in \cite{Roman_etal2020} and panel (b) of Fig.~\ref{fig:f5_comp}. This figure also shows an optical image in the $g$ band and the IRIS data, so they can be directly compared with each other. Note a good congruence of the cirrus contours in the separate subplots and a difference between the masked regions in panels (a) and (b).
\par
To ensure that our processing procedure is not affected by some internal flaws, we also exploit some IR fields for which $D$ has been measured in previous studies. We consider two of the five IRAS 100$\mu$m fields studied in~\cite{Dickman_etal1990}, namely the Chameleon and Taurus fields. We select an area within RA and Dec. coordinates as in \cite{Dickman_etal1990}. The reason why we consider only these two specific fields from~\cite{Dickman_etal1990} is because they are located inside a single IRAS plate and, thus, there is no need to compose several areas to produce a resulted image for our comparison analysis.
\par
There are several major points we should mention concerning the IR data. The IRIS data has a relatively low angular resolution with a pixel size of 90~arcsec. The IRIS PSF is much more extended than the PSF for the optical data, with the characteristic scale FWHM = 4.3 arcmins. As we show below, both these factors are important for measuring the fractal dimension. The \textit{Herschel} data has a much higher image resolution (a pixel size of 6 arcsec) and a much better PSF (FWHM = 18.1 arcsec).
\par
Below we consider two types of IR images: one with the optical mask applied and the other without any mask. The latter option should be considered to correctly compare the results of the present study with the results from the literature. In most previous studies a masking procedure was not actually carried out (see~\citealt{Bazell_Desert1988,Dickman_etal1990,Juvela_etal2018}). Instead, all the external sources were usually cut out from the analysis based on their size, that is, all contours, which were smaller than some fixed size, were excluded after the perimeters and the area of the clouds had been measured.
\begin{figure*}
\includegraphics[width=1.95\columnwidth]{imgs/Field5_comparison.png}
\caption{IR and optical data counterparts for Field\#5. All images show the same region. Panel (a) displays the data in the $g$ band with the 6~arcsec resolution and with the colourbar in counts. The red frame limits the part which is later shown in Fig.~\ref{fig:pipe}. Panel (b) showcases the \textit{Herschel} counterpart with the same resolution but with a different mask (the colourbar units are MJy/beam). Panel (c) represents the IRIS 100~$\mu$m data with the 1.5~arcmin resolution and in units MJy/sr. Finally, panel (d) shows the data in the $g$ band, rebinned to the 1.5~arcmin pixel size and convolved with the IRIS PSF, the units are counts. Note that the IRIS data in panel (c) is depicted without mask, which is exactly the same as the one in panel (d).}
\label{fig:f5_comp}
\end{figure*}
\section{Methods}
\label{sec:methods}
\subsection{Method description}
In order to compute the fractal dimension $D$, we employ a perimeter-area method following \cite{Bazell_Desert1988,Dickman_etal1990,Falgarone_etal1991,Hetem_Lepine1993,Vogelaar_Wakker1994,Sanchez_etal2005,Juvela_etal2018}. To measure the fractal dimension of a cloud, one should find the power index in the relation between the cloud perimeter $P$ and its surface area $A$, which has the following form \begin{equation}\label{eq:AP}
P =K\times A^{D/2}\,,
\end{equation}
where $D$ is the sought value of the fractal dimension and $K$ is the intercept coefficient. While the fractal dimension $D$ can be directly connected to some physical properties of the cloud (see Introduction), the intercept $K$ should characterise the general shape of the cloud. However, we note that, based on the results of previous studies~\citep{Dickman_etal1990,Vogelaar_Wakker1994}, the intercept values obtained for the real clouds can hardly be used to associate the shape of the molecular clouds with the shape of some simple geometric objects like ellipses or circles (see discussion section in~\citealt{Dickman_etal1990}).
\par
The above described method cannot be applied as is because the optical fields have a significant number of masked sources within individual clouds. In principle, this circumstance can change the actual fractal dimension in some non-trivial way. For example, suppose we have some data which satisfies eq.~(\ref{eq:AP}). If we start extracting ``holes'' from such clouds, then the resultant dependence will have a greater fractal dimension $D=2\log P/\log A$, because their area will decrease, while the perimeter will increase. Thus, some additional steps are needed to mitigate this effect.
The sequence of steps for $D$ estimation for each image is illustrated in Fig.~\ref{fig:pipe} and described below in the following nine-points list. Hereinafter, values of $A$ and $P$ are given in square pixel and pixel units, respectively, unless otherwise explicitly stated. Throughout the text all logarithms are natural. It is important to note here that some steps are optional, while other depend on subjective parameters. In the next subsection, we take them into account using Monte-Carlo simulations to verify whether the choice of the parameters can affect a fractal dimension measurement. The following list summarises the details of our algorithm:
(i) We select extended parts of the mask which are, at first sight, too large to be properly interpolated inside. These parts appear as large white areas in panel (b) of Fig.~\ref{fig:pipe} (also see panel (a) and (d) in Fig.~\ref{fig:f5_comp}). The second reason, why this step is necessary, is that such masked areas can, along with the image borders, shred up individual cirrus clouds or cut off some of their parts. If this happens, the $D$ value can change.
(ii) This step is optional and includes an interpolation of all mask parts which have not been previously selected in step (i). We use a linear interpolation method which produces an overall smooth image as shown in panel (b) in Fig.~\ref{fig:pipe}. Note a small number of artefacts for the ``holes'' of medium size.
(iii) Then we select a brightness contour level which corresponds to a lower boundary of the cirrus emission. All pixels of individual clouds, which we consider in the next steps, should have an equal or greater brightness. All the selected contours for the level $27$~mag arcsec$^{-2}$ in the $r$ band are displayed in panel (c) of Fig.~\ref{fig:pipe} by blue, light green and yellow colours and the difference between them is explained below.
(iv) This step is optional. We test the findings of \cite{Roman_etal2020} that cirrus can be filtered out from other sources by its optical colours. Thus, we apply an additional mask to select only those pixels that satisfy the $(r - i) < 0.43 \times (g - r) - 0.06$ condition (see equation (1) in \citealt{Roman_etal2020}). We note, however, that all emission regions from the analysed fields are expected to be actual cirrus clouds, thus these measurements are merely performed out of methodological interest, and the results obtained this way are not incorporated in the final $D$ estimation. We discuss this matter more thoroughly in Sec.~\ref{sec:color}.
(v) As mentioned above, additional ``holes'' inside cirrus clouds due to a mask can increase the fractal dimension. In this step, we decide to ``fill in'' all mask parts which are completely surrounded by the pixels that have been previously found to belong to some clouds. Hence, we assume that such mask parts are actual parts of the cirrus. If the ``holes'' have already been interpolated, then this step has no effect regardless of the choice to ``fill in'' or not. We note that individual clouds can still have ``empty holes'' inside their body. Such holes are not due to the masking of foreground sources. They appear as a result of the combined effect of the cloud geometry, projection effects, and the selected contour level (see figure~1 in \citealt{Sanchez_etal2007}). Panel (d) of Fig.~\ref{fig:pipe} showcases how the clouds, distinguished in panel (c) of the same figure, are transformed after applying the described procedure.
(vi) All the selected pixels are segmented to individual clouds using simple connectivity rules, where two pixels can be connected by a common side, but not by a common corner. Individual clouds are represented by different colours in panel (d). One can note the variety of shapes and sizes of individual segments.
(vii) The extracted individual clouds are then filtered in two different ways. First, we remove all noisy segments, the area $A$ of which is less than some small value, e.g. 5 square pixels (panel (e) of Fig.~\ref{fig:pipe}, shown in yellow). Second, we choose whether we should use all the segments which touch a border of the image or the extended mask parts selected in step (i) even by one pixel. This is explained by the fact that one can only analyse part of such a cut-off cloud. If we consider only part of a cloud, its fractal dimension $D_\mathrm{part}$ can differ from the fractal dimension $D$ of the whole cloud. This effect is investigated in greater detail in Sect.~\ref{sec:discussion}. The segments, removed after the first and second procedures, are shown in yellow and light green in panel (c) of Fig.~\ref{fig:pipe}, respectively.
(viii) For all remaining clouds, we find the area $A$ and perimeter $P$ using the method {\small regionprops} from the {\small skimage} package. The linear regression is fitted to data ($\log A$, $\log P$) using an ordinary least-squares method. The slope of the regression $D/2$ provides us with the desired fractal dimension value, as well as the intercept $K$. This is illustrated in panel (e) of Fig.~\ref{fig:pipe}. We also depict the points for small (red) and sliced (magenta) clouds, filtered out in the previous step. We note that the cut-off clouds follow the regression line fairly well. In contrast, the inclusion of ``small'' noisy clouds into the fit bends down the lower left tail of the regression line, which leads to an increase of the $D$ value. This is due to a much lower perimeter-to-area ratio of such noisy clouds with respect to larger ones.
(ix) Since the decision of which clouds should be filtered out because of their size is, strictly speaking, an arbitrary one, we introduce an additional parameter which we call the `tailcut' parameter. This parameter defines which portion of the small clouds in the regression should be removed from our analysis. The value of the parameter can be set in a range from 0 to 1 assuming that 0 corresponds to the lower limit of the $ \log A $ range, and 1 corresponds to the upper one. The `tailcut' parameter has a simple geometric interpretation. In panel (e) of Fig.~\ref{fig:pipe}, its value can be represented by a vertical line. By moving this line to the right, i.e. increasing the `tailcut', one can measure how stable the regression fit to this sort of filtering is. The example of this process is shown in panel (f) of the same figure. The error bars correspond to the errors of the linear regression fitting. We note that, for this selected part of the cirrus, the measurement of the fractal dimension is stable within the margin of error until we start filtering out clouds with an area $\log A > 5.5$, or $A > 250$ square pixels.
For the IR counterpart images from the \textit{Herschel} and IRIS surveys, step (iv) was excluded from the sequence, since we would like to use the IR data as is for better comparison with the literature.
Note, that, contrary to the previous studies, we decided not to use all the levels for a given field image to fit one single regression line. Instead, we measure the fractal dimension for each brightness level under consideration. This produces a less stable measurement, but, at the same time, gives us a better understanding of whether the fractal clouds have the same properties at different brightness levels in the optical bands.
\par
The presented method can be used as is. However, to estimate the effect of individual steps on a fractal dimension measurement and to estimate its uncertainties, we run a Monte Carlo (MC) simulation, the setup of which is described below.
\begin{figure*}
\includegraphics[width=1.95\columnwidth]{imgs/pipel_new.png}
\caption{An illustration of the method pipeline for part of Field\#5 shown in Fig.~\ref{fig:f5_comp} and for the selected contour level=$27$~mag arcsec$^{-2}$.
For the description of the individual panels see Sect.~3. In panel (c) blue colours show the selected contours for a further analysis, light green -- contours, which touch the mask, deep green -- large mask and image borders, yellow -- small filtered clouds with an area $A<30$~square pix. Different clouds in panel (d) are shown by different colours, the ``holes'' are filled (see text for details). In panel (e), blue squares correspond to individual clouds from (d), magenta circles are the clouds which touch the large mask (see step (vii) for details) and then filtered out, red points -- all small segments, which are filtered out by the `tailcut' parameter, depicted by the solid vertical line. Note that points from only part of the original Field\#5 are shown and, thus, the derived slope is just an example and not the real $D$ for the whole field.}
\label{fig:pipe}
\end{figure*}
\subsection{MC simulation}
\par
For each optical field, we generate a random set of parameters, according to the rules described below, and measure the fractal dimension $D$, intercept $K$, and their errors. The regression line cannot be fitted for some combinations of the parameters, e.g. when all clouds are filtered out. We consider such realisations unsuccessful and do not take them into account. We run 10000 successful realisations and estimate an average value of $D$ and its $1\sigma$ error. In Fig.~\ref{fig:mcparamdep}, one can see a typical example of how individual steps adjust the $D$ value for Field\#3. A detailed description for each of these steps is as follows.
First, we randomly choose which optical band from $g,r,i,z$ will be used in this particular realisation. In this step, we assume that all bands should equally contribute to the final estimation of the fractal dimension. However, in the course of this study, we found that the results, obtained for the $z$ band, significantly differ from others. This can be interpreted by the fact that the $z$ band is much shallower as compared to the $gri$ bands: the surface brightness limit $\mu_{lim}(3\sigma; 10\arcsec\times10\arcsec)=29.1,28.6,28.2,26.6$ mag\,arcsec$^{-2}$ for the $g$, $r$, $i$ and $z$ bands, respectively (see \citealt{Roman_etal2020}). Therefore, in our further discussion we do not take into account the results obtained for the $z$ band. This matter will be addressed in greater detail in Sec.~\ref{sec:diffbands}. Second, we choose with equal probability whether we should include the cut-off clouds or not. Next, the size of the extended mask parts, described in step (i), is randomly picked out from an interval of 200 square pixels to one third of the field area. Each mask segment, the area of which is larger than the selected size, is then considered to be too large to be interpolated.
We add a thin frame with a width of 5 pixels around each field. This frame is also used as an additional mask to filter out clouds in step (vii) if they touch it. Next, we select whether we should interpolate all the remaining mask parts or not (see step (ii)) with the equal probability. As regards the interpolation method, we found that the choice of the method used has a negligible or no effect on our results and, thus, we can use a linear interpolation for simplicity.
The brightness contour level is selected from a uniform distribution within a range from 0 to the maximum value. Since we work with images in different bands, the distribution range is not the same for all bands, with a trend of the maximum brightness to increase towards the $z$ band. We decided to use a wider range of brightnesses, which correspond to all individual bands and all possible fields (even if it is possible that some image has 0 pixels for a selected brightness level, because this situation is rare and we do not take it into account in the final measurement anyway). It is important to note that this is not true for the IRIS counterparts, where some fields have a zero intersection in the brightness levels. Next, we decide whether to filter out a cirrus or not to do that based on its colour with the probability $p=0.66$\footnote{We have shifted the probability distribution here because the colour filtering is an experimental step which has not been done before. Thus, we decided to collect more realisations without such filtering.} for the latter (see step iv). Also, we decide whether to fill in the inner mask ``holes'', as explained in step (v). The `tailcut' parameter from step (ix) is picked out randomly from the $[0.0, 0.95]$ range and then squared, in order to shift the distribution to filter out small clouds more frequently than larger ones. We note that we do not choose the size for the initial filtering of small clouds in step (vii) at random, but conservatively filter out all segments with $A \leq 5$~square pix to speed up the process. The reason behind this is that the `tailcut' parameter does the same filtering. Finally, we do not perform a regression fit if there are three or less points to fit. Such realisations are considered unsuccessful.
As was already mentioned in the previous section, for the IRIS and \textit{Herschel} data, the colour filtering is not applicable and, hence, it is not carried out in our MC simulation too. Another important difference from the optical data is that because of the worse resolution, only a constrained set of the parameters for the IRIS fields can lead to a successful measurement of the fractal dimension and, therefore, it is difficult to collect all 10000 realisations. Moreover, since some fields are extensively brighter and some images contain only a small number of clouds, selecting cirrus contours from the same distribution often results in an empty sample. Thus, for $100~\mu$m counterparts we decided to run the same number of realisations as for the optical bands, neglecting the fact whether a particular realisation is successful or not. All other parameters in the MC simulation for the IR counterparts remain the same.
\begin{figure*}
\includegraphics[width=1.95\columnwidth]{imgs/mc_param_dep.png}
\caption{Dependence of the fractal dimension $D$ on the MC parameters for Field\#3. Each parameter is illustrated by two subplots. The first subplot shows how parameter values affect the resulted $D$ estimation, while the second subplot represents the parameter distribution using the corresponding histogram. For discrete parameters, the data is presented as a $D$ density distribution for each option; for continuous parameters --- as standard deviation density contours, where individual dots are outliers. All histograms are normalised in such a way that the height of the largest bar equals unity and all density contours are drawn for the same bins as in the related histogram but slightly smoothed. For continuous parameters, the vertical limits in the left plot are the same as the horizontal ones in the right plot. In density plots, the blue line is for {\it False} option and the orange one is for {\it True}. The parameters have the following notation: `band' is for data in an optical band (the $z$ band is excluded, see Sec.~\ref{sec:diffbands} for details), `colour' is for colour filtering, `level' is for brightness contour in counts, `mask size' is for large extended mask size limits, `interp' is for option whether interpolate mask or not, `fill holes' parameter regulates whether we should fill the mask inside the contours or not, `tailcut' is the same as in the main text, `no touch' is to filter out clouds which are likely to be shaded by the extended mask, see step (vii).}
\label{fig:mcparamdep}
\end{figure*}
\section{Results}
\label{sec:results}
Table~\ref{table:main_parameters} and Fig.~\ref{fig:allmc3} summarise the results of our fractal dimension measurements from the MC simulation for the clouds from the selected fields.
We note that these table and figure present not only results of direct fractal measurements of optical data, but also results of other measurements that are important for the current analysis. Namely, the columns 5 and 6 present the fractal dimensions and intercepts for the optical and \textit{Herschel} data with the image resolution reduced to the IRAS data (panel (c) of Fig.~\ref{fig:allmc3}), while the columns 7--10 show the values of fractal dimension $D$ and intercepts $K$ for the IRIS counterpart with the optical or \textit{Herschel} mask applied where possible (columns 7 and 8, panel (b) of Fig.~\ref{fig:allmc3}, black crosses) and without the mask (columns 9 and 10, panel (b) of Fig.~\ref{fig:allmc3} magenta diamonds). The values presented in the columns 5--10 are provided to validate our approach and also to make comparison with previous studies more direct. In this section, we mainly discuss the new results of fractal dimension measurement that are done based on the optical data only (columns 3 and 4, panel (a) of Fig.~\ref{fig:allmc3}), while we postpone a detailed description of the other measurements and how they were carried out to the following section, where we analyse the various factors that contribute to a fractal dimension measurement.
\par
For optical data, the fractal dimension value, averaged across all fields, is $\langle D \rangle=1.69^{+0.05}_{-0.05}$ with a remarkably small characteristic spread of values $\sigma(D)=0.02$, if the outlier for Field\#1 is not taken into account ($\sigma(D)=0.07$ if it is taken into consideration). For the same clouds, measured in the IRIS fields, the fractal dimension is considerably smaller $\langle D \rangle=1.48^{+0.10}_{-0.07}$, $\sigma(D)=0.10$ with the optical mask applied and $\langle D \rangle=1.38^{+0.07}_{-0.06}$, $\sigma(D)=0.09$ without it. A typical measurement error obtained for all MC simulation realisations appeared to be smaller than 0.1 for both the optical and IR fields. Note, that the reported average values are calculated as a mean of the $D$ values from Table~\ref{tab:fd_result} without any correction for the individual size of a particular field or the number of clouds.
\par
First of all, we should emphasise that the measured fractal dimensions appear to be less than 2 for all the cirrus clouds under consideration. At the same time, this value does not change much if we consider different spatial scales. This claim is illustrated in panel (e) of Fig.~\ref{fig:pipe} for Field\#5, where it can be seen that the dependence of the perimeter on the area is clearly has small scatter from the best-fit line, and in Fig.~\ref{fig:lev_dep} for all fields, which illustrates that $D$ remains almost constant while the brightness of the contours and the scale vary. Therefore, we obtain that all the selected cirrus clouds are of a fractal nature. While this result is not conceptually new and has been known for a long time starting from~\cite{Bazell_Desert1988}, it is still important to note because, strictly speaking, 1) we analyse the clouds which have not been considered for measuring their fractal properties, and, most importantly, 2) we mainly concentrate on optical cirrus, whereas most of the previous studies were focused on IR data.
\par
Our value $\langle D \rangle=1.38^{+0.07}_{-0.06}$, averaged over all IRIS fields with no mask, seem to be slightly greater as compared to the average values obtained in previous works. For convenience, in Table~\ref{tab:previous} we collected all the results of fractal dimension measurements from previous studies where the same perimeter-area based method was used. As one can see, depending on the data, there is a range of values the fractal dimension can take. For example, \cite{Juvela_etal2018} obtained that the average fractal dimension of the dust clouds is about $D\approx1.2$ and spans a range from $1.05$ to $1.40$ based on the \textit{Herschel} data. \cite{Bazell_Desert1988} measured the fractal dimension specifically for cirrus clouds using IRAS data and found that the clouds have a fractal dimension of about $D\approx1.26$, with a range from $1.12$ to $1.4$. \citet{Vogelaar_Wakker1994} found for IRAS data that the fractal dimension of different clouds ranges from 1.2 to 1.6. While our average $\langle D \rangle$ seems to be slightly larger than those, the values, obtained for individual clouds, seem to fall into the same range of values. Two possible exceptions are Field\#2 and \#3 where the fractal dimension $D$ is about 1.5, which is around an upper limit of the values obtained in the literature, but, overall, our values of $D$ are still consistent with those from~\cite{Vogelaar_Wakker1994}.
\par
For the same IR data but with the mask applied, the $D$ values become greater by about $0.1$, but even in such a case they are still within an interval from 1.2 to 1.6. One possible outlier is Field\#2 where, for the IRIS image with the optical mask, $D=1.65$ is slightly larger than the mentioned upper limit 1.6 and is also the largest among all the fields under study. Taking into account that this field has the lowest number of successful realisations in MC and, as can be seen from Table~\ref{tab:fd_result}, the original image has a significantly lower $D\approx1.5$, we can address this outlier to an effect of optical mask influence. We discuss the impact of masking on the fractal dimension in greater detail in Sect.~\ref{sec:effects_that_change_D}.
\par
For both the Taurus and Chameleon test fields, we obtain the $D$ values which are greater than the previously measured ones: $1.36 \pm 0.06$ for Chameleon and $1.33 \pm 0.03$ for Taurus versus $1.280 \pm 0.016$ and $1.230 \pm 0.004$ in \cite{Dickman_etal1990}, respectively. However, all these measures are relatively close to each other. It is also important to note that the error budget of~\cite{Dickman_etal1990} takes into account only the error of regression fitting and, hence, is unreliably small. It is reasonable to expect that accounting for the differences, which are introduced by slightly different levels of selection or due to the filtering of small clouds (see discussion in Sect.~\ref{sec:tailcut}), should result in a larger margin of error. Thus, the results, measured here for the test images, are consistent with those from the literature.
\par
As to the intercepts, we found that the average values are~$\langle K \rangle=1.33$ and~$\langle K \rangle=1.63$ in the optical and IR, respectively, while the spread of the values across the different fields is also considerably small: $\sigma(K)=0.09$ and $\sigma(K)=0.24$ for the optical and IR data, respectively. Our values of $K$ are smaller than those found by~\cite{Vogelaar_Wakker1994} ($1.7<K<3$, see their table~3). Perhaps, the difference in the intercepts is associated with the fact that the measured fractal dimension is, on average, slightly larger in our work, but we do not pursue this question further.
\par
Concerning the striking difference between the results in the optical and IR, there are several possible reasons for that. First of all, there is a possibility that the measured fractal dimension values somehow reflect the way the fields were processed. The major factor here is obviously the mask as it substantially changes the geometry of small clouds and also affects larger ones, although to a smaller degree. As shown by~\cite{Sanchez_etal2005}, the image resolution (i.e. how many pixels are in a cloud of a fixed size) can also change the fractal dimension value depending on the actual 3D fractal dimension. Finally, there may be some physical reasons associated with the dust properties and dynamics. We thoroughly discuss this matter in the next Sect.~\ref{sec:effects_that_change_D} and in Sect.~\ref{sec:discussion}.
\par
Here we also note that the spread $\sigma$ of $D$ values for the clouds from different fields is rather small for both the optical and IR data. It is an interesting result since the fields are not connected in any way, see Fig.~\ref{fig:skypos}. Thus, in terms of the fractal properties, all the clouds observed seem to be similar to each other. This fact supports the idea that these cirrus clouds are close to each other in terms of the physical processes that shape them.
\par
\begin{table*}
\caption{Measured $D$ values for 8 examined optical fields and their IRIS counterparts. In addition, we include the results for \textit{Herschel} Field\#5 and two IRAS testing fields, Chameleon and Taurus. Second column lists an area of the image, columns (3) and (5) list the fractal dimensions and intercepts for the optical (or \textit{Herschel}) data; (5) and (6) lists the results for the optical data rebinned to 90 arcsec and reconvolved with the IRIS PSF; columns (7) and (8) list the results for the IRIS counterpart with the optical or \textit{Herschel} mask applied where possible, columns (9) and (10) --- for the same data without a mask. Column (11) is the difference between (5) and (7). All presented values of $D$ indicate a median for the MC realisations, the upper and lower error boundaries are 1~$\sigma$ from the corresponding percentiles.}
\label{table:main_parameters}
\begin{tabular}{c|c|c|c|c|c|c|c|c|c|c}
\hline
Field & $A$, deg$^2$ & $D$ & $K$ & $D_{90}$ & $K_{90}$ & $D_\mathrm{IRIS}$ & $K_\mathrm{IRIS}$ & $D_\mathrm{IRIS}^\mathrm{orig}$ & $K_\mathrm{IRIS}^\mathrm{orig}$ & $D_{90}-D_\mathrm{IRIS}$\\
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9) & (10) & (11)\\
\hline
\#1 & 2.0 & $1.88^{+0.02}_{-0.01}$ & 1.03 & $1.41^{+0.12}_{-0.12}$ & 1.83 & $1.49^{+0.16}_{-0.08}$ & 1.61 & $1.34^{+0.04}_{-0.05}$ & 2.02 & -0.09 \\[0.1cm]
\#2 & 0.8 & $1.66^{+0.06}_{-0.06}$ & 1.41 & $1.54^{+0.08}_{-0.04}$ & 1.49 & $1.65^{+0.09}_{-0.06}$ & 1.32 & $1.50^{+0.12}_{-0.10}$ & 1.61 & -0.10 \\[0.1cm]
\#3 & 1.8 & $1.70^{+0.04}_{-0.05}$ & 1.33 & $1.54^{+0.07}_{-0.09}$ & 1.53 & $1.59^{+0.18}_{-0.09}$ & 1.43 & $1.52^{+0.07}_{-0.06}$ & 1.53 & -0.06 \\[0.1cm]
\#4 & 2.0 & $1.66^{+0.06}_{-0.06}$ & 1.41 & $1.52^{+0.09}_{-0.09}$ & 1.53 & $1.57^{+0.08}_{-0.08}$ & 1.44 & $1.40^{+0.10}_{-0.07}$ & 1.76 & -0.05 \\[0.1cm]
\#5 & 3.0 & $1.66^{+0.05}_{-0.06}$ & 1.40 & $1.57^{+0.05}_{-0.04}$ & 1.43 & $1.43^{+0.05}_{-0.05}$ & 1.76 & $1.41^{+0.05}_{-0.05}$ & 1.75 & 0.14 \\[0.1cm]
\#16 & 1.0 & $1.64^{+0.04}_{-0.05}$ & 1.42 & $1.55^{+0.14}_{-0.07}$ & 1.46 & $1.40^{+0.04}_{-0.02}$ & 1.69 & $1.33^{+0.02}_{-0.02}$ & 2.01 & 0.15 \\[0.1cm]
\#join1 & 9.2 & $1.69^{+0.06}_{-0.05}$ & 1.31 & $1.63^{+0.05}_{-0.04}$ & 1.34 & $1.45^{+0.10}_{-0.08}$ & 1.68 & $1.35^{+0.06}_{-0.05}$ & 1.92 & 0.18 \\[0.1cm]
\#join2 & 6.8 & $1.67^{+0.05}_{-0.04}$ & 1.35 & $1.61^{+0.04}_{-0.03}$ & 1.38 & $1.32^{+0.09}_{-0.09}$ & 2.11 & $1.22^{+0.10}_{-0.05}$ & 2.35 & 0.29 \\[0.1cm]
\hline
Avg & - & $1.69^{+0.05}_{-0.05}$ & 1.33 & $1.54^{+0.08}_{-0.09}$ & 1.50 & $1.48^{+0.10}_{-0.07}$ & 1.63 & $1.38^{+0.07}_{-0.06}$ & 1.86 & -\\[0.1cm]
\hline
\textit{Herschel} \#5 & 3.0 & $1.65^{+0.07}_{-0.04}$ & 1.37 & $1.40^{+0.08}_{-0.05}$ & 1.87 & $1.43^{+0.05}_{-0.05}$ & 1.76 & $1.41^{+0.05}_{-0.05}$ & 1.75 & -0.03 \\[0.1cm]
Chameleon & - & - & - & - & - & - & - & $1.36^{+0.02}_{-0.02}$ & 1.91 & - \\[0.1cm]
Taurus & - & - & - & - & - & - & - & $1.33^{+0.03}_{-0.03}$ & 1.98 & -\\[0.1cm]
\end{tabular}
\\
\label{tab:fd_result}
\end{table*}
\begin{table*}
\caption{The results of fractal dimension measurements for cirrus clouds from our and previous studies.}
\begin{tabular}{c|c|c|c|c|c}
\hline
Study & Data & $\langle D\rangle$ & $D$ range & rough field size & contours size threshold \\
\hline
\cite{Bazell_Desert1988} & cirrus (IRAS) & 1.26 $\pm$ 0.04 & 1.12 - 1.40 & $\Delta \delta \sim 16^\circ$, $\Delta \alpha \sim 24^\circ$ & 187200, 8.5 \\
\cite{Dickman_etal1990} & molecular clouds (IRAS) &- & 1.174 - 1.278 & $\Delta \delta \sim 10^\circ$, $\Delta \alpha \sim 10^\circ$ & 187200, 8.5 \\
\cite{Falgarone_etal1991} & molecular clouds (CO lines) & $1.36 \pm 0.02$ & - & $\Delta \delta \sim 10^\circ$, $\Delta \alpha \sim 10^\circ$ & - \\
\cite{Vogelaar_Wakker1994} & cirrus (IRAS) & $1.42 \pm 0.02$ & 1.23-1.54 & $\sim 1 \square^\circ$ & 288000, 9.0 \\
\cite{Sanchez_etal2005} & Orion A (CO line) &1.35 & - & $\Delta \delta \sim 2^\circ$, $\Delta \alpha \sim 2^\circ$ & $\sim$777, 3.1 \\
\cite{Juvela_etal2018} & molecular clouds (\textit{Herschel}) & $1.25 \pm 0.07$ & 1.05 - 1.4 & $\sim 0.5 \square^\circ$ & 3960, 4.7\\
This study & SDSS Stripe82 & $1.69 \pm 0.07$ & 1.66-1.88 & $\sim (1-9)$ $\square^\circ$ & - \\
This study & IRIS & $1.38 \pm 0.09$ & 1.22-1.52 & $\sim (1-9)$ $\square^\circ$ & -\\
\hline
\end{tabular}
\\
{\raggedright \footnotesize{\textit{Notes}: The second column lists the type of the studied objects along with the data source. Note that for~\cite{Vogelaar_Wakker1994}, we include the results only for cirrus clouds, while in the referred work, the high-velocity molecular clouds were also studied. The third column lists the average values of the fractal dimension if they are provided by the authors or, otherwise, there is a dash symbol, while the fourth column lists the range of fractal dimensions of individual clouds studied in the refereed works. The dash symbol in the fourth column means that only one cloud was studied. The fifth column lists the rough values of the overall field sizes that were studied in the cited works. The sixth column lists the values of the contours size threshold that was used to filter out small clouds in physical units (arcsec$^2$) and in units of $\log A$ that are used in the present study, respectively.}\par}
\label{tab:previous}
\end{table*}
\begin{figure*}
\includegraphics[width=1.95\columnwidth]{imgs/allmc4.png}
\caption{Fractal dimension values obtained from our MC results for the different data. For each field the points show a median value and the vertical lines correspond to a 1$\sigma$ error. The horizontal segments show the result when the contours from all MC realisations are placed on the regression line (see text for details). Panel (a) shows all data with the 6~arcsec resolution, i.e. the optical data is shown in red, the optical data with colour filtering is depicted in cyan, and green is for \textit{Herschel} Field\#5 250~$\mu$m (see discussion in Sec.~\ref{sec:color}). Panel (b) contains the results for the IRIS counterparts (here, the magenta and grey colours depict the original data and the data with the optical mask, respectively). Panel (c) illustrates how $D$ changes for the optical (light blue) fields after rebinning to 90~arcsec and after applying the IRIS PSF (deep blue), the green point is for the single \textit{Herschel} field (see discussion in Sec.~\ref{sec:psf} for both (b) and (c) panels). Finally, panel (d) is the same as (b), except that it also depicts the effect of partial mask filling, which is shown using yellow-coded data points (see discussion in Sec.~\ref{sec:masking}). The red line in panel (c) corresponds to the optical data from (a), the blue line in panel (d) corresponds to the optical data with the IRIS PSF in (c).}
\label{fig:allmc3}
\end{figure*}
\begin{figure*}
\includegraphics[width=1.95\columnwidth]{imgs/level_dep.png}
\caption{Illustration of how the fractal dimension $D$ remains constant within the selected contour level change. The left subplot shows Field 2 as an example of different contours, where a brighter colour relates to a higher brightness. White pixels represent the mask used. The right panel shows how $D$ depends on the contour level, each line corresponds to an individual field, coloured dots show the results for Field\#2 and the same levels as on the left. A smaller subplot on the right side shows the logarithm of the number of clouds for the corresponding contour level.}
\label{fig:lev_dep}
\end{figure*}
\par
\section{Analysis of the factors that contribute to fractal dimension measurements}
\label{sec:effects_that_change_D}
One of the results obtained in the previous section is that the measured fractal dimension of the clouds, identified in the optical data, is substantially greater than the fractal dimension of their counterparts in the IR data. At the same time, the results of fractal dimension measurements are hard to interpret from a physical point of view if we do not know how the various factors, which are incorporated in the measurement procedure (the choice of the minimal brightness contour level to account for, the choice of the lower boundary of the region size, the masking), affect the results. Our MC simulation is useful to understand more clearly how the aforementioned factors, along with some others, can affect our fractal dimension measurements.
\par
\subsection{Analysis of the mask influence}
\label{sec:masking}
The main obvious difference between the clouds analysed here and the ones studied in the previous studies, listed in Table~\ref{tab:previous}, is the existence of masked areas which can overlap with the structure of an actual cloud.
Therefore, one can expect that masking should somewhat change the geometry of such a cloud. The question is \textit{to what degree} it can actually change the fractal dimension of the clouds in the fields under study.
\par
In our method, the existence of masked areas can affect a $D$ measurement in two different ways. First, the arbitrary decision, described in step (vii), to filter out all clouds, which touch the extended mask parts and are selected in step (i), can shift the resultant values of $D$ significantly.
\cite{Bazell_Desert1988} and \cite{Dickman_etal1990} also removed from their analysis those contours which touch (or intersect) the borders of the image. However, the extended mask parts of the fields under consideration can also lie completely inside. This is another difference (see an example of such a mask in panel (b) of Fig.~\ref{fig:pipe}).
Nevertheless, our MC analysis shows that the fractal dimension $D$ remains almost the same regardless of the decision on filtering and extended mask size selection, as illustrated in Fig.~\ref{fig:mcparamdep} for Field\#3 (parameters `mask size' and `no touch'). This is a surprising finding, since, in principle, rectangular and round blobs, which constitute the mask, can be found by chance completely inside the selected contour level, hence drastically decreasing the area $A$ while the perimeter $P$ becomes increasing.
\par
To study the effect when cloud regions touch the image borders or partially covered by a mask, we perform a simple experiment. First, we select more than 1500 `good' optical clouds that are not touched by extended masked regions or image borders from all fields and different contour levels. Then we randomly select some small number of clouds\footnote{This choice is motivated by the inner plot in Fig.~\ref{fig:lev_dep}, where the median of the natural logarithm $\log N \approx 4.5$, and, thus, $N \approx 90-100$ can be used as some characteristic number of clouds in a realisation, on average.} $n<100$ and measure the fractal dimension $D$ for them. After that, we dissect $m < n$ of these clouds by a straight line, whose slope and intercept was chosen randomly. The resulting pieces of the original clouds are then treated as new clouds. A set of such clouds, along with the remaining unsplit clouds, are characterised by a new fractal dimension $D_\mathrm{split}$. This process reliably simulates the situation when cirrus clouds are ``spoiled'' by the image boundaries or a mask. The values of $D-D_\mathrm{split}$ for 1000 realisations are shown in Fig.~\ref{fig:splitsim}. This figure can be used to determine how the effect of masking impacts the measurements. On average, the splitting into smaller clouds yields $D_\mathrm{split}$ which is larger by 0.05 than the original one. This can be explained by the fact that the number of small noisy clouds grows after the slicing. If we filter out those small clouds using the threshold `tailcut'=0.05, the resultant histogram by $D-D_\mathrm{split}$ becomes symmetric and centred at around zero with a standard deviation of just 0.03. These errors are too small to change the overall result of $D$ measurements. This answers the main question of this subsection.
\par
The second effect of masking in the optical relates to the ``holes'' it can produce in the interior of the clouds. As was already mentioned, if a data satisfies a condition $P \propto A^{D/2}$ and we apply an internal mask to it, then the resulting set of points $(\log A, \log P)$ will have a larger $D$. In our MC analysis, we can either interpolate such a mask in step (ii) before selecting a contour or fill the ``holes'' after the contour selection in step (v). Both these steps have almost no impact on $D$ estimation as can be seen from Fig.~\ref{fig:mcparamdep} (the parameters `interp' and `fill holes'). The reason behind this is that the ``holes'' are usually less than 5 pix wide and they can significantly change the position for only small clouds in the $\log A \div \log P$ plane, which are filtered out in most of the cases anyway. However, this does not hold true for the IRIS counterparts and rebinned optical fields, because due to lower resolution the sizes in pixels of all clouds in them are smaller than those in the original optical images. To verify this, we apply the mask, which we use for the optical data, to the IR images and carry out a MC analysis. We set a pixel masked only if more than half of the original (small) pixels within the rebinning window are masked. As an illustration, one can check the resulting mask for Field\#5 in panel (d) of Fig.~\ref{fig:f5_comp} and compare it with the original image in panel (a). It is clear from Fig.~\ref{fig:allmc3}, panel (b), that, on average, the fractal dimension $D$ is larger for the IRIS images with the mask applied (grey) than without it (magenta). We elaborate more on this result in panel (d) of the same figure, where we additionally show the results for the MC realisations with the filled ``holes'' only. As one can see, filling the ``holes'' lowers the $D$ value to the level which was measured before the masking. We note that these $D$ values are not entirely identical because the mask can also filter out some cirrus clouds, as discussed in the previous paragraph. Nevertheless, even if we can clearly see that the mask increases $D$ values for a low-resolution data, these values are shifted by less than $\approx0.1$ and still remain within the same margin of error. It is worth noting here that masking was rarely done in the previous studies. However, masking is also important for IR wavelengths because the cirrus emission can be contaminated by non-dust sources such as UltraLuminous InfraRed Galaxies (ULIRGs). This makes the obtained results valuable for future studies at different wavelengths.
\begin{figure}
\includegraphics[width=0.95\columnwidth]{imgs/split_simulation.png}
\caption{Histograms by the $D$ variations for our experiment when individual clouds are dissected by lines (see Sect.~\ref{sec:masking} for details). The right-shifted histogram with diagonal hatching corresponds to the cases when clouds with small sizes have been removed with the `tailcut'=0.05 applied. The histogram with dots hatching represents all cases.}
\label{fig:splitsim}
\end{figure}
\subsection{`Tailcut' parameter}
\label{sec:tailcut}
As mentioned in Sect.~\ref{sec:methods}, the lower boundary for the size of regions, we account for in linear regression fits, is parametrised by the `tailcut' parameter.
In~\cite{Bazell_Desert1988} and~\cite{Dickman_etal1990}, a typical size of the regions excluded from consideration was less than 52 arcmin$^2$~(13 square pix of the original IRIS resolution). In~\cite{Vogelaar_Wakker1994}, the authors found that depending on the choice of the lower limit for the region size, the results of a fractal dimension measurement will be different and that there are some ``stability regions'' in which the fractal dimension is independent of the selected value for region size. In subsequent studies, e.g.~\cite{Juvela_etal2018}, the excluded regions were much smaller in size (1.1 arcmin$^2$ or about 3 square pixels in their study). Our simulations allow us to find the exact way the fractal dimension depends on the selected value of the tailcut parameter. Note that this parameter is, by definition, dimensionless because it is a fraction from 0.0 to 1.0 where lower and upper bounds correspond to \textit{min} and \textit{max} of $\log A$, respectively. For any particular realisation these bounds are different and determined during the field processing. Thus, in each realisation one can indeed translate the `tailcut' fraction into some number of square pixels. Fig.~\ref{fig:mc} shows how the measured fractal dimension of the clouds changes with the value of the lower boundary size of the selected regions for the optical and IR data. For ease of presentation, we averaged the dependencies over all of our fields. As can be seen from the figure, for small values of the `tailcut' parameter, there is a plateau for the optical data over which the fractal dimension does not change much. After some threshold $\log A > 8$ for both the optical and IR data, the fractal dimension changes notably: it increases for the optical data and decreases in the IR. In general, we can conclude that the fractal dimension $D$ does depend on the size of the regions one decides to include in a fit. Due to the significantly better resolution of the optical data, there is a large ``stability region'' where one can safely measure the fractal dimension. The existence of such dependence is particularly important for IR data where one cannot find a similar stability region.
\par
Another small but important addition here is that a large value of the `tailcut' parameter can significantly affect the quality of $\log P \div \log A$ regression fitting. It is easy to imagine a situation when there are only a few points left and $D$ is not well constrained from such a limited data. Of course, this is not the only parameter responsible for such cases but the most important one. The quality of the MC regression lines is illustrated in Fig.~\ref{fig:linregerrors}. It is easy to see that, as the coefficient of determination $R^2$ suggests the total variation explained by the linear regression, it appeared to be lower for the optical data. The number of points is larger there, and, for exactly the same reason, the line fitting errors are greater for the IR fields. Nevertheless, these errors are small and the overall quality of the regression fits is good.
\begin{figure}
\includegraphics[width=0.95\columnwidth]{imgs/linreg_var_all.png}
\caption{Dependence of $D$ across all fields on threshold to filter out small clouds (i.e. `tailcut' parameter in each realisation). The colours represent individual sets of data, the solid line for each set depicts an average curve, the inner colour-filled spread corresponds to one standard deviation, whereas the outer spread shows the minimal and maximal values. The vertical depicts the 3~square pix size for the IRIS data resolution as a minimal used threshold for the filtered area.}
\label{fig:mc}
\end{figure}
\begin{figure}
\includegraphics[width=0.95\columnwidth]{imgs/linreg_errors.png}
\caption{The left panels show histograms for the coefficient of determination $R^2$ of the fitted MC regressions. The right panels show distributions of the regression slope fitting error obtained using an ordinary least squares method and multiplied by two, thus equal to the $D$ estimation uncertainty. The upper panels show the optical data, the second row is for the optical data convolved with the IRIS {PSF} and the bottom one is for the IRIS data. All histograms are normalized to unity size of the maximal bar for better visibility.}
\label{fig:linregerrors}
\end{figure}
\subsection{Image resolution and PSF}
\label{sec:psf}
The image resolution also affects the fractal dimension since the area and the perimeter are measured using discrete pixels. Therefore, the more pixels are contained in some selected cloud, the more accurate the estimate of its perimeter and area will be. Clearly, as the IR data has 15 times lower image resolution than the SDSS optical data we use, the image resolution can be a reason why the resultant fractal dimension is quite different for these two. To measure the effect of image resolution on fractal dimension measurements, we performed a simple test. Specifically, we rebin the optical data to the image resolution (pixel size) of the IR data. The intensity of the large resulting pixels is obtained by summing up the intensities of the original small pixels. The resultant values of the fractal dimension obtained from such reduced images are shown in Fig.~\ref{fig:allmc3}, panel (c) (light blue points). {As can be seen, the effect of the resolution appears to be almost negligible: the fractal dimension values only shift by $~0.05$, on average, to the lower end across the different fields.} This is comparable with the $D$ measurement error. In~\cite{Sanchez_etal2005}, an equally small shift in the values after decreasing the image resolution was observed in case of rather low volumetric fractal dimension $D_\mathrm{vol} \sim 1.2-2.0$ (see their figure 7). Perhaps our results indicate that we indeed deal with such clouds, but we cannot rule out the possibility that for real clouds the dependence of the projected $D$ on the volumetric $D_\mathrm{vol}$ and image resolution should be more complicated than that obtained by~\cite{Sanchez_etal2005} on the example of some simulated clouds.
\par
Apart from the image resolution, there is another important factor that distinguishes our IR and optical data, namely, the angular resolution, or the PSF. The effect of the PSF on fractal dimension measurements is hard to predict, but a larger PSF blurs out intensity gradients more effectively, and, therefore, leads to smoother contours of objects, in general. To estimate the consequences of the PSF differences for our optical and IR data, we applied the IRAS PSF with FWHM=4.3~arcmin to the optical data with the artificially lowered image resolution described in the previous paragraph (in other words, we convolved the optical data to the IR one). There is no need to consider the optical PSF here since the entire optical PSF is contained within less than one pixel of an IRIS image. The resultant fractal dimensions for all fields are shown in Fig.~\ref{fig:allmc3}, panel (c), and listed in Table~\ref{tab:fd_result}. It can be seen that the fractal dimension is actually quite dependent on the PSF size and it becomes significantly smaller (on average) for the degraded optical data as compared to the initial one. For Fields\#1-5, the fractal dimension of the clouds, observed in such degraded optical images, is now consistent with the results for the corresponding clouds in the IR~(see Table~\ref{tab:fd_result}, $D_{90}-D_{IRIS}$ column). After accounting for the differences in the PSF, a significant disagreement between the fractal dimensions for the degraded optical data and IR fields is essentially maintained for our largest fields, Field\#join1 and \#join2, and for Field\#16. For these fields, most likely, we trace different structures in the IR and optical due to the large-scale background subtraction in the SDSS (see Discussion). However, note that for these fields the differences between the fractal dimensions in the optical and IR become smaller too after convolution with IRIS PSF. To our knowledge, the influence of the PSF has not been previously considered in the context of fractal dimension measurements, but, apparently, it is one of the most important determining factors for measuring the fractal dimension of cirrus clouds. Therefore, this finding is of great importance for future studies aimed at exploration of the geometric properties of dust clouds in our Galaxy.
\par
It is important to note that both findings presented above are also valid for the \textit{Herschel} Field\#5 data, as can be seen for the green points in Fig.~\ref{fig:allmc3}. There is also no need to consider the \textit{Herschel} PSF for the above reason. Taking into account that the \textit{Herschel} $250\mu$m image corresponds to a slightly different part of the light spectrum as compared to {\it IRIS} 100~$\mu$m, but nonetheless, its analysis yields the same result. This makes our conclusion even more robust and emphasises the importance of the correct PSF conversion for a proper comparison between different data.
\subsection{Brightness contour level}
In~\cite{Bazell_Desert1988}, the authors calculated the fractal dimension of the clouds for different levels of brightness separately, while \cite{Dickman_etal1990} fitted a linear regression to the points obtained for a set of selected levels. They verified that their results remained the same if they considered only one specific level, although in that case the errors appeared larger. \citet{Vogelaar_Wakker1994} studied how the fractal dimension changes with the lowest contour level selected and found that the dimension of some clouds can depend on it.
\par
In the present work, we use our MC simulation to explore how the choice of the brightness contour level to distinguish cirrus clouds affects the resultant fractal dimension value. Fig.~\ref{fig:lev_dep} showcases how the fractal dimension value depends on the brightness level for all of the optical fields processed in the same way, i.e. they represent some subsample of MC realisations where all parameters remained the same. As can be seen, the choice of the brightness level indeed affects the resultant value, although to a small degree. On average, there is some linear trend with a higher brightness level corresponding to smaller values of the fractal dimension. The latter statement is also illustrated in Fig.~\ref{fig:mcparamdep}, the first panel in the second row. The reason why brighter contours resulted in a lower $D$ can be seen in Fig.~\ref{fig:pipe}, panel (e). With an increase of brightness, clouds become smaller, on average, and the left tail of the size distribution (shown by red points) starts to influence the result harder, while for such points the perimeter $P$ decreases faster than the area $A$. To be more confident in our results, we checked how the results, presented in Fig.~\ref{fig:allmc3}, would change if we included the structures, which are distinguished using all brightness level contours in one regression fit (see Sect.~\ref{sec:discussion} for more details). We plot the values obtained in this way in Fig.~\ref{fig:allmc3} --- depicted by horizontal segments of the same colour as the underlying data. It can be seen that accounting for all brightness levels in one fit does not change the values of the fractal dimension nor does it decrease the difference between the IR and optical data.
\subsection{Colour filtering}
\label{sec:color}
The results of~\cite{Roman_etal2020} suggest that cirrus clouds can be distinguished using optical colours~(see Sect.~\ref{sec:methods}, step (iv)). We apply this criterion for a part of the realisations in our MC simulation and check if filtering the pixels by their $(r-i)$ and $(g-r)$ optical colours has an impact on estimation of $D$. As can be seen from cyan points in Fig.~\ref{fig:allmc3}, the actual difference between the two sets of values is significant and approximately equals 0.2 for different fields. The colour filtering step shifts $D$ toward larger values, because it can produce a more perforated structure with additional holes, which leads to an overall increase of the $D$ value (see clarifications in Sect.~\ref{sec:masking}). Such behaviour is also illustrated in Fig.~\ref{fig:mcparamdep}, the third panel in the first row, where the two peaks corresponding to the realisations with and without colour filtering are clearly shifted by about 0.2 with respect to each other. For better understanding of how the procedure of colour filtering change the shape of the actual clouds, we refer the interested reader to figures 13-15 in \citet{Roman_etal2020}, where authors apply aforementioned criterion to images from Hyper Suprime-Cam Subaru Strategic Program data \citep{Aihara2018}. It is clear from those images that the colour filtering transforms the clouds into much less connected structures. Moreover, the resulted ``holes'' are very different from those produced by masking in size and number.
\par
Overall, the presented results show that the colour filtering has a strong impact on the fractal dimension measurements.
However, we should emphasise that there is room for some doubt as to whether this result really reflects the physical properties of the clouds under consideration. And, particularly, for the following reasons.
First, the calculation of $D$ through colour filtering has not been done before, thus we cannot compare our findings with previous works. Secondly, we do not know how good is the filtering relation in statistical terms. In fact, a given pixel could be outside the selected colour constraints due to low signal-to-noise ratio, but this does not mean that this pixel does not contain dust emission (see left subplot in fig.~12 in \citealt{Roman_etal2020}, where some grey error bars lie above the relation line). Probably, this affects clouds at the lowest surface brightness only, but this needs to be tested additionally. Thirdly, as mentioned earlier, the shape of the selected clouds after filtering differs significantly from those to which we do not apply the colour criterion (compare the clouds presented in Fig.~\ref{fig:lev_dep} and Fig.~\ref{fig:pipe} and those in figures 13-15 in \citealt{Roman_etal2020}). It is clear that the methods we use in our MC simulation will not help to solve connectivity and ``holes'' issues in such differently shaped clouds. Probably, an additional rebinning step can improve the situation by smoothing the statistical fluctuations, but this can lead to new biases and needs to be tested anyway. Taking into account all the reasons described above, we decided to not include the realisations with colour filtering into our final $D$ estimation and its error margins. Nevertheless, the result obtained here has a methodological importance, which, along with the colour relation itself, definitely deserves attention and an additional investigation in a separate study.
\par
The only case where the filtering has almost no effect at all is Field\#1, where the $D$ value is anomalously large and inconsistent with the others. Also, it is very stable, that is it shows almost no estimation error. This is even stranger since Field\#1 has the lowest $<g-r>$ colour among all, as shown in table~2 from \cite{Roman_etal2020}. Therefore, it has a more strict upper condition which should filter out, on average, more pixels and, thus, the filtering should affect the result even more. The reason for such a strange behaviour is that Field\#1 contains clouds of a significantly smaller size. This means that the point with the largest area $A$ of Field\#1 has an area that is more than an order of magnitude smaller as compared to the other fields. This addresses the issue under consideration since the colour filtering should affect large clouds the most.
\par
It is also interesting to note from Fig.~\ref{fig:allmc3} that $D$ values after the colour filtering are well consistent with each other between different fields, but whether this is an artificial result or not remains an open question. As already mentioned above, this is a potentially large and completely new area that requires further research.
\subsection{Different bands}
\label{sec:diffbands}
The values of the fractal dimension obtained for the optical data should be averaged over four available optical bands: $g,r,i,$ and $z$. However, for the individual bands, the values can be slightly different, so we should verify how large this difference is. The value of the fractal dimension $D$, averaged over all fields and realisations, is equal to $1.73, 1.74, 1.77$, and $1.89$, accordingly, in the same order as above and with a typical error of around 0.08. As can be seen, for all bands save the $z$ band, the values are somewhat consistent with each other, while for the $z$ band the average value is larger by about 0.15. The true reason for such a behaviour is hard to pinpoint accurately, but we should note that SDSS imaging in the $z$ band is substantially shallower than in the other bands, that is, the $z$ band surface brightness limit is $\mu_\mathrm{limit}\sim26.6$~mag arcsec$^{-2}$, as compared to $\mu_\mathrm{limit}\sim28.5$~mag arcsec$^{-2}$ \citep{Roman_etal2020}, on average, for the other bands. Consequently, there should be a smaller number of dim clouds in the $z$ band, which can be distinguished in principle. Perhaps, this circumstance leads to the observed increase of the fractal dimension in the $z$ band. We can also see by visual inspection that the data in this band have more artefacts and issues, especially where the original SDSS frames were composed. Since results in the $z$ band are inconsistent with those obtained for the other SDSS bands, we decided to exclude this band from our final results averaged over all bands under study (i.e. reported in Table~\ref{tab:fd_result} and shown in all figures).
\par
{\,}
\par
Let us briefly sum up the results of this section. Most of the results are illustrated in Fig.~\ref{fig:allmc3}. We have found that masking affects the results in two different ways. It can artificially increase $D$ values by ``making holes'' in the clouds. Secondly, extended mask parts and image boundaries can obscure or even split up individual clouds. Both these effects are found to be small ($\Delta D < 0.1$). Unlike masking, filtering of small clouds with the `tailcut' parameter have a great impact on $D$ estimation. We show that the optical data demonstrates the same fractal dimension within some `plateau', while for the IR counterparts, $D$ can only decrease. A change of image resolution only slightly reduces $D$, but convolution of the optical data with the IRAS PSF lowers the fractal dimension significantly. The obtained results are almost the same for different brightness contour levels, as expected for fractals. The calculation of $D$ through colour filtering was done for the first time in this work, and we found that it can significantly increase $D$ by a value around 0.2. This effect can be artificial due to several reasons listed above, and it needs further investigation. Finally, the $z$ band was found to be inconsistent with the other SDSS bands and was excluded from our final results. {Note that the cumulative effect of all aforementioned factors on $D$ should not be considered as a sum of individual factors because the correct uncertainties on $D$ have been specially estimated in our MC simulation.}
\section{Discussion}
\label{sec:discussion}
\begin{figure}
\includegraphics[width=0.95\columnwidth]{imgs/mean_fluxes.png}
\caption{Top panel shows average pixel fluxes in the $r$-band fields and their IRIS 100~$\mu$m counterparts. All images are rebinned to the same image and angular resolution, and processed with an identical optical mask. The red dashed lines, which start in the origin of the coordinates, show linear dependencies with different slopes. The bottom panel shows the pixel-by-pixel Pearson correlation coefficient for the same data.}
\label{fig:meanflux}
\end{figure}
\subsection{Optical versus IR fractal dimension}
The analysis, which we carried out in the previous section, convincingly shows that if we account for various differences between optical and IR data and the differences in the corresponding data processing, there are some fields (Fields\#5-6, \#join1, \#join2) where the fractal dimension values in the optical are considerably greater than those in the IR. For the other fields (Fields\#1-4), the measured fractal dimensions are quite close to each other. Below we discuss the possible reasons leading to the observed consistencies and inconsistencies.
\par
First of all, we assume that the light comes from the same dust grains in the optical and IR. However, different parts of their spectrum we observe are mainly associated with different generating mechanisms. In the case of the IR, it is a thermal emission while it is a scattering of incident light in the optical~\citep{Draine_2003}. The cirrus clouds are usually optically thin~\citep{Stark1995,Szomoru_Guhathakurta1999} and, hence, the flux should be proportional to the column density in both cases. However, there are various factors that can reduce the correlation between the flux and column density. For the thermal emission, there are variations in the dust temperature~\citep{Lehtinen_etal2007,Ienaka_etal2013}. For optical fluxes, it is important from where the photon comes and how it is scattered, i.e. we should know the phase function.
\cite{Seon_Witt2013} verified how the scattered flux correlates with the optical depth for different values of the phase factor $g$ and homogeneous optical depth values on the example of some simulated clouds where the source of radiation is located in the centre of the cloud. They found that there is a good correlation between the optical depth and the scattered fluxes maps in the case of complete forward scattering, i.e. $g=0.99$~(see their figure 3). However~\cite{Seon_Witt2013} also found that decreasing the $g$ factor (i.e. the case of more isotropic scattering) decreases the correlation between the fluxes and optical depth. For cirrus clouds, which are illuminated by the interstellar radiation field (and not a point source as considered by~\citealt{Seon_Witt2013}), we can assume that the scattering picture we observe should be closer to the case of isotropic scattering described by~\cite{Seon_Witt2013}. Nevertheless, from their figure~1 we can see that the correlation is still strong even in the case of $g\lesssim 0.5$. For the typical value of $g \sim 0.6-0.7$~\citep{Ienaka_etal2013}, the results of~\cite{Seon_Witt2013} suggest that if we measure some geometric property (fractal dimension) of the clouds in the optical, it can have a different value than if measured based on IR data (which is a better tracer of the dust density).
\par
These and several other reasons can lead to a situation that we indeed measure the fractal characteristics of objects, the images of which are different in some particular details at IR and optical wavelengths. This is especially true for some of the fields under consideration. This fact can be noticeable, for example, from the one-to-one visual comparison of lower panels in Fig.~\ref{fig:f5_comp} and is illustrated in Fig.~\ref{fig:meanflux}. The mean pixel fluxes in the optical and IR are compared in the upper panel. It can be seen that an IR emission increase does not always coincide with a significant optical intensity change. The same conclusion can be drawn based on the lower panel, where the pixel-by-pixel correlation coefficient is measured for the IR and optical fluxes. In only three cases, we found that the data behave in the same way, while in all other cases the correlation is weak. It is clear for our data that in the fields under consideration a direct relation between the images of the cirrus clouds, if exists, is at least nonlinear. In general, this is not a new result and has been known for a long time \citep{Guhathakurta1989}.
\par
Besides the underlying physical factors mentioned above, in our case, there are several additional reasons for the observed discrepancy between the IR and optical structures. First, it is partially due to the presence of bright unmasked sources such as, for example, ULIRG F22509-0040 in Field\#4. However, such sources are few. The second reason is large-scale subtraction of the background in the optical fields as an inevitable drawback of the SDSS data reduction pipeline. The background subtraction procedure implies a polynomial or spline approximation of the background and removes every diffuse emission with a scale larger than several arcmins (see section 4 in \citealt{Roman_etal2020}). As the result of this process, extended details of cirrus clouds (or large-scale diffuse emission) can potentially be removed from our optical images and, thus, not included in the $D$ evaluation. However, as we have already demonstrated with high confidence, the constancy of the fractal dimension value between different contours, as well as different parts of Stripe82, is great and, thus, it is reasonable to presume that, at least on small scales, it should be reliably measured and nonaffected by possible over-subtraction. This point can be additionally illustrated by the fact that one of the largest fields Field\#join2 demonstrates a strong correlation between optical and IR cirrus parts, as Fig.~\ref{fig:meanflux} suggests. It can be concluded that the issue about the correspondence of optical and IR data is a difficult one and needs additional research, which we are about to do in a separate study.
\par
We should note that, for dense molecular clouds with grains of sizes from $0.1$~$\mu$m to $10$~$\mu$m, there is also an interesting mechanism that can introduce the differences in the fractal dimension for the optical and IR data. It is known that the parameters of grain size distribution determine the extinction curve behaviour, as well as the scattering properties of a particular cloud~\citep{Weingartner_Draine2001} and, thus, govern the overall intensity level at different wavelengths. At the same time, based on the results of hydrodynamic simulations, various authors pointed out that for real observed clouds the larger grains can decouple from gas flows and tend to occupy denser areas, while small grains are typically well-coupled with the gas flows \citep{Hopkins_Lee2016,Tricco_etal2017,Mattsson_etal2019}. In terms of the fractal dimension, the smaller value of it means that, in general, the cloud has a more filamentary wispy structure. Hence, we found that some clouds in the IR demonstrate a more filamentary structure as compared to the optical. Qualitatively, this result agrees well with the results of~\cite{Hopkins_Lee2016}, \cite{Tricco_etal2017}, and \cite{Mattsson_etal2019}. We expect that larger dust grains, which are usually traced in the visible, inhabit denser parts of the clouds while small grains are better traced in low density filaments, which become more apparent in the IR. Although, we should admit that for a diffuse medium such as in a cirrus cloud, we expect that there is no such a wide spread by the grain size as considered by the mentioned authors. Further physical simulations of dust clouds are required to support this hypothesis.
\subsection{Comparison with previous studies}
\label{sec:comparison}
From Table~\ref{tab:previous}, it is interesting to note that in some earlier works~\cite{Bazell_Desert1988,Dickman_etal1990}, the dimension measured using IR data seem to be slightly smaller, on average, as compared to that obtained by~\cite{Vogelaar_Wakker1994} and the average value $\langle D\rangle=1.38^{+0.07}_{-0.06}$ in the present study for the IR data. Here we should also emphasise that our IR data have exactly the same image and angular resolution (PSF) as in~\cite{Bazell_Desert1988, Vogelaar_Wakker1994} and, therefore, some other factors should contribute to the observed inconsistency.
Based on the analysis presented in Sect.~\ref{sec:effects_that_change_D}, one can think of several possible reasons for that. First, we found that the fractal dimension value {depends on the initial filtering by the cloud size}. For the IR data, the exact value of this border seem to be one of the determining parameters. In Table~\ref{tab:previous}, we collected the values of exact sizes of the filtered regions used in the literature, where possible. As can be seen, in earlier works the authors filtered the clouds of rather large sizes. As we showed in Fig.~\ref{fig:mc}, such filtering indirectly leads to lower values of the fractal dimension $D$. The second reason is connected to the first one. As have been already noted, {smaller} clouds tend to have a smaller ratio of the perimeter and area logarithms than the larger ones. At this point, it is important to know how large the field size under study is and, according to that, how large, on average, the clouds under study are. Table~\ref{tab:previous} lists the rough values of the field sizes which were used for $D$ estimation in the previous studies. As can be seen, in earlier works, the authors used significantly larger fields (except for \citealt{Vogelaar_Wakker1994}, the results of which are most consistent with ours) than we analyse in this study. From Fig.~\ref{fig:mc}, one can expect that an increase {of the number of} of large clouds should decrease the fractal dimension.
One potentially major source of difference between this study and the previous ones is the way to choose which data to fit regression on. In our approach only one contour per realisation was used for this purpose, while in other studies starting from \cite{Dickman_etal1990} a set of shrinking contours were used with some increasing brightness interval between them. Since the interval choice is arbitrary, one needs to carry out a proper investigation of its effect on $D$. On the other side, the use of many contours at once should produce a more stable result because the same cloud is included several times and the points are placed more evenly in the $A$ range. In MC, the number of points is smaller and there is a chance that, for example, we pick just the brightest cores and do not fit the regression well due to a short data range. In order to examine if the derived results will be the same in case we use several contours instead, we carried out the following experiment. For a given Field all MC realisations were placed on the same $\log A \div \log P$ plane and then the $D$ value was measured, thus emulating a use of a huge set of different evenly distributed contours. We also use the conservative `tailcut' value 0.05 for filtering out the noisiest small clouds from the fitting. The results are presented in Fig.~\ref{fig:allmc3} by horizontal segments, where we compare the $D$ values from MC and from the described experiment. We note that the errors of the regression fitting with this number of points {are of the order of 0.01 and, thus, not shown.} For both the optical and IR data, in all cases with just a few exceptions we obtain that $D$ values lie within the uncertainty range of MC simulation and, often, they are close to its centre. Several exceptions are related to images with small resolution and can be addressed to situation when the numbers of clouds and successful realisations are small. Overall, the agreement is good and it is unlikely to be a source of disagreement in fractal dimension estimation. Note also that in the described experiment there is no tendency to produce a consistently greater or lower $D$ than MC shows.
\par
\section{Summary}
\label{sec:sum}
In the present study, we have carried out a comparative analysis of the \textcolor{black}{2D} fractal dimension $D$ for cirrus clouds in our Galaxy using IR and optical data. To the best of our knowledge, for the optical data, such a study has been conducted for the first time. We have considered 8 fields from SDSS Stripe82 which were used by~\cite{Roman_etal2020} to study the colour characteristics of the cirrus clouds in the optical. The corresponding IR counterparts have been taken from the IRIS 100~$\mu$m database. For one optical field, we have also considered the corresponding \textit{Herschel} 250$\mu$m counterpart. The deep optical fields and the mentioned \textit{Herschel} field were originally prepared by~\cite{Roman_etal2020} with accounting for the instrumental scattered light and masking of all the external sources, exposing only the diffuse emission.
\par
We have used a simple approach to compute the fractal dimension of the cirrus clouds based on the perimeter-area ratio of the cirrus contours. One of the aims was to compare our results for the IR with the results from the literature where the same approach was used. For each of the optical and IR fields, we distinguished the structures which are enclosed by some brightness contour level, calculated the corresponding perimeter $P$ and area $A$, taking into account the {cloud boundaries and possible masked inner pixels} for each of such structures, and then approximated the $\log P \div \log A$ dependence with a straight line. The angular coefficient of the line gives us the value of the fractal dimension for the {chosen} brightness level of the cirrus clouds observed in each field.
\par
In previous studies, the error on the obtained fractal dimension was assumed to be equal to the error of a linear regression fitting. In the present paper, we performed a MC simulation to estimate how the choice of some subjective parameters (e.g., the lower size limit of clouds to be filtered out, the brightness level to choose) affects the results and, based on that, estimated an error value on $D$.
\par
For IR data, we found that the average fractal dimension across all fields is $\langle D \rangle = 1.38^{+0.07}_{-0.06}$ with $\sigma(D)=0.09$. The obtained values are generally consistent with the results of previous studies~\citep{Bazell_Desert1988,Dickman_etal1990,Falgarone_etal1991,Vogelaar_Wakker1994,Hetem_Lepine1993,Sanchez_etal2005,Juvela_etal2018}. For our data, we found that there is a strong dependence of the fractal dimension value on the subjective parameters and, especially, on the lower size limit of the clouds to be filtered out. Accounting for this effect in our MC simulation yielded the significantly greater error $\sigma(D)$ as compared to those obtained in the literature.
\par
In the optical, we found that the average fractal dimension of the clouds is $D=1.69$ with the very small $\sigma(D)=0.02$ if one outlier is excluded. Consequently, the $D$ values in the optical appeared to to be considerably greater than those obtained for the IR data. We explored whether this difference arises due to the differences in the observational characteristics of the field under study, namely, the image resolution and the PSF properties. We came to conclusion that the PSF has a substantial effect on the fractal dimension while the image resolution itself does not significantly affect the result. Our finding that the PSF can significantly influence the measured fractal dimension is important from a methodological point of view: one should compare the results of fractal dimension measurements for different data (and simulations) only if they have been convolved to the same angular resolution.
\par
Concerning the effect of masking, based on our MC simulation and tests with the IRIS data, we found that the masking itself leads to an overall increase of $D$ by about 0.1, but it can be greater than that in some cases depending on data studied. As an example of such an exceptional case, the cirrus clouds in Field\#2 demonstrate an increase of the fractal dimension by about 0.15 after applying the optical mask.
\par
As to the threshold on the contour size to filter out small clouds, which is discussed, for example, in~\cite{Vogelaar_Wakker1994}, we found that for our IR data there is a strong dependence of the fractal dimension {on this parameter}, while for the optical data there is a plateau where the fractal dimension can be reliably measured.
\par
In total, we conclude that if we take into account the differences in the PSF, image resolution, and masking, half of the fields under study demonstrate equal fractal dimensions of the selected clouds in the optical and IR. For the other half, the fractal dimensions in the optical and IR remain inconsistent. We discussed various reasons for the observed phenomenon including the differences in flux generating mechanism at different wavelengths, presence of some bright unmasked sources in IR data, and the tendency of larger grains to decouple from smaller ones, proven in some model studies. We conclude that none of these reasons can be considered as dominant and additional studies are required to explain why the fractal dimensions of IR and optical cirrus can be essentially the same for some fields, while for others they can differ significantly.
\section*{Acknowledgements}
We acknowledge financial support from the Russian Science Foundation (grant no. 20-72-10052).
\par
We thank the anonymous referee for his/her
review and appreciate the comments, which contributed to improving the
quality of the article.
\par
Funding for the Sloan Digital Sky
Survey IV has been provided by the
Alfred P. Sloan Foundation, the U.S.
Department of Energy Office of
Science, and the Participating
Institutions.
\par
SDSS-IV acknowledges support and
resources from the Center for High
Performance Computing at the
University of Utah. The SDSS
website is www.sdss.org.
\par
SDSS-IV is managed by the
Astrophysical Research Consortium
for the Participating Institutions
of the SDSS Collaboration including
the Brazilian Participation Group,
the Carnegie Institution for Science,
Carnegie Mellon University, Center for
Astrophysics | Harvard \&
Smithsonian, the Chilean Participation
Group, the French Participation Group,
Instituto de Astrof\'isica de
Canarias, The Johns Hopkins
University, Kavli Institute for the
Physics and Mathematics of the
Universe (IPMU) / University of
Tokyo, the Korean Participation Group,
Lawrence Berkeley National Laboratory,
Leibniz Institut f\"ur Astrophysik
Potsdam (AIP), Max-Planck-Institut
f\"ur Astronomie (MPIA Heidelberg),
Max-Planck-Institut f\"ur
Astrophysik (MPA Garching),
Max-Planck-Institut f\"ur
Extraterrestrische Physik (MPE),
National Astronomical Observatories of
China, New Mexico State University,
New York University, University of
Notre Dame, Observat\'ario
Nacional / MCTI, The Ohio State
University, Pennsylvania State
University, Shanghai
Astronomical Observatory, United
Kingdom Participation Group,
Universidad Nacional Aut\'onoma
de M\'exico, University of Arizona,
University of Colorado Boulder,
University of Oxford, University of
Portsmouth, University of Utah,
University of Virginia, University
of Washington, University of
Wisconsin, Vanderbilt University,
and Yale University.
\par
This paper has used archival data
from the \textit{Herschel} mission. \textit{Herschel} is an ESA space observatory
with science instruments provided by European-led Principal Investigator consortia and with important participation from NASA.
\section*{Data availability}
The data underlying this article will be shared on reasonable request to the corresponding author.
\bibliographystyle{mnras}
|
\section{Introduction}
Evolutionary algorithms (EAs) are a big family of randomized search heuristics inspired from biological evolution.
Many evolutionary algorithms (EAs) such as genetic algorithm (GA) and evolutionary strategies (ES) use a crossover operator besides mutation and selection. Many empirical studies demonstrate that crossover which combines genes of two parents to generate new offspring is helpful to convergence of EAs. Theoretical results on runtime analysis validate the promising function of crossover in EAs~\citep{jansen2002analysis,kotzing2011crossover,corus2017standard,dang2017escaping,sudholt2017crossover,pinto2018simple, oliveto2020tight,lengler2020large,lehre2008crossover,doerr2012crossover,doerr2013more,sutton2021fixed}, whereas there are also some cases that crossover cannot be helpful~\citep{richter2008ignoble,antipov2021effect}.
Differential evolution (DE) algorithms implement crossover operations in a different way. By exchanging components of target vectors with donor vectors, continuous DE algorithms achieve competitive performance on a large variety of complicated problems~\citep{Das2011,DAS20161,SEPESYMAUCEC2019100428,BILAL2020103479}, and its competitiveness is to great extent attributed to the employed crossover operations~\citep{Lin2011}.
Besides the theoretical studies on continuous DE \citep{OPARA2019546} as well as the runtime analysis to reveal working principle of BDE~\citep{DOERR2020110}, there were no theoretical results reported on how crossover influences performance of discrete-coded DE algorithms.
Nevertheless, there is a gap between runtime analysis and practice. Since EAs belong to randomized search, their optimization time to reach an optimum is uncertain and could be even infinite in continuous optimization~\citep{CHEN2021200}. Due to this reason, optimization time is seldom used for evaluating the performance of EAs in computer simulation. While EAs stop after running finite generations, their performance is evaluated by solution quality such as the mean and median of the fitness value or approximation error~\citep{xu2020helper}. In theory, solution quality can be measured for given iteration budget by the expected fitness value~\citep{JANSEN201439} or approximation error~\citep{he2016analytic,WANG2021}, which contributes to the analysis framework named as fixed budget analysis. A fixed budget analysis on immune-inspired hypermutations leads to theoretical results that are very different from those of runtime analysis but consistent to the empirical results, which further demonstrates that the perspective of fixed budget computations
provides valuable information and additional insights for performance of randomized search heuristics~\citep{Jasen2014TEVC}.
In this paper, solution quality of an EA after running finite generations is measured by two metrics: the expected value of the approximation error and the error tail probability. The former measures solution quality, that is the fitness gap between a solution and optimum. The latter is the probability distribution of the error over error levels, which measures the probability of finding the optimum.
An EA is said to outperform another if for the former EA, its error and tail probability are smaller. Furthermore, an EA is said to asymptotically outperform another if for the former EA, its error and tail probability are smaller after a sufficiently large number of generations.
The research question of this paper is whether the binomial crossover operator can help reduce the approximation error. As a pioneering work on this topic, we investigate a $(1+1)EA_C$ that performs the binomial crossover on an individual and an offspring generated by mutation, and compare a $(1+1) EA$ without crossover and its variant $(1+1)EA_C$ on two classical problems, OneMax and Deceptive. By splitting the objective space into error levels, the analysis is performed based on the Markov chain models~\citep{he2003towards,He2016}. Given the two EAs, the comparison of their performance are drawn from the comparison of their transition probabilities, which are estimated by investigating the bits preferred to by evolutionary operations. Under some conditions, $(1+1)EA_C$ with binomial crossover outperforms $(1+1) EA$ on OneMax, but not on Deceptive; however, by adding an adaptive parameter mechanism arising from theoretical results, $(1+1)EA_C$ with binomial crossover outperforms $(1+1) EA$ on Deceptive too.
The rest of this paper is organized as follows. Section \ref{SecRelWork} reviews related theoretical work. Preliminary contents for our theoretical analysis are presented in Section \ref{SecPre}. Then, the influence of the binomial crossover on transition probabilities is investigated in Section \ref{SecTrans}. Section \ref{SecAsy} conducts an analysis of the asymptotic performance of EAs. To reveal how binomial crossover works on the performance of EAs for consecutive iterations, the OneMax problem and the Deceptive problem are investigated in Sections \ref{SecExpi} and \ref{SecExpr}, respectively. Finally, Section \ref{SecCon} presents the conclusions and discussions.
\section{Related Work}\label{SecRelWork}
\subsection{Theoretical Analysis of Crossover}
To understand how crossover influences the performance of EAs, Jansen \emph{et al.} \cite{jansen2002analysis} proved that an EA using crossover can reduce the expected optimization time from super-polynomial to a polynomial of small degree on the function Jump. K\"{o}tzing \emph{et al.}
\cite{kotzing2011crossover} investigated crossover-based EAs on the functions OneMax and Jump and showed the potential speedup by crossover when combined with a fitness-invariant bit shuffling operator in terms of optimization time. For a simple GA without shuffling, they found that the crossover probability has a drastic impact on the performance on Jump. Corus and Oliveto
\cite{corus2017standard} rigorously obtained an upper bound
on the runtime of standard steady state GAs to hillclimb the OneMax function and proved that the steady-state
EAs are 25\% faster than their mutation-only counterparts. Their analysis also suggests
that larger populations may be faster than populations of size 2. Dang \emph{et al.} \cite{dang2017escaping} revealed that the interplay between crossover and mutation may result in a sudden burst of diversity on the Jump test function and reduce the expected optimization time compared to mutation-only algorithms like the (1+1) EA.
For royal road functions and OneMax, Sudholt \cite{sudholt2017crossover} analyzed uniform crossover and k-point crossover and proved that crossover makes every $(\mu+\lambda)$ EA at least twice as fast as the fastest EA using only standard bit mutation. Pinto and Doerr
\cite{pinto2018simple} provided a simple proof of a crossover-based genetic algorithm (GA) outperforming any mutation-based black-box heuristic on the classic benchmark OneMax. Oliveto \emph{et al.} \cite{oliveto2020tight} obtained a tight lower bound on the expected runtime of the (2 + 1) GA on OneMax. Lengler and Meier
\cite{lengler2020large} studied the positive effect of using larger population sizes and crossover on Dynamic BinVal.
Besides artificial benchmark functions, several theoretical studies also show that crossover is helpful on non-artificial problems.
Lehre and Yao \cite{lehre2008crossover} proved that the use of crossover in the $(\mu+1)$ Steady State Genetic Algorithm may reduce the runtime from exponential to polynomial for some instance classes of the problem of computing unique input–output (UIO) sequences. Doerr \emph{et al.}
\cite{doerr2012crossover,doerr2013more} analyzed EAs on the all-pairs shortest path problem. Their results confirmed that the EA with a crossover operator is significantly faster than without in terms of the expected optimization time. Sutton
\cite{sutton2021fixed} investigated the effect of crossover on the closest string problem and proved that a multi-start $(\mu+1)$ GA required less randomized fixed-parameter tractable (FPT) time than that with disabled crossover.
However, there is some evidence that crossover is not always helpful. Richter \emph{et al.} \cite{richter2008ignoble} constructed Ignoble Trail functions and proved that mutation-based EAs optimize them more efficiently than GAs with crossover. The later need exponential optimization time. Antipov and Naumov
\cite{antipov2021effect}
compared crossover-based algorithms on RealJump functions with a slightly shifted optimum. The runtime of all considered algorithms increases on RealJump. The hybrid GA fails to find the shifted optimum with high probability.
\subsection{Theoretical Analysis of Differential Evolution Algorithms}
Although numerical investigations of DEs have been widely conducted, only a few theoretical studies paid attention to theoretical analysis, most of which are focused on continuous DEs~\citep{OPARA2019546}. By estimating the probability density function of generated individuals, Zhou \emph{et al.} \cite{Zhou2016Analysis} demonstrated that the selection mechanism of DE, which chooses mutually different parents for generation of donor vectors, sometimes does not work positively on performance of DE. Zaharie and Micota \cite{ZAHARIE20091126,zaharie2008statistical,Zaharie2017} investigated influence of the crossover rate on both the distribution of the number of mutated components
and the probability for a component to be taken from the mutant vector, as well as the influence of mutation and crossover on the diversity of intermediate population. Wang and Huang \cite{WANG20103263} attributed the DE to a one-dimensional stochastic model, and investigated how the probability distribution of population is connected to the mutation, selection and crossover operations of DE. Opara and Arabas \cite{OPARA201853} compared several variants of differential mutation using characteristics of their expected mutants’ distribution, which demonstrated that the classic mutation operators yield similar
search directions and differ primarily by the mutation range. Furthermore, they formalized the contour fitting notion and derived an analytical model that links the differential mutation operator with the adaptation of
the range and direction of search~\citep{OPARA2019100441}.
By investigating expected runtime of the binary differential evolution (BDE) proposed by Gong and Tuson \cite{Gong2007}, Doerr and Zhang \cite{DOERR2020110} performed a first fundamental analysis on the working principles of discrete-coded DE. It was shown that BDE optimizes the important decision variables, but is hard to find the optima for decision variables with small influence on the objective function. Since BDE generates trial vectors by implementing a binary variant of binomial crossover accompanied by the mutation operation, it has characteristics significantly different from classic EAs or estimation-of-distribution algorithms.
\subsection{Fixed Budget Analysis and Approximation Error}
To bridge the wide gap between theory and application, Jasen and Zarges \cite{JANSEN201439} proposed a fixed budget analysis (FBA) framework of randomized search heuristics (RSH), by which the fitness of random local search and (1+1)EA were investigated for given iteration budgets. Under the framework of FBA, they analyzed the any time performance of EAs and artificial immune systems on a proposed dynamic benchmark problem~\citep{Jansen2014}. Nallaperuma \emph{et al.} \cite{Nallaperuma2017} considered the well-known traveling salesperson problem (TSP) and derived the lower bounds of the expected fitness gain for a specified number of generations. Doerr \emph{et al.} \cite{Doerr2013} built a bridge between runtime analysis and FBA, by which a huge body of work and a large collection of tools for the analysis of the expected optimization time could meet the new challenges introduced by the new fixed budget perspective. Based on the Markov chain model of RSH, Wang \emph{et al.} \cite{WANG2021} constructed a general framework of FBA, by which they got the analytic expression of approximation error instead of asymptotic results of expected fitness values.
Considering that runtime analysis demonstrated that hypermutations tend to be inferior
on typical example functions, Jansen and Zarges \cite{Jasen2014TEVC} conducted an FBA to explain why artificial immune systems are popular in spite
of these proven drawbacks. Although the single point mutation in random local search (RLS) outperforms than the inversely fitness-proportional mutation (IFPM) and the somatic contiguous hypermutation (CHM) on OneMax in terms of expected optimization time, it was shown that IFPM and CHM could be better while FBA is performed by considering different starting points and varied iteration budgets. The results show that the traditional perspective of expected optimization time may be unable to
explain observed good performance that is due to limiting the length of runs. Therefore, the perspective of fixed budget computations
provides valuable information and additional insights.
\section{Preliminaries}\label{SecPre}
\subsection{Problems}
Consider a maximization problem
\begin{equation*
\max f(\mathbf{x}),\quad\mathbf{x}=(x_1,\dots,x_n)\in\{0,1\}^n,
\end{equation*}
Denote its optimal solution by $\mathbf{x}^*$ and optimal objective value by $f^*$.
The quality of a solution $\mathbf x$ is evaluated by its approximation error $e(\mathbf x):=\mid f(\mathbf x)-f^*\mid$. The error $e(\mathbf x)$ takes finite values, called error levels:
\begin{equation*
e(\mathbf x)\in\{e_0,e_1,\dots,e_L\},\quad 0= e_0\le e_1\le\dots\le e_L,
\end{equation*}
where $L$ is an non-negative integer. $\mathbf{x}$ is called \emph{at the level $i$} if $e(\mathbf{x})=e_i$, $i\in\{0,1,\dots,L\}$. The collection of solutions at level $i$ is denoted by $\mathcal{X}_i$.
Two instances, the uni-modal OneMax problem and the multi-modal Deceptive problem, are considered in this paper.
\begin{problem}\label{OneM}(\textbf{OneMax})
\begin{equation*}
\max f(\mathbf x)=\sum_{i=1}^nx_i,
\end{equation*}
where $\mathbf x=(x_1,\dots,x_n)\in \{0,1\}^n$.
\end{problem}
\begin{problem}\label{Dec}(\textbf{Deceptive})
\begin{equation*}
\max f(\mathbf x)=\left\{\begin{aligned}& \sum_{i=1}^nx_i, && \mbox{if }\sum_{i=1}^nx_i>n-1,\\ & n-1-\sum_{i=1}^nx_i, && \mbox{otherwise.} \end{aligned}\right.
\end{equation*}
where $\mathbf x=(x_1,\dots,x_n)\in \{0,1\}^n$.
\end{problem}
Both OneMax and Deceptive can be represented in the form
\begin{equation}\label{OP1}
\max f(\mid\mathbf{x}\mid),
\end{equation}
where $\mid\mathbf{x}\mid:=\sum_{i=1}^{n}x_i$. Error levels of (\ref{OP1}) take only $n+1$ values.
For the OneMax problem, both exploration and exploitation are helpful to convergence of EAs to the optimum, because exploration accelerates the convergence process and exploitation refines the precision of approximation solutions. However, for the Deceptive problem, local exploitation leads to convergence to the local optimum, but it in turn increases the difficulty to jump to the global optimum. That is, exploitation hinders convergence to the global optimum of the Deceptive problem, thus, the performance of EAs are dominantly influenced by their exploration ability.
\subsection{Evolutionary Algorithms}
\begin{algorithm}[ht]
\caption{$(1+1)EA$}\label{Alg1}
\begin{algorithmic}[1]
\STATE counter $t=0$;
\STATE randomly generate a solution $\mathbf x_0=(x_1,\dots,x_n)$;
\WHILE{the stopping criterion is not satisfied}
\FOR{$i=1,2,\dots,n$}
\STATE \begin{align}\label{Mut1}y_i=\left\{\begin{aligned}& 1-x_i, && \mbox{if } rnd_i<p_m,\\ & x_i, && \mbox{otherwise},\end{aligned}\right. \quad rnd_i\sim U[0,1];
\end{align
\ENDFOR
\IF{$f(\mathbf{y})\ge f(\mathbf{x}_t)$}
\STATE $\mathbf{x}_{t+1}=\mathbf{y}$;
\ENDIF
\STATE $t=t+1$;
\ENDWHILE
\end{algorithmic}
\end{algorithm}
For the sake of analysis on binomial crossover excluding influence of population, $(1+1)EA$ (presented by Algorithm \ref{Alg1}) without crossover is taken as the baseline algorithm in our study. Its candidate solutions are generated by the bitwise mutation with probability $p_m$. Binomial crossover is added to $(1+1)EA$, getting $(1+1)EA_C$ which is illustrated in Algorithm \ref{Alg3}. The $(1+1)EA_{MC}$ first performs bitwise mutation with probability $q_m$, and then applies binomial crossover with rate $C_R$ to generate a candidate solution for selection.
\begin{algorithm}[ht]
\caption{$(1+1)EA_C$}\label{Alg3}
\begin{algorithmic}[1]
\STATE counter $t=0$;
\STATE randomly generate a solution $\mathbf x_0=(x_1,\dots,x_n)$;
\WHILE{the stopping criterion is not satisfied}
\STATE set $rndi\sim U\{1,2,\dots,n\}$;
\FOR{$i=1,2,\dots,n$}
\STATE \begin{equation}\label{Mut2}v_i=\left\{\begin{aligned}& 1-x_i, && \mbox{if } rnd1_i<q_m,\\ & x_i, && \mbox{otherwise},\end{aligned}\right.\quad rnd1_i\sim U[0,1];\end{equation}
\STATE \begin{equation}\label{Cross}y_i=\left\{\begin{aligned}& v_i, && \mbox{if } i=rndi \mbox{ or } rnd2_i<C_R,\\ & x_i, && \mbox{otherwise},\end{aligned}\right.\quad rnd2_i\sim U[0,1]\end{equation}
\ENDFOR
\IF{$f(\mathbf{y})\ge f(\mathbf{x}_t)$}
\STATE $\mathbf{x}_{t+1}=\mathbf{y}$;
\ENDIF
\STATE $t=t+1$;
\ENDWHILE
\end{algorithmic}
\end{algorithm}
The EAs investigated in this paper can be modeled as homogeneous Markov chains~\citep{he2003towards,He2016}. Given the error vector
\begin{equation}\label{ErrVec}
\mathbf{\tilde{e}}=(e_0,e_1,\dots,e_L)',
\end{equation}
and the initial distribution
\begin{equation}\label{IniDis}
\mathbf{\tilde{q}}^{[0]}=(q_0^{[0]},q_1^{[0]},\dots,q_L^{[0]})'
\end{equation}
the transition matrix of $(1+1)EA$ and $(1+1)EA_C$ for the optimization problem (\ref{OP1}) can be written in the form
\begin{equation}\label{Trans}
\mathbf{\tilde{R}}=(r_{i,j})_{(L+1)\times (L+1)},
\end{equation}
where
\begin{equation*
r_{i,j}=\Pr\{\mathbf{x}_{t+1}\in\mathcal{X}_i\mid\mathbf{x}_{t}\in\mathcal{X}_j\},\quad i,j=0,\dots,L.
\end{equation*}
Recalling that the solutions are updated by the elitist selection, we know $\mathbf{\tilde{R}}$ is an upper triangular matrix that can be partitioned as
\begin{equation*
\mathbf{\tilde{R}}=\left(
\begin{array}{cc}
1 & \mathbf{r}_0 \\
\mathbf{0} & \mathbf{R} \\
\end{array}
\right),
\end{equation*}
where $\mathbf{R}$ is the transition submatrix depicting the transitions between non-optimal states.
\subsection{Transition Matrices}
By elitist selection, a candidate $\mathbf{y}$ replaces a solution $\mathbf{x}$ if and only if $f(\mathbf{y})\ge f(\mathbf{x})$, which is achieved if ``$l$ preferred bits'' of $\mathbf{x}$ are changed. If there are multiple solutions that are better than $\mathbf{x}$, there could be multiple choices for both the number of mutated bits $l$ and the location of ``$l$ preferred bits''.
\begin{example}\label{exam1}
For the OneMax problem, $e(\mathbf x)$ equals to the amount of `0'-bits in $\mathbf{x}$. Denoting $e(\mathbf{x})=j$ and $e(\mathbf{y})=i$, we know $\mathbf{y}$ replaces $\mathbf x$ if and only if $j\ge i$. Then, to generate a candidate $\mathbf y$ replacing $\mathbf x$, ``$l$ preferred bits'' can be confirmed as follows.
\begin{itemize}
\item If $i=j$, ``$l$ preferred bits'' consist of $l/2$ `1'-bits and $l/2$ `0'-bits, where $l$ is an even number that is not greater than $\min\{2j,2(n-j)\}$.
\item While $i<j$, ``$l$ preferred bits'' could be combinations of $j-i+k$ `0'-bits and $k$ `1'-bits ($l=j-i+2k$), where $0\le k\le\min\{i,n-j\}$. Here, $k$ is not greater than $i$, because $j-i+k$ could not be greater than $j$, the number of `0'-bits in $\mathbf{x}$. Meanwhile, $k$ does not exceed $n-j$, the number of `1'-bits in $\mathbf{x}$.
\end{itemize}
\end{example}
If an EA flips each bit with an identical probability, the probability of flipping $l$ bits are related to $l$ and independent of their locations. Denoting the probability of flipping $l$ bits by $P(l)$, we can confirm the connection between the transition probability $r_{i,j}$ and $P(l)$.
\subsubsection{Transition Probabilities for OneMax}
As presented in Example \ref{exam1}, transition from level $j$ to level $i$ ($i<j$) results from flips of $j-i+k$ `0'-bits and $k$ `1'-bits. Then,
\begin{equation}\label{POne}
{r}_{i,j}=\sum_{k=0}^{M}{C_{n-j}^{k}C_{j}^{k+\left( j-i \right)}}P(2k+j-i),
\end{equation}
where $M=\min \left\{ n-j,i \right\}$, $0\le i<j\le n$.
\subsubsection{Transition Probabilities for Deceptive}
According to definition of the Deceptive problem, we get the following map from $\mid\mathbf{x}\mid$ to $e(\mathbf{x})$.
\begin{equation}\label{DecMap}
\begin{matrix}
& \mid\mathbf{x}\mid: & 0 & 1 & \cdots & n-1 & n \\
& e(\mathbf{x}): & 1 & 2 & \cdots & n & 0 \\
\end{matrix}
\end{equation}
Transition from level $j$ to level $i$ ($0\le i<j\le n$) is attributed to one of the following cases.
\begin{itemize}
\item If $i\ge 1$, the amount of `1'-bits decreases from $j-1$ to $i-1$. This transition results from change of $j-i+k$ `1'-bits and $k$ `0'-bits, where
$0\le k\le \min\{n-j+1,i-1\}$;
\item if $i=0$, all of $n-j+1$ `0'-bits are flipped, and all of its `1'-bits keep unchanged.
\end{itemize}
Accordingly, we know
\begin{align}
& {r}_{i,j}=\left\{\begin{aligned}&\sum_{k=0}^{M}{C_{n-j+1}^{k}C_{j-1}^{k+\left( j-i \right)}}P(2k+j-i), && i\ge 1,\\
& P(n-j+1), && i=0, \end{aligned}
\right.\label{PDec}
\end{align}
where $M=\min \left\{ n-j+1,i-1 \right\}$.
\subsection{Performance Metrics}
We propose two metrics to evaluate the performance of EAs,
which are the expected approximation error (EAE) and the tail probability (TP) of EAs for $t$ consecutive iterations. The approximation error was considered in previous work~\cite{he2016analytic} but the tail probability is a new performance metric.
\begin{definition}
Let $\{\mathbf{x}_t,t=1,2\dots\}$ be the individual sequence of an individual-based EA. The expected approximation error (EAE) after $t$ consecutive iterations is
\begin{equation}
e^{[t]}=\mathbb{E}[e(\mathbf x_t)]=\sum_{i=0}^{L}e_i\Pr\{e(\mathbf x_t)=e_i\}.
\end{equation}
\end{definition}
EAE is the fitness gap between a solution and the optimum. It measures solution quality after running $t$ generations.
\begin{definition} Given $i>0$,
the tail probability (TP) of the approximation error that $e(\mathbf{x}_t)$ is greater than or equal to $e_i$ is defined as
\begin{equation
p^{[t]}(e_i)=\Pr\{e(\mathbf x_t)\ge e_i\}.
\end{equation}
\end{definition}
TP is the probability distribution of a found solution over non-optimal levels where $i >0$. The sum of TP is the probability of not finding the optimum.
Given two EAs $\mathcal{A}$ and $\mathcal{B}$, if both EAE and TP of Algorithm $\mathcal{A}$ are smaller than those of Algorithm $\mathcal{B}$ for any iteration budget, we say Algorithm $\mathcal{A}$ outperforms Algorithm $\mathcal{B}$ on problem (\ref{OP1}).
\begin{definition}\label{Def2}
Let $\mathcal{A}$ and $\mathcal{B}$ be two EAs applied to problem (\ref{OP1}).
\begin{enumerate}
\item Algorithm $\mathcal{A}$ {outperforms} $\mathcal{B}$, denoted by $\mathcal{A}\succsim \mathcal{B}$, if it holds that
\begin{itemize}
\item $e_{\mathcal A}^{[t]}- e_{\mathcal B}^{[t]} \le 0$, $\forall\,t>0$;
\item $p_{\mathcal A}^{[t]}(e_i)- p_{\mathcal B}^{[t]}(e_i)\le 0$, $\forall\,t>0$, $0<i<L$.
\end{itemize}
\item Algorithm $\mathcal{A}$ {asymptotically outperforms} $\mathcal{B}$ on problem (\ref{OP1}), denoted by $\mathcal{A}\succsim^{a} \mathcal{B}$, if it holds that
\begin{itemize}
\item $ \lim_{t\to \infty } e_{\mathcal A}^{[t]}- e_{\mathcal B}^{[t]} \le 0$;
\item $\lim_{t\to+\infty} p_{\mathcal A}^{[t]}(e_i)- p_{\mathcal B}^{[t]}(e_i) \le 0$.
\end{itemize}
\end{enumerate}
\end{definition}
The asymptotic outperformance is weaker than the outperformance.
\section{Comparison of Transition Probabilities of the Two EAs} \label{SecTrans}
In this section, we
compare transition probabilities of $(1+1)EA$ and $(1+1)EA_C$. According to the connection between $r_{i,j}$ and $P(l)$, comparison of transition probabilities can be conducted by considering the probabilities of flipping ``$l$ preferred bits''.
\subsection{Probabilities of Flipping Preferred Bits }
Denote probabilities of $(1+1)EA$ and $(1+1)EA_C$ of flipping ``$l$ preferred bits'' by $P_1(l,p_m)$ and $P_2(l,C_R,q_m)$, respectively. By (\ref{Mut1}), we know
\begin{align}
P_1(l,p_m)=(p_m)^l(1-p_m)^{n-l}. \label{P1
\end{align}
Since the mutation and the binomial crossover in Algorithm \ref{Alg3} are mutually independent, we can get the probability by considering the crossover first. When flipping ``$l$ preferred bits'' by the $(1+1)EA_C$, there are $l+k$ ($0\le k\le n-l$) bits of $\mathbf{y}$ set as $v_i$ by (\ref{Cross}), the probability of which is
\begin{equation*}
P_{C}(l+k,C_R)=\frac{l+k}{n}(C_R)^{l+k-1}(1-C_R)^{n-l-k}.
\end{equation*}
If only ``$l$ preferred bits'' are flipped, we know,
\begin{align}
P_2(l,C_R,q_m)&=\sum_{k=0}^{n-l}C_{n-l}^{k}P_C(l+k,C_R)(q_m)^l\left(1-q_m\right)^k \nonumber\\
&=\frac{1}{n}\left[l+(n-l)C_R-nq_mC_R\right](C_R)^{l-1}(q_m)^l\left(1-q_mC_R\right)^{n-l-1}.\label{P3}
\end{align}
Note that $(1+1)EA_C$ degrades to $(1+1)EA$ when $C_R=1$, and $(1+1)EA$ becomes the random search while $p_m=1$. Thus, we assume that $p_m$, $C_R$ and $q_m$ are located in $(0,1)$. For a fair comparison of transition probabilities, we consider the identical parameter setting
\begin{equation}\label{ParaSetting}
p_m=C_Rq_m=p,\quad 0<p<1.
\end{equation}
Then, we know $q_m=p/C_R$, and equation (\ref{P3}) implies
\begin{equation}\label{P3'}
{P_2(l,C_R,p/C_R)=\frac{1}{n}\left[(n-l)+\frac{l-np}{C_R}\right]p^l(1-p)^{n-l-1}}.
\end{equation}
Subtracting (\ref{P1}) from (\ref{P3'}), we have
\begin{align}
& {P_2(l,C_R,p/C_R)-P_1(l,p)
={\left\{\frac{1}{n}\left[(n-l)+\frac{l-np}{C_R}\right]-(1-p)\right\}p^l(1-p)^{n-l-1}}\nonumber\\
=&{\left(\frac{1}{C_R}-1\right)\left(\frac{l}{n}-p\right)p^l(1-p)^{n-l-1}}.\label{DP1P3}
\end{align}
From the fact that $0<C_R<1$, we conclude that $P_2(l,C_R,p/C_R)$ is greater than $P_1(l,p)$ if and only if $l>np$. That is, the introduction of the binomial crossover in the $(1+1)EA$ leads to the enhancement of exploration ability of the $(1+1)EA_C$. We get the following theorem for the case that $p\le \frac{1}{n}$.
\begin{theorem}\label{T2}
While $0<p\le \frac{1}{n}$, it holds for all $1\le l\le n$ that
$P_1(l,p)\le P_2(l,C_R,p/C_R).$
\end{theorem}
\begin{proof}
The result can be obtained directly from equation (\ref{DP1P3}) by setting $p\le\frac{1}{n}$.
\end{proof}
\subsection{Comparison of Transition Probabilities}
Given transition matrices from two EAs, one transition matrix dominating another is defined as follows.
\begin{definition}\label{Def1}
Let $\mathcal{A}$ and $\mathcal{B}$ be two EAs with an identical initialization mechanism. $\mathbf{\tilde{A}}=(a_{i,j})$ and $\mathbf{\tilde{B}}=(b_{i,j})$ are the transition matrices of $\mathcal{A}$ and $\mathcal{B}$, respectively.
It is said that $\mathbf{\tilde{A}}$ \textbf{dominates} $\mathbf{\tilde{B}}$, denoted by $\mathbf{\tilde{A}}\succeq\mathbf{\tilde{B}}$, if it holds that
\begin{enumerate}
\item $a_{i,j}\ge b_{i,j},\quad\forall \,0\le i<j\le L$;
\item $a_{i,j}> b_{i,j},\quad\exists\, 0\le i<j\le L$.
\end{enumerate}
\end{definition}
Denote the transition probabilities of $(1+1)EA$ and $(1+1)EA_C$ by $p_{i,j}$ and $s_{i,j}$, respectively. For the OneMax problem and Deceptive problem, we get the relation of transition dominance on the premise that $p_m=C_Rq_m=p\le \frac{1}{n}$.
\begin{theorem}\label{T4}
For $(1+1)EA$ and $(1+1)EA_C$, denote their transition matrices by $\mathbf{\tilde{P}}$ and $\mathbf{\tilde{S}}$, respectively. On the condition that $p_m=C_Rq_m=p\le \frac{1}{n}$, it holds for problem (\ref{OP1}) that
\begin{equation}\label{temp23}
\mathbf{\tilde{S}} \succeq \mathbf{\tilde{P}}.
\end{equation}
\end{theorem}
\begin{proof}
Denote the collection of all solutions at level $k$ by $\mathcal{S}(k)$, $k=0,1,\dots,n$. We prove the result by considering the transition probability
$$r_{i,j}=\Pr\{\mathbf{y}\in\mathcal{S}(i)\mid\mathbf{x}\in\mathcal{S}(j)\},\quad (i< j).$$
Since the function values of solutions are only related to the number of `1'-bits, the probability to generate a solution $\mathbf{y}\in\mathcal{S}(i)$ by performing mutation on $\mathbf{x}\in\mathcal{S}(j)$ depends on the Hamming distance
\begin{equation*}
l=H(\mathbf{x},\mathbf{y}).
\end{equation*}
Given $\mathbf{x}\in\mathcal{S}_j$, $\mathcal{S}(i)$ can be partitioned as
$$\mathcal{S}(i)=\bigcup_{l=1}^{L}\mathcal{S}_l(i),$$
where $\mathcal{S}_l(i)=\{\mathbf{y}\in\mathcal{S}(i)\mid H(\mathbf{x},\mathbf{y})=l\}$, and
$L$ is a positive integer that is smaller than or equal to $n$.
Accordingly, the probability to transfer from level $j$ to $i$ is confirmed as
\begin{align*}
r_{i,j}=\sum_{l=1}^{L}\Pr\{\mathbf{y}\in\mathcal{S}_l(i)\mid\mathbf{x}\in\mathcal{S}(j)\}=\sum_{l=1}^{L}\mid\mathcal{S}_l(i)\mid P(l),
\end{align*}
where $\mid\mathcal{S}_l(i)\mid$ is the size of $\mathcal{S}_l(i)$, $P(l)$ the probability to flip ``$l$ preferred bits''. Thus, we have
\begin{align}
& p_{i,j}=\sum_{l=1}^{L}\Pr\{\mathbf{y}\in\mathcal{S}_l(j)\mid\mathbf{x}\}=\sum_{l=1}^{L}\mid\mathcal{S}_l(j)\mid P_1(l,p),\label{r1}\\
& s_{i,j}=\sum_{l=1}^{L}\Pr\{\mathbf{y}\in\mathcal{S}_l(j)\mid\mathbf{x}\}=\sum_{l=1}^{L}\mid\mathcal{S}_l(j)\\mid P_2(l,C_R,p/C_R).\label{r3}
\end{align}
Since $p\le 1/n$, Theorem \ref{T2} implies that $$ P_1(l,p)\le P_2(l,C_R,p/C_R),\quad\forall\,\,1\le l\le n.$$
Combining it with (\ref{r1}) and (\ref{r3}) we know
\begin{equation}\label{temp21}
p_{i,j}\le s_{i,j}, \quad\forall\,\, 0\le i<j\le n.\end{equation}
Then, we get the result by Definition \ref{Def2}.
\end{proof}
\begin{example}\textbf{[Comparison of transition probabilities for the OneMax problem]}
Let $p_m=C_Rq_m=p\le\frac{1}{n}$. By (\ref{POne}), we have
\begin{align}
&{p_{i,j}}=\sum_{k=0}^M{C_{n-j}^{k}C_{j}^{k+\left( j-i \right)}}P_1(2k+j-i,p),\label{TP1}\\
&{s_{i,j}}=\sum_{k=0}^M{C_{n-j}^{k}C_{j}^{k+\left( j-i \right)}}P_2(2k+j-i,C_R,p/C_R).\label{TP3}
\end{align}
where $M={\min \left\{ n-j,i \right\}}$. Since $p\le 1/n$, Theorem \ref{T2} implies that $$ P_1(2k+j-i,p)\le P_2(2k+j-i,C_R,p/C_R),$$
and by (\ref{TP1}) and (\ref{TP3}) we have
$$p_{i,j}\le s_{i,j}, \quad\forall\,\, 0\le i<j\le n.$$
\end{example}
\begin{example}\label{exam3}
\textbf{[Comparison of transition probabilities for the Deceptive problem]}
Let $p_m=C_Rq_m=p\le\frac{1}{n}$. Equation (\ref{PDec}) implies that
\begin{equation}\label{TP11}
{p_{i,j}}=\left\{\begin{aligned}&\sum_{k=0}^{M}{C_{n-j+1}^{k}C_{j-1}^{k+\left( j-i \right)}}P_1(2k+j-i,p), && i>0,\\
& P_1(n-j+1,p), && i=0, \end{aligned}
\right.
\end{equation}
\begin{equation}\label{TP33}
{s_{i,j}}=\left\{\begin{aligned}& \sum_{k=0}^{M}{C_{n-j+1}^{k}C_{j-1}^{k+\left( j-i \right)}}P_2(2k+j-i,C_R,\frac{p}{C_R}), && i>0,\\
& P_2(n-j+1,C_R,p/C_R), && i=0, \end{aligned}
\right.
\end{equation}
where $M=\min \left\{ n-j+1,i-1 \right\}$. Similar to analysis of the Example 2, we know when $p\le 1/n$,
$$p_{i,j}\le s_{i,j}.$$
\end{example}
Nevertheless, if $p>\frac{1}{n}$, we cannot get Theorems \ref{T2}. Since
the differences among $p_{i,j}$ and $q_{i,j}$ depends on the characteristics of problem (\ref{OP1}), Theorem \ref{T4} does not hold, too.
\section{Analysis of Asymptotic Performance}\label{SecAsy}
Since the transition matrix of $(1+1)EA_C$ dominates that of $(1+1)EA$, this section proves that $(1+1)EA_C$ asymptotically outperforms $(1+1)EA$ using the average convergence rate~\cite{He2016,CHEN2021200}.
\begin{definition}
The average convergence rate (ACR) of an EA for $t$ generation is
\begin{equation
R_{EA}(t)=1-\left({e^{[t]}}/{e^{[0]}}\right)^{1/t}.
\end{equation}
\end{definition}
The following lemma presents the asymptotic characteristics of the ACR, by which
we get the result on the asymptotic performance of EAs.
\begin{lemma}\label{Pro6}
\textbf{\citep[Theorem 1]{He2016}}
Let $\mathbf{R}$ be the transition submatrix associated with a convergent EA. Under random initialization (i.e., the EA may start at any initial state with a positive probability), it holds
\begin{equation
\lim_{t\to +\infty}R_{EA}(t)=1-\rho(\mathbf{R}),
\end{equation}
where $\rho(\mathbf{R})$ is the spectral radius of $\mathbf{R}$.
\end{lemma}
\begin{lemma}\label{Pro7}
If $\mathbf{\tilde{A}}\succeq\mathbf{\tilde{B}}$, there exists $T>0$ such that
\begin{enumerate}
\item $e_{\mathcal{A}}^{[t]}\le e_{\mathcal{B}}^{[t]}$, $\forall\,t>T$;
\item $p_{\mathcal{A}}^{[t]}(e_i)\le p_{\mathcal{B}}^{[t]}(e_i)$, $\forall\,t>T, \, 1\le i\le L$.
\end{enumerate}
\end{lemma}
\begin{proof}
By Lemma \ref{Pro6}, we know $\forall\,\epsilon>0$, there exists $T>0$ such that
\begin{equation}\label{temp19}
e^{[0]}\left(\rho(\mathbf{R})-\epsilon\right)^t<e^{[t]} < e^{[0]}\left(\rho(\mathbf{R})+\epsilon\right)^t,\quad t>T.
\end{equation}
From the fact that the transition submatrix $\mathbf{R}$ of an RSH is upper triangular, we conclude
\begin{equation}\label{temp22}\rho(\mathbf{R})=\max\{r_{1,1},\dots, r_{L,L}\}.\end{equation}
Denote
$$\mathbf{\tilde{A}}=(a_{i,j})=\left(
\begin{array}{cc}
1 & \mathbf{a}_0 \\
\mathbf{0} & \mathbf{A} \\
\end{array}
\right),\quad \mathbf{\tilde{B}}=(b_{i,j})=\left(
\begin{array}{cc}
1 & \mathbf{b}_0 \\
\mathbf{0} & \mathbf{B} \\
\end{array}
\right).$$
While $\mathbf{\tilde{A}}\succeq\mathbf{\tilde{B}}$, it holds
$$ a_{j,j}=1-\sum_{i=0}^{j-1}a_{i,j}< 1-\sum_{i=0}^{j-1}b_{i,j}=b_{j,j},\,\,1\le j \le L.$$
Then, equation (\ref{temp22}) implies that
\begin{equation*}
\rho(\mathbf{A})< \rho (\mathbf{B}).
\end{equation*}
Applying it to (\ref{temp19}) for $\epsilon <\frac{1}{2}(\rho(\mathbf{B})-\rho(\mathbf{A}))$, we have
\begin{equation}\label{temp20}
e_{\mathcal{A}}^{[t]} < e^{[0]}\left(\rho(\mathbf{A})+\epsilon\right)^t<e^{[0]}\left(\rho(\mathbf{B})-\epsilon\right)^t<e_{\mathcal{B}}^{[t]},
\end{equation}
which proves the first conclusion.
Noting that the tail probability $p^{[t]}(e_i)$ can be taken as the expected approximation error of an optimization problem with error vector
$$\mathbf{e}=(\underbrace{0,\dots,0}_{i},1,\dots,1)',$$ by (\ref{temp20}) we have
$$p_{\mathcal{A}}^{[t]}(e_i)\le p_{\mathcal{B}}^{[t]}(e_i),\quad \forall\,t>T, \, 1\le i\le L.$$
The second conclusion is proven.
\end{proof}
Definition \ref{Def2} and Proposition \ref{Pro7} imply that dominance of transition matrix lead to the asymptotic outperformance. Then, we get the following theorem for comparing the asymptotic performance of $(1+1)EA$ and $(1+1)EA_C$.
\begin{theorem}\label{T5}
If $C_R=C_Rq_m=p\le \frac{1}{n}$, the $(1+1)EA_C$ {asymptotically outperforms} the $(1+1)EA$ on problem (\ref{OP1}).
\end{theorem}
\begin{proof}
The proof can be completed by applying Theorem \ref{T4} and Lemma \ref{Pro7}.
\end{proof}
On condition that $C_R=C_Rq_m=p\le \frac{1}{n}$, Theorem \ref{T5} indicates that after sufficiently many number of iterations, the $(1+1)EA_C$ can performs better on problem (\ref{OP1}) than the $(1+1)EA$.
A further question is whether the $(1+1)EA_C$ outperforms the $(1+1)EA$ for $ t < +\infty$. We answer the question in next sections.
\section{Comparison of the Two EAs on OneMax}\label{SecExpi}
In this section, we show that the outperformance introduced by binomial crossover can be obtained for the unimodel OneMax problem based on the following lemma~\citep{WANG2021}.
\begin{lemma}\label{Th_ES}\textbf{\citep[Theorem 3]{WANG2021}}
Let
\begin{align*}
&\mathbf{\tilde{e}}=(e_0,e_1,\dots,e_L)', \quad \mathbf{\tilde{v}}=(v_0,v_1,\dots,v_L)',
\end{align*}
where $0\le e_{i-1}\le e_{i}, i=1,\dots,L$, $v_i>0,i=0,1,\dots,L$. If transition matrices $\mathbf{\tilde{R}}$ and $\mathbf{\tilde{S}}$ satisfy
\begin{align}
\label{conC1}
&s_{j,j} \ge r_{j,j}, &&\forall\,\, j,
\\
\label{conC2}
& \sum^{i-1}_{l=0} (r_{l,j}-s_{l,j}) \ge 0 , &&\forall\,\,i<j,
\\
\label{conC3}
& \sum^{i}_{l=0} ( s_{l,j-1}- s_{l,j})\ge 0 , &&\forall\,\,i<j-1,
\end{align}
it holds
$$\mathbf{\tilde{e}}'\mathbf{\tilde{R}}^t\mathbf{\tilde{v}}\le \mathbf{\tilde{e}}'\mathbf{\tilde{S}}^t\mathbf{\tilde{v}}.$$
\end{lemma}
For the EAs investigated in this study, conditions (\ref{conC1})-(\ref{conC3}) are satisfied thanks to the monotonicity of transition probabilities.
\begin{lemma}\label{L2}
When $p\le 1/n$ ($n\ge 3$), $P_1(l,p)$ and $P_2(l,C_R,p/C_R)$ are monotonously decreasing in $l$.
\end{lemma}
\begin{proof}
When $p\le 1/n$, equations (\ref{P1}) and (\ref{P3}) imply that
\begin{align}
& \frac{P_1(l+1,p)}{P_1(l,p)}=\frac{p}{1-p} \le \frac{1}{n-1},\label{temp4}\\
& \frac{P_2(l+1,C_R,p/C_R)}{P_2(l,C_R,p/C_R)}=\frac{(l+1)(1-C_R)+nC_R(1-p/C_R)}{l(1-C_R)+nC_R(1-p/C_R)}\frac{p}{1-p}\le \frac{l+1}{l}\frac{p}{1-p}\le \frac{l+1}{l}\frac{1}{n-1},
\end{align}
all of which are not greater than $1$ when $n\ge 3$.
Thus, $P_1(l,p)$ and $P_2(l,C_R,p/C_R)$ are monotonously decreasing in $l$.
\end{proof}
\begin{lemma}\label{L3}
For the OneMax problem, $p_{i,j}$ and $s_{i,j}$ are monotonously decreasing in $j$.
\end{lemma}
\begin{proof}
We validate the monotonicity of $p_{i,j}$ for the $(1+1)EA$, and that of $q_{i,j}$ and $s_{i,j}$ can be confirmed in a similar way.
Let $0\le i< j< n$. By (\ref{TP1}) we know
\begin{align}
& p_{i,j+1}=\sum_{k=0}^M{C_{n-j-1}^{k}C_{j+1}^{i-k}}P_1(2k+j+1-i,p),\label{temp1}\\
& p_{i,j}= \sum_{k=0}^M{C_{n-j}^{k}C_{j}^{i-k}}P_1(2k+j-i,p),\label{temp2}
\end{align}
where $M={\min \left\{ n-j-1,i \right\}}$. Moreover, (\ref{temp4}) implies that
\begin{align*}
& \frac{C_{j+1}^{i-k}P_1(2k+j+1-i,p)}{C_{j}^{i-k}P_1(2k+j-i,p)}=\frac{j+1}{(j+1)-(i-k)}\frac{p}{1-p}\le\frac{j+1}{2}\frac{1}{n-1}<1,
\end{align*}
and we know
\begin{equation}\label{temp3}
{C_{j+1}^{i-k}P_1(2k+j+1-i,p)}<{C_{j}^{i-k}P_1(2k+j-i,p)}.
\end{equation}
Note that
\begin{align}\label{temp14}
& \min \left\{ n-j-1,i \right\}\ge \min \left\{ n-j,i \right\},\quad C_{n-j-1}^{k}<C_{n-j}^{k}.
\end{align}
From (\ref{temp1}), (\ref{temp2}), (\ref{temp3}) and (\ref{temp14}) we conclude that $$p_{i,j+1}<p_{i,j},\quad 0\le i< j< n.$$
Similarly, we can validate that
$$ s_{i,j+1}<s_{i,j},\quad 0\le i< j< n.$$
In conclusion, $p_{i,j}$ and $s_{i,j}$ are monotonously decreasing in $j$.
\end{proof}
\begin{theorem}\label{T6}
On condition that $p_m=C_Rq_m=p\le \frac{1}{n}$, it holds for the OneMax problem that
\begin{equation*}
(1+1)EA_C \succsim (1+1)EA.
\end{equation*}
\end{theorem}
\begin{proof}
Given the initial distribution $\mathbf{\tilde{q}}^{[0]}$ and transition matrix $\mathbf{\tilde{R}}$, the level distribution at iteration $t$ is confirmed by
\begin{equation}\label{Dis_t}
\mathbf{\tilde{q}}^{[t]}=\mathbf{\tilde{R}}^t\mathbf{\tilde{q}}^{[0]}.
\end{equation}
Denote
\begin{align*}
&\mathbf{\tilde{e}}=(e_0,e_1,\dots,e_L)', \quad \mathbf{\tilde{o}}_i=(\underbrace{0,\dots,0}_{i},1,\dots,1)'.
\end{align*}
By premultiplying (\ref{Dis_t}) with $\mathbf{\tilde{e}}$ and $\mathbf{\tilde{o}}_i$, respectively, we get
\begin{align}
& e^{[t]}=\mathbf{\tilde{e}}'\mathbf{\tilde{R}}^t\mathbf{\tilde{q}^{[0]}},\label{EAE}\\
& p^{[t]}(e_i)=\Pr\{e(\mathbf{x}_t)\}\ge e_i\}=\mathbf{\tilde{o}}_i'\mathbf{\tilde{R}}^t\mathbf{\tilde{q}^{[0]}}.\label{TP}
\end{align}
Meanwhile, By Theorem \ref{T4} we have
\begin{align
& q_{j,j}\le s_{j,j} \le p_{j,j},\label{temp5}\\
& \sum^{i-1}_{l=0} (q_{l,j}-s_{l,j}) \ge 0 ,\quad \sum^{i-1}_{l=0} (s_{l,j}-p_{l,j}) \ge 0 , \quad \forall\,\,i<j,\label{temp6
\end{align}
and Lemma \ref{L3} implies
\begin{align}\label{temp7}
\sum^{i}_{l=0} ( s_{l,j-1}- s_{l,j})\ge 0 , \quad \sum^{i}_{l=0} ( p_{l,j-1}- p_{l,j})\ge 0 &&\forall\,\,i<j-1.
\end{align}
Then, (\ref{temp5}), (\ref{temp6}) and (\ref{temp7}) validate satisfaction of conditions (\ref{conC1})-(\ref{conC3}), and by Lemma \ref{Th_ES} we know
\begin{align*}
& \mathbf{\tilde{e}}'\mathbf{\tilde{S}}^t\mathbf{\tilde{q}^{[0]}}\le \mathbf{\tilde{e}}'\mathbf{\tilde{P}}^t\mathbf{\tilde{q}^{[0]}}, &&\forall t>0;\\
& \mathbf{\tilde{o}}_i'\mathbf{\tilde{S}}^t\mathbf{\tilde{q}^{[0]}}\le \mathbf{\tilde{o}}_i'\mathbf{\tilde{P}}^t\mathbf{\tilde{q}^{[0]}}, &&\forall t>0,\,1\le i< n.
\end{align*}
Then, we get the conclusion by Definition \ref{Def2}.
\end{proof}
The above theorem demonstrates that the dominance of transition matrices introduced by the binomial crossover operator leads to the outperformance of $(1+1)EA_C$ on the unimodal problem OneMax.
\section{Comparison of the two EAs on Deceptive and Adaptive Parameter Strategy}\label{SecExpr}
In this section, we show that the outperformance of $(1+1)EA_C$ over $(1+1)EA$ may not always hold on Deceptive, and then, propose an adaptive strategy of parameter setting arising from the theoretical analysis.
\subsection{A counterexample for inconsistency between the transition dominance and the algorithm outperformance}
For the Deceptive problem, we present a counterexample to show even if the transition matrix of an EA dominates another EA, we cannot draw that the former EA outperform the later.
\begin{example}\label{Exam}
We construct two artificial Markov chains as the models of two EAs.
Let $EA_\mathcal{R}$ and $EA_\mathcal{S}$ be two EAs staring with an identical initial distribution
$$\mathbf{p}^{[0]}=\left(\frac{1}{n},\frac{1}{n},\dots,\frac{1}{n}\right)^t,$$
and the respective transition matrices are
\begin{align*}
\mathbf{\tilde{R}}=\begin{pmatrix}
1 & \frac{1}{n^3} & \frac{2}{n^3} & \dots &\frac{n}{n^3} \\
& 1-\frac{1}{n^3} & \frac{1}{n^2} & & \\
& & 1-\frac{1}{n^2}-\frac{2}{n^3} & \ddots & \\
& & & \ddots & \frac{n-1}{n^2}\\
& & & & 1-\frac{1}{n}\\
\end{pmatrix}
\end{align*}
and
\begin{align*}
\mathbf{\tilde{S}}=\begin{pmatrix}
1 & \frac{2}{n^3} & \frac{4}{n^3} &\dots &\frac{2n}{n^3} \\
& 1-\frac{2}{n^3} & \frac{1}{n^2}+\frac{1}{2n} & & \\
& & 1-\frac{n^2+2n+8}{2n^3} &\ddots & \\
& & &\ddots & \frac{n-1}{n^2}+\frac{n-1}{2n}\\
& & & & 1-\frac{n^2+n+2}{2n^2}\\
\end{pmatrix}.\\
\end{align*}
\end{example}
Obviously, it holds $\mathbf{\tilde{S}}\succeq \mathbf{\tilde{R}}$.
Through computer simulation, we get the curve of EAE difference of the two EAs in Figure~\ref{Test}(a) and the curve of TPs difference of the two EAs in Figure~\ref{Test}(b). From Figure~\ref{Test}(b), it is clear that $EA_\mathcal{R}$ does not always outperform $EA_\mathcal{S}$ because the difference of TPs is negative at the early stage of the iteration process but later positive.
\begin{figure}[!t]
\centering
\begin{minipage}[c]{0.48\textwidth}
\centering
\includegraphics[width=2.5in]{TestErr.eps}
\end{minipage}
\hspace{0.02\textwidth}
\begin{minipage}[c]{0.48\textwidth}
\centering
\includegraphics[width=2.5in]{TestPro.eps}
\end{minipage}
\caption{Simulation results of the difference of EAEs and TPs for the artificial case. }
\label{Test}
\end{figure}
\subsection{Numerical comparison for the Two EAs on Deceptive}
Now we turn to discuss $(1+1)EA$ and $(1+1)EA_C$ on Deceptive. We demonstrate $(1+1)EA_C$ may not outperform $(1+1)EA$ over all generations although the transition matrix of $(1+1)EA_C$ dominates that of $(1+1)EA$.
\begin{example}
In $(1+1)EA$ and $(1+1)EA_C$, set $$ p_m=C_Rq_m=\frac{1}{n}.$$
For $(1+1)EA_C$, let $q_m=\frac{1}{2}$, $C_R=\frac{2}{n}$. The numerical simulation results of EAEs and TPS for 5000 independent runs are depicted in Figure \ref{fig2}.
It is shown that when $n\ge 9$, both EAEs and TPS of $(1+1)EA$ could be smaller than those of $(1+1)EA_C$. This indicates that the dominance of transition matrix does not always guarantee the outperformance of the corresponding algorithm.
\end{example}
\begin{figure}[!t]
\centering
\begin{minipage}[c]{0.48\textwidth}
\centering
\includegraphics[width=2.5in]{ComErrDec.eps}
\end{minipage}
\hspace{0.02\textwidth}
\begin{minipage}[c]{0.48\textwidth}
\centering
\includegraphics[width=2.5in]{ComProDec.eps}
\end{minipage}
\caption{Numerical comparison on expected approximation errors (EAEs) and tail probabilities (TPs) of $(1+1)EA$ and $(1+1)EA_C$ applied to the Deceptive problem, where $n$ refers to the problem dimension. }
\label{fig2}
\end{figure}
With $p_m=C_Rq_m=p\le \frac{1}{n}$, although the binomial crossover leads to transition dominance of the $(1+1)EA_C$ over the $(1+1)EA$, the enhancement of exploitation plays a governing role in the iteration process. Thus, the imbalance of exploration and exploitation leads to a poor performance of $(1+1)EA_C$ on some stage of the iteration process.
\subsection{Comparisons on the Probabilities to Transfer from Non-optimal Statuses to the Optimal Status}
As shown in the previous two counterexamples, the outperformance of $(1+1)EA){MC}$ cannot be drawn from dominance of transition matrices. To enhance the performance of an EA, adaptive parameter settings should be incorporated to improve the exploration ability of an EA on Deceptive.
Local exploitation could result in convergence to the local optimal solution, global convergence of EAs on Deceptive is principally attributed to the direct transition from level $j$ to level $0$, which is quantified by the transition probability $r_{0,j}$. Thus, we investigate the impact of binomial crossover on the transition probability $r_{0,j}$, and accordingly, arrive at the strategies for adaptive regulations of the mutation rate and the crossover rate.
In the following, we first compare $p_{0,j}$ and $s_{0,j}$ by investigating their monotonicity, and then, get reasonable adaptive strategies that improve the performance of EAs on the Deceptive problem.
Substituting (\ref{P1}) and (\ref{P3}) to (\ref{TP11}) and (\ref{TP33}), respectively, we have
\begin{align
& p_{0,j} =P_1(n-j+1,p_m)=(p_m)^{n-j+1}(1-p_m)^{j-1}, \label{TP111}\\
& s_{0,j} =P_{3}(n-j+1,C_R,q_m)\nonumber\\
& = \frac{1}{n}\left[(j-1)(1-C_R)+nC_R(1-q_m)\right]C_R^{n-j}(q_m)^{n-j+1}\left(1-q_mC_R\right)^{j-2}. \label{TP333}
\end{align}
We first investigate the maximum values of $p_{0,j}$ to get the ideal performance of $(1+1)EA$ on the Deceptive problem.
\begin{theorem}\label{Pro1}
While
\begin{align}
& p^{\star}_m=\frac{n-j+1}{n},\label{temp10}
\end{align}
$p_{0,j}$ gets its maximum values
\begin{equation}
p_{0,j}^{max}=\left(\frac{n-j+1}{n}\right)^{n-j+1}\left(\frac{j-1}{n}\right)^{j-1}.
\end{equation}
\end{theorem}
\begin{proof}
By (\ref{TP111}), we know
\begin{align*}
{\frac{\partial}{\partial(p_m)}p_{0,j}=(n-j+1-np_m)p_m^{n-j}\left(1-p_m\right)^{j-2}}.
\end{align*}
While $ p_m=\frac{n-j+1}{n}$,
$p_{0,j}$ gets its maximum value
\begin{equation*}
p_{0,j}^{max}=P_1(n-j+1,\frac{n-j+1}{n})=\left(\frac{n-j+1}{n}\right)^{n-j+1}\left(\frac{j-1}{n}\right)^{j-1}.
\end{equation*}
\end{proof}
Influence of the binomial crossover on $s_{0,j}$ is investigated on condition that $p_m=q_m$. By regulating the crossover rate $C_R$, we can compare $p_{0,j}$ with the maximum value $s_{0,j}^{max}$ of $s_{0,j}$.
\begin{theorem}\label{Pro5}
On condition that $p_m=q_m$, the following results hold.
\begin{enumerate}
\item $p_{0,1}= s^{max}_{0,1}$.
\item If $q_m>\frac{n-1}{n}$, $p_{0,2}< s^{max}_{0,2}$; otherwise, $p_{0,2}= s^{max}_{0,2}$.
\item $\forall\,\, j\in\{3,\dots,n-1\}$, $p_{0,j}\le s^{max}_{0,j}$ if $q_m>\frac{n-j}{n-1}$; otherwise, $s^{max}_{0,j}=p_{0,j}$.
\item if $q_m>\frac{1}{n}$, $p_{0,n}< s^{max}_{0,n}$; otherwise, $s^{max}_{0,n}=p_{0,n}$.
\end{enumerate}
\end{theorem}
\begin{proof}
Note that $(1+1)EA_C$ degrades to $(1+1)EA$ when $C_R=1$. Then, if the maximum value $s_{0,j}^{max}$ of $s_{0,j}$ is obtained by setting $C_R=1$, we have $s_{0,j}^{max}=p_{0,j}$; otherwise, it holds $s_{0,j}^{max}>p_{0,j}$.
\begin{enumerate}
\item For the case that $\boldsymbol{j=1}$, equation (\ref{TP333}) implies
\begin{equation*}
{{s}_{0,1}}=q_{m}^{n}{{\left( C_R \right)}^{n-1}}.
\end{equation*}
Obviously, $s_{0,1}$ is monotonously increasing in $C_R$. It gets the maximum value while $C_R^{\star}=1$. Then, by (\ref{TP111}) we have $s^{max}_{0,1}= p_{0,1}$.
\item While $\boldsymbol{j=2}$, by (\ref{TP333}) we have
\begin{equation*}
\frac{\partial s_{0,2}}{\partial C_R}=\frac{n-1}{n}q_{m}^{n-1}{{\left( C_R \right)}^{n-3}}\left( n-2\text{+}\left( 1-n{{q}_{m}} \right)C_R \right).
\end{equation*}
\begin{itemize}
\item If $0<{{q}_{m}}\le \frac{n-1}{n}$, ${s}_{0,2}$ is monotonously increasing in $C_R$, and gets its maximum value while $C_R^{\star}=1$. For this case, we know $s^{max}_{0,2}= p_{0,2}$.
\item While $\frac{n-1}{n}<{q}_{m}<1$, ${s}_{0,2}$ gets its maximum value $s_{0,2}^{max}$ by setting
\begin{equation} C_R^{\star}=\frac{n-2}{nq_m-1}.\end{equation}
Then, we have $s_{0,2}^{max}>p_{0,2}$.
\end{itemize}
\item For the case that $\boldsymbol{3\le j\le n-1}$, we denote
\begin{equation*}
{s}_{0,j}=\frac{n-j+1}{n}q_{m}^{n-j+1}I_1+\frac{\left( j-1 \right)\left( 1-{{q}_{m}} \right)}{n}q_{m}^{n-j+1}I_2,
\end{equation*}
where
\begin{align*}
& I_1={{\left( C_R \right)}^{n-j}}{{\left( 1-{{q}_{m}}C_R \right)}^{j-1}}, \\
& I_2={{\left( C_R \right)}^{n-j+1}}{{\left( 1-{{q}_{m}}C_R \right)}^{j-2}}.
\end{align*}
Then,
\begin{equation*}
\frac{\partial I_1}{\partial C_R}={{\left( C_R \right)}^{n-j-1}}{{\left( 1-{{q}_{m}}C_R \right)}^{j-2}}\left( n-j-\left( n-1 \right){{q}_{m}}C_R \right),
\end{equation*}
\begin{equation*}
\frac{\partial I_2}{\partial C_R}={{\left( C_R \right)}^{n-j}}{{\left( 1-\frac{C_R}{n} \right)}^{j-3}}\left( n-j+1-\left( n-1 \right){{q}_{m}}C_R \right).
\end{equation*}
\begin{itemize}
\item While $0<q_m\le\frac{n-j}{n-1}$, both $I_1$ and $I_2$ are monotonously increasing in $C_R$. For this case, $s_{0,j}$ gets its maximum value when $C_R^{\star}=1$, and we have $s^{max}_{0,j}= p_{0,j}$.
\item If $\frac{n-j+1}{n-1}\le q_m\le 1$, $I_1$ gets its maximum value when $C_R=\frac{n-j}{(n-1)q_m}$, and $I_2$ gets its maximum value when $C_R=\frac{n-j+1}{(n-1)q_m}$. Then, $s_{0,j}$ could get its maximum value $s_{0,j}^{max}$ at some
\begin{equation}\label{temp8}
F^{\star}_R\in \left(\frac{n-j}{(n-1)q_m},\frac{n-j+1}{(n-1)q_m}\right).\end{equation}
Then, we kmow $s^{max}_{0,j}> p_{0,j}$.
\item If $\frac{n-j}{n-1}< q_m< \frac{n-j+1}{n-1}$, $I_1$ gets its maximum value when $C_R=\frac{n-j}{(n-1)q_m}$, and $I_2$ is monotonously increasing in $C_R$. Then, $s_{0,j}$ could get its maximum value $s_{0,j}^{max}$ at some
\begin{equation}\label{temp15} F^{\star}_R\in \left(\frac{n-j}{(n-1)q_m},1\right].\end{equation}
Then,
$s_{0,j}^{max}>p_{0,j}$.
\end{itemize}
\item While {$\boldsymbol{j=n}$}, equation (\ref{TP333}) implies that
\begin{align*}
\frac{\partial s_{0,n}}{\partial C_R}=\left( n-1 \right){{\left( 1-{{q}_{m}}C_R \right)}^{n-3}}\left( 1-2{{q}_{m}}-\left( n-1-n{{q}_{m}} \right){{q}_{m}}C_R \right).
\end{align*}
Denoting
\begin{equation*}g\left(q_m, C_R \right)=1-2{{q}_{m}}-\left( n-1-n{{q}_{m}} \right){{q}_{m}}C_R,\end{equation*}
we can confirm the sign of ${\partial s_{0,n}}/{\partial C_R}$ by considering
$$\frac{\partial}{\partial C_R}g\left(q_m, C_R \right)=-\left( n-1-n{{q}_{m}} \right){{q}_{m}}.$$
\begin{itemize}
\item While $0<{{q}_{m}}\le \frac{n-1}{n}$, $g\left(q_m, C_R \right)$ is monotonously decreasing in $C_R$, and its minimum value is
$$g\left(q_m, 1 \right)=\left( n{{q}_{m}}-1 \right)\left( {{q}_{m}}-1 \right).$$
The maximum value of $g\left(q_m, C_R \right)$ is
$$g\left(q_m, 0 \right)=1-2q_m.$$
\begin{enumerate}
\item If $0<{{q}_{m}}\le \frac{1}{n}$, we have
$$g\left(q_m, C_R \right)\ge g\left(q_m, 1 \right)>0.$$
Thus, $\frac{\partial s_{0,n}}{\partial C_R}\ge 0$, and ${{s}_{0,n}}$ is monotonously increasing in $C_R$. For this case, $s_{0,n}$ get its maximum value when $C_R^{\star}=1$, and we have $s_{0,n}^{max}=p_{0,n}$.
\item If $\frac{1}{n}<{{q}_{m}} \le \frac{1}{2}$, ${{s}_{0,n}}$ gets the maximum value $s_{0,n}^{max}$ when
$$ C_R^{\star}=\frac{1-2q_m}{q_m(n-1-nq_m)}.$$
Thus, $s_{0,n}^{max}>p_{0,n}$.
\item If $\frac{1}{2}<{{q}_{m}} \le \frac{n-1}{n}$, $g\left(q_m, 0 \right)<0$, and then, ${{s}_{0,n}}$ is monotonously decreasing in $C_R$. Then, its maximum value is obtained by setting $C_R^{\star}=0$. Then, we know $s^{max}_{0,n}>p_{0,n}$.
\end{enumerate}
\item While $ \frac{n-1}{n}< {{q}_{m}} \le 1$, $g\left(q_m, C_R \right)$ is monotonously increasing in $C_R$, and its maximum value
$$g\left(q_m, 1 \right)=\left( n{{q}_{m}}-1 \right)\left( {{q}_{m}}-1 \right)<0.$$
Then, ${{s}_{0,n}}$ is monotonously decreasing in $C_R$, and its maximum value is obtained by setting $C_R^{\star}=0$. Then, $s^{max}_{0,n}>p_{0,n}$.
\end{itemize}
In summary, $s^{max}_{0,n}>p_{0,n}$ while $q_m>\frac{1}{n}$; otherwise, $s^{max}_{0,n}=p_{0,n}$.
\end{enumerate}
\end{proof}
Theorems \ref{Pro1} and \ref{Pro5} presents the ``best'' settings to maximize the transition probabilities from non-optimal statues to the optimal level. Unfortunately, such results are not available if the global optima solution is unknown.
\subsection{Parameter Adaptive Strategy to Enhance Exploration of EAs}
We proposes a parameter Adaptive Strategy based on Hamming distance. Since the level index $j$ is equal to the Hamming distance between $\mathbf{x}$ and $\mathbf{x}^*$, improvement of level index $j$ is in deed equal to reduction of the Hamming distance obtained by replacing $\mathbf{x}$ with $\mathbf{y}$. Then, while the local exploitation leads to transition from level $j$ to a non-optimal level $i$, the practically adaptive strategy of parameters can be obtained according to the Hamming distance between $\mathbf{x}$ and $\mathbf{y}$.
For the Deceptive problem, consider two solutions $\mathbf{x}$ and $\mathbf{y}$ such that $f(\mathbf{y})>f(\mathbf{x})$, and denote their levels by $j$ and $i$, respectively. When the $(1+1)EA$ is located at the solution $\mathbf{x}$, equation (\ref{temp10}) implies that the ``best'' setting of mutation rate is $p^{\star}_m=\frac{n-j+1}{n}$. While the promising solution $\mathbf y$ is generated, the level transfers from $j$ to $i$, and the ``best'' setting change to $p^{\star}_m=\frac{n-i+1}{n}$. Let $H(\mathbf{x},\mathbf{y})$ denote the Hamming distance between $\mathbf{x}$ and $\mathbf{y}$. Noting that $H(\mathbf{x},\mathbf{y})\ge j-i$, we know the difference of ``best'' parameter settings is bounded from above by $\frac{H(\mathbf{x},\mathbf{y})}{n}$. Accordingly, the mutation rate of $(1+1)EA$ can be updated to
\begin{equation}\label{temp16}
p^{\prime}_m=p_m+\frac{H(\mathbf{x},\mathbf{y})}{n}.
\end{equation}
For the $(1+1)EA_C$, the parameter $q_m$ is adapted using the strategy consistent to that of $p_m$ to focus on influence of $C_R$. That is,
\begin{equation}\label{temp13}
q^{\prime}_m=q_m+\frac{H(\mathbf{x},\mathbf{y})}{n}.
\end{equation}
Since $s_{0,j}$ demonstrates different monotonicity for varied levels, one cannot get an identical strategy for adaptive setting of $C_R$. As a compromise, we would like to consider the case that $3\le j\le n-1$, which is obtained by random initialization with overwhelming probability.
According to the proof of Theorem \ref{Pro5}, we know $C_R$ should be set as great as possible for the case $q_m\in(0,\frac{n-j}{n-1}]$; while $q_m\in(\frac{n-j}{n-1}, 1]$, $C_R^{\star}$ is located in intervals whose boundary values are $\frac{n-j}{(n-1)q_m}$ and $\frac{n-j+1}{(n-1)q_m}$, given by (\ref{temp8}) and (\ref{temp15}), respectively. Then, while $q_m$ is updated by (\ref{temp13}), the update strategy of $C_R$ can be confirmed to satisfying that
$$ F^{\prime}_Rq^{\prime}_m=C_Rq_m+\frac{H(\mathbf{x},\mathbf{y})}{n-1}.$$
Accordingly, the adaptive setting of $C_R$ could be
\begin{equation}\label{temp18}
F^{\prime}_R=\left(C_Rq_m+\frac{H(\mathbf{x},\mathbf{y})}{n-1}\right)/q^{\prime}_m,
\end{equation}
where $q'_m$ is updated by (\ref{temp13}).
To demonstrate the promising function of the adaptive update strategy, we incorporate it to $(1+1)EA$ and $(1+1)EA_C$ to get the adaptive variant of these algorithms, and test their performance on the 12-20 dimensional Deceptive problems.
Parameters of EAs are initialized by (\ref{ParaSetting}), and adapted according to (\ref{temp16}), (\ref{temp13}) and (\ref{temp18}), respectively.
Since the adaptive strategy decreases stability of performances to a large extent, numerical simulation of the tail probability is implemented by 10,000 independent runs. To investigate the sensitivity of the adaptive strategy on initial values of $q_m$, the mutation rate $q_m$ in $(1+1)EA_C$ is initialized with values $\frac{1}{\sqrt{n}}$, $\frac{3}{2\sqrt{n}}$ and $\frac{2}{\sqrt{n}}$, where the three variants are denoted by $(1+1)EA_C^1$, $(1+1)EA_C^2$ and $(1+1)EA_C^3$, respectively.
The converging curves of averaged TPs are illustrated in Figure \ref{CompDecAdaptive}. Compared to the EAs with fixed parameters during the evolution process, the performance of adaptive EAs has significantly been improved. Meanwhile, the outperformance of $(1+1)EA_C$ introduced by the binomial crossover is greatly enhanced by the adaptive strategies, too.
\begin{figure*}[ht]
\centering
\includegraphics[width=7in]{ComProDecAdap.eps}
\caption{Numerical comparison on tail probabilities (TPs) of adaptive $(1+1)EA$, $(1+1)EA_C$ and $(1+1)EA_C$ applied to the Deceptive problem, where $n$ is the problem dimension. The $(1+1)EA_C^1$, $(1+1)EA_C^2$ and $(1+1)EA_C^3$ are three variants of $(1+1)EA_C$ with $q_m$ initialized as $\frac{1}{\sqrt n}$, $\frac{3}{2\sqrt n}$ and $\frac{2}{\sqrt n}$, respectively.}\label{CompDecAdaptive}
\end{figure*}
\section{Conclusions}\label{SecCon}
Under the framework of fixed budget analysis, we conduct a pioneering analysis of the influence of binomial crossover on the approximation error of EAs. The performance of an EA after running finite generations is measured by two metrics: the expected value of the approximation error and the error tail probability.
Using the two metrics, we make a case study of comparing the approximation error of $(1+1)EA$ and $(1+1)EA_C$ with binomial crossover.
Starting from the comparison on probability of flipping ``{\it l preferred bits}'', it is proven that under proper conditions, incorporation of binomial crossover leads to dominance of transition probabilities, that is, the probability of transferring to any promising status is improved. Accordingly, the asymptotic performance of $(1+1)EA_C$ is superior to that of $(1+1)EA$.
It is found that the dominance of transition probability guarantees that $(1+1)EA_C$ outperforms $(1+1)EA$ on OneMax in terms of both expected approximation error and tail probability. However, this dominance does leads to the outperformance on Deceptive. This means that using binomial crossover may reduce the approximation error on some problems but not on other problems.
For Deceptive, an adaptive strategy of parameter setting is proposed based on the monotonicity analysis of transition probabilities. Numerical simulations demonstrate that it can significantly improve the exploration ability of $(1+1)EA_C$ and $(1+1)EA$, and superiority of binomial crossover is further strengthened by the adaptive strategy. Thus, a problem-specific adaptive strategy is helpful for improving the performance of EAs.
Our future work will focus on a further study for adaptive setting of crossover rate in population-based EAs on more complex problems, as well as development of adaptive EAs improved by introduction of binomial crossover.
\section*{Acknowledgements}
This research was supported by the Fundamental Research Funds for the Central
Universities (WUT: 2020IB006).
|
\section{Introduction}
Pre-trained distributed representations of words (a.k.a. word embeddings) are ubiquitous tools in natural language processing (NLP).
Their utility is owing to the remarkable success in mapping semantic and syntactic relationships among words to linear relationships among real-valued vectors.
For instance, analogy generation using vector addition on word embeddings (e.g. Tokyo is to Japan as Paris is to France) was taken to be an early measure of word embedding quality.
In all kinds of related tasks, the vector space is known to encode semantic meaning surprisingly well \citep{pennington2014glove, Mik:2013:w2v, Mik:2013:linreg}.
However, harmful gender-biased properties of word embeddings are also known to exist.
Later is was observed that the same analogy generation property that produced the celebrated ``man is to king as woman is to queen" analogy would also predict ``man is to programmer as woman is to homemaker" \citep{Tolga:2016}. This observation sparked interest in developing debiased word embeddings.
Post-processing debiasing schemes are usually motivated by recognizing some intrinsic measure of bias in the embedding space, and then attempting to reduce that intrinsic bias. Early work (2016-2017) focused on the idea of a ``gender direction" vector within the embedding space, loosely defined as the difference vector between female and male attribute words.
It was noted that any non-zero projection of a word onto the gender direction (termed direct bias) implied that the word was more related to one gender over another.
In the case of ideally gender-neutral words (e.g. doctor, nurse, programmer, homemaker), this was viewed as an undesirable property.
The first debiasing methods, Hard Debias \citep{Tolga:2016} and Gender Neutral-GloVe \citep{Zhao2018:GNglove}, worked to minimize or eliminate the direct bias, and were shown to be successful in mitigating harmful analogies generated by word embeddings in relation to gender-stereotyped occupations.
An influential critique paper by \citet{Gonen:2019} demonstrated that minimizing direct bias did not eliminate bias in the vector space entirely. Rather, words that tended to cluster together due to gender bias (e.g. nurse, teacher, secretary, etc.) would still cluster together in the nullspace of the gender direction. Furthermore, the original bias could be recovered by classification techniques using only the debiased word embeddings as input. These observations were termed cluster and recoverability bias.
The next wave of debiasing methods (2019-present) focused on reducing cluster and recoverability bias while proposing new metrics to systematically quantify the indirect bias of the embedding space (e.g. the Gender-based Illicit Proximity Estimate, introduced by \citet{Kumar:2020:RAN}). While these new debiasing schemes do reduce indirect bias in multiple ways, there is a general lack of
connection to downstream applications such as coreference resolution, natural language inference (NLI) and sentiment analysis.
Current gender-bias evaluation tests (GBETs) in widespread use include the WinoBias test set \citep{Zhao:2018:WinoBias}, designed to measure bias in coreference resolution systems using stereotypical occupations as a probe, and the NLI test set \citep{Dev:2020:NLItest}, designed to measure stereotypical inferences again using occupations as the concept of interest.
More commonly used evaluations include the Word Embedding Association Test (WEAT) \citep{Caliskan:2017:weat}, and the analogy generation test SemBias \citep{Zhao2018:GNglove}. However these tests solely evaluate the vector properties of the word embeddings, without any connection to downstream applications.
Adding to the library of downstream GBETs is essential in building a robust understanding of gender bias in NLP applications \citep{Sun:2019:review}.
Here we introduce a new observation of gender-biased predictions in a downstream task, namely “marked attribute bias” in natural language inference, and develop corresponding GBETs. Marked attribute bias refers to the language model’s tendency to predict that “person” implies “man” (the default attribute), while simultaneously understanding that “person” does not necessarily imply “woman” (the marked attribute). Marked attribute bias was found to exist on explicitly defined gender words (e.g. man, woman, etc.), and persist on implicit gender words (e.g. names) as well as latent gender-carriers (e.g. stereotypical occupations).
An analysis of the currently available debiased embeddings reveals that none are able to successfully mitigate marked attribute bias. Furthermore, none of the currently proposed measures of intrinsic bias on the embedding space are predictive of the marked attribute effect. We define a new measure of intrinsic bias that was found to correlate with the marked attribute effect better than any currently available metric.
Using this insight, we introduce a new debiasing scheme: Multi-dimensional Information-weighted Soft Projection. Applying MISP to an existing debiased embedding achieves the lowest observed marked attribute bias error.
\\
\textbf{Summary of main contributions:}
\begin{enumerate}
\item We present a new observation of gender bias in a downstream NLP application: marked attribute bias (MAB). The MAB test sets are made available in order to expand the current set of GBETs.
\item An analysis of current debiasing methods and current intrinsic bias measures finds that none sufficiently mitigate the error, and likewise none sufficiently explain the effect. This observation creates two new open problems.
\item We propose a new measure for quantifying intrinsic bias on the embedding space: Multi-dimensional Information-weighted Direct Bias (MIDB). This measure was found to correlate meaningfully with the marked attribute effect.
\item We introduce a new debiasing scheme: Multi-dimensional Information-weighted Soft Projection. MISP-debiased embeddings obtain new best performance on the MAB test set.
\end{enumerate}
\section{Marked Attribute Bias in Natural Language Inference}
\subsection{Background: Natural Language Inference}
Natural language inference is one of the pillars of natural language understanding.
It is the task of determining whether a hypothesis sentence is (neutral, entailed, or contradicted) with respect to a premise sentence. For example:
\textbf{Premise:} A choir sings in the church.
\textbf{Hypothesis:} The church is filled with the sound of singing. (Correct prediction: Entail)
\citet{Dev:2020:NLItest} previously used NLI as a test case for gender bias with respect to occupations. For example, consider:
\textbf{Premise:} A doctor prepared a meal.
\textbf{Hypothesis 1:} A man prepared a meal. (N)
\textbf{Hypothesis 2:} A woman prepared a meal. (N)
This inference task essentially asks the question: is ``doctor" a subset of man/woman? I.e. if someone is a doctor, must they be a man?
While both hypothesis sentences should receive a neutral prediction (as ``doctor" does not imply any specific gender), hypothesis 1 will more likely receive an entailment, while hypothesis 2 will more likely receive a contradiction, given biased word embeddings. The corresponding GBET was published by \citet{Dev:2020:NLItest} and contains 1936512 sentence pairs in the form [A \textbf{occupation} \textit{verb object}] $\rightarrow$ [A \textbf{gender word} \textit{verb object}]. Throughout this paper, we will use the notation [Sentence A] $\rightarrow$ [Sentence B] to mean that premise Sentence A is paired with hypothesis Sentence B.
\subsection{Observation of marked vs. default attribute bias}
Marked vs.~default attribute bias occurs whenever a default attribute (e.g. male, white, etc.) is assumed, and a marked attribute has to be explicitly stated or becomes a defining trait. In the context of the natural language inference task, consider the sentence pair:
\textbf{Premise:} A person prepared a meal.
\textbf{Hypothesis 1:} He prepared a meal. (N)
\textbf{Hypothesis 2:} She prepared a meal. (N)
Due to the language model's\footnote{All NLI models mentioned throughout this paper are based on the Decomposable Attention Model \citep{parikh:2016:DAM} with intra-attention, trained on the Stanford Natural Language Inference training dataset \citep{bowman:2015:snli} (trained for 100 epochs; learning rate 0.025; weight decay 1e-5; dropout rate 0.2; 200 hidden units; approximately $10^4$ total model parameters). All the code and data needed to reproduce results mentioned in this paper are available at https://github.com/hillary-dawkins/MAB.} tendency to predict that ``person" implies a male (default) attribute, the first hypothesis sentence will have a prediction probability vector shifted towards Entail. However the same language model would tend towards a Neutral prediction for the second hypothesis, recognizing that ``person" does not necessarily imply female (the marked attribute).
To put it another way, this inference task essentially asks the question: is “person” a subset of man/woman?
When presented with a masculine form, the model answers: yes (entailment), a person must be a man. When presented with a feminine form, the model answers: not necessarily (neutral), a female has an attribute (gender) that not all persons have.
The name “Marked Attribute Bias” therefore derives from the observation that masculine forms are unmarked with respect to gender, whereas female forms carry a marked gender attribute.
In this particular example, the model trained with (original) GloVe\footnote{Taken as the GloVe embeddings trained on the Common Crawl corpus for 840B tokens; available at https://nlp.stanford.edu/projects/glove/. Results were not found to vary significantly among undebiased embeddings.} word embeddings \citep{pennington2014glove} gives a probability distribution $(N, E, C)$ of (0.0538, \textbf{0.929}, 0.0177) for hypothesis 1 and (\textbf{0.687}, 0.238, 0.0750) for hypothesis 2.
Note that although the MAB test construction appears similar to \citet{Dev:2020:NLItest}, it is actually measuring quite a distinct effect. The \citep{Dev:2020:NLItest} test set measures associations between gender and some concept of interest (occupations). The MAB test set measures something more general and pervasive; it measures how gender words carry meaning, independent of any concept of interest.
Achieving the correct prediction probability of $(N, E, C) = (1, 0, 0)$ on both sentences is difficult because it requires the language model to be attribute-aware (in this case gender-aware) while not using the gender attribute to alter predictions when it would be inappropriate to do so.
\section{Analysis of the current situation}
In order to investigate the presence of systematic marked attribute bias in natural language inference, we construct three types of tests: bias on explicit gender words, implicit gender carriers, and latent gender carriers. We wish to understand the depth and persistence of the marked attribute effect, as well as how it is handled by current debiasing methods. Firstly we provide a brief description of the current debiasing methods to be analyzed. Next we provide details of the test sets and report results.
\subsection{Debiased embeddings}
Within the scope of this paper, we focus on post-processing techniques applied to static word embeddings. These types of methods are computationally inexpensive, easy to concatenate, and are independent of the base embedding. In addition, we include GN-GloVe, one of the highly cited retraining methods. Notationally, we specify embeddings as (base embedding).method.
Where available, we use published debiased embeddings made available from the original authors of the corresponding method. Otherwise, we apply the method to the base GloVe embeddings.
The methods we will analyze include:
\textbf{Hard Debias\footnote{https://github.com/tolga-b/debiaswe} (GloVe\footnote{\label{gloveZ} The base (undebiased) embeddings are GloVe trained on the 2017 January Wikipedia dump (vocab contains 322,636 tokens). Available at https://github.com/uclanlp/gn\texttt{\_}glove.}.HD)} \citep{Tolga:2016}: The subset of gender-neutral words are projected onto the nullspace of the gender direction $\vec{g}$. Gender-neutral words are made equidistant to pairs of words in a defined equalization set.
\textbf{Gender-Neutral GloVe\footnote{https://github.com/uclanlp/gn\texttt{\_}glove} (GN-GloVe)} \citep{Zhao2018:GNglove}: Similar to hard debias, this method seeks to eliminate the direct bias. The embeddings are retrained from scratch using a modified version of GloVe’s original objective function. The gender information is sequestered to the final component of the word embedding. The gender-neutral portion of the word embedding is then defined as the first $d-1 = 299$ components, denoted \textbf{GN-GloVe$\pmb{(w_a)}$}.
\textbf{Gender-Preserving\footnote{https://github.com/kanekomasahiro/gp\texttt{\_}debias} (GloVe\footnotemark[4].GP)} \citep{Kaneko:2019}: This method seeks to eliminate harmful gender bias while retaining as much useful semantic gender information as possible.
\textbf{Double Hard Debias\footnote{https://github.com/uvavision/Double-Hard-Debias} (GloVe\footnotemark[4].DHD)} \citep{Wang:2020:DHD}: An extended version of the hard debias algorithm, based on the observation that frequency information encoded in the word embeddings convolutes the definition of the gender direction. Correctional pre-processing is applied prior to hard debiasing.
\textbf{Bias Alignment Model\footnote{https://github.com/anlausch/DEBIE} (GloVe\footnotemark[4].BAM)} \citep{Lauscher:2019}: Gender subspace matrices are defined by stacking explicit gender words. The projection that maps the embedding space to itself while approximately aligning the gender subspaces is learned and applied to all words. After alignment, gender information is not retained.
\textbf{Orthogonal Subspace Correction and Rectification\footnote{https://github.com/sunipa/OSCaR-Orthogonal-Subspace-Correction-and-Rectification} (GloVe\footnotemark[4].OSCaR)} \citep{Dev:2020:oscar}:
The rationale is that linear projective methods are too aggressive in modifying the entire embedding space.
OSCaR rectifies two concepts of interest (gender and occupations), such that these subspaces are orthogonal in the debiased space.
\textbf{Iterative Nullspace Linear Projection\footnote{https://github.com/shauli-ravfogel/nullspace\texttt{\_}projection. The projection matrix computed for our base GloVe embeddings is available at https://github.com/hillary-dawkins/MAB.} (GloVe\footnotemark[4].INLP)} \citep{Ravfogel:2020:INLP}: Rather than defining a gender direction, INLP \textit{learns} the most informative decision boundary for classifying gendered and gender-neutral words. All words are projected to the nullspace of the gender subspace, and the process proceeds iteratively until gender information is sufficiently erased.
A closely related method is the D$_4$ algorithm \citep{Davis:2020:D4}.
\textbf{Repulse Attract Neutralize Debias\footnote{https://github.com/TimeTraveller-San/RAN-Debias} (GloVe\footnotemark[4].RAN)} \citep{Kumar:2020:RAN}: Motivated by the persistence of implicit bias after debiasing through projective methods (observed as clustering and recoverability), RAN-debias attempts to address both direct bias and gender-based proximity bias.
\subsection{Explicit gender words test set and error definitions}
Firstly, we construct a test set where every sentence pair is of the form [A person \textit{verb object}] $\rightarrow$ [(A) \textbf{gender word} \textit{verb object}] (the correct inference is always neutral since a person can be of any gender). Verbs ($n = 27$) and objects ($n = 184$) are paired to create $n = 1968$ unique premise sentences\footnote{Verbs and objects are taken from \citep{Dev:2020:NLItest} word lists (https://github.com/sunipa/On-Measuring-and-Mitigating-Biased-Inferences-of-Word-Embeddings) and are paired using the same pairing rules.}. Gender words are taken to be \{man, woman, guy, girl, gentleman, lady, He, She\}, following \citep{Dev:2020:NLItest} with the addition of the pronouns, for a total test set $S$ of $|S| = 15744$ sentence pairs where hypotheses represent binary genders evenly (denoted $S_M$, $S_F$, $|S_M| = |S_F|$).
For every hypothesis sentence in the test set, the ideal prediction probability vector is $(N, E, C) = (1, 0, 0)$.
We could define the error on the test set as the average Euclidean distance from the ideal distribution:
\begin{equation}
\mathcal{E} = \frac{1}{|S|}\sum_{i \in S} \lVert (1,0,0) - (N, E, C)_i\rVert_2.
\end{equation}
This task, test set, and error definition are simple, and yet they encapsulate the central challenge of the debiasing field: to create attribute-aware (required to obtain the Neutral prediction) but attribute-unbiased embeddings.
A weaker, but still potentially desirable, condition might be to minimize the effect of gender while not requiring that the model be gender-aware. Typically, this means that all hypotheses tend towards an Entail prediction, regardless of gender. We could define the error as the average distance between probability vectors between genders:
\begin{equation}
d = \frac{1}{2|S|} \norm{\sum_{i \in S_M}(N,E,C)_i - \sum_{j \in S_F}(N,E,C)_j }_2.
\end{equation}
A gender-agnostic model could achieve zero error by this definition even with an accuracy of zero on the test set.
Table 1 shows the results for this test set on all the embeddings of interest. None of the debiased embeddings successfully mitigate the marked attribute error.
A similar test set shows that the effect persists on implicit gender words (e.g. names). Results are shown in the appendix.
\begin{table*}
\caption{Results of the marked attribute test set on \textbf{explicit gender words}. Due to varying results on gender nouns vs.~pronouns, results are shown separately for each case (M and F represent averages across the gender nouns). Some debiased embeddings are able to eliminate the distance across pronouns (really by definition since $\vec{she} \approx \vec{he}$ in these cases), but none are able to eliminate differences between the gender nouns significantly. Even when differences between genders are minimized, distance from the ideal distribution (error $\mathcal{E}$) remains or increases. This highlights the challenge of creating gender-aware but not gender-biased embeddings.}
\begin{tabularx}{\textwidth}{l | Xrrr | Xrrr | rr}
\hline
Emb.method & Gender & N & E & C & Gender & N & E & C & $d$ & $\mathcal{E}$ \\
\hline
GloVe & M & 0.7832 &0.1966 &0.0202& F &0.9449 &0.0401 &0.0149 & 0.225 & 0.182 \\
& he &0.0982 &0.8838 &0.0180 &she &0.6549 &0.3137 &0.0315 & 0.797 & 0.865 \\
\hline
GloVe.HD & M & 0.8306& 0.1329 &0.0365& F & 0.9269& 0.0499& 0.0232 & 0.128 & 0.155 \\
& he &0.2944 &0.6737 &0.0319 &she &0.5174 &0.4334 &0.0491 & 0.328 & 0.813 \\
\hline
GN-GloVe & M &0.6339 &0.3402 &0.0259 &F &0.9169 &0.0461 &0.0370 & 0.408 & 0.301 \\
&he &0.1767 &0.7968 &0.0265 &she &0.8223 &0.1405 &0.0373 & 0.921 & \textbf{0.688} \\
\hline
GN-GloVe($w_a$) & M & 0.8446& 0.1254& 0.0300& F & 0.9211& 0.0395 &0.0394 & \textbf{0.115} & \textbf{0.149} \\
& he & 0.1430 &0.8266 &0.0304 &she& 0.4237 &0.5367 &0.0396 & 0.404 & 0.990 \\
\hline
GloVe.DHD & M &0.7013 &0.2685 &0.0302 &F &0.9282 &0.0510 &0.0209 & 0.315 & 0.247 \\
& he &0.1566 &0.8187 &0.0247 &she &0.1597 &0.8139& 0.0264 & 0.006 & 1.173 \\
\hline
GloVe.GP &M &0.6172 &0.3521 &0.0306 &F &0.8777 &0.0693 &0.0530 & 0.385 & 0.336 \\
& he &0.2443 &0.7262 &0.0295 &she &0.6481& 0.3040 &0.0480 & 0.585 & 0.758 \\
\hline
GloVe.BAM & M &0.7983 &0.1703 &0.0314 &F &0.9329 &0.0447 &0.0224 & 0.184 & 0.175 \\
& he &0.1625 &0.8083 &0.0292 &she &0.6752 &0.2878 &0.0369 & 0.731 & 0.800 \\
\hline
GloVe.OSCaR & M & 0.8233& 0.1572& 0.0195& F &0.9431 &0.0400 &0.0169 & 0.168 & 0.154 \\
& he & 0.1482 &0.8292 &0.0226 &she &0.8428 &0.1278 &0.0294 & 0.987 & 0.697 \\
\hline
GloVe.RAN & M & 0.8055 &0.1686 &0.0260 &F &0.8994 &0.0701 &0.0305 & 0.136 & 0.193 \\
& he &0.1939 &0.7811 &0.0250 &she &0.5962 &0.3420 &0.0618 & 0.597 & 0.828 \\
\hline
GloVe.INLP & M &0.8298 &0.1537 &0.0166 &F &0.9204 &0.0633 &0.0164 & 0.128 & 0.167 \\
& he &0.1081 &0.8710 &0.0209 &she &0.1119 &0.8672 &0.0209 & \textbf{0.005} & 1.244 \\
\hline
\end{tabularx}
\label{tab:AllEmbs}
\end{table*}
\subsection{Latent gender carriers: Stereotyped occupations}
\begin{table*}
\caption{Results of marked attribute test set on \textbf{stereotypical occupations}. Each (N,E,C) probability vector is averaged over the 1968 unique premise sentences and the gender attribute words from each category (M or F) ($n = 31,488$ sentences for each gender). Smaller distances between the M and F vectors indicate less gender bias. The significance of the difference was evaluated using a permutation test; the alternate distance $d^*$ is computed for 10,000 randomly sampled partitions of the occupations into two groups. The significance value is the proportion of these samples to generate a distance $d^* > d$. This gives us an idea of whether the defined partition, based on gender, is a meaningful grouping. Smaller significance values indicate that the defined partition is non-random with respect to the distance.}
\begin{tabularx}{\textwidth}{l | X | X | rr}
\hline
Emb.method & M attribute (N, E, C) & F attribute (N, E, C) & Distance $d$ & Significance \\
\hline
GloVe& (0.6000, 0.3350, 0.0650) &(0.7378, 0.1711, 0.0910)& 0.216 & 0.0001 \\
GloVe.HD & (0.4975, 0.4500, 0.0525) &(0.6075, 0.3357, 0.0568) &0.159 & 0.0408 \\
GN-GloVe & (0.5026, 0.4434, 0.0540) &(0.7126, 0.2036, 0.0838) &0.320 & 0.0000 \\
GN-GloVe$(w_a)$ & (0.5309, 0.3915, 0.0776)& (0.6197, 0.2771, 0.1032)& 0.147 & 0.0478 \\
GloVe.DHD & (0.5285, 0.4126, 0.0589) &(0.6513, 0.2811, 0.0676) &0.180 & 0.0038 \\
GloVe.GP & (0.5016, 0.4380, 0.0604) &(0.6479, 0.2639, 0.0882) &0.229 & 0.0010 \\
GloVe.BAM & (0.6293, 0.3077, 0.0630) &(0.7116, 0.1972, 0.0912) &0.141 & 0.0060 \\
GloVe.OSCaR & (0.5577, 0.3901, 0.0522)& (0.6789, 0.2400, 0.0812)& 0.195 & 0.0036 \\
GloVe.RAN & (0.5393, 0.3933, 0.0674) &(0.5924, 0.3026, 0.1050) &0.112 & 0.0477 \\
GloVe.INLP & (0.5065, 0.4197, 0.0739) &(0.5465, 0.3949, 0.0587) &0.050 & 0.6595 \\
\hline
\end{tabularx}
\label{tab:Occs}
\end{table*}
Next, we would like to check if the gender-induced marked attribute bias can affect entities which should be gender neutral, but turn out to be hidden carriers of a gender attribute (e.g. stereotypical occupations). The same template [A person \textit{verb object}] $\rightarrow$ [A/An \textbf{occupation} \textit{verb object}] was used with the common vocabulary. Stereotypical occupations ($n=32$) were sourced from \citet{Tolga:2016}, and the SemBias test set. Examples are (doctor, engineer, boss, etc.~vs.~nurse, maid, homemaker, etc.). In total there are 62,976 sentence pairs in the test set\footnote{\label{wordset}The exact word set used to produce these results is available at https://github.com/hillary-dawkins/MAB.}.
Results are shown in Table 2. A permutation test is used to check if dividing the occupations into groups according to gender stereotypes produces a significant difference in the probability vectors (rather than dividing them randomly). As shown, the marked attribute effect persists on stereotypical occupations, especially on original embeddings. This is an important result because it highlights that unintended behaviour can appear in unexpected places due to a latent attribute.
Previously, GBETs have focused on how explicit gender words are treated under biased models. To our knowledge, this is the first GBET designed to analyze unintended behaviour on a latent attribute carrier.
Note that this task is easier to correct than the explicit gender words because occupation words have defining characteristics beyond gender. That is, a debiasing method such as Iterative Nullspace Projection can perform well by removing gender information entirely.
This does not mean that the challenge of having a gender-aware but gender-unbiased embedding is solved, but it does provide evidence that latent gender effects can be mitigated using linear projective methods.
The full extent of latent biased-attribute effects and possible mitigation strategies should be investigated further.
\section{Intrinsic bias measures}
How to define bias on an embedding space remains an active area of study.
In general, we seek to understand how the intrinsic or geometric properties of an embedding space translate to real observable bias in downstream tasks. Intrinsic properties are easy to compute quickly, whereas computing performance on downstream tasks requires us to train new models for every case. Understanding of the correlations between the two gives insight on how word embeddings should be debiased.
As a case study, let us focus on the marked attribute error $\mathcal{E}$ on the explicit gender words (shown in Table 1). Recall that this measure of bias is of interest because zero error corresponds to the gold standard: having an attribute-aware model, while simultaneously not using the gender attribute to make inappropriate inferences.
In this section, we look at 5 existing intrinsic bias measures: Direct Bias, Clustering, Recoverability, Gender-based Illicit Proximity Estimate (GIPE), and SemBias. We will investigate whether any of these measures are predictive of the marked attribute effect.
Recall that direct bias was the first measure to be proposed; it simply measures the average projection of word vectors onto a predefined gender direction.
Early methods (i.e. Hard Debias and GN-GloVe) defined bias in the embedding space completely as direct bias.
The idea of clustering and recoverability refer to a classifier's ability to correctly reassign gender labels to words, even after debiasing methods have been applied.
\citet{Gonen:2019}'s observation of clustering and recoverability sparked new interest in defining metrics for indirect bias on the embedding space.
Although clustering and recoverability do not provide well-defined measures of bias given an embedding space (as they depend training implementation - though they could be said to provide a lower bound), many new debiasing proposals will cite reduced clustering as a positive result. The effect on downstream applications is not well understood as of yet.
The Gender-based Illicit Proximity Estimate (GIPE) measures the extent of undue proximities in the embedding space due to a pervasive gender attribute.
Lastly, the SemBias analogy test set measures whether gender-biased analogies exist within the embedding space based on vector arithmetic properties.
Implementation details for each measure as well as the experimental set of embeddings ($n = 16$) are given in the appendix.
The average Direct Bias on the embedding space was found to have a Pearson correlation coefficient of 0.104 with the marked attribute error. The Clustering $v$-measure\footnote{With cluster size $n = 1500$ (which lead to the highest observed correlation); see appendix.} \citep{rosenberg:vmeas} achieved a correlation coefficient of 0.184. Recoverability was attempted using an SVM with a linear decision boundary, an SVM with a non-linear (radial basis function) kernel, logistic regression, and a simple 1-hidden-layer fully-connected network. All recoverability correlation results were comparable, but the best coefficient of 0.223 was achieved by logistic regression. The GIPE\footnote{Using an indirect bias threshold of $\theta = 0.05$, and number of nearest neighbours $n = 100$.} had a correlation coefficient of 0.432. The SemBias\footnote{The SemBias score was taken as the proportion of analogy examples in the test set for which the embedding space returns the correct definitional analogy.} test set had a correlation coefficient of 0.091.
The full correlation matrix between all intrinsic bias measures can be found in the appendix.
The results suggest that the marked attribute effect is not well correlated with any present notion of intrinsic bias, therefore we do not have a good understanding of how the word embedding properties contribute to this type of observable bias.
In seeking a potential solution, we make note of a new intrinsic bias measure, multi-dimensional information-weighted direct bias (MIDB), found to have a more meaningful correlation of 0.667 with the marked attribute error. We define the MIDB of a particular word $\vec{x}$ to be a weighted average over inner products with basis vectors of a multi-dimensional gender subspace:
\begin{align}
\text{MIDB}_d(x) = \sum_{i=1}^d a_i \langle g_i | x \rangle
\end{align}
where $\{ g_i\}$ form an orthonormal basis for the gender subspace, here defined as the first $d$ principal components summarizing difference vectors $\{ \delta_{jk}\}$. The difference vectors are taken as all pairwise differences\footnote{Using all pairwise differences creates a matrix with rank much less than the dimension of the matrix, however the rank is still much larger than $d$ (the number of principal components to extract) so it doesn't cause a problem.} between vectors in defined gender sets (here common names were used\footnotemark[13]): $\{ \delta_{jk}\} = \vec{f}_j - \vec{m}_k$, $f_j \in F_\text{names}$, $m_k \in M_\text{names}$ ($|M_\text{names}| = |F_\text{names}| = 100$).
The weighting $a_i$ is the proportion of variance explained by the $i^{th}$ principal component,
and $d$ is a hyperparameter controlling the number of dimensions to keep\footnote{On our set of experimental embeddings, $d = 4$ was empirically found to produce the 0.7 correlation result.}.
New proposals for defining a gender direction or subspace potentially have far reaching consequences in the landscape of intrinsic bias measures and their related debiasing schemes. In fact all of Clustering, Recoverability, GIPE, and SemBias use the classic uni-dimensional gender direction $\vec{g}$ within their definitions.
The weak observed correlation between DB and MIDB suggests that these subspaces are independent.
Swapping in a uniquely informative gender subspace to the existing indirect measures would produce a new family of intrinsic bias measures.
The observed utility of names in defining a meaningful gender subspace is encouraging because it opens an obvious avenue for this method to be applied to attributes of interest beyond gender (e.g. race or ethnicity).
\section{Multi-dimensional information-weighted soft projection}
\begin{table*}
\caption{Results for word similarity and analogy benchmarks. Results on the word analogy tasks are reported as percentage accuracy. Results on the word similarity tasks are reported as a Spearman correlation ($\times$ 100). Application of MISP does not alter the overall quality of word embeddings as measured by these classic test sets.}
\begin{tabularx}{\textwidth}{l | rrrr| rrrr}
\hline
Embedding.method & Sem & Syn & Google-Total & MSR & RG & MTurk & MEN & SL999 \\
\hline
GloVe & 80.48 & 62.76 & 70.80 & 51.49 & 75.29 & 64.27 & 72.19 & 34.86 \\
GloVe.MISP & 80.49 & 62.81 & 70.84 & 51.51 & 76.06 & 64.32 & 72.41 & 35.04 \\
GN-GloVe & 77.62 & 61.60 & 68.87 & 49.29 & 74.11 & 66.36 & 74.49 & 37.12 \\
GN-GloVe$(w_a)$ & 77.68 & 61.56 & 68.87 & 49.38 & 75.46 & 66.55 & 74.72 & 37.53 \\
GN-GloVe$(w_a)$.MISP & 77.68 & 61.59 & 68.89 & 49.26 & 75.49 & 66.45 & 74.76 & 37.60 \\
\hline
\end{tabularx}
\label{tab:Vanilla}
\end{table*}
In this section we motivate the above search for an informative intrinsic bias measure. As discussed, a greater understanding of how embedding properties influence observed bias can inform new debiasing techniques. Translating the idea of MIDB into a debiasing scheme yields Multi-dimensional Information-weighted Soft Projection (MISP).
In this debiasing procedure, we project all words into the nullspace of the multi-dimensional gender subspace, proportional to our belief that certain dimensions actually encode the latent idea of gender:
\begin{align}
\vec{w}_{deb} = \vec{w} - \sum_{i = 1}^{d} a_i \langle{g_i}|w\rangle |g_i\rangle
\end{align}
where $\vec{w}$ is the input embedding, $\vec{w}_{deb}$ is the debiased output embedding, and all other quantities are defined as in eq. (3).
As shown in Table 1, the GN-GloVe$(w_a)$ embeddings are currently the top performers on the explicit gender words test set, as measured by either error $\mathcal{E} = 0.149$, or distance $d = 0.115$.
Applying MISP to GN-GloVe$(w_a)$ embeddings (denoted GN-GloVe$(w_a)$.MISP), we achieve an error on the explicit gender words test set of $\mathcal{E} = 0.1107$, a 26\% error reduction over the previous best.
The distance $d$ between genders is reduced to $d = 0.08744$, a 21\% reduction over the previous best.
Successful concatenation suggests that this technique is distinct, and independently useful, from techniques that seek to minimize the traditional direct bias (including GN-GloVe). This observation is consistent with the weak observed correlation between direct bias and MIDB$_4$ on the experimental set of embeddings.
Computing the intrinsic bias measures Clustering, Recoverability, GIPE and SemBias on the newly created embedding space GN-GloVe$(w_a)$.MISP (compared to the base GN-GloVe$(w_a)$), we observe a clustering $v$-score of 0.498 (previously 0.497)\footnote{Where clustering size $n = 1500$.}, a recoverability accuracy of 0.992 (previously 0.993)\footnote{This is the highest accuracy achieved by any of the four classification methods tested; implementation details are in the appendix.}, a GIPE of 0.1169 (previously 0.1173)\footnote{Computed with indirect bias threshold $\theta = 0.03$, and number of nearest neighbours $n = 100$.}, and a SemBias score of 0.938 (previously 0.938)\footnote{Reported as the proportion of samples in the full test set to return the definitional analogy; higher scores are better.}.
The MISP method did not reduce bias by any of these measures, although this is not particularly surprising as it was designed to address the marked attribute effect (through MIDB). It is encouraging however that none of these bias measures were increased. In other words, there is no expected trade-off between the reduced marked attribute error and any previous debiasing work that relied on these measures. The SemBias result informs us that MISP did not reintroduce any harmful biased analogies, for example.
For reference, if we apply the analogous multi-dimensional hard debias method (i.e. equation (4) where all weights $a_i$ are set to 1), the output embeddings GN-Glove$(w_a)$.MHD do not successfully mitigate the marked attribute effect ($\mathcal{E} = 0.1501$, $d = 0.1603$). This suggests that the soft nature of the projection is a key ingredient.
Furthermore, we provide some evidence that specifically the information weighting of the soft projection is a good ingredient as follows. Recall that we are attenuating components of each basis vector according to our belief in that vector as a good gender direction.
The basis vectors are defined to be the first $d$ principal components, weighted by their corresponding variance explained. Therefore the first basis vector receives the greatest weight and so on.
To test the significance of this decision, we define alternative debiased embeddings by applying MISP where the weights get reassigned to the ``wrong" vector (for $d=4$, we have 23 alternative pairings).
We observe that \textbf{none} of the 23 alternatives obtain an error $\mathcal{E}$ less than the ``true" implementation of MISP. This suggests that weighting the components by order of information is a good ingredient. Values of $\mathcal{E}$ for the alternate embeddings can be found in the appendix. Model parameters for each case are made available in order to reproduce this argument on any extended version of the MAB test set.
Information weighting is an interesting idea because it could be applied to either defined or learned gender subspaces alike. For instance, if the basis vectors of a gender subspace are taken as the iteratively learned linear decision boundaries (as in INLP), we could investigate weighting each dimension by the accuracy $acc_i$ of classification on each iteration, as $a_i = (1-2acc_i)$. In this way, dimensions receive weights proportional to their ability to predict gender information.
When accuracy reaches 0.5, no gender information remains, the learned decision boundary is meaningless, and the basis vector receives zero weighting.
Finally, as with any debiasing method, we wish to verify that application of the method has not damaged the overall embedding quality. We assess the MISP embeddings on a handful of classic analogy and word semantic similarity benchmarks. The word similarity benchmarks measure how closely the word embeddings capture similarity between words compared to human annotation. We use the following datasets: RG \citep{RG}, MTurk \citep{MTurk}, MEN \citep{MEN}, and SimLex999 \citep{simlex}. The word analogy task measures how well the word embeddings capture semantic and syntactic relationships among words as vector properties. We report on the Google \citep{Mik:2013:goog}, and MSR \citep{Mik:2013:linreg} test sets. Results were obtained following the word embedding benchmark package\footnote{https://github.com/kudkudak/word-embeddings-benchmarks} \citep{Jastrzebski:2017}. As shown, application of MISP does not alter the overall word embedding quality.
\section{Conclusion}
This paper highlights a new observation of gender bias in a downstream setting: marked attribute bias in natural language inference.
The current inference is that ``person" implies male, while ``person" does not imply female.
Consequently, this inference is being baked into our models of natural language understanding.
The effect was shown to persist on explicitly defined gender words and on latent gender-attribute carriers.
Based on an assessment of the current debiasing landscape, none of the current debiasing methods satisfactorily mitigate the marked attribute error, and furthermore none of the intrinsic bias measures are useful at predicting the marked attribute effect.
By noticing a more meaningful correlation with a newly identified intrinsic bias measure, we propose a new debiasing scheme: multi-dimensional information-weighted soft projection (MISP).
This method introduces several concepts, including the use of a multi-dimensional defined gender subspace.
Previously, the concept of a defined gender subspace always appeared as a single dimension. The iterative nullspace projection method implicitly uses higher learned dimensions, however this requires learning a new decision boundary at every iteration, subject to the implementation of a training procedure. Furthermore, the learned dimensions were not used to define any bias metric, they were strictly used operationally for the debiasing procedure. MISP also introduces the idea of a soft or partial projection, where weights are informed by some measure of the dimension’s ability to capture the intended latent concept of a gender direction. Both of these ideas could be further explored and extended to create new notions of indirect bias, which in turn could inform more sophisticated debiasing procedures.
Multi-dimensional information-weighted soft projection applied to GN-GloVe$(w_a)$ produces new debiased embeddings that achieve the lowest error on the marked attribute bias test set, a 26\% reduction over the previous best, and a 45\% reduction over the original undebiased embeddings. Error reduction on this test set is thought to encapsulate the overall goal of producing gender-aware but gender-unbiased embeddings. Therefore, this method and its composite ingredients warrant further investigation.
Each of the marked attribute bias test sets are made available for further exploration and iteration on these ideas.
\section*{Acknowledgements}
We thank Daniel Gillis, Judi McCuaig, Stefan Kremer, and Graham Taylor for their insightful comments and discussions. We thank the anonymous reviewers for their comments and suggestions which improved the final version of this manuscript. This work is financially supported by the Government of Canada through the NSERC CGS-D program (CGSD3-518897-2018).
\bibliographystyle{acl_natbib}
|
\section*{INTRODUCTION}
When treatment effect modifiers have a different distribution among participants in a randomized trial compared to the target population of substantive interest, the average treatment effect estimated in the trial is not directly applicable to the target population. Methods for extending -- generalizing or transporting \cite{dahabreh2019commentaryonweiss} -- causal inferences from a trial to a target population of interest \cite{cole2010, westreich2017, dahabreh2018generalizing, dahabreh2020transportingStatMed, rudolph2017} rely on specifying models for either the probability of participation in the trial, the expectation of the outcome among trial participants, or both, to adjust for a set of covariates (typically measured at baseline) that is adequate to ensure conditional exchangeability between the randomized and non-randomized groups. Most work on extending causal inferences from trials has focused on methods for estimating average treatment effects in the entire target population or its non-randomized subset. Thus, analyses extending inferences from trials are characterized by the tension between the reason the methods are necessary (heterogeneous treatment effects over multiple covariates that have a different distribution between the randomized and non-randomized groups) and the output the methods produce (estimates of population-averaged treatment effects).
Here, we propose methods for estimating subgroup-specific effects that can be used to alleviate this tension, when one or a few key discrete effect modifiers of substantive interest can be specified. Subgroup-specific average treatment effects in the trial may not generalize to the target population when there is a difference in the distribution of effect modifiers between the randomized and non-randomized groups, after conditioning on the subgroup variables \cite{seamans2021generalizability}. For such cases, the methods we describe can account for multiple variables that have a different distribution between randomized and non-randomized groups, while producing inferences appropriate for subgroups defined in terms of the smaller set of key effect modifiers.
Recently, Mehrotra et al. \cite{mehrotra2021transporting} described a weighting method for estimating subgroup-specific average treatment effects in the non-randomized subset of the target population. Their method can estimate subgroup-specific treatment effects only in the non-randomized subset of the target population. Furthermore, it relies on a model for the probability of trial participation but cannot be used for inference when estimating that model using data-adaptive methods (e.g., machine learning) \cite{chernozhukov2018double}. We build upon their work \cite{mehrotra2021transporting}, previous work on heterogeneity of treatment effects in observational studies over discrete subgroups \cite{eeren2015, xie_estimating_2012, robertson2020assessing}, and previous work on extending inferences from trials \cite{dahabreh2018generalizing, dahabreh2020transportingStatMed}, to describe estimators for subgroup-specific potential outcome means and average treatment effects, for the entire target population and its non-randomized subset. The estimators for the non-randomized subset of the target population can be used when the trial is nested in a sample of the target population (i.e., nested trial designs) or when the trial is combined with a separately obtained sample of non-randomized individuals (i.e., non-nested trial designs) \cite{dahabreh2021studydesigns}. We describe estimators appropriate for both study designs, including g-formula estimators that rely on models of the outcome among trial participants, weighting estimators that rely on models for the probability of trial participation, and augmented weighting estimators that combine both types of models. Our augmented weighting estimators are model doubly robust, in the sense that they are consistent when either the model for participation or the model for the outcome is correctly specified, but not necessarily both. Furthermore, they allow for the use of data-adaptive estimation methods. We illustrate the methods using data from the Coronary Artery Surgery Study (CASS) \cite{olschewski1985} to compare the effect of surgery plus medical therapy versus medical therapy alone for chronic coronary artery disease in subgroups defined by history of myocardial infarction.
\section*{STUDY DESIGN, DATA, AND ESTIMANDS}
\paragraph{Study design:} Both nested or non-nested trial designs can be used to extend inferences from a trial to a target population \cite{dahabreh2021studydesigns}. In a nested trial design, the trial is embedded in a sample of the target population, by prospectively recruiting trial participants within a cohort (e.g., in comprehensive cohort studies \cite{olschewski1985,olschewski1992, schmoor1996}) or by data linkage methods (e.g., using similar approaches as when linking a trial to administrative data to extend follow-up duration \cite{fitzpatrick2018assessment}). In a non-nested trial design, a sample from the non-randomized subset of the target population is obtained separately from the trial sample.
In the nested trial design, it is possible to identify and estimate subgroup-specific causal effects in the entire target population and its non-randomized subset; in the non-nested trial design it is only possible to identify and estimate subgroup-specific causal effects in the non-randomized subset of the target population \cite{dahabreh2021studydesigns}. In a nested trial design, investigators may be interested both in estimating subgroups-specific effects in the entire target population or its non-randomized subset. Which estimand should be preferred in a nested trial design depends on the investigators' goals. For example, to inform clinical or policy decision-making, the effect in the entire target population is typically the most relevant estimand. In contrast, to compare treatment effects among the randomized and non-randomized groups, the effect in the non-randomized subset of the target population allows for a cleaner comparison against the effect in the trial.
In the remainder of the paper we work under the nested trial design because it is the design used in CASS. Nevertheless, our results for identifying and estimating subgroup-specific causal effects in the non-randomized subset of the target population can be extended to non-nested trial designs (see Web Appendix 1 for details) \cite{dahabreh2020transportingStatMed}.
\paragraph{Data:} Let $Y$ be an outcome measured at the end of the study (binary, continuous, or count); $A$ the assigned treatment strategy that takes values in a finite set $\mathcal A$; and $X$ a vector of covariates that takes values in $\mathcal{X}$. These covariates are typically measured before the decision to participate in the trial and before treatment assignment \cite{dahabreh2019identification} (or at least are not affected by trial participation and treatment assignment); we will refer to them as ``baseline covariates.'' Let $v$ be a proper subset of $\mathcal X$, $v \subset \mathcal X$. Then, $v$ defines a population subgroup consisting of individuals with $X \in v$. For example, we can define a collection of $k$ mutually exclusive and exhaustive non-empty subsets of $\mathcal{X}$, say $\{ v_1, \ldots, v_k \}$, that stratify the population into $k$ subgroups \cite{yang2021causal}.
We are interested in examining heterogeneity over subgroups defined in terms of only a few discrete (or discretized continuous) covariates, when conditioning on a large set of covariates is necessary to ensure exchangeability between randomized and non-randomized groups (in the sense formalized in the next section). For instance, in our CASS analysis, we define subgroups based on the presence (vs. absence) of a myocardial infarction in an individual's medical history.
The data collected in a nested trial design are realizations of the random tuple $(X_i, S_i, S_i A_i, S_i Y_i)$, with $i=1, \ldots, n$, where $n$ is the total sample size. In Web Appendix 1, we provide additional details regarding the data structure for non-nested trial designs.
\paragraph{Causal estimands:} Let $Y^{a}$ be the potential outcome under intervention to set treatment to $a$, for $a \in \mathcal A$ \cite{rubin1974, robins2000d}. When comparing two treatments, $a$ and $a^\prime$, both in $\mathcal A$, the subgroup-specific average treatment effect for subgroup $ X \in v$ in the target population is $\E[Y^{a} - Y^{a^\prime}| X \in v] = \E[Y^{a}| X \in v] - \E[Y^{a^\prime}| X \in v]$, and the subgroup-specific average treatment effect for subgroup $ X \in v$ in the non-randomized subset of the target population is $\E[Y^{a} - Y^{a^\prime}| X \in v, S = 0] = \E[Y^{a}| X \in v, S = 0] - \E[Y^{a^\prime}| X \in v, S = 0]$. These types of estimands are sometimes referred to as group-average treatment effects \cite{lechner2018modified}. Typically, in addition to these subgroup-specific average treatment effects, we are also interested in their component potential outcome means, $\E[Y^{a}| X \in v]$ and $\E[Y^{a}| X \in v, S=0]$, for each $a \in \mathcal A$.
\section*{IDENTIFICATION}
\subsection*{Identification of potential outcome means in the entire target population}
\paragraph{Identifiability conditions:}
We now list sufficient conditions under which the subgroup-specific potential outcome means and subgroup-specific average treatment effect in the target population are identifiable. These same conditions are also sufficient for identifying the overall potential outcome means and average treatment effects in the target population \cite{dahabreh2019generalizing, dahabreh2019identification}.
\noindent
\emph{(1) Consistency of potential outcomes:} for each unit $i$ in the target population and each $a \in \mathcal A$, if $A_i = a$, then $Y_i = Y^{a}_i$.
\noindent
\emph{(2) Mean exchangeability in the trial over $A$:} for every $a \in \mathcal A$ and every $x$ with positive density in the trial $f(x, S = 1 ) \neq 0$, $\E [ Y^{a} | X = x , S = 1, A =a] = \E [ Y^{a} | X = x, S = 1]$.
\noindent
\emph{(3) Positivity of treatment assignment:} $\Pr[A=a | X = x, S=1] > 0$, for each $a \in \mathcal A$ and each $x$ with positive density in the trial $f(x , S = 1) \neq 0$.
\noindent
\emph{(4) Mean exchangeability over $S$:} $\E [ Y^a | X = x , S = 1] = \E [ Y^a | X = x ] $, for every $x$ with positive density in the target population $f(x) \neq 0$ and for each $a \in \mathcal A$.
\noindent
\emph{(5) Positivity of trial participation:} $\Pr[S=1 | X = x] >0,$ for every $x$ such that $f(x) \neq 0.$
Note that if interest focuses on a single subgroup, then the positivity and exchangeability assumptions can be weakened to apply only to covariate values that can occur in the specific subgroup. In most applications, however, we are interested in some collection of subgroups that are mutually exclusive and exhaustive of $\mathcal X$, not just a single subgroup.
In addition, to focus on estimation issues, we consider the above identifiability conditions without reference to a fully elaborated structural model. Nevertheless, the conditions we invoke can be derived from structural equation models for the data generating mechanism, such as causal directed acyclic graphs and single-world intervention graphs \cite{richardson2013single, dahabreh2019identification, dahabreh2020benchmarking} or selection diagrams \cite{pearl2014}.
\paragraph{Reasoning about the identifiability conditions:} In applications, the identifiability conditions need to be examined in light of background knowledge and possibly subjected to sensitivity analyses \cite{dahabreh2019commentaryonweiss, dahabreh2019sensitivity}. Condition 1 requires the absence of ``hidden'' versions of treatment (or at least treatment variation irrelevance among versions) \cite{rubin1986, rubin2010reflections, vanderWeele2009}, trial engagement effects \cite{dahabreh2019identification, dahabreh2019generalizing}, and interference among individuals \cite{rubin1986, halloran1995causal}. This condition is largely untestable and judgements about its plausibility should be informed by substantive knowledge about the trial, the sampling of the target population, and the treatments of interest. Conditions 2 and 3 are likely to hold by design in marginally or conditionally randomized trials \cite{hernan2020causal}. Condition 4 is also untestable and judgements about its plausibility should be informed by substantive knowledge and sensitivity analyses \cite{dahabreh2019sensitivity}. Condition 5 is in principle testable, but can be challenging to assess when $X$ is high-dimensional \cite{petersen2012diagnosing}.
Furthermore, in applications, other complications, such as missing data, loss to follow-up, or non-adherence, may require additional conditions for identification. To focus on issues related to extending inferences from trials, we work in the setting with complete data, no losses to follow-up, and complete adherence. Nevertheless, the methods we describe can be naturally extended to address these complications \cite{dahabreh2019identification}.
\paragraph{Identification:} As we show in Web Appendix 2, under conditions 1 through 5, the subgroup-specific potential outcome mean in the target population with $X \in v$, $\E[Y^{a}| X \in v]$, for each $a \in \mathcal A$, is identified by
\begin{equation}\label{id_g_formula}
\begin{split}
\psi(v, a) \equiv \E\big[\E[Y | X, S = 1, A = a] \big| X \in v \big].
\end{split}
\end{equation}
Under the positivity conditions, $\psi(v, a)$ has an algebraically equivalent inverse probability weighting \cite{robins2000a, dahabreh2019relation} re-expression,
\begin{equation} \label{id_IP_weighting1}
\psi(v, a) = \dfrac{1}{\Pr[X \in v]} \E\left[ \dfrac{ I(S = 1, A = a, X \in v)Y}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]} \right],
\end{equation}
where $I(\cdot)$ denotes the indicator function, and the terms $\dfrac{ I(S = 1, A = a)}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]}$ inside the expectation can be viewed as population inverse probability of participation and treatment weights. Under positivity, we can replace the denominator in the first term of equation \eqref{id_IP_weighting1} with the expectation of these weights, because
\begin{equation*}
\Pr[X \in v] = \E\left[\dfrac{I(S = 1, A = a, X \in v)}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]}\right].
\end{equation*}
\subsection*{Identification of potential outcome means in the non-randomized subset}
\paragraph{Identifiability conditions:} To identify the potential outcome means in the non-randomized subset, we retain conditions 1 through 3 as listed above, but substitute the following for conditions 4 and 5 \cite{dahabreh2020transportingStatMed, dahabreh2019identification}:
\noindent
\emph{(4\textsuperscript{*}) Mean exchangeability over $S$:} $\E [ Y^a | X = x , S = 1] = \E [ Y^a| X = x, S=0 ] $, for every $x$ with positive density in the non-randomized subset of the target population $f(x, S = 0) \neq 0$ and for each $a \in \mathcal A$.
\noindent
\emph{(5\textsuperscript{*}) Positivity of trial participation:} $\Pr[S=1 | X = x] >0,$ for every $x$ such that $f(x, S=0) \neq 0.$
Conditions 1 through 3, along with 4\textsuperscript{*} and 5\textsuperscript{*}, are sufficient for identifying the subgroup-specific potential outcome means and average treatment effects in the non-randomized subset of the target population. The same conditions are also sufficient for identifying the overall potential outcome means and average treatment effects in the non-randomized subset of the target population \cite{dahabreh2020transportingStatMed, dahabreh2019identification}. Reasoning about conditions 4\textsuperscript{*} and 5\textsuperscript{*} is similar to that for conditions 4 and 5.
\paragraph{Identification:} As we show in Web Appendix 2, under conditions 1 through 3, 4\textsuperscript{*}, and 5\textsuperscript{*}, the subgroup-specific potential outcome mean in the non-randomized subset of the target population with $X \in v$, $\E[Y^{a}| X \in v, S = 0]$, for each $a \in \mathcal A$, is identified by
\begin{equation}\label{id_g_formula_S0}
\begin{split}
\phi(v, a) \equiv \E\big[\E[Y | X, S = 1, A = a] \big| X \in v , S = 0 \big].
\end{split}
\end{equation}
Under the positivity conditions, $\phi(v, a)$ has an algebraically equivalent inverse odds weighting \cite{robins2000a, dahabreh2019relation} re-expression,
\begin{equation} \label{id_IP_weighting1_S0}
\phi(v, a) = \dfrac{1}{\Pr[X \in v, S=0]} \E\left[ \dfrac{ I(S = 1, A = a, X \in v)\Pr[S=0|X]Y}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]} \right].
\end{equation}
Here, the terms $\dfrac{ I(S = 1, A = a)\Pr[S=0|X]}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]}$ inside the expectation can be viewed as population inverse odds of participation and inverse probability of treatment weights. Under positivity, we can replace the denominator in the first term of equation \eqref{id_IP_weighting1_S0}, with the expectation of these weights, because
\begin{equation*} \label{id_IP_weighting2_S0}
\Pr[X \in v, S=0] = \E\left[\dfrac{I(S = 1, A = a, X \in v)\Pr[S=0|X]}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]}\right].
\end{equation*}
\subsection*{Identification of treatment effects}
The previous results identify the subgroup-specific potential outcomes means in the target population or its non-randomized subset. It follows that we can identify subgroup-specific average treatment effects by taking differences and that we can identify measures of heterogeneity as contrasts of these subgroup-specific effects. For example, comparing treatments $a$ and $a^\prime$, under conditions 1 through 5, the subgroup-specific average treatment effect in the target population is identified by $\psi(v, a) - \psi(v,a^\prime)$. Similarly, under conditions 1 through 3, 4\textsuperscript{*}, and 5\textsuperscript{*}, the subgroup-specific average treatment effect in the non-randomized subset of the target population is identified by $\phi(v, a) - \phi(v,a^\prime)$.
\section*{ESTIMATION \& INFERENCE}
\subsection*{Estimation of potential outcome means in the entire target population}
\paragraph{Outcome modeling and standardization (g-formula):} The identification result in equation \eqref{id_g_formula} suggests the following outcome model-based estimator of $\psi(v,a)$,
\begin{equation} \label{g_formula_estimator}
\widehat {\psi}_{\text{\tiny{OM}}}(v,a) = \left\{ \sum_{i=1}^{n} I(X_i \in v) \right\}^{-1} \sum\limits_{i = 1}^{n} I(X_i \in v) \widehat g_{a}(X_i),
\end{equation}
where $\widehat g_{a}(X)$ is an estimator for $\E [Y | X, S=1, A = a]$. Typically, $\widehat g_{a}(X)$ is obtained using a parametric model for the conditional expectation of the outcome given treatment $A = a$ and the baseline covariates. If the parametric model is correctly specified, $\widehat g_{a}(X)$ is a consistent estimator for $\E [Y | X, S=1, A = a]$ and $\widehat {\psi}_{\text{\tiny{OM}}}(v,a)$ is a consistent estimator for $\psi(v,a)$ \cite{hernan2020causal, dahabreh2019relation}.
\paragraph{Inverse probability weighting:} The identification result in equation \eqref{id_IP_weighting1} suggests an inverse probability weighted estimator:
\begin{equation} \label{IPW_estimator1}
\widehat {\psi}_{\text{\tiny{IPW1}}}(v,a) = \left\{ \sum_{i=1}^{n} I(X_i \in v) \right\}^{-1} \sum\limits_{i = 1}^{n} \widehat w_{v,a}(X_i, S_i, A_i) Y_i,
\end{equation}
where the weight for the $i$th observation is defined as $$\widehat w_{v,a}(X_i, S_i, A_i) = \dfrac{I(S_i = 1, A_i = a,X_i \in v)}{\widehat p(X_i)\widehat e_a(X_i)},$$ $\widehat p(X)$ is an estimator for $\Pr[S=1|X],$ and $\widehat e_a(X)$ is an estimator for $\Pr[ A = a | X , S=1]$.
We can obtain another inverse probability weighting estimator by normalizing the weights \cite{hajek1971comment}:
\begin{equation} \label{IPW_estimator2}
\widehat {\psi}_{\text{\tiny{IPW2}}}(v,a) = \left\{ \sum_{i=1}^{n} \widehat w_{v,a}(X_i, S_i, A_i) \right\}^{-1} \sum\limits_{i = 1}^{n} \widehat w_{v,a}(X_i, S_i, A_i) Y_i.
\end{equation}
Typically, $\widehat p(X)$ is obtained using a parametric model for the probability of participation, conditional on baseline covariates. Similarly, $\widehat e_{a}(X)$ is often obtained using a parametric model for the probability of treatment. Typically, the probability of treatment is known, so $\widehat e_{a}(X)$ cannot be misspecified, and the known probability can be used in place of $\widehat e_{a}(X)$; but estimating $\widehat e_{a}(X)$ may correct for slight imbalances in the trial and improve efficiency \cite{lunceford2004, williamson2014variance}. Thus, in most trials, if the participation model is correctly specified, $\widehat p(X)$ is a consistent estimator for $\Pr[S=1|X]$, and both $\widehat {\psi}_{\text{\tiny{IPW1}}}(v,a)$ and $\widehat {\psi}_{\text{\tiny{IPW2}}}(v,a)$ are consistent estimators for $\psi(v,a)$ \cite{hernan2020causal, dahabreh2019relation}. In Web Appendix 3, we describe how an estimator equivalent to $\widehat {\psi}_{\text{\tiny{IPW2}}}(v,a)$ can be obtained by fitting a weighted saturated regression model for the outcome conditional on treatment and subgroup indicators.
\paragraph{Augmented inverse probability weighting:} The estimators above require models for either the expectation of the outcome or the probability of participation; we can obtain an estimator that uses both models:
\begin{align} \label{DR_estimator1}
\begin{split}
\widehat {\psi}_{\text{\tiny{AIPW1}}}(v,a) &\stackrel{\text{}}= \left\{ \sum_{i=1}^{n} I(X_i \in v) \right\}^{-1} \sum_{i=1}^{n} \Big\{ \widehat w_{v,a}(X_i, S_i, A_i) \left\{Y_i- \widehat g_{a}(X_i)\right\} + I(X_i \in v) \widehat g_{a}(X_i) \Big\},
\end{split}
\end{align}
where $\widehat w_{v,a}(X, S, A)$ and $\widehat g_a(X)$ are as defined above. We can also obtain a second augmented inverse probability weighting estimator by normalizing the weights:
\begin{align} \label{DR_estimator2}
\begin{split}
\widehat {\psi}_{\text{\tiny{AIPW2}}}(v,a) &\stackrel{\text{}}= \left\{ \sum_{i=1}^{n} \widehat w_{v,a}(X_i, S_i, A_i) \right\}^{-1} \sum_{i=1}^{n} \widehat w_{v,a}(X_i, S_i, A_i) \left\{Y_i- \widehat g_{a}(X_i)\right\} \\
&\quad\quad\quad + \left\{ \sum_{i=1}^{n} I(X_i \in v) \right\}^{-1} \sum_{i=1}^{n} I(X_i \in v) \widehat g_{a}(X_i).
\end{split}
\end{align}
When either $\widehat p(X)$ or $\widehat g_a(X)$ is obtained from a correctly specified model, $\widehat {\psi}_{\text{\tiny{AIPW1}}}$ and $\widehat {\psi}_{\text{\tiny{AIPW2}}}$ are consistent estimators for $\psi(v,a)$. Thus, $\widehat {\psi}_{\text{\tiny{AIPW1}}}$ and $\widehat {\psi}_{\text{\tiny{AIPW2}}}$ are model doubly robust, in the sense that they are consistent when either the model for the expectation of the outcome or the model for the probability of participation is correctly specified, but not necessarily both \cite{bang2005}. They are also rate doubly robust \cite{smucler2019unifying}: if data-adaptive methods are used to estimate $\widehat p(X)$ or $\widehat g_a(X)$ with a fast-enough rate \cite{chernozhukov2018double} (but slower than parametric), the estimators still converge at $\sqrt{n}$-rate. For completeness, in Web Appendix 4 we describe another estimator that relies on fitting a weighted multivariable regression model for the outcome, followed by standardization over the distribution of covariates in the target population \cite{robins2007, wooldridge2007}.
\subsection*{Estimation of potential outcome means in the non-randomized subset}
\paragraph{Outcome modeling and standardization (g-formula):} The identification result in equation \eqref{id_g_formula_S0} suggests the following outcome model-based estimator for $\phi(v, a)$:
\begin{equation} \label{g_formula_estimator_S0}
\widehat {\phi}_{\text{\tiny{OM}}}(v,a) = \left\{ \sum_{i=1}^{n} I(X_i \in v,S_i=0) \right\}^{-1} \sum\limits_{i = 1}^{n} I(X_i \in v,S_i=0) \widehat g_{a}(X_i).
\end{equation}
If the parametric model for the expectation of the outcome is correctly specified, $\widehat g_{a}(X)$ is a consistent estimator for $\E [Y | X, S=1, A = a]$ and $\widehat {\phi}_{\text{\tiny{OM}}}(v,a)$ is a consistent estimator for $\phi(v,a)$ \cite{hernan2020causal, dahabreh2019relation}.
\paragraph{Inverse odds weighting:} The identification result in \eqref{id_IP_weighting1_S0} suggests an inverse odds weighting estimator:
\begin{equation} \label{IPW_estimator1_S0}
\widehat {\phi}_{\text{\tiny{IOW1}}}(v,a) = \left\{ \sum_{i=1}^{n} I(X_i \in v,S_i=0) \right\}^{-1} \sum\limits_{i = 1}^{n} \widehat o_{v,a}(X_i, S_i,A_i) Y_i,
\end{equation}
where the inverse odds weight for the $i$th observation is defined as $$\widehat o_{v,a}(X_i,S_i,A_i) =\dfrac{I(S_i = 1, A_i = a,X_i \in v) \big\{ 1-\widehat p(X_i) \big\}}{\widehat p(X_i)\widehat e_a(X_i)}.$$ Of note, we refer to these weights as ``inverse odds weights'' because their defining characteristic is the term for the inverse of the odds of trial participation, $\{ 1-\widehat p(X_i) \}/\widehat p(X_i)$, even though they also contain a component for the inverse of the estimated probability of receiving the treatment actually received (in the trial), $1/\widehat e_a(X_i)$.
We can obtain another inverse odds weighting estimator by normalizing the weights \cite{hajek1971comment}:
\begin{equation} \label{IPW_estimator2_S0}
\widehat {\phi}_{\text{\tiny{IOW2}}}(v,a) = \left\{ \sum_{i=1}^{n} \widehat o_{v,a}(X_i, S_i,A_i) \right\}^{-1} \sum\limits_{i = 1}^{n} \widehat o_{v,a}(X_i, S_i,A_i) Y_i.
\end{equation}
If the parametric model for the probability of participation is correctly specified, $\widehat p(X)$ is a consistent estimator for $\Pr[S=1|X]$ and both $\widehat {\phi}_{\text{\tiny{IOW1}}}(v,a)$ and $\widehat {\phi}_{\text{\tiny{IOW2}}}(v,a)$ are consistent estimators for $\phi(v,a)$ \cite{hernan2020causal}. In Web Appendix 3 we describe how an estimator equivalent to $\widehat {\phi}_{\text{\tiny{IOW2}}}(v,a)$ can be obtained by fitting an inverse odds weighted saturated regression model for the outcome conditional on treatment and subgroup indicators; this estimator relates but is not identical to the one described in reference \cite{mehrotra2021transporting} (see Web Appendix 3 for details).
\paragraph{Augmented inverse odds weighting:} One augmented inverse odds weighting estimator of the subgroup-specific potential outcome mean is:
\begin{align} \label{DR_estimator1_S0}
\begin{split}
\widehat {\phi}_{\text{\tiny{AIOW1}}}(v,a) &\stackrel{\text{}}= \left\{ \sum_{i=1}^{n} I(X_i \in v, S_i=0) \right\}^{-1} \sum_{i=1}^{n} \Big\{ \widehat o_{v,a}(X_i, S_i,A_i) \big\{ Y_i- \widehat g_{a}(X_i) \big\} \\
&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad + I(X_i \in v,S_i=0) \widehat g_{a}(X_i) \Big\}.
\end{split}
\end{align}
We can obtain a second augmented inverse odds weighting estimator by normalizing the weights:
\begin{align} \label{DR_estimator2_S0}
\begin{split}
\widehat {\phi}_{\text{\tiny{AIOW2}}}(v,a) &\stackrel{\text{}}= \left\{ \sum_{i=1}^{n} \widehat o_{v,a}(X_i, S_i,A_i) \right\}^{-1} \sum_{i=1}^{n} \widehat o_{v,a}(X_i, S_i,A_i) \left\{Y_i- \widehat g_{a}(X_i)\right\} \\
&\quad\quad\quad + \left\{ \sum_{i=1}^{n} I(X_i \in v,S_i=0) \right\}^{-1} \sum_{i=1}^{n} I(X_i \in v,S_i=0) \widehat g_{a}(X_i).
\end{split}
\end{align}
These augmented inverse odds weighting estimators, like the augmented inverse probability weighting estimators discussed above, are also model and rate doubly robust (for related results see \cite{dahabreh2019generalizing}). Last, in Web Appendix 4, we describe another estimator that relies on fitting an inverse odds weighted multivariable regression model for the outcome, followed by standardization over the distribution of baseline covariates in the non-randomized subset of the target population.
\subsection*{Estimation of subgroup-specific treatment effects}
Subgroup-specific average treatment effects can be estimated by taking differences between pairs of the potential outcome mean estimators described above. For example, the subgroup-specific average treatment effect in the entire target population, comparing treatments $a$ and $a^\prime$, using the augmented inverse probability weighting estimator in equation \eqref{DR_estimator1}, can be estimated as $\widehat \psi_{\text{\tiny AIPW1}}(v, a) - \widehat \psi_{\text{\tiny AIPW1}}(v,a^\prime)$. Analogous treatment effect estimators can be obtained using each of the potential outcome mean estimators in equations \eqref{g_formula_estimator} through \eqref{DR_estimator2}, for the entire target population, or those in equations \eqref{g_formula_estimator_S0} through \eqref{DR_estimator2_S0}, for the non-randomized subset of the target population.
\subsection*{Modeling to estimate subgroup-specific treatment effects}
\paragraph{Borrowing strength across subgroups:} It is important to note that the methods we describe do not require fitting separate models for the expectation of the outcome or the probability of trial participation in each subgroup. The methods do require correct specification of these models, which may require specifying product terms (statistical interactions) between the subgroup defining variables and other covariates, but not necessarily separate modeling in each subgroup. For example, consider fitting a parametric regression model for the probability of trial participation conditional on covariates, separately in each subgroup. Fitting this model separately in each subgroup is practically equivalent to using all possible subgroup-covariate product terms in the regression. But if some of the coefficients of these product terms are zero (or very close to zero), then we can borrow strength across subgroups by fitting one model for participation that omits the unnecessary product terms. This is useful when, as is commonly the case in applications, some subgroups in the trial are small, making the approach of fitting each model separately by subgroup infeasible \cite{senn2008statistical}.
Thus, our identification and estimation results show that splitting the data by the subgroup-defining variables is not needed for subgroup-specific generalizability or transportability analyses. Although splitting the trial and target population sample by the subgroup-defining variables and performing the analysis in each subgroup separately is a natural approach (previously proposed estimators for extending inferences \cite{cole2010, westreich2017, dahabreh2018generalizing, dahabreh2020transportingStatMed} can be used with no modifications to estimate subgroup-specific effects) such a strategy can be infeasible when the subgroup is a small part of the trial or the target population and will often be inefficient.
\paragraph*{Data-adaptive estimation:} Parametric models are the most common approach for estimating the probability of trial participation and the expectation of the outcome among trial participants. Because the probability of participation and the expectation of the outcome are unknown and must be estimated, and parametric models are likely misspecified, it may be desirable to use data-adaptive modeling approaches (e.g., machine learning techniques). When using such data-adaptive approaches, the augmented weighting estimators presented above can still support valid inference because they can accommodate rates of convergence slower than the parametric rate, when estimating the probability of trial participation and the expectation of the outcome among trial participants \cite{chernozhukov2018double}. For many data-adaptive approaches it may also be necessary to modify our estimators to use cross-fitting strategies, as described in prior work \cite{chernozhukov2018double, chernozhukov2017double}. Of note, estimating the probability of treatment among trial participants is typically not necessary because the known-by-design randomization probability can be used, but we recommend estimating it with a parametric model because that model cannot be misspecified and its use may improve estimate precision.
\subsection*{Inference}
For all the estimators described above, inference using standard M-estimation methods \cite{stefanski2002} is possible (see references \cite{lunceford2004, williamson2014variance,yang2021causal} for examples). These methods can correctly account for the estimation of the working models for the probability of participation and the expectation of the outcome (and the probability of treatment, if needed) when obtaining measures of uncertainty. Furthermore, when using parametric models, we have found that numerical approximations to the usual M-estimation sandwich variance, such as those provided by the \texttt{geex} package \cite{saul2020calculus} in \texttt{R} \cite{currentRcitation}, work quite well in practice. Inference using non-parametric bootstrap methods \cite{efron1994introduction} is also easy to obtain and will often be preferred because of the simplicity of implementing resampling methods in most standard software packages.
\section*{SUBGROUP ANALYSES IN CASS}
\paragraph{CASS design and data:} The Coronary Artery Surgery Study (CASS) was a comprehensive cohort study \cite{olschewski1985} that compared coronary artery bypass grafting surgery plus medical therapy (hereafter ``surgery'') versus medical therapy alone for patients with chronic coronary artery disease. The data from a comprehensive cohort study, such as CASS, are consistent with the nested trial study design \cite{dahabreh2021studydesigns}. In CASS, among a total of 2099 trial-eligible patients, 780 agreed to participate in the randomized study; 1319 declined and were enrolled into an observational study of the same treatments. We excluded six patients for consistency with prior CASS analyses \cite{chaitman1990, olschewski1992} and in accordance with CASS data release notes. CASS recruited patients from August 1975 to May 1979, with follow-up until December 1996.
CASS collected a common set of baseline covariates using the same measurement approach among both the randomized and non-randomized groups. Additionally, the setting (in terms of doctors, location, and time) is similar between the randomized and non-randomized groups. In fact, the investigators who conducted CASS have, in prior work \cite{olschewski1992analysis}, jointly analyzed data from the randomized and non-randomized groups, suggesting that they would deem the consistency and exchangeability conditions (i.e., conditions 1 and 4, listed above) to be plausible \cite{dahabreh2020benchmarking}. Like most analyses attempting to draw causal inferences using data sources beyond a single randomized trial, our analyses of the CASS data invoke conditions that are not guaranteed to hold and are untestable. Nevertheless, we believe that the design of CASS makes the conditions reasonably plausible.
\paragraph{Estimators:} We applied the estimators presented above to estimate the 10-year risk of death (cumulative incidence proportion) and risk difference in subgroups defined by history of myocardial infarction, for the entire target population and its non-randomized subset. Because no patients were lost to follow-up during the first ten years of the study, the cumulative incidence proportion of death is a reasonable measure of risk. We provide code to implement the methods in Web Appendix 6.
\paragraph{Model specification:} We fit logistic regression models for the probability of trial participation, the probability of the outcome among individuals randomized in each treatment group, and the probability of treatment in the trial. All models adjusted for the main effects of covariates that we used in a previous analysis \cite{dahabreh2018generalizing} as well as an original CASS analysis that combined data from the randomized and observational components of the study \cite{olschewski1992}: age, severity of angina, history of previous myocardial infarction, percent obstruction of the proximal left anterior descending artery, left ventricular wall motion score, number of diseased vessels, and ejection fraction. For simplicity, we restricted our analyses to patients with complete data on the baseline covariates we adjusted for. In previous work \cite{dahabreh2018generalizing}, using weighting and multiple imputation methods under a missing-at-random assumption, we found that accounting for missing data did not appreciably affect estimates (and their uncertainty). To allow for more flexible modeling, we also used a generalized additive model (GAM) to estimate the probability of participation and the probability of the outcome. We used the \texttt{gam} function in the \texttt{mgcv} package (v1.8-34) \cite{Wood2011-ax, woodGAMbook} in \texttt{R} (v4.0.4) \cite{currentRcitation}, which fits a smoothness penalty using generalized cross-validation and we considered splines for age and ejection fraction.
\paragraph{Results:} We analyzed data from 1686 individuals; 731 randomized (368 to surgery and 363 to medical therapy) and 955 non-randomized. We summarize the baseline characteristics of individuals included in the analysis in Web Table 2 of Web Appendix 5. Estimates of the 10-year risk differences are shown for the target population of all trial-eligible patients (Table \ref{table_cass_analyses1}) and for its non-randomized subset (Table \ref{table_cass_analyses2}). We obtained non-parametric bootstrap-based standard errors and used them to calculate 95\% Wald-style confidence intervals for the 10-year mortality risk and risk difference (10,000 resampled datasets). We report results for the 10-year mortality risk for each treatment in each subgroup in Web Table 2. The results were similar across different estimators, suggesting that our modeling choices are reasonable \cite{robins2001}. Among patients without a history of previous myocardial infarction, we estimated a benefit of approximately 2\% in favor of medical therapy alone; among patients with a history of myocardial infarction we estimated a benefit of approximately 5\% in favor of surgery. The confidence intervals for both subgroups were wide and overlapping, reflecting considerable uncertainty in the subgroup analysis.
\section*{DISCUSSION}
We proposed g-formula, weighting, and augmented weighting estimators for subgroup-specific potential outcome means and average treatment effects in the entire target population and its non-randomized subset for nested trial designs. The estimators for the non-randomized subset of the target population are also applicable to non-nested trial designs, when the trial and non-randomized sample of observations are obtained separately \cite{dahabreh2021studydesigns}. Our augmented weighting estimators, which combine models for the probability of trial participation and the expectation of the outcome, may be preferred in practice because they offer robustness to model misspecification \cite{bang2005} and allow for valid inference when the models are estimated using data-adaptive methods \cite{chernozhukov2018double}, including machine learning techniques \cite{friedman2001elements, bishop2006pattern}.
The methods we describe should be useful for confirmatory or descriptive subgroup analyses \cite{varadhan2013framework} that examine heterogeneity of treatment effects over subgroups defined in terms of one or more key effect modifiers. Key effect modifiers are often identified in trial protocols on the basis of substantive knowledge (e.g., when pre-specifying subgroup analyses) \cite{dahabreh2017bookchapter}. Subgroup-specific treatment effects conditional on these effect modifiers are also of interest when extending inferences from trials to target populations. Estimates of these subgroup-specific treatment effects can be useful in the presence of strong effect modification, when the average treatment effect is not sufficient for guiding decisions in the target population, particularly in limited resource settings when we need to focus intervention efforts on subgroups that are most likely to benefit from treatment \cite{rothwell2005subgroup}. Furthermore, the methods we describe can form the basis of exploratory approaches for data-driven subgroup discovery (e.g., by adapting tree-based subgroup discovery approaches \cite{yang2021causal} to generalizability and transportability analyses).
\section*{ACKNOWLEDGEMENTS}
This work was supported in part by Agency for Healthcare Research and Quality (AHRQ) award R36HS028373-01 and Patient-Centered Outcomes Research Institute (PCORI) award ME-1306-03758. The content is solely the responsibility of the authors and does not necessarily represent the official views of PCORI, its Board of Governors, the PCORI Methodology Committee, or AHRQ. The data analyses in our paper used CASS research materials obtained from the NHLBI Biologic Specimen and Data Repository Information Coordinating Center. This paper does not necessarily reflect the opinions or views of the CASS or the NHLBI. Conflict of interest: none declared.
\clearpage
\bibliographystyle{ieeetr}
\renewcommand\refname{REFERENCES}
\section{Data and estimation in non-nested trial study designs}\label{appendix:study_design}
In a non-nested trial design, data from the trial participants and non-randomized individuals are collected separately. The data from the trial participants consist of $n_{\text{trial}}$ independent realizations of $(X, S=1, A, Y)$; the data from the sample of non-randomized individuals consist of $n_{\text{obs}}$ independent realizations of $(X, S=0)$.\footnote{In non-nested trial designs, the population of individuals with $S = 0$ is often referred to as \emph{the} target population and is not necessarily thought of as a subset of an overall target population.} Data from the trial participants and the non-randomized individuals are appended to form a composite dataset with sample size $n_{\text{trial}} + n_{\text{obs}}$. The data can be thought of as generated from a biased sampling model \cite{bickel1993efficient}, where randomized individuals and non-randomized individuals are sampled from an underlying population using possibly different, and unknown to the investigators, sampling fractions \cite{dahabreh2020transportingStatMed, dahabreh2021studydesigns}.
In the composite dataset, the estimators of $\phi(v,a)$ given in the main text, and this appendix, can, with minor modifications, be used to estimate subgroup-specific potential outcome means and average treatment effects in the population underlying the sample of non-randomized individuals. One necessary modification is that all conditional expectations and probabilities need to be estimated in the composite dataset; as such, they are estimated under the biased sampling model, mentioned above, and do not necessarily reflect the underlying population model. Another, is that all sums in the estimators are over the composite dataset, replacing $n$ with $n_{\text{trial}} + n_{\text{obs}}$.
\clearpage
\section{Identification results}\label{appendix:identification}
\subsection{Potential outcome means in the target population}
To derive the g-formula identification result, we re-write the subgroup-specific potential outcome mean for treatment $a$ using the observed data as follows:
\begin{equation*}
\begin{split}
\E[Y^a | X \in v]
&= \E\big[\E[ Y^a | X] \big| X \in v \big] \\
&= \E\big[\E[ Y^a | X, S=1] \big| X \in v \big] \\
&= \E\big[\E[ Y^a | X, S=1, A = a] \big| X \in v \big] \\
&= \E\big[\E[Y | X, S=1,A = a] \big| X \in v\big] \\
&\equiv \psi(v, a),
\end{split}
\end{equation*}
where the first step follows from the law of total expectation, the second by mean exchangeability over $S$, the third by conditional mean exchangeability over $A$ in the trial, the last by consistency, and all quantities are well-defined because of the positivity conditions.
We will now show that $\psi(v,a) \equiv \E\big[ \E[Y | X, S=1, A = a] | X \in v \big]$ has two inverse probability weighting re-expressions:
\begin{equation} \label{eq:ipw_re_expressions}
\begin{split}
\psi(v,a) &= \dfrac{1}{\Pr[X \in v]} \E\left[ \dfrac{ I(S = 1, A = a, X \in v)Y}{\Pr[S = 1| X] \Pr[A=a|X,S=1]} \right] \\
&= \Bigg\{ \E\left[\dfrac{I(S = 1, A = a, X \in v)}{\Pr[S = 1| X] \Pr[A=a|X,S=1]}\right] \Bigg\}^{-1} \E\left[ \dfrac{ I(S = 1, A = a, X \in v)Y}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]} \right].
\end{split}
\end{equation}
First, we note that under the positivity conditions,
\begin{equation*}
\begin{split}
\psi(v,a) &= \E\big[\E[Y | X, S=1, A = a] | X \in v\big] \\
&=\dfrac{1}{\Pr[X \in v]} \E\Big[I(X \in v) \E[Y|X, S=1, A=a] \Big] \\
&=\dfrac{1}{\Pr[X \in v]} \E\Bigg[ I(X \in v) \E\left[ \dfrac{ I(S = 1, A=a)Y}{\Pr[S = 1, A=a | X]} \Big| X \right] \Bigg] \\
&=\dfrac{1}{\Pr[X \in v]} \E\Bigg[ \E\left[ \dfrac{ I(S = 1, A=a,X \in v)Y}{\Pr[S = 1, A=a | X]} \Big| X \right] \Bigg] \\
&= \dfrac{1}{\Pr[X \in v]} \E\left[ \dfrac{ I(S = 1, A = a, X \in v)Y}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]} \right],
\end{split}
\end{equation*}
which establishes the first equality in \eqref{eq:ipw_re_expressions}.
Next, we note that
\begin{equation*}
\begin{split}
\E\left[\dfrac{I(S = 1, A = a,X \in v)}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]}\right] &= \E\Bigg[ \E\left[\dfrac{I(S = 1, A = a,X \in v)}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]} \Big| X \right] \Bigg] \\
&= \E\Bigg[ \dfrac{I(X \in v)}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]} \E[I(S=1, A = a) | X ] \Bigg] \\
&= \Pr[X \in v].
\end{split}
\end{equation*}
Thus, we can also write
\begin{equation*}
\psi(v,a) = \Bigg\{ \E\left[\dfrac{I(S = 1, A = a, X \in v)}{\Pr[S = 1 | X] \Pr[A = a | X , S = 1]}\right] \Bigg\}^{-1} \E\left[ \dfrac{ I(S = 1, A = a, X \in v)Y}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]} \right],
\end{equation*}
which establishes the second equality in \eqref{eq:ipw_re_expressions}.
\subsection{Potential outcome means in the non-randomized subset}
To derive the g-formula identification result, we re-write the subgroup-specific potential outcome mean for treatment $a$ using the observed data as follows:
\begin{equation*}
\begin{split}
\E[Y^a | X \in v, S=0]
&= \E\big[\E[ Y^a | X,S=0] \big| X \in v,S=0 \big] \\
&= \E\big[\E[ Y^a | X, S=1] \big| X \in v,S=0 \big] \\
&= \E\big[\E[ Y^a | X, S=1, A = a] \big| X \in v,S=0 \big] \\
&= \E\big[\E[Y | X, S=1,A = a] \big| X \in v, S=0\big] \\
&\equiv \phi(v, a),
\end{split}
\end{equation*}
where the first step follows from the law of total expectation, the second by mean exchangeability over $S$, the third by conditional mean exchangeability over $A$ in the trial, the last by consistency, and all quantities are well-defined because of the positivity condition.
We will now show that $\phi(v,a) \equiv \E\big[ \E[Y | X, S=1, A = a] | X \in v, S=0 \big]$ has two inverse probability weighting re-expressions:
\begin{equation} \label{eq:ipw_re_expressions_S0}
\begin{split}
\phi(v,a) &= \dfrac{1}{\Pr[X \in v,S=0]} \E\left[ \dfrac{ I(S = 1, A = a, X \in v) \Pr[S=0|X]Y}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]} \right] \\
&= \Bigg\{ \E\left[\dfrac{I(S = 1, A = a, X \in v)\Pr[S=0|X]}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]}\right] \Bigg\}^{-1} \E\left[ \dfrac{ I(S = 1, A = a, X \in v) \Pr[S=0|X]Y}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]} \right].
\end{split}
\end{equation}
First, we note that
\begin{equation*}
\begin{split}
\phi(v,a) &= \E\big[\E[Y | X, S=1, A = a] | X \in v, S=0\big] \\
&=\dfrac{1}{\Pr[X \in v, S=0]} \E\Big[I(X \in v,S=0) \E[Y|X, S=1, A=a] \Big] \\
&=\dfrac{1}{\Pr[X \in v,S=0]} \E\Bigg[ I(X \in v,S=0) \E\left[ \dfrac{ I(S = 1, A=a)Y}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]} \Big| X \right] \Bigg] \\
&=\dfrac{1}{\Pr[X \in v,S=0]} \E\Bigg[I(S=0) \E\left[\dfrac{ I(S = 1, A=a,X \in v)Y}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]} \Big| X \right] \Bigg] \\
&=\dfrac{1}{\Pr[X \in v,S=0]} \E\Bigg[\Pr[S=0|X] \E\left[\dfrac{ I(S = 1, A=a,X \in v)Y}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]} \Big| X \right] \Bigg] \\
&=\dfrac{1}{\Pr[X \in v,S=0]} \E\Bigg[ \E\left[\dfrac{ I(S = 1, A=a,X \in v) \Pr[S=0|X] Y}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]} \Big| X \right] \Bigg] \\
&= \dfrac{1}{\Pr[X \in v,S=0]} \E\left[ \dfrac{ I(S = 1, A = a, X \in v)\Pr[S=0|X] Y}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]} \right],
\end{split}
\end{equation*}
which establishes the first equality in \eqref{eq:ipw_re_expressions}. Next, we note that
\begin{equation*}
\begin{split}
\E\left[\dfrac{I(S = 1, A = a,X \in v)\Pr[S=0|X] }{\Pr[S = 1 | X] \Pr[A = a | X , S = 1]}\right] &= \E\Bigg[ \E\left[\dfrac{I(S = 1, A = a,X \in v)\Pr[S=0|X] }{\Pr[S = 1 | X] \Pr[A = a | X , S = 1]} \Big| X \right] \Bigg] \\
&= \E\Bigg[ \dfrac{I(X \in v) \Pr[S = 0 | X]}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]} \E[I(S=1, A = a) | X ] \Bigg] \\
&= \E\big[ I(X \in v) \Pr[S = 0 | X] \big] \\
&= \E\big[ \E[ I(X \in v , S = 0) | X] \big] \\
&= \Pr[X \in v,S=0].
\end{split}
\end{equation*}
Thus, we can also write
\begin{equation*}
\phi(v,a) = \Bigg\{ \E\left[\dfrac{I(S = 1, A = a, X \in v)\Pr[S = 0 | X]}{\Pr[S = 1 | X] \Pr[A = a | X , S = 1]}\right] \Bigg\}^{-1} \E\left[ \dfrac{ I(S = 1, A = a, X \in v)Y \Pr[S = 0| X]}{\Pr[S = 1 | X] \Pr[A = a | X, S = 1]} \right],
\end{equation*}
which establishes the second equality in \eqref{eq:ipw_re_expressions}.
\clearpage
\section{Weighted estimators using saturated regression models}\label{appendix:saturated_IPW_estimators}
\setcounter{figure}{0}
\setcounter{equation}{0}
\renewcommand{\figurename}{Web Figure}
\renewcommand{\tablename}{Web Table}
\subsection{Potential outcome means and treatment effects in the target population}
In the main text we provided the following equation for the inverse probability weighting estimator with normalized weights: \cite{hajek1971comment}:
\begin{equation*} \label{IPW_estimator2}
\widehat {\psi}_{\text{\tiny{IPW2}}}(v,a) = \left\{ \sum_{i=1}^{n} \widehat w_{v,a}(X_i, S_i, A_i) \right\}^{-1} \sum\limits_{i = 1}^{n} \widehat w_{v,a}(X_i, S_i, A_i) Y_i.
\end{equation*}
An estimator equivalent to $\widehat {\psi}_{\text{\tiny{IPW2}}}(v,a)$ can be obtained by fitting an appropriate saturated mean regression model for the outcome $Y$ conditional on indicators (dummy variables) for treatment $A$ and subgroup membership (for each subgroup $v \in \{v_1, \ldots, v_k\}$) and weights equal to the inverse of the product of the probability of trial participation times the probability of receiving the treatment actually received among trial participants for observations in the trial, and 0 otherwise \cite{robins2000a}. A saturated mean model should include main effects for treatment and the subgroup indicators, as well as all possible interactions between them. The regression can be estimated just among trial participants because non-randomized individuals receive weight zero.
As an example, consider a binary treatment $A$; two subgroups $v_1$ and $v_2$, such that $\mathcal X = v_1 \cup v_2$, and a continuous outcome $Y$. We estimate the following saturated linear regression model among trial participants, $$\E[Y | X, A , S = 1] = \alpha_0 + \alpha_1 A + \alpha_2 I(X \in v_1) + \alpha_3 A \times I(X \in v_1),$$ using estimated observation-level weights equal to $$ \dfrac{S_i}{\widehat p(X_i)} \times \left\{ \dfrac{A_i}{\widehat e_1(X_i)} + \dfrac{1 - A_i}{\widehat e_0(X_i)} \right\}.$$ To connect these weights with the notation in the main text, note that $$ \dfrac{S_i}{\widehat p(X_i)} \times \left\{ \dfrac{A_i}{\widehat e_1(X_i)} + \dfrac{1 - A_i}{\widehat e_0(X_i)} \right\} = \sum\limits_{v,a} \widehat w_{v,a}(X_i, S_i, A_i).$$ The estimated regression coefficients from this model, $\widehat \alpha_0, \ldots, \widehat \alpha_3$, can be used to obtain the estimators $\widehat \psi_{\text{\tiny IPW2}}(v,a)$, for $a=0,1$ and $v=0,1$, as follows:
\begin{equation*}
\begin{split}
\widehat \psi_{\text{\tiny IPW2}}(0,0) &= \widehat \alpha_0, \\
\widehat \psi_{\text{\tiny IPW2}}(0,1) &= \widehat \alpha_0 + \widehat \alpha_1, \\
\widehat \psi_{\text{\tiny IPW2}}(1,0) &= \widehat \alpha_0 + \widehat \alpha_2, \\
\widehat \psi_{\text{\tiny IPW2}}(1,1) &= \widehat \alpha_0 + \widehat \alpha_1 + \widehat \alpha_2 + \widehat \alpha_3.
\end{split}
\end{equation*}
Furthermore, the estimated coefficients can be used to estimate subgroup-specific average treatment effects. For example, the average treatment effect in the subgroup with $v=0$ can be estimated by $\widehat \psi_{\text{\tiny IPW2}}(0,1) - \widehat \psi_{\text{\tiny IPW2}}(0,0) = \widehat \alpha_1$. Similarly, the average treatment effect in the subgroup with $v = 1$, can be estimated by
$\widehat \psi_{\text{\tiny IPW2}}(1,1) - \widehat \psi_{\text{\tiny IPW2}}(1,0) = \widehat \alpha_1 + \widehat \alpha_3.$
To properly account for the estimation of the weights, inference should be based on resampling methods (e.g., the non-parametric bootstrap \cite{efron1994introduction}, estimating the weights and the saturated model in each bootstrap iteration \cite{robins2000a, lunceford2004}). An alternative is to jointly estimate the weights and the potential outcome means/average treatment effects (e.g., using standard stacked M-estimation methods \cite{stefanski2002, lunceford2004}). Using the robust (Huber-White) variance estimator \cite{huber1967behavior, white1996estimation} when fitting the weighted regression model with estimated weights does not explicitly account for the estimation of the weights but is expected to produce conservative estimates of uncertainty \cite{robins2000a}.
\subsection{Potential outcome means and treatment effects in the non-randomized subset of the target population\footnote{The approach we describe in this subsection is similar to the one in reference \cite{mehrotra2021transporting}, but we use weights that involve the inverse of the estimated probability of receiving the treatment actually received (among trial participants) to improve efficiency.}}
In the main text, we provided the following equation for the inverse odds weighting estimator with normalized weights:
\begin{equation*} \label{IPW_estimator2_S0}
\widehat {\phi}_{\text{\tiny{IOW2}}}(v,a) = \left\{ \sum_{i=1}^{n} \widehat o_{v,a}(X_i, S_i,A_i) \right\}^{-1} \sum\limits_{i = 1}^{n} \widehat o_{v,a}(X_i, S_i,A_i) Y_i.
\end{equation*}
Similar to the previous section of this Appendix, an estimator equivalent to $\widehat {\phi}_{\text{\tiny{IOW2}}}(v,a)$ can be obtained by fitting an appropriate saturated mean regression model for the outcome $Y$ conditional on indicators (dummy variables) for treatment $A$ and subgroup membership (for each subgroup $v \in \{v_1, \ldots, v_k\}$) and weights equal to the inverse of the product of the odds of trial participation times the probability of receiving the treatment actually received among trial participants for observations in the trial, and 0 otherwise \cite{robins2000a}. As above, a saturated mean model should include main effects for treatment and the subgroup indicators, as well as all possible interactions between them. The model can be estimated just among trial participants because non-randomized individuals receive weight zero. As an example, consider a binary treatment $A$; two subgroups $v_1$ and $v_2$, such that $\mathcal X = v_1 \cup v_2$, and a continuous outcome $Y$. We estimate the following saturated linear regression model among trial participants, $$\E[Y | X, A , S = 1] = \alpha_0 + \alpha_1 A + \alpha_2 I(X \in v_1) + \alpha_3 A \times I(X \in v_1),$$ using estimated observation-level weights equal to $$ \dfrac{S_i \times \{ 1- \widehat p(X_i) \} }{\widehat p(X_i)} \times \left\{ \dfrac{A_i}{\widehat e_1(X_i)} + \dfrac{1 - A_i}{\widehat e_0(X_i)} \right\}.$$ To connect these weights with the notation in the main text, note that $$ \dfrac{S_i \times \{ 1-\widehat p(X_i) \} }{\widehat p(X_i)} \times \left\{ \dfrac{A_i}{\widehat e_1(X_i)} + \dfrac{1 - A_i}{\widehat e_0(X_i)} \right\} = \sum\limits_{v,a} \widehat o_{v,a}(X_i, S_i, A_i).$$ The estimated regression coefficients from this model, $\widetilde \alpha_0, \ldots, \widetilde \alpha_3$, can be used to obtain the estimators $\widehat \psi_{\text{\tiny IOW2}}(v,a)$, for $a=0,1$ and $v=0,1$, as follows:
\begin{equation*}
\begin{split}
\widehat \phi_{\text{\tiny IOW2}}(0,0) &= \widetilde \alpha_0, \\
\widehat \phi_{\text{\tiny IOW2}}(0,1) &= \widetilde \alpha_0 + \widetilde \alpha_1, \\
\widehat \phi_{\text{\tiny IOW2}}(1,0) &= \widetilde \alpha_0 + \widetilde \alpha_2, \\
\widehat \phi_{\text{\tiny IOW2}}(1,1) &= \widetilde \alpha_0 + \widetilde \alpha_1 + \widetilde \alpha_2 + \widetilde \alpha_3.
\end{split}
\end{equation*}
Furthermore, these estimated coefficients can be used to estimate subgroup-specific average treatment effects. For example, the average treatment effect in the subgroup with $v=0$ can be estimated by $\widehat \phi_{\text{\tiny IPW2}}(0,1) - \widehat \phi_{\text{\tiny IPW2}}(0,0) = \widetilde \alpha_1$. Similarly, the average treatment effect in the subgroup with $v = 1$, can be estimated by
$\widehat \phi_{\text{\tiny IPW2}}(1,1) - \widehat \phi_{\text{\tiny IPW2}}(1,0) = \widetilde \alpha_1 + \widetilde \alpha_3.$
Inference can be carried out as described in the previous subsection of this appendix.
\clearpage
\section{Weighted multivariable regression estimators}\label{appendix:additional_estimators}
\setcounter{figure}{0}
\setcounter{equation}{0}
\renewcommand{\figurename}{Web Figure}
\renewcommand{\tablename}{Web Table}
\subsection{Potential outcome means in the target population}
A third doubly robust estimator relies on fitting a multivariable regression model for the outcome estimated using inverse probability weighting, followed by standardization over the distribution of baseline covariates \cite{robins2007, wooldridge2007}. The potential outcome mean is estimated as
\begin{equation} \label{DR_estimator3}
\widehat {\psi}_{\text{\tiny{AIPW3}}}(v,a) = \left\{ \sum_{i=1}^{n} I(X_i \in v) \right\}^{-1} \sum\limits_{i = 1}^{n} I(X_i \in v) g_{a}(X_i; \widehat\theta_a),
\end{equation}
where $g_{a}(X; \widehat \theta_a)$ is an estimator for $\E[Y | X, S=1, A = a]$, with a vector of estimated model parameters $\widehat\theta_a$, from an inverse probability weighted outcome regression, with weights $$\dfrac{S_i}{\widehat p(X_i)} \times \left\{ \dfrac{A_i}{\widehat e_1(X_i)} + \dfrac{1 - A_i}{\widehat e_0(X_i)} \right\}.$$ When the outcome is modeled in the linear exponential family with a canonical link, and estimation is by quasi-likelihood methods (e.g., linear or logistic regression) \cite{gourieroux1984}, this estimator has the double robustness property.
\subsection{Potential outcome means in the non-randomized subset of the target population}
A third doubly robust estimator relies on fitting a multivariable regression model for the outcome estimated using inverse odds weighting, followed by standardization over the distribution of baseline covariates \cite{robins2007, wooldridge2007}. The potential outcome mean is estimated as
\begin{equation} \label{DR_estimator3_S0}
\widehat {\phi}_{\text{\tiny{AIOW3}}}(v,a) = \left\{ \sum_{i=1}^{n} I(X_i \in v, S = 0) \right\}^{-1} \sum\limits_{i = 1}^{n} I(X_i \in v, S = 0) g_{a}(X_i; \widetilde\theta_a),
\end{equation}
where $g_{a}(X; \widetilde \theta_a)$ is an estimator for $\E[Y | X, S=1, A = a]$, with a vector of estimated model parameters $\widetilde\theta_a$, from an inverse odds weighted outcome regression, with weights $$\dfrac{S_i \times \{ 1-\widehat p(X_i) \} }{\widehat p(X_i)} \times \left\{ \dfrac{A_i}{\widehat e_1(X_i)} + \dfrac{1 - A_i}{\widehat e_0(X_i)} \right\}.$$ When the outcome is modeled in the linear exponential family with a canonical link, and estimation is by quasi-likelihood methods (e.g., linear or logistic regression) \cite{gourieroux1984}, this estimator has the double robustness property.
\clearpage
\setcounter{figure}{0}
\setcounter{equation}{0}
\section{Additional results from the CASS analyses}\label{appendix:cass_baselines}
\subsection*{Baseline covariates}
\begin{table}[H]
\caption{CASS baseline table (August 1975 to December 1996). $S=1 $ indicates randomized ($S=0$ indicates non-randomized) $A=1$ indicates surgical therapy ($A=0$ indicates medical therapy).}
\label{cass_baseline}
\centering
\begin{tabular}{@{}lcccc@{}}
\toprule
& $S=1, A=1$ & $S=1, A=0$ & $S=1$ & $S=0$ \\ \midrule
Number of patients & 368 & 363 & 731 & 955 \\
Age & 51.42 (7.24) & 50.92 (7.41) & 51.17 (7.32) & 50.89 (7.73) \\
History of angina & 285 (77.4) & 282 (77.7) & 567 (77.6) & 760 (79.6) \\
Ejection fraction & 60.86 (13.04) & 59.83 (12.78) & 60.35 (12.91) & 60.16 (12.25) \\
History of previous MI & 209 (56.8) & 228 (62.8) & 437 (59.8) & 549 (57.5) \\
LMCA percent obstruction & 4.27 (11.87) & 2.78 (9.55) & 3.53 (10.80) & 5.76 (14.50) \\
PLMA percent obstruction & 36.44 (38.04) & 34.89 (36.95) & 35.67 (37.49) & 39.14 (38.73) \\
Any diseased proximal vessels & 222 (60.3) & 230 (63.4) & 452 (61.8) & 608 (63.7) \\ \bottomrule
\end{tabular}
\caption*{MI= myocardial infarction; LMCA = left main coronary artery; PLMA= proximal left anterior artery. For continuous variables we report the mean (standard deviation); for binary variables we report the number of individuals (percentage).}
\end{table}
\clearpage
\subsection*{Estimates of potential outcome means}
We report estimates of the subgroup-specific mortality risk for each treatment group, for the entire target population and its non-randomized subset. All equation numbers in the captions refer to the main text.
\vspace{0.3in}
\begin{table}[ht!]
\footnotesize
\centering
\caption{Potential outcome mean estimates for subgroup analyses in the CASS study (August 1975 to December 1996), for previous myocardial infarction.}
\label{appendix_table_cass_analyses1}
\begin{tabular}{@{}llllll@{}}
\toprule
\multicolumn{1}{l}{Estimator} & \multicolumn{2}{c}{$v=1$} & \multicolumn{2}{c}{$v=0$} & \\
& \multicolumn{1}{c}{$a=1$} & \multicolumn{1}{c}{$a=0$} & \multicolumn{1}{c}{$a=1$} & \multicolumn{1}{c}{$a=0$} & \\ \midrule
\multicolumn{6}{l}{Trial-only population} \\ \midrule
TRIAL & 20.6\% (15.2, 26.1) & 25.0\% (19.5, 30.7) & 13.2\% (8.2, 18.6) & 11.9\% (6.6, 17.6) \\ \midrule
\multicolumn{6}{l}{Entire target population} \\ \midrule
OM & 20.6\% (14.9, 26.3) & 25.7\% (20.1, 31.3) & 13.6\% (8.2, 19.0) & 12.1\% (6.7, 17.5) & \\
IPW1 & 21.1\% (15.2, 26.9) & 25.7\% (20.1, 31.4) & 13.1\% (7.8, 18.3) & 12.0\% (6.5, 17.6) & \\
IPW2 & 20.8\% (15.1, 26.5) & 25.8\% (20.1, 31.4) & 13.3\% (8.0, 18.6) & 12.0\% (6.5, 17.5) & \\
AIPW1 & 20.6\% (14.9, 26.3) & 25.8\% (20.2, 31.4) & 13.6\% (8.1, 19.0) & 11.9\% (6.6, 17.2) & \\
AIPW2 & 20.6\% (14.9, 26.3) & 25.8\% (20.2, 31.4) & 13.6\% (8.1, 19.0) & 11.9\% (6.6, 17.2) & \\
AIPW1 (GAM) & 21.1\% (15.3, 26.9) & 25.8\% (20.2, 31.5) & 13.9\% (8.3, 19.6) & 11.7\% (6.4, 17.0) & \\
AIPW2 (GAM) & 21.1\% (15.3, 26.9) & 25.8\% (20.2, 31.5) & 13.9\% (8.3, 19.6) & 11.7\% (6.4, 17.0) & \\ \midrule
\multicolumn{6}{l}{Non-randomized subset of the target population} \\ \midrule
OM & 20.8\% (14.9, 26.7) & 25.5\% (19.7, 31.2) & 13.9\% (8.3, 19.4) & 12.4\% (6.9, 18.0) & \\
IOW1 & 21.4\% (15.3, 27.4) & 25.7\% (19.9, 31.5) & 13.1\% (7.8, 18.4) & 12.3\% (6.6, 17.9) & \\
IOW2 & 20.9\% (15.0, 26.8) & 25.8\% (20.0, 31.6) & 13.5\% (8.0, 18.9) & 12.2\% (6.6, 17.8) & \\
AIOW1 & 20.6\% (14.7, 26.6) & 25.6\% (19.8, 31.4) & 13.9\% (8.2, 19.6) & 12.2\% (6.8, 17.7) & \\
AIOW2 & 20.6\% (14.7, 26.6) & 25.6\% (19.8, 31.4) & 13.9\% (8.2, 19.6) & 12.2\% (6.8, 17.7) & \\
AIOW1 (GAM) & 21.4\% (15.0, 27.8) & 25.6\% (19.5, 31.7) & 14.3\% (8.3, 20.3) & 12.0\% (6.4, 17.5) & \\
AIOW2 (GAM) & 21.4\% (15.0, 27.8) & 25.6\% (19.5, 31.7) & 14.3\% (8.3, 20.3) & 12.0\% (6.4, 17.5) & \\ \bottomrule
\end{tabular}
\caption*{We obtained 95\% Wald-style confidence intervals using standard errors estimated from 10,000 bootstrap samples. CASS = Coronary Artery Surgery Study; $v=1$ indicates history of myocardial infarction ($v=0$ indicates no history of myocardial infarction); GAM = generalized additive models used to estimate the probability of trial participation and the expectation of the outcome, with a parametric model used to estimate the probability of treatment. For estimators in the table not labeled by GAM, we used parametric models. Here $a=1$ is surgical therapy; $a=0$ is medical therapy. See main text for descriptions of estimators.}
\end{table}
\clearpage
\setcounter{figure}{0}
\setcounter{equation}{0}
\section{Code and data}\label{appendix:code}
\paragraph{Code for empirical analyses:} We provide \texttt{R} \cite{currentRcitation} code to implement all the estimators described in the paper for the $\texttt{R}$ environment on GitHub: [link removed for peer review].
\begin{itemize}
\item[] \texttt{00\_subgroup\_generalizability\_source\_code\_tablecenters.R} contains the source code for all the estimators.
\item[] \texttt{01\_subgroup\_transportability.R} runs the source code, using a simulated dataset for illustration.
\end{itemize}
\paragraph{Data availability:} The CASS study data are not publicly available, but they can be obtained from the National Heart, Lung, and Blood Institute (NHLBI) Biologic Specimen and Data Repository Information Coordinating Center (\url{https://biolincc.nhlbi.nih.gov/studies/cass}; last accessed April 30, 2021).
\clearpage
\bibliographystyle{unsrt}
|
\subsection{Datasets}
The model is trained on two publicly available datasets that are benchmarks in the field: The ETH \cite{pellegrini2009you}, with subsets named ETH and HOTEL, and the UCY \cite{lerner2007crowds} datasets, with subsets named ZARA1, ZARA2, and UNIV. The trajectories are sampled at $0.4$ seconds intervals. The model observes 8 time steps, which corresponds to 3.2 seconds, and predicts the next 12 time steps, which corresponds to 4.8 seconds.
To further evaluate and demonstrate Grouptron's performance in densely populated scenarios, we create UNIV-N test sets, where $N$ is the minimum number of people present simultaneously at each time step in the test sets. Each UNIV-N test set contains all time steps that have at least $N$ people in the scene simultaneously from the original UNIV test set. In this way, we created test sets UNIV-40, UNIV-45, and UNIV-50 and, at each time step, there are at least 40, 45, and 50 people in the scene at the same time. These test sets are far more challenging than the original UNIV test set because of the more complex and dynamic interactions at different scales. We train the models on the original UNIV training set and evaluate the models on the UNIV-40, UNIV-45, and UNIV-50 test sets.
\subsection{Evaluation Metrics}
As in prior work \cite{alahi2016social,salzmann2020trajectron++,ivanovic2019trajectron} and more, we used the following metrics to evaluate our model:
\subsubsection{Final Displacement Error}
\begin{equation}
FDE = \frac{\sum_{i\in N}|| \hat{T^i_{T_F}}-T^i_{T_F} ||_2}{N},
\end{equation}
which is the $\textit{l}_2$ distance between the predicted final position and the ground truth final position with prediction horizon $T_F$.
\subsubsection{Average Displacement Error}
\begin{equation}
ADE = \frac{\sum_{i\in N}\sum_{t\in T_F}|| \hat{T^i_t}-T^i_t ||_2}{N\times T_F},
\end{equation}
where $N$ is the total number of pedestrians, $T_F$ is the number of future timesteps we want to predict for, $\hat{T^i_{t}}$ is the predicted trajectory for pedestrian $i$ at timestep $t$. $ADE$ is the mean $\textit{l}_2$ distance between the ground truth and predicted trajectories.
\subsection{Experiment settings}
The Grouptron model is implemented using PyTorch. The model is trained using an Intel I7 CPU and NVIDIA GTX 1080 Ti GPUs for 100 epochs. The batch size is 256. For the HOTEL, UNIV, ZARA1, and ZARA2 datasets, the output dimension for the group-level and scene level encoders are 16. For the ETH dataset, we set the output dimension of the group-level and scene-level encoders to be 8. This is because the ETH test set contains only 2 timesteps with at least 5 people in the scene, out of the total 1161 timesteps. In comparison, the training set contains 1910 timesteps, out of 4976 in total, with at least 5 people. Thus, to help the model learn generalizable representations in this case, we decrease the output dimension of the STGCNs to 8. The learning rate is set to $0.001$ initially and decayed exponentially every epoch with a decay rate of 0.9999. The model is trained using Adam gradient descent and gradients are clipped at $1.0$.
\subsection{Evaluation of the Group Clustering Algorithm}
\begin{figure}[t]
\centering
\includegraphics[width=0.75\columnwidth]{Figure/cluster.png}
\caption{An example of pedestrian group clustering. Pedestrians are divided into five groups. Different colors indicate different groups.}
\label{group}
\vspace{-3mm}
\end{figure}
In Fig. \ref{group}, it is shown that the groups created by the agglomerative clustering method are very close to the natural definition of pedestrian groups. We can see that pedestrians 5 and 6 are travelling in a highly correlated fashion and pedestrians 1, 2, 3, and 4's trajectories are highly similar as well. In both cases, the clustering algorithm is able to correctly cluster these pedestrians into their corresponding groups. This shows that by using agglomerative clustering based on Hausdorff distances, Grouptron is able to successfully generate naturally defined groups.
To quantitatively evaluate the groups generated by the algorithm, we selected 10 random time steps from the ETH training dataset. We invited 10 human volunteers to label groups for the time steps and used the agglomerative clustering method described in Section III-C to generate group clusters, respectively. For both human-generated groups and algorithm-generated groups, the number of groups to be formed is computed using Equation 1. We then compute the average Sørensen–Dice coefficient between human-generated and algorithm-generated groups. That is, we use
\begin{equation}
DSC = \frac{1}{T}\sum_{t\in T}\frac{1}{H}\sum_{h\in H}\frac{2\times |G_{h,t}\cap G_{a,t}|}{|G_{h,t}|+|G_{a,t}|},
\end{equation}
where $T$ is the number of timesteps, $H$ is the total number of human annotators, $G_{a, t}$ is the grouping created by the agglomerative clustering method for time step $t$, $G_{h,t}$ is the grouping created by humans for time step $t$, and $|G_{h}\cap G_{a}|$ measures how many of the groups by humans and the algorithm are exactly the same.
The Average Dice coefficient between human annotators and the algorithm is 0.72. The higher the Dice coefficient, the more similar are the groups created by the agglomerative clustering method and human annotators. Thus, The average Dice Coefficient of 0.72 indicates the groups output by agglomerative clustering method are really similar to human-generated ones.
\subsection{Quantitative Results}
We compare Grouptron's performance with state-of-the-art methods and common baselines in the field in terms of the FDE and ADE metrics, and the results are shown in Table \ref{table1}. Overall, Grouptron outperforms all state-of-the-art methods with considerable decrease in displacement errors. Since Grouptron is built on Trajectron++, we also compare the FDE and ADE values of Grouptron and Trajectron++. We find that Grouptron outperforms Trajectron++ on all 5 datasets by considerable margins. Particularly, on the ETH dataset, Grouptron achieves an FDE of 1.56m, which is 7.1\% better than the FDE value of 1.68m by Trajectron++. Furthermore, Grouptron achieves an FDE of 0.97m on the UNIV dataset. This is 9.3\% reduction in FDE error when compared with the FDE value of 1.07m by Trajectron++ on the same dataset.
Moreover, we compare Grouptron's performance with Trajectron++'s in dense crowds with the UNIV-N datasets in Table \ref{table2}. Overall, Grouptron outperforms Trajectron++ on all the UNIV-N test sets by enormous margins. In particular, Grouptron achieves an FDE of 1.04m and ADE of 0.40m on the UNIV-45 test set, which contains all timesteps from the original UNIV test set that have at least 45 pedestrians in the scene at the same time. This is 16.1\% in FDE improvement when compared with the FDE value of Trajectron++ and 13.0\% in ADE improvement when compared with the ADE value of Trajectron++ on the same test set.
Furthermore, we notice that the state-of-the-art method, Trajectron++, performs substantially worse as the number of pedestrians in the scene increases. Specifically, Trajectron++'s FDE increases from 1.07m to 1.25m as the minimum number of pedestrians in the scene increases from 1 to 50. In contrast, Grouptron's FDE remains relatively stable as the number of pedestrians increases. This shows that Grouptron performs much better and is more robust in densely populated scenarios.
\subsection{Qualitative Analysis}
\begin{figure}[t]
\vspace{3mm}
\centering
\includegraphics[scale=0.33]{Figure/res1.png}
\caption{Examples of Grouptron's predictions on the UNIV dataset. The predictions are the most likely trajectory predictions of the model. Green arrows indicate pedestrians' current positions and directions. Black dashed lines indicate trajectory histories. Grey dashed lines indicate ground truth future trajectories. Green lines are Grouptron's predicted future trajectories. }
\label{qualatative}
\vspace{-3mm}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[scale=0.30]{Figure/res_cmpare.png}
\caption{Comparisons of Grouptron's and Trajectron++'s distributions of 20 most likely future trajectories on examples of the UNIV dataset. Rows indicate different examples and columns represent different methods. Orange stars indicate pedestrians of interest. Yellow stars indicate their companions in the same pedestrian groups. Green arrows indicate pedestrians' current positions and directions. Black dashed lines indicate trajectory histories. Grey dashed lines indicate ground truth future trajectories. The comparisons show that Grouptron is able to produce predictions with higher quality and with better confidence levels.}
\label{qualatative2}
\vspace{-3mm}
\end{figure}
Fig. \ref{qualatative} shows Grouptron's most likely predictions for some examples of the UNIV dataset. Fig. \ref{qualatative}a shows two pedestrian groups crossing paths. We can see that Grouptron's predictions are consistent with the groups. Furthermore, it accurately predicts when and where the two groups' trajectories intersect. Fig. \ref{qualatative}b shows a case where pedestrians are forming groups and merging paths. Grouptron again successfully predicts the formation of this group. Fig. \ref{qualatative}c and Fig. \ref{qualatative}d show Grouptron's performance in densely populated scenes with more than 40 pedestrians. Even in these extremely challenging scenarios for state-of-the-art methods, Grouptron still produces predictions of high quality and the predictions are consistent with pedestrian groups. Furthermore, we can see that even when pedestrian groups are crossing paths or influencing each other, Grouptron successfully predicts these highly dynamic and complex scenarios.
In Fig. \ref{qualatative2}, we compare Grouptron's distributions of 20 most likely predictions with those of Trajectron++'s. Comparing Fig. \ref{qualatative2}a with \ref{qualatative2}b and \ref{qualatative2}c with \ref{qualatative2}d shows that Grouptron's predictions for the pedestrians of interest reflect the interactions within pedestrian groups more accurately. Furthermore, Grouptron's prediction distributions have much smaller ranges, indicating that it is much more confident with prediction outcomes.
\subsection{Problem Formulation}
\input{Formulation}
\subsection{Model Overview}
Rooted in the CVAE architecture in Trajectron++ \cite{salzmann2020trajectron++}, we design a more expressive multi-scale scene encoding structure, which actively takes into consideration the group-level and scene-level information for better representation of crowded scenes where groups of pedestrians are present.
Concretely, we leverage spatio-temporal graphs for each level to model information and interactions at the corresponding level. We refer to our model as Grouptron. Our model is illustrated in Fig. \ref{network}. In this subsection, we provide an overview of the architecture, and in
Section III-C, we elaborate on the details of the Grouptron model.
At the individual level, we construct spatio-temporal graphs for individual pedestrians. The graph is centered at the node whose trajectory we want to predict. We call it the ``current node''. Long short term memory (LSTM) networks \cite{hochreiter1997long} are used to encode this graph. We group the pedestrians with the agglomerative clustering algorithm based on Hausdorff distances \cite{atev2010clustering}. STGCN is used to encode dynamics within the groups.
At the scene level, spatio-temporal graphs are created to model dynamics among pedestrian groups and are encoded using a different STGCN. Lastly, the information across different scales is combined. A decoder is then used to obtain trajectory predictions and the model can output the most possible trajectory or the predicted trajectory distributions.
\subsection{Multi-Scale Scene Encoder}
\subsubsection{Individual-Level Encoder}
The first level of encoding is for the individual pedestrians. We represent information at the individual level using a spatio-temporal graph for the current node. The nodes include the current node and all other nodes that are in the perception range of the current node and nodes whose perception range covers the current node. The node states are the trajectories of the nodes. The edges are directional and there is an edge $e_{i,j}$ if pedestrian $i$ is in the perception range of pedestrian $j$.
To encode the current node's history trajectory, we use an LSTM network with hidden dimension 32. To encode the edges, we first perform an element-wise sum on the states of all neighboring nodes to form a single vector that represents all neighbors of the current node. This vector is then fed into the edge LSTM network which is an LSTM network with hidden dimension 8.
In this way, we obtain two vectors: a vector encoding the trajectory history of the current node and a vector encoding the representation of all the neighbors of the current node.
\subsubsection{Pedestrian Group Clustering}
To cluster nodes into groups based on trajectories, we propose to leverage the agglomerative clustering algorithm \cite{atev2010clustering}, which uses similarity scores based on Hausdorff distances between trajectories. The number of clusters (groups) to create for each scene is determined by:
\begin{equation}
C(N) = (N+1)/2,
\end{equation}
where C is the number of clusters and N is the total number of nodes to be clustered.
Furthermore, We only include nodes with an edge to or from the current node. This is because we only want to include nodes that can potentially influence the current node to avoid unhelpful information from nodes that are too far away from the current node.
\subsubsection{Group-Level Encoder}
For each group, we create a spatio-temporal graph consisting of $G_{g, t} = (V_{g,t}, E_{g, t})$, where $t$ is the time step and $g$ is the group id. $V_{g,t}= \{v_{i,t} \mid \forall i \in \{1,...,N_{g}\}\}$ are all the nodes in the group $g$. The node states are the trajectories of the represented pedestrians. $E_{g, t}=\{e^{i,j}_t \mid \forall i,j \in \{1,...,N_{g}\}$ are the set of edges between the nodes in the current group such that $e^{i,j}_t = 1$ to allow maximum interaction modeling within pedestrian groups.
After forming the aforementioned graphs for each group, they are then passed to the group-level trajectory encoder to obtain the encoded vectors for nodes in each group. The group-level trajectory encoder is an STGCN proposed in \cite{yan2018spatial} and used in \cite{mohamed2020social}. We set the convolution filter size to 3 and use the same weight parameters for all the groups.
We then average the encoded vectors of all nodes in each group to obtain the representations for the corresponding groups. That is, $E_{g} = \frac{1}{N_{g}} \sum_{i=1}^{N_{g}} E_{i}$, where $E_{g}$ is the encoded vector for group $g$, $E_{i}$ is the encoded vector for node $i$ in the output from the group-level trajectory encoder, and $N_g$ is the number of nodes in group $g$.
\subsubsection{Scene-Level Encoder}
After obtaining the encoded vectors for each group, a scene-level spatio-temporal graph with nodes representing groups is created. That is, $G_{scene, t} = (V_{scene,t}, E_{scene, t})$. $V_{scene,t}= \{v_{g, t} \mid \forall\,\text{g} \in \{1,...,G\}\}$, where G is the total number of groups and $t$ is the timestep. The state for each node is $E_{g}$ from the group-level trajectory encoder. $E_{scene, t}=\{e^{i,j}_t \mid \forall i,j \in \{1,...,G\}$ are the set of edges between the groups in the scene. Each $e^{i,j}$ is set to $1$ to allow maximum message passing between group nodes.
We then select the encoded vector corresponding to the last timestep and the group id of the current node as the scene-level encoding: $E_{scene} = E_{g, T}$, where $g$ is the group id of the current node we are encoding for and T is the total number of time steps.
\subsubsection{Multi-Scale Encoder Output}
The output of the multi-scale scene encoder is the concatenation of the following level encoded vectors: the output from the node history encoder, the output from individual-level edge encoder, and the output from the scene-Level encoder. That is, $ E_{multi} = [E_{his};E_{edge};E_{scene}]$, where $E_{his}$ is the encoded vector for the current node's history trajectory, $E_{edge}$ is the vector representing individual-level neighbors, $E_{scene}$ is the encoded vector from the scene-level encoder.
\subsection{Decoder Network}
Together with the latent variable $z$, $E_{multi}$ is passed to the decoder that is a Gated Recurrent Unit (GRU) \cite{chung2014empirical} with 128 dimensions. The output from the GRU decoder is then fed into dynamics integration modules as control actions to output the predicted trajectory distributions, or the single most-likley trajectory, depending on the task.
\subsection{Loss Functions}
We adopt the following objective function for the overall CVAE model:
\begin{equation}
\begin{aligned}
\max_{\phi,\theta,\psi} \sum_{i=1}^{N} & \mathbb{E}_{z \sim q_\phi(\cdot | \mathbf{x}_i,\mathbf{y}_i)} \left[\log p_\psi(\mathbf{y}_i | \mathbf{x}_i,z)\right] \\
& - \beta D_{KL} \left(q_\phi\left(z | \mathbf{x}_i,\mathbf{y}_i\right)||p_\theta\left(z|\mathbf{x}_i\right)\right)+\alpha I_q(\mathbf{x}|z),
\end{aligned}
\end{equation}
where $I_q$ is the mutual information between x and z under the distribution $q_\phi(\mathbf{x} | z)$. We follow the process given in \cite{zhao2019infovae} to compute $I_q$. We approximate $q_\phi\left(z | \mathbf{x}_i,\mathbf{y}_i\right)$ with $p_\theta\left(z|\mathbf{x}_i\right)$, and obtain the unconditioned latent distribution by summing out $\mathbf{x}_i$ over the batch.
\subsection{Human Trajectory Forecasting}
One of the pioneering works of human trajectory forecasting is the Social Force model \cite{helbing1995social}, which applies Newtonian forces to model human motion. Similar methods with strong priors have also been proposed \cite{antonini2006discrete}; yet, most of them rely on hand-crafted energy potential, such as relative distances and rules, to model human motion.
Recently, machine learning methods have been applied to the problem of human trajectory forecasting to obtain models with better performance.
One line of work is to formulate this problem as a deterministic time-series regression problem and then solve it using, e.g., Gaussian Process Regression (GPR) \cite{wang2007gaussian}, inverse reinforcement learning (IRL) \cite{lee2016predicting}, and recurrent neural networks (RNNs) \cite{alahi2016social,jain2016structural,vemula2018social}.
However, the issue of these deterministic regressors is that human behavior is rarely deterministic or unimodal. Hence, generative approaches have become the state-of-the-art trajectory forecasting methods, due to recent advancements in deep generative models \cite{sohn2015learning, goodfellow2014generative} and their ability of generating distributions of potential future trajectories (instead of a single future trajectory). Most of these methods use a recurrent neural network architecture with a latent variable model, such as a conditional variational auto-encoder (CVAE) \cite{salzmann2020trajectron++,lee2017desire,ma2019wasserstein,li2021spatio}, or a generative adversarial network (GAN) \cite{gupta2018social,kosaraju2019social,sadeghian2019sophie,zhao2019multi,li2019conditional} to encode multi-modality. Compared to previous work, we not only consider multi-modality from the perspective of a single agent, but also from the group level; we take into account the phenomenon that people usually move in groups. We show that our group-aware prediction has better understanding of the scenes and achieves better forecasting performance.
\subsection{Graph Convolutional Networks}
Of the methods mentioned above, RNN-based methods have achieved better performance. However, recurrent architectures are parameter inefficient and expensive in training \cite{bai2018empirical}. Besides, to handle spatial context, RNN-based methods need additional structures. Most of them use graph models to encode neighboring pedestrians' information since the topology of graphs is a natural way to represent interactions between pedestrians.
Graph convolutional networks (GCN) introduced in \cite{kipf2016semi} is more suitable for dealing with non-Euclidean data. The Social-BiGAT \cite{kosaraju2019social} introduces a graph attention network \cite{velivckovic2017graph} to model social interactions. GraphSAGE \cite{hamilton2017inductive} aggregates nodes and fuses adjacent nodes in different orders to extract node embeddings. To capture both the spatial and temporal information, Spatio-Temporal Graph Convolutional Networks (STGCN) extends the spatial
GCN to spatio-temporal GCN for skeleton-based action recognition \cite{yan2018spatial}. STGCN is adapted by Social-STGNN \cite{mohamed2020social} for trajectory forecasting, where trajectories are modeled by graphs with edges representing social interactions and weighted by the distances between pedestrians. A development related to our paper is dynamic multi-scale GNN (DMGNN) \cite{li2020dynamic}, which proposes a \textit{multi-scale graph} to model human body relations and extract features at multiple scales for motion prediction. There are two kinds of sub-graphs in the multi-scale graph:
\begin{enumerate*}[label=(\roman*)]
\item single-scale graphs, which connect body components at the same scales, and
\item cross-scale graphs, which form cross-scale connections among body components.
\end{enumerate*}
Based on the multi-scale graphs, a multi-scale graph computational unit is proposed to extract and fuse features across multiple scales. Motivated by this work, we adopt the multi-scale graph strategy for dense crowd forecasting which includes scene-level graphs, group-level graphs, and individual-level graphs.
\subsection{Group-aware Prediction}
People moving in groups (such as friends, family members, etc.) is a common phenomenon and people in each group tend to exhibit similar motion patterns. Motivated by this phenomenon, group-aware methods \cite{rudenko2020human} consider the possibility of human agents being in groups or formations to have more correlated motions than independent ones. They therefore can also model reactions of agents to the moving groups. Human agents can be assigned to different groups by clustering trajectories with similar motion patterns based on methods such as $k$-means clustering \cite{zhong2015learning}, support vector clustering \cite{lawal2016support}, coherent filtering \cite{bisagno2018group}, and spectral clustering methods \cite{atev2010clustering}.
\section{Introduction}
\input{Intro}
\section{Related Work}
\input{RelatedWork}
\section{Grouptron}
\input{Methodology}
\section{Experiments}
\input{Experiments}
\section{Conclusions}
\input{Conclusions}
\bibliographystyle{IEEEtran}
|
\section{Angular Momentum Optomechanical Interaction}
In this section we discuss the optomechanical Hamiltonian. Firstly, we derive the classical Hamiltonian that represents the mutual influence of the mechanical motion and the electromagnetic field in a bounded dielectric medium, which are characterized by the electric field $\mathbf{E}(\mathbf{x})$ and the mechanical displacement field $\mathbf{Q}(\mathbf{x})$, respectively.
For simplify we assume that $\mathbf{Q}(\mathbf{x})=\alpha \mathbf{u}(\mathbf{x})$, where $\mathbf{u}(\mathbf{x})$ is the mode function of the mechanical displacement, which satisfies the condition that $\max|\mathbf{u}|=1$ and $\alpha$ is the amplitude of the mechanical displacement. Usually $\alpha$ is much smaller than the scale of the mechanical oscillator.
By quantizing the classical electromagnetic Hamiltonian, we can derive the optomechanical Hamiltonian.
\begin{figure}[h]
\centering
\includegraphics[width=0.6\linewidth]{permittivity}
\caption{ A dielectric material with permittivity $ \epsilon_{1}$ in a volume $V_{1}$ (blue circle) is surrounded by a medium with permittivity $ \epsilon_{2}$ occupying the complementary volume $V_{2}$ (black circle). The perturbed volume is represented by the dashed red circle and $\mathbf{Q}$ is the corresponding mechanical displacement. $\partial V$ is the boundary between $V_{1}$ and $V_{2}$.}
\label{fig:permittivity}
\end{figure}
We consider a dielectric material with permittivity $\epsilon(\mathbf{x})=\epsilon_{1}$ in a volume $V_{1}$, which is localized in a surrounding medium with permittivity $\epsilon(\mathbf{x})=\epsilon_{2}$ occupying the complementary volume $V_{2}$. For a dielectric material in vacuum $\epsilon_{2}=\epsilon_{0}$. The unperturbed permittivity can be written as
\begin{equation}
\epsilon(\mathbf{x})=\epsilon_{2}+\left(\epsilon_{1}-\epsilon_{2}\right) \Theta(\mathbf{x})
\end{equation}
where $\Theta(\mathbf{x})$ is a step function defined by
\begin{equation}
\Theta(\mathbf{x})=\left\{\begin{array}{lll}
1 & \text { for } & \mathbf{x} \in V_{1} \\
0 & \text { for } & \mathbf{x} \in V_{2}
\end{array}\right.
\end{equation}
The mechanical motion of the dielectric material will influence the total permittivity $\epsilon(\mathbf
{x})$.
We assume that $\epsilon_{1}=\epsilon_{1}(\mathbf{x},\mathbf{Q}(\mathbf{x}))=\epsilon_{1}(\mathbf{x},\alpha)$, which is also dependent on $\mathbf{Q}(\mathbf{x})$. Notice that when $\alpha=0$, $\left.\epsilon_{1}(\mathbf{x},\alpha)\right|_{\alpha=0}=\epsilon_{1}(\mathbf{x},0)$, which is the unperturbed value.
The perturbed permittivity induced by mechanical displacement $\mathbf{Q}$ can be written as
\begin{equation}
\epsilon(\mathbf{x},\alpha)=\epsilon_{2}+\left[\epsilon_{1}(\mathbf{x},\mathbf{Q})-\epsilon_{2}\right] \Theta(\mathbf{x}+\mathbf{\mathbf{Q}})=\epsilon_{2}+\left[\epsilon_{1}(\mathbf{x},\mathbf{\alpha})-\epsilon_{2}\right] \Theta(\mathbf{x}+\alpha\mathbf{u}).
\end{equation}
Expanding $\epsilon(\mathbf{x},\alpha) $ to the first order of $\alpha$, $\epsilon(\mathbf{x},\alpha)$ is
\begin{equation}
\begin{aligned}
\epsilon(\mathbf{x}, \alpha) & \approx \epsilon_{2}+\left[\epsilon_{1}(\mathbf{x}, 0)-\epsilon_{2}+\left.\alpha \frac{\partial \epsilon_{1}(\mathbf{x}, \alpha)}{\partial \alpha}\right|_{\alpha=0}\right]\left[\Theta(\mathbf{x})+\alpha \mathbf{u}(\mathbf{x}) \cdot \nabla \Theta(\mathbf{x})\right] \\
& \approx \epsilon_{2}+\left[\epsilon_{1}(\mathbf{x}, 0)-\epsilon_{2}\right] \Theta(\mathbf{x})+\alpha\left(\epsilon_{1}(\mathbf{x}, 0)-\epsilon_{2}\right) \mathbf{u}(\mathbf{x}) \cdot \nabla \Theta(\mathbf{x})+\alpha \left.\frac{\partial \epsilon_{1}(\mathbf{x}, \alpha)}{\partial \alpha}\right|_{\alpha=0}\Theta(\mathbf{x}) \\
&=\epsilon(\mathbf{x})+\alpha\left(\epsilon_{M B}+\epsilon_{\text{other}}\right)
\end{aligned}
\end{equation}
where $\epsilon_{MB}=\left(\epsilon_{1}(\mathbf{x},0)-\epsilon_{2}\right) \mathbf{u}(\mathbf{x}) \cdot \nabla \Theta(\mathbf{x})$ is the moving boundary effect ($\mathbf{u} \cdot \nabla \Theta(\mathbf{x})$) induced permittivity and $\epsilon_{other}=\left.\dfrac{\partial\epsilon_{1}(\mathbf{x},\alpha)}{\partial\alpha}\right|_{\alpha=0} \Theta(\mathbf{x})$ is other effects induced permittivity. Considering the electromagnetic Hamiltonian $H_{\mathbf{em}}$ is $H_{\mathrm{em}}=1 / 2 \int_{V} d V\left( \mathbf{E}(\mathrm{x}) \cdot \epsilon(\mathrm{x}) \mathbf{E}(\mathrm{x})+ \mathbf{H}(\mathrm{x}) \cdot \mu(\mathrm{x}) \mathbf{H}(\mathrm{x})\right)$,
we can easily derive the optomechanical coupling Hamiltonian $H_{MB}$ and $H_{\text{other}}$,
\begin{equation}
H_{MB}=\frac{1}{2} \alpha \int d V \mathbf{E}(\mathbf{x})^{\dagger} \epsilon_{MB}\mathbf{E}(\mathbf{x}),
\end{equation}
\begin{equation}
H_{\text{other}}=\frac{1}{2} \alpha \int d V \mathbf{E}(\mathbf{x})^{\dagger} \epsilon_{\text{other}}\mathbf{E}(\mathbf{x}).
\end{equation}
In following subsections we will mainly discuss $H_{MB}$ and $H_{\text{other}}$.
\subsection{Moving Boundary Coupling Effect}
If the boundary of an optical resonator is perturbed by the motion of a mechanical resonator, the energy of the total electromagnetic field will be perturbed, which can be understood from the expression $\epsilon_{MB} \propto \mathbf{u}(\mathbf{x}) \cdot \Theta(\mathbf{x}) $. By means of simple mathematical transformations, $H_{MB}$ can be expressed as
\begin{equation}
H_{MB}=\frac{1}{2} \alpha \int_{\partial V} d \mathbf{A} \cdot \mathbf{u}(\mathbf{x})
\left\{-\mathbf{E}_{\|}(\mathbf{x})^{\dagger} \Delta\epsilon \mathbf{E}_{\|}(\mathbf{x}) + \mathbf{D}_{\perp}(\mathbf{x})^{\dagger} \Delta(\epsilon^{-1}) \mathbf{D}_{\perp}(\mathbf{x})\right\}
\end{equation}
where $\Delta \epsilon=\epsilon_{1}-\epsilon_{2}$, $\Delta\left(\epsilon^{-1}\right)=\epsilon_{1}^{-1}-\epsilon_{2}^{-1}$, $\mathbf{E}_{\|}$ is the electric field which is parallel with the surface of $V_{1}$ and $\mathbf{D_{\perp}}$ is the electric displacement field which is perpendicular with the surface of $V_{1}$ . The detailed calculation process can be found in references \cite{PhysRevE.65.066611,PhysRevA.94.053827}.
\subsection{Opto-elastic Coupling Effect and Optical Material Anisotropy Coupling Effect}
In this section we mainly discuss the permittivity $\epsilon_{other}=\left.\dfrac{\partial\epsilon_{1}(\mathbf{Q})}{\partial\alpha}\right|_{\alpha=0} \Theta(\mathbf{x})$. Function $\Theta(\mathbf{x})$ shows that $\epsilon_{\text{other}}$ can only be non-zero in the region of $V_{1}$, which means that Hamiltonian $H_{\text{other}}$ is concentrated in the region of optomechanical resonator.
\subsubsection{Optical Material Anisotropy Coupling Effect}
We assume that the material of the optical resonator is optical anisotropy, when mechanical displacement $\alpha=0$, the unperturbed permittivity $\epsilon_{1}$ can be expressed as a diagonal matrix
\begin{equation}
\epsilon_{1}(\mathbf{x},0)=
\begin{pmatrix}
\epsilon_{xx} & 0 &0 \\
0 & \epsilon_{yy} & 0 \\
0 & 0 & \epsilon_{zz}
\end{pmatrix},
\end{equation}
where $\epsilon_{ii}$ (i=x,y,z) is the permittivity along $i$ axis.
In our main article, the displacement vector $\mathbf{u}(\mathbf{x})$ of the torsional motion is given by $\mathbf{u}(\mathbf{x})=\theta(z) (r/R) \hat{\boldsymbol{\theta}}=$
$\theta(z)(-y \hat{\boldsymbol{x}}+x \hat{\boldsymbol{y}})/R$ in cylindrical and Cartesian coordinates, respectively. Here $R$ makes $\mathbf{u}$ satisfies the condition $\max|\mathbf{u}|=1$ and $\theta(z)$ is the the spatial mode function of torsional motion and satisfies the condition $\max|\theta(z)|=1$.
In lab frame, the rotation of the optical material will influence $\epsilon_{1}$,
\begin{equation}
\epsilon_{1}(\mathbf{x},\alpha)=R(\alpha \theta(z)) \epsilon_{1}(0) R(-\alpha \theta(z)),
\label{rotation per}
\end{equation}
where $R(\alpha)$ is a rotation matrix,
\begin{equation}
R(\alpha)=\left(\begin{array}{ccc}
\cos (\alpha) & -\sin (\alpha) & \\
\sin (\alpha) & \cos (\alpha) & \\
0 & 0 & 1
\end{array}\right).
\end{equation}
If $\alpha \ll 1$, $R(\alpha)$ can be expanded into $R(\alpha)=I+\alpha T$, where $T$ is
\begin{equation}
T=
\begin{pmatrix}
0 &-1 &0 \\
1 & 0& 0\\
0 &0 & 0
\end{pmatrix}.
\end{equation}
Expanding $\epsilon_{1}(\mathbf{x},\alpha)$ to the first order of $\alpha$, we have
\begin{equation}
\epsilon_{1}(\mathbf{x},\alpha) \approx \epsilon_{1}(\mathbf{x},0) + \alpha\theta(z) [T,\epsilon_{1}(\mathbf{x},0)]=\epsilon_{1}(\mathbf{x},0) + \alpha\theta(z) \delta \epsilon A
\end{equation}
where $\delta \epsilon=\epsilon_{x x}-\epsilon_{y y}$ and $A=\left(\begin{array}{lll}0 & 1 & 0 \\ 1 & 0 & 0 \\ 0 & 0 & 0\end{array}\right)$.
\subsubsection{Opto-elastic Coupling Effect}
The refractive index of most material will be influenced by opto-elastic effect when material is subject to strain. Mathematically, the opto-elastic effect is described by relating the change of the permittivity tensor $\delta \eta,$ to the material strain by Pockel's tensor $p$ \cite{PhysRevB.3.2778},
$
(\boldsymbol{\delta} \boldsymbol{\eta})_{i j}=\boldsymbol{p}_{i j k l} s_{k l}
$,
where $s_{mn}$ is the strain tensor of the material with $s_{m n}=\frac{1}{2}\left(\frac{\partial \boldsymbol{u}_{m}}{\partial \boldsymbol{x}_{n}}+\frac{\partial \boldsymbol{u}_{n}}{\partial \boldsymbol{x}_{m}}\right)$.
The modified permittivity tensor follows from $\epsilon^{OE}=\left(\epsilon^{-1}+\alpha\delta \eta\right)^{-1}$, where $\epsilon$ is the unmodified permittivity tensor.
Considering equation (\ref{rotation per}), the final expression of $\epsilon_{1}^{OE}(\mathbf{x},\alpha)$ is
\begin{equation}
\begin{split}
\epsilon_{1}^{OE}(\mathbf{x},\alpha)&=( \epsilon_{1}(\mathbf{x},\alpha)^{-1} + \alpha\delta \eta)^{-1}\\
&=\left( R(\alpha \theta) \epsilon_{1}^{-1}(\mathbf{x},0) R(-\alpha \theta) + \alpha \delta \eta\right)^{-1} \\
&\approx\left( (I+\alpha \theta T) \epsilon_{1}^{-1}(\mathbf{x},0) (I-\alpha \theta T) + \alpha \delta \eta\right)^{-1} \\
& \approx \left( \epsilon_{1}^{-1}(\mathbf{x},0)+ \alpha \theta (T \epsilon_{1}^{-1}(\mathbf{x},0) - \epsilon_{1}^{-1}(\mathbf{x},0) T) +\alpha \delta \eta \right)^{-1} \\
& \approx \epsilon_{1}(\mathbf{x},0) +\alpha \theta \delta\epsilon A-\alpha \epsilon_{1}(\mathbf{x},0) \delta \eta \epsilon_{1}(\mathbf{x},0).
\end{split}
\end{equation}
\subsubsection{Expression of $\epsilon_{\text{other}}$}
The final expression of $\epsilon_{other}$ is
\begin{equation}
\begin{split}
\epsilon_{\text{other}}&=\dfrac{\partial\epsilon_{1}(\mathbf{Q})}{\partial\alpha} \Theta(\mathbf{x})\\
&= \theta [T,\epsilon_{1}(\mathbf{x},0)]\Theta(\mathbf{x})- \epsilon_{1}(\mathbf{x},0) \delta \eta \epsilon_{1}(\mathbf{x},0)\Theta(\mathbf{x})\\
&=\epsilon_{MA}+\epsilon_{OE}.
\end{split}
\end{equation}
The first term $\epsilon_{MA}=\theta [T,\epsilon_{1}(\mathbf{x},0)]\Theta(\mathbf{x})=\theta \delta \epsilon A\Theta(\mathbf{x}) $ is related with the optical anisotropy of the material and the second term $\epsilon_{OE}=- \epsilon_{1}(\mathbf{x},0) \delta \eta \epsilon_{1}(\mathbf{x},0)\Theta(\mathbf{x})$ is related with opto-elastic effect. Therefore, the optomechanical Hamiltonian $H_{\text{other}}$ can be divided into two terms $H_{AM}$ and $H_{OE}$,
\begin{equation}
H_{OE}=-\frac{1}{2} \alpha \int_{V_{1}} d V \mathbf{E}(\mathbf{x})^{\dagger} \epsilon_{1}(\mathbf{x},0) \delta \eta \epsilon_{1}(\mathbf{x},0)\mathbf{E}(\mathbf{x}),
\end{equation}
\begin{equation}
H_{MA}=\frac{1}{2} \alpha \delta\epsilon \int_{V_{1}} d V \theta(z) \mathbf{E}(\mathbf{x})^{\dagger} A \mathbf{E}(\mathbf{x}).
\end{equation}
\subsection{Quantization of The Optomechanical Hamiltonian }
The quantized optomechanical Hamiltonian is obtained from
the classical one by replacing the classical amplitude $\alpha$ and the
electromagnetic fields by operators.
We define $x_{zp}=\sqrt{\hbar/(2m_{\mathrm{eff}}\Omega)}$ is the zero point linear displacement of the torsional oscillator, where $m_{\mathrm{eff}}=\int_{V}|\mathbf{u}(\mathbf{x})|^{2}\rho dV$ is the effective mass, and $\theta_{zp}=\sqrt{\hbar/(2I_{\mathrm{eff}}\Omega)}$ is the effective zero point angular displacement of the torsional oscillator and $I_{\mathrm{eff}}$ is the effective moment of inertia of the torsional oscillator.
For a two mode electric field $\mathbf{E}=\mathbf
E_{1} a_{1}+\mathbf{E}_{2} a_{2}+h.c.$ ($\vec{E}_{1}$ and $\vec{E}_{2}$ are not normalized), the original optomechanical Hamiltonian $H_{\text{int}}$ is
\begin{equation}
H_{\text{int}}=H_{MB}+H_{OE}+H_{MA}=\hbar(b+b^{\dagger})\sum_{i,j=1,2}a_{i}^{\dagger}a_{j} g_{ij}
\end{equation}
where $g_{ij}=g_{ijMB}+g_{ijOE}+g_{ijAM}$ and
\begin{equation}
\begin{aligned}
g_{ijMB}&=\frac{x_{zp}}{2} \int_{\partial V} d \mathbf{A} \cdot \mathbf{u}(\mathbf{x})
\left\{-\mathbf{E}_{i\|}(\mathbf{x})^{\dagger} \Delta\epsilon \mathbf{E}_{j\|}(\mathbf{x}) + \mathbf{D}_{i\perp}(\mathbf{x})^{\dagger} \Delta(\epsilon^{-1}) \mathbf{D}_{j\perp}(\mathbf{x})\right\},
\\g_{ijOE}&=-\frac{x_{zp}}{2} \int_{V_{1}} d V \mathbf{E}_{i}(\mathbf{x})^{\dagger} \epsilon_{1}(\mathbf{x},0) \delta \eta \epsilon_{1}(\mathbf{x},0)\mathbf{E}_{j}(\mathbf{x}),\\
g_{ijMA}&=\dfrac{1}{2}\theta_{zp}\delta\epsilon\int_{V_{1}}dV\theta(z)\left(E_{i y}(\mathbf{x}) E_{j x}^{*}(\mathbf{x})+E_{j y}^{*}(\mathbf{x}) E_{i x}(\mathbf{x})\right).
\label{gOE}
\end{aligned}
\end{equation}
Here we use "MB" to represent moving-boundary coupling effect, use "OE" to represent opto-elastic coupling effect and use "MA" to represent material anisotropy coupling effect.
\subsection{Discussion About the Torsional Oscillator}
The exact values of mode function $\theta(z)$ and resonance frequency $\Omega$ depends on the detailed geometrical structure of the torsional resonator, here we use a method to estimate the rough profile of $\theta(z)$.
Considering a suspended slender beam with two ends fixed, the corresponding wave equation for a cylinder with varying cross-section is known as Webster-type equation and takes the form \cite{PhysRevA.88.061801}
\begin{equation}
c_{t}^{-2} \partial_{t}^{2} \phi(t, z)-\partial_{z}^{2} \phi(t, z)-\left(\frac{\partial_{z} I_{p}(z)}{I_{p}(z)}\right) \partial_{z} \phi(t, z)=0
\label{Web}
\end{equation}
where $I_{p}(z)=\int_{A} r^{2} d A$ is the polar moment of inertia at the cross section $A(z)$ and $c_{t}$ is the phase velocity of the torsional wave.
To solve the wave equation, it is convenient to first separate it in time and space and construct solutions of the form $\phi(t, z)=\theta(z) \cos (\Omega t)$, so that only the differential equation in the axial coordinate remains.
In the simplest case
where the torsional oscillator is a homogeneous cylinder, we have $\partial_{z} I_{p}(z)=0$. Then the wave equation (\ref{Web}) will be reduced to its simplest form
\begin{equation}
\partial_{z}^{2} \theta( z) +k_{t}^{2} \theta( z)=0,
\end{equation}
with the well-known solutions:
\begin{equation}
\theta_{\mathrm{s}}(z)=\sin \left(k_{t} z\right), \quad \theta_{\mathrm{c}}(z)=\cos \left(k_{t} z\right), \quad \text { where } k_{t}=\Omega / c_{t}.
\end{equation}
At the ends of the torsional oscillator, $\theta(z)=0$, and we can derive the
resonance frequency $\Omega$ from the corresponding boundary condition.
However, the torsional oscillator also includes other structures besides the square suspended beam and the mode function $\theta(z)$ will be more complex than sine function or cosine function, and the effective moment of inertia $I_{\mathrm{eff}}$ will also larger than $I_{\mathrm{beam}}=
1 / 6 M a^{2} \int_{-L/2}^{L/2} \theta(z)^{2} d z= 1 / 12 \rho L a^{4}$. As a simple example, we consider
the fundamental mode of a suspended beam. The eigen wave vector $k_{t}$ of the sine mode $\sin(k_{t}z)$ is $2\pi n/L $ with boundary condition $\theta_{s}(L/2)=0$, where $n$ is an integer, and the wave vector $k_{t}$ of the cosine function mode $\cos(k_{t}z)$ is $2\pi (n+1/2)/L $ with boundary condition $\theta_{c}(L/2)=0$.
Therefore, the fundamental mode function $\theta(z)$ of the torsional oscillator should be an even function and the piece-wise mode function of the square beam should be a cosine function $\cos(k_{t}z)$.
In our calculation, we assume $\rho=2650\mathrm{kg/m^{3}}$, $c_{t}=5000\mathrm{m/s}$ \cite{RevModPhys.74.991}, $L=100 \mathrm{\mu m}$, $a=1\mathrm{\mu m}$, $\Omega=2\pi\times 500\mathrm{kHz}$, $I_{eff}=10\times 1 / 6 \rho a^{4} \int_{-L/2}^{L/2} \theta(z)^{2} d z= 10 / 12 \rho L a^{4}=4.4163\times 10^{-25} \mathrm{ kg\cdot m^{2}}$ and $k_{t}=\Omega/c_{t}=100\pi/\mathrm{m}$. The parameters we will use is listed in Table (\ref{para}).
\begin{table}[h]
\caption{Parameters used in calculation}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline\hline
$a/\mathrm{\mu m }$ & $L/\mathrm{\mu m }$ & $\rho/\mathrm{kg\cdot m^{-3} }$ & $\Omega/\mathrm{kHz}$&$c_{t}/\mathrm{km\cdot s^{-1}}$&$I_{\text{eff}}/\mathrm{kg\cdot m^{2}}$ &$k_{t}/\mathrm{m^{-1}}$ & $\beta_{1}/k$& $\beta_{2}/k$&$\epsilon_{xx} $ & $ \epsilon_{yy}$ & $\epsilon_{zz} $ \\ \hline
1& 100 & 2650&$2\pi \times 500$ & 5.0&$4.4163\times 10^{-25}$ &100$\pi$ & 1.2859 & 1.2926 &1.5326&1.5277 &1.5277 \\ \hline\hline
\end{tabular}
\label{para}
\end{table}
\subsection{Calculation of the Coupling Constants}
In this section we discuss the exact values of the coupling constants. When wavelength $\lambda=1550$nm, the relative permittivity tensor of $\alpha$-quartz are \cite{GHOS199995}
\begin{equation}
\epsilon(0)=\begin{pmatrix}
1.5326^{2}& 0 &0 \\
0& 1.5277^{2} & 0\\
0&0 & 1.5277^{2}
\end{pmatrix} ,
\end{equation}
and the corresponding mode profiles of the electric field $\mathbf{E}_{1}$ and the electric field $\mathbf{E}_{2}$ are plotted in Fig. (\ref{fig:em-field}).
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\linewidth]{"emfield"}
\caption{(a), (b) and (c) are the electric fields distribution of the TE-like mode. (d), (e) and (f) are the electric field distribution of the TM-like mode. It is obvious that $|\mathbf{E}_{i}|^{2}$ (i=x,y,z) is symmetric about $x$ axis and $y$ axis. }
\label{fig:em-field}
\end{figure}
\paragraph{$g_{11MB}\approx 0$ and $g_{22MB}\approx 0$} Because of the symmetry of the optical resonator, $|\mathbf{E}_{i\|}(\mathbf{x})|^{2}$ and $|\mathbf{D}_{i\perp}(\mathbf{x})|^{2}$ (i=x,y,z) are even functions about $x$ axis and $y$ axis simultaneously. However, torsional mode function $\mathbf{u}(\mathbf{x})$ is a odd function about $x$ axis and $y$ axis. Therefore, the integration $g_{iMB}$ will approach $0$.
\paragraph{$g_{12MB}\approx0 $} Form Fig. (\ref{fig:em-field}) we know that $\mathbf{E}_{1}(\mathbf{x})$ is a quasi-$x$ polarized electric field with $|E_{1y}| \ll |E_{1x}|$ and $\mathbf{E}_{2}(\mathbf{x})$ is a quasi-$y$ polarized electric field with $|E_{2x}| \ll |E_{2y}|$. In addition, $\mathbf{E}_{1\|}(\mathbf{x})^{\dagger} \Delta\epsilon \mathbf{E}_{2\|}(\mathbf{x})$ usually can be expanded into the summation of $E_{1m}(\mathbf{x})^{\dagger} \Delta\epsilon E_{2m}(\mathbf{x})$ ($m=x,y,z$). Considering the normalized values of $\mathbf{E}(\mathbf{x})$ and $\mathbf{E}(\mathbf{x})$,
\begin{equation}
g_{12MB} \propto \sum_{m=x,y,z} \frac{\int\left(- E_{1 m}^{*}\Delta \epsilon E_{2 m}+ D_{1 m}^{*}\Delta (\epsilon^{-1}) D_{2 m}\right)dA}{\sqrt{\int \mathbf{E}_{1} \cdot \epsilon \mathbf{E}_{1} d x d y} \cdot \sqrt{\int \mathbf{E}_{2} \cdot \epsilon_{2} \mathbf{E}_{2} d x d y}},
\end{equation}
Thus the value of $g_{MB}$ will approach $0$.
\paragraph{$g_{ijOE} \approx 0$} The exact values of $g_{ijOE}$ depends on the strain tensor $s_{mn}$, electric field $\mathbf{E}_{1}(\mathbf{x})$ and electric field $\mathbf{E}_{2}(\mathbf{x})$. When the mode function $\theta(z)$ of the torsional oscillator is a cosine function $\theta(z)=\cos(k_{t}z)$, the strain tensor $s_{mn}$ is
\begin{equation}
\mathbf{s}=\dfrac{1}{2}\sin(k_{t}z)\dfrac{k_{t}}{R}
\begin{pmatrix}
0 & 0& y \\
0 & 0 & -x \\
y & - x & 0
\end{pmatrix} .
\label{strain}
\end{equation}
In our main article, we assume that $\mathbf{E}_{i}(\mathbf{x})=\mathbf{E}_{i}(x,y)\cos(\beta_{i}z)$. Substituting Equ. (\ref{strain}) into Equ. (\ref{gOE}), we can easily derive the conclusion that
\begin{equation}
g_{ijOE} \propto \int_{-L/2}^{L/2} \sin(k_{t}z)\cos(\beta_{i}z)\cos(\beta_{j}z) dz=0.
\end{equation}
We need to note that we do not need the exact value of the Pockel's tensor p.
\paragraph{$g_{11MA} \approx 0$ and $g_{22MA} \approx 0$} The reason that $g_{iiAM} \approx 0$ is similar with the reason that $g_{12MB}=0$.
\paragraph{$g_{12MA} \neq 0$} Considering the normalized condition of $\mathbf{E}_{1}(\mathbf{x})$ and $\mathbf{E}_{2}(\mathbf{x})$, we can express $g_{12MA}$ as
\begin{equation}
\begin{aligned}
g_{12MA} &=\dfrac{\theta_{zp}}{2}\int_{V_{1}}dxdy\left(E_{1 y}(x,y) E_{2 x}^{*}(x,y)+E_{2 y}^{*}(x,y) E_{1 x}(x,y)\right) \times \int_{-L/2}^{L/2} \theta(z)\cos(\beta_{i}z)\cos(\beta_{i}z) dz \\
&=\frac{\theta_{zp}}{2}\left(\epsilon_{xx}-\epsilon_{yy}\right) \sqrt{\omega_{1} \omega_{2}}
\frac{\int\left(E_{2 x} E_{1 y}^{*}+E_{2 y} E_{1 x}^{*}\right) d x d y}{\sqrt{\int \mathbf{E}_{1} \cdot \epsilon \mathbf{E}_{1} d x d y} \cdot \sqrt{\int \mathbf{E}_{2} \cdot \epsilon_{2} \mathbf{E}_{2} d x d y}}
\times \frac{\int_{-L/2}^{L/2} \cos(k_{t}z) \cos \left(\beta_{1} z\right) \cos \left(\beta_{2} z\right) d z}{\sqrt{\int_{-L/2}^{L/2} \cos \left(\beta_{1} z\right)^{2} d z} \cdot \sqrt{\int_{-L/2}^{L/2} \cos \left(\beta_{2} z\right)^{2} d z}} \\
& \propto \theta_{z p} \sqrt{\omega_{1} \omega_{2}}\left(\frac{\delta \epsilon}{L}\right) \int_{-L / 2}^{L / 2} \theta(z) \cos \left(\beta_{1} z\right) \cos \left(\beta_{2} z\right) d z.
\end{aligned}
\label{g12AM}
\end{equation}
In the last step of Equ. (\ref{g12AM}), we use the approximation that $\int_{-L/2}^{L/2}\cos(\beta_{1} z)^{2}dz \approx \int_{-L/2}^{L/2}\cos(\beta_{2} z)^{2}dz \approx L/2$.
\paragraph{Discussion about $g_{12AM}$}
If $\theta(z) \propto \cos(k_{t}z)$, then the expression of $g_{12MA}$ will be
\begin{equation}
\begin{split}
g_{12AM} & \propto \dfrac{1}{L}\int_{-L / 2}^{L / 2} \cos(k_{t}z) \cos \left(\beta_{1} z\right) \cos \left(\beta_{2} z\right) d z \\&
= -\frac{\sin \left(\frac{1}{2} L\left(-\beta_{1}-\beta_{2}+k_{t}\right)\right)}{2L(\beta_{1}+\beta_{2}-k_{t})}+\frac{\sin \left(\frac{1}{2} L\left(\beta_{1}-\beta_{2}+k_{t}\right)\right)}{2L(\beta_{1}-\beta_{2}+k_{t})}+\frac{\sin \left(\frac{1}{2} L\left(-\beta_{1}+\beta_{2}+k_{t}\right)\right)}{2L(-\beta_{1}+\beta_{2}+k_{t})}+\frac{\sin \left(\frac{1}{2} L\left(\beta_{1}+\beta_{2}+k_{t}\right)\right)}{2L(\beta_{1}+\beta_{2}+k_{t})}.
\end{split}
\end{equation}
In this system, $\beta_{1}\sim 2\pi/\lambda$ and $\beta_{2}\sim 2\pi/\lambda$ where $\lambda$ is the wavelength of the electromagnetic field. If $L \geq 0.1 \mathrm{mm}$, $\beta_{1}L \gg 1$ and $\beta_{2}L \gg 1$. When $\beta_{1}-\beta_{2}=k_{t} \gg L^{-1}$, $\frac{\sin \left(\frac{1}{2} L\left(\beta_{1}-\beta_{2}+k_{t}\right)\right)}{(\beta_{1}-\beta_{2}+k_{t})L}=\dfrac{\sin(k_{t}L)}{2k_{t}L} \ll 1$ and $\frac{\sin \left(\frac{1}{2} L\left(\beta_{1}-\beta_{2}-k_{t}\right)\right)}{2(\beta_{1}-\beta_{2}-k_{t})L}=\dfrac{1}{4}$. Notice that $\theta_{zp}=\sqrt{\hbar/(2I_{\text{eff}}\Omega)}\propto L^{-1/2}$, we have $g\approx g_{12MA} \propto \delta \epsilon \theta_{z p} \propto \delta \epsilon L^{-1 / 2}$.
Using the parameters listed in Table.\ref{para}, the exact values of g is listed in the following table:
\begin{table}[h]
\caption{Table about $g$}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline\hline
$g_{12AM}/\mathrm{kHz }$ & $g_{11AM}/\mathrm{kHz }$ & $g_{22AM}/\mathrm{kHz }$ & $g_{12MB}/\mathrm{kHz }$ & $g_{11MB}/\mathrm{kHz }$ & $g_{22MB}/\mathrm{kHz }$ & $g_{ijOE}/\mathrm{kHz }$ \\ \hline
22& 0 & 0 & 0.081 & -0.01 & -0.01 &0 \\ \hline\hline
\end{tabular}
\label{g}
\end{table}
\subsection{A Simple Explanation For Our Optomechanical Hamiltonian}
Considering the energy of the torsional oscillator and neglecting the optomechanical interaction terms $a_{1}^{\dagger}a_{1}(b+b^{\dagger})$ and $a_{2}^{\dagger}a_{2}(b+b^{\dagger})$, the Hamiltonian of this system can be expressed as
\begin{equation}
H_{\mathrm{OM}} =\hbar \Omega b^{\dagger} b+\hbar \omega_{1} a_{1}^{\dagger} a_{1}+\hbar \omega_{2} a_{2}^{\dagger} a_{2}
+\hbar g\left(b+b^{\dagger}\right) a_{1}^{\dagger} a_{2}+\hbar g^{*}\left(b+b^{\dagger}\right) a_{2}^{\dagger} a_{1}
\label{HOM}
\end{equation}
where $g=g_{12MA}+g_{12OE}+g_{12MB} \approx g_{12AM}$.
\begin{figure}
\centering
\includegraphics[width=0.75\linewidth]{example}
\caption{A simple example for our optomechanical model. (a) An extremely thin anisotropic membrane interacts with the cavity photons. Under the influencing the cavity photons, the membrane will experience a mechanical torque $\tau_{\text{mech}}$. (b) We stack several optical anisotropic dielectric membranes in this cavity. The total mechanical torque $\tau_{\text{mech}}$ will be similar with the mechanical torque $\tau_{\text{mech}}$ in (a).}
\label{fig:example}
\end{figure}
In order to understand this optomechanical interaction further, we consider a more physical and simpler example (see Fig. \ref{fig:example}(a)): we use a laser beam with two orthogonal linear polarization modes $a_{x}$ and $a_{y}$ to interact with an optical anisotropic membrane with width $d \ll \lambda$ and then calculate the optomechanical Hamiltonian. According to the angular momentum conservation law \cite{novotny2012principles},
along the direction of the optical axis $z$, the mechanical torque $\mathbf{\tau}_{\mathrm{mech}}$ experienced by this membrane is
\begin{equation}
\mathbf{\tau}_{\mathrm{mech}}=
-\mathbf{\tau}_{\mathrm{field}}
-(\int_{\partial V} \stackrel{\leftrightarrow}{\mathbf{M}}(\mathbf{r}, t) \cdot \mathbf{n}\mathrm{d}S)_{z},
\end{equation}
where $\mathbf{\tau}_{\mathrm{field}}=\left(\frac{\mathrm{d}}{\mathrm{d} t} \frac{1}{c^{2}} \int_{V}\mathbf{r}\times(\mathbf{E} \times \mathbf{H}) \mathrm{d} V\right)_{z}$ is the electromagnetic torque, $\stackrel{\leftrightarrow}{\mathbf{M}}$ is optical angular momentum flux density and the integral $\int_{\partial V} \stackrel{\leftrightarrow}{\mathbf{M}}(\mathbf{r}, t) \cdot \mathbf{n}\mathrm{d} S$ is over the surface the membrane. In order to simple calculation, we define $\mathcal{M}_{zz}=\left(\int_{\partial V} \stackrel{\leftrightarrow}{\mathbf{M}}(\mathbf{r}, t) \cdot \mathbf{n}\mathrm{d} S\right)_{z}$. Following the calculation, we can divide $\mathcal{M}_{zz}$ into spin and orbit prats \cite{Barnett_2002},
\begin{equation}
\mathcal{M}_{z z}=\mathcal{M}_{z z}^{\text {spin }}+ \mathcal{M}_{z z}^{\text {orbit }}
\end{equation}
where
\begin{equation}
\mathcal{M}_{z z}^{\mathrm{spin}}=\frac{\epsilon_{0} c^{2}}{2 \omega} \operatorname{Re}\left[-\mathrm{i} \iint \rho \mathrm{d} \rho \mathrm{d} \phi\left(\mathcal{E}_{x} \mathcal{B}_{x}^{*}+\mathcal{E}_{y} \mathcal{B}_{y}^{*}\right)\right]
\end{equation}
and
\begin{equation}
\mathcal{M}_{z z}^{\text {orbit }}=\frac{\epsilon_{0} c^{2}}{4 \omega} \operatorname{Re}\left[-\mathrm{i} \iint \rho \mathrm{d} \rho \mathrm{d} \theta(-\mathcal{B}_{x}^{*} \frac{\partial}{\partial \theta} \mathcal{E}_{y}+\mathcal{E}_{y} \frac{\partial}{\partial \theta} \mathcal{B}_{x}^{*}
-\mathcal{E}_{x} \frac{\partial}{\partial \theta} \mathcal{B}_{y}^{*}+\mathcal{B}_{y}^{*} \frac{\partial}{\partial \theta} \mathcal{E}_{x})\right].
\end{equation}
Under paraxial approximation, the corresponding E.M. fields are
\begin{equation}
\mathcal{E}_{x}=F_{m}(\rho)e^{im\theta}T_{x}(z)\hat{a}_{x},
\quad
\mathcal{E}_{y}=F_{m}(\rho)e^{im\theta}T_{y}(z)\hat{a}_{y},
\end{equation}
\begin{equation}
\mathcal{B}_{x}=-\dfrac{F_{m}(\rho)e^{im\theta}}{i\omega}\dfrac{\partial{T_{y}(z)}}{\partial{z}}\hat{a}_{y},
\quad
\mathcal{B}_{y}=\dfrac{F_{m}(\rho)e^{im\theta}}{i\omega}\dfrac{\partial{T_{x}(z)}}{\partial{z}}\hat{a}_{x},
\end{equation}
where $m$ is quantum number of the orbital angular momentum, $F_{m}(\rho)$ is the radial mode function and $T_{x}(z)$, $T_{y}(z)$ are the longitudinal mode functions.
Without loss of generality, we assume that $\iint d\rho d\phi|F(\rho)|^{2}=1$. It is worth to notice that $\mathcal{B}_{x}$ is related with operator $\hat{a}_{y}$ and $\mathcal{B}_{y}$ is related with operator $\hat{a}_{x}$. Using these equations, we can derive the operator of optical spin angular momentum flux $\mathcal{M}_{zz}^{\text{spin}}$:
\begin{equation}
\begin{aligned}
\mathcal{M}_{z z}^{\mathrm{spin}}&=\frac{\epsilon_{0} c^{2}}{2 \omega} \operatorname{Re}\left[-\mathrm{i} \iint \rho \mathrm{d} \rho \mathrm{d} \theta\left(\mathcal{E}_{x} \mathcal{B}_{x}^{*}+\mathcal{E}_{y} \mathcal{B}_{y}^{*}\right)\right]\\
&=\frac{\epsilon_{0} c^{2}}{4 \omega^{2}} \left[
T_{y}(z)\dfrac{dT_{x}^{*}(z)}{dz}\hat{a}_{x}^{\dagger}\hat{a}_{y}-T_{x}(z)\dfrac{dT_{y}^{*}(z)}{dz}\hat{a}_{y}^{\dagger}\hat{a}_{x}
\right]+ h.c.
\label{spin}
\end{aligned}
\end{equation}
Similarly, the optical orbital angular momentum flux $\mathcal{M}_{zz}^{\text{orbit}}$ is
\begin{equation}
\begin{aligned}
\mathcal{M}_{zz}^{\text{orbit}}&=m\frac{\epsilon_{0} c^{2}}{2 \omega}\operatorname{Re}\left[
\iint\rho d\rho d\theta (-\mathcal{B}_{x}^{*}\mathcal{E}_{y}+\mathcal{E}_{x}\mathcal{B}_{y}^{*})
\right]\\
&=m\frac{\epsilon_{0} c^{2}}{2 \omega^{2}}
\operatorname{Re}\left[i\dfrac{dT_{y}^{*}(z)}{dz}T_{y}(z)\right]\hat{a}_{y}^{\dagger}\hat{a}_{y}+
m\frac{\epsilon_{0} c^{2}}{2 \omega^{2}}
\operatorname{Re}\left[i\dfrac{dT_{x}^{*}(z)}{dz}T_{x}(z)\right]\hat{a}_{x}^{\dagger}\hat{a}_{x}.
\end{aligned}
\label{orbit}
\end{equation}
Notice that spin angular momentum flux $\mathcal{M}_{zz}^{\text{spin}}$ involves the exchange interaction between two modes $\hat{a}_{x}$ and $\hat{a}_{y}$, while orbital angular momentum flux $\mathcal{M}_{zz}^{\text{orbit}}$ is proportional to photonic energy $\hat{a}_{x}^{\dagger}\hat{a}_{x}$ and $\hat{a}_{y}^{\dagger}\hat{a}_{y}$.
If the input beam does not carry orbital angular momentum, which means $m=0$, we can easily prove that $\mathcal{M}_{zz}^{\text{orbit}}=0$ but $\mathcal{M}_{zz}^{\text{spin}}$ is not influenced.
If the input beam is a single frequency E.M. field, we can easily prove that $\tau_{\mathrm{field}}=0$. Then considering the condition $d\ll \lambda$, the mechanical torque $\tau_{z}$ is
\begin{equation}
\iint dxdy( M_{zz}(x,y,0^{-}) -M_{zz}(x,y,0^{+}) ) =\tau_{\mathrm{mech}}.
\end{equation}
Here we assume the
dielectric membrane is in $z=0$.
The corresponding angular momentum optomechanical Hamiltonian is
\begin{equation}
H_{\mathrm{OM}}=-\hat{\tau}_{\mathrm{mech}}\hat{\theta}=-\hat{\tau}_{\mathrm{mech}}^{\mathrm{spin}}\hat{\theta}-\hat{\tau}_{\mathrm{mech}}^{\mathrm{orbit}}\hat{\theta}=H_{\text{OM}}^{\text{Spin}}+H_{\text{OM}}^{\text{Orbit}}
\label{OM H},
\end{equation}
where $H_{\text{OM}}^{\text{Spin}}=-\hat{\tau}_{\mathrm{mech}}^{\mathrm{spin}}\hat{\theta}$ and $H_{\text{OM}}^{\text{Orbit}}=-\hat{\tau}_{\mathrm{mech}}^{\mathrm{Orbit}}\hat{\theta}$. Considering Equ.(\ref{spin}) and Equ.(\ref{orbit}), we know that
\begin{equation}
H_{\text{OM}}^{\text{Spin}} \propto (g_{\text{spin}}\hat{a}_{x}^{\dagger}\hat{a}_{y} + g_{\text{spin}}^{*}\hat{a}_{y}^{\dagger}\hat{a}_{x})\hat{\theta},\quad H_{\text{OM}}^{\text{Orbit}} \propto m( g_{x}\hat{a}_{x}^{\dagger}\hat{a}_{x} +g_{y}\hat{a}_{y}^{\dagger}\hat{a}_{y} )\hat{\theta}.
\end{equation}
Essentially, our optomechanical system can be viewed as the stacking of several optical anisotropic dielectric membranes (Fig. \ref{fig:example}(b)). Therefore, the optomechanical Hamiltonian of our system also has the form $H_{\text{OM}}\propto (g a_{x}^{\dagger}a_{y} + h.c.)\hat{\theta}=(g a_{x}^{\dagger}a_{y} + h.c.)(b+b^{\dagger})$.
\section{Discussion about input-output relation}
\subsection{Mechanical Mode }
When the mechanical frequency $\Omega$ is much smaller than the optical pulse
bandwidth, $\Omega \ll \tau^{-1}$, the mechanical oscillator Hamiltonian $H_{M}=\hbar \Omega b^{\dagger}b$ and mechanical damping characterized by the mechanical damping rate $\gamma=\Omega/Q_{t}$ can be neglect in the time scale of a single optical pulse.
\subsection{Optical Mode $a_{2}$}
If we use the original form of $H_{OM}$ (\ref{HOM}), we can derive the Langevin equation of the optical mode $a_{2}$,
\begin{equation}
\Dot{a_{2}}=-\dfrac{\kappa}{2}a_{2}-\sqrt{2}g \theta_{M}a_{1}+\sqrt{\kappa}a_{2in}(t)
\label{a2 exa}
\end{equation}
where $\kappa$ is the decaying rate of the cavity. If the input mode $a_{2}(t)$ is a intense coherent laser pulse with duration $\tau$, we can neglect weak coupling term $\sqrt{2}g\theta_{M}a_{1}$. In our following discussion, we only consider the classical amplitude of mode $a_{2}$ and we define that $\langle a_{2} \rangle=\alpha_{2}$. When $\kappa \gg \tau^{-1}$, we can eliminate the dynamics of the intra-cavity field $a_{2}$ adiabatically, yielding a simple relation between the external and internal field amplitudes:
\begin{equation}
\frac{d \alpha_{2}(t)}{d t}\approx-\dfrac{\kappa}{2} \alpha_{2}(t)+\sqrt{\kappa} \alpha_{2\text{in}}(t) \approx 0 \quad \Rightarrow \quad \alpha_{2}(t)=\dfrac{2}{\sqrt{\kappa}} \alpha_{2\text{in}}(t).
\label{a2}
\end{equation}
We assume the temporal function of $a_{2\text{in}}(t)$ is $f(t)$ with an envelop of duration $\tau$. $f(t)$ satisfies the normalized condition $\int_{-\infty}^{\infty}f(t)^{2}dt=1$.
If the photon number of a single pulse is $N_{\text{in}}$, then $\alpha_{2\text{in}}$ becomes $\alpha_{2\text{in}}(t)=\sqrt{N_{\text{in}}}f(t)$. It is easy to verify that $\int_{-\infty}^{\infty}|\alpha_{2\text{in}}(t)|^{2}dt=N_{\text{in}}$, which shows the photon number of single pulse is $N_{\text{in}}$. At this case, $\alpha_{2}(t)$ can be exprerssed as
\begin{equation}
\alpha_{2 }(t)=\dfrac{2}{\sqrt{\kappa}} \alpha_{2\text{in}}(t)=2\sqrt{\dfrac{N_{\text{in}}}{\kappa}}f(t)=\alpha f(t),
\label{alpha2}
\end{equation}
where $\alpha=2\sqrt{\dfrac{N_{\text{in}}}{\kappa}}$ is the amplitude of $\alpha_{2 }(t)$.
\subsection{Input-Output Relation}
Neglecting phonon energy $H_{\mathrm{m}}=\hbar\Omega b^{\dagger}b$ and using $\alpha_{2}(t)$ to replace operator $a_{2}$, we can describe
the dynamics of the optical and mechanical quadrature operators by the set
of the following Langevin equations:
\begin{equation}
\begin{aligned}
\frac{d x_{L}(t)}{d t} &=-\dfrac{\kappa}{2} x_{L}(t)+\sqrt{\kappa} x_{L}^{\text{in}}(t), \\
\frac{d p_{L}(t)}{d t} &=-2 g\alpha_{2}(t) \theta_{M}-\dfrac{\kappa}{2} p_{L}(t)+\sqrt{\kappa} p_{L}^{\text{in}}(t) ,\\
\frac{d \theta_{M}}{d t} &=0, \\
\frac{d L_{M}}{d t} &=-2 g\alpha_{2}(t) x_{L}(t)
\end{aligned}
\end{equation}
where $x_{L}^{\text{in}}(t)$ and $p_{L}^{\text{in}}(t)$ are the quadratures of input noise, which satisfy the commutation relation $\left[x_{L}^{\text{in}}(t), p_{L}^{\text{in}}\left(t^{\prime}\right)\right]=i \delta\left(t-t^{\prime}\right)$.
Invoking adiabatic elimination of the optical intra-cavity field $\frac{d x_{L}(t)}{d t} \approx 0$ and $\frac{d p_{L}(t)}{d t} \approx 0$, we will have the following equations
\begin{equation}
x_{L}(t)=\dfrac{2}{\sqrt{\kappa}}x_{L}^{\text{in}}(t), \quad p_{L}(t)=\dfrac{2}{\sqrt{\kappa}}p_{L}^{\text{in}}(t)-\dfrac{4g\alpha_{2}(t)}{\kappa}\theta_{M}.
\label{ada}
\end{equation}
By means of the usually optical input-output relation $a_{1}^{\text{out}}(t)+a_{1}^{\text{in}}(t)=\sqrt{\kappa} a_{1}(t)$ we find:
\begin{equation}
\begin{aligned}
x_{L}^{\text{out}}(t) &=x_{L}^{\text{in}}(t), \\
p_{L}^{\text{out}}(t) &=p_{L}^{\text{in}}(t)-\frac{4 g \alpha_{2 }(t)}{\sqrt{\kappa}} \theta_{M}(t), \\
\frac{d \theta_{M}}{d t} &=0, \\
\frac{d L_{M}}{d t} &=-2 g\alpha_{2}(t)x_{L}(t),
\label{input 1}
\end{aligned}
\end{equation}
In order to derive the inout-output in our main article, we define time collective optical quandratures $x_{L}^{\text{in/out}}=\int_{-\infty}^{\infty}f(t)x_{L}^{\text{in/out}}(t)dt$, $p_{L}^{\text{in/out}}=\int_{-\infty}^{\infty}f(t)p_{L}^{\text{in/out}}(t)dt$. It is easy to verify that $x_{L}^{\text{in/out}}$ and $p_{L}^{\text{in/out}}$ also satisfy the standard commutation
$\left[x_{L}^{\text{in/out}}, p_{L}^{\text{in/out}}\right]=i$. Besides, we define $\theta_{M}^{\text{ in}}=\theta_{M}(t=-\infty)$ and $\theta_{M}^{\text{ out}}=\theta_{M}(t=+\infty)$. As $\alpha_{2 }(t)$ is a short pulse, $\theta_{M}^{\text{ in}}$ and $\theta_{M}^{\text{ out}}$ represent the $\theta_{M}$ quadrature before and after the optomechanical interaction, respectively.
Firstly, from equation $\frac{d \theta_{M}}{d t} =0$ we known that $\theta_{M}(t)=\theta_{M}^{\text{ in}}=\theta_{M}^{\text{ out}}$, which means that $\theta_{M}(t)$ is a constant operator during the optomechanical interaction.
Secondly, multipling the equation $x_{L}^{\text{out}}(t) =x_{L}^{\text{in}}(t)$ by $f(t)$ and integrating it over the whole time interval, we will obtain the equation $x_{L}^{\text {out }} =x_{L}^{\text {in }}$. Similarly, also multipling the equation $p_{L}^{\text{out}}(t) =p_{L}^{\text{in}}(t)-\frac{4 g \alpha_{2 }(t)}{\sqrt{\kappa}} \theta_{M}$ by $f(t)$ and integrating it over the whole time interval, we will have $p_{L}^{\text{out}}=p_{L}^{\text{in}}-\dfrac{4g}{\sqrt{\kappa}}\int_{-\infty}^{\infty}f(t)\alpha_{2 }(t)\theta_{M}(t)dt=p_{L}^{\text{in}}-\dfrac{4g\theta_{M}^{\text{in}}}{\sqrt{\kappa}}\int_{-\infty}^{\infty}f(t)\alpha_{2 }(t)dt$.
Defining $\chi=\dfrac{4g}{\sqrt{\kappa}}\int_{-\infty}^{\infty}f(t)\alpha_{2 }(t)dt$, and considering Equ. (\ref{alpha2}) $\alpha_{2}(t)=\alpha f(t)$, we have
\begin{equation}
\chi=\dfrac{4g}{\sqrt{\kappa}}\int_{-\infty}^{\infty}f(t)\alpha_{2 }(t)dt=\dfrac{4g}{\sqrt{\kappa}}\int_{-\infty}^{\infty}\alpha f(t)^{2}dt=\dfrac{4g\alpha}{\sqrt{\kappa}}=\dfrac{8g\sqrt{N_{\text{in}}}}{\kappa}=\dfrac{4g}{\sqrt{\kappa}}\sqrt{\int_{-\infty}^{\infty}|\alpha_{2}(t)|^{2}dt}.
\end{equation}
As a consequence, we have $p_{L}^{\text {out }} =p_{L}^{\text {in }}-\chi \theta_{M}^{\text {in }}$.
Thirdly, from equation $\frac{d L_{M}}{d t} =-2 g\alpha_{2}(t)x_{L}(t)$ we know that $L_{M}^{\text{ out}}-L_{M}^{\text{ in}}=-2g\int_{-\infty}^{\infty}\alpha_{2}(t)x_{L}(t)$. Using Equ. (\ref{ada}) $x_{L}(t)=\dfrac{2}{\sqrt{\kappa}}x_{L}^{\text{in}}(t)$, we will derive the equation $L_{M}^{\text{ out}}-L_{M}^{\text{ in}}=-4g/\sqrt{\kappa}\int_{-\infty}^{\infty}\alpha_{2}(t)x_{L}^{\text{in}}(t)dt=-\dfrac{4g\alpha}{\sqrt{\kappa}}x_{L}^{\text{in}}=-\chi x_{L}^{\text{in}}$.
Finally, we will derive this set of input-output relation
\begin{equation}
\begin{aligned}
x_{L}^{\text {out }} &=x_{L}^{\text {in }}, \\
p_{L}^{\text {out }} &=p_{L}^{\text {in }}-\chi \theta_{M}^{\text {in }}, \\
\theta_{M}^{\text {out }} &=\theta_{M}^{\text {in }}, \\
L_{M}^{\text {out }} &=L_{M}^{\text {in }}-\chi x_{L}^{\text {in }}.
\end{aligned}
\label{inout}
\end{equation}
\section{Wigner Function Formaluism}
For a continuous variable quantum system, if the initial states, the system evolution and the measurement process are all Gaussian, the whole system can advantageously be modeled using standard Gaussian formalism. However, in our system, the
the input optical state is non-Gaussian state and we must use other method to derive the Wigner function of the resulting mechanical state.
\subsection{General Definition of Wigner function}
For a $n$ modes Bosonic system, the Wigner function is defined by
\begin{equation}
W(\boldsymbol{r})=\frac{1}{(2 \pi)^{2 n}} \int \mathrm{d}^{2 n} \boldsymbol{\beta} \operatorname{Tr}\{\hat{\rho} \hat{D}(\boldsymbol{\beta})\} \exp \{-\mathrm{i} \boldsymbol{r}^{T} \cdot \Omega \boldsymbol{\beta}\},
\end{equation}
in which the $2 n \times 2 n$ matrix $\Omega$ is the $n$ -fold block diagonal matrix with diagonal blocks $ \varpi=\left(\begin{array}{rr}
0 & 1 \\
-1 & 0
\end{array}\right)$
and $\hat{D}(\boldsymbol{\beta})$ is the multi-mode displacement operator with $\boldsymbol{\beta}^{\mathrm{T}}=\left(\Re\left\{\beta_{1}\right\}, \Im\left\{\beta_{1}\right\}, \Re\left\{\beta_{2}\right\}, \ldots\right)$ giving the displacement of each mode.
Points in phase space are denoted $\boldsymbol{r}^{\mathrm{T}}=\left(x_{1}, p_{1}, x_{2}, \ldots\right)$ and $\boldsymbol{r} \cdot \Omega \boldsymbol{\beta}=\sum_{i=1}^{n} (-p_{i} \Re\left\{\beta_{i}\right\}+x_{i}\Im\left\{\beta_{i}\right\})$.
For a single mode state, displacement operator $\hat{D}(\boldsymbol{\beta})$ is defined as
$
\hat{D}(\boldsymbol{\beta})=\exp(\beta a^{\dagger}-\beta a).
$
If we define $\hat{x}=a^{\dagger}+a$, $\hat{p}=i(a^{\dagger}-a)$, and $ \hat{\boldsymbol{r}}=(\hat{x},\hat{p})$, then
$
\beta a^{\dagger}-\beta a=i (\Im\left\{\beta\right\} \hat{x}-\Re\left\{\beta\right\} \hat{p})=i\hat{\boldsymbol{r}}^{T} \cdot \Omega \boldsymbol{\beta}.
$
Therefore, displacement operator can also be expressed as
$\hat{D}(\boldsymbol{\beta})=\exp(i\hat{\boldsymbol{r}} \cdot \Omega \boldsymbol{\beta})$.
The Wigner function is Fourier dual to the characteristic function $\chi(\boldsymbol{\beta})=\operatorname{Tr}\{\hat{\rho} \hat{D}(\boldsymbol{\beta})\},$
\begin{equation}
\chi(\boldsymbol{\beta})=\int \mathrm{d}^{2 n} \boldsymbol{r} W(\boldsymbol{r}) \exp \{+\mathrm{i} \boldsymbol{r}^{T} \cdot \Omega \boldsymbol{\beta}\}
\end{equation}
and it is easy to verify that $\chi(0)=\operatorname{Tr}[\hat{\rho}]=1$.
\subsection{Time evolution of Wigner function}
If the time evolution of all the quadratures $\hat{r}$ of
a bosonic system can be described by a linear transformation, which means that
\begin{equation}
\hat{r}(T)=M\hat{r}(0),
\end{equation}
then the Wigner function and characteristic function of this system at time $T$ can also be totally described by matrix $M$ and the original density matrix $\hat{\rho}(0)$. In order to connect the Wigner function $W(\boldsymbol{r},T)$ of this system at time $T$ with the Wigner function $W(\boldsymbol{r},0)$, we consider the following calculation.
Firstly, we consider the time evolution of characteristic function.
At time $T$, the displacement operator is
\begin{equation}
\hat{D}(\boldsymbol{\beta},T)=\exp[i\hat{\boldsymbol{r}}(T)^{T} \cdot \Omega \boldsymbol{\beta}]=\exp[i(M\hat{\boldsymbol{r}}(0))^{T} \cdot \Omega \boldsymbol{\beta}]=\exp[i\beta^{T}\Omega^{T}M\hat{\boldsymbol{r}}(0)].
\end{equation}
If we define a vector $\gamma=\Omega M^{\mathrm{T}} \Omega \boldsymbol{\beta}$, then we will have the following relation
\begin{equation}
\begin{aligned} \boldsymbol{\beta}^{\mathrm{T}} \Omega^{\mathrm{T}} M \boldsymbol{\hat{r}}(0) &=-\boldsymbol{\beta}^{\mathrm{T}} \Omega^{\mathrm{T}} M \Omega \Omega^{\mathrm{T}} \boldsymbol{\hat{r}}(0) \\ &=\boldsymbol{\beta}^{\mathrm{T}}\left(-\Omega^{\mathrm{T}} M \Omega\right) \Omega^{\mathrm{T}} \boldsymbol{\hat{r}}(0) \\ &=\boldsymbol{\beta}^{\mathrm{T}}(\Omega M \Omega) \Omega^{\mathrm{T}} \boldsymbol{\hat{r}}(0) \\ &=\gamma^{\mathrm{T}} \Omega^{\mathrm{T}} \boldsymbol{\hat{r}}(0)
\\ &=\boldsymbol{\hat{r}}(0)^{T}\cdot \Omega \gamma.
\end{aligned}
\end{equation}
Therefore, at time $T$
\begin{equation}
\hat{D}(\boldsymbol{\beta},T)=\exp{[ \boldsymbol{\hat{r}}(0)^{T}\cdot \Omega \gamma]}=\hat{D}(\boldsymbol{\gamma},0).
\end{equation}
The corresponding characteristic is
\begin{equation}
\chi(\boldsymbol{\beta},T)=\operatorname{Tr}[\hat{\rho}(T)\hat{D}(\boldsymbol{\beta},0)]=\operatorname{Tr}[\hat{\rho}(0)\hat{D}(\boldsymbol{\beta},T)]=\chi(\boldsymbol{\gamma,0}).
\end{equation}
Secondly, we consider the time evolution of Wigner function. If we define $\Omega M^{T} \Omega=K$, then $\gamma=K\boldsymbol{\beta}$.
At time $T$,
\begin{equation}
W(\boldsymbol{r},T)=\frac{1}{(2 \pi)^{2 n}} \int \mathrm{d}^{2 n} \boldsymbol{\beta} \chi(\boldsymbol{\beta},T) \exp \{-\mathrm{i} \boldsymbol{r}^{T} \cdot \Omega \boldsymbol{\beta}\}=\frac{1}{(2 \pi)^{2 n}} \int \mathrm{d}^{2 n} \boldsymbol{\beta} \chi(K\boldsymbol{\beta},0) \exp \{-\mathrm{i} \boldsymbol{r}^{T} \cdot \Omega \boldsymbol{\beta}\}.
\end{equation}
Using coordination transformation $\beta^{\prime}=K\beta$, then
\begin{equation}
\boldsymbol{r}^{T} \cdot \Omega \beta = \boldsymbol{r}^{T} \cdot \Omega K^{-1} \beta^{\prime}= \boldsymbol{r}^{T} \cdot (M^{-1})^{T}\Omega^{-1} \beta^{\prime} =(M^{-1}\boldsymbol{r})^{T} \cdot \Omega \beta^{\prime}.
\end{equation}
Therefore, the Wigner function of the total system at time $T$ is
\begin{equation}
W(\boldsymbol{r},T) \propto \frac{1}{(2 \pi)^{2 n}} \int \mathrm{d}^{2 n} \boldsymbol{\beta}^{\prime} \chi(\boldsymbol{\beta}^{\prime},0) \exp \{-\mathrm{i} (M^{-1}\boldsymbol{r})^{T} \cdot \Omega \boldsymbol{\beta}^{\prime}\} = W(M^{-1}\boldsymbol{r},0).
\label{wig time}
\end{equation}
Here we use $"\propto"$ instead of $"="$ is that we do not calculate the Jacobin matrix during the process of coordination transformation $\beta^{\prime}=K\beta$.
\subsection{The Resulting Wigner Function }
For our optomechanical system, if we neglect all of the noises and only care about the torsional mode $b$ and optical mode $a_{1}$, the transformation matrix $M$ is
\begin{equation}
M=\begin{pmatrix}
1& 0 & 0 &0 \\
0& 1& -\chi &0 \\
0& 0 & 1 & 0 \\
-\chi& 0 & 0 & 1
\end{pmatrix},
\end{equation}
which is derived from the input-output relation ($\ref{inout}$).
We assume the initial quantum state $\rho(0)=\rho_{L}\otimes\rho_{M}$, where $\rho_{L}$ is the density matrix of the input optical mode $a_{1\text{in}}$ and $\rho_{M}$ is the density matrix of the mechanical mode. The initial Wigner function is
$W(\boldsymbol{r},0)=W_{L}(x_{L},p_{L})W_{M}(x_{M},p_{M})$. After the optomechanical interaction ($\ref{inout}$), the Wigner function of this system is
\begin{equation}
W(\boldsymbol{r},T)=W(M^{-1}\boldsymbol{r},0)=W_{L}(x_{L},p_{L}+\chi x_{M})W_{M}(x_{M},p_{M}+\chi x_{L}).
\end{equation}
If we do a quantum measurement on the subsystem $\rho_{L}$ with POVM measurement operator $\Pi$, then the resulting Wigner function is
\begin{equation}
W_{\text{meas}}(x_{M},p_{M}) \propto \int \mathrm{d}^{2} \boldsymbol{r}_{L}W(\boldsymbol{r})W_{\Pi}(x_{L},p_{L}),
\label{meas}
\end{equation}
where
\begin{equation}
W_{\Pi}(\boldsymbol{r})=\frac{1}{(2 \pi)^{2 n}} \int \mathrm{d}^{2 n} \boldsymbol{\beta} \operatorname{Tr}\{\Pi \hat{D}(\boldsymbol{\beta})\} \exp \{-\mathrm{i} \boldsymbol{r}^{T} \cdot \Omega \boldsymbol{\beta}\},
\end{equation}
which is the Wigner function of the measurement operator $\Pi$. In the field of quantum optics, the most commonly used measurements are photon number detection, homodyne detection and heterodyne detection.
\paragraph{Photon Number Detection}
The POVM operator for photon number resolved detection with detection efficiency $\eta$
is $\Pi=\Pi_{m}(\eta)=\eta^{m} \sum_{k=m}^{\infty}(1-\eta)^{k-m}
C_{m}^{k}
|k\rangle\langle k|$ where $C_{m}^{k}=\dfrac{m!}{k!(m-k)!}$. The corresponding Wigner function \cite{0503237}
\begin{equation}
W(X,Y)=\frac{1}{2\pi} \frac{(-1)^{m} \eta^{m}}{(2-\eta)^{1+m}} L_{m}\left(\frac{(X^2+Y^2)}{2-\eta}\right) \exp \left\{-\frac{ \eta}{2(2-\eta)}(X^2+Y^2)\right\}
\label{pho}
\end{equation}
where $L_{k}(x)$ is a Laguerre polynomials.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{com}
\caption{(a) Wigner Function of single phonon fock state. (b) Wigner function $W_{\text{sfcok}}(\theta_{M},L_{M})$. Here we set $V_{LL}=2000$ and $V_{\theta\theta}=0.2$. (c) Wigner function of cat state with $\alpha=2$. (d) Wigner function $W_{\text{scat}}(\theta_{M},L_{M})$. Here we also set $V_{LL}=2000$ and $V_{\theta\theta}=0.2$. }
\label{fig:com}
\end{figure}
\paragraph{Homodyne Detection}
If we use a photodetector with efficiency $\eta=1$ to detect the quadrature operator $X_{\phi}=a e^{-i\phi}+ a^{\dagger} e^{i\phi}$ with the measurement result $x$, then the corresponding POVM is
$
\Pi (x)=|x \rangle_{\phi} \langle x|
$,
where $X_{\phi} | x\rangle_{\phi}=x| x\rangle_{\phi}$
and the corresponding Wigner functions is given \cite{0503237}
\begin{equation}
W(X,P) \propto \delta\left(x-X\cos(\phi)-Y\sin(\phi)\right).
\end{equation}
In our article, after the ideal homodyne detection of the output optical mode, the Wigner function of the resulting mechanical state is
\begin{equation}
W_{\text{meas}}(\theta_{M},L_{M})\propto\int \mathrm{d}^{2}\boldsymbol{r}_{L} W_{L}(x_{L},p_{L}+\chi \theta_{M})W_{M}(\theta_{M},L_{M}+\chi x_{L})\delta(p_{L}-p)
\label{general w}
\end{equation}
where $p$ is the result of the homodyne detection and in our numerical calculation we will choose $p=0$.
\paragraph{Single Photon Pulse Input}
Here we consider a special and simple example: the original input optical state is a single photon fock state pulse with Wigner function $W_{L}(x,p)\propto \exp{(-\dfrac{1}{2}x^{2}-\dfrac{1}{2}p^{2})}(x^{2}+p^{2}-1)$
and the original mechanical state is a precooled thermal squeezed state $W_{M}(\theta_{M},L_{M})\propto \exp\left[ -\dfrac{\theta_{M}^{2}}{2V_{\theta \theta}} -\dfrac{L^{2}}{2V_{LL}} \right]$
where $V_{LL}$ and $V_{\theta\theta}$ are the vatiance of $L_{M}$ and $\theta_{M}$, respectively.
When $p=0$, the resulting mechanical state is
\begin{equation}
W_{\text{sfock}}(\theta_{M},L_{M}) \propto \exp{[-\dfrac{1}{2}(1+V_{LL}^{-1})\theta_{M}^{2}-\dfrac{1}{2}(1+V_{\theta\theta})^{-1}L_{M}^{2}]}
\left[(1+V_{\theta\theta})\theta_{M}^2+(1+V_{\theta\theta})^{-1}L_{M}^2-1
\right].
\end{equation}
If $V_{LL}\gg 1$, Wigner function $W_{\text{sfock}}(\theta_{M},L_{M})$ will near the Wigner function of squeezed single phonon fock state and
when $V_{LL} \to \infty$ and $V_{\theta\theta}\to 0$,
$W_{\text{meas}}(\theta_{M},L_{M})$ will near the Wigner function of single phonon fock state (see Fig. \ref{fig:com}).
\paragraph{Perfect Cat State Input}
Now we consider a more complex case: the input optical state is a perfect
even cat state $|Cat\rangle \propto |\alpha\rangle + |-\alpha\rangle$. Without loss of generality, we assume that $\alpha$ is positive real value. For this even cat state, the corresponding Wigner function is
\begin{equation}
W_{\alpha}(\boldsymbol{r})=\frac{\mathrm{e}^{-\frac{1}{2} r \cdot r}}{2 \pi\left(1 + \mathrm{e}^{-2|\alpha|^{2}}\right)}\left[\mathrm{e}^{-2|\alpha|^{2}} \cosh (2 \boldsymbol{r} \cdot \boldsymbol{\alpha}) + \cos (2 \boldsymbol{r} \cdot \varpi \boldsymbol{\alpha})\right],
\end{equation}
where $\boldsymbol{r}=(x_{L},p_{L})$ and $ \varpi=\left(\begin{array}{rr}
0 & 1 \\
-1 & 0
\end{array}\right)$. As for the mechanical state $W_{M}(\theta_{M},L_{M})$, we also assume that it is a thermal squeezed state.
Assuming the homodyne result $p=0$, the Wigner function of the resulting mechanical state is
\begin{equation}
W_{\text{scat}}(\theta_{M},L_{M}) \propto \exp{[-\dfrac{1}{2}(1+V_{LL}^{-1})
\theta_{M}^{2}-\dfrac{1}{2}(1+V_{\theta\theta})^{-1}L_{M}^{2}]}
\left[
\cos(\dfrac{2\alpha}{1+V_{\theta\theta}} L_{M}) + \exp{(-\dfrac{2\alpha^{2}}{1+V_{\theta\theta}})} \cosh(2\alpha\theta_{M})
\right].
\end{equation}
Notice that
when $V_{LL} \to \infty$ and $V_{\theta\theta}\to 0$,
$W_{\text{scat}}(\theta_{M},L_{M})$ will near the Wigner function of cat state (see Fig. \ref{fig:com}).
\begin{figure}[ht]
\centering
\includegraphics[width=0.85\linewidth]{process}
\caption{(a) The preparing process of the optical cat state with Wigner function $W_{\text{mcat}}(x_{1L},p_{1L})$. BS is a beam splitter with transmittance $T_{\text{tap}}$, $m$ represents the photon number resolving detection with $m$ photons detected and $W_{m}(x_{2L},p_{2L})$ is the corresponding Wigner function. (b) State transformation and preparing process of the resulting mechanical cat-like state with Wigner function $W(\theta_{M},L_{M})$. The input mechanical state $W_{M}(\theta_{M},L_{M})$ is a squeezed thermal state along $\hat{L}_{M}$ quadrature and $p_{1L}^{\text{out}}$ represents the homodyne detection on the output optical quadrature $\hat{p}_{1L}^{\text{out}}$. }
\label{fig:process}
\end{figure}
\paragraph{Catlike State Prepared by Generalized Photon Subtracted Method}
Following the calculation in ref \cite{PhysRevA.103.013710}, we discuss the Wigner function of the catlike state prepared by generalized photon subtracted method (see Fig. \ref{fig:process}). For a squeezed state with sqeezing parameter $r$ ($r$ is real number), the corresponding Wigner function $W_{r}(x_{L},p_{L})$ is
\begin{equation}
W_{r}(x_{L},p_{L}) \propto \exp \left[-\dfrac{x_{L}^{2}}{2\exp(2r)}-\dfrac{p_{L}^{2}}{2\exp(-2r)}\right].
\end{equation}
In the generalized photon subtracted method, the original optical input state are two unentangled squeezed with squeezing parameter $r_{1}$ and $r_{2}$. The corresponding Wigner function of the input state is
\begin{equation}
W_{\text{Lin}}(x_{1L},p_{1L},x_{2L},p_{2L}) \propto W_{r_{1}}(x_{1L},p_{1L})W_{r_{2}}(x_{2L},p_{2L}).
\end{equation}
The asymmetric beam splitter with transmtance $T_{tap}$ exchanges the quantum state of the two squeezed light. If we define $\hat{r}_{L}=(\hat{x}_{1L},\hat{p}_{1L},\hat{x}_{2L},\hat{p}_{2L})$, then after the interaction of the beam splitter, the resulting optical quadratures will be
\begin{equation}
\hat{r}_{L} \rightarrow M_{L}\hat{r}_{L},
\end{equation}
where $M_{L}$ is the transmission matrix of the beam splitter, and the exact form of $M_{L}$ is
\begin{equation}
M_{L}=
\begin{pmatrix}
\sqrt{T_{tap}} & 0 & \sqrt{1-T_{tap}} & 0 \\
0 & \sqrt{T_{tap}} & 0 &\sqrt{1-T_{tap}} \\
-\sqrt{1-T_{tap}} & 0 &\sqrt{T_{tap}} & 0 \\
0 & -\sqrt{1-T_{tap}} & 0 & \sqrt{T_{tap}}
\end{pmatrix}.
\end{equation}
According to the time evolution formalism of Wigner function ($\ref{wig time}$), after the beam splitter, the corresponding optical Wigner function $W_{Ltap}$ is
\begin{equation}
W_{\text{Ltap}}(x_{1L},p_{1L},x_{2L},p_{2L}) \propto W_{\text{Lin}}(M_{L}^{-1}r_{L}).
\end{equation}
Then we will do photon number resolved detection on the output port of squeezed light $r_{2}$. The corresponding measurement Wigner function $W_{\text{ m}}(x_{L2},p_{L2})$ is Equ. (\ref{pho}), notice that $W_{\text{m}}(x_{L2},p_{L2})$
depends on the detected photon number $m$. Following the photon number resolved detection, the photon on the other port will be projected into a catlike state, and the exact expression of the resulting Wigner function is
\begin{equation}
W_{\text{mcat}}(x_{1L},p_{1L}) \propto \int dx_{2L}dp_{2L} W_{\text{Ltap}}(x_{1L},p_{1L},x_{2L},p_{2L}) W_{\text{m}}(x_{L2},p_{L2}).
\end{equation}
Following these processes, we prapare a opticl catlike state. The Wigner function of the prepared optical catlike states are plotted in Fig. (\ref{fig:optical-cat}). When $m>1$, the resulting Wigner function shows the feature of catlike state.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{"opticalcat"}
\caption{Wigner functions of optical catlike state prepared by generalized photon subtraction method. In numerical calculation, we choose $r_{1}=-r_{2}=1.15$ and $T_{\text{tap}}=(e^{2r_{1}}-1)/(e^{2r_{1}}-e^{-2r_{1}})\approx0.909$. Here $m$ represents the detected photon number.}
\label{fig:optical-cat}
\end{figure}
By using this optical catlike sate pulse to pump the optical mode $a_{1}$ in our optomechanical system, we cat prepare our torsional oscillator in a catlike state, and the Wigner function of the resulting mechanical state is
\begin{equation}
W(\theta_{M},L_{M})\propto\int \mathrm{d}^{2}dx_{1L}dp_{1L} W_{\text{mcat}}(x_{1L},p_{1L}+\chi \theta_{M})W_{M}(\theta_{M},L_{M}+\chi x_{L})\delta(p_{1L}-p).
\end{equation}
For different detected photon number $m$, $W_{\text{mcat}}$ will be different, thus $W(\theta_{M},L_{M})$ will also be different.
\bibliographystyle{unsrt}
|
\section{Commitment with Certified Everlasting Hiding and Sum-Binding}\label{app:sum_binding}
In this appendix, we define and construct commitment with certified
everlasting hiding and statistical sum-binding.
\subsection{Definition}
\begin{definition}[Commitment with Certified Everlasting Hiding and Sum-Binding (Syntax)]\label{def:commitment_with_cd_sumbinding}
Let $\lambda$ be the security parameter, and let $p$, $q$, $r$ and $s$ be some polynomials.
Commitment with certified everlasting hiding and sum-binding consists of a tuple of algorithms $(\algo{Commit},\algo{Verify},\algo{Del},\algo{Cert})$
with message space $\mathcal{M}:=\{0,1\}$, commitment space $\mathcal{C}:=\mathcal{Q}^{\otimes p(\lambda)}$,
decommitment space $\mathcal{D}:=\{0,1\}^{q(\lambda)}$, key space $\mathcal{K}:=\{0,1\}^{r(\lambda)}$
and deletion certificate space $\mathcal{E}:= \{0,1\}^{s(\lambda)}$.
\begin{description}
\item[$\algo{Commit} (1^{\lambda},b)\rightarrow (\mathsf{com},\mathsf{d},\keys{ck})$:]
The commitment algorithm takes as input a security parameter $1^{\lambda}$ and a message $b\in\{0,1\}$, and outputs a commitment $\mathsf{com}\in\mathcal{C}$, a decommitment $\mathsf{d}\in\mathcal{D}$, and a key $\keys{ck}\in\mathcal{K}$.
\item[$\algo{Verify}(\mathsf{com},\mathsf{d},b)\rightarrow \top~or~\bot $:]
The verification algorithm takes as input $\mathsf{com}$, $\mathsf{d}$ and $b$,
and outputs $\top$ or $\bot$.
\item[$\algo{Del}(\mathsf{com})\rightarrow \mathsf{cert}$:]The deletion algorithm takes $\mathsf{com}$ as input, and outputs a certificate $\mathsf{cert}\in \mathcal{E}$.
\item[$\algo{Cert}(\mathsf{cert},\keys{ck})\rightarrow \top~or~\bot$:]The certification algorithm takes $\mathsf{cert}$ and $\keys{ck}$ as input,
and outputs $\top$ or $\bot$.
\end{description}
\end{definition}
\begin{definition}[Correctness]\label{def:correctness_bit_commitment}
There are two types of correctness, namely, decommitment correctness and deletion correctness.
\begin{description}
\item[Decommitment correctness:]
There exists a negligible function ${\mathsf{negl}}$ such that for any $\lambda\in\mathbb{N}$ and $b\in\{0,1\}$,
\begin{align}
\Pr[\algo{Verify} (\mathsf{com},\mathsf{d},b)=\top \mid (\mathsf{com},\mathsf{d},\keys{ck})\leftarrow\algo{Commit}(1^{\lambda},b)]\geq1-{\mathsf{negl}}(\lambda).
\end{align}
\item[Deletion correctness:]
There exists a negligible function ${\mathsf{negl}}$ such that for any $\lambda \in\mathbb{N}$ and $b\in\{0,1\}$,
\begin{align}
\Pr[\algo{Cert}(\mathsf{cert},\keys{ck})=\top \mid (\mathsf{com},\mathsf{d},\keys{ck})\leftarrow\algo{Commit}(1^\lambda,b),\mathsf{cert}\leftarrow\algo{Del}(\mathsf{com})]\geq1-{\mathsf{negl}}(\lambda).
\end{align}
\end{description}
\end{definition}
\begin{definition}[$\epsilon$-Sum-Binding]\label{def:sum_binding}
For any $\mathsf{com}$, $\mathsf{d}$, and $\mathsf{d}'$, it holds that
\begin{align}
\Pr[\algo{Verify}(\mathsf{com},\mathsf{d},0)=\top]+\Pr[\algo{Verify}(\mathsf{com},\mathsf{d}',1)=\top]\leq 1+\epsilon.
\end{align}
We call $\epsilon$-sum-binding just sum-binding if $\epsilon$ is negligible.
\end{definition}
\begin{definition}[Computational Hiding]\label{def:computationally_hiding}
Let $\Sigma\coloneqq(\algo{Commit},\algo{Verify},\algo{Del},\algo{Cert})$.
Let us consider the following security experiment $\expa{\Sigma,\mathcal{A}}{c}{hide}(\lambda,b)$ against
any QPT adversary $\mathcal{A}$.
\begin{enumerate}
\item The challenger computes $(\mathsf{com},\mathsf{d},\keys{ck})\leftarrow \algo{Commit}(1^{\lambda},b)$, and sends $\mathsf{com}$ to $\mathcal{A}$.
\item $\mathcal{A}$ outputs $b'\in\{0,1\}$.
\item The output of the experiment is $b'$.
\end{enumerate}
Computational hiding means that the following is satisfied for any QPT $\mathcal{A}$.
\begin{align}
\advb{\Sigma,\mathcal{A}}{c}{hide}(\lambda)\coloneqq
\left|\Pr[ \expa{\Sigma,\mathcal{A}}{c}{hide}(\lambda,0)=1]-
\Pr[\expa{\Sigma,\mathcal{A}}{c}{hide}(\lambda,1)=1]\right|\leq {\mathsf{negl}}(\lambda).
\end{align}
\end{definition}
\begin{definition}[Certified Everlasting Hiding]\label{def:everlasting_hiding}
Let $\Sigma\coloneqq(\algo{Commit},\algo{Verify},\algo{Del},\algo{Cert})$.
Let us consider the following security experiment $\expa{\Sigma,\mathcal{A}}{ever}{hide}(\lambda,b)$ against
$\mathcal{A}=(\mathcal{A}_1,\mathcal{A}_2)$ consisting of any QPT adversary $\mathcal{A}_1$ and any unbounded adversary $\mathcal{A}_2$.
\begin{enumerate}
\item The challenger computes $(\mathsf{com},\mathsf{d},\keys{ck})\leftarrow \algo{Commit}(1^{\lambda},b)$, and sends $\mathsf{com}$ to $\mathcal{A}_1$.
\item At some point, $\mathcal{A}_1$ sends $\mathsf{cert}$ to the challenger, and sends its internal state to $\mathcal{A}_2$.
\item The challenger computes $\algo{Cert}(\mathsf{cert},\keys{ck})$. If the output is $\top$, then the challenger outputs $\top$,
and sends $(\mathsf{d},\keys{ck})$ to $\mathcal{A}_2$.
Else, the challenger outputs $\bot$, and sends $\bot$ to $\mathcal{A}_2$.
\item $\mathcal{A}_2$ outputs $b'\in\{0,1\}$.
\item If the challenger outputs $\top$, then the output of the experiment is $b'$. Otherwise, the output of the experiment is $\bot$.
\end{enumerate}
We say that it is certified everlasting hiding if the following is satisfied for any $\mathcal{A}=(\mathcal{A}_1,\mathcal{A}_2)$.
\begin{align}
\advb{\Sigma,\mathcal{A}}{ever}{hide}(\lambda)\coloneqq
\left|\Pr[ \expa{\Sigma,\mathcal{A}}{ever}{hide}(\lambda,0)=1]-
\Pr[\expa{\Sigma,\mathcal{A}}{ever}{hide}(\lambda,1)=1]\right|\leq {\mathsf{negl}}(\lambda).
\end{align}
\end{definition}
\subsection{Construction}
Though the construction is essentially the same as that in \cref{sec:bit_comm_construction}, we give the full description for clarity.
Let $\lambda$ be the security parameter, and let $p$, $q$, $r$, $s$, $t$ and $u$ be some polynomials.
We construct a bit commitment with certified everlasting hiding and sum-binding,
$\Sigma_{\mathsf{ccd}}=(\algo{Commit},\algo{Verify},\algo{Del},\algo{Cert})$, with message space $\mathcal{M}=\{0,1\}$,
commitment space $\mathcal{C}=\mathcal{Q}^{\otimes p(\lambda)}\times \{0,1\}^{q(\lambda)}\times \{0,1\}^{r(\lambda)}$,
decommitment space $\mathcal{D}=\{0,1\}^{s(\lambda)}\times \{0,1\}^{t(\lambda)}$,
key space $\mathcal{K}=\{0,1\}^{r(\lambda)}$
and deletion certificate space $\mathcal{E}=\{0,1\}^{u(\lambda)}$
from the following primitives:
\begin{itemize}
\item Secret-key encryption with certified deletion
$\Sigma_{\mathsf{skcd}}=\mathsf{SKE}.(\algo{KeyGen},\algo{Enc},\algo{Dec},\algo{Del},\algo{Verify})$, with plaintext space $\mathcal{M}=\{0,1\}$, ciphertext space $\mathcal{C}=\mathcal{Q}^{\otimes p(\lambda)}$, key space $\mathcal{K}=\{0,1\}^{r(\lambda)}$, and deletion certificate space $\mathcal{E}=\{0,1\}^{u(\lambda)}$.
\item
Classical non-interactive commitment, $\Sigma_{\mathsf{com}}=\algo{Classical}.\algo{Commit}$,
with plaintext space $\{0,1\}^{s(\lambda)}$, randomness space $\{0,1\}^{t(\lambda)}$, and commitment space $\{0,1\}^{q(\lambda)}$.
\item
A hash function $H$ from $\{0,1\}^{s(\lambda)}$ to $\{0,1\}^{r(\lambda)}$ modeled as a quantumly-accessible random oracle.
\end{itemize}
The construction is as follows.
\begin{description}
\item[$\algo{Commit}(1^{\lambda},b)$:] $ $
\begin{itemize}
\item Generate $\mathsf{ske.sk}\leftarrow\mathsf{SKE}.\algo{KeyGen}(1^{\lambda})$, $R\leftarrow\{0,1\}^{s(\lambda)}$, $R'\leftarrow\{0,1\}^{t(\lambda)}$, and a hash function $H$ from $\{0,1\}^{s(\lambda)}$ to $\{0,1\}^{r(\lambda)}$.
\item Compute $\mathsf{ske}.\keys{CT}\leftarrow \SKE.\algo{Enc}(\mathsf{ske}.\keys{sk},b)$, $f\leftarrow \algo{Classical}.\algo{Commit}(R;R')$, and $h\coloneqq H(R)\oplus\mathsf{ske}.\keys{sk}$.
\item Output $\mathsf{com}\coloneqq (\mathsf{ske}.\keys{CT},f,h)$, $\mathsf{d}\coloneqq (R,R')$, and $\keys{ck}\coloneqq\mathsf{ske}.\keys{sk}$.
\end{itemize}
\item[$\algo{Verify}(\mathsf{com},\mathsf{d},b)$:] $ $
\begin{itemize}
\item Parse $\mathsf{com}=(\mathsf{ske}.\keys{CT},f,h)$ and $\mathsf{d}=(R,R')$.
\item Compute $\mathsf{ske}.\keys{sk}'\coloneqq H(R)\oplus h$.
\item Compute $b'\leftarrow\SKE.\algo{Dec}(\mathsf{ske}.\keys{sk}',\mathsf{ske}.\keys{CT})$.
\item Output $\top$ if $f=\algo{Classical}.\algo{Commit}(R;R')$ and $b'=b$, and output $\bot$ otherwise.
\end{itemize}
\item[$\algo{Del}(\mathsf{com})$:] $ $
\begin{itemize}
\item Parse $\mathsf{com}=(\mathsf{ske}.\keys{CT},f,h)$.
\item Compute $\mathsf{ske}.\mathsf{cert}\leftarrow \SKE.\algo{Del}(\mathsf{ske}.\keys{CT})$.
\item Output $\mathsf{cert}\coloneqq\mathsf{ske}.\mathsf{cert}$.
\end{itemize}
\item[$\algo{Cert}(\mathsf{cert},\keys{ck})$:] $ $
\begin{itemize}
\item Parse $\mathsf{cert}=\mathsf{ske}.\mathsf{cert}$ and $\keys{ck}= \mathsf{ske}.\keys{sk}$.
\item Output $\top/\bot\leftarrow \SKE.\algo{Verify}(\mathsf{ske}.\keys{sk},\mathsf{ske}.\mathsf{cert})$.
\end{itemize}
\end{description}
\paragraph{Correctness.}
The decommitment and deletion correctness easily follow from the correctness of $\Sigma_{\mathsf{skcd}}$.
\paragraph{Security.}
We prove the following three theorems.
\begin{theorem}\label{thm:sum_binding}
If $\Sigma_{\mathsf{com}}$ is perfect binding, then $\Sigma_{\mathsf{ccd}}$ is sum-binding.
\end{theorem}
\begin{theorem}\label{thm:everlasting_hiding}
If $\Sigma_{\mathsf{com}}$ is unpredictable and $\Sigma_\mathsf{skcd}$ is OT-CD secure, then $\Sigma_{\mathsf{ccd}}$ is certified everlasting hiding.
\end{theorem}
\begin{theorem}\label{thm:computationally_hiding}
If $\Sigma_{\mathsf{com}}$ is unpredictable and $\Sigma_{\mathsf{skcd}}$ is OT-CD secure, then $\Sigma_{\mathsf{ccd}}$ is computationally hiding.
\end{theorem}
\begin{proof}[Proof of \cref{thm:sum_binding}]
What we have to prove is that for any $\mathsf{com}$, $\mathsf{d}$, and $\mathsf{d}'$, it holds that
\begin{align}
\Pr[\algo{Verify}(\mathsf{com},\mathsf{d},0)=\top]+\Pr[\algo{Verify}(\mathsf{com},\mathsf{d}',1)=\top]\leq 1+{\mathsf{negl}}(\lambda).
\end{align}
Let $\mathsf{d}=(R_0,R_0')$, $\mathsf{d}'=(R_1,R_1')$,
and $\mathsf{com}=(\mathsf{ske}.\keys{CT},f,h)$.
Then,
\begin{align}
& \Pr[\algo{Verify}(\mathsf{com},\mathsf{d},0)=\top]+\Pr[\algo{Verify}(\mathsf{com},\mathsf{d}',1)=\top]\\
& = \Pr[0\leftarrow \SKE.\algo{Dec}(h\oplus H(R_0),\mathsf{ske}.\keys{CT})\wedge f=\algo{Classical}.\algo{Commit}(R_0;R_0')]\\
& + \Pr[1\leftarrow \SKE.\algo{Dec}(h\oplus H(R_1),\mathsf{ske}.\keys{CT})\wedge f=\algo{Classical}.\algo{Commit}(R_1;R_1')]\\
& \le \Pr[0\leftarrow \SKE.\algo{Dec}(h\oplus H(\tilde{R}),\mathsf{ske}.\keys{CT})\wedge f=\algo{Classical}.\algo{Commit}(\tilde{R};R_0')]\\
& + \Pr[1\leftarrow \SKE.\algo{Dec}(h\oplus H(\tilde{R}),\mathsf{ske}.\keys{CT})\wedge f=\algo{Classical}.\algo{Commit}(\tilde{R};R_1')]\\
& \le \Pr[0\leftarrow \SKE.\algo{Dec}(h\oplus H(\tilde{R}),\mathsf{ske}.\keys{CT})]
+ \Pr[1\leftarrow \SKE.\algo{Dec}(h\oplus H(\tilde{R}),\mathsf{ske}.\keys{CT})]\\
& = \Pr[0\leftarrow \SKE.\algo{Dec}(h\oplus H(\tilde{R}),\mathsf{ske}.\keys{CT}) \vee
1\leftarrow \SKE.\algo{Dec}(h\oplus H(\tilde{R}),\mathsf{ske}.\keys{CT})]\\
&\le1,
\end{align}
where we have used perfect binding of $\Sigma_{\mathsf{com}}$ in the second inequality.
\if0
Since $\algo{Classical}.\algo{Commit}$ is perfectly binding and $\algo{Verify}$ outputs $\bot$ if $f\neq \algo{Classical}.\algo{Commit}(R;R')$, for any $\mathsf{com}=(\mathsf{ske}.\keys{CT},f,h)$ there exists uniquely determined $R$ such that verification algorithm $\algo{Verify}$ outputs $\top$.
Therefore, without loss of generality we can assume that the element of $\mathsf{d}$ contains the unique $R$ mentioned above.
This is because the verification algorithm always outputs $\bot$ when $\mathsf{d}$ does not contain $R$ mentioned above and thus does not contribute to the probability $\Pr[\algo{Verify}(\mathsf{com},\mathsf{d},b)=\top]$.
Since $R$ is uniquely determined, $\mathsf{ske}.\keys{sk}'=h\oplus H(R)$ is also uniquely determined from $\mathsf{com}$.
If we define $p_b:=\Pr[\SKE.\algo{Dec}(\mathsf{ske}.\keys{sk}',\mathsf{ske}.\keys{CT})\to b]$, it is obvious that $p_{b}$ satisfies the following equation
\begin{align}
p_{0}+p_{1}\leq1.
\end{align}
Obliviously from our construction it holds that $p_{b}=\Pr[\algo{Verify}(\mathsf{com},\mathsf{d},b)=\top]$ and thus our construction satisfies sum-binding.
\fi
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:everlasting_hiding}]
It is the same as that of \cref{thm:everlasting_hiding_2}.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:computationally_hiding}]
It is the same as that of \cref{thm:computationally_hiding_2}.
\end{proof}
\section{Proof of \texorpdfstring{\cref{prop:abort}}{Proposition~\ref{prop:abort}}}\label{proof:abort}
\begin{proof}[Proof of \cref{prop:abort}]
We prove the proposition by contradiction.
Let $p$ be the probability that $\mathcal{S}^{(1)}$ does not abort.
Assume that the probability $p$ satisfies $|p-\frac{1}{m}|\geq \frac{1}{q(\lambda)}$ for a polynomial $q$.
Then, we can construct an adversary $\mathcal{B}$ that breaks the computational hiding of $\Sigma_{\mathsf{ccd}}$.
Let us describe how $\mathcal{B}$ works below.
\begin{enumerate}
\item $\mathcal{B}$ generates $c\leftarrow[m]$ and $x,z\leftarrow\{0,1\}^n$.
\item $\mathcal{B}$ sends $m_0\coloneqq \{x_i,z_i\}_{i\in S_c}$ and $m_1\coloneqq0^{10}$ to the challenger.
$\mathcal{B}$ receives commitments from the challenger which is either $\{\mathsf{com}_i(x_i),\mathsf{com}_i(z_i)\}_{i\in S_c}$ or $\{\mathsf{com}_i(0),\mathsf{com}_i(0)\}_{i\in S_c}$.
\item $\mathcal{B}$ generates $\{\mathsf{com}_i(0),\mathsf{com}_i(0)\}_{i\in \overline{S}_c}$.
\item $\mathcal{B}$ generates $X^xZ^z\sigma(c)Z^zX^x$.
$\mathcal{B}$ appends commitments and $\xi$ to the quantum state in the ascending order.
If the commitments for $i\in S_c$ are $\{\mathsf{com}_i(x_i),\mathsf{com}_i(z_i)\}_{i\in S_c}$,
$\mathcal{B}$ obtains $X^xZ^z\sigma(c)Z^zX^x\otimes \mathsf{com}(x^{S_c})\otimes\mathsf{com}(z^{S_c})\otimes \xi$.
If the commitments for $i\in S_c$ are $\{\mathsf{com}_i(0),\mathsf{com}_i(0)\}_{i\in S_c}$,
$\mathcal{B}$ obtains $X^xZ^z\sigma(c)Z^zX^x\otimes \mathsf{com}(0^n)\otimes\mathsf{com}(0^n)\otimes \xi$.
\item $\mathcal{B}$ runs $V_1^*$ on it to obtain $(c',\{\mathsf{cert}_{i,x}',\mathsf{cert}_{i,z}'\}_{i\in \overline{S}_{c'}})$.
$\mathcal{B}$ outputs 0 when $c\neq c'$.
$\mathcal{B}$ outputs 1 when $c=c'$.
\end{enumerate}
When $\mathcal{B}$ receives $\{\mathsf{com}_i(x_i),\mathsf{com}_i(z_i)\}_{i\in S_c}$ from the challenger,
it outputs 1 with probability $p$
since it simulates $\mathcal{S}^{(1)}$.
When $\mathcal{B}$ receives $\{\mathsf{com}_i(0),\mathsf{com}_i(0)\}_{i\in S_c}$ from the challenger, on the other hand,
it outputs 1 with probability $\frac{1}{m}$,
because $(X^xZ^z\sigma(c) Z^zX^x)\otimes \mathsf{com}(0^n)\otimes\mathsf{com}(0^n)\otimes \xi$ is independent of $c$.
(Note that $\sigma(c)$ is one-time padded by $x,z$.)
Therefore if there exists some polynomial $q$ such that $|p-\frac{1}{m}|\geq \frac{1}{q(\lambda)}$,
$\mathcal{B}$ can break the computational hiding of $\Sigma_{\mathsf{ccd}}$
from (the computational hiding version of) \cref{lemma:eachbit}.
\end{proof}
\section{Proof of \texorpdfstring{\cref{lem:computational_zero}}{Lemma~\ref{lem:computational_zero}}}\label{proof:computational_zero}
\begin{proof}[Proof of \cref{lem:computational_zero}]
This proof is similar to the proof of Lemma~\ref{lem:everlasting_zero}.
For a subset $S_c\subseteq[n]$ and $x,z\in\{0,1\}^n$, let us define $x^{S_c}\coloneqq (x^{S_c}_1,x^{S_c}_2,\cdots, x^{S_c}_n)$ and $z^{S_c}\coloneqq (z^{S_c}_1,z^{S_c}_2,\cdots,z^{S_c}_n)$,
where $x^{S_c}_{i}=x_{i}$ and $z^{S_c}_{i}=z_{i}$ for $i\in S_{c}$,
and $x^{S_c}_{i}=z^{S_c}_{i}=0$ for $i\notin S_{c}$.
For clarity, we describe how the interactive algorithm $\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}^*(\xi)\rangle (\mathsf{x})$ runs against a QPT verifier $\mathcal{V}^*$ with an input $\xi$,
where $\mathsf{w}$ is the witness and $\mathsf{x}$ is the instance.
\begin{description}
\item[$\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}^*(\xi)\rangle (\mathsf{x})$:] $ $
\begin{enumerate}
\item $\mathcal{P}$ generates $x,z\leftarrow \{0,1\}^n$, and computes
\begin{align}
&(\mathsf{com}_i(x_i),\mathsf{d}_i(x_i),\keys{ck}_i(x_i))\leftarrow\algo{Commit}(1^\lambda,x_i) \\
&(\mathsf{com}_i(z_i),\mathsf{d}_i(z_i),\keys{ck}_i(z_i))\leftarrow\algo{Commit}(1^\lambda,z_i)
\end{align}
for all $i\in [n]$.
$\mathcal{P}$ sends $\keys{msg}_1\coloneqq(X^xZ^z\rho_\mathsf{hist} Z^zX^x)\otimes \mathsf{com}(x)\otimes\mathsf{com}(z)$ to $\mathcal{V}^*$.
\item $\mathcal{V}^*$ appends $\xi$ to the received state, and
runs a QPT circuit $V_1^*$ on it to obtain $(c,\{\mathsf{cert}_{i,x}',\mathsf{cert}_{i,z}'\}_{i\in\overline{S}_c})$.
$\mathcal{V}^*$ sends $\keys{msg}_2\coloneqq(c,\{\mathsf{cert}_{i,x}',\mathsf{cert}_{i,z}'\}_{i\in\overline{S}_c})$ to $\mathcal{P}$.
\item $\mathcal{P}$ sends $\keys{msg}_3\coloneqq\{\mathsf{d}_i(x_i),\mathsf{d}_i(z_i)\}_{i\in S_c}$ to $\mathcal{V}^*$.
\item $\mathcal{V}^*$ appends $\keys{msg}_3$ to its state, and runs a QPT circuit $V_2^*$ on it. $\mathcal{V}^*$ outputs its state $\xi'$.
\end{enumerate}
\end{description}
Next let us define a simulator $\mathcal{S}^{(1)}$ as follows.
\begin{description}
\item [The simulator $\mathcal{S}^{(1)}(\mathsf{x},\mathcal{V}^*,\xi)$:] $ $
\begin{enumerate}
\item Pick $c\leftarrow [m]$ and $x,z\leftarrow \{0,1\}^n$.
Compute
\begin{align}
& (\mathsf{com}_i(x^{S_c}_i),\mathsf{d}_i(x^{S_c}_i),\keys{ck}_i(x^{S_c}_i))\leftarrow\algo{Commit}(1^\lambda,x^{S_c}_i)\\
& (\mathsf{com}_i(z^{S_c}_i),\mathsf{d}_i(z^{S_c}_i),\keys{ck}_i(z^{S_c}_i))\leftarrow\algo{Commit}(1^\lambda,z^{S_c}_i)
\end{align}
for all $i\in [n]$.
\item Generate $(X^xZ^z\sigma(c) Z^zX^x) \otimes \mathsf{com}(x^{S_c})\otimes \mathsf{com}(z^{S_c})\otimes \xi$, where $\sigma(c)\coloneqq \rho_\mathsf{sim}^{\mathsf{x},S_c}\otimes\left(\prod_{i\in \overline{S}_c}|0\rangle\langle0|_i\right)$.
Run $V_1^*$ on the state to obtain $(c',\{\mathsf{cert}_{i,x}',\mathsf{cert}_{i,z}'\}_{i\in\overline{S}_{c'}})$.
\item If $c'\neq c$, abort and output a fixed state $\eta$ and the flag state $\mathsf{fail}$.
\item Append $\{\mathsf{d}_i(x^{S_c}_i),\mathsf{d}_i(z^{S_c}_i)\}_{i\in S_c}$ to its quantum state, and run $V_2^*$ on the state.
$\mathcal{S}$ outputs the output state and the flag state $\mathsf{success}$.
\end{enumerate}
\end{description}
Let us also define other two simulators $\mathcal{S}^{(2)}$ and $\mathcal{S}^{(3)}$ as follows.
\begin{description}
\item [The modified simulator $\mathcal{S}^{(2)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\xi)$:] $ $
It is the same as $\mathcal{S}^{(1)}$ except that
$\sigma(c)$ is replaced with $\rho_{\mathsf{hist}}$.
\end{description}
\begin{description}
\item [The simulator $\mathcal{S}^{(3)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\xi)$:] $ $
$\mathcal{S}^{(3)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\cdot~)$ is the channel that postselects the output of
\\$\mathcal{S}^{(2)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\cdot~)$ on the non-aborting state. More precisely, if we
write $\mathcal{S}^{(2)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\rho_{in})=p \rho_{out}\otimes \mathsf{success}+(1-p)\eta\otimes \mathsf{fail}$,
where $p$ is the non-aborting probability,
$\mathcal{S}^{(3)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\rho_{in})= \rho_{out}$.
\end{description}
\cref{lem:computational_zero} is shown from the following \cref{prop:abort_comp,prop:indistinguishable_S_S'_comp,prop:indistinguishable_S'_V^*_comp} (whose proofs will be given later)
and quantum rewinding lemma(\cref{lemma:rewinding}), which is used to reduce the probability that $\mathcal{S}^{(1)}$ aborts to ${\mathsf{negl}}(\lambda)$.
In fact, from \cref{prop:abort_comp,lemma:rewinding}, there exists a quantum circuit $\mathcal{S}^{(0)}$ of size at most
$O(m\,{\rm poly}(n){\rm size}(\mathcal{S}^{(1)}))$
such that the probability $\mathcal{S}^{(0)}$ aborts is ${\mathsf{negl}}(\lambda)$,
and the output quantum states of $\mathcal{S}^{(0)}$ and $\mathcal{S}^{(1)}$ are ${\mathsf{negl}}(\lambda)$-close when they do not abort.
From \cref{prop:indistinguishable_S_S'_comp,prop:indistinguishable_S'_V^*_comp},
$\mathcal{S}^{(0)}$ is ${\mathsf{negl}}(\lambda)$-close to the run of the real protocol, which completes the proof.
\end{proof}
\begin{proposition}\label{prop:abort_comp}
If $\Sigma_{\mathsf{ccd}}$ is computational hiding, then the probability that $\mathcal{S}^{(1)}$ does not abort is $\frac{1}{m}\pm{\mathsf{negl}}(\lambda)$.
\end{proposition}
\begin{proposition}\label{prop:indistinguishable_S_S'_comp}
$\mathcal{S}^{(1)}(\mathsf{x},\mathcal{V}^*,\cdot~)\approx_{s}\mathcal{S}^{(2)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\cdot~)$
for any $\mathsf{x}\in A_{\mathsf{yes}}\cap\{0,1\}^\lambda$ and any $\mathsf{w}\in R_A(\mathsf{x})$.
\end{proposition}
\begin{proposition}\label{prop:indistinguishable_S'_V^*_comp}
If $\Sigma_{\mathsf{ccd}}$ is computational hiding,
$\mathcal{S}^{(3)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\cdot~)\approx_{c}\mathsf{OUT}_{\mathcal{V}^*}\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}^*(\cdot)\rangle (\mathsf{x})$.
\end{proposition}
\begin{proof}[Proof of \cref{prop:abort_comp}]
This proof is the same as the proof of \cref{prop:abort}.
\end{proof}
\begin{proof}[Proof of \cref{prop:indistinguishable_S_S'_comp}]
This proof is the same as the proof of \cref{prop:indistinguishable_S_S'_ever}.
\end{proof}
\begin{proof}[Proof of \cref{prop:indistinguishable_S'_V^*_comp}]
We prove the proposition by contradiction.
We construct an adversary $\mathcal{B}$ that breaks the security of the computationally hiding of $\Sigma_{\mathsf{ccd}}$ by assuming
the existence of a distinguisher $\mathcal{D}$ that distinguishes two states $\delta_0$ and $\delta_1$,
\begin{align}
&\delta_0\coloneqq(\mathsf{OUT}_{\mathcal{V}^*}\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}^*(\cdot)\rangle (\mathsf{x})\otimes I)\sigma\\
&\delta_1\coloneqq(\mathcal{S}^{(3)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\cdot~)\otimes I)\sigma,
\end{align}
with a certain state $\sigma$.
Let us describe how $\mathcal{B}$ works.
\begin{enumerate}
\item $\mathcal{B}$ generates $c\leftarrow[m]$ and $x,z\leftarrow\{0,1\}^n$.
\item $\mathcal{B}$ sends $m_0\coloneqq \{x_i,z_i\}_{i\in\overline{S}_c}$ and $m_1\coloneqq0^{2n-10}$ to the challenger.
$\mathcal{B}$ receives commitments from the challenger which is either $\{\mathsf{com}_i(x_i),\mathsf{com}_i(z_i)\}_{i\in \overline{S}_c}$ or $\{\mathsf{com}_i(0),\mathsf{com}_i(0)\}_{i\in \overline{S}_c}$.
\item $\mathcal{B}$ computes
\begin{align}
& (\mathsf{com}_i(x_i),\mathsf{d}_i(x_i),\keys{ck}_i(x_i))\leftarrow\algo{Commit}(1^\lambda,x_i)\\
& (\mathsf{com}_i(z_i),\mathsf{d}_i(z_i),\keys{ck}_i(z_i))\leftarrow\algo{Commit}(1^\lambda,z_i)
\end{align}
for $i\in S_c$ by itself.
\item $\mathcal{B}$ generates $X^xZ^z\rho_{\mathsf{hist}}Z^zX^x$.
$\mathcal{B}$ appends commitments and $\sigma$ to the quantum state.
If the commitments for ${i\in \overline{S}_c}$ are $\{\mathsf{com}_i(x_i),\mathsf{com}_i(z_i)\}_{i\in \overline{S}_c}$,
$\mathcal{B}$ obtains $X^xZ^z\rho_{\mathsf{hist}}Z^zX^x\otimes \mathsf{com}(x)\otimes\mathsf{com}(z)\otimes \sigma$.
If the commitments for ${i\in \overline{S}_c}$ are $\{\mathsf{com}_i(0),\mathsf{com}_i(0)\}_{i\in \overline{S}_c}$,
$\mathcal{B}$ obtains $X^xZ^z\rho_{\mathsf{hist}}Z^zX^x\otimes \mathsf{com}(x^{S_c})\otimes\mathsf{com}(z^{S_c})\otimes \sigma$.
\item $\mathcal{B}$ runs $V_1^*$ on it to obtain $(c',\{\mathsf{cert}_{i,x}',\mathsf{cert}_{i,z}'\}_{i\in \overline{S}_{c'}})$. $\mathcal{B}$ aborts when $c\neq c'$.
\item $\mathcal{B}$ appends $\{\mathsf{d}_i(x_i),\mathsf{d}_{i}(z_i)\}_{i\in S_c}$ to the post-measurement state and runs $V_2^*$ on it.
\item $\mathcal{B}$ passes the output state to $\mathcal{D}$.
\item When $\mathcal{D}$ outputs $b$, $\mathcal{B}$ outputs $b$.
\end{enumerate}
When $\mathcal{B}$ receives $\{\mathsf{com}_i(x_i),\mathsf{com}_i(z_i)\}_{i\in \overline{S}_c}$ from the challenger and it does not abort,
it simulates \\$\mathsf{OUT}_{\mathcal{V}^*}\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}^*(\cdot)\rangle(\mathsf{x})$.
Because $(X^xZ^z\rho_{\mathsf{hist}}Z^zX^x)\otimes \mathsf{com}(x)\otimes\mathsf{com}(z)\otimes \sigma$ is independent of $c$,
the probability that $\mathcal{B}$ does not abort is $\frac{1}{m}$.
Therefore, $\mathcal{B}$ can simulate ${\mathsf{OUT}}_{\mathcal{V}^*}\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}^*(\cdot)\rangle(\mathsf{x})$
with probability $\frac{1}{m}$.
When $\mathcal{B}$ receives $\{\mathsf{com}_i(0),\mathsf{com}_i(0)\}_{i\in\overline{S}_c}$ from the challenger and it does not abort,
it simulates $\mathcal{S}^{(3)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\cdot~)$.
The probability that $\mathcal{B}$ does not abort is $\frac{1}{m}\pm{\mathsf{negl}}(\lambda)$ from \cref{prop:abort_comp,prop:indistinguishable_S_S'_comp}.
Therefore, $\mathcal{B}$ can simulate
$\mathcal{S}^{(3)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\cdot~)$ with probability
$\frac{1}{m}\pm{\mathsf{negl}}(\lambda)$.
Therefore, if there exists the distinguisher $\mathcal{D}$ that distinguishes
$\delta_0$ and $\delta_1$,
$\mathcal{B}$ can distinguish $\{\mathsf{com}_i(x_i),\mathsf{com}_i(z_i)\}_{i\in \overline{S}_c}$ from $\{\mathsf{com}_i(0),\mathsf{com}_i(0)\}_{i\in \overline{S}_c}$.
From (the computational hiding version of) \cref{lemma:eachbit},
this contradicts the computational hiding of $\Sigma_{\mathsf{ccd}}$.
\end{proof}
\section{Commitment with Certified Everlasting Hiding and Classical-Extractor-Based Binding}\label{sec:everlasting_commitment}
In this section, we define and construct commitment with certified everlasting hiding and
statistical binding.
We adopt a non-standard syntax for the verification algorithm and a slightly involved definition for the binding, which we call the classical-extractor-based binding, that are tailored to our construction. This is
because they are convenient for our construction of certified everlasting zero-knowledge proof for $\compclass{QMA}$ given in \cref{sec:ZK}.
We can also construct one with a more standard syntax of verification and binding property, namely, the sum-binding, by essentially the same construction. The detail is given in \cref{app:sum_binding}.
\subsection{Definition}\label{sec:bitcomm_def}
\begin{definition}[Commitment with Certified Everlasting Hiding and Classical-Extractor-Based Binding (Syntax)]\label{def:commitment_with_cd_effective}
Let $\lambda$ be the security parameter and let $p$, $q$, $r$, $s$ and $t$ be some polynomials.
Commitment with certified everlasting hiding and classical-extractor-based binding consists of a tuple of algorithms $(\algo{Commit},\algo{Verify},\algo{Del},\algo{Cert})$
with message space $\mathcal{M}:=\{0,1\}^n$, commitment space $\mathcal{C}:=\mathcal{Q}^{\otimes p(\lambda)}\times \{0,1\}^{q(\lambda)}$,
decommitment space $\mathcal{D}:=\{0,1\}^{r(\lambda)}$, key space $\mathcal{K}:=\{0,1\}^{s(\lambda)}$ and
deletion certificate space $\mathcal{E}:= \{0,1\}^{t(\lambda)}$.
\begin{description}
\item[$\algo{Commit} (1^{\lambda},m)\rightarrow (\mathsf{com},\mathsf{d},\keys{ck})$:]
The commitment algorithm takes as input a security parameter $1^{\lambda}$ and a message $m\in\mathcal{M}$, and outputs a commitment $\mathsf{com}\in\mathcal{C} $,
a decommitment $\mathsf{d}\coloneqq(\mathsf{d}_1,\mathsf{d}_2)\in\mathcal{D}$ and a key $\keys{ck}\in\mathcal{K}$.
Note that $\mathsf{com}$ consists of a quantum state $\psi\in\mathcal{Q}^{\otimes p(\lambda)}$ and a classical bit string $f\in\{0,1\}^{q(\lambda)}$.
\item[$\algo{Verify}(\mathsf{com},\mathsf{d})\rightarrow m'~or~\bot $:]
The verification algorithm consists of two algorithms, $\algo{Verify}_1$ and $\algo{Verify}_2$.
It parses $\mathsf{d}=(\mathsf{d}_1,\mathsf{d}_2)$.
$\algo{Verify}_1$ takes $\mathsf{com}$ and $(\mathsf{d}_1,\mathsf{d}_2)$ as input, and outputs $\top$ or $\bot$.
$\algo{Verify}_2$ takes $\mathsf{com}$ and $\mathsf{d}_1$ as input, and outputs $m'$.
If the output of $\algo{Verify}_1$ is $\bot$, then the output of $\algo{Verify}$ is $\bot$.
Otherwise the output of $\algo{Verify}$ is $m'$.
\item[$\algo{Del}(\mathsf{com})\rightarrow \mathsf{cert}$:]The deletion algorithm takes $\mathsf{com}$ as input, and outputs a certificate $\mathsf{cert}\in \mathcal{E}$.
\item[$\algo{Cert}(\mathsf{cert},\keys{ck})\rightarrow \top~or~\bot$:]The certification algorithm takes $\mathsf{cert}$ and $\keys{ck}$ as input, and outputs $\top$ or $\bot$.
\end{description}
\end{definition}
\begin{definition}[Correctness]\label{def:correctness_bit_commitment_effective}
There are two types of correctness, namely, decommitment correctness and deletion correctness.
\begin{description}
\item[Decommitment correctness:]
There exists a negligible function ${\mathsf{negl}}$ such that for any $\lambda\in\mathbb{N}$ and $m\in\mathcal{M}$,
\begin{align}
\Pr[m\leftarrow\algo{Verify} (\mathsf{com},\mathsf{d})\mid(\mathsf{com},\mathsf{d},\keys{ck})\leftarrow\algo{Commit}(1^{\lambda},m)]\geq1-{\mathsf{negl}}(\lambda).
\end{align}
\item[Deletion correctness:]
There exists a negligible function ${\mathsf{negl}}$ such that for any $\lambda \in\mathbb{N}$ and $m\in\mathcal{M}$,
\begin{align}
\Pr[\top\leftarrow \algo{Cert}(\mathsf{cert},\keys{ck})\mid(\mathsf{com},\mathsf{d},\keys{ck})\leftarrow\algo{Commit}(1^\lambda,m),\mathsf{cert}\leftarrow\algo{Del}(\mathsf{com})]\geq1-{\mathsf{negl}}(\lambda).
\end{align}
\end{description}
\end{definition}
\begin{definition}[Classical-Extractor-Based Binding]\label{def:effevtive_binding}
There exists an unbounded-time deterministic algorithm $\mathsf{Ext}$ that
takes $f \in \{0,1\}^{q(\lambda)}$ of $\mathsf{com}$ as input, and outputs
$\mathsf{d}^{*}_1\leftarrow\mathsf{Ext}(f)$
such that
for any $\mathsf{com}$, any $\mathsf{d}_1\neq\mathsf{d}^*_{1}$, and any $\mathsf{d}_2$,
$
\Pr[\algo{Verify}(\mathsf{com},\mathsf{d}=(\mathsf{d}_1,\mathsf{d}_2))=\bot]= 1.
$
\end{definition}
\if0
\begin{definition}[Computationally hiding]
For any QPT adversary $\mathcal{A}$,
\begin{align}
&\advb{\Sigma,\mathcal{A}}{com}{hid}(\lambda)\coloneqq
\abs{\Pr[1\leftarrow \mathcal{A}(\mathsf{com},m_0,m_1)\mid (\mathsf{com},\mathsf{d},\keys{ck})\leftarrow\algo{Commit}(1^{\lambda},m_0)]\\&-\Pr[1\leftarrow\mathcal{A}(\mathsf{com},m_0,m_1) \mid (\mathsf{com},\mathsf{d},\keys{ck})\leftarrow\algo{Commit}(1^{\lambda},m_1)]}\leq{\mathsf{negl}}(\lambda).
\end{align}
\end{definition}
\fi
\begin{definition}[Computational Hiding]
Let $\Sigma\coloneqq(\algo{Commit},\algo{Verify},\algo{Del},\algo{Cert})$.
Let us consider the following security experiment $\expa{\Sigma,\mathcal{A}}{c}{hide}(\lambda,b)$ against
any QPT adversary $\mathcal{A}$.
\begin{enumerate}
\item $\mathcal{A}$ generates $(m_0,m_1)\in \mathcal{M}^2$ and sends them to the challenger.
\item The challenger computes $(\mathsf{com},\mathsf{d},\keys{ck})\leftarrow \algo{Commit}(1^{\lambda},m_b)$, and sends $\mathsf{com}$ to $\mathcal{A}$.
\item $\mathcal{A}$ outputs $b'\in\{0,1\}$.
\item The output of the experiment is $b'$.
\end{enumerate}
Computational hiding means that the following is satisfied for any QPT $\mathcal{A}$.
\begin{align}
\advb{\Sigma,\mathcal{A}}{c}{hide}(\lambda)\coloneqq
\left|\Pr[ \expa{\Sigma,\mathcal{A}}{c}{hide}(\lambda,0)=1]-
\Pr[\expa{\Sigma,\mathcal{A}}{c}{hide}(\lambda,1)=1]\right|\leq {\mathsf{negl}}(\lambda).
\end{align}
\end{definition}
\begin{definition}[Certified Everlasting Hiding]
Let $\Sigma\coloneqq(\algo{Commit},\algo{Verify},\algo{Del},\algo{Cert})$.
Let us consider the following security experiment $\expa{\Sigma,\mathcal{A}}{ever}{hide}(\lambda,b)$ against
$\mathcal{A}=(\mathcal{A}_1,\mathcal{A}_2)$ consisting of any QPT adversary $\mathcal{A}_1$ and any unbounded adversary $\mathcal{A}_2$.
\begin{enumerate}
\item $\mathcal{A}_1$ generates $(m_0,m_1)\in \mathcal{M}^2$ and sends it to the challenger.
\item The challenger computes $(\mathsf{com},\mathsf{d},\keys{ck})\leftarrow \algo{Commit}(1^{\lambda},m_b)$, and sends $\mathsf{com}$ to $\mathcal{A}_1$.
\item At some point, $\mathcal{A}_1$ sends $\mathsf{cert}$ to the challenger, and sends its internal state to $\mathcal{A}_2$.
\item The challenger computes $\algo{Cert}(\mathsf{cert},\keys{ck})$. If the output is $\top$, then the challenger outputs $\top$,
and sends $(\mathsf{d},\keys{ck})$ to $\mathcal{A}_2$.
Else, the challenger outputs $\bot$, and sends $\bot$ to $\mathcal{A}_2$.
\item $\mathcal{A}_2$ outputs $b'\in\{0,1\}$.
\item If the challenger outputs $\top$, then the output of the experiment is $b'$. Otherwise, the output of the experiment is $\bot$.
\end{enumerate}
We say that it is certified everlasting hiding if the following is satisfied for any
$\mathcal{A}=(\mathcal{A}_1,\mathcal{A}_2)$.
\begin{align}
\advb{\Sigma,\mathcal{A}}{ever}{hide}(\lambda)\coloneqq
\left|\Pr[ \expa{\Sigma,\mathcal{A}}{ever}{hide}(\lambda,0)=1]-
\Pr[\expa{\Sigma,\mathcal{A}}{ever}{hide}(\lambda,1)=1]\right|\leq {\mathsf{negl}}(\lambda).
\end{align}
\end{definition}
\begin{remark}
We remark that certified everlasting hiding does not imply computational hiding since it does not require anything if the adversary does not send a valid certificate.
\end{remark}
The following lemma will be used in the construction of the certified everlasting zero-knowledge proof for $\compclass{QMA}$
in \cref{sec:ZK}.
It is shown with the standard hybrid argument (see \cref{sec:proof_of_bit}).
It is also easy to see that a similar lemma holds for computational hiding.
\begin{lemma}\label{lemma:eachbit}
Let $\Sigma\coloneqq(\algo{Commit},\algo{Verify},\algo{Del},\algo{Cert})$ and $\mathcal{M}=\{0,1\}$.
Let us consider the following security experiment
$\expb{\Sigma,\mathcal{A}}{bit}{ever}{hide}(\lambda, b)$
against
$\mathcal{A}=(\mathcal{A}_1,\mathcal{A}_2)$ consisting of any QPT adversary $\mathcal{A}_1$ and any unbounded adversary $\mathcal{A}_2$.
\begin{enumerate}
\item $\mathcal{A}_1$ generates $(m^0,m^1)\in \{0,1\}^n\times\{0,1\}^n$ and sends it to the challenger.
\item The challenger computes
\begin{align}
(\mathsf{com}_i(m_i^b),\mathsf{d}_i(m_i^b),\keys{ck}_i(m_i^b))\leftarrow \algo{Commit}(1^{\lambda},m^b_i)
\end{align}
for each $i\in[n]$,
and sends $\{\mathsf{com}_i(m_i^b)\}_{i\in[n]}$ to $\mathcal{A}_1$.
Here, $m_i^b$ is the $i$-th bit of $m^b$.
\item At some point, $\mathcal{A}_1$ sends $\{\mathsf{cert}_i\}_{i\in[n]}$ to the challenger,
and sends its internal state to $\mathcal{A}_2$.
\item The challenger computes $\algo{Cert}(\mathsf{cert}_i,\keys{ck}_i(m_i^b))$ for each $i\in[n]$.
If the output is $\top$ for all $i\in[n]$, then the challenger outputs $\top$,
and sends $\{\mathsf{d}_i(m_i^b),\keys{ck}_i(m_i^b)\}_{i\in[n]}$ to $\mathcal{A}_2$.
Else, the challenger outputs $\bot$, and sends $\bot$ to $\mathcal{A}_2$.
\item $\mathcal{A}_2$ outputs $b'\in\{0,1\}$.
\item If the challenger outputs $\top$, then the output of the experiment is $b'$. Otherwise, the output of the experiment is $\bot$.
\end{enumerate}
If $\Sigma$ is certified everlasting hiding,
\begin{align}
\advc{\Sigma,\mathcal{A}}{bit}{ever}{hide}(\lambda)
\coloneqq
\left|\Pr[ \expb{\Sigma,\mathcal{A}}{bit}{ever}{hide}(\lambda,0)=1]-
\Pr[\expb{\Sigma,\mathcal{A}}{bit}{ever}{hide}(\lambda,1)=1]\right|\leq {\mathsf{negl}}(\lambda)
\end{align}
for any $\mathcal{A}=(\mathcal{A}_1,\mathcal{A}_2)$.
\end{lemma}
\if0
\begin{remark}
An intuitive explanation of this definition is as follows:
The verification algorithm $\algo{Verify}$ verifies
whether a commitment $(\mathsf{com},\mathsf{d}_1^*,\mathsf{d}_2)$ is in the correct format or not.
Since the decommitment $\mathsf{d}_1^*$ that passes the verification is unique, even the unbounded adversary cannot change $\mathsf{d}_1^*$.
Therefore, the probability of opening value $b$ does not change after the adversary sends the commitment $\mathsf{com}$.\taiga{Genmitu niha tigau.}
\end{remark}
\fi
\subsection{Construction}\label{sec:bit_comm_construction}
Let $\lambda$ be the security parameter, and let $p$, $q$, $r$, $s$, $t$ and $u$ be some polynomials.
We construct commitment with certified everlasting hiding and classical-extractor-based binding,
$\Sigma_{\mathsf{ccd}}=(\algo{Commit},\algo{Verify},\algo{Del},\algo{Cert})$, with message space $\mathcal{M}=\{0,1\}^n$,
commitment space $\mathcal{C}=\mathcal{Q}^{\otimes p(\lambda)}\times \{0,1\}^{q(\lambda)}\times \{0,1\}^{r(\lambda)}$,
decommitment space $\mathcal{D}=\{0,1\}^{s(\lambda)}\times \{0,1\}^{t(\lambda)}$,
key space $\mathcal{K}=\{0,1\}^{r(\lambda)}$
and deletion certificate space $\mathcal{E}=\{0,1\}^{u(\lambda)}$
from the following primitives:
\begin{itemize}
\item Secret-key encryption with certified deletion,
$\Sigma_{\mathsf{skcd}}=\mathsf{SKE}.(\algo{KeyGen},\algo{Enc},\algo{Dec},\algo{Del},\algo{Verify})$, with plaintext space $\mathcal{M}=\{0,1\}^n$, ciphertext space $\mathcal{C}=\mathcal{Q}^{\otimes p(\lambda)}$, key space $\mathcal{K}=\{0,1\}^{r(\lambda)}$, and deletion certificate space $\mathcal{E}=\{0,1\}^{u(\lambda)}$.
\item
Classical non-interactive commitment, $\Sigma_{\mathsf{com}}=\algo{Classical}.\algo{Commit}$, with plaintext space $\{0,1\}^{s(\lambda)}$, randomness space $\{0,1\}^{t(\lambda)}$, and commitment space $\{0,1\}^{q(\lambda)}$.
\item
A hash function $H$ from $\{0,1\}^{s(\lambda)}$ to $\{0,1\}^{r(\lambda)}$ modeled as a quantumly-accessible random oracle.
\end{itemize}
The construction is as follows.
\begin{description}
\item[$\algo{Commit}(1^{\lambda},m)$:] $ $
\begin{itemize}
\item Generate $\mathsf{ske.sk}\leftarrow\mathsf{SKE}.\algo{KeyGen}(1^{\lambda})$, $R\leftarrow\{0,1\}^{s(\lambda)}$, $R'\leftarrow\{0,1\}^{t(\lambda)}$, and a hash function $H$ from $\{0,1\}^{s(\lambda)}$ to $\{0,1\}^{r(\lambda)}$.
\item Compute $\mathsf{ske}.\keys{CT}\leftarrow \SKE.\algo{Enc}(\mathsf{ske}.\keys{sk},m)$, $f\leftarrow \algo{Classical}.\algo{Commit}(R;R')$, and $h\coloneqq H(R)\oplus\mathsf{ske}.\keys{sk}$.
\item Output $\mathsf{com}\coloneqq (\mathsf{ske}.\keys{CT},f,h)$, $\mathsf{d}_1\coloneqq R$, $\mathsf{d}_2\coloneqq R'$, and $\keys{ck}\coloneqq\mathsf{ske}.\keys{sk}$.
\end{itemize}
\item[$\algo{Verify}_1(\mathsf{com},\mathsf{d}_1,\mathsf{d}_2)$:] $ $
\begin{itemize}
\item Parse $\mathsf{com}=(\mathsf{ske}.\keys{CT},f,h)$, $\mathsf{d}_1= R$, and $\mathsf{d}_2=R'$.
\item Output $\top$ if $f=\algo{Classical}.\algo{Commit}(R;R')$, and output $\bot$ otherwise.
\end{itemize}
\item[$\algo{Verify}_2(\mathsf{com},\mathsf{d}_1)$:] $ $
\begin{itemize}
\item Parse $\mathsf{com}=(\mathsf{ske}.\keys{CT},f,h)$ and $\mathsf{d}_1=R$.
\item Compute $\mathsf{ske}.\keys{sk}'\coloneqq H(R)\oplus h$.
\item Output $m'\leftarrow \SKE.\algo{Dec}(\mathsf{ske}.\keys{sk}',\mathsf{ske}.\keys{CT})$.
\end{itemize}
\item[$\algo{Del}(\mathsf{com})$:] $ $
\begin{itemize}
\item Parse $\mathsf{com}=(\mathsf{ske}.\keys{CT},f,h)$.
\item Compute $\mathsf{ske}.\mathsf{cert}\leftarrow \SKE.\algo{Del}(\mathsf{ske}.\keys{CT})$.
\item Output $\mathsf{cert}\coloneqq\mathsf{ske}.\mathsf{cert}$.
\end{itemize}
\item[$\algo{Cert}(\mathsf{cert},\keys{ck})$:] $ $
\begin{itemize}
\item Parse $\mathsf{cert}=\mathsf{ske}.\mathsf{cert}$ and $\keys{ck}= \mathsf{ske}.\keys{sk}$.
\item Output $\top/\bot\leftarrow \SKE.\algo{Verify}(\mathsf{ske}.\keys{sk},\mathsf{ske}.\mathsf{cert})$.
\end{itemize}
\end{description}
\paragraph{Correctness.}
The decommitment and deletion correctness easily follow from the correctness of $\Sigma_{\mathsf{skcd}}$.
\paragraph{Security.}
We prove the following three theorems.
\begin{theorem}\label{thm:effective_binding}
If $\Sigma_{\mathsf{com}}$ is perfect binding, then $\Sigma_{\mathsf{ccd}}$ is classical-extractor-based binding.
\end{theorem}
\begin{theorem}\label{thm:everlasting_hiding_2}
If $\Sigma_{\mathsf{com}}$ is unpredictable and $\Sigma_{\mathsf{skcd}}$ is OT-CD secure, then $\Sigma_{\mathsf{ccd}}$ is certified everlasting hiding.
\end{theorem}
\begin{theorem}\label{thm:computationally_hiding_2}
If $\Sigma_{\mathsf{com}}$ is unpredictable and $\Sigma_{\mathsf{skcd}}$ is OT-CD secure, then $\Sigma_{\mathsf{ccd}}$ is computationally hiding.
\end{theorem}
\begin{proof}[Proof of Theorem~\ref{thm:effective_binding}]
Due to the perfect binding of $\Sigma_{\mathsf{com}}=\algo{Classical}.\algo{Commit}$, there exists a unique $\mathsf{d}_1^*$ such that $f=\algo{Classical}.\algo{Commit}(\mathsf{d}_1^*;\mathsf{d}_2)$ for a given $f$.
Let $\mathsf{Ext}$ be the algorithm that finds such $\mathsf{d}_1^*$ and outputs it.
(If there is no such $\mathsf{d}_1^*$, then $\mathsf{Ext}$ outputs $\bot$.)
Then, for any $\mathsf{com}=(\mathsf{ske}.\keys{CT},f,h)$, any $\mathsf{d}_1\neq \mathsf{d}_1^*$, and any $\mathsf{d}_2$,
\begin{eqnarray*}
\Pr[\algo{Verify}(\mathsf{com},\mathsf{d}=(\mathsf{d}_1,\mathsf{d}_2))=\bot]\ge\Pr[f\neq \algo{Classical}.\algo{Commit}(\mathsf{d}_1,\mathsf{d}_2)]=1,
\end{eqnarray*}
which completes the proof.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:everlasting_hiding_2}]
For clarity, we describe how the experiment works against an adversary $\mathcal{A}\coloneqq (\mathcal{A}_1,\mathcal{A}_2)$ consisting of any QPT adversary $\mathcal{A}_1$ and any quantum unbounded time adversary $\mathcal{A}_2$.
\begin{description}
\item[$\expa{\Sigma_{\mathsf{ccd}},\mathcal{A}}{ever}{hide}(\lambda,b)$:] This is the original experiment.
\begin{enumerate}
\item A uniformly random function $H$ from $\{0,1\}^{s(\lambda)}$ to $\{0,1\}^{r(\lambda)}$ is chosen. $\mathcal{A}_1$ and $\mathcal{A}_2$ can make arbitrarily many quantum queries to $H$ at any time in the experiment.
\item $\mathcal{A}_1$ chooses $(m_0,m_1)\leftarrow \mathcal{M}^2$, and sends $(m_0,m_1)$ to the challenger.
\item The challenger generates $\mathsf{ske}.\keys{sk}\leftarrow\SKE.\algo{KeyGen}(1^{\lambda})$, $R\leftarrow\{0,1\}^{s(\lambda)}$ and $R'\leftarrow\{0,1\}^{t(\lambda)}$.
The challenger computes $\mathsf{ske}.\keys{CT}\leftarrow\SKE.\algo{Enc}(\mathsf{ske}.\keys{sk},m_b)$, $f\coloneqq \algo{Classical}.\algo{Commit}(R;R')$ and $h\coloneqq H(R)\oplus \mathsf{ske}.\keys{sk}$, and sends $(\mathsf{ske}.\keys{CT},f,h)$ to $\mathcal{A}_1$.
\item $\mathcal{A}_1$ sends $\mathsf{ske}.\mathsf{cert}$ to the challenger and sends its internal state to $\mathcal{A}_2$.
\item If $\top\leftarrow\SKE.\algo{Verify}(\mathsf{ske}.\keys{sk},\mathsf{ske}.\mathsf{cert})$, the challenger outputs $\top$ and sends $(R,R',\mathsf{ske}.\keys{sk})$ to $\mathcal{A}_2$.
Otherwise, the challenger outputs $\bot$ and sends $\bot$ to $\mathcal{A}_2$.
\item $\mathcal{A}_2$ outputs $b'$.
\item If the challenger outputs $\top$, then the output of the experiment is $b'$.
Otherwise, the output of the experiment is $\bot$.
\end{enumerate}
What we have to prove is that
\begin{align}
\advb{\Sigma_{\mathsf{ccd}},\mathcal{A}}{ever}{hide}(\lambda)\coloneqq
\left|\Pr[\expa{\Sigma_{\mathsf{ccd}},\mathcal{A}}{ever}{hide}(\lambda,0)=1]-
\Pr[\expa{\Sigma_{\mathsf{ccd}},\mathcal{A}}{ever}{hide}(\lambda,1)=1]\right|\leq {\mathsf{negl}}(\lambda).
\end{align}
We define the following sequence of hybrids.
\item[$\sfhyb{1}{}(b)$:]
This is identical to $\expa{\Sigma_{\mathsf{ccd}},\mathcal{A}}{ever}{hide}(\lambda,b)$ except that the oracle given to $\mathcal{A}_1$ is replaced with $H_{R\rightarrow H'}$
which is $H$ reprogrammed according to $H'$ on an input $R$ where $H'$ is another independent uniformly random function. More formally, $H_{R\rightarrow H'}$ is defined by
\begin{align}
H_{R\rightarrow H'}(R^*) \coloneqq
\begin{cases}
H(R^*)~~~&(R^*\neq R)\\
H'(R^*)~~~&(R^*=R).
\end{cases}
\end{align}
We note that the challenger still uses $H$ to generate $h$, and the oracle which $\mathcal{A}_2$ uses is still $H$ similarly to the original experiment.
\item[$\sfhyb{2}{}(b)$:]
This is identical to $\sfhyb{1}{}(b)$
except for the following three points.
First, the challenger generates $h$ uniformly at random.
Second, the oracle given to $\mathcal{A}_1$ is replaced with $H'$ which is an independent uniformly random function.
Third, the oracle given to $\mathcal{A}_2$ is replaced with $H'_{R\rightarrow h\oplus \mathsf{ske}.\keys{sk}}$ which is $H'$ reprogrammed to $h\oplus \mathsf{ske}.\keys{sk}$ on an input $R$.
More formally, $H'_{R\rightarrow h\oplus \mathsf{ske}.\keys{sk}}$ is defined by
\begin{align}
H'_{R\rightarrow h\oplus \mathsf{ske}.\keys{sk}}(R^*)\coloneqq
\begin{cases}
H'(R^*)~~~&(R^*\neq R)\\
h\oplus \mathsf{ske}.\keys{sk} ~~~&(R^*=R).
\end{cases}
\end{align}
\end{description}
\begin{proposition}\label{prop:exp=hyb_1}
If $\Sigma_{\mathsf{com}}$ is unpredictable, then
\begin{align}
\abs{\Pr[\expa{\Sigma_{\mathsf{ccd}},\mathcal{A}}{ever}{hide}(\lambda,b)=1]- \Pr[\sfhyb{1}{}(b)=1]}\leq{\mathsf{negl}}(\lambda).
\end{align}
\end{proposition}
\begin{proof}
The proof is similar to \cite[Propositoin~5.8]{EPRINT:HMNY21}, but note that this time we have to consider
an unbounded adversary after the certificate is issued unlike the case of \cite{EPRINT:HMNY21}.
We assume that $\abs{\Pr[\expa{\Sigma_{\mathsf{ccd}},\mathcal{A}}{ever}{hide}(\lambda,b)=1]- \Pr[\sfhyb{1}{}(b)=1]}$ is non-negligible,
and construct an adversary $\mathcal{B}$ that breaks the unpredictability of $\Sigma_{\mathsf{com}}$.
For notational simplicity, we denote $\expa{\Sigma_{\mathsf{ccd}},\mathcal{A}}{ever}{hide}(\lambda,b)$ by $\sfhyb{0}{}(b)$.
We consider an algorithm $\widetilde{\mathcal{A}}$ that works as follows. $\widetilde{\mathcal{A}}$ is given an oracle $\mathcal{O}$,
which is either $H$ or $H_{R\rightarrow H'}$, and an input $z$ that consists of $R$ and the whole truth table of $H$, where $R\leftarrow\{0,1\}^{s(\lambda)}$,
and $H$ and $H'$ are uniformly random functions. $\widetilde{\mathcal{A}}$ runs $\sfhyb{0}{}(b)$ except that it uses its own oracle $\mathcal{O}$ to simulate $\mathcal{A}_1$'s random oracle queries.
On the other hand, $\widetilde{\mathcal{A}}$ uses $H$ to simulate $h$ and $\mathcal{A}_2$'s random oracle queries regardless of $\mathcal{O}$, which is possible because the truth table of $H$ is included in the input $z$.
By definition, we have
\begin{align}
\Pr[\sfhyb{0}{}(b)=1]=\Pr[\widetilde{\mathcal{A}}^{H}(R,H)=1]
\end{align}
and
\begin{align}
\Pr[\sfhyb{1}{}(b)=1]=\Pr[\widetilde{\mathcal{A}}^{H_{R\rightarrow H'}}(R,H)=1]
\end{align}
where $H$ in the input means the truth table of $H$.
We apply the one-way to hiding lemma (\cref{lemma:one-way_to_hiding}) to the above $\widetilde{\mathcal{A}}$.
Note that $\widetilde{\mathcal{A}}$ is inefficient, but the one-way to hiding lemma is applicable to inefficient algorithms. Then if we let $\widetilde{\mathcal{B}}$ be the algorithm that measures uniformly chosen query of $\widetilde{\mathcal{A}}$,
we have
\begin{align}
\abs{\Pr[\widetilde{\mathcal{A}}^H(R,H)=1]-\Pr[\widetilde{\mathcal{A}}^{H_{R\rightarrow H'}}(R,H)=1]}\leq 2q\sqrt{\Pr[\widetilde{\mathcal{B}}^{H_{R\rightarrow H'}}(R,H)=R]}.
\end{align}
By the assumption, the LHS is non-negligible, and thus $\Pr[\widetilde{\mathcal{B}}^{H_{R\rightarrow H'}}(R,H)=R]$ is non-negligible.
Let $\widetilde{\mathcal{B}}'$ be the algorithm that is the same as $\widetilde{\mathcal{B}}$ except that it does not take the truth table of $H$ as input,
and sets $h$ to be uniformly random string instead of setting $h\coloneqq H(R)\oplus \mathsf{ske}.\keys{sk}$. Then we have
\begin{align}
\Pr[\widetilde{\mathcal{B}}^{H_{R\rightarrow H'}}(R,H)=R]=\Pr[\widetilde{\mathcal{B}}'^{H_{R\rightarrow H'}}(R)=R].
\end{align}
The reason is as follows:
First, $\widetilde{\mathcal{B}}$ uses the truth table of $H$ only for generating $h\coloneqq H(R)\oplus\mathsf{ske}.\keys{sk}$,
because it halts before $\widetilde{\mathcal{B}}$ simulates $\mathcal{A}_2$.
Second, the oracle $H_{R\rightarrow H'}$ reveals no information about $H(R)$,
and thus $h$ can be independently and uniformly random.
Moreover, for any fixed $R$, when $H$ and $H'$ are uniformly random, $H_{R\rightarrow H'}$ is also a uniformly random function, and therefore we have
\begin{align}
\Pr[\widetilde{\mathcal{B}}'^{H_{R\rightarrow H'}}(R)=R]=\Pr[\widetilde{\mathcal{B}}'^H(R)=R].
\end{align}
Since $\Pr[\widetilde{\mathcal{B}}^{H_{R\rightarrow H'}}(R,H)=R]$ is non-negligible, $\Pr[\widetilde{\mathcal{B}}'^{H}(R)=R]$ is also non-negligible.
Recall that $\widetilde{\mathcal{B}}'^H$ is an algorithm that simulates $\sfhyb{0}{}(b)$ with the modification that $h$ is set to be uniformly random and measures randomly chosen $\mathcal{A}_1$'s query.
Then it is straightforward to construct an adversary $\mathcal{B}$ that breaks the unpredictability of $\Sigma_{\mathsf{com}}$ by using $\widetilde{\mathcal{B}}'$.
For clarity, let us give the description of $\mathcal{B}$ as follows.
$\mathcal{B}$ is given $\algo{Classical}.\algo{Commit}(R;R')$ from the challenger of
$\mathsf{Exp}_{\Sigma_{\mathsf{com}},\mathcal{B}}^{\mathsf{unpre}}(\lambda)$.
$\mathcal{B}$ chooses $i\leftarrow[q]$ and runs $\sfhyb{1}{}(b)$
until $\mathcal{A}_1$ makes $i$-th random oracle query
or $\mathcal{A}_1$ sends the internal state to $\mathcal{A}_2$,
where $\mathcal{B}$ embeds the problem instance $\algo{Classical}.\algo{Commit}(R;R')$ into those sent to $\mathcal{A}_1$ instead of generating it by itself.
$\mathcal{B}$ measures the $i$-th random oracle query by $\mathcal{A}_1$, and outputs the measurement outcome.
Note that $\mathcal{B}$ can efficiently simulate the random oracle $H$ by Zhandry's compressed oracle technique~\cite{C:Zhandry19}.
It is clear that the probability that $\mathcal{B}$ outputs $R$ is exactly $\Pr[\widetilde{\mathcal{B}}'^{H}(R)=R]$, which is non-negligible.
This contradicts the unpredictability of $\Sigma_{\mathsf{com}}$.
Therefore $\abs{\Pr[\sfhyb{0}{}(b)=1]-\Pr[\sfhyb{1}{}(b)=1]}$ is negligible.
\end{proof}
\begin{proposition}\label{prop:hyb_1=hyb_2}
$\Pr[\sfhyb{1}{}(b)=1]=\Pr[\sfhyb{2}{}(b)=1]$.
\end{proposition}
\begin{proof}
First, let us remind the difference between $\sfhyb{1}{}(b)$ and $\sfhyb{2}{}(b)$.
In $\sfhyb{1}{}(b)$, $\mathcal{A}_1$ receives $h=\mathsf{ske}.\keys{sk}\oplus H(R)$. Moreover, $\mathcal{A}_1$ can access to the random oracle $H_{R\rightarrow H'}$,
and $\mathcal{A}_2$ can access to the random oracle $H$.
On the other hand, in $\sfhyb{2}{}(b)$, $\mathcal{A}_1$ receives uniformly random $h$.
Moreover, $\mathcal{A}_1$ can access to the random oracle $H'$ instead of $H_{R\rightarrow H'}$,
and $\mathcal{A}_2$ can access to the random oracle $H'_{R\rightarrow h\oplus \mathsf{ske}.\keys{sk}}$ instead of $H$.
Let $\Pr[(h,H_{R\rightarrow H'},H)=(r,G,G') \mid \sfhyb{1}{}(b)]$
be the probability
that the adversary $\mathcal{A}_1$ in $\sfhyb{1}{}(b)$ receives a classical bit string $r$ as $h$,
random oracle which $\mathcal{A}_1$ can access to is $G$,
and random oracle which $\mathcal{A}_2$ can access to is $G'$.
Similarly, let us define $\Pr[(h,H',H'_{R\rightarrow h\oplus \mathsf{ske}.\keys{sk}})=(r,G,G') \mid \sfhyb{2}{}(b)]$
for $\sfhyb{2}{}(b)$.
What we have to show is that the following equation holds for any $(r,G,G')$
\begin{align}
\Pr[(h,H_{R\rightarrow H'},H)=(r,G,G') \mid \sfhyb{1}{}(b)]=\Pr[(h,H',H'_{R\rightarrow h\oplus \mathsf{ske}.\keys{sk}})=(r,G,G') \mid \sfhyb{2}{}(b)].
\end{align}
Since $h=\mathsf{ske}.\keys{sk}\oplus H(R)$ in $\sfhyb{1}{}(b)$, $H$ is a uniformly random function, and $h$ in $\sfhyb{2}{}(b)$ is uniformly generated,
\begin{align}
\Pr[h=r \mid \sfhyb{1}{}(b)]=\Pr[h=r \mid \sfhyb{2}{}(b)]
\end{align}
holds for any $r$.
For any classical bit string $r$ and any random oracle $G$, we have
\begin{align}
\Pr[H_{R\rightarrow H'}=G \mid h=r,\sfhyb{1}{}(b)]=\Pr[H'=G \mid h=r,\sfhyb{2}{}(b)].
\end{align}
This is shown as follows.
First, in $\sfhyb{1}{}(b)$, from the construction of $H_{R\rightarrow H'}$, $H_{R\rightarrow H'}(R)$ is independent from $h$ for any $R\in\{0,1\}^{s(\lambda)}$.
Furthermore, since $H$ and $H'$ is random oracle, $H_{R\rightarrow H'}(R)$ is uniformly random for any $R\in \{0,1\}^{s(\lambda)}$.
Second, in $\sfhyb{2}{}(b)$, from the construction of $H'$, $H'(R)$ is independent from $h$ for any $R\in\{0,1\}^{s(\lambda)}$.
Furthermore, since $H'$ is random oracle, $H'(R)$ is uniformly random for any $R\in \{0,1\}^{s(\lambda)}$.
Therefore, we have the above equation.
For any classical bit string $r$ and any random oracles $G$ and $G'$, we have
\begin{align}
\Pr[H=G' \mid (h,H_{R\rightarrow H'})=(r,G),\sfhyb{1}{}(b)]=\Pr[H'_{R\rightarrow h\oplus \mathsf{ske}.\keys{sk}}=G' \mid (h,H')=(r,G),\sfhyb{2}{}(b)].
\end{align}
This can be shown as follows.
First, in $\sfhyb{1}{}(b)$,
we obtain $H(R)=r\oplus \mathsf{ske}.\keys{sk}$,
because $h\coloneqq \mathsf{ske}.\keys{sk}\oplus H(R)$ and $h=r$.
Furthermore, from the definition of $H_{R\rightarrow H'}$, we obtain $H(R^*)=G(R^*)$ for $R^*\neq R$.
Second, in $\sfhyb{2}{}(b)$,
from the definition of $H'_{R\rightarrow h\oplus \mathsf{ske}.\keys{sk}}$, we have $H'_{R\rightarrow h\oplus \mathsf{ske}.\keys{sk}}(R)=r\oplus \mathsf{ske}.\keys{sk}$ and
$H'_{R\rightarrow h\oplus \mathsf{ske}.\keys{sk}}(R^*)=G(R^*)$ for $R^*\neq R$.
From all above discussions, we have
\begin{align}
\Pr[(h,H_{R\rightarrow H'},H)=(r,G,G') \mid \sfhyb{1}{}(b)]=\Pr[(h,H',H'_{R\rightarrow h\oplus \mathsf{ske}.\keys{sk}})=(r,G,G') \mid \sfhyb{2}{}(b)].
\end{align}
\end{proof}
\begin{proposition}\label{prop:hyb_2}
If $\Sigma_{\mathsf{skcd}}$ is OT-CD secure, then
\begin{align}
\abs{\Pr[\sfhyb{2}{}(1)=1]-\Pr[\sfhyb{2}{}(0)=1]}\leq{\mathsf{negl}}(\lambda).
\end{align}
\end{proposition}
\begin{proof}
To show this, we assume that $\abs{\Pr[\sfhyb{2}{}(1)=1]-\Pr[\sfhyb{2}{}(0)=1]}$ is non-negligible,
and construct an adversary $\mathcal{B}$ that breaks the OT-CD security of $\Sigma_{\mathsf{skcd}}$.
$\mathcal{B}$ plays the experiment $\expb{\Sigma_{\mathsf{skcd}},\mathcal{B}}{otsk}{cert}{del}(\lambda,b')$ for some $b'\in\{0,1\}$.
First, $\mathcal{B}$ sends $(m_0,m_1)\in \mathcal{M}^2$ to the challenger of $\expb{\Sigma_{\mathsf{skcd}},\mathcal{B}}{otsk}{cert}{del}(\lambda,b')$.
$\mathcal{B}$ receives $\mathsf{ske}.\keys{CT}$ from the challenger of $\expb{\Sigma_{\mathsf{skcd}},\mathcal{B}}{otsk}{cert}{del}(\lambda,b')$.
$\mathcal{B}$ generates $R\leftarrow\{0,1\}^{s(\lambda)}$, $R'\leftarrow\{0,1\}^{t(\lambda)}$ and $h\leftarrow\{0,1\}^{r(\lambda)}$, and computes $f\coloneqq\algo{Classical}.\algo{Commit}(R;R')$.
$\mathcal{B}$ sends $(\mathsf{ske}.\keys{CT},f,h)$ to $\mathcal{A}_1$.
$\mathcal{B}$ simulates the random oracle $H'$ given to $\mathcal{A}_1$ by itself.
At some point, $\mathcal{A}_1$ sends $\mathsf{ske}.\mathsf{cert}$ to $\mathcal{B}$, and sends the internal state to $\mathcal{A}_2$.
$\mathcal{B}$ passes $\mathsf{ske}.\mathsf{cert}$ to the challenger of $\expb{\Sigma_{\mathsf{skcd}},\mathcal{B}}{otsk}{cert}{del}(\lambda,b')$.
The challenger of
$\expb{\Sigma_{\mathsf{skcd}},\mathcal{B}}{otsk}{cert}{del}(\lambda,b')$
runs $\SKE.\algo{Verify}(\mathsf{ske}.\keys{sk},\mathsf{ske}.\mathsf{cert})\to\top/\bot$.
If it is $\top$,
the challenger sends $\mathsf{ske}.\keys{sk}$ to $\mathcal{B}$.
In that case, $\mathcal{B}$ outputs $\top$, and sends $(R,R',\mathsf{ske}.\keys{sk})$ to $\mathcal{A}_2$.
We denote this event by $\mathsf{Reveal_{sk}}(b')$.
$\mathcal{B}$ simulates $\mathcal{A}_2$, and outputs the output of $\mathcal{A}_2$.
On the other hand, if
$\SKE.\algo{Verify}(\mathsf{ske}.\keys{sk},\mathsf{ske}.\mathsf{cert})\to\bot$,
then the challenger sends $\bot$ to $\mathcal{B}$.
In that case,
$\mathcal{B}$ outputs $\bot$ and aborts.
Note that $\mathcal{B}$ can simulate the random oracle
$H'_{R\rightarrow h\oplus \mathsf{ske}.\keys{sk}}$
given to $\mathcal{A}_2$ when $\mathcal{B}$ does not abort,
because $\mathcal{B}$ receives $\mathsf{ske}.\keys{sk}$ from the challenger of $\expb{\Sigma_{\mathsf{skcd}},\mathcal{B}}{otsk}{cert}{del}(\lambda,b')$ when $\mathcal{B}$ does not abort.
\if0
From all above discussion, we can see that
\begin{itemize}
\item if $b'=0$, $\mathcal{B}$ perfectly simulates the challenger of $\sfhyb{2}{}(0)$ and
\item if $b'=1$, $\mathcal{B}$ perfectly simulates the challenger of $\sfhyb{2}{}(1)$
\end{itemize}
when $\mathcal{B}_{\mathsf{skcd}}$ does not abort.
\fi
Now we have
\begin{align}
&\advc{\Sigma_{\mathsf{skcd}},\mathcal{B}}{otsk}{cert}{del}(\lambda)\\
& \coloneqq\abs{\Pr[ \expb{\Sigma_{\mathsf{skcd}},\mathcal{B}}{otsk}{cert}{del}(\lambda, b')=1\mid b'=0] -
\Pr[ \expb{\Sigma_{\mathsf{skcd}},\mathcal{B}}{otsk}{cert}{del}(\lambda, b')=1\mid b'=1] }\\
&=\abs{\Pr[ \mathcal{B}=1\wedge\mathsf{Reveal_{sk}}(b') \mid b'=0] -
\Pr[ \mathcal{B}=1\wedge\mathsf{Reveal_{sk}}(b') \mid b'=1] }\\
&=\abs{\Pr[ \mathcal{A}_{2}=1\wedge\mathsf{Reveal_{sk}}(b') \mid b'=0] -
\Pr[ \mathcal{A}_{2}=1\wedge\mathsf{Reveal_{sk}}(b') \mid b'=1] }\\
&=|\Pr[\sfhyb{2}{}(b')=1\wedge\mathsf{Reveal_{sk}}(b')\mid b'=0]-\Pr[\sfhyb{2}{}(b')=1\wedge\mathsf{Reveal_{sk}}(b')\mid b'=1]|\\
&=|\Pr[\sfhyb{2}{}(b')=1\mid b'=0]-\Pr[\sfhyb{2}{}(b')=1\mid b'=1]|\\
&=|\Pr[\sfhyb{2}{}(0)=1]-\Pr[\sfhyb{2}{}(1)=1]|.
\end{align}
In the second equation, we have used the fact that $\expb{\Sigma_{\mathsf{skcd}},\mathcal{B}}{otsk}{cert}{del}(\lambda, b)=1$ if and only if $\mathcal{B}=1$ and the challenger of $\expb{\Sigma_{\mathsf{skcd}},\mathcal{B}}{otsk}{cert}{del}(\lambda,b')$ outputs $\top$.
In the third equation, we have used the fact that the output of $\mathcal{B}$ is equal to the output of $\mathcal{A}_2$ conditioned that $\mathsf{Reveal_{sk}}(b')$ occurs.
In the fourth equation, we have used the fact that $\mathcal{B}$ simulates the challenger of $\sfhyb{2}{}(b)$ when $\mathsf{Reveal_{sk}}(b)$ occurs.
In the fifth equation, we have used the fact that $\sfhyb{2}{}(b)=1$ only when $\mathsf{Reveal_{sk}}(b)$ occurs.
Since $|\Pr[\sfhyb{2}{}(0)=1]-\Pr[\sfhyb{2}{}(1)=1]|$ is non-negligible,
$\advc{\Sigma_{\mathsf{skcd}},\mathcal{B}}{otsk}{cert}{del}(\lambda)$ is non-negligible.
This contradicts the OT-CD security of $\Sigma_{\mathsf{skcd}}$.
\end{proof}
By \cref{prop:exp=hyb_1,prop:hyb_1=hyb_2,prop:hyb_2}, we immediately obtain
\cref{thm:everlasting_hiding_2}.
\end{proof}
\ifnum1=1
\begin{proof}[Proof of \cref{thm:computationally_hiding_2}]
This can be shown similarly to \cref{thm:everlasting_hiding_2}.
For the convenience of readers, we provide a proof in \cref{Sec:Computational_hiding}.
\end{proof}
\else
\input{Proof_C_hiding}
\fi
\section{Introduction}
\subsection{Background}\label{sec:background}
Zero-knowledge~\cite{SICOMP:GolMicRac89},
which roughly states that the verifier cannot learn anything beyond the validity of the statement,
is one of the most important concepts in cryptography and computer science.
The study of zero-knowledge has a long history in classical cryptography, and recently there have been many results
in quantum cryptography.
In known constructions of classical zero-knowledge protocols for $\compclass{NP}$, either of zero-knowledge or soundness holds only
against computationally bounded adversaries.
Indeed, achieving both statistical zero-knowledge and statistical soundness at the same time with classical verifier
is impossible for $\compclass{NP}$ unless the polynomial-time hierarchy collapses~\cite{STOC:Fortnow87}.
It is also believed to be impossible even with a quantum verifier~\cite{MW18}.
Broadbent and Islam~\cite{TCC:BroIsl20} recently suggested an idea of the novel compromise:
realizing ``everlasting zero-knowledge'' by using quantum encryption with certified deletion.
The everlasting security defined by Unruh~\cite{C:Unruh13} states that the protocol remains secure as long as
the adversary runs in polynomial-time during the execution of the protocol.
Quantum encryption with certified deletion introduced by Broadbent and Islam~\cite{TCC:BroIsl20} is a new quantum cryptographic primitive where
a classical message is encrypted into a quantum ciphertext,
and the receiver in possession of a quantum ciphertext can generate a classical certificate that shows that
the receiver has deleted the quantum ciphertext.
If the certificate is valid, the receiver can no longer decrypt the message even if it receives the secret key.
Broadbent and Islam’s idea is to use quantum commitment with a similar certified deletion security
to encrypt the first message from the prover to the verifier in the standard $\Sigma$-protocol.
Once the verifier issues the deletion certificate for all commitments that are not opened by the verifier's challenge,
even an unbounded verifier can no longer access the committed values of the unopened commitments.
They left the formal definition and the construction as future works.
There are many obstacles to realizing their idea.
First, their quantum encryption with certified deletion cannot be directly used in a $\Sigma$-protocol because
it does not have any binding property. Their ciphertext consists of a classical and quantum part. The classical part is $m\oplus u\oplus H(r)$,
where $m$ is the plaintext, $u$ and $r$ are random bit strings, and $H$ is a hash function. The quantum part is
a random BB84 states whose computational basis states encode $r$.
The decryption key is $u$ and the place of computational basis states that encode $r$,
and therefore it is not binding: by changing $u$, a different message can be obtained.
We therefore need to extend quantum encryption with certified deletion
in such a way that the statistical binding property is included.
Second, defining a meaningful notion of ``everlasting zero-knowledge proof'' itself is non-trivial.
In fact, everlasting zero-knowledge proofs for $\compclass{QMA}$ or even for $\compclass{NP}$ in the sense of Unruh's definition~\cite{C:Unruh13}
are unlikely to exist.\footnote{We mention that everlasting zero-knowledge \emph{arguments}, which only satisfy computational soundness, can exist. Indeed, any statistical zero-knowledge argument is everlasting zero-knowledge argument. One may think that the computational soundness is fine since that ensures everlasting soundness in the sense of Unruh's definition~\cite{C:Unruh13}. For practical purposes, this may be true. On the other hand, we believe that it is theoretically interesting to pursue (a kind of) everlasting zero-knowledge without compromising the soundness as is done in this paper.}
To see this, recall that the definition of quantum statistical zero-knowledge \cite{SIAM:Wat09,MW18} requires a simulator to simulate the view of a \emph{quantum polynomial-time} malicious verifier in a statistically indistinguishable manner. Therefore, everlasting zero-knowledge in the sense of Unruh's definition~\cite{C:Unruh13} is actually equivalent to quantum statistical zero-knowledge. On the other hand, as already mentioned, it is believed that quantum statistical zero-knowledge proofs for $\compclass{NP}$ do not exist \cite{MW18}. In particular, Menda and Watrous~\cite{MW18} constructed an oracle relative to which quantum statistical zero-knowledge proofs for (even a subclass of) $\compclass{NP}$ do not exist.
However, we notice that this argument does not go through for \emph{certified everlasting zero-knowledge}, where
the verifier can issue a classical certificate that shows that the verifier has deleted its information.
Once a valid certificate has been issued, even unbounded malicious verifier can no longer learn anything
beyond the validity of the statement.
The reason is that certified everlasting zero-knowledge does not imply statistical zero-knowledge since it does not ensure any security against a malicious verifier that refuses to provide a valid certificate of deletion.
Therefore, we have the following question.
\begin{center}
\emph{
Is it possible to define and construct a certified everlasting zero-knowledge proof for $\compclass{QMA}$?
}
\end{center}
\subsection{Our Results}
In this work, we define and construct the certified everlasting zero-knowledge proof for $\compclass{QMA}$.
This goal is achieved in the following four steps.
\begin{enumerate}
\item
We define a new quantum cryptographic primitive, which we call
{\it commitment with statistical binding and certified everlasting hiding}.
In this new commitment scheme, binding is statistical but hiding is computational.
However, the hiding property becomes statistical once the receiver has issued a valid certificate that
shows that the receiver has deleted the committed information.
\item
We construct commitment with statistical binding and certified everlasting hiding.
We use secret-key quantum encryption with certified deletion as the building block in a black box way .
This construction is secure in the quantum random oracle model~\cite{AC:BDFLSZ11}.
\item
We define a new notion of zero-knowledge proof, which we call \emph{the certified everlasting zero-knowledge proof for} $\compclass{QMA}$.
It is a computational zero-knowledge proof for $\compclass{QMA}$ with the following additional property. A verifier can issue a classical
certificate that shows that the verifier has deleted its information. If the certificate is valid,
even unbounded malicious verifier can no longer learn anything beyond the validity of the statement.
\item
We apply commitment with statistical binding and certified everlasting hiding
to the quantum $\Sigma$-protocol for $\compclass{QMA}$ by Broadbent and Grilo~\cite{FOCS:BroGri20}
to construct the certified everlasting zero-knowledge proof for $\compclass{QMA}$.
\end{enumerate}
We have three remarks on our results.
First, although our main results are the definition and the construction of the certified everlasting zero-knowledge
proof for $\compclass{QMA}$, our commitment with statistical binding and certified everlasting hiding itself is also of independent interest.
There will be many other useful applications beyond zero-knowledge.
In fact, it is known that binding and hiding cannot be made statistical at the same time
even in the quantum world~\cite{LC97,Mayers97},
and therefore our new commitment scheme provides a nice compromise.
Second, our new commitment scheme and the new zero-knowledge proof
are the first cryptographic applications of symmetric-key quantum encryption with certified deletion.
Although certified deletion is conceptually very interesting, there was no concrete construction of cryptographic
applications when it was first introduced~\cite{TCC:BroIsl20}.
One reason why the applications are limited is that
in cryptography it is not natural to consider the case when the receiver receives the private key later.
Hiroka et al.~\cite{EPRINT:HMNY21} recently extended the symmetric-key scheme by Broadbent and Islam~\cite{TCC:BroIsl20} to a public-key encryption scheme, an attribute-based encryption scheme, and a publicly verifiable scheme, which have opened many applications.
However, one disadvantage is that their security is the computational one unlike the symmetric-key scheme~\cite{TCC:BroIsl20}.
Therefore it was open whether there is any cryptographic application of the information-theoretically secure certified deletion scheme.
Our results provide the first cryptographic applications of it.
Interestingly, the setup of the symmetric-key scheme~\cite{TCC:BroIsl20}, where the receiver does not have the private key in advance,
nicely fits into the framework of the $\Sigma$-protocol, because the verifier (receiver) in the $\Sigma$-protocol does not have
the decryption key of the first encrypted message from the prover (sender).
Finally, note that certified everlasting zero-knowledge and certified everlasting hiding
seem to be impossible in the classical world, because a malicious adversary can copy its information.
In particular, certified everlasting zero-knowledge against classical verifiers clearly implies honest-verifier statistical zero-knowledge since an honest verifier runs in polynomial-time.\footnote{A similar argument does not work for quantum verifiers since the honest-verifier quantum statistical zero-knowledge~\cite{FOCS:Watrous02} requires a simulator
to simulate honest verifier's internal state \emph{at any point} of the protocol execution. This is not implied by certified everlasting zero-knowledge, which only requires security after generating a valid deletion certificate.} Moreover, it is known that $\compclass{HVSZK}=\compclass{SZK}$ where $\compclass{HVSZK}$ and $\compclass{SZK}$ are languages that have honest-verifier statistical zero-knowledge proofs and (general) statistical zero-knowledge proofs, respectively~\cite{STOC:GolSahVad98}.
Therefore, if certified everlasting zero-knowledge proofs for $\compclass{NP}$ with classical verification exist, we obtain
$\compclass{NP}\subseteq\compclass{HVSZK}=\compclass{SZK}$, which means the collapse of the polynomial-time hierarchy~\cite{STOC:Fortnow87}.
Though the above argument only works for protocols in the standard model, no construction of honest-verifier statistical zero-knowledge proofs for $\compclass{NP}$ is known in the random oracle model either.
Our results therefore add novel items to the list of quantum cryptographic primitives that can be achieved only in the quantum world.
\subsection{Technical Overview}
\paragraph{Certified everlasting zero-knowledge.}
As explained in \cref{sec:background}, everlasting zero-knowledge proofs for $\compclass{NP}$ (and for $\compclass{QMA}$) seem impossible even with quantum verifiers. Therefore, we introduce a relaxed notion of zero-knowledge which we call \emph{certified everlasting zero-knowledge} inspired by quantum encryption with certified deletion introduced by Broadbent and Islam~\cite{TCC:BroIsl20}.
Certified everlasting zero-knowledge ensures security against malicious verifiers that run in polynomial-time during the protocol and provide a valid certificate that sensitive information is ``deleted''.
(For the formal definition, see \cref{sec:def_everlasting_ZK}.)
The difference from everlasting zero-knowledge is that it does not ensure security against malicious verifiers that do not provide a valid certificate. We believe that this is still a meaningful security notion since if the verifier refuses to provide a valid certificate, the prover may penalize the verifier for cheating.
\paragraph{Quantum commitment with certified everlasting hiding.}
Our construction of certified everlasting zero-knowledge proofs is based on the idea sketched by Broadbent and
Islam~\cite{TCC:BroIsl20}.
(For the details of the construction, see \cref{sec:3_construction}.)
The idea is to implement a $\Sigma$-protocol using a commitment scheme with certified deletion. However, they did not give a construction or definition of commitment with certified deletion. First, we remark that the encryption with certified deletion in \cite{TCC:BroIsl20} cannot be directly used as a commitment. A natural way to use their scheme as a commitment scheme is to consider a ciphertext as a commitment. However, since different secret keys decrypt the same ciphertext into different messages, this does not satisfy the binding property as commitment.
A natural (failed) attempt to fix this problem is to add a classical commitment to the secret key of the encryption scheme with certified deletion making use of the fact that the secret key of the encryption with certified deletion in \cite{TCC:BroIsl20} is classical.
That is, a commitment to a message $m$ consists of
$$
(\keys{CT}=\algo{Enc}(\keys{sk},m),\mathsf{com}=\algo{Commit}(\keys{sk}))
$$
where $\algo{Enc}$ is the encryption algorithm of the scheme in \cite{TCC:BroIsl20}, $\keys{sk}$ is its secret key, and $\algo{Commit}$ is a statistically binding and computationally hiding classical commitment scheme.
This resolves the issue of binding since the secret key is now committed by the classical commitment scheme. On the other hand, we cannot prove a hiding property that is sufficiently strong for achieving certified everlasting zero-knowledge. It is not difficult to see that what we need here is \emph{certified everlasting hiding}, which ensures that once a receiver generates a valid certificate that it deleted the commitment in a polynomial-time, the hiding property is ensured even if the receiver runs in unbounded-time afterwards. Unfortunately, we observe that the above generic construction seems insufficient for achieving certified everlasting hiding.\footnote{One may think that we can just use statistically hiding commitment. However, such a commitment can only satisfy computational binding, which is not sufficient for achieving certified everlasting zero-knowledge \emph{proofs} rather than arguments.}
The reason is as follows: We want to reduce the certified everlasting hiding to the certified deletion security of $\algo{Enc}$. However, the security of $\algo{Enc}$ can be invoked only if $\keys{sk}$ is information theoretically hidden before the deletion. On the other hand, $\keys{sk}$ is committed by a statistically binding commitment in the above construction, and thus $\keys{sk}$ is information theoretically determined from the commitment.
Therefore, we have to somehow delete the information of $\keys{sk}$ from the commitment in some hybrid game in a security proof. A similar issue was dealt with by Hiroka et al. \cite{EPRINT:HMNY21} by using receiver non-committing encryption in the context of public key encryption with certified deletion. However, their technique inherently relies on the assumption that an adversary runs in polynomial-time \emph{even after the deletion}. Therefore, their technique is not applicable in the context of certified everlasting hiding.
To overcome the above issue, we rely on random oracles. We modify the above construction as follows:
$$
(\keys{CT}=\algo{Enc}(\keys{sk},m),\mathsf{com}=\algo{Commit}(R),H(R)\oplus \keys{sk})
$$
where $R$ is a sufficiently long random string and $H$ is a hash function modeled as a random oracle whose output length is the same as that of $\keys{sk}$.
We give an intuition on why the above issue is resolved with this modification. As explained in the previous paragraph, we want to delete the information of $\keys{sk}$ from the commitment in some hybrid game. By the computational hiding of commitment, a polynomial-time receiver cannot find $R$ from $\algo{Commit}(R)$. Therefore, it cannot get any information on $H(R)$ since otherwise we can ``extract'' $R$ from one of receiver's queries.
This argument can be made rigorous by using the one-way to hiding lemma~\cite{JACM:Unruh15,C:AmbHamUnr19}. Importantly, we only have to assume that the receiver runs in polynomial-time \emph{before the deletion} and do not need to assume anything about the running time after the deletion because we extract $R$ from one of the queries before the deletion.
Since $\keys{sk}$ is masked by $H(R)$, the receiver cannot get any information on $\keys{sk}$ either. Thus, we can simulate the whole commitment $(\keys{CT},\mathsf{com},H(R)\oplus \keys{sk})$ without using $\keys{sk}$, which resolves the issue and enables us to reduce certified everlasting hiding to certified deletion security of $\algo{Enc}$.
We remark that quantum commitments in general cannot satisfy the binding property in the classical sense.
Indeed, if a malicious sender generates a superposition of valid commitments on different messages $m_0$ and $m_1$, it can later open to $m_0$ or $m_1$ with probability $1/2$ for each. Defining a binding property for quantum commitments is non-trivial, and there have been proposed various flavors of definitions in the literature, e.g., \cite{TCC:CDMS04,C:DamFehSal04,C:DFRSS07,Yan20a,BB21}. It might be possible to adopt some of those definitions.
However, we choose to introduce a new definition, which we call
\emph{classical-extractor-based binding},
tailored to our construction because this is more convenient for our purpose.
Classical-extractor-based binding captures the property of our construction that the randomness $R$ is information-theoretically determined by the classical part $\mathsf{com}=\algo{Commit}(R)$ of a commitment, and the decommitment can be done by using the rest part of the commitment and $R$.\footnote{For this definition to make sense, we need to require that $\mathsf{com}=\algo{Commit}(R)$ is classical. This can be ensured if the honest receiver measures it as soon as receiving it even if only quantum communication channel is available.}
In particular, this roughly means that one can extract the committed message with an unbounded-time extractor before the sender decommits
This enables us to prove soundness for our certified everlasting zero-knowledge proofs in essentially the same manner as in the classical case.
The details of the construction and security proofs are explained in \cref{sec:bit_comm_construction}.
\paragraph{Certified everlasting zero-knowledge proof for QMA.}
Once we obtain a commitment scheme with certified everlasting hiding, the construction of certified everlasting zero-knowledge proofs is straightforward based on the idea sketched in \cite{TCC:BroIsl20}. Though they only considered a construction for $\compclass{NP}$, we observe that the idea can be naturally extended to a construction for $\compclass{QMA}$ since a ``quantum version'' of $\Sigma$-protocol for $\compclass{QMA}$ called $\Xi$-protocol is constructed by Broadbent and Grilo~\cite{FOCS:BroGri20}.
Below, we sketch the construction for clarity.
Let $A=(A_{\mathsf{yes}},A_{\mathsf{no}})$ be a promise problem in $\compclass{QMA}$.
\cite{FOCS:BroGri20} showed that for any $\mathsf{x}\in A_{\mathsf{yes}}$ and any corresponding witness $\mathsf{w}$,
it is possible to generate (in a quantum polynomial-time)
so-called the local simulatable history state $\rho_\mathsf{hist}$ from $\mathsf{w}$, which
satisfies the following two special properties (for details, see \cref{def:k-SimQMA}):
\begin{enumerate}
\item[(LS1)]
The verification can be done by measuring randomly chosen five qubits of $\rho_\mathsf{hist}$.
\item[(LS2)]
The classical description of any five-qubit reduced density matrix of $\rho_\mathsf{hist}$ can be obtained
in classical polynomial-time.
\end{enumerate}
With these properties, the quantum $\Sigma$-protocol of \cite{FOCS:BroGri20} is constructed as follows:
\begin{enumerate}
\item[{\bf 1. Commitment phase}:]
The prover randomly chooses $x,z\in\{0,1\}^n$,
and sends $(X^xZ^z \rho_\mathsf{hist} Z^zX^x)\otimes \mathsf{com}(x,z)$ to the verifier,
where
$X^xZ^z\coloneqq \prod_{i=1}^n X_i^{x_i}Z_i^{z_i}$,
$n$ is the number of qubits of $\rho_\mathsf{hist}$,
and $\mathsf{com}(x,z)$ is a classical commitment of $(x,z)$.
\item[{\bf 2. Challenge phase}:]
The verifier randomly chooses a subset $S\subset[n]$ of size $|S|=5$, and sends it to the prover.
\item[{\bf 3. Response phase}:]
The prover opens the commitment for $\{x_i,z_i\}_{i\in S}$.
\item[{\bf 4. Verification phase}:]
The verifier applies $\prod_{i\in S}X_i^{x_i}Z_i^{z_i}$ on the state and measures qubits in $S$.
\end{enumerate}
The correctness and the soundness come from the property (LS1),
and the zero-knowledge comes from the property (LS2).
If the classical commitment scheme used in the above construction is the one with statistical binding and
computational hiding, the quantum $\Sigma$-protocol is a computational zero-knowledge proof for $\compclass{QMA}$,
because the unbounded malicious verifier can open the commitment of $\{x_i,z_i\}_{i \in [n]\setminus S}$, and therefore can obtain the entire $\rho_\mathsf{hist}$.
If more than five qubits of $\rho_\mathsf{hist}$ are available to the malicious verifier, the zero-knowledge property no longer holds.
We construct the certified everlasting zero-knowledge proof for $\compclass{QMA}$ based on the quantum $\Sigma$-protocol.
Our idea is to use commitment with certified everlasting hiding and statistical binding
for the commitment of $(x,z)$ in the above construction of the quantum $\Sigma$-protocol.
If the verifier issues a valid deletion certificate for the commitment of $\{x_i,z_i\}_{i\in [n]\setminus S}$,
even unbounded malicious verifier can no longer learn $\{x_i,z_i\}_{i\in [n]\setminus S}$,
and therefore what it can access is only the five qubits of $\rho_\mathsf{hist}$.
This gives a proof for certified everlasting zero-knowledge.
Using classical-extractor-based binding, the proof of statistical soundness can be done almost in the same way as in \cite{FOCS:BroGri20}.
Recall that classical-extractor-based binding enables us to extract the committed message with an unbounded-time extractor before the sender decommits. Therefore, we can extract the committed $(x,z)$ from $\mathsf{com}(x,z)$. Since the extraction is done before the challenge phase, the extracted values do not depend on the challenge $S$. Then, it is easy to reduce the soundness of the scheme to that of the the original $\compclass{QMA}$ promise problem $A$.
The details of the construction is explained in \cref{sec:3_construction}.
\if0
\paragraph{Quantum encryption with certified deletion.}
Our starting point is the quantum encryption scheme with certified deletion by Broadbent and Islam~\cite{TCC:BroIsl20}.
Let us briefly explain their construction. (For details, see \cref{sec:crypt_tool}.)
Their quantum ciphertext consists of a classical part and a quantum part.
The classical part is a ``doubly encrypted" ciphertext, $\algo{Enc}_{\keys{sk}}(m)\oplus H(r)$, of the plaintext $m$, where $\algo{Enc}$ is a
classical symmetric key encryption scheme with the key $\keys{sk}$, $H$ is a two-universal hash function, and $r$ is a random bit string.
The quantum part is a tensor product of computational or Hadamard basis states. The computational basis states encode the information of $r$, and Hadamard basis states are just random ``dummy" states.
These computational and Hadamard basis states are randomly permutated.
Because computational basis states and Hadamard basis states are indistinguishable with each other, the receiver that possesses the quantum ciphertext cannot learn $r$,
and therefore it cannot learn $m$.
The deletion certificate is the Hadamard-basis measurement results on the all qubits of the quantum part.
If the all measurement results for the Hadamard basis states are consistent, it is accepted as a valid certificate.
Because the receiver cannot know which qubits contain the information of $r$ and which qubits are dummy Hadamard-basis states, it has to measure all qubits in the Hadamard basis
(and therefore has to destroy the information of $r$) to issue a valid certificate.
Once the information of $r$ is deleted, the receiver can no longer obtain the plaintext $m$ even if it receives $\keys{sk}$.
\paragraph{Bit commitment with certified everlasting hiding.}
We define the certified everlasting hiding with the following security game between a challenger and an adversary $\mathcal{A}=(\mathcal{A}_1,\mathcal{A}_2)$
that consists of a QPT adversary $\mathcal{A}_1$ and an unbounded adversary $\mathcal{A}_2$. (For details, see \cref{sec:bitcomm_def}.)
The challenger sends the commitment $\mathsf{com}$ to $\mathcal{A}_1$. At some point, $\mathcal{A}_1$ sends the deletion certificate $\mathsf{cert}$ to the challenger.
$\mathcal{A}_1$ sends its internal state to $\mathcal{A}_2$.
If the certificate is valid, the challenger sends its all information except for the committed value itself,
i.e., the decommiting key and the private key for the verification of the deletion certificate,
to $\mathcal{A}_2$.
$\mathcal{A}_2$ guesses the committed value.
The adversary wins if the certificate is valid and $\mathcal{A}_2$ outputs the correct committed value.
Let us explain the idea of our construction of the bit commitment with certified everlasting hiding.
(For details, see \cref{sec:bit_comm_construction}.)
The commitment algorithm runs as follows:
\begin{enumerate}
\item
Encrypt the plaintext $m$ that the sender wants to commit as $\SKE.\algo{Enc}(\mathsf{ske}.\keys{sk},m)\to \mathsf{ske}.\keys{CT}$,
where $\SKE.\algo{Enc}$, $\mathsf{ske}.\keys{sk}$, and $\mathsf{ske}.\keys{CT}$ are
the encryption algorithm,
the private key, and the ciphertext of the encryption with certified deletion $\Sigma_{\mathsf{skcd}}$, respectively.
\item
Generate a random bit string $R$, and commit it by using a classical non-interactive commitment scheme $\Sigma_{\mathsf{com}}$ with the perfect binding and computational hiding as
$\algo{Classical}.\algo{Commit}(R)\to f$,
where $\algo{Classical}.\algo{Commit}$ is the commiting algorithm of $\Sigma_{\mathsf{com}}$.
\item
Compute $h\coloneqq H(R)\oplus \mathsf{ske}.\keys{sk}$, where $H$ is a hash function.
\item
The commitment is $\mathsf{com}\coloneqq(\mathsf{ske}.\keys{CT},f,h)$.
\end{enumerate}
Intuitively, the computational hiding comes from the computational hiding of $\Sigma_{\mathsf{com}}$
and the (IND-CPA) security of $\Sigma_{\mathsf{skcd}}$:
$R$ cannot be learnt from $f$, and therefore $\mathsf{ske}.\keys{sk}$ cannot be learnt from $h$.
The receiver who does not know $\mathsf{ske}.\keys{sk}$ cannot learn the plaintext $m$ from $\mathsf{ske}.\keys{CT}$.
The binding is the statistical one, and it is guaranteed from the perfect binding of $\Sigma_{\mathsf{com}}$.
The certified everlasting hiding roughly comes from the certified deletion security of $\Sigma_{\mathsf{skcd}}$:
If the certificate is valid, $\mathcal{A}_2$ receives the decommitting key for $f$ and the private key for the verification
of the deletion certificate, from which $\mathcal{A}_2$ can learn $\mathsf{ske}.\keys{sk}$.
However, $\mathcal{A}_2$ still cannot know the value of the plaintext $m$ due to the certified deletion security of
$\Sigma_{\mathsf{skcd}}$.
\fi
\if0
\paragraph{Definition of quantum bit commitment with certified deletion}
A quantum bit commitment with certified deletion consists of the following algorithms.
\begin{description}
\item[$\algo{Commit} (1^{\lambda},b)\rightarrow (\mathsf{com},\mathsf{d},\keys{ck})$:]
This is a commitment algorithm that generates
\item[$\algo{Verify}(\mathsf{com},\mathsf{d},b)\rightarrow \top~or~\bot $:]
The verification algorithm takes as input $\mathsf{com}$, $\mathsf{d}$ and $b$
and outputs $\top$ or $\bot$.
\item[$\algo{Del}(\mathsf{com})\rightarrow \mathsf{cert}$:]The deletion algorithm takes as input $\mathsf{com}$ and outputs a certificate $\mathsf{cert}\in \mathcal{E}$.
\item[$\algo{Cert}(\mathsf{cert},\keys{ck})\rightarrow \top~or~\bot$:]The certification algorithm takes as input $\mathsf{cert}$ and $\keys{ck}$ and outputs $\top$ or $\bot$.
\end{description}
\fi
\subsection{Related Works}
\paragraph{Zero-knowledge for QMA.}
Zero-knowledge for $\compclass{QMA}$ was first constructed by
Broadbent, Ji, Song, and Watrous \cite{FOCS:BJSW16}.
Broadbent and Grilo \cite{FOCS:BroGri20} gave an elegant and simpler construction what they call the
$\Xi$-protocol (which is considered as a quantum version of the standard $\Sigma$-protocol) by using the local simulatability~\cite{FOCS:GriSloYue19}. Our construction is based on the $\Xi$-protocol.
Bitansky and Shmueli \cite{STOC:BitShm20} gave the first constant round zero-knowledge argument for $\compclass{QMA}$ with negligible soundness error.
Brakerski and Yuen \cite{BraYue} gave a construction of $3$-round \emph{delayed-input} zero-knowledge proof for $\compclass{QMA}$ where the prover needs to know the statement and witness only for generating its last message.
Chardouvelis and Malavolta~\cite{EPRINT:ChaMal21} constructed 4-round statistical zero-knowledge arguments for $\compclass{QMA}$ and 2-round zero-knowledge for $\compclass{QMA}$ in the timing model.
Regarding non-interactive zero-knowledge proofs or arguments (NIZK),
Kobayashi \cite{Kobayashi03} first studied (statistically sound and zero-knowledge) NIZKs in a model where the prover and verifier share Bell pairs, and gave a complete problem in this setting.
It is unlikely that the complete problem contains (even a subclass of) $\compclass{NP}$ \cite{MW18}, and thus even a NIZK for all $\compclass{NP}$ languages is unlikely to exist in this model.
Chailloux et al. \cite{TCC:CCKV08} showed that there exists a (statistically sound and zero-knowledge) NIZK for all languages in $\compclass{QSZK}$ in the help model where a trusted party generates a pure state \emph{depending on the statement to be proven} and gives copies of the state to both prover and verifier.
Recently, there are many constructions of NIZK proofs or arguments for $\compclass{QMA}$ in various kind of setup models and assumptions \cite{TCC:ACGH20,C:ColVidZha20,FOCS:BroGri20,C:Shmueli21,C:BCKM21a,EPRINT:MorYam21,BartusekMalavolta}.
\if0
\paragraph{Revocable quantum time-release encryption.}
Unruh~\cite{JACM:Unruh15} introduced
the revocable quantum timed-release encryption, which is
conceptually similar to quantum encryption with certified deletion.
In this primitive, a receiver can decrypt a quantum
ciphertext only after spending a certain amount of time $T$.
The receiver can also choose to return the ciphertext before
the time $T$ is over, in which case it is ensured that the message can no longer be recovered.
An essential difference from quantum encryption with certified deletion,
which was observed in \cite{TCC:BroIsl20}, is that the revocable quantum timed-release encryption does not have a mechanism to generate a classical certificate of deletion.
\takashi{Do we need to include this paragraph? This doesn't seem to be very related to this paper's topic.}
\fi
\paragraph{Quantum commitment.}
It is well-known that statistically binding and hiding commitments are impossible even with quantum communication \cite{LC97,Mayers97}.
On the other hand, there are a large body of literature on constructing quantum commitments assuming some computational assumptions, e.g., see the references in the introduction of \cite{Yan20a}. Among them, several works showed the possibility of using quantum commitments in constructions of zero-knowledge proofs and arguments \cite{ISAAC:YWLQ15,FUWYZ20,Yan20a,BB21}. However, they only consider replacing classical commitments with quantum commitments in classical constructions while keeping the same functionality and security level as the classical construction. In particular, none of them considers protocols for $\compclass{QMA}$ or properties that are classically impossible to achieve like our notion of the certified everlasting zero-knowledge.
\if0
\paragraph{Everlasting security.}
Unruh \cite{C:Unruh13} studied everlasting secure multi-party computation. The everlasting security defined in \cite{C:Unruh13} means that the protocol remains secure as long as each party runs in polynomial-time during the execution of the protocol. If we define everlasting zero-knowledge based on this approach, this would be stronger than our notion of certified everlasting zero-knowledge because the security should remain even if a cheating verifier fails to let the prover accept. However, we remark that everlasting zero-knowledge proofs for $\compclass{QMA}$ or even for $\compclass{NP}$ are unlikely to exist: If a protocol satisfies everlasting zero-knowledge, this in particular means that the protocol satisfies honest-verifier quantum statistical zero-knowledge \cite{FOCS:Watrous02} since an honest verifier runs in polynomial-time. Watrous \cite{SIAM:Wat09} showed $\mathbf{QSZK_{HV}}=\mathbf{QSZK}$ where $\mathbf{QSZK_{HV}}$ and $\mathbf{QSZK}$ are the classes of languages that have honest-verifier quantum statistical zero-knowledge proofs and those that have quantum statistical zero-knowledge (against malicious verifiers). On the other hand, it is believed that $\mathbf{NP}\not\subseteq\mathbf{QSZK}$ \cite{MW18}. Therefore, assuming $\mathbf{NP}\not\subseteq\mathbf{QSZK}$, there does not exist everlasting zero-knowledge proofs for $\mathbf{NP}$ (in the sense of \cite{C:Unruh13}). We remark that the above argument does not go through for certified everlasting zero-knowledge. The key point is that the honest-verifier quantum statistical zero-knowledge defined in \cite{FOCS:Watrous02} requires a simulator to simulate honest verifier's internal state \emph{at any point} of the protocol execution. On the other hand, certified everlasting zero-knowledge does not require this. It only requires a simulator to simulate the (possibly malicious) verifier's internal state \emph{after completing the protocol} conditioned on that the prover accepts. This is how we bypass the above impossibility.
Finally, we mention that everlasting zero-knowledge \emph{arguments}, which only satisfy computational soundness, can exist. Indeed, any statistical zero-knowledge argument is everlasting zero-knowledge argument. One may think that the computational soundness is fine since that ensures everlasting soundness in the sense of \cite{C:Unruh13}. For practical purposes, this may be true. On the other hand, we believe that it is theoretically interesting to pursue (a kind of) everlasting zero-knowledge without compromising the soundness as is done in this paper.
\fi
\section{Preliminaries}\label{sec:preliminaries}
\subsection{Notations}\label{sec:notation}
Here we introduce basic notations we will use.
In this paper, $x\leftarrow X$ denotes selecting an element from a finite set $X$ uniformly at random,
and $y\leftarrow A(x)$ denotes assigning to $y$ the output of a probabilistic or deterministic algorithm $A$ on an input $x$.
When we explicitly show that $A$ uses randomness $r$, we write $y\leftarrow A(x;r)$.
When $D$ is a distribution, $x\leftarrow D$ denotes sampling an element from $D$.
Let $[n]$ be the set $\{1,\dots,n\}$. Let $\lambda$ be a security parameter, and $y\coloneqq z$ denotes that $y$ is set, defined, or substituted by $z$.
For a bit string $s\in\{0,1\}^n$, $s_i$ denotes the $i$-th bit of $s$. QPT stands for quantum polynomial time.
PPT stands for (classical) probabilistic polynomial time.
For a subset $S\subseteq W$ of a set $W$, $\overline{S}$ is the complement of $S$, i.e.,
$\overline{S}\coloneqq W\setminus S$.
A function $f: \mathbb{N} \rightarrow \mathbb{R}$ is a negligible function if for any constant $c$, there exists $\lambda_0 \in \mathbb{N}$ such that for any $\lambda>\lambda_0$, $f(\lambda) < \lambda^{-c}$. We write $f(\lambda) \leq {\mathsf{negl}}(\lambda)$ to denote $f(\lambda)$ being a negligible function.
\subsection{Quantum Computation}\label{sec:quantum_computation}
We assume the familiarity with basics of quantum computation, and use standard notations.
Let us denote $\mathcal{Q}$ be the state space of a single qubit.
$I$ is the two-dimensional identity operator. For simplicity, we often write
$I^{\otimes n}$ as $I$ for any $n$ when the dimension of the identity operator is clear from the context.
For any single-qubit operator $O$, $O_i$ means an operator that applies $O$ on the $i$-th qubit and applies $I$ on all other qubits.
$X$ and $Z$ are the Pauli $X$ and $Z$ operators, respectively.
For any $n$-bit strings $x\coloneqq(x_1,x_2,\cdots,x_n)\in\{0,1\}^n$ and $z\coloneqq (z_1,z_2,\cdots,z_n)\in\{0,1\}^n$,
$X^x\coloneqq\prod_{i\in[n]}X_i^{x_i}$ and
$Z^z\coloneqq\prod_{i\in[n]}Z_i^{z_i}$.
For any subset $S$, $\mathrm{Tr}_S$ means the trace over all qubits in $S$.
For any quantum state $\rho$ and a bit string $s\in\{0,1\}^n$, $\rho\otimes s$ means $\rho\otimes|s\rangle\langle s|$.
The trace distance between two states $\rho$ and $\sigma$ is given by
$\frac{1}{2}\norm{\rho-\sigma}_{\tr}$, where $\norm{A}_{\tr}\coloneqq \mathrm{Tr} \sqrt{{\it A}^{\dagger}{\it A}}$ is the trace norm.
If $\frac{1}{2}\norm{\rho-\sigma}_{\tr}\leq \epsilon$, we say that $\rho$ and $\sigma$ are $\epsilon$-close.
If $\epsilon={\mathsf{negl}}(\lambda)$, then we say that $\rho$ and $\sigma$ are statistically indistinguishable.
Let $C_0$ and $C_1$ be quantum channels from $p$ qubits to $q$ qubits, where $p$ and $q$ are polynomials.
We say that they are computationally indistinguishable, and denote it by $C_0\approx_c C_1$
if there exists a negligible function ${\mathsf{negl}}$ such that
$|\Pr[D((C_0\otimes I)(\sigma))=1]-\Pr[D((C_1\otimes I)(\sigma))=1] |\leq {\mathsf{negl}}(\lambda)$
for any polynomial $k$, any $(p+k)$-qubit state $\sigma$, and any polynomial-size quantum circuit $D$ acting on
$q+k$ qubits.
We say that $C_0$ and $C_1$ are statistically indistinguishable, and denote it by $C_0\approx_s C_1$,
if $D$ is an unbounded algorithm.
\if0
\mor{kokokara}
For some $S\subseteq\{0,1\}^*$, let $\{\Psi_x\}_{x\in S}$ and $\{\Phi_x\}_{x\in S}$ be two families of quantum channels from $q(|x|)$ qubits to $r(|x|)$ qubits, where $q$ and $r$ are polynomials.
We say that these two families are computationally indistinguishable, and denote it by $\{\Psi_x\}_{x\in S}\approx_c \{\Phi_x\}_{x\in S}$
if there exists a negligible function ${\mathsf{negl}}$ such that for any polynomial $k$, any polynomial-size quantum circuit $Q$ acting on $r(|x|)+k(|x|)$ qubits, any $q(|x|)+k(|x|)$ qubit state $\sigma$\mor{even states that cannot be generated in QPT?}\taiga{for any de yoito omounodesuga. kakusin ha naidesu ga.} and for any $x\in S$ and $n\in\mathbb{N}$\taiga{Kore iranai?} the following holds
\begin{align}
|\Pr[Q(\Psi_x\otimes I)(\sigma)=1]-\Pr[Q(\Phi_x\otimes I)(\sigma)=1] |\leq {\mathsf{negl}}(n)\mbox{\taiga{${\mathsf{negl}}(|x|)$?}}.
\end{align}
\takashi{Strictly speaking, the order of quantifier should be "there exists a negligible function ${\mathsf{negl}}$ such that for all $x\in S$...".
It is problematic if we allow the negligible function to depend on $x$, e.g., see Section 2 of https://eprint.iacr.org/2016/046.pdf.}
\taiga{I have modified it. Is that ok?}
We say that these two families are statistically indistinguishable, and denote it by $\{\Psi_x\}_{x\in S}\approx_s \{\Phi_x\}_{x\in S}$
if for all circuit $Q$ acting on $r(|x|)+k(|x|)$ qubits, there exists a negligible function ${\mathsf{negl}}$ such that for all $x\in S$ and polynomial $k$, all state $\sigma$ on $q(|x|)+k(|x|)$ and $n\in\mathbb{N}$,
the following holds
\begin{align}
|\Pr[Q(\Psi_x\otimes I)(\sigma)=1]-\Pr[Q(\Phi_x\otimes I)(\sigma)=1] |\leq {\mathsf{negl}}(n).
\end{align}
\mor{kokomade}
\fi
\begin{lemma}[Quantum Rewinding Lemma~\cite{SIAM:Wat09}]\label{lemma:rewinding}
Let $Q$ be a quantum circuit that acts on an $n$-qubit state $|\psi\rangle$ and an $m$-qubit auxiliary state $|0^m\rangle$.
Let
$p(\psi)\coloneqq||(\langle 0|\otimes I)Q(|\psi\rangle \otimes |0^m\rangle)||^2$ and
$
|\phi(\psi)\rangle\coloneqq\frac{1}{\sqrt{p(\psi)}} (\langle 0|\otimes I) Q(|\psi\rangle\otimes |0^m\rangle).
$
Let $p_0,q\in(0,1)$ and $\epsilon\in (0,\frac{1}{2})$ such that $|p(\psi)-q|<\epsilon$, $p_0(1-p_0)<q(1-q)$, and $p_0<p(\psi)$.
Then there is a quantum circuit $R$ of size at most
$
O\left(\frac{\log(\frac{1}{\epsilon}){\rm size}(Q)}{p_0(1-p_0)}\right)
$
such that on input $|\psi\rangle$, $R$ computes a quantum state $\rho(\psi)$ that satisfies
$
\langle\phi(\psi)|\rho(\psi)|\phi(\psi)\rangle\geq 1-16\epsilon\frac{\log^2(\frac{1}{\epsilon})}{p_0^2(1-p_0)^2}.
$
\end{lemma}
\begin{lemma}[One-Way to Hiding Lemma \cite{C:AmbHamUnr19}]\label{lemma:one-way_to_hiding}
Let $S\subseteq \mathcal{X}$ be a random subset of $\mathcal{X}$. Let $G,H:\mathcal{X}\rightarrow\mathcal{Y}$ be random functions satisfying $\forall x\notin S$ $[G(x)=H(x)]$. Let $z$ be a random classical bit string.
($S,G,H,z$ may have an arbitrary joint distribution.)
Let $\mathcal{A}$ be an oracle-aided quantum algorithm that makes at most $q$ quantum queries.
Let $\mathcal{B}$ be an algorithm that on input $z$ chooses $i\leftarrow[q]$, runs $\mathcal{A}^{H}(z)$, measures $\mathcal{A}$'s $i$-th query, and outputs the measurement outcome.
Then we have
$
\abs{\Pr[\mathcal{A}^G(z)=1]-\Pr[\mathcal{A}^H(z)=1]}\leq2q\sqrt{\Pr[\mathcal{B}^H(z)\in S]}.
$
\end{lemma}
\subsection{QMA and $k$-$\compclass{SimQMA}$}
\begin{definition}[\compclass{QMA}]\label{def:QMA}
We say that a promise problem $A=(A_{\mathsf{yes}},A_{\mathsf{no}})$ is in $\compclass{QMA}$ if there exist a polynomial $p$, a QPT algorithm $V$,
and $0\le\beta<\alpha\le1$ with $\alpha-\beta\geq\frac{1}{{\mathrm{poly}}(|\mathsf{x}|)}$
such that
\begin{description}
\item[Completeness:]
For any $\mathsf{x}\in A_{\mathsf{yes}}$, there exists a quantum state $\mathsf{w}$ of $p(|\mathsf{x}|)$-qubit (called a witness) such that
\begin{align}
\Pr[V(\mathsf{x},\mathsf{w})=\top]\geq \alpha.
\end{align}
\item[Soundness:]
For any $\mathsf{x}\in A_{\mathsf{no}}$ and any quantum state $\mathsf{w}$ of $p(|\mathsf{x}|)$-qubit,
\begin{align}
\Pr[V(\mathsf{x},\mathsf{w})=\top]\leq \beta.
\end{align}
\end{description}
For any $\mathsf{x}\in A_{\mathsf{yes}}$, $R_A(\mathsf{x})$ is the (possibly infinite) set of all quantum states
$\mathsf{w}$ such that $\Pr[V(\mathsf{x},\mathsf{w})=\top]\geq \frac{2}{3}$.
\end{definition}
A complexity class of $k$-$\compclass{SimQMA}$ is introduced, and proven to be equal to $\compclass{QMA}$ in \cite{FOCS:BroGri20}.
\begin{definition}[$k$-\compclass{SimQMA}~\cite{FOCS:BroGri20}]\label{def:k-SimQMA}
A promise problem $A=(A_{\mathsf{yes}},A_{\mathsf{no}})$ is in $k$-$\compclass{SimQMA}$ with soundness $\beta(|\mathsf{x}|)\leq 1-\frac{1}{{\mathrm{poly}}(|\mathsf{x}|)}$,
if there exist polynomials $m$ and $n$ such that given $\mathsf{x}\in A_{\mathsf{yes}}$,
there is an efficient deterministic algorithm that computes $m(|\mathsf{x}|)$ $k$-qubit POVMs $\{\Pi_1,I-\Pi_1\},\dots,\{\Pi_{m(|\mathsf{x}|)},I-\Pi_{m(|\mathsf{x}|)}\}$ such that:
\begin{description}
\item [Simulatable completeness:]
If $\mathsf{x}\in A_{\mathsf{yes}}$, there exists an $n(|\mathsf{x}|)$-qubit state $\rho_{\mathsf{hist}}$, which we call a simulatable witness, such that for all $c\in[m]$,
$\mathrm{Tr}(\Pi_c\rho_{\mathsf{hist}})\geq 1-{\mathsf{negl}}(|\mathsf{x}|)$,
and
there exists a set of $k$-qubit density matrices $\{\rho_\mathsf{sim}^{\mathsf{x},S}\}_{S\subseteq[n(|\mathsf{x}|)],|S|=k}$
that can be computed in polynomial time from $\mathsf{x}$ and $\rho_{\mathsf{hist}}$
such that
$
||\mathrm{Tr}_{\overline{S}}(\rho_{\mathsf{hist}})-\rho_\mathsf{sim}^{\mathsf{x},S}||_{\rm tr}\leq {\mathsf{negl}}(|\mathsf{x}|).
$
\item [Soundness:]
If $\mathsf{x}\in A_{\mathsf{no}}$, for any $n(|\mathsf{x}|)$-qubit state $\rho$,
$
\frac{1}{m}\sum_{c\in[m]} \mathrm{Tr}(\Pi_{c}\rho)\leq \beta(|\mathsf{x}|).
$
\end{description}
\end{definition}
\subsection{Cryptographic Tools}\label{sec:crypt_tool}
In this section, we review cryptographic tools used in this paper.
\paragraph{Non-interactive commitment.}
\begin{definition}[Non-Interactive Commitment (Syntax)]\label{def:Non_interactive_commitments}
Let $\lambda$ be the security parameter and let $p$, $q$ and $r$ be some polynomials.
A (classical) non-interactive commitment scheme consists of a single PPT algorithm $\algo{Commit}$ with plaintext space $\mathcal{M}\coloneqq\{0,1\}^{p(\lambda)}$, randomness space $\{0,1\}^{q(\lambda)}$ and commitment space $\mathcal{C}\coloneqq \{0,1\}^{r(\lambda)}$ satisfying two properties:
\begin{description}
\item[Perfect binding:]For every $(r_0,r_1)\in\{0,1\}^{q(\lambda)}\times\{0,1\}^{q(\lambda)}$ and
$(m,m')\in \mathcal{M}^2$ such that $m\neq m'$,
we have that $\algo{Commit}(m;r_0)\neq \algo{Commit}(m';r_1)$, where $(\algo{Commit}(m;r_0), \algo{Commit}(m';r_1))\in \mathcal{C}^2$.
\item[Unpredictability:] Let $\Sigma\coloneqq\algo{Commit}$. For any QPT adversary $\mathcal{A}$, we define the following security experiment $\mathsf{Exp}_{\Sigma,\mathcal{A}}^{\mathsf{unpre}}(\lambda)$.
\begin{enumerate}
\item The challenger chooses $R\leftarrow \mathcal{M}$ and $R'\leftarrow\{0,1\}^{q(\lambda)}$, computes $\mathsf{com}\leftarrow\algo{Commit}(R;R')$, and sends $\mathsf{com}$ to $\mathcal{A}$.
\item $\mathcal{A}$ outputs $R^*$. The output of the experiment is $1$ if $R^*=R$. Otherwise, the output of the experiment is $0$.
\end{enumerate}
We say that the commitment is unpredictable if for any QPT adversary $\mathcal{A}$, it holds that
\begin{align}
\adva{\Sigma,\mathcal{A}}{unpre}(\lambda)\coloneqq \abs{\Pr[ \mathsf{Exp}_{\Sigma,\mathcal{A}}^{\mathsf{unpre}}(\lambda)=1] }\leq {\mathsf{negl}}(\lambda).
\end{align}
\end{description}
\end{definition}
\begin{remark}
The unpredictability is a weaker version of computational hiding. We define unpredictability instead of computational hiding since this suffices for our purpose.
\end{remark}
A non-interactive commitment scheme that satisfies the above definition exists assuming the existence of injective one-way functions or perfectly correct public key encryption \cite{EPRINT:LomSch19}. Alternatively, we can also instantiate it based on random oracles.
\paragraph{Quantum encryption with certified deletion.}
Broadbent and Islam~\cite{TCC:BroIsl20} introduced the notion of quantum encryption with certified deletion.
\begin{definition}[One-Time SKE with Certified Deletion (Syntax)]\label{def:sk_cert_del}
Let $\lambda$ be the security parameter and let $p$, $q$ and $r$ be some polynomials.
A one-time secret key encryption scheme with certified deletion consists of a tuple of algorithms $(\algo{KeyGen},\algo{Enc},\algo{Dec},\algo{Del},\algo{Verify})$ with
plaintext space $\mathcal{M}:=\{0,1\}^n$, ciphertext space $\mathcal{C}:= \mathcal{Q}^{\otimes p(\lambda)}$, key space $\mathcal{K}:=\{0,1\}^{q(\lambda)}$ and deletion certificate space $\mathcal{D}:= \{0,1\}^{r(\lambda)}$.
\begin{description}
\item[$\algo{KeyGen} (1^\lambda) \rightarrow \keys{sk}$:] The key generation algorithm takes as input the security parameter $1^\lambda$, and outputs a secret key $\keys{sk} \in \mathcal{K}$.
\item[$\algo{Enc}(\keys{sk},m) \rightarrow \keys{CT}$:] The encryption algorithm takes as input $\keys{sk}$ and a plaintext $m\in\mathcal{M}$, and outputs a ciphertext $\keys{CT}\in \mathcal{C}$.
\item[$\algo{Dec}(\keys{sk},\keys{CT}) \rightarrow m^\prime~or~\bot$:] The decryption algorithm takes as input $\keys{sk}$ and $\keys{CT}$, and outputs a plaintext $m^\prime \in \mathcal{M}$ or $\bot$.
\item[$\algo{Del}(\keys{CT}) \rightarrow \mathsf{cert}$:] The deletion algorithm takes as input $\keys{CT}$, and outputs a certification $\mathsf{cert}\in\mathcal{D}$.
\item[$\algo{Verify}(\keys{sk},\mathsf{cert})\rightarrow \top~or~\bot$:] The verification algorithm takes $\keys{sk}$ and $\mathsf{cert}$, and outputs $\top$ or $\bot$.
\end{description}
\end{definition}
\begin{definition}[Correctness for One-Time SKE with Certified Deletion]\label{def:sk_cd_correctness}
There are two types of correctness. One is decryption correctness and the other is verification correctness.
\begin{description}
\item[Decryption correctness:] There exists a negligible function ${\mathsf{negl}}$ such that for any $\lambda\in \mathbb{N}$ and $m\in\mathcal{M}$,
\begin{align}
\Pr\left[
\algo{Dec}(\keys{sk},\keys{CT})= m
\ \middle |
\begin{array}{ll}
\keys{sk}\leftarrow \algo{KeyGen}(1^\lambda)\\
\keys{CT} \leftarrow \algo{Enc}(\keys{sk},m)
\end{array}
\right]
\geq1-{\mathsf{negl}}(\lambda).
\end{align}
\item[Verification correctness:] There exists a negligible function ${\mathsf{negl}}$ such that for any $\lambda\in \mathbb{N}$ and $m\in\mathcal{M}$,
\begin{align}
\Pr\left[
\algo{Verify}(\keys{sk},\mathsf{cert})=\top
\ \middle |
\begin{array}{ll}
\keys{sk}\leftarrow \algo{KeyGen}(1^\lambda)\\
\keys{CT} \leftarrow \algo{Enc}(\keys{sk},m)\\
\mathsf{cert} \leftarrow \algo{Del}(\keys{CT})
\end{array}
\right]
\geq
1-{\mathsf{negl}}(\lambda).
\end{align}
\end{description}
\end{definition}
\begin{definition}[Certified Deletion Security for One-Time SKE]\label{def:sk_certified_del}
Let $\Sigma=(\algo{KeyGen}, \algo{Enc}, \algo{Dec}, \algo{Del}, \algo{Verify})$ be a secret key encryption with certified deletion.
We consider the following security experiment $\expb{\Sigma,\mathcal{A}}{otsk}{cert}{del}(\lambda,b)$.
\begin{enumerate}
\item The challenger computes $\keys{sk} \leftarrow \algo{KeyGen}(1^\lambda)$.
\item $\mathcal{A}$ sends $(m_0,m_1)\in\mathcal{M}^2$ to the challenger.
\item The challenger computes $\keys{CT}_b \leftarrow \algo{Enc}(\keys{sk},m_b)$ and sends $\keys{CT}_b$ to $\mathcal{A}$.
\item $\mathcal{A}$ sends $\mathsf{cert}$ to the challenger.
\item The challenger computes $\algo{Verify}(\keys{sk},\mathsf{cert})$. If the output is $\bot$, the challenger sends $\bot$ to $\mathcal{A}$.
If the output is $\top$, the challenger sends $\keys{sk}$ to $\mathcal{A}$.
\item $\mathcal{A}$ outputs $b'\in \{0,1\}$.
\end{enumerate}
We say that the $\Sigma$ is OT-CD secure if for any unbounded $\mathcal{A}$, it holds that
\begin{align}
\advc{\Sigma,\mathcal{A}}{otsk}{cert}{del}(\lambda)
\coloneqq \abs{\Pr[ \expb{\Sigma,\mathcal{A}}{otsk}{cert}{del}(\lambda, 0)=1] - \Pr[ \expb{\Sigma,\mathcal{A}}{otsk}{cert}{del}(\lambda, 1)=1] }\leq {\mathsf{negl}}(\lambda).
\end{align}
\end{definition}
Broadbent and Islam~\cite{TCC:BroIsl20} showed that one-time SKE scheme with certified deletion that satisfies the above correctness and security exists unconditionally.
\section{Certified Everlasting Zero-Knowledge Proof for QMA}\label{sec:ZK}
In this section, we define and construct the certified everlasting zero-knowledge proof for $\compclass{QMA}$.
In \cref{sec:def_everlasting_ZK}, we define the certified everlasting zero-knowledge proof for $\compclass{QMA}$.
We then construct a three round protocol with completeness-soundness gap $\frac{1}{{\mathrm{poly}}(\lambda)}$ in \cref{sec:3_construction},
and finally amplify the gap to $1-{\mathsf{negl}}(\lambda)$ with the sequential repetition in \cref{sec:Sequential_repetition}.
\subsection{Definition}\label{sec:def_everlasting_ZK}
We first define a quantum interactive protocol.
Usually, in zero-knowledge proofs or arguments, we do not consider prover's output. However, in this paper,
we also consider prover's output,
because we are interested in the certified everlasting zero-knowledge.
Furthermore, in this paper, we consider only an interactive proof, which means that a malicious prover is unbounded.
\begin{definition}[Quantum Interactive Protocol]\label{def:quantum_interactive_protocol}
A quantum interactive protocol is modeled as an interaction between QPT machines $\mathcal{P}$ referred as a prover and $\mathcal{V}$ referred as a verifier.
We denote by $\langle \mathcal{P}(x_P),\mathcal{V}(x_V)\rangle (x)$ an execution of the protocol where $x$ is a common input,
$x_P$ is $\mathcal{P}$'s private input, and $x_V$ is $\mathcal{V}$'s private input.
We denote by $\mathrm{OUT}_{\mathcal{V}}\langle \mathcal{P}(x_P),\mathcal{V}(x_V)\rangle(x)$ the final output of $\mathcal{V}$ in the execution.
An honest verifier's output is $\top$ indicating acceptance or $\bot$ indicating rejection, and a malicious verifier's output is an arbitrary quantum state.
We denote by $\mathrm{OUT}_{\mathcal{P}}\langle \mathcal{P}(x_P),\mathcal{V}(x_V)\rangle(x)$ the final output of $\mathcal{P}$ in the execution.
An honest prover's output is $\top$ indicating acceptance or $\bot$ indicating rejection.
We also define $\mathrm{OUT'}_{\mathcal{P},\mathcal{V}}\langle \mathcal{P}(x_{P}),\mathcal{V}(x_{V})\rangle(x)$ by
\begin{align}
\mathrm{OUT'}_{\mathcal{P},\mathcal{V}}\langle \mathcal{P}(x_P),\mathcal{V}(x_V)\rangle(x)
\coloneqq
\begin{cases}
\left(\top,\mathrm{OUT}_{\mathcal{V}}\langle \mathcal{P}(x_P),\mathcal{V}(x_V)\rangle(x)\right)~~~&(\mathrm{OUT}_{\mathcal{P}}\langle \mathcal{P}(x_P),\mathcal{V}(x_V)\rangle(x)=\top)\\
(\bot,\bot)~~~&(\mathrm{OUT}_{\mathcal{P}}\langle \mathcal{P}(x_P),\mathcal{V}(x_V)\rangle(x)\neq\top).
\end{cases}
\end{align}
\end{definition}
We next define a computational zero-knowledge proof for $\compclass{QMA}$, which is the standard definition.
\begin{definition}[Computational Zero-Knowledge Proof for QMA]\label{def:ZK}
A $c$-complete $s$-sound computational zero-knowledge proof for a $\compclass{QMA}$
promise problem $A=(A_{\mathsf{yes}},A_{\mathsf{no}})$ is a quantum interactive protocol between a QPT prover $\mathcal{P}$ and a QPT verifier $\mathcal{V}$ that satisfies the followings:
\begin{description}
\item[$c$-completeness:]
For any $\mathsf{x}\in A_{\mathsf{yes}}$ and any $\mathsf{w}\in R_{A}(\mathsf{x})$,
\begin{align}
\Pr[\mathsf{Out}_{\mathcal{V}}\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}), \mathcal{V}\rangle (\mathsf{x})=\top ]\ge c
\end{align}
for some polynomial $k$.
\item[$s$-soundness:]
For any $\mathsf{x}\in A_{\mathsf{no}}$ and
any unbounded-time prover $\mathcal{P}^*$,
\begin{align}
\Pr[\mathsf{Out}_{\mathcal{V}}\langle \mathcal{P}^*, \mathcal{V}\rangle (\mathsf{x})=\top]\leq s.
\end{align}
\item[Computational zero-knowledge:]
There exists a QPT algorithm $\mathcal{S}$ such that
\begin{align}
\mathsf{OUT}_{\mathcal{V}^*}\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}^*(\cdot) \rangle(\mathsf{x}) \approx_c \mathcal{S}(\mathsf{x},\mathcal{V}^*,\cdot~)
\end{align}
for any QPT malicious verifier $\mathcal{V}^*$,
any $\mathsf{x}\in A_{\mathsf{yes}}\cap \{0,1\}^\lambda$, any $\mathsf{w}\in R_{A}(\mathsf{x})$, and some polynomial $k$.
Note that $\mathsf{OUT}_{\mathcal{V}^*}\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}^*(\cdot) \rangle(\mathsf{x})$ and $\mathcal{S}(\mathsf{x},\mathcal{V}^*,\cdot~)$ are quantum channels
that map any quantum state $\xi$ to quantum states $\mathsf{OUT}_{\mathcal{V}^*}\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}^*(\xi) \rangle(\mathsf{x})$ and $\mathcal{S}(\mathsf{x},\mathcal{V}^*,\xi)$, respectively.
\end{description}
We just call it a computational zero-knowledge proof if it satisfies $(1-{\mathsf{negl}}(|\mathsf{x}|))$-completeness,
${\mathsf{negl}}(|\mathsf{x}|)$-soundness, and
computational zero-knowledge.
\end{definition}
\if0
\begin{lemma}[Gap Amplification]
If there exists a computational zero-knowledge proof for a $\compclass{QMA}$ promise problem $A$ with
$c$-completeness and $s$-soundness such that $c-s\ge\frac{1}{{\mathrm{poly}}(|\lambda|)}$,
then
there exists a computational zero-knowledge proof for $A$ (with $(1-{\mathsf{negl}}(|\lambda|))$-completeness
and ${\mathsf{negl}}(|\lambda|)$-soundness).
\end{lemma}
\begin{proof}
It is clear that the sequential repetition amplifies the gap between completeness and soundness.
It is also shown by the standard hybrid argument that computational zero-knowledge is preserved under the sequential
repetition~\cite{GoldOre94,SIAM:Wat09}.
\end{proof}
\takashi{The above lemma seems redundant since a stronger statement is proven in Theorem 4.14.}
\fi
We finally define a certified everlasting zero-knowledge proof for $\compclass{QMA}$, which is the main target of
this paper.
\begin{definition}[Certified Everlasting Zero-Knowledge Proof for QMA]\label{def:Everlasting_zeroknowledge}
A certified everlasting zero-knowledge proof for a $\compclass{QMA}$ promise problem $A=(A_{\mathsf{yes}},A_{\mathsf{no}})$ is
a computational zero-knowledge proof for $A$ (\cref{def:ZK}) that additionally satisfies the followings:
\begin{description}
\item[Prover's completeness:]
$\Pr[\mathrm{OUT}_{\mathcal{P}}\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}\rangle(\mathsf{x})=\top]\ge1-{\mathsf{negl}}(\lambda)$
for any $\mathsf{x}\in A_{\mathsf{yes}}\cap\{0,1\}^\lambda$ and any $\mathsf{w}\in R_A(\mathsf{x})$.
\item[Certified everlasting zero-knowledge:]
There exists a QPT algorithm $\mathcal{S}$ such that
\begin{align}
\mathrm{OUT'}_{\mathcal{P},\mathcal{V}^*}\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}^*(\cdot)\rangle (\mathsf{x}) \approx_{s}
\mathcal{S}(\mathsf{x},\mathcal{V}^*,\cdot~)
\end{align}
for any QPT malicious verifier $\mathcal{V}^*$,
any $\mathsf{x}\in A_{\mathsf{yes}}\cap \{0,1\}^\lambda$, any $\mathsf{w}\in R_{A}(\mathsf{x})$, and some polynomial $k$.
Note that $\mathsf{OUT'}_{\mathcal{P},\mathcal{V}^*}\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}^*(\cdot) \rangle(\mathsf{x})$ and $\mathcal{S}(\mathsf{x},\mathcal{V}^*,\cdot~)$ are quantum channels
that map any quantum state $\xi$ to quantum states
$\mathsf{OUT'}_{\mathcal{P},\mathcal{V}^*}\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}^*(\xi) \rangle(\mathsf{x})$ and $\mathcal{S}(\mathsf{x},\mathcal{V}^*,\xi)$, respectively.
\end{description}
\end{definition}
\begin{remark}
We remark that certified everlasting zero-knowledge does not imply computational zero-knowledge since it does not require anything if the prover does not output $\top$.
\end{remark}
\subsection{Construction of Three Round Protocol}\label{sec:3_construction}
In this section, we construct a three round protocol with completeness-soundness gap $\frac{1}{{\mathrm{poly}}(\lambda)}$.
In the next section, we will amplify its completeness-soundness gap by the sequential repetition.
In the following, $n$, $m$, $\Pi_{c}$, $\rho_\mathsf{hist}$, and $\rho_{\algo{Sim}}^{\mathsf{x},S}$ are given in \cref{def:k-SimQMA}.
Let $S_c\subseteq[n]$ be the set of qubits on which $\Pi_c$ acts non-trivially.
The three round protocol $\Sigma_{\Xi\mathsf{cd}}$ is constructed from commitment with certified everlasting hiding and classical-extractor-based binding, $\Sigma_{\mathsf{ccd}}=(\algo{Commit},\algo{Verify},\algo{Del},\algo{Cert})$.
\begin{description}
\item[The first action by the prover (commitment phase):] $ $
\begin{itemize}
\item Generate $x,z\leftarrow\{0,1\}^n$.
\item Compute
\begin{align}
& (\mathsf{com}_i(x_i),\mathsf{d}_i(x_i),\keys{ck}_i(x_i))\leftarrow\algo{Commit}(1^\lambda,x_i)\\
& (\mathsf{com}_i(z_i),\mathsf{d}_i(z_i),\keys{ck}_i(z_i))\leftarrow \algo{Commit}(1^\lambda,z_i)
\end{align}
for all $i\in[n]$.
\item Generate a simulatable witness $\rho_\mathsf{hist}$ for the instance $\mathsf{x}$ and generate $X^xZ^z\rho_\mathsf{hist} Z^zX^x$.
\item Send the first message (commitment),
$\keys{msg}_1\coloneqq(X^xZ^z\rho_\mathsf{hist} Z^zX^x)\otimes \mathsf{com}(x)\otimes \mathsf{com}(z)$, to the verifier,
where $\mathsf{com}(x)\coloneqq\bigotimes_{i=1}^{n} \mathsf{com}_i(x_i)$ and $\mathsf{com}(z)\coloneqq\bigotimes_{i=1}^n \mathsf{com}_i(z_i)$.
\end{itemize}
\item[The second action by the verifier (challenge phase):] $ $
\begin{itemize}
\item Generate $c\leftarrow [m]$.
\item Compute $\mathsf{cert}_i(x_i)\leftarrow\algo{Del}(\mathsf{com}_i(x_i))$ and $\mathsf{cert}_i(z_i)\leftarrow \algo{Del}(\mathsf{com}_i(z_i))$ for all $i\in \overline{S}_c$.
\item Send the second message (challenge), $\keys{msg}_2\coloneqq (c,\{\mathsf{cert}_i(x_i),\mathsf{cert}_i(z_i)\}_{i\in\overline{S}_c})$, to the prover.
\end{itemize}
\item[The third action by the prover (reply phase):] $ $
\begin{itemize}
\item Send the third message (reply), $\keys{msg}_3\coloneqq\{\mathsf{d}_i(x_i),\mathsf{d}_i(z_i)\}_{i\in S_{c}}$, to the verifier.
\item Output $\top$ if $\top\leftarrow\algo{Cert}(\mathsf{cert}_i(x_i),\keys{ck}_i(x_i))$ and $\top\leftarrow\algo{Cert}(\mathsf{cert}_i(z_i),\keys{ck}_i(z_i))$ for all $i\in\overline{S}_c$,
and output $\bot$ otherwise.
\end{itemize}
\item[The fourth action by the verifier (verification phase):] $ $
\begin{itemize}
\item Compute $x'_i\leftarrow\algo{Verify}(\mathsf{com}_i(x_i),\mathsf{d}_i(x_i))$ and $z'_i\leftarrow\algo{Verify}(\mathsf{com}_i(z_i),\mathsf{d}_i(z_i))$ for all $i\in S_c$. If $x_i'=\bot$ or $z_i'=\bot$ for at least one $i\in S_c$, output $\bot$ and abort.
\item Apply $X_i^{x'_i}Z_{i}^{z'_i}$ on the $i$-th qubit of $X^xZ^z\rho_{\mathsf{hist}}Z^zX^x$ for each $i\in S_c$, and
perform the POVM measurement $\{\Pi_c,I-\Pi_c\}$ on the state.
\item Output $\top$ if the result $\Pi_c$ is obtained, and output $\bot$ otherwise.
\end{itemize}
\end{description}
\begin{theorem}\label{thm:everlasting_zero_proof}
$\Sigma_{\Xi\mathsf{cd}}$ is a certified everlasting zero-knowledge proof for $\compclass{QMA}$ with $(1-{\mathsf{negl}}(\lambda))$-completeness and $\left(1-\frac{1}{{\mathrm{poly}}(\lambda)}\right)$-soundness.
\end{theorem}
This is shown from the following \cref{lem:completeness,lem:soundness,lem:everlasting_zero,lem:computational_zero}.
\begin{lemma}\label{lem:completeness}
$\Sigma_{\Xi\mathsf{cd}}$ satisfies the $(1-{\mathsf{negl}}(\lambda))$-completeness and prover's completeness.
\end{lemma}
\begin{lemma}\label{lem:soundness}
If $\Sigma_{\mathsf{ccd}}$ is classical-extractor-based binding, then $\Sigma_{\Xi\mathsf{cd}}$ satisfies $\left(1-\frac{1}{{\mathrm{poly}}(\lambda)}\right)$-soundness.
\end{lemma}
\begin{lemma}\label{lem:everlasting_zero}
If $\Sigma_{\mathsf{ccd}}$ is certified everlasting hiding and computational hiding, then $\Sigma_{\Xi\mathsf{cd}}$ satisfies certified everlasting zero-knowledge.
\end{lemma}
\begin{lemma}\label{lem:computational_zero}
If $\Sigma_{\mathsf{ccd}}$ is computational hiding, then $\Sigma_{\Xi\mathsf{cd}}$ satisfies computational zero-knowledge.
\end{lemma}
\begin{proof}[Proof of \cref{lem:completeness}]
It is clear from the definition of $k$-$\compclass{SimQMA}$ (\cref{def:k-SimQMA}) and the correctness of $\Sigma_{\mathsf{ccd}}$.
\end{proof}
\begin{proof}[Proof of \cref{lem:soundness}]
Let us show the soundness by analyzing the case for $\mathsf{x}\in A_{\mathsf{no}}$.
The prover sends the first message to the verifier.
The first message consists of three registers, $RS$, $RCX$, and $RCZ$.
The register $RCX$ further consists of $n$ registers $\{RCX_i\}_{i\in[n]}$.
The register $RCZ$ also consists of $n$ registers $\{RCZ_i\}_{i\in[n]}$.
If the prover is honest, $RS$ contains $X^xZ^z \rho_\mathsf{hist} Z^zX^x$, $RCX_i$ contains $\mathsf{com}_i(x_i)$, and
$RCZ_i$ contains $\mathsf{com}_i(z_i)$.
Let $\mathsf{com}_{i,x}'$ and $\mathsf{com}_{i,z}'$ be the (reduced) states of the registers $RCX_i$ and $RCZ_i$, respectively.
Let $f_{i,x}'$ and $f_{i,z}'$ be classical parts of $\mathsf{com}_{i,x}'$ and $\mathsf{com}_{i,z}'$, respectively.
\if0
Without loss of generality, we can assume that the first message from the prover to the verifier is a quantum state $\psi$ and
a classical bit string that is supposed to be the classical part of $\mathsf{com}(a)\otimes\mathsf{com}(e)$,
because we can assume that the verifier immediately performs computational basis measurements on the registers that are supposed to contain the classical information.
Let $y_i$ and $z_i$ be bit stings that are supposed to be $\mathsf{com}(a_i)$ and $\mathsf{com}(e_i)$, respectively.
\fi
The verifier generates $c\leftarrow[m]$, and issues the deletion certificate.
The verifier sends $c$ and the deletion certificate to the prover.
The verifier then receives $\{\mathsf{d}_1^{x,i},\mathsf{d}_2^{x,i},\mathsf{d}_1^{z,i},\mathsf{d}_2^{z,i}\}_{i\in S_c}$ from the prover.
For each $i\in[n]$, let us define $\mathsf{d}_1^{*,x,i}$ and $\mathsf{d}_1^{*,z,i}$ by
$\mathsf{d}_1^{*,x,i}\leftarrow \mathsf{Ext}(f_{i,x}')$ and
$\mathsf{d}_1^{*,z,i}\leftarrow \mathsf{Ext}(f_{i,z}')$, respectively.
Note that each $\mathsf{d}_1^{*,x,i}$ and $\mathsf{d}_1^{*,z,i}$ is independent of $c$, because
$\mathsf{com}_{i,x}'$ and $\mathsf{com}_{i,z}'$ are sent to the verifier before the verifier chooses $c$.
We have only to consider the case when
$\mathsf{d}_1^{x,i}=\mathsf{d}_1^{*,x,i}$
and
$\mathsf{d}_1^{z,i}=\mathsf{d}_1^{*,z,i}$
for all $i\in S_c$,
because of the following reason:
Due to the classical-extractor-based binding of $\Sigma_{\mathsf{ccd}}$,
$\algo{Verify}(\mathsf{com}_{i,x}',(\mathsf{d}_1^{x,i},\mathsf{d}_2^{x,i}))=\bot$
for any $\mathsf{d}_1^{x,i}\neq \mathsf{d}_1^{*,x,i}$ and any $\mathsf{d}_2^{x,i}$.
Similarly,
$\algo{Verify}(\mathsf{com}_{i,z}',(\mathsf{d}_1^{z,i},\mathsf{d}_2^{z,i}))=\bot$
for any $\mathsf{d}_1^{z,i}\neq \mathsf{d}_1^{*,z,i}$ and any $\mathsf{d}_2^{z,i}$.
Therefore, the prover who wants to make the verifier accept has to send
$\mathsf{d}_1^{x,i}=\mathsf{d}_1^{*,x,i}$
and
$\mathsf{d}_1^{z,i}=\mathsf{d}_1^{*,z,i}$ for all $i\in S_c$.
Let us define
\begin{align}
p(x,z)\coloneqq \Pr\left[\bigwedge_{i\in[n]}\left(\algo{Verify}_2 (\mathsf{com}_{i,x}',\mathsf{d}_1^{*,x,i})\to x_i \wedge \algo{Verify}_2(\mathsf{com}_{i,z}',\mathsf{d}_1^{*,z,i})\to z_i\right)\right].
\end{align}
Note that $p(x,z)$ is independent of $c$, because
$\{\mathsf{com}_{i,x}',\mathsf{com}_{i,z}'\}_{i\in[n]}$
and
$\{\mathsf{d}_1^{*,x,i},\mathsf{d}_1^{*,z,i}\}_{i\in [n]}$ are independent of $c$.
Let $\psi$ be the (reduced) state of the register $RS$.
The verifier's acceptance probability is
\begin{align}
&\frac{1}{m}\sum_{c\in[m]}\sum_{x,z\in\{0,1\}^n}p(x,z)\mathrm{Tr}\left[\Pi_c
\left(\prod_{i\in S_c}Z_i^{z_i}X_i^{x_i}\right)
\psi
\left(\prod_{i\in S_c}X_i^{x_i}Z_i^{z_i}\right)\right]\\
&=\frac{1}{m}\sum_{c\in[m]}\sum_{x,z\in\{0,1\}^n}p(x,z)\mathrm{Tr}\left[\Pi_c
\left(\prod_{i\in [n]}Z_i^{z_i}X_i^{x_i}\right)
\psi
\left(\prod_{i\in [n]}X_i^{x_i}Z_i^{z_i}\right)\right]\\
&=\frac{1}{m}\sum_{c\in[m]}
\mathrm{Tr}\left[\Pi_c
\sum_{x,z\in\{0,1\}^n}p(x,z)
\left(\prod_{i\in [n]}Z_i^{z_i}X_i^{x_i}\right)
\psi
\left(\prod_{i\in [n]}X_i^{x_i}Z_i^{z_i}\right)\right]\\
&\leq 1-\frac{1}{{\mathrm{poly}}(\lambda)},
\end{align}
where the last inequality comes from \cref{def:k-SimQMA}.
This completes the proof.
\end{proof}
\begin{proof}[Proof of \cref{lem:everlasting_zero}]
Let us show certified everlasting zero-knowledge.
For a subset $S_c\subseteq[n]$ and $x,z\in\{0,1\}^n$, let us define $x^{S_c}\coloneqq (x^{S_c}_1,x^{S_c}_2,\cdots, x^{S_c}_n)$ and $z^{S_c}\coloneqq (z^{S_c}_1,z^{S_c}_2,\cdots,z^{S_c}_n)$,
where $x^{S_c}_{i}=x_{i}$ and $z^{S_c}_{i}=z_{i}$ for $i\in S_{c}$,
and $x^{S_c}_{i}=z^{S_c}_{i}=0$ for $i\notin S_{c}$.
For clarity, we describe how the interactive algorithm $\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}^*(\xi)\rangle (\mathsf{x})$ runs against a QPT verifier $\mathcal{V}^*$ with an input $\xi$,
where $\mathsf{w}$ is the witness and $\mathsf{x}$ is the instance.
\begin{description}
\item[$\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}^*(\xi)\rangle (\mathsf{x})$:] $ $
\begin{enumerate}
\item $\mathcal{P}$ generates $x,z\leftarrow \{0,1\}^n$, and computes
\begin{align}
&(\mathsf{com}_i(x_i),\mathsf{d}_i(x_i),\keys{ck}_i(x_i))\leftarrow\algo{Commit}(1^\lambda,x_i) \\
&(\mathsf{com}_i(z_i),\mathsf{d}_i(z_i),\keys{ck}_i(z_i))\leftarrow\algo{Commit}(1^\lambda,z_i)
\end{align}
for all $i\in [n]$.
$\mathcal{P}$ sends $\keys{msg}_1\coloneqq(X^xZ^z\rho_\mathsf{hist} Z^zX^x)\otimes \mathsf{com}(x)\otimes\mathsf{com}(z)$ to $\mathcal{V}^*$.
\item $\mathcal{V}^*$ appends $\xi$ to the received state, and
runs a QPT circuit $V_1^*$ on it to obtain $(c,\{\mathsf{cert}_{i,x}',\mathsf{cert}_{i,z}'\}_{i\in\overline{S}_c})$.
$\mathcal{V}^*$ sends $\keys{msg}_2\coloneqq(c,\{\mathsf{cert}_{i,x}',\mathsf{cert}_{i,z}'\}_{i\in\overline{S}_c})$ to $\mathcal{P}$.
\item $\mathcal{P}$ sends $\keys{msg}_3\coloneqq\{\mathsf{d}_i(x_i),\mathsf{d}_i(z_i)\}_{i\in S_c}$ to $\mathcal{V}^*$.
\item $\mathcal{V}^*$ appends $\keys{msg}_3$ to its state, and runs a QPT circuit $V_2^*$ on it. $\mathcal{V}^*$ outputs its state $\xi'$.
\item $\mathcal{P}$ computes $\algo{Cert}(\mathsf{cert}_{i,x}',\keys{ck}_i(x_i))$ and $\algo{Cert}(\mathsf{cert}_{i,z}',\keys{ck}_i(z_i))$ for all $i\in\overline{S}_c$.
If all outputs are $\top$, then $\mathcal{P}$ outputs $\top$.
Otherwise, $\mathcal{P}$ outputs $\bot$.
\end{enumerate}
\end{description}
Next let us define a simulator $\mathcal{S}^{(1)}$ as follows.
\begin{description}
\item [The simulator $\mathcal{S}^{(1)}(\mathsf{x},\mathcal{V}^*,\xi)$:] $ $
\begin{enumerate}
\item Pick $c\leftarrow [m]$ and $x,z\leftarrow \{0,1\}^n$.
Compute
\begin{align}
& (\mathsf{com}_i(x^{S_c}_i),\mathsf{d}_i(x^{S_c}_i),\keys{ck}_i(x^{S_c}_i))\leftarrow\algo{Commit}(1^\lambda,x^{S_c}_i)\\
& (\mathsf{com}_i(z^{S_c}_i),\mathsf{d}_i(z^{S_c}_i),\keys{ck}_i(z^{S_c}_i))\leftarrow\algo{Commit}(1^\lambda,z^{S_c}_i)
\end{align}
for all $i\in [n]$.
\item Generate $(X^xZ^z\sigma(c) Z^zX^x) \otimes \mathsf{com}(x^{S_c})\otimes \mathsf{com}(z^{S_c})\otimes \xi$,
where $\sigma(c)\coloneqq \rho_\mathsf{sim}^{\mathsf{x},S_c}\otimes\left(\prod_{i\in \overline{S}_c}|0\rangle\langle0|_i\right)$.
Run $V_1^*$ on the state to obtain $(c',\{\mathsf{cert}_{i,x}',\mathsf{cert}_{i,z}'\}_{i\in\overline{S}_{c'}})$.
\item If $c'\neq c$, abort and output a fixed state $\eta$ and the flag state $\mathsf{fail}$.
\item Append $\{\mathsf{d}_i(x^{S_c}_i),\mathsf{d}_i(z^{S_c}_i)\}_{i\in S_c}$ to its quantum state, and run $V_2^*$ on the state to obtain $\xi'$.
\item Compute $\algo{Cert}(\mathsf{cert}_{i,x}',\keys{ck}_i(x^{S_c}_i))$ and $\algo{Cert}(\mathsf{cert}_{i,z}',\keys{ck}_i(z^{S_c}_i))$ for all $i\in \overline{S}_c$.
If all outputs are $\top$, then output the state $(\top,\xi')$. Otherwise, output $(\bot,\bot)$.
Also output the flag state $\mathsf{success}$.
\end{enumerate}
\end{description}
Let us also define other two simulators, $\mathcal{S}^{(2)}$ and $\mathcal{S}^{(3)}$, as follows.
\begin{description}
\item [The simulator $\mathcal{S}^{(2)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\xi)$:] $ $
It is the same as $\mathcal{S}^{(1)}$ except that
$\sigma(c)$ is replaced with $\rho_{\mathsf{hist}}$.
\end{description}
\begin{description}
\item [The simulator $\mathcal{S}^{(3)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\xi)$:] $ $
$\mathcal{S}^{(3)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\cdot~)$ is the channel that postselects the output of
\\$\mathcal{S}^{(2)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\cdot~)$ on the non-aborting state. More precisely, if we
write $\mathcal{S}^{(2)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\rho_{in})=p \rho_{out}\otimes \mathsf{success}+(1-p)\eta\otimes \mathsf{fail}$,
where $p$ is the non-aborting probability,
$\mathcal{S}^{(3)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\rho_{in})= \rho_{out}$.
\end{description}
\cref{lem:everlasting_zero} is shown from the following \cref{prop:abort,prop:indistinguishable_S_S'_ever,prop:indistinguishable_S'_V^*_ever}
(whose proofs will be given later) and quantum rewinding lemma (\cref{lemma:rewinding}),
which is used
to reduce the probability that $\mathcal{S}^{(1)}$ aborts to ${\mathsf{negl}}(\lambda)$.
In fact, from \cref{prop:abort,lemma:rewinding}, there exists a quantum circuit $\mathcal{S}^{(0)}$ of size at most
$O(m\,{\rm poly}(n){\rm size}(\mathcal{S}^{(1)}))$
such that the probability that $\mathcal{S}^{(0)}$ aborts is ${\mathsf{negl}}(\lambda)$, and
the output quantum states of $\mathcal{S}^{(0)}$ and $\mathcal{S}^{(1)}$ are ${\mathsf{negl}}(\lambda)$-close when they do not abort.
From \cref{prop:indistinguishable_S_S'_ever,prop:indistinguishable_S'_V^*_ever},
$\mathcal{S}^{(0)}$ is ${\mathsf{negl}}(\lambda)$-close to the real protocol,
which completes the proof.
\end{proof}
\begin{proposition}\label{prop:abort}
If $\Sigma_{\mathsf{ccd}}$ is computationally hiding, then
the probability that $\mathcal{S}^{(1)}$ does not abort is $\frac{1}{m}\pm{\mathsf{negl}}(\lambda)$.
\end{proposition}
\begin{proposition}\label{prop:indistinguishable_S_S'_ever}
$\mathcal{S}^{(1)}(\mathsf{x},\mathcal{V}^*,\cdot~)\approx_s\mathcal{S}^{(2)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\cdot~)$
for any $\mathsf{x}\in A_{\mathsf{yes}}\cap\{0,1\}^\lambda$
and any $\mathsf{w}\in R_A(\mathsf{x})$.
\end{proposition}
\begin{proposition}\label{prop:indistinguishable_S'_V^*_ever}
If $\Sigma_{\mathsf{ccd}}$ is certified everlasting hiding,
$\mathcal{S}^{(3)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\cdot~)\approx_s
\mathsf{OUT'}_{\mathcal{P},\mathcal{V}^*}\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}^*(\cdot)\rangle (\mathsf{x})$.
\end{proposition}
\begin{proof}[Proof of \cref{prop:abort}]
This can be shown similarly to \cite[Lemma~5.6]{FOCS:BroGri20}. For the convenience of readers, we provide a proof in~\cref{proof:abort}.
\end{proof}
\begin{proof}[Proof of \cref{prop:indistinguishable_S_S'_ever}]
It is clear from the local simulatability (\cref{def:k-SimQMA}) and
the definition of $x^{S_c}$ and $z^{S_c}$ (all $x_i^{S_c}$ and $z_i^{S_c}$ are 0 except for those in $i\in S_c$).
\if0
By the definition of $x^{S_c}$ and $z^{S_c}$, all $x_i^{S_c}$ and $z_i^{S_c}$ are 0 except for $i\in S_c$.
Therefore, the states of $\mathcal{V}^*$ simulated by
$\mathcal{S}^{(1)}$ and $\mathcal{S}^{(2)}$ are
$\mathrm{Tr}_{\overline{S}_c}(\rho_\mathsf{hist})$ and $\rho_\mathsf{sim}^{\mathsf{x},S_c}$, respectively.
Due to the local simulatability (\cref{def:k-SimQMA}), they are statistically indistinguishable from each other.
\fi
\end{proof}
\begin{proof}[Proof of \cref{prop:indistinguishable_S'_V^*_ever}]
We prove the proposition by contradiction.
We construct an adversary $\mathcal{B}$ that breaks the security of the certified
everlasting hiding of $\Sigma_{\mathsf{ccd}}$ by assuming
the existence of a distinguisher $\mathcal{D}$ that distinguishes two states $\delta_0$ and $\delta_1$,
\begin{align}
&\delta_0\coloneqq(\mathsf{OUT'}_{\mathcal{P},\mathcal{V}^*}\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}^*(\cdot)\rangle (\mathsf{x})\otimes I)\sigma\\
&\delta_1\coloneqq(\mathcal{S}^{(3)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\cdot~)\otimes I)\sigma,
\end{align}
with a certain state $\sigma$.
Let us describe how $\mathcal{B}$ works.
\begin{enumerate}
\item $\mathcal{B}$ generates $c\leftarrow[m]$ and $x,z\leftarrow\{0,1\}^n$.
\item $\mathcal{B}$ sends $m_0\coloneqq \{x_i,z_i\}_{i\in\overline{S}_c}$ and $m_1\coloneqq0^{2n-10}$ to the challenger of $\expb{\Sigma_{\mathsf{ccd}},\mathcal{B}}{bit}{ever}{hide}(\lambda,b)$.
$\mathcal{B}$ receives commitments from the challenger which is either $\{\mathsf{com}_i(x_i),\mathsf{com}_i(z_i)\}_{i\in \overline{S}_c}$ or $\{\mathsf{com}_i(0),\mathsf{com}_i(0)\}_{i\in \overline{S}_c}$.
\item $\mathcal{B}$ computes
\begin{align}
& (\mathsf{com}_i(x_i),\mathsf{d}_i(x_i),\keys{ck}_i(x_i))\leftarrow\algo{Commit}(1^\lambda,x_i)\\
& (\mathsf{com}_i(z_i),\mathsf{d}_i(z_i),\keys{ck}_i(z_i))\leftarrow\algo{Commit}(1^\lambda,z_i)
\end{align}
for $i\in S_c$ by itself.
\item $\mathcal{B}$ generates $X^xZ^z\rho_{\mathsf{hist}}Z^zX^x$.
$\mathcal{B}$ appends commitments and $\sigma$ to the quantum state.
If the commitments for ${i\in \overline{S}_c}$ are $\{\mathsf{com}_i(x_i),\mathsf{com}_i(z_i)\}_{i\in \overline{S}_c}$,
$\mathcal{B}$ obtains $(X^xZ^z\rho_{\mathsf{hist}}Z^zX^x)\otimes \mathsf{com}(x)\otimes\mathsf{com}(z)\otimes \sigma$.
If the commitments for ${i\in \overline{S}_c}$ are $\{\mathsf{com}_i(0),\mathsf{com}_i(0)\}_{i\in \overline{S}_c}$,
$\mathcal{B}$ obtains $(X^xZ^z\rho_{\mathsf{hist}}Z^zX^x)\otimes \mathsf{com}(x^{S_c})\otimes\mathsf{com}(z^{S_c})\otimes \sigma$.
\item $\mathcal{B}$ runs $V_1^*$ on it to obtain $(c',\{\mathsf{cert}_{i,x}',\mathsf{cert}_{i,z}'\}_{i\in \overline{S}_{c'}})$.
$\mathcal{B}$ aborts when $c\neq c'$.
\item $\mathcal{B}$ appends $\{\mathsf{d}_{i}(x_i),\mathsf{d}_{i}(z_i)\}_{i\in S_c}$ to the post-measurement state and runs $V_2^*$ on it to obtain $\sigma'$.
\item $\mathcal{B}$ sends $\{\mathsf{cert}_{i,x}',\mathsf{cert}_{i,z}'\}_{i\in \overline{S}_{c}}$ to the challenger of $\expb{\Sigma_{\mathsf{ccd}},\mathcal{B}}{bit}{ever}{hide}(\lambda,b)$,
and receives $\bot$ or
$\{\mathsf{d}_{i}(x_i),\mathsf{d}_{i}(z_i)\}_{i\in \overline{S}_c}$ and
$\{\keys{ck}_i(x_i),\keys{ck}_i(z_i)\}_{i\in \overline{S}_c}$ from the challenger.
\item $\mathcal{B}$ passes $(\bot,\bot)$ to $\mathcal{D}$ if $\mathcal{B}$ receives $\bot$ from the challenger, and passes $(\top,\sigma')$ to $\mathcal{D}$ otherwise.
\item When $\mathcal{D}$ outputs $b$, $\mathcal{B}$ outputs $b$.
\end{enumerate}
When $\mathcal{B}$ receives $\{\mathsf{com}_i(x_i),\mathsf{com}_i(z_i)\}_{i\in \overline{S}_c}$ from the challenger and it does not abort,
it simulates \\${\mathsf{OUT}}'_{\mathcal{P},\mathcal{V}^*}\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}^*(\cdot)\rangle(\mathsf{x})$.
Because $(X^xZ^z\rho_{\mathsf{hist}}Z^zX^x)\otimes \mathsf{com}(x)\otimes\mathsf{com}(z)\otimes \sigma$ is independent of $c$,
the probability that $\mathcal{B}$ does not abort is $\frac{1}{m}$.
Therefore,
$\mathcal{B}$ can simulate
${\mathsf{OUT}}'_{\mathcal{P},\mathcal{V}^*}\langle \mathcal{P}(\mathsf{w}^{\otimes k(|\mathsf{x}|)}),\mathcal{V}^*(\cdot)\rangle(\mathsf{x})$
with probability $\frac{1}{m}$.
When $\mathcal{B}$ receives $\{\mathsf{com}_i(0),\mathsf{com}_i(0)\}_{i\in \overline{S}_c}$ from the challenger and it does not abort,
it simulates $\mathcal{S}^{(3)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\cdot~)$.
The probability that
$\mathcal{B}$ does not abort is $\frac{1}{m}\pm{\mathsf{negl}}(\lambda)$ from
\cref{prop:abort,prop:indistinguishable_S_S'_ever}.
Therefore, $\mathcal{B}$ can simulate
$\mathcal{S}^{(3)}(\mathsf{x},\mathsf{w}^{\otimes k(|\mathsf{x}|)},\mathcal{V}^*,\cdot~)$ with probability
$\frac{1}{m}\pm{\mathsf{negl}}(\lambda)$.
Therefore, if there exists a distinguisher $\mathcal{D}$ that distinguishes
$\delta_0$ and $\delta_1$,
$\mathcal{B}$ can distinguish $\{\mathsf{com}_i(x_i),\mathsf{com}_i(z_i)\}_{i\in \overline{S}_c}$ from $\{\mathsf{com}_i(0),\mathsf{com}_i(0)\}_{i\in \overline{S}_c}$.
From \cref{lemma:eachbit},
this contradicts the certified everlasting hiding of $\Sigma_{\mathsf{ccd}}$.
\end{proof}
\begin{proof}[Proof of \cref{lem:computational_zero}]
Computational zero-knowledge can be proven similarly to \cite[Lemma~5.3]{FOCS:BroGri20} because our protocol is identical to theirs if we ignore the deletion certificates, which are irrelevant to the computational zero-knowledge property. For the convenience of readers, we provide a proof in \cref{proof:computational_zero}.
\end{proof}
\subsection{Sequential Repetition for Certified Everlasting Zero-Knowledge Proof for QMA}\label{sec:Sequential_repetition}
In this section, we amplify the completeness-soundness gap of the three-round protocol constructed in the previous
section by sequential repetition.
\begin{theorem}\label{thm:sequential_everlasting_zero_proof}
Let $\Sigma_{\Xi\mathsf{cd}}$ be a certified everlasting zero-knowledge proof for a $\compclass{QMA}$ promise problem $A$ with $\left(1-{\mathsf{negl}}(\lambda)\right)$-completeness
and $\left(1-\frac{1}{\rm poly(\lambda)}\right)$-soundness.
For any polynomial $N={\rm poly(\lambda)}$, let $\Sigma_{\Xi\mathsf{cd}}^N$ be the $N$-sequential repetition of $\Sigma_{\Xi\mathsf{cd}}$.
That is, $\mathcal{P}$ and $\mathcal{V}$ in $\Sigma_{\Xi\mathsf{cd}}^N$ run $\Sigma_{\Xi\mathsf{cd}}$ sequentially $N$ times.
Let $\mathcal{P}_j$ and $\mathcal{V}_j$ be the prover and the verifier in the $j$-th run of $\Sigma_{\Xi\mathsf{cd}}$,
respectively.
$\mathcal{P}$ in $\Sigma_{\Xi\mathsf{cd}}^N$ outputs $\top$ if $\mathcal{P}_j$ outputs $\top$ for all $j\in[N]$, and outputs $\bot$ otherwise.
$\mathcal{V}$ in $\Sigma_{\Xi\mathsf{cd}}^N$ outputs $\top$ if $\mathcal{V}_j$ outputs $\top$ for all $j\in[N]$, and outputs $\bot$ otherwise.
$\Sigma_{\Xi\mathsf{cd}}^{N}$ is a certified everlasting zero-knowledge proof for $A$ with $(1-{\mathsf{negl}}(\lambda))$-completeness and ${\mathsf{negl}}(\lambda)$-soundness.
\end{theorem}
\ifnum1=1
\begin{proof}
We provide a proof in \cref{Sec:sequential_everlasting}.
\end{proof}
\else
\input{Proof_sequential}
\fi
\section*{Acknowledgement}
TM is supported by
the JST Moonshot R\verb|&|D JPMJMS2061-5-1-1,
JST FOREST,
MEXT Q-LEAP,
and the Grant-in-Aid for Scientific Research (B) No.JP19H04066 of JSPS.
\fi
\input{reference}
\ifnum0=1
\else
\section{Proof of \texorpdfstring{\cref{lemma:eachbit}}{Lemma~\ref{lemma:eachbit}}}\label{sec:proof_of_bit}
Let us consider the following hybrids for $j\in\{0,1,...,n\}$.
\begin{description}
\item[$\sfhyb{j}{}$:] $ $
\begin{enumerate}
\item $\mathcal{A}_1$ generates $(m^0,m^1)\in \{0,1\}^n\times\{0,1\}^n$ and sends it to the challenger.
\item The challenger computes
\begin{align}
(\mathsf{com}_i(m_i^1),\mathsf{d}_i(m_i^1),\keys{ck}_i(m_i^1))\leftarrow \algo{Commit}(1^{\lambda},m^1_i)
\end{align}
for $i\in[j]$ and
\begin{align}
(\mathsf{com}_i(m_i^0),\mathsf{d}_i(m_i^0),\keys{ck}_i(m_i^0))\leftarrow \algo{Commit}(1^{\lambda},m^0_i)
\end{align}
for each $i\in\{j+1,...,n\}$,
and sends $\{\mathsf{com}_i(m_i^1)\}_{i\in[j]}$ and $\{\mathsf{com}_i(m_i^0)\}_{i\in\{j+1,...,n\}}$ to $\mathcal{A}_1$.
Here, $m_i^b$ is the $i$-th bit of $m^b$.
\item At some point, $\mathcal{A}_1$ sends $\{\mathsf{cert}_i\}_{i\in[n]}$ to the challenger,
and sends its internal state to $\mathcal{A}_2$.
\item The challenger computes $\algo{Cert}(\mathsf{cert}_i,\keys{ck}_i(m_i^1))$ for each $i\in[j]$
and
$\algo{Cert}(\mathsf{cert}_i,\keys{ck}_i(m_i^0))$ for each $i\in\{j+1,...,n\}$.
If the outputs are $\top$ for all $i\in[n]$, then the challenger outputs $\top$,
and sends
$\{\mathsf{d}_i(m_i^1),\keys{ck}_i(m_i^1)\}_{i\in[j]}$
and
$\{\mathsf{d}_i(m_i^0),\keys{ck}_i(m_i^0)\}_{i\in\{j+1,...,n\}}$
to $\mathcal{A}_2$.
Else, the challenger outputs $\bot$, and sends $\bot$ to $\mathcal{A}_2$.
\item $\mathcal{A}_2$ outputs $b'\in\{0,1\}$.
\item If the challenger outputs $\top$, then the output of the experiment is $b'$. Otherwise, the output of the experiment is $\bot$.
\end{enumerate}
\end{description}
It is clear that
$\sfhyb{0}{}=\expb{\Sigma,\mathcal{A}}{bit}{ever}{hide}(\lambda,0)$ and $\sfhyb{n}{}=\expb{\Sigma,\mathcal{A}}{bit}{ever}{hide}(\lambda,1)$.
Furthermore, we can show
\begin{align}
\left|\Pr[\sfhyb{j}{}=1]-\Pr[\sfhyb{j+1}{}=1]\right|\le{\mathsf{negl}}(\lambda)
\end{align}
for each $j\in\{0,1,...,n-1\}$. (Its proof is given below.)
From these facts, we obtain \cref{lemma:eachbit}.
Let us show the remaining one. To show it, let us assume that
$\left|\Pr[\sfhyb{j}{}=1]-\Pr[\sfhyb{j+1}{}=1]\right|$ is non-negligible.
Then, we can construct an adversary $\mathcal{B}$ that can break the certified everlasting hiding of
$\Sigma_{\mathsf{ccd}}$ as follows.
\begin{enumerate}
\item
$\mathcal{B}$ receives $(m^0,m^1)$ from $\mathcal{A}_1$, and computes
\begin{align}
(\mathsf{com}_i(m_i^1),\mathsf{d}_i(m_i^1),\keys{ck}_i(m_i^1))\leftarrow \algo{Commit}(1^{\lambda},m^1_i)
\end{align}
for $i\in[j]$
and
\begin{align}
(\mathsf{com}_i(m_i^0),\mathsf{d}_i(m_i^0),\keys{ck}_i(m_i^0))\leftarrow \algo{Commit}(1^{\lambda},m^0_i)
\end{align}
for $i\in\{j+2,...,n\}$.
\item
$\mathcal{B}$ sends $(m_{j+1}^0,m_{j+1}^1)$ to the challenger of
$\expa{\Sigma_{\mathsf{ccd}},\mathcal{B}}{ever}{hide}(\lambda,b')$,
and receives $\mathsf{com}_{j+1}(m_{j+1}^{b'})$ from the challenger.
\item
$\mathcal{B}$ sends
$\{\mathsf{com}_i(m_i^1)\}_{i\in[j]}$,
$\mathsf{com}_{j+1}(m_{j+1}^{b'})$, and
$\{\mathsf{com}_i(m_i^0)\}_{i\in\{j+2,...,n\}}$,
to $\mathcal{A}_1$.
\item
$\mathcal{A}_1$ sends $\{\mathsf{cert}_i\}_{i\in[n]}$ to $\mathcal{B}$, and sends its internal state to $\mathcal{A}_2$.
\item
$\mathcal{B}$ sends $\mathsf{cert}_{j+1}$ to the challenger of
$\expa{\Sigma_{\mathsf{ccd}},\mathcal{B}}{ever}{hide}(\lambda,b')$,
and receives
$(\mathsf{d}_{j+1}(m_{j+1}^{b'}),\keys{ck}_{j+1}(m_{j+1}^{b'}))$ or $\bot$ from the challenger.
If $\mathcal{B}$ receives $\bot$ from the challenger, it outputs $\bot$ and aborts.
\item
$\mathcal{B}$ sends all $\mathsf{d}_i$ and $\keys{ck}_i$ to $\mathcal{A}_2$.
\item
$\mathcal{A}_2$ outputs $b''$.
\item
$\mathcal{B}$ computes $\algo{Cert}$ for all $\mathsf{cert}_i$, and outputs $b''$ if all results are $\top$.
Otherwise, $\mathcal{B}$ outputs $\bot$.
\end{enumerate}
It is clear that $\Pr[\mathcal{B}\to1\mid b'=0]=\Pr[\sfhyb{j}{}=1]$ and
$\Pr[\mathcal{B}\to1\mid b'=1]=\Pr[\sfhyb{j+1}{}=1]$.
By assumption,
$|\Pr[\sfhyb{j}{}=1]-\Pr[\sfhyb{j+1}{}=1]|$
is non-negligible, and therefore
$|\Pr[\mathcal{B}\to1\mid b'=0]-\Pr[\mathcal{B}\to1\mid b'=1]|$ is non-negligible,
which contradict the certified everlasting hiding of $\Sigma_{\mathsf{ccd}}$.
|
\subsection{Experiment Setup}
\label{subsec:setup}
\paragraph{Datasets.} We perform experiments on the COCO-caption (COCO)~\cite{chen2015coco-cap} and Flickr30K (F30K)~\cite{young2014image} datasets. Each image of these two datasets is associated with five sentences. Flickr30K contains 31,000 images, and we use the same data split as~\cite{faghri2017vse++}, where there are 29,000 training images, 1000 test images, and 1000 validation images. COCO contains 123,287 images in total. For fast iteration, we use a subset training data C30K, which contains the same amount of images as the F30K. Note that C30K is a training split. We also trained models on the full COCO training split. For COCO dataset, the results are evaluated on COCO 1K test split~\cite{karpathy2015deep}. We use COCO 1k test split for both in-domain (models trained on either C30K or full COCO training split and evaluate on COCO-caption) and cross-dataset transfer (models trained on F30K and evaluate on COCO-caption) evaluation. For both F30K and COCO 1K test split, there are 5,000 text queries and 1,000 candidate images to be retrieved. We report recall@1 (R1) and recall@5 (R5) as the primary retrieval metric.
\paragraph{Compositional generalization evaluation.} To generate evaluations of compositional generalization, we use a method similar to that of \citet{shaw2020compositional} and \citet{keysers2020measuring} which maximizes \emph{compound divergence} between the distribution of \emph{compounds} in the evaluation set and in the training set. Here compounds are defined based on the predicates occurring in captions. Following this method, we first calculate the overall divergence of compounds from the evaluation data to the training data using predicates from all the sentences. Then, for each sentence in the evaluation data, we calculate a compound divergence with this specific example removed. We rank those sentences based on the difference of the compound divergence. Finally, we choose the top-K sentences with the largest compound divergence differences and its corresponding images to form the evaluation splits.
Using this method, we generate evaluation splits with 1,000 images and 5,000 text queries, COCO-MCD and F30K-MCD, to assess models trained on F30K and COCO, respectively. Therefore, these splits assess both compositional generalization and cross-dataset transfer. Defining such splits across datasets is also helpful to achieve greater compound divergence than is otherwise possible, given the small amount of available in-domain test data. More details are included in Appendix.
\paragraph{CRG construction.} We constructed two CRGs on the F30K and C30K datasets, using the procedure mentioned in \S~\ref{sec:crg}. The key statistics of the graph we generated as shown in Table~\ref{tab:crg_stats}.
\paragraph{Baselines and our approach.} We compare {\textsc{Composer}}\xspace to two strong baseline methods, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, ViLBERT~\cite{Lu2019ViLBERTPT} and VSE~\cite{kiros2014UVS}. We make sure all models are using the same object-centric visual features extracted from the Up-Down object detector~\cite{anderson2017updown} for fair comparison. For the texts, both ViLBERT and the re-implemented VSE use the pre-trained BERT model as initialization. For the {\textsc{Composer}}\xspace, we only initialize the predicate Transformer with the pre-trained BERT, which uses the first six layers. Note that the ViLBERT results are re-produced using the codebase from its author. ViLBERT is \textbf{not pre-trained} on any additional data of image-text pairs to prevent information leak in both cross-dataset evaluation and compositional generalization. Therefore, we used the pre-trained BERT models provided by HuggingFace to initalize the text stream of ViLBERT, and then followed the rest procedure in the original ViLBERT paper. Please refer to {Appendix} for complete details.
\input{tables/main_comparison_retrieval}
\begin{figure}[t]
\centering
\begin{tabular}{@{}c@{}c@{}}
\includegraphics[width=0.245\textwidth]{figs/abs_perf_with_cd} &
\includegraphics[width=0.245\textwidth]{figs/relative_perf_with_cd}
\end{tabular}
\caption{{\textsc{Composer}}\xspace's results on generalization splits of different compound divergence over text description (evaluated under the F30K$\rightarrow$COCO setting).}
\vspace{-1em}
\label{fig:compound_div}
\end{figure}
\subsection{Main Results}
\label{subsec:main}
We compare the {{\textsc{Composer}}\xspace} with ViLBERT~\cite{Lu2019ViLBERTPT} and VSE~\cite{kiros2014UVS} on F30k and COCO for in-domain, zero-shot cross-dataset transfer, and compositional generalization (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot F30K$\rightarrow$COCO-MCD). The notation A$\rightarrow$B means that the model is trained on A and evaluated on B. We report the results of sentence-to-image retrieval in the main paper and defer more ablation study results to the Appendix.
\paragraph{In-domain performance.}
Table~\ref{tab:main} presents the in-domain performance on both F30k and COCO datasets. First, we observe that both {\textsc{Composer}}\xspace and ViLBERT consistently outperform VSE, which is expected as ViLBERT contains a cross-modal transformer with stronger modeling capacity. Comparing to ViLBERT, the {{\textsc{Composer}}\xspace} performs on par.
\paragraph{Zero-shot cross-dataset transfer.}
We also consider zero-shot cross dataset transfer where we evaluate models on a dataset that is different from the training dataset. In this setting, the {{\textsc{Composer}}\xspace} outperforms ViBLERT and VSE significantly. Concretely, on the F30k$\rightarrow$COCO setting, the {{\textsc{Composer}}\xspace} improves R1 and R5 by $11.0\%$ and $7.0\%$ over the ViLBERT, relatively. There are $10.0\%$ and $4.2\%$ relative improvements on R1 and R5 on the other transfer direction.
\paragraph{Compositional generalization.}
On the max compound divergence (MCD) split, {\textsc{Composer}}\xspace outperforms baselines by a margin for both F30K and C30K trained models (shown as Table~\ref{tab:main}). To further characterize the performance on compositional generalization, we create 16 test splits on each dataset with different compound divergence (from $0.15$ to $0.31$, where $0.31$ is the max CD) and present the results in Figure~\ref{fig:compound_div}. With the increases of CD, we observe the performance of {\textsc{Composer}}\xspace and ViLBERT decreases. Compared to ViLBERT, we observe that {\textsc{Composer}}\xspace is relatively more robust to this distribution shift, as the relative performance improvement is increasing with CD increases.
\subsection{Analysis and Ablation Study}
\label{subsec:ablation}
We perform several ablation studies to analyze {{\textsc{Composer}}\xspace}, and provide qualitative results to demonstrate the model's interpretability.
\input{tables/ablation_primitive}
\input{tables/ablation_predicate_types}
\input{tables/ablation_dense}
\begin{figure*}[t]
\centering
\includegraphics[width=0.985\textwidth]{figs/main_composer_decision_example2}
\caption{Interpreting the {\textsc{Composer}}\xspace using visual-semantic alignment scores, formatted as $\texttt{[}\vct{s}_{\texttt{GT}}, \vct{s}_{\sc \texttt{Negative}}\texttt{]}$. The left figure corresponds to a correct example, and the right figure corresponds to an incorrect one.}
\label{fig:interpretation}
\vspace{-1em}
\end{figure*}
\paragraph{Is \texttt{CrossAtt} in primitive encoding useful?}
Table~\ref{tab:primitive} compares variants of {{\textsc{Composer}}\xspace} with and without \texttt{CrossAtt} for primitive encoding, and shows that \texttt{CrossAtt} improves all metrics in in-domain and cross-dataset evaluation.
\paragraph{Which modulator works better?}
We consider three modulators to combine input concepts with the syntax token embeddings for later composition, which are \texttt{Replace}, \texttt{MLP}, and \texttt{FiLM}. The \texttt{Replace} directly replaces the syntax embedding with the input concept embedding. This is an inferior approach by design as it ignores the relative position of each concept. \texttt{MLP} model applies multi-layer neural networks on the concatenated syntax and input concept embeddings. \texttt{FiLM} model uses the syntax embedding to infer the parameter of an affine transformation, which is then applied to the input concepts. We show the results in Table~\ref{tab:predicate_type}. {\texttt{Replace}} achieves the worst performance, indicating the importance of identifying the position of input concepts. {\textsc{Composer}}\xspace chooses {\texttt{FiLM}} as the modulator given its strong performance over all metrics.
\paragraph{Is MVSA supervision useful?}
We evaluate the influence of multi-level visual-semantic alignment on sentence and phrase to image retrieval. In the phrase-to-image experiments, we sample 5 non-sentence concepts from the CRG for each annotation in the corresponding test data and use them as the query to report results (in R1). Table~\ref{tab:dense} presents the results.
With the MVSA, {\textsc{Composer}}\xspace outperforms ViLBERT on both sentence and phrase-based retrieval by a noticeable margin, indicating the advantage of capturing mid-level alignment in our model design.
Secondly, MVSA improves both {\textsc{Composer}}\xspace and ViLBERT on the phrase to image retrieval over their counterparts. However, adding MVSA on ViLBERT leads to a degradation of sentence-to-image retrieval, showing that ViLBERT is incapable of mastering visual alignments for both sentences and phrases simultaneously. {\textsc{Composer}}\xspace with MVSA improves itself on both sentence and phrase, showing strong multi-granular visual-semantic alignment ability.
\input{tables/ablation_complexity}
\paragraph{Performance vs. complexity trade-off.}
We compare variants of {\textsc{Composer}}\xspace with different parameter and computation budgets, which uses different numbers of layers for the Predicate Transformer (PT) and Composition Transformer (CT). The results are shown in Table~\ref{tab:complexity}. First, We keep the size of CT fixed and vary the size of PT. It shows a marginal performance decrease occurring as the \# of layers of PT goes down. Then we keep the size of PT fixed and decrease the capacity of CT, which presents a significant performance drop, showing the essential role CT is playing. Besides having superior results, {\textsc{Composer}}\xspace has \textit{(at least 33\%) fewer parameters} than the ViLBERT model, which indicates a potential performance gain could be achieved with a larger {\textsc{Composer}}\xspace model.
For computation complexity, we observe that the full {\textsc{Composer}}\xspace model is 50\% less efficient to a ViLBERT model, due to its recursive nature. Meanwhile, we notice that the increase in the \# of CT layers contributes a significant amount to the total computation time as every two additional layers adds $\sim10G$ FLOPS.
\paragraph{Performance under different parsing qualities.}
CRG is generated based on constituent parser. We investigate the performance of {\textsc{Composer}}\xspace with CRG under different parsing qualites. Given a parsing tree, We randomly remove its branches randomly with a probability of 0.1, 0.3, or 0.5 to generate a tree with degraded parsing quality. We evaluate {\textsc{Composer}}\xspace on the resulting CRGs. We summarized the results in Table~\ref{tab:parsing_qualities}. When parsing quality drops, both in-domain and cross-dataset transfer performance drops. The performance degrades by 12\%, when half of the parse could be missing. We expect with better parsing quality, {\textsc{Composer}}\xspace can achieve stronger performance.
\input{tables/ablation_parsing_quality}
\paragraph{Interpreting {\textsc{Composer}}\xspace's decision.}
Despite the solid performance, {\textsc{Composer}}\xspace is also highly interpretable. Specifically, we visualize its alignment scores along with the concept composition procedure in Figure~\ref{fig:interpretation}. Empirically, we observe that most failures are caused by visually grounding mistakes at the primitive concepts level. The error then propagates “upwards” towards concept composition.
For instance, the left example shows that {\textsc{Composer}}\xspace is confusing between the ground truth and negative image when only the text of shared visual concept ``a bold man'' is presented. With more information are given, it gets clarified immediately as it notices that the target sentence is composed not only with the above subject, but also with the prepositional phrases ``by the beer pumps at the bar'' that reflects the visual environment.
\paragraph{Scalability to full COCO dataset.} Finally, we trained our model (PT=6, CT=5) on the full COCO training split and evaluated for both in-domain and cross-dataset transfer task. We use the same hyperparameters as C30K. However, {\textsc{Composer}}\xspace underperforms the ViLBERT in this setting, as it achieves 56.06\% and 44.24\% in R1 for the in-domain task (COCO$\rightarrow$COCO) and cross-dataset evaluation tasks (COCO$\rightarrow$F30k), while ViLBERT obtains 56.83\% and 46.62\%, respectively.
We hypothesize that this negative result is largely due to the limited model capacity of the proposed {\textsc{Composer}}\xspace, as it has relatively $33\%$ less parameters comparing to ViLBERT. Meanwhile, it is also observed that {{\textsc{Composer}}\xspace} performs worse than ViLBERT in fitting training data. We observe that doubling the training epoch would increase both in-domain and out-of-domain performance by 2\% relatively. Increasing the layer of Composition Transformer (CT) to 7 would also improves R1 by 2.5\% relatively. Further scaling up {\textsc{Composer}}\xspace may resolve this issue but requires more computational resources, and we leave this for future research.
\section{Introduction}
\label{sec:intro}
\input{intro}
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth]{figs/graph_example}
\caption{Concepts and their visual denotations organized by the Concept \& Relation Graph}
\label{fig:crg}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=0.975\textwidth]{figs/composer_architecture}
\caption{\small The overall design of the proposed {\textsc{Composer}}\xspace model.}
\label{fig:composer}
\end{figure*}
\section{Concept \& Relation Graph}
\label{sec:crg}
\input{crg}
\section{{\textsc{Composer}}\xspace: Recursive Modeling of the Compositional Structure}
\label{sec:method}
\input{method}
\section{Related Work}
\label{sec:related}
\input{related}
\section{Experiment}
\label{sec:exp}
\input{experiment}
\section{Conclusion}
\label{sec:conclusion}
\input{conclusion}
\section*{Acknowledgements} This work is partially supported by NSF Awards IIS-1513966/ 1632803/1833137, CCF-1139148, DARPA Award\#: FA8750-18-2-0117, FA8750-19-1-0504, DARPA-D3M - Award UCB-00009528, Google Research Awards, gifts from Facebook and Netflix, and ARO\# W911NF-12-1-0241 and W911NF-15-1-0484. We thank anonymous EMNLP reviewers for constructive feedback. Additionally, we would like to thank Jason Baldbridge for reviewing an early version of this paper, and Kristina Toutanova for helpful discussion.
{
\bibliographystyle{acl_natbib}
\subsection{Encoding Primitives and Predicates}
\label{subsec:encoding}
Given a paired image and sentence $({\vct{x}}, \vct{y})$, we parse the sentence as the tree of primitives and predicates $({\vct{x}}, \mat{U}, \mat{E})$. Here, we represent the image as a set of visual feature vectors $\{\vct{\phi}\}$, which are the object-centric features from an object detector~\citep{anderson2017updown}. Noted that we didn’t use structural information beyond object proposals/regions. Our {{\textsc{Composer}}\xspace} takes the primitives and predicates as input and output the visually grounded concept embeddings, with both the primitives and predicates as continuous vectors of different contextualization.
\paragraph{Representing primitives with visual context.}
The primitive concepts refer to tokens which can be visually grounded, and we represent them as word embeddings contextualized with visual features. As such, we use a one-layer Transformer with the $\texttt{CrossAtt}$ mechanism, where $K$, $V$, and $Q$ are linear transformations of $\vct{\phi}$, $\vct{\phi}$, and $\vct{u}$, respectively. This essentially uses the word embedding to query the visual features and outputs the grounded primitive embeddings $\hat{\mat{U}} = \{ \hat{\vct{u}} \}$. Note that the output is always a single vector for each primitive as it is a single word.
\paragraph{Representing predicates as neural templates.} A predicate $\vct{e}$ is a semantic $n$-place function that combines multiple concepts into one. We represent it as a \textbf{template sentence} with words and syntactic placeholders, such as ``\texttt{[NP]$_1$} running on \texttt{[NP]$_2$}'', where those syntactic placeholders denote the positions and types of arguments. We encode such template sentences via \texttt{SelfAtt} mechanism, using a multi-layer Predicate Transformer (PT). The output of this model is a contextualized sequence of the words and syntactic placeholders as $\hat{\vct{e}}$.
\begin{figure}[t]
\centering
\includegraphics[width=0.5\textwidth]{figs/composition}
\caption{Details of the composition procedure.}
\label{fig:composition}
\vspace{-1em}
\end{figure}
\subsection{Recursive Concept Composition}
\label{subsec:composition}
With the encoded primitives $\hat{\mat{U}}$ and predicates $\hat{\mat{E}}$, the {{\textsc{Composer}}\xspace} then performs multiple recursive composition steps to obtain the grounded concept embedding, $\vct{v}{({\vct{x}}, \vct{y})}$, representing the visual-linguistic embedding of the sentence and the image as shown in the Figure~\ref{fig:composer}. To further illustrate this process, we detail the composition function in below, as shown in Figure~\ref{fig:composition}.
\paragraph{Input concept modulation.}
We use a modulator to bind the arguments in the predicate to the input child concepts. Given a encoded predicate $\hat{\vct{e}}=$ $\{\texttt{[NP]}_1$, \texttt{running}, \texttt{on}, $\texttt{[NP]}_2$, \texttt{with}, $\texttt{[NP]}_3\}$ and a input concept $\vct{c}_1=$ \texttt{``a man''}, the modulator is a neural network that takes the concept embedding $\vct{c}_1$ and its corresponding syntactic placeholder $\texttt{[NP]}_1$ as input and outputs a modulated embedding. This embedding is then reassembled with the embeddings of non-arguments in the predicate and used for the later stage.
For example, the output sequence becomes $\{\texttt{Mod}(\texttt{[NP]}_1, \vct{c}_1)$, \texttt{running}, \texttt{on}, $\texttt{Mod}(\texttt{[NP]}_2, \vct{c}_2)$, \texttt{with}, $\texttt{Mod}(\texttt{[NP]}_3, \vct{c}_3)\}$ after the modulator processed each pair of input concept and syntactic placeholder.
Various choices of neural networks are available for this modulator, such as a Multi-Layer Perceptron (MLP) or a Feature-wise Linear Modulation (FiLM)~\cite{perez2018film}. {\textsc{Composer}}\xspace uses FiLM for its strong empirical performance.
\paragraph{Contextualization with visual context.} After concept modulation, we get a sequence of embeddings for non-argument words of the predicate and the binded child concepts, which is then fed as an input to a Composition Transformer (CT) model. This Transformer has multiple layers, with both \texttt{CrossAtt} layers that attends to the object-centric visual features and \texttt{SelfAtt} layers that contextualize between tokens. Please refer to {Appendix} for the detailed network architecture.
Given that our model is recursive by nature, the computation complexity of CT is proportional to the depth of the tree. We provide a comprehensive study in \S~\ref{subsec:ablation} to show the correlation between the parameter/complexity and model's performances.
\subsection{Learning {\textsc{Composer}}\xspace with Visual-Semantic Alignments}
\label{subsec:learning}
With the composed grounded concept embedding $\vct{v}({\vct{x}}, \vct{y})$, we use the visual-semantic alignment as the primary objective to learn {{\textsc{Composer}}\xspace}. To this end, we compute the alignment score by learning an additional linear regressor $\theta$:
\begin{equation}
\vct{s}({{\vct{x}}, \vct{y}}) = \theta^\top \cdot \vct{v}({\vct{x}}, \vct{y}) \propto p({\vct{x}}, \vct{y}), \nonumber
\end{equation}
where $p({\vct{x}}, \vct{y})$ is the probability that the sentence and image is a good match pair. Then we learn the sentence to image alignment by minimizing the negative log-likelihood (NLL):
\begin{align}
\ell_{\textsc{match}} = - \sum_i \log \frac{\exp({s({\vct{x}}_i, \vct{y}_i)})}{\sum_{(\hat{{\vct{x}}}, \hat{\vct{y}}) \sim \mathcal{D}_i}\exp({s(\hat{{\vct{x}}}, \hat{\vct{y}})})} \nonumber
\end{align}
with $\mathcal{D}_i = \{({\vct{x}}_i, \vct{y}_i)\}\cup\mathcal{D}^-_i$. To properly normalize the probability, it is necessary to sample a set of negative examples to contrast. Thus, we generate $\mathcal{D}^-_i$ using the strategy of \citet{Lu2019ViLBERTPT}.
\paragraph{Multi-level visual-semantic alignment (MVSA).} Since {\textsc{Composer}}\xspace composes grounded concepts recursively from the primitives, we obtain the embeddings of all the intermediate concepts automatically. Therefore, it is natural to extend the alignment learning objectives to all those intermediate concepts. We optimize the triplet hinge loss~\citep{kiros2014UVS}:
\begin{align}
\ell_{\textsc{MVSA}} = & \sum_i \sum_{\vct{c} \in \mat{C}_i} [\alpha - {s}{({\vct{x}}_i, \vct{c})} + {s}{({\vct{x}}_i, {\vct{c}}^-)}]_+ \nonumber \\
& + [\alpha - {s}{({\vct{x}}_i, \vct{c})} + {s}{({{\vct{x}}}^-_i, \vct{c})}]_+ \nonumber
\end{align}
where $[h]_+ = \text{max}(0, h)$ denotes the hinge loss and $\alpha$ is the margin to be tuned. We derive the negative concepts $\vct{c}^-$ from the negative sentences in the $\mathcal{D}^-_i$. We observe that negative concepts at word/phrase levels are noisier than the ones at sentence level because many are common objects presented in the positive image and lead to ambiguity in learning. Therefore, we choose hinge loss over NLL because it is more robust to label noises~\cite{biggio2011support}.
\paragraph{Learning to preserve orders in the tree.} Finally, we use an order-preserving objective proposed by \citet{zhang2020learning}, to ensure that a fine-grained concept (closer to sentence) can produce a more confident alignment score than a coarse-grained concept (closer to primitive):
\begin{equation}
\ell_{\textsc{order}} = \sum_i \sum_{\vct{e}_{jk}} [\beta - s({\vct{x}}_{i}, \vct{c}_j) + s({\vct{x}}_{i}, \vct{c}_k) ]_+ \nonumber
\end{equation}
Here, $\vct{e}_{jk}$ represents a predicate connecting the $\vct{c}_j$ and $\vct{c}_k$, with $\vct{c}_j$ to be the fine-grained parent concept which is closer to the sentence and $\vct{c}_k$ to be the coarse-grained child concept which is closer to the primitives. $\beta$ is the margin that sets the constraint on how hard the order of embeddings should be reserved.
The complete learning objective is a weighted combination of three individual losses defined above, with the loss weights $\lambda_1 = 1$ and $\lambda_2 = 1$:
\begin{equation}
\ell = \ell_{\textsc{match}} + \lambda_1 \cdot \ell_{\textsc{mvsa}} + \lambda_2 \cdot \ell_{\textsc{order}} \nonumber
\end{equation}
The details of model optimization and hyper-parameter setting are included in the Appendix.
\section*{Appendix}
\input{tables/supp_consitutiency_tree}
In the Appendix, we provide details omitted from the main text due to the limited space, including:
\begin{itemize}[topsep=1pt,parsep=0pt,partopsep=4pt,leftmargin=*,itemsep=2pt]
\item \S~\ref{supp:sec:primitive_crg} describes the implementation details for extracting primitives \& predicates from the constituency tree (\S~2 of the main text).
\item In \S~\ref{supp:sec:generate_compositional_split}, we describes the details of generating the compositional evaluation splits (\S~5.1 of the main text).
\item \S~\ref{supp:sec:implementation_details} contains training and architecture details for {\textsc{Composer}}\xspace and baselines (\S~5.1 of the main text).
\item \S~\ref{supp:sec:ablation_exps} includes the ablation studies on learning objectives and margin of MVSA (\S~5.2 of the main text).
\end{itemize}
\section{Extracting Primitives \& Predicates from the Constituency Tree}
\label{supp:sec:primitive_crg}
As mentioned in the main paper, we parse the sentence and convert it into a tree of concepts and primitives. Particularly, we first perform constituency parsing using the self-attention parser~\cite{kitaev2018constituency}. Table~\ref{tab:example_constituency} provides the visualization for two examples of the syntax sub-trees. Next, we perform a tree search (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, breadth-first search) on the constituency tree of the current input concept to extract the sub-concepts and predicate functions. Note that this step is applied recursively until we can no longer decompose a concept into any sub-concepts. On a single step of the extraction, we enumerate each node in the constituency tree of current input text expression and examine whether a constituent satisfies the criterion that defines the visually grounded concept.
The concept criterion defined for the Flickr30K and COCO dataset contains several principles: (1) If the constituent is a word, it is a primitive concept if its Part-of-Speech (POS) tag is one of the following: $\{$\texttt{[NN]},\texttt{[NNS]},\texttt{[NNP]},\texttt{[NNPS]}$\}$; (2) If the constituent is a phrase (with two words or more), it would be a concept when this constituent contains a primitive word (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, satisfying condition (1)) and its constituency tag is one of the following: $\{$\texttt{[S]}, \texttt{[SBAR]}, \texttt{[SBARQ]}, \texttt{[SQ]}, \texttt{[SINV]}, \texttt{[NP]}, \texttt{[NX]}$\}$. After all the concepts are extracted, we take the remaining words in the current input text expression as the predicate that combines those concepts and use the tag to represent syntactic blank. Concrete examples can be found in the Table~\ref{tab:example_constituency}. For instance, in the first example, we search the text ``two dogs are running on the grass'' and extract two noun constituents, ``two dogs'' and ``the grass'' as the concepts. We use the remaining text "\texttt{[NP]} is running on \texttt{[NP]}" as the predicate that indicates the semantic meaning of how these two sub-concepts composes into the original sentence.
\section{Details on Generation of Compositional Evaluation Splits}
\label{supp:sec:generate_compositional_split}
As mentioned in the main text, we generate compositional generalization (CG) splits with 1,000 images and 5,000 text queries, maximizing the Compound Divergence (MCD) as \citet{shaw2020compositional}\footnote{We adopt the released code here for the computing compound divergence: https://github.com/google-research/language/tree/master/language/nqg/tasks}, to assess models' capability in generalizing to the data with different predicate distribution. Concretely, we select Flickr30K training data to generate the F30K-MCD split. First, we remove all F30K test data that has unseen primitive concepts to the COCO training data. Next, we collect and count the predicates for each image among all the remaining data over the five associated captions. These predicates correspond to the ``compounds'' defined in ~\cite{keysers2020measuring, shaw2020compositional}, and the objective is to maximize the divergence between compound distribution of the evaluation data to the training data. As a result of this step, we end up with a data set formed with pairs of (image, predicates counts), which are then used for computing the overall compound divergence ($\texttt{CD}_{\textsc{all}}$) to the training dataset. Afterwards, we enumerate over each pair of data, and again compute the compound divergence to the training dataset but with this specific data is removed. We denote the change of compound divergence as $\Delta_i = \texttt{CD}_{i} - \texttt{CD}_{\textsc{all}}$, and use it as an additional score to associate every data. Finally, we sort all the data with regard to the difference of compound divergence $\Delta_i$, and use the top ranking one thousand examples as the maximum compound divergence (MCD) split. The process for generating the COCO-MCD split is symmetrical to the above process, except the data is collected from COCO val+test splits (as it is sufficiently large). Similarly, to generate different CDs for making Figure 4 of the main text, we can also make use of the above data sorted by $\Delta_i$. Concretely, we put a sliding window with 1,000 examples and enumerate over the sorted data to obtain a massive combination of data (we can take a stride to make this computation sparser.) For each window of data, we measure the compound divergence and only take the windows that are at the satisfaction to our criteria. In Figure 4, we keep the windows that has the closest CD values to desired X-axis values for plotting.
\section{Implementation Details of {\textsc{Composer}}\xspace and Baselines}
\label{supp:sec:implementation_details}
\paragraph{Visual feature pre-processing} We follow ViLBERT~\cite{Lu2019ViLBERTPT} that extracts the patch-based ResNet feature using the Bottom-Up Attention model. The image patch feature has a dimension of 2048. A 5-dimension position feature that describes the normalized up-top and bottom-down position is extracted alongside the image patch feature. Therefore, each image region is described by both the image patch feature and the position feature. We extracted features from up to 100 patches in one image.
\paragraph{Text pre-processiong} Following BERT~\cite{Devlin2019BERT}, we tokenize the text using the uncased WordPiece tokenizer. Specifically, we first lowercase the text and use the uncased tokenizer to extract tokens. The tokenizer has a vocabulary size of 30,522. The tokens are then transformed into word embeddings with 768 dimensions. Besides the word embedding, a 768-dimension position embedding is extracted. Both position embedding and word embedding are added together to represent the embedding of tokens.
\paragraph{Training details} We use Adam optimizer~\cite{kingma2014adam} to optimize the parameter of our model. All the models are trained with a mini-batch size of 64. We employ a warm-up training strategy as suggested by ViLBERT~\cite{Lu2019ViLBERTPT}. Specifically, the learning rate is linearly increasing from 0 to $4e-5$ in the first 2 epochs. Then the learning rate decays to $4e-6$ and $4e-7$ after 10 epochs and 15 epochs, respectively. The training stopped at 20 epochs.
\paragraph{Detials of baseline approaches.} The text encoder for both models contains 12 layers of transformers and is initialized from BERT pretrained model using the checkpoint provided by HuggingFace. For ViLBERT, we use the \texttt{[CLS]} embedding from the last layer as text representation $\vct{y}$. We use the average of contextualized text embedding from the last layer as $\vct{y}$ in the VSE model. The visual encoder of VSE contains an MLP model with the residual connection. It transforms the image patch feature into a joint image-text space. The output of the visual encoder is the mean of the transformed image patch features. Unlike VSE, ViLBERT contains 6 layers of transformers for the image encoder and 6 layers of the cross-modal transformer to model the text and image features jointly. We use the embedding of \texttt{[V-CLS]} token from the last layer of the image encoder as the image feature ${\vct{x}}$.
\paragraph{Details of {\textsc{Composer}}\xspace.} The composer contains four primary learning sub-modules: (1) the \texttt{CrossAtt} model in primitive encoding; (2) the Predicate Transformer (PT) model; (3) the modulator; (4) the Composition Transformer (CT). The details of this sub-modules are list as what follows:
\begin{itemize}[leftmargin=*,topsep=0pt,itemsep=0pt]
\item \textbf{Primitive encoding.} We implement the \texttt{CrossAtt} model as a one-layer multi-head cross-modal Transformer that contains 768 dimension with 12 attention heads. The query set $Q$ is the sub-word token embeddings of the primitive word, and the key and value set $K$ and $V$ are the union of sub-word token embeddings and the object-centric visual features (which is linearly transformed to have the same dimensionality). We use the average of the contextualized sub-word token embeddings as the final primitive encoding.
\item \textbf{Predicate Transformer (PT).} We use 6 layers text Transformers with 768 hidden dimension and 12 attention heads to instantiate the Predicate Transformer. This network is initialized with the first 6 layers of a pre-trained BERT model.
\item \textbf{Modulator.} We use FiLM~\cite{perez2018film} as the modulator. Specifically, it contains two MLP models with a hidden dimension size of 768 to generate the scale $\vct{a}$ and bias vectors $\vct{b}$, using the syntactic placeholders as input. The scale $\vct{a}$ and bias $\vct{b}$ are then used to transform the input concept embedding $\vct{c}$ as $\vct{a} \odot \vct{c} + \vct{b}$. Here $\odot$ represents the element-wise multiplication. This modulated concept embedding is then projected by another MLP with 768 hidden dimensions, and used for reassembling with the predicate sequence.
\item \textbf{Composition Transformer (CT).} We follow the architecture of ViLBERT~\cite{Lu2019ViLBERTPT} to design the Composition Transformer (shown in Figure~\ref{fig:ct_diagram}). Specifically, it has interleaved \texttt{SelfAtt} Transformer and \texttt{CrossAtt} Transformer in the network. For example, if we consider a three-layer Composition Transformer, we have a \texttt{SelfAtt} Transformer at the beginning for both modality, followed with a \texttt{CrossAtt} Transformer that interchanges the information between the modality, and then another \texttt{SelfAtt} Transformer that only operates on the text modality. The output embedding of this last text \texttt{SelfAtt} Transformer is then used for computing the visual-semantic alignment scores using the linear regressor $\vct{\theta}$. Thus, when we consider shallower or deeper network, we add or remove the two layers of interleaved \texttt{SelfAtt} and \texttt{CrossAtt} Transformers. The hidden dimension of \texttt{SelfAtt} Transformer is 768, and there is 12 attention heads.
The hidden dimension of \texttt{CrossAtt} Transformer is 1024, and there is 8 attention heads.
\end{itemize}
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{figs/ct_diagram}
\caption{Details of the Composition Transformer model.}
\label{fig:ct_diagram}
\end{figure}
\section{Additional Experiments on {\textsc{Composer}}\xspace}
\label{supp:sec:ablation_exps}
We report additional ablation studies that are omitted in the main paper due to space limitation. In this section, we study {{\textsc{Composer}}\xspace} performance under different MVSA objectives, Negative Log-Likelihood and Hinge loss. Then we study {{\textsc{Composer}}\xspace} performance under different margins of MVSA and Order objectives.
\input{tables/supp_ablation_order}
\paragraph{MVSA Objective.} The MVSA objectives can be implemented using NLL loss or Hinge loss. We study the performance of {{\textsc{Composer}}\xspace} under different losses for MVSA in Table~\ref{tab:supp_nll_hinge}. The models are trained with both MVSA and order objectives. We set the margin of order objectives $\beta=0.2$. For the hinge loss, we set the margin $\alpha=0.8$. {{\textsc{Composer}}\xspace} trained with hinge loss in MVSA achieves better performance than the NLL loss in all metrics across both in-domain and cross-dataset generalization settings. Therefore, for all the experiments training with MVSA, we use hinge loss instead.
\input{tables/supp_ablation_nll_vs_hinge}
\paragraph{Ablation study on $\alpha$ and $\beta$.} We study {{\textsc{Composer}}\xspace} performance on the different margin of MVSA and Order objectives. First, we fix the margin of order objectives $\beta$ and tune the margin for MVSA $\alpha$. {{\textsc{Composer}}\xspace} with a larger margin for MVSA achieves better R1 in-domain performance. Alternatively, by fixing the $\alpha$ and tuning $\beta$, {{\textsc{Composer}}\xspace} achieves the best R1 in-domain performance and best R5 in cross-dataset generalization setting with $\beta=0.2$.
|
\section{Introduction}
Distributed systems have become an ubiquitous part of modern computing, with systems continuing to grow in size and scope. As these systems grow larger, the need for topologies that allow for efficient operations like search and routing increases. To this end, many systems use \emph{overlay networks} to control the network topology. In overlay networks, connections are made using logical links, each of which consists of zero or more physical links. This use of logical links means program actions can add and delete edges in the network, allowing the system to maintain an arbitrary logical topology even when the physical topology may be fixed.
Many of these large systems operate in environments where faults are commonplace. Servers may crash, communication links may be damaged, and processes may join or leave the system frequently. This reality has increased the demand for fault tolerant overlay networks. One particularly strong type of fault tolerance is \emph{self-stabilization}, where a legal configuration is guaranteed to be reached after any transient fault. For overlay networks, this means a correct topology can always be built when starting from any configuration provided the network is not disconnected.
\subsection{Problem Overview}
Our current work focuses on self-stabilizing overlay networks. A self-stabilizing overlay network guarantees that program actions will build a legal topology even when the system starts in \emph{any} weakly-connected topology. Our interest, then, is in the design and analysis of algorithms that, when executed on an arbitrary initial weakly-connected topology, add and delete edges with program actions until a legal \emph{target topology} is reached.
Going further, we are interested in algorithms and analyses for general frameworks for overlay network creation. To date, most work has focused on algorithms for a single topology, or has been inefficient in terms of time or space complexity. Work focused on a specific topology is hard to generalize and derive insights from for expanding to other overlay network applications, while general frameworks with high time and space complexity may be too inefficient to be useful in practice. Our interest is in general frameworks for overlay network creation that allow efficient algorithms to be built while still being general enough to provide insights into the stabilization of arbitrary topologies.
\subsection{Main Results and Significance}
In this paper, we build upon the work of Berns~\cite{berns_avatar_15} to present a general algorithm for creating self-stabilizing overlay network protocols for a variety of target topologies, and provide several examples of the application of our general algorithm, including with a new overlay network topology. More specifically, our contributions are as follows:
\begin{itemize}
\item We update the analysis of Berns~\cite{berns_avatar_15} to show how their algorithm can be extended into an algorithm for \emph{any} target topology. Our updated analysis is the first to show a general framework for self-stabilizing overlay network creation that allows for efficient stabilization in both time and space.
\item As part of this updated analysis, we introduce several measures of complexity that are properties of the target topology itself. These measures are useful for two reasons. First, they allow us to analyze the general algorithm easily for a variety of topologies. Perhaps just as important, however, is that they provide valuable insight into how the selection of the target topology affects convergence in terms of both time and space. This insight can be quite useful for not only selecting a target topology, but also in designing new overlay network topologies.
\item We analyze several existing overlay network topologies, providing the necessary metrics for analysis in our general framework. This analysis provides an example of how our framework can be applied, and also helps provide concrete insight into factors affecting convergence and demonstrate how diameter, degree, and robustness are balanced when designing self-stabilizing overlay networks.
\item Using the insight gained from our earlier contribution, we define a new network topology which stabilizes with sublinear time and space complexity in our framework. This demonstrates how the framework can provide network designers with guidance to help them build new topologies that can stabilize efficiently with our approach. The design and analysis of new topologies targeted for efficient stabilization in this framework could be an interesting area of future study.
\end{itemize}
The key idea of our work is the extension of the algorithm of Berns~\cite{berns_avatar_15} to work with other topologies by defining the target topology and analyzing several relevant measures regarding this topology. This definition allows the creation and analysis of many self-stabilizing overlay networks without having to design the algorithm from scratch. Furthermore, our framework highlights the factors of the network that affect stabilization, allowing a designer to tune the topology to meet their needs.
\subsection{Related Work and Comparison}
The past few decades have seen a large body of work develop on overlay networks. Early work focused on defining \emph{structured} overlay networks, where a single correct configuration existed for a particular set of nodes. Examples of these networks include \textsc{Chord}~\cite{stoica_chord_01} and \textsc{Tapestry}~\cite{zhao_tapestry_2004}. These works often did not consider fault tolerance, or considered a weak model with limited possible failures.
As work expanded in overlay networks, so did work in various types of fault tolerance. One category of work considered \emph{self-healing} networks, where a particular network property could be maintained even during limited node deletions~\cite{hayes_forgiving-tree_2008,hayes_forgiving-graph_2009}. Several examples of this work even used virtual nodes~\cite{trehan_virtual_12}, although they were not used to create a specific embedding as done in this current work. Recently in \textsc{DConstructor}~\cite{gilbert_dconstructor_20}, the authors present a framework for building overlay networks. \textsc{DConstructor} works by forming clusters and merging these clusters together. However, as with the other examples, \textsc{DConstructor} is not self-stabilizing as it assumes all nodes begin in a single node cluster. Said in another way, \textsc{DConstructor} assumes an arbitrary initial \emph{topology}, but not an arbitrary initial state. This is also the same assumption in the work of G\"{o}tte et al.~\cite{gotte_time-optimal_podc21}, who presented an algorithm for transforming a constant-degree network into a tree in $\mathcal{O}(\log n)$ rounds.
Our work considers \emph{self-stabilizing overlay networks}, where the correct configuration is reached after an arbitrary number of transient faults that do not disconnect the network. There are several examples of these as well. The \textsc{Skip+} graph~\cite{jacob_skipplus_09} presents a self-stabilizing variant of the \textsc{Skip} graph~\cite{aspnes_skipgraph_03} with polylogarithmic convergence time, although the space requirements are linear for some configurations. Other examples of self-stabilizing overlay networks include \textsc{Re-Chord}~\cite{kniesburges_rechord_11}, a \textsc{Chord} variant with virtual nodes that requires $\mathcal{O}(n \log n)$ rounds to converge, and \textsc{MultiSkipGraph}, a skip graph variant which maintained a property called \emph{monotonic searchability} during convergence~\cite{luo_multiskipgraph_2019}.
To date, most work has been focused on the convergence to a particular topology, with algorithms and analysis all targeted at these specific instances. One exception to this is the \emph{Transitive Closure Framework}~\cite{berns_tcf_11}, which presents a general algorithm for creating any locally-checkable overlay network. Like our work, they also identify a general measure of interest for stabilization time which they call the \emph{detector diameter}. While this measure helps to bound the \emph{time} complexity of their algorithm, the space requirements are $\Theta(n)$ for any topology, limiting the applicability.
In \textsc{Avatar}~\cite{berns_avatar_15}, Berns presented both a locally-checkable definition of a network embedding for arbitrary topologies, and a self-stabilizing algorithm for building an embedded binary search tree with polylogarithmic time and space requirements. The work only considered a single topology, however, and did not offer insight into measures for arbitrary topologies. Our goal with this work is to build upon \textsc{Avatar} to address these issues.
\section{Preliminaries}
\label{section:prelims}
\subsection{Model of Computation}
We model our distributed system as an undirected graph $G = (V, E)$, with $n$ processes in $V$ communicating over the edges $E$. Each node $u \in V$ has a unique identifier $u.id \in \mathbb{N}$, which is stored as immutable data in $u$. Where clear from the context, we will use $u$ to represent the identifier of $u$.
Each node $u \in V$ has a \emph{local state} consisting of a set of variables and their values, along with its immutable identifier $u.id$. A node executes a \emph{program} whose actions modify the values of the variables in its local state. All nodes execute the same program. Nodes can also communicate with their neighbors. We use the \emph{synchronous message passing} model of computation~\cite{sync_mp_1998}, where computation proceeds in synchronous rounds. During each round, a node receives messages sent to it in the previous round from any node in its neighborhood $N(u) = \{v \in V:(u,v) \in E\}$, executes program actions to update its local state, and sends messages to any of its neighbors. We assume reliable communication channels with bounded delay, meaning a message is received by node $u$ in some round $i$ if and only if it was sent to $u$ in round $i - 1$.
In the overlay network model, nodes communicate over logical links that are part of a node's state, meaning a node may execute actions to create or delete edges in $G$. In any round, a node may delete any edge incident upon it, as well as create any edge to a node $v$ which has been ``introduced'' to it from some neighbor $w$, such that $(u,w)$ and $(w, v)$ are both in $E$. Said in another way, in a particular round a node may connect its neighbors to one another by direct logical links.
The goal for our computation is for nodes to execute actions to update their state (including modifying the topology by adding and deleting edges to other nodes) until a legal configuration is reached. A \emph{legal configuration} can be represented as a predicate over the state of the nodes in the system. In the overlay network model, links are part of a node's state, meaning a legal configuration is defined at least in part by the overlay network topology. The \emph{self-stabilizing overlay network problem} is to design an algorithm $\mathcal{A}$ such that when executing $\mathcal{A}$ on each node in a connected network with nodes in an arbitrary state, and allowing $\mathcal{A}$ to add and delete edges, eventually a legal configuration, including a predicate defined at least in part by the network's logical topology, is reached. This means that a self-stabilizing overlay network will always automatically restore a legal configuration (including reconfiguring the network topology) after \emph{any} transient failure so long as the network remains connected.
\subsection{Complexity Measures}
When designing self-stabilizing overlay network protocols, there are two measures of interest: the time required to build a correct configuration, and the space required to do so (in terms of a node's degree). In our model, we are concerned with the number of synchronous rounds that are required to reach a legal state. In particular, the maximum number of synchronous rounds required to build a legal topology when starting from any arbitrary configuration is called the \emph{convergence time}.
When measuring the space requirements, we use the \emph{degree expansion} measure from the original \textsc{Avatar} work~\cite{berns_avatar_15}, which is defined as the ratio of the maximum node degree of any node during convergence over the maximum node degree from the initial or final configuration. This measure is based upon the idea that if a node begins with a large degree in the initial configuration, or ends with a large degree in the final configuration, the overall algorithm cannot be expected to have a low degree during convergence. Instead, we are interested in the ``extra'' degree growth caused by the algorithm during convergence.
\section{Generalizing \textsc{Avatar}}
The original \textsc{Avatar} work~\cite{berns_avatar_15} provided two things: a definition of a locally-checkable embedding from any set of real nodes to a particular target topology, and a self-stabilizing algorithm for creating a specific binary tree topology. Below, we review these contributions and expand the analysis of the algorithm to show it can work for arbitrary topologies.
\subsection{\textsc{Avatar} Definition}
The \textsc{Avatar} network definition is simply a dilation-1 embedding between a \emph{guest network} and a \emph{host network}. More specifically, let $\mathcal{F}$ be a family of graphs such that, for each $N \in \mathbb{N}$, there is exactly one graph $F_N \in \mathcal{F}$ with node set $\{0, 1, \ldots, N-1\}$. We call $\mathcal{F}$ a \emph{full graph family}, capturing the notion that the family contains exactly one topology for each ``full'' set of nodes $\{0, 1, \ldots, N-1\}$ (relative to the identifiers). For any $N \in \mathbb{N}$ and $V \subseteq \{0, 1, \ldots, N-1\}$, $\textsc{Avatar}_{\mathcal{F}}(N,V)$ is a network with node set $V$ that realizes a dilation-1 embedding of $F_N \in \mathcal{F}$. The specific embedding is given below. We also show that, when given $N$, \textsc{Avatar} is locally checkable ($N$ can be viewed as an upper bound on the number of nodes in the system). It is on this full graph family \emph{guest network} that our algorithms shall execute, as we will show later.
\begin{definition}
Let $V \subseteq [N]$ be a node set $\{u_0, u_1, \ldots, u_{n-1}\}$, where $u_i < u_{i+1}$ for $0 \leq i < n-1$. Let the \emph{range} of a node $u_i$ be $\mathit{range}(u_i) = [u_i, u_{i+1})$ for $0 < i < n-1$. Let $\mathit{range}(u_0) = [0, u_1)$ and $\mathit{range}(u_{n-1}) = [u_{n-1}, N)$. $\textsc{Avatar}_{\mathcal{F}}(N, V)$ is a graph with node set $V$ and edge set consisting of two edge types:
\begin{description}
\item[Type 1:] $\{(u_i, u_{i+1}) | i=0,\ldots,n-2\}$
\item[Type 2:] $\{(u_i, u_j) | u_i \neq u_j \wedge \exists (a,b) \in E(F_N), a \in \mathit{range}(u_i) \wedge b \in \mathit{range}(u_j)\}$
\end{description}
\label{defn:avatar}
\end{definition}
When referring to a general \textsc{Avatar} network for any set of nodes, we will omit the $V$ and simply refer to $\textsc{Avatar}_{\mathcal{F}}(N)$.
As with the original work, to reason about \textsc{Avatar}, we consider two ``networks'': a \emph{host network} consisting of the \emph{real} nodes in $V$, and a \emph{guest network} consisting of the $N$ \emph{virtual} nodes from the target topology. Each real node in $V$ is the host of one or more virtual nodes in $N$. This embedding provides several advantages. First, it allows us to make many networks locally-checkable provided all nodes know $N$ in advance. Second, it provides a simple mechanism for which to reason about network behavior in the guest network. As the target $N$-node topology is fixed regardless of the actual set of real nodes $V$, the design and analysis of our algorithms is simplified by executing them on the guest network. Furthermore, since we are using a dilation-1 embedding, most metrics for performance regarding the guest network (e.g. diameter) still apply to the host network.
As an additional note, we shall assume that the guest nodes also use the synchronous message passing model described for the real nodes in the model section.
\subsubsection{A Note on $n$ versus $N$}
The work of \textsc{Avatar} does require all nodes know $N$, an upper bound on the number of nodes in the system. In cases where network membership is predictable, it may be possible for $n$ to be within a constant factor of $N$. From a practical standpoint, even in cases where $N$ and $n$ are significantly different, a polylogarithmic convergence time in $N$ may still be small enough (e.g. if we think of IPv6, $\mathcal{O}(\log N)$ is only 128).
\subsection{The \textsc{Avatar} Algorithm}
The original \textsc{Avatar} work was focused on the creation of a specific topology (a binary search tree) as the \emph{target topology}. Their algorithm followed a divide-and-conquer approach, separating nodes into clusters and then merging them together. One can think of their self-stabilizing algorithm as involving three different components:
\begin{enumerate}
\item \emph{Clustering}: The first step in the algorithm is for nodes to form clusters. These clusters begin as a single host node hosting a full $N$ node guest network of the target topology (\textsc{Cbt} in the original work). In the initial configuration, nodes may not be a part of a cluster, but since \textsc{Avatar} is locally checkable, all faulty configurations contain at least one node which detects the faulty configuration and will begin forming the single-node clusters. This fault detection and cluster creation will propagate through the network until eventually all nodes are members of $N$ node clusters of the target topology.
\item \emph{Matching}: The second step of the algorithm is to match together clusters so that they may merge together. To do this, the root node of a spanning tree defined on the cluster repeatedly polls the nodes of its cluster, asking them to either find neighboring clusters that are looking for merge partners (called the leader role), or to look at neighboring clusters that can assign them a merge partner (called the follower role). The role of leader or follower is randomly selected. Leader clusters will match together all of their followers for merging by adding edges between the roots of each cluster, creating a matching between clusters that may not be direct neighbors. This ability to create edges to match non-neighboring clusters allows more matches to occur, and thus more merges, and thus a faster convergence time.
\item \emph{Merging}: The algorithm then deals with the merging of matched clusters. To prevent degrees from growing too large, a cluster is only allowed to merge with at most one other cluster at a time. Once two clusters have matched from the previous step, the roots of the clusters connect as ``partners'' and update their successor pointers based upon the identifier of the host of the root of the other cluster. One node will have its responsible range become smaller, and this node will send all guest nodes that were in its old responsible range to its partner in the other cluster. The children of the root nodes are connected, and then they repeat the process of updating successor pointers and passing along guest nodes outside their new responsible range. Eventually this process reaches the leaves, at which point all nodes in both clusters have updated their responsible ranges and now form a new legal cluster of the target topology.
\end{enumerate}
As it turns out, the algorithm components from the original \textsc{Avatar} work do not depend upon the specific topology that is being built (the \emph{target topology}). While the analysis of complexity assumes a complete binary search tree, the algorithm components themselves simply rely upon an arbitrary target topology and a spanning tree defined upon that topology on which to execute PIF waves. We can therefore extend this algorithm to other topologies if we update the analysis and include several additional metrics. We define these metrics next after discussing the algorithm's intuition.
\subsection{Relevant Metrics}
To update the analysis of the original \textsc{Avatar} work for any target topology requires two measures of the target topology. The first of these is diameter of the target topology, which will be a factor in determining the convergence time. The second of these is a measure of a real node's degree inside the embedded target topology, which will be a factor in determining the degree expansion.
\subsubsection{Spanning Tree Diameter}
In the original \textsc{Avatar} algorithm, a spanning tree embedded onto the target topology was used to communicate and coordinate between nodes in a particular cluster. For our work, we will simply use a spanning tree with a root of (virtual) node 0 and consisting of the shortest path from node 0 to all other nodes. Obviously the diameter of this spanning tree is at most the diameter of the target topology, and we therefore shall use the diameter of the target topology as our first metric of interest. We denote the diameter of a particular target topology $T$ with $N$ nodes as $D(T_N)$.
As we shall see, this diameter measure will be key in determining the stabilization time. Intuitively, a low-diameter spanning tree results in faster communication within clusters, and therefore faster convergence than a higher diameter spanning tree.
\subsubsection{Maximum Degree of Embedding}
The other measure of interest has to do with the degree of the \emph{real} nodes when embedding the target topology. More formally, let the \emph{maximum degree of embedding $T_N$ in \textsc{Avatar}} be defined as the maximum degree of any node in $\textsc{Avatar}_T(N,V)$ for any node set $V \subseteq N$. Where clear from context, we will refer to this simply as the \emph{maximum degree of embedding} and denote it as $\Delta_\textsc{A}(T_N)$. Note the maximum degree of embedding is almost entirely determined by the target topology $T$, as there are only 2 edges in $\textsc{Avatar}_T(N)$ per node that are not present to realize a dilation-1 embedding of $T$.
The maximum degree of embedding is a critical measure for degree expansion as it determines how many additional edges a node may receive during the various stages of the algorithm. The clusters in the \textsc{Avatar} algorithm are $N$ node instances of the target topology $T$, and using metrics defined on $T_N$ is acceptable for running time. However, each time an edge is added to a virtual node within a cluster, we must consider the effects on the degree in the host network, not just the guest network.
Note that the definition of maximum degree of embedding considers \emph{any} possible subset of real nodes $V$. This differs from typical (non-stabilizing) overlay network results, where it is common to assume that identifiers are uniformly distributed, meaning that the ranges of each real node are of similar size. However, since we are building a self-stabilizing protocol, and each cluster is by itself an $N$ node instance of the target topology, the ranges of hosts inside clusters during convergence may be quite skewed, even when the final distribution of node identifiers is not.
\subsection{Overall Complexity}
If we are given the diameter and the maximum degree of embedding of an arbitrary target topology, we can then determine the convergence time and degree expansion of Berns' algorithm for the arbitrary target topology. We give the theorems for these measures below and provide brief proof sketches of each. As the updated analysis is heavily based on the original work, and the algorithms are from the original work, the full analysis is left for the appendix. Our contribution is not the original algorithm of Berns, but rather the observation that the algorithm works for any topology, the updated metrics for the analysis, and the examples and discussion that follow.
\begin{theorem}
\label{thm:avatar_convergence}
The algorithm of Berns~\cite{berns_avatar_15} defines a self-stabilizing overlay network for $\textsc{Avatar}_T$, for some full graph family target topology $T$, with convergence time of $\mathcal{O}(D(T_N) \cdot \log N)$ in expectation, where $D(T_N)$ is the diameter of the $N$ node topology $T_N$.
\end{theorem}
\begin{proofsketch}
A sketch of the steps for proving this theorem are as follows:
\begin{itemize}
\item In at most $\mathcal{O}(D(T_N))$ rounds, every node is a member of a cluster.
\item For any cluster, in an expected $\mathcal{O}(D(T_N))$ rounds, the cluster has completed a merge with another cluster, meaning the number of clusters has decreased by a constant fraction in $\mathcal{O}(D(T_N))$ rounds.
\item Reducing the number of clusters by a constant fraction needs to be done $\mathcal{O}(\log N)$ times before a single cluster remains.
\end{itemize}
\end{proofsketch}
\begin{theorem}
\label{thm:avatar_degree-expansion}
The algorithm of Berns~\cite{berns_avatar_15} defines a self-stabilizing overlay network for $\textsc{Avatar}_T$, for some full graph family target topology $T$, with degree expansion of $\mathcal{O}(\Delta_\textsc{A}(T_N) \cdot \log N)$ in expectation, where $\Delta_\textsc{A}(T_N)$ is the maximum degree of embedding of the $N$ node target topology $T_N$.
\end{theorem}
\begin{proofsketch}
To prove this, we consider the actions that might increase a node's degree.
\begin{itemize}
\item Regardless of the number of merges a node participates in, the node's degree will grow to at most $\mathcal{O}(\Delta_\textsc{A}(T_N))$ as the result of merge actions. By definition, a node's degree within its cluster after a merge cannot exceed $\Delta_\textsc{A}(T_N)$.
\item During the process of matching clusters together, a node's degree may grow by one for every child it has in the spanning tree. Since the node has at most $\Delta_\textsc{A}(T_N)$ children from other nodes, each time the node participates in the matching its degree grows by $\mathcal{O}(\Delta_\textsc{A}(T_N))$. In expectation, this matching happens $\mathcal{O}(\log N)$ times (see proof of Theorem \ref{thm:avatar_convergence}).
\end{itemize}
\end{proofsketch}
The implications of Theorems~\ref{thm:avatar_convergence} and \ref{thm:avatar_degree-expansion} are that we can simply define a target topology and analyze its diameter and maximum degree of embedding to have a self-stabilizing protocol for our target topology. We provide a few examples of this process in the following section.
\section{Examples}
In this section, we demonstrate how the selection of the target topology affects the complexity of our algorithm by considering several different topologies: the \textsc{Linear} network, a complete binary search tree (\textsc{Cbt}, taken from~\cite{berns_avatar_15}), and \textsc{Chord}~\cite{stoica_chord_01}.
\subsection{Linear}
As the name suggests, the \textsc{Linear} network consists of a line of nodes sorted by identifier. The formal desired end topology for the \textsc{Linear} network is given next.
\begin{definition}
The $\textsc{Linear}(N)$ network, for $N \in \mathbb{N}$, consists of nodes $V = \{0, 1, \ldots, N-1\}$ and edges $E = \{(i, i+1), i \in [0,N-2]\}$.
\end{definition}
\begin{lemma}
The diameter of an $N$ node \textsc{Linear} network is $\mathcal{O}(N)$.
\end{lemma}
Each virtual node has a degree of at most 2, and it is easy to see that each real node also has a degree of at most 2. Therefore, the maximum degree of embedding for \textsc{Linear} is $\mathcal{O}(1)$.
\begin{lemma}
The maximum degree of embedding of the \textsc{Linear} topology is $\mathcal{O}(1)$.
\end{lemma}
\begin{proof}
Note that for any particular $range(u) = [x, y]$, there are at most two external edges: $(x-1, x)$ and $(y, y+1)$. All other edges are between virtual nodes inside the range.
\end{proof}
The above lemmas combined with Theorems \ref{thm:avatar_convergence} and \ref{thm:avatar_degree-expansion} give us the following corollary.
\begin{corollary}
The self-stabilizing \textsc{Avatar} algorithm builds $\textsc{Avatar}_\textsc{Linear}(N)$ in an expected $\mathcal{O}(N \cdot \log N)$ rounds with an expected degree expansion of $\mathcal{O}(\log N)$.
\end{corollary}
Note the convergence time of this algorithm is a logarithmic factor slower than previous results~\cite{onus_linear_07}. This logarithmic factor comes from the ``cost of coordination'', as edges are only added when clusters have matched.
\subsection{Complete Binary Search Tree}
In the first \textsc{Avatar} paper, the author defined and analyzed an algorithm for one specific topology, the complete binary search tree (called \textsc{Cbt}). We formally define the desired end topology for \textsc{Cbt}, list the relevant measures for this topology below and omit the proofs, as those are contained in the work of Berns~\cite{berns_avatar_15}.
\begin{definition}
For $a \leq b$, let $\textsc{Cbt}[a,b]$ be a binary tree rooted at $\mathit{r} = \lfloor (b+a)/2 \rfloor$. Node $r$'s left cluster is $\textsc{Cbt}[a,r-1]$, and $r$'s right cluster is $\textsc{Cbt}[r+1,b]$. If $a > b$, then $\textsc{Cbt}[a,b] = \bot$. We define $\textsc{Cbt}(N) = \textsc{Cbt}[0,N-1]$.
\end{definition}
\begin{lemma}
The diameter of an $N$ node \textsc{Cbt} network is $\mathcal{O}(\log N)$.
\end{lemma}
\begin{lemma}
The maximum degree of embedding of an $N$ node \textsc{Cbt} network is $\mathcal{O}(\log N)$.
\end{lemma}
The above lemmas combined with Theorems \ref{thm:avatar_convergence} and \ref{thm:avatar_degree-expansion} give us the following corollary.
\begin{corollary}
The self-stabilizing \textsc{Avatar} algorithm builds a target topology of $\textsc{Avatar}_\textsc{Cbt}(N)$ in expected $\mathcal{O}(\log^2 N)$ rounds with $\mathcal{O}(\log^2 N)$ expected degree expansion.
\end{corollary}
Note the above corollary matches with the detailed proofs given in the original \textsc{Avatar} work. Unlike the original, however, we reached our conclusions based simply upon the metrics we defined earlier. This corollary, then, serves as a nice ``sanity check'' on the accuracy of our results.
\subsection{Chord}
Both \textsc{Linear} and \textsc{Cbt} are tree topologies, meaning they are fragile in the sense that a single node or link failure may partition the network. In this section, we consider a more robust topology. In particular, we apply the \textsc{Avatar} algorithm to an $N$-node \textsc{Chord} network~\cite{stoica_chord_01} as well. We define the network's desired topology as follows.
\begin{definition}
For any $N \in \mathbb{N}$, let $\textsc{Chord}(N)$ be a graph with nodes $[N]$ and edge set defined as follows. For every node $i$, $0 \leq i < N$, add to the edge set $(i, j)$, where $j = (i + 2^k) \mod N$, $0 \leq k < \log N - 1$. When $j = (i + 2^k)\mod N$, we say that $j$ is the $k$-th finger of $i$.
\end{definition}
The original \textsc{Chord} paper proves the following lemma regarding the network's diameter.
\begin{lemma}
The diameter of an $N$ node \textsc{Chord} network is $\mathcal{O}(\log N)$.
\end{lemma}
While the logarithmic diameter means we can efficiently build \textsc{Chord} in terms of time complexity, the results are not so hopeful in terms of degree complexity. In an $N$-node \textsc{Chord} network, every node has $\mathcal{O}(\log N)$ neighbors, some of which have identifiers up to $N/2$ away. The result of this is that if a real node has a range of size $N/2$, each virtual node in the range may potentially have a connection to a virtual node on a different host. This means that a real node in an embedded \textsc{Chord} network may have a maximum degree of embedding of $\mathcal{O}(N)$, as we show next.
\begin{lemma}
The maximum degree of embedding of the \textsc{Chord} topology is $\mathcal{O}(N)$.
\end{lemma}
\begin{proof}
To see this, consider a specific $n$ node embedding of \textsc{Chord} where $range(u) = [0, N/2)$. Note each node in $u$'s range is incident on at least one edge whose other endpoint is outside $range(u)$ -- specifically, the $\log N - 1$ \textsc{Chord} finger. As there are $N/2$ nodes in $u$'s range, we have at least $N/2$ edges with exactly one endpoint outside $range(u)$, and our lemma holds.
\end{proof}
Note this result is not a concern in the \emph{final} configuration if we assume identifiers are uniformly distributed (which is a reasonable and common assumption with overlay networks). However, we are working in a self-stabilizing setting where \emph{any} initial configuration is possible, and therefore one could imagine a scenario where the system reaches a configuration where a single cluster consisting of node $0$ is matched with an $N/2$ node cluster consisting of all real nodes from the range $[N/2, N)$.
Combining the above two lemmas with Theorem \ref{thm:avatar_convergence} and Theorem \ref{thm:avatar_degree-expansion} gives us the following corollary.
\begin{corollary}
\label{corollary:chord}
The self-stabilizing \textsc{Avatar} algorithm builds $\textsc{Avatar}_{\textsc{Chord}}(N)$ in expected $\mathcal{O}(\log^2 N)$ rounds and with expected degree expansion of $\mathcal{O}(N \cdot \log N)$.
\end{corollary}
Note we can actually improve the bound of the degree expansion, as it is at most $\mathcal{O}(N)$. As our theorems provide an upper bound, however, we leave them stated as is for simplicity.
\section{\textsc{SkipChord}}
\label{section:skipchord}
We showed above the role the diameter and maximum degree of embedding play in determining performance of the \textsc{Avatar} algorithm. One of the benefits of the general analysis of the \textsc{Avatar} algorithm is that it highlights the factors of the \emph{topology} that will affect convergence, allowing a network designer to select a topology based upon the problem requirements while weighing the impact of the topology on convergence time and degree expansion. One can even design \emph{new} topologies with an eye towards these metrics for use in \textsc{Avatar} embeddings.
In this section, we present a new network topology built specifically for embedding in the \textsc{Avatar} framework which we call \textsc{SkipChord}. This ring-based network is more robust than \textsc{Linear} and \textsc{Cbt} while avoiding the high degree requirements during stabilization of the standard \textsc{Chord} network. As we shall show, by balancing degree and robustness, we may achieve efficient stabilization while still having sublinear degree expansion.
\subsection{Definition}
As trees, the \textsc{Linear} and \textsc{Cbt} networks make poor choices for many fault-prone applications as they are easily disconnected by node or link failure. \textsc{Chord} represents a more robust choice, but suffers from a high maximum degree of embedding due to each of the $N$ nodes having a long link (to a neighbor with identifier $\mathcal{O}(N)$ away from itself). Our \textsc{SkipChord} network tries to balance this by limiting the number of fingers while still maintaining a topology more robust than a simple tree. We give the formal definition of \textsc{SkipChord} below.
\begin{definition}
For any $N \in \mathbb{N}$, let $\textsc{SkipChord}(N,s)$ be a graph with nodes $[N]$ and \emph{skip factor} $s$. The edge set for $\textsc{SkipChord}(N,s)$ is defined as follows:
\begin{itemize}
\item \emph{Ring edges}: For every node $i$, $0 \leq i < N$, add edges $(i-1 \mod N, i)$ and $(i, i+1 \mod N)$
\item \emph{Finger edges}: For every node $j$, where $j = sk$, for $k=0,1,\ldots,(N-s)/s$, add edge $(j, j + 2^{k \mod \log N} \mod N)$. We say the \emph{size} of this finger edge is $2^{k \mod \log N}$.
\end{itemize}
\end{definition}
This construction basically takes the $\log N$ fingers from each node in the original \textsc{Chord} and distributes them out over a range of nodes as determined by the \emph{skip factor} $s$. As we shall show, by ``skipping'' the fingers, fewer virtual nodes in a real node's range have ``long'' outgoing links, and therefore the number of edges to other real nodes is limited while not compromising efficient routing.
\begin{figure}
\centering
\subfigure[Six Fingers of Node $0$ in $\textsc{Chord}(64)$]{
\includegraphics[scale=0.25]{chord_neighborhood.png}
}\hfill
\subfigure[First Six ``Skipped'' Fingers for $\textsc{SkipChord}(64,2)$]{
\includegraphics[scale=0.25]{skipchord_neighborhood.png}
}
\caption{(a) The neighborhood of node $0$ in \textsc{Chord}, and (b) the corresponding six fingers in \textsc{SkipChord}. Note how the fingers in \textsc{SkipChord} are no longer all incident on node $0$ but instead have ``skipped'' ahead.}
\label{fig:skipchord}
\end{figure}
To better understand \textsc{SkipChord}, consider Figure~\ref{fig:skipchord}. The network on the top shows node $0$'s neighborhood for $\textsc{Chord}(64)$, while the network on the bottom shows the corresponding fingers ``skipped'' with a skip factor of 2 (i.e. a subset of the edges for $\textsc{SkipChord}(64,2)$). The first six fingers in \textsc{Chord} are all incident upon node $0$, while the first six fingers in \textsc{SkipChord} are distributed amongst nodes 0, 2, 4, 6, 8, and 10. As a result of this, each node in \textsc{SkipChord} has a much smaller degree than in \textsc{Chord}. This low degree result holds true for the degree of the embedding as well, as we show in the following section.
\subsection{Metrics}
We begin with a proof of the diameter of \textsc{SkipChord}. The intuition behind our result is simple: any node is at most $s \cdot \log N$ away from an edge that at least halves the distance from itself to any other node.
\begin{lemma}
The diameter of an $N$-node \textsc{SkipChord} network with skip factor $s$ (\textsc{SkipChord}$(N,s)$) is $\mathcal{O}(s \cdot \log^2 N)$.
\end{lemma}
\begin{proof}
Consider the hops required to halve the distance between two arbitrary nodes $u$ and $v$. In at most $\mathcal{O}(s \cdot \log N)$ hops from $u$ using ring edges, at least one finger of every size is reachable. One of these fingers will at least halve the distance to the other node $v$. As the distance to $v$ can be halved in $\mathcal{O}(s \cdot \log N)$ hops, and this halving will occur $\log N$ times before reaching $v$, our lemma holds.
\end{proof}
Next we consider the maximum degree of embedding. Here we see how spreading the fingers out over a set of nodes has resulted in a lower maximum degree of embedding.
\begin{lemma}
The maximum degree of embedding of an $N$-node \textsc{SkipChord} with a skip factor of $s$ (\textsc{SkipChord}$(N,s)$) is $\mathcal{O}(\frac{N}{s \cdot \log N})$.
\end{lemma}
\begin{proof}
Let an edge be called an \emph{external edge} for node $u$ if and only if it has exactly one endpoint in $range(u)$. Note that there are at most 2 external ring edges for any possible range for node $u$. We consider then the external finger edges in $range(u)$. Let $k$ be the largest finger that has exactly one endpoint in $range(u)$. There are at most $|range(u)| / s \cdot \log N$ such fingers, where $|range(u)|$ denotes the size of the range (for our embeddings, the number of nodes $u$ is hosting). For the $k-1$ fingers, there are at most half as many as the $k$ fingers with exactly one endpoint in $range(u)$. Similarly, for the $k-2$ fingers, there are at most $1/4$ as many as the $k$ fingers with exactly one endpoint in $range(u)$, and so on. Summing these together, we get $(|range(u)| / (s \cdot \log N)) + 1/2(|range(u)| / (s \cdot \log N)) + 1/4(|range(u)| / (s \cdot \log N)) + \ldots = 2(|range(u)| / (s \cdot \log N))$. Since $|range(u)|$ is at most $N$, our lemma holds.
\end{proof}
The above lemmas, combined with Theorem~\ref{thm:avatar_convergence} and Theorem~\ref{thm:avatar_degree-expansion} give us the following corollary.
\begin{corollary}
\label{corollary:skipchord}
The \textsc{Avatar} algorithm builds the $\textsc{SkipChord}(N,s)$ target topology in an expected $\mathcal{O}(s \cdot \log^3 N)$ rounds with an expected degree expansion of $\mathcal{O}(N/s)$.
\end{corollary}
Note that we can select a skip factor in such a way as to have efficient time and space complexity. For instance, if we select a skip factor of $\log N$, we have polylogarithmic convergence time, sublinear degree expansion, and a target topology that is more robust than the tree topologies of \textsc{Linear} and \textsc{Cbt}.
Given the fact that \textsc{SkipChord} can be built efficiently, there are multiple suitable applications for it. Like \textsc{Chord}, \textsc{SkipChord} can be used as a distributed hash table for storing and retrieving files, particularly in settings where transient failures may cause node and link failures, such as with unreliable Internet connections or in scenarios where a large number of nodes may be deployed and start up at the same time with the goal of forming a distributed hash table. If node identifiers are uniformly distributed, each real node would host a similar number of files as with the original \textsc{Chord} while having the added benefit of efficient stabilization after any transient faults.
\section{Discussion and Future Work}
Table~\ref{table:measures} summarizes the various measures of interest for the different target topologies.
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
\textbf{Topology} & \textbf{Diameter} & \textbf{Maximum Degree of Embedding} \\
\hline
\textsc{Linear} & $\mathcal{O}(N)$ & $\mathcal{O}(1)$ \\
\hline
\textsc{Cbt} & $\mathcal{O}(\log N)$ & $\mathcal{O}(\log N)$ \\
\hline
\textsc{Chord} & $\mathcal{O}(\log N)$ & $\mathcal{O}(N)$ \\
\hline
\textsc{SkipChord} & $\mathcal{O}(s \cdot \log^2 N)$ & $\mathcal{O}(N / (s\cdot \log N))$ \\
\hline
\end{tabular}
\end{center}
\caption{Relevant Metrics for Various Overlay Topologies}
\label{table:measures}
\end{table}
Besides providing a simple way to build and analyze self-stabilizing overlay networks, our analysis provides a set of parameters that a designer can tune to achieve a target level of efficiency. One application of our work, then, is in guiding the creation of new topologies that use the \textsc{Avatar} embedding and strive for low diameter \emph{and} low maximum degree of embedding while still maintaining other desirable properties like robustness to node or link failure. We have demonstrated this process with the creation of the \textsc{SkipChord} topology. It would be interesting to see how other topologies perform in this framework.
Future work could also consider variations to our framework. For instance, currently we have only considered spanning trees created by finding the shortest path from $0$ to every node. Note, however, that this would not have to be the case. One could imagine spanning trees of larger diameter but smaller maximum degree of embeddings that might help lower the degree complexity of the algorithm.
Finally, our framework can be used to better understand the upper and lower bounds for the work or degree expansion of self-stabilizing overlay network protocols. While our results deal entirely with network embeddings, it would be interesting to see if the provided insights help make general bounds for any network, embedded or not.
\bibliographystyle{splncs04}
|
\section{Introduction}
Diffusion MRI is traditionally performed using the well characterized pulsed gradient spin-echo (PGSE) sequence. While PGSE sequences enable efficient diffusion weighting, due to hardware constraints the range of accessible diffusion times remains limited to greater than $\sim$25 ms. Complimentary to PGSE, oscillating diffusion gradients initially proposed by Stepi{\v{s}}nik \cite{stepivsnik1981analysis,callaghan1995frequency}, are utilized in oscillating gradient spin-echo (OGSE) sequences. While the effective diffusion time is not well defined for oscillating gradients, it is generally accepted that shorter effective diffusion times can be achieved with increasing oscillation frequency $\omega$ when compared to PGSE \cite{does2003oscillating,xu2020probing,baron2014oscillating}. Consequently, OGSE has constituted a powerful tool in the investigation of time-dependent diffusion through monitoring of the ADC. As $\omega$ is increased, tracked water molecules diffuse shorter lengths and thereby have a reduced capacity to probe the surrounding environment; hence, the apparent diffusion coefficient (ADC) approaches the free diffusion limit. Deviations from this intrinsic limit, manifested as reductions in the ADC, can be instructive in providing information about physiological structures influencing diffusion such as cell membranes, cell density or extracellular components. Applications on this front include the probing of cell dimensions \cite{li2014fast,xu2014mapping,jiang2016quantification,xu2020magnetic,harkins2021simple,xu2021mri}, surface-to-volume ratios \cite{reynaud2016surface,novikov2011surface} and additional microstructural characteristics such as packing \cite{gore2010characterization,schachter2000measurements}, pore sizes \cite{gore2010characterization} and extracellular space \cite{reynaud2016pulsed}. Recently, OGSE sequences utilized by Arbabi et al. \cite{arbabi2020diffusion} have also been used to investigate structural disorder in the human brain, confirming the predicted short-range disorder model \cite{novikov2014revealing} by demonstrating the square-root dependence of the ADC on oscillation frequency.
It should be noted that most ADC investigations and diffusion tensor imaging (DTI) protocols fundamentally rely on Gaussian diffusion approximations \cite{jones2010diffusion}. However, local inhomogeneity of the tissue microenvironment results in deviations from Gaussian diffusion. Such deviations are quantified by extending the DTI treatment to estimate the directionally dependent diffusion kurtosis tensor, formally introduced in diffusion kurtosis imaging (DKI) \cite{jensen2005diffusional,wu2010mr}. Similar to the ADC, it has been proposed that the diffusion kurtosis is also a time-dependent quantity, approaching zero with vanishing effective diffusion times \cite{jensen2010mri}. Studies by Lee et al. \cite{lee2020vivo} have used PGSE to explore time-dependent kurtosis in human gray matter in the context of exchange, while small animal studies \cite{wu2018oscillating,aggarwal2020diffusion,zhang2021measurement,pyatigorskaya2014relationship} in addition to ex vivo studies \cite{jespersen2018diffusion,portnoy2013oscillating} have benefited from advanced hardware to observe time-dependent kurtosis more thoroughly. Notably, Aggarwal et al. \cite{aggarwal2020diffusion} used both PGSE and OGSE methods to characterize the time-dependence of diffusion kurtosis in mice while Wu et al. \cite{wu2018oscillating} demonstrated the use of OGSE kurtosis imaging of healthy and injured mouse brains. Moreover, both studies highlight the utility of comparing kurtosis measurements between PGSE and OGSE in the form of difference maps, indicating their sensitivity to demyelination \cite{aggarwal2020diffusion} as well as hypoxic-ischemic injury \cite{wu2018oscillating}.
Despite this however, at the time of writing only a single recent preliminary work by Yang et al. \cite{yang2020} has explored the in vivo kurtosis in the human brain using oscillating gradients. The technical challenges associated with translation to clinical imaging systems have severely limited kurtosis measurements in humans using OGSE. In particular, limitations of gradient systems (both slew rate and gradient amplitude) restrict the range of frequencies and b-values available for protocol design. The recent advent of high-performance gradient hardware (as used by Yang et al.) has expanded this parameter space and made kurtosis measurements more feasible. Such gradient insert coils significantly outperform modern clinical gradient systems, routinely realizing max amplitudes of 200 mT/m and slew rates in excess of 500 T/m/s \cite{foo2020highly,yang2020,weiger2018high,huang2021connectome} permitting both rapid slewing and increased gradient amplitudes to be achieved across a wider range of frequencies \cite{tan2020oscillating}. Unfortunately, such hardware is not widely accessible and thus there remains a need to accommodate kurtosis measurements with OGSE to clinically relevant systems to further explore its utility for investigating restricted diffusion. \\
\indent
In this work we demonstrate the first frequency-dependent kurtosis measurements in humans using OGSE with a modern clinical gradient system. Our measurements, in conjunction with PGSE acquisitions, enable the generation of maps demonstrating the frequency dispersion of both the ADC and the diffusion kurtosis. This protocol is enabled by a novel frequency tuned bipolar (FTB) oscillating gradient waveform that reduces the TE of the diffusion acquisition while retaining the intrinsically high b-values required for kurtosis imaging. In the sections that follow we present our waveform design, the optimization and validation of our acquisition protocol and preliminary results for diffusion kurtosis dispersion in healthy human subjects.
\section{Methods}
\subsection{Gradient Waveform Design}
Conventional OGSE is typically performed with trapezoidal cosine modulated waveforms (Figure \ref{fig:paper1}B) to maximize the achievable b-value \cite{van2014vivo}. However, recently proposed by Hennel et al. \cite{hennel2021improved}, the conventional cosine sequence can be modified by optimizing ramp times and reducing the spacing between the two diffusion gradients to the minimum allowable, thereby effectively consolidating the two diffusion gradients into a single waveform. These changes were shown to produce more selective power spectra and increased diffusion weighting capabilities \cite{hennel2021improved}. The framework presented by Hennel et al. enables the implementation of a non-integer number of periods introducing the $N = n + 1/2$ convention where N constitutes the total number of periods and $n$ is an integer greater than 1.
Our proposed FTB waveform achieves shorter diffusion weighting durations by utilizing only N = 1.5 net oscillation periods over both sides of the refocusing RF pulse via two bipolar gradient waveforms that are tuned to achieve the desired net frequency. This approach functions as a variation of OGSE that reduces the duration of the diffusion gradients and thereby significantly reduces the TE of the acquisition. A comparison between this new FTB implementation and typical N = 2 cosine modulated OGSE is presented in Figure \ref{fig:paper1}, where similar spectral selectivity is demonstrated between the two methods.
The waveform is constructed by initially determining the duration of the second lobe ($T$) using the expression:
\begin{equation}
T = \frac{1}{2} \left(\frac{1}{f}-\tau_{RF} - 2\tau \right)
\end{equation}
\noindent
where $\tau_{RF}$ is the separation required for the refocusing pulse, $\tau$ is the gradient rise time and $f$ is the target frequency in Hz. This equation stems from an assumption that the central lobe of the three lobe k-space waveform will dominate the net frequency content of the diffusion weighting (see middle column of Figure \ref{fig:paper1}). Constraining the integral of the zeroth moment to be zero to eliminate any DC spectral components, the duration of the first lobe ($L$) can then be calculated as:
\begin{equation}
L = \frac{1}{2} \left(-2T-\tau_{RF}-7\tau + \sqrt{8T^2 + 8T\tau_{RF} + \tau_{RF}^2 + 32T\tau + 14\tau_{RF}\tau + 33\tau^2} \right)
\end{equation}
This equation can also be further generalized to include additional periods, which we note then resembles the method presented by Hennel et al. \cite{hennel2021improved} with the added condition of encoding only non-zero spectral components. However, the finite truncation of the gradient waveform - more prominent for our abbreviated FTB rendition, imposes a frequency limit based on the minimum permitted separation time $\tau_{RF}$. In general this is not problematic for low frequencies when the period of the waveform is much larger than the separation time. However, when the period of the waveform approaches or exceeds the minimum separation time, the fidelity of the spectral components can become compromised as the gap between the gradients interrupts the periodicity. This can result in deviations from the target frequency in addition to significant spectral broadening. A useful empirical relationship for this limit proposed here is that this method is viable for frequencies that obey the relation $T$ $>$ $\tau_{RF}$ such that the period of the waveform is greater than the separation time between gradients. For a separation time of ~ 7 ms – the minimum permitted on our system, this limit is found to be $\sim$50 Hz.
\subsection{Monte Carlo Optimization}
The signal when kurtosis is introduced by the fourth order cumulant expansion can be written as \cite{jensen2005diffusional,jensen2010mri}:
\begin{equation}\label{sigeq}
S(b) = S_{0} e^{-bD + \frac{b^2D^2K}{6}} e^{-\frac{TE}{T_{2}}}
\end{equation}
where $b$ is the b-value, $D$ the apparent diffusion coefficient, $K$ the kurtosis, $T_{2}$ the spin-spin relaxation time and $TE$ the echo time. Monte Carlo simulations were used to optimize the SNR of the difference in kurtosis between PGSE and OGSE acquisitions ($\Delta K$); we define this quantity as:
\begin{equation}\label{dkeq}
\Delta K = K_{OGSE} - K_{PGSE} = K(\omega) - K(0)
\end{equation}
The signal curve for both PGSE and OGSE acquisitions was generated using equation \ref{sigeq} at three different b-values of 0, a maximum b-value that was varied and a third intermediate b-value that was equal to half of the maximum. In order to maintain feasibility for clinical systems, the maximum frequency was limited to 45 Hz. Gaussian noise was added to the calculated PGSE and OGSE signals upon which the magnitude of the noisy signals were then fitted to equation \ref{sigeq} with a non-negative least squares algorithm to recover ADC and kurtosis values. These values were then used to calculate $\Delta K$ according to equation \ref{dkeq}. This procedure was repeated for 2000 iterations per each frequency/b-value combination. Subsequently the SNR of $\Delta K$ was estimated as the mean of this set of values divided by the standard deviation. The b-values for OGSE and PGSE encoding were simulated for a gradient system with slew rate of 180 T/m/s and max gradient amplitude of 75 mT/m. To avoid the effects of higher order terms in the kurtosis signal expansion our simulations were limited to a maximum b-value of 2500 $s/mm^2$. The TE was chosen to be the minimum allowable for each frequency; accordingly the TE dependence is implicitly reflected through the varied frequency.
Accurate simulation of the diffusion signal required \emph{a priori} knowledge of the frequency dependence of the diffusivity $D(\omega)$ and kurtosis $K(\omega)$ to capture the time-dependence of both quantities. The model of the diffusion dispersion presented by Arbabi et al. \cite{arbabi2020diffusion} was used to infer $D(\omega)$ such that a unique diffusivity was assigned to each frequency. A similar relationship for $K(\omega)$ was also required, however no such characterization has yet been performed for the diffusion kurtosis. Rather, the multi-frequency measurements of mean kurtosis (MK) presented by Yang et al. \cite{yang2020} were fitted to a power law model. This model, reported here as $K(\omega) = 0.93-0.0016 \omega^{(0.78)}$ provided an empirical relationship between frequency and MK enabling simulation of the frequency dependent kurtosis.
The simulation was performed for FTB in addition to alternate OGSE encoding schemes including nominal N = 2 (Figure \ref{fig:paper1}B) and N = 2.5 OGSE to observe the effects of adding additional periods. In addition to $\Delta K$ the SNR of $\Delta ADC$ (defined similarly as $\Delta ADC$ $=$ $ADC_{OGSE}$ – $ADC_{PGSE}$), was also evaluated in a manner identical to $\Delta K$.
\subsection{In Vivo Protocol}
Five healthy participants (3 male, 2 female, mean age 26±4 years) were scanned on a head-only 7 Tesla MRI scanner (Siemens Magnetom 7T Plus, Erlangen Germany). The scanner was equipped with a gradient system capable of a maximum gradient amplitude and slew rate of 80 mT/m and 400 T/m/s respectively, however to maintain clinical feasibility, the maximum gradient amplitude was limited to 75 mT/m and slew rate to 125 T/m/s. Approval for this study was granted by the Institutional Review Board at Western University; written informed consent was obtained from each participant prior to scanning.
The in vivo protocol consisted of an optimal (as determined from Section 2.2) and sub-optimal scan each conducted twice in a test and re-test fashion: subjects were scanned to acquire the test data and were subsequently removed from the scanner, repositioned and returned after a short period to obtain the re-test acquisitions. Diffusion weighted images (DWIs) were acquired with two diffusion weighting schemes each with three shells using b-values of 0, 1250 and 2500 $s/mm^2$ and 0, 1000, 2000 $s/mm^2$ constituting the optimal and sub-optimal scans respectively. Images were acquired at each shell using both PGSE (f = 0 Hz) and 23 Hz frequency tuned bipolar OGSE encoding (see Figure \ref{fig:paper1}A). PGSE and OGSE acquisitions were integrated into one scan such that all acquisitions required for $\Delta K$ map generation could be performed in a single scan. In addition to human subjects, a multi ADC diffusion phantom (CaliberMRI, Boulder Colorado, USA) was also imaged with the optimized in vivo protocol to validate dispersion measurements (see Figure S1).
Diffusion weighting was applied along 4-directions in a tetrahedral scheme to enable maximum achievable b-values \cite{conturo1996}. The remaining acquisition details of both the optimal and sub-optimal scans were identical and were as follows: TE/TR = 91/6500 ms, FOV = 200 $\times$ 200 $mm^2$, matrix size = 100 x 100, 2174 Hz/Px bandwidth, 38 slices, 2 $mm$ isotropic resolution, 8 averages. Images were acquired with 6/8 partial Fourier phase encoding using a single shot EPI readout. The acquisition time per scan was ~14 minutes for a total scanning time of ~56 minutes not inclusive of the break between test/re-test sessions.
\subsection{Image Analysis}
Eddy current characterization was performed independently using a field-monitoring system (Skope MRT, Zurich Switzerland). Acquired k-space trajectories and higher-order field perturbations (up to third order) were utilized to correct for eddy current distortions through integration in an offline model-based image reconstruction algorithm. Principal component analysis based denoising \cite{veraart2016denoising} was also applied to the complex data during reconstruction. Following reconstruction, all diffusion-weighted images were processed with Gibbs ringing removal (MRtrix3) upon which FSL’s BET tool \cite{fslbet} was used to perform brain extraction and mask generation. Re-test images were registered to test images by applying rigid affine transforms generated from b = 0 $s/mm^2$ images from each acquisition; registration was performed using ANTs software \cite{ANTS}.
DWIs from each shell were directionally averaged and fitted on a voxel wise basis to the natural logarithm of equation \ref{sigeq} with a non-negative least squares algorithm to extract ADC and kurtosis parameters; we note the use of a non-negative least squares fitting algorithm eliminates the potential for implausible negative kurtosis values. From the fitted data, mean ADC and apparent kurtosis maps were generated for the PGSE and OGSE acquisitions separately. We note the apparent kurtosis - as used here, is formally distinct from the mean kurtosis, the latter being derived from the diffusion kurtosis tensor.
Dispersion maps of the ADC ($\Delta ADC$) and kurtosis ($\Delta K$) were generated for each subject as the difference between PGSE and OGSE acquisitions; as previously defined: $\Delta K$ = $K_{OGSE}$ – $K_{PGSE}$ and $\Delta ADC$ = $ADC_{OGSE}$ – $ADC_{PGSE}$.
\subsection{Re-test Analysis}
$\Delta K$ maps generated from registered re-test DWIs were compared to test $\Delta K$ maps through a Bland-Altman analysis \cite{BA1986} for both the optimal (b = 2500 $s/mm^2$) and sub-optimal (b = 2000 $s/mm^2$) protocols. A minimum kurtosis threshold of 0.9 was applied to isolate white matter voxels upon which volumes from each subject were combined and were used to generate Bland-Altman plots of $\Delta K$ measurements. Coefficients of variation (CoVs) and standard deviations were calculated for both the optimal and sub-optimal scans to infer relative differences in SNR.
\section{Results}
\subsection{Optimization Results}
Optimization results in Figure \ref{fig:paper2} indicate the optimal frequency of 23 Hz was most influenced by differences in maximum b-value between frequencies. The optimal protocol to acquire $\Delta K$ maps was found to consist of 23 Hz FTB OGSE with a b-value of 2500 $s/mm^2$ and a corresponding TE of 91 ms. The proposed FTB waveform (Figure \ref{fig:paper1}A) achieves higher SNR for both $\Delta K$ as well as $\Delta ADC$ maps compared to traditional N$\geq 2$ OGSE encoding as seen in Figure \ref{fig:paper2}B. We note the distinction in scaling in Figure \ref{fig:paper2}B, with $\Delta ADC$ maps having approximately 4 times higher SNR than $\Delta K$.
\subsection{In Vivo Results}
A trend of decreasing kurtosis and increasing ADC is observed when comparing OGSE to PGSE images as seen in Figure \ref{fig:paper3}. In frontal white matter ROIs (see Figure \ref{fig:paper3}C) mean differences between OGSE and PGSE of approximately 10\% and 14\% are observed across subjects in apparent kurtosis and ADC values respectively. These differences constitute the contrast of the dispersion maps generated from PGSE and OGSE shown in Figure \ref{fig:paper3}E and \ref{fig:paper3}F. Comparable image quality is observed across all subjects in both $\Delta K$ and $\Delta ADC$ maps as seen in Figure \ref{fig:paper4}. Consistent measurements of $\Delta K$ were also observed across participants in the genu, splenium and body of the corpus callosum and are presented in Figure \ref{fig:paper5}B. No statistically significant differences were observed between the regions.
The effect of applied diffusion weighting on $\Delta K$ is exhibited in Figure \ref{fig:paper6}, showing $\Delta K$ maps generated from both the optimal and sub-optimal protocols. A noticeable qualitative difference is observed when comparing the optimal and sub-optimal scans, with the lower b-value protocol exhibiting reduced SNR.
Bland-Altman plots comparing the test and re-test $\Delta K$ measurements across all WM volumes from each subject are shown in Figure \ref{fig:paper7}. Mean CoVs from $\Delta K$ maps across all subjects are reported as 0.60 and 0.69 for the optimal (b = 2500 $s/mm^2$) and sub-optimal (b = 2000 $s/mm^2$) scans respectively. The lower variation in the b = 2500 $s/mm^2$ images suggests a higher SNR of $\Delta K$ which is also qualitatively observed in Figure \ref{fig:paper6} and consistent with the optimization results of Figure \ref{fig:paper2}A. Moreover, mean CoVs for $\Delta ADC$ maps were considerably lower and calculated to be 0.50 and 0.54 for the optimal and sub-optimal protocols respectively, demonstrating increased SNR relative to $\Delta K$. This is supported by observations from Figure \ref{fig:paper2}B that suggest an increased SNR of $\Delta ADC$ maps when compared to $\Delta K$. The trends observed in the coefficients of variation were mirrored in the standard deviations, which are also shown on the Bland-Altman plots in Figure \ref{fig:paper7}.
\section{Discussion}
\subsection{Waveform Remarks}
This article demonstrates the first measurements of kurtosis in the human brain using oscillating gradients on a conventional gradient system. The work was facilitated by the introduction of a frequency tuned bipolar OGSE gradient waveform that significantly reduces the TE of the acquisition. By reducing the minimum number of oscillation periods to 1.5, significant gains in SNR were realized through the reduction of the echo time. Optimization results confirm this configuration to be ideal for measurements of $\Delta K$ and $\Delta ADC$ providing higher SNR for both compared to OGSE performed with additional periods (see Figure \ref{fig:paper2}B). This result agrees with previous findings from Arbabi et al. \cite{arbabi2020diffusion} that suggest a waveform with less than 2 periods would provide higher SNR than conventional N $\geq$ 2 OGSE for $\Delta ADC$ maps. Moreover, this reduction in echo time may also prove useful in future investigations by facilitating the implementation of lower oscillation frequencies without excessive signal loss previously incurred due to extensive TEs. This may find particular applications in future studies exploring PGSE and OGSE sensitivities with comparable diffusion times.
Despite these advantages, the shorter duration of the FTB waveform further limits the range of possible b-values and as such our method is likely best suited for lower oscillation frequencies ($f$ $<$ 40 Hz) where the reduction in b-value due to oscillation is negated by the longer period of the waveform. This effect further justifies the use of the tetrahedral direction scheme, which critically enables the maximum gradient amplitude to be exploited in each direction simultaneously thereby maximizing the applied diffusion weighting \cite{conturo1996}. While the tetrahedral scheme also prevents the use of DKI derived metrics associated with the diffusion kurtosis tensor and may introduce rotational variance \cite{nilsson2020tensor}, the directionally averaged kurtosis has been shown to be a useful and accurate reflection of MK derived from tensor fitting \cite{henriques2021toward}.
In addition, the method presented here inherently includes first-moment nulling to eliminate any spectral component at 0 Hz. While this change also provides flow-compensation to avoid perfusion effects, it may also be vital when performing $\Delta K$ or $\Delta ADC$ map generation using PGSE and OGSE acquisitions. Spectral overlap between the two acquisitions may result in signal loss when a difference is taken and hence eliminating overlapping components mitigates this potential issue.
\subsection{In Vivo Findings}
Generated $\Delta ADC$ maps exhibit comparable diffusion dispersion to previously reported studies. Using the model from Arbabi et al. \cite{arbabi2020diffusion} and the oscillation frequency of 23 Hz, the mean diffusion dispersion rate (across all subjects) can be estimated to be 10 $\mu m^2/s^{1/2}$. This is consistent with mean rate of 11 $\mu m^2/s^{1/2}$ reported by Arbabi et al. \cite{arbabi2020diffusion} in addition to the rate of $\sim$ 10 $\mu m^2/s^{1/2}$ inferred from Baron and Beaulieu's data \cite{baron2014oscillating, arbabi2020diffusion}.
In contrast to $\Delta ADC$ maps, $\Delta K$ maps demonstrate consistent differences between white (WM) and gray matter (GM) regions across all subjects as seen in Figure \ref{fig:paper4}B. The larger negative $\Delta K$ in WM suggests that, somewhat surprisingly, 23 Hz is a sufficiently high frequency to be in a regime where diffusion becomes increasingly Gaussian and kurtosis begins to vanish, similar to findings at short diffusion times in the ex vivo spinal cord and mouse brain \cite{jespersen2018diffusion,aggarwal2020diffusion}. Expectedly however, our frequency is not high enough to observe significant changes between regions of the corpus callosum where microstructural differences are more subtle. While the proximity to the ventricles may make measurements in the corpus callosum more susceptible to cerebrospinal fluid partial volume effects \cite{tetreault2020diffusion}, given the high diffusion weighting cerebrospinal fluid is not anticipated to contribute to these results. Moreover, lack of trends in the $\Delta K$ of the corpus callosum suggests higher frequencies may be required to observe finer structural changes even in dominant WM tracts \cite{reynaud2017time}. Conversely, at even lower frequencies (i.e., longer diffusion times), it is expected that the trend of kurtosis with frequency will reverse due to the effects of permeability, contributing to predicted \cite{aggarwal2020diffusion,fieremans2010monte} and observed \cite{aggarwal2020diffusion,jespersen2018diffusion,portnoy2013oscillating,pyatigorskaya2014relationship} non-monotonic behavior. Notably, this transition frequency may be larger here compared to previous ex vivo samples due to axon shrinkage that is known to occur with sample fixation. Given that the diffusion times accessible through OGSE are shorter than typical exchange times derived from the K\"{a}rger model, our measurements lie in the short diffusion time regime where exchange and permeability effects that act to reduce the apparent kurtosis likely do not have a significant role in these results \cite{lee2020vivo,nilsson2013role}. Moreover, for this reason a monotonic model is likely appropriate to describe $K(\omega)$ in the context of OGSE measurements.
As observed in Figure \ref{fig:paper6} even relatively small decreases in b-value (20\%) for the same frequency have noticeable impact on the SNR of produced $\Delta K$ maps. This effect is consistent with previous studies focusing on the optimization of kurtosis measurements \cite{poot2010optimal,gilani2016parameter} and suggests the optimal $\Delta K$ protocol will favor increasingly larger diffusion weighting. That said, future protocol designs should be attentive to the validity of equation \ref{sigeq} for $b$ $\geq$ 3000 $s/mm^2$ beyond which higher order effects begin to emerge. Such limitations may be avoided by recalling the relation proposed by Jensen et al. \cite{jensen2010mri}, indicating that equation \ref{sigeq} remains valid so long as the condition $b < \frac{3}{KD}$ is satisfied. These effects were assumed to be conservatively avoided in this study by limiting the b-value of the acquisitions to 2500 $s/mm^2$. However, the general increase in $\Delta K$ values for the sub-optimal (lower b-value) acquisition observed in Figure \ref{fig:paper6} may suggest some higher order effects are still present in our optimized protocol due to the increased b-value.
\subsection{Applications}
Observations of the frequency/time dependence of the kurtosis may constitute a novel biomarker, enabling further insight into physiological conditions influencing the degrees of diffusion restriction and heterogeneity - similar to the utility of $\Delta ADC$ \cite{arbabi2020diffusion,iima2019time,gao2020feasibility,wu2021time}. Aggarwal et al. \cite{aggarwal2020diffusion} observed significantly reduced $\Delta MK$ in regions of local demyelination due to increased permeability while Wu et al. \cite{wu2018oscillating} noted increases in $\Delta MK$ corresponding to regions of severe edema in a mouse model of hypoxic ischemic injury demonstrating the sensitivity of the frequency dependent kurtosis to various pathologies. Wu et al. \cite{wu2018oscillating} also noted significantly larger differences in kurtosis relative to differences in ADC between PGSE and OGSE, suggesting $\Delta K$ may be an equally prominent indicator of pathology.
Since our method does not require rapid slewing or extremely large gradient amplitudes our protocol can be easily adapted to full body scanners without exceeding the specifications of modern clinical gradient systems. However despite this, the method presented here is still likely ill-suited to fully characterize the relationship between kurtosis and frequency, $K(\omega)$. The range of frequencies remains limited by the b-values required for observing kurtosis while the tetrahedral scheme prohibits the calculation of the diffusion kurtosis tensor. Consequently, a multi-frequency investigation would benefit from the performance presented by recent advancements in gradient hardware \cite{foo2020highly,weiger2018high,huang2021connectome}. As a result, our optimized protocol does not aim to replace high-performance gradient measurements but rather to compliment them, providing an efficient method to make frequency dependent kurtosis measurements more efficient and accessible on a wider variety of systems.
The primary limitation of this study however, remains the low SNR of the generated $\Delta K$ maps. Since the results of Figure \ref{fig:paper7} demonstrate large single-voxel CoVs for both the optimal and sub-optimal protocols, the use of ROI-based analysis for investigations of diffusion kurtosis dispersion is recommended. Moreover, while our study avoids the use of high-performance gradients, it does benefit from the additional SNR that arises due to imaging at ultra-high field (7T). Equivalent measurements conducted at 3T would require more than twice as many averages thereby further elongating the scan time. However, our technique will benefit from recent and on-going advancements in the field including non-cartesian readouts such as single-shot spirals \cite{lee2021signal} or 3D multi-slab acquisitions \cite{wu2016slab} to provide increased SNR. Future work may also include the in-depth anatomical characterization of $\Delta K$ in addition to multi-frequency investigations of the time dependent kurtosis in the short diffusion-time regime that is accessible with oscillating gradients.
\section{Conclusion}
In this work we present a method for the generation of differential kurtosis maps on the basis of the kurtosis dispersion probed using a more efficient design for frequency-selective diffusion weighting. This frequency tuned bipolar gradient waveform demonstrates highest SNR for both $\Delta K$ and $\Delta ADC$ maps when compared to traditional OGSE encodings. Our findings demonstrate the feasibility of frequency dependent kurtosis measurements on modern clinical gradient systems. Moreover, our optimized protocol constitutes a viable method for exploring the increasingly compelling evidence from rodent studies suggesting the role of the frequency dependent kurtosis as a multi-purpose biomarker.
\section*{Acknowledgements}
This work was supported by the National Science and Engineering Research Council of Canada, the Canada First Research Excellence Fund to BrainsCAN and the Ontario Graduate Scholarship Program. The authors would like to thank Nico Arezza and Naila Rahman for helpful discussions regarding re-test analysis and image registration.
\singlespacing
\bibliographystyle{MRM}
|
\section{Introduction}
\label{intro}
The COVID-19 pandemic has caused severe suffering across the world in both public health and the economy.
These hardships have motivated researchers from various backgrounds to study the viral pathogenesis, epidemic spreading models, mitigation strategies\cite{cao2020covid,anderson2020will}, etc.
Besides the COVID-19 pandemic, it is relevant to build dynamic models to study viral spreading processes to predict future outbreaks and to design control algorithms to mitigate the epidemic \cite{nowzari2016epidemics}. One of the popular ways to capture
viral spreading processes is
by
using
network-based compartmental models\cite{mei2017epidemics_review}. In networked epidemic models, the infection rates, healing rates, and network structures all play important roles in determining the behaviors of the epidemic spreading processes. Recently, social factors such as human awareness\cite{Paarporn_tcss},
opinion interactions\cite{weihao2020_opinion},
etc., are being
taken into consideration when modeling epidemic spreading over networks. In this work, we will consider the classical networked Susceptible-Infected-Recovered (SIR) model coupled with opinion dynamics.
In previous works, researchers have studied the networked SIR models from different perspectives.
In \cite{mei2017epidemics_review}, the authors study
the dynamical behaviors of the networked SIR model, and
analyze
the threshold conditions for an epidemic to increase or decrease. In \cite{hota2020closed},
the authors leverage testing data to estimate the key parameters of the networked SIR model to design resource allocation methods to mitigate the epidemic.
As mentioned before,
people's beliefs in the seriousness of the epidemic
is one important social factor that
will have an impact on the
spreading process.
For example, \cite{weitz2020awareness} studies the correlations between the awareness-driven behaviors during the COVID-19 pandemic and the spreading of the COVID-19. Further, \cite{bhowmick2020influence} constructs a multiplex network with a networked SEIV model coupled with opinion dynamics, then explores the disease-free equilibrium. Inspired by the health-belief model developed by social scientists \cite{healthbelief}, where people's behavior in the pandemic will be influenced by their beliefs in the seriousness of the epidemic, \cite{weihao2020_opinion} and \cite{she2021network} develop a networked SIS model with cooperative opinion dynamics, and both cooperative and antagonistic opinion dynamics, respectively. The authors in \cite{weihao2020_opinion} study both the disease-free and non-disease-free equilibria of the model. Our previous work, \cite{she2021network}, defines an opinion-dependent reproduction number to explore further the effect of the antagonistic opinions in epidemic spreading. Based on the health-belief model, we will develop a networked SIR model coupled with cooperative opinion dynamics.
The main contributions of this work can be summarized as follows:
we define a networked SIR epidemic model coupled with opinion dynamics.
Then, we develop two concepts:
an effective reproduction number and a peak infection time.
We utilize the
effective reproduction number to explore epidemic spreading by studying the peak infection time.
In particular, different from the previous works \cite{weihao2020_opinion,she2021network}, where stability and convergence of the equilibria are the main focuses, this work emphasizes more on exploring the transient behavior of the epidemic, characterized by the effective reproduction number and the peak infection time.
Additionally, we further analyze the opinion states via the behavior of the epidemic.
We organize the paper as follows: In Section II, we introduce the networked SIR model coupled with opinion dynamics and formulate the problems of interest; Section III studies the equilibrium of the developed model. Based on the model, Section III defines the effective reproduction number and peak infection time.
Section III further explores the epidemic's dynamical behavior by relating the effective reproduction number and the peak infection time.
Section IV illustrates the results of the paper through simulations.
Section V concludes the paper and outlines research directions.
\vspace{-5ex}
\subsection*{Notation}
For any positive integer $n$, we use $[n]$ to denote the index set
$\left\{ 1,2,\ldots,n\right\} $. We view vectors as column vectors
and write $x^{\top}$ to denote the transpose of a column vector $x$.
For a vector $x$, we use $x_{i}$ to denote the $i$th entry.
For any matrix $M\in{\rm I\!R}^{n\times n}$, we use $[M]_{i,:}$, $[M]_{:,j}$,
$[M]_{ij}$, to denote its $i$th row, $j$th column and $ij$th
entry, respectively. We use $\tilde{M}=\text{diag}\left\{ m_{1},\ldots,m_{n}\right\} $
to represent a diagonal matrix $\tilde{M}\in{\rm I\!R}^{n\times n}$ with
$[\tilde{M}]_{ii}=m_{i}$, $\forall i\in\left[n\right]$. We use $\boldsymbol{0}$
and $\boldsymbol{1}_n$ to denote the vectors whose entries all equal 0 and 1,
respectively, and $I$ to denote the identity matrix. The dimensions
of the vectors and matrices are to be understood from the context.
For a real square matrix $M$, we use $\rho\left(M\right)$ and $\sigma\left(M\right)$
to denote
its spectral radius and spectral abscissa (the largest real part among its eigenvalues), respectively. For any two vectors $v,w\in{\rm I\!R}^{n}$,
we write $v\geq w$ if $v_{i}\geq w_{i}$,
and $v\gg w$ if $v_{i}>w_{i}$, $\forall i\in\left[n\right]$.
The comparison notations between vectors are used for matrices as well,
for instance, for $A,B\in{\rm I\!R}^{n\times n}$, $A\gg B$ indicates
that $A_{ij}>B_{ij}$, $\forall i,j\in\left[n\right]$.
Consider a directed graph $\mathcal{G}=\left(\mathcal{V},\mathcal{E}\right)$, with the node set $\mathcal{V}=\left\{ v_{1},\ldots,v_{n}\right\} $
and the edge set $\mathcal{E}\subseteq \mathcal{V}\times \mathcal{V}$. Let
matrix $A \in{\rm I\!R}^{n\times n}$, $[A]_{ij}=a_{ij}$, denote the adjacency matrix of $\mathcal{G}=\left(\mathcal{V},\mathcal{E}\right)$, where $a_{ij}\in{\rm I\!R}_{>0}$
if $\left(v_{j},v_{i}\right)\in \mathcal{E}$ and $a_{ij}=0$ otherwise.
Graph $\mathcal{G}$ does not allow self-loops, i.e., $a_{ii}=0,$ $\forall i\in \left[n\right]$. Let
$k_{i}=\sum_{j\in \mathcal{N}_{i}}\left|a_{ij}\right|$, where $\mathcal{N}_{i}=\left\{ \left.v_{j}\right\vert \left(v_{j},v_{i}\right)\in \mathcal{E}\right\}$
denotes the neighbor set of $v_{i}$ and $\left|a_{ij}\right|$ denotes
the absolute value of $a_{ij}$. The graph Laplacian of $\mathcal{G}$ is defined
as $L\triangleq \tilde{K}-A$, where $\tilde{K}\triangleq \text{diag}\left\{ k_{1},\ldots,k_{n}\right\}$.
\vspace{-4ex}
\section{Modeling and Problem Formulation}
In this section, we
introduce the networked SIR model coupled with opinion dynamics. We
also
formulate the problem to be analyzed in this work.
We start by defining a \textit{disease transmission network} $\mathcal{G}=\left(\mathcal{V},\mathcal{E}\right)$ as a weighted directed graph with a node set $\mathcal{V}=\left\{v_{1},\ldots,v_{n}\right\}$ representing $n$ disjoint communities and the edge set $\mathcal{E}\subseteq \mathcal{V}\times \mathcal{V}$ representing disease-transmitting contacts over $\mathcal{V}$. We denote the weight of each edge $(v_j, v_i)$ as $\beta_{ij}$.
Then, a basic continuous-time networked SIR model on graph $\mathcal{G}$, which was studied in \cite{mei2017epidemics_review}, can be defined as:
\vspace{-1ex}
\begin{subequations}
\begin{alignat}{3}
\dot{s_{i}}\left(t\right) &= -s_{i}\left(t\right)\sum_{j\in\mathcal{N}_{i}}\beta_{ij}x_{j}\left(t\right), \label{eq:S}\\
\dot{x_{i}}\left(t\right) &= s_{i}\left(t\right)\sum_{j\in\mathcal{N}_{i}}\beta_{ij}x_{j}\left(t\right) -\gamma_{i}x_{i}\left(t\right),\label{eq:I}\\
\dot{r}_{i}\left(t\right) &= \gamma_{i}x_{i}\left(t\right)\label{eq:R},
\end{alignat}
\end{subequations}
where $(s_i(t), x_i(t), r_i(t)) \in [0, 1]$, $\forall i\in[n]$ are the states indicating the proportion of susceptible, infected, and recovered population in community $i \in
[n]$ at time $t\geq0$, respectively. Moreover, $\beta_{ij} \in \mathbb{R}_{\geq 0}$ is the \textit{transmission rate} from community $j$ to $i$, and $\gamma_{i} \in \mathbb{R}_{\geq 0}$ is the \textit{recovery rate} of community $i$. Note that \eqref{eq:S}-\eqref{eq:R}
satisfy $s(t) + x(t) + r(t) = 1\ \forall t\geq t_0 \in \mathbb{R}_{\geq0}$ as a result of the assumption that $\exists t_0 \in \mathbb{R}_{\geq0}$ such that $s(t_0) + x(t_0) + r(t_0) = 1$ and $\dot{s}(t) + \dot{x}(t) + \dot{r}(t) = 0\ \forall t \in \mathbb{R}_{\geq0}$.
Similarly, we define
the
\textit{opinion spreading network} as a directed graph $\bar{\mathcal{G}}=\left(\mathcal{V},\bar{\mathcal{E}}\right)$, where the edge set $\bar{\mathcal{E}}\subseteq \mathcal{V}\times \mathcal{V}$ represents the opinion-disseminating interactions over the same $n$ communities. Each edge in the graph is weighted by $\bar{a}_{ij} \in \mathbb{R}_{\geq 0}$ indicating the opinion-disseminating influence from node $j$ to node $i$. Let $o_i(t)\in[0,1]$, $\forall i\in[n]$, $t\geq 0$, denote the belief of community $i$ on the severity of the epidemic at time $t$, where $o_i(t)=1$ indicates community $i$ considers the epidemic to be extremely serious, while $o_i(t)=0$ implies
community $i$ believes
the epidemic is not serious at all. We adapt
Abelson’s models of opinion dynamics from
\cite[Equation (10)]{proskurnikov2017tutorial}, where $\bar{\mathcal{N}}_{i}=\left\{ \left.v_{j}\right\vert \left(v_{j},v_{i}\right)\in \bar{\mathcal{E}}\right\}$:
\vspace{-1ex}
\begin{equation} \label{eq:O}
\dot{o}_{i}\left(t\right)=\sum_{j\in \mathcal{\bar{N}}_{i}}\bar{a}_{ij}\left(o_{j}\left(t\right)-o_{i}\left(t\right)\right).
\end{equation}
We assume that the $n$ communities share a homogeneous minimum incoming transmission rate $\beta_{\min}$ and a homogeneous recovery rate $\gamma_{\min}$, where $\beta_{\min}$ corresponds to the strongest belief of a community in the severity of the epidemic $o_i(t) = 1$, while $\gamma_{\min}$ corresponds to the weakest belief of a community in the severity of the epidemic $o_i(t) = 0$.
To couple the networked SIR model with the opinion dynamics, we employ the health-belief model, which is the best known and most widely used theory in health behavior research~\cite{healthbelief}. The health-belief model proposes people's beliefs\footnote{ In this article, beliefs, attitudes, and opinions are used interchangeably.} about health problems, perceived benefits of actions, and/or perceived barriers to actions that can explain their engagement, or lack thereof, in health-promoting behavior. Therefore, people's beliefs in their perceived susceptibility and/or in their perceived severity of the illness affect how susceptible they are and/or how effective they will be at healing from these epidemics. We
define a networked SIR model influenced by the opinion dynamics as:
\begin{subequations}
\small
\begin{alignat}{3}
\dot{s_{i}}\left(t\right) &= -s_{i}\left(t\right)\sum_{j\in\mathcal{N}_i}\left(\beta_{ij}-(\beta_{ij}-\beta_{\min})o_i(t)\right)x_{j}\left(t\right), \label{eq:S-O}\\
\dot{x_{i}}\left(t\right) &=s_{i}\left(t\right)\sum_{j\in\mathcal{N}_i}\left(\beta_{ij}-(\beta_{ij}-\beta_{\min})o_i(t)\right)x_{j}\left(t\right)\nonumber\\
& \ \ \ \ -\left(\gamma_{\min}+(\gamma_i-\gamma_{\min})o_{i}(t)\right)x_{i}\left(t\right).\label{eq:I-O}
\end{alignat}
\end{subequations}
In \eqref{eq:S-O} and \eqref{eq:I-O}, the transmission rate of community $i$,
is obtained through the linear interpolation between $\beta_{ij}$ and $\beta_{min}$ scaled by the level of community $i$'s belief in the seriousness of the epidemic, $o_{i}(t)$.
A higher $o_{i}(t)$ will lead to lower transmission rates for community $i$. A similar interpretation can apply to the healing rate of community $i$ which is scaled by the level of community $i$'s belief in the seriousness of the epidemic, with a higher opinion state
leading to a
higher healing rate of community $i$.
Notice that $(1-s_{i}(t))=x_{i}(t)+r_{i}(t)$, $t\geq0$, $\forall i\in[n]$, captures the proportion of the population that are infected/have been infected with the epidemic. Hence, $(1-s_{i}(t))$ captures the
\textit{infection level}
within community~$i$. By modifying the opinion dynamics in \eqref{eq:O} via cooperating the infection level
\vspace{-1ex}
\begin{equation}
\dot{o_{i}}\left(t\right) =\left(1-s_{i}\left(t\right)-o_{i}\left(t\right)\right)+\sum_{j\in\bar{\mathcal{N}}_i}\left(o_{j}\left(t\right)-o_{i}\left(t\right)\right), \label{eq:O-I}
\end{equation}
where a higher proportion of the infected plus recovered population within community $i$ will lead to
a stronger belief in the seriousness of the epidemic, and vice versa.
We have presented the epidemic-opinion model in \eqref{eq:S-O}-\eqref{eq:O-I}, then, we can state the problem of interest in this work. We are interested in exploring the mutual influence between the epidemic
spreading over
the graph $\mathcal{G}$ of
$n$ communities
in \eqref{eq:S-O} and \eqref{eq:I-O}, and the opinions of the $n$ communities about the epidemic captured by graph $\bar{\mathcal{G}}$ in \eqref{eq:O-I}.
In this paper, we will:
\begin{enumerate}
\item analyze the equilibria of the system in \eqref{eq:S-O}-\eqref{eq:O-I}.
In particular, we connect the opinion states at the equilibrium to the infection level of the communities;
\item define an
effective reproduction number
to characterize the spreading of the disease. In particular, we
explore the transient behavior of the epidemic-opinion model by leveraging peak infection time;
\item illustrate the results through simulations.
\end{enumerate}
The analysis presented in this work can provide insights for decision-makers who aim to analyze disease spreading
and its coupling with
the public's opinion towards the epidemic.
\label{section2}
\vspace{-5ex}
\section{Main Results}
\label{section3}
This section examines the mutual influence between the epidemic dynamics in \eqref{eq:S-O} and \eqref{eq:I-O}, and the opinion dynamics in \eqref{eq:O-I}.
Particularly,
we construct the compact form of the incorporated system to define an effective reproduction number to explore the peak infection time of the
model. We also
analyze the evolution of the epidemic
by using the effective reproduction number and peak infection time.
We write \eqref{eq:S-O}, \eqref{eq:I-O}, and \eqref{eq:O-I} in a compact
form as follows:
\vspace{-1ex}
\begin{subequations}
\small
\begin{alignat}{3}
\label{eq:S-O-C}
\dot{s}\left(t\right)&=-\left(\tilde{S}\left(t\right)\left(B-\tilde{O}\left(t\right)\left(B-B_{\min}\right)\right)\right)x\left(t\right),
\\
\dot{x}\left(t\right) & =\left(\tilde{S}\left(t\right)\left(B-\tilde{O}\left(t\right)\left(B-B_{\min}\right)\right)\right)x\left(t\right)\nonumber\\
& \ \ \ \ \ -\left(G_{\min}+\left(G-G_{\min}\right)\tilde{O}\left(t\right)\right)x\left(t\right),\label{eq:I-O-C}
\\
\label{eq:O-I-C}
\dot{o}\left(t\right)&=\left(\boldsymbol{1}_{n}-s\left(t\right)\right)-\left(\bar{L}+I_{n}\right)o\left(t\right),
\end{alignat}
\end{subequations}
where $\tilde{S}\left(t\right)=\text{diag}\left(s\left(t\right)\right)$, $\tilde{O}\left(t\right)=\text{diag}\left(o\left(t\right)\right)$,
$G_{\min}$ and $G$ are diagonal
matrices, with $\left[G_{\min}\right]_{ii}=\gamma_{\min}$,
and $[G]_{ii}=\gamma_{i}$, $\forall i \in[n]$. Note that $\bar{L}$ is the Laplacian matrix of the opinion spreading graph $\mathcal{\bar{G}}$.
By defining $B\left(o(t)\right)=(B-\tilde{O}\left(t\right)\left(B-B_{\min}\right))$, $G\left(o(t)\right)=(G_{\min}+\left(G-G_{\min}\right)\tilde{O}\left(t\right))$,
\vspace{-1ex}
\begin{subequations}
\small
\begin{alignat}{2}
\label{S-O-Final}
\dot{s}\left(t\right) &=-\left(\tilde{S}\left(t\right)B\left(o(t)\right)\right)x\left(t\right),
\\
\label{I-O-Final}
\dot{x}\left(t\right) & =\tilde{S}\left(t\right)B\left(o(t)\right)x\left(t\right)-G\left(o(t)\right)x\left(t\right).
\end{alignat}
\end{subequations}
For the epidemic spreading process, we assume that community $i$ can pass the virus to community $j$ through at least one directed path in the network $\mathcal{G}$, $\forall i, j \in[n]$, $i\neq j$. For the opinion spreading process, we assume that community $i$ can affect community $j$'s opinion through at least one directed path in $\mathcal{\bar{G}}$. Therefore, we have the following assumption for the epidemic and opinion spreading over the communities:
\begin{assumption}
\label{A1}
Suppose $\forall i\in[n]$,
$s_{i}(0), x_{i}(0), o_{i}(0)\in[0,1]$, $s_i(0)+x_i(0)+r_i(0)=1$,
$\gamma_{i}\ge\gamma_{{\rm min}}>0$, and $\beta_{ij}\ge\beta_{{\rm min}}>0$, $\forall j\in \mathcal{N}_{i}$. Further, both $\mathcal{G}$ and $\bar{\mathcal{G}}$ are strongly connected.
\end{assumption}
\vspace{-3ex}
\subsection{Equilibrium}
First we show the model in \eqref{eq:S-O-C}-\eqref{eq:O-I-C} is well-defined.
\begin{lemma}
\label{lem: well-defined}
For the epidemic-opinion model in \eqref{eq:S-O-C}-\eqref{eq:O-I-C}, if $(s_{i}(0), x_{i}(0), o_{i}(0))\in[0,1]$, and $s_i(0)+x_i(0)+r_i(0)=1$, then $(s_{i}(t), x_{i}(t), o_{i}(t))\in[0,1]$, $\forall t>0$, $\forall i\in[n]$.
\end{lemma}
The system defined in \eqref{eq:S-O-C}-\eqref{eq:O-I-C} is a group of polynomial ODEs defined over the compact set $[0,1]^{3n}$. Hence, the system
is Lipschitz over $[0,1]^{3n}$.
To prove Lemma \ref{lem: well-defined}, we can show that all gradient vector fields on the boundary of the set $[0,1]^{3n}$ are either pointing towards the set's interior or tangential to the boundary~\cite{blanchini2008set}.
The proof is similar to the proof of {\cite [Lemma 7]{she2021network}} and thus omitted here.
\begin{lemma}
\label{lem:decreasing}
If $s_{i}(0), x_{i}(0), o_{i}(0)\in[0,1]$, $\forall i\in [n]$, the susceptible states, $s_{i}(t)$, are
monotonically decreasing
\end{lemma}
\begin{proof}
Lemma \ref{lem: well-defined} indicates that $x(t)\in [0,1]^n$ $\forall t\geq0$. From Assumption \ref{A1}, the matrix $-(\tilde{S}(t)(B-\tilde{O}(t)(B-B_{\min})))$ is non-positive, therefore, the RHS of \eqref{eq:S-O-C} is always non-positive, which completes the proof.
\end{proof}
After considering the monotonicity of the susceptible population, we move to the next lemmas to study the equilibria of the epidemic-opinion model.
\begin{lemma}
\label{lem:Equi}
The equilibria of the epidemic-opinion model in \eqref{eq:S-O-C}-\eqref{eq:O-I-C} take the form $(s_{e},\boldsymbol{0},(\bar{L}+I_{n})^{-1}(\boldsymbol{1}_{n}-s_{e}))$, where $[s_e]_i\in[0,1]$ and $[(\bar{L}+I_{n})^{-1}(\boldsymbol{1}_{n}-s_{e}))]_i\in [0,1]$, $\forall i\in [n]$.
\end{lemma}
\begin{proof}
Based on Lemma \ref{lem:decreasing}, the RHS of \eqref{eq:S-O-C} is always less than or equal to zero. Therefore, the susceptible states are monotonically decreasing.
To compute the equilibria of the system in \eqref{eq:S-O-C}-\eqref{eq:O-I-C}, let $\dot{s}(t)=\dot{x}(t)=\dot{o}(t)=\boldsymbol{0}$, which leads to $x_e=\boldsymbol{0}$, and $s_e$ can be any point located in the set $[0,1]^n$. Substituting the equilibrium point $s_e$ into \eqref{eq:O-I-C} with $\dot{o}(t)=\boldsymbol{0}$,
\begin{equation}
\label{eq:equi_opi}
\boldsymbol{0}=(\boldsymbol{1}_{n}-s_e)-\left(\bar{L}+I_{n}\right)o_e.
\end{equation}
Since the row sums of the Laplacian matrix $L$ are zeros, the matrix $(\bar{L}+I_n)$ is a strictly diagonally dominant matrix, i.e., the matrix $(\bar{L}+I_n)$ is positive definite. Therefore,
\begin{equation}
\label{eq:unique}
o_{e}=\left(\bar{L}+I_{n}\right)^{-1}\left(\boldsymbol{1}_{n}-s_{e}\right).
\end{equation}
Now we have to show that for each $s_e\in [0,1]^n$, the solution $o_e\in[0,1]^n$. Rearrange
\eqref{eq:equi_opi} as follows:
\[
\left(\boldsymbol{1}_{n}-s_e\right)=\left(\bar{L}+I_{n}\right)o_e.
\]
If $s_e\in [0,1]^n$, we have $(\boldsymbol{1}_n-s_e)\in [0,1]^n$, which leads to $\left(\bar{L}+I_{n}\right)o_e\in [0,1]^n$.
We prove $[o_e]_i\in[0,1]$ $\forall i\in [n]$ by contradiction. Without loss of generality, suppose that $[o_e]_i>1$, with $[o_e]_i-a_i=1$, belongs to a solution to the $i$th row of the equation
$\left(\boldsymbol{1}_{n}-s_e\right)=\left(\bar{L}+I_{n}\right)o_e$.
As shown before, the matrix $\left(\bar{L}+I_{n}\right)$ is strictly diagonally dominant and $\bar{L}_{ii}+\sum_{j\in\mathcal{N}_{i}}\bar{L}_{ij}+1=1$, $\forall i\in[n]$. If the rest of the entries of the solution $[o_e]_j\leq [o_e]_i$, $j\neq i$, $j\in [n]$, we must have $[\left(I_{n}\right)o_e]_i>1$, which violates $(\boldsymbol{1}_n-s_e)\in[0,1]^n$. Hence, to ensure
the $i$th entry
$[\left(\bar{L}+I_{n}\right)o_e]_i\in[0,1]$,
there must be an entry $[o_e]_j>[o_e]_i$ in the solution $o_e$ such that $[o_e]_j-a_j=1$, where $a_j>a_i$, $j\in [n], j\neq i$.
Then, consider the $j$th row of the equation $\left(\boldsymbol{1}_{n}-s_e\right)=\left(\bar{L}+I_{n}\right)o_e$. To ensure the solution of the $j$th row belongs to $[0,1]$, based on the same analysis technique, there must exist
one entry $[o_e]_k$ such that $[o_e]_k>1$, with $[o_e]_k-a_k=1$, $a_k>a_j>a_i$. Following the same process to check the rest of the rows of the equation, we can conclude that, for the last equation left to check, there is no such entry left in $o_e$ satisfying the inequality condition, $a_l>\dots>a_k>a_j>a_i$, where $[o_e]_l$ corresponds to the last row left to be checked in the equation $(\boldsymbol{1}_{n}-s_e)=(\bar{L}+I_{n})o_e$, with
$[o_e]_l-a_l=1$. Therefore, each entry of $o_e$ as a solution to $\left(\boldsymbol{1}_{n}-s_e\right)=\left(\bar{L}+I_{n}\right)o_e$ is not greater than 1.
Using the same analysis technique, it can be shown that
each entry of $o_e$ as a solution to $\left(\boldsymbol{1}_{n}-s_e\right)=\left(\bar{L}+I_{n}\right)o_e$ is not smaller than 0. Therefore, the solution $o_e$ to the equation $\left(\boldsymbol{1}_{n}-s_e\right)=\left(\bar{L}+I_{n}\right)o_e$ must be located in $[0,1]^n$, which proves the statement.
\end{proof}
Lemma \ref{lem:Equi} shows that there are
infinite equilibria for the epidemic-opinion model captured by \eqref{eq:S-O-C}-\eqref{eq:O-I-C}.
In particular, \eqref{eq:unique} indicates that the opinion states of the communities at the equilibrium can be uniquely evaluated as a function of the
steady-state
susceptible population in the communities.
The following lemma further characterizes the condition that the communities reach a consensus on their opinions, i.e., the opinion states are the same when the epidemic disappears.
\begin{lemma}
\label{consensus}
The communities will reach consensus on their opinions if and only if all the communities have the same proportion of infections, captured by the equilibria $\left(s_{e},\boldsymbol{0},\boldsymbol{1}_n-s_{e}\right)$, where $[s_e]_i = [s_e]_j\ \forall i\neq j$.
\end{lemma}
\begin{proof}
First, we show the necessary condition. If all communities reach consensus at the equilibrium point, giving that $o_e$ is a consensus state, where $[o_e]_i=[o_e]_j$, $\forall i,j \in [n]$, $i\neq j$, then $Lo_e=\boldsymbol{0}$. From \eqref{eq:equi_opi}, $s_e=\boldsymbol{1}_n-o_e$, which implies that $[s_e]_i=1-[o_e]_i$, and $[s_e]_i=[s_e]_j$, $\forall i,j \in [n]$, $i\neq j$.
For the sufficient condition, if all communities have the same proportion of the infected population at the equilibria, we have $1-[s_e]_i=1-[s_e]_j$, $\forall i,j \in [n]$, $i\neq j$. Based on \eqref{eq:equi_opi}, we have $\boldsymbol{1}_n-s_e=(\bar{L}+I_n)o_e$.
Since $\left(\bar{L}+I_{n}\right)$ is a positive definite matrix, $o_e=\left(\bar{L}+I_{n}\right)^{-1}(\boldsymbol{1}_n-s_e)$ is the unique solution to the equation. Further, $L$ is positive semi-definite with only one zero eigenvalue, paired with the eigenvector $v$, where all the entries of $v$ are the same. Therefore, $\left(\bar{L}+I_{n}\right)v=v$,
giving that
$v=o_e=(\boldsymbol{1}_n-s_e)$ is the unique solution to the equation $\left(\bar{L}+I_{n}\right)o_e=\boldsymbol{1}_n-s_e$, completing
the proof.
\end{proof}
Lemma \ref{lem:Equi} and Lemma \ref{consensus} summarize the equilibria of the epidemic-opinion model from \eqref{eq:S-O-C} to \eqref{eq:O-I-C}. In particular, the lemmas
show that the communities' beliefs in the seriousness of the epidemic can reflect the
infection level.
More importantly, the communities will reach consensus on the seriousness of the epidemic if and only if the epidemic caused the same
proportion of infected population in all communities. Under this situation, the belief on the seriousness of the epidemic is proportional to the proportion of the recovered population in all communities, characterized by $o_e=\boldsymbol{1}_n-s_{e}$.
\begin{remark}
The communities can rarely reach a consensus of their opinions on the epidemic's severity since it will be implied by Lemma \ref{consensus} that every community has the same infection level, which is unusual. However, one exception is when every community is fully infected, then all communities will agree that the epidemic is utterly severe ($o_e = \boldsymbol{1}_n$).
\end{remark}
\vspace{-3ex}
\vspace{-1ex}
\subsection{Effective Reproduction Number}
\vspace{-1ex}
The effective reproduction number of the model characterizes the dynamical behavior of the system. We introduce the following lemmas before formally defining the notion.
\begin{lemma}
\label{lem:irr_non}\cite[Thm. 2.7, and Lemma 2.4]{varga2009matrix_book} Suppose
that M is an irreducible nonnegative matrix. Then:
\begin{enumerate}
\item M has a simple positive real eigenvalue equal to its spectral radius, $\rho(M)$;
\item There is a unique (up to scalar multiple) left eigenvector $v\gg\boldsymbol{0}$ (right eigenvector $w\gg \boldsymbol{0}$) pairing with $\rho(M)$;
\item $\rho(M)$ increases when any entry of M increases;
\item If N is also an irreducible nonnegative matrix with the same size and $M\geq N$, then $\rho(M)\geq\rho(N)$.
\end{enumerate}
\end{lemma}
\begin{lemma}
\label{lem:irr_spe}\cite[Prop. 1]{bivirus}
Suppose that $\varLambda$ is a negative diagonal matrix in ${\rm I\!R}^{n\times n}$
and $N$ is an irreducible nonnegative matrix in ${\rm I\!R}^{n\times n}$.
Let $M=\varLambda+N$. Then, $\sigma(M)<0$ if and only if $\rho(-\varLambda^{-1}N)<1$,
$\sigma(M)=0$ if and only if $\rho(-\varLambda^{-1}N)=1$, and $\sigma(M)>0$
if and only if $\rho(-\varLambda^{-1}N)~>~1$.
\end{lemma}
\begin{lemma}
\cite[Sec. 2.1 and Lemma 2.3]{varga2009matrix_book}
\label{lem:irr_M} Suppose that $M$ is an irreducible Metzler matrix. Then, $\sigma\left(M\right)$ is a simple eigenvalue of $M$ and
there exists a unique (up to scalar multiple) left eigenvector $x\gg\boldsymbol{0}$ (right eigenvector $y\gg\boldsymbol{0}$) such
that $x^\top M=\sigma\left(M\right)x$ ($My=\sigma\left(M\right)y$). %
\end{lemma}
\begin{lemma}
\label{lem:Metzler}
\cite[Sec. 1, Lemma 2]{cvetkovic2020stabilizing}
Suppose that $A, B\in{\rm I\!R}^{n\times n}$ are Metzler matrices. Then, $\sigma\left(A\right)\leq \sigma\left(B\right)$ if $A\leq B$.
\end{lemma}
\begin{definition}{[Effective
Reproduction Number $R_o(t)$]
Let $R_o(t)=\rho(G^{-1}(o(t))\tilde{S}(t)B(o(t)))$, $\forall t\geq0$, denote the
Effective
Reproduction Number, where $G(o(t))$, $\tilde{S}(t)$, and $B(o(t))$ are defined in \eqref{eq:I-O-C}.
\label{def:Rt} }
\end{definition}
Note that the
effective
reproduction number $R_o(t)$ depends not only on the proportion of the susceptible population $s(t)$, but also
on the variation of the opinion states \bs{$o\left(t\right)$}.
\begin{proposition}
\label{prop:Spe_r}The
Effective
Reproduction Number $R_o(t)$
\bs{has} the following
\bs{properties}:
\begin{enumerate}
\item If
\[
G^{-1}(o(t_1))\tilde{S}(t_1)B(o(t_1))\geq G^{-1}(o(t_0))\tilde{S}(t_0)B(o(t_0)),
\]
then
$R_o(t_1)\geq R_o(t_0)$;
\item $R_o(t)$ is strictly monotonically decreasing with respect to $s(t)$, $\forall t\geq 0$;
\item If $o\left(t_{0}\right)\leq o\left(t_{1}\right)$, $\forall t_0<t_1$, then
$R_o(t_1)\leq R_o(t_0)$.
\end{enumerate}
\end{proposition}
\begin{proof}
1) Based on Assumption \ref{A1}, and the definitions of
$G(o(t))$ and $\tilde{S}(t)$,
we conclude that $G(o(t))$ and $\tilde{S}(t)$
are
positive definite diagonal matrices,
and $B(o(t))$
is an irreducible nonnegative matrix, $\forall t\geq0$. Hence, $G^{-1}(o(t))\tilde{S}(t)B(o(t))$
is an irreducible nonnegative matrix. For statement 1),
if
\[
G^{-1}(o(t_1))\tilde{S}(t_1)B(o(t_1))\geq G^{-1}(o(t_0))\tilde{S}(t_0)B(o(t_0)),
\]
based on Lemma \ref{lem:irr_non},
\[
\rho(G^{-1}(o(t_1))\tilde{S}(t_1)B(o(t_1)))\geq \rho(G^{-1}(o(t_0))\tilde{S}(t_0)B(o(t_0))),
\]
which leads to $R_o(t_1)\geq R_o(t_0)$.
2) $R_o(t)$ is strictly monotonically decreasing with respect to $s(t)$ means that, when $o(t)$ is fixed,
a decrease in
$s(t)$ leads to
a decrease in
$R_o(t)$, $\forall t\geq0$. Without loss of generality, assume that $o(t_0)=o(t_1)$, and $t_0< t_1$.
From Lemma \ref{lem:decreasing}, the proportion of infected population for each community is monotonically decreasing. Thus, $s(t_1)\leq s(t_0)$, and $[\tilde{S}(t_1)]_{ii}\leq [\tilde{S}(t_0)]_{ii}$, $\forall i\in[n]$,
which leads to
\vspace{-2ex}
\small
\[
\left[G\left(o\left(t_{0}\right)\right)^{-1}\tilde{S}(t_0)B\left(o\left(t_{0}\right)\right)\right]_{i,:}\geq\left[G\left(o\left(t_{1}\right)\right)^{-1}\tilde{S}(t_1)B\left(o\left(t_{1}\right)\right)\right]_{i,:}
\]
\vspace{-1.5ex}
\normalsize
\noindent
where other entries of both matrices remain the same.
Based on statement 1) of this proposition, $R_o(t_1)\leq R_o(t_0)$, implying that $R_o(t)$ is monotonically decreasing with respect to $s(t)$.
3) When $o\left(t_{0}\right)\leq o\left(t_{1}\right)$, $\forall t_0<t_1$, we have
\[
[G^{-1}\left(o\left(t_{0}\right)\right)]_{ii}\geq [G^{-1}\left(o\left(t_{1}\right)\right)]_{ii},
\]
\[
[B\left(o\left(t_{0}\right)\right)]_{i,:}\geq [B\left(o\left(t_{1}\right)\right)]_{i,:},
\]
where other entries of the matrices $G^{-1}\left(o\left(t_{1}\right)\right)$
and $B\left(o\left(t_{1}\right)\right)$
are equal to $G^{-1}\left(o\left(t_{0}\right)\right)$
and $B\left(o\left(t_{1}\right)\right)$, respectively. Additionally, from Lemma \ref{lem:decreasing}, $\forall t_0<t_1$, we have $\tilde{S}_{ii}(t_1)\leq \tilde{S}_{ii}(t_0)$. Following
the same analysis technique from the proof of statement 2), we
have
that $R_o(t_1)\leq R_{o}(t_0)$.
\end{proof}
The effective reproduction number is influenced by both the opinion states and the proportion of the susceptible population. In particular, when the opinions are fixed, the susceptible proportion will always ensure
that
the effective reproduction number
decreases,
since the recovered population will not be infected again. The opinion states will also have an influence on the change of the effective reproduction number in both directions: higher opinion states will lead to a lower effective reproduction number, and vice versa. As we mentioned in Section II, communities with stronger beliefs in the seriousness of the epidemic will take actions to avoid infections, leading to a lower effective reproduction number, and vice versa.
Further, when all communities think the epidemic is extremely serious,
$o_{\max}(t)=\boldsymbol{1}_n, \forall t\geq 0$. When all communities think the epidemic is not worth treating at all during the pandemic, $o_{\min}(t)=\boldsymbol{0}$, $\forall t\geq 0$. Under
the two extreme
situations, the effective reproduction number
satisfies the following
result.
\begin{corollary}
\label{Prop:bounds}
For all $t\geq 0$, the effective reproduction number $R_{o}(t)$ satisfies $R_{\min}(t)
\leq R_{o}(t)\leq R_{\max}(t)$, where
\begin{align*}
R_{\min}(t)=\rho\left(G^{-1}\left(o_{\max}\right)\tilde{S}\left(t\right)B\left(o_{\max}\right)\right),
\end{align*}
\vspace{-1.5ex}
\noindent
and
\vspace{-1.5ex}
\begin{align*}
R_{\max}(t)=\rho\left(G^{-1}\left(o_{\min}\right)\tilde{S}(t)B\left(o_{\min}\right)\right).
\end{align*}
\end{corollary}
The proof of Corollary \ref{Prop:bounds} is similar to the proof of Proposition \ref{prop:Spe_r}, thus omitted here. Corollary \ref{Prop:bounds} indicates that, given any time $t$, if the proportion of the susceptible population of each community are the same, the effective reproduction number is determined by the opinion states, where stronger beliefs in the seriousness of the epidemic lead to a lower effective reproduction number, and vice versa. Compared to the classical
SIR model\cite{van2002reproduction}, where the effective reproduction number is monotonically decreasing with respect to the decreasement of the proportion of the susceptible population, under the influence of the opinions, the
$R_o(t)$ defined in this work may not monotonically decrease. Therefore, $R_o(t)$ can lead to
more diverse
behaviors in
the
epidemic spreading process. In order to analyze the dynamical behavior of the epidemic-opinion model, we define a
concept called peak infection time
to characterize the influence of the effective reproduction number $R_o(t)$ in determining the behavior of the epidemic.
\vspace{-1.5ex}
\subsection{Peak Infection Time}
To connect the effective reproduction number $R_o(t)$ to the behavior of the epidemic-opinion model, we denote
\begin{equation}\label{eq:sigma}
\sigma(t)=\sigma(\tilde{S}(t)B(o(t))-G(o(t)))
\end{equation}
and $p(t)$ as the spectral abscissa of $(\tilde{S}(t)B(o(t))-G(o(t)))$ and the corresponding normalized left eigenvector $\forall t\geq 0$, respectively. From Assumption \ref{A1} and Lemma \ref{lem:irr_M}, $(\tilde{S}(t)B(o(t))-G(o(t)))$ is an irreducible Metzler matrix,
thus
$\sigma(t)$, $\forall t\geq0$, is a positive real eigenvalue. Additionally, the normalized left eigenvector $p(t)$ satisfies $p(t)\gg \boldsymbol{0}$ and $p^\top(t)\boldsymbol{1}_n=1$, $\forall t\geq0$. Then, we define a weighted average of the epidemic states, for a given $t_1\in[t_0,t_2]$, $p^\top(t_1)x(t)$ as a metric to reflect the trend of the epidemic over the time interval $[t_0, t_2]$.
Based on the properties of $\sigma(t)$ and $p(t)$, we have $p^\top(t_1) x(t)\geq 0$, $\forall t \geq 0$ and $p^\top(t_1) x(t)= 0$ if and only if $x(t)=\boldsymbol{0}$.
Therefore, $p^\top(t_1) x(t)$ reflects the overall trend of the epidemic spreading over the time interval $[t_0,t_2]$, and $p^\top(t_1)x(t)=0$ if and only if the epidemic has died out.
\begin{definition}{\label{def:PT}} [Peak Infection Time $t_p$]
A
Peak Infection Time $t_p$ is defined as a turning point,
where $p^\top(t_p) x(t)$ is increasing for all $t\in[t_0,t_p)$ and $p^\top(t_p)x(t)$ is decreasing for all $t\in(t_p,t_1]$, for
sufficiently small time intervals $(t_p-t_0)>0$ and $(t_1-t_p)>0$.
\end{definition}
\noindent
The peak infection time describes
a point in time
where the weighted average of the infected proportions $p^\top(t_p) x(t)$
over the communities reaches a local peak value over the time interval $[t_0, t_1]$.
\begin{theorem}
\label{thm:1}
Given a peak infection time $t_p$, we have
$R_{o}(t_p)=1$, $R_{o}(t)>1$, for $t\in [t_0, t_p)$ and $R_{o}(t)<1$, for $t \in (t_p, t_1]$,
for $t_p-t_0>0$ and $t_1-t_p>0$
sufficiently small.
\end{theorem}
\begin{proof}
First we show that for a peak infection time $t_p$, we have $R_o(t_p)=1$.
Since $\frac{d}{dt}(p^{\top}(t_p)x(t))$ is a continuous function over the time interval $[t_0, t_1]$, based on Definition~\ref{def:PT}, $p^\top(t_p) x(t)$ is increasing for all $t \in [t_0,t_p)$ and
decreasing for all $t \in (t_p,t_1]$. Therefore, by continuity,
$\frac{d}{dt}(p^{\top}(t_p)x(t))=0$ at time $t_p$.
Using this fact and multiplying $p^\top(t_p)$ on both sides of \eqref{I-O-Final},
we have
\begin{align*}
0 &=
\frac{d}{dt}(p^{\top}(t_p)x(t))\Big|_{t=t_p}
\\%& =p^{\top}(t_p)(\tilde{S}\left(t\right)B\left(o(t)\right)x\left(t\right)-G\left(o(t)\right)x\left(t\right))\\
& =(p^{\top}(t_p))(\tilde{S}\left(t\right)B\left(o(t_p)\right)-G\left(o(t_p)\right))x\left(t_p\right)\\
& = \sigma(t_p)p^{\top}(t_p)x\left(t_p\right)
,
\end{align*}
where the third equality follows from the definition of $\sigma(t_p)$ in \eqref{eq:sigma}.
Recall that $p(t_p)x(t)>0$ unless $x(t)=x_e=\boldsymbol{0}$. Thus, for $\frac{d}{dt}(p^{\top}(t_p)x(t_p))=0$, we must have $\sigma(t_p)=0$. Therefore, from Definition \ref{def:Rt} and Lemma \ref{lem:irr_spe}, $R_o(t_p)=1$.
Now we consider the time interval $[t_0, t_p)$. Since $x(t)$ and $o(t)$ are continuous functions in~$t$,
$\sigma(t)$
is also continuous in~$t$.
Then, for a given time $\tau \in [t_0, t_p)$, since $[t_0, t_p)$ is a sufficiently small time interval, we have $p(\tau)\approx p(t_p)$, by continuity.
From Definition \ref{def:PT}, since $p^\top(t_p) x(t)$ is increasing for all $t\in[t_0,t_p)$, $\frac{d}{dt}(p^{\top}(t_p)x(t))>0$ for $t\in [t_0, t_p)$. Using this fact and
multiplying $p^\top(t_p)$ on both sides of \eqref{I-O-Final},
\begin{align*}
0&<\frac{d}{dt}(p^{\top}(t_p)x(t))\Big|_{t=\tau}\\
&=(p^{\top}(t_p))(\tilde{S}\left(\tau\right)B\left(o(\tau)\right)-G\left(o(\tau)\right))x\left(\tau\right)\\
&\approx (p^{\top}(\tau))(\tilde{S}\left(\tau\right)B\left(o(\tau)\right)-G\left(o(\tau)\right))x\left(\tau\right)\\
&=\sigma(\tau)(p^{\top}(\tau))x(t_\tau).
\end{align*}
Since $p^{\top}(t)x(\tau)>0$, for $x(\tau)>0$, we have that $\sigma(\tau)>0$. Therefore, by Lemma \ref{lem:irr_spe}, $R(\tau)>1$ for any time $\tau \in [t_0,t_p)$.
Following the same analysis techniques, we can show that
$R(\tau)<1$, for all $\tau\in(t_p,t_1]$, given a sufficiently small time interval $(t_p-t_1)>0$, completing the proof.
\end{proof}
Note that $R_o(t)=1$ is a necessary condition for the peak infection time, thus the condition does not guarantee that $t$ is the peak infection time.
From Proposition 1, $R_o(t)$ is not a monotonic function with respect to $t$. Consider the case that $R_o(t_1)=1$, if, for $\epsilon>0$, $R_o(t_1-\epsilon)<1$ and $R_o(t_1+\epsilon)>1$, the time $t_1$ is not the peak infection time. Additionally, from Lemma \cite[Sec. 2.1 and Lemma 2.3]{varga2009matrix_book}, $p(t_p)$ is unique for a peak infection time $t_p$.
For $\forall t\in [t_1,t_2]$, from Lemma \ref{lem:decreasing} and \eqref{I-O-Final}, we have
\begin{equation}
(\tilde{S}\left(t_1\right)B\left(o_{\min}\right)-G\left(o_{\min}\right))\geq (\tilde{S}\left(t\right)B\left(o(t)\right)-G\left(o(t)\right).
\label{eq:ineq}
\end{equation}
Based on Corollary \ref{Prop:bounds}, $R_{\min}\leq R_o(t)\leq R_{\max}(t)$, $\forall t\geq0$,
we define $\sigma_{\max}(t)=\sigma(\tilde{S}(t)B(o_{\min})-G^{-1}(o_{\min}))$, corresponding to $R_{\max}(t)$.
Since $(\tilde{S}(t)B(o(t))-G^{-1}(o(t))$ is a Metzler matrix $\forall t$, from Lemma \ref{lem:Metzler}, we have $\sigma_{\max}(t_1)\geq \sigma_{\max}(t)\geq \sigma(t)$, $\forall t\geq t_1$.
Then, we define $p_{\max}(t_1)$ corresponding to $\sigma_{\max}(t_1)$, and multiplying $p_{\max}(t_1)$, on both sides of \eqref{I-O-Final},
\vspace{-2ex}
\small
\begin{align*}
\frac{d}{dt}(p_{\max}^{\top}(t_1)x(t)) &=p_{\max}^{\top}(t_1)((\tilde{S}\left(t\right)B\left(o(t)\right)-G\left(o(t)\right))x\left(t\right)).
\end{align*}
\vspace{-1ex}
\normalsize
\noindent
Then, based on \eqref{eq:ineq},
\vspace{-2ex}
\small
\begin{align}
\frac{d}{dt}(p_{\max}^{\top}(t_1)x(t))
&\leq p_{\max}^{\top}(t_1)((\tilde{S}\left(t_1\right)B\left(o_{\min}\right)-G\left(o_{\min}\right))x\left(t\right)) \nonumber \\
& = \sigma_{\max}(t_1)p_{\max}^\top (t_1)x\left(t\right), \label{eq:cor}
\end{align}
\vspace{-1ex}
\normalsize
\noindent
which leads to
\[
p_{\max}^{\top}(t_1)x\left(t\right)\leq p_{\max}^{\top}(t_1)x(t_1)e^{(\sigma_{\max}(t_1))t},
\]
for any $t\geq t_1$.
The inequality listed above indicates that, when $\sigma_{\max}(t_1)<0$, the weighted average $p_{\max}^{\top}(t_1)x(t)$ will decrease exponentially fast to zero, $\forall t\geq t_1$, implying that $x(t)$ will decrease exponentially fast to zero. From Lemma~\ref{lem:irr_spe} and
Corollary~\ref{Prop:bounds}, $\sigma_{\max}(t_1)<0$ leads to $R_{\max}(t_1)<1$, which guarantees $R_{o}(t)<1$, $\forall t\geq t_1$. Hence, we have the following corollary, where we define $t_1$ as $t_f$.
\begin{corollary}
\label{cor:c1}
If $R_{\max}(t_f)<1$, there will exist no peak infection time in
$(t_f, \infty)$,
and $p_{\max}^{\top}(t_f)x(t)$ $\forall t\geq t_f$ will monotonically decrease to zero exponentially fast, indicating that the epidemic will die out exponentially fast.
\end{corollary}
\noindent
Corollary \ref{cor:c1} connects $R_o(t)$ to the behavior of the epidemic process. In particular, for $t_f=0$, at the beginning stages of the epidemic, even with every community ignoring the epidemic, we still have $R_{\max}(0)<1$ which means that the epidemic is serious, and will disappear quickly.
In addition to Corollary \ref{cor:c1}, Theorem \ref{thm:1} also implies that, if the effective reproduction number $R_{\min}(t)$ at the beginning stages of the epidemic is greater than $1$, before disappearing, there must exist at least one peak infection time.
This phenomenon is captured by the following corollary.
\begin{corollary}
\label{cor:c2}
If $R_{\min}(0)>1$, then
\begin{enumerate}
\item there will be at least one peak infection time $t_p$;
\item $p_{\min}^{\top}(t_0)x(t)$ will increase exponentially fast before reaching a peak infection time $t_p$;
\end{enumerate}
\end{corollary}
\begin{proof}
1) First we prove that when $R_{\min}(0)>1$, there will be at least one peak infection time $t_p$.
Since $\dot{s}(t)\leq \boldsymbol{0}$, and $s(t)$ is lower-bounded by $\boldsymbol{0}$, we must have
an equilibrium at $t=t_1$ when $\dot{s}(t_1)=\boldsymbol{0}$. Consider the case $s(t_1)=\boldsymbol{0}$ leading to $\dot{s}(t_1)=\boldsymbol{0}$. If $x(t_1)\neq \boldsymbol{0}$, from \eqref{I-O-Final}, we have $\dot{x}(t_1)=-G_o(t_1)x(t_1)< \boldsymbol{0}$, which violates Lemma~\ref{lem:Equi} that $(\boldsymbol{0},\boldsymbol{0},\boldsymbol{1}_n)$ is an equilibrium of the system. Therefore, in order to ensure $x(t)$ converges to $x(t_1)=\boldsymbol{0}$, based on Definition~\ref{def:PT} and Theorem \ref{thm:1},
for $\epsilon>0$, there must exist a moment $t_p$ where $t_p+\epsilon<t_1$, such that $R_o(t_p+\epsilon)<1$. Consider another case that $s(t_1)\neq\boldsymbol{0}$: to ensure $\dot{s}(t_1)=\boldsymbol{0}$, we must have $x(t_1)=\boldsymbol{0}$. Thus, for the same reason, there must exist a time $t_p+\epsilon<t_1$ such that $R_o(t_p+\epsilon)<1$. Additionally, Since $R_{\min}(0)>1$, we can conclude that for both cases, there exists a moment $t_p-\epsilon<t_p$ such that $R_o(t_p-\epsilon)>1$. Therefore, from Theorem \ref{thm:1}, we have proved statement~1).
2) Consider one peak infection time $t_p$, and a time interval $[t_0, t_p-\delta_t]$, for sufficiently small $\delta_t>0$. Based on Lemma~\ref{lem:Metzler}, we have
$\sigma_{\min}(t_0)=\sigma(\tilde{S}(t_0)B(o_{\max})-G^{-1}(o_{\max}))>0$, pairing with the normalized left eigenvector $p_{\min}^{\top}(t_0)$.
Multiplying $p_{\min}^\top(t_0)$ on both sides of \eqref{I-O-Final}, and evaluating the derivative at $t_0$,
\vspace{-1ex}
\begin{align*}
\frac{d}{dt}(p_{\min}^{\top}(t_0)x(t))|_{t=t_0} &
= \sigma_{\min}(t_0)p^{\top}(t_0)x\left(t_0\right),
\end{align*}
$\forall t \in [t_0, t_p-\delta_t]$.
Since
$\sigma_{\min}(t_0)>0$, $p_{\min}^{\top}(t_0)x(t)$ grows exponentially fast
for $t\in[t_0, t_p-\delta_t]$, completing the proof.
\end{proof}
\vspace{-1ex}
By combining Corollaries \ref{cor:c1} and \ref{cor:c2} with Theorem \ref{thm:1},
we can connect
the
behavior of the system
in
\eqref{eq:S-O-C}-\eqref{eq:O-I-C} to the peak infection time of the system in the following theorem.
\begin{theorem}
\label{thm:2}
For the epidemic-opinion model
in
\eqref{eq:S-O-C}-\eqref{eq:O-I-C},
the system will converge to an equilibrium $(s_{e},\boldsymbol{0},(\bar{L}+I_{n})^{-1}(\boldsymbol{1}_{n}-s_{e}))$, and the convergence is exponentially fast.
\end{theorem}
\begin{proof}
By combining Corollaries \ref{cor:c1} and \ref{cor:c2} with Theorem \ref{thm:1}, we can conclude that for any sufficiently small time interval $[t_0,t_1]$, we can find a weighted average of the infected epidemic state changes exponentially fast. Additionally, Corollary \ref{cor:c1} indicates that the epidemic state will converge to zero exponentially fast, after passing $t_f$.
Therefore, We can conclude that the epidemic state $x(t)$ always converges to $\boldsymbol{0}$ exponentially fast. From \eqref{eq:S-O-C}, under the condition that $x(t)$ converges to $\boldsymbol{0}$ exponentially fast, $s(t)$ will converge to an equilibrium point exponentially fast. Consider the convergence of the opinion states $o(t)$, in \eqref{eq:O-I-C}. The linear system
\vspace{-1.5ex}
\[
\dot{o}\left(t\right)=-\left(\bar{L}+I_{n}\right)o\left(t\right)
\]
\vspace{-1.5ex}
\noindent
converges to zero exponentially fast due to the fact that all the eigenvalues of the system matrix $-(\bar{L}+I_{n})$ are smaller than zero. Note that \eqref{eq:O-I-C} is input-to-state stable since the linear system mentioned above has a unique globally stable equilibrium at $o=\boldsymbol{0}$. Therefore,
treating $s(t)$ as an input to \eqref{eq:I-O-C} that
converges to $s_e$ exponentially fast, we have that $o(t)$ converges to $o_e$ exponentially fast\cite{khalil2002nonlinear}. Thus, we have proved the theorem.
\end{proof}
\noindent
Combined with Theorem \ref{thm:1}, Corollaries \ref{cor:c1} and \ref{cor:c2}, Theorem \ref{thm:2} implies that the epidemic will die out eventually,
but the effective reproduction number $R_o(t)$ will determine whether there will be an outbreak or the epidemic will die out directly.
\vspace{-2ex}
\section{Simulation}
\vspace{-1ex}
In the section, we will illustrate the main results developed in this work via simulations. Consider the epidemic coupled
with opinions spreading over ten communities.
The epidemic and opinion spreading network satisfies Assumption \ref{A1}, and share the same graph topology $\mathcal{G}$ as shown in Fig.~\ref{fig_graph}.
\begin{figure}
\begin{centering}
\includegraphics[trim = 7.5cm 7.2cm 4.8cm 6.15cm, clip, width=\columnwidth]{Graph_Structure.jpg}
\par\end{centering}
\centering{}\caption{The graph topology $\mathcal{G}$ of the simulated epidemic and opinion interactions
}
\label{fig_graph}
\end{figure}
\begin{figure}
\begin{centering}
\includegraphics[width=\columnwidth]{S1.jpg}
\par\end{centering}
\centering{}\caption{Typical evolution of epidemics and opinions}
\label{simulation1}
\vspace{-2ex}
\end{figure}
We set the initial condition $x(0)=0.01\times \boldsymbol{1}_n$, $s(0)=0.99\times \boldsymbol{1}_n$, and $o(0)= \boldsymbol{0}$. We also set the parameters $\beta_{\min}=0.2$, $\gamma_{\min}=0.07$, and each $\beta_{ij}$ is uniformly sampled from $[0.2,1]$.
Similarly, each $\gamma_i$ is uniformly sampled from $[0.07, 0.1]$.
We apply only unit edge weights to the opinion graph in all simulations.
Fig. \ref{simulation1}(a) shows that the proportion of the susceptible population in all communities decreases monotonically as claimed in Lemma~\ref{lem:decreasing}, and Fig. \ref{simulation1}(b) shows the evolution of the epidemic states, with the weighted average state $x_w(t)=p^\top(t_p)x(t)$ $\forall t\geq0$ being captured by the dashed line. Note that we use $p^\top(t_p)$ for the entire time interval. Furthermore, the trend of the weighted average of the epidemic states follows the changes of the effective reproduction number $R_o(t)$ in Fig. \ref{simulation1}(d): $x_w(t)=p^\top(t_p)x(t)$ is increasing when $R_o(t)>1$; $x_w(t)=p^\top(t_p)x(t)$ is decreasing when $R_o(t)<1$. Then, $x_w(t)=p^\top(t_p)x(t)$ reaches a local peak when $R_o(t)=1$. Fig. \ref{simulation1} (a)-(d)
illustrate the behavior of the system in \eqref{eq:S-O-C}-\eqref{eq:O-I-C} based on the effective reproduction number $R_o(t)$ and the peak infection time $t_p$, as we proved in Theorem~\ref{thm:1}, Corollaries~\ref{cor:c1} and \ref{cor:c2}. Additionally, Fig. \ref{simulation1}(c) shows that,
at the beginning stages of the epidemic, when no community considers the epidemic as a threat,
as the susceptible population, the beliefs in the seriousness of the epidemic will increase. Meanwhile,
as
the susceptible population decreases and
the opinion states increase, the effective reproduction number $R_o(t)$ decreases, which aligns with Proposition \ref{prop:Spe_r}. As stated in Theorem \ref{thm:2},
the states of the system converge to zero exponentially fast.
\begin{figure}
\begin{centering}
\includegraphics[width=\columnwidth]{S2_final.jpg}
\par\end{centering}
\centering{}\caption{States convergence with wide-spread initial opinions}
\vspace{-2ex}
\label{simulation2}
\end{figure}
Next, we will show the special case
where
the opinions reach consensus. As mentioned in Remark 1, the opinion states will reach consensus at the equilibrium if and only if all the communities have the same infection level.
We set $\beta_{\min}=0.1$ and $\gamma_{\min}=0.14$, while
$\beta_{ij}$ and $\gamma_i$ are uniformly sampled from $[0.1,0.6]$ and $[0.14, 0.30]$, respectively to generate plots in Fig. \ref{simulation2}.
In Fig. \ref{simulation2}(a) and (c), the initial conditions of the epidemic states are uniformly sampled from $[0,1]$. In Fig. \ref{simulation2}(b) we randomly sample the initial opinion states from $[0,1]$, and In Fig. \ref{simulation2}(d)
we set the initial opinion states as $o(0)=\boldsymbol{1}_n$.
Both Fig. \ref{simulation2}(a) and (c)
capture the extreme case where everyone in the population
becomes
infected, i.e., where the susceptible states converge to $s_e=\boldsymbol{0}$. Therefore, based on Lemma \ref{consensus}, the opinion states at the equilibrium will take the form $o_e= \boldsymbol{1}_n- s_e=\boldsymbol{1}_n$, captured by Fig. \ref{simulation2}(b) and (d). The simulations demonstrate that, when reaching agreement after the
epidemic dies out,
the evaluations on the seriousness of the epidemic can reflect the infection level.
Fig. \ref{S3} aims to show that the effective reproduction number $R_o(t)$
may not decrease monotonically, unlike the classical SIR model, as stated before.
We set $\beta_{\min}=0.01$ and $\gamma_{\min}=0.05$, while $\beta_{ij}$ and $\gamma_i$ are uniformly sampled from $[0.01,0.4]$ and $[0.05, 0.1]$, respectively.
We assume initial opinions $o(0)=\boldsymbol{1}_n$ as shown in Fig. \ref{S3}(c), and we sample the initial infected proportion for each community from $[0.3,0.6]$ randomly.
In Fig.~\ref{S3}(d),
since $R_{\min}(0)<1$, the weighted average $p^\top(t_p) x(t)$ decrease at the beginning stages of the outbreak. However,
the communities soon realize the epidemic is not as severe as they have evaluated as captured in Fig. \ref{S3}(c).
As the opinion states decrease, $R_o(t)$ increases, causing the weighted average $p^\top(t_p) x(t)$ to increase again, captured by Fig. \ref{S3}(b)-(d).
In
Fig.~\ref{S3}(d), we observe that there are two peak candidates where
$R_o(t)=1$; we rule out the first by Theorem~\ref{thm:1},
which states that peak infection time must satisfy
$R_{o}(t)>1$, for $t\in [t_0, t_p)$ and $R_{o}(t)<1$, for $t \in (t_p, t_1]$,
for $t_p-t_0>0$ and $t_1-t_p>0$
sufficiently small.
However,
the second
point where
$R_o(t)=1$ is a peak infection time,
consistent with
Fig. \ref{S3}(b) and (d).
Lastly, Fig. \ref{S3}(d) illustrates that $R_o(t)$ is not monotonically decreasing, compared to the effective reproduction number in the classical SIR model~\cite{van2002reproduction}.
\begin{figure}
\begin{centering}
\includegraphics[width=\columnwidth]{S3.jpg}
\par\end{centering}
\centering{}\caption{State dynamics with rebouncing $R_o(t)$}
\vspace{-2ex}
\label{S3}
\end{figure}
\vspace{-1.5ex}
\section{Conclusion}
\vspace{-1ex}
In this work, we develop a networked SIR model coupled with opinion dynamics to study
epidemic spreading processes over
multiple
communities. We
define
the effective reproduction number and peak infection time
to characterize
the
transient behavior of the epidemic. We also
study
the convergence time to the equilibria. Additionally, we discover that the opinion states at the equilibria can reflect the infection level of each community to some degree. The current work can be further extended to study the influence of the structures of the opinion spreading networks on the behavior of the system. Another potential future research direction is to
design control algorithms that influence the opinions
to change the behavior of the epidemic.
\vspace{-1.5ex}
\normalem
\bibliographystyle{IEEEtran}
|
\section{Motivation and results}
The purpose of this article is to initiate the study of the following question: given a closed biquotient $G/\!\!/ H$, is the tangent bundle $T(G/\!\!/ H)$ a biquotient vector bundle? Our motivation comes from the fact that a positive answer automatically endows $T(G/\!\!/ H)$ with a Riemannian metric of non-negative sectional curvature, whereas, with the current knowledge, a negative answer generally leaves open whether $T(G/\!\!/ H)$ admits such a metric.
The class of biquotients $G/\!\!/ H$ generalizes that of homogeneous spaces $G/H$. A common feature is that there is an associated $H$-principal bundle $H\to G\to G/\!\!/ H$. Each representation $V$ of $H$ naturally induces a vector bundle $G\times_H V$ over $G/\!\!/ H$. A vector bundle constructed in such a way will be called a \emph{biquotient (vector) bundle}. We refer to Section~\ref{sec:structure} for the precise definitions. Throughout this article, all Lie groups will be understood to be compact.
Given a biquotient $G/\!\!/ H$, one can find many other biquotient presentations $G'/\!\!/ H'$ with $G/\!\!/ H$ diffeomorphic to $G'/\!\!/ H'$. However, when we refer to biquotient bundles over the presentation $G/\!\!/ H$ we always mean a bundle of the form $G \times_H V$.
As is well known, the tangent bundle of a homogeneous space $G/H$ is isomorphic to a biquotient vector bundle $G\times_H V$ (see e.g. \cite[Theorem 18.16]{Michor}). We show that this no longer holds in general for other presentations.
\begin{theorem}\label{thm:presentations_of_HP}
The complex projective space $\mathbb{C}P^k$ for $k\geq 2$ and the quaternionic projective space $\mathbb{H}P^k$ for $k\geq 1$ both admit biquotient presentations $G/\!\!/ H
with $G$ simply connected such that the tangent bundle $T\mathbb{C}P^k$ or $T\mathbb{H}P^k$
is not isomorphic to any biquotient vector bundle $G\times_{H}V$ with respect to the presentation $G/\!\!/ H$.
\end{theorem}
The assumption that $G$ is simply connected in Theorem \ref{thm:presentations_of_HP} allows for the largest possible class of biquotient vector bundles; see Proposition \ref{prop:cover}.
As each of the manifolds in Theorem \ref{thm:presentations_of_HP} admits a presentation for which the tangent bundle is a biquotient bundle (one can take any homogeneous presentation), it is natural to ask whether every biquotient has some presentation for which the tangent bundle is a biquotient vector bundle. We find infinitely many examples indicating a negative answer to this question.
\begin{theorem}\label{thm:first_theorem}
In each dimension $4,6$ and $16n + 4$ with $n\geq 1$ there is a simply connected closed biquotient $M = G/\!\!/ H$ whose tangent bundle is not isomorphic to a biquotient vector bundle $G\times_H V$ for any presentation of $M$ as a biquotient. Moreover, in dimension $6$ there are infinitely many such examples of pairwise distinct homotopy type.
\end{theorem}
The examples found in dimensions $16n+4$ are rational quaternionic projective spaces of the form $Spin(4k+1)/\!\!/ (Spin(4k-1)\times SU(2))$ with $k\geq 3$ odd, and the proof involves an analysis of the representations of the group $Spin(4k-1)\times SU(2)$ together with the non-existence of almost complex structures on these spaces. The non-existence follows from an unpublished result of Hirzeburch; Michael Albanese has provided a proof of this result in Appendix \ref{appendix}. These biquotients were discovered by Eschenburg \cite{Es84}, and Kapovitch and Ziller showed in \cite{KZ04} that they have the same integral cohomology groups as $\mathbb{H} P^{2k-1}$ but a different ring structure.
The examples in dimensions $4$ and $6$ are of the form $G/\!\!/ T$ where $G$ is simply connected and $T$ is a torus with $\operatorname{rank} T =\operatorname{rank} G$, and include the connected sum $(S^3)^2\bq T^2\approx\mathbb{CP}^2\sharp\mathbb{CP}^2$ discovered by Totaro and Eschenburg's positively curved inhomogeneous $6$-manifold $SU(3)/\!\!/ T^2$. The proof uses the theory of characteristic classes. As
Hepworth proved in his thesis \cite[Proposition~5.3.9]{He05}, for any biquotient of the form $G/\!\!/ T$ (with $G$ not necessarily being simply connected nor $T$ satisfying $\operatorname{rank} T=\operatorname{rank} G$), the tangent bundle is stably isomorphic to a biquotient vector bundle. We do not know whether Hepworth's result holds for arbitrary biquotients, however we are able to give the following partial answer which only depends on the topology rather than on the presentation.
\begin{theorem}\label{thm:sufficient_conditions_stably_biquotient}
Let $M = G/\!\!/ H$ be a closed biquotient. If $\oplus_{i>0} H^{4i}(M,\mathbb{Q})=0$, then the tangent bundle $TM$ is stably isomorphic to a biquotient vector bundle $G\times_H V$ for some $H$-representation $V$.
\end{theorem}
Theorem~\ref{thm:sufficient_conditions_stably_biquotient} follows as an easy combination of various results. We use work of Singhof \cite{Si93} together with our results from Section~\ref{sec:structure} to conclude that the tangent bundle is stably isomorphic to the Whitney sum of a biquotient bundle and the inverse of another biquotient bundle. The topological condition in Theorem~\ref{thm:sufficient_conditions_stably_biquotient} was found by the authors in the recent paper \cite{DG21} to ensure that every biquotient bundle has an inverse which is also a biquotient bundle.
Next, we analyze the tangent bundle of simply connected closed biquotients $G/\!\!/ H$ of dimension at most $5$, which were classified by the first author and Pavlov \cite{De14, Pa04}. Recall that such a biquotient is diffeomorphic to one of the following spaces: $S^2$, $S^3$, $S^4$, $\mathbb{C} P^2$, $S^2\times S^2$, $\mathbb{C}P^2\#\pm\mathbb{C}P^2$, $S^5$, $S^2\times S^3$, the Wu space $SU(3)/SO(3)$ or the non-trivial $S^3$-bundle over $S^2$, denoted $S^3\,\widehat{\times}\, S^2$.
\begin{theorem}\label{thm:dims235}
Let $M$ be a simply connected closed biquotient of dimension at most $5$. Then the following hold:
\begin{itemize}
\item $M$ has at least one presentation $G/\!\!/ H$ for which $TM$ is isomorphic to a biquotient bundle $G\times_H V$ if and only if $M$ is not diffeomorphic to $\mathbb{C}P^2\#\mathbb{C}P^2$.
\item If $M = S^4$ or $\mathbb{C}P^2$, then $M$ has another presentation $G/\!\!/ H$ with $G$ simply connected for which $TM$ is not a biquotient vector bundle $G\times_H W$.
\item If $M$ is not diffeomorphic to any of $\mathbb{C}P^2\#\mathbb{C}P^2$, $S^4$ or $\mathbb{C}P^2$, then $TM$ is isomorphic to a biquotient vector bundle $G\times_H W$ for all presentations $G/\!\!/ H$ of $M$ with $G$ simply connected.
\end{itemize}
\end{theorem}
Our main interest in biquotient vector bundles $G\times_H V$ is that they carry metrics of non-negative (sectional) curvature, induced from any bi-invariant metric on $G$ and any orthogonal metric on $V$. Whether the tangent bundle of a closed biquotient admits a metric of non-negative curvature is a particular instance of the so-called Converse Question to the Soul Theorem of Cheeger and Gromoll \cite{CG72}, asking which vector bundles over a non-negatively curved closed manifold admit a complete metric of non-negative curvature.
Given the very limited number of existing tools to endow non-compact manifolds with non-negatively curved metrics, Theorem~\ref{thm:first_theorem} indicates that new ideas are required to construct such a metric on the tangent bundle of an arbitrary biquotient.
On the other hand, the positive results of Theorem~\ref{thm:sufficient_conditions_stably_biquotient} and \ref{thm:dims235} immediately yield, respectively, the following corollaries.
\begin{corollary}
For each closed biquotient $M = G/\!\!/ H$ with $\oplus_{i>0} H^{4i}(M,\mathbb{Q})=0$ there is some $k$ for which the product manifold $TM\times\mathbb{R}^k$ admits a metric of non-negative sectional curvature.
\end{corollary}
\begin{corollary}\label{cor:dims2345}
Let $M$ be a simply connected closed biquotient of dimension at most $5$. Then $TM$ admits a metric of non-negative sectional curvature, except possibly when $M$ is diffeomorphic to $\mathbb{C}P^2\#\mathbb{C}P^2$.
\end{corollary}
The inhomogeneous examples of Corollary~\ref{cor:dims2345} are $\mathbb{C}P^2\# \pm \mathbb{C}P^2$ and $S^3\, \widehat{\times}\, S^2$. When $M$ is diffeomorphic to $\mathbb{C}^2\#-\mathbb{C}P^2$, the fact that $TM$ admits a non-negatively curved metric can be proved using the results of \cite{GroveZiller}. When $M$ is $S^3\,\widehat{\times}\, S^2$, the fact that $TM$ admits a non-negatively cured metric seems to be new.
\textbf{Organization of the article. }In Section~\ref{sec:structure} we begin with some background information and then we prove some structure results regarding biquotient vector bundles. Specifically, we show that the largest class of biquotient vector bundles is obtained when $G$ is simply connected (Proposition~\ref{prop:cover}), and that the so-called \textit{reduced} presentation of a biquotient already gives rise to all biquotient vector bundles (Proposition~\ref{prop:reduced}). Armed with the structure results, we prove Theorem~\ref{thm:sufficient_conditions_stably_biquotient} at the end of Section 2. Section~\ref{S:examples_spin} is devoted to proving Theorem~\ref{thm:presentations_of_HP} and Theorem~\ref{thm:first_theorem} for dimensions of the form $16n+4$. Section~\ref{S:examples_torus} contains the proof of Theorem~\ref{thm:first_theorem} for dimensions $4$ and $6$. In Section~\ref{S:low_dims} we prove Theorem~\ref{thm:dims235}. During the proof of Theorem~\ref{thm:first_theorem}, an unpublished result of Hirzebruch is used. Towards closing this gap in the literature, Michael Albanese has supplied a proof of this result in Appendix~\ref{appendix}.
\textbf{Acknowledgements. } We thank Richard Hepworth for helpful conversations. We would also like to thank Manuel Amann, Michael Albanese and Wolfgang Ziller for helpful comments on an earlier draft of this article.
\section{Background and the set of biquotient vector bundles}\label{sec:structure}
\subsection{Definitions and notation}\label{SS:definitions}
Let us first recall the definition of biquotient, following the approach by Totaro in \cite[Lemma~1.1~(3)]{To02}. Let $G$ be a compact Lie group and $Z(G)$ its center, and let $Z\subseteq G\times G$ be the diagonal normal subgroup $Z\defeq \{(g,g) : g\in Z(G)\}$. Any homomorphism $f\colon H\to (G\times G)/Z$ can be written in the form $f(h)=[f_1(h),f_2(h)]$ and determines a well-defined two-sided $H$-action $\star$ on $G$ by the rule $h\star g=f_1(h)gf_2(h)^{-1}$. When this action is free, the orbit space, denoted by $G/\!\!/ H$, inherits a manifold structure and is called a \emph{biquotient}.
By construction, associated to a biquotient $G/\!\!/ H$ there is a principal $H$-bundle $H\to G\to G/\!\!/ H$. Each representation $V$ of $H$ induces a vector bundle $G\times_H V$ over $G/\!\!/ H$, defined as the quotient of $G\times V$ via the diagonal action by $H$ consisting of the $\star$-action on $G$ and the representation action on $V$; the projection map is given by $[g,v]\mapsto [g]$. A vector bundle constructed in such a way will be called a \emph{biquotient (vector) bundle}. A biquotient bundle is real or complex depending on the nature of the representation $V$.
As already mentioned in the introduction, given a biquotient $G/\!\!/ H$, one can find other biquotients $G'/\!\!/ H'$ with $G/\!\!/ H$ diffeomorphic to $G'/\!\!/ H'$. However, when we refer to biquotient bundles over $G/\!\!/ H$ we always mean a bundle of the form $G \times_H V$.
When the homomorphism $f\colon H\to (G\times G)/Z$ is given by a subgroup inclusion $H\subseteq \{e\}\times G\subseteq G\times G$, the resulting space is homogeneous and denoted by $G/H$.
More generally, when $H$ has the form $H=H_1\times H_2$ and the homomorphism $f\colon H\to (G\times G)/Z$ is given by subgroup inclusions $H_i\subseteq G$, we might denote the resulting space by $H_1\backslash G/H_2$ instead of $G/\!\!/ (H_1\times H_2)$.
\subsection{Structure results}
Our first structure result (Proposition~\ref{prop:cover} below) indicates that if we pull the biquotient struture $G/\!\!/ H$ back along a
covering $\pi : G'\rightarrow G$, the set of biquotient vector bundles can only increase.
We begin by noting that the product map $\pi\times \pi:G'\times G'\rightarrow G\times G$ descends to a covering $\psi:(G'\times G')/Z'\rightarrow (G\times G)/Z$, with $Z'$ denoting the diagonal center of $G'\times G'$. We include a proof for completeness.
\begin{proposition} The map $\psi$ is a covering map.
\end{proposition}
\begin{proof} Consider the diagram \begin{diagram} \ker (\pi \times \pi) & \rTo& \ker \psi\\ \dTo & & \dTo\\ G'\times G' &\rTo^{\eta} & (G'\times G')/Z'\\ \dTo_{\pi\times\pi} & & \dTo_{\psi} \\
G\times G&\rTo& (G\times G)/Z
\end{diagram}
We claim that $\eta|_{\ker (\pi\times \pi)}:\ker (\pi\times \pi)\rightarrow \ker \psi$ is a surjective map of Lie groups. Believing this claim, since $\ker (\pi\times \pi)$ is discrete, it will follow that $\ker \psi$ is discrete, so $\psi$ is a covering
We first argue that $\eta(\ker (\pi \times \pi))\subseteq \ker \psi$. To see this, simply note that if $(a,b)\in \ker (\pi \times \pi)$, then commutativity shows that $\psi(\eta(a,b))$ is the identity element, so $\eta(a,b)\in \ker \psi$.
We next argue that $\eta|_{\ker (\pi \times \pi)}:\ker (\pi\times \pi)\rightarrow \ker \psi$ is surjective. To see this, let $(a,b)Z'\in \ker \psi$. Thus, $(\pi\times \pi)(a,b)\in Z$. Since $\pi \times \pi$ is surjective, it maps $Z'$ surjectively onto $Z$, so there is an element $(z',z')\in Z'$ with $(\pi\times \pi)(a,b) = (\pi\times \pi)(z',z')$. Then the element $(a,b)(z',z')^{-1}\in \ker (\pi\times \pi)$ and $\eta((a,b)(z',z')^{-1}) = (a,b)Z'$.
\end{proof}
Next set $H' = \psi^{-1}(f(H))$, where $f:H\to (G\times G)/Z$ denotes the homomorphism defining the biquotient $G/\!\!/ H$. Since both $f,\psi$ are group homomorphisms, it follows that $H'$ is a subgroup of $(G'\times G')/Z'$, and the corresponding inclusion map induces a free biquotient action of $H'$ on $G'$, denoted by $\star'$. Moreover, $\pi$ induces a diffeomorphism $G'/\!\!/ H'\rightarrow G/\!\!/ H$ (see \cite[Lemma~3.1]{To02} for more details). With this in mind, we have the following proposition.
\begin{proposition}\label{prop:cover} Let $G/\!\!/ H$ be a biquotient, $\pi:G'\rightarrow G$ a covering map, and $H'$ as defined above so that $G/\!\!/ H$ is diffeomorphic to $G'/\!\!/ H'$. Then, every biquotient bundle $G\times_H V$ is isomorphic to a biquotient bundle of the form $G'\times_{H'} W$.
\end{proposition}
\begin{proof}
Let $\rho:= \psi|_{H'}:H'\rightarrow H$ denote the projection. Given a representation $V$ of $H$, we let $W$ be the same vector space endowed with an action by $H'$ by precomposing the $H$ action with $\rho$. That is, $h'\in H'$ acts via the action $\bullet'$ defined by the rule $h'\bullet' v = \rho(h')\bullet v$, where $\bullet$ denotes the $H$ action on $V$.
Then we claim that the map $\pi\times Id_V:G'\times W\rightarrow G\times V$ descends to a bundle isomorphism $\phi:G'\times_{H'} W\rightarrow G\times_H V$. Note first that $\phi$ covers the diffeomorphism $ G'/\!\!/ H'\rightarrow G/\!\!/ H$ induced by $\pi:G'\rightarrow G$, so $\phi$ is a bundle map, if it is well defined.
We observe that $\phi$ is well defined since:
\begin{align*}\phi(h'(g',v)H')&= \phi((h'\star' g', h'\bullet' v)H')\\ &= ((\pi\times Id_V)(h'\star'g', h'\bullet' v))H \\\ &= (\rho(h')\ast\pi(g'), \rho(h')\bullet v)H\\ &= \rho(h')(\pi(g'),v)H\\ &= (\pi(g'),v)H \\ &= \phi((g',v)H').\end{align*}
Because the composition $G'\times W\rightarrow G\times V\rightarrow G\times_H V$ is a surjective submersion, the same is true of $\phi$. Thus, we only need to show it is injective.
To that end, first note that if $\phi( (g_1^\prime, v_1) H') = \phi((g_2^\prime,v_2)H')$, then $(g_1^\prime,v_1)H'$ and $(g_2^\prime,v_2)H'$ must be in the same fiber above $G'/\!\!/ H'$. In particular, there is an $h'\in H'$ with $h'\ast g_1^\prime = g_2^\prime$. Applying $h'$ to $(g_1^\prime,v_1)$, we obtain a new representative $(g_2^\prime, w)H'$ of the orbit through $(g_1^\prime,v_1)$.
Now, $\phi((g_2^\prime, w)H') = \phi((g_2^\prime,v_2)H')$, so $(\pi(g_2^\prime), w)H = (\pi(g_2^\prime), v_2)H$. Thus, there is an $h\in H$ with $h(\pi(g_2^\prime), w) = (\pi(g_2^\prime), v_2)$. In particular, $h\ast \pi(g_2^\prime) = \pi(g_2^\prime)$. Since $H$ action on $G$ is free, $h = e$. In particular, $w = v_2$, so $\phi$ is injective.
\end{proof}
It is natural to wonder about the converse operation - passing from biquotient vector bundles on $G'/\!\!/ H'$ to biquotient vector bundles on $G /\!\!/ H$. The following proposition shows that this is not always possible.
\begin{proposition} The set of biquotient vector bundles over $S^2 = SO(3)/SO(2)$ is a proper subset of the biquotient vector bundles over $S^2 = Spin(3)/Spin(2) = SU(2)/U(1)$.
\end{proposition}
\begin{proof} We claim that the rank $2$ vector bundle over $S^2$ with Euler class $\pm 1$ is a biquotient vector bundle using the presentation $S^2 = SU(2)/U(1)$, but not when using the presentation $S^2 = SO(3)/SO(2)$.
Consider first the presentation with $S^2 = SU(2)/U(1)$. Consider the $V= \mathbb{R}^2$ with $U(1)$ acting as a unit speed rotation. We claim that $SU(2)\times_{U(1)} V$ has Euler class $\pm 1$. To see this, restrict $V$ to the unit sphere $S^1\subseteq V$. The action of $U(1)$ on $S^1$ is simply transitive, so we find that $SU(2)\times_{U(1)} S^1\cong SU(2)$. From the Gysin sequence applied to the bundle $S^1\rightarrow SU(2)\rightarrow S^2$, we see $e = \pm 1$.
Next, consider the presentation $S^2 = SO(3)/SO(2)$. The representations of $SO(2)$ on $V\cong \mathbb{R}^2$ are all given by rotation by some speed $k$. If $k = 0$, we obviously get the trivial bundle which has Euler class $0$. So, assume $k\neq 0$. Now, consider the unit sphere bundle $SO(3)\times_{SO(2)} S^1$. Because $k\neq 0$, the $SO(2)$ action on $S^1$ is transitively with stabilizer the $k$-th roots of $1$. But then, from \cite[Lemma 1.3]{KZ04}, we find that $SO(3)\times_{SO(2)}S^1\cong SO(3)/\mathbb{Z}_k$. In particular, the unit sphere bundle cannot be simply connected
\end{proof}
\begin{remark} In fact, it is easy to see that every vector bundle over $S^2 = SU(2)/S^1$ is a biquotient (homogeneous) vector bundle. In contrast, using the presentation $S^2 = SO(3)/SO(2)$, one only gets those with vanishing second Stiefel-Whitney class (evidently including $TS^2$).
\end{remark}
Recall that every compact Lie group $G$ has a cover of the form $G'\times T$ where $G'$ is simply connected and $T$ is a torus. Because of Proposition \ref{prop:cover}, in order to consider the largest possibly class of biquotient vector bundles, we will always consider the case where $G$ has this form.
In \cite[Lemma~3.3]{To02}, Totaro shows that every biquotient $M=G/\!\!/ H$ has a \textit{reduced} form. Specifically, if the restriction of the $H$ action to a normal subgroup $\tilde{H}\subseteq H$ acts transitively on any simple factor of $G$, then one can "cancel" these factors of $G$ and $H$ to find a presentation $\hat G/\!\!/ \hat H$ of $M$ with both $\hat G$ and $\hat H$ of smaller dimension respectively. We will refer to a biquotient $G/\!\!/ H$ as \textit{reduced} if no normal subgroup of $H$ acts transitively on any simple factor of $G$. The next proposition indicates that the reduced form of a biquotient gives the largest possible class of biquotient vector bundles.
\begin{proposition}\label{prop:reduced} Consider a homomorphism $H\to ((G_1\times G_2)\times (G_1\times G_2))/Z$ and suppose that the corresponding biquotient action of $H$ on $G_1\times G_2$ is free. Suppose the action is non-reduced in the sense that the projection of the $H$ action to $G_2$ is transitive. Let $\hat H$ denote the isotropy group at $e\in G_2$ of this transitive action. Then, the projection of $\hat H$ to $G_1$ gives a free action, and $(G_1\times G_2)/\!\!/ H$ is a canonically diffeomorphic to $G_1/\!\!/ \hat H$. Further, any biquotient vector bundle $(G_1\times G_2)\times_{H} V$ is isomorphic to a biquotient vector bundle $G_1\times_{\hat H} W$ for some appropriate $W$
\end{proposition}
\begin{proof} The statement about the isotropy group acting freely is simply \cite[Lemma 1.3]{ KZ04}, but we prove it because the proof will be useful for establishing the rest of this proposition. We use the notation $\ast$ for the $H$ action on $G_1\times G_2$, and $\ast_1$ for the projection of the $\hat H$ action to $G_1$.
Suppose $h\in \hat H\subseteq H$ and there is $g_1\in G_1$ for which $ h\ast_1 g_1 = g_1$. Because $h\in \hat H$, $h\ast(g_1,e) = (g_1,e)$. Since the action $\ast$ is free, we must have $h=e$, so the $\ast_1$ action is free.
For the statement about diffeomorphism types, consider the function $\phi:G_1/\!\!/ \hat H\rightarrow (G_1\times G_2)/\!\!/ H$ given by $\phi(g\hat H) = (g,e)H$. This is well defined since, for $h\in \hat H$, we have $\phi( h\ast_1 g \hat H ) = (h\ast_1 g, e)H = h\ast(g,e) H = (g,e)H$. Also, if $\phi( g\hat H) = \phi(g'\hat H)$, then $(g,e)H = (g',e)H$, so there is an $h\in H$ with $h\ast(g,e) = (g',e)$. But then $h\in \hat H$ and $h\ast_1 g= g'$, so $\phi$ is injective. To see it is surjective, consider an arbitrary point $(g_1,g_2)H\in (G_1\times G_2)/\!\!/ H$. Since $H$ acts transitively on $G_2$, there is an $h\in H$ for which $h\ast(g_1,g_2)$ has the form $(g_1,e) = \phi(g_1 \hat H)$. Also, $\phi$ is clearly smooth. One can check $\phi$ is constant rank without too much difficulty.
We now turn our attention to the statement about biquotient bundles. Let $V$ be any representation of $H$. Let $W$ denote $V$ with the action restricted to $\hat H$. Then we claim that $\phi\times Id_W$ induces an isomorphism $(G_1\times W)/ \hat H\rightarrow (G_1\times G_2\times V)/ H$. In fact, we can use the proof above with $G_1$ replaced by $G_1\times W$. Thus, we immediately conclude that $\phi\times Id_W$ is a diffeomorphism. It remains to see that it respects the bundle projections.
For $(G_1\times W)/\hat H$, the bundle projection is $\pi'((g_1,w)\hat H) = g_1 \hat H$. For $(G_1\times G_2 \times V)/H$, the bundle projection is $\pi((g_1,g_2,v)H)=(g_1,g_2)H$. Then we see that $(\pi \circ (\phi\times Id_W))((g_1,w)\hat H) = \pi((g_1,e,w)H) = (g_1,e)H$. On the other hand, $(\phi\circ \pi')((g_1,w)\hat H) = \phi(g_1 \hat H) = (g_1, e)\hat H$. Thus, $\pi\circ (\phi\times Id_W) = \phi \circ \pi'$, so the diffeomorphism respects the projections.
\end{proof}
\subsection{Alternative definition of biquotients}
In the literature, biquotients have an alternative definition where a homomorphism $H\rightarrow G\times G$ is used instead of a homomorphism to $(G\times G)/Z$. In this alternative definition, one is allowed more generally to consider effectively free actions, say with ineffective kernel $K \subseteq H$. Given a general representation $V$ of $H$, the space $G\times_H V$ need not be a smooth manifold owing to the fact that the isotropy group at a point in the zero-section contains $K$, while the isotropy group at a non-zero vector is in general smaller. Because of this, when using the alternative definition, one must restrict to representations of $H$ for which $K$ acts trivially.
In other words, in this alternative definition we do not, in general, get an $H$-principal bundle $H\to G\to G/\!\!/ H$ but rather an $H/K$-principal bundle $H/K\to G\to G/\!\!/ H$; thus it only make sense to form vector bundles using representations of $H/K$.
In order to compare which vector bundles one gets using either of the definitions we find, at least when $Z\cong Z(G)$ is finite, that the set of isomorphism classes of vector bundles over $G/\!\!/ H$ is independent of whether the action is defined by a homomorphism $H\rightarrow G\times G$ acting effectively freely or by a homomorphism $H\rightarrow (G\times G)/Z$ acting freely.
\begin{proposition} Suppose $Z\cong Z(G)$ is finite and $f: H\rightarrow (G\times G)/Z$ defines a free biquotient action. Let $H'$ denote the cover of $H$ corresponding to the subgroup $(f_\ast)^{-1}(\pi_1( (G\times G)/Z))$ of $\pi_1(H)$, so that there is a lift $\tilde{f}:H'\rightarrow G\times G$ and let $K$ denote the kernel of the projection $\pi :H'\rightarrow H$. Then every biquotient vector bundle of the form $G\times_H V$ is isomorphic to one of the form $G\times_{H'} W$ where $K\subseteq H'$ acts on $W$ trivially. Conversely, for any bundle of the form $G\times_{H'} W$ where $K$ acts trivially on $W$, there is an isomorphic bundle of the form $G\times_H V$.
\end{proposition}
\begin{proof} We use $\tilde{f}$ to identify $H'$ as a subset of $G\times G$. We first claim that $K$ is precisely the ineffective kernel of the $H'$ action on $G$. To see this, let $(h_1,h_2)\in K\subseteq H'\subseteq G\times G$. Then $(h_1,h_2)$ is a lift of the point $\pi(h_1,h_2) = e\in H$. Since $e$ acts trivially on $G$, $(h_1,h_2)$ must act trivially on $G$. Conversely, assume the point $(h_1,h_2)\in H'$ acts trivially on $G$. Since $(h_1,h_2)$ is a lift of $\pi(h_1,h_2)$, and the $H$ action on $G$ is free, we conclude that $\pi(h_1,h_2) = e\in H$. That is, $(h_1,h_2)\in K$.
Now, suppose $W$ is an $H'$ representation and that $K$ acts trivially. Then $K$ acts trivially on $G\times W$, so $G\times_{H'} W = G\times_{H'/K} V\cong G\times_H W$. Thus, to prove the proposition, given $W$, simply let $V = W$ with $H$ acting via the canonical isomorphism $H\cong H'/K$. Conversely, given $V$, let $W = V$ with $H'$ acting via its projection to $H$.
\end{proof}
\subsection{The stable class of the tangent bundle}
Let $G/\!\!/ H$ be a closed biquotient defined by a homomorphism $f$. As shown by Eschenburg, $G/\!\!/ H$ can be given the presentation $G\backslash(G\times G)/H$, where $G$ (resp. $H$) is embedded in $G\times G$ diagonally (resp. via the homomorphism $f$).
Let $\alpha_H$ denote the biquotient vector bundle (with respect to the presentation $G/\!\!/ H$) $G\times_H \mathfrak{h}$ with $\mathfrak{h}$ denoting the Lie algebra of $H$, with $H$ acting via the adjoint action. Let $\alpha_G$ denote the vector bundle $((G\times G)/H) \times_G \mathfrak{g}$, defined similarly. Building upon work of Singhof \cite{Si93}, Kerin shows in \cite[Lemma~6.2]{Ke11} that the tangent bundle $T(G/\!\!/ H)$ of a biquotient has the property that
$$T(G/\!\!/ H)\oplus \alpha_H\cong \alpha_G.$$
Rewriting $\alpha_G$ as $(G\times G)\times_{G\times H} \mathfrak{g}$, where the $H$ factor of $G\times H$ acts trivially on $\mathfrak{g}$, we see that it can be regarded as a biquotient vector bundle with respect to the presentation $G\backslash(G\times G)/H$. By Proposition~\ref{prop:reduced}, the bundle $\alpha_G$ is isomorphic to a biquotient vector bundle of the form $G\times_H \mathfrak{g}$ for some action of $H$ on $\mathfrak{g}$.
\begin{proof}[Proof of Theorem~\ref{thm:sufficient_conditions_stably_biquotient}]
It follows from the discussion above that $T(G/\!\!/ H)$ is stably isomorphic to $\alpha_G \oplus E$, where $E$ denotes any inverse of $\alpha_H$ (i.e.~any vector bundle such that the sum $E\oplus \alpha_H$ is isomorphic to a trivial vector bundle). From \cite[Theorem~1.3]{DG21} we know that the topological assumption implies that there an inverse $E_0$ which is isomorphic to a biquotient vector bundle $G\times_H V$. Since the map taking $H$-representations to biquotient vector bundles over $G/\!\!/ H$ is closed under taking Whitney sums, it follows that $\alpha_G \oplus E_0$ is a biquotient vector bundle.
\end{proof}
\begin{remark}
Suppose $G/\!\!/ H = G/H$ is homogeneous. When using Proposition~\ref{prop:reduced} to rewrite $\alpha_G=((G\times G)/H) \times_G \mathfrak{g}=(G\times (G/H))\times_G \mathfrak{g}$ as a bundle of the form $G\times_H \mathfrak{g}$, it turns out that the action of $H$ on $\mathfrak{g}$ is trivial. Indeed, $\alpha_G$ is in this case a trivial vector bundle. This should not be any surprise, as $G\times_H \mathfrak{h}$ is well known to be an inverse for the tangent bundle $T(G/H)$ (which equals $G\times_H \mathfrak{p}$, where $\mathfrak{p}$ denotes a complement of the Lie algebra of $H$ in $G$ endowed with the isotropy action of $H$, see e.g.~\cite[Section~18.16]{Michor}).
\end{remark}
\section{Examples using representation theory}\label{S:examples_spin}
The goal of this section is to find examples of biquotients for which the tangent bundle is not a biquotient vector bundle. The examples of this section are of the form $H_1\backslash G/H_2$ for certain subgroup inclusions $H_i\subseteq G$, see Subsection~\ref{SS:definitions}. We first fix some notation for the rest of the section.
We let $\Delta S^1\subseteq Spin(2n+2)$ be the image of the lift of the composition $S^1\rightarrow SO(2)\rightarrow SO(2n+2)$, where the first map is a double cover, and the second map is the block diagonal embedding $B\mapsto \operatorname{diag}(B,B, ...,B)$.
We denote by $\Delta SU(2)\subseteq Spin(4n+4)$ the image of the lift of the composition $SU(2)\rightarrow SO(4)\rightarrow SO(4n+4)$. Here the first map $SU(2)\rightarrow SO(4)$ is the natural inclusion, i.e. the one induced by regarding $\mathbb{C}^2$ as a real vector space isomorphic to $\mathbb{R}^4$. The second map $SO(4)\rightarrow SO(4n+4)$ is the block diagonal embedding $A\mapsto \operatorname{diag}(A,A,...,A)$.
Finally, a spin subgroup $Spin(k)$ of a spin group $Spin(k+m)$ will always be understood to be embedded via the lift of the standard block inclusion $A\mapsto \operatorname{diag}(A,1,...,1)$.
We shall begin with the examples of Theorem~\ref{thm:presentations_of_HP}. As shown by Eschenburg \cite[Table 101]{Es84}, the previous inclusions induce well-defined biquotients $\Delta S^1\backslash Spin(2n+2)/Spin(2n+1)$ and $\Delta SU(2)\backslash Spin(4n+4)/Spin(4n+3)$ that are diffeomorphic to $\mathbb{C}P^n$ and $\mathbb{H} P^n$ respectively. Let $\mathbb{K} \in \{\mathbb{C}, \mathbb{H}\}$. Except for the case of $\mathbb{C}P^1$, we will show that $T\mathbb{K}P^n$ is not a biquotient bundle for the above presentations. A more precise version of Theorem~\ref{thm:presentations_of_HP} is given by the following result.
\begin{theorem}\label{thm:pres_projective_spaces}
Consider the presentations $\mathbb{C}P^n \cong \Delta S^1\backslash Spin(2n+2)/Spin(2n+1)$ with $n\geq 2$ and $\mathbb{H}P^n\cong \Delta SU(2)\backslash Spin(4n+4)/Spin(4n+3)$ with $n\geq 1$. Then $T\mathbb{K}P^n$ is not a biquotient bundle with respect to these presentations.
\end{theorem}
Before giving its proof, we provide two lemmas that will be useful throughout this section.
First, we shall use the following well-known facts about the representations of the group $Spin(2n+1)$. Following \cite{BtD85}, let $\Lambda=\Lambda^1$ be the standard representation, let $\Lambda^i$ the $i$-th exterior product, and let $\Delta$ be the half-spin representation.
\begin{lemma}\label{LEM:rep_of_spin}
The non-trivial irreducible representation of $Spin(2n+1)$ with $n\geq 1$ of smallest (real) dimension is the standard one $\Lambda^1$, which is of (real) dimension $2n+1$. The non-trivial irreducible representation of $Spin(4n-1)$ with $n$ odd and $n\geq 3$ of second smallest (real) dimension is: $\Delta$ if $n=3$ (which is of real dimension $32$) and $\Lambda^2$ if $n\geq 5$ (which is of dimension $8n^2-6n+1$).
\end{lemma}
\begin{proof}
The fundamental irreducible representations are exactly $\Lambda^i$ with $ n\geq i\geq 1$ and $\Delta$. It follows from the Weyl dimension formula that the smallest dimensional representations of a simply connected compact Lie group are found among its fundamental representations.
Since $\dim \Lambda^i={2n+1 \choose i}\geq 2n+1$, it follows that $\dim \Lambda^{i}\leq \dim \Lambda^{i+1}$ for all $i$. Thus, in order to find the smallest and second smallest representations we are left to decide between $\Lambda^1$, $\Lambda^2$ and $\Delta$.
From \cite[Lemma~6.3, p.280]{BtD85}, we know that the representation $\Delta$ of $Spin(2n+1)$ has complex dimension $2^n$. Thus $\Delta$ is of smaller (real) dimension than $\Lambda$ if and only if the following two conditions are satisfied: $n=2$ and $\Delta$ is of real type. However, the latter fails to hold, since the representation $\Delta$ of $Spin(5)$ is well known to be not of real type but of quaternionic type \cite[Proposition~6.19, p.290]{BtD85}. This proves the first part of the statement.
As for the second smallest representation of $Spin(4n-1)$, we have to compare the (real) dimensions of $\Lambda^2$ and $\Delta$. The former equals $8n^2-6n+1$, which is smaller than $2^{2n-1}$ (i.e. the complex dimension of $\Delta$) if and only if $n\geq 5$. In the case of $n=3$ the representation $\Delta$ of $Spin(11)$ is of complex dimension $32$, however (see \cite[Proposition~6.19, p.290]{BtD85}) it is of real type and hence it can be considered as a real representation of (real) dimension $32$.
\end{proof}
Second, we will use the following bundle-theoretical result.
\begin{lemma}\label{LEM:euler_class_odd}
Let $E$ be an orientable vector bundle over a closed manifold $M$ such that all cohomology groups $H^i(M)$ are torsion-free. If $E$ splits as $E=E_1\oplus E_2$ with the subbundle $E_1$ of odd rank, then the Euler class $e(E)$ vanishes.
\end{lemma}
\begin{proof}
If $E = E_1\oplus E_2$ with $E_1$ odd rank, then the map which takes $(e_1,e_2)\mapsto (-e_1,e_2)$ is a bundle isomorphism which reverses orientation. The Euler class satisfies $e(E) = e(-E) = -e(E)$. Altogether we get that $2e(E) = 0$. The cohomological assumption on $M$ implies that $e(E)=0$.
\end{proof}
We are ready to give the proof of Theorem~\ref{thm:pres_projective_spaces}. For the rest of the section, recall that a manifold $M$ is said to be almost complex if its tangent bundle $TM$ is isomorphic to the realification of a complex vector bundle.
\begin{proof}[Proof of Theorem~\ref{thm:pres_projective_spaces}] Assume for a contradiction that we can find a representation $V$ of $H$ for which $T\mathbb{K}P^n\cong G\times_H V$. Here $(G,H)$ denotes $(Spin(2n+2), Spin(2n+1) \times S^1)$ when $\mathbb{K}=\mathbb{C}$ and $(Spin(4n+4), Spin(4n+3)\times SU(2))$ when $\mathbb{K}=\mathbb{H}$.
From the first part of Lemma~\ref{LEM:rep_of_spin} we know that the (real) dimension of the smallest non-trivial representation of the $Spin$ factor of $H$ is $2n+1$ for $\mathbb{C}$ (resp. $4n+3$ for $\mathbb{H}$), which is strictly larger than $\dim V=\dim \mathbb{C} P^n=2n$ (resp. $\dim V=\dim \mathbb{H} P^n = 4n$). In particular, in all cases, the $Spin$ factor of $H$ must act trivially on $V$. So, we can consider $V$ as a a representation of $S^1$ or $SU(2)$, respectively.
For the case $\mathbb{K} = \mathbb{C}$, the representation $V$ must split as a sum of line bundles, which implies a corresponding splitting of $T\mathbb{C}P^n$. However, there is no such such splitting, as shown by Glover, Homer, and Stong \cite[Theorem~1.1.(ii)]{GHS}.
We now turn to the case $\mathbb{K} = \mathbb{H}$. Since the Euler class $e(T\mathbb{H}P^n)$ is non-zero, $T\mathbb{H}P^n$ cannot contain any odd rank subbundles by Lemma~\ref{LEM:euler_class_odd}. Thus, all irreducible subrepresentations of $V$ are even dimensional. Since even dimensional irreducible representation of $SU(2)$ are complex, it follows that $V$ is a complex vector space, so $\mathbb{H}P^n$ is almost complex. However, Massey \cite{Massey} has shown that $\mathbb{H}P^n$ is not almost complex for any $n\geq 1$.
\end{proof}
Following a similar line of argument, we now give examples for which the tangent bundle is not a biquotient vector bundle for any presentation.
We denote by $\Delta SU(2)\subseteq Spin(4n+1)$ the image of the lift of the composition $SU(2)\rightarrow SO(4)\rightarrow SO(4n+1)$. Here the first map $SU(2)\rightarrow SO(4)$ is the same as above, while the second map $SO(4)\rightarrow SO(4n+1)$ is the block diagonal embedding $A\mapsto \operatorname{diag}(A,A,...,A,1)$. Again by work of Eschenburg \cite[Table 101]{Es84}, these inclusions induce biquotients:
$$M^{8n-4} := \Delta SU(2)\backslash Spin(4n+1)/Spin(4n-1)$$
We will show that, for $n\geq 3$ odd, no presentation of it as a biquotient has $TM$ a biquotient vector bundle. These are the examples of dimension $16n+4$ of Theorem~\ref{thm:first_theorem}.
To begin with, from Propositions \ref{prop:cover} and \ref{prop:reduced}, if $TM$ is a biquotient vector bundle, then it is for a reduced presentation with $G$ a product of a simply connected compact Lie group with a torus. Further, from \cite[Lemma~3.3]{To02}, the fact that it is reduced implies that the torus factor is trivial. Kapovitch and Ziller showed in \cite[Section~2]{KZ04} that the only reduced presentation $G/\!\!/ H$ of $M$ with $G$ simply connected is the one above. We note that Kapovitch and Ziller do not use the notion of ``reduced biquotient'', but their classification of presentations $G/\!\!/ H$ with $G$ simple is equivalent to classifying all such reduced presentations. Thus, it is enough to show that for this presentation, $TM$ is not a biquotient vector bundle.
A crucial step in the proof will be the following result of Hirzebruch relating the Euler characteristic $\chi(M)$ and the signature $\sigma(M)$ of an almost complex manifold whose dimension equals $0\pmod{4}$.
\begin{theorem}[Hirzebruch]\label{thm:hirzebruch}
Suppose $M$ is a closed $4m$-dimensional manifold which admits an almost complex structure, then $\chi(M) \equiv (-1)^m\sigma(M) \bmod 4$.
\end{theorem}
In Appendix \ref{appendix}, Michael Albanese has provided a proof of Theorem~\ref{thm:hirzebruch}.
\begin{proposition}
Let $n\geq 3$ odd. For the biquotient presentation above of $M^{8n-4}$, the tangent bundle $TM^{8n-4}$ is not a biquotient vector bundle.
\end{proposition}
\begin{proof} As shown in \cite{KZ04}, $M$ has the same integral cohomology groups of $\mathbb{H}P^{2n-1}$. Since $(8n-4)/2 = 4n-2 \equiv 2\pmod{4}$, it follows that $H^{4n-2}(M)=0$ and hence the signature $\sigma(M)$ vanishes for trivial reasons. On the other hand, $\chi(M) = 2n$. Thus, if $n$ is odd, then $\chi(M)\equiv 2\pmod{4}$. In particular, Theorem \ref{thm:hirzebruch} implies that $M$ cannot admit an almost complex structure for $n$ odd.
Assume for a contradiction that $TM$ is a biquotient vector bundle $G \times_H V$ for some odd $n\geq 3$ and certain representation $V$ of $H$, where $G=Spin(4n+1)$ and $H=SU(2)\times Spin(4n-1)$.
From the second part of Lemma~\ref{LEM:rep_of_spin}, we find that the only non-trivial irreducible representations of $Spin(4n-1)$ of dimension at most $8n-4=\dim M^{8n-4}$ is the standard representation $\Lambda^1$ (of real dimension $4n-1$).
Let $W_k$ denote the unique irreducible representation of $SU(2)$ of complex dimension $k$. Then the irreducible representation $W_k\otimes \Lambda^1$ has dimension $k(4n-1)$ so, if $W_k\otimes \Lambda^1$ is a subrepresentation of $V$, then we must have $k(4n-1) \leq 8n-4$ and hence $k=1$. But if $k=1$, then $TM$ has an odd rank subbundle induced by $W_1\otimes \Lambda^1$. This is a contradiction to Lemma~\ref{LEM:euler_class_odd}, since $\chi(M) = 2n$ and hence $e(TM)\neq 0$.
Thus, $V$ must be a representation of $SU(2)$ only. Again, because $TM$ cannot have any odd rank subbundles, all irreducible subrepresentations of $V$ are even dimensional. Since even dimensional irreducible representations of $SU(2)$ are complex, it follows that $V$ is a complex vector space, so $TM$ is almost complex. This is a contradiction.
\end{proof}
\section{Examples using characteristic classes}\label{S:examples_torus}
In this section, we complete the proof of Theorem \ref{thm:first_theorem} by constructing examples in dimensions $4$ and $6$, including infinitely many homotopy types in dimension $6$.
These biquotients will all be of the form $G/\!\!/ T^k$ for some biquotient action of a torus $T^k$ on a semi-simple compact Lie group $G$. The set of biquotient vector bundles over $G/\!\!/ T^k$ can be characterized as follows (see~\cite[Section~3]{DG21}).
\begin{proposition}\label{prop:biq_bundles}
A real vector bundle over $G/\!\!/ T$ is a biquotient bundle if and only if it is the realification of a sum of complex line bundles, possibly summed with a trivial real vector bundle.
\end{proposition}
In this section we will only consider $G/\!\!/ T^k$ where $T^k$ is of maximal rank. It follows that its dimension is even, say $2m$, and its Euler characteristic is positive. Moreover, we will consider only biquotient bundles of real rank $2m$. In other words, we only consider (realifications of) Whitney sums $L=\oplus L_i$ of $m$ complex line bundles $L_i$, i.e.~$1\leq i\leq m$. The relevant characteristic classes of the realification $rL$ will be the Pontryagin class $p_1$ and the Euler class $e$, which can be computed as:
\begin{align*}
p_1( rL) & = -c_2(L\oplus \bar L) = \sum_{i=1}^m c_1^2(L_i) && \in H^4(G/\!\!/ T^k) \\
e( rL ) & = c_{m}(L)=\prod_{i=1}^{m}c_1(L_i) && \in H^{2m}(G/\!\!/ T^k)
\end{align*}
where $\bar L$ denotes the conjugate bundle of $L$ and $c_i$ the $i$-th Chern class.
In the following subsections we will consider biquotients $N:=G/\!\!/ T^k$ and prove that $T N$ cannot be isomorphic to $rL$. This will be done in three steps
\begin{itemize}
\item we compute/collect/recall the necessary topological properties of $N$, including the values of $p_1(N)$ and $\pm e(N)$,
\item we compute $p_1(rL)$ and $e(rL)$ for any $L=\oplus_{i=1}^m L_i$ with $2m=\dim N$,
\item we convert the cohomological equations $p_1(rL) = p_1(TN)$ and $e(rL) = \pm e(TN)$ into numerical equations and show that they have no common solution.
\end{itemize}
We remark that any homogeneous space $G/T^k$ with $T^k$ a maximal torus in $G$ is stably parallelizable and hence $p_1=0$, a fact which no longer holds in general for biquotients $G/\!\!/ T^k$.
\subsection{The tangent bundle of $\mathbb{C}P^2\sharp \mathbb{C}P^2$} The manifold $N:=\mathbb{C}P^2\sharp \mathbb{C}P^2$ is a biquotient of the form $(S^3)^2/\!\!/ T^2$ \cite[p.~404]{To02}. Its cohomology ring equals $H^\ast(N) \cong \mathbb{Z}[u,v]/\langle u^2 - v^2, uv\rangle$ with both $|u|=|v| = 2$. The first Pontryagin class can be computed using Hirzebruch signature theorem and one finds out that $p_1(N)=6u^2$. Since $\chi(N) = 4$, it follows that the Euler class is $e(N) = \pm 4 u^2$.
Assume for a contradiction that $TN \cong r L =r (L_1\oplus L_2)$, where $L_i$ is determined by $c_1(L_i)=a_i u + b_i v$. Then one easily computes:
$$
p_1(r L)=(a_1^2 + b_1^2 + a_2^2 + b_2^2)u^2,\qquad e(r L)=(a_1 a_2 + b_1 b_2)u^2
$$
Since $TN \cong r L$ both classes $p_1,e$ must agree and hence we have the equations
\begin{align*}
6 &= a_1^2 + b_1^2 + a_2^2 + b_2^2\\
\pm 4 & = a_1 a_2 + b_1 b_2
\end{align*}
From the first equation it follows that, up to permutations and signs, the only integer solution of this equation is $(a_1,b_1,a_2,b_2) = (2,1,1,0)$. Assume without loss of generality that $b_2 = 0$. Hence the second equation simplifies to $\pm 4 = a_1 a_2$ with $(a_1,a_2)$ given up to permutation and signs as $(2,1)$ or $(1,1)$. In either case, $a_1a_2\neq 4$, so $TN$ is not a sum of line bundles.
Alternatively, one can argue that $TN$ is not a biquotient bundle as follows. If $TN$ was a biquotient bundle, then in particular it would be the realification of a complex vector bundle (see Proposition~\ref{prop:biq_bundles}). In other words, $\mathbb{C}P^2\sharp \mathbb{C}P^2$ would be an almost complex manifold, which leads to a contradiction as follows. The signature of $\mathbb{C}P^2\sharp \mathbb{C}P^2$ equals $2$ while its Euler characteristic equals $4$, hence Hirzebruch's Theorem~\ref{thm:hirzebruch} tells us that $\mathbb{C}P^2\sharp \mathbb{C}P^2$ cannot admit an almost complex structure.
\subsection{The tangent bundle of $SU(3)/\!\!/ T^2$}
Here we consider the inhomogeneous biquotient $N:=SU(3)/\!\!/ T^2$, which was discovered by Eschenburg. Its cohomology ring can be described as $H^\ast(N,\mathbb{Z})\cong\mathbb{Z}[x,y]/I$ where $I$ is the ideal $I = \langle x^3, y^2 + xy - x^2\rangle$ and with $|x| = |y| = 2$. In addition, $H^4(N,\mathbb{Z})\cong \mathbb{Z}^2$ has generators $x^2$ and $xy$, and $H^6(N,\mathbb{Z})$ has generator $x^2 y$ with $y^3 = 2x^2y$ and $xy^2 = -x^2y$. Further, $p_1(N) = 8x^2$ and $e(N) = \pm 6x^2 y$. All this information can be extracted e.g. from~\cite{EZ14}, where $SU(3)/\!\!/ T^2$ corresponds to $N_{-1}$ (see Section~4 and Proposition~5.4 therein).
Assume for a contradiction that $TN \cong r L =r (L_1\oplus L_2\oplus L_3)$, where $L_i$ is determined by $c_1(L_i) = a_i x + b_i y$. A simple calculation shows that
\begin{align*}
p_1(rL) &= \sum (a_i^2 + b_i^2)x^2 + \sum (2a_i b_i - b_i^2)xy \\
e(rL) &=(a_1a_2b_3 + a_1b_2a_3 + b_1a_2a_3 - a_1 b_2 b_3 - b_1a_2 b_3 - b_1b_2a_3 + 2b_1 b_2 b_3) x^2y
\end{align*}
Since $p_1(rL) = \sum c_1(L_i)^2$ is a sum of squares and $e(rL)$ is only defined up to sign, we may replace any $L_i$ with $-L_i$, with $c_1(-L_i) = -c_1(L_i)$. Thus, we may assume without loss of generality that all $b_i\geq 0$. Since $TN\cong rL$ we arrive to the numerical equations
\begin{align}
8 &= \sum (a_i^2 + b_i^2) \label{EQ:p_1_SU} \\
0 & = \sum (2a_i b_i - b_i^2) \label{EQ:p_11_SU} \\
\pm 6 &= a_1a_2b_3 + a_1b_2a_3 + b_1a_2a_3 - a_1 b_2 b_3 - b_1a_2 b_3 - b_1b_2a_3 + 2b_1 b_2 b_3 \label{EQ:e_SU}
\end{align}
with all $b_i\geq 0$.
From \eqref{EQ:p_1_SU} it is easy to see that, up to rearranging the order of the $a_i$ and $b_i$, the only possibilities are a) two of $\{a_i,b_j\}$ are $\pm 2$ with the others all zero, or b) one of $\{a_i,b_j\}$ is $\pm 2$, four are $\pm 1$, and one is zero.
For a), only two of $\{a_i, b_j\}$ are non-zero. Looking at \eqref{EQ:e_SU}, all terms in the right hand side involve three distinct elements of $\{a_i,b_j\}$, so all terms are zero and yield the contradiction $\pm 6=0$.
For b), we break into subcases depending on whether some $b_j = 0$, or some $a_i = 0$. So, first assume some $b_j = 0$, say, $b_1 = 0$. Then \eqref{EQ:e_SU} becomes $\pm 6 = a_1a_2 b_3 + a_1b_2a_3 -a_1b_2b_3$. If, say, $b_2 = \pm 2$, then all $a_i$ as well as $b_3$ must be odd. Then $a_1a_2 b_3 + a_1b_2a_3 -a_1b_2b_3$ is odd, so not equal to $\pm 6.$ Thus, we must have $b_2 = b_3 = 1$. Then \eqref{EQ:p_11_SU} simplifies to $1 = a_2 +a_3$. Since $a_2,a_3\in \{\pm 1, \pm 2\}$, we may assume $a_2 =\pm 2$ and $a_3 = -1$, and $a_1 = \pm 1$. Then \eqref{EQ:e_SU} becomes
$$\pm 6 = a_1a_2 b_3 + a_1b_2a_3 -a_1b_2b_3 = \pm 2 \pm 1 \pm 1,$$
which is a contradiction. This concludes the subcase with some $b_j = 0$.
It remains to rule out the subcase where some $a_i = 0$, say $a_1 = 0$. Rewrite \eqref{EQ:p_11_SU} as $\sum b_i^2 = 2 \sum a_i b_i$. If all $b_i$ are odd, then the left side is odd, while the right side is even. Thus at least one $b_i$ must be even. If $b_1 = 2$, then $\sum b_i^2 = 2 \sum a_i b_i$ gives $6 = 2 (0\cdot b_1 \pm 1 \pm 1)$, which has no solution. Thus, without loss of generality, we must have $b_2 = 2$. In this case $\sum b_i^2 = 2 \sum a_i b_i$ gives $6 = 2(2a_2 + a_3)$, hence $a_2 = a_3 = 1$ and consequently $b_1=b_3=1$. Substituting all this into \eqref{EQ:e_SU}, we find the contradiction
$$\pm 6 =b_1(a_2a_3 - a_2b_3-b_2a_3+2b_2b_3) = 1-1-2+4.$$
\subsection{The tangent bundle of the $R(p)$ biquotients}
For each integer $p$ we define a biquotient action of $T^3$ on $(S^3)^3$ as follows. Viewing $S^3\subseteq \mathbb{C}^2$ and denoting the $i$-th coordinate of $(S^3)^3$ by $(a_i,b_i)$, the action is defined by the rule:
$$(w_1,w_2,w_3)\ast(a_1,b_1,a_2,b_2,a_3,b_3) =(w_1a_1,w_1w_2^2b_2,w_2a_2,w_1w_2b_2,w_3a_3,w_1^pw_3b_3).$$
As shown in \cite[Proposition~4.22]{De17} (see also \cite[Proposition~3.5]{DG21}), the above action is free. For each $p\in\mathbb{Z}$, we denote the corresponding biquotient by $R(p)$ (note that it is denoted by $R(p,0)$ in \cite[Proposition~4.23]{De17} and by $R(A)$ in \cite[Section~3]{DG21}, where $A$ equals the matrix $A= \begin{bmatrix} 1 &2 & 0 \\ 1 & 1 & 0 \\ p & 0 & 1 \end{bmatrix}.$)
In \cite[Proposition~3.8]{DG21}, we use the cohomology ring of $H^\ast(R(p))$ to show that as $p$ varies over odd primes, infinitely many homotopy types of biquotients arise. In the present case, the biquotients of Theorem \ref{thm:first_theorem} will be of the form $R(2q)$, so this result does not directly apply. However, restricting to the case where $q$ is an odd prime, the proof of \cite[Proposition~3.8]{DG21} carries over verbatim until the very end, when the three equations should be reduced mod $q$ instead of mod $p$ to obtain the same contradiction. Thus, there are infinitely many homotopy types among these examples.
The following properties can be found in \cite[Section~2.4 and Proposition~4.26]{De17} (see also \cite[Proposition~3.6]{DG21}). The cohomology ring of $R(p)$ is given by $H^*(R(p)) \cong \mathbb{Z} [ u_1,u_2,u_3]/ I$ where all $|u_i| = 2$ and $I$ is the ideal $\langle u_1^2 + 2u_1 u_2, u_2^2 + u_1 u_2, u_i^2 + p u_1 u_3\rangle$. In particular, we observe that $u_2^2 = (u_1 + u_2)^2$ and $u_2 (u_1 + u_2) =0$ for any $p$. The Pontryagin class of $R(p)$ equals
$p_1(R(p))=(-6-2p^2)u_1u_2$. A fundamental class is $u_1 u_2 u_3$ and the Euler class equals $e(R(p))=\pm 8 u_1 u_2 u_3$. There are relations in degrees $4$ and $6$ as follows:
\begin{align*}
u_1^2 &=-2u_1u_2 && u_1^3 =u_2^3 =u_1^2u_2 =u_1u_2^2 = 0 \\
u_2^2 &=-u_1u_2 && u_3^2 =-2p^2u_1u_2u_3\\
u_3^2 &= -pu_1u_3 && u_1^2u_3 = -2u_1u_2u_3 \\
&&& u_1u_3^2 =2pu_1u_2u_3\\
&&& u_2^2 u_3 =-u_1u_2u_3\\
&&& u_2 u_3^2 =-pu_1u_2u_3
\end{align*}
From now on, suppose $p$ is even and $p\geq4$, and write $p=2q$ so that $q\geq 2$. In order to simplify the computations we consider an alternative description of $H^*(R(p))$ determined by the basis:
\begin{align*}
v_1 &:= u_1 + u_2 \\
v_2 &:= u_2 \\
v_3 &:= u_3 +q(v_1-v_2)=qu_1 + u_3
\end{align*}
The Pontryagin class hence equals $p_1(R(p))=(6+2p^2)v_1^2=(6+8q^2)v_1^2$. A fundamental class is $v_1^2v_3$ so that $e(R(p))=\pm 8 v_1^2v_3$. There are relations in degree $4$ and $6$ as follows:
\begin{align*}
v_2^2 & = v_1^2 && v_1^3 =v_2^3=v_1^2v_2=v_1v_2^2=v_1v_2v_3=v_1v_3^2=v_1v_3^2=0\\
v_1v_2&=0 && v_3^3 =2q^2v_1^2v_3\\
v_3^2 &= 2q^2 v_1^2 && v_2^2v_3 = v_1^2 v_3
\end{align*}
Let $L_i$ be the complex line bundle determined by $a_iv_1+b_iv_2+c_iv_3$ with $1\leq i\leq 3$ and define $L=L_1\oplus L_2\oplus L_3$. The relevant characteristic classes are
\begin{align*}
p_1(rL) =& \sum(a_i^2 + b_i^2 + 2q^2c_i^2)v_1^2 + 2\sum (a_ic_i) v_1v_3 + 2\sum (b_ic_i)v_2v_3 \\
e(rL) =& (a_1a_2c_3 + a_1c_2a_3 + c_1a_2a_3 + 2q^2 c_1c_2c_3 + b_1b_2c_3 + b_1c_2b_3 + c_1b_2b_3)v_1^2v_3
\end{align*}
Assume $TR(p)\cong rL$, so that we arrive to the numerical equations
\begin{align}
\pm 8 &= a_1a_2c_3 + a_1c_2a_3 + c_1a_2a_3 + 2q^2 c_1c_2c_3 + b_1b_2c_3 + b_1c_2b_3 + c_1b_2b_3 \label{EQ:R_p_euler} \\
6+8q^2 &= \sum(a_i^2 + b_i^2 + 2q^2c_i^2)\label{EQ:R_p_pont1} \\
0 &= \sum a_ic_i\label{EQ:R_p_pont2}\\
0 &= \sum b_ic_i\label{EQ:R_p_pont3}
\end{align}
We can rewrite \eqref{EQ:R_p_pont1} as
$$
2q^2(4- \sum c_i^2) = \sum(a_i^2 + b_i^2) - 6
$$
and since the right hand side is at least $ -6$ it follows that
$$
\sum c_i^2\leq 4 + \frac{3}{q^2}
$$
Since $q\geq 2$ by assumption and the $c_i$ are integers we obtain that $\sum c_i^2\leq 4$. Hence the only possibilities for $(c_1,c_2,c_3)$ up to ordering and signs are
$$
(0,0,0),\quad (2,0,0),\quad (1,1,0),\quad (1,1,1),\quad (1,0,0).
$$
The rest of the proof is dedicated to show that none of these possibilities can occur.
\begin{itemize}
\item[$(0,0,0)$] It cannot happen since \eqref{EQ:R_p_euler} yields the contradiction $\pm 8=0$.
\item[$(2,0,0)$] Assume without loss of generality that $c_1=\pm 2$ and $c_2=c_3=0$. Then from \eqref{EQ:R_p_pont2} and \eqref{EQ:R_p_pont3} we get that $a_1=b_1=0$. Then \eqref{EQ:R_p_euler} and \eqref{EQ:R_p_pont1} become:
$$
4= |a_2a_3 + b_2b_3|,\qquad 6=a_2^2+a_3^2+b_2^2+b_3^2.
$$
The last equation implies that the unique possibility for $(a_2,a_3,b_2,b_3)$ up to ordering and signs is $(2,1,1,0)$. None of the possible choices verifies the equation $4= |a_2a_3 + b_2b_3|$.
\item[$(1,1,0)$] Assume without loss of generality that $c_1,c_2\in\{\pm 1\}$ and $c_3=0$. Then \eqref{EQ:R_p_euler} becomes:
$$
\pm 8 = a_3(a_1c_2+ c_1a_2) + b_3(b_1c_2 + c_1b_2)
$$
We shall show both $a_1c_2 + a_2c_1$ and $b_1c_2 + b_2c_1$ are $0$, which gives the contradiction $\pm 8=0$. Let us just focus on $a_1c_2 + a_2c_1$, the case for $b's$ being essentially the same. If $c_1=c_2$, then \eqref{EQ:R_p_pont2} simplifies to $a_1 = -a_2$, and thus $a_1c_2 + a_2c_1 = a_1c_2 + a_2c_2 = c_2(a_1 + a_2) = 0$. If otherwise $c_1=-c_2$, then \eqref{EQ:R_p_pont2} simplifies to $a_1 = a_2$, thus $a_1 c_2 + a_2 c_1 = a_1 c_2 - a_2 c_2 = c_2(a_1 - a_2) = 0$. This completes the proof.
\item[$(1,1,1)$] We have that $c_i\in \{\pm 1\}$, so that $\frac{1}{c_i} = c_i$ and $c_i^2 = 1$. Then the equations \eqref{EQ:R_p_pont2} and \eqref{EQ:R_p_pont3} give
$$a_3 = -c_3(c_1 a_1 + c_2 a_2) \text{ and } b_3 = -c_3(c_1 b_1 + c_2 b_2).$$
Set $S = a_1^2 + a_2^2 + b_1^2 + b_2^2$. Substituting this in \eqref{EQ:R_p_pont1} we get
\begin{align*} 6 + 8q^2 &= S + a_3^2 + b_3^2 + 6q^2\\ &= S + a_1^2 + 2c_1c_2a_1 a_2 + a_2^2 + b_1^2 + 2c_1c_2 b_1b_2 + b_2^2+ 6q^2 \\ &= 2S + 2c_1c_2(a_1a_2 + b_1b_2)+ 6q^2.
\end{align*}
It follows that $c_1c_2 (3+q^2 - S) = a_1 a_2+b_1b_2$. Together with the fact that $(c_1 a_1 + c_2 a_2)(c_2 a_1 + c_1a_2) = c_1c_2(a_1^2+ a_2^2) + 2a_1 a_2$, equation \eqref{EQ:R_p_euler} can be written as:
\begin{align*} \pm 8 &= c_3(a_1 a_2 + b_1b_2) -c_3(c_1a_1+ c_2 a_2)(c_2 a_1 c_1 a_2) -c_3(c_1 b_1 + c_2b_2)(c_2 b_1 + c_1 b_2) + 2q^2 c_1 c_2 c_3\\ &= c_3(a_1 a_2 + b_1 b_2) -c_3 c_1 c_2 S - 2c_3(a_1 a_2 + b_1 b_2) + 2q^2 c_1 c_2 c_3 \\ &= -c_1(a_1 a_2 + b_1 b_2) -c_1c_2c_3(S-2q^2)\\ &= -c_3c_1c_2(3+q^2 - S) - c_1c_2c_3(S-2q^2)\\ &= -c_1c_2c_3(3+q^2 - S + S-2q^2)\\ &= -c_1 c_2 c_3(3-q^2).
\end{align*}
Thus, $q^2 = 3\pm 8$, giving an obvious contradiction.
\item[$(1,0,0)$] Assume without loss of generality that $c_1=\pm 1$ and $c_2=c_3=0$. Then from the equations \eqref{EQ:R_p_pont2} and \eqref{EQ:R_p_pont3} we get that $a_1=b_1=0$, and the equations \eqref{EQ:R_p_euler} and \eqref{EQ:R_p_pont1} become:
$$
8= |a_2a_3 + b_2b_3|,\qquad 6(q^2+1)=a_2^2+a_3^2+b_2^2+b_3^2
$$
Combining both equations we get that $(a_2+a_3)^2 +(b_2 + b_3)^2 = 6q^2 + 6\pm 16$. We find a contradiction by reducing mod $16$. To that end, note that a square mod $16$ equals $0,1,4,9$. Thus
$$6q^2 + 6 \pm 16 \in \{ 6, 12, 30, 60\} = \{6,12,14\} \mod 16$$
On the other hand, by inspection, the sum $(a_2+a_3)^2 +(b_2 + b_3)^2$ of two squares mod $16$ cannot be any of $6,12$ or $14$.
\end{itemize}
\subsection{Uniqueness of reduced presentations}
In the previous subsections we have shown that for one specific presentation of each of $\mathbb{C}P^2\sharp \mathbb{C}P^2$, $SU(3)/\!\!/ T^2$, and the $R(p)$ biquotients, the tangent bundle is not a biquotient vector bundle. We conclude this section with a proof that the tangent bundle cannot be a biquotient vector bundle for any presentation.
\begin{proposition}\label{prop:no_pres}
The tangent bundle of $\mathbb{C}P^2\sharp \mathbb{C}P^2$, $SU(3)/\!\!/ T^2$ or the $R(p)$ biquotients with $p$ even is not a biquotient vector bundle for any presentation as a biquotient.
\end{proposition}
\begin{proof}
By Propositions~\ref{prop:cover} and~\ref{prop:reduced} it is enough to consider reduced presentations $G/\!\!/ H$ with $G$ simply connected
For the biquotient presentations $G/\!\!/ H$ in the previous subsections, the proof that $T(G/\!\!/ H)$ is not a biquotient vector bundle for the standard presentation was done by showing that $T(G/\!\!/ H)$ is not a sum of complex line bundles. If we can show that any other reduced presentation of $G/\!\!/ H$ with $G$ simply connected must have $H$ a torus, then it will follow from Proposition~\ref{prop:biq_bundles} that $T(G/\!\!/ H)$ is not a biquotient vector bundle for any presentation.
For $\mathbb{C}P^2\sharp \mathbb{C}P^2$ \cite[Theorem 1.1]{De14} gives that the only reduced presentation has $H = T^2$.
For $SU(3)/\!\!/ T^2$,
notice that this space has the rational homotopy groups of $S^2\times \mathbb{C}P^2$, but that $\pi_4(SU(3)/\!\!/ T^2)$ is trivial. In \cite[Proposition 4.13]{De17}, it is shown that apart from quotients of $SU(3)$, any biquotient with the rational homotopy groups of $S^2\times \mathbb{C}P^2$ is a quotient of $S^3\times S^5$ by a $T^2$ action. In particular, such examples have $\pi_4\cong \pi_4(S^3)\cong \mathbb{Z}/2\mathbb{Z}$. Thus, the only reduced presentations of $SU(3)/\!\!/ T^2$ as $G/\!\!/ H$ have $G = SU(3)$ and $H = T^2$.
For the $R(p)$ biquotients, \cite[Proposition 3.14 Case 1]{De17} implies that any reduced presentation must have $G = (S^3)^3$ and $H = T^3$.\end{proof}
\subsection{Further examples}
Given the previous examples, we expect the existence of other biquotients of the form $G/\!\!/ T$ with $\operatorname{rank} T=\operatorname{rank} G$ for which the tangent bundle is not a biquotient bundle. However, our strategy uses in a crucial way the cohomology ring and the characteristic classes of the corresponding biquotients in order to obtain manageable numerical equations. As it turns out, the computations get significantly more complicated when the dimension increases.
We have investigated the case of the $8$-dimensional biquotients $Sp(2)/\!\!/ T^2$ and we discovered that there is one such biquotient, defined by the inclusion $(z,w)\mapsto (\operatorname{diag}(z,z),\operatorname{diag}(w,1))$, whose tangent bundle is not a biquotient bundle. The proof is longer than the previous ones, thus we omit most details and only give the following outline.
According to work of Eschenburg \cite[Corollary(b), p.8]{KZ06}, the biquotient of the previous paragraph is the unique inhomogeneous one of the form $Sp(2)/\!\!/ T^2$. Its cohomology ring and characteristic classes can be computed using the methods developed by Eschenburg and Singhof to obtain that $H^\ast(Sp(2)/\!\!/ T^2) \cong \mathbb{Z}[u, z]/\langle u^2-2z^2, z^4\rangle$ with $|u|=|z|=2$, and that $p_1(Sp(2)/\!\!/ T^2) = 12z^2$ and $e(N)=\pm 8 z^3 u$.
In order to arrive at a contradiction, one considers the realification of Whitney sums of four complex line bundles $L_i$ with $c_1(L_i)$ determined by $a_i u + b_i z$. Using the values of $p_1$ and $e$ of $Sp(2)/\!\!/ T^2$, one obtains three numerical equations with $8$ parameters $a_i,b_i$ each, which can be shown to have no common solution.
Additionally, and similarly to Proposition~\ref{prop:no_pres}, one can show that there is no other biquotient presentation $G/\!\!/ H$ for $Sp(2)/\!\!/ T^2$ with $G$ simple.
Finally, let us mention that $Sp(2)/\!\!/ T^2$ is easily seen to have the property that
$$ \text{for any }x_1,..., x_n \in H^2(M),\quad \sum x_i^2 = 0 \text{ if and only if } x_1 = x_2 = .... =x_n=0.$$
This is called Property $(\ast)$ in our previous paper \cite{DG21}, and its relevance is related to the non-existence of inverses among biquotient vector bundles. Moreover, and again according to work of Eschenburg \cite[Theorem~5.2]{KZ06}, there are exactly two inhomogeneous biquotients of the form $Sp(n)/\!\!/ T^n$ for any $n\geq 3$, and it can be shown that both of them satisfy Property $(\ast)$.
\section{Low dimensions}\label{S:low_dims}
This section contains the proof of Theorem~\ref{thm:dims235}. As mentioned in the introduction, a simply connected closed biquotient of dimension at most $5$ is diffeomorphic to one of the following: $S^2,S^3,S^4$, $\mathbb{C}P^2$, $S^2\times S^2$, $\mathbb{C}P^2\sharp \pm\mathbb{C}P^2$, $S^5$, $S^2\times S^3$, $SU(3)/SO(3)$ or $S^3\,\widehat{\times}\, S^2$.
The fact that the tangent bundle to $\mathbb{C}P^2\sharp \mathbb{C}P^2$ is not a biquotient bundle for any presentation is Proposition~\ref{prop:no_pres}. The fact that $\mathbb{C}P^2$ and $\mathbb{H}P^1 = S^4$ have presentations for which the tangent bundle is not a biquotient vector bundle is Theorem~\ref{thm:pres_projective_spaces}. Thus, we must show for the remaining examples $G/\!\!/ H$ that every presentation with $G$ simply connected, $T(G/\!\!/ H)$ is a biquotient vector bundle. From Proposition~\ref{prop:reduced}, we may assume the biquotient is reduced.
For $S^5$ and $SU(3)/SO(3)$, it follows from \cite{KZ04,De14} that the only presentations as reduced biquotients are homogeneous, where the result is clear.
In all remaining cases, by inspection, the only non-trivial rational homotopy groups occur in degrees $2$ and $3$. From \cite[Corollary 3.5]{De17}, the fact that $G/\!\!/ H$ is reduced implies that $H$ is a torus. From Proposition~\ref{prop:biq_bundles}, $T(G/\!\!/ H)$ is a biquotient bundle if and only if it is a sum of complex line bundles and trivial real bundles. For $S^2 = \mathbb{C}P^1$, $S^3$, and $S^2\times S^2$, this is obviously the case. For $\mathbb{C}P^2\sharp - \mathbb{C}P^2$, recall that this space is naturally an $S^2$-bundle over $S^2$. This induces a splitting of $T(\mathbb{C}P^2 \sharp - \mathbb{C}P^2)$ into (real) rank $2$ subbundles, i.e., a splitting into a sum of complex line bundles.
For $S^3\,\widehat{\times}\, S^2$ recall that it has the integral cohomology ring of $S^3\times S^2$. Denote its tangent bundle by $T$ and recall that $w_2(T)$ is the unique non-zero element in $H^2(S^3\,\widehat{\times}\, S^2;\mathbb{Z}_2)$. Take any complex line bundle $L$ with $c_1(L)$ an odd multiple of a generator of $H^2(S^3\,\widehat{\times}\, S^2)=\mathbb{Z}$, so that $w_2(rL)$ of the realification $rL$ is non-trivial. Denote by $3$ the trivial real bundle of rank $3$. It clearly follows that $w_2(rL\oplus 3)=w_2(T)\neq 0$. As $H^4(S^3\,\widehat{\times}\, S^2) = 0$, the characteristic classes $w_4$ and $p_1$ of $rL\oplus 3$ and $T$ trivially coincide. It follows from \cite[Theorem~1(ii)]{CV93} and the Remark in \cite[p.~755]{CV93}) that $rL\oplus 3$ is isomorphic to $T$. By Proposition~\ref{prop:biq_bundles}, $rL\oplus 3$ is a biquotient bundle and hence so is $T$.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.