An optimal transport approach for the Schrödinger bridge problem and convergence of Sinkhorn algorithm (Q2210656)

The paper is concerned with the \textit{Schrödinger Bridge problem}. Given two Polish spaces \(X,Y\), a cost function \(c:X\times Y \rightarrow \mathbb R\) and a positive number \(\epsilon>0\), this problem consists of minimizing, over all probability measures \(\gamma\) on \(X\times Y\) with fixed marginals \(\rho_1\) and \(\rho_2\), the Kullback-Leibler divergence between \(\gamma\) and the cost kernel \(\mathcal K = e^{-c/\epsilon}\rho_1\otimes \rho_2\). A solution to this problem is necessarily of the form \(a_\epsilon(x)K(x,y)b_\epsilon(y)\) for two functions \(a_\epsilon,b_\epsilon\) called the \textit{entropic potentials}. This problem is important in mathematical physics, probability, fluid mechanics, metric geometry, data science and optimal transport. With respect to the latter, its discrete counterpart, defining \emph{entropic regularization of optimal transport}, is of central importance in numerical computation of optimal transport distances. The paper is also concerned with the multi-marginal extension of the Schrödinger Bridge Problem, involving a collection of Polish spaces \(X_1,\dots ,X_m\) and a cost function \(c:X_1\times \ldots \times X_m \rightarrow \mathbb R\). The paper uses a dual formulation of the Schrödinger Bridge problem, providing a (non-linear) analogue of the Kantorovich dual in optimal transport theory, which has previously appeared in the literature. The paper also introduces a generalized Legendre transform determined by the cost function \(c\) and \(\epsilon\), which formally recovers the \(c\)-transform in optimal transport theory when \(\epsilon\rightarrow 0\). Using these, the authors provide the following new results: \begin{itemize} \item Regularity of the solution to the Schrödinger Bridge Problem, assuming regularity of the cost function. More precisely, if the cost function is bounded, then the entropic potentials are bounded. \item Convergence of the \textit{multi-marginal Iterative Proportional Fitting Procedure} (Sinkhorn Algorithm) to the solution of the multi-marginal Schrödinger Bridge Problem. \end{itemize}