A semiconcavity approach to stability of entropic plans and exponential convergence of Sinkhorn's algorithm

We study stability of optimizers and convergence of Sinkhorn's algorithm in the framework of entropic optimal transport. We show entropic stability for optimal plans in terms of the Wasserstein distance between their marginals under a semiconcavity assumption on the sum of the cost and one of the two entropic potentials. When employed in the analysis of Sinkhorn's algorithm, this result gives a natural sufficient condition for its exponential convergence, which does not require the ground cost to be bounded. By controlling from above the Hessians of Sinkhorn potentials in examples of interest, we obtain new exponential convergence results. For instance, for the first time we obtain exponential convergence for log-concave marginals and quadratic costs for all values of the regularization parameter. Moreover, the convergence rate has a linear dependence on the regularization: this behavior is sharp and had only been previously obtained for compact distributions arXiv:2407.01202. Other interesting new applications include subspace elastic costs [Cuturi et al. PMLR 202(2023)], weakly log-concave marginals, marginals with light tails, where, under reinforced assumptions, we manage to improve the rates obtained in arXiv:2311.04041, the case of unbounded Lipschitz costs, and compact Riemannian manifolds.

Non-asymptotic convergence bounds for Sinkhorn iterates and their gradients: a coupling approach

Computational optimal transport (OT) has recently emerged as a powerful framework with applications in various fields. In this paper we focus on a relaxation of the original OT problem, the entropic OT problem, which allows to implement efficient and practical algorithmic solutions, even in high dimensional settings. This formulation, also known as the Schr\"odinger Bridge problem, notably connects with Stochastic Optimal Control (SOC) and can be solved with the popular Sinkhorn algorithm. In the case of discrete-state spaces, this algorithm is known to have exponential convergence; however, achieving a similar rate of convergence in a more general setting is still an active area of research. In this work, we analyze the convergence of the Sinkhorn algorithm for probability measures defined on the $d$-dimensional torus $\mathbb{T}_L^d$, that admit densities with respect to the Haar measure of $\mathbb{T}_L^d$. In particular, we prove pointwise exponential convergence of Sinkhorn iterates and their gradient. Our proof relies on the connection between these iterates and the evolution along the Hamilton-Jacobi-Bellman equations of value functions obtained from SOC-problems. Our approach is novel in that it is purely probabilistic and relies on coupling by reflection techniques for controlled diffusions on the torus.