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

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Chiarini, Alberto, Conforti, Giovanni, Greco, Giacomo, Tamanini, Luca
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918153386196992
author Chiarini, Alberto
Conforti, Giovanni
Greco, Giacomo
Tamanini, Luca
author_facet Chiarini, Alberto
Conforti, Giovanni
Greco, Giacomo
Tamanini, Luca
contents We study stability of optimizers and convergence of Sinkhorn's algorithm for the entropic optimal transport problem. In the special case of the quadratic cost, our stability bounds imply that if one of the two entropic potentials is semiconcave, then the relative entropy between optimal plans is controlled by the squared Wasserstein distance between their marginals. 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, based on semiconcavity propagation results. 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. These optimal rates are also established in situations where one of the two marginals does not have subgaussian tails. Other interesting new applications include subspace elastic costs, 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 Lipschitz costs with bounded Hessian, and compact Riemannian manifolds.
format Preprint
id arxiv_https___arxiv_org_abs_2412_09235
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle A semiconcavity approach to stability of entropic plans and exponential convergence of Sinkhorn's algorithm
Chiarini, Alberto
Conforti, Giovanni
Greco, Giacomo
Tamanini, Luca
Probability
Optimization and Control
Machine Learning
49Q22, 68Q87, 68W40, 60E15, 90C25
We study stability of optimizers and convergence of Sinkhorn's algorithm for the entropic optimal transport problem. In the special case of the quadratic cost, our stability bounds imply that if one of the two entropic potentials is semiconcave, then the relative entropy between optimal plans is controlled by the squared Wasserstein distance between their marginals. 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, based on semiconcavity propagation results. 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. These optimal rates are also established in situations where one of the two marginals does not have subgaussian tails. Other interesting new applications include subspace elastic costs, 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 Lipschitz costs with bounded Hessian, and compact Riemannian manifolds.
title A semiconcavity approach to stability of entropic plans and exponential convergence of Sinkhorn's algorithm
topic Probability
Optimization and Control
Machine Learning
49Q22, 68Q87, 68W40, 60E15, 90C25
url https://arxiv.org/abs/2412.09235