Quantitative contraction rates for Sinkhorn's algorithm: beyond bounded costs and compact marginals

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Conforti, Giovanni, Durmus, Alain, Greco, Giacomo
Formato: Preprint
Publicado: 2023
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866911633609064448
author Conforti, Giovanni
Durmus, Alain
Greco, Giacomo
author_facet Conforti, Giovanni
Durmus, Alain
Greco, Giacomo
contents We show non-asymptotic exponential convergence of Sinkhorn iterates to the Schrödinger potentials, solutions of the quadratic Entropic Optimal Transport problem on $\mathbb{R}^ d$. Our results hold under mild assumptions on the marginal inputs: in particular, we only assume that they admit an asymptotically positive log-concavity profile, covering as special cases log-concave distributions and bounded smooth perturbations of quadratic potentials. Up to the authors' knowledge, these are the first results which establish exponential convergence of Sinkhorn's algorithm in a general setting without assuming bounded cost functions or compactly supported marginals.
format Preprint
id arxiv_https___arxiv_org_abs_2304_04451
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Quantitative contraction rates for Sinkhorn's algorithm: beyond bounded costs and compact marginals
Conforti, Giovanni
Durmus, Alain
Greco, Giacomo
Probability
Optimization and Control
49Q22, 90C25 (Primary) 49N05, 93E20, 47D07 (Secondary)
We show non-asymptotic exponential convergence of Sinkhorn iterates to the Schrödinger potentials, solutions of the quadratic Entropic Optimal Transport problem on $\mathbb{R}^ d$. Our results hold under mild assumptions on the marginal inputs: in particular, we only assume that they admit an asymptotically positive log-concavity profile, covering as special cases log-concave distributions and bounded smooth perturbations of quadratic potentials. Up to the authors' knowledge, these are the first results which establish exponential convergence of Sinkhorn's algorithm in a general setting without assuming bounded cost functions or compactly supported marginals.
title Quantitative contraction rates for Sinkhorn's algorithm: beyond bounded costs and compact marginals
topic Probability
Optimization and Control
49Q22, 90C25 (Primary) 49N05, 93E20, 47D07 (Secondary)
url https://arxiv.org/abs/2304.04451