Designing Algorithms for Entropic Optimal Transport from an Optimisation Perspective

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Srinivasan, Vishwak, Jiang, Qijia
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909691612758016
author Srinivasan, Vishwak
Jiang, Qijia
author_facet Srinivasan, Vishwak
Jiang, Qijia
contents In this work, we develop a collection of novel methods for the entropic-regularised optimal transport problem, which are inspired by existing mirror descent interpretations of the Sinkhorn algorithm used for solving this problem. These are fundamentally proposed from an optimisation perspective: either based on the associated semi-dual problem, or based on solving a non-convex constrained problem over subset of joint distributions. This optimisation viewpoint results in non-asymptotic rates of convergence for the proposed methods under minimal assumptions on the problem structure. We also propose a momentum-equipped method with provable accelerated guarantees through this viewpoint, akin to those in the Euclidean setting. The broader framework we develop based on optimisation over the joint distributions also finds an analogue in the dynamical Schrödinger bridge problem.
format Preprint
id arxiv_https___arxiv_org_abs_2507_12246
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Designing Algorithms for Entropic Optimal Transport from an Optimisation Perspective
Srinivasan, Vishwak
Jiang, Qijia
Optimization and Control
Probability
Machine Learning
In this work, we develop a collection of novel methods for the entropic-regularised optimal transport problem, which are inspired by existing mirror descent interpretations of the Sinkhorn algorithm used for solving this problem. These are fundamentally proposed from an optimisation perspective: either based on the associated semi-dual problem, or based on solving a non-convex constrained problem over subset of joint distributions. This optimisation viewpoint results in non-asymptotic rates of convergence for the proposed methods under minimal assumptions on the problem structure. We also propose a momentum-equipped method with provable accelerated guarantees through this viewpoint, akin to those in the Euclidean setting. The broader framework we develop based on optimisation over the joint distributions also finds an analogue in the dynamical Schrödinger bridge problem.
title Designing Algorithms for Entropic Optimal Transport from an Optimisation Perspective
topic Optimization and Control
Probability
Machine Learning
url https://arxiv.org/abs/2507.12246