Linear Convergence of Gradient Descent for Quadratically Regularized Optimal Transport

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: González-Sanz, Alberto, Nutz, Marcel, Valdevenito, Andrés Riveros
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914579768934400
author González-Sanz, Alberto
Nutz, Marcel
Valdevenito, Andrés Riveros
author_facet González-Sanz, Alberto
Nutz, Marcel
Valdevenito, Andrés Riveros
contents In optimal transport, quadratic regularization is an alternative to entropic regularization when sparse couplings or small regularization parameters are desired. Quadratic regularization penalizes transport couplings by the squared $L^2$ norm of their density, or equivalently by the $χ^2$ divergence. While a number of computational approaches have been shown to work in practice, the dual problem is not strongly convex and theoretical convergence results are scarce. We focus on the dual gradient descent algorithm in a continuous setting and establish linear convergence in $L^2$, that is, the $L^2$ distance between the iterates and the limiting potentials decreases exponentially fast. The proof is based on a spectral analysis of the linearized gradient descent operator at the optimum. We show that this operator is a strict contraction and that the nonlinear iteration inherits this property after a burn-in period.
format Preprint
id arxiv_https___arxiv_org_abs_2509_08547
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Linear Convergence of Gradient Descent for Quadratically Regularized Optimal Transport
González-Sanz, Alberto
Nutz, Marcel
Valdevenito, Andrés Riveros
Optimization and Control
Analysis of PDEs
Functional Analysis
Probability
49N10, 49N05, 90C25
In optimal transport, quadratic regularization is an alternative to entropic regularization when sparse couplings or small regularization parameters are desired. Quadratic regularization penalizes transport couplings by the squared $L^2$ norm of their density, or equivalently by the $χ^2$ divergence. While a number of computational approaches have been shown to work in practice, the dual problem is not strongly convex and theoretical convergence results are scarce. We focus on the dual gradient descent algorithm in a continuous setting and establish linear convergence in $L^2$, that is, the $L^2$ distance between the iterates and the limiting potentials decreases exponentially fast. The proof is based on a spectral analysis of the linearized gradient descent operator at the optimum. We show that this operator is a strict contraction and that the nonlinear iteration inherits this property after a burn-in period.
title Linear Convergence of Gradient Descent for Quadratically Regularized Optimal Transport
topic Optimization and Control
Analysis of PDEs
Functional Analysis
Probability
49N10, 49N05, 90C25
url https://arxiv.org/abs/2509.08547