Proximal methods in the Wasserstein space: links and differences from the Hilbert case
Proximal methods in the Wasserstein space: links and differences from the Hilbert case
Emanuele Naldi (Università di Genova)
Abstract: In this talk, we revisit the proximal point method, known in the Wasserstein setting as the Jordan–Kinderlehrer–Otto (JKO) scheme. While the classical JKO scheme assumes exact minimization at each step, practical implementations rely on approximate solutions due to computational constraints. We therefore investigate the asymptotic convergence properties of the inexact JKO scheme and its generalization, the proximal-gradient algorithm in the Wasserstein space. We analyze two types of inexactness: errors in Wasserstein distance and errors in functional evaluations. We establish rigorous convergence guarantees under controlled error conditions. Beyond the inexact setting, we also extend the convergence results by considering varying stepsizes. Our analysis expands previous approaches, providing new insights into discrete Wasserstein gradient flows. We finish the talk with a comparison to the Hilbert space setting, where the proximity operator is nonexpansive, a property that plays a central role in many classical convergence results. In the Wasserstein setting, the nonexpansivity of the proximity operator generally fails, even for geodesically convex functionals. We discuss the class of functions for which this property still holds and highlight potential directions for future research.
