Seminar: Sampling from non-smooth potentials using sub-gradient Langevin dynamics

AUEB STATISTICS SEMINARS SERIES 2025-26

Presenter: Iosif Lytras, Postdoctoral Researcher at Archimedes Research Unit, Athena Research Centre and National Technical University of Athens

Room Τ203, Troias Building

ABSTRACT

We examine the problem of sampling from non-smooth potentials. We propose a simple, efficient algorithm and provide novel state of the art convergence estimates for the convergence of the algorithm to the target measure in Wasserstein distance. By employing a novel local error framework we are able to obtain the first results for the convergence of the algorithm in KL divergence in the non-convex case. Finally, we present results for the performance of the algorithm as an optimizer for non-convex optimization problems.