[Non-asymptotic analysis of Langevin-type algorithms for non-log-concave non-smooth distributions] Talk by Dr.Shogo Nakakita, University of Tokyo

This talk will be held in a hybrid format, both in person at Room4 of RIKEN AIP (Nihonbashi office) and online by Zoom. Room4; only available to AIP researchers.

Speaker: Dr.Shogo Nakakita (https://www.shnakakita.org/), University of Tokyo

Abstract: We study the sampling problem from distributions without log-concavity or smoothness and give a non-asymptotic analysis for Langevin-type algorithms for the problem. Our analysis is based on a change of measures by the Liptser—Shiryaev approach and mollification of potential functions. Our result can be applied to a convergence analysis of the standard Langevin Monte Carlo method under weak smoothness with simple assumptions. We also apply the result to propose a Langevin-type algorithm for distributions without the continuous differentiability of the potentials.