Events

Sampling, spectral gaps and stochastic localization

Speaker: Ahmed el Alaoui

Location: 60 Fifth Avenue, Room 150

Date: Thursday, November 30, 2023

In this I will present techniques for efficiently sampling form a high-dimensional probability distribution. The stochastic localization process will take center stage. I will discuss ways of making the process algorithmic, with applications to sampling from mean-field spin glass Gibbs measures, and ways of proving spectral gaps, fast mixing of Markov chains, and related functional inequalities in situations where traditional methods fail. I will take the random field Ising model as an example to illustrate the difficulties.This talk is partly based on the following papers:
https://arxiv.org/abs/2310.08912
https://arxiv.org/abs/2311.06171