Guarantees in Wasserstein Distance for the Langevin Monte Carlo Algorithm

TL;DR

This paper analyzes the convergence rates of the Langevin Monte Carlo algorithm in Wasserstein distance, extending results beyond log-concave measures and exploring discretization schemes.

Contribution

It provides new convergence rate bounds for Langevin Monte Carlo in Wasserstein distance for broader classes of measures, including asymptotically log-concave, and compares discretization methods.

Findings

Convergence rates are established for asymptotically log-concave measures.

Sharper bounds lead to improved asymptotic convergence rates.

Ozaki's discretization does not significantly outperform Euler's scheme.

Abstract

We study the problem of sampling from a distribution $\target$ using the Langevin Monte Carlo algorithm and provide rate of convergences for this algorithm in terms of Wasserstein distance of order $2$. Our result holds as long as the continuous diffusion process associated with the algorithm converges exponentially fast to the target distribution along with some technical assumptions. While such an exponential convergence holds for example in the log-concave measure case, it also holds for the more general case of asymptoticaly log-concave measures. Our results thus extends the known rates of convergence in total variation and Wasserstein distances which have only been obtained in the log-concave case. Moreover, using a sharper approximation bound of the continuous process, we obtain better asymptotic rates than traditional results. We also look into variations of the Langevin Monte Carlo algorithm using other discretization schemes. In a first time, we look into the use of the Ozaki's discretization but are unable to obtain any significative improvement in terms of convergence rates compared to the Euler's scheme. We then provide a (sub-optimal) way to study more general schemes, however our approach only holds for the log-concave case.