Variance reduction for diffusions

TL;DR

This paper investigates how adding an antisymmetric drift to reversible diffusions can reduce asymptotic variance, thereby improving sampling efficiency from target distributions in Monte Carlo methods.

Contribution

It provides a general comparison framework showing that antisymmetric drifts generally decrease asymptotic variance in diffusions used for sampling.

Findings

Adding antisymmetric drift reduces asymptotic variance.

Theoretical results apply to diffusions on Euclidean space and Riemannian manifolds.

Extensions include strict inequalities and worst-case analysis.

Abstract

The most common way to sample from a probability distribution is to use Monte-Carlo methods. For distributions on a continuous state space, one can find diffusions with the target distribution as equilibrium measure, so that the state of the diffusion after a long time provides a good sample from the desired distribution. There exist many diffusions with a common equilibrium, and one would naturally like to choose those that make the convergence to equilibrium faster. One way to do this is to consider a reversible diffusion, and add to it an antisymmetric drift which preserves the invariant measure. It has been proven that, in general, the irreversible algorithm performs better than the reversible one, in that the spectral gap is larger. In the present work, asymptotic variance is used as the criterion to compare these algorithms. We first provide a general comparison result, and then apply it to the specific cases of a diffusion on $\mathbb{R}^d$, or on a compact Riemannian manifold. We prove that, in general, adding an antisymmetric drift to a reversible diffusion reduces the asymptotic variance. We also provide some extensions of this result concerning strict inequality, the worst-case analysis, and the behavior of the asymptotic variance when the drift goes to infinity.