Optimal scaling for the transient phase of the random walk Metropolis algorithm: The mean-field limit

TL;DR

This paper extends the understanding of the optimal scaling of the random walk Metropolis algorithm during its transient phase, demonstrating that the same mean-field diffusive limit applies regardless of initial distribution, which informs better proposal variance choices.

Contribution

It proves that the diffusive limit and optimal scaling for the transient phase match those near stationarity, even from non-equilibrium initial distributions, using mean-field analysis.

Findings

The same scaling applies during the transient phase as at equilibrium.

The limit process is a nonlinear McKean diffusion.

Provides insights for optimizing proposal variance for faster convergence.

Abstract

We consider the random walk Metropolis algorithm on $\mathbb{R}^n$ with Gaussian proposals, and when the target probability measure is the $n$-fold product of a one-dimensional law. In the limit $n\to\infty$, it is well known (see [Ann. Appl. Probab. 7 (1997) 110-120]) that, when the variance of the proposal scales inversely proportional to the dimension $n$ whereas time is accelerated by the factor $n$, a diffusive limit is obtained for each component of the Markov chain if this chain starts at equilibrium. This paper extends this result when the initial distribution is not the target probability measure. Remarking that the interaction between the components of the chain due to the common acceptance/rejection of the proposed moves is of mean-field type, we obtain a propagation of chaos result under the same scaling as in the stationary case. This proves that, in terms of the dimension $n$, the same scaling holds for the transient phase of the Metropolis-Hastings algorithm as near stationarity. The diffusive and mean-field limit of each component is a diffusion process nonlinear in the sense of McKean. This opens the route to new investigations of the optimal choice for the variance of the proposal distribution in order to accelerate convergence to equilibrium (see [Optimal scaling for the transient phase of Metropolis-Hastings algorithms: The longtime behavior Bernoulli (2014) To appear]).