53

Convergence of Adaptive Biasing Potential methods for diffusions

Abstract

We prove the consistency of an adaptive importance sampling strategy based on biasing the potential energy function VV of a diffusion process dXt0=V(Xt0)dt+dWtdX_t^0=-\nabla V(X_t^0)dt+dW_t; for the sake of simplicity, periodic boundary conditions are assumed, so that Xt0X_t^0 lives on the flat dd-dimensional torus. The goal is to sample its invariant distribution μ=Z1exp(V(x))dx\mu=Z^{-1}\exp\bigl(-V(x)\bigr)\,dx. The bias VtVV_t-V, where VtV_t is the new (random and time-dependent) potential function, acts only on some coordinates of the system, and is designed to flatten the corresponding empirical occupation measure of the diffusion XX in the large time regime. The diffusion process writes dXt=Vt(Xt)dt+dWtdX_t=-\nabla V_t(X_t)dt+dW_t, where the bias VtVV_t-V is function of the key quantity μt\overline{\mu}_t: a probability occupation measure which depends on the past of the process, {\it i.e.} on (Xs)s[0,t](X_s)_{s\in [0,t]}. We are thus dealing with a self-interacting diffusion. In this note, we prove that when tt goes to infinity, μt\overline{\mu}_t almost surely converges to μ\mu. Moreover, the approach is justified by the convergence of the bias to a limit which has an intepretation in terms of a free energy. The main argument is a change of variables, which formally validates the consistency of the approach. The convergence is then rigorously proven adapting the ODE method from stochastic approximation.

View on arXiv
Comments on this paper