13
9

Unbiased Estimation using a Class of Diffusion Processes

Abstract

We study the problem of unbiased estimation of expectations with respect to (w.r.t.) π\pi a given, general probability measure on (Rd,B(Rd))(\mathbb{R}^d,\mathcal{B}(\mathbb{R}^d)) that is absolutely continuous with respect to a standard Gaussian measure. We focus on simulation associated to a particular class of diffusion processes, sometimes termed the Schr\"odinger-F\"ollmer Sampler, which is a simulation technique that approximates the law of a particular diffusion bridge process {Xt}t[0,1]\{X_t\}_{t\in [0,1]} on Rd\mathbb{R}^d, dN0d\in \mathbb{N}_0. This latter process is constructed such that, starting at X0=0X_0=0, one has X1πX_1\sim \pi. Typically, the drift of the diffusion is intractable and, even if it were not, exact sampling of the associated diffusion is not possible. As a result, \cite{sf_orig,jiao} consider a stochastic Euler-Maruyama scheme that allows the development of biased estimators for expectations w.r.t.~π\pi. We show that for this methodology to achieve a mean square error of O(ϵ2)\mathcal{O}(\epsilon^2), for arbitrary ϵ>0\epsilon>0, the associated cost is O(ϵ5)\mathcal{O}(\epsilon^{-5}). We then introduce an alternative approach that provides unbiased estimates of expectations w.r.t.~π\pi, that is, it does not suffer from the time discretization bias or the bias related with the approximation of the drift function. We prove that to achieve a mean square error of O(ϵ2)\mathcal{O}(\epsilon^2), the associated cost is, with high probability, O(ϵ2log(ϵ)2+δ)\mathcal{O}(\epsilon^{-2}|\log(\epsilon)|^{2+\delta}), for any δ>0\delta>0. We implement our method on several examples including Bayesian inverse problems.

View on arXiv
Comments on this paper