Patchwork Sampling of Stochastic Differential Equations

TL;DR

This paper introduces Patchwork Sampling, a novel method for efficiently sampling stationary properties of stochastic differential equations, especially in rarely visited regions, by partitioning the state space into patches and combining local simulations.

Contribution

The paper extends the concept of truncated Markov chains to non-detailed balance processes and demonstrates a new sampling approach for complex stochastic systems.

Findings

Effective sampling in rarely visited regions of state space

Application to double-well potential systems and coupled particles

Extension of truncated Markov chain concepts to non-equilibrium processes

Abstract

We propose a method to sample stationary properties of solutions of stochastic differential equations, which is accurate and efficient if there are rarely visited regions or rare transitions between distinct regions of the state space. The method is based on a complete, non-overlapping partition of the state space into patches on which the stochastic process is ergodic. On each of these patches we run simulations of the process strictly truncated to the corresponding patch, which allows effective simulations also in rarely visited regions. The correct weight for each patch is obtained by counting the attempted transitions between all different patches. The results are patchworked to cover the whole state space. We extend the concept of truncated Markov chains which is originally formulated for processes which obey detailed balance to processes not fulfilling detailed balance. The method is illustrated by three examples, describing the one-dimensional diffusion of an overdamped particle in a double-well potential, a system of many globally coupled overdamped particles in double-well potentials subject to additive Gaussian white noise, and the overdamped motion of a particle on the circle in a periodic potential subject to a deterministic drift and additive noise. In the appendix we explain how other well-known Markov chain Monte Carlo algorithms can be related to truncated Markov chains.