Optimal control formulation of transition path problems for Markov Jump Processes

TL;DR

This paper formulates the transition path problem for Markov jump processes as an optimal control problem, providing explicit solutions and methods to realize transitions between metastable states with controlled transition rates.

Contribution

It introduces a stochastic optimal control framework for transition paths in Markov jump processes, including explicit formulas for optimal controls and measures, and handles unbounded terminal costs using Gamma convergence.

Findings

Explicit formula for the optimal path measure via the committor function.

A unified framework for finite and infinite horizon transition path problems.

Method to realize almost sure transitions through optimal control.

Abstract

Among various rare events, the effective computation of transition paths connecting metastable states in a stochastic model is an important problem. This paper proposes a stochastic optimal control formulation for transition path problems in an infinite time horizon for Markov jump processes on polish space. An unbounded terminal cost at a stopping time and a controlled transition rate for the jump process regulate the transition from one metastable state to another. The running cost is taken as an entropy form of the control velocity, in contrast to the quadratic form for diffusion processes. Using the Girsanov transformation for Markov jump processes, the optimal control problem in both finite time and infinite time horizon with stopping time fit into one framework: the optimal change of measures in the C\`adl\`ag path space via minimizing their relative entropy. We prove that the committor function, solved from the backward equation with appropriate boundary conditions, yields an explicit formula for the optimal path measure and the associated optimal control for the transition path problem. The unbounded terminal cost leads to a singular transition rate (unbounded control velocity), for which, the Gamma convergence technique is applied to pass the limit for a regularized optimal path measure. The limiting path measure is proved to solve a Martingale problem with an optimally controlled transition rate and the associated optimal control is given by Doob-h transformation. The resulting optimally controlled process can realize the transitions almost surely.