Lyapunov-Like Conditions for Tight Exit Probability Bounds through Comparison Theorems for SDEs

Abstract

Computing upper bounds on exit probabilities—the probability that a system reaches certain "bad" sets—may assist decision-making in control of stochastic systems. Existing analytical bounds for systems described by stochastic differential equations are quite loose, especially for low-probability events, which limits their applicability in practical situations. In this paper we analyze why existing bounds are loose, and conclude that it is a fundamental issue with the underlying techniques based on martingale inequalities. As an alternative, we give comparison results for stochastic differential equations that via a Lyapunov-like function allow exit probabilities of an n-dimensional system to be upper-bounded by an exit probability of a one-dimensional Ornstein-Uhlenbeck process. Even though no closed-form expression is known for the latter, it depends on three or four parameters and can be a priori tabulated for applications. We extend these ideas to the controlled setting and state a stochastic analogue of control barrier functions. The bounds are illustrated on numerical examples and are shown to be much tighter than those based on martingale inequalities.