Moment Propagation of Polynomial Systems Through Carleman Linearization for Probabilistic Safety Analysis

TL;DR

This paper introduces a method using Carleman linearization to approximate moments of stochastic polynomial systems, enabling efficient probabilistic safety analysis through convex optimization.

Contribution

It develops a novel moment propagation technique for stochastic polynomial systems using truncated Carleman linearization, with applications to safety analysis.

Findings

Accurate moment approximation with error bounds

Efficient online computation methods

Probabilistic safety regions via convex optimization

Abstract

We develop a method to approximate the moments of a discrete-time stochastic polynomial system. Our method is built upon Carleman linearization with truncation. Specifically, we take a stochastic polynomial system with finitely many states and transform it into an infinite-dimensional system with linear deterministic dynamics, which describe the exact evolution of the moments of the original polynomial system. We then truncate this deterministic system to obtain a finite-dimensional linear system, and use it for moment approximation by iteratively propagating the moments along the finite-dimensional linear dynamics across time. We provide efficient online computation methods for this propagation scheme with several error bounds for the approximation. Our results also show that precise values of certain moments at a given time step can be obtained when the truncated system is sufficiently large. Furthermore, we investigate techniques to reduce the offline computation load using reduced Kronecker power. Based on the obtained approximate moments and their errors, we also provide hyperellipsoidal regions that are safe for some given probability bound. Those bounds allow us to conduct probabilistic safety analysis online through convex optimization. We demonstrate our results on a logistic map with stochastic dynamics and a vehicle dynamics subject to stochastic disturbance.