Model reduction for fully nonlinear stochastic systems

TL;DR

This paper introduces a new model reduction method for fully nonlinear stochastic systems that avoids linearization and lifting, using structural properties to define Gramians for effective projection-based reduction.

Contribution

It develops a novel framework leveraging system nonlinearities directly to create Gramians, enabling balanced truncation for stochastic nonlinear systems without linearization.

Findings

Provides conditions for Gramians existence and stability

Derives explicit error bounds for reduced models

Demonstrates effective reduction on complex stochastic systems

Abstract

This paper presents a novel model order reduction framework tailored for fully nonlinear stochastic dynamics without lifting them to quadratic systems and without using linearization techniques. By directly leveraging structural properties of the nonlinearities -- such as local and one-sided Lipschitz continuity or one-sided linear growth conditions -- the approach defines generalized reachability and observability Gramians through Lyapunov-type differential operators. These Gramians enable projection-based reduction while preserving essential dynamics and stochastic characteristics. The paper provides sufficient conditions for the existence of these Gramians, including a Lyapunov-based mean square stability criterion, and derives explicit output error bounds for the reduced order models. Furthermore, the work introduces a balancing and truncation procedure for obtaining reduced systems and demonstrates how dominant subspaces can be identified from the spectrum of the Gramians. The theoretical findings are grounded in rigorous stochastic analysis, extending balanced truncation techniques to a broad class of nonlinear systems under stochastic excitation.