Paperlane

no. 17 · Wednesday 18 November 2026

Learning Locally Stable Neural ODEs and Their Regions of Attraction

Alice Harting, Karl Henrik Johansson, Sophie Tarbouriech and Matthieu Barreau

math.OC · v2

Abstract

We propose a class of neural ODEs that are hard-constrained to have a locally exponentially stable equilibrium point at the origin, with an exactly characterized region of attraction. The model obtains its dynamics by passing an arbitrary neural vector field through a projection layer defined by a jointly learned Lyapunov function. By design, this function satisfies a variant of the Zubov equation for the projected dynamics and is therefore a maximal Lyapunov function for the projected system. As a result, the projected dynamics are locally exponentially stable at the origin, and the region of attraction is exactly given by the 1-sublevel set of the Lyapunov function. We prove that this model class universally approximates any sufficiently smooth dynamics that are locally exponentially stable at the origin on their region of attraction, and propose an implementation that is consistent with the theory. The approach is validated experimentally on nonlinear systems with nonconvex regions of attraction.

fields
Optimization and Control
pdf ↗ page arXiv:2606.19109 ↗