Invertible networks learn Morse Lyapunov functions: hyperbolic equilibria get stability certificates on compact sets
Synopsis
For dynamical systems with a hyperbolic equilibrium, invertible neural networks can approximate a Morse Lyapunov function on compact sets and thereby certify global asymptotic stability. Previously, neural learning of Lyapunov functions guaranteed approximation only in norms that do not control derivatives, which is not enough for a Lyapunov inequality. The approach writes a Morse Lyapunov function as the squared norm of a diffeomorphism and approximates that diffeomorphism with a neural-ODE invertible network in a Sobolev norm.
Interpretation
The search space for stability certificates narrows from general smooth functions to Morse Lyapunov functions, whose Hessian is nonsingular at every critical point, and this class is shown to suffice for systems with a hyperbolic equilibrium. Systems were already known for which no polynomial, convex or locally real-analytic Lyapunov function exists, and planar systems were known to admit no Morse Lyapunov function at all. For vector fields with a hyperbolic equilibrium, a smooth Morse Lyapunov function exists that certifies global asymptotic stability; this existence result comes from a companion note, and the approximation theory is built on it.
A Morse Lyapunov function can be written as the squared norm of a diffeomorphism, so an invertible neural network can parametrize it. Coupling-based invertible networks had been shown to be universal approximators only in Lp or sup norm, and those norms do not control derivatives. Invertible networks built from neural ODEs are universal approximators of diffeomorphisms in a Sobolev norm, a result taken from Ishikawa et al. 2023.
Once the Sobolev approximation error is small enough, the network's output satisfies the Lyapunov inequality on a compact set and certifies global asymptotic stability there. Being close to a true Lyapunov function in sup norm does not guarantee a valid Lyapunov function, because a small perturbation can break strict decrease along trajectories. Theorem 5 gives a quantitative relation between the required approximation accuracy on a compact neighbourhood and the Lyapunov inequality, with a proof using the mean value theorem and convexity of the function.
Perspective
The discussion is restricted to smooth dynamical systems defined on the whole space with a hyperbolic equilibrium; the universal approximation part of Theorem 5 relies on neural-ODE results and therefore requires smoothness of order at least 2. In the non-hyperbolic case, Corollary 7 guarantees the Lyapunov inequality only on an annulus that excludes a ball around the equilibrium, corresponding to practical asymptotic stability. The numerical experiments cover only planar systems and illustrate these conclusions.
At a non-hyperbolic equilibrium the network is not forced to have the required structure near the equilibrium, so the guarantee holds only on an annulus. Architectures such as PLNet require choosing a suitable distortion parameter, which may be large, and its effect on convergence is not precisely characterized. The author lists converse theory for other architectures and rigorous numerical analysis as future work.
