Robustly Invertible Nonlinear Dynamics and the BiLipREN: From Inversion-Based Control to Generative Trajectory Modelling

What the math gives to ML

The paper provides a constructive mechanism for recurrent models whose forward dynamics and causal inverse are both contracting and strongly input-output monotone, yielding certified bi-Lipschitz behavior. Its transferable asset is a dissipativity and LMI parameterization that simultaneously bounds perturbation amplification, guarantees hidden-state forgetting, and makes input reconstruction well conditioned. The direct neural-network transfer is to replace unconstrained RNN or state-space blocks with BiLipREN-style implicit recurrent cells, using orthogonal layers for norm preservation and a trainable storage matrix for contraction certificates. This predicts measurable stability boundaries and exponential decay rates rather than only empirical benchmark improvements.