Finite-Time Chaos Diagnostics and Noise-Induced Basin Merging in a Two-Dimensional Map

What the math gives to ML

The paper provides a finite-horizon diagnostic for randomly perturbed discrete dynamical systems: products of independently sampled Jacobians yield finite-time Lyapunov exponents (FTLEs) whose centered and scaled values are approximately Gaussian, while increasing parameter noise can cause distinct attractor basins to merge at a critical noise amplitude. The transferable asset is estimating a distribution and confidence of finite-horizon instability rather than monitoring only an asymptotic or average exponent. The most promising neural-network use is monitoring recurrent or state-space models under stochastic parameter or input perturbations, and using an estimated basin-merging threshold to set robustness noise, scheduled perturbations, or safe operating limits.