Learning a quantitative criterion for distinguishing chaos from noise

What the math gives to ML

The paper presents a quantitative chaos-versus-noise diagnostic based on a fixed recurrent reservoir, a trained linear readout, and cross-prediction of future increments. Its transferable mechanism is a restricted dynamical model whose held-out predictability measures reproducible temporal structure while suppressing memorization of one noisy realization. This can be used beside neural sequence models as a data-quality gate, an early-stopping monitor, or a regularizer for learned dynamical systems. The main falsifiable signature is a positive gap between real-data and surrogate-data squared Pearson correlations, with the real-data score declining as prediction lag grows.