Lower bounds on entropy production from dynamical correlation functions

What the math gives to ML

The paper provides a constructive way to infer a lower bound on entropy production from time-reversal asymmetry in two-time correlation functions, even when the microscopic state is hidden and only coarse observables are available. The transferable mechanism is an irreversibility monitor: antisymmetric lagged correlations estimate probability currents, while a quadratic current bound certifies a minimum dissipation rate. In neural-network training, this can be applied to coarse-grained parameter and update states to detect circulating or unstable optimizer dynamics and trigger learning-rate reduction, momentum damping, or a trust-region reset. The key falsifiable prediction is that the correlation-asymmetry estimate rises before gradient explosions or loss divergence, and that adaptive control suppresses this rise.