Optimal Computation from Fluctuation Responses

ABSTRACT: We show how to design computation protocols that minimize thermodynamic cost while ensuring correct outcomes using fluctuation response relations (FRR) and machine learning. Unlike current approaches, our method simultaneously optimizes distributions and protocols, using FRR-derived gradients and learning iteratively from sampled trajectories. We design protocols for bit erasure in a double-well potential and translating harmonic traps. The framework extends to both over- damped and underdamped systems while achieving the theoretically optimal protocol and work costs comparable to relevant finite-time bounds.