Physical limits to stochastic classifiers

Abstract

Stochastic physical computers use random transitions between states to process inputs. We quantify their ability to solve distinct tasks as the control parameters of a device vary. For Markov jump devices, we bound the number of classification tasks attainable at a prescribed output margin, with each task defined as a binary labeling of fixed inputs. Combining geometric packing with fluctuation-response inequalities, we bound this number in terms of the control range and dimension, rate sensitivities, output fluctuations, and entropy production. For state and net-current observables, when controls change kinetics without changing thermodynamic forces and energies, the number of tasks is shaped by entropy production and output fluctuations. If the product of their maximum values over the allowed control settings and inputs is too small, the device can perform at most one task. An exactly solvable two-state switch illustrates how sufficient driving and observation time enable barrier tuning to select copying or inversion.