WorksPhysics and Astronomy
Logarithmic thermodynamic cost of prediction from rare observations
Abstract
A memory predicting changes in its environment from rare observations must spend energy on sampling and recording. We consider autonomous chemical memories driven by a sample-and-hold sensor observing an environment that switches at rate gamma, with sampling frequency f<<gamma. An ideal observer with access to the full observation history can retain predictive information proportional to f. However, retaining a finite fraction of this information in an autonomous physical memory requires work of order kBT ln(gamma/f) per sensor change. This additional thermodynamic cost arises from the competing requirements of suppressing false records during long intervals between observations and responding rapidly to new observations. We establish a universal low-power bound on predictive information, which grows at most as P/ln(P_ref/P) with power budget P, uniformly over any finite number of memory states and independently of network topology and memory relaxation times. We further demonstrate that memory architecture determines the attainable scaling. With bounded transition rates and readout from the memory alone, two-state memories are limited to P^2 scaling, whereas a three-state memory is sufficient to attain the optimal logarithmic scaling. Our results establish a fundamental thermodynamic distinction between acquiring information and maintaining a useful record of a changing environment.