Thermodynamics of computations with absolute irreversibility, unidirectional transitions, and stochastic computation times

Developing a physical theory of computation is an open challenging task at the interface of non-equilibrium thermodynamics and computer science. An important part of this task requires the examination of thermodynamic quantities in formal models of computers which execute computational tasks, such as automata or word-RAM models executing algorithms. This implies dealing with a number of difficulties such as stochastic halting times, unidirectional (possibly deterministic) transitions, and restricted initial conditions. Here, we present a framework which tackles all such difficulties by extending martingale theory of nonequilibrium thermodynamics to non-stationary Markovian dynamics with broken local detailed balance and absolute irreversibility. In doing so, we derive universal fluctuation relations and second-law-like inequalities that provide both lower and upper bounds for the intrinsic dissipation (mismatch cost) associated with any computation performed over arbitrary stochastic times, so long as it is implemented with a periodic process. We then provide universal equalities and inequalities for the acceptance probability of words of a given length by a computer in terms of thermodynamic quantities, and outline connections between computer science and stochastic resetting. We illustrate most of our results with exhaustive numerical simulations of fundamental models of computation from theoretical computer science, namely deterministic finite automata processing bit strings. Our results, while motivated from the computational context, are applicable far beyond it.