Inference in non-equilibrium systems from incomplete information: the case of linear systems and its pitfalls

TL;DR

This paper investigates the challenges of estimating the distance from equilibrium in linear systems using incomplete Gaussian data, highlighting fundamental limitations and proposing strategies involving multiple data series.

Contribution

It reveals the inherent difficulty of inferring non-equilibrium measures from scalar Gaussian data and suggests combining data from different parameters as a potential solution.

Findings

Scalar Gaussian data cannot reliably estimate entropy production.

Single-variable observations are indistinguishable from equilibrium processes.

Using multiple data series with known parameter relations can mitigate inference issues.

Abstract

Data from experiments and theoretical arguments are the two pillars sustaining the job of modelling physical systems through inference. In order to solve the inference problem, the data should satisfy certain conditions that depend also upon the particular questions addressed in a research. Here we focus on the characterization of systems in terms of a distance from equilibrium, typically the entropy production (time-reversal asymmetry) or the violation of the Kubo fluctuation-dissipation relation. We show how general, counter-intuitive and negative for inference, is the problem of the impossibility to estimate the distance from equilibrium using a series of scalar data which have a Gaussian statistics. This impossibility occurs also when the data are correlated in time, and that is the most interesting case because it usually stems from a multi-dimensional linear Markovian system where there are many time-scales associated to different variables and, possibly, thermal baths. Observing a single variable (or a linear combination of variables) results in a one-dimensional process which is always indistinguishable from an equilibrium one (unless a perturbation-response experiment is available). In a setting where only data analysis (and not new experiments) is allowed, we propose - as a way out - the combined use of different series of data acquired with different parameters. This strategy works when there is a sufficient knowledge of the connection between experimental parameters and model parameters. We also briefly discuss how such results emerge, similarly, in the context of Markov chains within certain coarse-graining schemes. Our conclusion is that the distance from equilibrium is related to quite a fine knowledge of the full phase space, and therefore typically hard to approximate in real experiments.