On spurious detection of linear response and misuse of the fluctuation-dissipation theorem in finite time series

TL;DR

This paper develops a statistical test to detect the breakdown of linear response in deterministic dynamical systems, highlighting the data requirements and pitfalls of using the fluctuation-dissipation theorem with limited or coarse data.

Contribution

It introduces a goodness-of-fit test for linear response detection and analyzes the data needs and potential errors when applying the fluctuation-dissipation theorem.

Findings

Large data sets (around 10^6 observations) are needed to reliably detect linear response breakdown.

Naive application of the fluctuation-dissipation theorem can lead to erroneous predictions with limited or coarse data.

Detection success depends on the observable, perturbation size, and data length.

Abstract

Using a sensitive statistical test we determine whether or not one can detect the breakdown of linear response given observations of deterministic dynamical systems. A goodness-of-fit statistics is developed for a linear statistical model of the observations, based on results on central limit theorems for deterministic dynamical systems, and used to detect linear response breakdown. We apply the method to discrete maps which do not obey linear response and show that the successful detection of breakdown depends on the length of the time series, the magnitude of the perturbation and on the choice of the observable. We find that in order to reliably reject the assumption of linear response for typical observables sufficiently large data sets are needed. Even for simple systems such as the logistic map, one needs of the order of $10^6$ observations to reliably detect the breakdown, if less observations are available one may be falsely led to conclude that linear response theory is valid. The amount of data required is larger the smaller the applied perturbation. For judiciously chosen observables the necessary amount of data can be drastically reduced, but requires detailed {\em{a priori}} knowledge about the invariant measure which is typically not available for complex dynamical systems. Furthermore we explore the use of the fluctuation-dissipation theorem (FDT) in cases with limited data length or coarse-graining of observations. The FDT, if applied naively to a system without linear response, is shown to be very sensitive to the details of the sampling method, resulting in erroneous predictions of the response.