Extreme value statistics for dynamical systems with noise

TL;DR

This paper investigates how noise influences the distribution of extreme values in dynamical systems, showing that noise can induce extreme value laws in regular systems and alter recurrence properties in chaotic systems, with implications for finite data analysis.

Contribution

It demonstrates the effect of additive noise on extreme value laws in both regular and chaotic dynamical systems, providing theoretical predictions and numerical validation.

Findings

Noise induces extreme value laws in regular systems regardless of noise intensity.

In chaotic systems, noise affects the extremal index and clustering of extreme events.

Finite data analysis must consider noise effects to avoid misinterpretation of dynamics.

Abstract

We study the distribution of maxima (Extreme Value Statistics) for sequences of observables computed along orbits generated by random transformations. The underlying, deterministic, dynamical system can be regular or chaotic. In the former case, we will show that by perturbing rational or irrational rotations with additive noise, an extreme value law appears, regardless of the intensity of the noise, while unperturbed rotations do not admit such limiting distributions. In the case of deterministic chaotic dynamics, we will consider observables specially designed to study the recurrence properties in the neighbourhood of periodic points. Hence, the exponential limiting law for the distribution of maxima is modified by the presence of the extremal index, a positive parameter not larger than one, whose inverse gives the average size of the clusters of extreme events. The theory predicts that such a parameter is unitary when the system is perturbed randomly. We perform sophisticated numerical tests to assess how strong is the impact of noise level, when finite time series are considered. We find agreement with the asymptotic theoretical results but also non-trivial behaviour in the finite range. In particular our results suggest that in many applications where finite datasets can be produced or analysed one must be careful in assuming that the smoothing nature of noise prevails over the underlying deterministic dynamics.