Dynamical large deviations of the fractional Ornstein-Uhlenbeck process

TL;DR

This paper investigates the large deviation properties of the fractional Ornstein-Uhlenbeck process, revealing a complex phase diagram of optimal paths and scaling behaviors, supported by advanced simulations and applicable to other Gaussian processes.

Contribution

It provides the first detailed analysis of dynamical large deviations for the non-Markovian fOU process, including a phase diagram and novel simulation techniques.

Findings

Identified three distinct scaling regimes of the optimal paths.

Derived a phase diagram on the (H, n) plane for the process.

Validated theoretical predictions with high-precision large-deviation simulations.

Abstract

The fractional Ornstein-Uhleneck (fOU) process is described by the overdamped Langevin equation $\dot{x}(t)+\gamma x=\sqrt{2 D}\xi(t)$, where $\xi(t)$ is the fractional Gaussian noise with the Hurst exponent $0<H<1$. For $H\neq 1/2$ the fOU process is non-Markovian but Gaussian, and it has either vanishing (for $H<1/2$), or divergent (for $H>1/2$) spectral density at zero frequency. For $H>1/2$, the fOU is long-correlated. Here we study dynamical large deviations of the fOU process and focus on the area $A_n=\int_{-T}^{T} x^n(t) dt$, $n=1,2,\ldots$ over a long time window $2T$. Employing the optimal fluctuation method, we determine the optimal path of the conditioned process, which dominates the large-$A_n$ tail of the probability distribution of the area, $\mathcal{P}(A_n,T)\sim \exp[-S(A_n,T)]$. We uncover a nontrivial phase diagram of scaling behaviors of the optimal paths and of the action $S(A_n\equiv 2 a_n T,T)\sim T^{\alpha(H,n)} a^{2/n}_n$ on the $(H,n)$ plane. The phase diagram includes three regions: (i) $H>1-1/n$, where $\alpha(H,n)=2-2H$, and the optimal paths are delocalized, (ii) $n=2$ and $H\leq \frac{1}{2}$, where $\alpha(H,n)=1$, and the optimal paths oscillate with an $H$-dependent frequency, and (iii) $H\leq 1-1/n$ and $n>2$, where $\alpha(H,n)=2/n$, and the optimal paths are strongly localized. We verify our theoretical predictions in large-deviation simulations of the fOU process. By combining the Wang-Landau Monte-Carlo algorithm with the circulant embedding method of generation of stationary Gaussian fields, we were able to measure probability densities as small as $10^{-170}$. We also generalize our findings to other stationary Gaussian processes with either diverging, or vanishing spectral density at zero frequency.