Large deviations at level 2.5 and for trajectories observables of diffusion processes : the missing parts with respect to their random-walks counterparts

TL;DR

This paper investigates the differences in large deviation principles at level 2.5 between diffusion processes and their approximating random walks, analyzing how these deviations behave in the limits of lattice spacing and time-step.

Contribution

It provides a detailed analysis of the qualitative differences in large deviations at level 2.5 for diffusion processes versus Markov chains, including the limits of random walks approximating diffusions.

Findings

Large deviations at level 2.5 differ qualitatively between diffusions and Markov chains.

The behavior of large deviations for random walks converging to diffusions is characterized in the limits of lattice spacing and time-step.

The projection of trajectory observables from random walks to diffusion processes is clarified.

Abstract

Behind the nice unification provided by the notion of the level 2.5 in the field of large deviations for time-averages over a long Markov trajectory, there are nevertheless very important qualitative differences between the meaning of the level 2.5 for diffusion processes on one hand, and the meaning of the level 2.5 for Markov chains either in discrete-time or in continuous-time on the other hand. In order to analyze these differences in detail, it is thus useful to consider two types of random walks converging towards a given diffusion process in dimension $d$ involving arbitrary space-dependent forces and diffusion coefficients, namely (i) continuous-time random walks on the regular lattice of spacing $b$ ; (ii) discrete-time random walks in continuous space with a small time-step $\tau$. One can then analyze how the large deviations at level 2.5 for these two types of random walks behave in the limits $b \to 0$ and $\tau \to 0$ respectively, in order to describe how the fluctuations of some empirical observables of the random walks are suppressed in the limit of diffusion processes. One can then also study the limits $b \to 0$ and $\tau \to 0$ for any trajectory observable of the random walks that can be decomposed on its empirical density and its empirical flows in order to see how it is projected on the appropriate trajectory observable of the diffusion process involving its empirical density and its empirical current.