On the relation between gradient flows and the large-deviation principle, with applications to Markov chains and diffusion (Q471066): Difference between revisions

The authors consider that the gradient flows are an important subclass of evolution equations. A gradient flow structure can be exploited to obtain additional information about the evolution, such as existence and stability of solutions or to derive limiting equations if some system parameter goes to zero. Moreover, gradient flows can provide additional physical and analytical insight, such as the maximum dissipation of entropy and energy, or the geometric structure induced by the dissipation distance. A special subclass is formed by gradient flows with respect to a Wasserstein-type metric. Another concern of the authors is to see exactly under which conditions a generalized gradient system can be constructed out of a given \(L\)-function. To stress the generality of the study, they switch to a more general setting, not necessarily related to large deviations. Then, in order to decrease the level of abstraction of the paper, the authors focus on a specific class of \(L\)-functions, namely those that arise from a large-deviation principle of an empirical process. In this way, they study the large deviations of the empirical process corresponding to a sequence of independent Markov processes. In the last part of the study we find some applications of the theory. The authors calculate the large deviations and show that these do not generate the Riemannian gradient structure, although the driving functional is a multiple of the relative entropy in both cases. However, they do generate a well-defined and symmetric generalized gradient structure, as it is shown in the first application. In the second application the authors consider the diffusion-drift equation. The connection between the large deviations of the empirical measure and the Wasserstein gradient flow of the free energy was already made for the discrete-time setting, and in the (continuous-time) \(L\)-function setting. Nevertheless, they discuss how the theory developed above can be used to reproduce this result.