Long-Time Large Deviation Asymptotics for Mean-Field Granular Media Equations

Abstract

We study long-time mean-field large deviations for granular-media diffusions. Under a dissipativity assumption, for arbitrary $T_N\to\infty$, the occupation measure of the $N$-particle empirical measure satisfies a large deviation principle on $\mathcal{P}(\mathcal{P}(\mathbb{R}^d))$ with speed $NT_N$ and rate function $\mathcal{I}(Γ)=\int_{\mathcal{P}(\mathbb{R}^d)}\mathfrak{i}(μ)Γ(dμ)$. Here $\mathfrak{i}(μ)$ is one quarter of the free-energy dissipation: $\mathfrak{i}(μ)=\frac14\left\|\nabla\frac{δ\mathcal{F}}{δμ}(μ)\right\|_{L^2(μ)}^2$, where $\mathcal{F}$ is the free energy of the model. The local cost $\mathfrak{i}(μ)$ can be interpreted as the per-unit-time Dawson--Gärtner action of the constant path at $μ$, and $\mathcal{I}$ averages these costs according to $Γ$. The main difficulty in the upper bound is that empirical measures are atomic, so their free energy is infinite. We address this by analyzing a regularized free-energy functional along the empirical-measure flow of controlled granular-media diffusions. The lower bound uses approximation by finitely supported occupation measures whose atoms have smooth densities with Gaussian tails. For quadratic potentials, we show that the variational problem associated with an atypically large overall space-time variance exhibits a transition between a Dirac minimizer at a Gaussian profile and non-Dirac minimizers supported on Gaussian profiles with different centers.