From large deviations to Wasserstein gradient flows in multiple dimensions (Q894519)

This paper deals with connection between large deviation rate functional and the Wasserstein gradient structure of the Fokker-Plank equation. Let \(\Psi\in C^2(\mathbb{R}^{d})\) be \(\lambda\)-convex for some \(\lambda\in \mathbb{R}\). The relative entropy functional \({\mathcal F}:\;{\mathcal P}_{2}(\mathbb{R}^{d})\to \mathbb{R}\cup\{+\infty\}\) is defined by \[ {\mathcal F}(\rho)=\int_{\mathbb{R}^{d}}f(x)\log f(x)\,dx+\int_{\mathbb{R}^{d}}\Psi(x)f(x)\,dx\,\, \text{ if}\,\, \rho(dx)=f(x)dx, {\mathcal F}(\rho)=0\,\,\text{ otherwise}, \] where \({\mathcal P}_2(\mathbb{R}^{d})=\{\rho\in{\mathcal P}(\mathbb{R}^{d}):\;\int| x|^2\rho(dx)<\infty\}\) denote the set of probability measures with finite second moment. Let \(D=C_{c}^{\infty}(\mathbb{R}^{d})\) be the space of test functions and let \(D'\) be the dual space of distributions. Given \(\rho\in {\mathcal P}(\mathbb{R}^{d})\), we define for \(s\in D':\;\| s\|^2_{-1,\rho}=\sup_{f\in D}\langle s,f\rangle^2/\int_{\mathbb{R}^{d}}|\nabla f|^2\,d\rho\). For fixed \(\rho_0\in {\mathcal P}_2(\mathbb{R}^{d})\) and \(\tau>0\) we defined functional \[ I_{\tau}(\bar\rho| \rho_0)=\inf_{(\rho_{t})_{t}\in AC^2(\rho_0,\bar\rho)}{1\over 4 \tau}\int_0^{1}\| \partial_{t}\rho_{t}-\tau\Delta\rho_{t}-\tau \text{div}(\rho_{t}\nabla\Psi)\|^2_{-1,\rho_{t}}dt, \] where \(AC^2(\rho_0,\rho_1)\) denotes the set of 2-absolutely continuous curves \((\rho_{t})_{t\in[0,1]}\) with values in \(({\mathcal P}_2(\mathbb{R}^{d}),W_2)\) and boundary conditions \(\rho|_{t=0}=\rho_0\) and \(\rho|_{t=1}=\rho_1\). The \(L^2\)-Wasserstein distance between \(\rho_0,\rho_1\in {\mathcal P}_2(\mathbb{R}^{d})\) is defined by \[ W_2(\rho_0,\rho_1)=\inf_{\pi\in \Gamma(\rho_0,\rho_1)}\left(\int_{\mathbb{R}^{d}\times \mathbb{R}^{d}}| x-y|^2\pi(dx,dy)\right)^{1/2}, \] where the infimum is taken over all coupling \(\pi\) of \(\rho_0\) and \(\rho_1\). The main result of this paper is following. For every \(\rho_0\in{\mathcal P}_2(\mathbb{R}^{d})\) such that \({\mathcal G}(\rho_0)<\infty\), we have \[ I_{\tau}(\cdot|\rho_0)-W_2^2(\rho_0,\cdot)/4\tau\rightarrow {1\over2}{\mathcal F}(\cdot)-{1\over2}{\mathcal F}(\rho_0), \,\, \text{as}\,\, \tau\to0, \] in the sense of \(\Gamma\)-convergence. Here \({\mathcal G}(\rho)=\int_{\{g>0\}}|\nabla g(x)|^2/g(x)\,d\nu(x)\), if \(\rho(dx)=g(x)\nu(dx)\), \(g\in W_{loc}^{1,1}(\mathbb{R}^{d})\), \({\mathcal G}(\rho)=+\infty\) otherwise.