Quantitative estimates of propagation of chaos for stochastic systems with \(W^{-1,\infty}\) kernels (Q1795189)

The authors consider stochastic systems of \(N\) indistinguishable point-particles which evolve according to the coupled stochastic differential equations \(dX_{i}=F(X_{i})dt+\frac{1}{N}\sum_{j\neq i}K(X_{i}-X_{j})dt+\sqrt{2\sigma N}dW_{t}^{i}\) , \(i=1,\ldots ,N\), in the Itô sense, where \(X_{i}\in \Pi ^{d}\), the \(d\)-dimensional torus, \(W_{i}\) are \(N\) independent standard Wiener processes in \(\mathbb{R}^{d}\), \(K\) is an interaction kernel which is given by the Biot--Savart law in dimension 2, \( K(x)=\alpha \frac{x^{\perp }}{\left| x\right| ^{2}}+K_{0}(x)\), \( x^{\perp }\) denoting the rotation of \(x\) by \(\pi /2\) and \(K_{0}\) being a smooth correction to periodize \(K\) on the torus. The authors introduce the equation \(\partial _{t}\rho _{N}+\sum_{i=1,\ldots ,N}\operatorname{div}_{x_{i}}(\rho _{N}(F(x_{i})+\frac{1}{N}\sum_{j=1,\ldots ,N}K(x_{i}-x_{j}))=\sum_{i=1,\ldots ,N}\sigma _{N}\Delta _{x_{i}}\rho _{N}\) and the notion of entropy solution associated to this problem. The purpose of the paper is to derive explicit and quantitative estimates comparing the above problem to the mean field limit \(\overline{\rho }\) solution of \( \partial _{t}\overline{\rho }+\operatorname{div}_{x}(\overline{\rho }[F+K\star _{x} \overline{\rho }])=\sigma \Delta \overline{\rho }\). The main result of the paper proves that if \(\operatorname{div}F\in L^{\infty }(\Pi ^{d})\), \(K\in \overset{. }{W}^{-1,\infty }(\Pi ^{d})\) with \(\operatorname{div} K\in \overset{.}{W}^{-1,\infty } \), \(\sigma _{N}\geq \sigma >0\), and if the solution \(\overline{\rho }\) of the last above equation satisfies \(\overline{\rho }\in L^{\infty }([0,T];W^{-2,p}(\Pi ^{d}))\) for any \(p<\infty \) and \(\inf \overline{\rho }>0 \), \(\int_{\Pi ^{d}}\overline{\rho }=1\), the entropy solution \(\rho _{N}\) of the above equation satisfies \(\mathcal{H}_{N}(\rho _{N},\overline{\rho } _{N})(t)\leq e^{\overline{M}(\left\| K\right\| +\left\| K\right\| ^{2})}(\mathcal{H}_{N}(\rho _{N}^{0},\overline{\rho }_{N}^{0})+ \frac{1}{N}\overline{M}(1+t(1+\left\| K\right\| ^{2}))\left| \sigma -\sigma _{N}\right| )\). Here the relative entropy \(\overline{ \rho }_{N}\) is defined through \(\overline{\rho }_{N}(t,x_{1},\ldots ,x_{N})=\Pi _{i=1,\ldots ,N}\overline{\rho }(t,x_{i})\) and \(\mathcal{H} _{N}(\rho _{N},\overline{\rho }_{N})(t)=\frac{1}{N}\int_{\Pi ^{dN}}\rho _{N}(t,X^{N})\log (\frac{\rho _{N}(t,X^{N})}{\overline{\rho }_{N}(t,X^{N})} )dX^{N}\). For the proof, the authors derive the equation satisfied by the relative entropy \(\overline{\rho }_{N}\) and they compute \(\frac{d}{dt} \mathcal{H}_{N}(\rho _{N},\overline{\rho }_{N})(t)\). From these estimates, the authors derive that if \(\mathcal{H}_{N}(\rho _{N}^{0},\overline{\rho } _{N}^{0})\rightarrow 0\) as \(N\rightarrow \infty \), then over any fixed time interval \([0,T]\) \(\mathcal{H}_{N}(\rho _{N},\overline{\rho } _{N})(t)\rightarrow 0\) as \(N\rightarrow \infty \). They also give a similar estimate on \(\mathcal{H}_{N}(\rho _{N},\overline{\rho }_{N})(t)\) in the case where \(\operatorname{div}F\in L^{\infty }(\Pi ^{d})\), \(\operatorname{div}K\in L^{\infty }(\Pi ^{d})\) and that either \(K\in L^{\infty }\) or for \(d\geq 2\), \( K(-x)=-K(x)\) with \(\left| x\right| K(x)\in L^{\infty }(\Pi ^{d})\).