Large deviations of Kac's conservative particle system and energy nonconserving solutions to the Boltzmann equation: a counterexample to the predicted rate function (Q6104012)

Given a system with \(N\) indistinguishable particles of mass \(N^{-1}\), of velocities \(V_{t}^{1},\ldots ,V_{t}^{N}\in \mathbb{R}^{d}\), the author introduces the normalized empirical measure \(\mu _{t}^{N}=N^{-1}\sum_{i}\delta _{V_{t}^{i}}\).\ For every (ordered) pair of particles with velocities \(v,v_{\star }\in\operatorname{supp}(\mu _{t}^{N})\), the velocities change according to \(v^{\prime }(v,v_{\star },\sigma )=v-((v-v_{\star })\cdot \sigma )\sigma \), \(v_{\star }^{\prime }(v,v_{\star },\sigma )=v_{\star }+((v-v_{\star })\cdot \sigma )\sigma \), at rate \( 2B(v-v_{\star },\sigma )/N\), where \(B:\mathbb{R}^{d}\times \mathbb{S} ^{d-1}\rightarrow \lbrack 0,\infty )\) is a collision kernel here chosen as \( B(v,\sigma )=1+\left\vert v\right\vert \) (regularized hard spheres), or \( B(v,\sigma )=1\) (Maxwell molecules). The Kac process may be written as a stochastic perturbation of the Boltzmann equation, for any bounded \(f\), through: \(\left\langle f,\mu _{t}^{N}\right\rangle =\left\langle f,\mu _{0}^{N}\right\rangle +\int_{0}^{t}\left\langle f,Q\mu _{s}^{N}\right\rangle ds+M_{t}^{N,f}\), where \(M_{t}^{N,f}\) is a martingale of quadratic variation \( [M^{N,f}]_{t}=O(N^{-1})\). The Gaussian moments are defined as: \(\mathcal{E} _{z}(\mu _{0}^{\star })=\int_{\mathbb{R}^{d}}e^{z\left\vert v\right\vert ^{2}}\mu _{0}^{\star }(dv)\in \lbrack 1,\infty ]\) and they are supposed to satisfy a Gaussian upper bound: there exists \(z_{1}>0\) such that \(\mathcal{E} _{z}(\mu _{0}^{\star })<\infty \), a Gaussian lower bound: there exists \( z_{2}<\infty \) such that \(\mathcal{E}_{z}(\mu _{0}^{\star })<\infty \) for \( z<z_{2}\), and \(\mathcal{E}_{z}(\mu _{0}^{\star })\rightarrow \infty \) as \( z\uparrow z_{2}\). Further \(\mu _{0}^{\star }\) is supposed to have a continuous density \(f_{0}^{\star }\) with respect to the Lebesgue measure, and for some \(z_{3}\in (0,\infty )\) and \(c>0\), \(f_{0}^{\star }\geq ce^{-z_{3}\left\vert v\right\vert ^{2}}\). To the Kac process, the author associates an auxiliary empirical flux \(w_{t}^{N}\) defined on the parameter space of collisions \(E=[0,T]\times \mathbb{R}^{d}\times \mathbb{R}^{d}\times \mathbb{S}^{d-1}\), setting \(w_{t}^{N}=0\) and \(w_{t}^{N}=w_{t-}^{N}+\frac{1}{N }\delta _{(t,v,v_{\star },\sigma )}\) at times \(t\) where there is a collision, choosing one possible assignment \((v,v_{\star },\sigma )\) uniformly at random between the \(4^{2}\) possible choices of collision parameters. The author then proposes a rate function previously identified by \textit{C. Léonard} in [Probab. Theory Relat. Fields 101, No. 1, 1--44 (1995; Zbl 0839.60031)]. For \(\mu _{\bullet }\in \mathcal{D}\), he defines \( \overline{m}_{\mu }\in \mathcal{M}(E)\) by \(\overline{m}_{\mu }(dt,dv,dv_{\star },d\sigma )=B(v-v_{\star })dt\mu _{t}(dv)\mu _{t}(dv_{\star })d\sigma \), where \(\mathcal{M}(E)\) is the space of finite Borel measures on \(E\).\ An element \((\mu _{\bullet },w)\in \mathcal{D}\times \mathcal{M}(E)\) is a measure-flux pair if \(w\ll \overline{m}_{\mu }\) and if they solve the continuity equation: for all \(0\leq t\leq T\), \(\mu _{t}=\mu _{0}+\int_{E}\Delta (v,v_{\star },\sigma )1_{s\leq t}w(ds,dv,dv_{\star },d\sigma )\). Here \(\mathcal{D}\) is the Skorokhod space defined as: \( \mathcal{D}=\{\mu _{\bullet }\in D([0,T],(\mathcal{P}_{2},W):\sup_{t\leq T}\left\vert v\right\vert ^{2},\mu _{t}<\infty \}\), equipped with a metric inducing the Skorokhod \(J_{1}\)-topology, \(\mathcal{P}_{2}\) being the space of probability measures on \(\mathbb{R}^{d}\) with finite second moment, equipped with the Monge-Kantorovich-Wasserstein distance. The dynamic cost of a trajectory \((\mu _{\bullet },w)\in \mathcal{D}\times \mathcal{M}(E)\) is defined as \(\mathcal{J}(\mu _{\bullet },w)=\int_{E}\tau (\frac{dw}{d \overline{m}_{\mu }})\overline{m}_{\mu }(ds,dv,dv_{\star },d\sigma )\) if \( (\mu _{\bullet },w)\) is a measure-flux pair, \(\infty \) else, where \(\tau :[0,\infty ]\rightarrow \lbrack 0,\infty ]\) is the function \(\tau (k)=k\log k-k+1\), and the full rate function is defined as: \(\mathcal{I}(\mu _{\bullet },w)=H(\mu _{0}\mid \mu _{0}^{\ast })+\mathcal{J}(\mu _{\bullet },w)\), with \(H(\mu _{0}\mid \mu _{0}^{\ast })=\int_{\mathbb{R}^{d}}\frac{ d\mu _{0}}{d\mu _{0}^{\ast }}\log(\frac{d\mu _{0}}{d\mu _{0}^{\ast }})\mu _{0}^{\ast }(dv)\) if \(\mu _{0}\ll \mu _{0}^{\ast }\); \(\infty \) else. The first main result proves that for \(N\geq 2\), for every Kac process and its flux \((\mu _{\bullet }^{N},w^{N})\), with particles drawn initially from \(\mu _{0}^{\star }\) satisfying the above hypotheses and \(\mathcal{I}\) the above rate function, then for all \(\mathcal{A}\subset \mathcal{D}\times \mathcal{M} (E)\) closed: \(\limsup_{N}\frac{1}{N}\log\mathbb{P}((\mu _{\bullet }^{N},w^{N})\in \mathcal{A})\leq -\inf\{\mathcal{I}(\mu _{\bullet },w):(\mu _{\bullet },w)\in \mathcal{A}\}\), and for all \(\mathcal{U}\subset \mathcal{D} \times \mathcal{M}(E)\) open: \(\liminf_{N}\frac{1}{N}\log\mathbb{P}((\mu _{\bullet }^{N},w^{N})\in \mathcal{U})\geq -\inf\{\mathcal{I}(\mu _{\bullet },w):(\mu _{\bullet },w)\in \mathcal{U}\cap \mathcal{R}\}\), where \(\mathcal{R }=\{(\mu _{\bullet },w)\in \mathcal{D}\times \mathcal{M}(E):\left\langle 1+\left\vert v\right\vert ^{2}+\left\vert v_{\ast }\right\vert ^{2},w\right\rangle <\infty \}\). The second main result proves estimates in the case where \(B\) is the regularized hard spheres kernel. For the proofs, the author gathers results concerning large deviations and related problems. He proves the upper bound using a variational formulation of the rate function \(\mathcal{I}\). He proves the lower bound using an approximation argument for paths belonging to \(\mathbb{R}\) and a standard ``left tilting'' argument. The proof of the last main result is based on properties of the Kac process and a careful analysis of Cramér bounds.