Lagrangian stochastic models with specular boundary condition (Q2338948): Difference between revisions

This paper treats the well-posedness of a conditional McKean Lagrangian stochastic model with specular boundary condition. This is a generalization of the authors' previous result in [Stochastic Processes Appl. 121, No. 12, 2751--2775 (2011; Zbl 1234.60061)], which deals with the well-posedness of McKean Lagrangian processes in the case where the domain \(D\) is the upper half-plane. In the present article, the authors consider the stochastic process \(\{ ( X_t, U_t): 0 \leqslant t \leqslant T \}\) for any arbitrary finite time \(T > 0\), whose time-evolution is given by \[ \left\{ \begin{aligned} X_t &= X_0 + \int_0^t U_s ds, \\ U_s &= U_0 + \int_0^t B[ X_s; \rho(s) ] ds + \sigma W_t + K_t, \\ K_t &= - \sum_{ 0 \leqslant s \leqslant t} 2 ( U_{s-} \cdot n_D( X_s) ) n_D( X_s) 1_{ \{ X_s \in \partial D \} }, \end{aligned} \right. \tag{1} \] where \(\rho(t)\) is the probability density of \(( X_t, U_t)\) for all \(t \in (0, T]\), \(\{ W_t \}\) is a standard \({\mathbb R}^d\)-Brownian motion, \(\sigma > 0\) is a diffusion constant, \(D\) is an open bounded domain of \({\mathbb R}^d\), and \(n_D\) is the outward normal unit vector of the boundary \(\partial D\). Here, the drift coefficient \(B\) is given by \[ B[x; \psi] := \left\{ \begin{aligned} \frac{ \int b(v) \psi( t,x, v) dv}{ \int \psi( t,x,v) d v }, & \qquad \text{whenever} \quad \int_{ {\mathbb R}^d } \psi( t,x, v) d v \neq 0 \\ 0, & \qquad \text{otherwise,} \end{aligned} \right. \tag{2} \] where \(b : {\mathbb R}^d \to {\mathbb R}^d\) is a given measurable function. (1) provides a Lagrangian model describing, at each time \(t\), the position \(X_t\) and the velocity \(U_t\) of a particle confined within \(\bar{D}\). Formally, the function \((t,x) \mapsto B[x; \rho(t)]\) in (1) corresponds to the conditional expectation \((t,x) \mapsto {\mathbb E}[ b( U_t ) | X_t = x ]\) and the velocity equation (1) is rewritten into \[ U_t = U_0 + \int_0^t {\mathbb E}[ b(U_s) | X_s ] ds + \sigma W_t + K_t. \tag{3} \] Here is the main result: \noindent Theorem. Let \(\partial D\) be a compact \(C^3\)-submanifold of \({\mathbb R}^d\). Under suitable assumptions on the initial data \(( X_0, U_0)\), boundedness of \(b\), and boundedness of a density \(\rho_0\) of the initial law \(\mu_0\), there exists a unique solution \(( X_t, U_t )\) in law to (1) in \[ \begin{aligned} \Pi_{\omega} &:= \{ {\mathbb Q} : \text{probability measure on the sample space} \\ & \qquad {\mathcal T} := C( [0, T], \bar{D} ) \times {\mathbb D} ( [0, T], {\mathbb R}^d ) \text{ such that, for all} \quad t \in [0, T], \\ & \qquad \qquad \qquad \qquad {\mathbb Q} \circ ( x(t), u(t) )^{-1} \in L^2 ( D \times {\mathbb R}^d; ( 1 + | u |^2 )^{ \alpha/2} du ) \}. \end{aligned} \] The intrinsic difficulty to the well-posedness of (1) lies in the study of the hitting times \(\{ \tau_n : n \geq 0 \}\) of the particle position \(\{ X_t \}\) on \(\partial D\), defined by \[ \left\{ \begin{aligned} \tau_n &= \inf \{ \tau_{n-1} < t \leqslant T : X_t \in \partial D \}, \quad n \geq 1, \\ \tau_0 &= 0. \end{aligned} \right. \] The leading idea in the proof consists in using PDE techniques (in addition to stochastic analysis) to construct the time-marginal densities of the nonlinear process, so that the authors may exhibit the conditional McKean Lagrangian stochastic model.