Diffusions and PDEs on Wasserstein Space
Feng‐Yu Wang
2019-01-01
Abstract:We propose a new type SDE, whose coefficients depend on the image of solutions, to investigate the diffusion process on the Wasserstein space $\mathcal P_2$ over $\mathbb R^d$, generated by the following time-dependent differential operator for $f\in C_b^2(\mathcal P_2)$: \begin{equation*} \begin{split}& \mathcal A_t f (\mu):= \ \frac 1 2\int_{\mathbb R^d\times \mathbb R^d} \big\langle \sigma (t,x,\mu)\sigma(t,y,\mu)^*, D^2 f(\mu)(x,y)\big\rangle \mu(d x)\mu(d y) \\ & + \int_{\mathbb R^d} \Big(\frac 1 2 \big\langle (\sigma\sigma^*)(t,x,\mu), \nabla\{Df(\mu)\}(x)\big\rangle +\big\langle b(t,x,\mu), Df(\mu)(x)\big\rangle \Big)\mu(d x),\ \mu\in \mathcal P_2, \end{split}\end{equation*} where $\langle \cdot,\cdot\rangle$ is the inner product on $\mathbb R^d$ or $\mathbb R^d\otimes\mathbb R^d$, $\nabla$ is the gradient operator on $\mathbb R^d$, $D$ is the intrinsic (or Lions) derivative on $\mathcal P_2$, and $$b: [0,\infty)\times \mathbb R^d\times\mathcal P_2\to \mathbb R^d,\ \ \sigma: [0,\infty)\times \mathbb R^d\times\mathcal P_2\to \mathbb R^{d}\otimes\mathbb R^m$$ are measurable. We study the exponential convergence of the diffusion process, and use the diffusion process to solve the following PDE $$(\partial_t+\mathcal A_t)U (t,\mu) + (V U)(t,\mu) +F(t,\mu)=0,\ \ (t,\mu)\in [0,T]\times \mathcal P_2,$$ where $V$ and $F$ are functions on $[0,T]\times \mathcal P_2.$ Moreover, the structure of the invariant probability measure is described.