Kolmogorov equations and weak order analysis for SPDEs with nonlinear diffusion coefficient (Q1791739): Difference between revisions

The authors consider the Kolmogorov equation associated to a stochastic partial differential equation \[ dX=(\Delta X+F_1(X)+\partial_\xi F_2(X))\,dt+\sigma(X)\,dW \] on $H=L^2(0,1)$ subject to homogeneous Dirichlet boundary conditions, where $F_1,F_2,\sigma\in C^3_b(\mathbb{R})$ and $W$ is a cylindrical Wiener process on $H$. One defines $G(\cdot)=F_1(\cdot)+\partial_xF_2(\cdot)$ and smoothed mappings $G_\delta(\cdot)=e^{\delta\Delta}G(e^{\delta\Delta}\cdot)$, $\sigma_\delta=e^{\delta\Delta}\sigma(e^{\delta\Delta}\cdot)e^{\delta\Delta}$, $\varphi_\delta(\cdot)=\varphi(e^{\delta\Delta}\cdot)$ for $\delta\in(0,1)$ and a fixed real function $\varphi$ defined on $L^p(0,1)$ for some $p\in[2,\infty)$ such that $\varphi_\delta\in C^3(H)$ and there exist $q\in[2,\infty)$, $K\ge 0$ and $C(p,q,K)$ satisfying \[\begin{aligned} |D\varphi_\delta(x)h_1|&\le C(p,q,K)(1+|x|_{L^p})^K|h_1|_{L^q} \\ |D^2\varphi_\delta(x)(h_1,h_2)|&\le C(p,q,K)(1+|x|_{L^p})^K|h_1|_{L^q}|h_2|_{L^q} \\ |D^3\varphi_\delta(x)(h_1,h_2,h_3)|&\le C(p,q,K)(1+|x|_{L^p})^K|h_1|_{L^q}|h_2|_{L^q}|h_3|_{L^q} \end{aligned}\] for every $x\in L^p(0,1)$, $h_1,h_2,h_3\in L^q(0,1)$ and $\delta\in(0,1)$. An approximating equation \[ dX^\delta=(\Delta X^\delta+G_\delta(X^\delta))\,dt+\sigma_\delta(X^\delta)\,dW \] is considered and the authors define $u_\delta(t,x)=\mathbb{E}\,[\varphi_\delta(X^\delta(t,x))]$ as an approximation of the solution $u_0(t,x)=\mathbb{E}\,[\varphi(X(t,x))]$ of the Kolmogorov equation. Then it is proved that, for every $\beta\in[0,1)$ and $T>0$, there exists $C_\beta(T)$ such that \[ |Du_\delta(t,x)h|\le\frac{C_\beta(T)}{t^\beta}(1+|x|_{L^{\text{max}\{p,2q\}}})^{K+1}|(-\Delta)^{-\beta}h|_{L^{2q}} \] holds for every $\delta\in[0,1)$, $t\in(0,T]$, $x\in L^p(0,1)$ and $h\in L^q(0,1)$. For second derivatives, it is proved that, for every $\beta,\gamma\in[0,\frac 12)$ and $T>0$, there exists $C_{\beta,\gamma}(T)$ such that \[ |D^2u_\delta(t,x)(h_1,h_2)|\le\frac{C_{\beta,\gamma}(T)}{t^{\beta+\gamma}}(1+|x|_{L^{\text{max}\{p,2q\}}})^{K+1}|(-\Delta)^{-\beta}h_1|_{L^{4q}}|(-\Delta)^{-\gamma}h_2|_{L^{4q}} \] holds for every $\delta\in[0,1)$, $t\in(0,T]$, $x\in L^p(0,1)$ and $h_1,h_2\in L^{4q}(0,1)$. An analogous estimate for the third derivative of $u_\delta$ is established too. \par In the second part of the paper, a discrete equation \[ X_{n+1}-X_n=\delta(\Delta X_{n+1}+G(X_n))+\sigma(X_n)(W((n+1)\delta)-W(n\delta)),\qquad X_0=x \] is considered for $0\le n<N$ and $\delta\in(0,T)$ such that $N=\frac T\delta$. The authors prove that, for every $\kappa\in(0,\frac 12)$ and $\delta_0\in(0,1)$, there exists $C_\kappa(T,\delta_0)$ such that \[ \left|\mathbb{E}\,\varphi(X(T))-\mathbb E\,\varphi(X_N)\right|\le C_\kappa(T,\delta_0)(1+|x|_{L^{\max\{p,8q\}}})^{K+3}\delta^{\frac 12-\kappa} \] holds for every $\delta\in(0,\delta_0)$ with $N=\frac T\delta\in\mathbb{N}$, and every $x\in L^p(0,1)\cap L^{8q}(0,1)$.