Weak convergence and invariant measure of a full discretization for parabolic SPDEs with non-globally Lipschitz coefficients (Q2021388)

In this paper, the authors propose a full discretization to approximate the invariant measure numerically for parabolic SPDEs with non-globally Lipschitz coefficients of the type: \[ d X(t) = ( A X(t) + F( X(t)) ) dt + d W(t), \quad t > 0, \qquad \text{with} \quad X(0) = X_0, \] where $A : Dom(A) \subset {\mathbb H} := L^2({\mathcal O}) \to {\mathbb H}$ is the Laplacian operator on ${\mathcal O} = [0, L]^d$, $d \leqslant 3$, $L > 0$, under homogeneous Dirichlet boundary condition, $F$ is the Nemytskii operator of a real-valued one-sided Lipschitz function $f$, i.e., $F(X)(\xi) = f( X(\xi))$, and $W(t)$, $t \geq 0$ is a generalized $Q$-Wiener process on a filtered probability space $( \Omega, {\mathcal F}, {\mathbb P}, \{ {\mathcal F}_t \}_{ t \geq 0} )$. Note that $Q$ is a bounded, linear, self-adjoint and positive definite operator on ${\mathbb H}$ and satisfies \[ \left\Vert ( - A)^{ \frac{ \beta -1}{2} } \right\Vert_{ {\mathcal L}_2^0 } < \infty \qquad \text{with} \quad 0 < \beta \leqslant 2 \] with ${\mathcal L}_2^0 := {\mathcal L}_2 ( U_0, {\mathbb H})$ (= the space of Hilbert-Schmidt operators), $U_0 = Q^{1/2} ( {\mathbb H} )$. One of the peculiar features in their work consists in derivation of a priori estimates and regularity estimates of the numerical solution via a variational approach and Malliavin calculus. Under certain hypotheses, they present the time-independent regularity estimates for the corresponding Kolmogorov equation and the time-independent weak convergence analysis for the full discretization: \[ X_{k+1}^N = S_{\delta t} X_k^N + \delta t S_{\delta t} P^N F( X_{k+1}^N ) + S_{\delta t} P^N \delta W_k, \] where $S_{\delta t} = ( I - A \delta t)^{-1}$, $N$ is the dimension of the spectral Galerkin projection space and $\delta t$ is the time-step size, $X_0^N = P^N X_0$ and $P^N \delta W_k = P^N ( W( (k+1) \delta t ) - W( k \delta t))$. The following is a result on weak convergence rate for (2) (see the paper): Theorem 1. Assume that $X_0$ is a sufficiently smooth function. Let $f$ be a cubic polynomial. Assume that $T > 0$ and \[ \delta t_0 \in ( 0, 1 \wedge 1/ \{ ( 2 \lambda_F - 2 \lambda_1) \vee 0 \} ). \] Then for any $\phi \in C_b^2({\mathbb H})$, there exists $C( T, X_0, Q, \phi) > 0$ such that for any $\delta t \in (0, \delta t_0 ]$, $K \delta t = T$, $K \in {\mathbb N}^T$ and $N \in {\mathbb N}^T$, \[ \vert \, {\mathbb E} [ \phi( X(T)) - \phi( X_K^N) ] \, \vert \leqslant C(T, X_0, Q, \phi) ( \delta t^{\gamma} + \lambda_N^{- \gamma} ), \] where $\lambda_1$ is the smallest eigenvalue of $-A$, and $\lambda_F$ is the one-sided Lipschitz constant of $F$. Furthermore, it is shown that the $V$-uniformly ergodic invariant measure of the original system is approximated by this full discretization with weak convergence rate. The following is about the time-independent weak error estimate: for any $\phi \in C_b^2({\mathbb H})$, there exists $C(X_0, Q, \phi) > 0$ such that for $\delta t \in (0, \delta t_0]$, $K \geq 2$ and $N \in {\mathbb N}^+$, \[ \vert \, {\mathbb E} [ \phi( X( K \delta t)) - \phi( X_K^N)] \, \vert \leqslant C(X_0, Q, \phi) \{ 1 + ( K \delta t)^{- \gamma} \} ( \delta t^{\gamma} + \lambda_N^{- \gamma} ) \] holds. Lastly, several examples are provided as well that numerical experiments verify theoretical findings. \par For other related works, see, e.g., [\textit{R. Anton} et al., IMA J. Numer. Anal. 40, No. 1, 247--284 (2020; Zbl 1470.80011)] for a fully discrete approximation of stochastic heat equation, and [\textit{D. Conus} et al., Ann. Appl. Probab. 29, No. 2, 653--716 (2019; Zbl 1477.65020)] for weak convergence rates of spectral Galerkin approximations for SPDEs.