Mean-Square Convergence Analysis of a Three-Level MGRIT Method for Linear Stochastic Evolution Equations with Additive Noise

Abstract

We investigate the mean-square convergence of a three-level multigrid reduction-in-time (MGRIT) method based on F-relaxation for linear stochastic evolution equations driven by additive Gaussian noise. The method combines exponential Euler time stepping on the finest grid with implicit Euler discretizations on the two coarser grids. The inexact solution of the intermediate-level full approximation scheme problem gives rise to coupled error recursions involving two interlevel propagator discrepancies. Under a common spectral decomposition of the drift and noise covariance operators, we establish a geometric contraction estimate in a modal mean-square norm, with a contraction factor uniform in the temporal grids and both coarsening factors. We further derive fixed-iteration temporal error bounds that are uniform in the first coarsening factor and explicitly quantify the dependence on the second. For a fixed iteration count $k$ and a fixed second coarsening factor, we prove an $O(ΔT_1^r)$ error bound relative to the finest-grid reference solution for $0<r\le k+1$ provided that $\mathcal R_{r,B}<\infty$. We also prove that the error-propagation operator is nilpotent on each fixed temporal grid, ensuring finite termination of the iteration. Numerical experiments illustrate the theoretical convergence behavior and compare temporal critical-path work under a unit-cost model.