Efficient adaptive multilevel stochastic Galerkin approximation using implicit a posteriori error estimation

TL;DR

This paper introduces an efficient adaptive multilevel stochastic Galerkin method for elliptic PDEs with infinitely many parameters, using implicit a posteriori error estimation to guide adaptive enrichment without tuning parameters.

Contribution

The paper presents a novel adaptive SGFEM algorithm with multilevel structure and implicit error estimation, improving efficiency and practicality in high-dimensional parameter spaces.

Findings

Error decay rate matches finite element method for parameter-free problems

Algorithm adaptively activates parameters based on error estimates

Numerical experiments confirm optimal performance

Abstract

Partial differential equations (PDEs) with inputs that depend on infinitely many parameters pose serious theoretical and computational challenges. Sophisticated numerical algorithms that automatically determine which parameters need to be activated in the approximation space in order to estimate a quantity of interest to a prescribed error tolerance are needed. For elliptic PDEs with parameter-dependent coefficients, stochastic Galerkin finite element methods (SGFEMs) have been well studied. Under certain assumptions, it can be shown that there exists a sequence of SGFEM approximation spaces for which the energy norm of the error decays to zero at a rate that is independent of the number of input parameters. However, it is not clear how to adaptively construct these spaces in a practical and computationally efficient way. We present a new adaptive SGFEM algorithm that tackles elliptic PDEs with parameter-dependent coefficients quickly and efficiently. We consider approximation spaces with a multilevel structure---where each solution mode is associated with a finite element space on a potentially different mesh---and use an implicit a posteriori error estimation strategy to steer the adaptive enrichment of the space. At each step, the components of the error estimator are used to assess the potential benefits of a variety of enrichment strategies, including whether or not to activate more parameters. No marking or tuning parameters are required. Numerical experiments for a selection of test problems demonstrate that the new method performs optimally in that it generates a sequence of approximations for which the estimated energy error decays to zero at the same rate as the error for the underlying finite element method applied to the associated parameter-free problem.