QMC algorithms with product weights for lognormal-parametric, elliptic PDEs

Abstract

We survey recent convergence rate bounds for single-level and multilevel QMC Finite Element (FE for short) algorithms for the numerical approximation of linear, second order elliptic PDEs in divergence form in a bounded, polygonal domain D. The diffusion coefficient a is assumed to be an isotropic, log-Gaussian random field (GRF for short) in D. The representation of the GRF Z= log a is assumed affine-parametric with i.i.d. standard normal random variables, and with locally supported functions ψj characterizing the spatial variation of the GRF Z. The goal of computation is the evaluation of expectations (i.e., of so-called “ensemble averages”) of (linear functionals of) the random solution. The QMC rules employed are randomly shifted lattice rules proposed in Nichols, Kuo (J Complex 30:444–468, 2014, [19]) as used and analyzed previously in a similar setting (albeit for globally in D supported spatial representation functions ψj as arise in Karhunen-Loève expansions) in Graham et al. (Numer Math 131:329–368, 2015, [9]), Kuo et al. (Math Comput 86:2827–2860, 2017, [14]). The multilevel QMC-FE approximation QL∗ analyzed here for locally supported ψj was proposed first in Kuo, Schwab, Sloan (Found Comput Math 15:411–449, 2015, [17]) for affine-parametric operator equations. As shown in Gantner, Herrmann, Schwab (SIAM J Numer Anal 56:111–135, 2018, [7]), Gantner, Herrmann, Schwab (Contemporary computational mathematics - a celebration of the 80th birthday of Ian Sloan. Springer, Cham, 2018, [6]), Herrmann, Schwab (QMC integration for lognormal-parametric, elliptic PDEs: local supports and product weights. Technical Report 2016-39, Seminar for Applied Mathematics, ETH Zürich, Switzerland, 2016, [10]), Herrmann, Schwab (Multilevel quasi-Monte Carlo integration with product weights for elliptic PDEs with lognormal coefficients. Technical Report 2017-19, Seminar for Applied Mathematics, ETH Zürich, Switzerland, 2017, [11]) localized supports of the ψj (which appear in multiresolution representations of GRFs Z of Lévy–Ciesielski type in D) allow for the use of product weights, originally proposed in construction of QMC rules in Sloan, Woźniakowski (J Complex 14:1–33, 1998, [23]) (cf. the survey (Dick, Kuo, Sloan in Acta Numer 22:133–288, 2013, [4]) and references there). The present results from Herrmann, Schwab (Multilevel quasi-Monte Carlo integration with product weights for elliptic PDEs with lognormal coefficients. Technical Report 2017-19, Seminar for Applied Mathematics, ETH Zürich, Switzerland, 2017, [11]) on convergence rates for the multilevel QMC FE algorithm allow for general polygonal domains D and for GRFs Z whose realizations take values in weighted spaces containing W1∞(D). Localized support assumptions on ψj are shown to allow QMC rule generation by the fast, FFT based CBC constructions in Nuyens, Cools (J Complex 22:4–28, 2006, [21]), Nuyens, Cools (Math Comput 75:903–920, 2006, [20]) which scale linearly in the integration dimension which, for multiresolution representations of GRFs, is proportional to the number of degrees of freedom used in the FE discretization in the physical domain D. We show numerical experiments based on public domain QMC rule generating software in Gantner (A generic c++ library for multilevel quasi-Monte Carlo. In: Proceedings of the Platform for Advanced Scientific Computing Conference, PASC ’16, ACM, New York, USA, pp 11:1–11:12 2016, [5]), Kuo, Nuyens (Found Comput Math 16:1631–1696, 2016, [13]).