betaFeedback

WorksMathematics

Convergence of an adaptive multilevel stochastic Galerkin FEM driven by a residual-based error estimator

Abstract

In this paper we analyze a residual-based a posteriori error estimation strategy and adaptive multilevel algorithms for a class of parametric elliptic PDEs, where the diffusion coefficient depends on countably many uncertainty parameters. Under a saturation assumption on the parametric approximation space, we propose a computable residual-based error estimator. We also introduce the notion of K-compatibility of finite element meshes, which ensures that only the triangulations with comparable mesh sizes are used, and incorporate this concept into the adaptive algorithm. We prove that the estimator is reliable and satisfies K-dependent stability and reduction properties. These results underpin our convergence analysis of the adaptive multilevel algorithm: we show that the algorithm drives both the a posteriori error estimator and the true approximation error to zero. Numerical experiments illustrate and confirm the theoretical findings.

Introduction

Model problem and state of the art

Let Ω⊂Rd\Omega\subset\mathbb{R}^{d}, d≥2d \ge2, be a bounded domain with Lipschitz boundary ∂Ω\partial\Omega and Λ:=∏m∈N[−1,1]\Lambda:= \prod_{m\in\mathbb{N}}[-1,1] be the infinite-dimensional hypercube. We consider the following parametric model problem: Given a volume force f:Ω→Rf:\Omega\to\mathbb{R}, find u⋆:Ω×Λ→Ru^{\star}:\Omega\times\Lambda\to\mathbb{R} such that

−div⁡(a∇u⋆)=fin Ω×Λsubject tou⋆=0on ∂Ω×Λ,(1)-\operatorname{div}(a\nabla u^{\star}) = f \quad\text{in } \Omega\times\Lambda\qquad\text{subject to} \qquad u^{\star}=0 \quad\text{on } \partial\Omega\times\Lambda, \tag*{(1)}

where the diffusion coefficient a:Ω×Λ→Ra:\Omega\times\Lambda\to\mathbb{R} is affine with respect to y=(ym)m∈N∈Λ\boldsymbol{y}=(y_m)_{m\in\mathbb{N}}\in\Lambda, i.e.,

a(x,y)=a0(x)+∑m∈Nam(x)ym.(2)a(x,\boldsymbol{y})=a_0(x)+\sum_{m\in\mathbb{N}}a_m(x)y_m. \tag*{(2)}

The boundary value problem (1) and other PDE-based models with stochastic inputs arise in various fields of science and engineering. Efficient numerical algorithms for solving these PDEs are paramount in the context of uncertainty quantification. The construction of such algorithms has attracted considerable attention over the last two decades in connection with, e.g., reduced basis schemes (e.g., [HUW13]), collocation methods (e.g., [BNT07]) and (quasi-)Monte Carlo sampling (e.g., [CST13]).

Stochastic Galerkin finite element methods (SGFEMs) are a useful alternative to the above mentioned sampling methods; we refer to [Xiu09; BNTT11; GWZ14] for details.

Acknowledgment. This research was funded by the Austrian Science Fund (FWF) projects 10.55776/F65 (SFB F65 “Taming complexity in PDE systems”), 10.55776/I6802 (international project I6802 “Functional error estimates for PDEs on unbounded domains”), and 10.55776/P33216 (standalone project P33216 “Computational nonlinear PDEs”). The generative AI tool Claude was used to check spelling and grammar in the manuscript, but was not involved in the development of the mathematical content or the software used in the numerical simulations. The authors take full responsibility for the content of the publication.

of the differences between SGFEM and other methods. One of the main advantages of intrusive Galerkin methods over non-intrusive sampling or collocation methods is the significantly lower number of function evaluations and solver steps; see, e.g., [HGP+08] and [Xiu09, Section 6].

SGFEM approximations combine generalized polynomial chaos (gPC) expansions in the stochastic variables/parameters with finite element discretizations of the expansion coefficients. The approximations (and the algorithms that generate them) are termed single-level if the same finite element mesh is used for all expansion coefficients, and multilevel otherwise. Single-level algorithms have emerged in [EGSZ15], while multilevel approximations have been considered in [CDS10; Git13], among other works. Flexibility is the main advantage of multilevel methods, especially in the context of adaptivity, since each spatial coefficient in the gPC expansion is allowed to reside in a finite element space associated with its own appropriately chosen mesh. Thus, multilevel algorithms can successfully tackle problems where some gPC expansion coefficients require a highly refined mesh to approximate the solution accurately, while a coarser mesh suffices for other gPC coefficients.

A posteriori error estimation strategies that steer adaptive multilevel refinement and the practical realization of the multilevel SGFEM have been considered, e.g., in [EGSZ14; CPB19; BPR21b; BPR21a]. In the latter work, under a saturation assumption on the full approximation space, convergence and rate-optimality of an adaptive multilevel algorithm are proven; in that work, the adaptive algorithm employs hierarchical parametric error indicators and two-level spatial error indicators together with combined Dörfler marking. Very recently, in [2], convergence was proven for an adaptive algorithm that utilized multilevel expansions of the diffusion coefficient and residual error indicators based on finite element frames. However, convergence of an adaptive multilevel algorithm employing residual-based error indicators as proposed in [EGSZ14] has so far remained an open problem.

Contributions

In this paper, we follow [EGSZ14] and employ residual-based error indicators to steer mesh refinement and parametric enrichment in an adaptive multilevel algorithm. At present, this type of a posteriori estimation for the parametric error component remains largely theoretical and is not directly accessible in computations. Under a saturation assumption on the parametric approximation space, we overcome this limitation.

The main difficulty in the analysis is that the spatial error indicator associated with a coefficient in the gPC expansion also depends on other gPC coefficients, which, in general, do not reside in the same finite element space. While this is not an issue for single-level approximations [EGSZ15; BPRR19], the analysis of multilevel methods requires greater care. Indeed, to ensure stability of the spatial estimator, one needs to control the ratio of the mesh sizes of triangulations associated with neighboring gPC coefficients. In [EGSZ14], this issue is addressed by introducing projection operators between different finite element spaces. However, this leads to additional projection errors exhibiting unstable behavior. We propose an alternative way to guarantee stability of the spatial estimator by introducing the notion of KK-compatibility for a set of triangulations, which ensures that their mesh sizes are compatible. We then follow [EGSZ15] and prove that the residual-based estimator is reliable and satisfies KK-dependent versions of the stability and reduction properties (see, e.g., [CKNS08; CFPP14]). Furthermore, for a suitable sequence (Kℓ)ℓ∈N(K_\ell)_{\ell\in\mathbb{N}} of compatibility degrees, we develop a novel adaptive algorithm with built-in KℓK_\ell-compatibility of all meshes employed at the ℓ\ell-th iteration of the algorithm.

In particular, we prove that the underlying multilevel refinement strategy steered by combined Dörfler marking maintains KℓK_{\ell}-compatibility of meshes throughout the adaptive loop. Finally, we show that the developed algorithm drives both the a posteriori error estimator and the true approximation error to zero.

Outline

The paper is organized as follows. Section 22 collects all necessary preliminaries and notation on function spaces, triangulations, multilevel structures, and SGFEM. In Section 33, we propose a computable and reliable residual-based a posteriori error estimator. We also propose an adaptive algorithm and, subsequently, formulate Theorem 44, the main result of this work. It states that the overall error and, under stronger assumptions on (Kℓ)ℓ∈N0(K_{\ell})_{\ell\in\mathbb{N}_0}, also the error estimator converge to zero. Sections 44–55 are devoted to the proof of Theorem 44. In Section 66, we develop multilevel refinement strategies which maintain KℓK_{\ell}-compatibility throughout the adaptive loop. Section 77 concludes the work with numerical experiments.

Preliminaries and Notation

General notation

Throughout this paper, let Ω\Omega be a bounded domain in Rd\mathbb{R}^{d}, d≥2d \ge2, with Lipschitz boundary ∂Ω\partial\Omega. By ∣⋅∣|\cdot| we denote, without any ambiguity, the absolute value of a scalar, the dd-dimensional Lebesgue volume of a measurable set, the (d−1)(d-1)-dimensional Hausdorff measure of a surface (piece), or the cardinality of a finite set. We write ∫U⋅ dx\int_U \cdot\,\mathrm{d}x for integration over a Lebesgue measurable set U⊆ΩU \subseteq\Omega. For a Lipschitz domain U⊆ΩU \subseteq\Omega, we denote by n∂U\boldsymbol{n}_{\partial U} the outward pointing unit normal vector on ∂U\partial U.

Throughout, we denote by Λ:=∏m∈N[−1,1]\Lambda:= \prod_{m\in\mathbb{N}}[-1,1] the infinite-dimensional hypercube and by B(Λ)\mathcal{B}(\Lambda) the Borel σ\sigma-algebra on Λ\Lambda. For m∈Nm \in\mathbb{N}, let πm\pi_m be a symmetric Borel probability measure on [−1,1][-1,1] and set π:=∏m∈Nπm\pi:= \prod_{m\in\mathbb{N}}\pi_m. Integration over [−1,1][-1,1] with respect to πm\pi_m is denoted by ∫[−1,1]⋅ dπm(ym)\int_{[-1,1]} \cdot\,\mathrm{d}\pi_m(y_m) and integration over Λ\Lambda with respect to π\pi is denoted by ∫Λ⋅ dπ(y)\int_{\Lambda}\cdot\,\mathrm{d}\pi(y).

For a Banach space X\mathcal{X}, we denote by ∥⋅∥X\|\cdot\|_{\mathcal{X}} the norm on X\mathcal{X} and by X∗\mathcal{X}^{*} the topological dual space of X\mathcal{X}. The duality pairing is denoted by ⟨⋅,⋅⟩X∗\langle\cdot,\cdot\rangle_{\mathcal{X}^{*}}.

For a countable set II, we introduce the standard space ℓ2(I)\ell^{2}(I) of square-summable sequences with the associated norm

∥(xi)i∈I∥ℓ2(I):=(∑i∈Ixi2)1/2for all (xi)i∈I∈ℓ2(I).\left\|(x_i)_{i\in I}\right\|_{\ell^{2}(I)} := \left(\sum_{i\in I}x_i^{2}\right)^{1/2}\quad\text{for all }(x_i)_{i\in I}\in\ell^{2}(I).

Discrete quantities are indicated by a subscript, where related objects are indexed by the same subscript, e.g., Tℓ\mathcal{T}_{\ell} denotes a triangulation of Ω\Omega with corresponding finite element functions φℓ\varphi_{\ell} etc. Similarly, we use the same subscript for quantities related to a multi-index, e.g., wℓ,νw_{\ell,\nu} is a discrete coefficient corresponding to a mesh Tℓ,ν\mathcal{T}_{\ell,\nu} and so forth. To abbreviate notation in proofs, we write A≲BA\lesssim B if and only if there holds A≤cBA\le cB with some constant c>0c>0 that does not depend on the iteration counter ℓ∈N0\ell\in\mathbb{N}_0. We write A≃BA\simeq B if A≲BA\lesssim B and B≲AB\lesssim A. Finally, we introduce the Kronecker delta

δAB:={1if A=B,0if A≠B,for any quantities A,B.\delta_{AB}:= \begin{cases} 1 & \text{if } A=B,\\ 0 & \text{if } A\ne B, \end{cases} \quad\text{for any quantities }A,B.

Function spaces

For a Lipschitz domain U⊆ΩU\subseteq\Omega, we denote by L2(U)L^{2}(U) the usual space of square-integrable functions on UU with the corresponding norm ∥⋅∥L2(U)\|\cdot\|_{L^{2}(U)} and by L∞(U)L^{\infty}(U) the space of all measurable functions vv on UU such that ∥v∥L∞(U):=ess sup⁡x∈U∣v(x)∣<\|v\|_{L^{\infty}(U)} := \operatorname*{ess\,sup}_{x\in U}|v(x)| < +∞+\infty. If it is clear from the context, we may write ∥⋅∥U\|\cdot\|_U instead of ∥⋅∥L2(U)\|\cdot\|_{L^2(U)}. For vector-valued spaces [L2(U)]n[L^2(U)]^n or [L∞(U)]n[L^\infty(U)]^n with n∈Nn \in\mathbb{N}, we will omit the superscript in unambiguous cases. We introduce the standard first-order Sobolev space

H1(Ω):={φ∈L2(Ω)∣∇φ∈L2(Ω)}H^1(\Omega) := \{\varphi\in L^2(\Omega) \mid\nabla\varphi\in L^2(\Omega)\}

with the norm ∥φ∥H1(Ω):=(∥φ∥L2(Ω)2+∥∇φ∥L2(Ω)2)1/2\|\varphi\|_{H^1(\Omega)} := (\|\varphi\|_{L^2(\Omega)}^2 + \|\nabla\varphi\|_{L^2(\Omega)}^2)^{1/2}.

Furthermore, we define the space of H1(Ω)H^1(\Omega)-functions with vanishing trace on ∂Ω\partial\Omega as

H01(Ω):={φ∈H1(Ω)∣φ∣∂Ω=0}.H_0^1(\Omega) := \{\varphi\in H^1(\Omega) \mid\varphi|_{\partial\Omega} = 0\}.

Additionally, we introduce the space

W1,∞(Ω):={φ∈L∞(Ω)∣∇φ∈L∞(Ω)}W^{1,\infty}(\Omega) := \{\varphi\in L^\infty(\Omega) \mid\nabla\varphi\in L^\infty(\Omega)\}

with the associated norm ∥φ∥W1,∞(Ω):=max⁡{∥φ∥L∞(Ω),∥∇φ∥L∞(Ω)}\|\varphi\|_{W^{1,\infty}(\Omega)} := \max\{\|\varphi\|_{L^\infty(\Omega)}, \|\nabla\varphi\|_{L^\infty(\Omega)}\}.

Bochner spaces and polynomial chaos

Given a Hilbert space H\mathcal{H} with norm ∥⋅∥H\|\cdot\|_{\mathcal{H}}, we denote by Lπ2(Λ;H)L^2_\pi(\Lambda;\mathcal{H}) the Bochner space of H\mathcal{H}-valued functions on Λ\Lambda with respect to the measure π\pi, i.e.,

Lπ2(Λ;H):={v:Λ→H | ∥v∥Lπ2(Λ;H):=(∫Λ∥v(y)∥H2 dπ(y))1/2<+∞}.L^2_\pi(\Lambda;\mathcal{H}) := \left\{v:\Lambda\to\mathcal{H} \,\middle|\, \|v\|_{L^2_\pi(\Lambda;\mathcal{H})} := \left(\int_\Lambda\|v(y)\|_{\mathcal{H}}^2 \,\mathrm{d}\pi(y)\right)^{1/2} < +\infty\right\}.

If H=R\mathcal{H} = \mathbb{R}, we write Lπ2(Λ)L^2_\pi(\Lambda) instead of Lπ2(Λ;R)L^2_\pi(\Lambda;\mathbb{R}). An orthonormal basis of Lπ2(Λ)L^2_\pi(\Lambda) is given by the following construction: For m∈Nm \in\mathbb{N}, let (Pmn)n∈N0(P_m^n)_{n\in\mathbb{N}_0} be the sequence of normalized orthogonal polynomials on [−1,1][-1,1] with respect to πm\pi_m, i.e., the degree of PmnP_m^n is nn and

∫[−1,1]Pmn(ym)Pmk(ym) dπm(ym)=δnk.(3)\int_{[-1,1]} P_m^n(y_m)P_m^k(y_m)\,\mathrm{d}\pi_m(y_m) = \delta_{nk}. \tag*{(3)}

References

  1. [1]R. A. Adams and J. J. F. Fournier. Sobolev Spaces. Academic Press, Amsterdam, 2nd edition, 2003.
  2. [2]M. Bachmayr, M. Eigel, H. Eisenmann, and I. Voulis. A convergent adaptive finite element stochastic Galerkin method based on multilevel expansions of random fields. SIAM J. Numer. Anal., 63(4):1776–1807, 2025. DOI: 10.1137/24M1649253.DOI
  3. [3]F. Bornemann, B. Erdmann, and R. Kornhuber. A posteriori error estimates for elliptic problems in two and three space dimensions. SIAM J. Numer. Anal., 33(3):1188–1204, 1996. DOI: 10.1137/0733059.DOI
  4. [4]J. Ballani and L. Grasedyck. Hierarchical tensor approximation of output quantities of parameter-dependent PDEs. SIAM/ASA J. Uncertain. Quantif., 3(1):852–872, 2015. DOI: 10.1137/140960980.DOI

Paper details

vlnc:W_6jnsgeke
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 1

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Claude

View statement
Looking for related papers…
Contents