betaFeedback

WorksMathematics

Polytopic symmetric interior-penalty approximation of elliptic SPDEs driven by spatial white noise

Abstract

We analyse symmetric interior-penalty discontinuous Galerkin approximations of elliptic stochastic partial differential equations driven by spatial Gaussian white noise on quasi-uniform polytopic meshes. The low regularity of the exact random field precludes the usual SIP consistency argument. We instead restrict the same continuum white-noise functional canonically to the discontinuous polynomial space and separate the error into stochastic forcing-restriction and deterministic discretisation components. For every fixed polynomial degree k≥1k\geq1, we prove the sharp two-sided estimate ∥Y−Yh∥L2(Ξ;L2(Ω))≃h2−d/2, \|Y-Y_h\|_{L^2(Ξ;L^2(Ω))} \simeq h^{2-d/2}, giving rates hh in two dimensions and h1/2h^{1/2} in three dimensions. A finite-dimensional lower bound establishes optimality. The analysis allows polytopic elements with arbitrarily many and arbitrarily small faces, including admissible non-convex elements. In two dimensions we also treat polygons with a single reentrant corner. Although the deterministic approximation exhibits an angle-dependent loss of regularity, the singular resolvent contribution is finite rank, and the sharp first-order stochastic rate is retained. The discontinuous structure yields an exact element-local sampler for the restricted white noise. Numerical experiments on irregular polygonal, non-convex polytopic and reentrant domains support the predicted convergence behaviour.

Introduction

Spatial Gaussian white noise is a canonical model for irregular random forcing and a fundamental ingredient in stochastic partial differential equation representations of Gaussian random fields. The connection between Matérn fields and elliptic SPDEs goes back to Whittle and Matérn and underpins the now standard SPDE approach to spatial statistics [27, 25, 24, 22].

We consider the representative problem

(−Δ+κ2)Y=Win Ω,Y∣∂Ω=0,(1)(-\Delta+\kappa^2)Y=\mathcal{W}\quad\text{in }\Omega,\qquad Y|_{\partial\Omega}=0, \tag*{(1)}

where W\mathcal{W} denotes spatial Gaussian white noise. The aim is to approximate YY by the classical symmetric interior-penalty discontinuous Galerkin method on general polytopic meshes, while retaining the original continuum white noise rather than replacing it by an independently regularised forcing.

Numerical approximation of SPDEs with additive or white-noise forcing has a substantial history. Early finite-element, finite-difference and lattice analyses include [1, 17, 20]. For conforming finite element methods, strong and weak convergence, fractional elliptic SPDEs and finite-dimensional representations of the driving noise have subsequently been studied in [28, 29, 2, 9, 3]. In particular, sharp strong rates of order h2−d/2h^{2-d/2} are known for conforming approximations of elliptic problems driven by spatial white noise [29, 9, 3]. Related operator and Hilbert–Schmidt formulations for Whittle–Matérn fields are developed in [11].

The representation and coupling of the Gaussian forcing form a separate computational question. Mass-matrix constructions for finite element white noise and multilevel couplings, including couplings between non-nested meshes, are developed in [12, 19, 6].

The principal analytical difficulty in (1) is the low spatial regularity of its solution. White noise has regularity below H−d/2(Ω)H^{-d/2}(\Omega), and the elliptic resolvent gives, in the regular case,

Y∈L2(Ξ;H2−d/2−ε(Ω))∀ε>0.Y\in L^2(\Xi;H^{2-d/2-\varepsilon}(\Omega))\qquad\forall\varepsilon>0.

Thus, in both dimensions considered here, the exact field need not belong to H01(Ω)H_0^1(\Omega) and the usual energy-space formulation is not the natural starting point. Instead, the SPDE is defined through the elliptic resolvent acting on an isonormal Gaussian process [2, 3, 23].

This lack of regularity is particularly relevant for an interior penalty method. Classical SIP consistency is usually expressed by inserting the exact solution into a broken bilinear form containing elementwise gradients and normal traces. Such quantities need not be defined for YY. Rather than assigning a DG differential structure to the rough random field, we avoid this step completely. We first restrict WW to the discontinuous polynomial space and solve the corresponding continuum elliptic problem. This produces the decomposition

Y−Yh=(Y−Y(h))+(Y(h)−Yh),Y-Y_h = \left(Y-Y^{(h)}\right)+\left(Y^{(h)}-Y_h\right),

in which the first term measures only restriction of the stochastic forcing, while the second is an ordinary deterministic SIP error with an L2(Ω)L^2(\Omega) right-hand side. The stochastic and deterministic parts can therefore be analysed using the tools appropriate to each.

DG discretisations of white-noise-driven problems have previously been considered in more structured settings. Existing analyses include LDG approximation of elliptic SPDEs in two and three dimensions with piecewise-constant regularisations of white noise on Cartesian meshes, where the stochastic error is controlled through spatial increments of the Green function [28], as well as finite element and DG approximations of white-noise-driven Helmholtz and scattering problems [14, 15]. Here we use the classical SIP method on general polytopic meshes, couple every discretisation to the same continuum white-noise functional, and seek sharp strong convergence rates without introducing a separate regularisation of the driving noise.

Our first main result establishes the sharp strong convergence rate

∥Y−Yh∥L2(Ξ;L2(Ω))≃h2−d/2(2)\lVert Y-Y_h\rVert_{L^2(\Xi;L^2(\Omega))} \simeq h^{2-d/2} \tag*{(2)}

for fixed polynomial degree on quasi-uniform polytopic meshes. A matching finite-dimensional lower bound shows that these rates are optimal for approximation spaces of comparable dimension.

Our second main result extends the analysis in two dimensions to polytopic domains with a single reentrant corner. Although the deterministic SIP approximation reflects the associated loss of elliptic regularity, we prove that the stochastic approximation retains the sharp first-order rate

∥Y−Yh∥L2(Ξ;L2(Ω))≃h.(3)\lVert Y-Y_h\rVert_{L^2(\Xi;L^2(\Omega))} \simeq h. \tag*{(3)}

The proof exploits the finite-dimensional structure of the corner singularity and shows that it does not alter the leading stochastic approximation error.

The analysis is carried out on general polytopic meshes. Elements are uniformly star-shaped with respect to a ball in their kernel, but may have arbitrarily many and arbitrarily small faces and may be non-convex. The resulting estimates are uniform with respect to the number and relative sizes of the faces [4, 8, 5, 6].

A further computational consequence of the discontinuous formulation is that the canonical restricted white-noise load can be sampled exactly from independent element-local Gaussian vectors. The same construction gives exact couplings across agglomerated mesh hierarchies and is also used in the multilevel Monte-Carlo study [10].

The numerical experiments examine each of these features. Direct continuum-error calculations on regular two- and three-dimensional domains test the rates in (2) across several polytopic mesh families. On reentrant domains, a corner-angle sweep tests the robustness of the first-order stochastic rate in (3) with respect to the strength of the corner singularity.

The remainder of the paper is organised as follows. Section 2 introduces the white-noise-driven elliptic problem and its spectral representation. Section 3 develops the polytopic SIP discretisation, the canonical white-noise restriction and the sharp strong convergence theory. Section 4 treats the reentrant-corner case. Section 5 presents the numerical experiments.

Problem setup

Let Ω⊂Rd\Omega\subset\mathbb{R}^d, d∈{2,3}d \in\{2,3\}, be a bounded polytopic Lipschitz domain, and let (Ξ,F,P)(\Xi,\mathcal{F},\mathbb{P}) be a complete probability space. We write Lp(Ξ;X)L^p(\Xi;X) for the usual Bochner spaces and denote expectation by E\mathbb{E}.

For κ>0\kappa>0, define

a(u,v):=(∇u,∇v)L2(Ω)+κ2(u,v)L2(Ω),u,v∈H01(Ω),(4)a(u,v):=(\nabla u,\nabla v)_{L^2(\Omega)}+\kappa^2(u,v)_{L^2(\Omega)},\qquad u,v\in H_0^1(\Omega), \tag*{(4)}

and let L\mathcal{L} be the positive self-adjoint operator in L2(Ω)L^2(\Omega) associated with aa. Thus, L\mathcal{L} is the Dirichlet realisation of −Δ+κ2-\Delta+\kappa^2. We consider the white-noise-driven problem

LY=W.(5)\mathcal{L}Y=W. \tag*{(5)}

The principal analysis is developed under the following deterministic regularity assumption.

Assumption 2.1 (Elliptic regularity). The Dirichlet realisation of L\mathcal{L} satisfies

D(L)=H2(Ω)∩H01(Ω)(6)D(\mathcal{L})=H^2(\Omega)\cap H_0^1(\Omega) \tag*{(6)}

and

∥z∥H2(Ω)≤Creg∥Lz∥L2(Ω)∀z∈D(L).(7)\|z\|_{H^2(\Omega)}\le C_{\mathrm{reg}}\|\mathcal{L}z\|_{L^2(\Omega)} \qquad\forall z\in D(\mathcal{L}). \tag*{(7)}

Assumption 2.1 holds, in particular, on bounded convex polytopic domain. The analysis uses (7) directly rather than convexity. Note in §4 we replace this assumption by an explicit corner-singularity decomposition on reentrant polygons.

Since Ω\Omega is bounded, L−1:L2(Ω)→L2(Ω)\mathcal{L}^{-1}:L^2(\Omega)\to L^2(\Omega) is compact. Consequently, L\mathcal{L} admits an L2L^2-orthonormal eigensystem {λj,ϕj}j≥1\{\lambda_j,\phi_j\}_{j\ge1} satisfying

Lϕj=λjϕj,(ϕi,ϕj)L2(Ω)=δij.(8)\mathcal{L}\phi_j=\lambda_j\phi_j,\qquad (\phi_i,\phi_j)_{L^2(\Omega)}=\delta_{ij}. \tag*{(8)}

By Weyl’s law, there exist constants cλ,Cλ>0c_\lambda,C_\lambda>0 such that

cλj2/d≤λj≤Cλj2/d,j≥1.(9)c_\lambda j^{2/d}\le\lambda_j\le C_\lambda j^{2/d}, \qquad j\ge1. \tag*{(9)}

For s≥0s\ge0, let HLs\mathcal{H}_{\mathcal{L}}^s be the Hilbert space

HLs:={v∈L2(Ω):∑j=1∞λjs∣(v,ϕj)L2(Ω)∣2<∞},(10)\mathcal{H}_{\mathcal{L}}^s:=\left\{v\in L^2(\Omega):\sum_{j=1}^{\infty}\lambda_j^s\left|(v,\phi_j)_{L^2(\Omega)}\right|^2<\infty\right\}, \tag*{(10)}

equipped with the norm

∥v∥HLs2:=∑j=1∞λjs∣(v,ϕj)L2(Ω)∣2.(11)\|v\|_{\mathcal{H}_{\mathcal{L}}^s}^2:=\sum_{j=1}^{\infty}\lambda_j^s\left|(v,\phi_j)_{L^2(\Omega)}\right|^2. \tag*{(11)}

In particular, HL0=L2(Ω)\mathcal{H}_{\mathcal{L}}^0=L^2(\Omega) and HL2=D(L)\mathcal{H}_{\mathcal{L}}^2=D(\mathcal{L}) with equivalent norms. Under Assumption 2.1, interpolation yields

HLs↪Hs(Ω),0≤s≤2.(12)\mathcal{H}_{\mathcal{L}}^s\hookrightarrow H^s(\Omega), \qquad0\le s\le2. \tag*{(12)}

Spatial white noise and the elliptic solution

We regard WW as an isonormal Gaussian process over L2(Ω)L^2(\Omega). It is a centred Gaussian linear mapping W:L2(Ω)→L2(Ξ)W:L^2(\Omega)\to L^2(\Xi) satisfying

E[W(v)W(w)]=(v,w)L2(Ω)∀v,w∈L2(Ω).(13)\mathbb{E}[W(v)W(w)]=(v,w)_{L^2(\Omega)} \qquad\forall v,w\in L^2(\Omega). \tag*{(13)}

In particular,

ξj:=W(ϕj),j≥1,(14)\xi_j:=W(\phi_j),\qquad j\ge1, \tag*{(14)}

are independent standard Gaussian random variables. The solution of (5) is defined spectrally by

Y:=∑j=1∞λj−1ξjϕj.(15)Y:=\sum_{j=1}^{\infty}\lambda_j^{-1}\xi_j\phi_j. \tag*{(15)}

Theorem 2.3 (Regularity of the white-noise solution). Let d∈{2,3}d\in\{2,3\} and suppose that the eigenvalues of L\mathcal{L} satisfy (9). Then the series (15) converges in L2(Ξ;L2(Ω))L^2(\Xi;L^2(\Omega)) and defines the unique L2(Ω)L^2(\Omega)-valued solution of (5) in the resolvent sense. Moreover,

Y∈L2(Ξ;HLs)∀s<2−d2,(16)Y\in L^2(\Xi;\mathcal{H}_{\mathcal{L}}^s) \qquad\forall s<2-\frac{d}{2}, \tag*{(16)}

whereas

E[∥Y∥HL2−d/22]=∞.(17)\mathbb{E}\left[\|Y\|_{\mathcal{H}_{\mathcal{L}}^{2-d/2}}^2\right]=\infty. \tag*{(17)}

If Assumption 2.1 also holds, then

Y∈L2(Ξ;Hs(Ω))∀0≤s<2−d2.(18)Y \in L^2(\Xi;H^s(\Omega)) \qquad\forall0 \le s < 2-\frac{d}{2}. \tag*{(18)}

Thus the available Sobolev regularity is H1−εH^{1-\varepsilon} in two dimensions and H1/2−εH^{1/2-\varepsilon} in three dimensions.

Proof. This is the case β=1\beta=1 and eigenvalue-growth exponent α=2/d\alpha=2/d of [3], Proposition 2.3 and Remark 2.4. Indeed, for every s∈Rs \in\mathbb{R},

E[∥Y∥HLs2]=∑j=1∞λjs−2.(19)\mathbb{E}\left[\lVert Y\rVert_{H^s_{\mathcal{L}}}^2\right]=\sum_{j=1}^{\infty}\lambda_j^{s-2}. \tag*{(19)}

By (9), this series has the same convergence behaviour as ∑j=1∞j2(s−2)/d\sum_{j=1}^{\infty}j^{2(s-2)/d}, and is therefore finite precisely when s<2−d/2s<2-d/2. The choice s=0s=0 gives convergence in L2(Ξ;L2(Ω))L^2(\Xi;L^2(\Omega)), while s=2−d/2s=2-d/2 gives the critical logarithmic divergence. Finally, (18) follows from (12). □

Polytopic symmetric interior-penalty discretisation

The convergence analysis separates restriction of the stochastic forcing from approximation of the elliptic resolvent. Let WhW_h be the canonical restriction of spatial white noise to a finite element space VhV_h, defined below, and let Y(h)Y^{(h)} solve LY(h)=Wh\mathcal{L}Y^{(h)}=W_h. Then

Y−Yh=(Y−Y(h))+(Y(h)−Yh).(20)Y-Y_h=\left(Y-Y^{(h)}\right)+\left(Y^{(h)}-Y_h\right). \tag*{(20)}

The first term measures restriction of the white noise, while the second is a deterministic SIP error with an L2L^2 right-hand side.

Polytopic meshes and discrete spaces

Let Th\mathcal{T}_h be a partition of Ω\Omega into non-overlapping open polytopes KK satisfying

Ω‾=⋃K∈ThK‾,K∩K′=∅for K≠K′.(21)\overline{\Omega}=\bigcup_{K\in\mathcal{T}_h}\overline{K},\qquad K\cap K'=\varnothing\quad\text{for }K\ne K'. \tag*{(21)}

For K∈ThK\in\mathcal{T}_h, let hK:=diam⁡(K)h_K:=\operatorname{diam}(K) and set

h:=max⁡K∈ThhK.(22)h:=\max_{K\in\mathcal{T}_h}h_K. \tag*{(22)}

Let Fh=Fhi∪Fhb\mathcal{F}_h=\mathcal{F}_h^i\cup\mathcal{F}_h^b be the set of open (d−1)(d-1)-dimensional mesh faces. For F=∂K−∩∂K+∈FhiF=\partial K^-\cap\partial K^+\in\mathcal{F}_h^i, fix a unit normal nF\boldsymbol{n}_F pointing from K−K^- to K+K^+; on boundary faces, nF\boldsymbol{n}_F is the outward unit normal. We use the face scale

hF:={2hK−hK+hK−+hK+,F=∂K−∩∂K+∈Fhi,hK,F⊂∂K∈Fhb.(23)h_F:= \begin{cases} \frac{2h_{K^-}h_{K^+}}{h_{K^-}+h_{K^+}}, & F=\partial K^-\cap\partial K^+\in\mathcal{F}_h^i,\\ h_K, & F\subset\partial K\in\mathcal{F}_h^b. \end{cases} \tag*{(23)}

For a piecewise scalar field vv, define on interior faces

⟦v⟧:=v−−v+,{ ⁣{v} ⁣}:=12(v−+v+),(24)\llbracket v\rrbracket:=v^--v^+,\qquad\{\!\{v\}\!\}:=\frac{1}{2}(v^-+v^+), \tag*{(24)}

and set ⟦v⟧:=v\llbracket v\rrbracket:=v and { ⁣{v} ⁣}:=v\{\!\{v\}\!\}:=v on boundary faces. Averages of vector fields are taken componentwise.

3.2. Assumption (Polytopic mesh regularity). There exists ϑ∗∈(0,1]\vartheta_*\in(0,1], independent of hh, such that every K∈ThK\in\mathcal{T}_h contains a ball B(xK,ρK)B(x_K,\rho_K) in its kernel satisfying

ρK≥ϑ∗hK.(25)\rho_K\ge\vartheta_*h_K. \tag*{(25)}

Equivalently,

B(xK,ρK)⊂K,[x,y]⊂K∀x∈B(xK,ρK), ∀y∈K.(26)B(x_K,\rho_K)\subset K,\qquad[x,y]\subset K\quad\forall x\in B(x_K,\rho_K),\ \forall y\in K. \tag*{(26)}

Moreover, the mesh family is quasi-uniform in that there exists cqu>0c_{\mathrm{qu}}>0, independent of hh, such that

cquh≤hK≤h∀K∈Th.(27)c_{\mathrm{qu}}h\le h_K\le h\qquad\forall K\in\mathcal{T}_h. \tag*{(27)}

Assumption 3.2 implies

∣K∣≃hKd(28)|K| \simeq h_K^d \tag*{(28)}

with constants depending only on ϑ∗\vartheta_* and dd. It also implies the geometric condition required by [4], Lemma 2.2. Indeed, triangulating each face and joining the resulting (d−1)(d-1)-simplices FTF_T to xKx_K gives

d∣T∣∣FT∣=dist⁡(xK,aff⁡(FT))≥ρK≥ϑ∗hK.(29)\frac{d|T|}{|F_T|}=\operatorname{dist}(x_K,\operatorname{aff}(F_T))\ge\rho_K\ge\vartheta_*h_K. \tag*{(29)}

Hence condition (7) of [4] holds uniformly. In particular, no uniform lower bound is required on the size of an individual face, and the number of faces per element may be arbitrarily large.

The quasi-uniformity condition (27) is actually not needed for the local estimates below; it will be used subsequently to pass from local mesh scales to hh and to obtain Nh≃h−dN_h \simeq h^{-d}.

Introduce the broken Sobolev space

H1(Th):={v∈L2(Ω):v∣K∈H1(K)∀K∈Th},(30)H^1(\mathcal{T}_h):=\left\{v\in L^2(\Omega):v|_K\in H^1(K)\quad\forall K\in\mathcal{T}_h\right\}, \tag*{(30)}

with broken gradient

(∇hv)∣K:=∇(v∣K).(31)(\nabla_hv)|_K:=\nabla(v|_K). \tag*{(31)}

3.3. Lemma (Trace, inverse and approximation estimates). Under the first part of Assumption 3.2.

∥v∥L2(∂K)2≲hK−1∥v∥L2(K)2+hK∥∇v∥L2(K)2∀v∈H1(K).(32)\|v\|_{L^2(\partial K)}^2\lesssim h_K^{-1}\|v\|_{L^2(K)}^2+h_K\|\nabla v\|_{L^2(K)}^2\qquad\forall v\in H^1(K). \tag*{(32)}

For every fixed polynomial degree k≥1k\ge1,

∥∇vh∥L2(K)≲hK−1∥vh∥L2(K),(33)\|\nabla v_h\|_{L^2(K)}\lesssim h_K^{-1}\|v_h\|_{L^2(K)}, \tag*{(33)}
∥vh∥L2(∂K)≲hK−1/2∥vh∥L2(K)(34)\|v_h\|_{L^2(\partial K)}\lesssim h_K^{-1/2}\|v_h\|_{L^2(K)} \tag*{(34)}

for all vh∈Pk(K)v_h\in\mathbb{P}_k(K). Consequently,

∑F⊂∂KhK∥τh⋅nF∥L2(F)2≲∥τh∥L2(K)2∀τh∈[Pk(K)]d.(35)\sum_{F\subset\partial K}h_K\|\tau_h\cdot n_F\|_{L^2(F)}^2\lesssim\|\tau_h\|_{L^2(K)}^2\qquad\forall\tau_h\in[\mathbb{P}_k(K)]^d. \tag*{(35)}

Let Πh\Pi_h be the elementwise L2L^2-orthogonal projection, defined by

(Πhv)∣K=ΠK(v∣K),ΠK:L2(K)⟶Pk(K).(\Pi_hv)|_K=\Pi_K(v|_K),\qquad\Pi_K:L^2(K)\longrightarrow\mathbb{P}_k(K).

Then, for every v∈H2(K)v\in H^2(K),

∥v−ΠKv∥L2(K)+hK∥∇(v−ΠKv)∥L2(K)+hK2∥D2(v−ΠKv)∥L2(K)+hK1/2∥v−ΠKv∥L2(∂K)+hK3/2∥∇(v−ΠKv)∥L2(∂K)≲hK2∥v∥H2(K).(36)\|v-\Pi_Kv\|_{L^2(K)}+h_K\|\nabla(v-\Pi_Kv)\|_{L^2(K)}+h_K^2\|D^2(v-\Pi_Kv)\|_{L^2(K)} +h_K^{1/2}\|v-\Pi_Kv\|_{L^2(\partial K)}+h_K^{3/2}\|\nabla(v-\Pi_Kv)\|_{L^2(\partial K)} \lesssim h_K^2\|v\|_{H^2(K)}. \tag*{(36)}

Consequently,

∥v−Πhv∥L2(Ω)≲h2∥v∥H2(Ω)∀v∈H2(Ω).(37)\|v-\Pi_hv\|_{L^2(\Omega)}\lesssim h^2\|v\|_{H^2(\Omega)}\qquad\forall v\in H^2(\Omega). \tag*{(37)}

All hidden constants are independent of the number and relative sizes of the faces.

Proof. By (29), the mesh condition of [4], Section 2.2, holds uniformly. Taking p=q=2p=q=2 in [4], Lemma 2.2 gives

∥v∥L2(∂K)2≲hK−1∥v∥L2(K)2+∥∇v∥L2(K)∥v∥L2(K).\|v\|_{L^2(\partial K)}^2\lesssim h_K^{-1}\|v\|_{L^2(K)}^2+\|\nabla v\|_{L^2(K)}\|v\|_{L^2(K)}.

Young’s inequality with the scale hKh_K yields (32).

The polynomial inverse and approximation estimates on uniformly star-shaped polytopes are standard; see, for example, [8]; [5]; [6]. Combining (32) with (33) gives (34). Moreover,

∑F⊂∂K∥τh⋅nF∥L2(F)2≤∥τh∥L2(∂K)2,\sum_{F\subset\partial K}\|\tau_h\cdot n_F\|_{L^2(F)}^2\le\|\tau_h\|_{L^2(\partial K)}^2,

so (35) follows immediately.

For (36), choose a polynomial pK∈Pk(K)p_K\in\mathbb{P}_k(K) satisfying the standard H2H^2 approximation estimates. The L2L^2 best-approximation property of ΠK\Pi_K, followed by the polynomial inverse estimate applied to pK−ΠKvp_K-\Pi_Kv, gives the three volume estimates. Applying (32) first to v−ΠKvv-\Pi_Kv and then to each component of ∇(v−ΠKv)\nabla(v-\Pi_Kv) gives the two boundary estimates. Summing the local L2L^2 estimate and using (27) proves (37). □\square

For fixed k≥1k \ge1, define

Vh:={vh∈L2(Ω):vh∣K∈Pk(K) ∀K∈Th}.(38)V_h := \left\{v_h \in L^2(\Omega) : v_h|_K \in\mathbb{P}_k(K)\ \forall K \in\mathcal{T}_h\right\}. \tag*{(38)}

By (28), (27) and the fixed polynomial degree,

Nh:=dim⁡Vh≃h−d.(39)N_h := \dim V_h \simeq h^{-d}. \tag*{(39)}

SIP form and discrete operator

Introduce the broken vector-polynomial space

Σh:={τh∈L2(Ω;Rd):τh∣K∈[Pk(K)]d ∀K∈Th}.(40)\Sigma_h := \left\{\boldsymbol{\tau}_h \in L^2(\Omega;\mathbb{R}^d) : \boldsymbol{\tau}_h|_K \in[\mathbb{P}_k(K)]^d\ \forall K \in\mathcal{T}_h\right\}. \tag*{(40)}

The lifting Rh:L2(Fh)→Σh\mathcal{R}_h : L^2(\mathcal{F}_h) \to\Sigma_h is defined by

(Rh(φ),τh)L2(Ω)=−∑F∈Fh(φ,{ ⁣{τh⋅nF} ⁣})L2(F)∀τh∈Σh.(41)\left(\mathcal{R}_h(\varphi),\boldsymbol{\tau}_h\right)_{L^2(\Omega)} = -\sum_{F\in\mathcal{F}_h}\left(\varphi,\{\!\{\boldsymbol{\tau}_h\cdot\boldsymbol{n}_F\}\!\}\right)_{L^2(F)} \qquad\forall\boldsymbol{\tau}_h\in\Sigma_h. \tag*{(41)}

By (35),

∥Rh(φ)∥L2(Ω)≲(∑F∈FhhF−1∥φ∥L2(F)2)1/2.(42)\|\mathcal{R}_h(\varphi)\|_{L^2(\Omega)} \lesssim\left(\sum_{F\in\mathcal{F}_h}h_F^{-1}\|\varphi\|_{L^2(F)}^2\right)^{1/2}. \tag*{(42)}

For γ>0\gamma>0, set

σF:=γk2hF.(43)\sigma_F := \frac{\gamma k^2}{h_F}. \tag*{(43)}

The SIP form ah:Vh×Vh→Ra_h: V_h\times V_h\to\mathbb{R} is

ah(wh,vh):=(∇hwh,∇hvh)L2(Ω)+κ2(wh,vh)L2(Ω)−∑F∈Fh({ ⁣{∇hwh⋅nF} ⁣},⟦vh⟧)L2(F)−∑F∈Fh({ ⁣{∇hvh⋅nF} ⁣},⟦wh⟧)L2(F)+∑F∈FhσF(⟦wh⟧,⟦vh⟧)L2(F).(44)\begin{aligned} a_h(w_h,v_h) := {}&(\nabla_h w_h,\nabla_h v_h)_{L^2(\Omega)}+\kappa^2(w_h,v_h)_{L^2(\Omega)}\\ &-\sum_{F\in\mathcal{F}_h}(\{\!\{\nabla_h w_h\cdot\boldsymbol{n}_F\}\!\},\llbracket v_h\rrbracket)_{L^2(F)}\\ &-\sum_{F\in\mathcal{F}_h}(\{\!\{\nabla_h v_h\cdot\boldsymbol{n}_F\}\!\},\llbracket w_h\rrbracket)_{L^2(F)}\\ &+\sum_{F\in\mathcal{F}_h}\sigma_F(\llbracket w_h\rrbracket,\llbracket v_h\rrbracket)_{L^2(F)}. \tag*{(44)} \end{aligned}

Since ∇hVh⊂Σh\nabla_h V_h\subset\Sigma_h, the two consistency terms in (44) may equivalently be represented through the lifting. For example,

−∑F∈Fh({ ⁣{∇hwh⋅nF} ⁣},⟦vh⟧)L2(F)=(Rh(⟦vh⟧),∇hwh)L2(Ω).-\sum_{F\in\mathcal{F}_h}(\{\!\{\nabla_h w_h\cdot\boldsymbol{n}_F\}\!\},\llbracket v_h\rrbracket)_{L^2(F)} =(\mathcal{R}_h(\llbracket v_h\rrbracket),\nabla_h w_h)_{L^2(\Omega)}.

Thus the lifting estimate (42) controls the face terms without any dependence on the number or relative sizes of the faces.

For vh∈Vhv_h\in V_h, define

∥vh∥DG,h2:=∥∇hvh∥L2(Ω)2+κ2∥vh∥L2(Ω)2+∑F∈FhσF∥⟦vh⟧∥L2(F)2.(45)\|v_h\|_{\mathrm{DG},h}^2 := \|\nabla_hv_h\|_{L^2(\Omega)}^2+\kappa^2\|v_h\|_{L^2(\Omega)}^2+\sum_{F\in\mathcal{F}_h}\sigma_F\|\llbracket v_h\rrbracket\|_{L^2(F)}^2. \tag*{(45)}

3.5. Lemma (SIP stability). For γ\gamma sufficiently large,

ah(vh,vh)≳∥vh∥DG,h2,∣ah(wh,vh)∣≲∥wh∥DG,h∥vh∥DG,h(46)a_h(v_h,v_h)\gtrsim\|v_h\|_{\mathrm{DG},h}^2,\qquad|a_h(w_h,v_h)|\lesssim\|w_h\|_{\mathrm{DG},h}\|v_h\|_{\mathrm{DG},h} \tag*{(46)}

for all wh,vh∈Vhw_h,v_h\in V_h, with constants independent of hh and of the number and relative sizes of the element faces.

Proof. Represent the consistency terms through Rh\mathcal{R}_h and apply (42). Cauchy–Schwarz and Young’s inequality then give coercivity for γ\gamma sufficiently large and continuity in the DG norm. □

Define Lh:Vh→VhL_h:V_h\to V_h by

(Lhwh,vh)L2(Ω)=ah(wh,vh)∀wh,vh∈Vh.(47)(L_hw_h,v_h)_{L^2(\Omega)}=a_h(w_h,v_h)\qquad\forall w_h,v_h\in V_h. \tag*{(47)}

Lemma 3.5 shows that LhL_h is symmetric positive definite.

Deterministic resolvent approximation

For the consistency analysis under Assumption 2.1, set

Vh♯:=(H2(Ω)∩H01(Ω))+Vh.(48)V_h^\sharp:= \left(H^2(\Omega) \cap H_0^1(\Omega)\right) + V_h. \tag*{(48)}

For w,v∈Vh♯w,v \in V_h^\sharp, let ah♯(w,v)a_h^\sharp(w,v) be given by the right-hand side of (44), with (wh,vh)(w_h,v_h) replaced by (w,v)(w,v). The face terms are well-defined by the H2H^2 trace regularity, and the restriction of ah♯a_h^\sharp to Vh×VhV_h \times V_h is precisely aha_h.

Set

∥v∥h,∗2:=∥∇hv∥L2(Ω)2+κ2∥v∥L2(Ω)2+∑F∈FhσF∥⟦v⟧∥L2(F)2+∑F∈FhhF∥{ ⁣{∇hv⋅nF} ⁣}∥L2(F)2.(49)\lVert v\rVert_{h,*}^2 := \lVert\nabla_h v\rVert_{L^2(\Omega)}^2 + \kappa^2 \lVert v\rVert_{L^2(\Omega)}^2 + \sum_{F \in\mathcal{F}_h} \sigma_F \lVert\llbracket v\rrbracket\rVert_{L^2(F)}^2 + \sum_{F \in\mathcal{F}_h} h_F \lVert\{\!\{ \nabla_h v \cdot n_F \}\!\}\rVert_{L^2(F)}^2. \tag*{(49)}

Cauchy–Schwarz gives

∣ah♯(w,v)∣≲∥w∥h,∗∥v∥h,∗∀w,v∈Vh♯.(50)\left|a_h^\sharp(w,v)\right| \lesssim\lVert w\rVert_{h,*}\lVert v\rVert_{h,*} \qquad\forall w,v \in V_h^\sharp. \tag*{(50)}

Moreover, Lemma 3.3 implies

∥vh∥h,∗≲∥vh∥DG,h∀vh∈Vh,(51)\lVert v_h\rVert_{h,*} \lesssim\lVert v_h\rVert_{\mathrm{DG},h} \qquad\forall v_h \in V_h, \tag*{(51)}
∥v−Πhv∥h,∗≲h∥v∥H2(Ω)∀v∈H2(Ω)∩H01(Ω).(52)\lVert v-\Pi_hv\rVert_{h,*} \lesssim h\lVert v\rVert_{H^2(\Omega)} \qquad\forall v \in H^2(\Omega)\cap H_0^1(\Omega). \tag*{(52)}

Lemma 3.7 (Deterministic resolvent approximation). Suppose Assumptions 2.1 and 3.2 hold, and let γ\gamma be sufficiently large for Lemma 3.5. For f∈L2(Ω)f \in L^2(\Omega), let z=L−1fz=\mathcal{L}^{-1}f and let zh∈Vhz_h \in V_h satisfy

ah(zh,vh)=(f,vh)L2(Ω)∀vh∈Vh.(53)a_h(z_h,v_h)=(f,v_h)_{L^2(\Omega)} \qquad\forall v_h \in V_h. \tag*{(53)}

Then

∥z−zh∥L2(Ω)≲h2∥f∥L2(Ω).(54)\lVert z-z_h\rVert_{L^2(\Omega)} \lesssim h^2\lVert f\rVert_{L^2(\Omega)}. \tag*{(54)}

Equivalently,

∥L−1−Lh−1Πh∥L(L2(Ω),L2(Ω))≲h2.(55)\left\|\mathcal{L}^{-1}-\mathcal{L}_h^{-1}\Pi_h\right\|_{\mathcal{L}(L^2(\Omega),L^2(\Omega))} \lesssim h^2. \tag*{(55)}

Proof. Elliptic regularity gives

∥z∥H2(Ω)≲∥f∥L2(Ω).\lVert z\rVert_{H^2(\Omega)} \lesssim\lVert f\rVert_{L^2(\Omega)}.

Since z∈H2(Ω)∩H01(Ω)z \in H^2(\Omega)\cap H_0^1(\Omega), elementwise integration by parts yields

ah♯(z,vh)=(f,vh)L2(Ω)∀vh∈Vh.(56)a_h^\sharp(z,v_h)=(f,v_h)_{L^2(\Omega)} \qquad\forall v_h \in V_h. \tag*{(56)}

Hence

ah♯(z−zh,vh)=0∀vh∈Vh.a_h^\sharp(z-z_h,v_h)=0 \qquad\forall v_h \in V_h.

Let

η:=z−Πhz,ξh:=Πhz−zh.\eta:= z-\Pi_hz,\qquad\xi_h := \Pi_hz-z_h.

Galerkin orthogonality gives

ah(ξh,vh)=−ah♯(η,vh)∀vh∈Vh.a_h(\xi_h,v_h)=-a_h^\sharp(\eta,v_h) \qquad\forall v_h \in V_h.

Taking vh=ξhv_h=\xi_h and using Lemma 3.5, (50), (51) and (52), we obtain

∥ξh∥DG,h≲h∥z∥H2(Ω).\lVert\xi_h\rVert_{\mathrm{DG},h}\lesssim h\lVert z\rVert_{H^2(\Omega)}.

Therefore, with e:=z−zh=η+ξhe:=z-z_h=\eta+\xi_h,

∥e∥h,∗≲h∥z∥H2(Ω).(57)\lVert e\rVert_{h,*}\lesssim h\lVert z\rVert_{H^2(\Omega)}. \tag*{(57)}

For the L2L^2 estimate, let

ψ:=L−1e.\psi:=\mathcal{L}^{-1}e.

By Assumption 2.1,

∥ψ∥H2(Ω)≲∥e∥L2(Ω).\lVert\psi\rVert_{H^2(\Omega)}\lesssim\lVert e\rVert_{L^2(\Omega)}.

Symmetry, consistency and Galerkin orthogonality give

∥e∥L2(Ω)2=ah♯(e,ψ)=ah♯(e,ψ−Πhψ).\lVert e\rVert_{L^2(\Omega)}^2=a_h^\sharp(e,\psi)=a_h^\sharp(e,\psi-\Pi_h\psi).

Hence, by (50), (52) and (57),

∥e∥L2(Ω)2≲∥e∥h,∗∥ψ−Πhψ∥h,∗≲h2∥z∥H2(Ω)∥ψ∥H2(Ω)≲h2∥f∥L2(Ω)∥e∥L2(Ω).\lVert e\rVert_{L^2(\Omega)}^2 \lesssim\lVert e\rVert_{h,*}\lVert\psi-\Pi_h\psi\rVert_{h,*} \lesssim h^2\lVert z\rVert_{H^2(\Omega)}\lVert\psi\rVert_{H^2(\Omega)} \lesssim h^2\lVert f\rVert_{L^2(\Omega)}\lVert e\rVert_{L^2(\Omega)}.

This proves (54).

Finally,

(f,vh)L2(Ω)=(Πhf,vh)L2(Ω)∀vh∈Vh.(f,v_h)_{L^2(\Omega)}=(\Pi_hf,v_h)_{L^2(\Omega)}\qquad\forall v_h\in V_h.

so zh=Lh−1Πhfz_h=L_h^{-1}\Pi_hf, and (55) follows. □

Canonical discrete white noise

The restriction of W\mathcal{W} to VhV_h is represented by the unique Gaussian random variable Wh∈L2(Ξ;Vh)\mathcal{W}_h\in L^2(\Xi;V_h) satisfying

(Wh,vh)L2(Ω)=W(vh)∀vh∈Vh.(58)(\mathcal{W}_h,v_h)_{L^2(\Omega)}=\mathcal{W}(v_h)\qquad\forall v_h\in V_h. \tag*{(58)}

Its covariance is

E[(Wh,vh)L2(Ω)(Wh,wh)L2(Ω)]=(vh,wh)L2(Ω)∀vh,wh∈Vh.(59)\mathbb{E}\left[(\mathcal{W}_h,v_h)_{L^2(\Omega)}(\mathcal{W}_h,w_h)_{L^2(\Omega)}\right]=(v_h,w_h)_{L^2(\Omega)}\qquad\forall v_h,w_h\in V_h. \tag*{(59)}

Let {ϕK,i}i=1Nk\{\phi_{K,i}\}_{i=1}^{N_k} be a basis of Pk(K)\mathbb{P}_k(K) and let

(MK)ij:=∫KϕK,iϕK,j dx(M_K)_{ij}:=\int_K\phi_{K,i}\phi_{K,j}\,\mathrm{d}x

be the local mass matrix. The corresponding load vector

bK:=(W(ϕK,1),…,W(ϕK,Nk))Tb_K:=\left(\mathcal{W}(\phi_{K,1}),\ldots,\mathcal{W}(\phi_{K,N_k})\right)^{\mathsf{T}}

is Gaussian with covariance MKM_K. It can therefore be sampled as

bK=MK1/2ξK,ξK∼N(0,I).(60)b_K=M_K^{1/2}\xi_K,\qquad\xi_K\sim\mathcal{N}(0,I). \tag*{(60)}

If cKc_K denotes the coefficient vector of Wh∣K\mathcal{W}_h|_K, then MKcK=bKM_Kc_K=b_K. Basis functions on distinct elements have disjoint support, so the vectors bKb_K are mutually independent. Thus (60) gives an exact element-local sampler for the continuum white-noise load restricted to VhV_h.

If {ψi}i=1Nh\{\psi_i\}_{i=1}^{N_h} is an L2L^2-orthonormal basis of VhV_h, then

Wh=∑i=1Nhζiψi,ζi∼N(0,1),\mathcal{W}_h=\sum_{i=1}^{N_h}\zeta_i\psi_i,\qquad\zeta_i\sim\mathcal{N}(0,1),

with independent coefficients. Consequently,

E[∥Wh∥L2(Ω)2]=Nh≃h−d.(61)\mathbb{E}\left[\lVert\mathcal{W}_h\rVert_{L^2(\Omega)}^2\right]=N_h\simeq h^{-d}. \tag*{(61)}

The discrete SPDE is: find Yh∈VhY_h\in V_h such that

ah(Yh,vh)=W(vh)=(Wh,vh)L2(Ω)∀vh∈Vh,(62)a_h(Y_h,v_h)=\mathcal{W}(v_h)=(\mathcal{W}_h,v_h)_{L^2(\Omega)}\qquad\forall v_h\in V_h, \tag*{(62)}

or, equivalently,

LhYh=Wh.(63)L_hY_h=\mathcal{W}_h. \tag*{(63)}

Restricted noise and strong convergence

Let Y(h)Y^{(h)} be the continuum solution driven by Wh\mathcal{W}_h:

LY(h)=Wh.(64)LY^{(h)}=\mathcal{W}_h. \tag*{(64)}

Since Wh∈L2(Ξ;L2(Ω))\mathcal{W}_h\in L^2(\Xi;L^2(\Omega)), Assumption 2.1 gives

Y(h)∈L2(Ξ;H2(Ω)∩H01(Ω)).(65)Y^{(h)}\in L^2\left(\Xi;H^2(\Omega)\cap H_0^1(\Omega)\right). \tag*{(65)}

Lemma (White-noise restriction error). For d∈{2,3}d \in\{2,3\},

E[∥Y−Y(h)∥L2(Ω)2]=∑j=1∞λj−2∥(I−Πh)ϕj∥L2(Ω)2.(66)\mathbb{E}\left[\left\|Y-Y^{(h)}\right\|_{L^2(\Omega)}^2\right]=\sum_{j=1}^{\infty}\lambda_j^{-2}\left\|(I-\Pi_h)\phi_j\right\|_{L^2(\Omega)}^2. \tag*{(66)}

Consequently,

∥Y−Y(h)∥L2(Ξ;L2(Ω))≲h2−d/2.(67)\left\|Y-Y^{(h)}\right\|_{L^2(\Xi;L^2(\Omega))}\lesssim h^{2-d/2}. \tag*{(67)}

Proof. For every eigenfunction ϕj\phi_j,

(Wh,ϕj)L2(Ω)=(Wh,Πhϕj)L2(Ω)=W(Πhϕj).(\mathcal{W}_h,\phi_j)_{L^2(\Omega)}=(\mathcal{W}_h,\Pi_h\phi_j)_{L^2(\Omega)}=\mathcal{W}(\Pi_h\phi_j).

Therefore

(Y−Y(h),ϕj)L2(Ω)=λj−1W((I−Πh)ϕj).\left(Y-Y^{(h)},\phi_j\right)_{L^2(\Omega)}=\lambda_j^{-1}\mathcal{W}((I-\Pi_h)\phi_j).

Parseval’s identity and the covariance of W\mathcal{W} give (66). By (37), Assumption 2.1 and the L2L^2 stability of Πh\Pi_h,

∥(I−Πh)ϕj∥L2(Ω)≲min⁡{1,h2λj}.\left\|(I-\Pi_h)\phi_j\right\|_{L^2(\Omega)}\lesssim\min\left\{1,h^2\lambda_j\right\}.

Hence

E[∥Y−Y(h)∥L2(Ω)2]≲∑j=1∞min⁡{λj−2,h4}.(68)\mathbb{E}\left[\left\|Y-Y^{(h)}\right\|_{L^2(\Omega)}^2\right]\lesssim\sum_{j=1}^{\infty}\min\left\{\lambda_j^{-2},h^4\right\}. \tag*{(68)}

Using (9) and splitting the sum at j≃h−dj\simeq h^{-d} gives

∑j≲h−dh4+∑j≳h−dλj−2≲h4−d.(69)\sum_{j\lesssim h^{-d}}h^4+\sum_{j\gtrsim h^{-d}}\lambda_j^{-2}\lesssim h^{4-d}. \tag*{(69)}

Taking square roots proves (67).

Proposition (Optimality over finite-dimensional spaces). Let SN⊂L2(Ω)S_N\subset L^2(\Omega) be any deterministic subspace with dim⁡SN=N\dim S_N=N, and let ZN∈L2(Ξ;SN)Z_N\in L^2(\Xi;S_N) be arbitrary. Then

E∥Y−ZN∥L2(Ω)2≥∑j>Nλj−2.(70)\mathbb{E}\left\|Y-Z_N\right\|_{L^2(\Omega)}^2\geq\sum_{j>N}\lambda_j^{-2}. \tag*{(70)}

Equality is attained when

SN=span⁡{ϕ1,…,ϕN}S_N=\operatorname{span}\{\phi_1,\ldots,\phi_N\}

and ZNZ_N is the L2L^2-orthogonal projection of YY onto SNS_N.

In particular, taking SN=VhS_N=V_h and N=NhN=N_h, under Assumption 3.2,

∥Y−Zh∥L2(Ξ;L2(Ω))≳h2−d/2∀Zh∈L2(Ξ;Vh).(71)\left\|Y-Z_h\right\|_{L^2(\Xi;L^2(\Omega))}\gtrsim h^{2-d/2}\qquad\forall Z_h\in L^2(\Xi;V_h). \tag*{(71)}

Proof. Let PN:L2(Ω)→SNP_N:L^2(\Omega)\to S_N be the L2L^2-orthogonal projection. Since ZNZ_N takes values in SNS_N,

∥Y−ZN∥L2(Ω)2=∥Y−PNY∥L2(Ω)2+∥PNY−ZN∥L2(Ω)2\left\|Y-Z_N\right\|_{L^2(\Omega)}^2=\left\|Y-P_NY\right\|_{L^2(\Omega)}^2+\left\|P_NY-Z_N\right\|_{L^2(\Omega)}^2

almost surely. Hence

E∥Y−ZN∥L2(Ω)2≥E∥Y−PNY∥L2(Ω)2.\mathbb{E}\left\|Y-Z_N\right\|_{L^2(\Omega)}^2\geq\mathbb{E}\left\|Y-P_NY\right\|_{L^2(\Omega)}^2.

The covariance operator of YY is L−2\mathcal{L}^{-2}, which is trace class for d∈{2,3}d\in\{2,3\}. Therefore,

E∥Y−PNY∥L2(Ω)2=Tr⁡((I−PN)L−2(I−PN))=Tr⁡((I−PN)L−2).\begin{aligned} \mathbb{E}\left\|Y-P_NY\right\|_{L^2(\Omega)}^2=\operatorname{Tr}\left((I-P_N)\mathcal{L}^{-2}(I-P_N)\right) \\ &=\operatorname{Tr}\left((I-P_N)\mathcal{L}^{-2}\right). \end{aligned}

Since PNP_N has rank NN, the variational characterisation of the eigenvalues of L−2\mathcal{L}^{-2} gives

Tr⁡((I−PN)L−2)≥∑j>Nλj−2.\operatorname{Tr}\left((I-P_N)\mathcal{L}^{-2}\right)\geq\sum_{j>N}\lambda_j^{-2}.

This proves (70). Equality is attained by the spectral space span⁡{ϕ1,…,ϕN}\operatorname{span}\{\phi_1,\ldots,\phi_N\}.

Finally, take SN=VhS_N = V_h and N=NhN = N_h. By (9),

∑j>Nhλj−2≳Nh1−4/d.\sum_{j>N_h} \lambda_j^{-2} \gtrsim N_h^{1-4/d}.

Since Assumption 3.2 gives Nh≃h−dN_h \simeq h^{-d},

Nh1−4/d≃h4−d.N_h^{1-4/d} \simeq h^{4-d}.

Taking square roots proves (71). □

Theorem 3.12 (Sharp strong L2L^2 convergence). Let d∈{2,3}d \in\{2,3\} and let k≥1k \ge1 be fixed, let Assumptions 2.1 and 3.2 hold and let γ\gamma be sufficiently large as in Lemma 3.5. Then, there exist constants c,C>0c,C>0, independent of hh, such that the solution YhY_h of (62) satisfies

ch2−d/2≤∥Y−Yh∥L2(Ξ;L2(Ω))≤Ch2−d/2.(72)c h^{2-d/2} \le\|Y-Y_h\|_{L^2(\Xi;L^2(\Omega))} \le C h^{2-d/2}. \tag*{(72)}

Thus the sharp strong L2L^2 rate is hh in two dimensions and h1/2h^{1/2} in three dimensions.

Proof. The decomposition (20) and Lemma 3.10 give

∥Y−Yh∥L2(Ξ;L2(Ω))≲h2−d/2+∥Y(h)−Yh∥L2(Ξ;L2(Ω)).\|Y-Y_h\|_{L^2(\Xi;L^2(\Omega))} \lesssim h^{2-d/2}+\left\|Y^{(h)}-Y_h\right\|_{L^2(\Xi;L^2(\Omega))}.

Applying Lemma 3.7 samplewise with f=Whf=\mathcal{W}_h, followed by (61), gives

∥Y(h)−Yh∥L2(Ξ;L2(Ω))≲h2∥Wh∥L2(Ξ;L2(Ω))≲h2−d/2.\left\|Y^{(h)}-Y_h\right\|_{L^2(\Xi;L^2(\Omega))} \lesssim h^2\|\mathcal{W}_h\|_{L^2(\Xi;L^2(\Omega))} \lesssim h^{2-d/2}.

This proves the upper bound. The lower bound follows from Proposition 3.11 with Zh=YhZ_h=Y_h. □

Extension to a reentrant polygon

To examine the impact of geometry, we now replace Assumption 2.1 by the regularity structure associated with a single reentrant corner. The deterministic SIPDG analysis for elliptic problems on polygons with corner singularities is well established. In particular, the Dirichlet corner exponent, low-regularity continuity and consistency of the SIP form, Galerkin orthogonality, energy quasi-optimality and adjoint-consistent L2L^2 duality are developed in [26]. We make use of the associated weighted-space framework. The points requiring attention here are instead that the meshes are the star-shaped polytopes of Assumption 3.2, and that the stochastic analysis requires the endpoint behaviour of the leading corner singularity.

Throughout this section, the penalty parameter is chosen sufficiently large for Lemma 3.5 to hold. Let d=2d=2 and let Ω\Omega be a bounded polygon with one reentrant corner x⋆x_\star of interior angle

ω⋆∈(π,2π).(73)\omega_\star\in(\pi,2\pi). \tag*{(73)}

All remaining vertices on ∂Ω\partial\Omega are assumed to be convex. We assume that the meshes are boundary-conforming and that x⋆x_\star is a mesh vertex, see Figure 1. Since ω⋆∈(π,2π)\omega_\star\in(\pi,2\pi),

Figure 1. Local geometry at the reentrant corner x⋆x_\star, with interior angle ω⋆\omega_\star and polar coordinates (r⋆,θ⋆)(r_\star,\theta_\star).

πω⋆∈(12,1).(74)\frac{\pi}{\omega_\star} \in\left(\frac{1}{2},1\right). \tag*{(74)}

Let χ⋆\chi_\star be a smooth radial cut-off supported near x⋆x_\star and equal to one in a neighbourhood of x⋆x_\star, and set

s⋆:=χ⋆(r⋆)r⋆π/ω⋆sin⁡(πθ⋆ω⋆).(75)s_\star:= \chi_\star(r_\star) r_\star^{\pi/\omega_\star}\sin\left(\frac{\pi\theta_\star}{\omega_\star}\right). \tag*{(75)}

Then s⋆∈H01(Ω)s_\star\in H_0^1(\Omega), Ls⋆∈L2(Ω)\mathcal{L}s_\star\in L^2(\Omega), and, near x⋆x_\star,

∣∇ms⋆∣≲r⋆π/ω⋆−m,m∈{0,1,2}.(76)|\nabla^m s_\star| \lesssim r_\star^{\pi/\omega_\star-m}, \qquad m \in\{0,1,2\}. \tag*{(76)}

Corner structure and deterministic SIP estimates

Write

T:=L−1.T := \mathcal{L}^{-1}.

The classical Dirichlet corner decomposition [21, 16] gives bounded maps

Treg:L2(Ω)⟶H2(Ω)∩H01(Ω),ℓ⋆∈(L2(Ω))∗,T_{\mathrm{reg}} : L^2(\Omega) \longrightarrow H^2(\Omega) \cap H_0^1(\Omega), \qquad\ell_\star\in(L^2(\Omega))^*,

such that

Tf=Tregf+ℓ⋆(f)s⋆,∥Tregf∥H2(Ω)+∣ℓ⋆(f)∣≲∥f∥L2(Ω).(77)Tf = T_{\mathrm{reg}}f + \ell_\star(f)s_\star, \qquad\|T_{\mathrm{reg}}f\|_{H^2(\Omega)} + |\ell_\star(f)| \lesssim\|f\|_{L^2(\Omega)}. \tag*{(77)}

Let p⋆∈L2(Ω)p_\star\in L^2(\Omega) denote the Riesz representative of ℓ⋆\ell_\star.

We shall also use

σj(Treg)≲j−1,j≥1.(78)\sigma_j(T_{\mathrm{reg}}) \lesssim j^{-1}, \qquad j \ge1. \tag*{(78)}

Indeed, Weyl’s law gives σj(T)=λj−1≲j−1\sigma_j(T) = \lambda_j^{-1} \lesssim j^{-1} in two dimensions, while T−TregT-T_{\mathrm{reg}} has rank one, so (78) follows from the finite-rank perturbation inequality for singular values.

Since π/ω⋆>1/2\pi/\omega_\star> 1/2, the gradient of s⋆s_\star has an L2L^2 trace on every mesh face. Hence, for each fixed mesh, ah♯a_h^\sharp extends naturally to

Vh,⋆:=(H2(Ω)∩H01(Ω))+span⁡{s⋆}+Vh.(79)\mathcal{V}_{h,\star} := \left(H^2(\Omega) \cap H_0^1(\Omega)\right) + \operatorname{span}\{s_\star\} + V_h. \tag*{(79)}

The uniform trace estimate required for the approximation analysis is proved in Lemma 4.2.

We may therefore define the SIP Ritz map

Gh:(H2(Ω)∩H01(Ω))+span⁡{s⋆}⟶VhG_h : \left(H^2(\Omega) \cap H_0^1(\Omega)\right) + \operatorname{span}\{s_\star\} \longrightarrow V_h

by

ah(Ghv,vh)=ah♯(v,vh)∀vh∈Vh.(80)a_h(G_hv,v_h) = a_h^\sharp(v,v_h) \qquad\forall v_h \in V_h. \tag*{(80)}

The usual consistency argument remains valid. To see the only nonstandard point, perform elementwise integration by parts after removing a circle of radius rr about x⋆x_\star. On that circle,

s⋆=O(rπ/ω⋆),∂ns⋆=O(rπ/ω⋆−1),s_\star= O\left(r^{\pi/\omega_\star}\right), \qquad\partial_n s_\star= O\left(r^{\pi/\omega_\star-1}\right),

and the boundary length is O(r)O(r). The additional contribution therefore vanishes as r↓0r \downarrow0. Consequently,

ah♯(Tf,vh)=(f,vh)L2(Ω)∀vh∈Vh,(81)a_h^\sharp(Tf,v_h) = (f,v_h)_{L^2(\Omega)} \qquad\forall v_h \in V_h, \tag*{(81)}

and, with Th=Lh−1ΠhT_h = L_h^{-1}\Pi_h,

Th=GhT.(82)T_h = G_hT. \tag*{(82)}

The following estimates collect the deterministic information needed for the stochastic analysis.

Lemma (Corner SIP estimates). For fixed polynomial degree k≥1k \ge1,

∥u−Ghu∥h,∗≲h∥u∥H2(Ω),u∈H2(Ω)∩H01(Ω),(83)\lVert u-G_hu\rVert_{h,*} \lesssim h\lVert u\rVert_{H^2(\Omega)}, \qquad u\in H^2(\Omega)\cap H_0^1(\Omega), \tag*{(83)}
∥s⋆−Πhs⋆∥h,∗+∥s⋆−Ghs⋆∥h,∗≲hπ/ω⋆,(84)\lVert s_\star-\Pi_hs_\star\rVert_{h,*}+\lVert s_\star-G_hs_\star\rVert_{h,*}\lesssim h^{\pi/\omega_\star}, \tag*{(84)}
∥s⋆−Ghs⋆∥L2(Ω)≲h2π/ω⋆.(85)\lVert s_\star-G_hs_\star\rVert_{L^2(\Omega)}\lesssim h^{2\pi/\omega_\star}. \tag*{(85)}

Proof. The estimate (83) follows from (52), coercivity, continuity and Galerkin orthogonality. We therefore concentrate on the singular function s⋆s_\star.

Introduce the metric corner patch

Nh⋆:={K∈Th:K∩B(x⋆,h)≠∅}.(86)\mathcal{N}_h^\star:=\{K\in\mathcal{T}_h:K\cap B(x_\star,h)\ne\varnothing\}. \tag*{(86)}

If K∈Nh⋆K\in\mathcal{N}_h^\star, then K⊂B(x⋆,2h)K\subset B(x_\star,2h). Moreover, Assumption 3.2, (28) and (27) give ∣K∣≳h2|K|\gtrsim h^2. Since the element interiors are disjoint,

#Nh⋆≲1.(87)\#\mathcal{N}_h^\star\lesssim1. \tag*{(87)}

On the other hand,

K∉Nh⋆⟹dist⁡(K,x⋆)≥h.(88)K\notin\mathcal{N}_h^\star\quad\Longrightarrow\quad\operatorname{dist}(K,x_\star)\ge h. \tag*{(88)}

These two properties separate the corner contribution from the regular part of the mesh.

We first estimate the contribution from Nh⋆\mathcal{N}_h^\star. Quasi-uniformity implies that, for K∈Nh⋆K\in\mathcal{N}_h^\star,

K⊂B(x⋆,ChK)K\subset B(x_\star,Ch_K)

with CC independent of hh. Hence (76) gives

∥s⋆∥L2(K)≲hK1+π/ω⋆,∥∇s⋆∥L2(K)≲hKπ/ω⋆.(89)\lVert s_\star\rVert_{L^2(K)}\lesssim h_K^{1+\pi/\omega_\star},\qquad\lVert\nabla s_\star\rVert_{L^2(K)}\lesssim h_K^{\pi/\omega_\star}. \tag*{(89)}

The only nonstandard estimate is the whole-boundary trace of ∇s⋆\nabla s_\star. Since π/ω⋆>1/2\pi/\omega_\star>1/2, choose an exponent μ\mu such that

43<μ<22−π/ω⋆,μ′:=μμ−1.(90)\frac{4}{3}<\mu<\frac{2}{2-\pi/\omega_\star},\qquad\mu':=\frac{\mu}{\mu-1}. \tag*{(90)}

Such a choice is possible precisely because π/ω⋆>1/2\pi/\omega_\star>1/2. The pointwise behaviour (76) then gives

∥∇s⋆∥Lμ′(K)≲hKπ/ω⋆−1+2/μ′,(91)\lVert\nabla s_\star\rVert_{L^{\mu'}(K)}\lesssim h_K^{\pi/\omega_\star-1+2/\mu'}, \tag*{(91)}
∥D2s⋆∥Lμ(K)≲hKπ/ω⋆−2+2/μ.(92)\lVert D^2s_\star\rVert_{L^\mu(K)}\lesssim h_K^{\pi/\omega_\star-2+2/\mu}. \tag*{(92)}

In particular, ∇s⋆∈[W1,μ(K)]2\nabla s_\star\in[W^{1,\mu}(K)]^2.

Apply [4], Lemma 2.2 with p=μp=\mu and q=2q=2 to each component of ∇s⋆\nabla s_\star. The mesh condition required there holds uniformly by (29). Since ∣K∣≃hK2|K|\simeq h_K^2, we obtain

∥∇s⋆∥L2(∂K)2≲hK−1∣K∣1−2/μ′∥∇s⋆∥Lμ′(K)(∥∇s⋆∥Lμ′(K)+∥D2s⋆∥Lμ(K)∥∇s⋆∥Lμ′(K))≲hK2π/ω⋆−1.\begin{aligned} \lVert\nabla s_\star\rVert_{L^2(\partial K)}^2 &\lesssim h_K^{-1}|K|^{1-2/\mu'}\lVert\nabla s_\star\rVert_{L^{\mu'}(K)} \left(\lVert\nabla s_\star\rVert_{L^{\mu'}(K)} +\lVert D^2s_\star\rVert_{L^\mu(K)}\lVert\nabla s_\star\rVert_{L^{\mu'}(K)}\right)\\ &\lesssim h_K^{2\pi/\omega_\star-1}. \end{aligned}

Thus

∥∇s⋆∥L2(∂K)≲hKπ/ω⋆−1/2.(93)\lVert\nabla s_\star\rVert_{L^2(\partial K)}\lesssim h_K^{\pi/\omega_\star-1/2}. \tag*{(93)}

Notice that this is a whole-boundary estimate; no lower bound on the size of an individual face is used.

Applying (32) directly to s⋆s_\star and using (89) similarly gives

∥s⋆∥L2(∂K)≲hKπ/ω⋆+1/2.(94)\lVert s_\star\rVert_{L^2(\partial K)}\lesssim h_K^{\pi/\omega_\star+1/2}. \tag*{(94)}

Let ΠK\Pi_K denote the local L2L^2 projection. Its L2L^2 stability, together with the polynomial inverse and trace estimates of Lemma 3.3, now yields

∥s⋆−ΠKs⋆∥L2(K)+hK∥∇(s⋆−ΠKs⋆)∥L2(K)+hK1/2∥s⋆−ΠKs⋆∥L2(∂K)+hK3/2∥∇(s⋆−ΠKs⋆)∥L2(∂K)≲hK1+π/ω⋆.(95)\begin{aligned} &\lVert s_\star-\Pi_Ks_\star\rVert_{L^2(K)} +h_K\lVert\nabla(s_\star-\Pi_Ks_\star)\rVert_{L^2(K)}\\ &\qquad+h_K^{1/2}\lVert s_\star-\Pi_Ks_\star\rVert_{L^2(\partial K)} +h_K^{3/2}\lVert\nabla(s_\star-\Pi_Ks_\star)\rVert_{L^2(\partial K)} \lesssim h_K^{1+\pi/\omega_\star}. \tag*{(95)} \end{aligned}

for every K∈Nh⋆K \in\mathcal{N}_h^\star.

We next consider the elements outside the metric corner patch. By (88), the singularity is separated from each such element by at least hh. Hence (36) and (76) give, for a fixed radius R>0R > 0 containing the support of χ⋆\chi_\star,

∑K∉Nh⋆hK2∥s⋆∥H2(K)2≲h2(1+∫hRr2π/ω⋆−3 dr)≲h2π/ω⋆.\sum_{K \notin\mathcal{N}_h^\star} h_K^2 \lVert s_\star\rVert_{H^2(K)}^2 \lesssim h^2 \left(1+\int_h^R r^{2\pi/\omega_\star-3}\,\mathrm{d}r\right) \lesssim h^{2\pi/\omega_\star}.

Here the fixed term accounts for the smooth cut-off region. Combining this estimate with (95), (87) and the definition of ∥⋅∥h,⋆\lVert\cdot\rVert_{h,\star} gives

∥s⋆−Πhs⋆∥h,⋆≲hπ/ω⋆.(96)\lVert s_\star-\Pi_hs_\star\rVert_{h,\star}\lesssim h^{\pi/\omega_\star}. \tag*{(96)}

To obtain the corresponding Ritz estimate, set

η:=s⋆−Πhs⋆,ξh:=Πhs⋆−Ghs⋆.\eta:= s_\star-\Pi_hs_\star,\qquad\xi_h := \Pi_hs_\star-G_hs_\star.

Consistency and Galerkin orthogonality imply

ah(ξh,vh)=−ah♯(η,vh)∀vh∈Vh.a_h(\xi_h,v_h)=-a_h^\sharp(\eta,v_h)\qquad\forall v_h\in V_h.

Taking vh=ξhv_h=\xi_h and using coercivity, (50) and (51) gives

∥ξh∥h,⋆≲∥η∥h,⋆.\lVert\xi_h\rVert_{h,\star}\lesssim\lVert\eta\rVert_{h,\star}.

Together with (96), this proves (84).

It remains to prove the L2L^2 estimate. Set

e⋆:=s⋆−Ghs⋆,ψ:=Te⋆=Trege⋆+ℓ⋆(e⋆)s⋆,e_\star:= s_\star-G_hs_\star,\qquad\psi:= Te_\star=T_{\mathrm{reg}}e_\star+\ell_\star(e_\star)s_\star,

and choose

ψh:=ΠhTrege⋆+ℓ⋆(e⋆)Πhs⋆.\psi_h := \Pi_hT_{\mathrm{reg}}e_\star+\ell_\star(e_\star)\Pi_hs_\star.

The regular approximation estimate, (96) and (77) imply

∥ψ−ψh∥h,⋆≲h∥Trege⋆∥H2(Ω)+hπ/ω⋆∣ℓ⋆(e⋆)∣≲hπ/ω⋆∥e⋆∥L2(Ω),\lVert\psi-\psi_h\rVert_{h,\star}\lesssim h\lVert T_{\mathrm{reg}}e_\star\rVert_{H^2(\Omega)}+h^{\pi/\omega_\star}\lvert\ell_\star(e_\star)\rvert \lesssim h^{\pi/\omega_\star}\lVert e_\star\rVert_{L^2(\Omega)},

where we used h≲hπ/ω⋆h\lesssim h^{\pi/\omega_\star} for 0<h≤10<h\leq1.

Finally, symmetry, consistency and Galerkin orthogonality give

∥e⋆∥L2(Ω)2=ah♯(e⋆,ψ−ψh)≲∥e⋆∥h,⋆∥ψ−ψh∥h,⋆≲h2π/ω⋆∥e⋆∥L2(Ω).\lVert e_\star\rVert_{L^2(\Omega)}^2=a_h^\sharp(e_\star,\psi-\psi_h) \lesssim\lVert e_\star\rVert_{h,\star}\lVert\psi-\psi_h\rVert_{h,\star} \lesssim h^{2\pi/\omega_\star}\lVert e_\star\rVert_{L^2(\Omega)}.

This proves (85). □

Hilbert–Schmidt resolvent error

The preceding estimates still reflect the deterministic corner loss. For the SPDE, however, the singular contribution in (77) is only rank one. This allows the regular infinite-dimensional part and the corner correction to be estimated separately.

Lemma (Reentrant-corner resolvent estimate). The continuous and discrete resolvents satisfy

∥T−Th∥L2(L2(Ω))≲h.(97)\lVert T-T_h\rVert_{L^2(L^2(\Omega))}\lesssim h. \tag*{(97)}

Proof. Let u∈H2(Ω)∩H01(Ω)u\in H^2(\Omega)\cap H^1_0(\Omega) and set

eu:=u−Ghu.e_u:=u-G_hu.

For g∈L2(Ω)g\in L^2(\Omega), adjoint consistency gives

(eu,g)L2(Ω)=ah♯(eu,Tg).(98)(e_u,g)_{L^2(\Omega)}=a_h^\sharp(e_u,Tg). \tag*{(98)}

Using the corner decomposition (77) and the Galerkin orthogonality

ah†(eu,vh)=0∀vh∈Vh,a_h^{\dagger}(e_u,v_h)=0 \qquad\forall v_h\in V_h,

we obtain

(eu,g)L2(Ω)=ah†(eu,Tregg−ΠhTregg)+ℓ⋆(g)ah†(eu,s⋆−Πhs⋆).(e_u,g)_{L^2(\Omega)}=a_h^{\dagger}(e_u,T_{\mathrm{reg}}g-\Pi_hT_{\mathrm{reg}}g)+\ell_\star(g)a_h^{\dagger}(e_u,s_\star-\Pi_hs_\star).

The first term on the right is a bounded linear functional of gg. Hence, by the Riesz representation theorem, there exists a unique Bhu∈L2(Ω)B_hu\in L^2(\Omega) such that

(Bhu,g)L2(Ω):=ah†(eu,Tregg−ΠhTregg).(B_hu,g)_{L^2(\Omega)}:=a_h^{\dagger}(e_u,T_{\mathrm{reg}}g-\Pi_hT_{\mathrm{reg}}g).

Set also

qh(u):=ah†(eu,s⋆−Πhs⋆).q_h(u):=a_h^{\dagger}(e_u,s_\star-\Pi_hs_\star).

Lemma 4.2 and (77) give

∥Bhu∥L2(Ω)≲h2∥u∥H2(Ω),∣qh(u)∣≲h1+π/ω⋆∥u∥H2(Ω).(99)\lVert B_hu\rVert_{L^2(\Omega)}\lesssim h^2\lVert u\rVert_{H^2(\Omega)},\qquad|q_h(u)|\lesssim h^{1+\pi/\omega_\star}\lVert u\rVert_{H^2(\Omega)}. \tag*{(99)}

Let p⋆∈L2(Ω)p_\star\in L^2(\Omega) be the Riesz representative of ℓ⋆\ell_\star, so that

ℓ⋆(g)=(g,p⋆)L2(Ω).\ell_\star(g)=(g,p_\star)_{L^2(\Omega)}.

Since the preceding identity holds for every g∈L2(Ω)g\in L^2(\Omega),

u−Ghu=Bhu+qh(u)p⋆.u-G_hu=B_hu+q_h(u)p_\star.

Applying this with u=Tregfu=T_{\mathrm{reg}}f and using (77) together with (82), we obtain

(T−Th)f=BhTregf+qh(Tregf)p⋆+ℓ⋆(f)(s⋆−Ghs⋆).(100)(T-T_h)f=B_hT_{\mathrm{reg}}f+q_h(T_{\mathrm{reg}}f)p_\star+\ell_\star(f)(s_\star-G_hs_\star). \tag*{(100)}

We first consider the infinite-dimensional term BhTregB_hT_{\mathrm{reg}}. By (99),

∥BhTreg∥L(L2(Ω))≲h2.\lVert B_hT_{\mathrm{reg}}\rVert_{\mathcal{L}(L^2(\Omega))}\lesssim h^2.

From

u−Ghu=Bhu+qh(u)p⋆u-G_hu=B_hu+q_h(u)p_\star

we have, for every f∈L2(Ω)f\in L^2(\Omega),

(BhTreg−Treg)f=−GhTregf−qh(Tregf)p⋆.(B_hT_{\mathrm{reg}}-T_{\mathrm{reg}})f=-G_hT_{\mathrm{reg}}f-q_h(T_{\mathrm{reg}}f)p_\star.

The first term takes values in VhV_h, while the second takes values in span⁡{p⋆}\operatorname{span}\{p_\star\}. Consequently,

rank⁡(BhTreg−Treg)≤Nh+1.\operatorname{rank}(B_hT_{\mathrm{reg}}-T_{\mathrm{reg}})\leq N_h+1.

We use the standard finite-rank perturbation inequality for singular values: if AA and BB are compact operators and

rank⁡(A−B)≤r,\operatorname{rank}(A-B)\leq r,

then

σr+j(A)≤σj(B),j≥1.\sigma_{r+j}(A)\leq\sigma_j(B),\qquad j\geq1.

Applying this with

A=BhTreg,B=Treg,r=Nh+1,A=B_hT_{\mathrm{reg}},\qquad B=T_{\mathrm{reg}},\qquad r=N_h+1,

and using (78), we obtain

σNh+1+j(BhTreg)≤σj(Treg)≲j−1,j≥1.\sigma_{N_h+1+j}(B_hT_{\mathrm{reg}})\leq\sigma_j(T_{\mathrm{reg}})\lesssim j^{-1},\qquad j\geq1.

Combining this with the operator-norm bound above and Nh≃h−2N_h\simeq h^{-2} yields

∥BhTreg∥L2(L2(Ω))2≲Nhh4+∑j=1∞min⁡{h4,j−2}≲h2.(101)\lVert B_hT_{\mathrm{reg}}\rVert_{\mathcal{L}_2(L^2(\Omega))}^2\lesssim N_hh^4+\sum_{j=1}^{\infty}\min\{h^4,j^{-2}\}\lesssim h^2. \tag*{(101)}

Hence

∥BhTreg∥L2(L2(Ω))≲h.\lVert B_hT_{\mathrm{reg}}\rVert_{\mathcal{L}_2(L^2(\Omega))}\lesssim h.

The remaining two terms in (100) have rank one. Using (99), (77) and (85), we obtain

∥f↦qh(Tregf)p⋆∥L2(L2(Ω))≲h1+π/ω⋆,\lVert f\mapsto q_h(T_{\mathrm{reg}}f)p_\star\rVert_{\mathcal{L}_2(L^2(\Omega))}\lesssim h^{1+\pi/\omega_\star},

and

∥f↦ℓ⋆(f)(s⋆−Ghs⋆)∥L2(L2(Ω))≲h2π/ω⋆.\left\|f \mapsto\ell_\star(f)(s_\star-G_hs_\star)\right\|_{\mathcal{L}_2(L^2(\Omega))} \lesssim h^{2\pi/\omega_\star}.

Since

πω⋆>12,\frac{\pi}{\omega_\star}>\frac{1}{2},

both exponents are strictly larger than one. Thus the two rank-one corner terms are of higher order than the O(h)O(h) Hilbert–Schmidt contribution from BhTregB_hT_{\mathrm{reg}}, and (97) follows. □

Theorem 4.5 (Sharp convergence on a reentrant polygon). Let d=2d=2 and let Ω\Omega have the corner geometry described above. Under Assumption 3.2, there exist constants c,C>0c,C>0, independent of hh, such that

ch≤∥Y−Yh∥L2(Ξ;L2(Ω))≤Ch.(102)ch \le\|Y-Y_h\|_{L^2(\Xi;L^2(\Omega))} \le Ch. \tag*{(102)}

Proof. Let {ej}j≥1\{e_j\}_{j\ge1} be any L2(Ω)L^2(\Omega)-orthonormal basis and set ζj:=W(ej)\zeta_j:=\mathcal{W}(e_j). Since the continuous and discrete problems are driven by the same isonormal process,

Y=∑j=1∞ζjTej,Yh=∑j=1∞ζjThej,Y=\sum_{j=1}^{\infty}\zeta_jTe_j,\qquad Y_h=\sum_{j=1}^{\infty}\zeta_jT_he_j,

and therefore

∥Y−Yh∥L2(Ξ;L2(Ω))=∥T−Th∥L2(L2(Ω)).(103)\|Y-Y_h\|_{L^2(\Xi;L^2(\Omega))}=\|T-T_h\|_{\mathcal{L}_2(L^2(\Omega))}. \tag*{(103)}

Lemma 4.4 gives the upper bound. The lower bound follows from Proposition 3.11, since YhY_h takes values in VhV_h and Nh≃h−2N_h\simeq h^{-2}. □

Remark 4.6 (Why the corner does not change the stochastic rate). A global deterministic regularity argument only sees H1+π/ω⋆−ε(Ω)H^{1+\pi/\omega_\star-\varepsilon}(\Omega) regularity and hence reflects the loss caused by the reentrant corner. The operator decomposition (99) is sharper. The infinite-dimensional regular contribution has Hilbert–Schmidt error O(h)O(h), whereas the two corner corrections are rank one and have orders

O(h1+π/ω⋆)andO(h2π/ω⋆).O\left(h^{1+\pi/\omega_\star}\right) \qquad\text{and}\qquad O\left(h^{2\pi/\omega_\star}\right).

Both are o(h)o(h) because ω⋆<2π\omega_\star<2\pi. Thus the reentrant corner degrades deterministic approximation rates without changing the sharp first-order strong rate of the white-noise problem.

Numerical experiments

The numerical experiments examine the two main convergence results. On the unit square and cube, the explicit Dirichlet eigensystem is used to estimate the continuum strong error and test the dimension-dependent rate in Theorem 3.12. We then consider a family of reentrant polygonal domains with varying corner angle and test the first-order stochastic rate of Theorem 4.5 using a coupled fine-level reference.

The numerical implementation is based on the polytopic discontinuous Galerkin finite element library [18]. Throughout, κ=4\kappa=4 and VhV_h consists of discontinuous piecewise-linear polynomials. We take γ=14\gamma=14 in two dimensions and γ=24\gamma=24 in three dimensions. Linear systems are solved by ILU-preconditioned BiCGSTAB with relative tolerance 2×10−92\times10^{-9}. For each mesh, the SIP matrix and preconditioner are assembled once and reused for all modal right-hand sides or stochastic realisations.

Continuum strong-error estimator on (0,1)d(0,1)^d

For Ω=(0,1)d\Omega=(0,1)^d, the eigensystem of L\mathcal{L} is

ϕn(x)=2d/2∏i=1dsin⁡(πnixi),λn=π2∥n∥22+κ2,n∈Nd.\phi_n(x)=2^{d/2}\prod_{i=1}^{d}\sin(\pi n_i x_i),\qquad \lambda_n=\pi^2\lVert n\rVert_2^2+\kappa^2,\qquad n\in\mathbb{N}^d.

For each mode, let uh,n∈Vhu_{h,n}\in V_h satisfy

ah(uh,n,vh)=(ϕn,vh)L2(Ω)∀vh∈Vh.a_h(u_{h,n},v_h)=(\phi_n,v_h)_{L^2(\Omega)} \qquad\forall v_h\in V_h.

Expanding the continuous and discrete random fields with respect to the same isonormal process gives

E[∥Y−Yh∥L2(Ω)2]=∑n∈Ndeh,n2,eh,n2:=∥λn−1ϕn−uh,n∥L2(Ω)2.(104)\mathbb{E}\left[\|Y-Y_h\|_{L^2(\Omega)}^2\right] =\sum_{n\in\mathbb{N}^d}e_{h,n}^2,\qquad e_{h,n}^2:=\left\|\lambda_n^{-1}\phi_n-u_{h,n}\right\|_{L^2(\Omega)}^2. \tag*{(104)}

We evaluate the low-frequency part of (104) explicitly and estimate the remaining infinite sum by importance sampling. Write

m(n):=∣n∣∞.m(\mathbf{n}) := \lvert\mathbf{n} \rvert_\infty.

For a hierarchy with mesh scales h1,…,hLh_1,\ldots,h_L, the tail proposal is the equal mixture

q(n):=1L∑ℓ=1Lqℓ(n),qℓ(n)∝min⁡{hℓ4,m(n)−4},m(n)>m0.q(\mathbf{n}) := \frac{1}{L}\sum_{\ell=1}^{L}q_\ell(\mathbf{n}), \qquad q_\ell(\mathbf{n}) \propto\min\left\{h_\ell^4,m(\mathbf{n})^{-4}\right\}, \qquad m(\mathbf{n})>m_0.

Each component resolves the transition m(n)≃hℓ−1m(\mathbf{n}) \simeq h_\ell^{-1} and retains an infinite m−4m^{-4} tail. For independent samples Nq∼q\mathbf{N}_q \sim q, we use

E^h2:=∑m(n)≤m0eh,n2+1Ntail∑q=1Ntaileh,Nq2q(Nq)(105)\widehat{\mathcal{E}}_h^2 := \sum_{m(\mathbf{n})\le m_0} e_{h,\mathbf{n}}^2 + \frac{1}{N_{\mathrm{tail}}}\sum_{q=1}^{N_{\mathrm{tail}}}\frac{e_{h,\mathbf{N}_q}^2}{q(\mathbf{N}_q)} \tag*{(105)}

The same sampled modes are used across every mesh level and mesh family in a fixed dimension. In two dimensions we take m0=3m_0=3 and Ntail=32N_{\mathrm{tail}}=32, while in three dimensions we take m0=2m_0=2 and Ntail=12N_{\mathrm{tail}}=12. Error bars show one bootstrap standard error of the RMS estimator, obtained from 2000 resamples of the importance-sampled tail contributions with the explicitly summed low-frequency contribution held fixed.

For the convergence plots we use NK−1/dN_K^{-1/d} as the mesh scale, where NKN_K is the number of computational elements. Under Assumption 3.2, NK−1/d≃hN_K^{-1/d}\simeq h.

Experiment 1: regular polytopic meshes

Two dimensions. For d=2d=2, Theorem 3.12 gives the sharp rate O(h)O(h). We consider four quasi-uniform mesh families: Voronoi, wave-distorted Voronoi, checkerboard-distorted Voronoi and jittered quadrilateral meshes. Each hierarchy contains six refinement levels, with NK=1282N_K=128^2 elements on the finest level. Figure 2 shows representative stochastic fields on the four mesh families.

Representative two-dimensional stochastic fields on the four quasi-uniform polytopic mesh families

Figure 2. Representative two-dimensional stochastic fields on the four quasi-uniform polytopic mesh families.

Figure 3 shows the continuum RMS error computed from (105), plotted against NK−1/2N_K^{-1/2}. The dotted line shows the first-order reference slope.

Figure 3. Continuum RMS L2L^2 error on the four two-dimensional polytopic mesh families. The horizontal axis is NK−1/2N_K^{-1/2} and the dotted line shows the sharp rate O(h)O(h) from Theorem 3.12. Error bars show one bootstrap standard error of the modal estimator.

Three dimensions. For d=3d = 3, Theorem 3.12 gives the sharp rate O(h1/2)O(h^{1/2}). We consider regular hexahedral, jittered hexahedral and non-convex polyhedral mesh families. Each hierarchy contains five refinement levels. The regular and jittered families have NK=323N_K = 32^3 elements on the finest level, while the non-convex family is generated from a 48348^3 background partition.

The non-convex computational elements are connected agglomerates of tetrahedra, with a single affine polynomial defined on each resulting polyhedron. Figure 4 illustrates the computational geometry.

(a) Cut-away polyhedral mesh

(b) Representative non-convex cell

Figure 4. Non-convex polyhedral geometry used in the three-dimensional experiment. The displayed boundaries are those of the computational polyhedra.

Figure 5 shows the continuum RMS error computed from (104), plotted against NK−1/3N_K^{-1/3}. The dotted line shows the O(h1/2)O(h^{1/2}) reference slope.

Figure 5. Continuum RMS L2L^2 error on the regular hexahedral, jittered hexahedral and non-convex polyhedral mesh families. The horizontal axis is NK−1/3N_K^{-1/3} and the dotted line shows the sharp rate O(h1/2)O(h^{1/2}) from Theorem 3.12. Error bars show one bootstrap standard error of the modal estimator.

Experiment 2: stochastic reentrant problem

We consider a family of polygonal domains with

α∈{45∘,60∘,75∘,90∘,105∘,120∘,135∘}.(106)\alpha\in\{45^\circ, 60^\circ, 75^\circ, 90^\circ, 105^\circ, 120^\circ, 135^\circ\}. \tag*{(106)}

The corresponding reentrant angle is

ω⋆=2π−α,(107)\omega_\star= 2\pi- \alpha, \tag*{(107)}

so π/ω⋆\pi/\omega_\star varies from 4/74/7 to 4/54/5 across the angle sweep.

The reentrant vertex is preserved exactly at every refinement level. The meshes are constructed from a perturbed Cartesian background, cut to the polygonal domain and agglomerated into general polygons. The resulting hierarchies contain many-sided and non-convex elements. Figure 6 shows representative stochastic fields for four openings.

Representative stochastic field for $\alpha = 45^\circ$

(a) α=45∘\alpha= 45^\circ

Representative stochastic field for $\alpha = 75^\circ$

(b) α=75∘\alpha= 75^\circ

Representative stochastic field for $\alpha = 105^\circ$

(c) α=105∘\alpha= 105^\circ

Representative stochastic field for $\alpha = 135^\circ$

(d) α=135∘\alpha= 135^\circ

Figure 6. Representative stochastic fields on the reentrant polygonal domains for four mouth openings spanning the angle sweep.

For every mouth opening, the stochastic hierarchy contains five reported levels followed by two additional refinement levels. The final level is generated from a 256×256256 \times256 background partition and is used as the coupled reference. For each angle we use M=24M = 24 independent white-noise realisations.

For each realisation, the same continuum white-noise functional is restricted to every discrete space in the hierarchy. Thus the solutions Yh(q)Y_h^{(q)} and Yhref(q)Y_{h_{\mathrm{ref}}}^{(q)} are driven by the same Gaussian realisation. On each reported level we compute the coupled fine-reference RMS difference

E^h,ref:=(1M∑q=1M∥Yh(q)−Yhref(q)∥L2(Ω)2)1/2.(108)\widehat{E}_{h,\mathrm{ref}} := \left(\frac{1}{M}\sum_{q=1}^{M}\left\|Y_h^{(q)}-Y_{h_{\mathrm{ref}}}^{(q)}\right\|_{L^2(\Omega)}^2\right)^{1/2}. \tag*{(108)}

For each angle, the stochastic convergence rate is estimated from a least-squares fit of log⁡E^h,ref\log\widehat{E}_{h,\mathrm{ref}} against log⁡NK−1/2\log N_K^{-1/2}, omitting the coarsest reported level. Its sampling uncertainty is estimated by resampling the MM coupled realisations simultaneously across all mesh levels and repeating the slope fit. We use 2000 paired bootstrap resamples.

Figure 7 reports the resulting stochastic convergence. Panel (A) shows the coupled fine-reference RMS differences for representative α\alpha, plotted against NK−1/2N_K^{-1/2} together with a first-order reference slope. Panel (B) shows the estimated exponent across the complete angle sweep.

Figure 7. Strong stochastic convergence on the reentrant polygonal domains. Panel (A) shows the coupled fine-reference RMS L2L^2 difference against NK−1/2N_K^{-1/2} for representative mouth openings; the dotted line shows the first-order rate. Panel (B) shows the fitted stochastic exponent for all seven mouth openings. Error bars show bootstrap standard errors of the RMS differences in panel (A) and of the fitted exponent in panel (B).

The experiments recover the two convergence behaviours established in the analysis. On regular polygonal and polyhedral domains, the continuum estimator exhibits the dimension-dependent rate h2−d/2h^{2-d/2} across several polytopic mesh families. Across the reentrant angle sweep, the coupled stochastic approximation exhibits the first-order rate of eq:4.5.

Acknowledgments

TP is supported by the EPSRC programme grant Mathematics of Radiation Transport (MaThRad) EP/W026899/2 which is gratefully acknowledged.

References

¹ Institute for Mathematical Innovation, University of Bath, Bath, UK. ² Department of Mathematical Sciences, University of Bath, Bath, UK.

References

  1. [1]E. J. Allen, S. J. Novosel, and Z. Zhang. “Finite Element and Difference Approximation of Some Linear Stochastic Partial Differential Equations”. In: Stochastics and Stochastics Reports 64.1–2 (1998), pp. 117–142. DOI: 10.1080/17442509808834159.DOI
  2. [2]D. Bolin, K. Kirchner, and M. Kovács. “Weak convergence of Galerkin approximations for fractional elliptic stochastic PDEs with spatial white noise”. In: BIT Numerical Mathematics 58.4 (2018), pp. 881–906. DOI: 10.1007/s10543-018-0719-8.DOI
  3. [3]D. Bolin, K. Kirchner, and M. Kovács. “Numerical Solution of Fractional Elliptic Stochastic PDEs with Spatial White Noise”. In: IMA Journal of Numerical Analysis 40.2 (2020), pp. 1051–1073. DOI: 10.1093/imanum/dry091.
  4. [4]M. Botti and L. Mascotto. “Trace inequalities for piecewise W^{1,p} functions over general polytopic meshes”. In: arXiv preprint arXiv:2512.09752 (2025).arxiv.org/abs/2512.09752
  5. [5]A. Cangiani, Z. Dong, E. H. Georgoulis, and P. Houston. “hp-Version Discontinuous Galerkin Methods for Advection-Diffusion-Reaction Problems on Polytopic Meshes”. In: ESAIM: Mathematical Modelling and Numerical Analysis 50.3 (2016), pp. 699–725. DOI: 10.1051/m2an/2015059.DOI
  6. [6]A. Cangiani, Z. Dong, and E. H. Georgoulis. “hp-Version Discontinuous Galerkin Methods on Essentially Arbitrarily-Shaped Elements”. In: Mathematics of Computation 91.333 (2022), pp. 1–35. DOI: 10.1090/mcom/3667.DOI
  7. [8]A. Cangiani, E. H. Georgoulis, and P. Houston. “hp-Version Discontinuous Galerkin Methods on Polygonal and Polyhedral Meshes”. In: Mathematical Models and Methods in Applied Sciences 24.10 (2014), pp. 2009–2041. DOI: 10.1142/S0218202514500146.DOI
  8. [9]Y. Cao, J. Hong, and Z. Liu. “Well-Posedness and Finite Element Approximations for Elliptic SPDEs with Gaussian Noises”. In: Communications in Mathematical Research 36.2 (2020), pp. 113–127. DOI: 10.4208/cmr.2020–0006.
  9. [10]V. Cronholm, M. Evans, E. H. Georgoulis, and T. Pryer. “Multilevel Monte Carlo for Polytopic DG Discretisations with Spatially Represented Random Coefficients”. Manuscript. 2026.
  10. [11]S. G. Cox and K. Kirchner. “Regularity and convergence analysis in Sobolev and Hölder spaces for generalized Whittle–Matérn fields”. In: Numerische Mathematik 146.4 (2020), pp. 819–873. DOI: 10.1007/s00211–020–01151–x.
  11. [12]M. Croci, M. B. Giles, M. E. Rognes, and P. E. Farrell. “Efficient White Noise Sampling and Coupling for Multilevel Monte Carlo with Non-nested Meshes”. In: SIAM/ASA Journal on Uncertainty Quantification 6.4 (2018), pp. 1630–1655. DOI: 10.1137/18M1175239.DOI
  12. [14]Y. Cao, R. Zhang, and K. Zhang. “Finite Element and Discontinuous Galerkin Method for Stochastic Helmholtz Equation in Two- and Three-Dimensions”. In: Journal of Computational Mathematics 26.5 (2008), pp. 702–715.
  13. [15]Y. Cao, R. Zhang, and K. Zhang. “Finite element method and discontinuous Galerkin method for stochastic scattering problem of Helmholtz type in R^d (d= 2, 3)”. In: Potential Analysis 28.4 (2008), pp. 301–319.
  14. [16]M. Dauge. Elliptic Boundary Value Problems on Corner Domains: Smoothness and Asymptotics of Solutions. Vol. 1341. Lecture Notes in Mathematics. Berlin, Heidelberg: Springer, 1988. ISBN: 978-3-540-50169-5. DOI: 10.1007/BFb0086682.DOI
  15. [17]Q. Du and T. Zhang. “Numerical Approximation of Some Linear Stochastic Partial Differential Equations Driven by Special Additive Noises”. In: SIAM Journal on Numerical Analysis 40.4 (2002), pp. 1421–1445. DOI: 10.1137/S0036142901387956.DOI
  16. [18]M. Evans and T. Pryer. “Reyna: A minimal overhead polytopal discontinuous Galerkin finite element library”. In: Journal of Open Source Software 11.118 (2026), p. 9292.
  17. [19]H. R. Fairbanks, U. Villa, and P. S. Vassilevski. “Multilevel Hierarchical Decomposition of Finite Element White Noise with Application to Multilevel Markov Chain Monte Carlo”. In: SIAM Journal on Scientific Computing 43.5 (2021), S293–S316. DOI: 10.1137/20M1349606.DOI
  18. [20]I. Gyöngy and T. Martínez. “On Numerical Solution of Stochastic Partial Differential Equations of Elliptic Type”. In: Stochastics 78.4 (2006), pp. 213–231. DOI: 10.1080/17442500600805047.DOI
  19. [21]P. Grisvard. Elliptic Problems in Nonsmooth Domains. Vol. 24. Monographs and Studies in Mathematics. Boston, MA: Pitman Advanced Publishing Program, 1985. ISBN: 0-273-08647-2.
  20. [22]F. Lindgren, D. Bolin, and H. Rue. “The SPDE approach for Gaussian and non-Gaussian fields: 10 years and still running”. In: Spatial Statistics 50 (2022), p. 100599. DOI: 10.1016/j.spasta.2022.100599.DOI
  21. [23]S. Lasanen, L. Roininen, and J. M. J. Huttunen. “Elliptic Boundary Value Problems with Gaussian White Noise Loads”. In: Stochastic Processes and their Applications 128.11 (2018), pp. 3607–3627. DOI: 10.1016/j.spa.2017.11.007.DOI
  22. [24]F. Lindgren, H. Rue, and J. Lindström. “An Explicit Link Between Gaussian Fields and Gaussian Markov Random Fields: The Stochastic Partial Differential Equation Approach”. In: Journal of the Royal Statistical Society: Series B 73.4 (2011), pp. 423–498. DOI: 10.1111/j.1467-9868.2011.00777.x.DOI
  23. [25]B. Matérn. Spatial Variation. 2nd ed. Vol. 36. Lecture Notes in Statistics. New York, NY: Springer, 1986. DOI: 10.1007/978-1-4615-7892-5.DOI
  24. [26]F. Müller, D. Schötzau, and C. Schwab. “Symmetric Interior Penalty Discontinuous Galerkin Methods for Elliptic Problems in Polygons”. In: SIAM Journal on Numerical Analysis 55.5 (2017), pp. 2490–2521. DOI: 10.1137/17M1120634.DOI
  25. [27]P. Whittle. “On Stationary Processes in the Plane”. In: Biometrika 41.3–4 (1954), pp. 434–449. DOI: 10.1093/biomet/41.3-4.434.DOI
  26. [28]R.-M. Yao and L.-J. Bo. “Discontinuous Galerkin method for elliptic stochastic partial differential equations on two and three dimensional spaces”. In: Science in China Series A: Mathematics 50.11 (2007), pp. 1661–1672. DOI: 10.1007/s11425-007-0103-7.DOI
  27. [29]Z. Zhang, B. Rozovskii, and G. E. Karniadakis. “Strong and weak convergence order of finite element methods for stochastic PDEs with spatial white noise”. In: Numerische Mathematik 134.1 (2016), pp. 61–89. DOI: 10.1007/s00211-015-0768-8.DOI

Paper details

vlnc:W_5p5w2jap
Version
v1 · latest
Published
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents