Abstract
We study the question of approximability for the inverse of the FEM stiffness matrix for (scalar) second order elliptic boundary value problems by blockwise low rank matrices such as those given by the \(\mathcal {H}\)-matrix format introduced by Hackbusch (Computing 62(2):89–108, 1999). We show that exponential convergence in the local block rank \(r\) can be achieved. We also show that exponentially accurate \(LU\)-decompositions in the \(\mathcal {H}\)-matrix format are possible for the stiffness matrices arising in the FEM. Our analysis avoids any coupling of the block rank \(r\) to the mesh width \(h\). We also cover fairly general boundary conditions of mixed Dirichlet–Neumann–Robin boundary conditions.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
The format of \({\mathcal H}\)-matrices was introduced in [27] as blockwise low-rank matrices that permit storage, application, and even a full (approximate) arithmetic with log-linear complexity, [20, 22, 28]. This data-sparse format is well suited to represent at high accuracy matrices arising as discretizations of many integral operators, for example, those appearing in boundary integral equation methods. Also the sparse matrices that are obtained when discretizing differential operator by means of the finite element method (FEM) are amenable to a treatment by \(\mathcal {H}\)-matrices; in fact, they feature a lossless representation. Since the \(\mathcal {H}\)-matrix format comes with an arithmetic that provides algorithms to invert matrices as well as to compute \(LU\)-factorizations, approximations of the inverses of FEM matrices or their \(LU\)-factorizations are available computationally. Immediately, the question of accuracy and/or complexity comes into sight. On the one hand, the complexity of the \(\mathcal {H}\)-matrix inversion can be log-linear if the \(\mathcal {H}\)-matrix structure including the block ranks is fixed [20, 22, 28]. Then, however, the accuracy of the resulting approximate inverse is not completely clear. On the other hand, the accuracy of the inverse can be controlled by means of an adaptive arithmetic (going back at least to [20]); the computational cost at which this error control comes, is problem-dependent and not completely clear. Therefore, a fundamental question is how well the inverse can be approximated in a selected \(\mathcal {H}\)-matrix format, irrespective of algorithmic considerations. This question is answered in the present paper for FEM matrices arising from the discretization of second order elliptic boundary value problems.
It was first observed numerically in [20] that the inverse of the finite element (FEM) stiffness matrix corresponding to the Dirichlet problem for elliptic operators with bounded coefficients can be approximated in the format of \(\mathcal {H}\)-matrices with an error that decays exponentially in the block rank employed. Using properties of the continuous Green’s function for the Dirichlet problem [4] proves this exponential decay in the block rank, at least up to the discretization error. The work [6] improves on the result [4] in several ways, in particular, by proving a corresponding approximation result in the framework of \({\mathcal H}^2\)-matrices; we do not go into the details of \(\mathcal {H}^2\)-matrices here and merely mention that \(\mathcal {H}^2\)-matrices are a refinement of the concept of \(\mathcal {H}\)-matrices with better complexity properties, [7, 18, 29, 30].
Whereas the analysis of [4, 6] is based on the solution operator on the continuous level (i.e., by studying the Green’s function), the approach taken in the present article is to work on the discrete level. This seemingly technical difference has several important ramifications: First, the exponential approximability in the block rank shown here is not limited by the discretization error as in [4, 6]. Second, in contrast to [4, 6], where the block rank \(r\) and the mesh width \(h\) are coupled by \(r \sim \vert {\log h}\vert \), our estimates are explicit in both \(r\) and \(h\). Third, a unified treatment of a variety of boundary conditions is possible and indeed worked out by us. Fourth, our approach paves the way for a similar approximability result for discretizations of boundary integral operators, [16, 17]. Additionally, we mention that we also allow here the case of higher order FEM discretizations.
The last theoretical part of this paper (Sect. 5) shows that the \({\mathcal H}\)-matrix format admits \({\mathcal H}\)-\(LU\)-decompositions or \({\mathcal H}\)-Cholesky factorizations with exponential accuracy in the block rank. This is achieved, following [3, 11], by exploiting that the off-diagonal blocks of certain Schur complements are low-rank. Such an approach is closely related to the concepts of hierarchically semiseparable matrices (see, for example, [36, 42, 43] and references therein) and recursive skeletonization (see [26, 32]) and their arithmetic. In fact, several multilevel “direct” solvers for PDE discretizations have been proposed in the recent past, [19, 31, 38, 39]. These solvers take the form of (approximate) matrix factorizations. A key ingredient to their efficiency is that certain Schur complement blocks are compressible since they are low-rank. Thus, our analysis in Sect. 5 could also be of value for the understanding of these algorithms. We close by stressing that our analysis in Sect. 5 of \({\mathcal H}\)-\(LU\)-decompositions makes very few assumptions on the actual ordering of the unknowns and does not explore beneficial features of special orderings. It is well-known in the context of classical direct solvers that the ordering of the unknowns has a tremendous impact on the fill-in in factorizations. One of the most successful techniques for discretizations of PDEs are multilevel nested dissection strategies, which permit to identify large matrix blocks that will not be filled during the factorization. An in-depth complexity analysis for the \({\mathcal H}\)-matrix arithmetic for such ordering strategies can be found in [25]. The recent works [19, 31] and, in a slightly different context, [5], owe at least parts of their efficiency to the use of nested dissection techniques.
2 Main results
Let \(\Omega \subset \mathbb {R}^{d}\), \(d \in \{2,3\}\), be a bounded polygonal (for \(d = 2\)) or polyhedral (for \(d= 3\)) Lipschitz domain with boundary \(\Gamma :=\partial \Omega \). We consider differential operators of the form
where \(\mathbf {b} \in L^\infty (\Omega ; \mathbb {R}^{d}),\beta \in L^{\infty }(\Omega )\), and \(\mathbf {C} \in L^\infty (\Omega ; \mathbb {R}^{d\times d})\) is pointwise symmetric with
with certain constants \(c_1, c_2 >0\).
For \(f\in L^2(\Omega )\), we consider the mixed boundary value problem
where \(n\) denotes the outer normal vector to the surface \(\Gamma \), \(\alpha \in L^{\infty }(\Gamma _{\mathcal {R}})\), \(\alpha >0 \) and \(\Gamma = \overline{\Gamma _D} \; \cup \; \overline{\Gamma _N} \; \cup \; \overline{\Gamma _{\mathcal {R}}}\), with the pairwise disjoint and relatively open subsets \(\Gamma _D,\Gamma _N,\Gamma _{\mathcal {R}}\). With the trace operator \( \gamma _0^{\text {int}}\) we define \(H^1_0(\Omega ,\Gamma _D):= \{u\in H^1(\Omega ): \gamma _0^{\text {int}}u = 0 \;\text {on}\; \Gamma _D\}\). The bilinear form \(a: H^1_0(\Omega ,\Gamma _D) \times H^1_0(\Omega ,\Gamma _D) \rightarrow \mathbb {R}\) corresponding to (3) is given by
We additionally assume that the coefficients \(\alpha ,\mathbf {C},\mathbf {b},\beta \) are such that the coercivity
of the bilinear form \(a(\cdot ,\cdot )\) holds. Then, the Lax–Milgram Lemma implies the unique solvability of the weak formulation of our model problem.
For the discretization, we assume that \(\Omega \) is triangulated by a quasiuniform mesh \({\mathcal T}_h=\{T_1,\dots ,T_N\}\) of mesh width \(h := \max _{T_j\in \mathcal {T}_h}\mathrm{diam}(T_j)\), and the Dirichlet \(\Gamma _D\), Neumann \(\Gamma _N\), and Robin \(\Gamma _{\mathcal {R}}\)-parts of the boundary are resolved by the mesh \(\mathcal {T}_h\). The elements \(T_j \in \mathcal {T}_h\) are triangles (\(d=2\)) or tetrahedra (\(d =3\)), and we assume that \(\mathcal {T}_h\) is regular in the sense of Ciarlet. The nodes are denoted by \(x_i \in \mathcal {N}_h\), for \(i=1,\ldots ,N\). Moreover, the mesh \(\mathcal {T}_h\) is assumed to be \(\gamma \)-shape regular in the sense of \( h \sim \mathrm{diam}(T_j) \le \gamma \,|T_j|^{1/d}\) for all \(T_j\in \mathcal {T}_h\). In the following, the notation \(\lesssim \) abbreviates \(\le \) up to a constant \(C>0\) which depends only on \(\Omega \), the dimension \(d\), and \(\gamma \)-shape regularity of \(\mathcal {T}_h\). Moreover, we use \(\simeq \) to abbreviate that both estimates \(\lesssim \) and \(\gtrsim \) hold.
We consider the Galerkin discretization of the bilinear form \(a(\cdot ,\cdot )\) by continuous, piecewise polynomials of fixed degree \(p \ge 1\) in \(S^{p,1}_0({\mathcal T}_h,\Gamma _D) := S^{p,1}({\mathcal T}_h) \cap H^1_0(\Omega ,\Gamma _D)\) with \(S^{p,1}({\mathcal T}_h) = \{u \in C(\Omega )\,:\, u|_{T_j} \in \mathcal {P}_p, \, \forall \, T_j \in \mathcal {T}_h\}\), where \(\mathcal {P}_p\) denotes the space of polynomials of degree \(p\). We choose a basis of \(S^{p,1}_0({\mathcal T}_h,\Gamma _D)\), which is denoted by \({\mathcal B}_h:= \{\psi _j\, :\, j = 1,\dots , N\}\). Given that our results are formulated for matrices, assumptions on the basis \({\mathcal B}_h\) need to be imposed. For the isomorphism \(\mathcal {J}:\mathbb {R}^N\rightarrow S^{p,1}_0({\mathcal T}_h,\Gamma _D)\), \(\mathbf {x} \mapsto \sum _{j=1}^Nx_j\psi _j\), we require
Remark 1
Standard bases for \(p=1\) are the classical hat functions satisfying \(\psi _j(x_i) = \delta _{ij}\) and for \(p\ge 2\) we refer to, e.g., [13, 34, 40].
The Galerkin discretization of (4) results in a positive definite matrix \(\mathbf A\in \mathbb R^{N\times N}\) with
Our goal is to derive an \(\mathcal {H}\)-matrix approximation \(\mathbf {B}_{\mathcal {H}}\) of the inverse matrix \(\mathbf {B} =\mathbf {A}^{-1}\). An \({\mathcal H}\)-matrix \(\mathbf {B}_{\mathcal {H}}\) is a blockwise low rank matrix based on the concept of “admissibility”, which we now introduce:
Definition 1
(Bounding boxes and \(\eta \)-admissibility) A cluster \(\tau \) is a subset of the index set \(\mathcal {I} = \{1,\ldots ,N\}\). For a cluster \(\tau \subset \mathcal {I}\), we say that \(B_{R_{\tau }} \subset \mathbb {R}^d\) is a bounding box if:
-
(i)
\(B_{R_{\tau }}\) is a hyper cube with side length \(R_{\tau }\),
-
(ii)
\( \mathrm{* }{supp} \psi _j \subset B_{R_{\tau }}\) for all \( j \in \tau \).
For \(\eta > 0\), a pair of clusters \((\tau ,\sigma )\) with \(\tau ,\sigma \subset \mathcal {I}\) is \(\eta \)-admissible, if there exist boxes \(B_{R_{\tau }}\), \(B_{R_{\sigma }}\) satisfying (i)–(ii) such that
Definition 2
(blockwise rank-r matrices) Let \(P\) be a partition of \({\mathcal I} \times {\mathcal I}\) and \(\eta >0\) an admissibility parameter. A matrix \({\mathbf B}_{\mathcal {H}} \in \mathbb {R}^{N \times N}\) is said to be a blockwise rank- \(r\) matrix, if for every \(\eta \)-admissible cluster pair \((\tau ,\sigma ) \in P\), the block \({\mathbf B}_{\mathcal {H}}|_{\tau \times \sigma }\) is a rank-\(r\) matrix, i.e., it has the form \({\mathbf B}_{\mathcal {H}}|_{\tau \times \sigma } = {\mathbf X}_{\tau \sigma } {\mathbf Y}^T_{\tau \sigma }\) with \(\mathbf {X}_{\tau \sigma } \in \mathbb {R}^{\vert {\tau }\vert \times r}\) and \(\mathbf {Y}_{\tau \sigma } \in \mathbb {R}^{\vert {\sigma }\vert \times r}\). Here and below, \(\vert {\sigma }\vert \) denotes the cardinality of a finite set \(\sigma \), and \(\mathbf {B}|_{\tau \times \sigma } \in \mathbb {R}^{\left| \tau \right| \times \left| \sigma \right| }\) denotes the \(\tau \times \sigma \) subblock of a matrix \(\mathbf {B}\) as a \(\left| \tau \right| \times \left| \sigma \right| \)-matrix.
The following theorems are the main results of this paper. Theorem 1 shows that admissible blocks can be approximated by rank-\(r\) matrices:
Theorem 1
Fix the admissibility parameter \(\eta > 0\), \(q\in (0,1)\). Let the cluster pair \((\tau ,\sigma )\) be \(\eta \)-admissible. Then, for \(k \in \mathbb {N}\) there are matrices \(\mathbf {X}_{\tau \sigma } \in \mathbb {R}^{\vert {\tau }\vert \times r}\), \(\mathbf {Y}_{\tau \sigma } \in \mathbb {R}^{\vert {\sigma }\vert \times r}\) of rank \(r \le C_\mathrm{dim} (2+\eta )^dq^{-d}k^{d+1}\) such that
The constants \(C_\mathrm{apx},C_\mathrm{dim}>0\) depend only on the boundary value problem (3), \(\Omega \), \(d\), \(p\), and the \(\gamma \)-shape regularity of the quasiuniform triangulation \(\mathcal {T}_h\).
The approximations for the individual blocks can be combined to gauge the approximability of \({\mathbf A}^{-1}\) by blockwise rank-\(r\) matrices. Particularly satisfactory estimates are obtained if the blockwise rank-\(r\) matrices have additional structure. To that end, we introduce the following definitions.
Definition 3
(Cluster tree) A cluster tree with leaf size \(n_\mathrm{leaf} \in \mathbb {N}\) is a binary tree \(\mathbb {T}_{\mathcal {I}}\) with root \(\mathcal {I}\) such that for each cluster \(\tau \in \mathbb {T}_{\mathcal {I}}\) the following dichotomy holds: either \(\tau \) is a leaf of the tree and \(\vert {\tau }\vert \le n_\mathrm{leaf}\), or there exist sons \(\tau '\), \(\tau '' \in \mathbb {T}_{\mathcal {I}}\), which are disjoint subsets of \(\tau \) with \(\tau = \tau ' \cup \tau ''\). The level function \(\mathrm{level}: \mathbb {T}_{\mathcal {I}} \rightarrow \mathbb {N}_0\) is inductively defined by \(\mathrm{level}(\mathcal {I}) = 0\) and \(\mathrm{level}(\tau ') := \mathrm{level}(\tau ) + 1\) for \(\tau '\) a son of \(\tau \). The depth of a cluster tree is \(\mathrm{depth}(\mathbb {T}_{\mathcal {I}}) := \max _{\tau \in \mathbb {T}_{\mathcal {I}}}\mathrm{level}(\tau )\).
Definition 4
(Far field, near field, and sparsity constant) A partition \(P\) of \(\mathcal {I} \times \mathcal {I}\) is said to be based on the cluster tree \(\mathbb {T}_{\mathcal {I}}\), if \(P \subset \mathbb {T}_{\mathcal {I}}\times \mathbb {T}_{\mathcal {I}}\). For such a partition \(P\) and fixed admissibility parameter \(\eta > 0\), we define the far field and the near field as
The sparsity constant \(C_\mathrm{sp}\), introduced in [20], of such a partition is defined by
The following Theorem 2 shows that the matrix \(\mathbf {A}^{-1}\) can be approximated by blockwise rank-\(r\) matrices at an exponential rate in the block rank \(r\):
Theorem 2
Fix the admissibility parameter \(\eta > 0\). Let a partition \(P\) of \({\mathcal {I}} \times {\mathcal {I}}\) be based on a cluster tree \(\mathbb {T}_{\mathcal {I}}\). Then, there is a blockwise rank-\(r\) matrix \(\mathbf {B}_{\mathcal {H}}\) such that
The constants \(C_\mathrm{apx},b>0\) depend only on the boundary value problem (3), \(\Omega \), \(d\), \(p\), and the \(\gamma \)-shape regularity of the quasiuniform triangulation \(\mathcal {T}_h\).
Remark 2
Typical clustering strategies such as the “geometric clustering” described in [28] and applied to quasiuniform meshes with \(\mathcal {O}(N)\) elements lead to fairly balanced cluster trees \(\mathbb {T}_{\mathcal {I}}\) of depth \(\mathcal {O}(\log N)\) and feature a sparsity constant \(C_\mathrm{sp}\) that is bounded uniformly in \(N\). We refer to [28] for the fact that the memory requirement to store \(\mathbf {B}_{\mathcal {H}}\) is \(\mathcal {O}((r+n_\mathrm{leaf}) N \log N)\).
Remark 3
With the estimate \(\frac{1}{\left\| \mathbf {A}^{-1} \right\| _2} \lesssim N^{-1}\) from [14, Theorem 2], we get a bound for the relative error
Let us conclude this section with an observation concerning the admissibility condition (8). If the operator \(L\) is symmetric, i.e., \(\mathbf {b}=\mathbf {0}\), then the admissibility condition (8) can be replaced by the weaker admissibility condition
This follows from the fact that Proposition 1 only needs an admissibility criterion of the form \(\mathrm{diam}B_{R_{\tau }} \le \eta \, \mathrm{dist}(B_{R_{\tau }},B_{R_{\sigma }})\). Due to the symmetry of \(L\), deriving a block approximation for the block \(\tau \times \sigma \) is equivalent to deriving an approximation for the block \(\sigma \times \tau \). Therefore, we can interchange roles of the boxes \(B_{R_{\tau }}\) and \(B_{R_{\sigma }}\), and as a consequence the weaker admissibility condition (12) is sufficient. We summarize this observation in the following corollary.
Corollary 1
In the symmetric case \(\mathbf {b} = \mathbf {0}\), the results from Theorems 1 and 2 hold verbatim with the weaker admissibility criterion (12) instead of (8).
3 Low-dimensional approximation of the Galerkin solution on admissible blocks
In terms of functions and function spaces, the question of approximating the matrix block \(\mathbf {A}^{-1}|_{\tau \times \sigma }\) by a low-rank factorization \(\mathbf {X}_{\tau \sigma }\mathbf {Y}_{\tau \sigma }^T\) can be rephrased as one of how well one can approximate locally the solution of certain variational problems. More precisely, we consider, for data \(f\) supported by \(B_{R_\sigma }\cap \Omega \), the problem to find \(\phi _h \in S^{p,1}_0({\mathcal T}_h,\Gamma _D)\) such that
We remark in passing that existence and uniqueness of \(\phi _h\) follow from coercivity of \(a(\cdot ,\cdot )\). The question of approximating the matrix block \(\mathbf {A}^{-1}|_{\tau \times \sigma }\) by a low-rank factorization is intimately linked to the question of approximating \(\phi _h|_{B_{R_\tau }\cap \Omega }\) from low-dimensional spaces. The latter problem is settled in the affirmative in the following proposition for \(\eta \)-admissible cluster pairs \((\tau ,\sigma )\):
Proposition 1
Let \((\tau ,\sigma )\) be a cluster pair with bounding boxes \(B_{R_\tau }\), \(B_{R_\sigma }\). Assume that \( \eta \mathrm{dist }(B_{R_\tau },B_{R_\sigma }) \ge \mathrm{* }{diam}(B_{R_\tau })\) for a fixed admissibility parameter \(\eta > 0\). Fix \(q \in (0,1)\). Let \(\Pi ^{L^2}: L^2(\Omega ) \rightarrow S^{p,1}_0(\mathcal {T}_h,\Gamma _D)\) be the \(L^2(\Omega )\)-orthogonal projection. Then, for each \(k\in \mathbb {N}\) there exists a space \(V_k\subset S^{p,1}_0({\mathcal T}_h,\Gamma _D)\) with \(\dim V_k\le C_\mathrm{dim} (2+\eta )^d q^{-d}k^{d+1}\) such that for arbitrary \(f \in L^{2}(\Omega )\) with \(\mathrm{* }{supp} f \subset B_{R_\sigma }\cap \Omega \subset F_{\tau }:=\{x \in \Omega : \eta \mathrm{* }{dist}(x,B_{R_{\tau }})\ge \mathrm{* }{diam}(B_{R_{\tau }})\}\), the solution \(\phi _h\) of (13) satisfies
The constant \(C_\mathrm{box}>0\) depends only on the boundary value problem (3) and \(\Omega \), while \(C_\mathrm{dim}>0\) additionally depends on \(p\), \(d\), and the \(\gamma \)-shape regularity of the quasiuniform triangulation \(\mathcal {T}_h\).
The proof of Proposition 1 will be given at the end of this section. The basic steps are as follows: First, one observes that \(\mathrm{* }{supp} f \subset B_{R_{\sigma }}\cap \Omega \) together with the admissibility condition \(\mathrm{dist}(B_{R_{\tau }},B_{R_{\sigma }})\ge \eta ^{-1}\mathrm{diam}(B_{R_{\tau }}) > 0\) imply the orthogonality condition
Second, this observation will allow us to prove a Caccioppoli-type estimate (Lemma 2) in which stronger norms of \(\phi _h\) are estimated by weaker norms of \(\phi _h\) on slightly enlarged regions. Third, we proceed as in [4, 6] by iterating an approximation result (Lemma 3) derived from Scott–Zhang interpolation of the Galerkin solution \(\phi _h\). This iteration argument accounts for the exponential convergence (Lemma 4).
3.1 The space \({\mathcal H}_h(D,\omega )\) and a Caccioppoli type estimate
It will be convenient to introduce, for index sets \(\rho \subset \mathcal {I}\), the set
we will implicitly assume henceforth that such sets are unions of elements. Let \(D\subset \mathbb {R}^d\) be a bounded open set and \(\omega \subset \Omega \) be of the form (16). The orthogonality property that we have identified in (15) is captured by the following space \({\mathcal H}_{h}(D,\omega )\):
For the proof of Proposition 1 and subsequently Theorems 1 and 2, we will only need the special case \(\omega = \Omega \); the general case \({\mathcal H}_h(D,\omega )\) with \(\omega \ne \Omega \) will be required in our analysis of \(LU\)-decompositions in Sect. 5.2.
Clearly, the finite dimensional space \(\mathcal {H}_{h}(D,\omega )\) is a closed subspace of \(H^1(D\cap \omega )\), and we have \(\phi _h \in \mathcal {H}_{h}(B_{R_{\tau }},\Omega )\) for the solution \(\phi _h\) of (13) with \(\mathrm{* }{supp} f \subset B_{R_{\sigma }}\cap \Omega \) and bounding boxes \(B_{R_{\tau }},B_{R_{\sigma }}\) that satisfy the \(\eta \)-admissibility criterion (8). Since multiplications of elements of \({\mathcal H}_h(D,\omega )\) with cut-off function and trivial extensions to \(\Omega \) appear repeatedly in the sequel, we note the following very simple lemma:
Lemma 1
Let \(\omega \) be a union of elements, \(D \subset \mathbb {R}^d\) be bounded and open, and \(\eta \in W^{1,\infty }(\mathbb {R}^d)\) with \(\mathrm{* }{supp} \eta \subset D\). For \(u \in {\mathcal H}_h(D,\omega )\) define the function \(\eta u\) pointwise on \(\Omega \) by \((\eta u)(x):= \eta (x) u(x)\) for \(x \in D\cap \omega \) and \((\eta u)(x) = 0\) for \(x \not \in D \cap \omega \). Then
-
(i)
\(\eta u \in H^1_0(\Omega ;\Gamma _D)\)
-
(ii)
\(\mathrm{* }{supp} (\eta u) \subset \overline{D\cap \omega }\)
-
(iii)
If \(\eta \in S^{q,1}({\mathcal T}_h)\), then \(\eta u \in S^{p+q,1}_0({\mathcal T}_h, \Gamma _D)\).
Proof
We only illustrate (i). Given \(u \in {\mathcal H}_h(D,\omega )\) there exists by definition a function \(\widetilde{u} \in S^{p,1}_0({\mathcal T}_h,\Gamma _D)\) with \(\mathrm{* }{supp} \widetilde{u} \subset \overline{\omega }\). By the support properties of \(\eta \) and \(\widetilde{u}\), the function \(\eta u\) coincides with \(\eta \widetilde{u}\). As the product of an \(H^1(\Omega )\)-function and a Lipschitz continuous function, the function \(\eta \widetilde{u}\) is in \(H^1(\Omega )\).
A main tool in our proofs is the nodal interpolation operator \(J_h: C(\overline{\Omega }) \rightarrow S^{p,1}_0({\mathcal T}_h, \Gamma _D)\). Since \(p+1>\frac{d}{2}\), the interpolation operator \(J_h\) has the following local approximation property for continuous, \(\mathcal {T}_h\)-piecewise \(H^{p+1}\)-functions \(u \in C(\overline{\Omega })\cap H^{p+1}_{\text {pw}}(\mathcal {T}_h,\omega ) := \{u\in L^2(\omega ) : u|_{T} \in H^{p+1}(T)\, \forall \, T \in \mathcal {T}_h\}\)
The constant \(C>0\) depends only on \(\gamma \)-shape regularity of the quasiuniform triangulation \(\mathcal {T}_h\), the dimension \(d\), and the polynomial degree \(p\). In particular, it is independent of the choice of the set \(\omega \).
In the following, we will construct approximations on nested boxes and therefore introduce the notion of concentric boxes.
Definition 5
(Concentric boxes) Two boxes \(B_R\), \(B_{R'}\) of side length \(R,R'\) are said to be concentric, if they have the same barycenter and \(B_R\) can be obtained by a stretching of \(B_{R'}\) by the factor \(R/R'\) taking their common barycenter as the origin.
For a box \(B_R\) with side length \(R \le 2 \,\mathrm{diam}(\Omega )\), we introduce the norm
which is, for fixed \(h\), equivalent to the \(H^1\)-norm.
The following lemma states a discrete Caccioppoli-type estimate for functions in \(\mathcal {H}_h(B_{(1+\delta )R},\omega )\), where \(B_{(1+\delta )R}\) and \(B_R\) are concentric boxes. In contrast to the classical, continuous Caccioppoli inequality, an additional assumption on the size parameters \(R,\delta \) of the box \(B_{(1+\delta )R}\) compared to the mesh size \(h\) has to be made.
Lemma 2
Let \(\delta \in (0,1)\), \(R\in (0,2\mathrm{* }{diam}(\Omega ))\) such that \(\frac{h}{R} \le \frac{\delta }{4}\) and let \(\omega \subseteq \Omega \) be of the form (16). Let \(B_{R}\), \(B_{(1+\delta )R}\) be two concentric boxes. Let \(u\in \mathcal {H}_h(B_{(1+\delta )R},\omega )\). Then, there exists a constant \(C_{\text {reg}} > 0\), which depends only on the boundary value problem (3), \(\Omega \), \(d\), \(p\), and the \(\gamma \)-shape regularity of the quasiuniform triangulation \(\mathcal {T}_h\), such that
Proof
Let \(\eta \in S^{1,1}(\mathcal {T}_h)\) be a piecewise affine cut-off function with \(\mathrm{* }{supp} \eta \subset B_{(1+\delta /2)R}\cap \overline{\Omega }\), \(\eta \equiv 1\) on \(B_{R}\cap \omega \), \(0\le \eta \le 1\), and \(\left\| \nabla \eta \right\| _{L^{\infty }(B_{(1+\delta )R}\cap \Omega )} \lesssim \frac{1}{\delta R}\). By Lemma 1 we have \(\eta ^2 u \in S^{p+2,1}_0({\mathcal T}_h,\Gamma _D) \subset H^1_0(\Omega ;\Gamma _D)\) and
Recall that \(h\) is the maximal element diameter and \(4h\le \delta R\). Hence, for the nodal interpolation operator \(J_h\), we have \(\mathrm{* }{supp} J_h (\eta ^2 u) \subset \overline{B_{(1+\delta ) R}}\); in view of the locality of the nodal interpolation, we furthermore have \(\mathrm{* }{supp} J_h (\eta ^2 u) \subset \overline{\omega }\) so that
With the coercivity of the bilinear form \(a(\cdot ,\cdot )\) and \(\frac{1}{\delta R} \lesssim \frac{1}{\delta ^2R^2}\), since \(\delta < 1\) and \(R \le 2\mathrm{* }{diam}(\Omega )\), we have
Recall from (21) that \(\mathrm{* }{supp} J_h (\eta ^2 u) \subset \overline{B}\). The orthogonality relation (17) in the definition of the space \(\mathcal {H}_h(B,\omega )\) therefore implies
The approximation property (18) and the support properties of \(\eta ^2 u\) lead to
By \(D^k u\), we denote the derivative \(D^k u := \frac{\partial ^{\left| k \right| }u}{\partial x_1^{k_1}\cdot \dots \partial x_d^{k_d}}\) with the multi-index \(k\in \mathbb {N}_0^d\) with \(\left| k \right| = \sum _{i=1}^dk_i\). Since, for each \(T \subset B\) we have \(u|_T \in {\mathcal P}_p\), we get \(D^{k}u|_T = 0\) for all multi-indices \(k\in \mathbb {N}_0^d\) with \(\left| k \right| = p+1\). The assumption \(\eta \in S^{1,1}(\mathcal {T}_h)\) implies \(D^k \eta |_T = 0\) for all \(k\in \mathbb {N}_0^d\) with \(\left| k \right| \ge 2\). With the Leibniz product rule, the right-hand side of (24) can therefore be estimated by
where the suppressed constant depends on \(p\). The inverse inequality \(\displaystyle \left| \eta u \right| _{H^p(T)} \lesssim h^{-p+1}\left\| \nabla (\eta u) \right\| _{L^2(T)} \), see, e.g., [12], leads to
The same line of reasoning leads to
In order to derive an estimate for the boundary term in (23), we need a second smooth cut-off function \(\widetilde{\eta }\) with \(\mathrm{* }{supp} \widetilde{\eta }\subset \overline{B_{(1+\delta )R}}\) and \(\widetilde{\eta }\equiv 1\) on \(\mathrm{* }{supp} (J_h(\eta ^2 u) - \eta ^2 u)\) and \(\Vert \nabla \widetilde{\eta }\Vert _{L^\infty (B_{(1+\delta )R})} \lesssim \frac{1}{\delta R}\). By Lemma 1 we can define the function \(\widetilde{\eta }u \in H^1(\Omega )\) with the support property \(\displaystyle \mathrm{* }{supp} \widetilde{\eta }u \subset \overline{B_{(1+\delta )R} \cap \omega } = \overline{B} \) and therefore
Then, we get
The multiplicative trace inequality (see, e.g., [9, inequality (1.6.2)]) for \(\Omega \) and the estimate (27) gives
The multiplicative trace inequality for \(\Omega \) and the estimates (25)–(26) imply
Therefore,
Young’s inequality and \(h/(\delta R) \le 1/4\), as well as \(\delta \in (0,1), R\le 2\mathrm{* }{diam}(\Omega )\) implying \(\frac{1}{\delta R}\lesssim \frac{1}{\delta ^2R^2}\), allow us to conclude (rather generously)
Inserting the estimates (25), (26), (28) into (23) and with Young’s inequality, we get with (22b) that
Moving the term \(\frac{1}{2} \left\| \eta \nabla u \right\| _{L^2(B)}^2\) to the left-hand side and inserting this estimate in (22a), we conclude the proof.
3.2 Low-dimensional approximation in \({\mathcal {H}_{h}}(D,\omega )\)
In this subsection, we will derive a low dimensional approximation of the Galerkin solution by Scott–Zhang interpolation on a coarser grid.
We need to be able to extend functions defined on \(B_{(1+2\delta )R}\cap \omega \) to \({\mathbb R}^d\). To this end, we use an extension operator \(E: H^1(\Omega ) \rightarrow H^1(\mathbb {R}^d)\), see, e.g., [1, Theorem 4.32], which satisfies \(E u = u\) on \(\Omega \) and the \(H^1\)-stability estimate
For a function \(u \in \mathcal {H}_h(B_{(1+2\delta )R},\omega )\) and a cut-off function \(\eta \in C_0^{\infty }(B_{(1+2\delta )R})\) with \(\mathrm{* }{supp} \eta \subset B_{(1+\delta )R}\), \(\eta \equiv 1\) on \(B_{R}\) we can define the function \(\eta u \in H^1(\Omega )\) with the aid of Lemma 1. We note the support property \(\mathrm{* }{supp} (\eta u) \subset \overline{B_{(1+2\delta )R} \cap \omega }\), due to \(\mathrm{* }{supp} u \subset \overline{\omega }\). Therefore, the extension of \(\eta u\) to \(\Omega \) by zero is in \(H^1(\Omega )\). Therefore, we have
Moreover, let \(\Pi _{h,R} : (H^1(B_R\cap \omega ),\left| \! \left| \! \left| \cdot \right| \! \right| \! \right| _{h,R}) \rightarrow (\mathcal {H}_h(B_R,\omega ), \left| \! \left| \! \left| \cdot \right| \! \right| \! \right| _{h,R})\) be the orthogonal projection, which is well-defined since \(\mathcal {H}_h(B_R,\omega )\subset H^1(B_R\cap \omega )\) is a closed subspace.
In the following lemma, we use a Scott–Zhang projection \(I_H: H^1(\Omega ) \rightarrow S^{1,1}({\mathcal K}_H)\) of the form introduced in [41] for a quasiuniform grid \(\mathcal {K}_H\) with mesh width \(H\). By
we denote the element patch of \(K\), which contains \(K\) and all elements \(K' \in \mathcal {K}_H\) that have a common node with \(K\). Then, \(I_H\) has the following local approximation property for \(u \in H^1(\omega _K)\)
The constant \(C>0\) depends only on \(\gamma \)-shape regularity of the quasiuniform triangulation \(\mathcal {K}_H\) and the dimension \(d\).
Lemma 3
Let \(\delta \in (0,1)\), \(R\in (0,2\mathrm{* }{diam}(\Omega ))\), \(B_R\), \(B_{(1+\delta )R}\), and \(B_{(1+2\delta )R}\) be concentric boxes, and let \(\omega \subseteq \Omega \) of the form (16) and \(u\in \mathcal {H}_h(B_{(1+2\delta )R},\omega )\). Assume \(\frac{h}{R}\le \frac{\delta }{4}\). Let \(\mathcal {K}_H \) be an (infinite) \(\gamma \)-shape regular triangulation of \(\mathbb {R}^d\) and assume \(\frac{H}{R} \le \frac{\delta }{4}\) for the corresponding mesh width \(H\). Let \(\eta \in C_0^{\infty }(B_{(1+2\delta )R})\) be a cut-off function satisfying \(\mathrm{* }{supp} \eta \subset B_{(1+\delta )R}\), \(\eta \equiv 1\) on \(B_{R}\), and \(\left\| \nabla \eta \right\| _{L^{\infty }(B_{(1+2\delta )R})} \lesssim \frac{1}{\delta R}\). Moreover, let \(I_H: H^1(\mathbb {R}^d) \rightarrow S^{1,1}(\mathcal {K}_H)\) be the Scott–Zhang projection and \(E:H^1(\Omega ) \rightarrow H^1(\mathbb {R}^d)\) be an \(H^1\)-stable extension operator. Then, there exists a constant \(C_\mathrm{app} > 0\), which depends only on the boundary value problem (3), \(\Omega \), \(d\), \(p\), the \(\gamma \)-shape regularity of the quasiuniform triangulation \(\mathcal {T}_h\), and \(E\) such that
-
(i)
\((u-\Pi _{h,R}I_H E (\eta u))|_{B_{R}\cap \omega } \in \mathcal {H}_h(B_{R},\omega )\);
-
(ii)
\(\left| \! \left| \! \left| u-\Pi _{h,R}I_H E(\eta u) \right| \! \right| \! \right| _{h,R} \le C_\mathrm{app}\frac{1+2\delta }{\delta } (\frac{h}{R}+\frac{H}{R})\left| \! \left| \! \left| u \right| \! \right| \! \right| _{h,(1+2\delta )R}\);
-
(iii)
\(\dim W\le C_\mathrm{app}(\frac{(1+2\delta )R}{H})^d\), where \(W:=\Pi _{h,R}I_H E \mathcal {H}_h(B_{(1+2\delta )R},\omega ) \).
Proof
The statement (iii) follows from the fact that \(\dim I_H(E \mathcal {H}_h(B_{(1+2\delta )R},\omega ))|_{B_{(1+\delta )R}} \lesssim ((1+2\delta )R/H)^d\). For \(u \in \mathcal {H}_h(B_{(1+2\delta )R},\omega )\), we have \(u|_{B_R \cap \omega } \in \mathcal {H}_h(B_{R},\omega )\) as well and hence \(\Pi _{h,R}(u|_{B_{R}\cap \omega }) = u|_{B_{R}\cap \omega }\), which gives (i). It remains to prove (ii): The assumption \(\frac{H}{R} \le \frac{\delta }{4}\) implies \(\bigcup \{K \in \mathcal {K}_H \;:\; \omega _K \cap B_{R} \ne \emptyset \} \subseteq B_{(1+\delta )R}\). The locality and the approximation properties (30) of \(I_H\) yield
We apply Lemma 2 with \(\widetilde{R} = (1+\delta )R\) and \(\widetilde{\delta } = \frac{\delta }{1+\delta }\). Note that \((1+\widetilde{\delta })\widetilde{R} = (1+2\delta )R\), and \(\frac{h}{\widetilde{R}}\le \frac{\widetilde{\delta }}{4}\) follows from \(4h \le \delta R = \widetilde{\delta }\widetilde{R}\). Hence, we obtain with (29)
which concludes the proof.
By iterating this approximation result on suitable concentric boxes, we can derive a low-dimensional subspace in the space \(\mathcal {H}_h\) and the bestapproximation in this space converges exponentially, which is stated in the following lemma.
Lemma 4
Let \(C_\mathrm{app}\) be the constant of Lemma 3. Let \(q,\kappa \in (0,1)\), \(R\in (0,2\mathrm{* }{diam}(\Omega ))\), \(k \in \mathbb {N}\) and \(\omega \subseteq \Omega \) be of the form (16). Assume
Then, there exists a subspace \(V_k\) of \(S^{p,1}_0(\mathcal {T}_h,\Gamma _D)|_{B_R \cap \omega }\) with dimension
such that for every \(u \in \mathcal {H}_{h}(B_{(1+\kappa )R},\omega )\)
The constant \(C_\mathrm{dim}>0\) depends only on the boundary value problem (3), \(\Omega ,d\), and the \(\gamma \)-shape regularity of the quasiuniform triangulation \(\mathcal {T}_h\).
Proof
We iterate the approximation result of Lemma 3 on boxes \(B_{(1+\delta _j)R}\), with \(\delta _j := \kappa \frac{k-j}{k}\) for \(j=0,\dots ,k\). We note that \(\kappa = \delta _0>\delta _1 > \dots >\delta _k = 0\). We choose \(H = \frac{\kappa q R}{8k\max \{C_\mathrm{app},1\}}\).
If \(h \ge H\), then we select \(V_k = \mathcal {H}_h(B_R,\omega )\). Due to the choice of \(H\) we have \(\dim V_k \lesssim (\frac{R}{h})^d \lesssim k(\frac{R}{H})^d \simeq C_\mathrm{dim} (\frac{1 + \kappa ^{-1}}{q})^dk^{d+1}\).
If \(h < H\), we apply Lemma 3 with \(\widetilde{R} = (1+\delta _j)R\) and \(\widetilde{\delta }_j = \frac{1}{2k(1+\delta _j)} < \frac{1}{2}\). Note that \(\delta _{j-1} = \delta _j +\frac{1}{k}\) gives \((1+\delta _{j-1})R=(1+2\widetilde{\delta }_j)\widetilde{R}\). The assumption \(\frac{H}{\widetilde{R}}\le \frac{1}{8k(1+\delta _j)} = \frac{\widetilde{\delta }_j}{4}\) is fulfilled due to our choice of \(H\). For \(j=1\), Lemma 3 provides an approximation \(w_1\) in a subspace \(W_1\) of \(\mathcal {H}_h(B_{(1+\delta _1)R},\omega )\) with \(\dim W_1 \le C(\frac{(1+\kappa )R}{H})^d\) such that
Since \(u-w_1 \in \mathcal {H}_h(B_{(1+\delta _1)R},\omega )\), we can use Lemma 3 again and get an approximation \(w_2\) of \(u-w_1\) in a subspace \(W_2\) of \(\mathcal {H}_h(B_{(1+\delta _2)R},\omega )\) with \(\dim W_2\le C(\frac{(1+\kappa )R}{H})^d\). Arguing as for \(j=1\), we get
Continuing this process \(k-2\) times leads to an approximation \(v := \sum _{j=1}^kw_i\) in the space \(V_k := \sum _{j=1}^{k}W_j\) of dimension \(\dim V_k \le Ck(\frac{(1+\kappa )R}{H})^d=C_\mathrm{dim} (\frac{1 + \kappa ^{-1}}{q})^dk^{d+1}\).
Now we are able to prove the main result of this section.
Proof
of Proposition 1 Choose \(\kappa = \frac{1}{1+\eta }\). By assumption, we have \(\mathrm{dist }(B_{R_\tau },B_{R_\sigma }) \ge \eta ^{-1} \mathrm{* }{diam} B_{R_\tau } = \sqrt{d} \eta ^{-1} R_{\tau }\). In particular, this implies
The Galerkin solution \(\phi _h\) satisfies \(\phi _h|_{B_{(1+\delta )R}\cap \Omega } \in \mathcal {H}_h(B_{(1+\delta )R},\Omega )\). The coercivity (5) of the bilinear form \(a(\cdot ,\cdot )\) implies
Furthermore, with \(\frac{h}{R_{\tau }}<1\), we get
and we have a bound on the right-hand side of (32). We are now in the position to define the space \(V_k\), for which we distinguish two cases.
Case 1: The condition (31) is satisfied with \(R = R_{\tau }\). With the space \(V_k\) provided by Lemma 4 we get
and the dimension of \(V_k\) is bounded by \(\dim V_k \le C ((2 + \eta ) q^{-1})^d k^{d+1}.\)
Case 2: The condition (31) is not satisfied, i.e., we have \(\frac{h}{R_{\tau }} \ge \frac{\kappa q}{8 k \max \{1,C_\mathrm{app}\}}\).
Then, we select \(V_k:= \{v|_{B_{R_{\tau }}\cap \Omega } : v \in S^{p,1}_0({\mathcal T}_h,\Gamma _D)\}\), and the minimum in (14) is obviously zero. By choice of \(\kappa \), the dimension of \(V_k\) is bounded by
which concludes the proof of the non trivial statement in (14). The other estimate follows directly from the \(L^2(\Omega )\)-stability of the \(L^2(\Omega )\)-orthogonal projection.
Remark 4
A result similar to Proposition 1 holds for the pure Neumann problem, i.e., \(\Gamma = \Gamma _{\mathcal {N}}\) as well. In this case, the matrix \(\mathbf {A}\) is not invertible and therefore either a stabilized Galerkin discretization or a saddle point formulation has to be chosen to deal with the one-dimensional kernel. The general ideas underlying Proposition 1 can be utilized, see [15, 17] for details.
4 Proof of main results
We use the approximation of \(\phi _h\) from the low dimensional spaces given in Proposition 1 to construct a blockwise low-rank approximation of \(\mathbf {A}^{-1}\) and in turn an \(\mathcal {H}\)-matrix approximation of \(\mathbf {A}^{-1}\). In fact, we will only use a FEM-isomorphism to transfer Proposition 1 to the matrix level, which follows the lines of [6, Theorem 2].
Proof
of Theorem 1 If \(C_\mathrm{dim} (2+\eta )^d q^{-d} k^{d+1} \ge \min (\vert {\tau }\vert ,\vert {\sigma }\vert )\), we use the exact matrix block \(\mathbf {X}_{\tau \sigma }=\mathbf {A}^{-1}|_{\tau \times \sigma }\) and \(\mathbf {Y}_{\tau \sigma } = \mathbf {I} \in \mathbb {R}^{\vert {\sigma }\vert \times \vert {\sigma }\vert }\).
If \(C_\mathrm{dim} (2+\eta )^d q^{-d} k^{d+1} < \min (\vert {\tau }\vert ,\vert {\sigma } \vert )\), let \(\lambda _i:L^2(\Omega ) \rightarrow \mathbb {R}\) be continuous linear functionals on \(L^2(\Omega )\) satisfying \(\lambda _i(\psi _j) = \delta _{ij}\). We define \(\mathbb {R}^{\tau }:= \{\mathbf {x} \in \mathbb {R}^N \, : \, x_i = 0 \; \forall \; i \notin \tau \}\) and the mappings
For \(\mathbf {x} \in \mathbb {R}^{\tau }\), (6) leads to the stability estimate
Let \(V_k\) be the finite dimensional subspace from Proposition 1.
Because of (33) and the \(L^2\)-stability of \(\mathcal {J}_{\mathcal {I}} \Lambda _{\mathcal {I}}\), the adjoint \(\Lambda _{\mathcal {I}}^* : \mathbb {R}^N \rightarrow L^2(\Omega )' \simeq L^2(\Omega )\) of \(\Lambda _{\mathcal {I}}\) satisfies
Moreover, if \(\mathbf {b} = (\left\langle f,\psi _i \right\rangle )_{i\in {\mathcal {I}}}\), we have \((\Lambda _{\mathcal {I}}^* \mathbf {b})(\psi _i) = b_i = \left\langle f,\psi _i \right\rangle = \langle \Pi ^{L^2}f,\psi _i \rangle \). Therefore, \(f\) and \(\Lambda _{\mathcal {I}}^* \mathbf {b} = \Pi ^{L^2} f\) have the same Galerkin approximation.
Let \(V_k\) be the finite dimensional subspace from Proposition 1. We define \(\mathbf {X}_{\tau \sigma }\) as an orthogonal basis of the space \(\mathcal {V}_{\tau } := \{\Lambda _{\tau } v : v \in V_k \}\). Then, the rank of \(\mathbf {X}_{\tau \sigma }\) is bounded by \(\dim V_k \le C_\mathrm{dim}(2+\eta )^d q^{-d} k^{d+1}\).
The estimate (33) and the approximation result from Proposition 1 provide the error estimate
Since \(\mathbf {X}_{\tau \sigma } \mathbf {X}_{\tau \sigma }^T\) is the orthogonal projection from \(\mathbb {R}^N\) onto \(\mathcal {V}_{\tau }\), we get that \(z:=\mathbf {X}_{\tau \sigma } \mathbf {X}_{\tau \sigma }^T \Lambda _{\tau } \phi _h\) is the best approximation of \(\Lambda _{\tau } \phi _h\) in \(\mathcal {V}_{\tau }\) and arrive at
If we define \(\mathbf {Y}_{\tau ,\sigma } := \mathbf {A}^{-1}|_{\tau \times \sigma }^{T}\mathbf {X}_{\tau \sigma }\), we get \(z = \mathbf {X}_{\tau \sigma } \mathbf {Y}_{\tau \sigma }^T \mathbf {b}\), since \(\Lambda _{\tau } \phi _h = \mathbf {A}^{-1}|_{\tau \times \sigma }\mathbf {b}\).
The following lemma gives an estimate for the global spectral norm by the local spectral norms, which we will use in combination with Theorem 1 to derive our main result, Theorem 2.
Lemma 5
[7, 20, 28, Lemma 6.5.8] Let \(\mathbf {M} \in \mathbb {R}^{N\times N}\) and \(P\) be a partitioning of \({\mathcal {I}}\times {\mathcal {I}}\). Then,
Now we are able to prove our main result, Theorem 2.
Proof
of Theorem 2 For each admissible cluster pair \((\tau ,\sigma )\), Theorem 1 provides matrices \(\mathbf {X}_{\tau \sigma } \in \mathbb {R}^{\vert {\tau }\vert \times r}\), \(\mathbf {Y}_{\tau \sigma } \in \mathbb {R}^{r \times \vert {\sigma } \vert }\), so that we can define the \(\mathcal {H}\)-matrix \(\mathbf {V}_{\mathcal {H}}\) by
On each admissible block \((\tau ,\sigma ) \in P_{\text {far}}\), we can use the blockwise estimate of Theorem 1 and get for \(q\in (0,1)\)
On inadmissible blocks, the error is zero by definition. Therefore, Lemma 5 concludes the proof, since
Since in Theorem 1 we have \(r \le C_\mathrm{dim}(2+\eta )^dq^{-d}k^{d+1}\), defining \(b = -\frac{\ln (q)}{C_\mathrm{dim}^{1/(d+1)}}q^{d/(d+1)}(2+\eta )^{-d/(1+d)} > 0\), we obtain \(q^k = e^{-br^{1/(d+1)}}\) and hence
which concludes the proof.
5 Hierarchical \(LU\)-decomposition
In [3] the existence of an (approximate) \(\mathcal {H}\)-\(LU\) decomposition, i.e., a factorization of the form \(\mathbf {A} \approx \mathbf {L}_{\mathcal {H}}\mathbf {U}_{\mathcal {H}}\) with lower and upper triangular \(\mathcal {H}\)-matrices \(\mathbf {L}_{\mathcal {H}}\) and \(\mathbf {U}_{\mathcal {H}}\), was asserted for finite element matrices \(\mathbf {A}\) corresponding to the Dirichlet problem for elliptic operators with \(L^{\infty }\)-coefficients. In [25] this result was extended to the case, where the block structure of the \(\mathcal {H}\)-matrix is constructed by domain decomposition clustering methods, instead of the standard geometric bisection clustering.
Algorithms for computing an \(\mathcal {H}\)-\(LU\) decomposition have been proposed repeatedly in the literature, e.g., [2, 37] and numerical evidence for their usefulness put forward; we mention here that \(\mathcal {H}\)-\(LU\) decomposition can be employed for black box preconditioning in iterative solvers, [2, 21, 23, 24, 35]. An existence result for \(\mathcal {H}\)-\(LU\) factorization is then an important step towards a mathematical understanding of the good performance of these schemes.
The main steps in the proof of [3] are to approximate certain Schur complements of \(\mathbf {A}\) by \(\mathcal {H}\)-matrices and to show a recursion formula for the Schur complement. Using these two observations an approximation of the exact \(LU\)-factors for the Schur complements, and consequently for the whole matrix, can be derived recursively.
Since the construction of the approximate \(LU\)-factors is completely algebraic, once we know that the Schur complements have an \(\mathcal {H}\)-matrix approximation of arbitrary accuracy, we will show that we can provide such an approximation and only sketch the remaining steps. Details can be found in [3, 25].
Our main result, Theorem 2, shows the existence of an \(\mathcal {H}\)-matrix approximation to the inverse FEM stiffness matrix with arbitrary accuracy, whereas previous results achieve accuracy up to the finite element error. In fact, both [3, 25] assume, in order to derive an \(\mathcal {H}\)-\(LU\) decomposition, that approximations to the inverse with arbitrary accuracy exist. Thus, due to our main result this assumption is fulfilled for inverse finite element matrices for elliptic operators with various boundary conditions.
Since we are in the setting of the Lax–Milgram Lemma, we get that the, in general, non symmetric matrix \(\mathbf {A}\) is positive definite in the sense that \(\mathbf {x}^T\mathbf {A}\mathbf {x}>0\) for all \(\mathbf {x} \ne 0\). Therefore, \(\mathbf {A}\) has an \(LU\)-decomposition \(\mathbf {A}=\mathbf {L}\mathbf {U}\), where \(\mathbf {L}\) is a lower triangular matrix and \(\mathbf {U}\) is an upper triangular matrix, independently of the numbering of the degrees of freedom, i.e., every other numbering of the basis functions permits an \(LU\)-decomposition as well (see, e.g., [33, Cor. 3.5.6]). By classical linear algebra (see, e.g., [33, Cor. 3.5.6]), this implies that for any \(n \le N\) and index set \(\rho := \{1,\ldots ,n\}\), the matrix \(\mathbf {A}|_{\rho \times \rho }\) is invertible.
We start with the approximation of appropriate Schur complements.
5.1 Schur complements
One way to approximate the Schur complement for a finite element matrix is to follow the lines of [3, 25] by using \(\mathcal {H}\)-arithmetics and the sparsity of the finite element matrix. We present a different way of deriving such a result, which is more in line with our procedure in Sect. 3. It relies on interpreting Schur complements as FEM stiffness matrices from constrained spaces.
Lemma 6
Let \((\tau ,\sigma )\) be an admissible cluster pair and \(\rho := \{i\in \mathcal {I} : i < \min (\tau \cup \sigma )\}\). Define the Schur complement \(\mathbf {S}(\tau ,\sigma ) = \mathbf {A}|_{\tau \times \sigma } - \mathbf {A}|_{\tau \times \rho } (\mathbf {A}|_{\rho \times \rho })^{-1}\mathbf {A}|_{\rho \times \sigma }\). Then, there exists a rank-\(r\) matrix \(\mathbf {S}_{\mathcal {H}}(\tau ,\sigma )\) such that
where the constant \(C_\mathrm{sc}>0\) depends only on the boundary value problem (3), \(\Omega \), \(p\), \(d\), and the \(\gamma \)-shape regularity of the quasiuniform triangulation \(\mathcal {T}_h\).
Proof
We define \(\omega _{\rho } = \mathrm{interior}(\bigcup _{i\in \rho }\mathrm{* }{supp} \psi _i) \subset \Omega \) and let \(B_{R_{\tau }},B_{R_{\sigma }}\) be bounding boxes for the clusters \(\tau \), \(\sigma \) with (8). Our starting point is the well-known observation that the Schur complement matrix \(\mathbf {S}(\tau ,\sigma )\) can be understood in terms of an orthogonalization with respect to the degrees of freedom in \(\rho \). That is, for \(\mathbf {u}\in \mathbb {R}^{\vert {\tau }\vert },\mathbf {w}\in \mathbb {R}^{\vert {\sigma }\vert }\) a direct calculation (see, e.g., [10] for the essentials) shows
with \(w = \sum _{j=1}^{\vert {\sigma }\vert }\mathbf {w}_j\psi _{j_{\sigma }}\), where the index \(j_{\sigma }\) denotes the \(j\)-th basis function corresponding to the cluster \(\sigma \), and the function \(\widetilde{u} \in S^{p,1}_0(\mathcal {T}_h,\Gamma _D)\) is defined by \(\widetilde{u} = \sum _{j=1}^{\vert {\tau }\vert }\mathbf {u}_j\psi _{j_{\tau }} + u_{\rho }\) with \(\mathrm{* }{supp} u_{\rho } \subset \overline{\omega _{\rho }}\) such that
The key to approximate the Schur complement \(\mathbf {S}(\tau ,\sigma )\) is to approximate the function \(\widetilde{u}\). We will provide such an approximation by applying the techniques from the previous sections with the use of the orthogonality (35).
Since \(\mathrm{* }{supp} \widetilde{u} \subset B_{R_{\tau }} \cup \overline{\omega _{\rho }}\), we get for \(w\) with \(\mathrm{* }{supp} w \subset B_{R_{\sigma }}\) that
Therefore, we only need to approximate \(\widetilde{u}\) on the intersection \(B_{R_{\sigma }}\cap \omega _{\rho }\). This support property and the orthogonality (35) imply that \(\widetilde{u} \in \mathcal {H}_h(B_{(1+\delta )R_{\sigma }},\omega _{\rho })\).
Therefore, Lemma 2 can be applied to \(\widetilde{u}\). As a consequence, Lemma 4 provides a low dimensional space \( V_k\), where the choice \(\kappa =\frac{1}{\eta +1}\) bounds the dimension of \(V_k\) by \(\dim V_k \le C_\mathrm{dim}(2+\eta )^dq^{-d}k^{d+1}\). Moreover, the best approximation \(\widetilde{v} = \Pi _{V_k}\widetilde{u} \in V_k\) to \(\widetilde{u}\) in the space \(V_k\) satisfies
This implies
Since \(\mathrm{* }{supp}(\widetilde{u}-u) = \mathrm{* }{supp}(u_{\rho }) \subset \overline{\omega _{\rho }}\) with \(u = \sum _{j=1}^{\vert {\tau }\vert }\mathbf {u}_j\psi _{j_{\tau }}\), the coercivity (5) and orthogonality (35) lead to
Consequently, we get with an inverse estimate and (33) that
The linear mapping \(\mathcal {E}: u\mapsto \widetilde{v}\) with \(\dim \text {ran}\; \mathcal {E} \le C_\mathrm{dim} (2+\eta )^d q^{-d}k^{d+1}\) has a matrix representation \(\mathbf {u}\mapsto \mathbf {B}\mathbf {u}\), where the rank of \(\mathbf {B}\) is bounded by \(C_\mathrm{dim} (2+\eta )^d q^{-d}k^{d+1}\). Therefore, we get that \(a(\mathcal {E}u,w) = \mathbf {u}^T\mathbf {B}^T\mathbf {A}|_{\tau \times \sigma }\mathbf {w}\). The definition \(\mathbf {S_{\mathcal {H}}}(\tau ,\sigma ) := \mathbf {B}^T\mathbf {A}|_{\tau \times \sigma }\) leads to a matrix \(\mathbf {S_{\mathcal {H}}}(\tau ,\sigma )\) of rank \(r \le C_\mathrm{dim} (2+\eta )^d q^{-d}k^{d+1}\) such that
and the estimate \(\frac{1}{\left\| \mathbf {A} \right\| _2}\lesssim h^{2-d}\) from [14, Theorem 2] finishes the proof.
We refer to the next subsection for the existence of the inverse \(\mathbf {S}(\tau ,\tau )^{-1}\) of the Schur complement \(\mathbf {S}(\tau ,\tau )\). We proceed to approximate it by blockwise rank-\(r\) matrices. With the representation of the Schur complement from (34), we get that for a given right-hand side \(f \in L^2(\Omega )\), solving \(\mathbf {S}(\tau ,\tau )\mathbf {u} = \mathbf {f}\) with \(\mathbf {f} \in \mathbb {R}^{\vert {\tau }\vert }\) defined by \(\mathbf f_i = \left\langle f,\psi _{i_{\tau }} \right\rangle \), is equivalent to solving \(a(\widetilde{u},w) = \left\langle f,w \right\rangle \) for all \(w \in S^{p,1}_0(\mathcal {T}_h,\Gamma _D)\) with \(\mathrm{* }{supp} w \subset \overline{\omega _{\tau }}\). Let \(\tau _1\times \sigma _1 \subset \tau \times \tau \) be an \(\eta \)-admissible subblock. For \(f \in L^2(\Omega )\) with \(\mathrm{* }{supp} f \subset B_{R_{\sigma _1}}\), we get the orthogonality
Therefore, we have \(\widetilde{u} \in \mathcal {H}_h(B_{R_{\tau _1}},\omega _{\tau })\) and our results from Sect. 3 can be applied to approximate \(\widetilde{u}\) on \(B_{R_{\tau _1}}\cap \omega _{\tau }\). As in Sect. 4, this approximation can be used to construct a rank-\(r\) factorization of the subblock \(\mathbf {S}(\tau ,\tau )^{-1}|_{\tau _1\times \sigma _1}\), which is stated in the following theorem.
Theorem 3
Let \(\tau \subset \mathcal {I}\) and \(\rho := \{i\in \mathcal {I} : i < \min (\tau )\}\) and \(\tau _1\times \sigma _1 \subset \tau \times \tau \) be \(\eta \)-admissible. Define the Schur complement \(\mathbf {S}(\tau ,\tau ) = \mathbf {A}|_{\tau \times \tau } - \mathbf {A}|_{\tau \times \rho } (\mathbf {A}|_{\rho \times \rho })^{-1}\mathbf {A}|_{\rho \times \tau }\). Then, there exist rank-\(r\) matrices \(\mathbf {X}_{\tau _1\sigma _1} \in \mathbb {R}^{\vert {\tau _1}\vert \times r}\), \(\mathbf {Y}_{\tau _1\sigma _1} \in \mathbb {R}^{\vert {\sigma _1}\vert \times r}\) such that
The constants \(C_\mathrm{apx},b>0\) depend only on the boundary value problem (3), \(\Omega \), \(d\), \(p\), and the \(\gamma \)-shape regularity of the quasiuniform triangulation \(\mathcal {T}_h\).
5.2 Existence of \(\mathcal {H}\)-\(LU\) decomposition
In this subsection, we will use the approximation of the Schur complement from the previous section to prove the existence of an (approximate) \(\mathcal {H}\)-\(LU\) decomposition. We start with a hierarchical relation of the Schur complements \(\mathbf {S}(\tau ,\tau )\).
The Schur complements \(\mathbf {S}(\tau ,\tau )\) for a block \(\tau \in \mathbb {T}_{\mathcal {I}}\) can be derived from the Schur complements of its sons by
where \(\tau _1,\tau _2\) are the sons of \(\tau \). A proof of this relation can be found in [3, Lemma 3.1]. One should note that the proof does not use any properties of the matrix \(\mathbf {A}\) other than invertibility and existence of an \(LU\)-decomposition. Moreover, we have by definition of \(\mathbf {S}(\tau ,\tau )\) that \(\mathbf {S}(\mathcal {I},\mathcal {I}) = \mathbf {A}\).
If \(\tau \) is a leaf, we get the \(LU\)-decomposition of \(\mathbf {S}(\tau ,\tau )\) by the classical \(LU\)-decomposition, which exists since \(\mathbf {A}\) has an \(LU\)-decomposition. If \(\tau \) is not a leaf, we use the hierarchical relation of the Schur complements to define an \(LU\)-decomposition of the Schur complement \(\mathbf {S}(\tau ,\tau )\) by
with \(\mathbf {S}(\tau _1,\tau _1) = \mathbf {L}(\tau _1)\mathbf {U}(\tau _1)\), \(\mathbf {S}(\tau _2,\tau _2) = \mathbf {L}(\tau _2)\mathbf {U}(\tau _2)\) and indeed get \(\mathbf {S}(\tau ,\tau ) = \mathbf {L}(\tau ) \mathbf {U}(\tau )\). Moreover, the uniqueness of the \(LU\)-decomposition of \(\mathbf {A}\) implies that due to \(\mathbf {L}\mathbf {U} = \mathbf {A} = \mathbf {S}(\mathcal {I},\mathcal {I}) = \mathbf {L}(\mathcal {I})\mathbf {U}(\mathcal {I})\), we have \(\mathbf {L} = \mathbf {L}(\mathcal {I})\) and \(\mathbf {U} = \mathbf {U}(\mathcal {I})\).
The existence of the inverses \(\mathbf {L}(\tau _1)^{-1}\) and \(\mathbf {U}(\tau _1)^{-1}\) follows by induction over the levels, since on a leaf the existence is clear and the matrices \(\mathbf {L}(\tau )\), \(\mathbf {U}(\tau )\) are block triangular matrices. Consequently, the inverse of \(\mathbf {S}(\tau ,\tau )\) exists.
Moreover, the restriction of the lower triangular part \(\mathbf {S}(\tau _2,\tau _1)\mathbf {U}(\tau _1)^{-1}\) of the matrix \(\mathbf {L}(\tau )\) to a subblock \(\tau _2'\times \tau _1'\) with \(\tau _i'\) a son of \(\tau _i\) satisfies
and the upper triangular part of \(\mathbf {U}(\tau )\) satisfies a similar relation.
The following Lemma shows that the spectral norm of the inverses \(\mathbf {L}(\tau )^{-1}\), \(\mathbf {U}(\tau )^{-1}\) can be bounded by the norm of the inverses \(\mathbf {L}(\mathcal {I})^{-1}\), \(\mathbf {U}(\mathcal {I})^{-1}\).
Lemma 7
For \(\tau \in \mathbb {T}_{\mathcal {I}}\), let \(\mathbf {L}(\tau )\), \(\mathbf {U}(\tau )\) be given by (37). Then,
Proof
We only show the result for \(\mathbf {L}(\tau )\). With the block structure of (37) we get the inverse
So, we get by choosing \(\mathbf {x}\) such that \(\mathbf {x}_i = 0\) for \(i \in \tau _1\) that
The same argument for \((\mathbf {L}(\tau )^{-1})^T\) leads to
Thus, we have \(\left\| \mathbf {L}(\tau )^{-1} \right\| _2 \ge \max _{i=1,2}\left\| \mathbf {L}(\tau _1)^{-1} \right\| _2\) and as a consequence \(\max _{\tau \in \mathbb {T}_{\mathcal {I}}}\left\| \mathbf {L}(\tau )^{-1} \right\| _2 = \left\| \mathbf {L}(\mathcal {I})^{-1} \right\| _2\).
We can now formulate the existence result for an \(\mathcal {H}\)-\(LU\) decomposition.
Theorem 4
Let \(\mathbf {A} = \mathbf {L}\mathbf {U}\) with \(\mathbf {L},\mathbf {U}\) being lower and upper triangular matrices. There exist lower and upper triangular blockwise rank-\(r\) matrices \(\mathbf {L_{\mathcal {H}}},\mathbf {U_{\mathcal {H}}}\) such that
where \(C_\mathrm{LU} = C_\mathrm{sp}C_\mathrm{apx}(\kappa _2(\mathbf {U})+\kappa _2(\mathbf {L}))\), with the constant \(C_\mathrm{apx}\) from Theorem 1 and the spectral condition numbers \(\kappa _2(\mathbf {U})\), \(\kappa _2(\mathbf {L})\).
Proof
With Lemma 6, we get a low rank approximation of an admissible subblock \(\tau '\times \sigma '\) of the lower triangular part of \(\mathbf {L}(\tau )\) by
Since \(\mathbf {S}_{\mathcal {H}}(\tau ',\sigma ')\mathbf {U}(\sigma ')^{-1}\) is a rank-\(r\) matrix, Lemma 5 immediately provides an \(\mathcal {H}\)-matrix approximation \(\mathbf {L}_{\mathcal {H}}\) of the \(LU\)-factor \(\mathbf {L}(\mathcal {I}) = \mathbf {L}\). Therefore, with Lemma 7 we get
and in the same way an \(\mathcal {H}\)-matrix approximation \(\mathbf {U}_{\mathcal {H}}\) of \(\mathbf {U}(\mathcal {I}) = \mathbf {U}\) with
Since \(\mathbf {A}=\mathbf {L}\mathbf {U}\), the triangle inequality finally leads to
and the estimate \(\left\| \mathbf {A} \right\| _2 \le \left\| \mathbf {L} \right\| _2\left\| \mathbf {U} \right\| _2\) finishes the proof.
In the symmetric case, we may use the weaker admissibility condition (12) instead of (8) and obtain a result analogously to that of Theorem 4 for the Cholesky decomposition.
Corollary 2
Let \(\mathbf {b} = \mathbf {0}\) in (1) so that the resulting Galerkin matrix \(\mathbf {A}\) is symmetric and positive definite. Let \(\mathbf {A} = \mathbf {C}\mathbf {C}^T\) with \(\mathbf {C}\) being a lower triangular matrix with positive diagonal entries \(\mathbf {C}_{jj} > 0\). There exists a lower triangular blockwise rank-\(r\) matrix \(\mathbf {C_{\mathcal {H}}}\) such that
where \(C_\mathrm{Ch} = 2C_\mathrm{sp}C_\mathrm{apx}\sqrt{\kappa _2(\mathbf {A})}\), with the constant \(C_\mathrm{apx}\) from Theorem 1 and the spectral condition number \(\kappa _2(\mathbf {A})\).
Proof
Since \(\mathbf {A}\) is symmetric and positive definite, the Schur complements \(\mathbf {S}(\tau ,\tau )\) are symmetric and positive definite as well and therefore we get \(\mathbf {D}(\tau )\mathbf {L}(\tau ) = \mathbf {C}(\tau )\) in (37), where \(\mathbf {D}(\tau )\) is a diagonal matrix, such that \(\mathbf {D}(\tau )^2\) contains the diagonal elements of \(\mathbf {U}(\tau )\). Moreover, we have \(\left\| \mathbf {A} \right\| _2=\left\| \mathbf {C} \right\| _2^2\) and \(\kappa _2(\mathbf {C}) = \left\| \mathbf {C}^{-1} \right\| _2\left\| \mathbf {C} \right\| _2 = \sqrt{\kappa _2(\mathbf {A})}\).
6 Numerical examples
In this section, we present numerical examples in two and three dimensions to confirm our theoretical estimates derived in the previous sections. Since numerical examples for the Dirichlet case have been studied before, e.g., in [4, 20], we will focus on mixed Dirichlet–Neumann and pure Neumann problems in two and three dimensions.
With the choice \(\eta =2\) for the admissibility parameter in (8), the clustering is done by the standard geometric clustering algorithm, i.e., by splitting bounding boxes in half until they are admissible or smaller than the constant \(n_{\text {leaf}}\), which we choose as \(n_{\text {leaf}} = 25\) for our computations. An approximation to the inverse Galerkin matrix is computed by using the bestapproximation via singular value decomposition. Throughout, we use the C-library HLiB [8] developed at the Max-Planck-Institute for Mathematics in the Sciences.
6.1 2D-diffusion
We consider the unit square \(\Omega = (0,1)^2\). The boundary \(\Gamma = \partial \Omega \) is divided into the Neumann part \(\Gamma _{D} := \{\mathbf {x} \in \Gamma \,:\, \mathbf {x}_1 = 0 \vee \mathbf {x}_2 = 0 \}\) and the Dirichlet part \(\Gamma _{\mathcal {N}} = \Gamma \backslash \overline{\Gamma _{D}}\). We consider the bilinear form \(a(\cdot ,\cdot ):H^1_0(\Omega ,\Gamma _D)\times H^1_0(\Omega ,\Gamma _D) \rightarrow \mathbb {R}\) corresponding to the mixed Dirichlet–Neumann Poisson problem
and use a lowest order Galerkin discretization in \(S^{1,1}_0(\mathcal {T}_h,\Gamma _D)\).
As a second example, we study pure Neumann boundary conditions, i.e., \(\Gamma =\Gamma _{\mathcal {N}}\), and use the bilinear form \(a_{\mathcal {N}}(\cdot ,\cdot ): H^1(\Omega ) \times H^1(\Omega )\rightarrow \mathbb {R}\) corresponding to the stabilized Neumann Poisson problem
and a lowest order Galerkin discretization in \(S^{1,1}(\mathcal {T}_h)\).
In Fig. 1, we compare the decrease of the upper bound \(\left\| \mathbf {I}-\mathbf {A}\mathbf {B}_{\mathcal {H}} \right\| _2\) of the relative error with the increase in the block-rank for a fixed number \(N = 262,144\) of degrees of freedom, where the largest block of \(\mathbf {B}_{\mathcal {H}}\) has a size of 32,768.
We observe exponential convergence in the block rank, where the convergence rate is \(\exp (-br)\), which is even faster than the rate of \(\exp (-br^{1/3})\) guaranteed by Theorem 2.
6.2 3D-diffusion
We consider the unit cube \(\Omega = (0,1)^3\) with the Dirichlet boundary \(\Gamma _{D} := \{\mathbf {x} \in \Gamma \,:\, \exists i\in \{1,2,3\} : \mathbf {x}_i = 0\}\) and the Neumann part \(\Gamma _{\mathcal {N}} = \Gamma \backslash \overline{\Gamma _{D}}\).
Again, we consider the bilinear forms (40) and (41) corresponding to the weak formulations of the Dirichlet–Neumann Poisson problem and the stabilized Neumann problem.
In Fig. 2, we compare the decrease of \(\left\| \mathbf {I}-\mathbf {A}\mathbf {B}_{\mathcal {H}} \right\| _2\) with the increase in the block-rank for a fixed number \(N = 32,768\) of degrees of freedom, where the largest block of \(\mathbf {B}_{\mathcal {H}}\) has a size of 4,096.
Comparing the results with our theoretical bound from Theorem 2, we empirically observe a rate of \(e^{-br^{1/2}}\) instead of \(e^{-br^{1/4}}\). Moreover, whether we study mixed boundary conditions or pure Neumann boundary conditions does not make any difference, as both model problems lead to similar computational results.
6.3 Convection-diffusion
Finally, we study a convection-diffusion problem on the L-shaped domain \(\Omega = (0,1)\times (0,\frac{1}{2}) \cup (0,\frac{1}{2})\times [\frac{1}{2},1)\). The boundary \(\Gamma = \partial \Omega \) is divided into the Neumann part \(\Gamma _{\mathcal {N}} := \{\mathbf {x} \in \Gamma \,:\, \mathbf {x}_2 = 0 \; \vee \; \mathbf {x}_1 = 1 \}\) and the Dirichlet part \(\Gamma _D = \Gamma \backslash \overline{\Gamma _{\mathcal {N}}}\).
We consider the bilinear form \(a(\cdot ,\cdot ):H^1_0(\Omega ,\Gamma _D)\times H^1_0(\Omega ,\Gamma _D) \rightarrow \mathbb {R}\) corresponding to the mixed Dirichlet–Neumann Poisson problem
with \(c = 10^{-2}\) and \(\mathbf {b}(x_1,x_2) = (-x_2,x_1)^T\) and use a lowest order Galerkin discretization in \(S^{1,1}_0(\mathcal {T}_h,\Gamma _D)\).
In Fig. 3, we observe exponential convergence of the upper bound \(\left\| \mathbf {I}-\mathbf {A}\mathbf {B}_{\mathcal {H}} \right\| _2\) of the relative error with respect to the increase in the block-rank for a fixed number \(N = 196,352\) of degrees of freedom, where the largest block of \(\mathbf {B}_{\mathcal {H}}\) has a size of 24,544.
References
Adams, R.A.: Sobolev spaces, Pure and Applied Mathematics, vol. 65. Academic Press, New York (1975)
Bebendorf, M.: Hierarchical LU decomposition-based preconditioners for BEM. Computing 74(3), 225–247 (2005)
Bebendorf, M.: Why finite element discretizations can be factored by triangular hierarchical matrices. SIAM J. Numer. Anal. 45(4), 1472–1494 (2007)
Bebendorf, M., Hackbusch, W.: Existence of \(\cal {H}\)-matrix approximants to the inverse FE-matrix of elliptic operators with \(L^{\infty }\)-coefficients. Numer. Math. 95(1), 1–28 (2003)
Bennighof, J.K., Lehoucq, R.B.: An automated multilevel substructuring method for eigenspace computation in linear elastodynamics. SIAM J. Sci. Comput. 25(6), 2084–2106 (electronic) (2004). doi:10.1137/S1064827502400650
Börm, S.: Approximation of solution operators of elliptic partial differential equations by \(\cal {H}\)- and \({\cal {H}}^2\)-matrices. Numer. Math. 115(2), 165–193 (2010)
Börm, S.: Efficient Numerical Methods for Non-local Operators, EMS Tracts in Mathematics, vol. 14. European Mathematical Society (EMS), Zürich (2010)
Börm, S., Grasedyck, L.: H-Lib—a library for \(\cal {H}\)-and \({\cal {H}}^2\)-matrices. http://www.hlib.org (1999)
Brenner, S., Scott, L.: The Mathematical Theory of Finite Element Methods, Texts in Applied Mathematics, vol. 15. Springer, New York (2002)
Brenner, S.C.: The condition number of the Schur complement in domain decomposition. Numer. Math. 83(2), 187–203 (1999)
Chandrasekaran, S., Dewilde, P., Gu, M., Somasunderam, N.: On the numerical rank of the off-diagonal blocks of Schur complements of discretized elliptic PDEs. SIAM J. Matrix Anal. Appl. 31(5), 2261–2290 (2010). doi:10.1137/090775932
Dahmen, W., Faermann, B., Graham, I.G., Hackbusch, W., Sauter, S.A.: Inverse inequalities on non-quasiuniform meshes and application to the mortar element method. Math. Comput. 73, 1107–1138 (2001)
Demkowicz, L., Kurtz, J., Pardo, D., Paszyński, M., Rachowicz, W., Zdunek, A.: Computing with \(hp\)-adaptive finite elements, vol. 2. Chapman & Hall/CRC Applied Mathematics and Nonlinear Science Series, Frontiers: three dimensional elliptic and Maxwell problems with applications. Chapman & Hall/CRC, Boca Raton (2008)
Ern, A., Guermond, J.L.: Evaluation of the condition number in linear systems arising in finite element approximations. M2AN Math. Model. Numer. Anal. 40(1), 29–48 (2006)
Faustmann, M.: \({\cal {H}}\)-matrix approximantion of inverses of FEM and BEM matrices. doctoral thesis, work in progress, Vienna (2015)
Faustmann, M., Melenk, J.M., Praetorius, D.: Existence of \({\cal {H}}\)-matrix approximation to the inverse of BEM matrices: the simple layer operator. Institute for Analysis and Scientific Computing, Vienna University of Technology, Wien, Tech. Rep. in preparation (2013)
Faustmann, M., Melenk, J.M., Praetorius, D.: Existence of \({\cal {H}}\)-matrix approximants to the inverses of BEM matrices: the hyper singular integral operator. Work in progress (2014)
Giebermann, K.: Multilevel approximation of boundary integral operators. Computing 67(3), 183–207 (2001). doi:10.1007/s006070170005
Gillman, A., Martinsson, P.: A direct solver with \(O(N)\) complexity for variable coefficient elliptic PDEs discretized via a high-order composite spectral collocation method. Tech. rep. (2013). arXiv:1302.5995 [math.NA]
Grasedyck, L.: Theorie und Anwendungen Hierarchischer Matrizen. Doctoral thesis, Kiel (2001) (in German)
Grasedyck, L.: Adaptive recompression of \({\cal H}\)-matrices for BEM. Computing 74(3), 205–223 (2005)
Grasedyck, L., Hackbusch, W.: Construction and arithmetics of \({\cal H}\)-matrices. Computing 70(4), 295–334 (2003)
Grasedyck, L., Hackbusch, W., Kriemann, R.: Performance of \({\cal H}\)-LU preconditioning for sparse matrices. Comput. Methods Appl. Math. 8(4), 336–349 (2008)
Grasedyck, L., Kriemann, R., Le Borne, S.: Parallel black box \({\cal H}\)-LU preconditioning for elliptic boundary value problems. Comput. Vis. Sci. 11(4–6), 273–291 (2008). doi:10.1007/s00791-008-0098-9
Grasedyck, L., Kriemann, R., Le Borne, S.: Domain decomposition based \({\cal H}\)-LU preconditioning. Numer. Math. 112(4), 565–600 (2009)
Greengard, L., Gueyffier, D., Martinsson, P.G., Rokhlin, V.: Fast direct solvers for integral equations in complex three-dimensional domains. Acta Numer. 18, 243–275 (2009). doi:10.1017/S0962492906410011
Hackbusch, W.: A sparse matrix arithmetic based on \({\cal H}\)-matrices. Introduction to \({\cal H}\)-matrices. Computing 62(2), 89–108 (1999)
Hackbusch, W.: Hierarchische Matrizen: Algorithmen und Analysis. Springer, Dordrecht (2009)
Hackbusch, W., Börm, S.: \({\cal H}^2\)-matrix approximation of integral operators by interpolation. Appl. Numer. Math. 43(1–2), 129–143 (2002). doi:10.1016/S0168-9274(02)00121-6 (19th Dundee Biennial Conference on Numerical Analysis)
Hackbusch, W., Khoromskij, B., Sauter, S.A.: On \({\cal H}^2\)-matrices. In: Lectures on applied mathematics (Munich, 1999), pp. 9–29. Springer, Berlin (2000)
Ho, K., Ying, L.: Hierarchical interpolative factorization for elliptic operators: differential equations. Tech. rep. (2013). arXiv:1307.2895 [math.NA]
Ho, K.L., Greengard, L.: A fast direct solver for structured linear systems by recursive skeletonization. SIAM J. Sci. Comput. 34(5), A2507–A2532 (2012). doi:10.1137/120866683
Horn, R.A., Johnson, C.R.: Matrix Analysis, 2nd edn. Cambridge University Press, Cambridge (2013)
Karniadakis, G., Sherwin, S.: Spectral/hp Element Methods for CFD. Oxford University Press, Oxford (1999)
Le Borne, S., Grasedyck, L.: \({\cal H}\)-matrix preconditioners in convection-dominated problems. SIAM J. Matrix Anal. Appl. 27(4), 1172–1183 (electronic) (2006). doi:10.1137/040615845
Li, S., Gu, M., Wu, C.J., Xia, J.: New efficient and robust HSS Cholesky factorization of SPD matrices. SIAM J. Matrix Anal. Appl. 33(3), 886–904 (2012). doi:10.1137/110851110
Lintner, M.: The eigenvalue problem for the 2D Laplacian in \({\cal H}\)- matrix arithmetic and application to the heat and wave equation. Computing 72(3–4), 293–323 (2004)
Martinsson, P.G.: A fast direct solver for a class of elliptic partial differential equations. J. Sci. Comput. 38(3), 316–330 (2009). doi:10.1007/s10915-008-9240-6
Schmitz, P.G., Ying, L.: A fast direct solver for elliptic problems on general meshes in 2D. J. Comput. Phys. 231(4), 1314–1338 (2012). doi:10.1016/j.jcp.2011.10.013
Schwab, C.: \(p\)- and \(hp\)-finite element methods. Theory and applications in solid and fluid mechanics. Numerical Mathematics and Scientific Computation. The Clarendon Press, Oxford University Press, New York (1998)
Scott, L.R., Zhang, S.: Finite element interpolation of nonsmooth functions satisfying boundary conditions. Math. Comput. 54(190), 483–493 (1990)
Xia, J.: Efficient structured multifrontal factorization for general large sparse matrices. SIAM J. Sci. Comput. 35(2), A832–A860 (2013). doi:10.1137/120867032
Xia, J., Chandrasekaran, S., Gu, M., Li, X.S.: Superfast multifrontal method for large structured linear systems of equations. SIAM J. Matrix Anal. Appl. 31(3), 1382–1411 (2009). doi:10.1137/09074543X
Author information
Authors and Affiliations
Corresponding author
Rights and permissions
About this article
Cite this article
Faustmann, M., Melenk, J.M. & Praetorius, D. \(\mathcal {H}\)-matrix approximability of the inverses of FEM matrices. Numer. Math. 131, 615–642 (2015). https://doi.org/10.1007/s00211-015-0706-9
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00211-015-0706-9