betaFeedback

WorksMathematics

A robust and adaptive GenEO-type domain decomposition preconditioner for H(curl)\mathbf{H}(\mathbf{curl}) problems in three-dimensional general topologies

Abstract

In this paper we design, analyse and test domain decomposition methods for linear systems of equations arising from conforming finite element discretisations of positive Maxwell-type equations, namely for H(curl)\mathbf{H}(\mathbf{curl}) problems. It is well known that convergence of domain decomposition methods rely heavily on the efficiency of the coarse space used in the second level. We design adaptive coarse spaces that complement a near-kernel space made from the gradient of scalar functions. The new class of preconditioner is inspired by the idea of subspace decomposition, but based on spectral coarse spaces, and is specially designed for curl-conforming discretisations of Maxwell's equations in heterogeneous media on general domains which may have holes. We also address the practical robustness of various solvers in the case of non-trivial topologies and/or high aspect ratio of the domain.

Motivation

In this work we focus on the efficient solution of the following Maxwell problem:

∇×(μ−1∇×E)+γεE=fin Ω,E×n=0on ∂Ω.(1)\nabla\times\left(\mu^{-1}\nabla\times E\right) + \gamma\varepsilon E = f \quad\text{in } \Omega, \qquad E \times n = 0 \quad\text{on } \partial\Omega. \tag*{(1)}

Here EE is the vector-valued electric field, ff is a source term, while coefficients μ\mu and ε\varepsilon are electromagnetic parameters which are uniformly bounded and strictly positive but which we allow to be heterogeneous. Further, γ>0\gamma> 0 is a positive parameter which is allowed to be very small. Let Ω\Omega denote the computational domain (possibly non-convex and not simply connected) and nn denote the outward normal to ∂Ω\partial\Omega. For ease of presentation we consider the Dirichlet case in (1) but analogous results apply for other boundary conditions, as examples in our numerical results will show. Upon appropriate discretisation we must solve a linear system AE=fAE = f. This problem, although positive definite, remains challenging from the solution methods point of view. To date, the reference approach for tackling this problem is the celebrated algorithm of Hiptmair–Xu from [27], which was identified by the U.S. Department of Energy in a 2008 report as one of the top ten recent breakthroughs in computational science.1 This preconditioner is based upon a splitting of the space (giving a Helmholtz decomposition), here H0(curl⁡,Ω)H_0(\operatorname{curl},\Omega) for Ω\Omega convex, by isolating the kernel of the curl

Funding: The first two authors gratefully acknowledge support from the EPSRC grant EP/S004017/1. operator

H0(curl⁡,Ω)=(H01(Ω))3+grad⁡H01(Ω).\mathbf{H}_0(\operatorname{curl},\Omega) = (H_0^1(\Omega))^3 + \operatorname{grad} H_0^1(\Omega).

The auxiliary space H01(Ω)H_0^1(\Omega) then uses a nodal (i.e., Lagrangian) discretisation. The matrix version of the preconditioner is given as

MAMS−1=diag⁡(A)−1+P(L~+γQ~)−1PT+γ−1CL−1CT,\mathcal{M}_{\mathrm{AMS}}^{-1} = \operatorname{diag}(A)^{-1} + P(\widetilde{L}+\gamma\widetilde{Q})^{-1}P^{T} + \gamma^{-1}CL^{-1}C^{T},

where L~\widetilde{L} is the nodal discretisation of the μ−1\mu^{-1}-weighted vector Laplacian operator, Q~\widetilde{Q} is the nodal ε\varepsilon-weighted vector mass matrix, PP is the matrix form of the nodal interpolation operator between the Nédélec space and nodal element space, LL is the nodal ε\varepsilon-weighted scalar Laplacian, and CC is the “gradient matrix”, which is exactly the null space matrix of the discretisation of the curl–curl operator. This approach is known by several different names, the Hiptmair–Xu (HX) preconditioner, the (nodal) auxiliary space preconditioner (ASP), and the auxiliary-space Maxwell solver (AMS); here we will use the abbreviation AMS, as in [35] where a parallel implementation is considered and tested along with similar variants. The robustness of the AMS preconditioner with respect to the mesh size, heterogeneities in the coefficients and the topology of the domain has been thoroughly studied from both theoretical and practical points of view; see, e.g., [26, 27, 25, 55, 56, 35, 24, 1, 28, 29]. Nonetheless, to the best of our knowledge there is no theoretical guarantee that the AMS preconditioner is efficient for elongated domains with holes, such as the one exhibited in Figure 1 (see § 4). Our numerical results in § 4 show that, indeed, the iteration counts of AMS increase in such cases; see Table 2 and Table 6.

The beam test configuration with holes for $N = 8$ subdomains, showing views from the side and front of the beam

Figure 1. The beam test configuration with holes for N=8N = 8 subdomains, showing views from the side (left) and front (right) of the beam. The subdomain partitioning is also shown on the left with subdomains indicated by the different colours.

NN8163264128256
#dofs113K226K451K901K1800K3600K
NK size18K36K72K144K288K576K
SNK size24K49K99K198K397K794K
GenEO size (Dirichlet)000000
GenEO size (mixed)184290186378762
Dirichlet — AMS4471169364660813
Dirichlet — AS243659797258
Dirichlet — AS-SNK141515151515
Dirichlet — AS-SNK-GenEO141515151515
Dirichlet — AS-NK171818181717
Dirichlet — AS-NK-GenEO171818181717
mixed — AMS43671133215881302
mixed — AS3761106173294557
mixed — AS-SNK3662109202383554
mixed — AS-SNK-GenEO232425262727
mixed — AS-NK3662111203386588
mixed — AS-NK-GenEO242525283037

Table 2. A weak scalability study detailing iteration counts required for the homogeneous beam problem with holes (non-trivial topology) and Dirichlet (top) or mixed (bottom) boundary conditions.

γ\gamma10−510^{-5}10−410^{-4}10−310^{-3}10−210^{-2}10−110^{-1}11101010210^{2}
GenEO size (Dirichlet)00000000
GenEO size (mixed)76276276276276276200
DirichletAMS4945698133429838168
DirichletAS545458844623137
DirichletAS-SNK151515151514129
DirichletAS-SNK-GenEO151515151514129
DirichletAS-NK171717171614129
DirichletAS-NK-GenEO171717171614129
mixedAMS2057219313023038633146
mixedAS61857955746315046157
mixedAS-SNK77276155444414445179
mixedAS-SNK-GenEO282727262320179
mixedAS-NK776765588444144451710
mixedAS-NK-GenEO5149372524201710

Table 6. Iteration counts required when varying the parameter γ\gamma for the homogeneous beam problem with holes (non-trivial topology), Dirichlet (top) or mixed (bottom) boundary conditions, and N=256N = 256 subdomains. Here we have #dofs=3600K\#\mathrm{dofs} = 3600\mathrm{K}, NK size = 576K, and SNK size = 794K.

A natural question is then whether we can propose a preconditioner that would remedy this issue. We choose here to consider this question from the point of view of domain decomposition methods, especially in order to address non-trivial topologies. Overlapping Schwarz methods were considered in [53, 48], for which a best condition number bound was very recently proved in [37]. Non-overlapping and substructuring methods are considered in [31, 32, 54, 15], with some analysis in the case of jumps in the physical parameters provided in [30], although under somewhat restrictive assumptions. Further, a non-overlapping BDDC algorithm is constructed in [16] and gives improved bounds in relation to the ratio between coarse and fine meshes but still with dependence on the physical parameters akin to [54]; see also results in [57]. Within overlapping domain decomposition methods, in order to obtain a robust approach a second level is typically required. A good way to provide appropriate information is via spectral coarse spaces, exemplified by the GenEO approach [52, 17], which use local generalised eigenvalue problems (GEVPs) and can provide robustness for highly heterogeneous problems. The approach has also seen theoretical extensions to indefinite and nonself-adjoint problem [7, 8] as well as to saddle point systems [44]. It has also been adapted for wave propagation in the form of high-frequency heterogeneous Helmholtz problems [10, 6]; see also the more general overview for wave propagation problems in [9]. We additionally mention an alternative approach called MS-GFEM [2, 41, 40], which uses optimal local approximations spaces through use of local eigenproblems defined on generalised harmonic spaces and has also been applied to the high-frequency heterogeneous Helmholtz problem [39].

Given the success of spectral coarse spaces, such an approach is natural to consider here in the case of the positive Maxwell equation. Note that, since the operator is symmetric positive definite, the GenEO theory developed in [52] applies directly, but in practice a difficulty arises in solving the local eigenproblems required to build the coarse space. Indeed, the purpose of the local GEVPs is to capture the local components of the near-kernel of the global problem. In our present case of the Maxwell problem (1) and for small γ\gamma, the gradients of H1H^1 functions contribute to the sought coarse space since their curl is zero. There are many such functions and an iterative eigenvalue solver has difficulties finding all of them. Thus, we a priori include the gradients of H1H^1 functions in the coarse space, but even this coarse space is not always sufficient. For instance, when the domain has a non-trivial topology (e.g., holes in the domain), which features naturally in various practical applications, the kernel of the curl operator is larger than the gradient of H1H^1 functions. It can be computed but requires some effort; see [49]. Moreover, even when the complete kernel of the curl operator is known, it may not be sufficient to build a good coarse space, such as when there are jumps in the coefficients.

With this in mind, our proposed solution is to start from some approximate knowledge of appropriate coarse space vectors, namely through the gradients of H1H^1 functions, and enrich this space by locally solving GEVPs in each subdomain. This ensures the size of the enrichment is small and so computable by the eigensolver. Our approach yields an adaptive coarse space in the GenEO spirit, but with modification required to account for the large near-kernel of the global problem. Overall, our main contributions in this work are:

  1. A general framework for the extension of the GenEO domain decomposition method to cases where the analytic near-kernel of the problem is large.

  2. An application of the framework to the special case of the Maxwell system.

  3. Extensive numerical tests and comparisons with the AMS preconditioner which show robustness with respect to:

    • the geometry and its topology;

    • how close the problem is to being singular due to a small zeroth-order term given by the value of γ\gamma.

Our method ensures robustness across a wide range of challenging scenarios in terms of iteration counts, as seen in the numerical results of § 4, and to our knowledge this is the first method to be provably robust with respect to topology of the domain and heterogeneities in the coefficients; this comes at a price of a coarse space which is large since it has to contain all gradients of scalar functions, plus a small contribution of local eigenvectors coming from a spectral GenEO coarse space. Nevertheless, the coarse space size is typically five times smaller than the original problem and the gradient part of the coarse matrix retains the sparsity pattern of the underlying finite element discretization. Further investigation is needed in order to provide an efficient way of solving the coarse problem.

Note that the preconditioner we propose can also be applied to time dependent Maxwell’s equations, with the parameter γ=(Δt)−1\gamma=(\Delta t)^{-1} in this case, where Δt\Delta t is the time discretisation step. Additionally, the positive Maxwell operator is used in [3] to build preconditioners in the context of time-periodic eddy current problems. Another class of related problems is the time-harmonic Maxwell system where the parameter γ=−ω2\gamma=-\omega^2 can be negative and ε\varepsilon complex valued. A different set of complications arise depending on whether we have high-frequency or low-frequency problems, as determined by the value of the angular frequency ω\omega. For this reason, the construction of fast iterative solvers for time-harmonic Maxwell systems is a problem of great current interest. Typically in the high-frequency case, the solution is very oscillatory and the size of the discretised problem increases with the frequency. The nature of the equations becomes very different from an elliptic problem and the method we propose is not suitable. The design of scalable coarse spaces that remain effective under strong indefiniteness is a very active field of research; see, e.g., [39, 9] and references therein. Also, in this case, a domain decomposition preconditioner can be built from the absorptive case, relying on an idea originating in [18] but see also [42]; this is often called “shifted Laplacian preconditioning”. In [20], the authors propose a two-level domain decomposition preconditioner for the Helmholtz equation based on adding absorption. These ideas are extended to Maxwell’s equation in [5] through a combination of techniques inspired from the indefinite case but also from the positive definite one.

In the low-frequency regime, i.e., when the angular frequency ω≪1\omega\ll1, the problem is close to being singular and to deal with this problem, one option utilised in [21, 36] is to consider a mixed form of these equations, which ensures stability and well-posedness, by introducing an artificial variable (Lagrange multiplier) pp to obtain the system

∇×(μ−1∇×E)−ω2εE+ε∇p=fin Ω,(2)\nabla\times\left(\mu^{-1}\nabla\times E\right)-\omega^{2}\varepsilon E+\varepsilon\nabla p=f \quad\text{in }\Omega, \tag*{(2)}
∇⋅(εE)=0in Ω,(3)\nabla\cdot(\varepsilon E)=0 \quad\text{in }\Omega, \tag*{(3)}
E×n=0on ∂Ω,(4)E\times n=0 \quad\text{on }\partial\Omega, \tag*{(4)}
p=0on ∂Ω.(5)p=0 \quad\text{on }\partial\Omega. \tag*{(5)}

If we discretise the system (1.2) using Nédélec elements for EE and Lagrange nodal elements for pp we obtain the following saddle point system

K(Ep):=(K−ω2MBT−B0)(Ep)=(f0),(6)\mathcal{K}\begin{pmatrix} E \\ p \end{pmatrix}:= \begin{pmatrix} K-\omega^{2}M & B^{T} \\ -B & 0 \end{pmatrix} \begin{pmatrix} E \\ p \end{pmatrix} = \begin{pmatrix} f \\ 0 \end{pmatrix}, \tag*{(6)}

where K∈Rn×nK\in\mathbb{R}^{n\times n} corresponds to the discretisation of ∇×(μ−1∇×)\nabla\times(\mu^{-1}\nabla\times) and has a large kernel of dimension mm, M∈Rn×nM\in\mathbb{R}^{n\times n} is the ε\varepsilon-weighted mass matrix computed on the edge-element space, and B∈Rm×nB\in\mathbb{R}^{m\times n} is a full row rank matrix corresponding to the divergence constraint. As a consequence, the stiffness matrix K\mathcal{K} is symmetric but indefinite.

A block preconditioner is proposed in [21], taking the form

PM,L:=(PM00L),PM:=K+γM,γ=1−ω2>0,(7)\mathcal{P}_{M,L}:=\begin{pmatrix} \mathcal{P}_{M} & 0 \\ 0 & L \end{pmatrix},\qquad\mathcal{P}_{M}:=K+\gamma M,\quad\gamma=1-\omega^{2}>0, \tag*{(7)}

where LL is the nodal ε\varepsilon-weighted Laplace operator; the challenge now being solving efficiently with PM\mathcal{P}_{M}. Here PM\mathcal{P}_{M} represents the discretisation of the positive Maxwell problem (1) and thus will be amenable to our methodology. We note that in the limit case ω=0\omega=0, the leading block in (6) is maximally rank deficient and the approach of [19] applies. This further inspired related work that extends to the case of block preconditioners for ω\omega small [38]. We also mention earlier work on mixed approximation of magnetostatic problems in [50]. More recent literature on solving general saddle point systems can be found in, e.g., [51, 47].

Since a key part of many Maxwell-type solvers is the solution of a positive Maxwell problem (1), as in (7) and, for example, [13, 14], in what follows we will focus on the efficient solution to the discretisation of this system, given by

AU:=(K+γM)U=f,(8)A\mathbf{U}:=(K+\gamma M)\mathbf{U}=f, \tag*{(8)}

where we now use U\mathbf{U} as our unknown. We reiterate that the matrix KK has a large kernel (since all gradients of H1H^{1} functions are part of the kernel of the curl operator) and so designing efficient preconditioners for this problem can be challenging when γ\gamma is small, especially in a general case with a non-trivial topology.

The remainder of this work is laid out as follows. In § 2 we introduce notation and preliminaries required for our theory. We then prove condition number bounds for our newly developed two-level additive Schwarz preconditioner in § 3. A selection of illustrative numerical results are then presented in § 4. These detail examples with heterogeneity and domains with holes and show the robustness of our approach in all cases, even for elongated domains with holes for which the iteration counts of AMS increase with the difficulty of the test case. Finally, we summarise our conclusions in § 5.

Notation and preliminaries

In the work that follows we will use the idea of subspace decomposition from the auxiliary space preconditioner to build a new domain decomposition preconditioner. The latter will contain a second level with two types of components: one arising from the part of the kernel of matrix KK that is readily available (i.e., related to the gradient of H1H^1 functions in our case), and the other stemming from local spectral information, which will help tackle the heterogeneous nature of the problem and the potential for holes in the domain.

Both cases can be treated from a theoretical point of view in the general framework of the fictitious space lemma (FSL) of Nepomnyaschikh [45], which can be considered the Lax–Milgram theorem of domain decomposition. For the sake of completeness, we briefly introduce the problem formulation and notation used (which is similar to that in [43]), along with some theoretical results that will be useful in § 3. We then detail the FSL, presented in the form originating in Griebel and Oswald [22], before applying it within our context.

We suppose that our problem is posed on an arbitrary finite domain Ω⊂R3\Omega\subset\mathbb{R}^{3} and is given in weak form. Namely, for an appropriate Hilbert space HH, of functions on Ω\Omega, we wish to find u∈Hu \in H such that

a(u,v)=l(v),for all v∈H,a(u,v)=l(v), \qquad\text{for all } v \in H,

where a(⋅,⋅)a(\cdot,\cdot) is a symmetric coercive bilinear form on HH and l(⋅)l(\cdot) is a linear functional on HH. In our case of the positive Maxwell problem (1), highlighting the functions as vectorial by using bold, our bilinear form is

a(u,v)=∫Ωμ−1(∇×u)⋅(∇×v)+γεu⋅v dx.(9)a(\mathbf{u},\mathbf{v})=\int_{\Omega}\mu^{-1}(\nabla\times\mathbf{u})\cdot(\nabla\times\mathbf{v})+\gamma\varepsilon\mathbf{u}\cdot\mathbf{v}\,\mathrm{d}x. \tag*{(9)}

while the linear functional is

l(v)=∫Ωf⋅v dx.l(\mathbf{v})=\int_{\Omega}\mathbf{f}\cdot\mathbf{v}\,\mathrm{d}x.

We assume the problem is discretised on a mesh Th\mathcal{T}_h by an appropriate finite element method, say using Nédélec elements, to obtain the linear system (15), with symmetric positive definite system matrix AA.

Consider an overlapping domain decomposition Ω=⋃1≤i≤NΩi\Omega=\bigcup_{1\leq i\leq N}\Omega_i where the domain Ω\Omega is decomposed into NN subdomains, which we assume to conform to the mesh Th\mathcal{T}_h. We will later need the use of local Neumann matrices on each subdomain Ωi\Omega_i. These are given by restricting the bilinear form (9) to be only over Ωi\Omega_i without enforcing any new boundary conditions on the interior of Ω\Omega, that is, we define AiNeuA_i^{\mathrm{Neu}} to be the equivalent discretisation of

aΩi(u,v)=∫Ωiμ−1(∇×u)⋅(∇×v)+γεu⋅v dx,a_{\Omega_i}(\mathbf{u},\mathbf{v})=\int_{\Omega_i}\mu^{-1}(\nabla\times\mathbf{u})\cdot(\nabla\times\mathbf{v})+\gamma\varepsilon\mathbf{u}\cdot\mathbf{v}\,\mathrm{d}x,

where the boundary conditions from the underlying problem, i.e., from (1), are imposed on ∂Ωi∩∂Ω\partial\Omega_i \cap\partial\Omega. Note that, since γ>0\gamma> 0, the Neumann matrices are SPD; this will ensure a necessary eigenproblem (to be defined in GEVP 3.4) is always well-posed.

Let the degrees of freedom (dofs) for the discretised problem (8) be ordered into a set N\mathcal{N} of size nn. For each subdomain Ωi\Omega_i we suppose the local dofs contained within the subdomain are given by Ni⊂N\mathcal{N}_i \subset\mathcal{N}, with size nin_i. We define the restriction operators from the set N\mathcal{N} to Ni\mathcal{N}_i, respectively for each ii, as Ri:Rn→RniR_i:\mathbb{R}^{n}\to\mathbb{R}^{n_i}. The corresponding matrices RiTR_i^T, which constitute extension by zero outside of Ni\mathcal{N}_i, are called prolongation (or extension) operators. We also require a partition of unity to glue together solutions on each subdomain Ωi\Omega_i. At the algebraic level this is given by non-negative diagonal matrices Di∈Rni×niD_i \in\mathbb{R}^{n_i \times n_i} such that

∑i=1NRiTDiRi=I,(10)\sum_{i=1}^{N} R_i^T D_i R_i = I, \tag*{(10)}

where I∈Rn×nI \in\mathbb{R}^{n \times n} is the identity. We further introduce several constants related to the domain decomposition which will be useful later on. Firstly, we define the maximum multiplicity of the interaction between subdomains plus one as

k0:=max⁡1≤i≤N#{j∣RjARiT≠0},(11)k_0 := \max_{1 \le i \le N} \#\{j \mid R_j A R_i^T \ne0\}, \tag*{(11)}

where #\# gives the size of a set. Note that this is bounded from below by the number of colours required to colour a decomposition so that no two neighbouring subdomains share the same colour and is typically a small constant for standard decomposition techniques. Secondly, we define the maximum multiplicity of subdomain intersection, k1k_1, as the largest number of subdomains whose intersection has non-zero measure. We now present two auxiliary lemmas from [17], related to these constants, that will be used later in our theory in § 3.

Lemma 2.1 (Lemma 7.9 in [17]). Suppose that we have a set of rectangular matrices Qi∈Rn×niQ_i \in\mathbb{R}^{n \times n_i}, for 1≤i≤N1 \le i \le N, and that A∈Rn×nA \in\mathbb{R}^{n \times n} is symmetric positive definite. Define the constant

k~0:=max⁡1≤i≤N#{j∣QjAQiT≠0}.\widetilde{k}_0 := \max_{1 \le i \le N} \#\{j \mid Q_j A Q_i^T \ne0\}.

Consider the vectors Ui∈Rni\mathbf{U}_i \in\mathbb{R}^{n_i} for 1≤i≤N1 \le i \le N, then we have the following estimate

(∑i=1NQiUi)TA(∑i=1NQiUi)≤k~0∑i=1NUiT(QiTAQi)Ui.\left(\sum_{i=1}^{N} Q_i\mathbf{U}_i\right)^T A \left(\sum_{i=1}^{N} Q_i\mathbf{U}_i\right) \le\widetilde{k}_0 \sum_{i=1}^{N} \mathbf{U}_i^T(Q_i^T A Q_i)\mathbf{U}_i.

Lemma 2.2 (Lemma 7.13 in [17]). Suppose that the matrix A∈Rn×nA \in\mathbb{R}^{n \times n} stems from the symmetric coercive bilinear form a(⋅,⋅)a(\cdot,\cdot) on a domain Ω\Omega discretised on a mesh Th\mathcal{T}_h. Further, assume a domain decomposition as above, which yield equivalent local Neumann matrices AiNeuA_i^{\mathrm{Neu}}, 1≤i≤N1 \le i \le N, and let k1k_1 be defined, as previously, as the maximum multiplicity of subdomain intersection. Consider a vector U∈Rn\mathbf{U} \in\mathbb{R}^{n}, then we have the bound

∑i=1N(RiU)TAiNeu(RiU)≤k1UTAU.\sum_{i=1}^{N}(R_i\mathbf{U})^T A_i^{\mathrm{Neu}}(R_i\mathbf{U}) \le k_1\mathbf{U}^T A\mathbf{U}.

Our theory will build on the additive Schwarz domain decomposition approach. When used as a preconditioner it takes the form

MAS−1=∑i=1NRiTAi−1Ri,(12)M_{\mathrm{AS}}^{-1} = \sum_{i=1}^{N} R_i^T A_i^{-1} R_i, \tag*{(12)}

where Ai=RiARiTA_i = R_i A R_i^T is the (local) Dirichlet matrix on Ωi\Omega_i, namely where Dirichlet conditions are implicitly assumed one extra layer of mesh elements away from ∂Ωi\∂Ω\partial\Omega_i\backslash\partial\Omega. Since such a one-level approach is typically not scalable (though see [11]) we must enhance convergence through use of a second-level correction, i.e., a coarse space. This derives from a collection of column vectors ZZ, which should form a basis for a space V0V_0, called the coarse space. This further yields a coarse space operator E=ZTAZE = Z^T A Z along with the coarse correction operator ZE−1ZTZE^{-1}Z^T. The second level can be incorporated in a number of ways; we consider one that maintains the symmetry of (12) given by

MAS,2−1=ZE−1ZT+(I−P0)MAS−1(I−P0T),(13)M_{\mathrm{AS},2}^{-1} = ZE^{-1}Z^T + (I-P_0)M_{\mathrm{AS}}^{-1}(I-P_0^T), \tag*{(13)}

where P0=ZE−1ZTAP_0 = ZE^{-1}Z^T A is the AA-orthogonal projection on the coarse space V0V_0. Note that vectors in the coarse space are typically denoted with a subscript zero, which relates to the coarse space as opposed to a subdomain Ωi\Omega_i, i≥1i \ge1. We suppose that the size of the coarse space is n0n_0 and a generic vector in V0V_0 is written as U0∈Rn0U_0 \in\mathbb{R}^{n_0} in a similar way that Ui∈RniU_i \in\mathbb{R}^{n_i} refers to a local vector on the subdomain Ωi\Omega_i, 1≤i≤N1 \le i \le N.

We now turn to the main theoretical result that we employ, the fictitious space lemma (FSL), before detailing how this fits within our current context.

Lemma 2.3 (Fictitious space lemma, as presented by Griebel and Oswald [22]). Consider two Hilbert spaces HH and HDH_D with the respective inner products (⋅,⋅)(\cdot,\cdot) and (⋅,⋅)D(\cdot,\cdot)_D. Further, suppose that we have symmetric coercive bilinear forms a:H×H→Ra:H \times H \to\mathbb{R} and b:HD×HD→Rb:H_D \times H_D \to\mathbb{R}. The operators AA and BB are defined as follows:

  • Let A:H→HA:H \to H be such that (Au,v)=a(u,v)(Au,v) = a(u,v) for all u,v∈Hu,v \in H;

  • Let B:HD→HDB:H_D \to H_D be such that (BuD,vD)D=b(uD,vD)(Bu_D,v_D)_D = b(u_D,v_D) for all uD,vD∈HDu_D,v_D \in H_D.

Suppose that we have a linear surjective operator R:HD→HR:H_D \to H verifying the following properties.

  • Continuity: ∃cR>0\exists c_R > 0 such that ∀uD∈HD\forall u_D \in H_D we have

a(RuD,RuD)≤cRb(uD,uD).a(Ru_D,Ru_D) \le c_R b(u_D,u_D).
  • Stable decomposition: ∃cT>0\exists c_T > 0 such that ∀u∈H ∃uD∈HD\forall u \in H\ \exists u_D \in H_D with RuD=uRu_D = u and

cTb(uD,uD)≤a(RuD,RuD)=a(u,u).c_T b(u_D,u_D) \le a(Ru_D,Ru_D) = a(u,u).

Consider the adjoint operator R∗:H→HDR^*:H \to H_D given by (RuD,u)=(uD,R∗u)D(Ru_D,u) = (u_D,R^*u)_D for all uD∈HDu_D \in H_D and u∈Hu \in H. Then for all u∈Hu \in H we have the spectral estimate

cTa(u,u)≤a(RB−1R∗Au,u)≤cRa(u,u).c_T a(u,u) \le a(RB^{-1}R^*Au,u) \le c_R a(u,u).

Thus, the eigenvalues of the preconditioned operator RB−1R∗ARB^{-1}R^*A are bounded from below by cTc_T and from above by cRc_R and we have the spectral condition number estimate

κ2(RB−1R∗A)≤cT−1cR.\kappa_2(RB^{-1}R^*A) \le c_T^{-1}c_R.

Remark 2.4. In this lemma we have two Hilbert spaces (that are linked by the surjective operator R\mathcal{R}) and two symmetric coercive bilinear forms. The first one comes from our problem while the second is for the preconditioner. Under the assumptions of continuity and stable decomposition, the spectral estimate tells us that the spectral condition number of the preconditioned problem is bounded solely in terms of the constants cRc_R and cTc_T. The matrix A∈Rn×nA \in\mathbb{R}^{n \times n} is typically a symmetric positive definite matrix, stemming from the symmetric coercive bilinear form a(⋅,⋅)a(\cdot,\cdot) on Ω\Omega corresponding to the problem to solve. Our main assumption is that the matrix AA, defining the linear system we wish to solve, has a large near-kernel which can cause problems numerically. In other words, AA is a small perturbation from KK, a matrix with a large kernel. Thus, we would like to deal with the near-kernel directly in the coarse space in order that we can primarily work in the orthogonal complement, avoiding the ill-conditioning in AA from the near-kernel. The near-kernel in the case of the positive Maxwell problem (1) on a trivial topology domain Ω\Omega is given by the gradient of all H1(Ω)H^1(\Omega) functions, namely the kernel of the curl--curl operator.

Definition 2.5 (Split near-kernel coarse space). Let GG be the easily computable part of the near-kernel of AA; namely, in our case, vectors stemming from gradients of H1H^1 functions but not those related to any holes in the domain. We define the split near-kernel (SNK) coarse space VG⊂RnV_G \subset\mathbb{R}^n as the vector space spanned by the sequence (RiTDiGi)1≤i≤N(R_i^T D_i G_i)_{1\leq i\leq N}, where Gi:=RiGG_i := R_iG, so that G⊂VGG \subset V_G. Here we have split GG into contributions on subdomains and note that dim⁡VG>dim⁡G\dim V_G > \dim G in general; we will see that such a splitting is necessary in numerical results. The coarse space matrix Z∈Rn×n0Z \in\mathbb{R}^{n\times n_0} is a rectangular matrix whose columns are a basis of V0:=VGV_0 := V_G. The coarse space matrix is then defined in the usual way by E=ZTAZE = Z^T A Z.

Note that the space VGV_G is the equivalent of the Nicolaides coarse space [46] for the Laplace problem or the rigid body modes coarse space for the linear elasticity problem; see [17].

We now need to define all the other ingredients in the FSL (Lemma 2.3). Our (discrete) problem space is H=RnH = \mathbb{R}^n while the second Hilbert space HDH_D, for the preconditioning, is the product space of vectors stemming from the nin_i degrees of freedom on the local subdomains Ωi\Omega_i along with the n0n_0 coarse space vectors

HD:=Rn0×∏i=1NRni.(14)H_D := \mathbb{R}^{n_0} \times\prod_{i=1}^{N} \mathbb{R}^{n_i}. \tag*{(14)}

The bilinear form b(⋅,⋅)b(\cdot,\cdot) is chosen to be given by the sum of local bilinear forms bi(Ui,Vi):=(AiUi,Vi)b_i(U_i,V_i):=(A_iU_i,V_i) and a coarse space contribution

b(U,V):=(EU0,V0)+∑i=1Nbi(Ui,Vi),for all U,V∈HD,(15)b(U,V) := (EU_0,V_0) + \sum_{i=1}^{N} b_i(U_i,V_i), \qquad\text{for all } U,V \in H_D, \tag*{(15)}

where the product space vectors are U=(U0,(Ui)1≤i<N)U=(U_0,(U_i)_{1\leq i<N}) and V=(V0,(Vi)1≤i<N)V=(V_0,(V_i)_{1\leq i<N}). Finally, the surjective operator RAS:HD→H\mathcal{R}_{AS}:H_D \to H corresponding the additive Schwarz method is given by

RAS(U):=ZU0+(I−P0)∑i=1NRiTUi,(16)\mathcal{R}_{AS}(U) := ZU_0 + (I-P_0)\sum_{i=1}^{N} R_i^T U_i, \tag*{(16)}

where P0=ZE−1ZTAP_0 = ZE^{-1}Z^T A is the AA-orthogonal projection on the coarse space V0V_0.

Definition 2.6 (Orthogonal projections and orthogonal complements). Let the matrices Bi∈Rni×niB_i \in\mathbb{R}^{n_i \times n_i} correspond to the local symmetric coercive bilinear forms bi(⋅,⋅)b_i(\cdot,\cdot), for i≥1i \ge1. Within Rni\mathbb{R}^{n_i} we define the bib_i-orthogonal complement of GiG_i

Gi⊥Bi:={Ui∈Rni∣∀Vi∈Gi, bi(Ui,Vi)=0}.(17)G_i^{\perp_{B_i}} := \{U_i \in\mathbb{R}^{n_i} \mid\forall V_i \in G_i,\ b_i(U_i,V_i)=0\}. \tag*{(17)}

As such, let bGi⊥Bib_{G_i^{\perp_{B_i}}} denote the restriction of bib_i to Gi⊥BiG_i^{\perp_{B_i}}.

We have a representation formula bGi⊥Bib_{G_i^{\perp_{B_i}}}:

Lemma 2.7. The Riesz representation theorem gives the existence of a unique isomorphism BGi⊥Bi:Gi⊥Bi→Gi⊥BiB_{G_i^{\perp_{B_i}}}:G_i^{\perp_{B_i}}\to G_i^{\perp_{B_i}} into itself so that

(BGi⊥BiUi,Vi):=bGi⊥Bi(Ui,Vi), ∀Ui,Vi∈Gi⊥Bi.(B_{G_i^{\perp_{B_i}}}U_i,V_i):=b_{G_i^{\perp_{B_i}}}(U_i,V_i),\ \forall U_i,V_i\in G_i^{\perp_{B_i}}.

Further, we let ξ0i\xi_{0i} be the bib_i-orthogonal projection from Rni\mathbb{R}^{n_i} on GiG_i parallel to Gi⊥BiG_i^{\perp_{B_i}}. This also enables us to express the inverse of BGi⊥BiB_{G_i^{\perp_{B_i}}}, which we will denote by Bi†B_i^\dagger, through the following formula

Bi†=(I−ξ0i)Bi−1.(18)B_i^\dagger=(I-\xi_{0i})B_i^{-1}. \tag*{(18)}

Proof. In order to check formula (18), we have to show that

BGi⊥Bi(I−ξ0i)Bi−1y=y, ∀y∈Gi⊥Bi.(19)B_{G_i^{\perp_{B_i}}}(I-\xi_{0i})B_i^{-1}y=y,\ \forall y\in G_i^{\perp_{B_i}}. \tag*{(19)}

Let z∈Gi⊥Biz\in G_i^{\perp_{B_i}}, using the fact that I−ξ0iI-\xi_{0i} is the bib_i-orthogonal projection on Gi⊥BiG_i^{\perp_{B_i}}, we have

(BGi⊥Bi(I−ξ0i)Bi−1y,z)=bi((I−ξ0i)Bi−1y,z)=bi(Bi−1y,z)=(y,z).(B_{G_i^{\perp_{B_i}}}(I-\xi_{0i})B_i^{-1}y,z)=b_i((I-\xi_{0i})B_i^{-1}y,z)=b_i(B_i^{-1}y,z)=(y,z).

Since this equality holds for any z∈Gi⊥Biz\in G_i^{\perp_{B_i}}, this proves that (19) holds and thus that Bi†B_i^\dagger in (18) is the inverse of BGi⊥BiB_{G_i^{\perp_{B_i}}}. □\square

Additive Schwarz method using the FSL

In this section we will, in the first instance, apply the FSL to the two-level additive Schwarz algorithm which uses the split near-kernel (SNK) coarse space from Definition 2.5, the setup of which was detailed in the previous section. We then improve the condition number bound obtained to give a robust approach by enriching the underlying coarse space with appropriate generalised eigenfunctions from the orthogonal complements Gi⊥BiG_i^{\perp_{B_i}}.

A two-level additive Schwarz method with the split near-kernel coarse space

In order to apply the FSL (Lemma 2.3) we first need to check its assumptions. These results are obtained in the following two lemmas.

Lemma 3.1 (Surjectivity and continuity of RASR_{\mathrm{AS}}). The operator RASR_{\mathrm{AS}} defined in (16) is surjective and continuous with continuity constant cR=k0c_R=k_0, as defined in (11).

Proof. First, we note that for any U∈HU\in H, due to the partition of unity property (10), we have

U=P0U+(I−P0)U=P0U+(I−P0)∑i=1NRiTDiRiU.U=P_0U+(I-P_0)U=P_0U+(I-P_0)\sum_{i=1}^{N}R_i^{T}D_iR_iU.

Now since P0U∈VGP_0\mathbf{U}\in V_G there exists U0\mathbf{U}_0 such that ZU0=P0UZ\mathbf{U}_0=P_0\mathbf{U}. Therefore, we have

U=RAS(U0,(DiRiU)1≤i≤N).\mathbf{U}=\mathcal{R}_{\mathrm{AS}}\left(\mathbf{U}_0,(D_iR_i\mathbf{U})_{1\leq i\leq N}\right).

and hence RAS\mathcal{R}_{\mathrm{AS}} is surjective.

To prove continuity, we are required to estimate a constant cRc_R such that for all U=(U0,(Ui)1≤i≤N)∈HD\mathbf{U}=(\mathbf{U}_0,(\mathbf{U}_i)_{1\leq i\leq N})\in H_D we have

a(RAS(U),RAS(U))≤cRb(U,U)=cR[(EU0,U0)+∑i=1N(RiARiTUi,Ui)].a(\mathcal{R}_{\mathrm{AS}}(\mathbf{U}),\mathcal{R}_{\mathrm{AS}}(\mathbf{U}))\leq c_R b(\mathbf{U},\mathbf{U})=c_R\left[(E\mathbf{U}_0,\mathbf{U}_0)+\sum_{i=1}^{N}(R_iAR_i^T\mathbf{U}_i,\mathbf{U}_i)\right].

In this regard, using twice the AA-orthogonality of I−P0I-P_0 along with Lemma 2.1, we have the estimate

a(RAS(U),RAS(U))=∥ZU0+(I−P0)∑i=1NRiTUi∥A2=∥ZU0∥A2+∥(I−P0)∑i=1NRiTUi∥A2≤(EU0,U0)+k0∑i=1N∥RiTUi∥A2≤k0[(EU0,U0)+∑i=1N∥RiTUi∥A2].\begin{aligned} a(\mathcal{R}_{\mathrm{AS}}(\mathbf{U}),\mathcal{R}_{\mathrm{AS}}(\mathbf{U})) &=\left\|Z\mathbf{U}_0+(I-P_0)\sum_{i=1}^{N}R_i^T\mathbf{U}_i\right\|_A^2\\ &=\|Z\mathbf{U}_0\|_A^2+\left\|(I-P_0)\sum_{i=1}^{N}R_i^T\mathbf{U}_i\right\|_A^2\\ &\leq(E\mathbf{U}_0,\mathbf{U}_0)+k_0\sum_{i=1}^{N}\|R_i^T\mathbf{U}_i\|_A^2\\ &\leq k_0\left[(E\mathbf{U}_0,\mathbf{U}_0)+\sum_{i=1}^{N}\|R_i^T\mathbf{U}_i\|_A^2\right]. \end{aligned}

Thus the estimate of the constant of continuity of RAS\mathcal{R}_{\mathrm{AS}} can be chosen as cR:=k0≥1c_R:=k_0\geq1, as defined in (2.3).

Lemma 3.2 (Stable decomposition with RAS\mathcal{R}_{\mathrm{AS}}). The operator RAS\mathcal{R}_{\mathrm{AS}} verifies the stable decomposition property, i.e., ∃cT>0\exists c_T>0 such that ∀U∈H ∃U∈HD\forall\mathbf{U}\in H\ \exists\mathcal{U}\in H_D with RAS(U)=U\mathcal{R}_{\mathrm{AS}}(\mathcal{U})=\mathbf{U} and

cTb(U,U)≤a(RAS(U),RAS(U))=a(U,U).c_Tb(\mathcal{U},\mathcal{U})\leq a(\mathcal{R}_{\mathrm{AS}}(\mathcal{U}),\mathcal{R}_{\mathrm{AS}}(\mathcal{U}))=a(\mathbf{U},\mathbf{U}).

In particular, such a constant cTc_T is given by

cT:=11+k1τ0,(20)c_T:=\frac{1}{1+k_1\tau_0}, \tag*{(20)}

where τ0\tau_0 will be defined in (3.3) and with k1k_1 as defined in Lemma 2.2.

Proof. Note that for U∈H\mathbf{U}\in H, we have

U=P0U+(I−P0)U=P0U+(I−P0)∑i=1NRiTDiRiU\mathbf{U}=P_0\mathbf{U}+(I-P_0)\mathbf{U}=P_0\mathbf{U}+(I-P_0)\sum_{i=1}^{N}R_i^TD_iR_i\mathbf{U}
U=P0U+(I−P0)∑i=1NRiTDi(I−ξ0i)RiU+(I−P0)∑i=1NRiTDiξ0iRiU.\mathbf{U}=P_0\mathbf{U}+(I-P_0)\sum_{i=1}^{N}R_i^TD_i(I-\xi_{0i})R_i\mathbf{U}+(I-P_0)\sum_{i=1}^{N}R_i^TD_i\xi_{0i}R_i\mathbf{U}.

Let us expand on the last line and observe that, since P0U∈VGP_0\mathbf{U}\in V_G, there exists U0\mathbf{U}_0 such that ZU0=P0UZ\mathbf{U}_0=P_0\mathbf{U} and, further, the third term is zero since ∑i=1NRiTDiξ0iRiU∈VG\sum_{i=1}^{N}R_i^TD_i\xi_{0i}R_i\mathbf{U}\in V_G. Therefore, we have

U=RAS(U),U=(U0,(Di(I−ξ0i)RiU)1≤i≤N).\mathbf{U}=\mathcal{R}_{\mathrm{AS}}(\mathcal{U}),\quad\mathcal{U}=\left(\mathbf{U}_0,\left(D_i(I-\xi_{0i})R_i\mathbf{U}\right)_{1\leq i\leq N}\right).

Determining the stability of this decomposition consists in the estimation of a constant cT>0c_T > 0 such that

cT[(EU0,U0)+∑j=1N(RjAjRjTDj(I−ξ0j)RjU,Dj(I−ξ0j)RjU)]≤a(U,U).c_T\left[\left(E\mathbf{U}_0,\mathbf{U}_0\right)+\sum_{j=1}^{N}\left(R_jA_jR_j^T D_j(I-\xi_{0j})R_j\mathbf{U},D_j(I-\xi_{0j})R_j\mathbf{U}\right)\right]\leq a(\mathbf{U},\mathbf{U}).

Introducing AjNeuA_j^{\mathrm{Neu}} and, in the second step, applying Lemma 2.2 we have

∑j=1N(RjAjRjTDj(I−ξ0j)RjU,Dj(I−ξ0j)RjU)≤τ0∑j=1N(AjNeuRjU,RjU)≤τ0k1a(U,U),(21)\begin{aligned} \sum_{j=1}^{N}\left(R_jA_jR_j^T D_j(I-\xi_{0j})R_j\mathbf{U},D_j(I-\xi_{0j})R_j\mathbf{U}\right)\leq\tau_0\sum_{j=1}^{N}\left(A_j^{\mathrm{Neu}}R_j\mathbf{U},R_j\mathbf{U}\right) \\ &\leq\tau_0 k_1a(\mathbf{U},\mathbf{U}), \tag*{(21)} \end{aligned}

where we define

τ0:=max⁡1≤j≤Nmax⁡Vj∈Rnj(RjAjRjTDj(I−ξ0j)Vj,Dj(I−ξ0j)Vj)(AjNeuVj,Vj),(22)\tau_0:=\max_{1\leq j\leq N}\max_{\mathbf{V}_j\in\mathbb{R}^{n_j}}\frac{\left(R_jA_jR_j^T D_j(I-\xi_{0j})\mathbf{V}_j,D_j(I-\xi_{0j})\mathbf{V}_j\right)}{\left(A_j^{\mathrm{Neu}}\mathbf{V}_j,\mathbf{V}_j\right)}, \tag*{(22)}

so that the first inequality holds. Now by applying (21), we obtain

b(U,U)=(EU0,U0)+∑j=1N(RjAjRjTDj(I−ξ0j)RjU,Dj(I−ξ0j)RjU)≤a(P0U,P0U)+τ0k1a(U,U)≤(1+k1τ0)a(U,U).\begin{aligned} b(\mathbf{U},\mathbf{U})=\left(E\mathbf{U}_0,\mathbf{U}_0\right)+\sum_{j=1}^{N}\left(R_jA_jR_j^T D_j(I-\xi_{0j})R_j\mathbf{U},D_j(I-\xi_{0j})R_j\mathbf{U}\right) \\ &\leq a(P_0\mathbf{U},P_0\mathbf{U})+\tau_0 k_1a(\mathbf{U},\mathbf{U})\leq(1+k_1\tau_0)a(\mathbf{U},\mathbf{U}). \end{aligned}

Thus the stable decomposition estimate holds with the constant cTc_T in (20).

Theorem 3.3 (Convergence of additive Schwarz with the split near-kernel coarse space). The spectral condition number estimate for the two-level additive Schwarz preconditioner MAS,SNK−1:=RASB−1RAS∗M_{\mathrm{AS,SNK}}^{-1}:=R_{\mathrm{AS}}B^{-1}R_{\mathrm{AS}}^* is

κ(MAS,SNK−1A)≤(1+k1τ0)k0.\kappa(M_{\mathrm{AS,SNK}}^{-1}A)\leq(1+k_1\tau_0)k_0.

The proof is a direct consequence of Lemmas 3.1 and 3.2, which give the necessary criterion to apply the FSL (Lemma 2.3).

We note that this spectral condition number depends on the constants k0k_0 and k1k_1 which are fixed and bounded for a given decomposition into subdomains. However, the remaining constant τ0\tau_0 may be very large due to the shape and size of the domain or the heterogeneities in the coefficients of the problem, leading to a large condition number. The fix is to enlarge the coarse space by a generalised eigenvalue problem (GEVP) induced by the formula (22). More precisely, we introduce in the following a generalised eigenvalue problem and the related GenEO coarse space.

The GenEO coarse space

We start by introducing the relevant GEVP that will yield the appropriate GenEO eigenvectors that should be added in addition to the split near-kernel coarse space.

GEVP 3.4 (Generalised eigenvalue problem for the lower bound). For each of the subdomains 1≤j≤N1\leq j\leq N, we introduce the generalised eigenvalue problem:

Find (Vjk,λjk)∈Rnj∖{0}×R(\mathbf{V}_{jk},\lambda_{jk})\in\mathbb{R}^{n_j}\setminus\{0\}\times\mathbb{R} such that

(I−ξ0jT)DjRjAjRjTDj(I−ξ0j)Vjk=λjkAjNeuVjk.(23)(I-\xi_{0j}^{T})D_jR_jA_jR_j^TD_j(I-\xi_{0j})\mathbf{V}_{jk}=\lambda_{jk}A_j^{\mathrm{Neu}}\mathbf{V}_{jk}. \tag*{(23)}

Let τ>0\tau> 0 be a user-defined threshold, we define Vj,GenEOτ⊂RnV^{\tau}_{j,\mathrm{GenEO}} \subset\mathbb{R}^{n} as the vector space spanned by the family of vectors (RjTDj(I−ξ0j)Vjk)λjk>τ(R_j^T D_j(I-\xi_{0j})V_{jk})_{\lambda_{jk}>\tau} corresponding to eigenvalues larger than τ\tau. Further, let VGenEOτV^{\tau}_{\mathrm{GenEO}} be the vector space spanned by the collection over all subdomains of the vector spaces (Vj,GenEOτ)1≤j≤N(V^{\tau}_{j,\mathrm{GenEO}})_{1\leq j\leq N}.

Remark 3.5. The classical GenEO coarse space is built from the following GEVP:

Find (Vjk,λjk)∈Rnj∖{0}×R such that\text{Find }(\mathbf{V}_{jk},\lambda_{jk})\in\mathbb{R}^{n_j}\setminus\{0\}\times\mathbb{R}\text{ such that}
DjRjARjTDjVjk=λjkAjNeuVjk,(24)D_jR_jAR_j^TD_j\mathbf{V}_{jk}=\lambda_{jk}A_j^{\mathrm{Neu}}\mathbf{V}_{jk}, \tag*{(24)}

and selecting eigenvectors with λjk>τ\lambda_{jk}>\tau. Our experience in this instance is that there are too many large eigenvalues such that the eigensolver has difficulty finding them. This is addressed by the filtering technique introduced above. As a result, the cost remains comparable to that of classical GenEO approaches applied to scalar problems, such as Darcy flow or elasticity. In this context, the overhead is not a practical limitation, even in large-scale runs (see also [33, 44]). Each eigenproblem is local to a subdomain, thus embarrassingly parallel, and its cost depends only on the subdomain size, which can be adjusted by the user.

In the theory that follows for the stable decomposition estimate but not in the algorithm itself, we will make use of the projection πj\pi_j which we now define.

Definition 3.6 (Projection onto GenEO eigenspace). Let πj\pi_j be the projection from Rnj\mathbb{R}^{n_j} on

Vj,τ:=span⁡{Vjk∣λjk>τ}V_{j,\tau}:=\operatorname{span}\{\mathbf{V}_{jk}\mid\lambda_{jk}>\tau\}

parallel to the complement span⁡{Vjk∣λjk≤τ}\operatorname{span}\{\mathbf{V}_{jk}\mid\lambda_{jk}\leq\tau\}.

The key to understanding GEVP 3.4 is the following bound, which can be derived directly from [17 Lemma 7.7].

Lemma 3.7. Consider GEVP 3.4 and the projection πj\pi_j from Definition 3.6. Then for all Uj∈RnjU_j\in\mathbb{R}^{n_j} we have

(RjARjTDj(I−ξ0j)(I−πj)Uj,Dj(I−ξ0j)(I−πj)Uj)≤τ(AjNeuUj,Uj),(25)\left(R_jAR_j^TD_j(I-\xi_{0j})(I-\pi_j)U_j,D_j(I-\xi_{0j})(I-\pi_j)U_j\right)\leq\tau\left(A_j^{\mathrm{Neu}}U_j,U_j\right), \tag*{(25)}

where we reiterate that τ>0\tau>0 is a user-defined threshold within the GEVP.

We can now build the coarse space V0V_0 from the split near-kernel space VGV_G in Definition 2.5 and the GenEO space VGenEOτV^{\tau}_{\mathrm{GenEO}} in GEVP 3.4 by defining the vector space sum

V0:=VG+VGenEOτ.(26)V_0:=V_G+V^{\tau}_{\mathrm{GenEO}}. \tag*{(26)}

The coarse space V0V_0 is spanned by the columns of a full rank rectangular matrix ZZ with n0n_0 columns. We can now define the abstract framework for the two-level additive Schwarz preconditioner with the split near-kernel and GenEO coarse space. The definition of HDH_D and the bilinear form b(⋅,⋅)b(\cdot,\cdot) follow the same as in (14) and (15), only with the definition of the space V0V_0 being different. We denote by RAS,2:HD→H\mathcal{R}_{\mathrm{AS},2}:H_D\to H the linear operator linking the new product Hilbert space HDH_D to HH, defined identically to that in (16) except that ZZ and P0P_0 now stem from the newly enhanced coarse space (26). In order to apply the FSL (Lemma 2.3) we must check the three key assumptions as before. Surjectivity and continuity of RAS,2\mathcal{R}_{\mathrm{AS},2} follows identically to that for RAS\mathcal{R}_{\mathrm{AS}} as in Lemma 3.1, thus, we need only consider the stable decomposition property.

Lemma 3.8 (Stable decomposition with RAS,2\mathcal{R}_{\mathrm{AS},2}). The operator RAS,2:HD→H\mathcal{R}_{\mathrm{AS},2}: H_D \to H verifies the stable decomposition property required by Lemma 2.3 where, for a user-defined threshold τ>0\tau> 0, the constant can be taken as

cT:=11+k1τ(27)c_T := \frac{1}{1+k_1\tau} \tag*{(27)}

Proof. Consider U∈H\mathbf{U} \in H, then following a similar argument to before we have

U=P0U+(I−P0)∑i=1NRiTDiRiU=P0U+(I−P0)∑i=1NRiTDi(I−ξ0i)RiU+(I−P0)∑i=1NRiTDiξ0iRiU⏟=0=P0U+(I−P0)∑i=1NRiTDi(I−ξ0i)(I−πi)RiU+(I−P0)∑i=1NRiTDi(I−ξ0i)πiRiU⏟=0.\begin{aligned} \mathbf{U} &= P_0\mathbf{U} + (I-P_0)\sum_{i=1}^{N}R_i^T D_iR_i\mathbf{U}\\ &= P_0\mathbf{U} + (I-P_0)\sum_{i=1}^{N}R_i^T D_i(I-\xi_{0i})R_i\mathbf{U} + (I-P_0)\underbrace{\sum_{i=1}^{N}R_i^T D_i\xi_{0i}R_i\mathbf{U}}_{=0}\\ &= P_0\mathbf{U} + (I-P_0)\sum_{i=1}^{N}R_i^T D_i(I-\xi_{0i})(I-\pi_i)R_i\mathbf{U}\\ &\quad+ (I-P_0)\underbrace{\sum_{i=1}^{N}R_i^T D_i(I-\xi_{0i})\pi_iR_i\mathbf{U}}_{=0}. \end{aligned}

Now the very last term is zero since RiTDi(I−ξ0i)πiRiU∈VGenEOT⊂V0R_i^T D_i(I-\xi_{0i})\pi_iR_i\mathbf{U} \in V^{T}_{\mathrm{GenEO}} \subset V_0 for all 1≤i≤N1 \leq i \leq N. Let U0∈Rn0\mathbf{U}_0 \in\mathbb{R}^{n_0} be such that ZU0=P0UZ\mathbf{U}_0=P_0\mathbf{U}, then we can choose the decomposition

U=RAS,2(U0,(Di(I−ξ0i)(I−πi)RiU)1≤i≤N).\mathbf{U}=\mathcal{R}_{\mathrm{AS},2}\left(\mathbf{U}_0,\left(D_i(I-\xi_{0i})(I-\pi_i)R_i\mathbf{U}\right)_{1\leq i\leq N}\right).

With this decomposition, using the GEVP bound (25) from Lemma 3.7 and then Lemma 2.2, we have

(EU0,U0)+∑j=1N(RjARjTDj(I−ξ0j)(I−πj)RjUDj(I−ξ0j)(I−πj)RjU)≤a(ZU0,ZU0)+τ∑j=1N(AjNeuRjU,RjU)≤a(P0U,P0U)+τk1a(U,U)≤(1+k1τ)a(U,U).\begin{aligned} &(E\mathbf{U}_0,\mathbf{U}_0)+\sum_{j=1}^{N}\left(R_jAR_j^T D_j(I-\xi_{0j})(I-\pi_j)R_j\mathbf{U}D_j(I-\xi_{0j})(I-\pi_j)R_j\mathbf{U}\right)\\ &\qquad\leq a(Z\mathbf{U}_0,Z\mathbf{U}_0)+\tau\sum_{j=1}^{N}\left(A_j^{\mathrm{Neu}}R_j\mathbf{U},R_j\mathbf{U}\right)\\ &\qquad\leq a(P_0\mathbf{U},P_0\mathbf{U})+\tau k_1a(\mathbf{U},\mathbf{U})\\ &\qquad\leq(1+k_1\tau)a(\mathbf{U},\mathbf{U}). \end{aligned}

Thus the stable decomposition property holds with constant (27). \hfill □\square

This provides the necessary machinery to apply the FSL (Lemma 2.3). Defining the resulting preconditioner as MAS,2−1M_{\mathrm{AS},2}^{-1}, namely being

MAS,2−1=RAS,2B−1RAS,2∗=Z(ZTAZ)−1ZT+(I−P0)∑i=1NRiTBi−1Ri(I−P0T)=Z(ZTAZ)−1ZT+(I−P0)∑i=1NRiT(RiARiT)−1Ri(I−P0T),\begin{aligned} M_{\mathrm{AS},2}^{-1} &=\mathcal{R}_{\mathrm{AS},2}B^{-1}\mathcal{R}_{\mathrm{AS},2}^{*}\\ &=Z(Z^TAZ)^{-1}Z^T+(I-P_0)\sum_{i=1}^{N}R_i^TB_i^{-1}R_i(I-P_0^T)\\ &=Z(Z^TAZ)^{-1}Z^T+(I-P_0)\sum_{i=1}^{N}R_i^T(R_iAR_i^T)^{-1}R_i(I-P_0^T), \end{aligned}

we arrive at our main theorem as a corollary of the results just proved.

Theorem 3.9 (Convergence of additive Schwarz with the split near-kernel and GenEO coarse space). For a given user-defined constant τ>0\tau> 0, namely the eigenvalue threshold within GEVP 3.4, the spectral condition number estimate for the two-level additive Schwarz preconditioner MAS,2−1M_{\mathrm{AS},2}^{-1}, with coarse space stemming from the split near-kernel (Definition 2.5) and GenEO eigenvectors (GEVP 3.4), is given by

κ(MAS,2−1A)≤(1+k1τ)k0.\kappa(M_{\mathrm{AS},2}^{-1}A) \leq(1 + k_1\tau)k_0.

Numerical results

In this section we study two broad challenges, that of domains with a growing number of holes and that of increasing heterogeneity within the physical parameters ε\varepsilon and μ\mu in the problem, and compare both one-level and two-level methods along with AMS. We use different colours in our tabulated results to indicate the performance of a given method; namely, green for low and robust iteration counts, orange for relatively low but not robust iteration counts, and red for iteration counts which become large and lack robustness. The reader can find a summary of the trends observed in our experiments in Table 11 as part of the conclusions.

ExperimentGeometryBCsTableAMSASAS-SNKAS-SNK-GenEOAS-NKAS-NK-GenEO
weak scalabilityno holesDirichlet1greengreengreengreenorangeorange
mixedgreenredgreengreenorangeorange
holesDirichlet2redorangegreengreengreengreen
mixedredredredgreenredgreen
increasing holesfixed domainDirichlet3greenorangeorangegreenorangeorange
strong scalabilityholesDirichlet4orangeredorangegreenorangeorange
mixedredredorangegreenredred
varying γ\gammano holesDirichlet5greenredgreengreengreengreen
holesDirichlet6redorangegreengreengreengreen
mixedredredredgreenredorange
heterogeneity varying ε\varepsilonsimpleDirichlet7greenredgreengreenorangeorange
mixedgreenredgreengreenorangeorange
complexDirichlet9greenredgreengreenorangeorange
mixedgreenredgreengreenorangeorange
heterogeneity varying μ\musimpleDirichlet8greenredredgreenredgreen
mixedgreenredredgreenredgreen
complexDirichlet10orangeredredgreenredgreen
mixedorangeredredgreenredgreen

Table 11. A summary of the trends exhibited in our numerical results for the different preconditioners. Recall that green indicates low and robust iteration counts, orange implies relatively low but not robust iteration counts, while red signals that iteration counts become large and lack robustness. Note: AMS is highly effective for simple geometries without holes, benefiting from available optimised implementations. In contrast, GenEO-type methods adapt automatically to the discretisation and are more robust in complex or evolving simulation settings.

In subsection 4.1, we first present weak scaling experiments for a reference test case consisting of a 3D beam with square cross section having size N2×1×1\frac{N}{2} \times1 \times1, where NN is the number of subdomains, which may additionally have holes; see Figure 1 for the configuration with holes. In addition, we perform a strong scalability study on a 3D beam of fixed size, as reported in Table 4. We then study the dependence on γ\gamma in subsection 4.2, where γ\gamma varies from 10−510^{-5} to 10210^{2} on a complex geometric configuration. Based on these results, we select γ=10−3\gamma= 10^{-3} for the other experiments provided since it is representative of a difficult case. In subsection 4.3, we move to study the impact of varying coefficients ε\varepsilon and μ\mu. More details on each experiment will be given below.

NN8163264128256
#dofs226K226K226K226K226K226K
NK size36K36K36K36K36K36K
SNK size43K50K58K66K82K101K
GenEO size (Dirichlet)000000
GenEO size (mixed)326149433610
DirichletAMS977168717799
DirichletAS5283111121126136
DirichletAS-SNK192124252323
DirichletAS-SNK-GenEO192124252323
DirichletAS-NK212428303132
DirichletAS-NK-GenEO212428303132
mixedAMS60677298106104
mixedAS6592104121163200
mixedAS-SNK648771354025
mixedAS-SNK-GenEO262930292625
mixedAS-NK6489102113155181
mixedAS-NK-GenEO386595109156181

Table 4. A strong scalability study detailing iteration counts required for the homogeneous beam problem of size 8×1×18 \times1 \times1 with holes (non-trivial topology) and Dirichlet (top) or mixed (bottom) boundary conditions. The geometry is fixed and a METIS partitioning is used to define NN subdomains.

The beam is discretised using a Cartesian mesh with mesh spacing h=116h = \frac{1}{16}, on which we use Nédélec edge elements of lowest order; this leads to linear systems of size ranging from 121,696 dofs, for N=8N = 8, to 3,869,472 dofs, for N=256N = 256. In all test configurations, the right-hand side of equation (1) is defined as a constant volume source term, f=(1,1,1)\boldsymbol{f} = (1,1,1). To solve the linear systems we use right-preconditioned GMRES6 along with a relative stopping criterion of 10−610^{-6}. Aside from in the strong scalability study, the subdomain partitioning is uniform along the length of the beam, yielding subdomains with constant size as NN varies. One layer of elements is added in each direction to define the overlapping decomposition; this results in an overlap region of two elements width between neighbouring subdomains. The strong scalability tests are performed similarly but with an arbitrary mesh partitioning into subdomains given by METIS [34]. The GenEO portion of the coarse space is constructed by selecting eigenvectors of (23) corresponding to eigenvalues larger than τ=10\tau= 10.

All our numerical tests were performed with the free open-source domain specific language FreeFEM [23] and for each test configuration we compare the iteration counts of the following methods:

  • AMS: the auxiliary-space Maxwell solver, originally introduced by Hiptmair and Xu [27] and available via PETSc [4];

  • AS: the one-level additive Schwarz preconditioner;

  • AS-SNK: the two-level Schwarz preconditioner with a coarse space consisting of the split near-kernel (SNK) from Definition 2.5;

  • AS-SNK-GenEO: the two-level Schwarz preconditioner with a coarse space consisting of SNK plus VGenEOV_{\mathrm{GenEO}} with the eigenvectors computed by (23);

  • AS-NK: the two-level Schwarz preconditioner with a coarse space consisting of the “global” near-kernel (NK), i.e., simply GG from Definition 2.5;

  • AS-NK-GenEO: the two-level Schwarz preconditioner with a coarse space consisting of NK plus VGenEOV_{\mathrm{GenEO}}.

Note that AS-SNK can be seen as a generalisation of the Nicolaides coarse space for the Poisson problem [46] to the positive Maxwell system. Indeed, for a Poisson problem, the near-kernel is made of constant functions and the Nicolaides coarse space consists in splitting it amongst the subdomains and multiplying each local component by the partition of unity (PoU). For a positive Maxwell problem in trivial topology, the near-kernel is made of the gradients of H1H^1 functions and the SNK coarse space also splits this amongst the subdomains and multiplies by the PoU, hence the name SNK. Theoretical results can be found in [38, 37] and references therein. The AS-SNK-GenEO approach is the full two-level preconditioner MAS,2−1M_{\mathrm{AS},2}^{-1} with the GenEO-enriched coarse space we have designed and analysed in this work. For comparison, we also report iteration counts for the latter two variants of the two-level preconditioner listed above which are not covered by our theory: AS-NK and AS-NK-GenEO. In each table exhibiting scalability results, we report the number of unknowns (dofs) of the linear system, the size of the “global” near-kernel (NK), the size of the split near-kernel (SNK), and the global number of GenEO eigenvectors entering the coarse space.

Influence of the geometry

In a first test configuration, with results collected in Table 1 (top), we choose the constant coefficients μ=ε=1\mu= \varepsilon= 1 and enforce homogeneous Dirichlet boundary conditions E×n=0\mathbf{E} \times\mathbf{n} = \mathbf{0} on all six faces of the beam. We can see that all methods are robust, including the one-level method (AS) due to the Dirichlet boundary conditions surrounding the elongated geometry. Note that the adaptive GenEO coarse space is empty as it automatically captures the fact that there is no need to enrich the near-kernel coarse space to reach the prescribed condition number.

NN8163264128256
#dofs122K243K484K968K1935K3869K
NK size19K37K74K148K296K592K
SNK size25K50K101K203K406K813K
GenEO size (Dirichlet)000000
GenEO size (mixed)000000
DirichletAMS777777
DirichletAS121515151514
DirichletAS-SNK141515151515
DirichletAS-SNK-GenEO141515151515
DirichletAS-NK141515151515
DirichletAS-NK-GenEO141515151515
mixedAMS111111121212
mixedAS20274169124231
mixedAS-SNK151617171818
mixedAS-SNK-GenEO151617171818
mixedAS-NK192432414242
mixedAS-NK-GenEO192432414242

Table 1. A weak scalability study detailing iteration counts required for the homogeneous beam problem with Dirichlet (top) or mixed (bottom) boundary conditions.

In a second test configuration, Neumann boundary conditions are imposed on two of the four lateral faces of the beam (chosen to be opposing) and the other faces of the beam retain a Dirichlet boundary condition. Results using these mixed boundary conditions are reported in Table 1 (bottom) where we can see that AMS is still robust, while the one-level method is not, with iteration counts ranging from 20 to 231. The SNK coarse space is sufficient to have a robust method, albeit with the iteration count mildly increasing from 15 to 18, so that again the GenEO coarse space is empty. The increase is a little more marked for the global NK coarse space, with iteration counts ranging from 19 to 42, suggesting that treatment, such as the splitting we perform for the SNK, is necessary to obtain a fully robust preconditioner. Nonetheless, note that the size of the NK coarse space is smaller than that of the SNK coarse space and that iteration counts appear to stabilise in the low forties, which is not prohibitively high.

For a third test configuration, we add various holes that extend across the domain. Four square holes extend along the whole length of the beam; see Figure 1 (right). Further, two square holes perpendicularly cross each subdomain, from one lateral side to the opposite side, and also connect to two of the four holes along the length; see Figure 1 (left). Since we perform a weak scalability test, the number of holes increases with the number of subdomains NN. Homogeneous Dirichlet boundary conditions E×n=0\mathbf{E} \times\mathbf{n} = \mathbf{0} are enforced on the whole boundary of the domain. Results are gathered in Table 2 (top). As we can see, with this configuration the AMS preconditioner is no longer robust, resulting in a significant degradation in performance, with iteration counts increasing from 44 to 813. Note that, in this case, we both elongate the domain while increasing the complexity of the topology. To further analyse this behaviour, we consider a fixed cube but with more and more holes in it; see Figure 2. Results are detailed in Table 3. The AMS approach is now scalable again and the decrease in the iteration counts from 26 to 18 is in agreement with the decrease in the problem size from 881k dofs to 696k dofs. These last two tables are in agreement with the theory of the AMS preconditioner, for which it is proved in [25] to be robust with respect to the topology for a fixed sized domain; cf. Table 3. But such a robustness result was not proved when both the domain is elongated and the topology complexity is increased. Our numerical results in Table 2 suggest that it is indeed not the case. A possible fix might be to compute the full kernel of the curl operator for the geometry with holes which is larger than the gradients of H1H^1 functions. This can be done using homology computation tools, see [49], but remains computationally expensive and it is unclear how such an approach can be incorporated into the AMS preconditioner.

The unit cube test configuration with an increasing number of crossing holes in x-y directions: 8 (top left), 18 (top right), 72 (bottom left), and 288 (bottom right) holes.

Figure 2. The unit cube test configuration with an increasing number of crossing holes in x-y directions: 8 (top left), 18 (top right), 72 (bottom left), and 288 (bottom right) holes.

#holes81872288
#dofs891K883K842K696K
NK size132K132K129K120K
SNK size221K219K210K185K
GenEO size0000
AMS26191518
AS83716961
AS-SNK36333228
AS-SNK-GenEO36333228
AS-NK50474942
AS-NK-GenEO50474942

Table 3. A study detailing iteration counts required for the unit cube problem with an increasing number of holes (see Figure 2) and Dirichlet boundary conditions. METIS is used to partition the domain into N=64N = 64 subdomains.

As for the one-level AS preconditioner, it also shows a lack of robustness here, while all two-level domain decomposition methods display nearly constant iteration counts and accordingly the GenEO coarse space is empty.

We return to considering the weak scaling setup with holes, as with Table 2 (top), but now change the boundary conditions imposed on the boundaries created by holes from Dirichlet to Neumann. As shown in Table 2 (bottom), for this configuration the AMS preconditioner is again no longer robust, resulting in a significant degradation in its performance, with iteration counts rising from 43 to 1302. The one-level AS preconditioner shows a similar lack of robustness, with iteration counts reaching 557. In contrast, the full AS-SNK-GenEO preconditioner shows nearly constant iteration counts, increasing only from 23 to 27, and hence good robustness. In this scenario, with mixed boundary conditions, the near-kernel coarse space needs to be enriched and GenEO selects around three vectors per subdomain to enter the coarse space. As for the two other variants, we can see that the global near-kernel (NK) alone is not enough. When adding the GenEO eigenvectors this leads to much better results but with iteration counts a little higher than with AS-SNK-GenEO and somewhat less favourable robustness.

In Table 4 we move to consider a strong scalability study using a beam of size 8×1×18 \times1 \times1, with the same hole configuration as in Table 2 but note that now the number of holes is fixed. We vary the number of subdomains from N=8N = 8 to N=256N = 256 with this fixed problem size and geometry. In contrast to our previous experiments, where strip-wise partitioning was performed, here we use the automatic graph partitioner METIS [34] to decompose the mesh. We observe in case of Dirichlet conditions that, due to the fixed number of holes compared to the weak scaling experiment of Table 2, the iteration counts of AMS are stable but still relatively high (between 68 and 99). Moreover, the only scalable methods with low iteration counts are AS-SNK and AS-SNK-GenEO, which utilise the split near-kernel; GenEO not being required here.

When Neumann boundary conditions (instead of Dirichlet boundary conditions) are imposed on the holes, we first observe that the iteration counts of AMS and AS now show a more substantial growth. Again, the only scalable methods for increasing NN are AS-SNK and AS-SNK-GenEO, which incorporate the split near-kernel (unlike in the case of Table 2 (bottom) where AS-SNK shows the same bad behaviour as AS and AS-NK). Note that here the iteration counts of AS-SNK improve as the subdomains get smaller and, accordingly, the contribution of GenEO decreases to only 10 eigenvectors in total for 256 subdomains, whereas in Table 2 (mixed) the size of the GenEO coarse space is 3(N−2)3(N - 2). This can be explained by the fact that, previously, all interior subdomains contained the same number of holes as NN increased, which is not the case here since subdomains are getting smaller and so there are less holes contained in the subdomains. We also observe in Table 4 (along with results in our other experiments) that the SNK space is strictly larger than the NK space, which contains only gradients of functions. As the number of subdomains increases, the relative size of SNK compared to NK also increases (recall that the dimension of NK is fixed here) and, at some point, the SNK coarse space is nearly sufficient by itself. This difference is illustrated by the iteration counts of AS-NK and AS-NK-GenEO in Table 4 (bottom), which grow to 181 for N=256N = 256, and further reveals that the NK space alone does not contain the relevant information on the presence of holes.

Varying the parameter

In this section, we now consider the effect of varying the parameter γ\gamma in (9), which controls the size of the zeroth-order term, from 10−510^{-5} to 10210^{2} for a decomposition into N=256N = 256 subdomains. Table 5 considers the same beam problem and setup used in Table 1 (top) with N=256N = 256; the size of SNK and NK spaces staying the same. We see that, although the problem gets closer to being singular, all methods perform very well even for very small values of γ\gamma.

γ\gamma10−510^{-5}10−410^{-4}10−310^{-3}10−210^{-2}10−110^{-1}11101010210^{2}
GenEO size00000000
AMS77777875
AS141414141414116
AS-SNK151515151514129
AS-SNK-GenEO151515151514129
AS-NK1515151515141210
AS-NK-GenEO1515151515141210

Table 5. Iteration counts required when varying the parameter γ\gamma for the homogeneous beam problem (no holes) with Dirichlet boundary conditions and N=256N = 256 subdomains. Here we have #dofs=3869K\#\mathrm{dofs} = 3869\mathrm{K}, NK size = 592K, and SNK size = 813K.

In Table 6 we now add in holes to the beam and consider both Dirichlet and mixed boundary conditions, as in Table 2, again focusing on N=256N = 256. The smaller the value γ>0\gamma> 0 takes, the harder the problem is. For the Dirichlet case this is seen by the AMS and one-level AS methods struggling. Otherwise, all our two-level methods perform very well and, accordingly, the GenEO portion of the coarse space is empty.

Table 6 (bottom) exhibits the mixed boundary conditions scenario with holes. The exacting challenge as γ\gamma shrinks is reflected in the iteration counts, which fail to be robust for all methods except AS-SNK-GenEO, in which case the iteration counts tend to stabilise at around 28 iterations, and to a lesser extent AS-NK-GenEO, where the iteration counts are under 60. Note that for large values of γ\gamma (namely γ>1\gamma> 1), the GenEO portion of the coarse space is empty, otherwise its size remains at a constant

762 vectors for this particular problem.

Heterogeneities

In this section we detail experiments with heterogeneous problems similar to those in [27]. The parameter γ\gamma is kept constant at γ=10−3\gamma= 10^{-3} in order to retain a challenging problem. For the results in Table 7, the beam is made of eight alternating layers with the parameter ε\varepsilon alternating between a value of unity and a value εalt\varepsilon_{\mathrm{alt}} which we vary, namely ranging between 10−410^{-4} and 10410^{4}, while μ\mu is kept uniform throughout at a constant value μ=1\mu= 1. In Table 7 (top), we consider Dirichlet boundary conditions on all six faces of the beam. We see that all methods perform very well and that, except for the first column, there is no contribution of the GenEO methodology to the coarse space.

εalt=10−4\varepsilon_{\mathrm{alt}}=10^{-4}εalt=10−3\varepsilon_{\mathrm{alt}}=10^{-3}εalt=10−2\varepsilon_{\mathrm{alt}}=10^{-2}εalt=10−1\varepsilon_{\mathrm{alt}}=10^{-1}εalt=1\varepsilon_{\mathrm{alt}}=1εalt=10\varepsilon_{\mathrm{alt}}=10εalt=102\varepsilon_{\mathrm{alt}}=10^{2}εalt=103\varepsilon_{\mathrm{alt}}=10^{3}εalt=104\varepsilon_{\mathrm{alt}}=10^{4}
GenEO size (Dirichlet)971100000000
GenEO size (mixed)000000000
Dirichlet AMS7777777810
Dirichlet AS141414141414151517
Dirichlet AS-SNK151515151515151515
Dirichlet AS-SNK-GenEO151515151515151515
Dirichlet AS-NK151515151515151515
Dirichlet AS-NK-GenEO151515151515151515
mixed AMS131313121211111110
mixed AS2882332332332312282026521
mixed AS-SNK181818181818181714
mixed AS-SNK-GenEO181818181818181714
mixed AS-NK434343424238332516
mixed AS-NK-GenEO434343424238332516

Table 7. Iteration counts required when varying ε\varepsilon for the heterogeneous beam problem with alternating layers, Dirichlet (top) or mixed (bottom) boundary conditions, and N=256N = 256 subdomains. Here we have #dofs = 3869K, NK size = 592K, and SNK size = 813K.

In Table 7 (bottom) we change the boundary conditions from Dirichlet everywhere to have Neumann conditions on two opposing lateral faces of the beam and, as ε\varepsilon tends to zero, the problem demands a good preconditioner. AMS, AS-SNK and AS-SNK-GenEO all yield very robust iteration counts, however, AS-NK and AS-NK-GenEO require iterations counts that increase from 16 to 43 as ε\varepsilon gets smaller. The one-level

AS method is not robust at all, with iteration counts reaching 288. Note that in this problem instance the GenEO coarse space is again empty and thus not required.

In the setup for Table 8, the beam is still made of eight alternating layers but now ε\varepsilon is fixed throughout at ε=1\varepsilon= 1 and, instead, μ\mu takes alternate values of unity and a value μalt\mu_{\mathrm{alt}} between 10−410^{-4} and 10410^{4}. In Table 8 (top), Dirichlet boundary conditions are imposed on all faces of the beam whereas in Table 8 (bottom) they are again changed to Neumann boundary conditions on two opposing lateral faces of the beam. Numerical results suggest that both small and large extreme values for μ\mu make the problem harder. In this scenario only AMS and the approaches enriched by GenEO, namely AS-SNK-GenEO and AS-NK-GenEO, have robust iteration counts. All other methods have iteration counts that can reach over 300 for the hardest problems. In contrast to varying ε\varepsilon, we now see that the GenEO portion of the coarse space is essential and grows as the problem becomes more challenging, in particular when μalt\mu_{\mathrm{alt}} increases to become very large. In our most extreme case of μalt=104\mu_{\mathrm{alt}} = 10^{4} we require, on average, approximately 300 eigenvectors per subdomain.

μalt\mu_{\mathrm{alt}}10−410^{-4}10−310^{-3}10−210^{-2}10−110^{-1}11010210^{2}10310^{3}10410^{4}
GenEO size (Dir.)25K12K1272000152812K25K
GenEO size (mixed)256025601272000152611K77K
Dirichlet AMS131211979111214
Dirichlet AS3942137625142576211380
Dirichlet AS-SNK41423421152254150335
Dirichlet AS-SNK-GenEO191816211522181820
Dirichlet AS-NK46443421152255154340
Dirichlet AS-NK-GenEO321917211522181920
mixed AMS131313131212121314
mixed AS581407275236231233264378523
mixed AS-SNKOOM403624182658159339
mixed AS-SNK-GenEO191918241826191920
mixed AS-NK48474443424267174354
mixed AS-NK-GenEO444443434242424341

Table 8. Iteration counts required when varying μ\mu for the heterogeneous beam problem with alternating layers, Dirichlet (top) or mixed (bottom) boundary conditions, and N=256N = 256 subdomains. Here we have #dofs=3869K\#\mathrm{dofs} = 3869\mathrm{K}, NKNK size =592K= 592\mathrm{K}, and SNKSNK size =813K= 813\mathrm{K}. The out-of-memory (OOM) is not due to the size of the coarse space alone (indeed, enlarging it with the GenEO vectors removed this issue), but rather to limitations of the coarse solver used, which exceeded the available system memory in this case.

In our final set of experiments we use a more geometrically complex heterogeneity. The problem is again defined in the whole beam but this time with ε\varepsilon or μ\mu varying from unity in the location of the holes used in previous tests, akin to those shown in Figure 1 but now with a fixed N=256N = 256 subdomains. As before, we vary separately ε\varepsilon and μ\mu to distinguish their effects and label the value taken in the holes as εholes\varepsilon_{\mathrm{holes}} or μholes\mu_{\mathrm{holes}}, respectively. We again consider pure Dirichlet and mixed boundary condition cases, as described previously. Results in Table 9 detail the behaviour when εholes\varepsilon_{\mathrm{holes}} varies between 10−410^{-4} and 10410^{4} and μ\mu is kept constant at μ=1\mu= 1. We see that the performance of each method is rather similar to that observed in Table 7, suggesting that the precise geometric variation of ε\varepsilon is not too important.

εholes\varepsilon_{\mathrm{holes}}10−410^{-4}10−310^{-3}10−210^{-2}10−110^{-1}11010210^{2}10310^{3}10410^{4}
GenEO size (Dirichlet)102300000000
GenEO size (mixed)000000000
Dirichlet AMS777777789
Dirichlet AS141414141414172562
Dirichlet AS-SNK151515151515151515
Dirichlet AS-SNK-GenEO151515151515151515
Dirichlet AS-NK151515151515151515
Dirichlet AS-NK-GenEO151515151515151515
mixed AMS121212121212131212
mixed AS232232232232231231226180135
mixed AS-SNK181818181817171615
mixed AS-SNK-GenEO181818181817171615
mixed AS-NK424242424241383324
mixed AS-NK-GenEO424242424241383324

Table 9. Iteration counts required when varying ε\varepsilon for the heterogeneous beam problem with complex geometric heterogeneity which varies in the location of holes akin to Figure 1, Dirichlet (top) or mixed (bottom) boundary conditions, and N=256N = 256 subdomains. Here we have #dofs=3869K\#\mathrm{dofs} = 3869\mathrm{K}, NKNK size =592K= 592\mathrm{K}, and SNKSNK size =813K= 813\mathrm{K}.

In contrast, when varying μholes\mu_{\mathrm{holes}} between 10−410^{-4} and 10410^{4} in Table 10, with a fixed ε=1\varepsilon= 1, we now observe some degradation in the performance of AMS, suggesting that this approach is not fully robust to complex heterogeneities. Our approach, AS-SNK-GenEO, yields the lowest iteration counts in the most challenging regime of large μholes\mu_{\mathrm{holes}} and provides good behaviour overall, with iteration counts stabilising at around 25 once the GenEO space is required. Note that the size of the GenEO coarse space adapts to the difficulty of the problem and thus ultimately increases with the value of μholes\mu_{\mathrm{holes}}. More precisely, the number of GenEO vectors is zero for moderate μholes\mu_{\mathrm{holes}} until it reaches 10210^{2}, at which point the size of the GenEO space increases up to 15K15\mathrm{K} once μholes=104\mu_{\mathrm{holes}} = 10^{4}, which equates to 62 eigenvectors per subdomain on average. While AS-NK-GenEO also performs well in terms of robustness to varying μ\mu, both the one-level AS method and those without the enrichment of GenEO, that is AS-SNK

μholes\mu_{\mathrm{holes}}10−410^{-4}10−310^{-3}10−210^{-2}10−110^{-1}11101010210^{2}10310^{3}10410^{4}
GenEO size (Dirichlet)25K5610000076276215K
GenEO size (mixed)000000762101615K
AMS13121187102256158
AS16161616141735108333
AS-SNK15151515151735105322
AS-SNK-GenEO121115151517182325
AS-NK15151515151735105322
AS-NK-GenEO171315151517182325
AMS181816131212203243
AS216216216219231241247299448
AS-SNK18181818182345119287
AS-SNK-GenEO181818181823262425
AS-NK39393939424453132299
AS-NK-GenEO393939394244454543

Table 10. Iteration counts required when varying μ\mu for the heterogeneous beam problem with complex geometric heterogeneity which varies in the location of holes akin to Figure 1, Dirichlet (top) or mixed (bottom) boundary conditions, and N=256N = 256 subdomains. Here we have #dofs=3869K\#\mathrm{dofs} = 3869\mathrm{K}, NKNK size =592K= 592\mathrm{K}, and SNKSNK size =813K= 813\mathrm{K}.

and AS-NK, become much poorer for large values of μholes\mu_{\mathrm{holes}}.

Conclusions

In this work, we have both designed and tested a new two-level domain decomposition preconditioner for positive Maxwell equations inspired by the idea of subspace decomposition, but based on spectral coarse spaces. This was motivated by the AMS preconditioner of Hiptmair and Xu, although our approach is different in spirit. Our method performs well irrespective of the topology of the domain and which boundary conditions are imposed, as can be seen in a summary of trends exhibited by our experiments in Table 11, where only our approach performs well in all scenarios. The numerical results corroborate our theory and show robustness also with respect to heterogeneous coefficients.

It is worth emphasising that the AMS preconditioner is highly efficient when applied to problems with simple geometries and no topological complexity (e.g., without holes). In such scenarios, AMS may well be preferable to domain decomposition methods, as it benefits from optimised implementations available in libraries such as PETSc. However, the performance of AMS as well as related multigrid techniques depend heavily on the choice of prolongation operators used, which in turn depend on the discretisation scheme. This can limit their applicability in evolving or heterogeneous simulation environments. In contrast, the AS-SNK-GenEO method is designed for robustness precisely in more challenging regimes, such as non-trivial topologies, mixed boundary conditions, and strong heterogeneities, where AMS may deteriorate significantly. A key strength of GenEO-based methods lies in their flexibility: the coarse space is constructed automatically through local generalised eigenvalue problems, adapting naturally to the problem at hand and to the discretisation. This makes our approach particularly useful in scenarios where no corresponding AMS theory or robust implementation exists, or when the discretisation scheme may change over the lifetime of the software. In this sense, AMS and GenEO-type methods are complementary, and hybrid or adaptive strategies could be explored in future work to combine their strengths.

While several variants of our preconditioner are considered, namely an underlying additive Schwarz method with a split near-kernel approach or a global near-kernel approach, both with or without GenEO enrichment, other forms could also be defined and then analysed at the theoretical level, such as those based on an underlying SORAS method or through the use of inexact solvers at the coarse level, as detailed in the preliminary exposition [12].

Presently, we have shown an approach which is essentially a proof of concept, with many directions for further improvement in terms of efficiency. In particular, the coarse space is typically rather sizeable since it contains all gradients of scalar functions, plus a small contribution of local eigenvectors coming from a spectral GenEO coarse space. Nevertheless, the coarse space size is typically five times smaller than the original problem and the gradient part of the coarse matrix retains the sparsity pattern of the underlying finite element discretization. This should enable the design of fast coarse problem solvers. One possible solution is to employ inexact or hierarchical coarse solvers. Such strategies have already received some theoretical backing [12] and will be explored in future work.

Our proposed preconditioner must ultimately be tested on possibly highly irregular or heterogeneous problems arising from concrete applications. Its effectiveness should also be investigated for low-frequency Maxwell problems following the framework of [36], where it can be applied within a block preconditioner.

Finally, we note also that the theoretical framework in § 3 is written in an abstract way and is not specific to the positive Maxwell problem. Our approach could thus be applied numerically to any other situation where the discrete problem has a coefficient matrix with a large near-kernel that can be efficiently computed.

Acknowledgments. This work was granted access to the HPC resources of IRENE@TGCC under the allocation 2023-067730 granted by GENCI. We also thank Clemens Pechstein (Dassault Systèmes) for his helpful comments on the manuscript and his insights on the AMS preconditioner.

Data access statement. Data sharing is not applicable to this article as no datasets were generated or analysed during this study.

References

References

  1. [1]B. Ayuso De Dios, R. Hiptmair, and C. Pagliantini, Auxiliary space preconditioners for SIP-DG discretizations of H(curl)-elliptic problems with discontinuous coefficients, IMA J. Numer. Anal., 37 (2017), pp. 646–686, https://doi.org/10.1093/imanum/drw018.
  2. [2]I. Babuska and R. Lipton, Optimal local approximation spaces for generalized finite element methods with application to multiscale problems, Multiscale Model. Simul., 9 (2011), pp. 373–406, https://doi.org/10.1137/100791051.
  3. [3]F. Bachinger, U. Langer, and J. Schöberl, Efficient solvers for nonlinear time-periodic eddy current problems, Comput. Vis. Sci., 9 (2006), pp. 197–207, https://doi.org/10.1007/s00791-006-0023-z.
  4. [4]S. Balay, M. F. Adams, S. Benson, J. Brown, P. Brune, K. Buschelman, E. Constantinescu, L. Dalcin, A. Dener, V. Eijkhout, J. Faibussowitsch, W. D. Gropp, V. Hapla, T. Isaac, P. Jolivet, D. Karpeev, D. Kaushik, M. G. Knepley, F. Kong, S. Kruger, D. A. May, L. C. McInnes, R. T. Mills, L. Mitchell, T. Munson, J. E. Roman, K. Rupp, P. Sanan, J. Sarich, B. F. Smith, S. Zampini, H. Zhang, H. Zhang, and J. Zhang, PETSc/TAO Users Manual, Tech. Report ANL-21/39 - Revision 3.20, Argonne National Laboratory, 2023, https://doi.org/10.2172/1968587.
  5. [5]M. Bonazzoli, V. Dolean, L. G. Graham, E. A. Spence, and P.-H. Tournier, Domain decomposition preconditioning for the high-frequency time-harmonic Maxwell equations with absorption, Math. Comp., 88 (2019), pp. 2559–2604, https://doi.org/10.1090/mcom/3447.
  6. [6]N. Bootland and V. Dolean, Can DtN and GenEO coarse spaces be sufficiently robust for heterogeneous Helmholtz problems?, Math. Comput. Appl., 27 (2022), p. 35, https://doi.org/10.3390/mca27030035.
  7. [7]N. Bootland, V. Dolean, I. G. Graham, C. Ma, and R. Scheichl, GenEO coarse spaces for heterogeneous indefinite elliptic problems, in Domain Decomposition Methods in Science and Engineering XXVI, S. C. Brenner, E. Chung, A. Klawonn, F. Kwok, J. Xu, and J. Zou, eds., Springer, 2022, pp. 117–125, https://doi.org/10.1007/978-3-030-95025-5_10.
  8. [8]N. Bootland, V. Dolean, I. G. Graham, C. Ma, and R. Scheichl, Overlapping Schwarz methods with GenEO coarse spaces for indefinite and nonself-adjoint problems, IMA J. Numer. Anal., 43 (2023), pp. 1839–1936, https://doi.org/10.1093/imanum/drac036.
  9. [9]N. Bootland, V. Dolean, P. Jolivet, F. Nataf, S. Operto, and P.-H. Tournier, Several ways to achieve robustness when solving wave propagation problems, in Domain Decomposition Methods in Science and Engineering XXVI, S. C. Brenner, E. Chung, A. Klawonn, F. Kwok, J. Xu, and J. Zou, eds., Springer, 2023, pp. 17–28, https://doi.org/10.1007/978-3-030-95025-5_2.
  10. [10]N. Bootland, V. Dolean, P. Jolivet, and P.-H. Tournier, A comparison of coarse spaces for Helmholtz problems in the high frequency regime, Comput. Math. Appl., 98 (2021), pp. 239–253, https://doi.org/10.1016/j.camwa.2021.07.011.
  11. [11]N. Bootland, V. Dolean, A. Kyrakidis, and J. Pestana, Analysis of parallel Schwarz algorithms for time-harmonic problems using block Toeplitz matrices, Electron. Trans. Numer. Anal., 55 (2022), pp. 112–141, https://doi.org/10.1553/etna_vol55s112.
  12. [12]N. Bootland, V. Dolean, F. Nataf, and P.-H. Tournier, Two-level DDM preconditioners for positive Maxwell equations, 2020, https://arxiv.org/abs/2012.02388.
  13. [13]M. Cessenat, Mathematical Methods in Electromagnetism: Linear Theory and Applications, World Scientific, 1996, https://doi.org/10.1142/9789814447157_0002.
  14. [14]P. Ciarlet, Jr and J. Zou, Fully discrete finite element approaches for time-dependent Maxwell’s equations, Numer. Math., 82 (1999), pp. 193–219, https://doi.org/10.1007/s002110050417.
  15. [15]C. R. Dohrmann and O. B. Widlund, An iterative substructuring algorithm for two-dimensional problems in H(curl), SIAM J. Numer. Anal., 50 (2012), pp. 1004–1028, https://doi.org/10.1137/100818145.
  16. [16]C. R. Dohrmann and O. B. Widlund, A BDDC algorithm with deluxe scaling for three-dimensional H(curl) problems, Comm. Pure Appl. Math., 69 (2016), pp. 745–770, https://doi.org/10.1002/cpa.21574.
  17. [17]V. Dolean, P. Jolivet, and F. Nataf, An Introduction to Domain Decomposition Methods: algorithms, theory and parallel implementation, SIAM, 2015, https://doi.org/10.1137/1.9781611974065.
  18. [18]Y. A. Erlangga, C. Vuik, and C. W. Oosterlee, On a class of preconditioners for solving the Helmholtz equation, Appl. Numer. Math., 50 (2004), pp. 409–425, https://doi.org/10.1016/j.apnum.2004.01.009.
  19. [19]R. Estrin and C. Greif, On nonsingular saddle-point systems with a maximally rank deficient leading block, SIAM J. Matrix Anal. Appl., 36 (2015), pp. 367–384, https://doi.org/10.1137/140989996.
  20. [20]I. G. Graham, E. A. Spence, and E. Vainikko, Domain decomposition preconditioning for high-frequency Helmholtz problems with absorption, Math. Comp., 86 (2017), pp. 2089–2127, https://doi.org/10.1090/mcom/3190.
  21. [21]C. Greif and D. Schötzau, Preconditioners for the discretized time-harmonic Maxwell equations in mixed form, Numer. Linear Algebra Appl., 14 (2007), pp. 281–297, https://doi.org/10.1002/nla.515.
  22. [22]M. Griebel and P. Oswald, On the abstract theory of additive and multiplicative Schwarz algorithms, Numer. Math., 70 (1995), pp. 163–180, https://doi.org/10.1007/s002110050115.
  23. [23]F. Hecht, New development in FreeFem++, J. Numer. Math., 20 (2012), pp. 251–265, https://doi.org/10.1515/jnum-2012-0013.
  24. [24]R. Hiptmair and C. Pechstein, Discrete regular decompositions of tetrahedral discrete 1-forms, in Maxwell’s Equations: Analysis and Numerics, U. Langer, D. Pauly, and S. Repin, eds., De Gruyter, 2019, pp. 199–258, https://doi.org/10.1515/9783110543612-007.
  25. [25]R. Hiptmair and C. Pechstein, A review of regular decompositions of vector fields: continuous, discrete, and structure-preserving, in Spectral and High Order Methods for Partial Differential Equations ICOSAHOM 2018, S. J. Sherwin, D. Moxey, J. Peiró, P. E. Vincent, and C. Schwab, eds., Springer, 2020, pp. 45–60, https://doi.org/10.1007/978-3-030-39647-3_3.
  26. [26]R. Hiptmair, G. Widmer, and J. Zou, Auxiliary space preconditioning in H_0(curl; Ω), Numer. Math., 103 (2006), pp. 435–459, https://doi.org/10.1007/s00211-006-0683-0.
  27. [27]R. Hiptmair and J. Xu, Nodal auxiliary space preconditioning in H(curl) and H(div) spaces, SIAM J. Numer. Anal., 45 (2007), pp. 2483–2509, https://doi.org/10.1137/060660588.
  28. [28]Q. Hu, Convergence of the Hiptmair–Xu preconditioner for Maxwell’s equations with jump coefficients (i): Extensions of the regular decomposition, SIAM J. Numer. Anal., 59 (2021), pp. 2500–2535, https://doi.org/10.1137/20M1320365.
  29. [29]Q. Hu, Convergence of the Hiptmair–Xu preconditioner for H(curl)-elliptic problems with jump coefficients (ii): Main results, SIAM J. Numer. Anal., 61 (2023), pp. 2434–2459, https://doi.org/10.1137/22M1501593.
  30. [30]Q. Hu, S. Shu, and J. Zou, A discrete weighted Helmholtz decomposition and its application, Numer. Math., 125 (2013), pp. 153–189, https://doi.org/10.1007/s00211-013-0536-6.
  31. [31]Q. Hu and J. Zou, A nonoverlapping domain decomposition method for Maxwell’s equations in three dimensions, SIAM J. Numer. Anal., 41 (2003), pp. 1682–1708, https://doi.org/10.1137/S0036142901396909.
  32. [32]Q. Hu and J. Zou, Substructure preconditioners for saddle-point problems arising from Maxwell’s equations in three dimensions, Math. Comp., 73 (2004), pp. 35–61, https://doi.org/10.1090/S0025-5718-03-01541-2.
  33. [33]P. Jolivet, F. Hecht, F. Nataf, and C. Prud’Homme, Scalable domain decomposition preconditioners for heterogeneous elliptic problems, in Proceedings of the International Conference on High Performance Computing, Networking, Storage and Analysis, 2013, pp. 1–11.
  34. [34]G. Karypis and V. Kumar, A fast and high quality multilevel scheme for partitioning irregular graphs, SIAM J. Sci. Comput., 20 (1998), pp. 359–392, https://doi.org/10.1137/S1064827595287997.
  35. [35]T. V. Kolev and P. S. Vassilevski, Parallel auxiliary space AMG for H(curl) problems, J. Comput. Math., (2009), pp. 604–623, https://doi.org/10.4208/jcm.2009.27.5.013.
  36. [36]D. Li, C. Greif, and D. Schötzau, Parallel numerical solution of the time-harmonic Maxwell equations in mixed form, Numer. Linear Algebra Appl., 19 (2012), pp. 525–539, https://doi.org/10.1002/nla.782.
  37. [37]Q. Liang, X. Xu, and S. Zhang, On a sharp estimate of overlapping Schwarz methods in H(curl; Ω) and H(div; Ω), IMA J. Numer. Anal., (2024), https://doi.org/10.1093/imanum/drae022.
  38. [38]Y. Liang, H. Xiang, S. Zhang, and J. Zou, Preconditioners and their analyses for edge element saddle-point systems arising from time-harmonic Maxwell’s equations, Numer. Algorithms, 86 (2021), pp. 281–302, https://doi.org/10.1007/s11075-020-00889-7.
  39. [39]C. Ma, C. Alber, and R. Scheichl, Error estimates for discretized generalized FEMs with locally optimal spectral approximations, SIAM J. Numer. Anal., 61 (2023), pp. 1546–1584, https://doi.org/10.1137/21M1466748.
  40. [40]C. Ma and R. Scheichl, Error estimates for discretized generalized FEMs with locally optimal spectral approximations, Math. Comp., 91 (2022), pp. 2539–2569, https://doi.org/10.1090/mcom/3755.
  41. [41]C. Ma, R. Scheichl, and T. Dodwell, Novel design and analysis of generalized finite element methods based on locally optimal spectral approximations, SIAM J. Numer. Anal., 60 (2022), pp. 244–273, https://doi.org/10.1137/21M1406179.
  42. [42]M. Magolu Monga Made, Incomplete factorization-based preconditionings for solving the Helmholtz equation, Internat. J. Numer. Methods Engrg., 50 (2001), pp. 1077–1101, https://doi.org/10.1002/1097-0209(20010220)50:5<1077::AID-NME653>3.0.CO;2-P.
  43. [43]F. Nataf, Mathematical analysis of robustness of two-level domain decomposition methods with respect to inexact coarse solves, Numer. Math., 144 (2020), pp. 811–833, https://doi.org/10.1007/s00211-020-01102-6.
  44. [44]F. Nataf and P.-H. Tournier, A GenEO domain decomposition method for saddle point problems, Comptes Rendus. Mécanique, 351 (2023), pp. 1–18, https://doi.org/10.5802/crmeca.175.
  45. [45]S. V. Nepomnyashchikh, Mesh theorems on traces, normalizations of function traces and their inversion, Russian J. Numer. Anal. Math. Modelling, 6 (1991), pp. 223–242, https://doi.org/10.1515/rnam.1991.6.3.223.
  46. [46]R. A. Nicolaides, Definition of conjugate gradients with applications to boundary value problems, SIAM J. Numer. Anal., 24 (1987), pp. 355–365, https://doi.org/10.1137/0724027.
  47. [47]Y. Notay, Convergence of some iterative methods for symmetric saddle point linear systems, SIAM J. Matrix Anal. Appl., 40 (2019), pp. 122–146, https://doi.org/10.1137/18M1208836.
  48. [48]J. E. Pasciak and J. Zhao, Overlapping Schwarz methods in H(curl) on polyhedral domains, J. Numer. Math., 10 (2002), pp. 221–234, https://doi.org/10.1515/JNMA.2002.221.
  49. [49]M. Pellikka, S. Suuriniemi, L. Kettunen, and C. Geuzaine, Homology and cohomology computation in finite element modeling, SIAM J. Sci. Comput., 35 (2013), pp. B1195–B1214, https://doi.org/10.1137/130906556.
  50. [50]I. Perugia, V. Simoncini, and M. Arioli, Linear algebra methods in a mixed approximation of magnetostatic problems, SIAM J. Sci. Comput., 21 (1999), pp. 1085–1101, https://doi.org/10.1137/S1064827598333211.
  51. [51]J. Pestana and A. J. Wathen, Natural preconditioning and iterative methods for saddle point systems, SIAM Rev., 57 (2015), pp. 71–91, https://doi.org/10.1137/130394921.
  52. [52]N. Spillane, V. Dolean, P. Hauret, F. Nataf, C. Pechstein, and R. Scheichl, Abstract robust coarse spaces for systems of PDEs via generalized eigenproblems in the overlaps, Numer. Math., 126 (2014), pp. 741–770, https://doi.org/10.1007/s00211-013-0576-y.
  53. [53]A. Toselli, Overlapping Schwarz methods for Maxwell’s equations in three dimensions, Numer. Math., 86 (2000), pp. 733–752, https://doi.org/10.1007/PL00005417.
  54. [54]A. Toselli, Dual-primal FETI algorithms for edge finite-element approximations in 3D, IMA J. Numer. Anal., 26 (2006), pp. 96–130, https://doi.org/10.1093/imanum/dri023.
  55. [55]J. Xu, L. Chen, and R. H. Nochetto, Optimal multilevel methods for H(grad), H(curl), and H(div) systems on graded and unstructured grids, in Multiscale, Nonlinear and Adaptive Approximation, R. DeVore and A. Kunoth, eds., Springer, 2009, pp. 599–659, https://doi.org/10.1007/978-3-642-03413-8_14.
  56. [56]J. Xu and Y. Zhu, Robust preconditioner for H(curl) interface problems, in Domain Decomposition Methods in Science and Engineering XIX, Y. Huang, R. Kornhuber, O. Widlund, and J. Xu, eds., Springer, 2011, pp. 173–180, https://doi.org/10.1007/978-3-642-11304-8_18.
  57. [57]S. Zampini, P. Vassilevski, V. Dobrev, and T. Kolev, Balancing domain decomposition by constraints algorithms for curl-conforming spaces of arbitrary order, in Domain Decomposition Methods in Science and Engineering XXIV, P. E. Bjørstad, S. C. Brenner, L. Halpern, H. H. Kim, R. Kornhuber, T. Rahman, and O. B. Widlund, eds., Springer, 2018, pp. 103–116, https://doi.org/10.1007/978-3-319-93873-8_8.

Paper details

vlnc:W_q8tatzzr
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
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