Abstract
We consider Galerkin approximations of eigenvalue problems for holomorphic Fredholm operator functions for which the operators do not have the structure “coercive+compact”. In this case the regularity (in the vocabulary of discrete approximation schemes) of Galerkin approximations is not unconditionally satisfied and the question of convergence is delicate. We report a technique to prove regularity of approximations which is applicable to a wide range of eigenvalue problems. The technique is based on the knowledge of a suitable Test function operator. In particular, we introduce the concepts of weak T-coercivity and T-compatibility and prove that for weakly T-coercive operators, T-compatibility of Galerkin approximations implies their regularity. Our framework can be successfully applied to analyze e.g. complex scaling/perfectly matched layer methods, problems involving sign-changing coefficients due to meta-materials and also (boundary element) approximations of Maxwell-type equations. We demonstrate the application of our framework to the Maxwell eigenvalue problem for a conductive material.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
In this article we consider the approximation of eigenvalue problems for holomorphic Fredholm operator functions of the following form: find \(\lambda \in \Lambda \subset \mathbb {C}\) and non-homogeneous eigenelements u in a Hilbert space X such that \(A(\lambda )u=0\), where it is assumed that \(A(\cdot ):\Lambda \rightarrow L(X)\) is an operator function which depends holomorphically on \(\lambda \) and that for all \(\lambda \in \Lambda \) the operator \(A(\lambda ):X \rightarrow X\) is linear and bounded, and satisfies a so-called weak T-coercivity condition. The analysis of approximations of eigenvalue problems for holomorphic Fredholm operator has a long history and was contributed by e.g. Anselone, Grigorieff, Jeggle, Karma, Stummel, Treuden, Vainikko and Wendland (in alphabetic order). It is usually performed in the framework of discrete approximation schemes [30] and regular approximations of operator functions [1, 18]. This tradition roots in the seventies [19, 25, 33, 34]. In this framework a complete convergence analysis and asymptotic error estimates for eigenvalues are given by Karma in [26, 27]. If the discrete approximation scheme is chosen as a Galerkin scheme, then the assumptions of [26, 27] reduce to a single non-trivial assumption: the regular approximation property (see Definition 5 for the meaning of regularity). If the operators are of the form “coercive+compact”, the regularity of Galerkin approximations is unconditionally satisfied [19, (32)]. However, if the operator values are not of this kind the question of reliable eigenvalue approximations is very delicate. This can already be observed for linear eigenvalue problems, see e.g. [2, 4]. Thus it is little known how to prove regularity of approximations for non-injective Fredholm operators being not weakly coercive. In Theorem 1 we report a new condition on the Galerkin spaces to ensure the regularity of Galerkin approximations such that [26, 27] can be applied. This condition is stronger than the classical regularity condition. However, it is satisfied for a wide variety of applications. We combine our approach with the results of [26, 27] in Theorem 3. On the side, we report in Theorem 3 vii) new asymptotic error estimates on eigenspaces (which are not provided by [26, 27]). Further, we demonstrate how to apply our framework to the Maxwell eigenvalue problem for a conductive material.
As preparation for the forthcoming concept of weakly T-coercive operators (operator functions) we remind the reader how Fredholmness of operators is usually established. In the case of coercive operators Fredholmness is trivial. The same holds for weakly coercive operators A, i.e. A is a compact perturbation of a coercive operator. Alteratively we may construct an isomorphism T such that \(T^*A\) is weakly coercive (\(T^*\) denotes the adjoint operator of T), which yields the Fredholmness of A. The name “T-coercivity” originates from Bonnet-Ben Dhia, Ciarlet, Zwölf [6]. The notion was introduced to analyze differential operators with sign-changing coefficients in the principal part which occur e.g. in the modeling of meta materials. The technique is also applied in the analysis of interior transmission eigenvalue problems, see e.g. [11, 12]. However, as far as we know the concept goes back at least to an article [8] of Buffa, Costabel, Schwab on Maxwell’s equations (wherein \(T=\theta \)). For an operator A to be (weakly) T-coercive means that \(T^*A\) is already (weakly) coercive. However, in eigenvalue problems the operators will be in general not bijective (precisely at the eigenvalues). Thus the nomenclature of T-coercivity is not meaningful for our purposes and we will rely on the term weak T-coercivity. In general the Galerkin spaces will not be T-invariant and hence one cannot reproduce the above analysis on the approximation level. An invariance condition is indeed not necessary, but can be relaxed. We will make precise in which sense the Galerkin spaces have to interact with the operator T to ensure regularity. It will turn out that the existence of bounded linear operators \(T_n\) from the Galerkin spaces \(X_n\) to themselves such that
is sufficient. We call this property “T-compatibility”. The previous “norm” \(\Vert \cdot \Vert _n\) was termed “discrete norm” by Descloux, Nassif and Rappaz [14, 15] wherein it was used in a different but familiar context. In our context it was already employed by Hohage and Nannen [24] for the analysis of complex scaling/perfectly matched layer and Hardy space infinite element methods in cylindrical waveguides; and also by Bonnet-Ben Dhia, Ciarlet and Carvalho [5, 10] for the analysis of finite element methods for equations which involve meta materials. Both works [5, 24] prove weak T-coercivity and T-compatibility. Thus our results can directly be applied to improve the results of [24] and to establish convergence results for approximations of the eigenvalue problems related to [5]. Note that the negative material parameters in meta materials are e.g. of the kind \(1-\frac{1}{\omega ^2}\) with \(\omega ^2\) being the eigenvalue parameter. Hence such eigenvalue problems are indeed non-linear. Recently we applied our framework to study the approximation of such problems in [22], wherein we analyze the approximation of plasmonic electromagnetic eigenvalue problems in 2D.
However, the original motivation for this article was to provide a framework for the convergence analysis of boundary element discretizations of boundary integral formulations of Maxwell eigenvalue problems. Indeed the results of Unger [32] are based on an earlier version of this article. Although in the non-dispersive case the Maxwell eigenvalue problem is of linear nature, its formulation as boundary integral equation becomes non-linear due to the dependency of the fundamental solution on the frequency.
This article is an extract of the thesis [20]. Therein (see also [23]) the framework is applied to establish convergence results for radial complex scaling/perfectly matched layer methods for scalar resonance problems in homogeneous open systems. Although these eigenvalue problems are linear, classical theory [3, 4] can’t be applied.
Indeed our presented framework is of interest even for linear eigenvalue problems for which [3, 4] could be applied. It serves as an alternative and has the advantage that no transformation of the problem to a standard stencil \(A-\lambda I\) by means of a solution operator is necessary. We want to mention that we apply our framework in this sense also in [21] for electromagnetic Steklov eigenvalue problems.
The remainder of this article is structured as follows. In Sect. 2 we introduce the notion of weak T-coercivity and T-compatibility. In Theorem 1 we prove that T-compatibility implies regularity. In Sect. 3 we report in Theorem 3 convergence results for \(T(\cdot )\)-compatible Galerkin approximations of eigenvalue problems for weakly \(T(\cdot )\)-coercive holomorphic operator functions. In Sect. 4 we apply our framework to the Maxwell eigenvalue problem for a conductive material.
2 Weak T-coercivity and T-compatibility
Let X be a Hilbert space over \(\mathbb {C}\). We denote its scalar product as \(\langle \cdot ,\cdot \rangle _X\) and its associated norm as \(\Vert \cdot \Vert _X\). Further, let L(X) be the space of bounded linear operators from X to X with operator norm \(\Vert A\Vert _{L(X)}:=\sup _{u\in X{\setminus }\{0\}}\Vert Au\Vert _X/\Vert u\Vert _X\) for \(A\in L(X)\). For an operator \(A\in L(X)\) we denote its adjoint operator by \(A^*\in L(X)\), i.e. \(\langle u,A^*v\rangle _X = \langle Au,v\rangle _X\) for all \(u,v\in X\). For a closed subspace \(X_n\subset X\) let \(L(X_n)\) be the space of bounded linear operators from \(X_n\) to \(X_n\) with norm \(\Vert A_n\Vert _{L(X_n)}:=\sup _{u_n\in X_n{\setminus }\{0\}}\Vert A_nu_n\Vert _X/\Vert u_n\Vert _X\) for \(A_n\in L(X_n)\) and denote \(P_n\) the orthogonal projection from X to \(X_n\). Henceforth we assume that \((X_n)_{n\in \mathbb {N}}\) is a sequence of closed subspaces of X such that \(P_n\) converges point-wise to the identity, i.e. \(\lim _{n\rightarrow \infty } \Vert u-P_nu\Vert _X=0\) for each \(u\in X\). Further we denote \((A_n:=P_nA|_{X_n}\in L(X_n))_{n\in \mathbb {N}}\) the Galerkin approximation of A.
Definition 1
Let \(A,T\in L(X)\) and T be bijective. The operator A is called
-
1.
Coercive, if \(\inf _{u\in X{\setminus }\{0\}}|\langle Au,u\rangle _X|/\Vert u\Vert _X^2>0\),
-
2.
Weakly coercive, if there exists a compact operator \(K\in L(X)\) such that \(A+K\) is coercive,
-
3.
T-Coercive if \(T^*A\) is coercive,
-
4.
Weakly T-coercive if \(T^*A\) is weakly coercive.
Due to the Lemma of Lax-Milgram every coercive operator is invertible. Every weakly T-coercive operator is Fredholm with index zero. For a (weakly) coercive operator A it is true that the Galerkin approximations \(A_n=P_nA|_{X_n}\in L(X_n)\) inherit the (weak) coercivity, while for (weakly) T-coercive operators it is in general wrong.
We note that if \(T^*A\) is weakly coercive, then \(AT^{-1}\) is so too. Vice-versa, if AT is weakly coercive, then so is \(T^{-*}A\). Hence we could alternatively define A to be (weakly) right T-coercive, if AT is (weakly) coercive. However, we stick to the former variant because it is more convenient.
For an operator \(T\in L(X)\) or \(T\in L(X_n),\) or a sum of such we define the “discrete norm”
Note that we use \(\Vert \cdot \Vert _n\) simultaneously for operators defined on different spaces \(X, X_n\), and hence \(\Vert \cdot \Vert _n\) is only a norm on \(L(X_n)\), but not on L(X) (if \(X_n\ne X\)).
Definition 2
Consider \(T\in L(X)\) and \((T_n\in L(X_n))_{n\in \mathbb {N}}\). We say that \((T_n)_{n\in \mathbb {N}}\) converges to T in discrete norm, if \(\lim _{n\rightarrow \infty } \Vert T-T_n\Vert _n=0\).
We define in the following what we mean by T-compatible approximations of weakly T-coercive operators.
Definition 3
Let \(A\in L(X)\) be weakly T-coercive. Then we call the sequence of Galerkin approximations \((A_n)_{n\in \mathbb {N}}\) to be T-compatible, if \((A_n)_{n\in \mathbb {N}}\) is a sequence of index-zero Fredholm operators and there exists a sequence of index-zero Fredholm operators \((T_n\in L(X_n))_{n\in \mathbb {N}}\) such that \(T_n\) converges to T in discrete norm: \(\lim _{n\rightarrow \infty }\Vert T-T_n\Vert _n=0\).
Definition 4
A sequence \((u_n \in X)_{n\in \mathbb {N}}\) is said to be compact, if for every subsequence exists a converging subsubsequence.
Definition 5
A sequence \((A_n\in L(X_n))_{n\in \mathbb {N}}\) is called regular, if for every bounded sequence \((u_n\in X_n)_{n\in \mathbb {N}}\) the compactness of \((A_nu_n)_{n\in \mathbb {N}}\) already implies the compactness of \((u_n)_{n\in \mathbb {N}}\).
Next we briefly elaborate on the notion of regularity. To this end we recall that for a bijective operator \(A\in L(X)\) its Galerkin approximation \((A_n)_{n\in \mathbb {N}}\) is called stable, if there exists \(n_0>0\) such that \(A_n\) is invertible for each \(n>n_0\) and \(\sup _{n>n_0}\Vert A_n^{-1}\Vert _{L(X_n)} <\infty \). It is well known that for stable Galerkin approximations the solution \(u_n\in X_n\) to \(A_nu_n=P_nf\) converges to the solution \(u\in X\) of \(Au=f\). However for the approximation of eigenvalue problems it is necessary to approximate also non-bijective operators. In this case the notion of stability is not meaningful. E.g. if \(u\in \ker A{\setminus }\{0\}\), then
and hence \((A_n)_{n\in \mathbb {N}}\) cannot be stable. Thus it is necessary to introduce a generalized notion of stability, i.e. regularity. Indeed, the regularity of approximations of bijective operators implies their stability: Assume that \((A_n)_{n\in \mathbb {N}}\) is not stable. Thus there exists a sequence \((u_n\in X_n)_{n\in \mathbb {N}}\) with \(\Vert u_n\Vert _X=1\) for each \(n\in \mathbb {N}\) such that \(\lim _{n\rightarrow \infty }\Vert A_nu_n\Vert _X=0\). If \((A_n)_{n\in \mathbb {N}}\) is regular, there exists a subsequence \((n(m))_{m\in \mathbb {N}}\) and \(u\in X\) such that \(\lim _{m\rightarrow \infty }u_{n(m)}=u\). It follows \(Au\) \(=\) \(\lim _{m\rightarrow \infty } A_{n(m)}u_{n(m)}\) \(=\) 0. Since A is bijective, it follows \(u=0\) which is a contradiction to \(\Vert u\Vert _X=\lim _{m\rightarrow \infty }\Vert u_{n(m)}\Vert _X=1\).
Our next goal is to prove in Theorem 1 that T-compatible Galerkin approximations of weakly T-coercive operators are regular. In preparation we formulate the next two lemmata.
Lemma 1
Let \(T\in L(X){\setminus }\{0\}\) and \((T_n\in L(X_n))_{n\in \mathbb {N}}\) be a sequence of operators with \(\lim _{n\rightarrow \infty } \Vert T-T_n\Vert _n=0\). Then there exist a constant \(c>0\) and an index \(n_0\in \mathbb {N}\) such that
for all \(n>n_0\). If T is bijective and \(T_n\) is Fredholm with index zero for each \(n\in \mathbb {N}\), then there exist a constant \(c>0\) and an index \(n_0\in \mathbb {N}\) such that \(T_n\) is also bijective for all \(n>n_0\) and
Proof
Let \(u_n\in X_n\). With the triangle inequality we deduce
and hence
Since \(\lim _{n\rightarrow \infty }\Vert T-T_n\Vert _n=0\) the right hand side of the previous inequality is bounded. Similar, with the inverse triangle inequality we deduce
and hence
Since \(\lim _{n\rightarrow \infty }\Vert T\Vert _n=\Vert T\Vert _{L(X)}>0\) and \(\lim _{n\rightarrow \infty }\Vert T-T_n\Vert _n=0\) it follows that there exist \(n_0>0\) and \(c>0\) such that \(\Vert T_n\Vert _n>c\) for all \(n>n_0\). For the last claim let \(n_0>0\) be such that \(\Vert T-T_n\Vert _n<\frac{1}{2\Vert T^{-1}\Vert _{L(X)}}\) for all \(n>n_0\). With the inverse triangle inequality and
it follows
for all \(n>n_0\). We deduce that \(T_n\) is injective. Since \(T_n\) is a Fredholm operator with index zero its bijectivity follows. The norm estimate holds due to \(\inf _{u_n\in X_n, \Vert u_n\Vert _X=1} \Vert T_nu_n\Vert _X\) \(=\) \(1/\Vert T_n^{-1}\Vert _{L(X_n)}\). \(\square \)
Lemma 2
Let \(A\in L(X)\) be weakly T-coercive and \(K\in L(X)\) be compact such that \(T^*A+K\) is coercive. Let \((A_n)_{n\in \mathbb {N}}\) be a T-compatible Galerkin approximation of A. Then there exist \(n_0\in \mathbb {N}\) and \(c>0\), such that \(A_n+P_nT^{-*}K|_{X_n}\in L(X_n)\) is invertible and
for all \(n>n_0\). The constant c can be chosen as
for any \(\epsilon >0\).
Proof
Let n be large enough such that \(T_n\) is bijective (see Lemma 1). We compute
with coercivity constant
Since \(\Vert T_n\Vert _{L(X_n)}\) is uniformly bounded from above and below (see Lemma 1) and \(T_n\) converges to T in discrete norm by assumption, it follows the existence of \(n_0\in \mathbb {N}\) and \(c>0\) such that
for all \(n>n_0\). Hence \(A_n+P_nT^{-*}K|_{X_n}\) is injective. Since \(A_n\) is Fredholm with index zero and K is compact, \(A_n+P_nT^{-*}K|_{X_n}\) is Fredholm with index zero too. Thus \(A_n+P_nT^{-*}K|_{X_n}\) is bijective. The norm estimate follows now from
for any bijective \(B_n\in L(X_n)\). \(\square \)
Theorem 1
Let \(A\in L(X)\) be weakly T-coercive and \((A_n)_{n\in \mathbb {N}}\) be a T-compatible Galerkin approximation. Then \((A_n)_{n\in \mathbb {N}}\) is regular.
Proof
Without loss of generality let \((u_n\in L(X_n))_{n\in \mathbb {N}}\) be a bounded sequence, \((A_nu_n)_{n\in \mathbb {N}}\) and \(u'\in X\) be such that \(\lim _{n\rightarrow \infty }A_nu_n=u'\). Let \(K\in L(X)\) be compact such that \(T^*A+K\) is coercive. Let \({\tilde{A}}:=A+T^{-*}K\) and \({\tilde{A}}_n:=P_n{\tilde{A}}|_{X_n}\). Since K is compact and \((u_n)_{n\in \mathbb {N}}\) is bounded, there exist a subsequence \((u_{n(m)})_{m\in \mathbb {N}}\) and \(u''\in X\) such that \(\lim _{m\rightarrow \infty }T^{-*}Ku_{n(m)}=u''\). It follows
Due to Lemma 2 there exist \(c>0\) and \(m_0\in \mathbb {N}\), such that for all \(m>m_0\) the operator \({\tilde{A}}_{n(m)}\) is invertible and \(\Vert {\tilde{A}}_{n(m)}^{-1}\Vert _{L(X_{n(m)})}\le c\). For \(m>m_0\) we compute
The first term on the right hand side of the latter inequality converges to zero, as previously discussed. The third term converges to zero, because \((P_{n(m)})_{m\in \mathbb {N}}\) converges point-wise to the identity. The second term can be estimated as
and converges to zero, because \((P_{n(m)})_{m\in \mathbb {N}}\) converges point-wise to the identity. Hence \(\lim _{m\rightarrow \infty } u_{n(m)} = {\tilde{A}}^{-1}(u'+u'')\). \(\square \)
3 T-Compatible approximation of holomorphic eigenvalue problems
First let us recall the general theory on holomorphic (Fredholm) operator functions. We refer the reader e.g. to [17] and [28, Appendix]. Let \(\Lambda \subset \mathbb {C}\) be an open, connected and non-empty subset of \(\mathbb {C}\). Let \(A(\cdot ):\Lambda \rightarrow L(X)\) be an operator function. An operator function \(A(\cdot )\) is called holomorphic, if it is complex differentiable. An operator function \(A(\cdot )\) is called Fredholm, if \(A(\lambda )\) is Fredholm for each \(\lambda \in \Lambda \). We denote the resolvent set and spectrum of an operator function \(A(\cdot ):\Lambda \rightarrow L(X)\) as
For an operator function \(A(\cdot ):\Lambda \rightarrow L(X)\) we denote by \(A^*(\cdot )\) the operator function defined by \(A^*(\lambda ):=A(\lambda )^*\) for each \(\lambda \in \Lambda \) and by \(A^{-1}(\cdot ):\rho \big (A(\cdot )\big )\rightarrow L(X)\) the operator function defined by \(A^{-1}(\lambda ):=A(\lambda )^{-1}\) for each \(\lambda \in \rho \big (A(\cdot )\big )\). In addition, we note that for a holomorphic operator function \(A(\cdot ):\Lambda \rightarrow L(X)\) the operator function defined by \(\lambda \mapsto A^*(\overline{\lambda })\) is holomorphic as well. Further, we denote by \(A^{(j)}(\cdot ):\Lambda \rightarrow L(X)\) the \(j^{th}\) derivative of a holomorphic operator function \(A(\cdot ):\Lambda \rightarrow L(X)\). It is well known (see e.g. [16, Theorem 8.2]) that for a holomorphic Fredholm operator function \(A(\cdot ):\Lambda \rightarrow L(X)\) such that \(A(\lambda )\) is bijective for at least one \(\lambda \in \Lambda \), the spectrum \(\sigma \big (A(\cdot )\big )\) is discrete, has no accumulation points in \(\Lambda \) and every \(\lambda \in \sigma \big (A(\cdot )\big )\) is an eigenvalue. That is, there exists \(u\in X\) such that \(A(\lambda )u=0\). In this case we call \(u\) an eigenelement. An ordered collection of elements \((u_0,u_1,\dots ,u_{m-1})\) in X is called a Jordan chain at \(\lambda \) if \(u_0\) is an eigenelement corresponding to \(\lambda \) and if
The elements of a Jordan chain are called generalized eigenelements and the closed linear hull of all generalized eigenelements of \(A(\cdot )\) at \(\lambda \) is called the generalized eigenspace \(G(A(\cdot ),\lambda )\) for \(A(\cdot )\) at \(\lambda \). For an eigenelement \(u\in \ker A(\lambda ){\setminus }\{0\}\) we denote by \(\varkappa (A(\cdot ),\lambda ,u)\) the maximal length of a Jordan chain at \(\lambda \) beginning with \(u\) and
The maximal length of a Jordan chain \(\varkappa (A(\cdot ),\lambda )\) is always finite, see e.g. [28, Lemma A.8.3]. Next we generalize Definitions 1, 3, 5 and Theorem 1 to operator functions. For an operator function \(A(\cdot ):\Lambda \rightarrow L(X)\) we call \((A_n(\cdot ):=P_nA(\cdot )|_{X_n}:\Lambda \rightarrow L(X_n))_{n\in \mathbb {N}}\) the Galerkin approximation of \(A(\cdot )\).
Definition 6
Let \(A(\cdot ), T(\cdot ):\Lambda \rightarrow L(X)\) be operator functions and \(\rho \big (T(\cdot )\big )=\Lambda \). \(A(\cdot )\) is (weakly) (\(T(\cdot )\)-) coercive, if \(A(\lambda )\) is (weakly) (\(T(\lambda )\)-)coercive for each \(\lambda \in \Lambda \).
Definition 7
Let \(A(\cdot ):\Lambda \rightarrow L(X)\) be weakly \(T(\cdot )\)-coercive. The sequence of Galerkin approximations \((A_n(\cdot ))_{n\in \mathbb {N}}\) is \(T(\cdot )\)-compatible, if \((A_n(\lambda ))_{n\in \mathbb {N}}\) is \(T(\lambda )\) compatible for each \(\lambda \in \Lambda \).
Definition 8
Let \(A(\cdot ):\Lambda \rightarrow L(X)\) be an operator function. The sequence of Galerkin approximations \((A_n(\cdot ))_{n\in \mathbb {N}}\) is regular, if \((A_n(\lambda ))_{n\in \mathbb {N}}\) is regular for each \(\lambda \in \Lambda \)
Theorem 2
Let \(A(\cdot ):\Lambda \rightarrow L(X)\) be weakly \(T(\cdot )\)-coercive and \((A_n(\cdot ))_{n\in \mathbb {N}}\) be a \(T(\cdot )\)-compatible Galerkin approximation. Then \((A_n(\cdot ))_{n\in \mathbb {N}}\) is regular.
Proof
Follows from Theorem 1. \(\square \)
Our forthcoming results will heavily rely on [26, 27], which’s theory is formulated in a very generalized sense of approximations, i.e. discrete approximation schemes. However, we will need only the restricted case of Galerkin approximations. In particular the theory of [26, 27] applies the following terms. Let U, V, \(U_n, V_n\), \(n\in \mathbb {N}\) be Banach spaces and \(p_n:U\rightarrow U_n\), \(q_n:V\rightarrow V_n\). Note that it is neither required that \(U_n\subset U\) nor \(V_n\subset V\) and that these spaces are only connected through the mappings \(p_n\) and \(q_n\). Further note that it is neither required that \(p_n\) and \(q_n\) are linear or bounded. Instead one assumes the following compatibility conditions
-
(a1)
\(\lim _{n\rightarrow \infty }\Vert p_n u\Vert _{U_n}=\Vert u\Vert _U\) for all \(u\in U\),
-
(a2)
\(\lim _{n\rightarrow \infty }\Vert q_n v\Vert _{V_n}=\Vert v\Vert _V\) for all \(v\in V\),
-
(a3)
\(\lim _{n\rightarrow \infty }\Vert p_n (\alpha u+\alpha ' u')-(\alpha p_n u+\alpha 'p_n u')\Vert _{U_n}=0\) for all \(u,u'\in U\), \(\alpha , \alpha '\in \mathbb {C}\),
-
(a4)
\(\lim _{n\rightarrow \infty }\Vert q_n (\alpha v+\alpha ' v')-(\alpha q_n v+\alpha 'q_n v')\Vert _{V_n}=0\) for all \(v,v'\in V\), \(\alpha , \alpha '\in \mathbb {C}\).
Then the convergence of the sequence \((u_n)_{n\in \mathbb {N}}\), \(u_n\in U_n\) to \(u\in U\) is defined as
and like-wise for \(v_n\in V_n\) and \(v\in V\)
Similarly a sequence \((u_n \in U_n)_{n\in \mathbb {N}}\) is defined to be compact, if for every subsequence exists in turn a converging subsubsequence in the above sense. Consider now the approximation of \(A\in L(U,V)\) by a sequence of operators \((B_n\in L(U_n,V_n))_{n\in \mathbb {N}}\). Then, by definition,
-
(b1)
\((B_n)_{n\in \mathbb {N}}\) approximates A, if for each \(u\in U\) it holds \(\lim _{n\rightarrow \infty }B_np_nu=Au\),
-
(b2)
\((B_n)_{n\in \mathbb {N}}\) is regular, if for every bounded sequence \((u_n\in U_n)_{n\in \mathbb {N}}\) the compactness of \((B_nu_n)_{n\in \mathbb {N}}\) implies the compactness of \((u_n\in U_n)_{n\in \mathbb {N}}\).
Consider now a holomorphic operator function \(A(\cdot ):\Lambda \rightarrow L(U,V)\) and its approximation by a sequence \((B_n(\cdot ):\Lambda \rightarrow L(U_n,V_n))_{n\in \mathbb {N}}\). Then, by definition,
-
(c1)
\((B_n(\cdot ))_{n\in \mathbb {N}}\) is equibounded, if for every compact \({\tilde{\Lambda }}\subset \Lambda \) exists \(c>0\) such that \(\Vert B_n(\lambda )\Vert _{L(U_n,V_n)}\le c\) for all \(\lambda \in {\tilde{\Lambda }}\), \(n\in \mathbb {N}\).
In our particular case of interest U and V coincide and are, in addition, a Hilbert space X, the approximation spaces \(U_n\) and \(V_n\) coincide and are, in addition, a subspace \(X_n\subset X\), \(p_n\) and \(q_n\) coincide and are, in addition, the orthogonal projection \(P_n\), and \(B_n(\cdot )\) is the Galerkin approximation \(A_n(\cdot )=P_nA(\cdot )|_{X_n}\). It follows with basic properties of the orthogonal projection \(P_n\) that (a1)-(a4) and (b1) are satisfied. Since \(A(\cdot )\) is holomorphic it is continuous and thus (c1) follows easily too.
Theorem 3
Let \(\Lambda \subset \mathbb {C}\) be open, connected and non-empty, X be a Hilbert space and L(X) be the space of bounded linear operators from X to X. Let \(A(\cdot ):\Lambda \rightarrow L(X)\) be a holomorphic weakly \(T(\cdot )\)-coercive operator function (see Definition 6) with non-empty resolvent set \(\rho \big (A(\cdot )\big )\ne \emptyset \). Let \((X_n)_{n\in \mathbb {N}}\) be a sequence of closed subspaces of X with orthogonal projections \(P_n\) onto \(X_n\), such that \((P_n)_{n\in \mathbb {N}}\) converges point-wise to the identity, i.e. \(\lim _{n\rightarrow \infty }\Vert u-P_nu\Vert _X=0\) for each \(u\in X\). Let \(A_n(\cdot ):\Lambda \rightarrow L(X_n)\) be the Galerkin approximation of \(A(\cdot )\) defined by \(A_n(\lambda ):=P_nA(\lambda )|_{X_n}\) for each \(\lambda \in \Lambda \). Assume that \(A_n(\cdot )\) is \(T(\cdot )\)-compatible (see Definition 7). Then the following results hold.
-
i)
For every eigenvalue \(\lambda _0\) of \(A(\cdot )\) exists a sequence \((\lambda _n)_{n\in \mathbb {N}}\) converging to \(\lambda _0\) with \(\lambda _n\) being an eigenvalue of \(A_n(\cdot )\) for almost all \(n\in \mathbb {N}\).
-
ii)
Let \((\lambda _n, u_n)_{n\in \mathbb {N}}\) be a sequence of normalized eigenpairs of \(A_n(\cdot )\), i.e.
$$\begin{aligned} A_n(\lambda _n)u_n=0, \end{aligned}$$and \(\Vert u_n\Vert _X=1\), so that \(\lambda _n\rightarrow \lambda _0\in \Lambda \), then
-
a)
\(\lambda _0\) is an eigenvalue of \(A(\cdot )\),
-
b)
\((u_n)_{n\in \mathbb {N}}\) is a compact sequence and its cluster points are normalized eigenelements of \(A(\lambda _0)\).
-
a)
-
iii)
For every compact \({\tilde{\Lambda }}\subset \rho (A)\) the sequence \((A_n(\cdot ))_{n\in \mathbb {N}}\) is stable on \({\tilde{\Lambda }}\), i.e. there exist \(n_0\in \mathbb {N}\) and \(c>0\) such that \(\Vert A_n(\lambda )^{-1}\Vert _{L(X_n)}\le c\) for all \(n>n_0\) and all \(\lambda \in {\tilde{\Lambda }}\).
-
iv)
For every compact \({\tilde{\Lambda }}\subset \Lambda \) with \({\tilde{\Lambda }}\cap \sigma \big (A(\cdot )\big )=\{\lambda _0\}\) and rectifiable boundary \(\partial {\tilde{\Lambda }}\subset \rho \big (A(\cdot )\big )\) exists an index \(n_0\in \mathbb {N}\) such that
$$\begin{aligned} {\text {dim}}G(A(\cdot ),\lambda _0) = \sum _{\lambda _n\in \sigma \left( A_n(\cdot )\right) \cap {\tilde{\Lambda }}} {\text {dim}}G(A_n(\cdot ),\lambda _n). \end{aligned}$$for all \(n>n_0\), whereby \(G(B(\cdot ),\lambda )\) denotes the generalized eigenspace of an operator function \(B(\cdot )\) at \(\lambda \in \Lambda \). Let \({\tilde{\Lambda }}\subset \Lambda \) be a compact set with rectifiable boundary \(\partial {\tilde{\Lambda }}\subset \rho \big (A(\cdot )\big )\), \({\tilde{\Lambda }}\cap \sigma \big (A(\cdot )\big )=\{\lambda _0\}\) and
$$\begin{aligned} \delta _n&:=\max _{\begin{array}{c} u_0\in G(A(\cdot ),\lambda _0)\\ \Vert u_0\Vert _X\le 1 \end{array}} \, \inf _{u_n\in X_n} \Vert u_0-u_n\Vert _X,\\ \delta _n^*&:=\max _{\begin{array}{c} u_0\in G(A^*(\overline{\cdot }),\lambda _0)\\ \Vert u_0\Vert _X\le 1 \end{array}} \, \inf _{u_n\in X_n} \Vert u_0-u_n\Vert _X, \end{aligned}$$whereby \(\overline{\lambda _0}\) denotes the complex conjugate of \(\lambda _0\) and \(A^*(\cdot )\) the adjoint operator function of \(A(\cdot )\) defined by \(A^*(\lambda ):=A(\lambda )^*\) for each \(\lambda \in \Lambda \). Then there exist \(n_0\in \mathbb {N}\) and \(c>0\) such that for all \(n>n_0\)
-
v)
$$\begin{aligned} |\lambda _0-\lambda _n|\le c(\delta _n\delta _n^*)^{1/\varkappa \left( A(\cdot ),\lambda _0\right) } \end{aligned}$$
for all \(\lambda _n\in \sigma \big (A_n(\cdot )\big )\cap {\tilde{\Lambda }}\), whereby \(\varkappa \left( A(\cdot ),\lambda _0\right) \) denotes the maximal length of a Jordan chain of \(A(\cdot )\) at the eigenvalue \(\lambda _0\).
-
vi)
$$\begin{aligned} |\lambda _0-\lambda _n^\mathrm {mean}|\le c\delta _n\delta _n^* \end{aligned}$$
whereby \(\lambda _n^\mathrm {mean}\) is the weighted mean of all the eigenvalues of \(A_n(\cdot )\) in \({\tilde{\Lambda }}\)
$$\begin{aligned} \lambda _n^\mathrm {mean}:=\sum _{\lambda \in \sigma \left( A_n(\cdot )\right) \cap {\tilde{\Lambda }}}\lambda \, \frac{{\text {dim}}G(A_n(\cdot ),\lambda )}{{\text {dim}}G(A(\cdot ),\lambda _0)}. \end{aligned}$$ -
vii)
$$\begin{aligned} \inf _{u_0\in \ker A(\lambda _0)} \Vert u_n-u_0\Vert _X&\le c \Big (|\lambda _n-\lambda _0|+ \max _{\begin{array}{c} u'_0\in \ker A(\lambda _0)\\ \Vert u_0'\Vert _X\le 1 \end{array}} \inf _{u'_n\in X_n} \Vert u'_0-u'_n\Vert _X\Big )\\&\le c\big (c(\delta _n\delta _n^*)^{1/\varkappa \left( A(\cdot ),\lambda _0\right) } + \delta _n\big ) \end{aligned}$$
for all \(\lambda _n\in \sigma \big (A_n(\cdot )\big )\cap {\tilde{\Lambda }}\) and all \(u_n\in \ker A_n(\lambda _n)\) with \(\Vert u_n\Vert _X=1\).
Proof
Proof of i)–iii). The first three claims i)–iii) follow from [26, Theorem 2]. In particular the assumptions of [26, Theorem 2] are
-
(d1)
\(\rho \big (A(\cdot )\big )\) is non-empty,
-
(d2)
\(A_n(\lambda )\) is Fredholm with index zero for each \(\lambda \in \Lambda , n\in \mathbb {N}\),
-
(d3)
\(\big (A_n(\cdot )\big )_{n\in \mathbb {N}}\) is equibounded (i.e. (c1) holds),
-
(d4)
\(\big (A_n(\lambda )\big )_{n\in \mathbb {N}}\) approximates \(A(\lambda )\) for each \(\lambda \in \Lambda \) (i.e. (b1) holds),
-
(d5)
\(\big (A_n(\lambda )\big )_{n\in \mathbb {N}}\) is regular for each \(\lambda \in \Lambda \) (i.e. (b2) holds).
The first two assumptions are also assumptions of this theorem. The second two assumptions are satisfied, because our particular discrete approximation scheme is a Galerkin scheme. The fifth assumptions follows from Theorem 2. To enable a more self-sufficient reading of this article we include the explicit proofs of i)-iii) at this point.
Direct proof of ii). Since \(\lim _{n\rightarrow \infty } \Vert A(\lambda _0)-A(\lambda _n)\Vert _{L(X)}=0\) and by assumption \(A_{n}(\lambda _n)u_{n}=0\), it follows \(\lim _{n\rightarrow \infty } A_{n}(\lambda _0)u_{n}=0\). Due to the regularity of \(A_n(\lambda _0)\), there exist \(u\in X\) and a subsequence \((n(m))_{m\in \mathbb {N}}\) such that \(\lim _{m\rightarrow \infty } \Vert u-u_{n(m)}\Vert _X=0\). Since \(\Vert u_{n(m)}\Vert _X=1\) for each \(m\in \mathbb {N}\), it follows \(\Vert u\Vert _X=1\). Together with the point-wise convergence of \(A_{n(m)}(\lambda _0)\) we deduce
Hence \(A(\lambda _0)u=0\) and \(\Vert u\Vert _X=1\), i.e. \((\lambda _0,u)\) is a normalized eigenpair of \(A(\cdot )\). The very same technique allows to choose for any arbitrary subsequence \((n(m))_{m\in \mathbb {N}}\) a normalized \(u\in X\) and a subsequence \((m(k))_{k\in \mathbb {N}}\) such that \(\lim _{k\rightarrow \infty } u_{n(m(k))}=u\) and \(A(\lambda _0)u=0\). Hence \((u_n)_{n\in \mathbb {N}}\) is a compact sequence its cluster points are normalized eigenelements.
Direct proof of iii). Assume the contrary. Then there exist a sequence \((\lambda _n\in {\tilde{\Lambda }})_{n\in \mathbb {N}}\) and a normalized sequence \((u_n\in X)_{n\in \mathbb {N}}\) with
Since \({\tilde{\Lambda }}\) is compact there exist \(\lambda _0\in {\tilde{\Lambda }}\) and a subsequence \((n(m))_{m\in \mathbb {N}}\) such that \(\lim _{m\rightarrow \infty }\lambda _{n(m)}=\lambda _0\). As \(A_n(\lambda _0)\) is regular there exist normalized \(u\in X\) and a subsequence \((m(k))_{k\in \mathbb {N}}\) such that \(\lim _{k\rightarrow \infty } u_{n(m(k))}=u\). Thus
due to the continuity of \(A(\cdot )\). Hence \(\lambda _0\in {\tilde{\Lambda }}\) is an eigenvalue to \(A(\cdot )\), which is a contradiction to \({\tilde{\Lambda }}\subset \rho \big (A(\cdot )\big )\).
Direct proof of i). Assume the contrary. Then there exist \(\delta >0\) and a subsequence \((n(m))_{m\in \in \mathbb {N}}\) such that \(B_\delta :=\{\lambda \in \mathbb {C}:|\lambda -\lambda _0|<\delta \}\subset \rho \big (A_{n(m)}(\cdot )\big )\). In addition we choose \(\delta \) small enough such that \(\Gamma :=\{\lambda \in \mathbb {C}:|\lambda -\lambda _0|=\delta \}\subset \rho \big (A(\cdot )\big )\). Due to 3) there exist constants \(m_0, c>0\) such that \(\Gamma \subset \rho \big (A_{n(m)}(\cdot )\big )\) and \(\Vert A_{n(m)}^{-1}(\lambda )\Vert _{L(X_{n(m)})}\le c\) for all \(m>m_0\), \(\lambda \in \Gamma \). Since \(A_{n(m)}^{-1}(\cdot )\) is holomorphic in \(B_\delta \) the principle of maximum of modulus yields \(\Vert A_{n(m)}^{-1}(\lambda )\Vert _{L(X_{n(m)})}\le c\) for all \(\lambda \in B_\delta \), \(m>m_0\). Thus for \(u\in \ker A(\lambda _0)\), \(\Vert u\Vert _X=1\) and \(m>m_0\) it follows
and hence \(\lim _{m\rightarrow \infty } \Vert P_{n(m)}u\Vert _X=0\), which is a contradiction to \(\Vert u\Vert _X=1\).
Proof of iv)-vi). The second three claims 3)-3) follow from [26, Theorem 3] and [27, Theorem 2, Theorem 3]. In particular the assumptions of [26, Theorem 3] are in addition to those of [26, Theorem 2] the existence of \(c>0\) and \(r_n\in L(X_n,X)\), \(n\in \mathbb {N}\) such that \(\Vert r_n\Vert _{L(X_n,X)}\le c\) for all \(n\in \mathbb {N}\) and \(\lim _{n\rightarrow \infty } r_nP_n u=u\) for each \(u\in X\). Since our particular discrete approximation scheme is a Galerkin scheme, we can simply choose \(r_n\) to be the inclusion operator \(u_n\mapsto u_n\). [27, Theorem 2] assumes (a1)-(a4), (d1)-(d5) and two mappings \(p_n'\), \(q_n'\) which we can choose as \(P_n\). For [27, Theorem 3] we can again choose \(r_n\) as the inclusion map and \(q_n'=P_n\). The proofs of [26, Theorem 3] and [27, Theorem 2, Theorem 3] are a bit technical and essentially consist of two steps. First one constructs operator functions which act on finite dimensional spaces and which locally have the same spectrum and Jordan chains as \(A(\cdot )\) and \(A_n(\cdot )\). In a second step one proves the desired claims for matrix functions.
Proof of vii). The last claim vii) is essentially Theorem 4.3.7 of [31]. However, [31, Theorem 4.3.7] considers only operator functions \(A(\cdot )\) of a special category, which prohibits its direct application for our purpose. Thus we repeat the proof of [31, Theorem 4.3.7] whereby we achieve some simplification and completion. Without loss of generality we assume that \({\tilde{\Lambda }}\cap \sigma (A_n(\cdot ))\ne \emptyset \) for all \(n\in \mathbb {N}\). Let \((\lambda _n,u_n)_{n\in \mathbb {N}}\) be a sequence of normalized eigenpairs of \(A_n(\cdot )\) with \(\lim _{n\rightarrow \infty }\lambda _n=\lambda \) such that
Note that \(u_n\) exists since \(\{{\tilde{u}}_n:{\tilde{\lambda }}_n\in {\tilde{\Lambda }}\cap \sigma (A_n(\cdot )), {\tilde{u}}_n\in \ker A_n({\tilde{\lambda }}_n)\}\) is finite dimensional for each \(n\in \mathbb {N}\). This explicit choice ensures the independence of the forthcoming constant \(c_0\) on \({\tilde{u}}_n\in \{{\tilde{u}}_n:{\tilde{\lambda }}_n\in {\tilde{\Lambda }}\cap \sigma (A_n(\cdot )), {\tilde{u}}_n\in \ker A_n({\tilde{\lambda }}_n)\}\). We introduce two auxiliary sequences \((v_n\in \ker A(\lambda ))_{n\in \mathbb {N}}\) and \((w_n\in \ker A(\lambda ))_{n\in \mathbb {N}}\) which satisfy
and
Note that \(v_n\) and \(w_n\) exist since \(ker A(\lambda )\) and \(P_n ker A(\lambda )\) are finite dimensional. The introduction of \((w_n)_{n\in \mathbb {N}}\) allows us to estimate
We estimate the two terms in the right hand side of (2) separately.
1. part: \((w_n)_{n\in \mathbb {N}}\) is bounded. It follows from ii) that
Further, we estimate
The first term in the right hand side of (4) tends to zero due to (3). Since \(\ker A(\lambda )\) is finite dimensional it follows that \(\lim _{n\rightarrow \infty }(I-P_n)|_{\ker A(\lambda )}=0\) in operator norm and due to \(v_n\in \ker A(\lambda )\) the second term in the right hand side of (4) tends to zero as well. Hence
and since \((u_n)_{n\in \mathbb {N}}\) is bounded it follows that \((P_nw_n)_{n\in \mathbb {N}}\) is bounded too. Due to \(w_n\in \ker A(\lambda )\) and \(\lim _{n\rightarrow \infty }(I-P_n)|_{\ker A(\lambda )}=0\) in operator norm it follows that \((w_n)_{n\in \mathbb {N}}\) is bounded too.
2. part: estimation of \(\Vert w_n-P_nw_n\Vert _X\). Let \(c_0:=\sup _{n\in \mathbb {N}} \Vert w_n\Vert _X\). Thence
3. part: estimate \(\Vert u_n-P_nw_n\Vert _X \le c_1 \Vert A_n(\lambda _n)(u_n-P_nw_n)\Vert _X\). We show that there exists a constant \(c_1>0\) such that
holds for sufficiently large \(n\in \mathbb {N}\). Assume the contrary, then there exists a subsequence \((n(m))_{m\in \mathbb {N}}\) such that
holds for all \(m\in \mathbb {N}\). From (8) we deduce that
Since \(A_n(\cdot )\) is regular there exist \(y\in \ker A(\lambda )\) and a subsequence \((m(k))_{k\in \mathbb {N}}\) such that
So we conclude with the abbreviation \(\epsilon _k:=\Vert u_{n(m(k))}-P_{n(m(k))}w_{n(m(k))} \Vert _X\)
for \(k\rightarrow \infty \) and hence \(\epsilon _k=0\) for sufficiently large k. However, \(\epsilon _k\ne 0\) due to (8), which is a contradiction. Hence (7) holds.
4. part: estimation of \(\Vert A_n(\lambda _n)(u_n-P_nw_n)\Vert _X\). Due to \(u_n\in \ker A_n(\lambda _n)\) and \(A_n(\cdot )=P_n A(\cdot )\) we can estimate
Since \(A(\cdot )\) is holomorphic and \((\Vert w_n\Vert _X)_{n\in \mathbb {N}}\) is bounded, there exists a constant \(c_2\) such that
Due to \(w_n\in \ker A(\lambda )\) we can estimate the first term in the right hand side of (9) as
5. part: conclusion. Finally the combination of (1), (2), (6), (7), (9), (10) and (11) yield the existence of a constant \(c>0\) such that
for all \(\lambda _n\in \sigma \big (A_n(\cdot )\big )\cap {\tilde{\Lambda }}\) and all \(u_n\in \ker A_n(\lambda _n)\) with \(\Vert u_n\Vert _X=1\). \(\square \)
Remark 1
We remark that the proof of Theorem 3 vii) actually requires only the regularity of \(A_n(\cdot )\) and doesn’t apply the \(T(\cdot )\)-compatibility condition. Further, Theorem 3 assumes implicitly that the operators \(A_n(\lambda )\) are Fredholm with index zero. However, for most applications \(X_n\) will be finite dimensional and hence in these cases this assumption will be trivial.
4 Example of application
In this section we present an example how to apply our theory to the approximation of an electromagnetic eigenvalue problem for a conductive material by means of Nédélec finite elements. Since the related stencil is quadratic in the eigenvalue parameter, the convergence of approximations is not covered by convenient theory such as [2, 7, 9, 13]. To the best of our knowledge the obtained results are new. The applied technique essentially follows [21].
4.1 Maxwell eigenvalue problem for a conductive material
Let \(\Omega \subset \mathbb {R}^3\) be a bounded Lipschitz polyhedron with outward unit vector \(\nu \) and \(\omega \in \mathbb {C}\) be the temporal frequency. Let \(\epsilon \) be the electric permittivity, \(\mu \) be the magnetic permeability and \(\gamma \) be the conductivity of a heterogeneous and anisotropic material, i.e. \(\epsilon ,\mu ^{-1},\gamma \in \big (L^\infty (\Omega )\big )^{3x3}\) are real symmetric matrix functions such that
We note that we label the conductivity with the symbol \(\gamma \) instead of the convenient symbol \(\sigma \) to avoid a conflict with the symbol \(\sigma \big (A(\cdot )\big )\) for the spectrum of an operator function. We denote the hermitian \(L^2(\Omega )\) and \((L^2(\Omega ))^3\) scalar products as \(\langle \cdot ,\cdot \rangle \). Let \(\nabla , {\text {div}}\) and \({{\,\mathrm{curl}\,}}\) be the differential operators defined by
Then we introduce the Hilbert space \(X=H_0({{\,\mathrm{curl}\,}};\Omega )\)
Then we consider the Maxwell eigenvalue problem [29] to find \((\omega ,u)\in \mathbb {C}\times X{\setminus }\{0\}\) such that
Here the frequency \(\omega \) takes the role of the spectral parameter \(\lambda \). The weak formulation of (12) is to find \((\omega ,u)\in \mathbb {C}\times X{\setminus }\{0\}\) such that \(a(\omega ;u,u')=0\) for all \(u'\in X\) with the sesquilinear form
Let \(A(\cdot ):\mathbb {C}\rightarrow L(X)\) be defined by \(\langle A(\omega )u,u'\rangle _X=a(\omega ;u,u')\) for all \(\omega \in \mathbb {C}, u,u'\in X\). Then the operator formulation of the eigenvalue problem is to find \((\omega ,u)\in \mathbb {C}\times X{\setminus }\{0\}\) such that \(A(\omega )u=0\).
4.2 Weak \(T(\cdot )\)-coercivity
In order to construct a suitable \(T(\cdot )\)-operator function we follow e.g. [7] and introduce the orthogonal Helmholtz decomposition \(X=V\oplus ^{\bot _X} W\) with subspaces
and respective orthogonal projections \(P_V\) and \(P_W\). To achieve a more compact presentation we abbreviate \(v:=P_Vu\), \(w:=P_Wu\), \(v':=P_Vu'\) and \(w':=P_Wu'\) for \(u,u'\in X\). A well known and important property of V is its compact embedding into \((L^2(\Omega ))^3\), see, e.g. [35]. Subsequently we define
It easily follows that \(T(\cdot )\) is bijective with inverse \(T^{-1}(\omega ):=P_V-\omega ^2P_W\). With the former notation it holds \(\langle T^*(\omega )A(\omega )u,u'\rangle _X=a(\omega ;v+w,v'-\omega ^{-2}w')\). Then we compute \(T^*(\omega )A(\omega )=A_1+A_2\) with \(A_1,A_2\in L(X)\) defined by
for all \(u,u'\in X\). Note that several terms in this decomposition vanished due to \({\text {div}}w={\text {div}}w'=0\). For \(\omega \in \Lambda :=\mathbb {C}{\setminus }\{-it:t\ge 0\}\) it holds \(\mathfrak {R}(i\omega ^{-1})\ge 0\) and hence
i.e. \(A_1\) is coercive. The former equality holds since due to \({\text {div}}w=0\) we have \(\Vert w\Vert _{(L^2(\Omega ))^3}=\Vert w\Vert _X\) and since V and W are orthogonal it holds \(\Vert v\Vert _X^2+\Vert w\Vert _X^2=\Vert u\Vert _X^2\). On the other hand it follows from the compact embedding \(V\hookrightarrow (L^2(\Omega ))^3\) that \(A_2\) is coercive. Thus \(A(\cdot ):\Lambda \rightarrow L(X)\) is weakly \(T(\cdot )\)-coercive. For \(\omega \in \{-it:t\in [0,\Vert \gamma \Vert _{(L^\infty (\Omega ))^{3\times 3}}/c_\epsilon ]\}\) it is possible that \(\epsilon +i\omega ^{-1}\gamma \) is not uniformly coercive. For such frequencies we cannot expect that \(A(\omega )\) is Fredholm. Further, note that \(A(\omega )\) is already coercive and hence bijective for all \(\omega \in \{it:t>0\}\). Thus the resolvent set of \(A(\cdot )\) is not empty.
4.3 \(T(\cdot )\)-compatible approximation
In order to prove the convergence of discretizations by means of Theorem 3, it remains to discuss the \(T(\cdot )\)-compatibility of approximations. It is well known that naive discretizations of Maxwell eigenvalue problems can lead to erroneous results and the application of special (Nédélec) finite element spaces is necessary. In this section we will show that if the Galerkin spaces \(X_n\) are chosen as Nédélec spaces, then indeed the corresponding approximations are \(T(\cdot )\)-compatible. The research on the underlying theory of Nédélec elements led to many different conditions on the finite element spaces and an overview on their equivalence can be found in [13]. Here we choose the notion of \(L^2\)-uniformly bounded cochain projections [2] to formulate our theory. To this end we introduce the de Rham Complex
and consider for each \(n\in \mathbb {N}\) a subcomplex of (13)
with finite dimensional spaces \(Y^k_n\subset Y^k\), \(k=1,\dots ,4\) such that the corresponding orthogonal projections \(P_{Y_n^k}\), \(k=1,\dots ,4\) converge point-wise to the identity for \(n\rightarrow \infty \). Further, let \(Z^1:=Z^4:=L^2(\Omega )\) and \(Z^2:=Z^3:=(L^2(\Omega ))^3\). We say that the sequence of subcomplexes (14) admits \(L^2\)-uniformly bounded commuting cochain projections \(\pi _n^k\), \(k=1,\dots ,4\), if the following commuting diagram
is satisfied, the projections are surjective and of the form \(\pi _n^k = {\tilde{\pi }}_n^k E_k\) with embedding operators \(E_k\in L\big (Y^k,Z^k\big )\) and projections \({\tilde{\pi }}_n^k \in L\big (Z^k,Y_n^k\big )\), \(k=1,\dots ,4\), such that \(c_{{\tilde{\pi }}}:=\sup _{n\in \mathbb {N}, k=1,\dots ,4} \big \{ \Vert {\tilde{\pi }}_n^k\Vert _{L(Z^k)},\, \big \} < \infty \). We refer to [2] for the systematic construction of such finite element spaces \((Y_n^k)_{n\in \mathbb {N}}\) and the proof of existence of respective projections \((\pi ^k_n)_{n\in \mathbb {N}}\). In this context the index n corresponds to the sequence of decreasing mesh width parameters \(h_n\rightarrow 0+\). We note that \({\tilde{\pi }}_n^k\rightarrow I\) point-wise in \(Z^k\) for \(n\rightarrow \infty \). Indeed, for \(u\in Y^k\) it follows
Further, \(\pi _n^2P_V\) converges to \(P_V\) in discrete norm. Indeed, we for \(u\in X_n\)
because \({{\,\mathrm{curl}\,}}P_W =0\), \(P_V+P_W=I\) and \({{\,\mathrm{curl}\,}}u \in Y_n^3\). Thus
Since \(E_2P_V\) is compact and \(\lim _{n\rightarrow \infty }{\tilde{\pi }}_n^2=I\) point-wise in \((L^2(\Omega ))^3)\) it follows \(\lim _{n\rightarrow \infty } \Vert P_V-\pi _n^2P_V\Vert _n=0\). Due to
it follows for \(T_n(\omega ):=\pi _n^2T(\omega )|_{X_n}\) with the triangle inequality that
Hence Theorem 3 can be applied for Nédélec approximations to the eigenvalue problem (12) and the respective convergence properties i)-vii) hold. Note that the obtained quadratic matrix eigenvalue problem can subsequently be linearized with the auxiliary variable \({\tilde{u}}_n:=\omega u_n\) and the resulting linear eigenvalue can be solved with an Arnoldi algorithm.
References
Anselone, P.M., Treuden, M.L.: Regular operator approximation theory. Pacific J. Math. 120(2), 257–268 (1985). http://projecteuclid.org/euclid.pjm/1102703407
Arnold, D.N., Falk, R.S., Winther, R.: Finite element exterior calculus: from Hodge theory to numerical stability. Bull. Am. Math. Soc. (N.S.) 47(2), 281–354 (2010). https://doi.org/10.1090/S0273-0979-10-01278-4
Babuška, I., Osborn, J.: Eigenvalue problems. In: Finite element methods (part 1), handbook of numerical analysis, vol. 2, pp. 641–787. Elsevier, Amsterdam (1991)
Boffi, D.: Finite element approximation of eigenvalue problems. Acta Numer. 19, 1–120 (2010). https://doi.org/10.1017/S0962492910000012
Bonnet-BenDhia, A.S., Carvalho, C., Ciarlet, P.: Mesh requirements for the finite element approximation of problems with sign-changing coefficients. Numer. Math. 138(4), 801–838 (2018). https://doi.org/10.1007/s00211-017-0923-5
Bonnet-BenDhia, A.S., Ciarlet, P., Zwölf, C.M.: Time harmonic wave diffraction problems in materials with sign-shifting coefficients. J. Comput. Appl. Math. 234(6), 1912–1919 (2010). https://doi.org/10.1016/j.cam.2009.08.041
Buffa, A.: Remarks on the discretization of some noncoercive operator with applications to heterogeneous maxwell equations. SIAM J. Numer. Anal. 43(1), 1–18 (2005). https://doi.org/10.1137/S003614290342385X
Buffa, A., Costabel, M., Schwab, C.: Boundary element methods for Maxwell’s equations on non-smooth domains. Numer. Math. 92(4), 679–710 (2002). https://doi.org/10.1007/s002110100372
Caorsi, S., Fernandes, P., Raffetto, M.: On the convergence of Galerkin finite element approximations of electromagnetic eigenproblems. SIAM J. Numer. Anal 38(2), 580–607 (2000). https://doi.org/10.1137/S0036142999357506
Carvalho, C.: Etude mathématique et numérique de structures plasmonique avec des coins. Ph.D. thesis (2015)
Chesnel, L.: Interior transmission eigenvalue problem for maxwell’s equations: the T-coercivity as an alternative approach. Inverse Probl. (2012). https://doi.org/10.1088/0266-5611/28/6/065005
Chesnel, L.: Étude de quelques problèmes de transmission avec changement de signe. application aux métamatériaux. Ph.D. thesis (2012)
Christiansen, S.H., Winther, R.: On variational eigenvalue approximation of semidefinite operators. IMA J. Numer. Anal. 33(1), 164–189 (2013). https://doi.org/10.1093/imanum/drs002
Descloux, J., Nassif, N., Rappaz, J.: On spectral approximation. part 1. the problem of convergence. ESAIM: Mathematical Modelling and Numerical Analysis - Modélisation Mathématique et Analyse Numérique 12(2), 97–112 (1978). http://eudml.org/doc/193319
Descloux, J., Nassif, N., Rappaz, J.: On spectral approximation. part 2. error estimates for the galerkin method. ESAIM: Mathematical Modelling and Numerical Analysis - Modélisation Mathématique et Analyse Numérique 12(2), 113–119 (1978). http://eudml.org/doc/193315
Gohberg, I., Goldberg, S., Kaashoek, M.A.: Classes of linear operators. Vol. I, Operator Theory: Advances and Applications, vol. 49. Birkhäuser Verlag, Basel (1990). https://doi.org/10.1007/978-3-0348-7509-7
Gohberg, I., Leiterer, J.: Holomorphic operator functions of one variable and applications, Operator Theory: Advances and Applications, vol. 192. Birkhäuser Verlag, Basel (2009). https://doi.org/10.1007/978-3-0346-0126-9. Methods from complex analysis in several variables
Grigorieff, R.D.: Zur Theorie linearer approximationsregulärer Operatoren. I. II. Math. Nachr. 55, 233–249 (1973)
Grigorieff, R.D., Jeggle, H.: Approximation von Eigenwertproblemen bei nichtlinearer Parameterabhängigkeit. Manuscripta Mathematica 10(3), 245–271 (1973). https://doi.org/10.1007/BF01332768
Halla, M.: Analysis of radial complex scaling methods for scalar resonance problems in open systems. Ph.D. thesis, Technische Universität Wien (2019). https://repositum.tuwien.ac.at/urn:nbn:at:at-ubtuw:1-131893
Halla, M.: Electromagnetic Steklov eigenvalues approximation analysis. ESAIM: MAN 55(1), 57–76 (2021)
Halla, M.: On the approximation of dispersive electromagnetic eigenvalue problems in 2d. preprint (2021). https://arxiv.org/abs/2104.09318
Halla, M.: Analysis of radial complex scaling methods: scalar resonance problems. SIAM J. Numer. Anal. (2021 accepted)
Hohage, T., Nannen, L.: Convergence of infinite element methods for scalar waveguide problems. BIT Numer. Math. 55(1), 215–254 (2015). https://doi.org/10.1007/s10543-014-0525-x
Jeggle, H., Wendland, W.: On the discrete approximation of eigenvalue problems with holomorphic parameter dependence. Proc. Roy. Soc. Edinburgh Sect. A 78(1–2), 1–29 (1977)
Karma, O.: Approximation in eigenvalue problems for holomorphic Fredholm operator functions. I. Numer. Funct. Anal. Optim. 17(3–4), 365–387 (1996). https://doi.org/10.1080/01630569608816699
Karma, O.: Approximation in eigenvalue problems for holomorphic Fredholm operator functions. II. (Convergence rate). Numer. Funct. Anal. Optim. 17(3–4), 389–408 (1996). https://doi.org/10.1080/01630569608816700
Kozlov, V., Mazya, V.: Differential equations with operator coefficients with applications to boundary value problems for partial differential equations Springer Monographs in Mathematics. Springer-Verlag, Berlin (1999)
Monk, P.: Finite element methods for Maxwell’s equations. numerical mathematics and scientific computation. Oxford University Press, New York (2003)
Stummel, F.: Diskrete Konvergenz linearer Operatoren. I. Math. Ann. 190, 45–92 (1970)
Unger, G.: Analysis of boundary element methods for laplacian eigenvalue problems. Ph.D. thesis, TU Graz, Graz, Austria (2009). https://www.applied.math.tugraz.at/~gunger/diss.pdf
Unger, G.: Convergence analysis of a galerkin boundary element method for electromagnetic eigenvalue problems. SN PDE (2020 accepted). https://www.numerik.math.tugraz.at/~gunger/ConvergenceMaxwellBEM.pdf
Vainikko, G.: Funktionalanalysis der Diskretisierungsmethoden. B. G. Teubner Verlag, Leipzig (1976). Mit Englischen und Russischen Zusammenfassungen, Teubner-Texte zur Mathematik
Vaĭnikko, G.M., Karma, O.O.: The rate of convergence of approximation methods for an eigenvalue problem in which the parameter occurs nonlinearly. Ž. Vyčisl. Mat. i Mat. Fiz. 14, 1393–1408, 1628 (1974)
Weber, C.: A local compactness theorem for Maxwell’s equations. Math. Methods Appl. Sci. 2(1), 12–25 (1980). https://doi.org/10.1002/mma.1670020103
Funding
Open Access funding enabled and organized by Projekt DEAL.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Halla, M. Galerkin approximation of holomorphic eigenvalue problems: weak T-coercivity and T-compatibility. Numer. Math. 148, 387–407 (2021). https://doi.org/10.1007/s00211-021-01205-8
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00211-021-01205-8