Abstract
This paper considers computing partial eigenpairs of differential eigenvalue problems (DEPs) such that eigenvalues are in a certain region on the complex plane. Recently, based on a “solvethendiscretize” paradigm, an operator analogue of the FEAST method has been proposed for DEPs without discretization of the coefficient operators. Compared to conventional “discretizethensolve” approaches that discretize the operators and solve the resulting matrix problem, the operator analogue of FEAST exhibits much higher accuracy; however, it involves solving a large number of ordinary differential equations (ODEs). In this paper, to reduce the computational costs, we propose operation analogues of Sakurai–Sugiuratype complex momentbased eigensolvers for DEPs using higherorder complex moments and analyze the error bound of the proposed methods. We show that the number of ODEs to be solved can be reduced by a factor of the degree of complex moments without degrading accuracy, which is verified by numerical results. Numerical results demonstrate that the proposed methods are over five times faster compared with the operator analogue of FEAST for several DEPs while maintaining almost the same high accuracy. This study is expected to promote the “solvethendiscretize” paradigm for solving DEPs and contribute to faster and more accurate solutions in realworld applications.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
This paper considers solving differential eigenvalue problems (DEPs)
with boundary conditions, where \(\mathcal {A}\) and \({\mathscr{B}}\) are linear, ordinary differential operators acting on functions from a Hilbert space \({\mathscr{H}}\) and Ω is a prescribed simply connected open set. This type of problems appears in various fields such as physics [1, 2] and materials science [3,4,5]. Here, λ_{i} and u_{i} is an eigenvalue and the corresponding eigenfunction, respectively. We assume that the boundary Γ of Ω is a rectifiable, simple closed curve and that the spectrum of (1) is discrete and does not intersect Γ, while only m finite eigenvalues counting multiplicities are in Ω. We also assume that there are eigenfunctions of (1) that form a basis for the invariant subspace associated with λ_{i} ∈Ω.
A conventional way to solve (1) is to discretize the operators \(\mathcal {A}\) and \({\mathscr{B}}\) and solve the resulting matrix eigenvalue problem using some matrix eigensolver, e.g., the QZ and Krylov subspace methods [6]. Fine discretization can reduce discretization error but lead to the formation of large matrix eigenvalue problems. Owing to parallel efficiency, complex momentbased eigensolvers are practical choices for large eigenvalue problems such as Sakurai–Sugiura’ s approach [7] and FEAST eigensolvers [1]. This class of eigensolvers constructs an approximation of the target invariant subspace using a contour integral and computes an approximation of the target eigenpairs using a projection onto the subspace. Because of the high efficiency of parallel computation of the contour integral [2, 3], which is the most timeconsuming part, complex momentbased eigensolvers have attracted considerable attention.
In contrast to the above “discretizethensolve” paradigm, a “solvethendiscretize” paradigm emerged, motivated by mathematical software Chebfun [8]. Chebfun enables highly adaptive computation with operators and functions in the same manner as matrices and functions. This paradigm has extended numerical linear algebra techniques in finite dimensional spaces to infinitedimensional spaces [9,10,11,12,13,14]. Under the circumstances, an operator analogue of FEAST was recently developed [15] for solving (1) and dealing with operators \(\mathcal {A}\) and \({\mathscr{B}}\) without their discretization.^{Footnote 1} On one hand, the operator analogue of FEAST exhibits much higher accuracy than methods based on the traditional “discretizethensolve” paradigm. On the other hand, a large number of ordinary differential equations (ODEs) must be solved for the construction of invariant subspaces, which is computationally expensive, although the method can be efficiently parallelized.
In this paper, we propose operation analogues of Sakurai–Sugiura’s approach for DEPs (1) in the “solvethendiscretize” paradigm. The difference between the operator analogue of FEAST and the proposed methods lies in the order of complex moments used: the operator analogue of FEAST used only complex moments of order zero, whereas the proposed methods use complex moments of higher order. The difference enables the proposed methods to reduce the number of ODEs to be solved by a factor of the degree of complex moments without degrading accuracy. The proposed methods can be extended to higher dimensions in a straightforward manner for simple geometries.
The remainder of this paper is organized as follows. Section 2 briefly introduces the complex momentbased matrix eigensolvers. In Section 3, we propose operation analogues of Sakurai–Sugiura’s approach for DEPs (1). We also introduce a subspace iteration technique and analyze an error bound. Numerical experiments are reported in Section 4. The paper concludes with Section 5.
We use the following notations for quasimatrices. Let \(V = [v_{1}, v_{2}, \dots , v_{L}]\), \(W = [w_{1}, w_{2}, {\dots } ,w_{L}]\): \(\mathbb {C}^{L} \rightarrow {\mathscr{H}}\) be quasimatrices, whose columns are functions defined on an interval [a, b], \(a, b \in \mathbb {R}\). Then, we define the range of V by \({\mathscr{R}}(V) = \{ y \in {\mathscr{H}} \mid y = V {\boldsymbol {x}}, {\boldsymbol {x}} \in \mathbb {C}^{L} \}\). In addition, the L × L matrix X, whose (i, j) element is \(X_{ij} = (v_{i}, w_{j})_{{\mathscr{H}}}\), is expressed as X = V^{H}W. Here, V^{H} is the conjugate transpose of a quasimatrix V such that its rows are the complex conjugates of functions \(v_{1}, v_{2}, {\dots } , v_{L}\).
2 Complex momentbased matrix eigensolvers
The complex momentbased eigensolvers proposed by Sakurai and Sugiura [7] are intended for solving matrix generalized eigenvalue problems:
where zB − A is nonsingular in a boundary Γ of the target region Ω. These eigensolvers use Cauchy’s integral formula to form complex moments. Complex moments can extract the target eigenpairs from random vectors or matrices.
We denote the k th order complex moment by
where π is the circular constant, i is the imaginary unit, and Γ is a positively oriented closed Jordan curve of which Ω is the interior. Then, the complex moment M_{k} applied to a matrix \(V \in \mathbb {C}^{n \times L}\) serves as a filter that stops undesired eigencomponents in the column vectors of V from passing through. To achieve this role of a complex moment, we introduce a transformation matrix \(S \in \mathbb {C}^{n \times LM}\)
where \(V \in \mathbb {C}^{n \times L}\) and M − 1 is the largest order of complex moments. Note that L and M are regarded as parameters. The special case M = 1 in S reduces to FEAST [1, (3)]. Thus, the range \({{\mathscr{R}}}(S)\) of S forms the eigenspace of interest (see, e.g., [16, Theorem 1]).
Practical algorithms of the complex momentbased eigensolvers approximate the contour integral of the transformation matrix \(\widehat {S}_{k} \simeq S_{k}\) of (2) using a quadrature rule
where \(z_{j}, \omega _{j} \in \mathbb {C}\) \((j = 1, 2, \dots , N)\) are quadrature points and the corresponding weights, respectively.
The most timeconsuming part of complex momentbased eigensolvers involves solving linear systems at each quadrature point. These linear systems can be independently solved so that the eigensolvers have good scalability, as demonstrated in [2, 3]. For this reason, complex momentbased eigensolvers have attracted considerable attention, particularly in physics [1, 2], materials science [3,4,5], power systems [17], data science [18], and so on. Currently, there are several methods, including direct extensions of Sakurai and Sugiura’s approach [19,20,21,22,23,24,25], the FEAST eigensolver [1] developed by Polizzi, and its improvements [2, 26, 27]. We refer to the study by [23] and the references therein, for relationship among typical complex momentbased methods: the methods using the Rayleigh–Ritz procedure [19, 21], the methods using Hankel matrices [7, 20], the method using the communication avoiding Arnoldi procedure [24], FEAST eigensolver [1], and so on.
3 Complex momentbased methods
In the “solvethendiscretize” paradigm, an operator analogue of the FEAST method was proposed [15] for solving (1) without requiring discretization of the operators \(\mathcal {A}\) and \({\mathscr{B}}\). The operator analogue of FEAST (contFEAST) is a simple extension of the matrix FEAST eigensolver and is based on an accelerated subspace iteration only with complex moments of order zero (see Algorithm 1). In each iteration, contFEAST requires solving a large number of ODEs to construct a subspace. In this study, to reduce computational costs, we propose operator analogues of Sakurai–Sugiuratype complex momentbased eigensolvers: contSSRR, contSSHankel, and contSSCAA using complex moments of higher order.
3.1 Complex moment subspace and its properties
For the differential eigenvalue problem (1), spectral projectors \(\mathcal {P}_{i}\) and \(\mathcal {P}_{\Omega }\) associated with a finite eigenvalue λ_{i} and the target eigenvalues λ_{i} ∈Ω are defined as
respectively, where Γ_{i} is a positively oriented closed Jordan curve in which λ_{i} lies and contour paths Γ_{i} and Γ_{j} do not intersect each other for i≠j (see [28, pp.178–179] for the case of \({\mathscr{B}}=\mathcal {I}\)). Here, spectral projectors \(\mathcal {P}_{i}\) satisfy
where δ_{ij} is the Kronecker delta.
Analogously to the complex momentbased eigensolvers for matrix eigenvalue problems, we define the k th order complex moment as
and the transformation quasimatrix as
for \(k = 0, 1, \dots , M1\), where M − 1 is the highest order of complex moments and \(V: \mathbb {C}^{L} \rightarrow {\mathscr{H}}\) is a quasimatrix. Here, L is a parameter. Note the identity \(\mathcal {P}_{i} = {\mathscr{M}}_{0}\) for Γ = Γ_{i}. Then, the range \({\mathscr{R}}(S)\) has the following properties.
Theorem 1
The columns of S defined in (5) form a basis of the target eigenspace \(\mathcal {X}_{\Omega }\) corresponding to Ω, i.e.,
if rank(S) = m, where m is the number of eigenvalues, counting multiplicity, in Ω of (1).
Proof
Cauchy’s integral formula shows
Therefore, from the definitions of S and S_{k}, the quasimatrix S can be written as
which provides (6) if rank(S) = m. □
Remark 1
Theorem 1 shows that the target eigenpairs of (1) can be obtained by using a projection method onto \({\mathscr{R}}(S)\).
Theorem 2
Let S_{0} and S be defined as in (5). Then, the range \({\mathscr{R}}(S)\) and the block Krylov subspace
are the same, i.e.,
where
Here, \(\mathcal {P}_{i}\) is defined in (3). Moreover, the eigenvalue problem of linear operator \(\mathcal {C}\)
has the same finite eigenpairs as \(\mathcal {A}u_{i} = \lambda _{i}{\mathscr{B}}u_{i}\).
Proof
The quasimatrix S_{k} is written as
which provides (7). Hence, the eigenspace of \(\mathcal {C}\) and that of (1) are the same. □
Remark 2
Theorem 2 shows that several techniques for block Krylov subspace can be used to form \({\mathscr{R}}(S)\) and the target eigenpairs of (1) can be obtained by solving (8).
Theorems 1 and 2 are used to derive methods in Section 3.2 and provide an error bound in Section 3.3.
3.2 Derivations of methods
Using Theorems 1 and 2, based on the complex momentbased eigensolvers, SSRR, SSHankel, and SSCAA, we develop complex momentbased differential eigensolvers for solving (1) without the discretization of operators \(\mathcal {A}\) and \({\mathscr{B}}\). The proposed methods are projection methods based on \({\mathscr{R}}(S)\), which is a larger subspace than \({\mathscr{R}}(S_{0})\) used in contFEAST (Algorithm 1).
In practice, we numerically deal with operators, functions, and the contour integrals. The contour integral in (4) is approximated using the quadrature rule:
where \(z_{j}, \omega _{j} \in \mathbb {C}\) \((j = 1, 2, \dots , N)\) are quadrature points and the corresponding weights, respectively. As well as contFEAST, we avoid discretizing the operators, but we construct polynomial approximations on the basis of the invariant subspace by approximately solving ODEs of the form
with boundary conditions. Note that the number of ODEs to be solved does not depend on the degree of complex moments M.
For real operators \(\mathcal {A}\) and \({\mathscr{B}}\), if quadrature points and the corresponding weights are set symmetric about the real axis, \((z_{j}, \omega _{j}) = (\overline {z}_{j+N/2}, \overline {\omega }_{j+N/2}), j = 1,2,\dots , N/2\), we can halve the number of ODEs to be solved as follows:
As another efficient computation technique for real selfadjoint problems, we can avoid complex ODEs using the real rational filtering technique [29] for matrix eigenvalue problems. Using the real rational filtering technique, quasimatrix S is approximated by (9) with the N Chebyshev points of the first kind and the corresponding barycentric weights,
for \(j = 1,2, \dots , N\), where γ and ρ are the center and radius of the target interval. Note that \(z_{j}, \omega _{j} \in \mathbb {R}\) for \(j = 1,2, \dots , N\).
3.2.1 ContSSRR method
An operator analogue of the complex momentbased method using the Rayleigh–Ritz procedure for matrix eigenvalue problems [19, 21] is presented. Theorem 1 shows that the target eigenpairs of (1) can be obtained by a Rayleigh–Ritz procedure based on \({\mathscr{R}}(S)\), i.e.,
where (λ_{i},u_{i}) = (𝜃_{i},St_{i}). We approximate this Rayleigh–Ritz procedure using an \({\mathscr{H}}\)orthonormal basis of the approximated subspace \({\mathscr{R}}(\widehat {S})\). Here, to reduce computational costs and improve numerical stability, we use a lowrank approximation of quasimatrix \(\widehat {S}\) based on its truncated singular value decomposition (TSVD) [10], i.e.,
where \({\Sigma }_{\text {S1}} \in \mathbb {R}^{d \times d}\) is a diagonal matrix whose diagonal entries are the d largest singular values such that σ_{d}/σ_{1} ≥ δ ≥ σ_{d+ 1}/σ_{1} \((\sigma _{i} \geq \sigma _{i+1}, i = 1, 2, \dots , d)\) and \(U_{\text {S1}}: \mathbb {C}^{d} \rightarrow {\mathscr{H}}\) and \(W_{\text {S1}} \in \mathbb {C}^{LM \times d}\) are columnorthonormal (quasi)matrices corresponding to the left and right singular vectors, respectively.
Thus, the target problem (1) is reduced to a ddimensional matrix generalized eigenvalue problem
where the approximated eigenpairs are computed as \((\widehat {\lambda }_{i}, \widehat {u}_{i}) = (\theta _{i}, U_{\text {S1}} {\boldsymbol {t}}_{i})\). The procedure of the contSSRR method is summarized in Algorithm 2.
3.2.2 ContSSHankel method
An operator analogue of the complex momentbased method using Hankel matrices for matrix eigenvalue problems [7, 20] is presented. Let \(\mu _{k} \in \mathbb {C}^{L \times L}\) be a reduced complex moment of order k defined as
with \(\widetilde {V}: \mathbb {C}^{L} \rightarrow {\mathscr{H}}\). We also define block Hankel matrices
which have the following property.
Theorem 3
If rank(H_{M}) = rank(HM<) = m, where m is the number of eigenvalues in Ω of (1), then the nonsingular part of a matrix pencil zH_{M} − HM< and \(z  \mathcal {P}_{\Omega }\) have the same spectrum, where \(\mathcal {P}_{\Omega }\) is defined in (3).
Proof
The complex moment μ_{k} can be written as
where \(\mathcal {C}\) is defined in Theorem 2. Letting
the block Hankel matrices are written as
which proves Theorem 3. □
Theorem 3 shows that the target eigenpairs of (1) can be computed via a matrix eigenvalue problem:
Note that from the equivalence
this approach can be regarded as a Petrov–Galerkintype projection for (8), which has the same finite eigenpairs as \(\mathcal {A}u_{i} = \lambda _{i}{\mathscr{B}}u_{i}\). In practice, block Hankel matrices HM< and H_{M} are approximated by block Hankel matrices \(\widehat {H}_{M}^<\) and \(\widehat {H}_{M}\) whose block (i, j) entries are \(\widehat {\mu }_{i+j+1}\) and \(\widehat {\mu }_{i+j}\), respectively, where
for \(k = 1, 2, \dots , 2M1\). To reduce computational costs and improve numerical stability, we use a lowrank approximation of \(\widehat {H}_{M}\) based on TSVD, i.e.,
where \({\Sigma }_{\text {H1}} \in \mathbb {R}^{d \times d}\) is a diagonal matrix whose diagonal entries are the d largest singular values such that σ_{d}/σ_{1} ≥ δ ≥ σ_{d+ 1}/σ_{1} \((\sigma _{i} \geq \sigma _{i+1}, i = 1, 2, \dots , d)\) and \(U_{\text {H1}}, W_{\text {H1}} \in \mathbb {C}^{LM \times d}\) are columnorthonormal matrices corresponding to the left and right singular vectors, respectively.
Then, the target problem (1) is reduced to a ddimensional standard matrix eigenvalue problem of the form
where the approximated eigenpairs can be computed as \((\widehat {\lambda }_{i}, \widehat {u}_{i}) = (\theta _{i}, \widehat {S} W_{\text {H1}} {\Sigma }_{\text {H1}}^{1} {\boldsymbol {t}}_{i})\). The procedure of the contSSHankel method is summarized in Algorithm 3.
3.2.3 ContSSCAA method
An operator analogue of the complex momentbased method using the communication avoiding Arnoldi procedure for matrix eigenvalue problems [24] is presented. Theorem 2 shows that the target eigenpairs of (1) can be obtained by using a block Arnoldi method with \({{\mathscr{R}}}(S) = {\mathscr{K}}_{M}(\mathcal {C},S_{0})\) for (8), which has the same finite eigenpairs as \(\mathcal {A}u_{i} = \lambda _{i}{\mathscr{B}}u_{i}\). In this algorithm, a quasimatrix \(Q : \mathbb {C}^{LM} \rightarrow {\mathscr{H}}\) whose columns form an orthonormal basis of \({{\mathscr{R}}}(S) = {\mathscr{K}}_{M}(\mathcal {C},S_{0})\) and a block Hessenberg matrix \(T_{M} = Q^{\textsf {H}} \mathcal {C} Q\) are constructed, and the target eigenpairs are computed by solving the standard matrix eigenvalue problem
Therefore, we have (λ_{i},u_{i}) = (𝜃_{i},Qt_{i}).
Further, we consider using a block version of the communicationavoiding Arnoldi procedure [30]. Let \(S_{+}= [S_{0}, S_{1}, \dots , S_{M}]: \mathbb {C}^{L(M+1)} \rightarrow {\mathscr{H}}\) be a quasimatrix. From Theorem 2, we have
Here, based on the concept of a block version of the communicationavoiding Arnoldi procedure, using the QR factorizations
where Q = Q_{+}(:,1 : LM),R = R_{+}(1 : LM,1 : LM), the block Hessenberg matrix T_{M} is obtained by
In practice, we approximate the block Hessenberg matrix T_{M} by
where
are the QR factorizations of \(\widehat {S}_{+}\) and \(\widehat {S}\), respectively, \(\widehat {Q} = \widehat {Q}_{+}(:,1\!:\!LM)\), and \(\widehat {R}=\widehat {R}_{+}(1\!:\!LM,1\!:\!LM)\), and use a lowrank approximation of \(\widehat {S}\) based on TSVD, i.e.,
where \({\Sigma }_{\text {R1}} \in \mathbb {R}^{d \times d}\) is a diagonal matrix whose diagonal entries are the d largest singular values σ_{d}/σ_{1} ≥ δ ≥ σ_{d+ 1}/σ_{1} \((\sigma _{i} \geq \sigma _{i+1}, i = 1,2, \dots , d)\) and \(U_{\text {R1}}, W_{\text {R1}} \in \mathbb {C}^{LM \times d}\) are columnorthonormal matrices corresponding to the left and right singular vectors of \(\widehat {R}\), respectively.
Then, the target problem (1) is reduced to a ddimensional matrix standard eigenvalue problem of the form
The approximate eigenpairs are obtained as \((\widehat {\lambda }_{i},\widehat {u}_{i}) = (\theta _{i}, \widehat {Q} U_{\text {R1}} {\boldsymbol {t}}_{i})\). The coefficient matrix \(U_{\text {R1}}^{\textsf H} \widehat {T}_{M} U_{\text {R1}}\) is efficiently computed by
The procedure of the contSSCAA method is summarized in Algorithm 4.
3.3 Subspace iteration and error bound
We consider improving the accuracy of the eigenpairs via a subspace iteration technique, as in the matrix version of complex momentbased eigensolvers. We construct \(\widehat {S}_{0}^{(\ell 1)}\) via the following iteration step:
with the initial quasimatrix \(\widehat {S}_{0}^{(0)} = V\). Then, instead of \(\widehat {S}\) in each method, we use \(\widehat {S}^{(\ell )}\) constructed from \(\widehat {S}_{0}^{(\ell 1)}\) by
The orthonormalization of the columns of \(\widehat {S}_{0}^{(\nu )}\) in each iteration may improve the numerical stability.
Now, we analyze the error bound of the proposed methods with the subspace iteration technique as introduced in (13) and (14). We assume that \({\mathscr{B}}\) is invertible and all the eigenvalues are isolated. Then, we have
where \(\mathcal {P}_{i}\) is a spectral projector associated with λ_{i} s.t. \(\mathcal {P}_{i} \mathcal {P}_{j} = \delta _{ij} \mathcal {P}_{i}\) [28, VII Section 6]. We also assume that the numerical quadrature satisfies
Under the above assumptions, for \(k = 0, 1, \dots , M1\), the quasimatrix \(\widehat {S}_{k}^{(\ell )}\) can be written as
where
From the definitions of \(\mathcal {C}\) and \(\mathcal {F}\), these operators are commutative, \(\mathcal {C}\mathcal {F} = \mathcal {F}\mathcal {C}\). Therefore, we have
Here, f_{N}(λ_{i}) is called the filter function; it is used for error analyses of complex momentbased matrix eigensolvers [16, 23, 26, 27]. Figure 1 shows the magnitude of the filter function \(\lvert f_{N}(\lambda _{i}) \rvert \) for the Npoint trapezoidal rule with N = 16,32, and 64 for the unit circle region Ω. Here, note that the oscillations at \(\lvert f_{N}(\lambda ) \rvert \approx 10^{16}\) are due to roundoff errors. The filter function has \(\lvert f_{N}(\lambda ) \rvert \approx 1\) inside Ω, \(\lvert f_{N}(\lambda ) \rvert \approx 0\) far from Ω, and \(0 < \lvert f_{N}(\lambda ) \rvert < 1\) outside but near the region. Therefore, \(\mathcal {F}\) is a bounded linear operator.
Applying [15, Theorem 5.1] to (15) under the above assumptions, we have the following theorem for an error bound of the proposed methods.
Theorem 4
Let (λ_{i},u_{i}) be exact finite eigenpairs of the differential eigenvalue problem \(\mathcal {A} u_{i} = \lambda _{i} {\mathscr{B}} u_{i}, i = 1, 2, \dots , LM\). Assume that the filter function f_{N}(λ_{i}) is ordered by decreasing magnitude \(\lvert f_{N}(\lambda _{i}) \rvert \geq \lvert f_{N}(\lambda _{i+1}) \rvert \). We define as an orthogonal projector onto the subspaces \({\mathscr{R}}(\widehat {S}^{(\ell )} )\) and the spectral projector with an invariant subspace span{u_{1},u_{2},…,u_{LM}} by \(\mathcal {P}^{(\ell )}\) and \(\mathcal {P}_{LM}\), respectively. Assume that \(\mathcal {P}_{LM} \widehat {S}^{(0)}\) has full rank, where \(\widehat {S}_{0}\) is defined in (14). Then, for each eigenfunction \(u_{i}, i = 1, 2, \dots \), LM, there exists a unique function \(s_{i} \in \mathcal {K}_{M}^{\square }(\mathcal {C},V)\) such that \(\mathcal {P}_{LM} s_{i} = u_{i}\). Thus, the following inequality is satisfied:
where α is a constant and \(\beta _{i} = \ u_{i}  s_{i} \_{{\mathscr{H}}}\).
Theorem 4 indicates that, using a sufficiently large number of columns LM in the transformation quasimatrix \(\widehat {S}\) such that \(\lvert f_{N}(\lambda _{LM+1}) \rvert ^{\ell } \approx 0\), the proposed methods achieve high accuracy for the target eigenpairs even if N is small and some eigenvalues exist outside but near the region.
3.4 Summary and advantages over existing methods
We summarize the proposed methods and present their advantages over existing methods.
3.4.1 Summary of the proposed methods
ContSSRR is a Rayleigh–Ritztype projection method that explicitly solves (1); on the other hand, contSSHankel and contSSCAA are a Petrov–Galerkintype projection method and block Arnoldi method that implicitly solve (8), respectively. If the computational cost for explicit projection, i.e., \(U_{\text {S1}}^{\textsf H} \mathcal {A} U_{\text {S1}}\) and \(U_{\text {S1}}^{\textsf H} {\mathscr{B}} U_{\text {S1}}\) in contSSRR, is large, contSSHankel and contSSCAA can be more efficient than contSSRR.
Orthogonalization of basis functions is required for contSSRR (step 3 of Algorithm 2) and contSSCAA(step 5 of Algorithm 4) for accuracy but not performed in contSSHankel. This is the advantage of contSSHankel regarding computational costs over other methods. In addition, this is advantageous for contSSHankel when applied to DEPs over a domain for which it is difficult to construct accurate orthonormal bases in such as triangles and tetrahedra domains.
As well as contFEAST, since solving LN ODEs (10) is the mosttime consuming part of the proposed methods and is fully parallelizable, the proposed methods can be efficiently parallelized.
3.4.2 Advantages over contFEAST
ContFEAST is a subspace iteration method based on the L dimensional subspace \({\mathscr{R}}(\widehat {S}_{0})\) for (1). Instead, the proposed methods are projection methods based on the LM dimensional subspace \({\mathscr{R}}(\widehat {S})\). From Theorem 4, we can also observe that the proposed methods using higherorder complex moments achieve higher accuracy than contFEAST, since \( \lvert f_{N}(\lambda _{LM+1}) \rvert < \lvert f_{N}(\lambda _{L+1})\rvert \). In other words, the proposed methods can use a smaller number of initial functions L than contFEAST to achieve almost the same accuracy. Since the number of ODEs to solve is LN in each iteration, the reduction of L drastically reduces the computational costs.
Therefore, the proposed methods exhibit smaller elapsed time than contFEAST, while maintaining almost the same high accuracy, as experimentally verified in Section 4.
3.4.3 Advantages over complex momentbased matrix eigensolvers
Methods using a “solvethendiscretize” approach, including the proposed methods and contFEAST, automatically preserves the normality or selfadjointness of the problems with respect to a relevant Hilbert space \({\mathscr{H}}\). In addition, the stability analysis in [15] shows that the sensitivity of the eigenvalues is preserved by Rayleigh–Ritztype projection methods with an \({\mathscr{H}}\)orthonormal basis for selfadjoint DEPs, but can be increased by methods using a “discretizethensolve” approach. As well as contFEAST, contSSRR follows this result.
Based on these properties, the proposed methods exhibit much higher accuracy than the complex momentbased matrix eigensolvers using a “solvethendiscretize” approach, as experimentally verified in Section 4.
4 Numerical experiments
In this section, we evaluate the performances of the proposed methods, contSSRR (Algorithm 2), contSSHankel (Algorithm 3), and contSSCAA (Algorithm 4), and compare them with that of contFEAST (Algorithm 1) for solving DEPs (1). Although the target problem of this paper is DEPs with ordinary differential operators, here we apply the proposed methods to DEPs with partial differential operators and evaluate their effectiveness.
The compared methods use the Npoint trapezoidal rule to approximate the contour integrals. In Sections 4.1–4.4 (Experiments I–IV) for ordinary differential operators, the quadrature points for the Npoint trapezoidal are on an ellipse with center γ, major axis ρ, and aspect ratio α, i.e.,
The corresponding weights are set as
Here, for real problems, we used (11) to reduce the number of ODEs to be solved. In Section 4.5 (Experiment V) for partial differential operators, we used the real rational filtering technique (12) to avoid complex partial differential equations (PDEs). For the proposed methods, we set δ = 10^{− 14} for the threshold of the lowrank approximation. In all the methods, we set \(V: \mathbb {C}^{L} \rightarrow {\mathscr{H}}\) to a random quasimatrix, whose columns are randomly generated functions represented by using 32 Chebyshev points on the same domain with the target problem.
Methods were implemented using MATLAB and Chebfun [8]. ODEs and PDEs were solved by using the “∖” command of Chebfun. All the numerical experiments were performed on a serial computer with the Microsoft (R) Windows (R) 10 Pro Operating System, an 11th Gen Intel(R) Core(TM) i71185G7 @ 3.00GHz CPU, and 32GB RAM.
4.1 Experiment I: proof of concept
For a proof of concept of the proposed methods, i.e., to show an advantage of the proposed method in the “solvethendiscretize” paradigm over a “discretizethensolve” approach in terms of accuracy, we tested on the onedimensional Laplace eigenvalue problem
Note that the true eigenpairs are \((\lambda _{i}, u_{i}) = (i^{2}, {\sin \limits } (ix)), i \in \mathbb {Z}_{+}\). We computed four eigenpairs such that λ_{i} ∈ [0,20].
First, we apply standard “discretizethensolve” approaches for solving (16) in which the coefficient operator is discretized by a threepoint central difference. The obtained matrix eigenvalue problem of size n is
and its eigenvalues can be written as
Note that we have \(\lim _{n \rightarrow \infty } \lambda _{i}^{(n)} = i^{2}\). We computed the eigenvalues using (18) with increasing n. We also applied SSRR [21] with (L, M, N) = (3,2,16) to the discretized problem (17) for each n.
The absolute errors of approximate eigenvalues computed by the “discretizethensolve” approaches are shown in Fig. 2. The errors decrease with increasing n; however, the errors turn to increase at n ≈ 10^{5} when using (18) due to rounding error and n ≈ 10^{6} for SSRR due to quadrature and rounding errors. The error reaches a minimum approximately 10^{− 8} and 10^{− 10} when using (18) and SSRR, respectively.
Next, we apply contSSRR with (L, M, N) = (3,2,16) to (16). Here, we set (γ, ρ, α) = (10,10,1) for the contour path. The obtained eigenvalues and eigenfunction are shown in Table 1 and Fig. 3, respectively. In contrast to the “discretizethensolve” approach, the proposed method achieves much higher accuracy (absolute errors are approximately 10^{− 14}), which shows effectiveness of the “solvethendiscretize” approach over the “discretizethensolve” approach. This is one of the greatest advantages of the proposed method over complex momentbased matrix eigensolvers.
These results demonstrate that the proposed methods work well for solving DEPs without discretization of the coefficient operator.
4.2 Experiment II: parameter dependence
We evaluate the parameter dependence of contSSRR with the subspace iteration technique and contFEAST on the convergence. We computed the same four eigenpairs λ_{i} ∈ [0,20] of (16) using the same (γ, ρ, α) = (10,10,1) for the contour path as used in Section 4.1. We evaluate the convergence of contSSRR with (L, N) = (4,4) varying M = 1,2,3,4 and with (M, N) = (1,4) varying L = 4,8,12,16, and contFEAST with N = 4 varying L = 4,8,12,16.
We show the residual history of each method in Fig. 4(a)–(c) regarding residual norm \(\ r_{i} \_{{\mathscr{H}}} = \ \mathcal {A} \widehat {u}_{i}  \widehat {\lambda }_{i} {\mathscr{B}} \widehat {u}_{i} \_{{\mathscr{H}}}\). The convergences of contFEAST and contSSRR with M = 1 are almost identical and improve with an increase in L (Fig. 4(a) and (b)). We also observed that, in contSSRR, increasing M also improves the convergence to the same degree as increasing L (Fig. 4(c)).
We also show in Fig. 4(d) the theoretical convergence rate obtained from Theorem 4, i.e., \(\max \limits _{\lambda _{i} \in {\Omega }} \lvert f_{N}(\lambda _{LM+1}) / f_{N}(\lambda _{i}) \rvert \), and the evaluated convergence rate of each method, i.e., the ratio of the residual norm between the first and second iterations. Although, in contSSRR, increasing L indicates a slightly smaller convergence rate than increasing M, both evaluated convergence rates are almost the same as the theoretical convergence rate obtained from Theorem 4.
These results demonstrate that the proposed method with M ≥ 2 achieves fast convergence even with a small value of L. This contributes to the reduction in elapsed time, which will be shown in Section 4.3.
4.3 Experiment III: performance for realworld problems
Next, we evaluate the performances of the proposed methods without iteration (ℓ = 1) and compare them with those of contFEAST and the “eigs” function [31] in Chebfun for the following six eigenvalue problems: two for computing real outermost eigenvalues, two for computing real interior eigenvalues, and two for computing complex eigenvalues.

Real Outermost: Mathieu eigenvalue problem [32]:
$$ \left(  \frac{ \mathrm{d}^{2} }{ \mathrm{d}x^{2} } + 2 q \cos(2x) \right) u = \lambda u, \quad u(0) = u(\pi/2) = 0 $$with q = 2. It has only real eigenvalues. We computed 15 eigenpairs corresponding to outermost eigenvalues λ_{i} ∈ [0,1000].

Real Outermost: Schrödinger eigenvalue problem [33, Chapter 6]:
$$ \left(  \frac{\hbar}{2m} \frac{ \mathrm{d}^{2} }{ \mathrm{d}x^{2} } + V(x) \right) u = \lambda u, \quad u(1) = u(1) = 0 $$with a doublewell potential V (x) = 1.5,x ∈ [− 0.2,0.3], where we set \(\hbar /2m = 0.01\). It has only real eigenvalues. We computed 19 eigenpairs corresponding to outermost eigenvalues λ_{i} ∈ [0,10].

Real Interior: Bessel eigenvalue problem [34]:
$$ \left( x^{2} \frac{ \mathrm{d}^{2} }{ \mathrm{d}x^{2} } + x \frac{ \mathrm{d} }{ \mathrm{d}x}  \alpha^{2} \right) u =  \lambda x^{2} u, \quad u(0) = u(1) = 0 $$with α = 1. It has only real eigenvalues. We computed 11 eigenpairs corresponding to interior eigenvalues λ_{i} ∈ [500,3000].

Real Interior: Sturm–Liouvilletype eigenvalue problem:
$$ \left(  \frac{ \mathrm{d}^{2} }{ \mathrm{d}x^{2} } + x^{2} \right) u = \lambda \cosh(x) u, \quad u(1) = u(1) = 0, $$which is used in [15]. It has only real eigenvalues. We computed 12 eigenpairs corresponding to interior eigenvalues λ_{i} ∈ [200,1000].

Complex: Orr–Sommerfeld eigenvalue problem [35]:
$$ \begin{array}{@{}rcl@{}} & \left\{ \frac{1}{ Re } \left( \frac{ \mathrm{d}^{2} }{ \mathrm{d}x^{2} }  \alpha^{2}\right)^{2}  \mathrm{i} \alpha \left[ U \left( \frac{ \mathrm{d}^{2}}{ \mathrm{d}x^{2}}  \alpha^{2} \right) + U^{\prime\prime} \right] \right\} u = \lambda \left( \frac{ \mathrm{d}^{2}}{ \mathrm{d}x^{2}}  \alpha^{2} \right) u, \\ & u(1) = u(1) = 0 \end{array} $$with α = 1 and U = 1 − x^{2}. We solved two cases with Re = 1000 and Re = 2000. They have complex eigenvalues. We computed 18 eigenpairs for Re = 1000 and 28 eigenpairs for Re = 2000 shown in Fig. 5.
Tables 2 and 3 give the contour path and values of parameters for each problem. The “eigs” function in Chebfun with parameters k and σ computes k closest eigenvalues to σ and the corresponding eigenfunctions. We set the parameters k and σ to the number of input functions L of contFEAST in Table 3 and the center of contour path γ in Table 2, respectively.
Figures 6 and 7 show the residual norms \(\ r_{i} \_{{\mathscr{H}}} = \ \mathcal {A} \widehat {u}_{i}  \widehat {\lambda }_{i} {\mathscr{B}} \widehat {u}_{i} \_{{\mathscr{H}}}\) for each problem and Fig. 8 shows the elapsed times for each problem. In Fig. 8, “Solve ODEs,” “Orthonormalization,” “Matrix Eig,” and “MISC” denote the elapsed times for solving ODEs (10), orthonormalization of the column vectors of \(\widehat {S}\), construction and solution of the matrix eigenvalue problem, and other parts including computation of the contour integral, respectively.
First, we discuss the accuracy of the presented methods. For the real outermost and interior problems (Fig. 6), the residual norms of contFEAST decrease with more iterations reaching \(\ r_{i} \_{{\mathscr{H}}} \approx 10^{10}\) at ℓ = 3 for the target eigenpairs. ContSSRR and contSSCAA demonstrate almost the same high accuracy (\(\ r_{i} \_{{\mathscr{H}}} \approx 10^{10}\)) as contFEAST with ℓ = 3; on the other hand, contSSHankel shows lower accuracy than the others except for the Bessel eigenvalue problem. The residual norms for the eigenvalues outside the target region tend to be large depending on the distance from the target region. The experimental results exhibit a similar trend for both outermost and interior problems.
For the complex problems (Fig. 7), the residual norms of contFEAST stagnate at \(\ r_{i} \_{{\mathscr{H}}} \approx 10^{7}\) for Re = 1000 and \(\ r_{i} \_{{\mathscr{H}}} \approx 10^{6}\) for Re = 2000 in ℓ = 2. ContSSRR achieves almost the same accuracy as contFEAST; on the other hand, SSHankel and contSSCAA are less accurate than contFEAST and contSSRR.
Next, we discuss the elapsed times of the methods (Fig. 8). For the complex momentbased methods, most of the elapsed time is spent on solving the ODEs. The total elapsed time of contFEAST increases in proportion to the number of iterations ℓ. Although contSSRR and contSSCAA account for larger portions of elapsed time for orthonormalization of the basis functions of \({\mathscr{R}}(\widehat {S})\) because they use a larger dimensional subspace (Section 3.4.2), the proposed methods exhibit much less total elapsed times than contFEAST. The proposed methods are over eight times faster than contFEAST with ℓ = 3 for real problems and over four times faster than contFEAST with ℓ = 2 for complex problems, while maintaining almost the same high accuracy.
We also compare the performance of the proposed methods with that of the “eigs” function in Chebfun. As shown in Fig. 8, the “eigs” function is much faster than the proposed methods and contFEAST. On the other hand, Figs. 6 and 7 show that the “eigs” function exhibits significant losses of accuracy in several cases (\(\ r_{i} \_{{\mathscr{H}}} \approx 10^{5}\) for the Bessel eigenvalue problem, \(\ r_{i} \_{{\mathscr{H}}} \approx 10^{4}\) for the Orr–Sommerfeld eigenvalue problems with Re = 1000, and \(\ r_{i} \_{{\mathscr{H}}} \approx 10^{2}\) for the Orr–Sommerfeld eigenvalue problems with Re = 2000) and is unrobust in accuracy relative to the complex momentbased methods.
4.4 Experiment IV: parallel performance
As demonstrated in Section 4.3, the most timeconsuming part of the complex momentbased methods is the solutions of LN ODEs (10). Since these LN ODEs can be solved independently, the methods are expected to have high parallel performance.
Here, we estimated the strong scalability of the methods by using the following performance model. We assume that the elapsed time \(T_{\text {ODE}}^{(j)}\) for solving ODEs (10) depends on the quadrature point z_{j} but is independent of the righthand side \({\mathscr{B}} v_{i}\). We also assume that the elapsed time T_{QP} for other computation at each quadrature point is independent of the quadrature point z_{j}. In addition, we let T_{other} be the elapsed time for computation of other parts in each method, respectively. The LN ODEs are solved in parallel by P processes, computations at quadrature points are parallelized in \(\min \limits (P,N)\) processes, and other parts are computed in serial.
Then, using the measured elapsed times \(T_{\text {ODE}}^{(j)}, T_{\text {QP}}\), and T_{other}, we estimate the total elapsed time T_{total}(P) of each method in P processes as
where \({\mathscr{J}}_{p}\) is the index set of quadrature points equally assigned to each process p and ⌈⋅⌉ denotes the ceiling function.
We estimated the strong scalability of methods for solving the Orr–Sommerfeld eigenvalue problem with Re = 2000. We used the same parameter values as in Section 4.3. Figure 9 shows the estimated time and strong scalability of methods. This result demonstrates that all the methods exhibit highly parallel performance. The proposed methods, especially contSSHankel, are much faster than contFEAST even with a large number of processes P, although contFEAST shows slightly better scalability than the proposed methods.
4.5 Experiment V: performance for partial differential operators
The complex momentbased methods can be extended to partial differential operators in a straightforward manner in which L PDEs are solved regarding each quadrature point.
Here, we evaluate the performances of the proposed methods without iteration (ℓ = 1) and compare them with that of contFEAST for two real selfadjoint problems:

2D Laplace eigenvalue problem:
$$  \frac{\hbar}{2m} \left( \frac{ \partial^{2} }{ \partial x^{2} } + \frac{ \partial^{2} }{ \partial y^{2} } \right) u = \lambda u $$in a domain [0,π] × [0,π] with zero Dirichlet boundary condition. The true eigenvalues are \({i_{x}^{2}} + {i_{y}^{2}}\) with \(i_{x}, i_{y} \in \mathbb {Z}_{+}\). We computed 4 eigenpairs, counting multiplicity, corresponding to λ_{i} ∈ [0,9]. Note that the target eigenvalues are 2,5, and 8, where the eigenvalue 5 has multiplicity 2.

2D Schrödinger eigenvalue problem:
$$ \left[  \frac{\hbar}{2m} \left( \frac{ \partial^{2} }{ \partial x^{2} } + \frac{ \partial^{2} }{ \partial y^{2} } \right) + V(x,y) \right] u = \lambda u $$in a domain [− 1,1] × [− 1,1] with a potential V (x) = 0.1(x + 0.4)^{2} + 0.1(y − 0.8)^{2} and zero Dirichlet boundary condition, where we set \(\hbar /2m = 0.01\). We computed 5 eigenpairs corresponding to λ_{i} ∈ [0.15,0.4].
For both problems, we set (L, M, N) = (2,4,24) for the proposed methods and (L, N) = (6,24) for contFEAST.
Table 4 gives the obtained eigenvalues of contSSRR for the 2D Laplace eigenvalue problem. In addition, residual norms \(\ r_{i} \_{{\mathscr{H}}} = \ \mathcal {A} \widehat {u}_{i}  \widehat {\lambda }_{i} {\mathscr{B}} \widehat {u}_{i} \_{{\mathscr{H}}}\) for each problem are presented in Fig. 10 and the elapsed times for each problem are presented in Fig. 11.
We observed from Table 4 and Fig. 10 that, as in the case of ordinary differential operators, the proposed methods work well for solving DEPs with partial differential operators even for a nonsimple case (2DLaplacian eigenvalue problem). In addition, the proposed methods exhibit much lower elapsed times than contFEAST (see Fig. 11), although the elapsed times for orthonormalization of the column vectors of \(\widehat {S}\) and construction of the matrix eigenvalue problem are relatively larger than the cases of ordinary differential operators in Section 4.3.
4.6 Summary of numerical experiments
From the numerical experiments, we observed the following:

As well as contFEAST, the proposed methods in the “solvethendiscretize” paradigm exhibit a much higher accuracy than the “discretizethensolve” approach for solving DEPs (1).

Using higherorder complex moments improves the accuracy as well as increasing the number of input functions L.

Thanks to the higherorder complex moments, the proposed methods are over eight times faster for real problems and more than four times faster for complex problems compared with contFEAST while maintaining almost the same high accuracy.
5 Conclusion
In this paper, based on the “solvethendiscretize” paradigm, we propose operation analogues of the Sakurai–Sugiura’s approach, contSSHankel, contSSRR, and contSSCAA, for DEPs (1), without discretization of operators \(\mathcal {A}\) and \({\mathscr{B}}\). Theoretical and numerical results indicate that the proposed methods significantly reduce the number of ODEs to solve and elapsed time by using higherorder complex moments while maintaining almost the same high accuracy as contFEAST.
As well as contFEAST, the proposed methods based on the “solvethendiscretize” paradigm exhibit much higher accuracy than methods based on the traditional “discretizethensolve” paradigm. This study successfully reduced the computational costs of contFEAST and is expected to promote the “solvethendiscretize” paradigm for solving differential eigenvalue problems and contribute to faster and more accurate solutions in realworld applications.
This paper did not intend to investigate a practical parameter setting, rounding error analysis and parallel performance evaluation. In future, we will develop the proposed methods and evaluate the parallel performance specifically for higher dimensional problems. Furthermore, based on the concept in [36], we will rigorously evaluate the truncation error of the quadrature and numerical errors in the proposed methods and investigate a verified computation method based on the proposed methods for differential eigenvalue problems.
References
Polizzi, E.: A density matrixbased algorithm for solving eigenvalue problems. Phys. Rev. B 79, 115112 (2009). https://doi.org/10.1103/physrevb.79.115112
Kestyn, J., Kalantzis, V., Polizzi, E., Saad, Y.: PFEAST: A high performance sparse eigenvalue solver using distributedmemory linear solvers. In: SC’16 Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis. https://doi.org/10.1109/SC.2016.15, pp 178–189 (2016)
Iwase, S., Futamura, Y., Imakura, A., Sakurai, T., Ono, T.: Efficient and scalable calculation of complex band structure using Sakurai–Sugiura method. In: SC’17 Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis. https://doi.org/10.1145/3126908.3126942, pp 1–12 (2017)
Huang, T.M., Liao, W., Lin, W.W., Wang, W.: An efficient contour integral based eigensolver for 3D dispersive photonic crystal. J. Comput. Appl. Math. 395, 113581 (2021). https://doi.org/10.1016/j.cam.2021.113581
Kurz, S., Schöps, S., Unger, G., Wolf, F.: Solving Maxwell’s eigenvalue problem via isogeometric boundary elements and a contour integral method. Mathematical Methods in the Applied Sciences 44(13), 10790–10803 (2021). https://doi.org/10.1002/mma.7447
Chatelin, F.: Eigenvalues of Matrices: Revised Edition. SIAM, Philadelphia (2012)
Sakurai, T., Sugiura, H.: A projection method for generalized eigenvalue problems using numerical integration. J. Comput. Appl. Math. 159(1), 119–128 (2003). https://doi.org/10.1016/S03770427(03)00565X
Driscoll, T.A., Hale, N., Trefethen, L.N.: Chebfun Guide. Pafnuty Publications, Oxford (2014)
Battles, Z., Trefethen, L.N.: An extension of MATLAB to continuous functions and operators. SIAM J. Sci. Comput. 25(5), 1743–1770 (2004). https://doi.org/10.1137/s1064827503430126
Trefethen, L.N.: Householder triangularization of a quasimatrix. IMA J. Numer. Anal. 30(4), 887–897 (2010). https://doi.org/10.1093/imanum/drp018
Olver, S., Townsend, A.: A practical framework for infinitedimensional linear algebra. In: 2014 First Workshop for High Performance Technical Computing in Dynamic Languages. https://doi.org/10.1109/HPTCDL.2014.10, pp 57–62 (2014)
Townsend, A., Trefethen, L.N.: Continuous analogues of matrix factorizations. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences 471(2173), 20140585 (2015). https://doi.org/10.1098/rspa.2014.0585
Gilles, M.A., Townsend, A.: Continuous analogues of Krylov subspace methods for differential operators. SIAM J. Numer. Anal. 57(2), 899–924 (2019). https://doi.org/10.1137/18M1177810
Mohr, S., Nakatsukasa, Y., Urzúatorres, C.: Full operator preconditioning and the accuracy of solving linear systems. arXiv:2105.07963(2021)
Horning, A., Townsend, A.: FEAST For differential eigenvalue problems. SIAM J. Numer. Anal. 58(2), 1239–1262 (2020). https://doi.org/10.1137/19M1238708
Imakura, A., Du, L., Sakurai, T.: Error bounds of Rayleigh–Ritz type contour integralbased eigensolver for solving generalized eigenvalue problems. Numer. Algorithms 71(1), 103–120 (2016). https://doi.org/10.1007/s1107501599874
Tzounas, G., Dassios, I., Liu, M., Milano, F.: Comparison of numerical methods and opensource libraries for eigenvalue analysis of largescale power systems. Appl. Sci. 10(21). https://doi.org/10.3390/app10217592 (2020)
Imakura, A., Matsuda, M., Ye, X., Sakurai, T.: Complex momentbased supervised eigenmap for dimensionality reduction. In: Proceedings of the AAAI Conference on Artificial Intelligence. https://doi.org/10.1609/aaai.v33i01.33013910, pp 3910–3918 (2019)
Sakurai, T., Tadano, H.: CIRR: A Rayleigh–Ritz type method with counter integral for generalized eigenvalue problems. Hokkaido Math. J 36, 745–757 (2007). https://doi.org/10.14492/hokmj/1272848031
Ikegami, T., Sakurai, T., Nagashima, U.: A filter diagonalization for generalized eigenvalue problems based on the Sakurai–Sugiura projection method. J. Comput. Appl. Math. 233(8), 1927–1936 (2010). https://doi.org/10.1016/j.cam.2009.09.029
Ikegami, T., Sakurai, T.: Contour integral eigensolver for nonHermitian systems: a Rayleigh–Ritztype approach. Taiwan. J. Math., 825–837. https://doi.org/10.11650/twjm/1500405869 (2010)
Imakura, A., Du, L., Sakurai, T.: A block Arnolditype contour integral spectral projection method for solving generalized eigenvalue problems. Appl. Math. Lett. 32, 22–27 (2014). https://doi.org/10.1016/j.aml.2014.02.007
Imakura, A., Du, L., Sakurai, T.: Relationships among contour integralbased methods for solving generalized eigenvalue problems. Jpn. J. Ind. Appl. Math. 33(3), 721–750 (2016). https://doi.org/10.1007/s131600160224x
Imakura, A., Sakurai, T.: Block Krylovtype complex momentbased eigensolvers for solving generalized eigenvalue problems. Numer. Algorithms 75(2), 413–433 (2017). https://doi.org/10.1007/s1107501602415
Imakura, A., Futamura, Y., Sakurai, T.: Structurepreserving technique in the block SS–Hankel method for solving Hermitian generalized eigenvalue problems. In: International Conference on Parallel Processing and Applied Mathematics, pp 600–611. Springer, Cham (2017). https://doi.org/10.1007/9783319780245_52
Tang, P.T.P., Polizzi, E.: FEAST As a subspace iteration eigensolver accelerated by approximate spectral projection. SIAM J. Matrix Anal. Appl. 35 (2), 354–390 (2014). https://doi.org/10.1137/13090866X
Güttel, S., Polizzi, E., Tang, P.T.P., Viaud, G.: Zolotarev quadrature rules and load balancing for the FEAST eigensolver. SIAM J. Sci. Comput. 37(4), 2100–2122 (2015). https://doi.org/10.1137/140980090
Kato, T.: Perturbation Theory for Linear Operators, 2nd edn., vol. 132. Springer, Berlin, Heidelberg (1995). https://doi.org/10.1007/9783642662829
Austin, A.P., Trefethen, L.N.: Computing eigenvalues of real symmetric matrices with rational filters in real arithmetic. SIAM J. Sci. Comput. 37(3), 1365–1387 (2015). https://doi.org/10.1137/140984129
Hoemmen, M.: Communicationavoiding Krylov subspace methods. Technical Report UCB/EECS201037, University of California, Berkeley (2010)
Driscoll, T.A., Bornemann, F., Trefethen, L.N.: The chebop system for automatic solution of differential equations. BIT Numer. Math. 48(4), 701–723 (2008). https://doi.org/10.1007/s1054300801984
Mathieu, E.: Mémoire sur le mouvement vibratoire d’une membrane de forme elliptique. Journal de Mathématiques Pures et Appliquées 13, 137–203 (1868)
Trefethen, L.N., Birkisson, A., Driscoll, T.A.: Exploring ODEs. Society for Industrial and Applied Mathematics, Philadelphia, PA (2017). https://doi.org/10.1137/1.9781611975161
Watson, G.N.: A treatise on the theory of Bessel functions. Cambridge mathematical library. Cambridge University Press, New York, NY (1995)
Schmid, P.J., Henningson, D.S.: Stability and transition in shear flows. Springer, New York, NY (2001). https://doi.org/10.1007/9781461301851
Imakura, A., Morikuni, K., Takayasu, A.: Verified partial eigenvalue computations using contour integrals for Hermitian generalized eigenproblems. J. Comput. Appl. Math. 369, 112543 (2020). https://doi.org/10.1016/j.cam.2019.112543
Acknowledgements
The authors would like to thank the reviewers for their careful reading and constructive comments.
Funding
This work was supported in part by the Japan Society for the Promotion of Science (JSPS), GrantsinAid for Scientific Research (Nos. JP18K13453, JP19KK0255, JP20K14356, and JP21H03451).
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
The authors declare no competing interests.
Additional information
Dedicated to Claude Brezinski on the occasion of his eightieth birthday.
Data availability
The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Akira Imakura, Keiichi Morikuni, and Akitoshi Takayasu contributed equally to this work.
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
Imakura, A., Morikuni, K. & Takayasu, A. Complex momentbased methods for differential eigenvalue problems. Numer Algor 92, 693–721 (2023). https://doi.org/10.1007/s1107502201456y
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s1107502201456y