betaFeedback

WorksMathematics

Kernel Compensation and Quasi-Newton Iteration for Nonlinear Non-Hermitian Eigenproblems in Photonic Crystals

Abstract

We extend the kernel-compensated eigensolver developed in our previous work from Hermitian to non-Hermitian nonlinear eigenproblems arising in dispersive photonic crystals. A mimetic finite-difference discretization preserves the discrete curl–divergence complex, and three-dimensional discrete Fourier transforms provide matrix-free realizations of the curl and divergence blocks. The isotropic inverse dielectric coefficient is stored through the material-edge indices, so its value and derivatives can be updated without assembling a Maxwell matrix. Kernel compensation is characterized by a fixed unitary block diagonalization, which remains exact for the nonlinear problem. At each outer step, a right Generalized Davidson iteration solves the non-Hermitian linearized eigenproblem. The exact derivative of a simple eigenvalue branch is expressed through its left and right eigenvectors. To avoid a second eigensolve, we replace the left eigenvector by the available normalized right eigenvector and obtain a right-Rayleigh quasi-Newton update. We prove that the exact left–right Newton iteration is locally quadratically convergent at simple roots, whereas the proposed iteration with exact inner eigenpairs has an explicit linear error factor and is at least quadratic when the derivative defect vanishes. A spectral regularity estimate bounds the gradient appearing in the derivative defect.

Introduction

Frequency-dependent permittivities turn the Maxwell eigenvalue problem in photonic crystals into a nonlinear eigenvalue problem (NEP). Material absorption leads to complex frequencies and non-Hermitian matrices [5, 21]. The resulting three-dimensional problems combine material nonlinearity with a large curl kernel. Iterative frequency updates require repeated solutions of large eigenproblems, so both the nonlinear iteration and the treatment of the kernel affect the computational cost [8, 12].

Several methods address the frequency dependence. Auxiliary-field and linearization formulations replace the NEP by a larger linear problem [5, 20], while spectral indicator methods treat frequency-dependent permittivities through finite-element discretization [25]. For general analytic NEPs, infinite Arnoldi uses an infinite-dimensional linear operator [14], NLEIGS uses rational interpolation and structured linearization [9], and contour integration recovers eigenvalues inside a prescribed contour from resolvent moments [1]. Quasi-Newton analyses identify the effect of Jacobian approximations [13]; inexact Newton analyses relate inner accuracy to local convergence [3, 23]. Our analysis concerns the particular derivative approximation used in repeated frozen-frequency Maxwell eigensolves.

The Yee discretization preserves the discrete curl–divergence relation [26]. FFT-based null-space-free methods exploit the curl structure to remove zero eigenvalues [10]. Tiexiang Li and collaborators developed this approach in FAME, including CPU and GPU implementations [19], and extended it to anisotropic crystals in oblique coordinates [24]. Their structure-preserving complex-band method fixes a real frequency and computes a complex wave-vector component through a palindromic quadratic eigenvalue formulation [11]. That formulation differs from the present NEP, which fixes a real Bloch wave vector and seeks complex frequencies in a dispersive material. For dispersive metallic crystals, Huang, Lin, and Mehrmann combined a Newton-type iteration with nonequivalence deflation and FFT-based null-space-free eigensolves [12]. Their scalar Newton update uses a left eigenvector as well as the right eigenpair.

Our earlier work introduced kernel compensation with mimetic finite differences [15]. A discrete divergence penalty treats the curl kernel while retaining the original vector space. We subsequently developed anisotropic discretizations and a GPU-accelerated Hermitian solver [17], and combined kernel compensation with fixed-point and Newton iterations for nonlinear Hermitian problems [16]. In that setting, the normalized right eigenvector gives the exact eigenvalue derivative, and the Hermitian min–max principle separates the compensated spectrum.

The non-Hermitian extension requires different spectral and convergence arguments. The min–max principle is unavailable for complex spectra, and a simple eigenvalue derivative involves both left and right eigenvectors [7]. We retain the mimetic DFT operators and use a right Generalized Davidson eigensolver [2, 22] at each frozen frequency. The quasi-Newton update uses the available normalized right eigenvector in place of the left eigenvector and avoids an additional adjoint eigensolve. Its derivative defect is expressed through the gradient of the material Rayleigh quotient and bounded using the relative spectral gap of the Hermitian part.

The contributions of this work are as follows.

  • A frequency-independent unitary transformation block-diagonalizes the compensated non-Hermitian matrix, separating the discrete curl image, divergence image, and harmonic spaces, and proves that kernel compensation preserves the original nonlinear eigenpairs under an explicit penalty condition.

  • The spectral analysis of the preconditioner is extended to complex material coefficients: the two-sided preconditioned matrix is normal, its spectrum lies on a complex line segment, and its condition number admits a material-dependent bound independent of the grid size.

  • The right-Rayleigh update reuses the computed right eigenvector. For exact inner eigenpairs, we express its linear error factor in terms of the gradient of the material Rayleigh quotient. A relative-gap estimate for the Hermitian part bounds the derivative defect and gives a sufficient condition for local contraction at a simple root. We implement the method in complex double precision with matrix-free CUDA operators. For the SC_CURV/SiC tests with 64 ≤N≤\le N \le 128, the GPU implementation achieves speedups of 69.6–73.0 over the single-core, single-thread CPU implementation.

The paper is organized as follows. § 2 gives the mimetic discretization and matrix-free operators. § 3 establishes the kernel compensation formulation and preconditioning results. § 4 presents the eigensolver, the quasi-Newton update, and the local convergence analysis. § 6 reports complex band structures, convergence histories, and CPU and GPU timings. § 7 concludes the paper.

Discretization

Model Problem and Mimetic Finite Difference Scheme

The model problem is defined over the primitive cell

Ω:={x=∑i=13xiai:xi∈[0,1], i=1,2,3}(1)\Omega:= \left\{ \mathbf{x} = \sum_{i=1}^{3} x_i \mathbf{a}_i : x_i \in[0,1],\ i = 1,2,3 \right\} \tag*{(1)}

where {ai}i=13\{\mathbf{a}_i\}_{i=1}^{3} forms a basis in R3\mathbb{R}^{3}. For a Bloch wave vector k\mathbf{k}, the original magnetic field satisfies

HPQ(x+ai)=exp⁡(ik⋅ai)HPQ(x).\mathbf{H}^{\mathrm{PQ}}(\mathbf{x}+\mathbf{a}_i)=\exp(\mathrm{i}\mathbf{k}\cdot\mathbf{a}_i)\mathbf{H}^{\mathrm{PQ}}(\mathbf{x}).

The phase transformation H(x):=exp⁡(−ik⋅x)HPQ(x)\mathbf{H}(\mathbf{x}):=\exp(-\mathrm{i}\mathbf{k}\cdot\mathbf{x})\mathbf{H}^{\mathrm{PQ}}(\mathbf{x}) makes the transformed field periodic. The corresponding differential operator is

∇k:=∇+ik.(2)\nabla_{\mathbf{k}} := \nabla+ \mathrm{i}\mathbf{k}. \tag*{(2)}

The magnetic-field formulation with this shifted operator is

{∇k×ε−1(x,ω)∇k×H=ω2Hin Ω,∇k⋅H=0in Ω,H(x+ai)=H(x),i=1,2,3.(3)\begin{cases} \nabla_{\mathbf{k}} \times\varepsilon^{-1}(\mathbf{x},\omega)\nabla_{\mathbf{k}} \times\mathbf{H} = \omega^{2}\mathbf{H} & \text{in }\Omega,\\ \nabla_{\mathbf{k}} \cdot\mathbf{H} = 0 & \text{in }\Omega,\\ \mathbf{H}(\mathbf{x}+\mathbf{a}_{i}) = \mathbf{H}(\mathbf{x}), \quad i=1,2,3. \tag*{(3)} \end{cases}

The dielectric coefficient ε\varepsilon satisfies the spatial periodicity condition: ε(x+ai,ω)=ε(x,ω)\varepsilon(\mathbf{x}+\mathbf{a}_{i},\omega)=\varepsilon(\mathbf{x},\omega). ε−1\varepsilon^{-1} is piecewise defined in Ω\Omega:

ε−1(x,ω)={η(ω),x∈Ω1,1,x∈Ω∖Ω1,(4)\varepsilon^{-1}(\mathbf{x},\omega)= \begin{cases} \eta(\omega), & \mathbf{x}\in\Omega_{1},\\ 1, & \mathbf{x}\in\Omega\setminus\Omega_{1}, \end{cases} \tag*{(4)}

where Ω1\Omega_{1} denotes the material subregion. The function η\eta is holomorphic in a pole-free neighborhood of each frequency under consideration. We consider frequencies in Ωsol⊂C\Omega_{\mathrm{sol}}\subset\mathbb{C} such that

Ωsol⊂{ω: ω≠0, 0<Re⁡η(ω)<1}.(5)\Omega_{\mathrm{sol}}\subset\{\omega:\ \omega\ne0,\ 0<\operatorname{Re}\eta(\omega)<1\}. \tag*{(5)}

We map the primitive cell to the unit cube and divide each coordinate direction into NN intervals of width h=1Nh=\frac{1}{N}. As in our previous mimetic finite-difference discretizations [15, 16, 17], the Yee scheme [26] places the three electric components at edge centers, the three magnetic components at face centers, and the divergence value at the cell center, as is illustrated in Table 1.

Physical QuantityFunctionsApproximate Vectors
Electric FieldE=(E1,E2,E3)\mathbf{E}=(E^{1},E^{2},E^{3})(Ei−12,j,k1,Ei,j−12,k2,Ei,j,k−123)\left(E^{1}_{i-\frac{1}{2},j,k},E^{2}_{i,j-\frac{1}{2},k},E^{3}_{i,j,k-\frac{1}{2}}\right)
Magnetic FieldH=(H1,H2,H3)\mathbf{H}=(H^{1},H^{2},H^{3})(Hi,j−12,k−121,Hi−12,j,k−122,Hi−12,j−12,k3)\left(H^{1}_{i,j-\frac{1}{2},k-\frac{1}{2}},H^{2}_{i-\frac{1}{2},j,k-\frac{1}{2}},H^{3}_{i-\frac{1}{2},j-\frac{1}{2},k}\right)
Divergenceψ\psi(ψi−12,j−12,k−12)\left(\psi_{i-\frac{1}{2},j-\frac{1}{2},k-\frac{1}{2}}\right)

Table 1. Degrees of freedom of the discrete electromagnetic fields and divergence image on the Yee grid with NN intervals in each direction. The indices satisfy i,j,k∈{1,2,…,N}i,j,k\in\{1,2,\ldots,N\}.

The symbol ψ\psi denotes the divergence. The discrete electric and magnetic fields each contain 3N33N^{3} degrees of freedom (DoFs), whereas the discrete divergence contains N3N^{3} DoFs. The outer discrete curl maps edge DoFs to face DoFs; its conjugate transpose gives the inner curl. The discrete inverse dielectric matrix Mη(ω)\mathcal{M}_{\eta}(\omega) maps edge DoFs to edge DoFs.

Discrete Matrix Operators

The matrix representations of the discrete operators have circulant structures. We introduce two N×NN\times N circulant matrices:

D1:=1h(10⋯0−1−11⋱00⋱⋱⋱⋮⋮−1100⋯0−11),D0:=12(10⋯0111⋱00⋱⋱⋱⋮⋮1100⋯011).(6)\mathbf{D}_{1}:=\frac{1}{h} \begin{pmatrix} 1 & 0 & \cdots& 0 & -1\\ -1 & 1 & \ddots& 0\\ 0 & \ddots& \ddots& \ddots& \vdots\\ \vdots& -1 & 1 & 0\\ 0 & \cdots& 0 & -1 & 1 \end{pmatrix}, \qquad \mathbf{D}_{0}:=\frac{1}{2} \begin{pmatrix} 1 & 0 & \cdots& 0 & 1\\ 1 & 1 & \ddots& 0\\ 0 & \ddots& \ddots& \ddots& \vdots\\ \vdots& 1 & 1 & 0\\ 0 & \cdots& 0 & 1 & 1 \end{pmatrix}. \tag*{(6)}

The matrix D1\mathbf{D}_{1} approximates a first derivative between staggered locations, while D0\mathbf{D}_{0} averages the zeroth-order Bloch term to the same target DoFs. The higher-order mimetic stencils in [16, 17] replace the circulant coefficients while preserving the factorizations and commutativity used below. The matrix representation in 3D is assembled by Kronecker products between Ds\mathbf{D}_{s} (s=0,1s=0,1) and identity matrices:

Ds,1=IN2⊗Ds,Ds,2=IN⊗Ds⊗IN,Ds,3=Ds⊗IN2.(7)\mathcal{D}_{s,1}=I_{N^{2}}\otimes\mathbf{D}_{s},\qquad \mathcal{D}_{s,2}=I_{N}\otimes\mathbf{D}_{s}\otimes I_{N},\qquad \mathcal{D}_{s,3}=\mathbf{D}_{s}\otimes I_{N^{2}}. \tag*{(7)}

For a general primitive cell, the matrix A:=(a1 a2 a3)\mathbf{A}:=(\mathbf{a}_1\ \mathbf{a}_2\ \mathbf{a}_3) collects the lattice vectors, and bij:=(A−1)ijb_{ij}:=(\mathbf{A}^{-1})_{ij} denotes the (i,j)(i,j)-th entry of its inverse. After mapping Ω\Omega to the unit cube, we define the transformed blocks as in [15, 17]:

Di:=∑j=13bijD1,j+ikiD0,i,i=1,2,3.(8)D_i:=\sum_{j=1}^{3}b_{ij}D_{1,j}+\mathrm{i}k_iD_{0,i},\qquad i=1,2,3. \tag*{(8)}

The discrete curl and divergence operators A\mathcal{A} and B\mathcal{B} are defined by

References

  1. [1]W.-J. Beyn. An integral method for solving nonlinear eigenvalue problems. 𝘓𝘪𝘯𝘦𝘢𝘳 𝘈𝘭𝘨𝘦𝘣𝘳𝘢 𝘢𝘯𝘥 𝘪𝘵𝘴 𝘈𝘱𝘱𝘭𝘪𝘤𝘢𝘵𝘪𝘰𝘯𝘴, 436(10):3839–3863, 2012.
  2. [2]E. R. Davidson. The iterative calculation of a few of the lowest eigenvalues and corresponding eigenvectors of large real-symmetric matrices. 𝘑𝘰𝘶𝘳𝘯𝘢𝘭 𝘰𝘧 𝘊𝘰𝘮𝘱𝘶𝘵𝘢𝘵𝘪𝘰𝘯𝘢𝘭 𝘗𝘩𝘺𝘴𝘪𝘤𝘴, 17(1):87–94, 1975.DOI
  3. [3]R. S. Dembo, S. C. Eisenstat, and T. Steihaug. Newton methods. 𝘚𝘐𝘈𝘔 𝘑𝘰𝘶𝘳𝘯𝘢𝘭 𝘰𝘯 𝘕𝘶𝘮𝘦𝘳𝘪𝘤𝘢𝘭 𝘈𝘯𝘢𝘭𝘺𝘴𝘪𝘴, 19(2):400–408, 1982.
  4. [4]G. C. R. Devarapu and S. Foteinopoulou. Mid-IR near-perfect absorption with a SiC photonic crystal with angle-controlled polarization selectivity. 𝘖𝘱𝘵𝘪𝘤𝘴 𝘌𝘹𝘱𝘳𝘦𝘴𝘴, 20(12):13040–13054, 2012.DOI
  5. [5]C. Effenberger, D. Kressner, and C. E. Englström. Linearization techniques for band structure calculations in absorbing photonic crystals. 𝘐𝘯𝘵𝘦𝘳𝘯𝘢𝘵𝘪𝘰𝘯𝘢𝘭 𝘑𝘰𝘶𝘳𝘯𝘢𝘭 𝘧𝘰𝘳 𝘕𝘶𝘮𝘦𝘳𝘪𝘤𝘢𝘭 𝘔𝘦𝘵𝘩𝘰𝘥𝘴 𝘪𝘯 𝘌𝘯𝘨𝘪𝘯𝘦𝘦𝘳𝘪𝘯𝘨, 89(2):180–191, 2012.
  6. [6]S. C. Eisenstat and I. C. F. Ipsen. Three absolute perturbation bounds for matrix eigenvalues imply relative bounds. 𝘚𝘐𝘈𝘔 𝘑. 𝘔𝘢𝘵𝘳𝘪𝘹 𝘈𝘯𝘢𝘭., 20(1):149–158, 1998.DOI
  7. [7]A. Greenbaum, R.-C. Li, and M. L. Overton. First-order perturbation theory for eigenvalues and eigenvectors. 𝘚𝘐𝘈𝘔 𝘙𝘦𝘷𝘪𝘦𝘸, 62(2):463–482, 2020.arxiv.org/abs/1903.00785
  8. [8]S. Güttel and F. Tisseur. The nonlinear eigenvalue problem. 𝘈𝘤𝘵𝘢 𝘕𝘶𝘮𝘦𝘳𝘪𝘤𝘢, 26:1–94, 2017.DOI
  9. [9]S. Güttel, R. Van Beeumen, K. Meerbergen, and W. Michiels. NLEIGS: A class of fully rational Krylov methods for nonlinear eigenvalue problems. 𝘚𝘐𝘈𝘔 𝘑𝘰𝘶𝘳𝘯𝘢𝘭 𝘰𝘯 𝘚𝘤𝘪𝘦𝘯𝘵𝘪𝘧𝘪𝘤 𝘊𝘰𝘮𝘱𝘶𝘵𝘪𝘯𝘨, 36(6):A2842–A2864, 2014.
  10. [10]T.-M. Huang, H.-E. Hsieh, W.-W. Lin, and W. Wang. Eigendecomposition of the discrete double-curl operator with application to fast eigensolver for three-dimensional photonic crystals. 𝘚𝘐𝘈𝘔 𝘑𝘰𝘶𝘳𝘯𝘢𝘭 𝘰𝘯 𝘔𝘢𝘵𝘳𝘪𝘹 𝘈𝘯𝘢𝘭𝘺𝘴𝘪𝘴 𝘢𝘯𝘥 𝘈𝘱𝘱𝘭𝘪𝘤𝘢𝘵𝘪𝘰𝘯𝘴, 34(2):369–391, 2013.DOI
  11. [11]T.-M. Huang, T. Li, J.-W. Lin, W.-W. Lin, and H. Tian. Structure-preserving methods for computing complex band structures of three dimensional photonic crystals. Journal of Scientific Computing, 83(2):35, 2020.DOI
  12. [12]T.-M. Huang, W.-W. Lin, and V. Mehrmann. A Newton-type method with nonequivalence deflation for nonlinear eigenvalue problems arising in photonic crystal modeling. SIAM Journal on Scientific Computing, 38(2):B191–B218, 2016.DOI
  13. [13]E. Jarlebring, A. Koskela, and G. Mele. Disguised and new quasi-Newton methods for nonlinear eigenvalue problems. Numerical Algorithms, 79:311–335, 2018.arxiv.org/abs/1702.08492
  14. [14]E. Jarlebring, A. Kressner, and K. Meerbergen. A linear eigenvalue algorithm for the nonlinear eigenvalue problem. Numerische Mathematik, 122(1):169–195, 2012.
  15. [15]C. Jin, Y. Xia, and Y. Xu. Kernel compensation method for Maxwell eigenproblem in photonic crystals with mimetic finite difference discretizations. Numerical Methods for Partial Differential Equations, 41(2):e23171, 2025.DOI
  16. [16]C. Jin and H. Xie. Kernel compensation-based Newton iteration for nonlinear Hermitian eigenproblems in photonic crystals. Mathematicа Numerica Sinica, 2026. In press.arxiv.org/abs/2611.25791
  17. [17]C. Jin and H. Xie. A robust GPU-accelerated kernel compensation solver with novel discretization for photonic crystals in anisotropic media. Preprint submitted to Computer Methods in Applied Mechanics and Engineering, 2026.DOI
  18. [18]L. Lu, L. Fu, J. D. Joannopoulos, and M. Soljačić. Weyl points and line nodes in gyroid photonic crystals. Nature Photonics, 7:294–299, 2013.DOI
  19. [19]X.-L. Lyu, T. Li, T.-M. Huang, J.-W. Lin, W.-W. Lin, and S. Wang. FAME: Fast algorithms for Maxwell’s equations for three-dimensional photonic crystals. ACM Transactions on Mathematical Software, 47(3):26:1–26:24, 2021.DOI
  20. [20]A. Raman and S. Fan. Photonic band structure of dispersive metamaterials formulated as a Hermitian eigenvalue problem. Physical Review Letters, 104(8):087401, 2010.
  21. [21]M. M. Sigalas, C. M. Soukoulis, T. C. Chan, and K. M. Ho. Electromagnetic-wave propagation through dispersive and absorptive photonic-band-gap materials. Physical Review B, 49(16):1080–1087, 1994.pubmed.ncbi.nlm.nih.gov/10009955
  22. [22]G. L. G. Sleijpen and H. A. van der Vorst. A Jacobi–Davidson iteration method for linear eigenvalue problems. SIAM Journal on Matrix Analysis and Applications, 17(2):401–425, 1996.
  23. [23]D. B. Szyld and F. Xue. Local convergence analysis of several inexact Newton-type algorithms for general nonlinear eigenvalue problems. Numerische Mathematik, 123:333–362, 2013.DOI
  24. [24]H. Tian, T. Li, X.-L. Lyu, and W.-W. Lin. Band structure calculations of three-dimensional anisotropic photonic crystals in the oblique coordinate system. SIAM Journal on Scientific Computing, 45(4):B440–B466, 2023.DOI
  25. [25]W. Xiao and J. Sun. Band structure calculation of photonic crystals with frequency-dependent permittivities. Journal of the Optical Society of America A, 38(5):628–633, 2021.DOI
  26. [26]K. S. Yee. Numerical solution of initial boundary value problems involving Maxwell’s equations in isotropic media. IEEE Transactions on Antennas and Propagation, 14(3):302–307, 1966.DOI

Paper details

vlnc:W_wprd44st
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

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

  • Mathematics

Codex

View statement
Looking for related papers…
Contents