Elsevier

Journal of Computational Physics

Volume 376, 1 January 2019, Pages 1232-1252
Journal of Computational Physics

On a refinement-free Calderón multiplicative preconditioner for the electric field integral equation

https://doi.org/10.1016/j.jcp.2018.10.009Get rights and content

Highlights

  • A well-conditioned electric field integral equation (EFIE).

  • EFIE system is Hermitian and positive definite.

  • No second discretization of the EFIE with dual basis functions.

  • No need for cycle detection on multiply connected geometries.

Abstract

We present a Calderón preconditioner for the electric field integral equation (EFIE), which does not require a barycentric refinement of the mesh and which yields a Hermitian, positive definite (HPD) system matrix allowing for the usage of the conjugate gradient (CG) solver. The resulting discrete equation system is immune to the low-frequency and the dense-discretization breakdown and, in contrast to existing Calderón preconditioners, no second discretization of the EFIE operator with Buffa–Christiansen (BC) functions is necessary. This preconditioner is obtained by leveraging on spectral equivalences between (scalar) integral operators, namely the single layer and the hypersingular operator known from electrostatics, on the one hand, and the Laplace–Beltrami operator on the other hand. Since our approach incorporates Helmholtz projectors, there is no search for global loops necessary and thus our method remains stable on multiply connected geometries. The numerical results demonstrate the effectiveness of this approach for both canonical and realistic (multi-scale) problems.

Introduction

The electric field integral equation (EFIE), which is used for solving electromagnetic scattering and radiation problems, results in an ill-conditioned linear system of equations when discretized with Rao–Wilton–Glisson (RWG) basis functions—the common choice in standard codes. The ill-conditioning stems from two issues: the low-frequency breakdown, which is due to the different scaling of the vector and the scalar potential in frequency, and the dense-discretization breakdown, which is due to the fact that the vector and the scalar potential are pseudo-differential operators of negative and positive order [1], a property that leads in general to ill-conditioned system matrices. Altogether, the condition number of the EFIE system matrix grows as (kh)2, where k is the wavenumber and h is the average edge length of the mesh, leading to slowly or non-converging iterative solvers [2].

The low-frequency breakdown has been overcome in the past by using explicit quasi-Helmholtz decompositions such as the loop-star or the loop-tree decomposition [3], [4], [5], [6], [7]. While these decompositions cure the low-frequency breakdown, the dense-discretization breakdown persists and is even worsened to a 1/h3 scaling of the condition number in the case of the loop-star decomposition [8]. Some other low-frequency stable methods have been proposed in the past such as the augmented EFIE, the rearranged loop-tree decomposition, or the augmented EFIE with normally constrained magnetic field and static charge extraction [9], [10], [11], yet all of them suffer from h-ill-conditioning.

Different strategies have been presented to overcome the dense-discretization breakdown. A first class of techniques relies on algebraic strategies such as the incomplete LU factorization [12], [13], sparse approximate inverse [14], [15], or near-range preconditioners [16]. While they improve the conditioning, the condition number still grows with decreasing h.

In recent years more elaborate explicit quasi-Helmholtz decompositions have been presented, the so-called hierarchical basis preconditioners for both structured [17], [18], [19], [20] and unstructured meshes [21], [22] (in the mathematical community, the hierarchical basis preconditioners are typically termed multilevel or prewavelet preconditioners). The best of these methods can yield a condition number that only grows logarithmically in 1/h [20], [22] meaning that there is still space left for improvement.

In fact, Calderón identity-based preconditioners yield a condition number that has an upper bound independent from h [23], [24], [25], [26], [27]. In the static limit, however, the Calderón strategies stop working due to numerical cancellation in both the right-hand side excitation vector and the unknown current, since solenoidal and non-solenoidal components scale differently in k [28]. Explicit quasi-Helmholtz decompositions do not suffer from this cancellation since the solenoidal and non-solenoidal components are stored separately. To make Calderón preconditioners stable in the static limit, one could combine the Calderón multiplicative preconditioner (CMP) with an explicit quasi-Helmholtz decomposition.

Such an approach has a severe downside: if the geometry is multiply connected, then the quasi-harmonic global loop functions have to be added to the basis of the decomposition [6]. In contrast to loop, star, or tree functions, (or any of the hierarchical bases mentioned here), the construction of the global loops becomes costly if the genus g is proportional to the number of unknowns N resulting in the overall complexity O(N2log(N)), where N is the number of unknowns (see, for example, the discussion in [29]). In order to avoid the construction of the global loops, a modified CMP has been presented which leverages on an implicit quasi-Helmholtz decomposition based on projectors [30]. These projectors require the application of the inverse primal (i.e., cell-based) and the inverse dual (i.e., vertex-based) graph Laplacian, a task for which blackbox-like preconditioners such as algebraic multigrid methods can be used for rapidly obtaining the inverse.

In order to avoid the inversion of graph Laplacians, one could think that an alternative might be the Calderón preconditioner combined with an explicit loop-star quasi-Helmholtz decomposition as described in the penultimate paragraph, at least if no global loops are present. However, the inverse Gram matrices appearing in such a scheme are all spectrally equivalent to discretized Laplace–Beltrami operators. Different from the graph Laplacians, these Gram matrices are not symmetric since the loop-star basis is applied to a mixed Gram matrix, that is, Buffa–Christiansen (BC) functions are used as expansion and rotated RWG functions are used as testing functions [8], [30]. In general, this complication makes it more challenging to stably invert these Gram matrices compared with the graph Laplacians of the Helmholtz projectors since, for example, many algebraic multigrid preconditioners require symmetric matrices.

The work in [25], [30] demonstrated a Calderón scheme that can be relatively easily integrated in existing codes. Instead of discretizing the operator on the standard mesh with RWG and BC functions, only a single discretization with RWG functions on the barycentrically refined mesh is necessary. The disadvantage of this approach is that the memory consumption as well as the costs for a single-matrix vector product are increased by a factor of six.

In this work, we propose a refinement-free Calderón multiplicative preconditioner (RF-CMP) for the EFIE. In contrast to existing Calderón preconditioners, no BC functions are employed, so that a standard discretization of the EFIE with RWG functions can be used. What is more, we get a system matrix, which is Hermitian, positive definite (HPD). We obtain this result by leveraging on spectral equivalences between the single layer and the hypersingular operator known from electrostatics on the one hand and the Laplace–Beltrami operator on the other hand. Similar to [30], graph Laplacians need to be inverted. Since the new system matrix is HPD, we are allowed to employ the conjugate gradient (CG) solver. Different from other Krylov subspace methods, it guarantees convergence and has the least computational overhead. The numerical results corroborate the new formulation. Preliminary results have been presented at conferences [31], [32].

The paper is structured as follows. Section 2 sets the background and notation; Section 3 introduces the new formulation and provides the theoretical apparatus. Numerical results demonstrating the effectiveness of the new approach are shown in Section 4.

Section snippets

Notation and background

In the following, we denote quantities residing in R3, such as the electric field E, with an italic, bold, serif font. For any other vectors and matrices we use an italic, bold, sans-serif font, and we distinguish matrices from vectors by using capital letters for matrices and minuscules for vectors. The expression ab has to be read as aCb, where C>0 is a constant independent of the mesh parameter h (i.e., the average edge length). Furthermore, ab means that ab and ba holds. Let fnXf and g

New formulation

The formulation for which we are going to show that the system matrix is well-conditioned in the static limit reads where the outer matrix

is with and the middle matrix
is using and with the conjugate transpose
. The unknown vector
is recovered by
.

The use of the imaginary unit +i in the definition of

is motivated for the same reason as it was for
: to prevent the numerical cancellation due to different scaling of the solenoidal and

Numerical results

For the implementation of the preconditioner, we did not use the wavenumber k directly to cure the low-frequency breakdown. Instead, we used the following definitions where These norms are estimated using the power iteration algorithm. Typically the condition number obtained by using norms is lower than usingα=k,β=1/k,γ=k; thereby, the number of iterations used by a Krylov subspace method is reduced (and this saving usually outweighs the costs for estimating the norms).

First, we considered a

Conclusion

We presented a preconditioner for the EFIE that yields a Hermitian, positive definite, and well-conditioned system matrix. Due to the applicability of the CG method, there is a—at least theoretically—guaranteed convergence and the complexity for obtaining such a solution is O(NlogN) if a fast method is used to compress the EFIE system matrix (and if it is assumed that the algebraic multigrid methods allow to solve Laplacian systems in O(N) complexity). Preliminary results indicate that an

Acknowledgement

This work has been funded in part by the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation program (ERC project 321, grant No. 724846).

References (48)

  • S.B. Adrian et al.

    A hierarchical preconditioner for the electric field integral equation on unstructured meshes based on primal and dual Haar bases

    J. Comput. Phys.

    (2017)
  • R. Hiptmair

    Operator preconditioning

    Comput. Math. Appl.

    (2006)
  • J.-C. Nédélec

    Acoustic and Electromagnetic Equations: Integral Representations for Harmonic Problems

    (2001)
  • Y. Saad

    Iterative Methods for Sparse Linear Systems

    (2003)
  • D.R. Wilton et al.

    On improving the stability of the electric field integral equation at low frequencies

  • W.-L. Wu et al.

    A study of two numerical solution procedures for the electric field integral equation

    Appl. Comput. Electromagn. Soc. J.

    (1995)
  • M. Burton et al.

    A study of a recent, moment-method algorithm that is accurate to very low frequencies

    Appl. Comput. Electromagn. Soc. J.

    (1995)
  • G. Vecchi

    Loop-star decomposition of basis functions in the discretization of the EFIE

    IEEE Trans. Antennas Propag.

    (1999)
  • G. Miano et al.

    A surface integral formulation of Maxwell equations for topologically complex conducting domains

    IEEE Trans. Antennas Propag.

    (2005)
  • F. Andriulli

    Loop-star and loop-tree decompositions: analysis and efficient algorithms

    IEEE Trans. Antennas Propag.

    (2012)
  • J.-S. Zhao et al.

    Integral equation solution of Maxwell's equations from zero frequency to microwave frequencies

    IEEE Trans. Antennas Propag.

    (2000)
  • Z.G. Qian et al.

    An augmented electric field integral equation for high-speed interconnect analysis

    Microw. Opt. Technol. Lett.

    (2008)
  • J. Cheng et al.

    Augmented EFIE with normally constrained magnetic field and static charge extraction

    IEEE Trans. Antennas Propag.

    (2015)
  • K. Sertel et al.

    Incomplete LU preconditioner for FMM implementation

    Microw. Opt. Technol. Lett.

    (2000)
  • B. Carpentieri et al.

    Symmetric inverse-based multilevel ILU preconditioning for solving dense complex non-Hermitian systems in electromagnetics

    Prog. Electromagn. Res.

    (2012)
  • B. Carpentieri et al.

    Combining fast multipole techniques and an approximate inverse preconditioner for large electromagnetism calculations

    SIAM J. Sci. Comput.

    (2005)
  • X.-M. Pan et al.

    Sparse approximate inverse preconditioner for multiscale dynamic electromagnetic problems

    Radio Sci.

    (2014)
  • O. Wiedenmann et al.

    The effect of near-zone preconditioning on electromagnetic integral equations of first and second kind

    Adv. Radio Sci.

    (2013)
  • F. Vipiana et al.

    A multiresolution method of moments for triangular meshes

    IEEE Trans. Antennas Propag.

    (2005)
  • F.P. Andriulli et al.

    A multiresolution approach to the electric field integral equation in antenna problems

    SIAM J. Sci. Comput.

    (2007)
  • R.-S. Chen et al.

    A multiresolution curvilinear Rao–Wilton–Glisson basis function for fast analysis of electromagnetic scattering

    IEEE Trans. Antennas Propag.

    (2009)
  • F.P. Andriulli et al.

    Solving the EFIE at low frequencies with a conditioning that grows only logarithmically with the number of unknowns

    IEEE Trans. Antennas Propag.

    (2010)
  • F.P. Andriulli et al.

    Hierarchical bases for nonhierarchic 3-D triangular meshes

    IEEE Trans. Antennas Propag.

    (2008)
  • S.H. Christiansen et al.

    A preconditioner for the electric field integral equation based on Calderon formulas

    SIAM J. Numer. Anal.

    (2002)
  • Cited by (28)

    • Manufactured solutions for the method-of-moments implementation of the electric-field integral equation

      2021, Journal of Computational Physics
      Citation Excerpt :

      While the quadrature error can be driven low enough that mesh-refinement studies become feasible, the presence of this additional error often places a lower bound on the discretization error that can be obtained, and therefore limits the scope of the mesh-refinement study, in addition to being undesirable for the aforementioned reasons. To further complicate matters, for the MoM implementation of the EFIE, the discretized equations can easily become ill-conditioned [39]. For the MoM implementation of the EFIE, numerical error is introduced from three sources:

    • On almost automorphic type solutions of abstract integral equations, a Bohr-Neugebauer type property and some applications

      2021, Journal of Mathematical Analysis and Applications
      Citation Excerpt :

      In the last direction, we mention its applications in the following topics (among others): the heat conduction in materials with memory, transport phenomena in Biological Systems, optimal replacement problems in engineering and production economics, and so on, see for instance the references [31,39,25]. Of course, this justifies the enormous progress in the qualitative and numerical study of them, for the numerical treatment, see the references [5,42,38,40,41] (and references cited therein). In the context of differential, integral or integro-differential equations, it is common to start with the study of periodic problems (as a first instance), and then go across its natural functional generalizations.

    • On the efficiency of nested GMRES preconditioners for 3D acoustic and elastodynamic H-matrix accelerated Boundary Element Methods

      2020, Computers and Mathematics with Applications
      Citation Excerpt :

      The inverse of the preconditioner is computed with an iterative solver and thus leads to an inner–outer preconditioned solver which in practice can be solved through a nested-GMRES method. Contrary to the current trend in this field [33–38], we do not resort to Calderón preconditioning. As a result, we do not have to compute an additional integral operator.

    View all citing articles on Scopus
    View full text