High-Frequency Preconditioners for Electromagnetic Integral Equations Based on Helmholtz Regularizations

Technical Summary: High-Frequency Preconditioners for Electromagnetic Integral Equations Based on Helmholtz Regularizations

Problem Statement
The numerical solution of the Electric Field Integral Equation (EFIE) for perfectly electrically conducting (PEC) objects using the Boundary Element Method (BEM) faces significant computational challenges due to ill-conditioning. This ill-conditioning manifests in three distinct regimes:

1. Low-frequency breakdown: Occurs when frequency decreases while discretization density remains constant.
2. Dense-discretization breakdown: Occurs when frequency is constant but the mesh is refined.
3. High-frequency breakdown: Occurs when frequency increases alongside discretization density.

While iterative solvers like GMRES are preferred for large systems, they fail to converge efficiently when the system matrix is ill-conditioned. Previous work (specifically [3]) introduced a stabilization strategy for the EFIE based on quasi-Helmholtz decomposition and shifted Helmholtz operators to address these regimes simultaneously. However, the implementation of this strategy requires the pseudo-inversion of a discretized shifted Helmholtz operator ($H_S$). Standard pseudo-inversion techniques for this operator suffer from the same dense-discretization and high-frequency breakdowns, leading to unsatisfactory computational complexity and preventing the overall EFIE solution from achieving quasi-linear complexity.

Methodology
This paper proposes a novel preconditioning strategy specifically designed to regularize the shifted Helmholtz operator matrix ($H_S$) across all three challenging regimes. The core of the methodology involves applying left- and right-multiplicative preconditioning matrices constructed from a single-layer operator ($S_{k_m}$) associated with a complex wavenumber kernel ($k_m$).

The preconditioned system is formulated as:
$$H_{S}^{prec} = S G_{\tilde{\lambda}p}^{-1} H_S G_{\tilde{\lambda}p}^{-T} S$$
where $S$ is the discretized single-layer operator using patch basis functions, and $G_{\tilde{\lambda}p}$ is a mixed Gram matrix mapping dual pyramid basis functions to patch functions.

The theoretical justification relies on the pseudo-differential orders of the operators involved. The Laplace-Beltrami operator in $H$ has an order of +2, while the single-layer operator $S$ has an order of -1. The composition $S_{k_m} H S_{k_m}$ results in a net pseudo-differential order of 0 ($-1 + 2 - 1$), which theoretically suggests a mesh-independent condition number. Furthermore, the use of a complex wavenumber ($k_m = k - 0.4jk^{1/3}d^{-2/3}$) is critical to prevent spurious resonances and quasi-null eigenvalues, particularly in the high-frequency regime.

The authors validate this approach through a spherical harmonics analysis of the continuous operators on a sphere. This analysis demonstrates that the eigenvalues of the preconditioned operator remain bounded in both the elliptic region (evanescent modes, dense discretization) and the hyperbolic region (propagative modes, high frequency), ensuring stability as the frequency or mesh density changes.

Key Contributions

• Novel Preconditioning Strategy: The paper introduces a specific preconditioning technique for the shifted Helmholtz operator that utilizes a single-layer operator with a complex wavenumber.
• Stabilization Across Regimes: The proposed method stabilizes both the matrix condition number and the number of GMRES iterations in the dense-discretization and high-frequency regimes, addressing limitations of standard pseudo-inversion techniques.
• Complexity Reduction: By stabilizing the preconditioned system, the pseudo-inversion of the shifted Helmholtz operator can be achieved with quasi-linear complexity. This, in turn, enables the numerical solution of the EFIE with the same quasi-linear complexity when coupled with fast solvers like the Fast Multipole Method (FMM).
• Theoretical and Numerical Validation: The work provides a rigorous spectral analysis on a sphere and corroborates these findings with numerical experiments on a spherical geometry.

Results
Numerical experiments were conducted on a sphere ($a=1$ m) using the GMRES algorithm with a relative residual tolerance of $10^{-6}$.

• Dense-Discretization Regime: With a fixed frequency ($k=4$ m$^{-1}$) and varying mesh density ($\lambda/h$ from 5 to 25), the preconditioned system ($H_S^{prec}$) exhibited a stable condition number and a quasi-constant number of iterations. In contrast, the original shifted Helmholtz operator ($H_S$) showed a sharp increase in both metrics as the mesh was refined.
• High-Frequency Regime: With a fixed mesh density ($\lambda/h = 10$) and varying frequency ($k$ from 1 to 15 m$^{-1}$), the preconditioned system maintained stable iteration counts and condition numbers. Conversely, the unpreconditioned system's metrics increased significantly as the electrical size ($ka$) grew.
• Spectral Analysis: Eigenvalue analysis confirmed that the spectrum of the preconditioned operator remains bounded and does not introduce spurious resonances, even as frequency increases.

Significance
The paper claims that this work improves the efficiency of the matrix-vector products required for the EFIE formulation introduced in [3]. By successfully regularizing the shifted Helmholtz operator, the authors enable the solution of the EFIE with quasi-linear complexity, overcoming the computational bottlenecks associated with dense discretization and high frequencies. The authors note that while the current work focuses on theoretical validation on canonical geometries (the sphere), future work will assess the approach on a broader class of practically relevant scatterers and integrate fast algorithms like FMM to fully realize the acceleration potential.