Spectral Filtering of 3D Integral Operators Using Modified Green's Functions

Technical Summary: Spectral Filtering of 3D Integral Operators Using Modified Green's Functions

Problem Statement
The discretization of integral operators, such as the Electric Field Integral Equation (EFIE), via the Boundary Element Method (BEM) results in dense systems of equations. These systems impose high memory and computational costs, necessitating the use of iterative or fast direct solvers. However, the convergence speed and precision of these solvers are heavily dependent on the spectral properties of the underlying integral operators. While operator filtering has proven effective for regularizing and compressing operators in two-dimensional (2D) spaces (e.g., 2D EFIE), extending these strategies to three-dimensional (3D) problems remains a challenge. Existing 3D approaches often rely on quasi-Helmholtz decompositions, but there is a need for analytical filtering methods based on the spectral truncation of integral operator kernels specifically tailored for 3D geometries.

Methodology
The authors propose a spectral filtering strategy for 3D integral operators associated with the EFIE. The core of the methodology involves modifying the Green's function kernel, $g_k(|r-r'|)$, by truncating its spectral representation.

1. Spectral Representation: Leveraging the spherical symmetry of the Green's function, the authors utilize the Spherically Symmetric Fourier Transform (SSFT), also known as the spherical Hankel transform, to map the kernel into the spectral domain.

   • Static Case ($k=0$): The spectral representation is derived directly from the Fourier transform of the static Green's function ($G_0(s) = 1/s^2$).
   • Dynamic Case ($k>0$): The spectral representation is obtained using the Sokhotsky-Plemelj theorem, separating the real part (Cauchy principal value) and the imaginary part (Dirac delta function representing propagating waves).

2. Spectral Truncation: A filtering parameter $\alpha$ (a cutoff spatial frequency) is introduced. The integral representation of the Green's function in the spectral domain is truncated at $\alpha$.

   • For the static case, this yields a filtered kernel expressed via the sine integral function: $g^\alpha_0(|r-r'|) = \frac{1}{2\pi^2|r-r'|} \text{Si}(\alpha|r-r'|)$.
   • For the dynamic case, the truncated spectral representation is evaluated numerically to define the filtered kernel $g^\alpha_k$.

3. Operator Definition and Discretization: New filtered integral operators, $S^\alpha_k$ (single-layer) and $N^\alpha_k$ (hypersingular), are defined using the modified kernels. These operators are subsequently discretized using standard BEM techniques (Galerkin method with RWG and dual pyramid basis functions) to generate matrices $S^\alpha_k$ and $N^\alpha_k$.

4. Spectral Analysis: The authors develop a semi-analytical procedure to compute the eigenvalues of the filtered operators over a unit sphere. By applying the Funk-Hecke theorem, they derive an integral expression for the eigenvalues of the filtered single-layer operator, allowing for a direct comparison with the unfiltered operator's known eigenvalues.

Key Contributions and Results

• Analytical Foundation: The paper establishes an analytical foundation for 3D operator filtering by deriving the spectral representation of the Green's function via spherical Hankel transforms and demonstrating how spectral truncation leads to well-defined modified kernels.
• Semi-Analytical Eigenvalue Computation: A procedure is outlined to compute the eigenvalues of boundary integral operators with modified kernels, facilitating a comparison between standard and filtered operators before discretization.
• Spectral Properties: Numerical and semi-analytical results demonstrate that the filtering technique effectively alters the spectral properties of the operators:
  • Eigenvalues: The eigenvalues of the filtered single-layer operator ($S^\alpha_k$) are shown to differ significantly from the unfiltered operator ($S_k$), particularly in the high-frequency range, as illustrated for a unit sphere.
  • Singular Values: In the static case ($k=0$), the singular values of the discretized filtered matrix $S^\alpha_0$ exhibit a different distribution compared to the standard matrix $S_0$.
  • Calderón Preconditioning: The study extends the analysis to Calderón-preconditioned EFIE operators ($G^{-T} S^\alpha_k G^{-1} N$). Results indicate that the filtering procedure remains effective within this preconditioning framework, modifying the singular value distribution in a manner that suggests potential benefits for solver performance.

Significance and Claims
The paper claims to extend the applicability of operator filtering from 2D to 3D spaces by utilizing the geometric properties of the Green's function. The authors posit that this approach provides a rigorous analytical basis for regularizing 3D integral operators through spectral domain truncation. The work provides evidence that this filtering strategy impacts the spectral properties of both continuous integral operators and their discrete boundary element counterparts in both static and dynamic regimes. The results suggest that the technique is viable for improving the spectral characteristics of operators used in iterative solvers, including those employing Calderón preconditioning, thereby offering a pathway to more efficient solutions for 3D electromagnetic scattering and radiation problems.