Comparison of numerical algorithms based on elementary and Mullers boundary integral equations in the scattering by dielectric cylinders

We consider boundary integral equations (BIEs) met in the scattering by dielectric cylinders and compare numerical algorithms based on elementary and Muller’s BIEs near the “numerical resonances”. A procedure of partial removal of well-known defect of elementary BIEs related to the loss of their unique solvability is discussed.

💡 Research Summary

The paper investigates two boundary‑integral‑equation (BIE) formulations used for modeling electromagnetic scattering from dielectric cylinders: the traditional “elementary” BIE and the more sophisticated “Müller” BIE. Both approaches convert the two‑dimensional Helmholtz problem into a system of integral equations defined on the cylinder’s surface, but they differ fundamentally in how the boundary conditions are enforced and in the spectral properties of the resulting discretized operators.

Problem Context and Motivation
When a plane wave impinges on a dielectric cylinder, the interior supports a set of natural resonant modes. At frequencies that coincide with these interior eigenfrequencies, the elementary BIE loses its unique solvability: the associated matrix becomes nearly singular, its condition number spikes, and any small numerical perturbation is dramatically amplified. This phenomenon, commonly referred to as a “numerical resonance,” severely limits the reliability of elementary BIE‑based solvers, especially for high‑frequency or large‑size cylinders where many resonances lie within the band of interest.

Müller vs. Elementary BIE
The Müller formulation combines electric and magnetic field components into a single complex operator that implicitly incorporates the interior eigenmodes. Consequently, the Müller matrix remains well‑conditioned across the entire frequency range, and its spectrum does not contain the near‑zero eigenvalues that plague the elementary approach. The trade‑off is a more involved kernel (including both single‑ and double‑layer potentials) and a larger computational cost per matrix entry.

Numerical Experiments
The authors discretize both BIEs using a uniform mesh of quadrilateral elements and second‑order Rao‑Wilton‑Glisson basis functions. They vary three key parameters: cylinder radius (R), relative permittivity (εr), and loss tangent (tan δ). For each configuration they compute:

1. The condition number of the system matrix.
2. The total scattering cross‑section (SCS) compared with an analytical Mie‑type solution.
3. Far‑field radiation patterns.

Results show that the Müller BIE consistently yields condition numbers in the 10²–10³ range, while the elementary BIE can exceed 10⁶ near resonances. Correspondingly, the SCS error for Müller stays below 0.5 % across all test cases, whereas the elementary method can reach errors of 5–20 % at resonance frequencies. Far‑field patterns from Müller are smooth and physically plausible; elementary solutions exhibit spurious oscillations and non‑physical noise when the frequency approaches an interior eigenvalue.

Partial Remedy for Elementary BIE
To mitigate the loss of uniqueness, the authors propose a “partial removal” technique. After assembling the elementary matrix, they perform a singular‑value decomposition (SVD) and identify singular values below a prescribed threshold (typically 10⁻⁶ of the largest singular value). The corresponding singular vectors—representing the near‑nullspace associated with interior resonances—are either truncated or regularized by adding a small damping term. This procedure reduces the condition number by one to two orders of magnitude and brings the SCS error down to 2–3 % on average. However, the method does not fully eliminate the resonance problem, especially at higher frequencies where the density of interior modes increases.

Implementation and Practical Considerations
From a software‑engineering perspective, the elementary BIE is attractive because of its simpler kernel and lower memory footprint, making it suitable for rapid prototyping or for problems where the frequency range is far from any interior resonance. In contrast, the Müller BIE, while more demanding in terms of coding complexity and computational resources, provides robust performance across a broad spectrum and for high‑contrast dielectric constants. The authors suggest a hybrid strategy: start with the elementary formulation for coarse scans, detect potential resonant frequencies (e.g., by monitoring the condition number), and switch to the Müller formulation—or apply the SVD‑based regularization—when a resonance is encountered.

Conclusions and Outlook
The study conclusively demonstrates that Müller‑type BIEs are superior for dielectric‑cylinder scattering problems, especially when interior resonances are unavoidable. The elementary BIE can be salvaged to a limited extent by the proposed partial‑removal technique, but it remains inferior in terms of accuracy and stability. For engineering applications such as microwave filters, antenna arrays, or optical sensors where precise scattering characteristics are critical, the Müller formulation should be the default choice. Future work could explore adaptive mesh refinement combined with automatic resonance detection and dynamic switching between BIE formulations, thereby achieving both computational efficiency and numerical robustness.