Two-grid discretizations and a local finite element scheme for a non-selfadjoint Stekloff eigenvalue problem

Hai Bi, Yu Zhang, Yidu Yang

This paper, recently published by researchers from Guizhou Normal University, tackles a computationally demanding problem in spectral geometry and inverse scattering: the non-selfadjoint Stekloff eigenvalue problem. By leveraging "two-grid" and "local" finite element methods, the authors demonstrate how to achieve high-precision results for complex physical systems at a fraction of the usual computational cost.

TL;DR

The authors develop efficient numerical schemes for a challenging class of non-selfadjoint Stekloff eigenvalue problems that lack standard ellipticity. Their approach uses a coarse-grid eigenvalue solve followed by fine-grid linear corrections, significantly reducing CPU time while maintaining optimal accuracy, even in the presence of geometric singularities.

What Problem Does This Paper Address?

Stekloff eigenvalue problems are essential in physics, particularly in inverse scattering, where researchers try to determine the properties of an object (like its refractive index) based on how it scatters waves.

Most existing mathematical tools for these problems assume the system is "selfadjoint" (symmetric) and "$H^1$-elliptic" (well-behaved energy). However, real-world scattering often involves complex refractive indices, leading to non-selfadjoint problems. These are much harder to solve because:

  1. The eigenvalues can be complex.
  2. Standard finite element method (FEM) convergence proofs often fail.
  3. Solving large-scale eigenvalue problems on very fine meshes is computationally "expensive" in terms of memory and time.

Methods

The authors propose three distinct schemes to bypass the "curse of dimensionality" in fine-mesh calculations:

  • Two-Grid Discretization (Schemes 1 & 2): Instead of solving a massive eigenvalue problem on a fine mesh (size $w$), they solve it on a much coarser mesh (size $H$). They then use that coarse solution to solve two simple linear boundary value problems on the fine mesh.
  • The Innovation: Scheme 2 is particularly clever because it reformulates the fine-grid step to ensure the resulting matrices are symmetric and definite, allowing for much faster solvers.
  • Local Finite Element Scheme (Scheme 3): For domains with sharp corners or slits (singularities), global refinement is wasteful. This scheme solves the problem on a coarse grid, performs a "mesoscopic" global correction, and then applies a final high-resolution correction only in the local region containing the singularity.
  • Theoretical Framework: They utilize Gårding’s inequality to handle the lack of $H^1$-ellipticity and prove convergence in the "negative" Sobolev space $H^{-1/2}(\partial\Omega)$, which is the natural home for Stekloff boundary data.

Datasets

As a theoretical and computational mathematics paper, the "datasets" consist of standard benchmark geometric domains used to test FEM robustness:

  • Square Domain ($\Omega_S$): A smooth case to test optimal convergence rates.
  • L-shaped Domain ($\Omega_L$): Introduces a re-entrant corner, creating a singularity in the solution.
  • Slit Domain ($\Omega_{SL}$): The most extreme case, featuring a crack or slit that severely tests local refinement schemes.
  • Refractive Indices: Tested with both real ($n=4$) and complex ($n=4+4i$) values.

Key Insights and Ideas

  • The "Correction" Power: The paper reinforces the idea that you don't need to solve the eigenvalue problem on a fine grid. Solving the linear source problem on a fine grid is enough to "boost" the accuracy of a coarse eigenvalue to fine-grid levels.
  • Symmetry Recovery: Even though the original problem is non-selfadjoint, Scheme 2 shows that we can still utilize symmetric solvers for the most computationally intensive part of the algorithm.
  • Singularity Handling: The local scheme proves that error from a geometric singularity (like a sharp corner) can be "localized." You don't need a super-fine mesh everywhere to get a highly accurate global eigenvalue; you just need it where the geometry is "sharp."

Key Results

  • Efficiency: In numerical tests, Scheme 2 was significantly faster than direct solving. For a square domain, direct solving took ~110 seconds, while the two-grid Scheme 2 achieved similar accuracy in only ~42 seconds.
  • Accuracy: The authors proved and numerically verified optimal error estimates. For the square domain, they achieved $O(w^2)$ accuracy for eigenvalues.
  • Robustness: The local correction scheme (Scheme 3) successfully recovered accuracy in the L-shaped and Slit domains, where standard global FEM would typically struggle or require massive amounts of memory.

Limitations and Future Work

  • Neumann Eigenvalue Avoidance: The current analysis assumes that the wavenumber $k^2$ is not a Neumann eigenvalue of the domain. If it were, the source problem would not have a unique solution, requiring a different analytical approach.
  • Ascent of Eigenvalues: The most detailed error estimates assume the "ascent" of the eigenvalue is 1 (meaning the algebraic and geometric multiplicities are equal). While common, higher-ascent cases require more complex generalized eigenfunction analysis.
  • Parallelization: While the authors mention that the local scheme can be parallelized for multiple singular points, they did not implement or benchmark a parallel version in this specific paper.

Summary updated

Abstract

In this paper, for a new Stekloff eigenvalue problem which is non-selfadjoint and not $H^1$-elliptic, we establish and analyze two kinds of two-grid discretization scheme and a local finite element scheme. We present the error estimates of approximations of two-grid discretizations. We also prove a local error estimate which is suitable for the case that the local refined region contains singular points lying on the boundary of domain. Numerical experiments are reported finally to show the efficiency of our schemes.
BibTeX
@misc{bi2018two,
  author = {Hai Bi and Yu Zhang and Yidu Yang},
  title = {{Two-grid discretizations and a local finite element scheme for a non-selfadjoint Stekloff eigenvalue problem}},
  year = {2018},
  eprint = {1806.05062},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1806.05062},
}
Figure