Preconditioning Frequency in Wave Simulations

• Preconditioning frequency is the study of designing methods that mitigate frequency-dependent spectral pathologies in linear and nonlinear operator equations.
• It involves operator-theoretic, algebraic, and data-driven strategies to enhance iterative solver convergence across low- and high-frequency regimes.
• These techniques are crucial for stabilizing numerical methods in wave propagation, such as Helmholtz and Maxwell’s equations, ensuring mesh- and frequency-independent performance.

Preconditioning frequency refers to the study and design of preconditioning strategies for linear and nonlinear operator equations whose conditioning and solver convergence depend strongly on the physical or discretization frequency parameter (e.g., wavenumber $k$ or angular frequency $\omega$). This topic is central to computational wave propagation, electromagnetics, and high-frequency numerical acoustics, where both low- and high-frequency regimes can dramatically affect the efficiency and robustness of iterative solvers. The subject subsumes not only the design of operator-based or algebraic preconditioners for frequency-domain discretizations (e.g., boundary or volume integral equations, finite elements for the Helmholtz or Maxwell’s equations) but also the analysis of how such preconditioners stabilize the spectral properties of the system across different frequency and mesh refinement regimes.

1. Conditioning Pathologies in Frequency-Domain Operators

Discretizations of wave propagation equations such as the Helmholtz and Maxwell's equations, acoustic transmission boundary-integral equations, and related mixed-field formulations inherit frequency-dependent spectral pathologies. As frequency increases, the spectral radius and indefinite part of system matrices also rise, leading to slower convergence or stagnation of Krylov and multigrid solvers, especially in absence of absorption. Conversely, in the low-frequency limit, operators such as EFIE, PMCHWT, and CFIE suffer from breakdowns associated to vanishing diagonal blocks, leading to deteriorating condition numbers, spurious null-spaces, and loss of accuracy (Dély et al., 2020, Giunzioni et al., 2024).

Typical unpreconditioned system behaviors include:

• For EFIE/PMCHWT, $\operatorname{cond}(Z(\omega)) \sim 1/(h k)^2$ as $k\to 0$, diverging for small frequency (Rahmouni et al., 2020, Guzman et al., 2016).
• For CFIE, eigenvalues tend to zero or infinity with $k$ in both static and high-frequency regimes even at fixed $kh$ (Chhim et al., 2020).
• In finite or boundary element form, ill-conditioning also deteriorates with mesh refinement ($h \to 0$) due to the singular integral structure and lack of spectral separation.

2. Preconditioning Strategies and Frequency-Robust Design

A range of frequency-robust preconditioning techniques has been developed to neutralize these spectral pathologies. These methods can be classified broadly into operator-theoretic, algebraic, and data-driven approaches with precise frequency-scaling, mode projection, or normalization mechanisms.

A. Operator-Aware and Integral Preconditioners

• Helmholtz-type shifted Laplacian preconditioners introduce absorption or complex shifts to move the spectrum away from the origin, enabling optimal domain decomposition or multigrid convergence for $\varepsilon\sim k^2$ (Graham et al., 2015, Spence, 18 Apr 2025).
• On-surface radiation condition (OSRC) preconditioners utilize pseudo-differential operator approximations (surface square-root operators) to mimic Dirichlet-to-Neumann or Neumann-to-Dirichlet maps, yielding spectra clustered near unity for all frequencies (Wout et al., 2021).
• Helmholtz-domain frame or phase-space adapted preconditioners diagonalize the symbol in ray coordinates or tight frames, providing uniform iteration counts with respect to $k$ (Stolk, 2010).

B. Hierarchical and Spectral Projection Approaches

• Loop-Star, quasi-Helmholtz projectors, and algebraic spectral filters split the current or field unknowns into solenoidal and irrotational components and apply frequency- and mesh-scaled normalization to each block. The quasi-Helmholtz strategy, e.g., ensures no frequency-induced condition blow-up even for multiply connected topologies and diverse conduction regimes (Dély et al., 2020, Giunzioni et al., 2024).
• Hierarchical basis preconditioners scale hierarchical levels by powers of $\omega$ to equilibrate block-wise scaling (e.g., $\alpha(\omega) = \beta(\omega) = \omega^{1/2}$ for PMCHWT and EFIE), yielding uniformly bounded condition numbers versus both frequency and mesh (Guzman et al., 2016).

C. Spectral-Clustering, Hybrid, and Learning-Based Preconditioning

• Circulant and block-circulant FFT-based preconditioners exploit underlying Toeplitz structures and diagonalize the major part of the system in frequency space, ensuring frequency-independent iteration counts for very long photonic structures (Groth et al., 2019).
• DeepONet-based and operator-aware neural preconditioners leverage the spectral bias of neural networks (favoring low-frequency modes) to correct or project out components either missed by classical relaxations or over-emphasized by the operator, restoring balanced convergence across frequencies in PINNs (He et al., 1 Mar 2026, Kopaničáková et al., 2024).

3. Theoretical Analysis: Frequency-Uniform Conditioning

Several analytic frameworks and condition number bounds now underpin the efficacy of frequency-preconditioned formulations.

• Coercivity and norm/field-of-values estimates for domain decomposition and shifted-Laplacian preconditioners provide strict sufficient conditions for frequency-independent GMRES convergence (e.g., $|\varepsilon|\sim k^2$ and subdomain sizes $H\sim k^{-1}$ yield $O(1)$ iterations for the Helmholtz equation) (Graham et al., 2015).
• Spectral equivalence and cancellation of frequency-scaling arises when preconditioners mimic the symbol of the original operator: e.g., preconditioning by the Green's function of a reference operator in PINNs cancels high-frequency amplification, restoring per-mode learning rates to those of direct $L^2$ approximation (He et al., 1 Mar 2026).
• Closed-form frequency-dependent scaling in PMCHWT and quasi-Helmholtz methods produces uniform Gershgorin bounds, eliminating $O(\omega^{-p})$ blow-up. For instance, quasi-Helmholtz preconditioning yields $\operatorname{cond}(Z_P)=O(1)$ as both $\omega\to 0$ and $\sigma\to\infty$ (Giunzioni et al., 2024).

Illustrative table: frequency-scaling of condition numbers for PMCHWT (toroidal scatterer, $\sigma=1$ S/m):

$f$ [Hz]	cond($Z$)	cond($\bar{Z}$)	cond($Z_P$)
$1\text{e}0$	$2.3\text{e}3$	$1.9\text{e}3$	$12$
$1\text{e}-1$	$2.1\text{e}4$	$1.7\text{e}4$	$13$
$1\text{e}-2$	$2.0\text{e}5$	$1.6\text{e}5$	$14$
$1\text{e}-3$	$2.0\text{e}6$	$1.5\text{e}6$	$15$

4. Practical Preconditioning Schemes: Algorithms and Applications

Frequency-robust preconditioners have been systematically tested across multiple regimes and physical models.

• PMCHWT stabilization: Algebraic quasi-Helmholtz projectors with regime-adaptive scaling coefficients deliver $O(1)$ conditioning across static, eddy-current, and high-conductivity transitions, extending seamlessly to multiply connected and multiscale geometries without combinatorial overhead (Giunzioni et al., 2024).
• EFIE/CFIE normalization: Helmholtz-type operators and balanced block preconditioning, employing surface Laplace–Beltrami regularizations with tailored complex shifts, freeze the spectral envelope of the operator around unity, independent of $k$ and $h$ (Chhim et al., 2020, Rahmouni et al., 2020).
• Helmholtz equation: Multilevel domain decomposition with coarse grid corrections and shifted absorption ensures nearly flat iteration counts in large-scale, high-frequency seismic inversion and wave simulation (Graham et al., 2015, Dolean et al., 2020).
• Learning-based and hybrid multilevel solvers: DeepONet trunk-basis corrections and standard smoothers combine multigrid and data-driven strengths—low-frequency errors are targeted by neural networks exploiting spectral bias, while classical relaxation damps unresolved high frequencies, yielding mesh-independent iterations even for a wide range of parameteric PDEs (Kopaničáková et al., 2024).

5. Spectral and Mode-Wise Interpretation

Frequency preconditioning modifies the per-mode response of the linear or nonlinear problem, equilibrating the amplification or suppression inherent to high- or low-frequency components by the original operator:

• In the spectral domain, operator-aware preconditioning can be interpreted as flattening the gradient flow or error-propagation rates across all Fourier or spherical harmonic components, leading to mode-wise learning or solver speeds that are frequency-independent (He et al., 1 Mar 2026, Stolk, 2010).
• Frame-based or phase-space preconditioners, through tight frames constructed to diagonalize the Helmholtz symbol, achieve bounded spectral condition numbers uniformly in frequency, confirmed numerically by iteration plateaus as $\omega\to\infty$ (Stolk, 2010).
• Graph Laplacian spectral filter preconditioners for EFIE achieve band-wise frequency neutralization by explicit dyadic scaling, resulting in frequency- and mesh-independent condition numbers (e.g., cond~3 independent of up to 40 decades of frequency and mesh refinement) (Rahmouni et al., 2020).

6. Extensions, Implementation, and Computational Scaling

Robustness to frequency is maintained in practical implementations via algebraic multigrid or fast-matrix algorithms. Major features include:

• Compatibility with fast multipole methods, $\mathcal{H}$-matrices, and adaptive cross-approximation for large-scale integral equations (Giunzioni et al., 2024).
• Memory and computational overheads are offset by blockwise or sparse application of projectors, with per-iteration costs scaling at most logarithmically with system size, or $O(n\log n)$ for frame and circulant approaches (Stolk, 2010, Groth et al., 2019).
• In hybrid schemes, the frequency-dependent split between coarse neural and fine classical correction can be adjusted according to observed error-propagation or spectral gap (Kopaničáková et al., 2024).

7. Future Directions and Open Problems

Despite strong progress, several challenges and opportunities remain:

• High-frequency integral equation preconditioning in presence of spurious resonances or highly inhomogeneous coefficients is not yet fully solved and continues to motivate operator-adaptive and domain-decomposition innovations (Dély et al., 2020).
• In data-driven and PINN solvers, practical frequency-aware preconditioning strategies must address curse-of-dimensionality effects in sampling and kernel evaluation (He et al., 1 Mar 2026).
• For multiphysics and constrained-inverse problems, integration of frequency-scaled preconditioners for multiple operators simultaneously (e.g., joint diffusion/Helmholtz) remains an area of active research.

In sum, preconditioning frequency is both a demanding theoretical problem and an essential practical engineering consideration; robust methodologies must address the explicit frequency dependence of the operator spectrum, and effective preconditioners exploit operator symbol, mode decomposition, and spectral equilibria to deliver mesh- and frequency-independent convergence across the full range of models encountered in computational science and engineering.