Some finite element methods compute the vibration frequencies of a domain (the eigenvalues of an elliptic operator such as the Laplacian) far more accurately than standard theory predicts. The trick is to arrange for the leading discretization errors to cancel. One way is to blend two numerical integration rules. Another is to add small penalty terms to the matrices. On a uniform grid over an interval, square or cube, these tricks raise the eigenvalue error rate from the textbook $\mathcal{O}(h^{2p})$ to $\mathcal{O}(h^{2p+2})$, or even to $\mathcal{O}(h^8)$ for linear elements, at essentially no extra cost. Here $h$ is the mesh size and $p$ the polynomial degree. Every published demonstration, however, uses exactly those idealized synthetic test beds. The cancellation depends on symmetry, and real meshes on real shapes lack it. Nobody has measured how much of the gain survives on graded meshes, curved domains, or variable material coefficients. Nor does anyone know whether the loss is gradual or sudden.
This note makes one observation and builds on it. When a curved or non-uniform element is mapped back to the reference element, the Laplacian becomes an operator with spatially variable coefficients, namely the metric of the mapping. The loss of superconvergence on irregular geometry is therefore the same failure the authors of Isogeometric spectral approximation for elliptic differential operators already observed for a variable Schrödinger potential, where "the two-extra orders of convergence are not ensured." We propose to measure this failure systematically. We then propose to repair it with element-local blending weights that respond to how fast the pulled-back coefficient varies inside each element, and we test the repaired method on geometries that are not unit squares.
Core Claim
Our hypothesis comes in two parts, and both can be falsified. The first part is a degradation law. Suppose the mesh distortion is smooth, so that neighbouring elements differ in size and shape by a relative amount of order $h$, as with graded meshes and smooth isoparametric maps. Then a single globally fixed blending weight should lose only part of its gain, because the uncancelled error term scales with the coefficient gradient times one extra power of $h$. Suppose instead the distortion is rough, with node jitter of fixed relative size $\delta$ that does not shrink under refinement. Then the extra orders should vanish entirely, and the error should fall back to $\mathcal{O}(h^{2p})$ with a constant that grows with $\delta$. The second part is a remedy. On each element $e$, let $a_e(\xi)$ be the pulled-back coefficient, which is $\kappa(x(\xi))/J_e(\xi)$ in one dimension, where $\kappa$ is the physical coefficient and $J_e$ the Jacobian of the element map. We choose that element's blending weight $\alpha_e$ from a precomputed table indexed by the dimensionless gradient
$$ g_e = \frac{h_e\, |\nabla a_e|}{a_e}, $$
where $h_e$ is the element size. The table is built offline by requiring that, on a patch whose coefficient varies linearly, the discrete dispersion relation match the continuous one through the next order. We expect this to restore $\mathcal{O}(h^{2p+2})$ on smoothly distorted meshes and smooth coefficients. We expect it to do nothing for rough jitter or corner singularities, and we treat those regimes as stated limits rather than targets.
The idea differs from the closest existing work in a specific way. Dispersion optimized quadratures for isogeometric analysis and Generalized soft finite element method for elliptic eigenvalue problems derive one closed-form weight for uniform constant-coefficient elements; we supply the method with a weight that varies from element to element. G-Adaptivity learns where to move mesh nodes to reduce forward-problem error; we leave the mesh alone and change only how its integrals are evaluated, for eigenvalue problems. Diagnosing Failure Modes of Neural Operators Across Diverse PDE Families motivates the evaluation design: it found that the method most accurate on clean synthetic data degraded most under structured shifts. We ask whether the same holds for classical superconvergent solvers.
Research Program
The first stage is diagnostic. We implement standard Gauss–Legendre integration, globally blended quadrature, the boundary-penalized outlier-free variant from Outlier removal for isogeometric spectral approximation with the optimally-blended quadratures, and the mass- and stiffness-penalized method (GSFEMBQ) from the generalized soft finite element paper. All run in one isogeometric and finite element code. We then sweep three controlled stress axes. The first is smooth grading of strength $s$. The second is rough node jitter of amplitude $\delta$. The third is curvature through exact geometry maps. For each setting we record the observed convergence order of the first several eigenvalues and the error across the fraction of the spectrum that is accurately resolved. We also record the largest discrete eigenvalue, which sets the stable explicit time step, and the matrix condition number. Together these give an empirical map of where each published guarantee holds.
The second stage builds the local-weight table. In one dimension it depends on $p$ and on the scalar $g_e$. In two dimensions it depends on the anisotropy of the element metric and on its gradient. We fit it once per $p$ from small patch dispersion problems, never from the target problem. We derive the one-dimensional leading-order correction by Taylor expansion wherever it is tractable, and check it against the numerical table. For GSFEMBQ at $p \ge 2$, the authors already say closed forms are out of reach, so the table is purely numerical.
The third stage checks that the table does not undo existing benefits. The boundary penalties must still remove outliers, the stiffness reduction must survive, and eigenfunction errors must stay at their optimal rate.
Baselines
The naive reference point is standard Gauss–Legendre Galerkin discretization at the same degree, mesh and degrees of freedom. On uniform meshes it attains the classical $\mathcal{O}(h^{2p})$ rate reported by the dispersion-optimized quadrature and generalized soft element papers. Any method that cannot beat it on distorted meshes has lost its reason to exist.
The strongest existing method is globally blended quadrature, combined with boundary penalties for outlier removal and, for linear elements, the penalized GSFEMBQ scheme. On uniform tensor meshes these report $\mathcal{O}(h^{2p+2})$ (Dispersion optimized quadratures for isogeometric analysis) and $\mathcal{O}(h^8)$ for $p=1$ (Generalized soft finite element method for elliptic eigenvalue problems). The outlier-removal paper reports condition number reductions of 32%, 60% and 75% for $p=3,4,5$. For distorted, curved or variable-coefficient meshes, no reported result was found on Lacuna for any of these methods; that absence is the gap we fill. We run them with their published weights, unmodified, on identical meshes. Beating them means recovering an observed order at least one higher on smoothly distorted meshes. It also means doing so while keeping outlier-free spectra and the reported condition number reductions within a few percent.
The ablation that isolates our ingredient is an oracle single global weight. For each mesh, we tune one $\alpha$ by minimizing error against a fine reference solution, which is information unavailable in practice. If the oracle global weight matches the local table, spatial variation of the weight is unnecessary and our central claim fails. We also run the local table with $g_e$ replaced by a random permutation across elements. This checks that the gain comes from matching the weight to the local coefficient gradient and not from averaging effects.
Why not just…?
Why not just refine the mesh or raise the polynomial degree with standard quadrature? At fixed degrees of freedom, standard quadrature is stuck at $\mathcal{O}(h^{2p})$. The dispersion-optimized quadrature paper reports that blended quadrature has the same assembly and solve time as a single rule, so any extra orders we retain come for free. We compare at matched degrees of freedom and matched wall-clock time. If $p$-refinement wins on curved domains at equal cost, we will report that plainly.
Why not just apply Richardson extrapolation to two standard solutions? This is the most serious threat. Combining eigenvalues from two nested meshes also cancels the leading error term. It is cheap, and it does not care about geometry, provided the refinement is nested and the computation is in the asymptotic regime. We run it at equal total cost. Our expected advantages are specific and testable. Extrapolation needs a second, finer solve. It needs each eigenvalue matched to the right mode across the two meshes, which becomes fragile higher in the spectrum. Most importantly, it leaves the stiff matrices and outliers untouched, so the explicit time-step limit does not improve. If extrapolation matches our accuracy on low eigenvalues, the honest contribution narrows to the high-spectrum and stability metrics, and we will frame it that way.
Why not just move the mesh nodes, classically or with a learned mover, to make the mesh more uniform? G-Adaptivity reports 21.01% error reduction in 88 ms against 12.69% in 3,780 ms for Monge–Ampère relocation on its Poisson square task. That result concerns forward-problem error, not eigenvalue order. In addition, relocation cannot make a curved domain's pulled-back metric constant. We include Monge–Ampère-smoothed meshes with global blending as a baseline, to measure whether smoothing alone recovers the lost orders.
Proposed Experiments
The known-answer problems come first. These are the 1D, 2D and 3D Laplace benchmarks from the cited papers, now under grading and jitter; the variable-coefficient problem with $\kappa(x)=e^{x+x^2}$ against a fine $N^h=1000$ reference, as in the generalized soft element paper; and the Schrödinger operator with a Pöschl–Teller potential. We add the quarter-annulus and the disk, both exactly representable in isogeometric analysis, whose eigenvalues are known through Bessel function zeros. They are the cleanest curved-geometry test where an analytic answer exists. As a deliberate negative control, we use the L-shaped domain from Verified eigenvalue evaluation for Laplacian over polygonal domain of arbitrary shape. Its corner singularity caps convergence regardless of quadrature, and the method must not appear to beat that cap. We use that paper's verified bounds as ground truth for the first eigenvalues.
To move beyond manufactured shapes, we take airfoil cross-sections from AirfRANS and use each profile as the boundary of an acoustic cavity. We compare against a heavily refined high-order reference. The file offers no measured vibration data for such shapes, so this remains a realistic-geometry test with no physical measurement behind it, and we say so.
The decisive checks are three. First, on smoothly graded meshes and on the annulus, the observed order under global blending should fall strictly between $2p$ and $2p+2$, and local weights should restore it close to $2p+2$. Second, under rough jitter, both methods should drop to $2p$. This confirms the stated limit, and an unexpected gain there would point to an error in our analysis. Third, the oracle global weight should fall measurably short of the local table on the airfoil cavities. If the first or third check fails, the claim that geometry-aware local weights are needed is refuted.
References
- Isogeometric spectral approximation for elliptic differential operators
- Dispersion optimized quadratures for isogeometric analysis
- Generalized soft finite element method for elliptic eigenvalue problems
- G-Adaptivity: optimised graph-based mesh relocation for finite element methods
- Diagnosing Failure Modes of Neural Operators Across Diverse PDE Families
- Outlier Removal for Isogeometric Spectral Approximation with the Optimally-Blended Quadratures
- Verified eigenvalue evaluation for Laplacian over polygonal domain of arbitrary shape
- AirfRANS: High Fidelity Computational Fluid Dynamics Dataset for Approximating Reynolds-Averaged Navier–Stokes Solutions