REVIEW 3 major objections 5 minor 8 references
Stable Computation of Laplacian Eigenfunctions Corresponding to Clustered Eigenvalues
T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read The paper proves that stably computing eigenfunctions for clustered Laplacian eigenvalues reduces to solving a small generalized eigenproblem formed from shape difference quotients, and that this reduction is stable under discretization…
desk verdict Promising method, but the printed main theorem is wrong: the proof yields a different bilinear form for b̃, so the paper needs a correction before the results can be fully trusted. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The object that carries the argument is the shape difference quotient $D_t\lambda_i=(\lambda_i^t-\lambda_i^0)/t$, together with the piecewise-affine pullback map $\Phi_t$ from the perturbed domain back to the unperturbed one. The two bilinear forms $\tilde a_t$ and $\tilde b_t$ in Lemma 3.1 convert the difference of the two eigenvalue problems, one on $K^t$ and one on $K^0$, into a finite-dimensional eigenproblem on the degenerate eigenspace $E$. Solving $M_t\sigma=\mu N_t\sigma$ gives the difference quotients as eigenvalues and the stabilized eigenfunction coefficients as eigenvectors, so the small eigenproblem is the device that turns an unstable eigenvector computation into a stable one.
What would settle it
A decisive test is to run Algorithm 1 on a pair of domains whose Laplacian eigenvalues are closer than $10^{-8}$ but do not come from splitting a repeated eigenvalue; the paper's theory predicts failure of the span assumption or of the difference-quotient gap, which would confirm that condition (3.1) is the active mechanism.
Extended reading notes
Core claim
The central claim is that the ill-posed task of computing eigenfunctions for clustered eigenvalues becomes well-posed when reformulated in terms of eigenvalue difference quotients. On a polygonal domain with a repeated eigenvalue (condition (3.1)), each pulled-back eigenfunction $\tilde u_i^t$ on the unperturbed domain satisfies the variational identity $\tilde a_t(\tilde u_i^t,w)=D_t\lambda_i\,\tilde b_t(\tilde u_i^t,w)$ for every $w$ in the degenerate eigenspace $E$. Consequently the $M$ difference quotients $D_t\lambda_i$ are exactly the eigenvalues of a small $M\times M$ generalized eigenproblem $M_t\sigma=\mu N_t\sigma$, and the components of each eigenvector express the true eigenfunction as a linear combination of the approximate pulled-back modes. The paper demonstrates on a perturbed rectangle and a nearly equilateral triangle that this recovers the symmetry of the exact eigenfunctions even when standard finite element computation fails.
Load-bearing premise
The load-bearing premise is that the clustered eigenvalues come from perturbing a genuinely multiple eigenvalue on the unperturbed domain, and that the pulled-back approximate eigenfunctions still span that entire eigenspace; if either part fails, the small eigenproblem's output no longer represents the true eigenfunctions.
Editorial extensions
If this is right
- For any perturbation family where the unperturbed eigenvalues are multiple and the shape difference quotients remain separated, stabilized eigenfunctions can be computed at accuracy independent of the eigenvalue gap.
- On the two model problems, the method preserves the expected antisymmetry of the second and third eigenfunctions even at perturbation sizes $\epsilon=10^{-10}$ (rectangle) and $\epsilon=10^{-6}$ (triangle).
- The eigenvalue differences $D_t\lambda_i$ obtained from the small eigenproblem are useful shape-sensitivity data by themselves, not just intermediate quantities.
- Because the analysis uses only the variational form of the eigenvalue problem, the same argument applies to general differential operators with clustered eigenvalues caused by domain perturbation.
Reading between the lines
- The authors leave implicit that the same pullback and difference-quotient machinery could support certified a posteriori error bounds for the reconstructed eigenfunctions, since each stabilized function is an explicit linear combination whose residual can be evaluated directly.
- I expect the real smallness parameter of the method is the separation of the first-order directional derivatives of the multiple eigenvalue, not the perturbation size itself; the reported gap of 75.76 in the difference quotients for the triangle is consistent with stability persisting at arbitrarily small perturbation.
- A natural extension test is to apply the approach to clusters that do not arise from a split multiple eigenvalue; the theory predicts degradation, so such a test would separate the mechanism from a generic fix.
- The small eigenproblem can be assembled directly from the elementwise affine maps used in a finite element discretization, so the added cost over an existing eigenvalue solver is minor.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This manuscript proposes a method for computing Dirichlet Laplacian eigenfunctions corresponding to tightly clustered eigenvalues caused by perturbing a domain that has a multiple eigenvalue. The authors introduce a piecewise-affine perturbation map Φ_t between a reference polygon K0 and a perturbed polygon Kt, pull back the perturbed eigenfunctions, and define bilinear forms ã_t and b̃_t so that each eigenvalue difference quotient D_tλ_i and the pulled-back eigenfunction satisfy a finite-dimensional generalized eigenproblem (Theorem 3.1). The resulting Algorithm 1 solves a small M×M matrix eigenproblem to obtain stable linear combinations of approximate eigenfunctions. Numerical experiments on a near-rectangle and a near-equilateral triangle report that the method preserves symmetry properties for perturbations as small as 10^-10 and 10^-6.
Significance. The underlying idea is attractive: replacing the ill-conditioned recovery of eigenfunctions in a tight cluster by a small generalized eigenproblem whose eigenvalues are shape difference quotients is a potentially useful reformulation. The variational derivation is mostly clear, and the algorithm is parameter-free in the sense that no data-fitting is involved. However, as printed, the central theorem contains a definition of b̃_t that is inconsistent with the proof, and the advertised stability is not backed by a quantitative theorem; the numerical evidence is limited to a single mesh size and a single quality metric. With the b̃_t correction and an error analysis, the contribution could be a solid practical method.
major comments (3)
- [Section 3, Lemma 3.1 and Theorem 3.1] The definition of b̃_t is inconsistent with the derivation. After subtracting (3.6) from (3.5), equations (3.7)-(3.8) give ã_t(ũ_i,u) = D_tλ_i Σ_j |det S_{j,t}| (ũ_i,u), not D_tλ_i Σ_j d_{j,t}(ũ_i,u). A uniform-scaling check makes the failure explicit: for S_{j,t}=(1+t)I, |det S_{j,t}|=(1+t)^2 and d_{j,t}=2+t, so for u∈E the printed identity (3.2) would require D_tλ_i(2+t)‖u‖² = -λ(2+t)‖u‖², which is false for general t>0. The theorem and Algorithm Step 3 should use b̃_t(u,v)=Σ_j |det S_{j,t}|(u,v), or an equivalent correction. This is load-bearing because the generalized eigenproblem (3.9) is the basis of the proposed algorithm.
- [Section 1 and Section 5] The central robustness claim is not supported by a quantitative theorem. The introduction states that if the first-order variations of eigenvalues remain separated, the algorithm is robust to discretization and rounding errors even for arbitrarily close eigenvalues, but no error estimate is proved. A natural statement would be a Davis-Kahan-type bound for the small eigenproblem (3.9) that relates the error in the reconstructed eigenfunctions to the gap in the difference quotients, the mesh size h, and the perturbation size. The numerical evidence in Section 5 is limited to one mesh size h=1/64 and the antisymmetry measure A_i; it does not demonstrate convergence as h→0 or compare against exact eigenfunctions. Please add a quantitative stability result and corresponding convergence/error experiments.
- [Section 3, paragraph before Theorem 3.1] The assumption dim(eEt)=M is stated to be 'expected to hold' rather than proved, but Theorem 3.1 and Algorithm 1 depend on this dimension condition. If it fails, the coefficients σ_i obtained from (3.9) need not represent the true pulled-back eigenfunctions ũ_i. The authors should prove this condition under explicit hypotheses (for instance, from invertibility of the pullback and linear independence of {u_t_i}) or state a precise criterion that guarantees it in the finite-dimensional discretized setting.
minor comments (5)
- [Example 1] The displayed formulas λ2=4π²/(1+ε)^2 and λ3=4π² are not the Dirichlet eigenvalues of the rectangle (0,1+ε)×(0,1); the difference quotients in Table 1 correspond to λ2=π²(1+4/(1+ε)^2) and λ3=π²(4+1/(1+ε)^2). Please correct these formulas and the surrounding text.
- [Algorithm 1, Step 2] The text says 'Use FRM to solve the Dirichlet eigenvalue problem' but this should read 'Use FEM'; the same step also refers to 'perturbated DomainK0' where K0 is the unperturbed domain.
- [Throughout] The word 'perturbated' is used repeatedly and should be 'perturbed'.
- [Theorem 3.1] The claim that the eigenvalues 'appear in ascending order along the generalized spectrum' should be stated more carefully; with the corrected positive definite b̃_t this ordering is natural, but the text should explicitly say that the D_tλ_i are ordered increasingly with i.
- [Section 6, Remark 6.1] The statement that the gap μ3-μ2=75.76 is stable across perturbation directions would be more convincing if accompanied by a table of the computed μ2 and μ3 values for cases (A)-(D) rather than a single number in a remark.
Circularity Check
No significant circularity: the difference quotients and stabilized eigenfunctions are outputs of the derived small generalized eigenproblem, not inputs fitted to reproduce the target; the proof is self-contained and the numerical validation uses an external antisymmetry benchmark.
full rationale
The derivation chain is self-contained. Lemma 3.1 starts from the pulled-back weak form of the perturbed eigenvalue problem (3.3) and the unperturbed eigenvalue equation (3.4), subtracts one from the other, and obtains the variational identity (3.2) for each exact pulled-back eigenfunction. The small generalized eigenproblem (3.9) in Theorem 3.1 is then a purely algebraic reformulation: if the M pulled-back eigenfunctions are expanded in the chosen basis, their coefficient vectors satisfy the pencil equation, and the eigenvalues of that pencil are the difference quotients D_t lambda_i. No parameter is fitted to the target eigenfunctions; the difference quotients and the coefficient vectors are outputs of the computation. The numerical antisymmetry measures A_i are computed from the known reflection symmetry of the true eigenfunctions and therefore serve as an independent external check rather than as an input to the algorithm. The only self-citation, Remark 3.1 pointing to the authors' earlier paper [6] for the concept of the shape difference quotient, is not load-bearing because the present paper gives a full derivation of the required identity rather than importing it. The unproved dimension assumption dim(eEt)=M is explicitly stated as a hypothesis of Theorem 3.1 and acknowledged as 'expected to hold'; it is a limitation on applicability, not a circular step. A separate mathematical inconsistency may exist in the printed definition of btilde_t relative to the proof, but that is a correctness concern, not circularity, and does not affect the circularity score.
Assumptions & free parameters
assumptions (4)
- standard math The Laplacian on a bounded polygonal domain has a discrete spectrum with an orthonormal eigenbasis (spectral theorem for compact self-adjoint operators).
- domain assumption On the unperturbed domain K0, the eigenvalues lambda_n^p through lambda_N^p coincide (multiple eigenvalue).
- domain assumption The pulled-back perturbed eigenfunctions span a space of dimension M = N-n+1 (dim eEt = M).
- domain assumption The difference quotients D_t lambda_i are sufficiently separated so that the small generalized eigenproblem is well posed.
Cite this review
Pith. "Pith review of Stable Computation of Laplacian Eigenfunctions Corresponding to Clustered Eigenvalues." pith.science (2026). https://pith.science/paper/GNUWJXHO
@misc{pith2026250607340,
author = {Pith},
title = {Pith review of: Stable Computation of Laplacian Eigenfunctions Corresponding to Clustered Eigenvalues},
year = {2026},
howpublished = {\url{https://pith.science/paper/GNUWJXHO}},
note = {Machine review of arXiv:2506.07340}
}
read the original abstract
The accurate computation of eigenfunctions corresponding to tightly clustered Laplacian eigenvalues remains an extremely difficult problem. In this paper, using the shape difference quotient of eigenvalues, we propose a stable computation method for the eigenfunctions of clustered eigenvalues caused by domain perturbation.
Figures
Figures from the paper (2 more)
Reference graph
Works this paper leans on
- [1]
-
[2]
B. Parlett, The Symmetric Eigenvalue Problem, Classics in Applied Mathematics, Society for Industrial and Applied Mathematics, 1998
work page 1998
-
[3]
X. Liu, T. Vejchodsk´ y, Fully computable a posteriori error bounds for eigenfunctions, Numer. Math. (2022)
work page 2022
-
[4]
O. Takeshi, K. Aishima, Iterative refinement for symmetric eigenvalue decomposition ii: clustered eigenvalues., Japan Journal of Industrial and Applied Mathematics (2019) 435–459
work page 2019
-
[5]
E. J. Haug, B. Rousselet, Design sensitivity analysis in structural mechanics. ii. eigenvalue variations, Journal of Structural Mechanics 8 (2) (1980) 161–186
work page 1980
-
[6]
R. Endo, X. Liu, Rigorous estimation for the difference quotients of multiple eigenvalues (2025). arXiv:2305.14063
work page Pith review arXiv 2025
-
[7]
B. J. McCartin, Eigenstructure of the equilateral triangle, part i: The dirichlet problem, Siam Review 45 (2) (2003) 267–287
work page 2003
-
[8]
X. Liu, Guaranteed Computational Methods for Self-Adjoint Differential Eigenvalue Problems, SpringerBriefs in Mathematics, Springer Nature Singapore, 2024. 9
work page 2024
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.