- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes two differentiable neural surrogates for the first ten Dirichlet eigenvalues of planar domains, aimed at shape optimization. The first treats star-shaped domains: the boundary is a truncated Fourier series, and three small MLPs predict area-normalized eigenvalues, with rotation/reflection averaging at test time The second uses the torsion/landscape function as the input: a three-channel image of a smoothed indicator, the normalized landscape, and its squared gradient is passed to a U-Net whose decoder emits ten functions that are orthonormalized by a differentiable Gram--Schmidt step, with eigenvalue ratios read off Rayleigh quotients. The landscape model is the headline result, reaching about mean relative error over ten eigenvalues and outperforming author-reimplemented FNO () and DeepONet (). The authors motivate the landscape representation with a stability inequality relating landscape functions to eigenvalue perturbations.
Critical strengths
- The landscape-as-learned-input representation is new, and the ablation is clean. Feeding the torsion/landscape function to a learned geometry-to-spectrum map is a new contribution. The same-backbone comparison shows that the input representation, not the architecture, is doing most of the work. This is the contribution the paper should be built around.
- The surrogate is amortized and differentiable, so it is actually usable for optimization. It predicts ten eigenvalues in one pass and exposes gradients for arbitrary functions of them. The case studies recover the disk and two-disk optima to within and , which is a credible check that the tool works.
Critical weaknesses
-
The abstract does not match the paper's own tables. It says FNO is at , but Table 5 says . It says the Fourier model reaches , but the retained model reaches at . It says replacing the landscape with an SDF worsened optimization, but no SDF optimization experiment was run. It calls the MPEG-7 error , which is the median; the mean is . The corrected numbers are still decent, which is exactly why this is damaging: it looks like the headline text was not checked against the results.
-
The main theoretical justification is not actually established. The stability inequality is presented as "mathematically proved," but there is no citation, no proof, and no value or dependence for . The cited David et al. paper is about the landscape on for random Schrodinger operators and does not contain this statement, and Arnold et al. (2019) describe the landscape-to-spectrum method as only partially justified. Since this inequality is the stated reason for using the landscape representation, the motivation is missing.
-
Several technical claims do not match the implementation. Eq. (3) is said to ensure positivity of the border function, but summing the printed per-mode bound gives about , which is greater than , so it does not imply ; the released code uses a total-amplitude budget instead. The contributions say the Fourier model uses translation invariance, while Section 3.1 says translations are explicitly not imposed. The "single forward pass for arbitrary raster domains" claim only holds for the landscape model; the Fourier model takes 30 coefficients and averages over 12 rotations and reflections for three networks.
-
The novelty is narrower than the title suggests, and the reference list is thin and padded. Neural, differentiable shape optimization of spectral and geometric functionals has already been done by Martinet and Ftouhi (2026, arXiv:2602.14881), and An and del Castillo (2025, arXiv:2507.07073) already learn Laplace-Beltrami spectra from geometry. NEO (arXiv:2605.24390) already does amortized eigenspectrum prediction, and Rowan et al. (2025) already combine Rayleigh quotients with Gram-Schmidt. The paper's 17 references are thin for the claims and include tangential items that do not support the argument. The defensible contribution is the amortized Dirichlet-spectrum surrogate with a landscape input, and the paper should be framed around that.
Questions for authors
-
The stability inequality is the stated reason for using the landscape representation, but it has no citation, no proof, and no value or dependence for . What is the precise theorem and reference, and what does depend on? The cited David et al. paper appears to be about the landscape on for random Schrodinger operators, not this Dirichlet eigenvalue statement.
-
Eq. (3) is said to ensure , but summing the printed per-mode bound gives about , which is larger than . Was the implemented constraint the per-mode bound or a total-amplitude budget, and what exactly was used? Also, why do the contributions claim translation invariance when Section 3.1 says translations are not imposed?
-
How do you distinguish the contribution from Martinet and Ftouhi (2026, arXiv:2602.14881), An and del Castillo (2025, arXiv:2507.07073), NEO (arXiv:2605.24390), and Rowan et al. (2025), who already combine Rayleigh quotients with Gram-Schmidt?
-
Tables 2 to 6 report no seed spread, and the optimization uses the best of ten starts without a success rate. Given that some headline gaps are small, for example landscape versus DeepONet , how do you establish that these differences are stable rather than run-to-run variation? Can you report per-seed errors and optimization success rates?
3 comments
I thank the authors for their rebuttal, which addressed some of my concerns.
Even though I believe that the idea is worth pursuing and the results are interesting, I will maintain my rejection recommendation, as I believe that the work could really benefit from a fundamental rewriting.
I would also advise the authors to submit the rebuttal earlier than 15 hours before the end of the discussion window.
We thank you for the follow-up comment and for acknowledging that the core idea is worth pursuing and the results are interesting.
We apologize for posting the revised manuscript late in the discussion period. Training the PointNet on our dataset alongside the new optimization experiments took significant compute time. Should the paper be accepted, we commit to revising the presentation and framing of our contribution for the camera-ready version.
We thank Reviewer Y7Yz for the constructive comments and questions. Before answering the questions, we want to clarify the problem we address in the paper is variational spectral shape optimization. In more detail, "the methods consist in deforming the domain Ω to minimize an objective functional F(λ1(Ω),…,λk(Ω)) under area constraint Antunes & Freitas (2012). This variational problem is different from spectral shape analysis in geometry processing Reuter et al. (2006); Yigit & Wilson (2025); Velich et al. (2026) or Kac’s inverse problem Kac (1966); Gordon et al. (1992), which either compute Laplace-Beltrami descriptors on surfaces for shape matching or reconstruct a shape from a target spectrum.” (please refer to the updated introduction).
Questions:
We added the specific mathematical statement with the explicit constant and the precise references. The inequality, as now stated, is a direct consequence of Theorem 3.4, Bucur (2003) and Theorem 3.1, Cheng & Yang (2007).
We removed the claim about translation invariance. We also corrected the paragraph describing the database.
We hope we clarify our contribution in the contributions section with
“ Additionally, Martinet & Ftouhi (2026) used invertible neural networks as surrogates of the shapes in their exploration of Blaschke-Santaló diagrams rather than the functionals themselves, their work focuses on convex bodies parametrized by a transformation of the disk making use of analytic formulae for the shape functionals and their derivatives. ”
Moreover, following the point cloud representation in geometry processing (Marin et al., 2020; Pang et al., 2024; An & del Castillo, 2025); Yang et al. (2026), we trained a PointNet on the same dataset. Please refer the last paragraph before the conclusion. Briefly, PointNet performs well on simply connected and convex geometries (triangle 5.12%, polygon 5.65%) but fail for more topologically complex shape (several connected components 18.18%, non convex 34.93% compared to 0.97% and 2.55%).
Finally, Rowan et al. (2025) optimize a neural network through the Rayleigh quotient for a single fixed domain, requiring a full training whenever the geometry changes. Our methods predict the first eigenvalues and eigenfunctions in a single forward pass and are differentiable with respect to domain deformations.
We thank the Reviewer for raising this point. To clarify, the dispersion in the updated Table 5 corresponds to intra test set standard deviation across the test set. Retraining completely every model on N=100000 domains was not feasible within the rebuttal window. Our model achieves not only a lower mean relative error (0.92% vs. 1.48% for DeepONet and 2.42% for FNO) but also has a low inter shape standard deviation (0.86% vs. 1.61% for DeepONet and 1.77% for FNO).