acceptodds
Under review as a conference paper at ICLR 2027

BernNet Revisited: Adaptive Spectral Warping and Decomposed Regularisation for Heterophilic Graphs

Abstract

Spectral graph neural networks on heterophilic graphs face two structural limitations: fixed polynomial bandwidth, which allocates capacity to uninformative frequency regions, and homogeneous regularization, which cannot simultaneously capture smooth trends and sharp spectral transitions. This issue is further compounded by self-information dilution under high-order neighborhood aggregation. We present BernNetII, a framework that resolves each limitation through principled modifications to the Bernstein spectral filter: a learnable spectral warp that adaptively concentrates polynomial resolution on the graph's effective frequency band; and a morphological decomposition of filter coefficients into smooth and step components with individually tailored regularisation, augmented by a residual bypass that preserves node-level features under high-order propagation. We establish a Rademacher complexity bound with separate capacity control for each component and prove convergence of the joint proximal gradient optimisation. Experiments on eight node classification benchmarks demonstrate that BernNetII consistently outperforms or matches state-of-the-art methods, validating its effectiveness across diverse settings. Our code is available at https://github.com/AI-FZY/BernNetII.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Spectral Graph Neural Networks, Heterophilic Graphs, Bernstein Polynomial Filters, Graph Signal Processing, Generalization Bounds

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_LgumOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces BernNetII, which aims to improve BernNet for heterophilic graph learning. Three modifications are applied: coefficient decomposition, learnable spectral warp, and residual bypass. The filter coefficients are decomposed into smooth and step components with separate regularization. The spectral warp rescales the evaluation domain of the Bernstein basis, while the residual bypass retains self representation alongside the propagated representations. The authors provide theoretical analyses of the properties of the decomposition, spectral warp, convergence and generalization bound. They then evaluate BernNetII on eight node-classification benchmarks under fully supervised and semi-supervised settings.

Critical strengths

The proposed modifications are simple and effective, although I have concerns about their justification and empirical contribution (see Critical Weaknesses).

Critical weaknesses

Here are my concerns for each of the proposed modifications.

  • The motivation for the sharp step component is insufficiently supported. The learned spectral responses visualized in the original BernNet (Figure 4 of [1]) and ChebNetII (Figure 6 of [2]) exhibit, even for heterophilic datasets, smooth curves rather than step-like patterns. The authors should provide evidence supporting the need for step-like spectral responses in real-world classification tasks, for example by citing relevant prior work or providing convincing explanations or analysis. It would be also great if the authors show whether BernNetII actually learns such responses, e.g., providing visualizations of the learned spectral responses of BernNetII.

  • The theoretical result does not establish the claimed benefits of spectral warping. Section 3.2 states that Proposition 4 shows that the warp reduces approximation error by focusing polynomial degrees on the effective frequency band. However, the proposition only establishes that the filter’s total variation remains bounded by the coefficients’ discrete total variation, which does not demonstrate improved approximation accuracy or more efficient allocation of filter domains. The claimed benefit requires further theoretical analysis or empirical evidence.

  • The improvements beyond the residual modification of BernNet are unclear. The residual bypass is straightforward to add to the original BernNet, making BernNet with the same learnable bypass an essential baseline. Without the bypass, the most substantial improvement over BernNet is concentrated only on Chameleon, while the gains on Squirrel and Texas are small, and performance decreases on Cornell, Cora, and Pubmed. Although the component-removal ablations suggest that warp and decomposition contribute within the full model, they do not establish their value over BernNet with a bypass. A matched comparison is needed to determine whether the two main proposed modifications provide benefits beyond the baseline.

[1] BernNet: Learning Arbitrary Graph Spectral Filters via Bernstein Approximation. NeurIPS 2021.

[2] Convolutional Neural Networks on Graphs with Chebyshev Approximation, Revisited. NeurIPS 2022.

Questions for authors

Could the authors address the concerns raised in Critical Weaknesses?

5 comments

W1. The label spectra of real datasets contain sharp transitions, and the coefficients learned by BernNetII contain the corresponding jumps. As a model-independent check, we projected the node labels onto the Laplacian eigenvectors and compared the result with randomly permuted labels (Appendix G). On Cora the label energy drops by more than an order of magnitude within λ≤0.2\lambda\leq0.2, and on Chameleon and Squirrel the label spectrum has a narrow notch near λ≈1\lambda\approx1. This agrees with the stochastic block model, where the classes are recovered from a few eigenvectors, so the ideal response is one on a few bands and zero elsewhere (Rohe et al., 2011). The curves in Fig. 4 of [1] and Fig. 6 of [2] look smooth because the slope of a polynomial of degree 10 is bounded by Markov's inequality (DeVore and Lorentz, 1993), whether or not the target has a step. In the Bernstein basis a sharp transition appears as a jump in the coefficients, and a jump in β\boldsymbol\beta gives a single transition without overshoot whose oscillation is at most ∥β∥DTV\lVert\boldsymbol\beta\rVert_{DTV} (Theorem 1). BernNetII learns a steep low-pass filter on Cora and a band-stop filter on Chameleon and Squirrel whose stopband lies on the notch (Figure 2), with the forward-pass coefficients θ4+=θ5+=θ6+=0\theta_4^{+}=\theta_5^{+}=\theta_6^{+}=0 in all ten runs on Chameleon and in seven of ten on Squirrel. In Table 6, removing the decomposition lowers accuracy on seven of the eight datasets, by 3.09 points on Squirrel.

W2. Proposition 4 guarantees that the regulariser remains effective under the warp, and the benefit of the warp is supported by the ablation in Table 6. Theorem 1 bounds the oscillation of the filter by the DTV of its coefficients for the original Bernstein basis. The warp replaces the argument λ/2\lambda/2 by sλ/2s\lambda/2 and therefore changes the basis functions, and Proposition 4 proves that the same bound holds for every ss. The warped basis spans the same polynomials for every ss, so the benefit of the warp is established empirically. In Table 6, removing the warp lowers accuracy on seven of the eight datasets, by 3.01 points on Squirrel. The sentence after Proposition 4 has been revised accordingly.

W3. Appendix F now includes BernNet with the same bypass, initialisation, seeds and splits as BernNetII. BernNetII exceeds this baseline on seven of the eight datasets, by 3.82 points on Squirrel and by 1.3 to 1.9 points on Chameleon, Texas, Cornell and Actor. The same table removes the warp and the decomposition separately, and each removal lowers accuracy on seven of the eight datasets. The warp and the decomposition therefore add value beyond the bypass.

References

DeVore, R. A., & Lorentz, G. G. (1993). Constructive approximation (Vol. 303). Springer. Rohe, K., Chatterjee, S., & Yu, B. (2011). Spectral clustering and the high-dimensional stochastic blockmodel. The Annals of Statistics, 39(4), 1878-1915.

I generally accept the authors' explanations for spectral warping and the residual bypass. However, in my view, both are closer to practical tricks than substantial contributions. Spectral warping is mainly supported by empirical results, while its theoretical motivation remains largely intuitive. As for the residual bypass, I see it as a fairly standard technique rather than a novel contribution. Therefore, I do not consider either of them a main contribution of this paper.

In my view, the most interesting and potentially significant contribution is the proposed decomposition of spectral filter coefficients into smooth and step components. After reading the authors' additional explanations, I find this idea plausible and potentially useful. However, I am still not fully convinced about how broadly applicable the assumption of step-like spectral structures is in real-world tasks, or whether the proposed decomposition actually learns the intended step-like components.

For instance, the model mainly tries to achieve this separation through different regularization terms. However, the two components are still jointly learned and further coupled through the curvature-alignment penalty, so it is unclear whether they can truly separate global smooth trends from localized sharp variations. For example, on Squirrel (Figure 2), the label spectrum is relatively smooth except for a narrow notch near λ≈1\lambda\approx1, whereas the learned filter exhibits a much broader suppression region and substantial variations across other frequencies. Although the label spectrum and learned filter need not have the same shape, this result raises questions about whether the proposed decomposition works as intended.

I will keep my score unchanged, but I would encourage the authors to reconsider the priorities and emphasis of the paper. In particular, I suggest placing less emphasis on spectral warping and the residual bypass, and investigating the proposed decomposition more thoroughly, as I believe this is a genuinely interesting research question. For example, the following questions may be worth exploring:

  • Why do sharp spectral variations occur at different frequencies across datasets? Are there any underlying patterns, and how are they related to graph homophily and heterophily?
  • How can we design a better decomposition that effectively separates global smooth trends from localized sharp variations?

P.S. I think the paper may confuse two different things: heterophily is often associated with stronger high-frequency graph signals, but this does not necessarily mean that the spectral filter itself needs sharp variations. For example, a smooth high-pass filter can also emphasize high-frequency signals.

Thank you for the careful reading and for identifying the decomposition as the most interesting part of the work. We agree that it is the central idea. Proposition 3, Theorem 5 and Proposition 6 all concern it, and the final version will state this priority in the introduction.

Separation under joint learning. The two components do separate, and the split is fixed by the regularisers, as in morphological component analysis (Starck et al., 2005). Because the forward pass depends on α\boldsymbol{\alpha} and β\boldsymbol{\beta} only through θ=α+β\boldsymbol{\theta}=\boldsymbol{\alpha}+\boldsymbol{\beta}, the loss cannot distinguish two splits of the same θ\boldsymbol{\theta}. For ν>0\nu > 0 and given weights of RcaR_{\mathrm{ca}}, exactly one split of each θ\boldsymbol{\theta} minimises the regularisers (Section 3.1), so any stationary point reached by training has this split. The weights of RcaR_{\mathrm{ca}} are computed with stop-gradient, so RcaR_{\mathrm{ca}} only weights the penalty on β\boldsymbol{\beta} and does not mix the components. This split places each kind of structure in the intended component. By Proposition 3, a sufficiently large jump in θ\boldsymbol{\theta} cannot stay entirely in α\boldsymbol{\alpha}, and by the optimality condition of the fused lasso, β=0\boldsymbol{\beta}=\boldsymbol{0} whenever every second difference of θ\boldsymbol{\theta} is at most γ/(4μ)\gamma/(4\mu) in absolute value. Accordingly, the learned β\boldsymbol{\beta} on Cora and Citeseer is a single jump between k=0k=0 and k=1k=1 (Appendix G).

Width of the stopband on Squirrel. The wide stopband on Squirrel is a limit of the polynomial degree, not a failure of the decomposition. The notch is one bin, 0.050.05 in λ\lambda, wide, whereas a single jump in the Bernstein coefficients already gives a transition about 0.70.7 wide for K=10K=10 and s=1s=1, from 10% to 90% of its height (Appendix G). A filter of this degree therefore cannot follow the notch, and the stopband formed by the zero coefficients θ4+\theta_4^{+}, θ5+\theta_5^{+} and θ6+\theta_6^{+} has a width set by this resolution. Within this resolution, the response follows the label spectrum. The label energy is near or below the random-label band only in the first bin and around λ=1\lambda=1, and these are exactly the places where the response is lowest.

Smooth or sharp filters (P.S.). The sharp variations in BernNetII are steps at the edges of narrow bands that carry the class information, not the high-frequency emphasis of a smooth high-pass filter, and these steps are needed. The key point is the width of the informative band, not its frequency. The stochastic block model analysis in Section 2.4 shows that the response that separates the classes is one on a few bands and zero elsewhere, and Appendix G finds such narrow bands in five of the eight label spectra. Because a narrow band must be passed while its neighbours are suppressed, the filter has to change within a short interval of λ\lambda at the edges of the band, which a smooth filter cannot do, whether the band lies at low or high frequency. This is why steps also appear on the homophilic Cora and Citeseer, and why the band-stop target of Chameleon and Squirrel, with class information on both sides of λ=1\lambda=1, cannot be met by a high-pass filter. When the informative band is wide, a smooth filter suffices, and BernNetII represents it without the step component. The high-pass filter sλ/2s\lambda/2, for example, has the coefficients θk=k/K\theta_k=k/K with zero second differences and is represented with β=0\boldsymbol{\beta}=\boldsymbol{0} at no regularisation cost.

We hope this resolves your remaining questions, and we would be grateful if you would take it into account in your final assessment.

References

Starck, J. L., Elad, M., & Donoho, D. L. (2005). Image decomposition via the combination of sparse representations and a variational approach. IEEE transactions on image processing, 14(10), 1570-1582.

Reviewer_QefrOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes BernNetII, a spectral graph neural network designed for node classification on heterophilic graphs. Building on BernNet, it introduces three modifications: a learnable spectral warp to adjust the effective spectral resolution, a decomposition of filter coefficients into smooth and step components with separate regularization, and a residual bypass to preserve node features during high-order propagation. The authors provide theoretical analyses of filter regularization, optimization convergence, and generalization, and evaluate the method on eight node-classification benchmarks. The results show improvements over existing methods on several heterophilic datasets, particularly in semi-supervised settings.

Critical strengths
  1. The paper identifies several meaningful limitations of existing Bernstein-based spectral GNNs. The authors consider not only the difficulty of capturing both smooth spectral trends and sharp transitions, but also the fixed distribution of spectral resolution and the dilution of node features under high-order propagation. Addressing these issues jointly provides a reasonable motivation for the proposed framework.
  2. The coefficient decomposition is an interesting and well-motivated design. By separating the filter coefficients into smooth and step components, the method applies curvature regularization and fused-lasso regularization to different components. This offers a principled way to accommodate spectral responses with different structural characteristics, particularly when discriminative signals involve sharp transitions.
  3. The paper provides theoretical analyses to support several aspects of the proposed method. These include a total-variation bound for Bernstein spectral filters, a condition under which coefficient decomposition is preferable to a trivial single-component representation, and a Rademacher complexity bound. Such analyses help clarify the properties and potential benefits of the proposed parameterization.
  4. The experimental results demonstrate promising performance, particularly on heterophilic graphs. BernNetII achieves competitive results in both fully supervised and semi-supervised node classification, with substantial improvements over BernNet on some datasets.
Critical weaknesses
  1. The motivation and effectiveness of the spectral warp require further clarification. The proposed transformation replaces the Bernstein basis argument \lambda/2 with s\lambda/2, where s\in(0,1) is a learnable scalar. This changes how many basis functions have their peaks within the realized spectral domain, but it does not explicitly learn an arbitrary target frequency interval.
  2. The theoretical justification does not fully cover the complete proposed framework. Proposition 3 analyzes a simplified objective containing the curvature and total-variation penalties, without the curvature-alignment term introduced in the full model. Moreover, Theorem 5 establishes convergence under a fixed-step joint proximal gradient scheme, whereas the practical implementation uses Adam to update several smooth variables. The theoretical convergence guarantee does not directly apply to the optimization procedure used in the experiments.
  3. The experimental comparison could be more comprehensive. Although the paper includes several established spectral GNNs, the evaluation would benefit from comparisons with more recent polynomial spectral filtering methods, such as JacobiCon, OptBasisGNN, UniFilter, and Specformer.
Questions for authors
  1. How does the proposed scalar transformation s\lambda/2 adapt to an informative frequency band whose location and width vary across graphs? Can the authors provide a theoretical explanation or empirical visualization showing that the learned warp allocates spectral resolution more effectively than the original Bernstein basis and existing adaptive spectral filters?
  2. Can the authors establish stronger theoretical results for the complete objective, including the curvature-alignment term? Given that Adam is used for several variables in the actual implementation, what convergence guarantees hold for the optimization procedure used in the experiments?
  3. Can the authors compare BernNetII against JacobiCon, OptBasisGNN, UniFilter, and Specformer under consistent experimental settings? Such comparisons would help establish whether the proposed method offers substantive advantages over existing polynomial spectral filtering approaches.
2 comments

W1 / Q1. The warp rescales the frequency axis. With the same coefficients, the warped filter gs(λ)g_s(\lambda) takes the value of the unwarped filter at sλs\lambda, so the peak of the kk-th basis function lies at λ=2k/(sK)\lambda=2k/(sK) instead of 2k/K2k/K. A jump of β\boldsymbol\beta produces a band edge. At s=1s=1 this edge can take only KK positions, and under the warp it moves continuously from its position λ′\lambda' to λ′/s\lambda'/s. The position and the width of an individual band are determined by the indices at which β\boldsymbol\beta jumps. JacobiConv and OptBasisGNN also adapt the basis, and different polynomial bases have the same expressive power (Wang and Zhang, 2022), so they and the warp represent the same filters. The warp differs in that Proposition 4 keeps the oscillation of the filter bounded by ∥θ∥DTV\lVert\boldsymbol\theta\rVert_{DTV} for every ss. A Jacobi basis or a learned orthogonal basis has no such bound, because with equal coefficients the DTV is zero while the filter, a sum of basis polynomials of degree up to KK, is not constant. In Table 6, removing the warp lowers accuracy on seven of the eight datasets, by 3.01 points on Squirrel. Section 3.2 now describes the warp in these terms.

W2 / Q2. The argument of Proposition 3 extends to the full objective (11). The curvature-alignment term is zero at the trivial split (θ,0)(\boldsymbol\theta,\mathbf 0), so it only adds Rca(θs,θstep)R_{\mathrm{ca}}(\boldsymbol\theta_s,\boldsymbol\theta_{\mathrm{step}}) to the cost of the split (θs,θstep)(\boldsymbol\theta_s,\boldsymbol\theta_{\mathrm{step}}), and the conclusion holds with γ∥θstep∥DTV\gamma\lVert\boldsymbol\theta_{\mathrm{step}}\rVert_{DTV} replaced by γ∥θstep∥DTV+Rca(θs,θstep)\gamma\lVert\boldsymbol\theta_{\mathrm{step}}\rVert_{DTV}+R_{\mathrm{ca}}(\boldsymbol\theta_s,\boldsymbol\theta_{\mathrm{step}}) in the threshold (9). The term also makes the decomposition unique. For a given θ\boldsymbol\theta and given weights, the sum of the three regularisers is strongly convex in β\boldsymbol\beta when ν>0\nu > 0, so exactly one split minimises it. Both results are stated after Proposition 3 and proved in Appendix B, and in Table 6 removing RcaR_{\mathrm{ca}} lowers accuracy on six of the eight datasets. Theorem 5 is stated for the fixed-step scheme, as Section 3.3 says. It shows that the non-smooth fused-lasso term is compatible with joint training, since its proximal step (14) is exact and costs O(K)\mathcal O(K). The experiments update the smooth variables with Adam, the usual choice for polynomial spectral GNNs including BernNet and ChebNetII. Adam does not ensure the decrease of the objective on which the proof relies, so the Adam variant lies outside the scope of the theorem and is supported empirically. Section 3.3 now states this scope.

W3 / Q3. Tables 2 and 3 now include JacobiConv, OptBasisGNN, Specformer and UniFilter, evaluated with their official implementations and published configurations on the same graphs, splits and seeds. Under full supervision BernNetII is best on Texas, Cornell, Actor and Citeseer, the new baselines are better on Chameleon and Squirrel, and UniFilter is best on Pubmed. Under sparse supervision BernNetII is best on five of the eight datasets and exceeds all four new baselines on Chameleon and Squirrel. BernNetII needs no eigendecomposition, which Specformer requires, and its coefficient regularisation is complementary to methods that learn the basis.

Reference Wang, X., & Zhang, M. (2022). How powerful are spectral graph neural networks. In International Conference on Machine Learning (pp. 23341-23362). PMLR.

Reviewer_QefrOpenReview ↗

Repy

Thank the authors for the reply. The score is maintained.

Reviewer_wJNPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework called BernNetII, which addresses two limitations of spectral graph neural networks through targeted improvements to the Bernstein spectral filter. Experimental results demonstrate the effectiveness of the proposed method.

Critical strengths

1.The proposed method is interesting, as it improves model performance by modifying the Bernstein spectral filter.

2.The paper provides sufficient theoretical justifications and proofs.

Critical weaknesses

1.Figure 1 only illustrates the Bernstein filter, without showing whether similar phenomena also occur with other spectral filters.

2.The authors only discuss why Chebyshev polynomials are unsuitable for coefficient decomposition, but do not consider other filter bases, such as those used in GPR-GNN and JacobiConv.

3.The authors do not explain the role and significance of Theorem 1.

4.The baselines are relatively outdated, published before 2022, and the evaluation lacks comparisons with recent methods.

5.The experimental results show that the proposed method does not achieve the best performance on all datasets, and the performance gains are relatively modest.

6.The experimental section occupies a relatively small portion of the paper.

Questions for authors

1.Do other spectral filters exhibit the phenomenon illustrated in Figure 1?

2.Besides Chebyshev polynomials, could the authors consider other filter bases?

3.What is the role and significance of Theorem 1?

4.How does the proposed method compare with state-of-the-art baseline methods?

5.Since the proposed method does not achieve the best performance on all datasets, does this indicate that its applicability is limited?

1 comment

W1 / Q1. The Gibbs phenomenon in Figure 1 does not depend on the Bernstein basis, and other spectral filters exhibit it as well. It occurs whenever a function with a jump is approximated by a finite expansion in smooth functions, and it is documented for Fourier series and for Chebyshev, Legendre and other orthogonal polynomial expansions (Gottlieb and Shu, 1997). ChebNet and ChebNetII use the Chebyshev basis, GPR-GNN the monomial basis and JacobiConv the Jacobi basis, and in every case the filter is a polynomial of degree KK in λ\lambda, like the Bernstein filter in Figure 1. Changing the basis therefore does not remove the oscillation, and BernNetII instead changes how the coefficients are parameterised and regularised. We added this remark to the discussion of Figure 1.

W2 / Q2. BernNetII separates the smooth part and the step part in the coefficients of the filter, without eigendecomposition. This requires that the shape of the coefficient sequence carries over to the filter, which holds in the Bernstein basis. Theorem 1 bounds the oscillation of the filter by the DTV of its coefficients, and a monotone coefficient sequence gives a monotone filter. The monomial basis of GPR-GNN, the Jacobi basis of JacobiConv and the Chebyshev basis lack this property. With all coefficients equal, the DTV is zero and the Bernstein filter is constant, because the basis functions sum to one. In the other bases the filter is then the sum of the basis polynomials, for example 1+(1−λ)+⋯+(1−λ)K1+(1-\lambda)+\dots+(1-\lambda)^K in GPR-GNN, which has degree KK under any rescaling and is not constant. A fused-lasso penalty on those coefficients would therefore not control the oscillation of the filter. Section 2.2 now makes this argument for GPR-GNN and JacobiConv.

W3 / Q3. Theorem 1 states that the total variation of the filter is at most the DTV of its coefficients, ∑k∣θk+1−θk∣\sum_k|\theta_{k+1}-\theta_k|. It connects the coefficients, which the model learns and regularises, to the oscillation of the filter. By Theorem 1 the fused-lasso penalty γ∥β∥DTV\gamma\lVert\boldsymbol\beta\rVert_{DTV} limits the oscillation that β\boldsymbol\beta produces, which is why Section 2.3 calls the theorem "the mathematical foundation of our regularisation design". Proposition 4 extends the bound to the warped filter, and Proposition 6 uses the same quantity to measure the capacity of the step component. The bound is tight, with equality for monotone coefficients. A paragraph after Theorem 1 now states these points.

W4 / Q4. Tables 2 and 3 now include four recent spectral GNNs, JacobiConv (ICML 2022), OptBasisGNN (ICML 2023), Specformer (ICLR 2023) and UniFilter (ICML 2024), evaluated with their official implementations and published configurations on the same graphs, splits and seeds. Under full supervision BernNetII is best on Texas, Cornell, Actor and Citeseer, and the new baselines are better on Chameleon and Squirrel. Under sparse supervision BernNetII is best on five of the eight datasets.

W5 / Q5. No method is best on all datasets, including the four new baselines, which reflects different assumptions about the shape of the spectral filter. BernNetII is best on four of the eight datasets under full supervision and on five under sparse supervision, where it exceeds ChebNetII by 8.04 and 7.12 points on Texas and Cornell. It needs no eigendecomposition, unlike Specformer, and keeps the propagation cost of BernNet.

W6. The experimental part has been extended. Tables 2 and 3 include four more baselines, Appendix F reports the ablation of every component together with BernNet plus the bypass, and Appendix G compares the learned responses with the label spectrum of each dataset.

Reference

Gottlieb, D., & Shu, C. W. (1997). On the Gibbs phenomenon and its resolution. SIAM Review, 39(4), 644-668.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank the reviewers for their comments. The revision adds four recent spectral GNNs to Tables 2 and 3, a component ablation that includes BernNet with the same bypass (Appendix F), and a comparison of the learned responses with the label spectrum of each dataset (Appendix G). On the theory, the response that separates the classes of a heterophilic graph has sharp transitions, and a low-degree polynomial fitted to it shows the Gibbs phenomenon in every basis. BernNetII represents the jumps by a separate coefficient vector in the Bernstein basis, where Theorem 1 bounds the oscillation of the filter by the DTV of its coefficients. This bound is tight and does not hold in the monomial, Chebyshev and Jacobi bases. Proposition 4 guarantees that the regulariser remains effective under the warp, whose benefit is shown by the ablation. The argument of Proposition 3 extends to the full objective, and the curvature-alignment term makes the decomposition unique. Theorem 5 covers the fixed-step scheme. The numbering of theorems, propositions and equations is unchanged.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.