acceptodds
Under review as a conference paper at ICLR 2027

Oscillatory Neural Dynamics over Sheaves

Abstract

Effective long-range propagation remains a central challenge in graph neural networks, as increasing model's propagation depth does not guarantee that distant information remains influential. Sheaf neural networks enrich graph propagation through matrix-valued transport between stalks, still this expressivity alone does not automatically imply effective long-range communication. We introduce ONDA, a long-range graph learning framework based on operator-valued information waves. Stalk-valued representations evolve through second-order dynamics governed by learned sheaf transport operators, combining wave-like propagation with expressive local geometry. We characterize long-range influence through a stalkwise sensitivity analysis and show that, in the conservative self-adjoint setting, the dynamics preserve node information. Across long-range propagation, severe graph bottlenecks, graph transfer, and heterophilic benchmarks, ONDA consistently improves over scalar wave propagation, diffusive sheaf baselines, and state-of-the-art models, demonstrating the benefit of coupling wave dynamics with matrix-valued transport.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
sheaf neural networks, long-range propagation, graph neural networks

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_UdeGOpenReview ↗
Rating
3
Confidence
4
Informativeness
3(+1)
Presentation
3
Soundness
3
Read review
Paper summary

The paper extends two lines of research that explore the information propagation of graph neural networks (GNNs) to overcome the limitations of message passing neural networks (MPNNs). Combining the line of work on using wave-like equations on graphs, and enriching the graph structure to that of a sheaf structure, the paper explores the benefits of such a combination. The paper justifies the construction theoretically, and evaluates the proposed architecture mainly on synthetic benchmarks, but also on a real world benchmark in the appendix.

Critical strengths
  1. Combining wave equations and sheaves is a very natural next step in exploring alternative information propagation on graphs. It clearly builds on the literature and connects two research directions that have not been connected so far.
  2. The paper is very well written, the theory is well-justified, and the experiments are informative.
Critical weaknesses
  1. The theory is relatively straightforward extensions of the theory on sheaves and on wave equations on graphs.
  2. The experiments do not clearly identify and separate how combining the strength of both sheaves and waves distinguishes the architecture to sheaf heat diffusion and to graph wave equation. I would love to see a synthetic task where graph wave equation succeeds but sheaf heat equation fails, one task where graph wave equation fails but sheaf heat equation fails, and finally one task task where both graph wave equation and sheaf heat equation fails. The Table 2 suggests that sheaves do not work there, but it was introduced in BuNNs (Bamberger et. al) to argue that they can overcome oversquashing. Similarly, I would be curious to see the performance of BuNNs on the ECHO tasks.
  3. The authors did not cite the most relevant prior work: "Surfing on the Neural Sheaf" by Suk et. al (Neurips workshop 2022, link below). The reference is a very short workshop extended abstract and do not weaken the contributions of the current work. However, the work should be cited, and discussed in the main text.

link: https://openreview.net/forum?id=xOXFkyRzTlu

Questions for authors
  1. The theory is nice but assumes identity weights and no nonlinearity. It would be nice of proposition 1 could be extended to at least include the learnable weights, and ideally also the nonlinearity (i.e. closed form Jacobian computation for one block in the continuous time setting).
  2. How did the time t parameter influence the synthetic task performance? Each synthetic task chose is intended to test for long-range propagation and/or oversquashing. The parameter that influences this seems to be the propagation time t. I would be curious to see an ablation of the performance of various t on each (or some) of the tasks.
  3. It would be helpful to see the empirical runtime of running a forward as a function of step size and propagation time t, and how it compares to wave-only equation.
  4. What I found interesting in the theory is that this architecture overcomes over-smoothing without requiring assumptions on the sheaf structure. Namely, Bodnar et al requires the sheaf to be path-independence for the long time result to hold, this property does not generally hold. BuNNs use flat vector bundles, which guarantee that this property holds. I wonder if path-dependence/independence has any impact in this theory here. Do you have any thoughts on this?
8 comments
AuthorsOpenReview ↗

Response 1/3

We thank the Reviewer for their thoughtful feedback and for recognizing the clarity of the presentation, the informativeness of our experiments, and the theoretical motivations of our work. We further appreciate that the reviewer found our work to be closing the gap between two previously unconnected fields and that our architecture overcomes over-smoothing without requiring assumptions on the sheaf structure. We address the Reviewer’s concerns below.

W1. We thank the Reviewer for the comment and we would like to clarify the nature of our contribution, which we believe is not a straightforward extension. This perspective was also recognized by Reviewer PGhW, who described our contribution as “a potentially useful research direction” and noted that our work “can help researchers to follow a similar path, modifying how nodes exchange information within sheaf-based architectures”.

Specifically, extending the wave equation from graphs to the sheaf setting requires introducing a different class of operators and mathematical tools, leading to substantially different and more involved theoretical development than in conventional GNNs. Our contribution therefore requires characterizing how matrix-valued transport interacts with oscillatory dynamics and influences long-range information propagation. We think the seemingly “straightforward extension” instead reflects our efforts to establish a clear and intuitive connection between the mathematical frameworks of sheaf theory and wave equations, despite the underlying technical complexity. In the paper, we complement this formulation with an extensive analysis of source-to-target sensitivity, which, to the best of our knowledge, has not previously been developed in the sheaf setting. Specifically, we investigate how matrix-valued transport shapes information propagation, with particular attention to the relationship between continuous-time dynamics and the discrete updates underlying ONDA. During the rebuttal phase, we have further strengthened this analysis through the Jacobian interpretation (lines 987-1025), extensions incorporating learnable weights, learned velocity initialization and nonlinear transformations between blocks (lines 1068-1133), and additional sensitivity bounds (lines 1350-1472). These developments are presented in the revised theoretical appendix and are also discussed in our responses below.

W2. We thank the reviewer for the raised comment, which allows us to further discuss and show the benefits of the combination of sheaves and the wave equation, compared to their use in isolation. To show the importance of both components, we consider the following three settings to illustrate the differences in behavior between NSD (isolating the sheaf component), SONAR (isolating the wave component on a non-sheaf architecture), and ONDA (combining both):

  1. The graph transfer task with n=50 serves as a situation wherein the advantage of wave propagation compared to sheaf diffusive baselines becomes clear. On this task, where the source-to-target distance is 50 hops, SONAR successfully solves the transfer task, whereas NSD does not. This provides a setting where the scalar wave model achieves effective long-range communication that NSD struggles to maintain. Furthermore, the wave-like dynamics of ONDA also allow it to perform well in this task, outperforming SONAR.
  2. To show the complementary role of learned sheaf transport , we refer to the results on the SSSP task from the ECHO benchmark, on which sheaf diffusion is outperforming scalar wave propagation. Concretely, Table 1 shows that NSD achieves a lower MAE than SONAR on single-source shortest-path prediction: 0.231±0.0380.231\pm0.038 versus 0.527±0.2540.527\pm0.254. ONDA further reduces the error to 0.085±0.0190.085\pm0.019. This illustrates a task where the sheaf diffusion baseline performs better than the scalar wave baseline, while their combination in ONDA yields a further improvement.
  3. For the case where both baselines struggle, while ONDA succeeds convincingly, we can consider the results from the Barbell task. Here, both SONAR and NSD obtain test MSEs of 0.809±0.0680.809\pm0.068 and 0.825±0.0320.825\pm0.032 at N=50, respectively, while ONDA achieves 0.083±0.0850.083\pm0.085. Thus, ONDA satisfies the task’s solving criterion of MSE below 0.25, while neither baseline does. This comparison supports the benefit of combining oscillatory dynamics with learned sheaf transport under a severe communication bottleneck. Regarding the request to compare performance with BuNN on ECHO, unfortunately, the BuNN implementation is not publicly available, preventing us from directly evaluating it in our experiments.
AuthorsOpenReview ↗

Response 2/3

Q1. We thank the Reviewer for this suggestion, which has helped us improve the theoretical analysis and clarify its connection to the architecture. Following this suggestion, we have extended the theoretical discussion in Appendix A. Here, we summarize the new statements and findings.

First, Appendix A.4 makes the Jacobian interpretation of our source-to-target sensitivity precise. Lemma 3 establishes that this sensitivity is exactly the corresponding block of the Jacobian, describing how the target state changes under a perturbation of the source’s initial state. This provides the foundation for extending the analysis to additional components of the model. In Appendix A.5, we augment the analysis, as suggested by the Reviewer, to include multiple feature channels, learnable weights W1W_1 acting on stalk coordinates, W2W_2 mixing feature channels, and the learned velocity initialization X˙(0)=X‾WV\dot X(0)=\overline{X}W_V. Proposition 3 then provides an exact expression for the continuous-time Jacobian of a propagation block and its expansion over graph walks. The result preserves the interpretation established in Proposition 1: source-to-target sensitivity is expressed through compositions of local transport operators along walks. The learned propagation weights enter these operators, while the input-dependent velocity initialization contributes an additional term to the Jacobian. This extension therefore retains the original analytical structure while incorporating additional architectural components.

Finally, Corollary 1 incorporates the nodewise MLPs applied between propagation blocks, which introduce nonlinear transformations in ONDA. Whenever these maps are differentiable, their Jacobians compose with the propagation Jacobians from Proposition 3 through the chain rule. Repeating this composition gives the Jacobian of the resulting sequence of propagation blocks and nonlinear maps. This result illustrates how the source-to-target sensitivity analysis extends to the whole learnable model.

Q2. We thank the reviewer for pointing out the importance of the propagation time in ONDA, which intuitively describes how long nodes were able to exchange information. We remark that, in the continuous formulation, it describes the elapsed time within a propagation block, while the discrete model approximates these dynamics through LL updates of step size hh, giving a propagation horizon of exactly T=LhT=Lh. Increasing hh at fixed LL extends this horizon without changing the number of local aggregation steps, whereas increasing LL at fixed hh increases both. Since hh also determines the time discretization, we study these choices through two complementary performance ablations. On ECHO, we vary the step size while keeping the number of updates and all other hyperparameters fixed at the configuration selected for each task. Table 12 shows that increasing the step size from 0.0010.001 to 0.20.2 reduces the mean test MAE from 3.6303.630 to 0.1060.106 on SSSP and from 5.5725.572 to 0.8720.872 on eccentricity. Diameter prediction exhibits a different pattern: the error decreases from 4.8214.821 at h=0.001h=0.001 to 1.2331.233 at h=0.01h=0.01, then increases to 1.3941.394 at h=0.2h=0.2. Thus, extending the propagation horizon through a larger step size does not improve performance uniformly across tasks, showing that small variations in the step size have little impact on the performance, while small variations in the step size have little impact on the performance.

In Table 11, we complement the above comparison with a controlled experiment on the Barbell task with N=10N=10 nodes in each clique. We vary the step size hh and the number of updates LL. Results are means over five training seeds at validation-selected checkpoints, with forcing and dissipation disabled. The comparison across step sizes illustrates that LL must be considered together with hh. For L=16L=16, step sizes h=0.2,0.5h=0.2, 0.5, and 1.01.0 give horizons of T=3.2,8T=3.2, 8, and T=16T=16, with MSEs 0.8775,0.13830.8775, 0.1383, and 0.02930.0293, respectively.The substantial variation in error across these settings highlights the importance of choosing the step size hh and the number of propagation steps LL jointly, rather than attributing performance to TT alone. For your convenience we report Table 11 below.

Step size h ↓h\, \downarrow / updates L →L\, \rightarrow6810121416
1.00.32800.11820.07240.04450.03440.0293
0.50.91160.70190.34540.21610.17970.1383
0.20.94780.94290.93970.92020.91490.8775
AuthorsOpenReview ↗

Response 3/3

Q3. We thank the Reviewer for this helpful suggestion to extend the computational evaluation. Table 9 (in the original manuscript, now Table 13) already provides a comparison between ONDA and the wave-based SONAR model, including forward inference latency. Both models use approximately 100k trainable parameters, four blocks, and 12 propagation updates per block.

For a fixed graph and model configuration, the number of updates LL determines the propagation cost, since it determines how many times the propagation operator is applied. As discussed in Q2, together with the step size hh, it determines the horizon T=LhT=Lh. Changing hh while keeping LL fixed changes this horizon without changing the number of operations, i.e., the computational cost of a forward/backward pass. Moreover, LL plays the role of the number of local aggregation steps, providing a natural basis for comparison with the message-passing depth of other architectures.

To complement the existing comparison, we report additional inference-time measurements on ogbn-arxiv, using a single block with increasing numbers of updates (LL) for ONDA, SONAR, and corresponding numbers of message-passing layers for NSD and CSNN. The results are reported in Table 14 in Appendix D and can also be found below for your convenience. Here we clearly observe that, although runtime differs based on the type of restriction maps, ONDA scales graciously while increasing the iterations, and at L=8L=8, every ONDA variant uses less training time and memory than its corresponding NSD variant and CSNN.

computational costs on ogbn-arxiv wrt LL

ModelL=1 Mem.L=1 TrainL=2 Mem.L=2 TrainL=3 Mem.L=3 TrainL=5 Mem.L=5 TrainL=8 Mem.L=8 Train
SONAR2193.923.62661.332.63127.042.14061.559.85461.286.4
Diag-NSD3511.766.65689.0131.47866.2196.912218.9325.718750.2535.7
Ortho-NSD4136.9124.16859.3257.29580.3376.015025.1626.023191.8985.0
General-NSD4693.2213.08226.1425.611759.6640.718827.21068.229427.61731.6
CSNN2478.274.73738.7148.24998.4222.77517.7370.311295.2591.5
Diag-ONDA3512.467.94139.790.24767.1112.46021.8158.57903.8225.3
Ortho-ONDA3995.6125.74423.0161.24788.3197.45479.0269.86513.0378.1
General-ONDA4693.2216.65320.4266.45948.5317.77203.3420.99084.7575.7

Q4. We thank the Reviewer for this insightful observation. As discussed in [1], over-smoothing is closely related to contractive propagation dynamics and vanishing gradients. ONDA addresses this through non-dissipative oscillatory dynamics, which prevent the systematic attenuation of information, as established in Section 4. Therefore, as correctly observed by the Reviewer, ONDA avoids the over-smoothing mechanism associated with dissipative propagation. In diffusion-based SNNs, path independence plays an important role because the long-time behavior is determined by the kernel of the sheaf Laplacian, whose dimension can be related to the underlying transport structure (see for instance Bodnar et al., 2022, Lemma 6). In contrast, ONDA's non-decaying behavior does not rely on the existence of a nontrivial kernel, since its oscillatory dynamics preserve contributions from positive-eigenvalue modes, as becomes clear from Proposition 2, which shows that the positive-eigenvalue contributions do not decay. Thus, persistence of a nontrivial source-to-target sensitivity does not rely on a nontrivial harmonic space or on path-independent transport, and can hold even when the harmonic space is trivial This is a fundamental distinction between ONDA and dissipative models. Path dependence can still influence the spectral structure and the resulting information propagation patterns, but it does not alter the non-decaying nature of nonzero source-to-target sensitivities established in Section 4.


We hope these clarifications, revisions and additional experiments have addressed the Reviewer’s concerns.

Reviewer_PGhWOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ONDA, a coupling of the Neural Sheaf Diffusion model with oscillatory dynamics. Augmenting the dynamics with state-dependent dissipation and external forcing, ONDA achieves SOTA results on the ECHO-Synth and Barbell-clique benchmarks, and competitive performance on the Graph Transfer tasks.

Critical strengths
  • The proposed model achieves strong results on the long-range and over-squashing benchmarks, highlighting its practical relevance for this type of task.

  • Modifying the underlying propagation mechanism of sheaf-based architectures is a potentially useful research direction. This paper can help researchers to follow a similar path, modifying how nodes exchange information within sheaf-based architectures

Critical weaknesses
  • The theoretical development seems incomplete: it is shown that source-to-target responses do not converge to zero as t→∞t \to \infty, yet there seems to be no guarantee preventing these quantities to become arbitrarily small, no bounds are provided and it is hard to understand how these results connect to the model itself. The experiments do not establish that the theoretical properties explain the performance.
  • Scalability seems a major concern for ONDA: the NSD model is already a heavy-cost architecture, and introducing a more intricate propagation paradigm can make the new model hard to use in practice. The computational comparison with SONAR is useful, but the manuscript lacks comparison to other sheaf neural networks and experiments on larger datasets.
Questions for authors
  • Could you add a comparison against BuNN on the long-range and over-squashing benchmarks?
  • In the abstract, the authors claim that ONDA improves over SOTA on the heterophilic benchmarks, which is not true.
  • Does CSNN also employ different restriction map types as hyperparameters? Would it be possible to make the same ablation from Table 6 with CSNN?
  • Could authors explain the relationship between the source-to-target response maps and the Jacobian usually used in sensitivity bounds? - Is it possible to derive bounds similar to the ones in [1, Lemma 1] and [2, Theorem 3.4]? This would help to connect the theoretical development with the practical results.
  • Could authors provide an ablation removing (i) the state-dependent dissipation, (ii) the external forcing, and (iii) removing both? This would help me to better understand how these additional terms improve the model.
  • Could you provide experiments on the TreeNeighborsMatch task? It would also be useful to add NSD and CSNN in the Graph Transfer task.
  • Could you provide runtimes for NSD and CSNN in table 9?
  • I would like to see the runtime and peak memory on larger datasets, such as the ones from [3] and [4].

I highly encourage the authors to engage. I am willing to raise my score during the rebuttal phase.

[1] Understanding over-squashing and bottlenecks on graphs via curvature, ICLR 2022.

[2] SONAR: Long-Range Graph Propagation Through Information Waves, NeurIPS 2025.

[3] Large Scale Learning on Non-Homophilous Graphs: New Benchmarks and Strong Simple Methods, NeurIPS 2021.

[4] Open Graph Benchmark: Datasets for Machine Learning on Graphs, NeurIPS 2020.

[5] Bundle Neural Network for message diffusion on graphs, ICLR 2025.

[6] Cooperative Sheaf Neural Networks, ICLR 2026.

10 comments
AuthorsOpenReview ↗

Response 1/4

We thank the reviewer for recognizing the strong results of our model on long-range interaction and over-squashing benchmarks. We also appreciate the reviewer's acknowledgment of the potential of rethinking propagation mechanisms in Sheaf Neural Networks as a useful research direction. We address their concerns below and hope these clarifications will help improve their assessment towards acceptance.

W1. We thank the Reviewer for the important comment. As correctly pointed out, our theoretical results establish that the sensitivity does not asymptotically converge to zero. Indeed, due to the oscillatory nature of the dynamics, interference between different spectral components can lead to arbitrarily small responses at particular steps. However, the proof of Theorem 1 (Appendix A.7) establishes a stronger quantitative result. Equation 31 of the proof shows that the asymptotic average of the squared source-to-target sensitivity is strictly positive whenever the response is nonzero at some step. This ensures that the sensitivity cannot remain arbitrarily small as propagation depth increases, as otherwise the average would converge to a null value. Hence, there are infinitely many steps at which the source-to-target sensitivity norm is strictly positive.

To further connect our theoretical findings with the empirical behavior of ONDA, we conducted an additional sensitivity analysis on a line graph with a diameter of 50, calculating the norm of the source-to-target sensitivity between node 0 and nodes at increasing distances (i.e., 1, 10, 20, 40, and 50 hops), comparing ONDA against a standard GCN (which we report in Appendix C.3, and below for your convenience).

DistanceONDAGCN
117.30841.73664
104.592732.23517e-06
2014.27461.49012e-06
3011.71342.98023e-06
409.796130
502.846950

The results reveal a clear distinction between the two architectures. GCN exhibits rapidly vanishing sensitivity, dropping to values on the order of 10−610^{-6} already at distance 10 and eventually reaching zero across all distances. In contrast, ONDA maintains substantial source-to-target sensitivity across all evaluated distances, with no systematic decay to arbitrarily small values.

These findings are consistent with our theoretical analysis and provide additional empirical evidence that ONDA mitigates vanishing sensitivity to distant nodes, thereby supporting the connection between ONDA’s oscillatory propagation mechanism and its strong performance on long-range benchmarks.

W2, Q7, Q8. We thank the reviewer for these suggestions. To provide a more comprehensive computational comparison, we extended our runtime experiments to include NSD and CSNN on ogbn-arxiv (a substantially larger graph than those considered in Table 9 of the original manuscript) in Tables 14 and 15. Specifically, we run two comparisons. In the first, we compare the models under matched propagation steps and equal hidden dimension (which here is given by the product of the stalk dimension d and feature channels c), ensuring that each architecture has a matching number of propagation updates. The results show that ONDA achieves comparable or lower runtimes than NSD and CSNN.

Since the parameter counts in the above experiment can vary between the model families, we additionally compared the same three models under a shared parameter count regime. Each model has 3 layers, and stalk dimension d=4, with the number of feature channels adjusted per model to land on comparable parameter counts. The results in Table 15 show that, relative to the corresponding NSD variant, each ONDA variant reduces training time and peak allocated memory. Moreover, ONDA with diagonal and orthogonal restriction maps also use less time and memory than CSNN, while General-ONDA uses less memory but takes more training time per epoch.

Together, these experiments demonstrate that ONDA's improved long-range propagation capabilities do not come at the expense of substantially increased computational cost, supporting its practical scalability to larger graphs.

AuthorsOpenReview ↗

Response 2/4

Q1. We thank the Reviewer for suggesting this comparison. Unfortunately, the BuNN implementation is not publicly available, preventing us from directly evaluating it in our experiments. Nevertheless, we can discuss the results reported by Bamberger et al. (2025) on the Barbell benchmark, which we independently reimplemented following their experimental protocol, as the original benchmark implementation is also not publicly available.

Although the two implementations differ in the randomly generated node features, Bamberger et al. (2025) report an MSE of 0.01 ± 0.07 for BuNN at N=10, while ONDA achieves 0.0007 ± 0.0003 under our experimental setting. While these results are not directly comparable, both models achieve errors well below the 0.25 threshold conventionally used to indicate successful information transfer across the bottleneck. Indeed, the main purpose of this benchmark is to assess whether a model can successfully overcome the information bottleneck, rather than to rank models that already solve the task.

If the Reviewer considers including BuNN's published results beneficial, we would be happy to add them to the comparison, clearly distinguishing them from independently reproduced results. However, we note that BuNN's reported experiments only consider N=10, whereas we additionally evaluate the more complex settings of N=20 and N=50, for which directly comparable BuNN results are unavailable.

Q2. We thank the reviewer for pointing out this imprecise formulation. Our intention was to highlight that ONDA achieves highly competitive performance on heterophilic datasets while consistently outperforming several families of considered models, i.e., MPNNs, Graph Transformers, and Heterophily-Designated GNNs. We have revised the abstract to avoid further confusion.

Q3. We thank the reviewer for this suggestion. We would like to clarify that CSNN employs a single, specifically designed restriction-map parametrization (which they name as conformal restriction map), designed to enable cooperative message passing. For this reason, such an ablation would not be possible since CSNN does not consider the same restriction-map families investigated in our ablation, namely diagonal, orthogonal, low-rank, and general maps, as interchangeable architectural hyperparameters. In particular, CSNN introduces separate source and target conformal restriction maps, allowing nodes to independently control whether they propagate or receive information. This parametrization is central to the cooperative behavior introduced by Ribeiro et al. (2026).

AuthorsOpenReview ↗

Response 3/4

Q4. We thank the Reviewer for these insightful questions, which have helped us strengthen the theoretical analysis and clarify its connection to the model. Following the suggestions, we have extended the theoretical appendix to address both points.

First, we added a detailed discussion in lines 987-1025 that explains the precise relationship between the source-to-target sensitivity used throughout the paper and the usual Jacobian sensitivity analysis. Specifically, we differentiate the node states at a fixed time or depth with respect to their initialization, holding the sheaf geometry and initial velocity fixed. Equation 18 shows that restricting this differential to perturbations supported at the source and projecting its output onto the target yields the corresponding Jacobian block, thereby explaining the precise relationship between the differential and the source-to-target sensitivity.

For the inquiries involving the stability bounds, we have added two such bounds in Appendix A.8. First, Proposition 4 bounds how strongly an initial perturbation at a source node can affect a target node after a given number of propagation updates. This statement incorporates learnable weights, learned velocity initialization, state-dependent nodewise dissipation, and linear forcing into the analysis. The bound obtained in Equation 33 combines local contributions across successive updates, making the roles of propagation depth, step size, and learned sheaf transport explicit. Interestingly, this result follows the same general structure as [2, Theorem 3.4], while accounting for the matrix-valued transport present within the ONDA architecture. Finally, Corollary 3 specializes the analysis to conservative propagation at depth equal to the source-to-target distance. The resulting bound in Equation 35 is a sum of products along shortest paths, and clearly echoes the statement [1, Lemma 1]. In that result, Topping et al. bound source-to-target sensitivity by combining bounds on the derivatives of the message and update functions with a power of the normalized adjacency matrix. This matrix power admits a graph-walk expansion: each entry sums products of local adjacency weights along walks between the source and target. In ONDA, the local factors are the operator norms of the sheaf Laplacian blocks coupling neighboring stalks, each scaled by the square of the integration step size. This explicitly relates the sensitivity estimate to the matrix-valued transport provided by the sheaf.

Q5. We thank the reviewer for this insightful suggestion. We conducted an additional ablation study on the Graph Transfer benchmark, considering all four possible configurations: the complete ONDA model, without dissipation, without forcing, and without either component.

We report the results on the three graph topologies (Line, Ring, and Crossed-Ring), which exhibit different information propagation characteristics, below.

DissipationForcingLine MSE ↓Ring MSE ↓Crossed-Ring MSE ↓
YesYes1.698107e-070.0031795.397578e-07
YesNo8.498423e-070.0014861.021358e-06
NoYes1.400111e-070.0045482.043218e-04
NoNo8.417039e-070.0016758.981296e-05

The results highlight that the contribution of dissipation and forcing depends on the graph topology. On Line graphs, where information propagates along a unique path between source and target, forcing is particularly beneficial, potentially helping sustain the signal over long distances. On Ring graphs, where waves can travel along multiple routes and recirculate, dissipation becomes more useful for attenuating unwanted oscillations. Finally, on Crossed-Ring graphs, the presence of additional paths and cycles creates more complex propagation patterns, for which combining dissipation and forcing yields the best results. In this setting, dissipation can suppress unwanted oscillations, while forcing provides additional flexibility to sustain or reinforce relevant signals. Importantly, ONDA achieves low errors even without either component across all three topologies, demonstrating that the conservative oscillatory sheaf dynamics are already effective at propagating information over long distances. The additional forcing and dissipation terms provide task-dependent flexibility, rather than being necessary for successful information transfer.

We included this discussion in Appendix C.4.

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

The authors proposed ONDA, a sheaf neural network (SNN) in which stalk-valued node states evolve through second-order oscillatory dynamics governed by a learned sheaf Laplacian, for long-range propagation on graphs. The influence between connected nodes in ONDA never vanishes over depth. In experiments, ONDA outperforms scalar oscillatory GNNs, diffusive SNNs, and some baselines on three synthetic benchmarks for long-range propagation, graph bottleneck, and graph transfer, and on five heterophilic datasets.

Critical strengths
  1. The theoretical analysis is self-contained. It derives the closed-form solution of the conservative form of ONDA, the pathwise and spectral representations of the source-to-target sensitivity (Propositions 1–2), and the discrete source-to-target sensitivity under a stability condition with its convergence to the continuous counterpart (Theorem 1, Corollary 1), with complete proofs in Appendix A.

  2. The empirical evaluation is extensive. It covers long-range propagation, graph bottleneck, graph transfer up to 50 hops, and heterophilic tasks, with a restriction-map ablation against NSD (Table 6) and a matched-parameter cost comparison with SONAR (Table 9). ONDA reduces the eccentricity error from 2.393 to 0.885 (Table 1) and is the only model solving Barbell at N = 50 (Table 2).

Critical weaknesses
  1. The presentation of the proposed methods uses inconsistent terms. For example, the "operator-valued information waves" in the Abstract and Sec. 1 are never formally defined, and the influence between nodes is named "cross-influence" (Abstract), "source-to-target influence" (Eq. 8), "source-to-target response" (Proposition 1), and "source-to-target sensitivity" (Proposition 2). Meanwhile, this problem also happens in the theoretical claim, where the Abstract states that "the cross-influence never vanishes", yet Theorem 1 shows that the response is zero for fewer steps than the graph distance and only "does not converge to zero" afterwards (L.343–344), making the claim not demonstrated. Besides, the references to Table 2 and Table 6 are swapped in Sec. 5.2 and Appendix C.1.
Questions for authors

None

2 comments
AuthorsOpenReview ↗

Response 1/1

We thank the Reviewer for their constructive feedback and for recognizing the self-contained theoretical analysis, the extensive empirical evaluation, and that ONDA outperforms scalar oscillatory GNNs, diffusive SNNs, and some baselines on three synthetic benchmarks for long-range propagation, graph bottleneck, and graph transfer, and on five heterophilic datasets. We appreciate their careful reading and the opportunity to improve the precision and consistency of the paper. We address their concerns below.

On the terminology

We thank the reviewer for their observation with respect to terminology, which we clarified in the revised manuscript to avoid unnecessary ambiguity. Accordingly, we have standardized the terminology to source-to-target sensitivity throughout the manuscript, referring to both its continuous formulation Jv←uF(t)J^{\mathcal F}_{v\leftarrow u}(t) and its discrete counterpart Jv←uF[ℓ,h]J^{\mathcal F}_{v\leftarrow u}[\ell,h].

We have also clarified the formal definition of operator-valued information waves after Equation (7). Specifically, the term operator-valued reflects that, at each time tt, CF(t)\mathcal C_{\mathcal F}(t) is an operator that maps the initial perturbations to their propagated values, while wave refers to the fact that this family satisfies the sheaf wave equation C¨F(t)+LFCF(t)=0\ddot{\mathcal C}_{\mathcal F}(t) + L_{\mathcal F} \mathcal C_{\mathcal F}(t) = 0. Consequently, applying the family to any initial perturbation produces a cochain-valued solution of the sheaf wave equation, with the learned sheaf determining how these perturbations are transported and transformed between stalks. With the initial velocity held fixed, its action on a perturbation supported at node uu gives ΔX(t)=CF(t)ιuW\Delta\mathbf X(t)=\mathcal C_{\mathcal F}(t)\iota_u\mathbf W. The source-to-target sensitivity is the corresponding block πvCF(t)ιu\pi_v\mathcal C_{\mathcal F}(t)\iota_u.

On the vanishing claim

We thank the Reviewer for the comment, and we have now revised the formulation to distinguish non-convergence to zero from being nonzero at every propagation step.

Specifically, under the setting analyzed in Section 4 and under the stability condition h2λmax⁡(LF)<4h^2\lambda_{\max}(L_{\mathcal F})<4, Theorem 1 establishes two compatible properties. For nodes at graph distance rr, the discrete sensitivity is zero for ℓ<r\ell<r. If the sensitivity is nonzero at some step, it does not converge to zero as ℓ→∞\ell \to \infty. Individual later zeros remain possible because oscillatory contributions allow for cancelling interference patterns. However, it will never converge to zero or be zero indefinitely.

We have therefore aligned the Abstract and the discussion following Theorem 1 with the precise non-decaying statement. In particular, we replaced “does not vanish” with “does not converge to zero” in the discussion of the discrete dynamics. These revisions preserve the mathematical result while correcting the interpretation suggested by the initial prose.

On table references

We thank the reviewer for noticing the typo; we fixed it in the revised paper.


We hope these clarifications and revisions have addressed the Reviewer’s concerns.

Dear Reviewer DWtp,

As the rebuttal discussion period is approaching its end, we would like to kindly follow up on our response.

We have revised the manuscript to clarify the terminology and formal definition of operator-valued information waves, refine the interpretation of our theoretical guarantees regarding non-vanishing sensitivity, and correct the reported typo.

We hope these revisions have addressed your concerns and would appreciate any further feedback. We would be happy to clarify any remaining points and hope you might reconsider your score in light of these improvements.

Best regards,

The Authors

Comments on the paper

1 comment
Area_Chair_6Ls8OpenReview ↗

Fairness of the Experimental Setup and Comparisons

Dear Reviewers and Authors,

The discussion on this paper has become quite heated. The main point of contention is whether the experimental setup and the comparisons are fair.

I would like to invite all reviewers to carefully check this issue and share their assessment, as it is critical to the evaluation of this paper.

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.