acceptodds
Under review as a conference paper at ICLR 2027

Equivariant Sheaf Neural Networks: Learning Geometric Transport on Graphs

Abstract

Equivariant graph neural networks constrain how geometric features transform, but efficient first-order architectures remain limited in how vector information can be transported across graph edges. We introduce ESNN, an Equivariant Sheaf Neural Network that learns directed, matrix-valued transport between neighboring vector features while preserving exact Euclidean equivariance. Rather than increasing representation order, ESNN retains first-order scalar and vector features and places additional geometric expressivity in the edge interactions themselves. Its transport mechanisms range from isotropic transformations to feature-conditioned rotations and anisotropic radial–tangential actions, allowing vector messages to be transformed according to the local geometry before aggregation. We also equip ESNN with a controlled symmetry-relaxation mechanism for systems with a preferred ambient direction, preserving exact equivariance to the remaining symmetry while recovering full -equivariance as a special case. Across diverse geometric learning problems spanning physical dynamics, point clouds, and molecular systems, ESNN improves predictive performance over first-order equivariant baselines, learns a latent preferred direction from data, and remains robust to unseen rotations. These results establish edge-wise geometric transport as a source of equivariant expressivity that does not require higher-order representations.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Sheaf Neural Networks, Equivariance, Graph

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_7GdcOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces Equivariant Sheaf Neural Networks (ESNNs), a generalized geometric graph learning framework that enriches edge interactions through learned, directed matrix-valued transport while retaining scalar and first-order vector features. Each transport combines a spatially covariant operator (S) with invariant channel mixing opertaor (M), preserving Euclidean equivariance.

Critical strengths

The paper presents a concise framework that separates equivariant spatial transformations from invariant channel mixing. This provides an understandable way to construct geometry-dependent edge interactions while retaining scalar and first-order vector features. The accompanying equivariance guarantees and geometric interpretations make the framework useful for researchers designing structured message-passing architectures.

Critical weaknesses
  1. The framework is interesting and clearly formulated, but I found the additional insight provided by the proposed transport families somewhat limited. Section 4 explains their geometric actions, yet offers relatively little guidance on when these differences are useful in practice. For example, Diagonal Transport performs best on charged N-body dynamics, whereas Radial–Tangential Transport is strongest on DeformingPlate. A more detailed discussion connecting each map’s inductive bias to the structure of the tasks and the observed results would help clarify their respective advantages and guide the choice of transport for a new application. Given that the framework builds naturally on existing sheaf and equivariant constructions, such insights would strengthen the paper’s conceptual contribution.

  2. The empirical advantages over other baselines are somewhat task-dependent. For example, MeshGraphNet achieves lower mean errors on Airfoil and lower full-rollout error on CylinderFlow, despite ESNN’s improvements over EGNN. However, the authors do not claim state-of-the-art performance, and I do not view these mixed results as undermining the paper’s central contribution. This is a minor limitation concerning the breadth of the empirical benefits.

Questions for authors
  1. Table 1(b) compares ESNN variants with no, fixed, or learned directional conditioning. Could the authors include a comparison with an existing subequivariant method, such as SGNN (Han et al., 2022), on the same gravity-augmented N-body task? In particular, supplying the same known gravity direction to the baseline and ESNN-Fixed would allow a comparison under matched information. This would help assess ESNN’s relative performance against an established method designed to accommodate the reduced symmetry.

[1] Learning Physical Dynamics with Subequivariant Graph Neural Networks

  1. I appreciated the distinctions drawn in Appendix C between directed transport, classical sheaf diffusion, and connection-sheaf constructions. Building on this discussion, I would be interested in whether a bundle/connection viewpoint offers further intuition for the learned ESNN transports, particularly when channel mixing and gating make the full maps non-orthogonal. Could this perspective help explain the roles of the different transport families? This is a discussion point rather than a concern affecting my assessment.
6 comments

We thank the Reviewer for the thoughtful assessment and for recognizing the clear separation of equivariant spatial actions from invariant channel mixing, the equivariance guarantees, and the framework's usefulness for designing structured message-passing architectures. We appreciate the request for more practical guidance and a comparison with subequivariant methods. We address both weaknesses and questions below.

W1 — Choosing geometric transport maps

The choice of transport reflects the geometric structure of the interactions being modeled:

  • Identity: provides a reference without learned geometric transport.
  • Diagonal: performs isotropic spatial modulation and channel mixing, which may suffice when directional information is already captured by other components of the architecture.
  • Orthogonal: introduces feature-conditioned spatial reorientation, potentially useful when neighboring vector signals have different orientations and must be combined before aggregation.
  • Radial–Tangential: independently transforms components parallel and perpendicular to an edge, naturally suited to interactions involving distinct longitudinal and transverse responses.

Theorem 4.1 characterizes displacement-conditioned linear O(n)O(n)-equivariant transport as Tr(V)=P∥VA+P⊥VBT_r(V)=P_\parallel VA+P_\perp VB, showing that radial–tangential anisotropy can be modeled while preserving Euclidean equivariance.

The experimental results provide further insight into how these geometric inductive biases relate to the underlying physical interactions:

  • DeformingPlate: This benchmark models the quasi-static deformation of a hyperelastic plate under prescribed actuator motion. Deformation produces different responses along and transverse to local mesh connections, making direction-dependent transport particularly relevant. Radial–Tangential could explicitly distinguish these components, allowing the model to learn different transformations for longitudinal and transverse vector information. Consistent with this inductive bias, it reduces full-rollout RMSE by 37% relative to Diagonal, despite having 2.7% fewer parameters.

  • Airfoil: The compressible flow around an airfoil exhibits strong spatial variations in velocity direction, particularly near its curved boundary and in the wake. Orthogonal Transport can reorient incoming vector features according to the local geometric context before aggregation, potentially facilitating the propagation of information across differently oriented flow regions. Consistent with this interpretation, Orthogonal reduces full-rollout RMSE by approximately 20% relative to Diagonal at nearly identical parameter counts. Radial–Tangential achieves the lowest ESNN full-rollout error, indicating that directional anisotropy can also be beneficial in this setting.

  • CylinderFlow: Unlike Airfoil, this benchmark considers transient incompressible flow around a cylinder, with evolving wake structures and vortex shedding. The results reveal an interesting dependence on prediction horizon: Orthogonal achieves the lowest ESNN errors at one and 50 steps, while Radial–Tangential performs best over the full rollout. This suggests that spatial reorientation and direction-dependent modulation may offer different advantages for short-horizon prediction and long-term propagation, although the results do not isolate the underlying mechanism.

  • Charged N-body: Electrostatic forces act along relative displacement directions, which are already incorporated into the coordinate updates of the architecture. This may explain why the simpler Diagonal Transport performs best in the original five-body setting (MSE 0.0053). However, central-force interactions are not necessarily isotropic in their response to perturbations: radial and transverse displacements generally produce different responses. Radial–Tangential can represent this distinction explicitly and achieves lower mean errors than Diagonal in the additional 20-, 50-, and 100-body experiments (Appendix E.1, Table 5; see Reviewer gFYH, Q1).

Together, these findings suggest a practical principle: the appropriate transport depends on which geometric transformations are useful for the underlying interactions and which are already represented elsewhere in the architecture. Isotropic transport may suffice when directional information is already available; spatial reorientation may be beneficial for varying vector orientations; and radial–tangential transport provides an explicit inductive bias when longitudinal and transverse responses differ. The task-dependent results also indicate that greater geometric flexibility does not guarantee better performance. We have added these considerations to Section 4 and Section 7 (Q3).

We continue in the next comment.

W2 — Task-dependent performance

We appreciate the Reviewer's observation that the mixed results do not undermine the central contribution. As discussed in W1, ESNN provides a general framework for geometric transport rather than a universally optimal transport architecture.

While the specialized MeshGraphNet baseline remains stronger on some benchmarks (Table 2), our additional parameter-matched multi-channel EGNN comparisons help isolate the contribution of structured transport. At virtually identical parameter counts, Radial–Tangential reduces full-rollout RMSE by 67% on CylinderFlow and 56% on DeformingPlate (Section 7, Q3; Appendix E.4, Table 6; see Reviewer gFYH, Q2).

These results reinforce the distinction between the value of geometric transport as an architectural inductive bias and the task-specific performance of a particular model. They support ESNN's central contribution without requiring it to outperform specialized architectures on every benchmark.

Q1 — Comparison with SGNN

We conducted a direct comparison with SGNN (Han et al., 2022) under matched directional information, together with a supplementary check of the learned-direction setting. Known gravity direction. We compared SGNN with ESNN-Orthogonal-Fixed on the same 20-body gravity-augmented task, supplying the correct gravity axis to both architectures. We retained the evaluation protocol of the original manuscript, extending the dataset to 20 bodies. The results are coordinate-averaged test position MSE (mean ± sample SD over three runs):

ModelParametersTest MSE ↓
SGNN (known axis)186,4530.0079 ± 0.0005
ESNN-Orthogonal-Fixed191,3490.0105 ± 0.0002

SGNN achieves lower prediction error in this setting, consistent with its more specialized treatment of symmetry breaking. Specifically, SGNN incorporates the prescribed gravity direction into its geometric vector basis and Gram-invariant features, enabling gravity-conditioned interactions throughout message passing and node updates, supported by theoretical universality guarantees for subequivariant functions. ESNN instead incorporates directional conditioning into its general geometric transport framework, without introducing dedicated subequivariant representations. As established in Theorem 6.1, this formulation recovers full E(n)E(n)-equivariance when directional conditioning is disabled, while retaining equivariance under the stabilizer of the preferred direction when it is active.

Withheld gravity direction — supplementary check. We additionally considered a separate five-body experiment in which the gravity axis was not supplied. The input was taken before the first force update, preventing the observed velocities from revealing the direction through previous dynamics.

As expected, stock SGNN, which assumes a prescribed preferred direction, performs poorly without this information (MSE 8.0895 ± 0.0080). In contrast, ESNN-Diagonal-Learned achieves 0.0229 ± 0.0029 MSE, inferring the gravity axis from training trajectories and recovering its orientation within 0.14° up to sign across three runs, without explicit supervision of the direction. This is consistent with the original five-body ablation in Table 1(b).

This supplementary check is intended to illustrate the difference between the original architectures, rather than establish a performance advantage over SGNN.

Overall, SGNN offers a specialized architecture for systems with known symmetry breaking, whereas ESNN integrates fully equivariant, fixed-direction, and learned-direction behavior within the same geometric transport framework.

We report both comparisons here, distinguishing the primary matched-information experiment from the supplementary withheld-direction check.

We continue in the next comment.

Q2 — Bundle/connection interpretation of non-orthogonal transport

We appreciate the Reviewer's interest in this aspect of our work. The bundle/connection viewpoint provides useful geometric intuition for understanding the different transport families.

As discussed in Appendix C, ESNN associates each node with a vector-feature fiber Rn⊗Rcv\mathbb{R}^n\otimes\mathbb{R}^{c_v}, connected through learned directed transport maps. In this interpretation, the spatial operators determine how vector information is geometrically transformed between neighboring fibers, while channel mixing and gating control how that information is combined and transmitted (Definition 3.1).

This perspective also clarifies the roles of the transport families introduced in Section 4. Identity transfers vector features unchanged; Diagonal introduces isotropic modulation and channel mixing; Orthogonal allows feature-conditioned spatial reorientation; and Radial–Tangential independently transforms longitudinal and transverse components. Each therefore imposes a different geometric structure on the interactions between neighboring fibers.

Unlike parallel transport in a metric connection, these maps need not preserve the full fiber norm. For instance, even when Orthogonal Transport uses a spatial rotation, its channel mixing and gating can amplify or attenuate the transmitted features. More generally, ESNN can be interpreted as learning directed, potentially non-isometric transport between local feature spaces, allowing geometric alignment and feature modulation to be learned jointly.

Appendix C.3 formalizes this relationship: Proposition C.1 identifies the adjoint-consistent specialization, while Proposition C.2 recovers connection-sheaf transport when the complete fiber maps are orthogonal and reverse transports are their adjoints. Thus, classical connection-sheaf constructions arise as a special case of the broader transport formulation considered by ESNN.

This perspective helps explain the geometric inductive biases of the proposed transport families.

We hope these clarifications and revisions address the Reviewer’s concerns, and we would be grateful if they would consider reassessing their evaluation and score based on the updated text and the above comments.

Reviewer_6kmLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ESNN, a message-passing network for 3D graphs where each node has scalar features and vector features. The new part is how a vector is passed from one node to its neighbour. Instead of copying it, the model applies a learned map, which the authors call transport. They give four versions. Identity passes the vector unchanged, Diagonal scales it, Orthogonal rotates it with a rotation built from the two nodes' vectors, and Radial-Tangential uses separate maps along and across the edge. Experiments cover charged N-body with and without gravity, three MeshGraphNets tasks, ModelNet40 under rotations and QM9. ESNN beats EGNN wherever EGNN is included, and beats MeshGraphNet by a large margin on DeformingPlate.

Critical strengths
  1. the paper is honest about their scope. It does not claim gauge equivariance or state of the art, and it reports single runs and unfavourable results openly.
  2. the mesh based task shows very promising results
Critical weaknesses
  1. only mesh tasks test the learned transport fairly. On modelnet40, identity is as good as learned versions from the single run in table 2. To show that the transport helps, you need to compare it with Identity while keeping everything else the same. The tables only do that on the mesh tasks.
  2. Once the claims are narrowed, the contribution is incremental. The paper gives no result showing that ESNN can represent something these models cannot. A short paragraph comparing the transports with PaiNN, GMN and Han et al. would help a lot.

[1] Jiaqi Han, Wenbing Huang, Hengbo Ma, Jiachen Li, Joshua B. Tenenbaum, and Chuang Gan. Learning Physical Dynamics with Subequivariant Graph Neural Networks. Advances in Neural Information Processing Systems (NeurIPS) 35, pages 26256–26268, 2022.

Questions for authors
  1. could you share the five per-run values behind the ESNN numbers of Table 3, and say how the ± values were computed?
  2. what does Identity get on charged N-body with the same size and training as Diagonal, over five runs?
  3. on Airfoil or DeformingPlate, how do Diagonal and Orthogonal or Radial-Tangential compare at the same model size? Or, is there something an ESNN layer can compute that a simpler first-order layer cannot?
  4. On the gravity task: is the input the first frame or a later one, and is the error summed or averaged over coordinates?
8 comments

We thank the Reviewer for their careful assessment and for recognizing the honest scope of our work and the promising results on mesh-based dynamics. We appreciate the suggestions for clarifying the contribution and strengthening the experimental comparisons. Below, we address the weaknesses and questions raised.

W1 — Evidence for learned transport beyond Identity

We acknowledge that the advantage of learned transport was harder to establish from the original manuscript. However, we would also like to highlight that better-performing configurations did not necessarily have more parameters, nor did the more expressive transport families systematically require larger models. As detailed in Q2 (charged N-body) and Q3 (MeshGraphNet), the parameter counts and corresponding results provide further evidence of this observation.

Following the Reviewer's suggestions, we conducted additional parameter-matched experiments to better assess the contribution of geometric transport independently of model capacity and the availability of multiple vector channels. These include comparisons on charged N-body and with a stronger multi-channel EGNN on all three mesh benchmarks. The latter show substantial improvements at virtually identical parameter counts (Section 7, Q3; Appendix E.4, Table 6; see Q3 and our response to Reviewer gFYH, Q2).

The additional N-body experiments retain the original evaluation protocol, extending the datasets to 20, 50, and 100 bodies (Appendix E.1, Tables 4–5; Reviewer gFYH, Q1). The multi-channel mesh comparison likewise follows the original evaluation protocol, using three seeds for both MC-EGNN and ESNN-Radial–Tangential (Appendix E.4, Table 6; Reviewer gFYH, Q2).

W2 — Novelty and relationship to PaiNN, GMN, and SGNN

We thank the Reviewer for the opportunity to better clarify the novelty of ESNN and its distinctions from existing architectures.

ESNN builds on the matrix-valued message-passing formulation of Sheaf Neural Networks (Hansen & Gebhart, 2020; Bodnar et al., 2022), where restriction maps induce node-to-node couplings of the form ρi→e∗ρj→e\rho_{i\to e}^{*}\rho_{j\to e}. Building on directed sheaf transport, ESNN introduces Euclidean-equivariant geometric transport between vector-valued node features. Its relationship to classical sheaf constructions, including adjoint-consistent and connection-sheaf specializations, is established in Section 2.1 and Appendix C.

The main theoretical contribution is to characterize the geometric structure of these learned transport maps. Definition 3.1 combines covariant spatial operators with invariant channel mixing, with equivariance established in Proposition 3.2. More importantly, Theorem 4.1 characterizes every linear, displacement-conditioned O(n)O(n)-equivariant transport (with invariant scalar conditioning) as:

Tr(V)=P∥VA+P⊥VBT_r(V)=P_{\parallel}VA+P_{\perp}VB

This identifies the general radial–tangential form and shows how anisotropic transport (A≠BA\neq B) extends isotropic channel mixing (A=BA=B) while preserving equivariance. Orthogonal Transport further incorporates feature-conditioned spatial reorientation (Appendix B.3). Q3 illustrates the resulting distinction at the local-operator level.

Regarding the cited architectures:

  • PaiNN: combines vector features with displacement-aligned messages, whereas ESNN explicitly transforms incoming vector features through geometric edge operators, including independent radial and tangential actions.

  • GMN: constructs geometric interactions using multiple covariant vectors and Gram-invariant coefficients. ESNN instead directly parameterizes symmetry-constrained transport operators, making their geometric structure explicit.

  • SGNN (Han et al.): incorporates a preferred direction to model subequivariance. ESNN preserves full Euclidean equivariance by default and supports optional symmetry relaxation through a supplied or learned global direction (Theorem 6.1).

The novelty therefore lies in the sheaf-based formulation and mathematical characterization of equivariant geometric transport, together with transport families whose different inductive biases are evaluated experimentally.

We have made these distinctions clearer in Section 2 (Related Work).

We continue in the next comment.

Q1 — Individual runs and uncertainty in Table 2

As requested, we report below the five individual run values underlying each ESNN result in Table 2.

The reported ±\pm values correspond to the sample standard deviation across five independent runs, computed as:

\qquad \mathrm{SD}(m)=\sqrt{\frac{1}{4}\sum_{s=1}^{5}(m_s-\bar m)^2}$$ All results are test RMSE in Table 2 units ($\times 10^{-3}$). The individual values are shown at the precision used in the manuscript. **CylinderFlow** | Model | Horizon | Run 1 | Run 2 | Run 3 | Run 4 | Run 5 | Mean ± SD | |---|---|---:|---:|---:|---:|---:|---:| | ESNN-Id | 1-Step | 5.58 | 5.12 | 5.09 | 5.17 | 5.80 | 5.35 ± 0.32 | | ESNN-Id | 50-Step | 21.7 | 19.3 | 21.3 | 20.8 | 19.9 | 20.6 ± 1.0 | | ESNN-Id | Full | 88.27 | 81.70 | 80.43 | 83.03 | 87.01 | 84.09 ± 3.4 | | ESNN-Diag | 1-Step | 2.66 | 2.68 | 2.64 | 3.04 | 2.83 | 2.77 ± 0.17 | | ESNN-Diag | 50-Step | 8.9 | 9.4 | 8.6 | 8.7 | 8.4 | 8.8 ± 0.4 | | ESNN-Diag | Full | 51.12 | 49.39 | 48.30 | 52.19 | 53.20 | 50.84 ± 2.0 | | ESNN-Ortho | 1-Step | 2.64 | 2.50 | 2.34 | 2.59 | 2.73 | 2.56 ± 0.15 | | ESNN-Ortho | 50-Step | 8.9 | 8.4 | 7.9 | 8.5 | 8.8 | 8.5 ± 0.4 | | ESNN-Ortho | Full | 54.41 | 56.95 | 54.79 | 58.27 | 52.68 | 55.42 ± 2.2 | | ESNN-RadTan | 1-Step | 3.30 | 2.87 | 3.07 | 3.33 | 3.23 | 3.16 ± 0.19 | | ESNN-RadTan | 50-Step | 10.0 | 9.4 | 10.7 | 10.2 | 10.6 | 10.2 ± 0.5 | | ESNN-RadTan | Full | 49.47 | 50.23 | 48.50 | 52.96 | 47.79 | 49.79 ± 2.0 | **DeformingPlate** | Model | Horizon | Run 1 | Run 2 | Run 3 | Run 4 | Run 5 | Mean ± SD | |---|---|---:|---:|---:|---:|---:|---:| | ESNN-Id | 1-Step | 0.17 | 0.14 | 0.15 | 0.14 | 0.15 | 0.15 ± 0.01 | | ESNN-Id | 50-Step | 1.1 | 1.4 | 1.3 | 1.3 | 1.3 | 1.3 ± 0.1 | | ESNN-Id | Full | 7.3 | 8.0 | 8.1 | 7.3 | 7.8 | 7.7 ± 0.4 | | ESNN-Diag | 1-Step | 0.16 | 0.17 | 0.19 | 0.17 | 0.17 | 0.17 ± 0.01 | | ESNN-Diag | 50-Step | 1.4 | 1.5 | 1.4 | 1.5 | 1.2 | 1.4 ± 0.1 | | ESNN-Diag | Full | 9.3 | 8.9 | 9.3 | 9.9 | 8.6 | 9.2 ± 0.5 | | ESNN-Ortho | 1-Step | 0.14 | 0.12 | 0.14 | 0.13 | 0.12 | 0.13 ± 0.01 | | ESNN-Ortho | 50-Step | 1.0 | 1.1 | 1.2 | 1.2 | 1.1 | 1.1 ± 0.1 | | ESNN-Ortho | Full | 7.8 | 8.4 | 8.1 | 8.6 | 7.6 | 8.1 ± 0.4 | | ESNN-RadTan | 1-Step | 0.10 | 0.07 | 0.08 | 0.07 | 0.08 | 0.08 ± 0.01 | | ESNN-RadTan | 50-Step | 1.1 | 1.1 | 1.0 | 0.9 | 0.9 | 1.0 ± 0.1 | | ESNN-RadTan | Full | 5.4 | 6.2 | 5.7 | 5.7 | 6.0 | 5.8 ± 0.3 | **Airfoil** | Model | Horizon | Run 1 | Run 2 | Run 3 | Run 4 | Run 5 | Mean ± SD | |---|---|---:|---:|---:|---:|---:|---:| | ESNN-Id | 1-Step | 4596 | 4667 | 4866 | 4948 | 4642 | 4744 ± 154 | | ESNN-Id | 50-Step | 8005 | 7623 | 7299 | 7326 | 7641 | 7579 ± 287 | | ESNN-Id | Full | 16451 | 18027 | 16058 | 16949 | 16980 | 16893 ± 740 | | ESNN-Diag | 1-Step | 2427 | 2439 | 2356 | 2369 | 2473 | 2413 ± 49 | | ESNN-Diag | 50-Step | 4940 | 4875 | 5139 | 4831 | 5045 | 4966 ± 126 | | ESNN-Diag | Full | 11186 | 11149 | 10950 | 10486 | 11299 | 11014 ± 321 | | ESNN-Ortho | 1-Step | 1485 | 1443 | 1443 | 1432 | 1491 | 1459 ± 27 | | ESNN-Ortho | 50-Step | 2878 | 2727 | 2748 | 2852 | 2830 | 2807 ± 66 | | ESNN-Ortho | Full | 8771 | 8563 | 9084 | 8589 | 9098 | 8821 ± 259 | | ESNN-RadTan | 1-Step | 1567 | 1574 | 1519 | 1509 | 1531 | 1540 ± 29 | | ESNN-RadTan | 50-Step | 2441 | 2438 | 2399 | 2304 | 2348 | 2386 ± 59 | | ESNN-RadTan | Full | 8446 | 8943 | 8307 | 8785 | 8519 | 8600 ± 259 | ### Q2 — Identity versus Diagonal at the same size and training In the original charged N-body experiments, Identity has 463,372 parameters and achieves a test MSE of 0.0084, whereas Diagonal uses 222,548 parameters and achieves 0.0053. This suggests that Diagonal's advantage is not associated with a larger parameter budget. We additionally evaluated a parameter-matched Identity configuration (220,012 parameters), obtaining a test MSE of $0.007483 \pm 0.000837$. As shown in Appendix E.1 (Table 4) and our response to Reviewer gFYH (Q1), **Identity continues to exhibit higher mean error than Diagonal and Radial–Tangential at comparable parameter counts**. We continue in the next comment.

Q3 — Comparable-capacity mesh models and geometric expressivity

The original mesh results (Section 7, Q3; Table 2) already provide two comparisons at closely matched parameter counts:

  • Airfoil: Diagonal has 6.095M parameters and Orthogonal 6.125M (0.49% difference). Orthogonal improves all three prediction horizons, reducing full-rollout RMSE from 11,014 to 8,821 (approximately 20%).

  • DeformingPlate: Radial–Tangential has 9.021M parameters, compared with 9.272M for Diagonal (2.7% fewer), while reducing full-rollout RMSE from 9.2 to 5.8 (approximately 37%).

These results indicate that the additional geometric transformations can improve prediction at comparable, or even smaller, parameter budgets.

The distinction can also be expressed mathematically. For a fixed edge context, isotropic transport acts as Tiso(V)=VMT_{\mathrm{iso}}(V)=VM, whereas Radial–Tangential Transport takes the form

TRT(V)=P∥VA+P⊥VBT_{\mathrm{RT}}(V)=P_{\parallel}VA+P_{\perp}VB

The two operators agree for every input VV if and only if A=B=MA=B=M.

For a concrete example, consider one vector channel with v=e1v=e_1, r^=(e1+e2)/2\hat r=(e_1+e_2)/\sqrt{2}, A=1A=1, and B=0B=0. Radial–Tangential Transport produces

TRT(v)=12(e1+e2)T_{\mathrm{RT}}(v)=\frac12(e_1+e_2)

introducing a component along e2e_2 that isotropic channel mixing of the same input vector cannot produce. Orthogonal Transport similarly permits spatial reorientation conditioned on the endpoint features.

Q4 — Gravity input frame and error definition

The gravity experiment uses the first stored trajectory frame, corresponding to simulation step 100, and predicts the state at step 1100, giving a horizon of 1,000 integration steps.

The reported position MSE is averaged over spatial coordinates, particles, and test examples:

=\frac{1}{3BN}\sum_{b=1}^{B}\sum_{i=1}^{N} \|\hat x_{b,i}-x_{b,i}\|_2^2$$ We have clarified the input-frame convention in Appendix E.2. We hope these clarifications and revisions address the Reviewer’s concerns, and we would be grateful if they would consider reassessing their evaluation and score based on the updated text and the above comments.
Reviewer_gFYHOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Equivariant Sheaf Neural Networks (ESNN), a framework that enhances equivariant graph neural networks through more expressive edge-wise transport operators for vector features, including diagonal, orthogonal, and radial–tangential transport. The authors also introduce a mechanism for relaxing Euclidean equivariance in the presence of a preferred ambient direction and evaluate the proposed framework on particle dynamics, mesh-based physical simulations, point-cloud classification, and molecular property prediction. The results show improvements over EGNN on several tasks, although the novelty of the proposed architecture and the extent to which the improvements can be attributed to the new transport mechanisms remain unclear.

Critical strengths

The paper presents a coherent framework for incorporating more expressive edge-wise interactions into equivariant graph neural networks while maintaining the desired symmetry properties. The proposed transport mechanisms are mathematically motivated, and the experiments cover a relatively diverse set of geometric learning tasks, including particle dynamics, physical simulations, and point-cloud classification. The study provides a useful exploration of alternative equivariant transport parameterizations, although their advantages over existing approaches require stronger empirical validation.

Critical weaknesses
  1. Insufficiently controlled experiments. Despite the extensive experiments, the comparisons do not adequately control for model capacity, making it difficult to determine whether the performance gains arise from the proposed transport mechanisms or differences in the number of learnable parameters and overall architecture. Moreover, the five-body N-body benchmark is relatively small and potentially susceptible to overfitting or exploiting task-specific patterns; experiments with larger systems (e.g., 20, 50, and 100 bodies) would provide more convincing evidence of the method's effectiveness and scalability.
  2. Imbalanced presentation. The main text devotes substantial space to theoretical analysis, while the experimental section is comparatively brief and does not sufficiently explain the experimental settings, implementation details, and evaluation protocols. More comprehensive descriptions and controlled analyses are needed to improve the reproducibility of the experiments and substantiate the paper's empirical claims.
Questions for authors

My main concern is whether the reported performance improvements genuinely stem from the proposed architectural design, rather than simply from increased model capacity or selective model/configuration choices. I encourage the authors to address this concern through the following experiments and clarifications:

  1. Parameter counts and scalability on N-body tasks. Please report the number of learnable parameters for all baselines evaluated on the N-body benchmarks, or conduct additional experiments with parameter-matched ESNN and EGNN models. Furthermore, please evaluate the models on larger systems with 20, 50, and 100 bodies to provide stronger evidence of the effectiveness and scalability of the proposed approach.
  2. A multi-channel EGNN baseline. Please modify the edge message in EGNN to support multi-channel feature transformations, following an implementation similar to GMN, and compare this variant against ESNN under matched parameter budgets. This comparison would help clarify whether the proposed transport mechanisms provide benefits beyond those achievable by increasing the capacity of a conventional EGNN.
7 comments

We thank the Reviewer for their careful assessment and for recognizing ESNN as a coherent framework with mathematically motivated, symmetry-preserving transport mechanisms, evaluated across diverse geometric tasks. We appreciate the suggestions for strengthening the experimental controls and clarifying the evaluation protocol. Below, we address the two weaknesses and the corresponding questions.

W1 — Controlled experiments and model capacity

We appreciate the Reviewer's observation that the original experiments would benefit from more explicit controls for model capacity. While parameter counts were not explicitly reported in the main text, the original five-body charged N-body experiments already provide some relevant evidence. ESNN-Identity uses 463,372 parameters and achieves a test MSE of 0.0084, whereas Diagonal uses 222,548 parameters (MSE 0.0053) and Radial–Tangential uses 186,383 parameters (MSE 0.0057).

These results suggest that the improvements are not simply associated with larger parameter counts, as the more geometrically expressive transport families can achieve better performance with fewer parameters.

To address the Reviewer's concern more directly, we conducted additional parameter-matched experiments on five-body charged N-body, larger systems with 20, 50, and 100 particles, and all three mesh benchmarks using a stronger multi-channel EGNN baseline. We report the corresponding results in Q1 and Q2.

W2 — Experimental presentation and reproducibility

In the revised manuscript, we have expanded the experimental discussion in Section 7 (Q1 and Q3), making the role of the additional controls and their relationship to the original results more explicit. We have also added the larger N-body evaluations in Appendix E.1 (Tables 4–5) and the multi-channel EGNN comparisons in Appendix E.4 (Table 6).

The corresponding descriptions specify the model parameter counts and evaluation conventions. The task inputs, preprocessing, training configurations, and common evaluation protocols are documented in Appendix E, including Tables 8–9.

Q1 — Parameter counts and scalability on N-body

As requested, we performed additional experiments on 5, 20, 50, and 100 particles, comparing ordinary EGNN with ESNN at closely matched parameter budgets. We retained the evaluation protocol of the original manuscript, extending the datasets to 20, 50, and 100 particles. Models were trained separately for each system size using fixed dataset splits and the same prediction horizon and evaluation protocol within each setting.

The following tables report test MSE (mean ± sample SD).

Five-body charged N-body — Parameter-matched comparison

ModelParametersTest MSE ↓
EGNN222,8930.0072 ± 0.0007
ESNN-Identity220,0120.0075 ± 0.0008
ESNN-Diagonal222,5480.0060 ± 0.0006
ESNN-Orthogonal221,2120.0068 ± 0.0008
ESNN-RadTan222,6500.0061 ± 0.0006

Larger charged N-body systems — Parameter-matched comparison

BodiesEGNN (222,893 params)ESNN-Diagonal (222,548)ESNN-RadTan (222,650)
20 (3 seeds)0.0102 ± 0.00140.0096 ± 0.00050.0084 ± 0.0009
50 (3 seeds)0.0161 ± 0.00350.0104 ± 0.00070.0102 ± 0.0010
100 (3 seeds)0.0265 ± 0.01240.0124 ± 0.00040.0118 ± 0.0009

Radial–Tangential achieves lower mean test error than ordinary EGNN at all four system sizes. At 20, 50, and 100 particles, the reductions are approximately 17%, 37%, and 56%, respectively. Diagonal also improves over ordinary EGNN at all three larger sizes.

These findings extend the original five-body evaluation and show that the advantages over ordinary EGNN persist in larger particle systems. We have included the larger-system results and experimental settings in Appendix E.1 (Tables 4–5).

We continue in the next comment.

Q2 — Multi-channel EGNN baseline

To investigate whether the observed improvements could be obtained simply by introducing additional vector channels, we implemented a GMN-inspired multi-channel EGNN and compared it with ESNN-Radial–Tangential on all three mesh benchmarks.

We retained the evaluation protocol of the original manuscript, using three independent seeds for both MC-EGNN and ESNN-Radial–Tangential.

The models are matched to within 0.32% in trainable parameters. Test rollout RMSE (×10⁻³) is reported as mean ± standard deviation:

TaskParams: MC-EGNN / ESNN1-Step: MC-EGNN / ESNN50-Step: MC-EGNN / ESNNFull: MC-EGNN / ESNN
CylinderFlow1.945M / 1.939M5.98 ± 0.17 / 3.20 ± 0.1227.0 ± 0.7 / 10.4 ± 0.4147.33 ± 10.4 / 48.6 ± 0.8
Airfoil1.506M / 1.505M7033 ± 343 / 1539 ± 2513941 ± 185 / 2396 ± 47Diverged (2/3 runs) / 8424 ± 108
DeformingPlate9.014M / 9.021M0.32 ± 0.04 / 0.09 ± 0.012.3 ± 0.3 / 1.0 ± 0.112.9 ± 0.8 / 5.7 ± 0.3

ESNN-Radial–Tangential achieves lower error across the reported finite-horizon comparisons. In particular, full-rollout RMSE is 67% lower on CylinderFlow and 56% lower on DeformingPlate. On Airfoil, the first-50-step error is reduced by approximately 83%, while the multi-channel baseline diverges in two of three full rollouts. These results provide additional evidence for the usefulness of explicitly structured geometric transport. Across the three mesh benchmarks, the improvements persist against a multi-channel baseline with identical parameter counts, suggesting that the observed gains are not solely attributable to model size or the inclusion of multiple vector channels (Section 7, Q3; Appendix E.4, Table 6).

We hope these clarifications and revisions address the Reviewer’s concerns, and we would be grateful if they would consider reassessing their evaluation and score based on the updated text and the above comments.

Dear Reviewer gFYH,

We would like to kindly follow up on our responses and the revised manuscript. In particular, we have conducted the additional experiments you suggested, including parameter-matched comparisons on larger N-body systems (20, 50, and 100 particles) and a multi-channel EGNN baseline across the three mesh benchmarks.

We would greatly appreciate hearing whether these additional results address your main concerns regarding experimental controls and model capacity, or whether any aspects would benefit from further clarification.

Thank you again for your constructive feedback and consideration.

Best regards, The Authors

Comments on the paper

5 comments
AuthorsOpenReview ↗

General Comment Summary

Dear Reviewers, ACs, and PCs,

We thank the reviewers for their careful assessments and constructive feedback. We appreciate the recognition of ESNN's coherent and mathematically motivated equivariant framework and diverse evaluation [gFYH], the honest scope of our contributions and promising mesh-based results [6kmL], and the clear spatial–channel decomposition and equivariance guarantees [7Gdc].

The reviews highlighted two central questions: whether the observed improvements arise from the proposed geometric transport mechanisms rather than model capacity, and how the different transport families relate to existing architectures and the geometry of the underlying tasks.

To address these points, we conducted additional controlled experiments and clarified the theoretical and practical contributions of ESNN. We summarize the main results and revisions below, with detailed evidence in our individual responses.

Revisions and Updates

Experimental Results and Controls

  • [gFYH; 6kmL] Parameter-matched comparisons and larger N-body systems. We extended the evaluation with approximately 220k-parameter EGNN and ESNN configurations, including the Identity reference, and experiments on 20-, 50-, and 100-body charged systems. At these larger sizes, ESNN-Radial–Tangential achieves 17%, 37%, and 56% lower mean test MSE than EGNN, respectively. These results extend the evidence beyond the original five-body setting and show that the improvements are not simply associated with larger parameter counts.

  • [gFYH; 6kmL; 7Gdc] Stronger multi-channel baseline. We implemented a GMN-inspired multi-channel EGNN and compared it with ESNN-Radial–Tangential across all three MeshGraphNet benchmarks, using the same evaluation protocol and three independent seeds. With trainable parameter counts matched to within 0.32%, Radial–Tangential achieves 67% and 56% lower full-rollout RMSE on CylinderFlow and DeformingPlate, respectively. On Airfoil, the multi-channel baseline diverges in two of three full rollouts. These comparisons provide additional evidence that the benefits of structured transport are not solely attributable to model capacity or the availability of multiple vector channels.

  • [7Gdc] Comparison with SGNN. In our response to Reviewer 7Gdc (Q1), we report a matched-information comparison with SGNN (Han et al., 2022) on gravity-augmented N-body. SGNN achieves lower error when the gravity direction is supplied. On the other hand, in a separate experiment with the direction withheld, ESNN's learned-direction mechanism recovers the gravity axis from data. These results clarify the capabilities and relative strengths of the two approaches.

Theoretical Contribution and Geometric Interpretation

  • [6kmL] Expressivity and relationship to existing architectures. We clarify how ESNN relates to PaiNN, GMN, SGNN, and classical sheaf constructions. In particular, Theorem 4.1 characterizes linear displacement-conditioned equivariant transport, showing how radial–tangential transformations extend isotropic channel mixing. We further illustrate this distinction through a concrete local-operator example.

  • [7Gdc] Geometric inductive biases. We have expanded the discussion of when each transport family may be useful, connecting spatial reorientation to varying flow directions and radial–tangential anisotropy to longitudinal and transverse deformation. We also clarify the bundle/connection interpretation of the learned transports, including the role of non-orthogonal channel mixing and gating.

Experimental Clarity and Reproducibility

  • [gFYH; 6kmL] Evaluation protocols and run statistics. We provide the five individual runs underlying the ESNN results in Table 2 and clarify that the reported uncertainty is the sample standard deviation across independent runs. We also clarify the gravity input-frame convention and the coordinate-averaged MSE definition. The added experiments retain the original evaluation protocols, with the relevant changes in system size and number of seeds explicitly reported.

Manuscript Revisions

We have strengthened the experimental discussion in Section 7 (Q1 and Q3), clarified the transport families and their inductive biases in Section 4, and expanded the related-work discussion. Additional experimental results and implementation details are provided in Appendix E, including the larger N-body evaluations (E.1, Tables 4–5), the clarified gravity protocol (E.2), and the multi-channel EGNN comparisons (E.4, Table 6), with the corresponding training configurations in E.6 (Table 9).

The individual reviewer responses provide the corresponding results, theoretical arguments, and further clarifications.

We continue in the next comment.

Request for Feedback

We sincerely thank the reviewers once again for their thoughtful and constructive feedback, which has helped us substantially strengthen the manuscript. We believe that the additional experiments and clarifications prompted by their suggestions strengthen the empirical evidence and clarify the distinctive contribution of equivariant geometric transport.

We would be grateful if the reviewers could consider the revised manuscript and our individual responses and revisit their assessments in light of the additional evidence and clarifications. We appreciate their time and careful consideration and would welcome any further questions or suggestions.

Best regards,

The Authors

AuthorsOpenReview ↗

General Comment Summary - Updated

General Rebuttal — Updated Summary

Dear Reviewers, ACs, and PCs,

We thank the reviewers for their careful assessments, feedback, and engagement throughout the discussion period. We appreciate the recognition of ESNN's coherent and mathematically motivated equivariant framework and diverse evaluation [gFYH], the honest scope of our contributions and promising mesh-based results [6kmL], and the clear spatial–channel decomposition and equivariance guarantees [7Gdc].

The reviews raised two central questions: whether the observed improvements arise from learned geometric transport rather than differences in model capacity or vector-channel counts, and how ESNN's theoretical and architectural contributions relate to existing equivariant models, particularly PaiNN, GMN, and SGNN.

We addressed these concerns through additional controlled experiments, mathematical clarifications, and manuscript revisions. Following our initial rebuttal, we conducted further experiments and refined the comparisons in response to the reviewers' subsequent questions. We summarize the complete set of results and revisions below.

Revisions and Updates

Experimental Results and Controls

  • [gFYH; 6kmL] Parameter-matched comparisons and larger N-body systems. We extended our evaluation to include EGNN and ESNN configurations with approximately 220k parameters, including Identity Transport, and charged systems with 20, 50, and 100 particles. At these larger sizes, ESNN-Radial–Tangential achieves 17%, 37%, and 56% lower mean test MSE than EGNN, respectively. These results extend the original five-body evaluation and show that the improvements cannot be explained solely by larger parameter budgets (Appendix E.1, Tables 4–5).

  • [6kmL] Identity–Diagonal comparison with matched vector channels. Following the Reviewer's additional question, we clarified that the reported parameter-matched Identity and Diagonal configurations use zero and one vector channel, respectively, while sharing the same training settings and evaluation protocol. Importantly, we had also evaluated Identity with one vector channel, matching Diagonal in this respect. This configuration achieved a test MSE of 0.007547 ± 0.000879, compared with 0.0060 ± 0.0006 for Diagonal. Thus, Diagonal maintains its advantage even when the vector-channel count is matched, providing further evidence that the improvement is not simply attributable to additional vector channels. We retained the original Identity configuration in Table 4 for consistency and reported this control in our follow-up response.

  • [gFYH; 6kmL; 7Gdc] Stronger multi-channel EGNN baseline. We implemented a GMN-inspired multi-channel EGNN (MC-EGNN) and evaluated it against ESNN-Radial–Tangential on all three MeshGraphNet benchmarks using three independent seeds. With parameter counts matched to within 0.32%, ESNN-Radial–Tangential achieves 67% and 56% lower full-rollout RMSE on CylinderFlow and DeformingPlate, respectively. On Airfoil, MC-EGNN diverges in two of three full rollouts. These experiments provide stronger evidence that ESNN's improvements are not merely a consequence of model capacity or the availability of multiple vector channels (Section 7, Q3; Appendix E.4, Table 6).

  • [gFYH] Additional vector-count controls. In response to a further concern, we evaluated MC-EGNN with three vector channels on Airfoil and two on CylinderFlow and DeformingPlate, matching the respective numbers of ESNN-Radial–Tangential vector groups while keeping parameter counts within 0.56% of the corresponding ESNN models. The performance gap persists: ESNN-Radial–Tangential retains lower mean one-step and 50-step errors across all three tasks and lower full-rollout error on DeformingPlate. The new MC-EGNN configurations also exhibit additional nonfinite full rollouts on Airfoil and CylinderFlow. These controls provide further evidence that the original differences are not simply explained by MC-EGNN's choice of vector-channel count. Complete results were provided in our follow-up response to Reviewer gFYH.

  • [7Gdc] Comparison with SGNN. We performed a matched-information comparison with SGNN (Han et al., 2022) on gravity-augmented N-body dynamics. SGNN performs better when the gravity direction is explicitly supplied, while a separate experiment shows that ESNN can recover the preferred direction from data when it is withheld. These experiments clarify the role of available symmetry information and the respective strengths of the two approaches (Reviewer 7Gdc, Q1).

We continue in the next comment

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.