- 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
-
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.
-
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
- 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
- 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
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):
| Model | Parameters | Test MSE ↓ |
|---|---|---|
| SGNN (known axis) | 186,453 | 0.0079 ± 0.0005 |
| ESNN-Orthogonal-Fixed | 191,349 | 0.0105 ± 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 -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 , 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.
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:
Theorem 4.1 characterizes displacement-conditioned linear O(n)-equivariant transport as Tr(V)=P∥VA+P⊥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.